1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 /* 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved. 23 * Use is subject to license terms. 24 */ 25 26 #include <sys/types.h> 27 #include <sys/systm.h> 28 #include <sys/ddi.h> 29 #include <sys/sysmacros.h> 30 #include <sys/archsystm.h> 31 #include <sys/vmsystm.h> 32 #include <sys/machparam.h> 33 #include <sys/machsystm.h> 34 #include <sys/machthread.h> 35 #include <sys/cpu.h> 36 #include <sys/cmp.h> 37 #include <sys/elf_SPARC.h> 38 #include <vm/vm_dep.h> 39 #include <vm/hat_sfmmu.h> 40 #include <vm/seg_kpm.h> 41 #include <sys/cpuvar.h> 42 #include <sys/cheetahregs.h> 43 #include <sys/us3_module.h> 44 #include <sys/async.h> 45 #include <sys/cmn_err.h> 46 #include <sys/debug.h> 47 #include <sys/dditypes.h> 48 #include <sys/prom_debug.h> 49 #include <sys/prom_plat.h> 50 #include <sys/cpu_module.h> 51 #include <sys/sysmacros.h> 52 #include <sys/intreg.h> 53 #include <sys/clock.h> 54 #include <sys/platform_module.h> 55 #include <sys/machtrap.h> 56 #include <sys/ontrap.h> 57 #include <sys/panic.h> 58 #include <sys/memlist.h> 59 #include <sys/bootconf.h> 60 #include <sys/ivintr.h> 61 #include <sys/atomic.h> 62 #include <sys/taskq.h> 63 #include <sys/note.h> 64 #include <sys/ndifm.h> 65 #include <sys/ddifm.h> 66 #include <sys/fm/protocol.h> 67 #include <sys/fm/util.h> 68 #include <sys/fm/cpu/UltraSPARC-III.h> 69 #include <sys/fpras_impl.h> 70 #include <sys/dtrace.h> 71 #include <sys/watchpoint.h> 72 #include <sys/plat_ecc_unum.h> 73 #include <sys/cyclic.h> 74 #include <sys/errorq.h> 75 #include <sys/errclassify.h> 76 #include <sys/pghw.h> 77 #include <sys/clock_impl.h> 78 79 #ifdef CHEETAHPLUS_ERRATUM_25 80 #include <sys/xc_impl.h> 81 #endif /* CHEETAHPLUS_ERRATUM_25 */ 82 83 ch_cpu_logout_t clop_before_flush; 84 ch_cpu_logout_t clop_after_flush; 85 uint_t flush_retries_done = 0; 86 /* 87 * Note that 'Cheetah PRM' refers to: 88 * SPARC V9 JPS1 Implementation Supplement: Sun UltraSPARC-III 89 */ 90 91 /* 92 * Per CPU pointers to physical address of TL>0 logout data areas. 93 * These pointers have to be in the kernel nucleus to avoid MMU 94 * misses. 95 */ 96 uint64_t ch_err_tl1_paddrs[NCPU]; 97 98 /* 99 * One statically allocated structure to use during startup/DR 100 * to prevent unnecessary panics. 101 */ 102 ch_err_tl1_data_t ch_err_tl1_data; 103 104 /* 105 * Per CPU pending error at TL>0, used by level15 softint handler 106 */ 107 uchar_t ch_err_tl1_pending[NCPU]; 108 109 /* 110 * For deferred CE re-enable after trap. 111 */ 112 taskq_t *ch_check_ce_tq; 113 114 /* 115 * Internal functions. 116 */ 117 static int cpu_async_log_err(void *flt, errorq_elem_t *eqep); 118 static void cpu_log_diag_info(ch_async_flt_t *ch_flt); 119 static void cpu_queue_one_event(ch_async_flt_t *ch_flt, char *reason, 120 ecc_type_to_info_t *eccp, ch_diag_data_t *cdp); 121 static int cpu_flt_in_memory_one_event(ch_async_flt_t *ch_flt, 122 uint64_t t_afsr_bit); 123 static int clear_ecc(struct async_flt *ecc); 124 #if defined(CPU_IMP_ECACHE_ASSOC) 125 static int cpu_ecache_line_valid(ch_async_flt_t *ch_flt); 126 #endif 127 int cpu_ecache_set_size(struct cpu *cp); 128 static int cpu_ectag_line_invalid(int cachesize, uint64_t tag); 129 int cpu_ectag_pa_to_subblk(int cachesize, uint64_t subaddr); 130 uint64_t cpu_ectag_to_pa(int setsize, uint64_t tag); 131 int cpu_ectag_pa_to_subblk_state(int cachesize, 132 uint64_t subaddr, uint64_t tag); 133 static void cpu_flush_ecache_line(ch_async_flt_t *ch_flt); 134 static int afsr_to_afar_status(uint64_t afsr, uint64_t afsr_bit); 135 static int afsr_to_esynd_status(uint64_t afsr, uint64_t afsr_bit); 136 static int afsr_to_msynd_status(uint64_t afsr, uint64_t afsr_bit); 137 static int afsr_to_synd_status(uint_t cpuid, uint64_t afsr, uint64_t afsr_bit); 138 static int synd_to_synd_code(int synd_status, ushort_t synd, uint64_t afsr_bit); 139 static int cpu_get_mem_unum_synd(int synd_code, struct async_flt *, char *buf); 140 static void cpu_uninit_ecache_scrub_dr(struct cpu *cp); 141 static void cpu_scrubphys(struct async_flt *aflt); 142 static void cpu_payload_add_aflt(struct async_flt *, nvlist_t *, nvlist_t *, 143 int *, int *); 144 static void cpu_payload_add_ecache(struct async_flt *, nvlist_t *); 145 static void cpu_ereport_init(struct async_flt *aflt); 146 static int cpu_check_secondary_errors(ch_async_flt_t *, uint64_t, uint64_t); 147 static uint8_t cpu_flt_bit_to_plat_error(struct async_flt *aflt); 148 static void cpu_log_fast_ecc_error(caddr_t tpc, int priv, int tl, uint64_t ceen, 149 uint64_t nceen, ch_cpu_logout_t *clop); 150 static int cpu_ce_delayed_ec_logout(uint64_t); 151 static int cpu_matching_ecache_line(uint64_t, void *, int, int *); 152 static int cpu_error_is_ecache_data(int, uint64_t); 153 static void cpu_fmri_cpu_set(nvlist_t *, int); 154 static int cpu_error_to_resource_type(struct async_flt *aflt); 155 156 #ifdef CHEETAHPLUS_ERRATUM_25 157 static int mondo_recover_proc(uint16_t, int); 158 static void cheetah_nudge_init(void); 159 static void cheetah_nudge_onln(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, 160 cyc_time_t *when); 161 static void cheetah_nudge_buddy(void); 162 #endif /* CHEETAHPLUS_ERRATUM_25 */ 163 164 #if defined(CPU_IMP_L1_CACHE_PARITY) 165 static void cpu_dcache_parity_info(ch_async_flt_t *ch_flt); 166 static void cpu_dcache_parity_check(ch_async_flt_t *ch_flt, int index); 167 static void cpu_record_dc_data_parity(ch_async_flt_t *ch_flt, 168 ch_dc_data_t *dest_dcp, ch_dc_data_t *src_dcp, int way, int word); 169 static void cpu_icache_parity_info(ch_async_flt_t *ch_flt); 170 static void cpu_icache_parity_check(ch_async_flt_t *ch_flt, int index); 171 static void cpu_pcache_parity_info(ch_async_flt_t *ch_flt); 172 static void cpu_pcache_parity_check(ch_async_flt_t *ch_flt, int index); 173 static void cpu_payload_add_dcache(struct async_flt *, nvlist_t *); 174 static void cpu_payload_add_icache(struct async_flt *, nvlist_t *); 175 #endif /* CPU_IMP_L1_CACHE_PARITY */ 176 177 int (*p2get_mem_info)(int synd_code, uint64_t paddr, 178 uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep, 179 int *segsp, int *banksp, int *mcidp); 180 181 /* 182 * This table is used to determine which bit(s) is(are) bad when an ECC 183 * error occurs. The array is indexed by an 9-bit syndrome. The entries 184 * of this array have the following semantics: 185 * 186 * 00-127 The number of the bad bit, when only one bit is bad. 187 * 128 ECC bit C0 is bad. 188 * 129 ECC bit C1 is bad. 189 * 130 ECC bit C2 is bad. 190 * 131 ECC bit C3 is bad. 191 * 132 ECC bit C4 is bad. 192 * 133 ECC bit C5 is bad. 193 * 134 ECC bit C6 is bad. 194 * 135 ECC bit C7 is bad. 195 * 136 ECC bit C8 is bad. 196 * 137-143 reserved for Mtag Data and ECC. 197 * 144(M2) Two bits are bad within a nibble. 198 * 145(M3) Three bits are bad within a nibble. 199 * 146(M3) Four bits are bad within a nibble. 200 * 147(M) Multiple bits (5 or more) are bad. 201 * 148 NO bits are bad. 202 * Based on "Cheetah Programmer's Reference Manual" rev 1.1, Tables 11-4,11-5. 203 */ 204 205 #define C0 128 206 #define C1 129 207 #define C2 130 208 #define C3 131 209 #define C4 132 210 #define C5 133 211 #define C6 134 212 #define C7 135 213 #define C8 136 214 #define MT0 137 /* Mtag Data bit 0 */ 215 #define MT1 138 216 #define MT2 139 217 #define MTC0 140 /* Mtag Check bit 0 */ 218 #define MTC1 141 219 #define MTC2 142 220 #define MTC3 143 221 #define M2 144 222 #define M3 145 223 #define M4 146 224 #define M 147 225 #define NA 148 226 #if defined(JALAPENO) || defined(SERRANO) 227 #define S003 149 /* Syndrome 0x003 => likely from CPU/EDU:ST/FRU/BP */ 228 #define S003MEM 150 /* Syndrome 0x003 => likely from WDU/WBP */ 229 #define SLAST S003MEM /* last special syndrome */ 230 #else /* JALAPENO || SERRANO */ 231 #define S003 149 /* Syndrome 0x003 => likely from EDU:ST */ 232 #define S071 150 /* Syndrome 0x071 => likely from WDU/CPU */ 233 #define S11C 151 /* Syndrome 0x11c => likely from BERR/DBERR */ 234 #define SLAST S11C /* last special syndrome */ 235 #endif /* JALAPENO || SERRANO */ 236 #if defined(JALAPENO) || defined(SERRANO) 237 #define BPAR0 152 /* syndrom 152 through 167 for bus parity */ 238 #define BPAR15 167 239 #endif /* JALAPENO || SERRANO */ 240 241 static uint8_t ecc_syndrome_tab[] = 242 { 243 NA, C0, C1, S003, C2, M2, M3, 47, C3, M2, M2, 53, M2, 41, 29, M, 244 C4, M, M, 50, M2, 38, 25, M2, M2, 33, 24, M2, 11, M, M2, 16, 245 C5, M, M, 46, M2, 37, 19, M2, M, 31, 32, M, 7, M2, M2, 10, 246 M2, 40, 13, M2, 59, M, M2, 66, M, M2, M2, 0, M2, 67, 71, M, 247 C6, M, M, 43, M, 36, 18, M, M2, 49, 15, M, 63, M2, M2, 6, 248 M2, 44, 28, M2, M, M2, M2, 52, 68, M2, M2, 62, M2, M3, M3, M4, 249 M2, 26, 106, M2, 64, M, M2, 2, 120, M, M2, M3, M, M3, M3, M4, 250 #if defined(JALAPENO) || defined(SERRANO) 251 116, M2, M2, M3, M2, M3, M, M4, M2, 58, 54, M2, M, M4, M4, M3, 252 #else /* JALAPENO || SERRANO */ 253 116, S071, M2, M3, M2, M3, M, M4, M2, 58, 54, M2, M, M4, M4, M3, 254 #endif /* JALAPENO || SERRANO */ 255 C7, M2, M, 42, M, 35, 17, M2, M, 45, 14, M2, 21, M2, M2, 5, 256 M, 27, M, M, 99, M, M, 3, 114, M2, M2, 20, M2, M3, M3, M, 257 M2, 23, 113, M2, 112, M2, M, 51, 95, M, M2, M3, M2, M3, M3, M2, 258 103, M, M2, M3, M2, M3, M3, M4, M2, 48, M, M, 73, M2, M, M3, 259 M2, 22, 110, M2, 109, M2, M, 9, 108, M2, M, M3, M2, M3, M3, M, 260 102, M2, M, M, M2, M3, M3, M, M2, M3, M3, M2, M, M4, M, M3, 261 98, M, M2, M3, M2, M, M3, M4, M2, M3, M3, M4, M3, M, M, M, 262 M2, M3, M3, M, M3, M, M, M, 56, M4, M, M3, M4, M, M, M, 263 C8, M, M2, 39, M, 34, 105, M2, M, 30, 104, M, 101, M, M, 4, 264 #if defined(JALAPENO) || defined(SERRANO) 265 M, M, 100, M, 83, M, M2, 12, 87, M, M, 57, M2, M, M3, M, 266 #else /* JALAPENO || SERRANO */ 267 M, M, 100, M, 83, M, M2, 12, 87, M, M, 57, S11C, M, M3, M, 268 #endif /* JALAPENO || SERRANO */ 269 M2, 97, 82, M2, 78, M2, M2, 1, 96, M, M, M, M, M, M3, M2, 270 94, M, M2, M3, M2, M, M3, M, M2, M, 79, M, 69, M, M4, M, 271 M2, 93, 92, M, 91, M, M2, 8, 90, M2, M2, M, M, M, M, M4, 272 89, M, M, M3, M2, M3, M3, M, M, M, M3, M2, M3, M2, M, M3, 273 86, M, M2, M3, M2, M, M3, M, M2, M, M3, M, M3, M, M, M3, 274 M, M, M3, M2, M3, M2, M4, M, 60, M, M2, M3, M4, M, M, M2, 275 M2, 88, 85, M2, 84, M, M2, 55, 81, M2, M2, M3, M2, M3, M3, M4, 276 77, M, M, M, M2, M3, M, M, M2, M3, M3, M4, M3, M2, M, M, 277 74, M, M2, M3, M, M, M3, M, M, M, M3, M, M3, M, M4, M3, 278 M2, 70, 107, M4, 65, M2, M2, M, 127, M, M, M, M2, M3, M3, M, 279 80, M2, M2, 72, M, 119, 118, M, M2, 126, 76, M, 125, M, M4, M3, 280 M2, 115, 124, M, 75, M, M, M3, 61, M, M4, M, M4, M, M, M, 281 M, 123, 122, M4, 121, M4, M, M3, 117, M2, M2, M3, M4, M3, M, M, 282 111, M, M, M, M4, M3, M3, M, M, M, M3, M, M3, M2, M, M 283 }; 284 285 #define ESYND_TBL_SIZE (sizeof (ecc_syndrome_tab) / sizeof (uint8_t)) 286 287 #if !(defined(JALAPENO) || defined(SERRANO)) 288 /* 289 * This table is used to determine which bit(s) is(are) bad when a Mtag 290 * error occurs. The array is indexed by an 4-bit ECC syndrome. The entries 291 * of this array have the following semantics: 292 * 293 * -1 Invalid mtag syndrome. 294 * 137 Mtag Data 0 is bad. 295 * 138 Mtag Data 1 is bad. 296 * 139 Mtag Data 2 is bad. 297 * 140 Mtag ECC 0 is bad. 298 * 141 Mtag ECC 1 is bad. 299 * 142 Mtag ECC 2 is bad. 300 * 143 Mtag ECC 3 is bad. 301 * Based on "Cheetah Programmer's Reference Manual" rev 1.1, Tables 11-6. 302 */ 303 short mtag_syndrome_tab[] = 304 { 305 NA, MTC0, MTC1, M2, MTC2, M2, M2, MT0, MTC3, M2, M2, MT1, M2, MT2, M2, M2 306 }; 307 308 #define MSYND_TBL_SIZE (sizeof (mtag_syndrome_tab) / sizeof (short)) 309 310 #else /* !(JALAPENO || SERRANO) */ 311 312 #define BSYND_TBL_SIZE 16 313 314 #endif /* !(JALAPENO || SERRANO) */ 315 316 /* 317 * Types returned from cpu_error_to_resource_type() 318 */ 319 #define ERRTYPE_UNKNOWN 0 320 #define ERRTYPE_CPU 1 321 #define ERRTYPE_MEMORY 2 322 #define ERRTYPE_ECACHE_DATA 3 323 324 /* 325 * CE initial classification and subsequent action lookup table 326 */ 327 static ce_dispact_t ce_disp_table[CE_INITDISPTBL_SIZE]; 328 static int ce_disp_inited; 329 330 /* 331 * Set to disable leaky and partner check for memory correctables 332 */ 333 int ce_xdiag_off; 334 335 /* 336 * The following are not incremented atomically so are indicative only 337 */ 338 static int ce_xdiag_drops; 339 static int ce_xdiag_lkydrops; 340 static int ce_xdiag_ptnrdrops; 341 static int ce_xdiag_bad; 342 343 /* 344 * CE leaky check callback structure 345 */ 346 typedef struct { 347 struct async_flt *lkycb_aflt; 348 errorq_t *lkycb_eqp; 349 errorq_elem_t *lkycb_eqep; 350 } ce_lkychk_cb_t; 351 352 /* 353 * defines for various ecache_flush_flag's 354 */ 355 #define ECACHE_FLUSH_LINE 1 356 #define ECACHE_FLUSH_ALL 2 357 358 /* 359 * STICK sync 360 */ 361 #define STICK_ITERATION 10 362 #define MAX_TSKEW 1 363 #define EV_A_START 0 364 #define EV_A_END 1 365 #define EV_B_START 2 366 #define EV_B_END 3 367 #define EVENTS 4 368 369 static int64_t stick_iter = STICK_ITERATION; 370 static int64_t stick_tsk = MAX_TSKEW; 371 372 typedef enum { 373 EVENT_NULL = 0, 374 SLAVE_START, 375 SLAVE_CONT, 376 MASTER_START 377 } event_cmd_t; 378 379 static volatile event_cmd_t stick_sync_cmd = EVENT_NULL; 380 static int64_t timestamp[EVENTS]; 381 static volatile int slave_done; 382 383 #ifdef DEBUG 384 #define DSYNC_ATTEMPTS 64 385 typedef struct { 386 int64_t skew_val[DSYNC_ATTEMPTS]; 387 } ss_t; 388 389 ss_t stick_sync_stats[NCPU]; 390 #endif /* DEBUG */ 391 392 uint_t cpu_impl_dual_pgsz = 0; 393 #if defined(CPU_IMP_DUAL_PAGESIZE) 394 uint_t disable_dual_pgsz = 0; 395 #endif /* CPU_IMP_DUAL_PAGESIZE */ 396 397 /* 398 * Save the cache bootup state for use when internal 399 * caches are to be re-enabled after an error occurs. 400 */ 401 uint64_t cache_boot_state; 402 403 /* 404 * PA[22:0] represent Displacement in Safari configuration space. 405 */ 406 uint_t root_phys_addr_lo_mask = 0x7fffffu; 407 408 bus_config_eclk_t bus_config_eclk[] = { 409 #if defined(JALAPENO) || defined(SERRANO) 410 {JBUS_CONFIG_ECLK_1_DIV, JBUS_CONFIG_ECLK_1}, 411 {JBUS_CONFIG_ECLK_2_DIV, JBUS_CONFIG_ECLK_2}, 412 {JBUS_CONFIG_ECLK_32_DIV, JBUS_CONFIG_ECLK_32}, 413 #else /* JALAPENO || SERRANO */ 414 {SAFARI_CONFIG_ECLK_1_DIV, SAFARI_CONFIG_ECLK_1}, 415 {SAFARI_CONFIG_ECLK_2_DIV, SAFARI_CONFIG_ECLK_2}, 416 {SAFARI_CONFIG_ECLK_32_DIV, SAFARI_CONFIG_ECLK_32}, 417 #endif /* JALAPENO || SERRANO */ 418 {0, 0} 419 }; 420 421 /* 422 * Interval for deferred CEEN reenable 423 */ 424 int cpu_ceen_delay_secs = CPU_CEEN_DELAY_SECS; 425 426 /* 427 * set in /etc/system to control logging of user BERR/TO's 428 */ 429 int cpu_berr_to_verbose = 0; 430 431 /* 432 * set to 0 in /etc/system to defer CEEN reenable for all CEs 433 */ 434 uint64_t cpu_ce_not_deferred = CPU_CE_NOT_DEFERRED; 435 uint64_t cpu_ce_not_deferred_ext = CPU_CE_NOT_DEFERRED_EXT; 436 437 /* 438 * Set of all offline cpus 439 */ 440 cpuset_t cpu_offline_set; 441 442 static void cpu_delayed_check_ce_errors(void *); 443 static void cpu_check_ce_errors(void *); 444 void cpu_error_ecache_flush(ch_async_flt_t *); 445 static int cpu_error_ecache_flush_required(ch_async_flt_t *); 446 static void cpu_log_and_clear_ce(ch_async_flt_t *); 447 void cpu_ce_detected(ch_cpu_errors_t *, int); 448 449 /* 450 * CE Leaky check timeout in microseconds. This is chosen to be twice the 451 * memory refresh interval of current DIMMs (64ms). After initial fix that 452 * gives at least one full refresh cycle in which the cell can leak 453 * (whereafter further refreshes simply reinforce any incorrect bit value). 454 */ 455 clock_t cpu_ce_lkychk_timeout_usec = 128000; 456 457 /* 458 * CE partner check partner caching period in seconds 459 */ 460 int cpu_ce_ptnr_cachetime_sec = 60; 461 462 /* 463 * Sets trap table entry ttentry by overwriting eight instructions from ttlabel 464 */ 465 #define CH_SET_TRAP(ttentry, ttlabel) \ 466 bcopy((const void *)&ttlabel, &ttentry, 32); \ 467 flush_instr_mem((caddr_t)&ttentry, 32); 468 469 static int min_ecache_size; 470 static uint_t priv_hcl_1; 471 static uint_t priv_hcl_2; 472 static uint_t priv_hcl_4; 473 static uint_t priv_hcl_8; 474 475 void 476 cpu_setup(void) 477 { 478 extern int at_flags; 479 extern int cpc_has_overflow_intr; 480 481 /* 482 * Setup chip-specific trap handlers. 483 */ 484 cpu_init_trap(); 485 486 cache |= (CACHE_VAC | CACHE_PTAG | CACHE_IOCOHERENT); 487 488 at_flags = EF_SPARC_32PLUS | EF_SPARC_SUN_US1 | EF_SPARC_SUN_US3; 489 490 /* 491 * save the cache bootup state. 492 */ 493 cache_boot_state = get_dcu() & DCU_CACHE; 494 495 /* 496 * Due to the number of entries in the fully-associative tlb 497 * this may have to be tuned lower than in spitfire. 498 */ 499 pp_slots = MIN(8, MAXPP_SLOTS); 500 501 /* 502 * Block stores do not invalidate all pages of the d$, pagecopy 503 * et. al. need virtual translations with virtual coloring taken 504 * into consideration. prefetch/ldd will pollute the d$ on the 505 * load side. 506 */ 507 pp_consistent_coloring = PPAGE_STORE_VCOLORING | PPAGE_LOADS_POLLUTE; 508 509 if (use_page_coloring) { 510 do_pg_coloring = 1; 511 } 512 513 isa_list = 514 "sparcv9+vis2 sparcv9+vis sparcv9 " 515 "sparcv8plus+vis2 sparcv8plus+vis sparcv8plus " 516 "sparcv8 sparcv8-fsmuld sparcv7 sparc"; 517 518 /* 519 * On Panther-based machines, this should 520 * also include AV_SPARC_POPC too 521 */ 522 cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2; 523 524 /* 525 * On cheetah, there's no hole in the virtual address space 526 */ 527 hole_start = hole_end = 0; 528 529 /* 530 * The kpm mapping window. 531 * kpm_size: 532 * The size of a single kpm range. 533 * The overall size will be: kpm_size * vac_colors. 534 * kpm_vbase: 535 * The virtual start address of the kpm range within the kernel 536 * virtual address space. kpm_vbase has to be kpm_size aligned. 537 */ 538 kpm_size = (size_t)(8ull * 1024 * 1024 * 1024 * 1024); /* 8TB */ 539 kpm_size_shift = 43; 540 kpm_vbase = (caddr_t)0x8000000000000000ull; /* 8EB */ 541 kpm_smallpages = 1; 542 543 /* 544 * The traptrace code uses either %tick or %stick for 545 * timestamping. We have %stick so we can use it. 546 */ 547 traptrace_use_stick = 1; 548 549 /* 550 * Cheetah has a performance counter overflow interrupt 551 */ 552 cpc_has_overflow_intr = 1; 553 554 #if defined(CPU_IMP_DUAL_PAGESIZE) 555 /* 556 * Use Cheetah+ and later dual page size support. 557 */ 558 if (!disable_dual_pgsz) { 559 cpu_impl_dual_pgsz = 1; 560 } 561 #endif /* CPU_IMP_DUAL_PAGESIZE */ 562 563 /* 564 * Declare that this architecture/cpu combination does fpRAS. 565 */ 566 fpras_implemented = 1; 567 568 /* 569 * Setup CE lookup table 570 */ 571 CE_INITDISPTBL_POPULATE(ce_disp_table); 572 ce_disp_inited = 1; 573 } 574 575 /* 576 * Called by setcpudelay 577 */ 578 void 579 cpu_init_tick_freq(void) 580 { 581 /* 582 * For UltraSPARC III and beyond we want to use the 583 * system clock rate as the basis for low level timing, 584 * due to support of mixed speed CPUs and power managment. 585 */ 586 if (system_clock_freq == 0) 587 cmn_err(CE_PANIC, "setcpudelay: invalid system_clock_freq"); 588 589 sys_tick_freq = system_clock_freq; 590 } 591 592 #ifdef CHEETAHPLUS_ERRATUM_25 593 /* 594 * Tunables 595 */ 596 int cheetah_bpe_off = 0; 597 int cheetah_sendmondo_recover = 1; 598 int cheetah_sendmondo_fullscan = 0; 599 int cheetah_sendmondo_recover_delay = 5; 600 601 #define CHEETAH_LIVELOCK_MIN_DELAY 1 602 603 /* 604 * Recovery Statistics 605 */ 606 typedef struct cheetah_livelock_entry { 607 int cpuid; /* fallen cpu */ 608 int buddy; /* cpu that ran recovery */ 609 clock_t lbolt; /* when recovery started */ 610 hrtime_t recovery_time; /* time spent in recovery */ 611 } cheetah_livelock_entry_t; 612 613 #define CHEETAH_LIVELOCK_NENTRY 32 614 615 cheetah_livelock_entry_t cheetah_livelock_hist[CHEETAH_LIVELOCK_NENTRY]; 616 int cheetah_livelock_entry_nxt; 617 618 #define CHEETAH_LIVELOCK_ENTRY_NEXT(statp) { \ 619 statp = cheetah_livelock_hist + cheetah_livelock_entry_nxt; \ 620 if (++cheetah_livelock_entry_nxt >= CHEETAH_LIVELOCK_NENTRY) { \ 621 cheetah_livelock_entry_nxt = 0; \ 622 } \ 623 } 624 625 #define CHEETAH_LIVELOCK_ENTRY_SET(statp, item, val) statp->item = val 626 627 struct { 628 hrtime_t hrt; /* maximum recovery time */ 629 int recovery; /* recovered */ 630 int full_claimed; /* maximum pages claimed in full recovery */ 631 int proc_entry; /* attempted to claim TSB */ 632 int proc_tsb_scan; /* tsb scanned */ 633 int proc_tsb_partscan; /* tsb partially scanned */ 634 int proc_tsb_fullscan; /* whole tsb scanned */ 635 int proc_claimed; /* maximum pages claimed in tsb scan */ 636 int proc_user; /* user thread */ 637 int proc_kernel; /* kernel thread */ 638 int proc_onflt; /* bad stack */ 639 int proc_cpu; /* null cpu */ 640 int proc_thread; /* null thread */ 641 int proc_proc; /* null proc */ 642 int proc_as; /* null as */ 643 int proc_hat; /* null hat */ 644 int proc_hat_inval; /* hat contents don't make sense */ 645 int proc_hat_busy; /* hat is changing TSBs */ 646 int proc_tsb_reloc; /* TSB skipped because being relocated */ 647 int proc_cnum_bad; /* cnum out of range */ 648 int proc_cnum; /* last cnum processed */ 649 tte_t proc_tte; /* last tte processed */ 650 } cheetah_livelock_stat; 651 652 #define CHEETAH_LIVELOCK_STAT(item) cheetah_livelock_stat.item++ 653 654 #define CHEETAH_LIVELOCK_STATSET(item, value) \ 655 cheetah_livelock_stat.item = value 656 657 #define CHEETAH_LIVELOCK_MAXSTAT(item, value) { \ 658 if (value > cheetah_livelock_stat.item) \ 659 cheetah_livelock_stat.item = value; \ 660 } 661 662 /* 663 * Attempt to recover a cpu by claiming every cache line as saved 664 * in the TSB that the non-responsive cpu is using. Since we can't 665 * grab any adaptive lock, this is at best an attempt to do so. Because 666 * we don't grab any locks, we must operate under the protection of 667 * on_fault(). 668 * 669 * Return 1 if cpuid could be recovered, 0 if failed. 670 */ 671 int 672 mondo_recover_proc(uint16_t cpuid, int bn) 673 { 674 label_t ljb; 675 cpu_t *cp; 676 kthread_t *t; 677 proc_t *p; 678 struct as *as; 679 struct hat *hat; 680 uint_t cnum; 681 struct tsb_info *tsbinfop; 682 struct tsbe *tsbep; 683 caddr_t tsbp; 684 caddr_t end_tsbp; 685 uint64_t paddr; 686 uint64_t idsr; 687 u_longlong_t pahi, palo; 688 int pages_claimed = 0; 689 tte_t tsbe_tte; 690 int tried_kernel_tsb = 0; 691 mmu_ctx_t *mmu_ctxp; 692 693 CHEETAH_LIVELOCK_STAT(proc_entry); 694 695 if (on_fault(&ljb)) { 696 CHEETAH_LIVELOCK_STAT(proc_onflt); 697 goto badstruct; 698 } 699 700 if ((cp = cpu[cpuid]) == NULL) { 701 CHEETAH_LIVELOCK_STAT(proc_cpu); 702 goto badstruct; 703 } 704 705 if ((t = cp->cpu_thread) == NULL) { 706 CHEETAH_LIVELOCK_STAT(proc_thread); 707 goto badstruct; 708 } 709 710 if ((p = ttoproc(t)) == NULL) { 711 CHEETAH_LIVELOCK_STAT(proc_proc); 712 goto badstruct; 713 } 714 715 if ((as = p->p_as) == NULL) { 716 CHEETAH_LIVELOCK_STAT(proc_as); 717 goto badstruct; 718 } 719 720 if ((hat = as->a_hat) == NULL) { 721 CHEETAH_LIVELOCK_STAT(proc_hat); 722 goto badstruct; 723 } 724 725 if (hat != ksfmmup) { 726 CHEETAH_LIVELOCK_STAT(proc_user); 727 if (hat->sfmmu_flags & (HAT_BUSY | HAT_SWAPPED | HAT_SWAPIN)) { 728 CHEETAH_LIVELOCK_STAT(proc_hat_busy); 729 goto badstruct; 730 } 731 tsbinfop = hat->sfmmu_tsb; 732 if (tsbinfop == NULL) { 733 CHEETAH_LIVELOCK_STAT(proc_hat_inval); 734 goto badstruct; 735 } 736 tsbp = tsbinfop->tsb_va; 737 end_tsbp = tsbp + TSB_BYTES(tsbinfop->tsb_szc); 738 } else { 739 CHEETAH_LIVELOCK_STAT(proc_kernel); 740 tsbinfop = NULL; 741 tsbp = ktsb_base; 742 end_tsbp = tsbp + TSB_BYTES(ktsb_sz); 743 } 744 745 /* Verify as */ 746 if (hat->sfmmu_as != as) { 747 CHEETAH_LIVELOCK_STAT(proc_hat_inval); 748 goto badstruct; 749 } 750 751 mmu_ctxp = CPU_MMU_CTXP(cp); 752 ASSERT(mmu_ctxp); 753 cnum = hat->sfmmu_ctxs[mmu_ctxp->mmu_idx].cnum; 754 CHEETAH_LIVELOCK_STATSET(proc_cnum, cnum); 755 756 if ((cnum < 0) || (cnum == INVALID_CONTEXT) || 757 (cnum >= mmu_ctxp->mmu_nctxs)) { 758 CHEETAH_LIVELOCK_STAT(proc_cnum_bad); 759 goto badstruct; 760 } 761 762 do { 763 CHEETAH_LIVELOCK_STAT(proc_tsb_scan); 764 765 /* 766 * Skip TSBs being relocated. This is important because 767 * we want to avoid the following deadlock scenario: 768 * 769 * 1) when we came in we set ourselves to "in recover" state. 770 * 2) when we try to touch TSB being relocated the mapping 771 * will be in the suspended state so we'll spin waiting 772 * for it to be unlocked. 773 * 3) when the CPU that holds the TSB mapping locked tries to 774 * unlock it it will send a xtrap which will fail to xcall 775 * us or the CPU we're trying to recover, and will in turn 776 * enter the mondo code. 777 * 4) since we are still spinning on the locked mapping 778 * no further progress will be made and the system will 779 * inevitably hard hang. 780 * 781 * A TSB not being relocated can't begin being relocated 782 * while we're accessing it because we check 783 * sendmondo_in_recover before relocating TSBs. 784 */ 785 if (hat != ksfmmup && 786 (tsbinfop->tsb_flags & TSB_RELOC_FLAG) != 0) { 787 CHEETAH_LIVELOCK_STAT(proc_tsb_reloc); 788 goto next_tsbinfo; 789 } 790 791 for (tsbep = (struct tsbe *)tsbp; 792 tsbep < (struct tsbe *)end_tsbp; tsbep++) { 793 tsbe_tte = tsbep->tte_data; 794 795 if (tsbe_tte.tte_val == 0) { 796 /* 797 * Invalid tte 798 */ 799 continue; 800 } 801 if (tsbe_tte.tte_se) { 802 /* 803 * Don't want device registers 804 */ 805 continue; 806 } 807 if (tsbe_tte.tte_cp == 0) { 808 /* 809 * Must be cached in E$ 810 */ 811 continue; 812 } 813 if (tsbep->tte_tag.tag_invalid != 0) { 814 /* 815 * Invalid tag, ingnore this entry. 816 */ 817 continue; 818 } 819 CHEETAH_LIVELOCK_STATSET(proc_tte, tsbe_tte); 820 idsr = getidsr(); 821 if ((idsr & (IDSR_NACK_BIT(bn) | 822 IDSR_BUSY_BIT(bn))) == 0) { 823 CHEETAH_LIVELOCK_STAT(proc_tsb_partscan); 824 goto done; 825 } 826 pahi = tsbe_tte.tte_pahi; 827 palo = tsbe_tte.tte_palo; 828 paddr = (uint64_t)((pahi << 32) | 829 (palo << MMU_PAGESHIFT)); 830 claimlines(paddr, TTEBYTES(TTE_CSZ(&tsbe_tte)), 831 CH_ECACHE_SUBBLK_SIZE); 832 if ((idsr & IDSR_BUSY_BIT(bn)) == 0) { 833 shipit(cpuid, bn); 834 } 835 pages_claimed++; 836 } 837 next_tsbinfo: 838 if (tsbinfop != NULL) 839 tsbinfop = tsbinfop->tsb_next; 840 if (tsbinfop != NULL) { 841 tsbp = tsbinfop->tsb_va; 842 end_tsbp = tsbp + TSB_BYTES(tsbinfop->tsb_szc); 843 } else if (tsbp == ktsb_base) { 844 tried_kernel_tsb = 1; 845 } else if (!tried_kernel_tsb) { 846 tsbp = ktsb_base; 847 end_tsbp = tsbp + TSB_BYTES(ktsb_sz); 848 hat = ksfmmup; 849 tsbinfop = NULL; 850 } 851 } while (tsbinfop != NULL || 852 ((tsbp == ktsb_base) && !tried_kernel_tsb)); 853 854 CHEETAH_LIVELOCK_STAT(proc_tsb_fullscan); 855 CHEETAH_LIVELOCK_MAXSTAT(proc_claimed, pages_claimed); 856 no_fault(); 857 idsr = getidsr(); 858 if ((idsr & (IDSR_NACK_BIT(bn) | 859 IDSR_BUSY_BIT(bn))) == 0) { 860 return (1); 861 } else { 862 return (0); 863 } 864 865 done: 866 no_fault(); 867 CHEETAH_LIVELOCK_MAXSTAT(proc_claimed, pages_claimed); 868 return (1); 869 870 badstruct: 871 no_fault(); 872 return (0); 873 } 874 875 /* 876 * Attempt to claim ownership, temporarily, of every cache line that a 877 * non-responsive cpu might be using. This might kick that cpu out of 878 * this state. 879 * 880 * The return value indicates to the caller if we have exhausted all recovery 881 * techniques. If 1 is returned, it is useless to call this function again 882 * even for a different target CPU. 883 */ 884 int 885 mondo_recover(uint16_t cpuid, int bn) 886 { 887 struct memseg *seg; 888 uint64_t begin_pa, end_pa, cur_pa; 889 hrtime_t begin_hrt, end_hrt; 890 int retval = 0; 891 int pages_claimed = 0; 892 cheetah_livelock_entry_t *histp; 893 uint64_t idsr; 894 895 if (cas32(&sendmondo_in_recover, 0, 1) != 0) { 896 /* 897 * Wait while recovery takes place 898 */ 899 while (sendmondo_in_recover) { 900 drv_usecwait(1); 901 } 902 /* 903 * Assume we didn't claim the whole memory. If 904 * the target of this caller is not recovered, 905 * it will come back. 906 */ 907 return (retval); 908 } 909 910 CHEETAH_LIVELOCK_ENTRY_NEXT(histp); 911 CHEETAH_LIVELOCK_ENTRY_SET(histp, lbolt, LBOLT_WAITFREE); 912 CHEETAH_LIVELOCK_ENTRY_SET(histp, cpuid, cpuid); 913 CHEETAH_LIVELOCK_ENTRY_SET(histp, buddy, CPU->cpu_id); 914 915 begin_hrt = gethrtime_waitfree(); 916 /* 917 * First try to claim the lines in the TSB the target 918 * may have been using. 919 */ 920 if (mondo_recover_proc(cpuid, bn) == 1) { 921 /* 922 * Didn't claim the whole memory 923 */ 924 goto done; 925 } 926 927 /* 928 * We tried using the TSB. The target is still 929 * not recovered. Check if complete memory scan is 930 * enabled. 931 */ 932 if (cheetah_sendmondo_fullscan == 0) { 933 /* 934 * Full memory scan is disabled. 935 */ 936 retval = 1; 937 goto done; 938 } 939 940 /* 941 * Try claiming the whole memory. 942 */ 943 for (seg = memsegs; seg; seg = seg->next) { 944 begin_pa = (uint64_t)(seg->pages_base) << MMU_PAGESHIFT; 945 end_pa = (uint64_t)(seg->pages_end) << MMU_PAGESHIFT; 946 for (cur_pa = begin_pa; cur_pa < end_pa; 947 cur_pa += MMU_PAGESIZE) { 948 idsr = getidsr(); 949 if ((idsr & (IDSR_NACK_BIT(bn) | 950 IDSR_BUSY_BIT(bn))) == 0) { 951 /* 952 * Didn't claim all memory 953 */ 954 goto done; 955 } 956 claimlines(cur_pa, MMU_PAGESIZE, 957 CH_ECACHE_SUBBLK_SIZE); 958 if ((idsr & IDSR_BUSY_BIT(bn)) == 0) { 959 shipit(cpuid, bn); 960 } 961 pages_claimed++; 962 } 963 } 964 965 /* 966 * We did all we could. 967 */ 968 retval = 1; 969 970 done: 971 /* 972 * Update statistics 973 */ 974 end_hrt = gethrtime_waitfree(); 975 CHEETAH_LIVELOCK_STAT(recovery); 976 CHEETAH_LIVELOCK_MAXSTAT(hrt, (end_hrt - begin_hrt)); 977 CHEETAH_LIVELOCK_MAXSTAT(full_claimed, pages_claimed); 978 CHEETAH_LIVELOCK_ENTRY_SET(histp, recovery_time, \ 979 (end_hrt - begin_hrt)); 980 981 while (cas32(&sendmondo_in_recover, 1, 0) != 1) 982 ; 983 984 return (retval); 985 } 986 987 /* 988 * This is called by the cyclic framework when this CPU becomes online 989 */ 990 /*ARGSUSED*/ 991 static void 992 cheetah_nudge_onln(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when) 993 { 994 995 hdlr->cyh_func = (cyc_func_t)cheetah_nudge_buddy; 996 hdlr->cyh_level = CY_LOW_LEVEL; 997 hdlr->cyh_arg = NULL; 998 999 /* 1000 * Stagger the start time 1001 */ 1002 when->cyt_when = cpu->cpu_id * (NANOSEC / NCPU); 1003 if (cheetah_sendmondo_recover_delay < CHEETAH_LIVELOCK_MIN_DELAY) { 1004 cheetah_sendmondo_recover_delay = CHEETAH_LIVELOCK_MIN_DELAY; 1005 } 1006 when->cyt_interval = cheetah_sendmondo_recover_delay * NANOSEC; 1007 } 1008 1009 /* 1010 * Create a low level cyclic to send a xtrap to the next cpu online. 1011 * However, there's no need to have this running on a uniprocessor system. 1012 */ 1013 static void 1014 cheetah_nudge_init(void) 1015 { 1016 cyc_omni_handler_t hdlr; 1017 1018 if (max_ncpus == 1) { 1019 return; 1020 } 1021 1022 hdlr.cyo_online = cheetah_nudge_onln; 1023 hdlr.cyo_offline = NULL; 1024 hdlr.cyo_arg = NULL; 1025 1026 mutex_enter(&cpu_lock); 1027 (void) cyclic_add_omni(&hdlr); 1028 mutex_exit(&cpu_lock); 1029 } 1030 1031 /* 1032 * Cyclic handler to wake up buddy 1033 */ 1034 void 1035 cheetah_nudge_buddy(void) 1036 { 1037 /* 1038 * Disable kernel preemption to protect the cpu list 1039 */ 1040 kpreempt_disable(); 1041 if ((CPU->cpu_next_onln != CPU) && (sendmondo_in_recover == 0)) { 1042 xt_one(CPU->cpu_next_onln->cpu_id, (xcfunc_t *)xt_sync_tl1, 1043 0, 0); 1044 } 1045 kpreempt_enable(); 1046 } 1047 1048 #endif /* CHEETAHPLUS_ERRATUM_25 */ 1049 1050 #ifdef SEND_MONDO_STATS 1051 uint32_t x_one_stimes[64]; 1052 uint32_t x_one_ltimes[16]; 1053 uint32_t x_set_stimes[64]; 1054 uint32_t x_set_ltimes[16]; 1055 uint32_t x_set_cpus[NCPU]; 1056 uint32_t x_nack_stimes[64]; 1057 #endif 1058 1059 /* 1060 * Note: A version of this function is used by the debugger via the KDI, 1061 * and must be kept in sync with this version. Any changes made to this 1062 * function to support new chips or to accomodate errata must also be included 1063 * in the KDI-specific version. See us3_kdi.c. 1064 */ 1065 void 1066 send_one_mondo(int cpuid) 1067 { 1068 int busy, nack; 1069 uint64_t idsr, starttick, endtick, tick, lasttick; 1070 uint64_t busymask; 1071 #ifdef CHEETAHPLUS_ERRATUM_25 1072 int recovered = 0; 1073 #endif 1074 1075 CPU_STATS_ADDQ(CPU, sys, xcalls, 1); 1076 starttick = lasttick = gettick(); 1077 shipit(cpuid, 0); 1078 endtick = starttick + xc_tick_limit; 1079 busy = nack = 0; 1080 #if defined(JALAPENO) || defined(SERRANO) 1081 /* 1082 * Lower 2 bits of the agent ID determine which BUSY/NACK pair 1083 * will be used for dispatching interrupt. For now, assume 1084 * there are no more than IDSR_BN_SETS CPUs, hence no aliasing 1085 * issues with respect to BUSY/NACK pair usage. 1086 */ 1087 busymask = IDSR_BUSY_BIT(cpuid); 1088 #else /* JALAPENO || SERRANO */ 1089 busymask = IDSR_BUSY; 1090 #endif /* JALAPENO || SERRANO */ 1091 for (;;) { 1092 idsr = getidsr(); 1093 if (idsr == 0) 1094 break; 1095 1096 tick = gettick(); 1097 /* 1098 * If there is a big jump between the current tick 1099 * count and lasttick, we have probably hit a break 1100 * point. Adjust endtick accordingly to avoid panic. 1101 */ 1102 if (tick > (lasttick + xc_tick_jump_limit)) 1103 endtick += (tick - lasttick); 1104 lasttick = tick; 1105 if (tick > endtick) { 1106 if (panic_quiesce) 1107 return; 1108 #ifdef CHEETAHPLUS_ERRATUM_25 1109 if (cheetah_sendmondo_recover && recovered == 0) { 1110 if (mondo_recover(cpuid, 0)) { 1111 /* 1112 * We claimed the whole memory or 1113 * full scan is disabled. 1114 */ 1115 recovered++; 1116 } 1117 tick = gettick(); 1118 endtick = tick + xc_tick_limit; 1119 lasttick = tick; 1120 /* 1121 * Recheck idsr 1122 */ 1123 continue; 1124 } else 1125 #endif /* CHEETAHPLUS_ERRATUM_25 */ 1126 { 1127 cmn_err(CE_PANIC, "send mondo timeout " 1128 "(target 0x%x) [%d NACK %d BUSY]", 1129 cpuid, nack, busy); 1130 } 1131 } 1132 1133 if (idsr & busymask) { 1134 busy++; 1135 continue; 1136 } 1137 drv_usecwait(1); 1138 shipit(cpuid, 0); 1139 nack++; 1140 busy = 0; 1141 } 1142 #ifdef SEND_MONDO_STATS 1143 { 1144 int n = gettick() - starttick; 1145 if (n < 8192) 1146 x_one_stimes[n >> 7]++; 1147 else 1148 x_one_ltimes[(n >> 13) & 0xf]++; 1149 } 1150 #endif 1151 } 1152 1153 void 1154 syncfpu(void) 1155 { 1156 } 1157 1158 /* 1159 * Return processor specific async error structure 1160 * size used. 1161 */ 1162 int 1163 cpu_aflt_size(void) 1164 { 1165 return (sizeof (ch_async_flt_t)); 1166 } 1167 1168 /* 1169 * Tunable to disable the checking of other cpu logout areas during panic for 1170 * potential syndrome 71 generating errors. 1171 */ 1172 int enable_check_other_cpus_logout = 1; 1173 1174 /* 1175 * Check other cpus logout area for potential synd 71 generating 1176 * errors. 1177 */ 1178 static void 1179 cpu_check_cpu_logout(int cpuid, caddr_t tpc, int tl, int ecc_type, 1180 ch_cpu_logout_t *clop) 1181 { 1182 struct async_flt *aflt; 1183 ch_async_flt_t ch_flt; 1184 uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs; 1185 1186 if (clop == NULL || clop->clo_data.chd_afar == LOGOUT_INVALID) { 1187 return; 1188 } 1189 1190 bzero(&ch_flt, sizeof (ch_async_flt_t)); 1191 1192 t_afar = clop->clo_data.chd_afar; 1193 t_afsr = clop->clo_data.chd_afsr; 1194 t_afsr_ext = clop->clo_data.chd_afsr_ext; 1195 #if defined(SERRANO) 1196 ch_flt.afar2 = clop->clo_data.chd_afar2; 1197 #endif /* SERRANO */ 1198 1199 /* 1200 * In order to simplify code, we maintain this afsr_errs 1201 * variable which holds the aggregate of AFSR and AFSR_EXT 1202 * sticky bits. 1203 */ 1204 t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) | 1205 (t_afsr & C_AFSR_ALL_ERRS); 1206 1207 /* Setup the async fault structure */ 1208 aflt = (struct async_flt *)&ch_flt; 1209 aflt->flt_id = gethrtime_waitfree(); 1210 ch_flt.afsr_ext = t_afsr_ext; 1211 ch_flt.afsr_errs = t_afsr_errs; 1212 aflt->flt_stat = t_afsr; 1213 aflt->flt_addr = t_afar; 1214 aflt->flt_bus_id = cpuid; 1215 aflt->flt_inst = cpuid; 1216 aflt->flt_pc = tpc; 1217 aflt->flt_prot = AFLT_PROT_NONE; 1218 aflt->flt_class = CPU_FAULT; 1219 aflt->flt_priv = ((t_afsr & C_AFSR_PRIV) != 0); 1220 aflt->flt_tl = tl; 1221 aflt->flt_status = ecc_type; 1222 aflt->flt_panic = C_AFSR_PANIC(t_afsr_errs); 1223 1224 /* 1225 * Queue events on the async event queue, one event per error bit. 1226 * If no events are queued, queue an event to complain. 1227 */ 1228 if (cpu_queue_events(&ch_flt, NULL, t_afsr_errs, clop) == 0) { 1229 ch_flt.flt_type = CPU_INV_AFSR; 1230 cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR, 1231 (void *)&ch_flt, sizeof (ch_async_flt_t), ue_queue, 1232 aflt->flt_panic); 1233 } 1234 1235 /* 1236 * Zero out + invalidate CPU logout. 1237 */ 1238 bzero(clop, sizeof (ch_cpu_logout_t)); 1239 clop->clo_data.chd_afar = LOGOUT_INVALID; 1240 } 1241 1242 /* 1243 * Check the logout areas of all other cpus for unlogged errors. 1244 */ 1245 static void 1246 cpu_check_other_cpus_logout(void) 1247 { 1248 int i, j; 1249 processorid_t myid; 1250 struct cpu *cp; 1251 ch_err_tl1_data_t *cl1p; 1252 1253 myid = CPU->cpu_id; 1254 for (i = 0; i < NCPU; i++) { 1255 cp = cpu[i]; 1256 1257 if ((cp == NULL) || !(cp->cpu_flags & CPU_EXISTS) || 1258 (cp->cpu_id == myid) || (CPU_PRIVATE(cp) == NULL)) { 1259 continue; 1260 } 1261 1262 /* 1263 * Check each of the tl>0 logout areas 1264 */ 1265 cl1p = CPU_PRIVATE_PTR(cp, chpr_tl1_err_data[0]); 1266 for (j = 0; j < CH_ERR_TL1_TLMAX; j++, cl1p++) { 1267 if (cl1p->ch_err_tl1_flags == 0) 1268 continue; 1269 1270 cpu_check_cpu_logout(i, (caddr_t)cl1p->ch_err_tl1_tpc, 1271 1, ECC_F_TRAP, &cl1p->ch_err_tl1_logout); 1272 } 1273 1274 /* 1275 * Check each of the remaining logout areas 1276 */ 1277 cpu_check_cpu_logout(i, NULL, 0, ECC_F_TRAP, 1278 CPU_PRIVATE_PTR(cp, chpr_fecctl0_logout)); 1279 cpu_check_cpu_logout(i, NULL, 0, ECC_C_TRAP, 1280 CPU_PRIVATE_PTR(cp, chpr_cecc_logout)); 1281 cpu_check_cpu_logout(i, NULL, 0, ECC_D_TRAP, 1282 CPU_PRIVATE_PTR(cp, chpr_async_logout)); 1283 } 1284 } 1285 1286 /* 1287 * The fast_ecc_err handler transfers control here for UCU, UCC events. 1288 * Note that we flush Ecache twice, once in the fast_ecc_err handler to 1289 * flush the error that caused the UCU/UCC, then again here at the end to 1290 * flush the TL=1 trap handler code out of the Ecache, so we can minimize 1291 * the probability of getting a TL>1 Fast ECC trap when we're fielding 1292 * another Fast ECC trap. 1293 * 1294 * Cheetah+ also handles: TSCE: No additional processing required. 1295 * Panther adds L3_UCU and L3_UCC which are reported in AFSR_EXT. 1296 * 1297 * Note that the p_clo_flags input is only valid in cases where the 1298 * cpu_private struct is not yet initialized (since that is the only 1299 * time that information cannot be obtained from the logout struct.) 1300 */ 1301 /*ARGSUSED*/ 1302 void 1303 cpu_fast_ecc_error(struct regs *rp, ulong_t p_clo_flags) 1304 { 1305 ch_cpu_logout_t *clop; 1306 uint64_t ceen, nceen; 1307 1308 /* 1309 * Get the CPU log out info. If we can't find our CPU private 1310 * pointer, then we will have to make due without any detailed 1311 * logout information. 1312 */ 1313 if (CPU_PRIVATE(CPU) == NULL) { 1314 clop = NULL; 1315 ceen = p_clo_flags & EN_REG_CEEN; 1316 nceen = p_clo_flags & EN_REG_NCEEN; 1317 } else { 1318 clop = CPU_PRIVATE_PTR(CPU, chpr_fecctl0_logout); 1319 ceen = clop->clo_flags & EN_REG_CEEN; 1320 nceen = clop->clo_flags & EN_REG_NCEEN; 1321 } 1322 1323 cpu_log_fast_ecc_error((caddr_t)rp->r_pc, 1324 (rp->r_tstate & TSTATE_PRIV) ? 1 : 0, 0, ceen, nceen, clop); 1325 } 1326 1327 /* 1328 * Log fast ecc error, called from either Fast ECC at TL=0 or Fast 1329 * ECC at TL>0. Need to supply either a error register pointer or a 1330 * cpu logout structure pointer. 1331 */ 1332 static void 1333 cpu_log_fast_ecc_error(caddr_t tpc, int priv, int tl, uint64_t ceen, 1334 uint64_t nceen, ch_cpu_logout_t *clop) 1335 { 1336 struct async_flt *aflt; 1337 ch_async_flt_t ch_flt; 1338 uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs; 1339 char pr_reason[MAX_REASON_STRING]; 1340 ch_cpu_errors_t cpu_error_regs; 1341 1342 bzero(&ch_flt, sizeof (ch_async_flt_t)); 1343 /* 1344 * If no cpu logout data, then we will have to make due without 1345 * any detailed logout information. 1346 */ 1347 if (clop == NULL) { 1348 ch_flt.flt_diag_data.chd_afar = LOGOUT_INVALID; 1349 get_cpu_error_state(&cpu_error_regs); 1350 set_cpu_error_state(&cpu_error_regs); 1351 t_afar = cpu_error_regs.afar; 1352 t_afsr = cpu_error_regs.afsr; 1353 t_afsr_ext = cpu_error_regs.afsr_ext; 1354 #if defined(SERRANO) 1355 ch_flt.afar2 = cpu_error_regs.afar2; 1356 #endif /* SERRANO */ 1357 } else { 1358 t_afar = clop->clo_data.chd_afar; 1359 t_afsr = clop->clo_data.chd_afsr; 1360 t_afsr_ext = clop->clo_data.chd_afsr_ext; 1361 #if defined(SERRANO) 1362 ch_flt.afar2 = clop->clo_data.chd_afar2; 1363 #endif /* SERRANO */ 1364 } 1365 1366 /* 1367 * In order to simplify code, we maintain this afsr_errs 1368 * variable which holds the aggregate of AFSR and AFSR_EXT 1369 * sticky bits. 1370 */ 1371 t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) | 1372 (t_afsr & C_AFSR_ALL_ERRS); 1373 pr_reason[0] = '\0'; 1374 1375 /* Setup the async fault structure */ 1376 aflt = (struct async_flt *)&ch_flt; 1377 aflt->flt_id = gethrtime_waitfree(); 1378 ch_flt.afsr_ext = t_afsr_ext; 1379 ch_flt.afsr_errs = t_afsr_errs; 1380 aflt->flt_stat = t_afsr; 1381 aflt->flt_addr = t_afar; 1382 aflt->flt_bus_id = getprocessorid(); 1383 aflt->flt_inst = CPU->cpu_id; 1384 aflt->flt_pc = tpc; 1385 aflt->flt_prot = AFLT_PROT_NONE; 1386 aflt->flt_class = CPU_FAULT; 1387 aflt->flt_priv = priv; 1388 aflt->flt_tl = tl; 1389 aflt->flt_status = ECC_F_TRAP; 1390 aflt->flt_panic = C_AFSR_PANIC(t_afsr_errs); 1391 1392 /* 1393 * XXXX - Phenomenal hack to get around Solaris not getting all the 1394 * cmn_err messages out to the console. The situation is a UCU (in 1395 * priv mode) which causes a WDU which causes a UE (on the retry). 1396 * The messages for the UCU and WDU are enqueued and then pulled off 1397 * the async queue via softint and syslogd starts to process them 1398 * but doesn't get them to the console. The UE causes a panic, but 1399 * since the UCU/WDU messages are already in transit, those aren't 1400 * on the async queue. The hack is to check if we have a matching 1401 * WDU event for the UCU, and if it matches, we're more than likely 1402 * going to panic with a UE, unless we're under protection. So, we 1403 * check to see if we got a matching WDU event and if we're under 1404 * protection. 1405 * 1406 * For Cheetah/Cheetah+/Jaguar/Jalapeno, the sequence we care about 1407 * looks like this: 1408 * UCU->WDU->UE 1409 * For Panther, it could look like either of these: 1410 * UCU---->WDU->L3_WDU->UE 1411 * L3_UCU->WDU->L3_WDU->UE 1412 */ 1413 if ((t_afsr_errs & (C_AFSR_UCU | C_AFSR_L3_UCU)) && 1414 aflt->flt_panic == 0 && aflt->flt_priv != 0 && 1415 curthread->t_ontrap == NULL && curthread->t_lofault == NULL) { 1416 get_cpu_error_state(&cpu_error_regs); 1417 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 1418 aflt->flt_panic |= 1419 ((cpu_error_regs.afsr & C_AFSR_WDU) && 1420 (cpu_error_regs.afsr_ext & C_AFSR_L3_WDU) && 1421 (cpu_error_regs.afar == t_afar)); 1422 aflt->flt_panic |= ((clop == NULL) && 1423 (t_afsr_errs & C_AFSR_WDU) && 1424 (t_afsr_errs & C_AFSR_L3_WDU)); 1425 } else { 1426 aflt->flt_panic |= 1427 ((cpu_error_regs.afsr & C_AFSR_WDU) && 1428 (cpu_error_regs.afar == t_afar)); 1429 aflt->flt_panic |= ((clop == NULL) && 1430 (t_afsr_errs & C_AFSR_WDU)); 1431 } 1432 } 1433 1434 /* 1435 * Queue events on the async event queue, one event per error bit. 1436 * If no events are queued or no Fast ECC events are on in the AFSR, 1437 * queue an event to complain. 1438 */ 1439 if (cpu_queue_events(&ch_flt, pr_reason, t_afsr_errs, clop) == 0 || 1440 ((t_afsr_errs & (C_AFSR_FECC_ERRS | C_AFSR_EXT_FECC_ERRS)) == 0)) { 1441 ch_flt.flt_type = CPU_INV_AFSR; 1442 cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR, 1443 (void *)&ch_flt, sizeof (ch_async_flt_t), ue_queue, 1444 aflt->flt_panic); 1445 } 1446 1447 /* 1448 * Zero out + invalidate CPU logout. 1449 */ 1450 if (clop) { 1451 bzero(clop, sizeof (ch_cpu_logout_t)); 1452 clop->clo_data.chd_afar = LOGOUT_INVALID; 1453 } 1454 1455 /* 1456 * We carefully re-enable NCEEN and CEEN and then check if any deferred 1457 * or disrupting errors have happened. We do this because if a 1458 * deferred or disrupting error had occurred with NCEEN/CEEN off, the 1459 * trap will not be taken when NCEEN/CEEN is re-enabled. Note that 1460 * CEEN works differently on Cheetah than on Spitfire. Also, we enable 1461 * NCEEN/CEEN *before* checking the AFSR to avoid the small window of a 1462 * deferred or disrupting error happening between checking the AFSR and 1463 * enabling NCEEN/CEEN. 1464 * 1465 * Note: CEEN and NCEEN are only reenabled if they were on when trap 1466 * taken. 1467 */ 1468 set_error_enable(get_error_enable() | (nceen | ceen)); 1469 if (clear_errors(&ch_flt)) { 1470 aflt->flt_panic |= ((ch_flt.afsr_errs & 1471 (C_AFSR_EXT_ASYNC_ERRS | C_AFSR_ASYNC_ERRS)) != 0); 1472 (void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs, 1473 NULL); 1474 } 1475 1476 /* 1477 * Panic here if aflt->flt_panic has been set. Enqueued errors will 1478 * be logged as part of the panic flow. 1479 */ 1480 if (aflt->flt_panic) 1481 fm_panic("%sError(s)", pr_reason); 1482 1483 /* 1484 * Flushing the Ecache here gets the part of the trap handler that 1485 * is run at TL=1 out of the Ecache. 1486 */ 1487 cpu_flush_ecache(); 1488 } 1489 1490 /* 1491 * This is called via sys_trap from pil15_interrupt code if the 1492 * corresponding entry in ch_err_tl1_pending is set. Checks the 1493 * various ch_err_tl1_data structures for valid entries based on the bit 1494 * settings in the ch_err_tl1_flags entry of the structure. 1495 */ 1496 /*ARGSUSED*/ 1497 void 1498 cpu_tl1_error(struct regs *rp, int panic) 1499 { 1500 ch_err_tl1_data_t *cl1p, cl1; 1501 int i, ncl1ps; 1502 uint64_t me_flags; 1503 uint64_t ceen, nceen; 1504 1505 if (ch_err_tl1_paddrs[CPU->cpu_id] == 0) { 1506 cl1p = &ch_err_tl1_data; 1507 ncl1ps = 1; 1508 } else if (CPU_PRIVATE(CPU) != NULL) { 1509 cl1p = CPU_PRIVATE_PTR(CPU, chpr_tl1_err_data[0]); 1510 ncl1ps = CH_ERR_TL1_TLMAX; 1511 } else { 1512 ncl1ps = 0; 1513 } 1514 1515 for (i = 0; i < ncl1ps; i++, cl1p++) { 1516 if (cl1p->ch_err_tl1_flags == 0) 1517 continue; 1518 1519 /* 1520 * Grab a copy of the logout data and invalidate 1521 * the logout area. 1522 */ 1523 cl1 = *cl1p; 1524 bzero(cl1p, sizeof (ch_err_tl1_data_t)); 1525 cl1p->ch_err_tl1_logout.clo_data.chd_afar = LOGOUT_INVALID; 1526 me_flags = CH_ERR_ME_FLAGS(cl1.ch_err_tl1_flags); 1527 1528 /* 1529 * Log "first error" in ch_err_tl1_data. 1530 */ 1531 if (cl1.ch_err_tl1_flags & CH_ERR_FECC) { 1532 ceen = get_error_enable() & EN_REG_CEEN; 1533 nceen = get_error_enable() & EN_REG_NCEEN; 1534 cpu_log_fast_ecc_error((caddr_t)cl1.ch_err_tl1_tpc, 1, 1535 1, ceen, nceen, &cl1.ch_err_tl1_logout); 1536 } 1537 #if defined(CPU_IMP_L1_CACHE_PARITY) 1538 if (cl1.ch_err_tl1_flags & (CH_ERR_IPE | CH_ERR_DPE)) { 1539 cpu_parity_error(rp, cl1.ch_err_tl1_flags, 1540 (caddr_t)cl1.ch_err_tl1_tpc); 1541 } 1542 #endif /* CPU_IMP_L1_CACHE_PARITY */ 1543 1544 /* 1545 * Log "multiple events" in ch_err_tl1_data. Note that 1546 * we don't read and clear the AFSR/AFAR in the TL>0 code 1547 * if the structure is busy, we just do the cache flushing 1548 * we have to do and then do the retry. So the AFSR/AFAR 1549 * at this point *should* have some relevant info. If there 1550 * are no valid errors in the AFSR, we'll assume they've 1551 * already been picked up and logged. For I$/D$ parity, 1552 * we just log an event with an "Unknown" (NULL) TPC. 1553 */ 1554 if (me_flags & CH_ERR_FECC) { 1555 ch_cpu_errors_t cpu_error_regs; 1556 uint64_t t_afsr_errs; 1557 1558 /* 1559 * Get the error registers and see if there's 1560 * a pending error. If not, don't bother 1561 * generating an "Invalid AFSR" error event. 1562 */ 1563 get_cpu_error_state(&cpu_error_regs); 1564 t_afsr_errs = (cpu_error_regs.afsr_ext & 1565 C_AFSR_EXT_ALL_ERRS) | 1566 (cpu_error_regs.afsr & C_AFSR_ALL_ERRS); 1567 if (t_afsr_errs != 0) { 1568 ceen = get_error_enable() & EN_REG_CEEN; 1569 nceen = get_error_enable() & EN_REG_NCEEN; 1570 cpu_log_fast_ecc_error((caddr_t)NULL, 1, 1571 1, ceen, nceen, NULL); 1572 } 1573 } 1574 #if defined(CPU_IMP_L1_CACHE_PARITY) 1575 if (me_flags & (CH_ERR_IPE | CH_ERR_DPE)) { 1576 cpu_parity_error(rp, me_flags, (caddr_t)NULL); 1577 } 1578 #endif /* CPU_IMP_L1_CACHE_PARITY */ 1579 } 1580 } 1581 1582 /* 1583 * Called from Fast ECC TL>0 handler in case of fatal error. 1584 * cpu_tl1_error should always find an associated ch_err_tl1_data structure, 1585 * but if we don't, we'll panic with something reasonable. 1586 */ 1587 /*ARGSUSED*/ 1588 void 1589 cpu_tl1_err_panic(struct regs *rp, ulong_t flags) 1590 { 1591 cpu_tl1_error(rp, 1); 1592 /* 1593 * Should never return, but just in case. 1594 */ 1595 fm_panic("Unsurvivable ECC Error at TL>0"); 1596 } 1597 1598 /* 1599 * The ce_err/ce_err_tl1 handlers transfer control here for CE, EMC, EDU:ST, 1600 * EDC, WDU, WDC, CPU, CPC, IVU, IVC events. 1601 * Disrupting errors controlled by NCEEN: EDU:ST, WDU, CPU, IVU 1602 * Disrupting errors controlled by CEEN: CE, EMC, EDC, WDC, CPC, IVC 1603 * 1604 * Cheetah+ also handles (No additional processing required): 1605 * DUE, DTO, DBERR (NCEEN controlled) 1606 * THCE (CEEN and ET_ECC_en controlled) 1607 * TUE (ET_ECC_en controlled) 1608 * 1609 * Panther further adds: 1610 * IMU, L3_EDU, L3_WDU, L3_CPU (NCEEN controlled) 1611 * IMC, L3_EDC, L3_WDC, L3_CPC, L3_THCE (CEEN controlled) 1612 * TUE_SH, TUE (NCEEN and L2_tag_ECC_en controlled) 1613 * L3_TUE, L3_TUE_SH (NCEEN and ET_ECC_en controlled) 1614 * THCE (CEEN and L2_tag_ECC_en controlled) 1615 * L3_THCE (CEEN and ET_ECC_en controlled) 1616 * 1617 * Note that the p_clo_flags input is only valid in cases where the 1618 * cpu_private struct is not yet initialized (since that is the only 1619 * time that information cannot be obtained from the logout struct.) 1620 */ 1621 /*ARGSUSED*/ 1622 void 1623 cpu_disrupting_error(struct regs *rp, ulong_t p_clo_flags) 1624 { 1625 struct async_flt *aflt; 1626 ch_async_flt_t ch_flt; 1627 char pr_reason[MAX_REASON_STRING]; 1628 ch_cpu_logout_t *clop; 1629 uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs; 1630 ch_cpu_errors_t cpu_error_regs; 1631 1632 bzero(&ch_flt, sizeof (ch_async_flt_t)); 1633 /* 1634 * Get the CPU log out info. If we can't find our CPU private 1635 * pointer, then we will have to make due without any detailed 1636 * logout information. 1637 */ 1638 if (CPU_PRIVATE(CPU) == NULL) { 1639 clop = NULL; 1640 ch_flt.flt_diag_data.chd_afar = LOGOUT_INVALID; 1641 get_cpu_error_state(&cpu_error_regs); 1642 set_cpu_error_state(&cpu_error_regs); 1643 t_afar = cpu_error_regs.afar; 1644 t_afsr = cpu_error_regs.afsr; 1645 t_afsr_ext = cpu_error_regs.afsr_ext; 1646 #if defined(SERRANO) 1647 ch_flt.afar2 = cpu_error_regs.afar2; 1648 #endif /* SERRANO */ 1649 } else { 1650 clop = CPU_PRIVATE_PTR(CPU, chpr_cecc_logout); 1651 t_afar = clop->clo_data.chd_afar; 1652 t_afsr = clop->clo_data.chd_afsr; 1653 t_afsr_ext = clop->clo_data.chd_afsr_ext; 1654 #if defined(SERRANO) 1655 ch_flt.afar2 = clop->clo_data.chd_afar2; 1656 #endif /* SERRANO */ 1657 } 1658 1659 /* 1660 * In order to simplify code, we maintain this afsr_errs 1661 * variable which holds the aggregate of AFSR and AFSR_EXT 1662 * sticky bits. 1663 */ 1664 t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) | 1665 (t_afsr & C_AFSR_ALL_ERRS); 1666 1667 pr_reason[0] = '\0'; 1668 /* Setup the async fault structure */ 1669 aflt = (struct async_flt *)&ch_flt; 1670 ch_flt.afsr_ext = t_afsr_ext; 1671 ch_flt.afsr_errs = t_afsr_errs; 1672 aflt->flt_stat = t_afsr; 1673 aflt->flt_addr = t_afar; 1674 aflt->flt_pc = (caddr_t)rp->r_pc; 1675 aflt->flt_priv = (rp->r_tstate & TSTATE_PRIV) ? 1 : 0; 1676 aflt->flt_tl = 0; 1677 aflt->flt_panic = C_AFSR_PANIC(t_afsr_errs); 1678 1679 /* 1680 * If this trap is a result of one of the errors not masked 1681 * by cpu_ce_not_deferred, we don't reenable CEEN. Instead 1682 * indicate that a timeout is to be set later. 1683 */ 1684 if (!(t_afsr_errs & (cpu_ce_not_deferred | cpu_ce_not_deferred_ext)) && 1685 !aflt->flt_panic) 1686 ch_flt.flt_trapped_ce = CE_CEEN_DEFER | CE_CEEN_TRAPPED; 1687 else 1688 ch_flt.flt_trapped_ce = CE_CEEN_NODEFER | CE_CEEN_TRAPPED; 1689 1690 /* 1691 * log the CE and clean up 1692 */ 1693 cpu_log_and_clear_ce(&ch_flt); 1694 1695 /* 1696 * We re-enable CEEN (if required) and check if any disrupting errors 1697 * have happened. We do this because if a disrupting error had occurred 1698 * with CEEN off, the trap will not be taken when CEEN is re-enabled. 1699 * Note that CEEN works differently on Cheetah than on Spitfire. Also, 1700 * we enable CEEN *before* checking the AFSR to avoid the small window 1701 * of a error happening between checking the AFSR and enabling CEEN. 1702 */ 1703 if (ch_flt.flt_trapped_ce & CE_CEEN_NODEFER) 1704 set_error_enable(get_error_enable() | EN_REG_CEEN); 1705 if (clear_errors(&ch_flt)) { 1706 (void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs, 1707 NULL); 1708 } 1709 1710 /* 1711 * Panic here if aflt->flt_panic has been set. Enqueued errors will 1712 * be logged as part of the panic flow. 1713 */ 1714 if (aflt->flt_panic) 1715 fm_panic("%sError(s)", pr_reason); 1716 } 1717 1718 /* 1719 * The async_err handler transfers control here for UE, EMU, EDU:BLD, 1720 * L3_EDU:BLD, TO, and BERR events. 1721 * Deferred errors controlled by NCEEN: UE, EMU, EDU:BLD, L3_EDU:BLD, TO, BERR 1722 * 1723 * Cheetah+: No additional errors handled. 1724 * 1725 * Note that the p_clo_flags input is only valid in cases where the 1726 * cpu_private struct is not yet initialized (since that is the only 1727 * time that information cannot be obtained from the logout struct.) 1728 */ 1729 /*ARGSUSED*/ 1730 void 1731 cpu_deferred_error(struct regs *rp, ulong_t p_clo_flags) 1732 { 1733 ushort_t ttype, tl; 1734 ch_async_flt_t ch_flt; 1735 struct async_flt *aflt; 1736 int trampolined = 0; 1737 char pr_reason[MAX_REASON_STRING]; 1738 ch_cpu_logout_t *clop; 1739 uint64_t ceen, clo_flags; 1740 uint64_t log_afsr; 1741 uint64_t t_afar, t_afsr, t_afsr_ext, t_afsr_errs; 1742 ch_cpu_errors_t cpu_error_regs; 1743 int expected = DDI_FM_ERR_UNEXPECTED; 1744 ddi_acc_hdl_t *hp; 1745 1746 /* 1747 * We need to look at p_flag to determine if the thread detected an 1748 * error while dumping core. We can't grab p_lock here, but it's ok 1749 * because we just need a consistent snapshot and we know that everyone 1750 * else will store a consistent set of bits while holding p_lock. We 1751 * don't have to worry about a race because SDOCORE is set once prior 1752 * to doing i/o from the process's address space and is never cleared. 1753 */ 1754 uint_t pflag = ttoproc(curthread)->p_flag; 1755 1756 bzero(&ch_flt, sizeof (ch_async_flt_t)); 1757 /* 1758 * Get the CPU log out info. If we can't find our CPU private 1759 * pointer then we will have to make due without any detailed 1760 * logout information. 1761 */ 1762 if (CPU_PRIVATE(CPU) == NULL) { 1763 clop = NULL; 1764 ch_flt.flt_diag_data.chd_afar = LOGOUT_INVALID; 1765 get_cpu_error_state(&cpu_error_regs); 1766 set_cpu_error_state(&cpu_error_regs); 1767 t_afar = cpu_error_regs.afar; 1768 t_afsr = cpu_error_regs.afsr; 1769 t_afsr_ext = cpu_error_regs.afsr_ext; 1770 #if defined(SERRANO) 1771 ch_flt.afar2 = cpu_error_regs.afar2; 1772 #endif /* SERRANO */ 1773 clo_flags = p_clo_flags; 1774 } else { 1775 clop = CPU_PRIVATE_PTR(CPU, chpr_async_logout); 1776 t_afar = clop->clo_data.chd_afar; 1777 t_afsr = clop->clo_data.chd_afsr; 1778 t_afsr_ext = clop->clo_data.chd_afsr_ext; 1779 #if defined(SERRANO) 1780 ch_flt.afar2 = clop->clo_data.chd_afar2; 1781 #endif /* SERRANO */ 1782 clo_flags = clop->clo_flags; 1783 } 1784 1785 /* 1786 * In order to simplify code, we maintain this afsr_errs 1787 * variable which holds the aggregate of AFSR and AFSR_EXT 1788 * sticky bits. 1789 */ 1790 t_afsr_errs = (t_afsr_ext & C_AFSR_EXT_ALL_ERRS) | 1791 (t_afsr & C_AFSR_ALL_ERRS); 1792 pr_reason[0] = '\0'; 1793 1794 /* 1795 * Grab information encoded into our clo_flags field. 1796 */ 1797 ceen = clo_flags & EN_REG_CEEN; 1798 tl = (clo_flags & CLO_FLAGS_TL_MASK) >> CLO_FLAGS_TL_SHIFT; 1799 ttype = (clo_flags & CLO_FLAGS_TT_MASK) >> CLO_FLAGS_TT_SHIFT; 1800 1801 /* 1802 * handle the specific error 1803 */ 1804 aflt = (struct async_flt *)&ch_flt; 1805 aflt->flt_id = gethrtime_waitfree(); 1806 aflt->flt_bus_id = getprocessorid(); 1807 aflt->flt_inst = CPU->cpu_id; 1808 ch_flt.afsr_ext = t_afsr_ext; 1809 ch_flt.afsr_errs = t_afsr_errs; 1810 aflt->flt_stat = t_afsr; 1811 aflt->flt_addr = t_afar; 1812 aflt->flt_pc = (caddr_t)rp->r_pc; 1813 aflt->flt_prot = AFLT_PROT_NONE; 1814 aflt->flt_class = CPU_FAULT; 1815 aflt->flt_priv = (rp->r_tstate & TSTATE_PRIV) ? 1 : 0; 1816 aflt->flt_tl = (uchar_t)tl; 1817 aflt->flt_panic = ((tl != 0) || (aft_testfatal != 0) || 1818 C_AFSR_PANIC(t_afsr_errs)); 1819 aflt->flt_core = (pflag & SDOCORE) ? 1 : 0; 1820 aflt->flt_status = ((ttype == T_DATA_ERROR) ? ECC_D_TRAP : ECC_I_TRAP); 1821 1822 /* 1823 * If the trap occurred in privileged mode at TL=0, we need to check to 1824 * see if we were executing in the kernel under on_trap() or t_lofault 1825 * protection. If so, modify the saved registers so that we return 1826 * from the trap to the appropriate trampoline routine. 1827 */ 1828 if (aflt->flt_priv && tl == 0) { 1829 if (curthread->t_ontrap != NULL) { 1830 on_trap_data_t *otp = curthread->t_ontrap; 1831 1832 if (otp->ot_prot & OT_DATA_EC) { 1833 aflt->flt_prot = AFLT_PROT_EC; 1834 otp->ot_trap |= OT_DATA_EC; 1835 rp->r_pc = otp->ot_trampoline; 1836 rp->r_npc = rp->r_pc + 4; 1837 trampolined = 1; 1838 } 1839 1840 if ((t_afsr & (C_AFSR_TO | C_AFSR_BERR)) && 1841 (otp->ot_prot & OT_DATA_ACCESS)) { 1842 aflt->flt_prot = AFLT_PROT_ACCESS; 1843 otp->ot_trap |= OT_DATA_ACCESS; 1844 rp->r_pc = otp->ot_trampoline; 1845 rp->r_npc = rp->r_pc + 4; 1846 trampolined = 1; 1847 /* 1848 * for peeks and caut_gets errors are expected 1849 */ 1850 hp = (ddi_acc_hdl_t *)otp->ot_handle; 1851 if (!hp) 1852 expected = DDI_FM_ERR_PEEK; 1853 else if (hp->ah_acc.devacc_attr_access == 1854 DDI_CAUTIOUS_ACC) 1855 expected = DDI_FM_ERR_EXPECTED; 1856 } 1857 1858 } else if (curthread->t_lofault) { 1859 aflt->flt_prot = AFLT_PROT_COPY; 1860 rp->r_g1 = EFAULT; 1861 rp->r_pc = curthread->t_lofault; 1862 rp->r_npc = rp->r_pc + 4; 1863 trampolined = 1; 1864 } 1865 } 1866 1867 /* 1868 * If we're in user mode or we're doing a protected copy, we either 1869 * want the ASTON code below to send a signal to the user process 1870 * or we want to panic if aft_panic is set. 1871 * 1872 * If we're in privileged mode and we're not doing a copy, then we 1873 * need to check if we've trampolined. If we haven't trampolined, 1874 * we should panic. 1875 */ 1876 if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) { 1877 if (t_afsr_errs & 1878 ((C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS) & 1879 ~(C_AFSR_BERR | C_AFSR_TO))) 1880 aflt->flt_panic |= aft_panic; 1881 } else if (!trampolined) { 1882 aflt->flt_panic = 1; 1883 } 1884 1885 /* 1886 * If we've trampolined due to a privileged TO or BERR, or if an 1887 * unprivileged TO or BERR occurred, we don't want to enqueue an 1888 * event for that TO or BERR. Queue all other events (if any) besides 1889 * the TO/BERR. Since we may not be enqueing any events, we need to 1890 * ignore the number of events queued. If we haven't trampolined due 1891 * to a TO or BERR, just enqueue events normally. 1892 */ 1893 log_afsr = t_afsr_errs; 1894 if (trampolined) { 1895 log_afsr &= ~(C_AFSR_TO | C_AFSR_BERR); 1896 } else if (!aflt->flt_priv) { 1897 /* 1898 * User mode, suppress messages if 1899 * cpu_berr_to_verbose is not set. 1900 */ 1901 if (!cpu_berr_to_verbose) 1902 log_afsr &= ~(C_AFSR_TO | C_AFSR_BERR); 1903 } 1904 1905 /* 1906 * Log any errors that occurred 1907 */ 1908 if (((log_afsr & 1909 ((C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS) & ~C_AFSR_ME)) && 1910 cpu_queue_events(&ch_flt, pr_reason, log_afsr, clop) == 0) || 1911 (t_afsr_errs & (C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS)) == 0) { 1912 ch_flt.flt_type = CPU_INV_AFSR; 1913 cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR, 1914 (void *)&ch_flt, sizeof (ch_async_flt_t), ue_queue, 1915 aflt->flt_panic); 1916 } 1917 1918 /* 1919 * Zero out + invalidate CPU logout. 1920 */ 1921 if (clop) { 1922 bzero(clop, sizeof (ch_cpu_logout_t)); 1923 clop->clo_data.chd_afar = LOGOUT_INVALID; 1924 } 1925 1926 #if defined(JALAPENO) || defined(SERRANO) 1927 /* 1928 * UE/RUE/BERR/TO: Call our bus nexus friends to check for 1929 * IO errors that may have resulted in this trap. 1930 */ 1931 if (t_afsr & (C_AFSR_UE|C_AFSR_RUE|C_AFSR_TO|C_AFSR_BERR)) { 1932 cpu_run_bus_error_handlers(aflt, expected); 1933 } 1934 1935 /* 1936 * UE/RUE: If UE or RUE is in memory, we need to flush the bad 1937 * line from the Ecache. We also need to query the bus nexus for 1938 * fatal errors. Attempts to do diagnostic read on caches may 1939 * introduce more errors (especially when the module is bad). 1940 */ 1941 if (t_afsr & (C_AFSR_UE|C_AFSR_RUE)) { 1942 /* 1943 * Ask our bus nexus friends if they have any fatal errors. If 1944 * so, they will log appropriate error messages. 1945 */ 1946 if (bus_func_invoke(BF_TYPE_UE) == BF_FATAL) 1947 aflt->flt_panic = 1; 1948 1949 /* 1950 * We got a UE or RUE and are panicking, save the fault PA in 1951 * a known location so that the platform specific panic code 1952 * can check for copyback errors. 1953 */ 1954 if (aflt->flt_panic && cpu_flt_in_memory(&ch_flt, C_AFSR_UE)) { 1955 panic_aflt = *aflt; 1956 } 1957 } 1958 1959 /* 1960 * Flush Ecache line or entire Ecache 1961 */ 1962 if (t_afsr & (C_AFSR_UE | C_AFSR_RUE | C_AFSR_EDU | C_AFSR_BERR)) 1963 cpu_error_ecache_flush(&ch_flt); 1964 #else /* JALAPENO || SERRANO */ 1965 /* 1966 * UE/BERR/TO: Call our bus nexus friends to check for 1967 * IO errors that may have resulted in this trap. 1968 */ 1969 if (t_afsr & (C_AFSR_UE|C_AFSR_TO|C_AFSR_BERR)) { 1970 cpu_run_bus_error_handlers(aflt, expected); 1971 } 1972 1973 /* 1974 * UE: If the UE is in memory, we need to flush the bad 1975 * line from the Ecache. We also need to query the bus nexus for 1976 * fatal errors. Attempts to do diagnostic read on caches may 1977 * introduce more errors (especially when the module is bad). 1978 */ 1979 if (t_afsr & C_AFSR_UE) { 1980 /* 1981 * Ask our legacy bus nexus friends if they have any fatal 1982 * errors. If so, they will log appropriate error messages. 1983 */ 1984 if (bus_func_invoke(BF_TYPE_UE) == BF_FATAL) 1985 aflt->flt_panic = 1; 1986 1987 /* 1988 * We got a UE and are panicking, save the fault PA in a known 1989 * location so that the platform specific panic code can check 1990 * for copyback errors. 1991 */ 1992 if (aflt->flt_panic && cpu_flt_in_memory(&ch_flt, C_AFSR_UE)) { 1993 panic_aflt = *aflt; 1994 } 1995 } 1996 1997 /* 1998 * Flush Ecache line or entire Ecache 1999 */ 2000 if (t_afsr_errs & 2001 (C_AFSR_UE | C_AFSR_EDU | C_AFSR_BERR | C_AFSR_L3_EDU)) 2002 cpu_error_ecache_flush(&ch_flt); 2003 #endif /* JALAPENO || SERRANO */ 2004 2005 /* 2006 * We carefully re-enable NCEEN and CEEN and then check if any deferred 2007 * or disrupting errors have happened. We do this because if a 2008 * deferred or disrupting error had occurred with NCEEN/CEEN off, the 2009 * trap will not be taken when NCEEN/CEEN is re-enabled. Note that 2010 * CEEN works differently on Cheetah than on Spitfire. Also, we enable 2011 * NCEEN/CEEN *before* checking the AFSR to avoid the small window of a 2012 * deferred or disrupting error happening between checking the AFSR and 2013 * enabling NCEEN/CEEN. 2014 * 2015 * Note: CEEN reenabled only if it was on when trap taken. 2016 */ 2017 set_error_enable(get_error_enable() | (EN_REG_NCEEN | ceen)); 2018 if (clear_errors(&ch_flt)) { 2019 /* 2020 * Check for secondary errors, and avoid panicking if we 2021 * have them 2022 */ 2023 if (cpu_check_secondary_errors(&ch_flt, t_afsr_errs, 2024 t_afar) == 0) { 2025 aflt->flt_panic |= ((ch_flt.afsr_errs & 2026 (C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS)) != 0); 2027 } 2028 (void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs, 2029 NULL); 2030 } 2031 2032 /* 2033 * Panic here if aflt->flt_panic has been set. Enqueued errors will 2034 * be logged as part of the panic flow. 2035 */ 2036 if (aflt->flt_panic) 2037 fm_panic("%sError(s)", pr_reason); 2038 2039 /* 2040 * If we queued an error and we are going to return from the trap and 2041 * the error was in user mode or inside of a copy routine, set AST flag 2042 * so the queue will be drained before returning to user mode. The 2043 * AST processing will also act on our failure policy. 2044 */ 2045 if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) { 2046 int pcb_flag = 0; 2047 2048 if (t_afsr_errs & 2049 (C_AFSR_ASYNC_ERRS | C_AFSR_EXT_ASYNC_ERRS & 2050 ~(C_AFSR_BERR | C_AFSR_TO))) 2051 pcb_flag |= ASYNC_HWERR; 2052 2053 if (t_afsr & C_AFSR_BERR) 2054 pcb_flag |= ASYNC_BERR; 2055 2056 if (t_afsr & C_AFSR_TO) 2057 pcb_flag |= ASYNC_BTO; 2058 2059 ttolwp(curthread)->lwp_pcb.pcb_flags |= pcb_flag; 2060 aston(curthread); 2061 } 2062 } 2063 2064 #if defined(CPU_IMP_L1_CACHE_PARITY) 2065 /* 2066 * Handling of data and instruction parity errors (traps 0x71, 0x72). 2067 * 2068 * For Panther, P$ data parity errors during floating point load hits 2069 * are also detected (reported as TT 0x71) and handled by this trap 2070 * handler. 2071 * 2072 * AFSR/AFAR are not set for parity errors, only TPC (a virtual address) 2073 * is available. 2074 */ 2075 /*ARGSUSED*/ 2076 void 2077 cpu_parity_error(struct regs *rp, uint_t flags, caddr_t tpc) 2078 { 2079 ch_async_flt_t ch_flt; 2080 struct async_flt *aflt; 2081 uchar_t tl = ((flags & CH_ERR_TL) != 0); 2082 uchar_t iparity = ((flags & CH_ERR_IPE) != 0); 2083 uchar_t panic = ((flags & CH_ERR_PANIC) != 0); 2084 char *error_class; 2085 2086 /* 2087 * Log the error. 2088 * For icache parity errors the fault address is the trap PC. 2089 * For dcache/pcache parity errors the instruction would have to 2090 * be decoded to determine the address and that isn't possible 2091 * at high PIL. 2092 */ 2093 bzero(&ch_flt, sizeof (ch_async_flt_t)); 2094 aflt = (struct async_flt *)&ch_flt; 2095 aflt->flt_id = gethrtime_waitfree(); 2096 aflt->flt_bus_id = getprocessorid(); 2097 aflt->flt_inst = CPU->cpu_id; 2098 aflt->flt_pc = tpc; 2099 aflt->flt_addr = iparity ? (uint64_t)tpc : AFLT_INV_ADDR; 2100 aflt->flt_prot = AFLT_PROT_NONE; 2101 aflt->flt_class = CPU_FAULT; 2102 aflt->flt_priv = (tl || (rp->r_tstate & TSTATE_PRIV)) ? 1 : 0; 2103 aflt->flt_tl = tl; 2104 aflt->flt_panic = panic; 2105 aflt->flt_status = iparity ? ECC_IP_TRAP : ECC_DP_TRAP; 2106 ch_flt.flt_type = iparity ? CPU_IC_PARITY : CPU_DC_PARITY; 2107 2108 if (iparity) { 2109 cpu_icache_parity_info(&ch_flt); 2110 if (ch_flt.parity_data.ipe.cpl_off != -1) 2111 error_class = FM_EREPORT_CPU_USIII_IDSPE; 2112 else if (ch_flt.parity_data.ipe.cpl_way != -1) 2113 error_class = FM_EREPORT_CPU_USIII_ITSPE; 2114 else 2115 error_class = FM_EREPORT_CPU_USIII_IPE; 2116 aflt->flt_payload = FM_EREPORT_PAYLOAD_ICACHE_PE; 2117 } else { 2118 cpu_dcache_parity_info(&ch_flt); 2119 if (ch_flt.parity_data.dpe.cpl_off != -1) 2120 error_class = FM_EREPORT_CPU_USIII_DDSPE; 2121 else if (ch_flt.parity_data.dpe.cpl_way != -1) 2122 error_class = FM_EREPORT_CPU_USIII_DTSPE; 2123 else 2124 error_class = FM_EREPORT_CPU_USIII_DPE; 2125 aflt->flt_payload = FM_EREPORT_PAYLOAD_DCACHE_PE; 2126 /* 2127 * For panther we also need to check the P$ for parity errors. 2128 */ 2129 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 2130 cpu_pcache_parity_info(&ch_flt); 2131 if (ch_flt.parity_data.dpe.cpl_cache == CPU_PC_PARITY) { 2132 error_class = FM_EREPORT_CPU_USIII_PDSPE; 2133 aflt->flt_payload = 2134 FM_EREPORT_PAYLOAD_PCACHE_PE; 2135 } 2136 } 2137 } 2138 2139 cpu_errorq_dispatch(error_class, (void *)&ch_flt, 2140 sizeof (ch_async_flt_t), ue_queue, aflt->flt_panic); 2141 2142 if (iparity) { 2143 /* 2144 * Invalidate entire I$. 2145 * This is required due to the use of diagnostic ASI 2146 * accesses that may result in a loss of I$ coherency. 2147 */ 2148 if (cache_boot_state & DCU_IC) { 2149 flush_icache(); 2150 } 2151 /* 2152 * According to section P.3.1 of the Panther PRM, we 2153 * need to do a little more for recovery on those 2154 * CPUs after encountering an I$ parity error. 2155 */ 2156 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 2157 flush_ipb(); 2158 correct_dcache_parity(dcache_size, 2159 dcache_linesize); 2160 flush_pcache(); 2161 } 2162 } else { 2163 /* 2164 * Since the valid bit is ignored when checking parity the 2165 * D$ data and tag must also be corrected. Set D$ data bits 2166 * to zero and set utag to 0, 1, 2, 3. 2167 */ 2168 correct_dcache_parity(dcache_size, dcache_linesize); 2169 2170 /* 2171 * According to section P.3.3 of the Panther PRM, we 2172 * need to do a little more for recovery on those 2173 * CPUs after encountering a D$ or P$ parity error. 2174 * 2175 * As far as clearing P$ parity errors, it is enough to 2176 * simply invalidate all entries in the P$ since P$ parity 2177 * error traps are only generated for floating point load 2178 * hits. 2179 */ 2180 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 2181 flush_icache(); 2182 flush_ipb(); 2183 flush_pcache(); 2184 } 2185 } 2186 2187 /* 2188 * Invalidate entire D$ if it was enabled. 2189 * This is done to avoid stale data in the D$ which might 2190 * occur with the D$ disabled and the trap handler doing 2191 * stores affecting lines already in the D$. 2192 */ 2193 if (cache_boot_state & DCU_DC) { 2194 flush_dcache(); 2195 } 2196 2197 /* 2198 * Restore caches to their bootup state. 2199 */ 2200 set_dcu(get_dcu() | cache_boot_state); 2201 2202 /* 2203 * Panic here if aflt->flt_panic has been set. Enqueued errors will 2204 * be logged as part of the panic flow. 2205 */ 2206 if (aflt->flt_panic) 2207 fm_panic("%sError(s)", iparity ? "IPE " : "DPE "); 2208 2209 /* 2210 * If this error occurred at TL>0 then flush the E$ here to reduce 2211 * the chance of getting an unrecoverable Fast ECC error. This 2212 * flush will evict the part of the parity trap handler that is run 2213 * at TL>1. 2214 */ 2215 if (tl) { 2216 cpu_flush_ecache(); 2217 } 2218 } 2219 2220 /* 2221 * On an I$ parity error, mark the appropriate entries in the ch_async_flt_t 2222 * to indicate which portions of the captured data should be in the ereport. 2223 */ 2224 void 2225 cpu_async_log_ic_parity_err(ch_async_flt_t *ch_flt) 2226 { 2227 int way = ch_flt->parity_data.ipe.cpl_way; 2228 int offset = ch_flt->parity_data.ipe.cpl_off; 2229 int tag_index; 2230 struct async_flt *aflt = (struct async_flt *)ch_flt; 2231 2232 2233 if ((offset != -1) || (way != -1)) { 2234 /* 2235 * Parity error in I$ tag or data 2236 */ 2237 tag_index = ch_flt->parity_data.ipe.cpl_ic[way].ic_idx; 2238 if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) 2239 ch_flt->parity_data.ipe.cpl_ic[way].ic_way = 2240 PN_ICIDX_TO_WAY(tag_index); 2241 else 2242 ch_flt->parity_data.ipe.cpl_ic[way].ic_way = 2243 CH_ICIDX_TO_WAY(tag_index); 2244 ch_flt->parity_data.ipe.cpl_ic[way].ic_logflag = 2245 IC_LOGFLAG_MAGIC; 2246 } else { 2247 /* 2248 * Parity error was not identified. 2249 * Log tags and data for all ways. 2250 */ 2251 for (way = 0; way < CH_ICACHE_NWAY; way++) { 2252 tag_index = ch_flt->parity_data.ipe.cpl_ic[way].ic_idx; 2253 if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) 2254 ch_flt->parity_data.ipe.cpl_ic[way].ic_way = 2255 PN_ICIDX_TO_WAY(tag_index); 2256 else 2257 ch_flt->parity_data.ipe.cpl_ic[way].ic_way = 2258 CH_ICIDX_TO_WAY(tag_index); 2259 ch_flt->parity_data.ipe.cpl_ic[way].ic_logflag = 2260 IC_LOGFLAG_MAGIC; 2261 } 2262 } 2263 } 2264 2265 /* 2266 * On an D$ parity error, mark the appropriate entries in the ch_async_flt_t 2267 * to indicate which portions of the captured data should be in the ereport. 2268 */ 2269 void 2270 cpu_async_log_dc_parity_err(ch_async_flt_t *ch_flt) 2271 { 2272 int way = ch_flt->parity_data.dpe.cpl_way; 2273 int offset = ch_flt->parity_data.dpe.cpl_off; 2274 int tag_index; 2275 2276 if (offset != -1) { 2277 /* 2278 * Parity error in D$ or P$ data array. 2279 * 2280 * First check to see whether the parity error is in D$ or P$ 2281 * since P$ data parity errors are reported in Panther using 2282 * the same trap. 2283 */ 2284 if (ch_flt->parity_data.dpe.cpl_cache == CPU_PC_PARITY) { 2285 tag_index = ch_flt->parity_data.dpe.cpl_pc[way].pc_idx; 2286 ch_flt->parity_data.dpe.cpl_pc[way].pc_way = 2287 CH_PCIDX_TO_WAY(tag_index); 2288 ch_flt->parity_data.dpe.cpl_pc[way].pc_logflag = 2289 PC_LOGFLAG_MAGIC; 2290 } else { 2291 tag_index = ch_flt->parity_data.dpe.cpl_dc[way].dc_idx; 2292 ch_flt->parity_data.dpe.cpl_dc[way].dc_way = 2293 CH_DCIDX_TO_WAY(tag_index); 2294 ch_flt->parity_data.dpe.cpl_dc[way].dc_logflag = 2295 DC_LOGFLAG_MAGIC; 2296 } 2297 } else if (way != -1) { 2298 /* 2299 * Parity error in D$ tag. 2300 */ 2301 tag_index = ch_flt->parity_data.dpe.cpl_dc[way].dc_idx; 2302 ch_flt->parity_data.dpe.cpl_dc[way].dc_way = 2303 CH_DCIDX_TO_WAY(tag_index); 2304 ch_flt->parity_data.dpe.cpl_dc[way].dc_logflag = 2305 DC_LOGFLAG_MAGIC; 2306 } 2307 } 2308 #endif /* CPU_IMP_L1_CACHE_PARITY */ 2309 2310 /* 2311 * The cpu_async_log_err() function is called via the [uc]e_drain() function to 2312 * post-process CPU events that are dequeued. As such, it can be invoked 2313 * from softint context, from AST processing in the trap() flow, or from the 2314 * panic flow. We decode the CPU-specific data, and take appropriate actions. 2315 * Historically this entry point was used to log the actual cmn_err(9F) text; 2316 * now with FMA it is used to prepare 'flt' to be converted into an ereport. 2317 * With FMA this function now also returns a flag which indicates to the 2318 * caller whether the ereport should be posted (1) or suppressed (0). 2319 */ 2320 static int 2321 cpu_async_log_err(void *flt, errorq_elem_t *eqep) 2322 { 2323 ch_async_flt_t *ch_flt = (ch_async_flt_t *)flt; 2324 struct async_flt *aflt = (struct async_flt *)flt; 2325 uint64_t errors; 2326 extern void memscrub_induced_error(void); 2327 2328 switch (ch_flt->flt_type) { 2329 case CPU_INV_AFSR: 2330 /* 2331 * If it is a disrupting trap and the AFSR is zero, then 2332 * the event has probably already been noted. Do not post 2333 * an ereport. 2334 */ 2335 if ((aflt->flt_status & ECC_C_TRAP) && 2336 (!(aflt->flt_stat & C_AFSR_MASK))) 2337 return (0); 2338 else 2339 return (1); 2340 case CPU_TO: 2341 case CPU_BERR: 2342 case CPU_FATAL: 2343 case CPU_FPUERR: 2344 return (1); 2345 2346 case CPU_UE_ECACHE_RETIRE: 2347 cpu_log_err(aflt); 2348 cpu_page_retire(ch_flt); 2349 return (1); 2350 2351 /* 2352 * Cases where we may want to suppress logging or perform 2353 * extended diagnostics. 2354 */ 2355 case CPU_CE: 2356 case CPU_EMC: 2357 /* 2358 * We want to skip logging and further classification 2359 * only if ALL the following conditions are true: 2360 * 2361 * 1. There is only one error 2362 * 2. That error is a correctable memory error 2363 * 3. The error is caused by the memory scrubber (in 2364 * which case the error will have occurred under 2365 * on_trap protection) 2366 * 4. The error is on a retired page 2367 * 2368 * Note: AFLT_PROT_EC is used places other than the memory 2369 * scrubber. However, none of those errors should occur 2370 * on a retired page. 2371 */ 2372 if ((ch_flt->afsr_errs & 2373 (C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS)) == C_AFSR_CE && 2374 aflt->flt_prot == AFLT_PROT_EC) { 2375 2376 if (page_retire_check(aflt->flt_addr, NULL) == 0) { 2377 if (ch_flt->flt_trapped_ce & CE_CEEN_DEFER) { 2378 2379 /* 2380 * Since we're skipping logging, we'll need 2381 * to schedule the re-enabling of CEEN 2382 */ 2383 (void) timeout(cpu_delayed_check_ce_errors, 2384 (void *)(uintptr_t)aflt->flt_inst, 2385 drv_usectohz((clock_t)cpu_ceen_delay_secs 2386 * MICROSEC)); 2387 } 2388 2389 /* 2390 * Inform memscrubber - scrubbing induced 2391 * CE on a retired page. 2392 */ 2393 memscrub_induced_error(); 2394 return (0); 2395 } 2396 } 2397 2398 /* 2399 * Perform/schedule further classification actions, but 2400 * only if the page is healthy (we don't want bad 2401 * pages inducing too much diagnostic activity). If we could 2402 * not find a page pointer then we also skip this. If 2403 * ce_scrub_xdiag_recirc returns nonzero then it has chosen 2404 * to copy and recirculate the event (for further diagnostics) 2405 * and we should not proceed to log it here. 2406 * 2407 * This must be the last step here before the cpu_log_err() 2408 * below - if an event recirculates cpu_ce_log_err() will 2409 * not call the current function but just proceed directly 2410 * to cpu_ereport_post after the cpu_log_err() avoided below. 2411 * 2412 * Note: Check cpu_impl_async_log_err if changing this 2413 */ 2414 if (page_retire_check(aflt->flt_addr, &errors) == EINVAL) { 2415 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, 2416 CE_XDIAG_SKIP_NOPP); 2417 } else { 2418 if (errors != PR_OK) { 2419 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, 2420 CE_XDIAG_SKIP_PAGEDET); 2421 } else if (ce_scrub_xdiag_recirc(aflt, ce_queue, eqep, 2422 offsetof(ch_async_flt_t, cmn_asyncflt))) { 2423 return (0); 2424 } 2425 } 2426 /*FALLTHRU*/ 2427 2428 /* 2429 * Cases where we just want to report the error and continue. 2430 */ 2431 case CPU_CE_ECACHE: 2432 case CPU_UE_ECACHE: 2433 case CPU_IV: 2434 case CPU_ORPH: 2435 cpu_log_err(aflt); 2436 return (1); 2437 2438 /* 2439 * Cases where we want to fall through to handle panicking. 2440 */ 2441 case CPU_UE: 2442 /* 2443 * We want to skip logging in the same conditions as the 2444 * CE case. In addition, we want to make sure we're not 2445 * panicking. 2446 */ 2447 if (!panicstr && (ch_flt->afsr_errs & 2448 (C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS)) == C_AFSR_UE && 2449 aflt->flt_prot == AFLT_PROT_EC) { 2450 if (page_retire_check(aflt->flt_addr, NULL) == 0) { 2451 /* Zero the address to clear the error */ 2452 softcall(ecc_page_zero, (void *)aflt->flt_addr); 2453 /* 2454 * Inform memscrubber - scrubbing induced 2455 * UE on a retired page. 2456 */ 2457 memscrub_induced_error(); 2458 return (0); 2459 } 2460 } 2461 cpu_log_err(aflt); 2462 break; 2463 2464 default: 2465 /* 2466 * If the us3_common.c code doesn't know the flt_type, it may 2467 * be an implementation-specific code. Call into the impldep 2468 * backend to find out what to do: if it tells us to continue, 2469 * break and handle as if falling through from a UE; if not, 2470 * the impldep backend has handled the error and we're done. 2471 */ 2472 switch (cpu_impl_async_log_err(flt, eqep)) { 2473 case CH_ASYNC_LOG_DONE: 2474 return (1); 2475 case CH_ASYNC_LOG_RECIRC: 2476 return (0); 2477 case CH_ASYNC_LOG_CONTINUE: 2478 break; /* continue on to handle UE-like error */ 2479 default: 2480 cmn_err(CE_WARN, "discarding error 0x%p with " 2481 "invalid fault type (0x%x)", 2482 (void *)aflt, ch_flt->flt_type); 2483 return (0); 2484 } 2485 } 2486 2487 /* ... fall through from the UE case */ 2488 2489 if (aflt->flt_addr != AFLT_INV_ADDR && aflt->flt_in_memory) { 2490 if (!panicstr) { 2491 cpu_page_retire(ch_flt); 2492 } else { 2493 /* 2494 * Clear UEs on panic so that we don't 2495 * get haunted by them during panic or 2496 * after reboot 2497 */ 2498 cpu_clearphys(aflt); 2499 (void) clear_errors(NULL); 2500 } 2501 } 2502 2503 return (1); 2504 } 2505 2506 /* 2507 * Retire the bad page that may contain the flushed error. 2508 */ 2509 void 2510 cpu_page_retire(ch_async_flt_t *ch_flt) 2511 { 2512 struct async_flt *aflt = (struct async_flt *)ch_flt; 2513 (void) page_retire(aflt->flt_addr, PR_UE); 2514 } 2515 2516 /* 2517 * Return true if the error specified in the AFSR indicates 2518 * an E$ data error (L2$ for Cheetah/Cheetah+/Jaguar, L3$ 2519 * for Panther, none for Jalapeno/Serrano). 2520 */ 2521 /* ARGSUSED */ 2522 static int 2523 cpu_error_is_ecache_data(int cpuid, uint64_t t_afsr) 2524 { 2525 #if defined(JALAPENO) || defined(SERRANO) 2526 return (0); 2527 #elif defined(CHEETAH_PLUS) 2528 if (IS_PANTHER(cpunodes[cpuid].implementation)) 2529 return ((t_afsr & C_AFSR_EXT_L3_DATA_ERRS) != 0); 2530 return ((t_afsr & C_AFSR_EC_DATA_ERRS) != 0); 2531 #else /* CHEETAH_PLUS */ 2532 return ((t_afsr & C_AFSR_EC_DATA_ERRS) != 0); 2533 #endif 2534 } 2535 2536 /* 2537 * The cpu_log_err() function is called by cpu_async_log_err() to perform the 2538 * generic event post-processing for correctable and uncorrectable memory, 2539 * E$, and MTag errors. Historically this entry point was used to log bits of 2540 * common cmn_err(9F) text; now with FMA it is used to prepare 'flt' to be 2541 * converted into an ereport. In addition, it transmits the error to any 2542 * platform-specific service-processor FRU logging routines, if available. 2543 */ 2544 void 2545 cpu_log_err(struct async_flt *aflt) 2546 { 2547 char unum[UNUM_NAMLEN]; 2548 int synd_status, synd_code, afar_status; 2549 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 2550 2551 if (cpu_error_is_ecache_data(aflt->flt_inst, ch_flt->flt_bit)) 2552 aflt->flt_status |= ECC_ECACHE; 2553 else 2554 aflt->flt_status &= ~ECC_ECACHE; 2555 /* 2556 * Determine syndrome status. 2557 */ 2558 synd_status = afsr_to_synd_status(aflt->flt_inst, 2559 ch_flt->afsr_errs, ch_flt->flt_bit); 2560 2561 /* 2562 * Determine afar status. 2563 */ 2564 if (pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT)) 2565 afar_status = afsr_to_afar_status(ch_flt->afsr_errs, 2566 ch_flt->flt_bit); 2567 else 2568 afar_status = AFLT_STAT_INVALID; 2569 2570 synd_code = synd_to_synd_code(synd_status, 2571 aflt->flt_synd, ch_flt->flt_bit); 2572 2573 /* 2574 * If afar status is not invalid do a unum lookup. 2575 */ 2576 if (afar_status != AFLT_STAT_INVALID) { 2577 (void) cpu_get_mem_unum_synd(synd_code, aflt, unum); 2578 } else { 2579 unum[0] = '\0'; 2580 } 2581 2582 /* 2583 * Do not send the fruid message (plat_ecc_error_data_t) 2584 * to the SC if it can handle the enhanced error information 2585 * (plat_ecc_error2_data_t) or when the tunable 2586 * ecc_log_fruid_enable is set to 0. 2587 */ 2588 2589 if (&plat_ecc_capability_sc_get && 2590 plat_ecc_capability_sc_get(PLAT_ECC_ERROR_MESSAGE)) { 2591 if (&plat_log_fruid_error) 2592 plat_log_fruid_error(synd_code, aflt, unum, 2593 ch_flt->flt_bit); 2594 } 2595 2596 if (aflt->flt_func != NULL) 2597 aflt->flt_func(aflt, unum); 2598 2599 if (afar_status != AFLT_STAT_INVALID) 2600 cpu_log_diag_info(ch_flt); 2601 2602 /* 2603 * If we have a CEEN error , we do not reenable CEEN until after 2604 * we exit the trap handler. Otherwise, another error may 2605 * occur causing the handler to be entered recursively. 2606 * We set a timeout to trigger in cpu_ceen_delay_secs seconds, 2607 * to try and ensure that the CPU makes progress in the face 2608 * of a CE storm. 2609 */ 2610 if (ch_flt->flt_trapped_ce & CE_CEEN_DEFER) { 2611 (void) timeout(cpu_delayed_check_ce_errors, 2612 (void *)(uintptr_t)aflt->flt_inst, 2613 drv_usectohz((clock_t)cpu_ceen_delay_secs * MICROSEC)); 2614 } 2615 } 2616 2617 /* 2618 * Invoked by error_init() early in startup and therefore before 2619 * startup_errorq() is called to drain any error Q - 2620 * 2621 * startup() 2622 * startup_end() 2623 * error_init() 2624 * cpu_error_init() 2625 * errorq_init() 2626 * errorq_drain() 2627 * start_other_cpus() 2628 * 2629 * The purpose of this routine is to create error-related taskqs. Taskqs 2630 * are used for this purpose because cpu_lock can't be grabbed from interrupt 2631 * context. 2632 */ 2633 void 2634 cpu_error_init(int items) 2635 { 2636 /* 2637 * Create taskq(s) to reenable CE 2638 */ 2639 ch_check_ce_tq = taskq_create("cheetah_check_ce", 1, minclsyspri, 2640 items, items, TASKQ_PREPOPULATE); 2641 } 2642 2643 void 2644 cpu_ce_log_err(struct async_flt *aflt, errorq_elem_t *eqep) 2645 { 2646 char unum[UNUM_NAMLEN]; 2647 int len; 2648 2649 switch (aflt->flt_class) { 2650 case CPU_FAULT: 2651 cpu_ereport_init(aflt); 2652 if (cpu_async_log_err(aflt, eqep)) 2653 cpu_ereport_post(aflt); 2654 break; 2655 2656 case BUS_FAULT: 2657 if (aflt->flt_func != NULL) { 2658 (void) cpu_get_mem_unum_aflt(AFLT_STAT_VALID, aflt, 2659 unum, UNUM_NAMLEN, &len); 2660 aflt->flt_func(aflt, unum); 2661 } 2662 break; 2663 2664 case RECIRC_CPU_FAULT: 2665 aflt->flt_class = CPU_FAULT; 2666 cpu_log_err(aflt); 2667 cpu_ereport_post(aflt); 2668 break; 2669 2670 case RECIRC_BUS_FAULT: 2671 ASSERT(aflt->flt_class != RECIRC_BUS_FAULT); 2672 /*FALLTHRU*/ 2673 default: 2674 cmn_err(CE_WARN, "discarding CE error 0x%p with invalid " 2675 "fault class (0x%x)", (void *)aflt, aflt->flt_class); 2676 return; 2677 } 2678 } 2679 2680 /* 2681 * Scrub and classify a CE. This function must not modify the 2682 * fault structure passed to it but instead should return the classification 2683 * information. 2684 */ 2685 2686 static uchar_t 2687 cpu_ce_scrub_mem_err_common(struct async_flt *ecc, boolean_t logout_tried) 2688 { 2689 uchar_t disp = CE_XDIAG_EXTALG; 2690 on_trap_data_t otd; 2691 uint64_t orig_err; 2692 ch_cpu_logout_t *clop; 2693 2694 /* 2695 * Clear CEEN. CPU CE TL > 0 trap handling will already have done 2696 * this, but our other callers have not. Disable preemption to 2697 * avoid CPU migration so that we restore CEEN on the correct 2698 * cpu later. 2699 * 2700 * CEEN is cleared so that further CEs that our instruction and 2701 * data footprint induce do not cause use to either creep down 2702 * kernel stack to the point of overflow, or do so much CE 2703 * notification as to make little real forward progress. 2704 * 2705 * NCEEN must not be cleared. However it is possible that 2706 * our accesses to the flt_addr may provoke a bus error or timeout 2707 * if the offending address has just been unconfigured as part of 2708 * a DR action. So we must operate under on_trap protection. 2709 */ 2710 kpreempt_disable(); 2711 orig_err = get_error_enable(); 2712 if (orig_err & EN_REG_CEEN) 2713 set_error_enable(orig_err & ~EN_REG_CEEN); 2714 2715 /* 2716 * Our classification algorithm includes the line state before 2717 * the scrub; we'd like this captured after the detection and 2718 * before the algorithm below - the earlier the better. 2719 * 2720 * If we've come from a cpu CE trap then this info already exists 2721 * in the cpu logout area. 2722 * 2723 * For a CE detected by memscrub for which there was no trap 2724 * (running with CEEN off) cpu_log_and_clear_ce has called 2725 * cpu_ce_delayed_ec_logout to capture some cache data, and 2726 * marked the fault structure as incomplete as a flag to later 2727 * logging code. 2728 * 2729 * If called directly from an IO detected CE there has been 2730 * no line data capture. In this case we logout to the cpu logout 2731 * area - that's appropriate since it's the cpu cache data we need 2732 * for classification. We thus borrow the cpu logout area for a 2733 * short time, and cpu_ce_delayed_ec_logout will mark it as busy in 2734 * this time (we will invalidate it again below). 2735 * 2736 * If called from the partner check xcall handler then this cpu 2737 * (the partner) has not necessarily experienced a CE at this 2738 * address. But we want to capture line state before its scrub 2739 * attempt since we use that in our classification. 2740 */ 2741 if (logout_tried == B_FALSE) { 2742 if (!cpu_ce_delayed_ec_logout(ecc->flt_addr)) 2743 disp |= CE_XDIAG_NOLOGOUT; 2744 } 2745 2746 /* 2747 * Scrub memory, then check AFSR for errors. The AFAR we scrub may 2748 * no longer be valid (if DR'd since the initial event) so we 2749 * perform this scrub under on_trap protection. If this access is 2750 * ok then further accesses below will also be ok - DR cannot 2751 * proceed while this thread is active (preemption is disabled); 2752 * to be safe we'll nonetheless use on_trap again below. 2753 */ 2754 if (!on_trap(&otd, OT_DATA_ACCESS)) { 2755 cpu_scrubphys(ecc); 2756 } else { 2757 no_trap(); 2758 if (orig_err & EN_REG_CEEN) 2759 set_error_enable(orig_err); 2760 kpreempt_enable(); 2761 return (disp); 2762 } 2763 no_trap(); 2764 2765 /* 2766 * Did the casx read of the scrub log a CE that matches the AFAR? 2767 * Note that it's quite possible that the read sourced the data from 2768 * another cpu. 2769 */ 2770 if (clear_ecc(ecc)) 2771 disp |= CE_XDIAG_CE1; 2772 2773 /* 2774 * Read the data again. This time the read is very likely to 2775 * come from memory since the scrub induced a writeback to memory. 2776 */ 2777 if (!on_trap(&otd, OT_DATA_ACCESS)) { 2778 (void) lddphys(P2ALIGN(ecc->flt_addr, 8)); 2779 } else { 2780 no_trap(); 2781 if (orig_err & EN_REG_CEEN) 2782 set_error_enable(orig_err); 2783 kpreempt_enable(); 2784 return (disp); 2785 } 2786 no_trap(); 2787 2788 /* Did that read induce a CE that matches the AFAR? */ 2789 if (clear_ecc(ecc)) 2790 disp |= CE_XDIAG_CE2; 2791 2792 /* 2793 * Look at the logout information and record whether we found the 2794 * line in l2/l3 cache. For Panther we are interested in whether 2795 * we found it in either cache (it won't reside in both but 2796 * it is possible to read it that way given the moving target). 2797 */ 2798 clop = CPU_PRIVATE(CPU) ? CPU_PRIVATE_PTR(CPU, chpr_cecc_logout) : NULL; 2799 if (!(disp & CE_XDIAG_NOLOGOUT) && clop && 2800 clop->clo_data.chd_afar != LOGOUT_INVALID) { 2801 int hit, level; 2802 int state; 2803 int totalsize; 2804 ch_ec_data_t *ecp; 2805 2806 /* 2807 * If hit is nonzero then a match was found and hit will 2808 * be one greater than the index which hit. For Panther we 2809 * also need to pay attention to level to see which of l2$ or 2810 * l3$ it hit in. 2811 */ 2812 hit = cpu_matching_ecache_line(ecc->flt_addr, &clop->clo_data, 2813 0, &level); 2814 2815 if (hit) { 2816 --hit; 2817 disp |= CE_XDIAG_AFARMATCH; 2818 2819 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 2820 if (level == 2) 2821 ecp = &clop->clo_data.chd_l2_data[hit]; 2822 else 2823 ecp = &clop->clo_data.chd_ec_data[hit]; 2824 } else { 2825 ASSERT(level == 2); 2826 ecp = &clop->clo_data.chd_ec_data[hit]; 2827 } 2828 totalsize = cpunodes[CPU->cpu_id].ecache_size; 2829 state = cpu_ectag_pa_to_subblk_state(totalsize, 2830 ecc->flt_addr, ecp->ec_tag); 2831 2832 /* 2833 * Cheetah variants use different state encodings - 2834 * the CH_ECSTATE_* defines vary depending on the 2835 * module we're compiled for. Translate into our 2836 * one true version. Conflate Owner-Shared state 2837 * of SSM mode with Owner as victimisation of such 2838 * lines may cause a writeback. 2839 */ 2840 switch (state) { 2841 case CH_ECSTATE_MOD: 2842 disp |= EC_STATE_M; 2843 break; 2844 2845 case CH_ECSTATE_OWN: 2846 case CH_ECSTATE_OWS: 2847 disp |= EC_STATE_O; 2848 break; 2849 2850 case CH_ECSTATE_EXL: 2851 disp |= EC_STATE_E; 2852 break; 2853 2854 case CH_ECSTATE_SHR: 2855 disp |= EC_STATE_S; 2856 break; 2857 2858 default: 2859 disp |= EC_STATE_I; 2860 break; 2861 } 2862 } 2863 2864 /* 2865 * If we initiated the delayed logout then we are responsible 2866 * for invalidating the logout area. 2867 */ 2868 if (logout_tried == B_FALSE) { 2869 bzero(clop, sizeof (ch_cpu_logout_t)); 2870 clop->clo_data.chd_afar = LOGOUT_INVALID; 2871 } 2872 } 2873 2874 /* 2875 * Re-enable CEEN if we turned it off. 2876 */ 2877 if (orig_err & EN_REG_CEEN) 2878 set_error_enable(orig_err); 2879 kpreempt_enable(); 2880 2881 return (disp); 2882 } 2883 2884 /* 2885 * Scrub a correctable memory error and collect data for classification 2886 * of CE type. This function is called in the detection path, ie tl0 handling 2887 * of a correctable error trap (cpus) or interrupt (IO) at high PIL. 2888 */ 2889 void 2890 cpu_ce_scrub_mem_err(struct async_flt *ecc, boolean_t logout_tried) 2891 { 2892 /* 2893 * Cheetah CE classification does not set any bits in flt_status. 2894 * Instead we will record classification datapoints in flt_disp. 2895 */ 2896 ecc->flt_status &= ~(ECC_INTERMITTENT | ECC_PERSISTENT | ECC_STICKY); 2897 2898 /* 2899 * To check if the error detected by IO is persistent, sticky or 2900 * intermittent. This is noticed by clear_ecc(). 2901 */ 2902 if (ecc->flt_status & ECC_IOBUS) 2903 ecc->flt_stat = C_AFSR_MEMORY; 2904 2905 /* 2906 * Record information from this first part of the algorithm in 2907 * flt_disp. 2908 */ 2909 ecc->flt_disp = cpu_ce_scrub_mem_err_common(ecc, logout_tried); 2910 } 2911 2912 /* 2913 * Select a partner to perform a further CE classification check from. 2914 * Must be called with kernel preemption disabled (to stop the cpu list 2915 * from changing). The detecting cpu we are partnering has cpuid 2916 * aflt->flt_inst; we might not be running on the detecting cpu. 2917 * 2918 * Restrict choice to active cpus in the same cpu partition as ourselves in 2919 * an effort to stop bad cpus in one partition causing other partitions to 2920 * perform excessive diagnostic activity. Actually since the errorq drain 2921 * is run from a softint most of the time and that is a global mechanism 2922 * this isolation is only partial. Return NULL if we fail to find a 2923 * suitable partner. 2924 * 2925 * We prefer a partner that is in a different latency group to ourselves as 2926 * we will share fewer datapaths. If such a partner is unavailable then 2927 * choose one in the same lgroup but prefer a different chip and only allow 2928 * a sibling core if flags includes PTNR_SIBLINGOK. If all else fails and 2929 * flags includes PTNR_SELFOK then permit selection of the original detector. 2930 * 2931 * We keep a cache of the last partner selected for a cpu, and we'll try to 2932 * use that previous partner if no more than cpu_ce_ptnr_cachetime_sec seconds 2933 * have passed since that selection was made. This provides the benefit 2934 * of the point-of-view of different partners over time but without 2935 * requiring frequent cpu list traversals. 2936 */ 2937 2938 #define PTNR_SIBLINGOK 0x1 /* Allow selection of sibling core */ 2939 #define PTNR_SELFOK 0x2 /* Allow selection of cpu to "partner" itself */ 2940 2941 static cpu_t * 2942 ce_ptnr_select(struct async_flt *aflt, int flags, int *typep) 2943 { 2944 cpu_t *sp, *dtcr, *ptnr, *locptnr, *sibptnr; 2945 hrtime_t lasttime, thistime; 2946 2947 ASSERT(curthread->t_preempt > 0 || getpil() >= DISP_LEVEL); 2948 2949 dtcr = cpu[aflt->flt_inst]; 2950 2951 /* 2952 * Short-circuit for the following cases: 2953 * . the dtcr is not flagged active 2954 * . there is just one cpu present 2955 * . the detector has disappeared 2956 * . we were given a bad flt_inst cpuid; this should not happen 2957 * (eg PCI code now fills flt_inst) but if it does it is no 2958 * reason to panic. 2959 * . there is just one cpu left online in the cpu partition 2960 * 2961 * If we return NULL after this point then we do not update the 2962 * chpr_ceptnr_seltime which will cause us to perform a full lookup 2963 * again next time; this is the case where the only other cpu online 2964 * in the detector's partition is on the same chip as the detector 2965 * and since CEEN re-enable is throttled even that case should not 2966 * hurt performance. 2967 */ 2968 if (dtcr == NULL || !cpu_flagged_active(dtcr->cpu_flags)) { 2969 return (NULL); 2970 } 2971 if (ncpus == 1 || dtcr->cpu_part->cp_ncpus == 1) { 2972 if (flags & PTNR_SELFOK) { 2973 *typep = CE_XDIAG_PTNR_SELF; 2974 return (dtcr); 2975 } else { 2976 return (NULL); 2977 } 2978 } 2979 2980 thistime = gethrtime(); 2981 lasttime = CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime); 2982 2983 /* 2984 * Select a starting point. 2985 */ 2986 if (!lasttime) { 2987 /* 2988 * We've never selected a partner for this detector before. 2989 * Start the scan at the next online cpu in the same cpu 2990 * partition. 2991 */ 2992 sp = dtcr->cpu_next_part; 2993 } else if (thistime - lasttime < cpu_ce_ptnr_cachetime_sec * NANOSEC) { 2994 /* 2995 * Our last selection has not aged yet. If this partner: 2996 * . is still a valid cpu, 2997 * . is still in the same partition as the detector 2998 * . is still marked active 2999 * . satisfies the 'flags' argument criteria 3000 * then select it again without updating the timestamp. 3001 */ 3002 sp = cpu[CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id)]; 3003 if (sp == NULL || sp->cpu_part != dtcr->cpu_part || 3004 !cpu_flagged_active(sp->cpu_flags) || 3005 (sp == dtcr && !(flags & PTNR_SELFOK)) || 3006 (pg_plat_cpus_share(sp, dtcr, PGHW_CHIP) && 3007 !(flags & PTNR_SIBLINGOK))) { 3008 sp = dtcr->cpu_next_part; 3009 } else { 3010 if (sp->cpu_lpl->lpl_lgrp != dtcr->cpu_lpl->lpl_lgrp) { 3011 *typep = CE_XDIAG_PTNR_REMOTE; 3012 } else if (sp == dtcr) { 3013 *typep = CE_XDIAG_PTNR_SELF; 3014 } else if (pg_plat_cpus_share(sp, dtcr, PGHW_CHIP)) { 3015 *typep = CE_XDIAG_PTNR_SIBLING; 3016 } else { 3017 *typep = CE_XDIAG_PTNR_LOCAL; 3018 } 3019 return (sp); 3020 } 3021 } else { 3022 /* 3023 * Our last selection has aged. If it is nonetheless still a 3024 * valid cpu then start the scan at the next cpu in the 3025 * partition after our last partner. If the last selection 3026 * is no longer a valid cpu then go with our default. In 3027 * this way we slowly cycle through possible partners to 3028 * obtain multiple viewpoints over time. 3029 */ 3030 sp = cpu[CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id)]; 3031 if (sp == NULL) { 3032 sp = dtcr->cpu_next_part; 3033 } else { 3034 sp = sp->cpu_next_part; /* may be dtcr */ 3035 if (sp->cpu_part != dtcr->cpu_part) 3036 sp = dtcr; 3037 } 3038 } 3039 3040 /* 3041 * We have a proposed starting point for our search, but if this 3042 * cpu is offline then its cpu_next_part will point to itself 3043 * so we can't use that to iterate over cpus in this partition in 3044 * the loop below. We still want to avoid iterating over cpus not 3045 * in our partition, so in the case that our starting point is offline 3046 * we will repoint it to be the detector itself; and if the detector 3047 * happens to be offline we'll return NULL from the following loop. 3048 */ 3049 if (!cpu_flagged_active(sp->cpu_flags)) { 3050 sp = dtcr; 3051 } 3052 3053 ptnr = sp; 3054 locptnr = NULL; 3055 sibptnr = NULL; 3056 do { 3057 if (ptnr == dtcr || !cpu_flagged_active(ptnr->cpu_flags)) 3058 continue; 3059 if (ptnr->cpu_lpl->lpl_lgrp != dtcr->cpu_lpl->lpl_lgrp) { 3060 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = ptnr->cpu_id; 3061 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime; 3062 *typep = CE_XDIAG_PTNR_REMOTE; 3063 return (ptnr); 3064 } 3065 if (pg_plat_cpus_share(ptnr, dtcr, PGHW_CHIP)) { 3066 if (sibptnr == NULL) 3067 sibptnr = ptnr; 3068 continue; 3069 } 3070 if (locptnr == NULL) 3071 locptnr = ptnr; 3072 } while ((ptnr = ptnr->cpu_next_part) != sp); 3073 3074 /* 3075 * A foreign partner has already been returned if one was available. 3076 * 3077 * If locptnr is not NULL it is a cpu in the same lgroup as the 3078 * detector, is active, and is not a sibling of the detector. 3079 * 3080 * If sibptnr is not NULL it is a sibling of the detector, and is 3081 * active. 3082 * 3083 * If we have to resort to using the detector itself we have already 3084 * checked that it is active. 3085 */ 3086 if (locptnr) { 3087 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = locptnr->cpu_id; 3088 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime; 3089 *typep = CE_XDIAG_PTNR_LOCAL; 3090 return (locptnr); 3091 } else if (sibptnr && flags & PTNR_SIBLINGOK) { 3092 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = sibptnr->cpu_id; 3093 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime; 3094 *typep = CE_XDIAG_PTNR_SIBLING; 3095 return (sibptnr); 3096 } else if (flags & PTNR_SELFOK) { 3097 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_id) = dtcr->cpu_id; 3098 CPU_PRIVATE_VAL(dtcr, chpr_ceptnr_seltime) = thistime; 3099 *typep = CE_XDIAG_PTNR_SELF; 3100 return (dtcr); 3101 } 3102 3103 return (NULL); 3104 } 3105 3106 /* 3107 * Cross call handler that is requested to run on the designated partner of 3108 * a cpu that experienced a possibly sticky or possibly persistnet CE. 3109 */ 3110 static void 3111 ce_ptnrchk_xc(struct async_flt *aflt, uchar_t *dispp) 3112 { 3113 *dispp = cpu_ce_scrub_mem_err_common(aflt, B_FALSE); 3114 } 3115 3116 /* 3117 * The associated errorqs are never destroyed so we do not need to deal with 3118 * them disappearing before this timeout fires. If the affected memory 3119 * has been DR'd out since the original event the scrub algrithm will catch 3120 * any errors and return null disposition info. If the original detecting 3121 * cpu has been DR'd out then ereport detector info will not be able to 3122 * lookup CPU type; with a small timeout this is unlikely. 3123 */ 3124 static void 3125 ce_lkychk_cb(ce_lkychk_cb_t *cbarg) 3126 { 3127 struct async_flt *aflt = cbarg->lkycb_aflt; 3128 uchar_t disp; 3129 cpu_t *cp; 3130 int ptnrtype; 3131 3132 kpreempt_disable(); 3133 if (cp = ce_ptnr_select(aflt, PTNR_SIBLINGOK | PTNR_SELFOK, 3134 &ptnrtype)) { 3135 xc_one(cp->cpu_id, (xcfunc_t *)ce_ptnrchk_xc, (uint64_t)aflt, 3136 (uint64_t)&disp); 3137 CE_XDIAG_SETLKYINFO(aflt->flt_disp, disp); 3138 CE_XDIAG_SETPTNRID(aflt->flt_disp, cp->cpu_id); 3139 CE_XDIAG_SETPTNRTYPE(aflt->flt_disp, ptnrtype); 3140 } else { 3141 ce_xdiag_lkydrops++; 3142 if (ncpus > 1) 3143 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, 3144 CE_XDIAG_SKIP_NOPTNR); 3145 } 3146 kpreempt_enable(); 3147 3148 errorq_commit(cbarg->lkycb_eqp, cbarg->lkycb_eqep, ERRORQ_ASYNC); 3149 kmem_free(cbarg, sizeof (ce_lkychk_cb_t)); 3150 } 3151 3152 /* 3153 * Called from errorq drain code when processing a CE error, both from 3154 * CPU and PCI drain functions. Decide what further classification actions, 3155 * if any, we will perform. Perform immediate actions now, and schedule 3156 * delayed actions as required. Note that we are no longer necessarily running 3157 * on the detecting cpu, and that the async_flt structure will not persist on 3158 * return from this function. 3159 * 3160 * Calls to this function should aim to be self-throtlling in some way. With 3161 * the delayed re-enable of CEEN the absolute rate of calls should not 3162 * be excessive. Callers should also avoid performing in-depth classification 3163 * for events in pages that are already known to be suspect. 3164 * 3165 * We return nonzero to indicate that the event has been copied and 3166 * recirculated for further testing. The caller should not log the event 3167 * in this case - it will be logged when further test results are available. 3168 * 3169 * Our possible contexts are that of errorq_drain: below lock level or from 3170 * panic context. We can assume that the cpu we are running on is online. 3171 */ 3172 3173 3174 #ifdef DEBUG 3175 static int ce_xdiag_forceaction; 3176 #endif 3177 3178 int 3179 ce_scrub_xdiag_recirc(struct async_flt *aflt, errorq_t *eqp, 3180 errorq_elem_t *eqep, size_t afltoffset) 3181 { 3182 ce_dispact_t dispact, action; 3183 cpu_t *cp; 3184 uchar_t dtcrinfo, disp; 3185 int ptnrtype; 3186 3187 if (!ce_disp_inited || panicstr || ce_xdiag_off) { 3188 ce_xdiag_drops++; 3189 return (0); 3190 } else if (!aflt->flt_in_memory) { 3191 ce_xdiag_drops++; 3192 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, CE_XDIAG_SKIP_NOTMEM); 3193 return (0); 3194 } 3195 3196 dtcrinfo = CE_XDIAG_DTCRINFO(aflt->flt_disp); 3197 3198 /* 3199 * Some correctable events are not scrubbed/classified, such as those 3200 * noticed at the tail of cpu_deferred_error. So if there is no 3201 * initial detector classification go no further. 3202 */ 3203 if (!CE_XDIAG_EXT_ALG_APPLIED(dtcrinfo)) { 3204 ce_xdiag_drops++; 3205 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, CE_XDIAG_SKIP_NOSCRUB); 3206 return (0); 3207 } 3208 3209 dispact = CE_DISPACT(ce_disp_table, 3210 CE_XDIAG_AFARMATCHED(dtcrinfo), 3211 CE_XDIAG_STATE(dtcrinfo), 3212 CE_XDIAG_CE1SEEN(dtcrinfo), 3213 CE_XDIAG_CE2SEEN(dtcrinfo)); 3214 3215 3216 action = CE_ACT(dispact); /* bad lookup caught below */ 3217 #ifdef DEBUG 3218 if (ce_xdiag_forceaction != 0) 3219 action = ce_xdiag_forceaction; 3220 #endif 3221 3222 switch (action) { 3223 case CE_ACT_LKYCHK: { 3224 caddr_t ndata; 3225 errorq_elem_t *neqep; 3226 struct async_flt *ecc; 3227 ce_lkychk_cb_t *cbargp; 3228 3229 if ((ndata = errorq_elem_dup(eqp, eqep, &neqep)) == NULL) { 3230 ce_xdiag_lkydrops++; 3231 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, 3232 CE_XDIAG_SKIP_DUPFAIL); 3233 break; 3234 } 3235 ecc = (struct async_flt *)(ndata + afltoffset); 3236 3237 ASSERT(ecc->flt_class == CPU_FAULT || 3238 ecc->flt_class == BUS_FAULT); 3239 ecc->flt_class = (ecc->flt_class == CPU_FAULT) ? 3240 RECIRC_CPU_FAULT : RECIRC_BUS_FAULT; 3241 3242 cbargp = kmem_alloc(sizeof (ce_lkychk_cb_t), KM_SLEEP); 3243 cbargp->lkycb_aflt = ecc; 3244 cbargp->lkycb_eqp = eqp; 3245 cbargp->lkycb_eqep = neqep; 3246 3247 (void) timeout((void (*)(void *))ce_lkychk_cb, 3248 (void *)cbargp, drv_usectohz(cpu_ce_lkychk_timeout_usec)); 3249 return (1); 3250 } 3251 3252 case CE_ACT_PTNRCHK: 3253 kpreempt_disable(); /* stop cpu list changing */ 3254 if ((cp = ce_ptnr_select(aflt, 0, &ptnrtype)) != NULL) { 3255 xc_one(cp->cpu_id, (xcfunc_t *)ce_ptnrchk_xc, 3256 (uint64_t)aflt, (uint64_t)&disp); 3257 CE_XDIAG_SETPTNRINFO(aflt->flt_disp, disp); 3258 CE_XDIAG_SETPTNRID(aflt->flt_disp, cp->cpu_id); 3259 CE_XDIAG_SETPTNRTYPE(aflt->flt_disp, ptnrtype); 3260 } else if (ncpus > 1) { 3261 ce_xdiag_ptnrdrops++; 3262 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, 3263 CE_XDIAG_SKIP_NOPTNR); 3264 } else { 3265 ce_xdiag_ptnrdrops++; 3266 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, 3267 CE_XDIAG_SKIP_UNIPROC); 3268 } 3269 kpreempt_enable(); 3270 break; 3271 3272 case CE_ACT_DONE: 3273 break; 3274 3275 case CE_ACT(CE_DISP_BAD): 3276 default: 3277 #ifdef DEBUG 3278 cmn_err(CE_PANIC, "ce_scrub_post: Bad action '%d'", action); 3279 #endif 3280 ce_xdiag_bad++; 3281 CE_XDIAG_SETSKIPCODE(aflt->flt_disp, CE_XDIAG_SKIP_ACTBAD); 3282 break; 3283 } 3284 3285 return (0); 3286 } 3287 3288 /* 3289 * We route all errors through a single switch statement. 3290 */ 3291 void 3292 cpu_ue_log_err(struct async_flt *aflt) 3293 { 3294 switch (aflt->flt_class) { 3295 case CPU_FAULT: 3296 cpu_ereport_init(aflt); 3297 if (cpu_async_log_err(aflt, NULL)) 3298 cpu_ereport_post(aflt); 3299 break; 3300 3301 case BUS_FAULT: 3302 bus_async_log_err(aflt); 3303 break; 3304 3305 default: 3306 cmn_err(CE_WARN, "discarding async error %p with invalid " 3307 "fault class (0x%x)", (void *)aflt, aflt->flt_class); 3308 return; 3309 } 3310 } 3311 3312 /* 3313 * Routine for panic hook callback from panic_idle(). 3314 */ 3315 void 3316 cpu_async_panic_callb(void) 3317 { 3318 ch_async_flt_t ch_flt; 3319 struct async_flt *aflt; 3320 ch_cpu_errors_t cpu_error_regs; 3321 uint64_t afsr_errs; 3322 3323 get_cpu_error_state(&cpu_error_regs); 3324 3325 afsr_errs = (cpu_error_regs.afsr & C_AFSR_ALL_ERRS) | 3326 (cpu_error_regs.afsr_ext & C_AFSR_EXT_ALL_ERRS); 3327 3328 if (afsr_errs) { 3329 3330 bzero(&ch_flt, sizeof (ch_async_flt_t)); 3331 aflt = (struct async_flt *)&ch_flt; 3332 aflt->flt_id = gethrtime_waitfree(); 3333 aflt->flt_bus_id = getprocessorid(); 3334 aflt->flt_inst = CPU->cpu_id; 3335 aflt->flt_stat = cpu_error_regs.afsr; 3336 aflt->flt_addr = cpu_error_regs.afar; 3337 aflt->flt_prot = AFLT_PROT_NONE; 3338 aflt->flt_class = CPU_FAULT; 3339 aflt->flt_priv = ((cpu_error_regs.afsr & C_AFSR_PRIV) != 0); 3340 aflt->flt_panic = 1; 3341 ch_flt.afsr_ext = cpu_error_regs.afsr_ext; 3342 ch_flt.afsr_errs = afsr_errs; 3343 #if defined(SERRANO) 3344 ch_flt.afar2 = cpu_error_regs.afar2; 3345 #endif /* SERRANO */ 3346 (void) cpu_queue_events(&ch_flt, NULL, afsr_errs, NULL); 3347 } 3348 } 3349 3350 /* 3351 * Routine to convert a syndrome into a syndrome code. 3352 */ 3353 static int 3354 synd_to_synd_code(int synd_status, ushort_t synd, uint64_t afsr_bit) 3355 { 3356 if (synd_status == AFLT_STAT_INVALID) 3357 return (-1); 3358 3359 /* 3360 * Use the syndrome to index the appropriate syndrome table, 3361 * to get the code indicating which bit(s) is(are) bad. 3362 */ 3363 if (afsr_bit & 3364 (C_AFSR_MSYND_ERRS | C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) { 3365 if (afsr_bit & C_AFSR_MSYND_ERRS) { 3366 #if defined(JALAPENO) || defined(SERRANO) 3367 if ((synd == 0) || (synd >= BSYND_TBL_SIZE)) 3368 return (-1); 3369 else 3370 return (BPAR0 + synd); 3371 #else /* JALAPENO || SERRANO */ 3372 if ((synd == 0) || (synd >= MSYND_TBL_SIZE)) 3373 return (-1); 3374 else 3375 return (mtag_syndrome_tab[synd]); 3376 #endif /* JALAPENO || SERRANO */ 3377 } else { 3378 if ((synd == 0) || (synd >= ESYND_TBL_SIZE)) 3379 return (-1); 3380 else 3381 return (ecc_syndrome_tab[synd]); 3382 } 3383 } else { 3384 return (-1); 3385 } 3386 } 3387 3388 int 3389 cpu_get_mem_sid(char *unum, char *buf, int buflen, int *lenp) 3390 { 3391 if (&plat_get_mem_sid) 3392 return (plat_get_mem_sid(unum, buf, buflen, lenp)); 3393 else 3394 return (ENOTSUP); 3395 } 3396 3397 int 3398 cpu_get_mem_offset(uint64_t flt_addr, uint64_t *offp) 3399 { 3400 if (&plat_get_mem_offset) 3401 return (plat_get_mem_offset(flt_addr, offp)); 3402 else 3403 return (ENOTSUP); 3404 } 3405 3406 int 3407 cpu_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp) 3408 { 3409 if (&plat_get_mem_addr) 3410 return (plat_get_mem_addr(unum, sid, offset, addrp)); 3411 else 3412 return (ENOTSUP); 3413 } 3414 3415 /* 3416 * Routine to return a string identifying the physical name 3417 * associated with a memory/cache error. 3418 */ 3419 int 3420 cpu_get_mem_unum(int synd_status, ushort_t flt_synd, uint64_t flt_stat, 3421 uint64_t flt_addr, int flt_bus_id, int flt_in_memory, 3422 ushort_t flt_status, char *buf, int buflen, int *lenp) 3423 { 3424 int synd_code; 3425 int ret; 3426 3427 /* 3428 * An AFSR of -1 defaults to a memory syndrome. 3429 */ 3430 if (flt_stat == (uint64_t)-1) 3431 flt_stat = C_AFSR_CE; 3432 3433 synd_code = synd_to_synd_code(synd_status, flt_synd, flt_stat); 3434 3435 /* 3436 * Syndrome code must be either a single-bit error code 3437 * (0...143) or -1 for unum lookup. 3438 */ 3439 if (synd_code < 0 || synd_code >= M2) 3440 synd_code = -1; 3441 if (&plat_get_mem_unum) { 3442 if ((ret = plat_get_mem_unum(synd_code, flt_addr, flt_bus_id, 3443 flt_in_memory, flt_status, buf, buflen, lenp)) != 0) { 3444 buf[0] = '\0'; 3445 *lenp = 0; 3446 } 3447 3448 return (ret); 3449 } 3450 3451 return (ENOTSUP); 3452 } 3453 3454 /* 3455 * Wrapper for cpu_get_mem_unum() routine that takes an 3456 * async_flt struct rather than explicit arguments. 3457 */ 3458 int 3459 cpu_get_mem_unum_aflt(int synd_status, struct async_flt *aflt, 3460 char *buf, int buflen, int *lenp) 3461 { 3462 /* 3463 * If we come thru here for an IO bus error aflt->flt_stat will 3464 * not be the CPU AFSR, and we pass in a -1 to cpu_get_mem_unum() 3465 * so it will interpret this as a memory error. 3466 */ 3467 return (cpu_get_mem_unum(synd_status, aflt->flt_synd, 3468 (aflt->flt_class == BUS_FAULT) ? 3469 (uint64_t)-1 : ((ch_async_flt_t *)aflt)->flt_bit, 3470 aflt->flt_addr, aflt->flt_bus_id, aflt->flt_in_memory, 3471 aflt->flt_status, buf, buflen, lenp)); 3472 } 3473 3474 /* 3475 * Return unum string given synd_code and async_flt into 3476 * the buf with size UNUM_NAMLEN 3477 */ 3478 static int 3479 cpu_get_mem_unum_synd(int synd_code, struct async_flt *aflt, char *buf) 3480 { 3481 int ret, len; 3482 3483 /* 3484 * Syndrome code must be either a single-bit error code 3485 * (0...143) or -1 for unum lookup. 3486 */ 3487 if (synd_code < 0 || synd_code >= M2) 3488 synd_code = -1; 3489 if (&plat_get_mem_unum) { 3490 if ((ret = plat_get_mem_unum(synd_code, aflt->flt_addr, 3491 aflt->flt_bus_id, aflt->flt_in_memory, 3492 aflt->flt_status, buf, UNUM_NAMLEN, &len)) != 0) { 3493 buf[0] = '\0'; 3494 } 3495 return (ret); 3496 } 3497 3498 buf[0] = '\0'; 3499 return (ENOTSUP); 3500 } 3501 3502 /* 3503 * This routine is a more generic interface to cpu_get_mem_unum() 3504 * that may be used by other modules (e.g. the 'mm' driver, through 3505 * the 'MEM_NAME' ioctl, which is used by fmd to resolve unum's 3506 * for Jalapeno/Serrano FRC/RCE or FRU/RUE paired events). 3507 */ 3508 int 3509 cpu_get_mem_name(uint64_t synd, uint64_t *afsr, uint64_t afar, 3510 char *buf, int buflen, int *lenp) 3511 { 3512 int synd_status, flt_in_memory, ret; 3513 ushort_t flt_status = 0; 3514 char unum[UNUM_NAMLEN]; 3515 uint64_t t_afsr_errs; 3516 3517 /* 3518 * Check for an invalid address. 3519 */ 3520 if (afar == (uint64_t)-1) 3521 return (ENXIO); 3522 3523 if (synd == (uint64_t)-1) 3524 synd_status = AFLT_STAT_INVALID; 3525 else 3526 synd_status = AFLT_STAT_VALID; 3527 3528 flt_in_memory = (*afsr & C_AFSR_MEMORY) && 3529 pf_is_memory(afar >> MMU_PAGESHIFT); 3530 3531 /* 3532 * Get aggregate AFSR for call to cpu_error_is_ecache_data. 3533 */ 3534 if (*afsr == (uint64_t)-1) 3535 t_afsr_errs = C_AFSR_CE; 3536 else { 3537 t_afsr_errs = (*afsr & C_AFSR_ALL_ERRS); 3538 #if defined(CHEETAH_PLUS) 3539 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) 3540 t_afsr_errs |= (*(afsr + 1) & C_AFSR_EXT_ALL_ERRS); 3541 #endif /* CHEETAH_PLUS */ 3542 } 3543 3544 /* 3545 * Turn on ECC_ECACHE if error type is E$ Data. 3546 */ 3547 if (cpu_error_is_ecache_data(CPU->cpu_id, t_afsr_errs)) 3548 flt_status |= ECC_ECACHE; 3549 3550 ret = cpu_get_mem_unum(synd_status, (ushort_t)synd, t_afsr_errs, afar, 3551 CPU->cpu_id, flt_in_memory, flt_status, unum, UNUM_NAMLEN, lenp); 3552 if (ret != 0) 3553 return (ret); 3554 3555 if (*lenp >= buflen) 3556 return (ENAMETOOLONG); 3557 3558 (void) strncpy(buf, unum, buflen); 3559 3560 return (0); 3561 } 3562 3563 /* 3564 * Routine to return memory information associated 3565 * with a physical address and syndrome. 3566 */ 3567 int 3568 cpu_get_mem_info(uint64_t synd, uint64_t afar, 3569 uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep, 3570 int *segsp, int *banksp, int *mcidp) 3571 { 3572 int synd_status, synd_code; 3573 3574 if (afar == (uint64_t)-1) 3575 return (ENXIO); 3576 3577 if (synd == (uint64_t)-1) 3578 synd_status = AFLT_STAT_INVALID; 3579 else 3580 synd_status = AFLT_STAT_VALID; 3581 3582 synd_code = synd_to_synd_code(synd_status, synd, C_AFSR_CE); 3583 3584 if (p2get_mem_info != NULL) 3585 return ((p2get_mem_info)(synd_code, afar, 3586 mem_sizep, seg_sizep, bank_sizep, 3587 segsp, banksp, mcidp)); 3588 else 3589 return (ENOTSUP); 3590 } 3591 3592 /* 3593 * Routine to return a string identifying the physical 3594 * name associated with a cpuid. 3595 */ 3596 int 3597 cpu_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp) 3598 { 3599 int ret; 3600 char unum[UNUM_NAMLEN]; 3601 3602 if (&plat_get_cpu_unum) { 3603 if ((ret = plat_get_cpu_unum(cpuid, unum, UNUM_NAMLEN, lenp)) 3604 != 0) 3605 return (ret); 3606 } else { 3607 return (ENOTSUP); 3608 } 3609 3610 if (*lenp >= buflen) 3611 return (ENAMETOOLONG); 3612 3613 (void) strncpy(buf, unum, buflen); 3614 3615 return (0); 3616 } 3617 3618 /* 3619 * This routine exports the name buffer size. 3620 */ 3621 size_t 3622 cpu_get_name_bufsize() 3623 { 3624 return (UNUM_NAMLEN); 3625 } 3626 3627 /* 3628 * Historical function, apparantly not used. 3629 */ 3630 /* ARGSUSED */ 3631 void 3632 cpu_read_paddr(struct async_flt *ecc, short verbose, short ce_err) 3633 {} 3634 3635 /* 3636 * Historical function only called for SBus errors in debugging. 3637 */ 3638 /*ARGSUSED*/ 3639 void 3640 read_ecc_data(struct async_flt *aflt, short verbose, short ce_err) 3641 {} 3642 3643 /* 3644 * Clear the AFSR sticky bits. The routine returns a non-zero value if 3645 * any of the AFSR's sticky errors are detected. If a non-null pointer to 3646 * an async fault structure argument is passed in, the captured error state 3647 * (AFSR, AFAR) info will be returned in the structure. 3648 */ 3649 int 3650 clear_errors(ch_async_flt_t *ch_flt) 3651 { 3652 struct async_flt *aflt = (struct async_flt *)ch_flt; 3653 ch_cpu_errors_t cpu_error_regs; 3654 3655 get_cpu_error_state(&cpu_error_regs); 3656 3657 if (ch_flt != NULL) { 3658 aflt->flt_stat = cpu_error_regs.afsr & C_AFSR_MASK; 3659 aflt->flt_addr = cpu_error_regs.afar; 3660 ch_flt->afsr_ext = cpu_error_regs.afsr_ext; 3661 ch_flt->afsr_errs = (cpu_error_regs.afsr & C_AFSR_ALL_ERRS) | 3662 (cpu_error_regs.afsr_ext & C_AFSR_EXT_ALL_ERRS); 3663 #if defined(SERRANO) 3664 ch_flt->afar2 = cpu_error_regs.afar2; 3665 #endif /* SERRANO */ 3666 } 3667 3668 set_cpu_error_state(&cpu_error_regs); 3669 3670 return (((cpu_error_regs.afsr & C_AFSR_ALL_ERRS) | 3671 (cpu_error_regs.afsr_ext & C_AFSR_EXT_ALL_ERRS)) != 0); 3672 } 3673 3674 /* 3675 * Clear any AFSR error bits, and check for persistence. 3676 * 3677 * It would be desirable to also insist that syndrome match. PCI handling 3678 * has already filled flt_synd. For errors trapped by CPU we only fill 3679 * flt_synd when we queue the event, so we do not have a valid flt_synd 3680 * during initial classification (it is valid if we're called as part of 3681 * subsequent low-pil additional classification attempts). We could try 3682 * to determine which syndrome to use: we know we're only called for 3683 * CE/RCE (Jalapeno & Serrano) and CE/EMC (others) so the syndrome to use 3684 * would be esynd/none and esynd/msynd, respectively. If that is 3685 * implemented then what do we do in the case that we do experience an 3686 * error on the same afar but with different syndrome? At the very least 3687 * we should count such occurences. Anyway, for now, we'll leave it as 3688 * it has been for ages. 3689 */ 3690 static int 3691 clear_ecc(struct async_flt *aflt) 3692 { 3693 ch_cpu_errors_t cpu_error_regs; 3694 3695 /* 3696 * Snapshot the AFSR and AFAR and clear any errors 3697 */ 3698 get_cpu_error_state(&cpu_error_regs); 3699 set_cpu_error_state(&cpu_error_regs); 3700 3701 /* 3702 * If any of the same memory access error bits are still on and 3703 * the AFAR matches, return that the error is persistent. 3704 */ 3705 return ((cpu_error_regs.afsr & (C_AFSR_MEMORY & aflt->flt_stat)) != 0 && 3706 cpu_error_regs.afar == aflt->flt_addr); 3707 } 3708 3709 /* 3710 * Turn off all cpu error detection, normally only used for panics. 3711 */ 3712 void 3713 cpu_disable_errors(void) 3714 { 3715 xt_all(set_error_enable_tl1, EN_REG_DISABLE, EER_SET_ABSOLUTE); 3716 3717 /* 3718 * With error detection now turned off, check the other cpus 3719 * logout areas for any unlogged errors. 3720 */ 3721 if (enable_check_other_cpus_logout) { 3722 cpu_check_other_cpus_logout(); 3723 /* 3724 * Make a second pass over the logout areas, in case 3725 * there is a failing CPU in an error-trap loop which 3726 * will write to the logout area once it is emptied. 3727 */ 3728 cpu_check_other_cpus_logout(); 3729 } 3730 } 3731 3732 /* 3733 * Enable errors. 3734 */ 3735 void 3736 cpu_enable_errors(void) 3737 { 3738 xt_all(set_error_enable_tl1, EN_REG_ENABLE, EER_SET_ABSOLUTE); 3739 } 3740 3741 /* 3742 * Flush the entire ecache using displacement flush by reading through a 3743 * physical address range twice as large as the Ecache. 3744 */ 3745 void 3746 cpu_flush_ecache(void) 3747 { 3748 flush_ecache(ecache_flushaddr, cpunodes[CPU->cpu_id].ecache_size, 3749 cpunodes[CPU->cpu_id].ecache_linesize); 3750 } 3751 3752 /* 3753 * Return CPU E$ set size - E$ size divided by the associativity. 3754 * We use this function in places where the CPU_PRIVATE ptr may not be 3755 * initialized yet. Note that for send_mondo and in the Ecache scrubber, 3756 * we're guaranteed that CPU_PRIVATE is initialized. Also, cpunodes is set 3757 * up before the kernel switches from OBP's to the kernel's trap table, so 3758 * we don't have to worry about cpunodes being unitialized. 3759 */ 3760 int 3761 cpu_ecache_set_size(struct cpu *cp) 3762 { 3763 if (CPU_PRIVATE(cp)) 3764 return (CPU_PRIVATE_VAL(cp, chpr_ec_set_size)); 3765 3766 return (cpunodes[cp->cpu_id].ecache_size / cpu_ecache_nway()); 3767 } 3768 3769 /* 3770 * Flush Ecache line. 3771 * Uses ASI_EC_DIAG for Cheetah+ and Jalapeno. 3772 * Uses normal displacement flush for Cheetah. 3773 */ 3774 static void 3775 cpu_flush_ecache_line(ch_async_flt_t *ch_flt) 3776 { 3777 struct async_flt *aflt = (struct async_flt *)ch_flt; 3778 int ec_set_size = cpu_ecache_set_size(CPU); 3779 3780 ecache_flush_line(aflt->flt_addr, ec_set_size); 3781 } 3782 3783 /* 3784 * Scrub physical address. 3785 * Scrub code is different depending upon whether this a Cheetah+ with 2-way 3786 * Ecache or direct-mapped Ecache. 3787 */ 3788 static void 3789 cpu_scrubphys(struct async_flt *aflt) 3790 { 3791 int ec_set_size = cpu_ecache_set_size(CPU); 3792 3793 scrubphys(aflt->flt_addr, ec_set_size); 3794 } 3795 3796 /* 3797 * Clear physical address. 3798 * Scrub code is different depending upon whether this a Cheetah+ with 2-way 3799 * Ecache or direct-mapped Ecache. 3800 */ 3801 void 3802 cpu_clearphys(struct async_flt *aflt) 3803 { 3804 int lsize = cpunodes[CPU->cpu_id].ecache_linesize; 3805 int ec_set_size = cpu_ecache_set_size(CPU); 3806 3807 3808 clearphys(aflt->flt_addr, ec_set_size, lsize); 3809 } 3810 3811 #if defined(CPU_IMP_ECACHE_ASSOC) 3812 /* 3813 * Check for a matching valid line in all the sets. 3814 * If found, return set# + 1. Otherwise return 0. 3815 */ 3816 static int 3817 cpu_ecache_line_valid(ch_async_flt_t *ch_flt) 3818 { 3819 struct async_flt *aflt = (struct async_flt *)ch_flt; 3820 int totalsize = cpunodes[CPU->cpu_id].ecache_size; 3821 int ec_set_size = cpu_ecache_set_size(CPU); 3822 ch_ec_data_t *ecp = &ch_flt->flt_diag_data.chd_ec_data[0]; 3823 int nway = cpu_ecache_nway(); 3824 int i; 3825 3826 for (i = 0; i < nway; i++, ecp++) { 3827 if (!cpu_ectag_line_invalid(totalsize, ecp->ec_tag) && 3828 (aflt->flt_addr & P2ALIGN(C_AFAR_PA, ec_set_size)) == 3829 cpu_ectag_to_pa(ec_set_size, ecp->ec_tag)) 3830 return (i+1); 3831 } 3832 return (0); 3833 } 3834 #endif /* CPU_IMP_ECACHE_ASSOC */ 3835 3836 /* 3837 * Check whether a line in the given logout info matches the specified 3838 * fault address. If reqval is set then the line must not be Invalid. 3839 * Returns 0 on failure; on success (way + 1) is returned an *level is 3840 * set to 2 for l2$ or 3 for l3$. 3841 */ 3842 static int 3843 cpu_matching_ecache_line(uint64_t faddr, void *data, int reqval, int *level) 3844 { 3845 ch_diag_data_t *cdp = data; 3846 ch_ec_data_t *ecp; 3847 int totalsize, ec_set_size; 3848 int i, ways; 3849 int match = 0; 3850 int tagvalid; 3851 uint64_t addr, tagpa; 3852 int ispanther = IS_PANTHER(cpunodes[CPU->cpu_id].implementation); 3853 3854 /* 3855 * Check the l2$ logout data 3856 */ 3857 if (ispanther) { 3858 ecp = &cdp->chd_l2_data[0]; 3859 ec_set_size = PN_L2_SET_SIZE; 3860 ways = PN_L2_NWAYS; 3861 } else { 3862 ecp = &cdp->chd_ec_data[0]; 3863 ec_set_size = cpu_ecache_set_size(CPU); 3864 ways = cpu_ecache_nway(); 3865 totalsize = cpunodes[CPU->cpu_id].ecache_size; 3866 } 3867 /* remove low order PA bits from fault address not used in PA tag */ 3868 addr = faddr & P2ALIGN(C_AFAR_PA, ec_set_size); 3869 for (i = 0; i < ways; i++, ecp++) { 3870 if (ispanther) { 3871 tagpa = PN_L2TAG_TO_PA(ecp->ec_tag); 3872 tagvalid = !PN_L2_LINE_INVALID(ecp->ec_tag); 3873 } else { 3874 tagpa = cpu_ectag_to_pa(ec_set_size, ecp->ec_tag); 3875 tagvalid = !cpu_ectag_line_invalid(totalsize, 3876 ecp->ec_tag); 3877 } 3878 if (tagpa == addr && (!reqval || tagvalid)) { 3879 match = i + 1; 3880 *level = 2; 3881 break; 3882 } 3883 } 3884 3885 if (match || !ispanther) 3886 return (match); 3887 3888 /* For Panther we also check the l3$ */ 3889 ecp = &cdp->chd_ec_data[0]; 3890 ec_set_size = PN_L3_SET_SIZE; 3891 ways = PN_L3_NWAYS; 3892 addr = faddr & P2ALIGN(C_AFAR_PA, ec_set_size); 3893 3894 for (i = 0; i < ways; i++, ecp++) { 3895 if (PN_L3TAG_TO_PA(ecp->ec_tag) == addr && (!reqval || 3896 !PN_L3_LINE_INVALID(ecp->ec_tag))) { 3897 match = i + 1; 3898 *level = 3; 3899 break; 3900 } 3901 } 3902 3903 return (match); 3904 } 3905 3906 #if defined(CPU_IMP_L1_CACHE_PARITY) 3907 /* 3908 * Record information related to the source of an Dcache Parity Error. 3909 */ 3910 static void 3911 cpu_dcache_parity_info(ch_async_flt_t *ch_flt) 3912 { 3913 int dc_set_size = dcache_size / CH_DCACHE_NWAY; 3914 int index; 3915 3916 /* 3917 * Since instruction decode cannot be done at high PIL 3918 * just examine the entire Dcache to locate the error. 3919 */ 3920 if (ch_flt->parity_data.dpe.cpl_lcnt == 0) { 3921 ch_flt->parity_data.dpe.cpl_way = -1; 3922 ch_flt->parity_data.dpe.cpl_off = -1; 3923 } 3924 for (index = 0; index < dc_set_size; index += dcache_linesize) 3925 cpu_dcache_parity_check(ch_flt, index); 3926 } 3927 3928 /* 3929 * Check all ways of the Dcache at a specified index for good parity. 3930 */ 3931 static void 3932 cpu_dcache_parity_check(ch_async_flt_t *ch_flt, int index) 3933 { 3934 int dc_set_size = dcache_size / CH_DCACHE_NWAY; 3935 uint64_t parity_bits, pbits, data_word; 3936 static int parity_bits_popc[] = { 0, 1, 1, 0 }; 3937 int way, word, data_byte; 3938 ch_dc_data_t *dcp = &ch_flt->parity_data.dpe.cpl_dc[0]; 3939 ch_dc_data_t tmp_dcp; 3940 3941 for (way = 0; way < CH_DCACHE_NWAY; way++, dcp++) { 3942 /* 3943 * Perform diagnostic read. 3944 */ 3945 get_dcache_dtag(index + way * dc_set_size, 3946 (uint64_t *)&tmp_dcp); 3947 3948 /* 3949 * Check tag for even parity. 3950 * Sum of 1 bits (including parity bit) should be even. 3951 */ 3952 if (popc64(tmp_dcp.dc_tag & CHP_DCTAG_PARMASK) & 1) { 3953 /* 3954 * If this is the first error log detailed information 3955 * about it and check the snoop tag. Otherwise just 3956 * record the fact that we found another error. 3957 */ 3958 if (ch_flt->parity_data.dpe.cpl_lcnt == 0) { 3959 ch_flt->parity_data.dpe.cpl_way = way; 3960 ch_flt->parity_data.dpe.cpl_cache = 3961 CPU_DC_PARITY; 3962 ch_flt->parity_data.dpe.cpl_tag |= CHP_DC_TAG; 3963 3964 if (popc64(tmp_dcp.dc_sntag & 3965 CHP_DCSNTAG_PARMASK) & 1) { 3966 ch_flt->parity_data.dpe.cpl_tag |= 3967 CHP_DC_SNTAG; 3968 ch_flt->parity_data.dpe.cpl_lcnt++; 3969 } 3970 3971 bcopy(&tmp_dcp, dcp, sizeof (ch_dc_data_t)); 3972 } 3973 3974 ch_flt->parity_data.dpe.cpl_lcnt++; 3975 } 3976 3977 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 3978 /* 3979 * Panther has more parity bits than the other 3980 * processors for covering dcache data and so each 3981 * byte of data in each word has its own parity bit. 3982 */ 3983 parity_bits = tmp_dcp.dc_pn_data_parity; 3984 for (word = 0; word < 4; word++) { 3985 data_word = tmp_dcp.dc_data[word]; 3986 pbits = parity_bits & PN_DC_DATA_PARITY_MASK; 3987 for (data_byte = 0; data_byte < 8; 3988 data_byte++) { 3989 if (((popc64(data_word & 3990 PN_DC_DATA_PARITY_MASK)) & 1) ^ 3991 (pbits & 1)) { 3992 cpu_record_dc_data_parity( 3993 ch_flt, dcp, &tmp_dcp, way, 3994 word); 3995 } 3996 pbits >>= 1; 3997 data_word >>= 8; 3998 } 3999 parity_bits >>= 8; 4000 } 4001 } else { 4002 /* 4003 * Check data array for even parity. 4004 * The 8 parity bits are grouped into 4 pairs each 4005 * of which covers a 64-bit word. The endianness is 4006 * reversed -- the low-order parity bits cover the 4007 * high-order data words. 4008 */ 4009 parity_bits = tmp_dcp.dc_utag >> 8; 4010 for (word = 0; word < 4; word++) { 4011 pbits = (parity_bits >> (6 - word * 2)) & 3; 4012 if ((popc64(tmp_dcp.dc_data[word]) + 4013 parity_bits_popc[pbits]) & 1) { 4014 cpu_record_dc_data_parity(ch_flt, dcp, 4015 &tmp_dcp, way, word); 4016 } 4017 } 4018 } 4019 } 4020 } 4021 4022 static void 4023 cpu_record_dc_data_parity(ch_async_flt_t *ch_flt, 4024 ch_dc_data_t *dest_dcp, ch_dc_data_t *src_dcp, int way, int word) 4025 { 4026 /* 4027 * If this is the first error log detailed information about it. 4028 * Otherwise just record the fact that we found another error. 4029 */ 4030 if (ch_flt->parity_data.dpe.cpl_lcnt == 0) { 4031 ch_flt->parity_data.dpe.cpl_way = way; 4032 ch_flt->parity_data.dpe.cpl_cache = CPU_DC_PARITY; 4033 ch_flt->parity_data.dpe.cpl_off = word * 8; 4034 bcopy(src_dcp, dest_dcp, sizeof (ch_dc_data_t)); 4035 } 4036 ch_flt->parity_data.dpe.cpl_lcnt++; 4037 } 4038 4039 /* 4040 * Record information related to the source of an Icache Parity Error. 4041 * 4042 * Called with the Icache disabled so any diagnostic accesses are safe. 4043 */ 4044 static void 4045 cpu_icache_parity_info(ch_async_flt_t *ch_flt) 4046 { 4047 int ic_set_size; 4048 int ic_linesize; 4049 int index; 4050 4051 if (CPU_PRIVATE(CPU)) { 4052 ic_set_size = CPU_PRIVATE_VAL(CPU, chpr_icache_size) / 4053 CH_ICACHE_NWAY; 4054 ic_linesize = CPU_PRIVATE_VAL(CPU, chpr_icache_linesize); 4055 } else { 4056 ic_set_size = icache_size / CH_ICACHE_NWAY; 4057 ic_linesize = icache_linesize; 4058 } 4059 4060 ch_flt->parity_data.ipe.cpl_way = -1; 4061 ch_flt->parity_data.ipe.cpl_off = -1; 4062 4063 for (index = 0; index < ic_set_size; index += ic_linesize) 4064 cpu_icache_parity_check(ch_flt, index); 4065 } 4066 4067 /* 4068 * Check all ways of the Icache at a specified index for good parity. 4069 */ 4070 static void 4071 cpu_icache_parity_check(ch_async_flt_t *ch_flt, int index) 4072 { 4073 uint64_t parmask, pn_inst_parity; 4074 int ic_set_size; 4075 int ic_linesize; 4076 int flt_index, way, instr, num_instr; 4077 struct async_flt *aflt = (struct async_flt *)ch_flt; 4078 ch_ic_data_t *icp = &ch_flt->parity_data.ipe.cpl_ic[0]; 4079 ch_ic_data_t tmp_icp; 4080 4081 if (CPU_PRIVATE(CPU)) { 4082 ic_set_size = CPU_PRIVATE_VAL(CPU, chpr_icache_size) / 4083 CH_ICACHE_NWAY; 4084 ic_linesize = CPU_PRIVATE_VAL(CPU, chpr_icache_linesize); 4085 } else { 4086 ic_set_size = icache_size / CH_ICACHE_NWAY; 4087 ic_linesize = icache_linesize; 4088 } 4089 4090 /* 4091 * Panther has twice as many instructions per icache line and the 4092 * instruction parity bit is in a different location. 4093 */ 4094 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 4095 num_instr = PN_IC_DATA_REG_SIZE / sizeof (uint64_t); 4096 pn_inst_parity = PN_ICDATA_PARITY_BIT_MASK; 4097 } else { 4098 num_instr = CH_IC_DATA_REG_SIZE / sizeof (uint64_t); 4099 pn_inst_parity = 0; 4100 } 4101 4102 /* 4103 * Index at which we expect to find the parity error. 4104 */ 4105 flt_index = P2ALIGN(aflt->flt_addr % ic_set_size, ic_linesize); 4106 4107 for (way = 0; way < CH_ICACHE_NWAY; way++, icp++) { 4108 /* 4109 * Diagnostic reads expect address argument in ASI format. 4110 */ 4111 get_icache_dtag(2 * (index + way * ic_set_size), 4112 (uint64_t *)&tmp_icp); 4113 4114 /* 4115 * If this is the index in which we expect to find the 4116 * error log detailed information about each of the ways. 4117 * This information will be displayed later if we can't 4118 * determine the exact way in which the error is located. 4119 */ 4120 if (flt_index == index) 4121 bcopy(&tmp_icp, icp, sizeof (ch_ic_data_t)); 4122 4123 /* 4124 * Check tag for even parity. 4125 * Sum of 1 bits (including parity bit) should be even. 4126 */ 4127 if (popc64(tmp_icp.ic_patag & CHP_ICPATAG_PARMASK) & 1) { 4128 /* 4129 * If this way is the one in which we expected 4130 * to find the error record the way and check the 4131 * snoop tag. Otherwise just record the fact we 4132 * found another error. 4133 */ 4134 if (flt_index == index) { 4135 ch_flt->parity_data.ipe.cpl_way = way; 4136 ch_flt->parity_data.ipe.cpl_tag |= CHP_IC_TAG; 4137 4138 if (popc64(tmp_icp.ic_sntag & 4139 CHP_ICSNTAG_PARMASK) & 1) { 4140 ch_flt->parity_data.ipe.cpl_tag |= 4141 CHP_IC_SNTAG; 4142 ch_flt->parity_data.ipe.cpl_lcnt++; 4143 } 4144 4145 } 4146 ch_flt->parity_data.ipe.cpl_lcnt++; 4147 continue; 4148 } 4149 4150 /* 4151 * Check instruction data for even parity. 4152 * Bits participating in parity differ for PC-relative 4153 * versus non-PC-relative instructions. 4154 */ 4155 for (instr = 0; instr < num_instr; instr++) { 4156 parmask = (tmp_icp.ic_data[instr] & 4157 CH_ICDATA_PRED_ISPCREL) ? 4158 (CHP_ICDATA_PCREL_PARMASK | pn_inst_parity) : 4159 (CHP_ICDATA_NPCREL_PARMASK | pn_inst_parity); 4160 if (popc64(tmp_icp.ic_data[instr] & parmask) & 1) { 4161 /* 4162 * If this way is the one in which we expected 4163 * to find the error record the way and offset. 4164 * Otherwise just log the fact we found another 4165 * error. 4166 */ 4167 if (flt_index == index) { 4168 ch_flt->parity_data.ipe.cpl_way = way; 4169 ch_flt->parity_data.ipe.cpl_off = 4170 instr * 4; 4171 } 4172 ch_flt->parity_data.ipe.cpl_lcnt++; 4173 continue; 4174 } 4175 } 4176 } 4177 } 4178 4179 /* 4180 * Record information related to the source of an Pcache Parity Error. 4181 */ 4182 static void 4183 cpu_pcache_parity_info(ch_async_flt_t *ch_flt) 4184 { 4185 int pc_set_size = CH_PCACHE_SIZE / CH_PCACHE_NWAY; 4186 int index; 4187 4188 /* 4189 * Since instruction decode cannot be done at high PIL just 4190 * examine the entire Pcache to check for any parity errors. 4191 */ 4192 if (ch_flt->parity_data.dpe.cpl_lcnt == 0) { 4193 ch_flt->parity_data.dpe.cpl_way = -1; 4194 ch_flt->parity_data.dpe.cpl_off = -1; 4195 } 4196 for (index = 0; index < pc_set_size; index += CH_PCACHE_LSIZE) 4197 cpu_pcache_parity_check(ch_flt, index); 4198 } 4199 4200 /* 4201 * Check all ways of the Pcache at a specified index for good parity. 4202 */ 4203 static void 4204 cpu_pcache_parity_check(ch_async_flt_t *ch_flt, int index) 4205 { 4206 int pc_set_size = CH_PCACHE_SIZE / CH_PCACHE_NWAY; 4207 int pc_data_words = CH_PC_DATA_REG_SIZE / sizeof (uint64_t); 4208 int way, word, pbit, parity_bits; 4209 ch_pc_data_t *pcp = &ch_flt->parity_data.dpe.cpl_pc[0]; 4210 ch_pc_data_t tmp_pcp; 4211 4212 for (way = 0; way < CH_PCACHE_NWAY; way++, pcp++) { 4213 /* 4214 * Perform diagnostic read. 4215 */ 4216 get_pcache_dtag(index + way * pc_set_size, 4217 (uint64_t *)&tmp_pcp); 4218 /* 4219 * Check data array for odd parity. There are 8 parity 4220 * bits (bits 57:50 of ASI_PCACHE_STATUS_DATA) and each 4221 * of those bits covers exactly 8 bytes of the data 4222 * array: 4223 * 4224 * parity bit P$ data bytes covered 4225 * ---------- --------------------- 4226 * 50 63:56 4227 * 51 55:48 4228 * 52 47:40 4229 * 53 39:32 4230 * 54 31:24 4231 * 55 23:16 4232 * 56 15:8 4233 * 57 7:0 4234 */ 4235 parity_bits = PN_PC_PARITY_BITS(tmp_pcp.pc_status); 4236 for (word = 0; word < pc_data_words; word++) { 4237 pbit = (parity_bits >> (pc_data_words - word - 1)) & 1; 4238 if ((popc64(tmp_pcp.pc_data[word]) & 1) ^ pbit) { 4239 /* 4240 * If this is the first error log detailed 4241 * information about it. Otherwise just record 4242 * the fact that we found another error. 4243 */ 4244 if (ch_flt->parity_data.dpe.cpl_lcnt == 0) { 4245 ch_flt->parity_data.dpe.cpl_way = way; 4246 ch_flt->parity_data.dpe.cpl_cache = 4247 CPU_PC_PARITY; 4248 ch_flt->parity_data.dpe.cpl_off = 4249 word * sizeof (uint64_t); 4250 bcopy(&tmp_pcp, pcp, 4251 sizeof (ch_pc_data_t)); 4252 } 4253 ch_flt->parity_data.dpe.cpl_lcnt++; 4254 } 4255 } 4256 } 4257 } 4258 4259 4260 /* 4261 * Add L1 Data cache data to the ereport payload. 4262 */ 4263 static void 4264 cpu_payload_add_dcache(struct async_flt *aflt, nvlist_t *nvl) 4265 { 4266 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 4267 ch_dc_data_t *dcp; 4268 ch_dc_data_t dcdata[CH_DCACHE_NWAY]; 4269 uint_t nelem; 4270 int i, ways_to_check, ways_logged = 0; 4271 4272 /* 4273 * If this is an D$ fault then there may be multiple 4274 * ways captured in the ch_parity_log_t structure. 4275 * Otherwise, there will be at most one way captured 4276 * in the ch_diag_data_t struct. 4277 * Check each way to see if it should be encoded. 4278 */ 4279 if (ch_flt->flt_type == CPU_DC_PARITY) 4280 ways_to_check = CH_DCACHE_NWAY; 4281 else 4282 ways_to_check = 1; 4283 for (i = 0; i < ways_to_check; i++) { 4284 if (ch_flt->flt_type == CPU_DC_PARITY) 4285 dcp = &ch_flt->parity_data.dpe.cpl_dc[i]; 4286 else 4287 dcp = &ch_flt->flt_diag_data.chd_dc_data; 4288 if (dcp->dc_logflag == DC_LOGFLAG_MAGIC) { 4289 bcopy(dcp, &dcdata[ways_logged], 4290 sizeof (ch_dc_data_t)); 4291 ways_logged++; 4292 } 4293 } 4294 4295 /* 4296 * Add the dcache data to the payload. 4297 */ 4298 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1D_WAYS, 4299 DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL); 4300 if (ways_logged != 0) { 4301 nelem = sizeof (ch_dc_data_t) / sizeof (uint64_t) * ways_logged; 4302 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1D_DATA, 4303 DATA_TYPE_UINT64_ARRAY, nelem, (uint64_t *)dcdata, NULL); 4304 } 4305 } 4306 4307 /* 4308 * Add L1 Instruction cache data to the ereport payload. 4309 */ 4310 static void 4311 cpu_payload_add_icache(struct async_flt *aflt, nvlist_t *nvl) 4312 { 4313 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 4314 ch_ic_data_t *icp; 4315 ch_ic_data_t icdata[CH_ICACHE_NWAY]; 4316 uint_t nelem; 4317 int i, ways_to_check, ways_logged = 0; 4318 4319 /* 4320 * If this is an I$ fault then there may be multiple 4321 * ways captured in the ch_parity_log_t structure. 4322 * Otherwise, there will be at most one way captured 4323 * in the ch_diag_data_t struct. 4324 * Check each way to see if it should be encoded. 4325 */ 4326 if (ch_flt->flt_type == CPU_IC_PARITY) 4327 ways_to_check = CH_ICACHE_NWAY; 4328 else 4329 ways_to_check = 1; 4330 for (i = 0; i < ways_to_check; i++) { 4331 if (ch_flt->flt_type == CPU_IC_PARITY) 4332 icp = &ch_flt->parity_data.ipe.cpl_ic[i]; 4333 else 4334 icp = &ch_flt->flt_diag_data.chd_ic_data; 4335 if (icp->ic_logflag == IC_LOGFLAG_MAGIC) { 4336 bcopy(icp, &icdata[ways_logged], 4337 sizeof (ch_ic_data_t)); 4338 ways_logged++; 4339 } 4340 } 4341 4342 /* 4343 * Add the icache data to the payload. 4344 */ 4345 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1I_WAYS, 4346 DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL); 4347 if (ways_logged != 0) { 4348 nelem = sizeof (ch_ic_data_t) / sizeof (uint64_t) * ways_logged; 4349 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L1I_DATA, 4350 DATA_TYPE_UINT64_ARRAY, nelem, (uint64_t *)icdata, NULL); 4351 } 4352 } 4353 4354 #endif /* CPU_IMP_L1_CACHE_PARITY */ 4355 4356 /* 4357 * Add ecache data to payload. 4358 */ 4359 static void 4360 cpu_payload_add_ecache(struct async_flt *aflt, nvlist_t *nvl) 4361 { 4362 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 4363 ch_ec_data_t *ecp; 4364 ch_ec_data_t ecdata[CHD_EC_DATA_SETS]; 4365 uint_t nelem; 4366 int i, ways_logged = 0; 4367 4368 /* 4369 * Check each way to see if it should be encoded 4370 * and concatinate it into a temporary buffer. 4371 */ 4372 for (i = 0; i < CHD_EC_DATA_SETS; i++) { 4373 ecp = &ch_flt->flt_diag_data.chd_ec_data[i]; 4374 if (ecp->ec_logflag == EC_LOGFLAG_MAGIC) { 4375 bcopy(ecp, &ecdata[ways_logged], 4376 sizeof (ch_ec_data_t)); 4377 ways_logged++; 4378 } 4379 } 4380 4381 /* 4382 * Panther CPUs have an additional level of cache and so 4383 * what we just collected was the L3 (ecache) and not the 4384 * L2 cache. 4385 */ 4386 if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) { 4387 /* 4388 * Add the L3 (ecache) data to the payload. 4389 */ 4390 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L3_WAYS, 4391 DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL); 4392 if (ways_logged != 0) { 4393 nelem = sizeof (ch_ec_data_t) / 4394 sizeof (uint64_t) * ways_logged; 4395 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L3_DATA, 4396 DATA_TYPE_UINT64_ARRAY, nelem, 4397 (uint64_t *)ecdata, NULL); 4398 } 4399 4400 /* 4401 * Now collect the L2 cache. 4402 */ 4403 ways_logged = 0; 4404 for (i = 0; i < PN_L2_NWAYS; i++) { 4405 ecp = &ch_flt->flt_diag_data.chd_l2_data[i]; 4406 if (ecp->ec_logflag == EC_LOGFLAG_MAGIC) { 4407 bcopy(ecp, &ecdata[ways_logged], 4408 sizeof (ch_ec_data_t)); 4409 ways_logged++; 4410 } 4411 } 4412 } 4413 4414 /* 4415 * Add the L2 cache data to the payload. 4416 */ 4417 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L2_WAYS, 4418 DATA_TYPE_UINT8, (uint8_t)ways_logged, NULL); 4419 if (ways_logged != 0) { 4420 nelem = sizeof (ch_ec_data_t) / 4421 sizeof (uint64_t) * ways_logged; 4422 fm_payload_set(nvl, FM_EREPORT_PAYLOAD_NAME_L2_DATA, 4423 DATA_TYPE_UINT64_ARRAY, nelem, (uint64_t *)ecdata, NULL); 4424 } 4425 } 4426 4427 /* 4428 * Initialize cpu scheme for specified cpu. 4429 */ 4430 static void 4431 cpu_fmri_cpu_set(nvlist_t *cpu_fmri, int cpuid) 4432 { 4433 char sbuf[21]; /* sizeof (UINT64_MAX) + '\0' */ 4434 uint8_t mask; 4435 4436 mask = cpunodes[cpuid].version; 4437 (void) snprintf(sbuf, sizeof (sbuf), "%llX", 4438 (u_longlong_t)cpunodes[cpuid].device_id); 4439 (void) fm_fmri_cpu_set(cpu_fmri, FM_CPU_SCHEME_VERSION, NULL, 4440 cpuid, &mask, (const char *)sbuf); 4441 } 4442 4443 /* 4444 * Returns ereport resource type. 4445 */ 4446 static int 4447 cpu_error_to_resource_type(struct async_flt *aflt) 4448 { 4449 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 4450 4451 switch (ch_flt->flt_type) { 4452 4453 case CPU_CE_ECACHE: 4454 case CPU_UE_ECACHE: 4455 case CPU_UE_ECACHE_RETIRE: 4456 case CPU_ORPH: 4457 /* 4458 * If AFSR error bit indicates L2$ Data for Cheetah, 4459 * Cheetah+ or Jaguar, or L3$ Data for Panther, return 4460 * E$ Data type, otherwise, return CPU type. 4461 */ 4462 if (cpu_error_is_ecache_data(aflt->flt_inst, 4463 ch_flt->flt_bit)) 4464 return (ERRTYPE_ECACHE_DATA); 4465 return (ERRTYPE_CPU); 4466 4467 case CPU_CE: 4468 case CPU_UE: 4469 case CPU_EMC: 4470 case CPU_DUE: 4471 case CPU_RCE: 4472 case CPU_RUE: 4473 case CPU_FRC: 4474 case CPU_FRU: 4475 return (ERRTYPE_MEMORY); 4476 4477 case CPU_IC_PARITY: 4478 case CPU_DC_PARITY: 4479 case CPU_FPUERR: 4480 case CPU_PC_PARITY: 4481 case CPU_ITLB_PARITY: 4482 case CPU_DTLB_PARITY: 4483 return (ERRTYPE_CPU); 4484 } 4485 return (ERRTYPE_UNKNOWN); 4486 } 4487 4488 /* 4489 * Encode the data saved in the ch_async_flt_t struct into 4490 * the FM ereport payload. 4491 */ 4492 static void 4493 cpu_payload_add_aflt(struct async_flt *aflt, nvlist_t *payload, 4494 nvlist_t *resource, int *afar_status, int *synd_status) 4495 { 4496 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 4497 *synd_status = AFLT_STAT_INVALID; 4498 *afar_status = AFLT_STAT_INVALID; 4499 4500 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFSR) { 4501 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFSR, 4502 DATA_TYPE_UINT64, aflt->flt_stat, NULL); 4503 } 4504 4505 if ((aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFSR_EXT) && 4506 IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) { 4507 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFSR_EXT, 4508 DATA_TYPE_UINT64, ch_flt->afsr_ext, NULL); 4509 } 4510 4511 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFAR_STATUS) { 4512 *afar_status = afsr_to_afar_status(ch_flt->afsr_errs, 4513 ch_flt->flt_bit); 4514 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFAR_STATUS, 4515 DATA_TYPE_UINT8, (uint8_t)*afar_status, NULL); 4516 } 4517 4518 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_AFAR) { 4519 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_AFAR, 4520 DATA_TYPE_UINT64, aflt->flt_addr, NULL); 4521 } 4522 4523 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PC) { 4524 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PC, 4525 DATA_TYPE_UINT64, (uint64_t)aflt->flt_pc, NULL); 4526 } 4527 4528 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TL) { 4529 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TL, 4530 DATA_TYPE_UINT8, (uint8_t)aflt->flt_tl, NULL); 4531 } 4532 4533 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TT) { 4534 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TT, 4535 DATA_TYPE_UINT8, flt_to_trap_type(aflt), NULL); 4536 } 4537 4538 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PRIV) { 4539 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PRIV, 4540 DATA_TYPE_BOOLEAN_VALUE, 4541 (aflt->flt_priv ? B_TRUE : B_FALSE), NULL); 4542 } 4543 4544 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_ME) { 4545 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ME, 4546 DATA_TYPE_BOOLEAN_VALUE, 4547 (aflt->flt_stat & C_AFSR_ME) ? B_TRUE : B_FALSE, NULL); 4548 } 4549 4550 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SYND_STATUS) { 4551 *synd_status = afsr_to_synd_status(aflt->flt_inst, 4552 ch_flt->afsr_errs, ch_flt->flt_bit); 4553 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SYND_STATUS, 4554 DATA_TYPE_UINT8, (uint8_t)*synd_status, NULL); 4555 } 4556 4557 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SYND) { 4558 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SYND, 4559 DATA_TYPE_UINT16, (uint16_t)aflt->flt_synd, NULL); 4560 } 4561 4562 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_ERR_TYPE) { 4563 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ERR_TYPE, 4564 DATA_TYPE_STRING, flt_to_error_type(aflt), NULL); 4565 } 4566 4567 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_ERR_DISP) { 4568 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_ERR_DISP, 4569 DATA_TYPE_UINT64, aflt->flt_disp, NULL); 4570 } 4571 4572 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L2) 4573 cpu_payload_add_ecache(aflt, payload); 4574 4575 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_COPYFUNCTION) { 4576 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_COPYFUNCTION, 4577 DATA_TYPE_UINT8, (uint8_t)aflt->flt_status & 0xff, NULL); 4578 } 4579 4580 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_HOWDETECTED) { 4581 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_HOWDETECTED, 4582 DATA_TYPE_UINT8, (uint8_t)(aflt->flt_status >> 8), NULL); 4583 } 4584 4585 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_INSTRBLOCK) { 4586 fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_INSTRBLOCK, 4587 DATA_TYPE_UINT32_ARRAY, 16, 4588 (uint32_t *)&ch_flt->flt_fpdata, NULL); 4589 } 4590 4591 #if defined(CPU_IMP_L1_CACHE_PARITY) 4592 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L1D) 4593 cpu_payload_add_dcache(aflt, payload); 4594 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L1I) 4595 cpu_payload_add_icache(aflt, payload); 4596 #endif /* CPU_IMP_L1_CACHE_PARITY */ 4597 4598 #if defined(CHEETAH_PLUS) 4599 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_L1P) 4600 cpu_payload_add_pcache(aflt, payload); 4601 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAGS_TLB) 4602 cpu_payload_add_tlb(aflt, payload); 4603 #endif /* CHEETAH_PLUS */ 4604 /* 4605 * Create the FMRI that goes into the payload 4606 * and contains the unum info if necessary. 4607 */ 4608 if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_RESOURCE) { 4609 char unum[UNUM_NAMLEN] = ""; 4610 char sid[DIMM_SERIAL_ID_LEN] = ""; 4611 int len, ret, rtype, synd_code; 4612 uint64_t offset = (uint64_t)-1; 4613 4614 rtype = cpu_error_to_resource_type(aflt); 4615 switch (rtype) { 4616 4617 case ERRTYPE_MEMORY: 4618 case ERRTYPE_ECACHE_DATA: 4619 4620 /* 4621 * Memory errors, do unum lookup 4622 */ 4623 if (*afar_status == AFLT_STAT_INVALID) 4624 break; 4625 4626 if (rtype == ERRTYPE_ECACHE_DATA) 4627 aflt->flt_status |= ECC_ECACHE; 4628 else 4629 aflt->flt_status &= ~ECC_ECACHE; 4630 4631 synd_code = synd_to_synd_code(*synd_status, 4632 aflt->flt_synd, ch_flt->flt_bit); 4633 4634 if (cpu_get_mem_unum_synd(synd_code, aflt, unum) != 0) 4635 break; 4636 4637 ret = cpu_get_mem_sid(unum, sid, DIMM_SERIAL_ID_LEN, 4638 &len); 4639 4640 if (ret == 0) { 4641 (void) cpu_get_mem_offset(aflt->flt_addr, 4642 &offset); 4643 } 4644 4645 fm_fmri_mem_set(resource, FM_MEM_SCHEME_VERSION, 4646 NULL, unum, (ret == 0) ? sid : NULL, offset); 4647 fm_payload_set(payload, 4648 FM_EREPORT_PAYLOAD_NAME_RESOURCE, 4649 DATA_TYPE_NVLIST, resource, NULL); 4650 break; 4651 4652 case ERRTYPE_CPU: 4653 /* 4654 * On-board processor array error, add cpu resource. 4655 */ 4656 cpu_fmri_cpu_set(resource, aflt->flt_inst); 4657 fm_payload_set(payload, 4658 FM_EREPORT_PAYLOAD_NAME_RESOURCE, 4659 DATA_TYPE_NVLIST, resource, NULL); 4660 break; 4661 } 4662 } 4663 } 4664 4665 /* 4666 * Initialize the way info if necessary. 4667 */ 4668 void 4669 cpu_ereport_init(struct async_flt *aflt) 4670 { 4671 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 4672 ch_ec_data_t *ecp = &ch_flt->flt_diag_data.chd_ec_data[0]; 4673 ch_ec_data_t *l2p = &ch_flt->flt_diag_data.chd_l2_data[0]; 4674 int i; 4675 4676 /* 4677 * Initialize the info in the CPU logout structure. 4678 * The I$/D$ way information is not initialized here 4679 * since it is captured in the logout assembly code. 4680 */ 4681 for (i = 0; i < CHD_EC_DATA_SETS; i++) 4682 (ecp + i)->ec_way = i; 4683 4684 for (i = 0; i < PN_L2_NWAYS; i++) 4685 (l2p + i)->ec_way = i; 4686 } 4687 4688 /* 4689 * Returns whether fault address is valid for this error bit and 4690 * whether the address is "in memory" (i.e. pf_is_memory returns 1). 4691 */ 4692 int 4693 cpu_flt_in_memory(ch_async_flt_t *ch_flt, uint64_t t_afsr_bit) 4694 { 4695 struct async_flt *aflt = (struct async_flt *)ch_flt; 4696 4697 return ((t_afsr_bit & C_AFSR_MEMORY) && 4698 afsr_to_afar_status(ch_flt->afsr_errs, t_afsr_bit) == 4699 AFLT_STAT_VALID && 4700 pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT)); 4701 } 4702 4703 /* 4704 * Returns whether fault address is valid based on the error bit for the 4705 * one event being queued and whether the address is "in memory". 4706 */ 4707 static int 4708 cpu_flt_in_memory_one_event(ch_async_flt_t *ch_flt, uint64_t t_afsr_bit) 4709 { 4710 struct async_flt *aflt = (struct async_flt *)ch_flt; 4711 int afar_status; 4712 uint64_t afsr_errs, afsr_ow, *ow_bits; 4713 4714 if (!(t_afsr_bit & C_AFSR_MEMORY) || 4715 !pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT)) 4716 return (0); 4717 4718 afsr_errs = ch_flt->afsr_errs; 4719 afar_status = afsr_to_afar_status(afsr_errs, t_afsr_bit); 4720 4721 switch (afar_status) { 4722 case AFLT_STAT_VALID: 4723 return (1); 4724 4725 case AFLT_STAT_AMBIGUOUS: 4726 /* 4727 * Status is ambiguous since another error bit (or bits) 4728 * of equal priority to the specified bit on in the afsr, 4729 * so check those bits. Return 1 only if the bits on in the 4730 * same class as the t_afsr_bit are also C_AFSR_MEMORY bits. 4731 * Otherwise not all the equal priority bits are for memory 4732 * errors, so return 0. 4733 */ 4734 ow_bits = afar_overwrite; 4735 while ((afsr_ow = *ow_bits++) != 0) { 4736 /* 4737 * Get other bits that are on in t_afsr_bit's priority 4738 * class to check for Memory Error bits only. 4739 */ 4740 if (afsr_ow & t_afsr_bit) { 4741 if ((afsr_errs & afsr_ow) & ~C_AFSR_MEMORY) 4742 return (0); 4743 else 4744 return (1); 4745 } 4746 } 4747 /*FALLTHRU*/ 4748 4749 default: 4750 return (0); 4751 } 4752 } 4753 4754 static void 4755 cpu_log_diag_info(ch_async_flt_t *ch_flt) 4756 { 4757 struct async_flt *aflt = (struct async_flt *)ch_flt; 4758 ch_dc_data_t *dcp = &ch_flt->flt_diag_data.chd_dc_data; 4759 ch_ic_data_t *icp = &ch_flt->flt_diag_data.chd_ic_data; 4760 ch_ec_data_t *ecp = &ch_flt->flt_diag_data.chd_ec_data[0]; 4761 #if defined(CPU_IMP_ECACHE_ASSOC) 4762 int i, nway; 4763 #endif /* CPU_IMP_ECACHE_ASSOC */ 4764 4765 /* 4766 * Check if the CPU log out captured was valid. 4767 */ 4768 if (ch_flt->flt_diag_data.chd_afar == LOGOUT_INVALID || 4769 ch_flt->flt_data_incomplete) 4770 return; 4771 4772 #if defined(CPU_IMP_ECACHE_ASSOC) 4773 nway = cpu_ecache_nway(); 4774 i = cpu_ecache_line_valid(ch_flt); 4775 if (i == 0 || i > nway) { 4776 for (i = 0; i < nway; i++) 4777 ecp[i].ec_logflag = EC_LOGFLAG_MAGIC; 4778 } else 4779 ecp[i - 1].ec_logflag = EC_LOGFLAG_MAGIC; 4780 #else /* CPU_IMP_ECACHE_ASSOC */ 4781 ecp->ec_logflag = EC_LOGFLAG_MAGIC; 4782 #endif /* CPU_IMP_ECACHE_ASSOC */ 4783 4784 #if defined(CHEETAH_PLUS) 4785 pn_cpu_log_diag_l2_info(ch_flt); 4786 #endif /* CHEETAH_PLUS */ 4787 4788 if (CH_DCTAG_MATCH(dcp->dc_tag, aflt->flt_addr)) { 4789 dcp->dc_way = CH_DCIDX_TO_WAY(dcp->dc_idx); 4790 dcp->dc_logflag = DC_LOGFLAG_MAGIC; 4791 } 4792 4793 if (CH_ICTAG_MATCH(icp, aflt->flt_addr)) { 4794 if (IS_PANTHER(cpunodes[aflt->flt_inst].implementation)) 4795 icp->ic_way = PN_ICIDX_TO_WAY(icp->ic_idx); 4796 else 4797 icp->ic_way = CH_ICIDX_TO_WAY(icp->ic_idx); 4798 icp->ic_logflag = IC_LOGFLAG_MAGIC; 4799 } 4800 } 4801 4802 /* 4803 * Cheetah ECC calculation. 4804 * 4805 * We only need to do the calculation on the data bits and can ignore check 4806 * bit and Mtag bit terms in the calculation. 4807 */ 4808 static uint64_t ch_ecc_table[9][2] = { 4809 /* 4810 * low order 64-bits high-order 64-bits 4811 */ 4812 { 0x46bffffeccd1177f, 0x488800022100014c }, 4813 { 0x42fccc81331ff77f, 0x14424f1010249184 }, 4814 { 0x8898827c222f1ffe, 0x22c1222808184aaf }, 4815 { 0xf7632203e131ccf1, 0xe1241121848292b8 }, 4816 { 0x7f5511421b113809, 0x901c88d84288aafe }, 4817 { 0x1d49412184882487, 0x8f338c87c044c6ef }, 4818 { 0xf552181014448344, 0x7ff8f4443e411911 }, 4819 { 0x2189240808f24228, 0xfeeff8cc81333f42 }, 4820 { 0x3280008440001112, 0xfee88b337ffffd62 }, 4821 }; 4822 4823 /* 4824 * 64-bit population count, use well-known popcnt trick. 4825 * We could use the UltraSPARC V9 POPC instruction, but some 4826 * CPUs including Cheetahplus and Jaguar do not support that 4827 * instruction. 4828 */ 4829 int 4830 popc64(uint64_t val) 4831 { 4832 int cnt; 4833 4834 for (cnt = 0; val != 0; val &= val - 1) 4835 cnt++; 4836 return (cnt); 4837 } 4838 4839 /* 4840 * Generate the 9 ECC bits for the 128-bit chunk based on the table above. 4841 * Note that xor'ing an odd number of 1 bits == 1 and xor'ing an even number 4842 * of 1 bits == 0, so we can just use the least significant bit of the popcnt 4843 * instead of doing all the xor's. 4844 */ 4845 uint32_t 4846 us3_gen_ecc(uint64_t data_low, uint64_t data_high) 4847 { 4848 int bitno, s; 4849 int synd = 0; 4850 4851 for (bitno = 0; bitno < 9; bitno++) { 4852 s = (popc64(data_low & ch_ecc_table[bitno][0]) + 4853 popc64(data_high & ch_ecc_table[bitno][1])) & 1; 4854 synd |= (s << bitno); 4855 } 4856 return (synd); 4857 4858 } 4859 4860 /* 4861 * Queue one event based on ecc_type_to_info entry. If the event has an AFT1 4862 * tag associated with it or is a fatal event (aflt_panic set), it is sent to 4863 * the UE event queue. Otherwise it is dispatched to the CE event queue. 4864 */ 4865 static void 4866 cpu_queue_one_event(ch_async_flt_t *ch_flt, char *reason, 4867 ecc_type_to_info_t *eccp, ch_diag_data_t *cdp) 4868 { 4869 struct async_flt *aflt = (struct async_flt *)ch_flt; 4870 4871 if (reason && 4872 strlen(reason) + strlen(eccp->ec_reason) < MAX_REASON_STRING) { 4873 (void) strcat(reason, eccp->ec_reason); 4874 } 4875 4876 ch_flt->flt_bit = eccp->ec_afsr_bit; 4877 ch_flt->flt_type = eccp->ec_flt_type; 4878 if (cdp != NULL && cdp->chd_afar != LOGOUT_INVALID) 4879 ch_flt->flt_diag_data = *cdp; 4880 else 4881 ch_flt->flt_diag_data.chd_afar = LOGOUT_INVALID; 4882 aflt->flt_in_memory = 4883 cpu_flt_in_memory_one_event(ch_flt, ch_flt->flt_bit); 4884 4885 if (ch_flt->flt_bit & C_AFSR_MSYND_ERRS) 4886 aflt->flt_synd = GET_M_SYND(aflt->flt_stat); 4887 else if (ch_flt->flt_bit & (C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) 4888 aflt->flt_synd = GET_E_SYND(aflt->flt_stat); 4889 else 4890 aflt->flt_synd = 0; 4891 4892 aflt->flt_payload = eccp->ec_err_payload; 4893 4894 if (aflt->flt_panic || (eccp->ec_afsr_bit & 4895 (C_AFSR_LEVEL1 | C_AFSR_EXT_LEVEL1))) 4896 cpu_errorq_dispatch(eccp->ec_err_class, 4897 (void *)ch_flt, sizeof (ch_async_flt_t), ue_queue, 4898 aflt->flt_panic); 4899 else 4900 cpu_errorq_dispatch(eccp->ec_err_class, 4901 (void *)ch_flt, sizeof (ch_async_flt_t), ce_queue, 4902 aflt->flt_panic); 4903 } 4904 4905 /* 4906 * Queue events on async event queue one event per error bit. First we 4907 * queue the events that we "expect" for the given trap, then we queue events 4908 * that we may not expect. Return number of events queued. 4909 */ 4910 int 4911 cpu_queue_events(ch_async_flt_t *ch_flt, char *reason, uint64_t t_afsr_errs, 4912 ch_cpu_logout_t *clop) 4913 { 4914 struct async_flt *aflt = (struct async_flt *)ch_flt; 4915 ecc_type_to_info_t *eccp; 4916 int nevents = 0; 4917 uint64_t primary_afar = aflt->flt_addr, primary_afsr = aflt->flt_stat; 4918 #if defined(CHEETAH_PLUS) 4919 uint64_t orig_t_afsr_errs; 4920 #endif 4921 uint64_t primary_afsr_ext = ch_flt->afsr_ext; 4922 uint64_t primary_afsr_errs = ch_flt->afsr_errs; 4923 ch_diag_data_t *cdp = NULL; 4924 4925 t_afsr_errs &= ((C_AFSR_ALL_ERRS & ~C_AFSR_ME) | C_AFSR_EXT_ALL_ERRS); 4926 4927 #if defined(CHEETAH_PLUS) 4928 orig_t_afsr_errs = t_afsr_errs; 4929 4930 /* 4931 * For Cheetah+, log the shadow AFSR/AFAR bits first. 4932 */ 4933 if (clop != NULL) { 4934 /* 4935 * Set the AFSR and AFAR fields to the shadow registers. The 4936 * flt_addr and flt_stat fields will be reset to the primaries 4937 * below, but the sdw_addr and sdw_stat will stay as the 4938 * secondaries. 4939 */ 4940 cdp = &clop->clo_sdw_data; 4941 aflt->flt_addr = ch_flt->flt_sdw_afar = cdp->chd_afar; 4942 aflt->flt_stat = ch_flt->flt_sdw_afsr = cdp->chd_afsr; 4943 ch_flt->afsr_ext = ch_flt->flt_sdw_afsr_ext = cdp->chd_afsr_ext; 4944 ch_flt->afsr_errs = (cdp->chd_afsr_ext & C_AFSR_EXT_ALL_ERRS) | 4945 (cdp->chd_afsr & C_AFSR_ALL_ERRS); 4946 4947 /* 4948 * If the primary and shadow AFSR differ, tag the shadow as 4949 * the first fault. 4950 */ 4951 if ((primary_afar != cdp->chd_afar) || 4952 (primary_afsr_errs != ch_flt->afsr_errs)) { 4953 aflt->flt_stat |= (1ull << C_AFSR_FIRSTFLT_SHIFT); 4954 } 4955 4956 /* 4957 * Check AFSR bits as well as AFSR_EXT bits in order of 4958 * the AFAR overwrite priority. Our stored AFSR_EXT value 4959 * is expected to be zero for those CPUs which do not have 4960 * an AFSR_EXT register. 4961 */ 4962 for (eccp = ecc_type_to_info; eccp->ec_desc != NULL; eccp++) { 4963 if ((eccp->ec_afsr_bit & 4964 (ch_flt->afsr_errs & t_afsr_errs)) && 4965 ((eccp->ec_flags & aflt->flt_status) != 0)) { 4966 cpu_queue_one_event(ch_flt, reason, eccp, cdp); 4967 cdp = NULL; 4968 t_afsr_errs &= ~eccp->ec_afsr_bit; 4969 nevents++; 4970 } 4971 } 4972 4973 /* 4974 * If the ME bit is on in the primary AFSR turn all the 4975 * error bits on again that may set the ME bit to make 4976 * sure we see the ME AFSR error logs. 4977 */ 4978 if ((primary_afsr & C_AFSR_ME) != 0) 4979 t_afsr_errs = (orig_t_afsr_errs & C_AFSR_ALL_ME_ERRS); 4980 } 4981 #endif /* CHEETAH_PLUS */ 4982 4983 if (clop != NULL) 4984 cdp = &clop->clo_data; 4985 4986 /* 4987 * Queue expected errors, error bit and fault type must match 4988 * in the ecc_type_to_info table. 4989 */ 4990 for (eccp = ecc_type_to_info; t_afsr_errs != 0 && eccp->ec_desc != NULL; 4991 eccp++) { 4992 if ((eccp->ec_afsr_bit & t_afsr_errs) != 0 && 4993 (eccp->ec_flags & aflt->flt_status) != 0) { 4994 #if defined(SERRANO) 4995 /* 4996 * For FRC/FRU errors on Serrano the afar2 captures 4997 * the address and the associated data is 4998 * in the shadow logout area. 4999 */ 5000 if (eccp->ec_afsr_bit & (C_AFSR_FRC | C_AFSR_FRU)) { 5001 if (clop != NULL) 5002 cdp = &clop->clo_sdw_data; 5003 aflt->flt_addr = ch_flt->afar2; 5004 } else { 5005 if (clop != NULL) 5006 cdp = &clop->clo_data; 5007 aflt->flt_addr = primary_afar; 5008 } 5009 #else /* SERRANO */ 5010 aflt->flt_addr = primary_afar; 5011 #endif /* SERRANO */ 5012 aflt->flt_stat = primary_afsr; 5013 ch_flt->afsr_ext = primary_afsr_ext; 5014 ch_flt->afsr_errs = primary_afsr_errs; 5015 cpu_queue_one_event(ch_flt, reason, eccp, cdp); 5016 cdp = NULL; 5017 t_afsr_errs &= ~eccp->ec_afsr_bit; 5018 nevents++; 5019 } 5020 } 5021 5022 /* 5023 * Queue unexpected errors, error bit only match. 5024 */ 5025 for (eccp = ecc_type_to_info; t_afsr_errs != 0 && eccp->ec_desc != NULL; 5026 eccp++) { 5027 if (eccp->ec_afsr_bit & t_afsr_errs) { 5028 #if defined(SERRANO) 5029 /* 5030 * For FRC/FRU errors on Serrano the afar2 captures 5031 * the address and the associated data is 5032 * in the shadow logout area. 5033 */ 5034 if (eccp->ec_afsr_bit & (C_AFSR_FRC | C_AFSR_FRU)) { 5035 if (clop != NULL) 5036 cdp = &clop->clo_sdw_data; 5037 aflt->flt_addr = ch_flt->afar2; 5038 } else { 5039 if (clop != NULL) 5040 cdp = &clop->clo_data; 5041 aflt->flt_addr = primary_afar; 5042 } 5043 #else /* SERRANO */ 5044 aflt->flt_addr = primary_afar; 5045 #endif /* SERRANO */ 5046 aflt->flt_stat = primary_afsr; 5047 ch_flt->afsr_ext = primary_afsr_ext; 5048 ch_flt->afsr_errs = primary_afsr_errs; 5049 cpu_queue_one_event(ch_flt, reason, eccp, cdp); 5050 cdp = NULL; 5051 t_afsr_errs &= ~eccp->ec_afsr_bit; 5052 nevents++; 5053 } 5054 } 5055 return (nevents); 5056 } 5057 5058 /* 5059 * Return trap type number. 5060 */ 5061 uint8_t 5062 flt_to_trap_type(struct async_flt *aflt) 5063 { 5064 if (aflt->flt_status & ECC_I_TRAP) 5065 return (TRAP_TYPE_ECC_I); 5066 if (aflt->flt_status & ECC_D_TRAP) 5067 return (TRAP_TYPE_ECC_D); 5068 if (aflt->flt_status & ECC_F_TRAP) 5069 return (TRAP_TYPE_ECC_F); 5070 if (aflt->flt_status & ECC_C_TRAP) 5071 return (TRAP_TYPE_ECC_C); 5072 if (aflt->flt_status & ECC_DP_TRAP) 5073 return (TRAP_TYPE_ECC_DP); 5074 if (aflt->flt_status & ECC_IP_TRAP) 5075 return (TRAP_TYPE_ECC_IP); 5076 if (aflt->flt_status & ECC_ITLB_TRAP) 5077 return (TRAP_TYPE_ECC_ITLB); 5078 if (aflt->flt_status & ECC_DTLB_TRAP) 5079 return (TRAP_TYPE_ECC_DTLB); 5080 return (TRAP_TYPE_UNKNOWN); 5081 } 5082 5083 /* 5084 * Decide an error type based on detector and leaky/partner tests. 5085 * The following array is used for quick translation - it must 5086 * stay in sync with ce_dispact_t. 5087 */ 5088 5089 static char *cetypes[] = { 5090 CE_DISP_DESC_U, 5091 CE_DISP_DESC_I, 5092 CE_DISP_DESC_PP, 5093 CE_DISP_DESC_P, 5094 CE_DISP_DESC_L, 5095 CE_DISP_DESC_PS, 5096 CE_DISP_DESC_S 5097 }; 5098 5099 char * 5100 flt_to_error_type(struct async_flt *aflt) 5101 { 5102 ce_dispact_t dispact, disp; 5103 uchar_t dtcrinfo, ptnrinfo, lkyinfo; 5104 5105 /* 5106 * The memory payload bundle is shared by some events that do 5107 * not perform any classification. For those flt_disp will be 5108 * 0 and we will return "unknown". 5109 */ 5110 if (!ce_disp_inited || !aflt->flt_in_memory || aflt->flt_disp == 0) 5111 return (cetypes[CE_DISP_UNKNOWN]); 5112 5113 dtcrinfo = CE_XDIAG_DTCRINFO(aflt->flt_disp); 5114 5115 /* 5116 * It is also possible that no scrub/classification was performed 5117 * by the detector, for instance where a disrupting error logged 5118 * in the AFSR while CEEN was off in cpu_deferred_error. 5119 */ 5120 if (!CE_XDIAG_EXT_ALG_APPLIED(dtcrinfo)) 5121 return (cetypes[CE_DISP_UNKNOWN]); 5122 5123 /* 5124 * Lookup type in initial classification/action table 5125 */ 5126 dispact = CE_DISPACT(ce_disp_table, 5127 CE_XDIAG_AFARMATCHED(dtcrinfo), 5128 CE_XDIAG_STATE(dtcrinfo), 5129 CE_XDIAG_CE1SEEN(dtcrinfo), 5130 CE_XDIAG_CE2SEEN(dtcrinfo)); 5131 5132 /* 5133 * A bad lookup is not something to panic production systems for. 5134 */ 5135 ASSERT(dispact != CE_DISP_BAD); 5136 if (dispact == CE_DISP_BAD) 5137 return (cetypes[CE_DISP_UNKNOWN]); 5138 5139 disp = CE_DISP(dispact); 5140 5141 switch (disp) { 5142 case CE_DISP_UNKNOWN: 5143 case CE_DISP_INTERMITTENT: 5144 break; 5145 5146 case CE_DISP_POSS_PERS: 5147 /* 5148 * "Possible persistent" errors to which we have applied a valid 5149 * leaky test can be separated into "persistent" or "leaky". 5150 */ 5151 lkyinfo = CE_XDIAG_LKYINFO(aflt->flt_disp); 5152 if (CE_XDIAG_TESTVALID(lkyinfo)) { 5153 if (CE_XDIAG_CE1SEEN(lkyinfo) || 5154 CE_XDIAG_CE2SEEN(lkyinfo)) 5155 disp = CE_DISP_LEAKY; 5156 else 5157 disp = CE_DISP_PERS; 5158 } 5159 break; 5160 5161 case CE_DISP_POSS_STICKY: 5162 /* 5163 * Promote "possible sticky" results that have been 5164 * confirmed by a partner test to "sticky". Unconfirmed 5165 * "possible sticky" events are left at that status - we do not 5166 * guess at any bad reader/writer etc status here. 5167 */ 5168 ptnrinfo = CE_XDIAG_PTNRINFO(aflt->flt_disp); 5169 if (CE_XDIAG_TESTVALID(ptnrinfo) && 5170 CE_XDIAG_CE1SEEN(ptnrinfo) && CE_XDIAG_CE2SEEN(ptnrinfo)) 5171 disp = CE_DISP_STICKY; 5172 5173 /* 5174 * Promote "possible sticky" results on a uniprocessor 5175 * to "sticky" 5176 */ 5177 if (disp == CE_DISP_POSS_STICKY && 5178 CE_XDIAG_SKIPCODE(disp) == CE_XDIAG_SKIP_UNIPROC) 5179 disp = CE_DISP_STICKY; 5180 break; 5181 5182 default: 5183 disp = CE_DISP_UNKNOWN; 5184 break; 5185 } 5186 5187 return (cetypes[disp]); 5188 } 5189 5190 /* 5191 * Given the entire afsr, the specific bit to check and a prioritized list of 5192 * error bits, determine the validity of the various overwrite priority 5193 * features of the AFSR/AFAR: AFAR, ESYND and MSYND, each of which have 5194 * different overwrite priorities. 5195 * 5196 * Given a specific afsr error bit and the entire afsr, there are three cases: 5197 * INVALID: The specified bit is lower overwrite priority than some other 5198 * error bit which is on in the afsr (or IVU/IVC). 5199 * VALID: The specified bit is higher priority than all other error bits 5200 * which are on in the afsr. 5201 * AMBIGUOUS: Another error bit (or bits) of equal priority to the specified 5202 * bit is on in the afsr. 5203 */ 5204 int 5205 afsr_to_overw_status(uint64_t afsr, uint64_t afsr_bit, uint64_t *ow_bits) 5206 { 5207 uint64_t afsr_ow; 5208 5209 while ((afsr_ow = *ow_bits++) != 0) { 5210 /* 5211 * If bit is in the priority class, check to see if another 5212 * bit in the same class is on => ambiguous. Otherwise, 5213 * the value is valid. If the bit is not on at this priority 5214 * class, but a higher priority bit is on, then the value is 5215 * invalid. 5216 */ 5217 if (afsr_ow & afsr_bit) { 5218 /* 5219 * If equal pri bit is on, ambiguous. 5220 */ 5221 if (afsr & (afsr_ow & ~afsr_bit)) 5222 return (AFLT_STAT_AMBIGUOUS); 5223 return (AFLT_STAT_VALID); 5224 } else if (afsr & afsr_ow) 5225 break; 5226 } 5227 5228 /* 5229 * We didn't find a match or a higher priority bit was on. Not 5230 * finding a match handles the case of invalid AFAR for IVC, IVU. 5231 */ 5232 return (AFLT_STAT_INVALID); 5233 } 5234 5235 static int 5236 afsr_to_afar_status(uint64_t afsr, uint64_t afsr_bit) 5237 { 5238 #if defined(SERRANO) 5239 if (afsr_bit & (C_AFSR_FRC | C_AFSR_FRU)) 5240 return (afsr_to_overw_status(afsr, afsr_bit, afar2_overwrite)); 5241 else 5242 #endif /* SERRANO */ 5243 return (afsr_to_overw_status(afsr, afsr_bit, afar_overwrite)); 5244 } 5245 5246 static int 5247 afsr_to_esynd_status(uint64_t afsr, uint64_t afsr_bit) 5248 { 5249 return (afsr_to_overw_status(afsr, afsr_bit, esynd_overwrite)); 5250 } 5251 5252 static int 5253 afsr_to_msynd_status(uint64_t afsr, uint64_t afsr_bit) 5254 { 5255 return (afsr_to_overw_status(afsr, afsr_bit, msynd_overwrite)); 5256 } 5257 5258 static int 5259 afsr_to_synd_status(uint_t cpuid, uint64_t afsr, uint64_t afsr_bit) 5260 { 5261 #ifdef lint 5262 cpuid = cpuid; 5263 #endif 5264 #if defined(CHEETAH_PLUS) 5265 /* 5266 * The M_SYND overwrite policy is combined with the E_SYND overwrite 5267 * policy for Cheetah+ and separate for Panther CPUs. 5268 */ 5269 if (afsr_bit & C_AFSR_MSYND_ERRS) { 5270 if (IS_PANTHER(cpunodes[cpuid].implementation)) 5271 return (afsr_to_msynd_status(afsr, afsr_bit)); 5272 else 5273 return (afsr_to_esynd_status(afsr, afsr_bit)); 5274 } else if (afsr_bit & (C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) { 5275 if (IS_PANTHER(cpunodes[cpuid].implementation)) 5276 return (afsr_to_pn_esynd_status(afsr, afsr_bit)); 5277 else 5278 return (afsr_to_esynd_status(afsr, afsr_bit)); 5279 #else /* CHEETAH_PLUS */ 5280 if (afsr_bit & C_AFSR_MSYND_ERRS) { 5281 return (afsr_to_msynd_status(afsr, afsr_bit)); 5282 } else if (afsr_bit & (C_AFSR_ESYND_ERRS | C_AFSR_EXT_ESYND_ERRS)) { 5283 return (afsr_to_esynd_status(afsr, afsr_bit)); 5284 #endif /* CHEETAH_PLUS */ 5285 } else { 5286 return (AFLT_STAT_INVALID); 5287 } 5288 } 5289 5290 /* 5291 * Slave CPU stick synchronization. 5292 */ 5293 void 5294 sticksync_slave(void) 5295 { 5296 int i; 5297 int tries = 0; 5298 int64_t tskew; 5299 int64_t av_tskew; 5300 5301 kpreempt_disable(); 5302 /* wait for the master side */ 5303 while (stick_sync_cmd != SLAVE_START) 5304 ; 5305 /* 5306 * Synchronization should only take a few tries at most. But in the 5307 * odd case where the cpu isn't cooperating we'll keep trying. A cpu 5308 * without it's stick synchronized wouldn't be a good citizen. 5309 */ 5310 while (slave_done == 0) { 5311 /* 5312 * Time skew calculation. 5313 */ 5314 av_tskew = tskew = 0; 5315 5316 for (i = 0; i < stick_iter; i++) { 5317 /* make location hot */ 5318 timestamp[EV_A_START] = 0; 5319 stick_timestamp(×tamp[EV_A_START]); 5320 5321 /* tell the master we're ready */ 5322 stick_sync_cmd = MASTER_START; 5323 5324 /* and wait */ 5325 while (stick_sync_cmd != SLAVE_CONT) 5326 ; 5327 /* Event B end */ 5328 stick_timestamp(×tamp[EV_B_END]); 5329 5330 /* calculate time skew */ 5331 tskew = ((timestamp[EV_B_END] - timestamp[EV_B_START]) 5332 - (timestamp[EV_A_END] - timestamp[EV_A_START])) 5333 / 2; 5334 5335 /* keep running count */ 5336 av_tskew += tskew; 5337 } /* for */ 5338 5339 /* 5340 * Adjust stick for time skew if not within the max allowed; 5341 * otherwise we're all done. 5342 */ 5343 if (stick_iter != 0) 5344 av_tskew = av_tskew/stick_iter; 5345 if (ABS(av_tskew) > stick_tsk) { 5346 /* 5347 * If the skew is 1 (the slave's STICK register 5348 * is 1 STICK ahead of the master's), stick_adj 5349 * could fail to adjust the slave's STICK register 5350 * if the STICK read on the slave happens to 5351 * align with the increment of the STICK. 5352 * Therefore, we increment the skew to 2. 5353 */ 5354 if (av_tskew == 1) 5355 av_tskew++; 5356 stick_adj(-av_tskew); 5357 } else 5358 slave_done = 1; 5359 #ifdef DEBUG 5360 if (tries < DSYNC_ATTEMPTS) 5361 stick_sync_stats[CPU->cpu_id].skew_val[tries] = 5362 av_tskew; 5363 ++tries; 5364 #endif /* DEBUG */ 5365 #ifdef lint 5366 tries = tries; 5367 #endif 5368 5369 } /* while */ 5370 5371 /* allow the master to finish */ 5372 stick_sync_cmd = EVENT_NULL; 5373 kpreempt_enable(); 5374 } 5375 5376 /* 5377 * Master CPU side of stick synchronization. 5378 * - timestamp end of Event A 5379 * - timestamp beginning of Event B 5380 */ 5381 void 5382 sticksync_master(void) 5383 { 5384 int i; 5385 5386 kpreempt_disable(); 5387 /* tell the slave we've started */ 5388 slave_done = 0; 5389 stick_sync_cmd = SLAVE_START; 5390 5391 while (slave_done == 0) { 5392 for (i = 0; i < stick_iter; i++) { 5393 /* wait for the slave */ 5394 while (stick_sync_cmd != MASTER_START) 5395 ; 5396 /* Event A end */ 5397 stick_timestamp(×tamp[EV_A_END]); 5398 5399 /* make location hot */ 5400 timestamp[EV_B_START] = 0; 5401 stick_timestamp(×tamp[EV_B_START]); 5402 5403 /* tell the slave to continue */ 5404 stick_sync_cmd = SLAVE_CONT; 5405 } /* for */ 5406 5407 /* wait while slave calculates time skew */ 5408 while (stick_sync_cmd == SLAVE_CONT) 5409 ; 5410 } /* while */ 5411 kpreempt_enable(); 5412 } 5413 5414 /* 5415 * Cheetah/Cheetah+ have disrupting error for copyback's, so we don't need to 5416 * do Spitfire hack of xcall'ing all the cpus to ask to check for them. Also, 5417 * in cpu_async_panic_callb, each cpu checks for CPU events on its way to 5418 * panic idle. 5419 */ 5420 /*ARGSUSED*/ 5421 void 5422 cpu_check_allcpus(struct async_flt *aflt) 5423 {} 5424 5425 struct kmem_cache *ch_private_cache; 5426 5427 /* 5428 * Cpu private unitialization. Uninitialize the Ecache scrubber and 5429 * deallocate the scrubber data structures and cpu_private data structure. 5430 */ 5431 void 5432 cpu_uninit_private(struct cpu *cp) 5433 { 5434 cheetah_private_t *chprp = CPU_PRIVATE(cp); 5435 5436 ASSERT(chprp); 5437 cpu_uninit_ecache_scrub_dr(cp); 5438 CPU_PRIVATE(cp) = NULL; 5439 ch_err_tl1_paddrs[cp->cpu_id] = NULL; 5440 kmem_cache_free(ch_private_cache, chprp); 5441 cmp_delete_cpu(cp->cpu_id); 5442 5443 } 5444 5445 /* 5446 * Cheetah Cache Scrubbing 5447 * 5448 * The primary purpose of Cheetah cache scrubbing is to reduce the exposure 5449 * of E$ tags, D$ data, and I$ data to cosmic ray events since they are not 5450 * protected by either parity or ECC. 5451 * 5452 * We currently default the E$ and D$ scan rate to 100 (scan 10% of the 5453 * cache per second). Due to the the specifics of how the I$ control 5454 * logic works with respect to the ASI used to scrub I$ lines, the entire 5455 * I$ is scanned at once. 5456 */ 5457 5458 /* 5459 * Tuneables to enable and disable the scrubbing of the caches, and to tune 5460 * scrubbing behavior. These may be changed via /etc/system or using mdb 5461 * on a running system. 5462 */ 5463 int dcache_scrub_enable = 1; /* D$ scrubbing is on by default */ 5464 5465 /* 5466 * The following are the PIL levels that the softints/cross traps will fire at. 5467 */ 5468 uint_t ecache_scrub_pil = PIL_9; /* E$ scrub PIL for cross traps */ 5469 uint_t dcache_scrub_pil = PIL_9; /* D$ scrub PIL for cross traps */ 5470 uint_t icache_scrub_pil = PIL_9; /* I$ scrub PIL for cross traps */ 5471 5472 #if defined(JALAPENO) 5473 5474 /* 5475 * Due to several errata (82, 85, 86), we don't enable the L2$ scrubber 5476 * on Jalapeno. 5477 */ 5478 int ecache_scrub_enable = 0; 5479 5480 #else /* JALAPENO */ 5481 5482 /* 5483 * With all other cpu types, E$ scrubbing is on by default 5484 */ 5485 int ecache_scrub_enable = 1; 5486 5487 #endif /* JALAPENO */ 5488 5489 5490 #if defined(CHEETAH_PLUS) || defined(JALAPENO) || defined(SERRANO) 5491 5492 /* 5493 * The I$ scrubber tends to cause latency problems for real-time SW, so it 5494 * is disabled by default on non-Cheetah systems 5495 */ 5496 int icache_scrub_enable = 0; 5497 5498 /* 5499 * Tuneables specifying the scrub calls per second and the scan rate 5500 * for each cache 5501 * 5502 * The cyclic times are set during boot based on the following values. 5503 * Changing these values in mdb after this time will have no effect. If 5504 * a different value is desired, it must be set in /etc/system before a 5505 * reboot. 5506 */ 5507 int ecache_calls_a_sec = 1; 5508 int dcache_calls_a_sec = 2; 5509 int icache_calls_a_sec = 2; 5510 5511 int ecache_scan_rate_idle = 1; 5512 int ecache_scan_rate_busy = 1; 5513 int dcache_scan_rate_idle = 1; 5514 int dcache_scan_rate_busy = 1; 5515 int icache_scan_rate_idle = 1; 5516 int icache_scan_rate_busy = 1; 5517 5518 #else /* CHEETAH_PLUS || JALAPENO || SERRANO */ 5519 5520 int icache_scrub_enable = 1; /* I$ scrubbing is on by default */ 5521 5522 int ecache_calls_a_sec = 100; /* E$ scrub calls per seconds */ 5523 int dcache_calls_a_sec = 100; /* D$ scrub calls per seconds */ 5524 int icache_calls_a_sec = 100; /* I$ scrub calls per seconds */ 5525 5526 int ecache_scan_rate_idle = 100; /* E$ scan rate (in tenths of a %) */ 5527 int ecache_scan_rate_busy = 100; /* E$ scan rate (in tenths of a %) */ 5528 int dcache_scan_rate_idle = 100; /* D$ scan rate (in tenths of a %) */ 5529 int dcache_scan_rate_busy = 100; /* D$ scan rate (in tenths of a %) */ 5530 int icache_scan_rate_idle = 100; /* I$ scan rate (in tenths of a %) */ 5531 int icache_scan_rate_busy = 100; /* I$ scan rate (in tenths of a %) */ 5532 5533 #endif /* CHEETAH_PLUS || JALAPENO || SERRANO */ 5534 5535 /* 5536 * In order to scrub on offline cpus, a cross trap is sent. The handler will 5537 * increment the outstanding request counter and schedule a softint to run 5538 * the scrubber. 5539 */ 5540 extern xcfunc_t cache_scrubreq_tl1; 5541 5542 /* 5543 * These are the softint functions for each cache scrubber 5544 */ 5545 static uint_t scrub_ecache_line_intr(caddr_t arg1, caddr_t arg2); 5546 static uint_t scrub_dcache_line_intr(caddr_t arg1, caddr_t arg2); 5547 static uint_t scrub_icache_line_intr(caddr_t arg1, caddr_t arg2); 5548 5549 /* 5550 * The cache scrub info table contains cache specific information 5551 * and allows for some of the scrub code to be table driven, reducing 5552 * duplication of cache similar code. 5553 * 5554 * This table keeps a copy of the value in the calls per second variable 5555 * (?cache_calls_a_sec). This makes it much more difficult for someone 5556 * to cause us problems (for example, by setting ecache_calls_a_sec to 0 in 5557 * mdb in a misguided attempt to disable the scrubber). 5558 */ 5559 struct scrub_info { 5560 int *csi_enable; /* scrubber enable flag */ 5561 int csi_freq; /* scrubber calls per second */ 5562 int csi_index; /* index to chsm_outstanding[] */ 5563 uint64_t csi_inum; /* scrubber interrupt number */ 5564 cyclic_id_t csi_omni_cyc_id; /* omni cyclic ID */ 5565 cyclic_id_t csi_offline_cyc_id; /* offline cyclic ID */ 5566 char csi_name[3]; /* cache name for this scrub entry */ 5567 } cache_scrub_info[] = { 5568 { &ecache_scrub_enable, 0, CACHE_SCRUBBER_INFO_E, 0, 0, 0, "E$"}, 5569 { &dcache_scrub_enable, 0, CACHE_SCRUBBER_INFO_D, 0, 0, 0, "D$"}, 5570 { &icache_scrub_enable, 0, CACHE_SCRUBBER_INFO_I, 0, 0, 0, "I$"} 5571 }; 5572 5573 /* 5574 * If scrubbing is enabled, increment the outstanding request counter. If it 5575 * is 1 (meaning there were no previous requests outstanding), call 5576 * setsoftint_tl1 through xt_one_unchecked, which eventually ends up doing 5577 * a self trap. 5578 */ 5579 static void 5580 do_scrub(struct scrub_info *csi) 5581 { 5582 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5583 int index = csi->csi_index; 5584 uint32_t *outstanding = &csmp->chsm_outstanding[index]; 5585 5586 if (*(csi->csi_enable) && (csmp->chsm_enable[index])) { 5587 if (atomic_add_32_nv(outstanding, 1) == 1) { 5588 xt_one_unchecked(CPU->cpu_id, setsoftint_tl1, 5589 csi->csi_inum, 0); 5590 } 5591 } 5592 } 5593 5594 /* 5595 * Omni cyclics don't fire on offline cpus, so we use another cyclic to 5596 * cross-trap the offline cpus. 5597 */ 5598 static void 5599 do_scrub_offline(struct scrub_info *csi) 5600 { 5601 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5602 5603 if (CPUSET_ISNULL(cpu_offline_set)) { 5604 /* 5605 * No offline cpus - nothing to do 5606 */ 5607 return; 5608 } 5609 5610 if (*(csi->csi_enable) && (csmp->chsm_enable[csi->csi_index])) { 5611 xt_some(cpu_offline_set, cache_scrubreq_tl1, csi->csi_inum, 5612 csi->csi_index); 5613 } 5614 } 5615 5616 /* 5617 * This is the initial setup for the scrubber cyclics - it sets the 5618 * interrupt level, frequency, and function to call. 5619 */ 5620 /*ARGSUSED*/ 5621 static void 5622 cpu_scrub_cyclic_setup(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, 5623 cyc_time_t *when) 5624 { 5625 struct scrub_info *csi = (struct scrub_info *)arg; 5626 5627 ASSERT(csi != NULL); 5628 hdlr->cyh_func = (cyc_func_t)do_scrub; 5629 hdlr->cyh_level = CY_LOW_LEVEL; 5630 hdlr->cyh_arg = arg; 5631 5632 when->cyt_when = 0; /* Start immediately */ 5633 when->cyt_interval = NANOSEC / csi->csi_freq; 5634 } 5635 5636 /* 5637 * Initialization for cache scrubbing. 5638 * This routine is called AFTER all cpus have had cpu_init_private called 5639 * to initialize their private data areas. 5640 */ 5641 void 5642 cpu_init_cache_scrub(void) 5643 { 5644 int i; 5645 struct scrub_info *csi; 5646 cyc_omni_handler_t omni_hdlr; 5647 cyc_handler_t offline_hdlr; 5648 cyc_time_t when; 5649 5650 /* 5651 * save away the maximum number of lines for the D$ 5652 */ 5653 dcache_nlines = dcache_size / dcache_linesize; 5654 5655 /* 5656 * register the softints for the cache scrubbing 5657 */ 5658 cache_scrub_info[CACHE_SCRUBBER_INFO_E].csi_inum = 5659 add_softintr(ecache_scrub_pil, scrub_ecache_line_intr, 5660 (caddr_t)&cache_scrub_info[CACHE_SCRUBBER_INFO_E], SOFTINT_MT); 5661 cache_scrub_info[CACHE_SCRUBBER_INFO_E].csi_freq = ecache_calls_a_sec; 5662 5663 cache_scrub_info[CACHE_SCRUBBER_INFO_D].csi_inum = 5664 add_softintr(dcache_scrub_pil, scrub_dcache_line_intr, 5665 (caddr_t)&cache_scrub_info[CACHE_SCRUBBER_INFO_D], SOFTINT_MT); 5666 cache_scrub_info[CACHE_SCRUBBER_INFO_D].csi_freq = dcache_calls_a_sec; 5667 5668 cache_scrub_info[CACHE_SCRUBBER_INFO_I].csi_inum = 5669 add_softintr(icache_scrub_pil, scrub_icache_line_intr, 5670 (caddr_t)&cache_scrub_info[CACHE_SCRUBBER_INFO_I], SOFTINT_MT); 5671 cache_scrub_info[CACHE_SCRUBBER_INFO_I].csi_freq = icache_calls_a_sec; 5672 5673 /* 5674 * start the scrubbing for all the caches 5675 */ 5676 mutex_enter(&cpu_lock); 5677 for (i = 0; i < CACHE_SCRUBBER_COUNT; i++) { 5678 5679 csi = &cache_scrub_info[i]; 5680 5681 if (!(*csi->csi_enable)) 5682 continue; 5683 5684 /* 5685 * force the following to be true: 5686 * 1 <= calls_a_sec <= hz 5687 */ 5688 if (csi->csi_freq > hz) { 5689 cmn_err(CE_NOTE, "%s scrub calls_a_sec set too high " 5690 "(%d); resetting to hz (%d)", csi->csi_name, 5691 csi->csi_freq, hz); 5692 csi->csi_freq = hz; 5693 } else if (csi->csi_freq < 1) { 5694 cmn_err(CE_NOTE, "%s scrub calls_a_sec set too low " 5695 "(%d); resetting to 1", csi->csi_name, 5696 csi->csi_freq); 5697 csi->csi_freq = 1; 5698 } 5699 5700 omni_hdlr.cyo_online = cpu_scrub_cyclic_setup; 5701 omni_hdlr.cyo_offline = NULL; 5702 omni_hdlr.cyo_arg = (void *)csi; 5703 5704 offline_hdlr.cyh_func = (cyc_func_t)do_scrub_offline; 5705 offline_hdlr.cyh_arg = (void *)csi; 5706 offline_hdlr.cyh_level = CY_LOW_LEVEL; 5707 5708 when.cyt_when = 0; /* Start immediately */ 5709 when.cyt_interval = NANOSEC / csi->csi_freq; 5710 5711 csi->csi_omni_cyc_id = cyclic_add_omni(&omni_hdlr); 5712 csi->csi_offline_cyc_id = cyclic_add(&offline_hdlr, &when); 5713 } 5714 register_cpu_setup_func(cpu_scrub_cpu_setup, NULL); 5715 mutex_exit(&cpu_lock); 5716 } 5717 5718 /* 5719 * Indicate that the specified cpu is idle. 5720 */ 5721 void 5722 cpu_idle_ecache_scrub(struct cpu *cp) 5723 { 5724 if (CPU_PRIVATE(cp) != NULL) { 5725 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc); 5726 csmp->chsm_ecache_busy = ECACHE_CPU_IDLE; 5727 } 5728 } 5729 5730 /* 5731 * Indicate that the specified cpu is busy. 5732 */ 5733 void 5734 cpu_busy_ecache_scrub(struct cpu *cp) 5735 { 5736 if (CPU_PRIVATE(cp) != NULL) { 5737 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc); 5738 csmp->chsm_ecache_busy = ECACHE_CPU_BUSY; 5739 } 5740 } 5741 5742 /* 5743 * Initialization for cache scrubbing for the specified cpu. 5744 */ 5745 void 5746 cpu_init_ecache_scrub_dr(struct cpu *cp) 5747 { 5748 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc); 5749 int cpuid = cp->cpu_id; 5750 5751 /* initialize the number of lines in the caches */ 5752 csmp->chsm_ecache_nlines = cpunodes[cpuid].ecache_size / 5753 cpunodes[cpuid].ecache_linesize; 5754 csmp->chsm_icache_nlines = CPU_PRIVATE_VAL(cp, chpr_icache_size) / 5755 CPU_PRIVATE_VAL(cp, chpr_icache_linesize); 5756 5757 /* 5758 * do_scrub() and do_scrub_offline() check both the global 5759 * ?cache_scrub_enable and this per-cpu enable variable. All scrubbers 5760 * check this value before scrubbing. Currently, we use it to 5761 * disable the E$ scrubber on multi-core cpus or while running at 5762 * slowed speed. For now, just turn everything on and allow 5763 * cpu_init_private() to change it if necessary. 5764 */ 5765 csmp->chsm_enable[CACHE_SCRUBBER_INFO_E] = 1; 5766 csmp->chsm_enable[CACHE_SCRUBBER_INFO_D] = 1; 5767 csmp->chsm_enable[CACHE_SCRUBBER_INFO_I] = 1; 5768 5769 cpu_busy_ecache_scrub(cp); 5770 } 5771 5772 /* 5773 * Un-initialization for cache scrubbing for the specified cpu. 5774 */ 5775 static void 5776 cpu_uninit_ecache_scrub_dr(struct cpu *cp) 5777 { 5778 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(cp, chpr_scrub_misc); 5779 5780 /* 5781 * un-initialize bookkeeping for cache scrubbing 5782 */ 5783 bzero(csmp, sizeof (ch_scrub_misc_t)); 5784 5785 cpu_idle_ecache_scrub(cp); 5786 } 5787 5788 /* 5789 * Called periodically on each CPU to scrub the D$. 5790 */ 5791 static void 5792 scrub_dcache(int how_many) 5793 { 5794 int i; 5795 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5796 int index = csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_D]; 5797 5798 /* 5799 * scrub the desired number of lines 5800 */ 5801 for (i = 0; i < how_many; i++) { 5802 /* 5803 * scrub a D$ line 5804 */ 5805 dcache_inval_line(index); 5806 5807 /* 5808 * calculate the next D$ line to scrub, assumes 5809 * that dcache_nlines is a power of 2 5810 */ 5811 index = (index + 1) & (dcache_nlines - 1); 5812 } 5813 5814 /* 5815 * set the scrub index for the next visit 5816 */ 5817 csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_D] = index; 5818 } 5819 5820 /* 5821 * Handler for D$ scrub inum softint. Call scrub_dcache until 5822 * we decrement the outstanding request count to zero. 5823 */ 5824 /*ARGSUSED*/ 5825 static uint_t 5826 scrub_dcache_line_intr(caddr_t arg1, caddr_t arg2) 5827 { 5828 int i; 5829 int how_many; 5830 int outstanding; 5831 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5832 uint32_t *countp = &csmp->chsm_outstanding[CACHE_SCRUBBER_INFO_D]; 5833 struct scrub_info *csi = (struct scrub_info *)arg1; 5834 int scan_rate = (csmp->chsm_ecache_busy == ECACHE_CPU_IDLE) ? 5835 dcache_scan_rate_idle : dcache_scan_rate_busy; 5836 5837 /* 5838 * The scan rates are expressed in units of tenths of a 5839 * percent. A scan rate of 1000 (100%) means the whole 5840 * cache is scanned every second. 5841 */ 5842 how_many = (dcache_nlines * scan_rate) / (1000 * csi->csi_freq); 5843 5844 do { 5845 outstanding = *countp; 5846 for (i = 0; i < outstanding; i++) { 5847 scrub_dcache(how_many); 5848 } 5849 } while (atomic_add_32_nv(countp, -outstanding)); 5850 5851 return (DDI_INTR_CLAIMED); 5852 } 5853 5854 /* 5855 * Called periodically on each CPU to scrub the I$. The I$ is scrubbed 5856 * by invalidating lines. Due to the characteristics of the ASI which 5857 * is used to invalidate an I$ line, the entire I$ must be invalidated 5858 * vs. an individual I$ line. 5859 */ 5860 static void 5861 scrub_icache(int how_many) 5862 { 5863 int i; 5864 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5865 int index = csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_I]; 5866 int icache_nlines = csmp->chsm_icache_nlines; 5867 5868 /* 5869 * scrub the desired number of lines 5870 */ 5871 for (i = 0; i < how_many; i++) { 5872 /* 5873 * since the entire I$ must be scrubbed at once, 5874 * wait until the index wraps to zero to invalidate 5875 * the entire I$ 5876 */ 5877 if (index == 0) { 5878 icache_inval_all(); 5879 } 5880 5881 /* 5882 * calculate the next I$ line to scrub, assumes 5883 * that chsm_icache_nlines is a power of 2 5884 */ 5885 index = (index + 1) & (icache_nlines - 1); 5886 } 5887 5888 /* 5889 * set the scrub index for the next visit 5890 */ 5891 csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_I] = index; 5892 } 5893 5894 /* 5895 * Handler for I$ scrub inum softint. Call scrub_icache until 5896 * we decrement the outstanding request count to zero. 5897 */ 5898 /*ARGSUSED*/ 5899 static uint_t 5900 scrub_icache_line_intr(caddr_t arg1, caddr_t arg2) 5901 { 5902 int i; 5903 int how_many; 5904 int outstanding; 5905 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5906 uint32_t *countp = &csmp->chsm_outstanding[CACHE_SCRUBBER_INFO_I]; 5907 struct scrub_info *csi = (struct scrub_info *)arg1; 5908 int scan_rate = (csmp->chsm_ecache_busy == ECACHE_CPU_IDLE) ? 5909 icache_scan_rate_idle : icache_scan_rate_busy; 5910 int icache_nlines = csmp->chsm_icache_nlines; 5911 5912 /* 5913 * The scan rates are expressed in units of tenths of a 5914 * percent. A scan rate of 1000 (100%) means the whole 5915 * cache is scanned every second. 5916 */ 5917 how_many = (icache_nlines * scan_rate) / (1000 * csi->csi_freq); 5918 5919 do { 5920 outstanding = *countp; 5921 for (i = 0; i < outstanding; i++) { 5922 scrub_icache(how_many); 5923 } 5924 } while (atomic_add_32_nv(countp, -outstanding)); 5925 5926 return (DDI_INTR_CLAIMED); 5927 } 5928 5929 /* 5930 * Called periodically on each CPU to scrub the E$. 5931 */ 5932 static void 5933 scrub_ecache(int how_many) 5934 { 5935 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5936 int i; 5937 int cpuid = CPU->cpu_id; 5938 int index = csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_E]; 5939 int nlines = csmp->chsm_ecache_nlines; 5940 int linesize = cpunodes[cpuid].ecache_linesize; 5941 int ec_set_size = cpu_ecache_set_size(CPU); 5942 5943 /* 5944 * scrub the desired number of lines 5945 */ 5946 for (i = 0; i < how_many; i++) { 5947 /* 5948 * scrub the E$ line 5949 */ 5950 ecache_flush_line(ecache_flushaddr + (index * linesize), 5951 ec_set_size); 5952 5953 /* 5954 * calculate the next E$ line to scrub based on twice 5955 * the number of E$ lines (to displace lines containing 5956 * flush area data), assumes that the number of lines 5957 * is a power of 2 5958 */ 5959 index = (index + 1) & ((nlines << 1) - 1); 5960 } 5961 5962 /* 5963 * set the ecache scrub index for the next visit 5964 */ 5965 csmp->chsm_flush_index[CACHE_SCRUBBER_INFO_E] = index; 5966 } 5967 5968 /* 5969 * Handler for E$ scrub inum softint. Call the E$ scrubber until 5970 * we decrement the outstanding request count to zero. 5971 * 5972 * Due to interactions with cpu_scrub_cpu_setup(), the outstanding count may 5973 * become negative after the atomic_add_32_nv(). This is not a problem, as 5974 * the next trip around the loop won't scrub anything, and the next add will 5975 * reset the count back to zero. 5976 */ 5977 /*ARGSUSED*/ 5978 static uint_t 5979 scrub_ecache_line_intr(caddr_t arg1, caddr_t arg2) 5980 { 5981 int i; 5982 int how_many; 5983 int outstanding; 5984 ch_scrub_misc_t *csmp = CPU_PRIVATE_PTR(CPU, chpr_scrub_misc); 5985 uint32_t *countp = &csmp->chsm_outstanding[CACHE_SCRUBBER_INFO_E]; 5986 struct scrub_info *csi = (struct scrub_info *)arg1; 5987 int scan_rate = (csmp->chsm_ecache_busy == ECACHE_CPU_IDLE) ? 5988 ecache_scan_rate_idle : ecache_scan_rate_busy; 5989 int ecache_nlines = csmp->chsm_ecache_nlines; 5990 5991 /* 5992 * The scan rates are expressed in units of tenths of a 5993 * percent. A scan rate of 1000 (100%) means the whole 5994 * cache is scanned every second. 5995 */ 5996 how_many = (ecache_nlines * scan_rate) / (1000 * csi->csi_freq); 5997 5998 do { 5999 outstanding = *countp; 6000 for (i = 0; i < outstanding; i++) { 6001 scrub_ecache(how_many); 6002 } 6003 } while (atomic_add_32_nv(countp, -outstanding)); 6004 6005 return (DDI_INTR_CLAIMED); 6006 } 6007 6008 /* 6009 * Timeout function to reenable CE 6010 */ 6011 static void 6012 cpu_delayed_check_ce_errors(void *arg) 6013 { 6014 if (!taskq_dispatch(ch_check_ce_tq, cpu_check_ce_errors, arg, 6015 TQ_NOSLEEP)) { 6016 (void) timeout(cpu_delayed_check_ce_errors, arg, 6017 drv_usectohz((clock_t)cpu_ceen_delay_secs * MICROSEC)); 6018 } 6019 } 6020 6021 /* 6022 * CE Deferred Re-enable after trap. 6023 * 6024 * When the CPU gets a disrupting trap for any of the errors 6025 * controlled by the CEEN bit, CEEN is disabled in the trap handler 6026 * immediately. To eliminate the possibility of multiple CEs causing 6027 * recursive stack overflow in the trap handler, we cannot 6028 * reenable CEEN while still running in the trap handler. Instead, 6029 * after a CE is logged on a CPU, we schedule a timeout function, 6030 * cpu_check_ce_errors(), to trigger after cpu_ceen_delay_secs 6031 * seconds. This function will check whether any further CEs 6032 * have occurred on that CPU, and if none have, will reenable CEEN. 6033 * 6034 * If further CEs have occurred while CEEN is disabled, another 6035 * timeout will be scheduled. This is to ensure that the CPU can 6036 * make progress in the face of CE 'storms', and that it does not 6037 * spend all its time logging CE errors. 6038 */ 6039 static void 6040 cpu_check_ce_errors(void *arg) 6041 { 6042 int cpuid = (int)(uintptr_t)arg; 6043 cpu_t *cp; 6044 6045 /* 6046 * We acquire cpu_lock. 6047 */ 6048 ASSERT(curthread->t_pil == 0); 6049 6050 /* 6051 * verify that the cpu is still around, DR 6052 * could have got there first ... 6053 */ 6054 mutex_enter(&cpu_lock); 6055 cp = cpu_get(cpuid); 6056 if (cp == NULL) { 6057 mutex_exit(&cpu_lock); 6058 return; 6059 } 6060 /* 6061 * make sure we don't migrate across CPUs 6062 * while checking our CE status. 6063 */ 6064 kpreempt_disable(); 6065 6066 /* 6067 * If we are running on the CPU that got the 6068 * CE, we can do the checks directly. 6069 */ 6070 if (cp->cpu_id == CPU->cpu_id) { 6071 mutex_exit(&cpu_lock); 6072 cpu_check_ce(TIMEOUT_CEEN_CHECK, 0, 0, 0); 6073 kpreempt_enable(); 6074 return; 6075 } 6076 kpreempt_enable(); 6077 6078 /* 6079 * send an x-call to get the CPU that originally 6080 * got the CE to do the necessary checks. If we can't 6081 * send the x-call, reschedule the timeout, otherwise we 6082 * lose CEEN forever on that CPU. 6083 */ 6084 if (CPU_XCALL_READY(cp->cpu_id) && (!(cp->cpu_flags & CPU_QUIESCED))) { 6085 xc_one(cp->cpu_id, (xcfunc_t *)cpu_check_ce, 6086 TIMEOUT_CEEN_CHECK, 0); 6087 mutex_exit(&cpu_lock); 6088 } else { 6089 /* 6090 * When the CPU is not accepting xcalls, or 6091 * the processor is offlined, we don't want to 6092 * incur the extra overhead of trying to schedule the 6093 * CE timeout indefinitely. However, we don't want to lose 6094 * CE checking forever. 6095 * 6096 * Keep rescheduling the timeout, accepting the additional 6097 * overhead as the cost of correctness in the case where we get 6098 * a CE, disable CEEN, offline the CPU during the 6099 * the timeout interval, and then online it at some 6100 * point in the future. This is unlikely given the short 6101 * cpu_ceen_delay_secs. 6102 */ 6103 mutex_exit(&cpu_lock); 6104 (void) timeout(cpu_delayed_check_ce_errors, 6105 (void *)(uintptr_t)cp->cpu_id, 6106 drv_usectohz((clock_t)cpu_ceen_delay_secs * MICROSEC)); 6107 } 6108 } 6109 6110 /* 6111 * This routine will check whether CEs have occurred while 6112 * CEEN is disabled. Any CEs detected will be logged and, if 6113 * possible, scrubbed. 6114 * 6115 * The memscrubber will also use this routine to clear any errors 6116 * caused by its scrubbing with CEEN disabled. 6117 * 6118 * flag == SCRUBBER_CEEN_CHECK 6119 * called from memscrubber, just check/scrub, no reset 6120 * paddr physical addr. for start of scrub pages 6121 * vaddr virtual addr. for scrub area 6122 * psz page size of area to be scrubbed 6123 * 6124 * flag == TIMEOUT_CEEN_CHECK 6125 * timeout function has triggered, reset timeout or CEEN 6126 * 6127 * Note: We must not migrate cpus during this function. This can be 6128 * achieved by one of: 6129 * - invoking as target of an x-call in which case we're at XCALL_PIL 6130 * The flag value must be first xcall argument. 6131 * - disabling kernel preemption. This should be done for very short 6132 * periods so is not suitable for SCRUBBER_CEEN_CHECK where we might 6133 * scrub an extended area with cpu_check_block. The call for 6134 * TIMEOUT_CEEN_CHECK uses this so cpu_check_ce must be kept 6135 * brief for this case. 6136 * - binding to a cpu, eg with thread_affinity_set(). This is used 6137 * in the SCRUBBER_CEEN_CHECK case, but is not practical for 6138 * the TIMEOUT_CEEN_CHECK because both need cpu_lock. 6139 */ 6140 void 6141 cpu_check_ce(int flag, uint64_t pa, caddr_t va, uint_t psz) 6142 { 6143 ch_cpu_errors_t cpu_error_regs; 6144 uint64_t ec_err_enable; 6145 uint64_t page_offset; 6146 6147 /* Read AFSR */ 6148 get_cpu_error_state(&cpu_error_regs); 6149 6150 /* 6151 * If no CEEN errors have occurred during the timeout 6152 * interval, it is safe to re-enable CEEN and exit. 6153 */ 6154 if (((cpu_error_regs.afsr & C_AFSR_CECC_ERRS) | 6155 (cpu_error_regs.afsr_ext & C_AFSR_EXT_CECC_ERRS)) == 0) { 6156 if (flag == TIMEOUT_CEEN_CHECK && 6157 !((ec_err_enable = get_error_enable()) & EN_REG_CEEN)) 6158 set_error_enable(ec_err_enable | EN_REG_CEEN); 6159 return; 6160 } 6161 6162 /* 6163 * Ensure that CEEN was not reenabled (maybe by DR) before 6164 * we log/clear the error. 6165 */ 6166 if ((ec_err_enable = get_error_enable()) & EN_REG_CEEN) 6167 set_error_enable(ec_err_enable & ~EN_REG_CEEN); 6168 6169 /* 6170 * log/clear the CE. If CE_CEEN_DEFER is passed, the 6171 * timeout will be rescheduled when the error is logged. 6172 */ 6173 if (!((cpu_error_regs.afsr & cpu_ce_not_deferred) | 6174 (cpu_error_regs.afsr_ext & cpu_ce_not_deferred_ext))) 6175 cpu_ce_detected(&cpu_error_regs, 6176 CE_CEEN_DEFER | CE_CEEN_TIMEOUT); 6177 else 6178 cpu_ce_detected(&cpu_error_regs, CE_CEEN_TIMEOUT); 6179 6180 /* 6181 * If the memory scrubber runs while CEEN is 6182 * disabled, (or if CEEN is disabled during the 6183 * scrub as a result of a CE being triggered by 6184 * it), the range being scrubbed will not be 6185 * completely cleaned. If there are multiple CEs 6186 * in the range at most two of these will be dealt 6187 * with, (one by the trap handler and one by the 6188 * timeout). It is also possible that none are dealt 6189 * with, (CEEN disabled and another CE occurs before 6190 * the timeout triggers). So to ensure that the 6191 * memory is actually scrubbed, we have to access each 6192 * memory location in the range and then check whether 6193 * that access causes a CE. 6194 */ 6195 if (flag == SCRUBBER_CEEN_CHECK && va) { 6196 if ((cpu_error_regs.afar >= pa) && 6197 (cpu_error_regs.afar < (pa + psz))) { 6198 /* 6199 * Force a load from physical memory for each 6200 * 64-byte block, then check AFSR to determine 6201 * whether this access caused an error. 6202 * 6203 * This is a slow way to do a scrub, but as it will 6204 * only be invoked when the memory scrubber actually 6205 * triggered a CE, it should not happen too 6206 * frequently. 6207 * 6208 * cut down what we need to check as the scrubber 6209 * has verified up to AFAR, so get it's offset 6210 * into the page and start there. 6211 */ 6212 page_offset = (uint64_t)(cpu_error_regs.afar & 6213 (psz - 1)); 6214 va = (caddr_t)(va + (P2ALIGN(page_offset, 64))); 6215 psz -= (uint_t)(P2ALIGN(page_offset, 64)); 6216 cpu_check_block((caddr_t)(P2ALIGN((uint64_t)va, 64)), 6217 psz); 6218 } 6219 } 6220 6221 /* 6222 * Reset error enable if this CE is not masked. 6223 */ 6224 if ((flag == TIMEOUT_CEEN_CHECK) && 6225 (cpu_error_regs.afsr & cpu_ce_not_deferred)) 6226 set_error_enable(ec_err_enable | EN_REG_CEEN); 6227 6228 } 6229 6230 /* 6231 * Attempt a cpu logout for an error that we did not trap for, such 6232 * as a CE noticed with CEEN off. It is assumed that we are still running 6233 * on the cpu that took the error and that we cannot migrate. Returns 6234 * 0 on success, otherwise nonzero. 6235 */ 6236 static int 6237 cpu_ce_delayed_ec_logout(uint64_t afar) 6238 { 6239 ch_cpu_logout_t *clop; 6240 6241 if (CPU_PRIVATE(CPU) == NULL) 6242 return (0); 6243 6244 clop = CPU_PRIVATE_PTR(CPU, chpr_cecc_logout); 6245 if (cas64(&clop->clo_data.chd_afar, LOGOUT_INVALID, afar) != 6246 LOGOUT_INVALID) 6247 return (0); 6248 6249 cpu_delayed_logout(afar, clop); 6250 return (1); 6251 } 6252 6253 /* 6254 * We got an error while CEEN was disabled. We 6255 * need to clean up after it and log whatever 6256 * information we have on the CE. 6257 */ 6258 void 6259 cpu_ce_detected(ch_cpu_errors_t *cpu_error_regs, int flag) 6260 { 6261 ch_async_flt_t ch_flt; 6262 struct async_flt *aflt; 6263 char pr_reason[MAX_REASON_STRING]; 6264 6265 bzero(&ch_flt, sizeof (ch_async_flt_t)); 6266 ch_flt.flt_trapped_ce = flag; 6267 aflt = (struct async_flt *)&ch_flt; 6268 aflt->flt_stat = cpu_error_regs->afsr & C_AFSR_MASK; 6269 ch_flt.afsr_ext = cpu_error_regs->afsr_ext; 6270 ch_flt.afsr_errs = (cpu_error_regs->afsr_ext & C_AFSR_EXT_ALL_ERRS) | 6271 (cpu_error_regs->afsr & C_AFSR_ALL_ERRS); 6272 aflt->flt_addr = cpu_error_regs->afar; 6273 #if defined(SERRANO) 6274 ch_flt.afar2 = cpu_error_regs->afar2; 6275 #endif /* SERRANO */ 6276 aflt->flt_pc = NULL; 6277 aflt->flt_priv = ((cpu_error_regs->afsr & C_AFSR_PRIV) != 0); 6278 aflt->flt_tl = 0; 6279 aflt->flt_panic = 0; 6280 cpu_log_and_clear_ce(&ch_flt); 6281 6282 /* 6283 * check if we caused any errors during cleanup 6284 */ 6285 if (clear_errors(&ch_flt)) { 6286 pr_reason[0] = '\0'; 6287 (void) cpu_queue_events(&ch_flt, pr_reason, ch_flt.afsr_errs, 6288 NULL); 6289 } 6290 } 6291 6292 /* 6293 * Log/clear CEEN-controlled disrupting errors 6294 */ 6295 static void 6296 cpu_log_and_clear_ce(ch_async_flt_t *ch_flt) 6297 { 6298 struct async_flt *aflt; 6299 uint64_t afsr, afsr_errs; 6300 ch_cpu_logout_t *clop; 6301 char pr_reason[MAX_REASON_STRING]; 6302 on_trap_data_t *otp = curthread->t_ontrap; 6303 6304 aflt = (struct async_flt *)ch_flt; 6305 afsr = aflt->flt_stat; 6306 afsr_errs = ch_flt->afsr_errs; 6307 aflt->flt_id = gethrtime_waitfree(); 6308 aflt->flt_bus_id = getprocessorid(); 6309 aflt->flt_inst = CPU->cpu_id; 6310 aflt->flt_prot = AFLT_PROT_NONE; 6311 aflt->flt_class = CPU_FAULT; 6312 aflt->flt_status = ECC_C_TRAP; 6313 6314 pr_reason[0] = '\0'; 6315 /* 6316 * Get the CPU log out info for Disrupting Trap. 6317 */ 6318 if (CPU_PRIVATE(CPU) == NULL) { 6319 clop = NULL; 6320 ch_flt->flt_diag_data.chd_afar = LOGOUT_INVALID; 6321 } else { 6322 clop = CPU_PRIVATE_PTR(CPU, chpr_cecc_logout); 6323 } 6324 6325 if (clop && ch_flt->flt_trapped_ce & CE_CEEN_TIMEOUT) { 6326 ch_cpu_errors_t cpu_error_regs; 6327 6328 get_cpu_error_state(&cpu_error_regs); 6329 (void) cpu_ce_delayed_ec_logout(cpu_error_regs.afar); 6330 clop->clo_data.chd_afsr = cpu_error_regs.afsr; 6331 clop->clo_data.chd_afar = cpu_error_regs.afar; 6332 clop->clo_data.chd_afsr_ext = cpu_error_regs.afsr_ext; 6333 clop->clo_sdw_data.chd_afsr = cpu_error_regs.shadow_afsr; 6334 clop->clo_sdw_data.chd_afar = cpu_error_regs.shadow_afar; 6335 clop->clo_sdw_data.chd_afsr_ext = 6336 cpu_error_regs.shadow_afsr_ext; 6337 #if defined(SERRANO) 6338 clop->clo_data.chd_afar2 = cpu_error_regs.afar2; 6339 #endif /* SERRANO */ 6340 ch_flt->flt_data_incomplete = 1; 6341 6342 /* 6343 * The logging/clear code expects AFSR/AFAR to be cleared. 6344 * The trap handler does it for CEEN enabled errors 6345 * so we need to do it here. 6346 */ 6347 set_cpu_error_state(&cpu_error_regs); 6348 } 6349 6350 #if defined(JALAPENO) || defined(SERRANO) 6351 /* 6352 * FRC: Can't scrub memory as we don't have AFAR for Jalapeno. 6353 * For Serrano, even thou we do have the AFAR, we still do the 6354 * scrub on the RCE side since that's where the error type can 6355 * be properly classified as intermittent, persistent, etc. 6356 * 6357 * CE/RCE: If error is in memory and AFAR is valid, scrub the memory. 6358 * Must scrub memory before cpu_queue_events, as scrubbing memory sets 6359 * the flt_status bits. 6360 */ 6361 if ((afsr & (C_AFSR_CE|C_AFSR_RCE)) && 6362 (cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_CE)) || 6363 cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_RCE)))) { 6364 cpu_ce_scrub_mem_err(aflt, B_TRUE); 6365 } 6366 #else /* JALAPENO || SERRANO */ 6367 /* 6368 * CE/EMC: If error is in memory and AFAR is valid, scrub the memory. 6369 * Must scrub memory before cpu_queue_events, as scrubbing memory sets 6370 * the flt_status bits. 6371 */ 6372 if (afsr & (C_AFSR_CE|C_AFSR_EMC)) { 6373 if (cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_CE)) || 6374 cpu_flt_in_memory(ch_flt, (afsr & C_AFSR_EMC))) { 6375 cpu_ce_scrub_mem_err(aflt, B_TRUE); 6376 } 6377 } 6378 6379 #endif /* JALAPENO || SERRANO */ 6380 6381 /* 6382 * Update flt_prot if this error occurred under on_trap protection. 6383 */ 6384 if (otp != NULL && (otp->ot_prot & OT_DATA_EC)) 6385 aflt->flt_prot = AFLT_PROT_EC; 6386 6387 /* 6388 * Queue events on the async event queue, one event per error bit. 6389 */ 6390 if (cpu_queue_events(ch_flt, pr_reason, afsr_errs, clop) == 0 || 6391 (afsr_errs & (C_AFSR_CECC_ERRS | C_AFSR_EXT_CECC_ERRS)) == 0) { 6392 ch_flt->flt_type = CPU_INV_AFSR; 6393 cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_INVALID_AFSR, 6394 (void *)ch_flt, sizeof (ch_async_flt_t), ue_queue, 6395 aflt->flt_panic); 6396 } 6397 6398 /* 6399 * Zero out + invalidate CPU logout. 6400 */ 6401 if (clop) { 6402 bzero(clop, sizeof (ch_cpu_logout_t)); 6403 clop->clo_data.chd_afar = LOGOUT_INVALID; 6404 } 6405 6406 /* 6407 * If either a CPC, WDC or EDC error has occurred while CEEN 6408 * was disabled, we need to flush either the entire 6409 * E$ or an E$ line. 6410 */ 6411 #if defined(JALAPENO) || defined(SERRANO) 6412 if (afsr & (C_AFSR_EDC | C_AFSR_CPC | C_AFSR_CPU | C_AFSR_WDC)) 6413 #else /* JALAPENO || SERRANO */ 6414 if (afsr_errs & (C_AFSR_EDC | C_AFSR_CPC | C_AFSR_WDC | C_AFSR_L3_EDC | 6415 C_AFSR_L3_CPC | C_AFSR_L3_WDC)) 6416 #endif /* JALAPENO || SERRANO */ 6417 cpu_error_ecache_flush(ch_flt); 6418 6419 } 6420 6421 /* 6422 * depending on the error type, we determine whether we 6423 * need to flush the entire ecache or just a line. 6424 */ 6425 static int 6426 cpu_error_ecache_flush_required(ch_async_flt_t *ch_flt) 6427 { 6428 struct async_flt *aflt; 6429 uint64_t afsr; 6430 uint64_t afsr_errs = ch_flt->afsr_errs; 6431 6432 aflt = (struct async_flt *)ch_flt; 6433 afsr = aflt->flt_stat; 6434 6435 /* 6436 * If we got multiple errors, no point in trying 6437 * the individual cases, just flush the whole cache 6438 */ 6439 if (afsr & C_AFSR_ME) { 6440 return (ECACHE_FLUSH_ALL); 6441 } 6442 6443 /* 6444 * If either a CPC, WDC or EDC error has occurred while CEEN 6445 * was disabled, we need to flush entire E$. We can't just 6446 * flush the cache line affected as the ME bit 6447 * is not set when multiple correctable errors of the same 6448 * type occur, so we might have multiple CPC or EDC errors, 6449 * with only the first recorded. 6450 */ 6451 #if defined(JALAPENO) || defined(SERRANO) 6452 if (afsr & (C_AFSR_CPC | C_AFSR_CPU | C_AFSR_EDC | C_AFSR_WDC)) { 6453 #else /* JALAPENO || SERRANO */ 6454 if (afsr_errs & (C_AFSR_CPC | C_AFSR_EDC | C_AFSR_WDC | C_AFSR_L3_CPC | 6455 C_AFSR_L3_EDC | C_AFSR_L3_WDC)) { 6456 #endif /* JALAPENO || SERRANO */ 6457 return (ECACHE_FLUSH_ALL); 6458 } 6459 6460 #if defined(JALAPENO) || defined(SERRANO) 6461 /* 6462 * If only UE or RUE is set, flush the Ecache line, otherwise 6463 * flush the entire Ecache. 6464 */ 6465 if (afsr & (C_AFSR_UE|C_AFSR_RUE)) { 6466 if ((afsr & C_AFSR_ALL_ERRS) == C_AFSR_UE || 6467 (afsr & C_AFSR_ALL_ERRS) == C_AFSR_RUE) { 6468 return (ECACHE_FLUSH_LINE); 6469 } else { 6470 return (ECACHE_FLUSH_ALL); 6471 } 6472 } 6473 #else /* JALAPENO || SERRANO */ 6474 /* 6475 * If UE only is set, flush the Ecache line, otherwise 6476 * flush the entire Ecache. 6477 */ 6478 if (afsr_errs & C_AFSR_UE) { 6479 if ((afsr_errs & (C_AFSR_ALL_ERRS | C_AFSR_EXT_ALL_ERRS)) == 6480 C_AFSR_UE) { 6481 return (ECACHE_FLUSH_LINE); 6482 } else { 6483 return (ECACHE_FLUSH_ALL); 6484 } 6485 } 6486 #endif /* JALAPENO || SERRANO */ 6487 6488 /* 6489 * EDU: If EDU only is set, flush the ecache line, otherwise 6490 * flush the entire Ecache. 6491 */ 6492 if (afsr_errs & (C_AFSR_EDU | C_AFSR_L3_EDU)) { 6493 if (((afsr_errs & ~C_AFSR_EDU) == 0) || 6494 ((afsr_errs & ~C_AFSR_L3_EDU) == 0)) { 6495 return (ECACHE_FLUSH_LINE); 6496 } else { 6497 return (ECACHE_FLUSH_ALL); 6498 } 6499 } 6500 6501 /* 6502 * BERR: If BERR only is set, flush the Ecache line, otherwise 6503 * flush the entire Ecache. 6504 */ 6505 if (afsr_errs & C_AFSR_BERR) { 6506 if ((afsr_errs & ~C_AFSR_BERR) == 0) { 6507 return (ECACHE_FLUSH_LINE); 6508 } else { 6509 return (ECACHE_FLUSH_ALL); 6510 } 6511 } 6512 6513 return (0); 6514 } 6515 6516 void 6517 cpu_error_ecache_flush(ch_async_flt_t *ch_flt) 6518 { 6519 int ecache_flush_flag = 6520 cpu_error_ecache_flush_required(ch_flt); 6521 6522 /* 6523 * Flush Ecache line or entire Ecache based on above checks. 6524 */ 6525 if (ecache_flush_flag == ECACHE_FLUSH_ALL) 6526 cpu_flush_ecache(); 6527 else if (ecache_flush_flag == ECACHE_FLUSH_LINE) { 6528 cpu_flush_ecache_line(ch_flt); 6529 } 6530 6531 } 6532 6533 /* 6534 * Extract the PA portion from the E$ tag. 6535 */ 6536 uint64_t 6537 cpu_ectag_to_pa(int setsize, uint64_t tag) 6538 { 6539 if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation)) 6540 return (JG_ECTAG_TO_PA(setsize, tag)); 6541 else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) 6542 return (PN_L3TAG_TO_PA(tag)); 6543 else 6544 return (CH_ECTAG_TO_PA(setsize, tag)); 6545 } 6546 6547 /* 6548 * Convert the E$ tag PA into an E$ subblock index. 6549 */ 6550 int 6551 cpu_ectag_pa_to_subblk(int cachesize, uint64_t subaddr) 6552 { 6553 if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation)) 6554 return (JG_ECTAG_PA_TO_SUBBLK(cachesize, subaddr)); 6555 else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) 6556 /* Panther has only one subblock per line */ 6557 return (0); 6558 else 6559 return (CH_ECTAG_PA_TO_SUBBLK(cachesize, subaddr)); 6560 } 6561 6562 /* 6563 * All subblocks in an E$ line must be invalid for 6564 * the line to be invalid. 6565 */ 6566 int 6567 cpu_ectag_line_invalid(int cachesize, uint64_t tag) 6568 { 6569 if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation)) 6570 return (JG_ECTAG_LINE_INVALID(cachesize, tag)); 6571 else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) 6572 return (PN_L3_LINE_INVALID(tag)); 6573 else 6574 return (CH_ECTAG_LINE_INVALID(cachesize, tag)); 6575 } 6576 6577 /* 6578 * Extract state bits for a subblock given the tag. Note that for Panther 6579 * this works on both l2 and l3 tags. 6580 */ 6581 int 6582 cpu_ectag_pa_to_subblk_state(int cachesize, uint64_t subaddr, uint64_t tag) 6583 { 6584 if (IS_JAGUAR(cpunodes[CPU->cpu_id].implementation)) 6585 return (JG_ECTAG_PA_TO_SUBBLK_STATE(cachesize, subaddr, tag)); 6586 else if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) 6587 return (tag & CH_ECSTATE_MASK); 6588 else 6589 return (CH_ECTAG_PA_TO_SUBBLK_STATE(cachesize, subaddr, tag)); 6590 } 6591 6592 /* 6593 * Cpu specific initialization. 6594 */ 6595 void 6596 cpu_mp_init(void) 6597 { 6598 #ifdef CHEETAHPLUS_ERRATUM_25 6599 if (cheetah_sendmondo_recover) { 6600 cheetah_nudge_init(); 6601 } 6602 #endif 6603 } 6604 6605 void 6606 cpu_ereport_post(struct async_flt *aflt) 6607 { 6608 char *cpu_type, buf[FM_MAX_CLASS]; 6609 nv_alloc_t *nva = NULL; 6610 nvlist_t *ereport, *detector, *resource; 6611 errorq_elem_t *eqep; 6612 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 6613 char unum[UNUM_NAMLEN]; 6614 int synd_code; 6615 uint8_t msg_type; 6616 plat_ecc_ch_async_flt_t plat_ecc_ch_flt; 6617 6618 if (aflt->flt_panic || panicstr) { 6619 eqep = errorq_reserve(ereport_errorq); 6620 if (eqep == NULL) 6621 return; 6622 ereport = errorq_elem_nvl(ereport_errorq, eqep); 6623 nva = errorq_elem_nva(ereport_errorq, eqep); 6624 } else { 6625 ereport = fm_nvlist_create(nva); 6626 } 6627 6628 /* 6629 * Create the scheme "cpu" FMRI. 6630 */ 6631 detector = fm_nvlist_create(nva); 6632 resource = fm_nvlist_create(nva); 6633 switch (cpunodes[aflt->flt_inst].implementation) { 6634 case CHEETAH_IMPL: 6635 cpu_type = FM_EREPORT_CPU_USIII; 6636 break; 6637 case CHEETAH_PLUS_IMPL: 6638 cpu_type = FM_EREPORT_CPU_USIIIplus; 6639 break; 6640 case JALAPENO_IMPL: 6641 cpu_type = FM_EREPORT_CPU_USIIIi; 6642 break; 6643 case SERRANO_IMPL: 6644 cpu_type = FM_EREPORT_CPU_USIIIiplus; 6645 break; 6646 case JAGUAR_IMPL: 6647 cpu_type = FM_EREPORT_CPU_USIV; 6648 break; 6649 case PANTHER_IMPL: 6650 cpu_type = FM_EREPORT_CPU_USIVplus; 6651 break; 6652 default: 6653 cpu_type = FM_EREPORT_CPU_UNSUPPORTED; 6654 break; 6655 } 6656 6657 cpu_fmri_cpu_set(detector, aflt->flt_inst); 6658 6659 /* 6660 * Encode all the common data into the ereport. 6661 */ 6662 (void) snprintf(buf, FM_MAX_CLASS, "%s.%s.%s", 6663 FM_ERROR_CPU, cpu_type, aflt->flt_erpt_class); 6664 6665 fm_ereport_set(ereport, FM_EREPORT_VERSION, buf, 6666 fm_ena_generate_cpu(aflt->flt_id, aflt->flt_inst, FM_ENA_FMT1), 6667 detector, NULL); 6668 6669 /* 6670 * Encode the error specific data that was saved in 6671 * the async_flt structure into the ereport. 6672 */ 6673 cpu_payload_add_aflt(aflt, ereport, resource, 6674 &plat_ecc_ch_flt.ecaf_afar_status, 6675 &plat_ecc_ch_flt.ecaf_synd_status); 6676 6677 if (aflt->flt_panic || panicstr) { 6678 errorq_commit(ereport_errorq, eqep, ERRORQ_SYNC); 6679 } else { 6680 (void) fm_ereport_post(ereport, EVCH_TRYHARD); 6681 fm_nvlist_destroy(ereport, FM_NVA_FREE); 6682 fm_nvlist_destroy(detector, FM_NVA_FREE); 6683 fm_nvlist_destroy(resource, FM_NVA_FREE); 6684 } 6685 /* 6686 * Send the enhanced error information (plat_ecc_error2_data_t) 6687 * to the SC olny if it can process it. 6688 */ 6689 6690 if (&plat_ecc_capability_sc_get && 6691 plat_ecc_capability_sc_get(PLAT_ECC_ERROR2_MESSAGE)) { 6692 msg_type = cpu_flt_bit_to_plat_error(aflt); 6693 if (msg_type != PLAT_ECC_ERROR2_NONE) { 6694 /* 6695 * If afar status is not invalid do a unum lookup. 6696 */ 6697 if (plat_ecc_ch_flt.ecaf_afar_status != 6698 AFLT_STAT_INVALID) { 6699 synd_code = synd_to_synd_code( 6700 plat_ecc_ch_flt.ecaf_synd_status, 6701 aflt->flt_synd, ch_flt->flt_bit); 6702 (void) cpu_get_mem_unum_synd(synd_code, 6703 aflt, unum); 6704 } else { 6705 unum[0] = '\0'; 6706 } 6707 plat_ecc_ch_flt.ecaf_sdw_afar = ch_flt->flt_sdw_afar; 6708 plat_ecc_ch_flt.ecaf_sdw_afsr = ch_flt->flt_sdw_afsr; 6709 plat_ecc_ch_flt.ecaf_afsr_ext = ch_flt->afsr_ext; 6710 plat_ecc_ch_flt.ecaf_sdw_afsr_ext = 6711 ch_flt->flt_sdw_afsr_ext; 6712 6713 if (&plat_log_fruid_error2) 6714 plat_log_fruid_error2(msg_type, unum, aflt, 6715 &plat_ecc_ch_flt); 6716 } 6717 } 6718 } 6719 6720 void 6721 cpu_run_bus_error_handlers(struct async_flt *aflt, int expected) 6722 { 6723 int status; 6724 ddi_fm_error_t de; 6725 6726 bzero(&de, sizeof (ddi_fm_error_t)); 6727 6728 de.fme_version = DDI_FME_VERSION; 6729 de.fme_ena = fm_ena_generate_cpu(aflt->flt_id, aflt->flt_inst, 6730 FM_ENA_FMT1); 6731 de.fme_flag = expected; 6732 de.fme_bus_specific = (void *)aflt->flt_addr; 6733 status = ndi_fm_handler_dispatch(ddi_root_node(), NULL, &de); 6734 if ((aflt->flt_prot == AFLT_PROT_NONE) && (status == DDI_FM_FATAL)) 6735 aflt->flt_panic = 1; 6736 } 6737 6738 void 6739 cpu_errorq_dispatch(char *error_class, void *payload, size_t payload_sz, 6740 errorq_t *eqp, uint_t flag) 6741 { 6742 struct async_flt *aflt = (struct async_flt *)payload; 6743 6744 aflt->flt_erpt_class = error_class; 6745 errorq_dispatch(eqp, payload, payload_sz, flag); 6746 } 6747 6748 /* 6749 * This routine may be called by the IO module, but does not do 6750 * anything in this cpu module. The SERD algorithm is handled by 6751 * cpumem-diagnosis engine instead. 6752 */ 6753 /*ARGSUSED*/ 6754 void 6755 cpu_ce_count_unum(struct async_flt *ecc, int len, char *unum) 6756 {} 6757 6758 void 6759 adjust_hw_copy_limits(int ecache_size) 6760 { 6761 /* 6762 * Set hw copy limits. 6763 * 6764 * /etc/system will be parsed later and can override one or more 6765 * of these settings. 6766 * 6767 * At this time, ecache size seems only mildly relevant. 6768 * We seem to run into issues with the d-cache and stalls 6769 * we see on misses. 6770 * 6771 * Cycle measurement indicates that 2 byte aligned copies fare 6772 * little better than doing things with VIS at around 512 bytes. 6773 * 4 byte aligned shows promise until around 1024 bytes. 8 Byte 6774 * aligned is faster whenever the source and destination data 6775 * in cache and the total size is less than 2 Kbytes. The 2K 6776 * limit seems to be driven by the 2K write cache. 6777 * When more than 2K of copies are done in non-VIS mode, stores 6778 * backup in the write cache. In VIS mode, the write cache is 6779 * bypassed, allowing faster cache-line writes aligned on cache 6780 * boundaries. 6781 * 6782 * In addition, in non-VIS mode, there is no prefetching, so 6783 * for larger copies, the advantage of prefetching to avoid even 6784 * occasional cache misses is enough to justify using the VIS code. 6785 * 6786 * During testing, it was discovered that netbench ran 3% slower 6787 * when hw_copy_limit_8 was 2K or larger. Apparently for server 6788 * applications, data is only used once (copied to the output 6789 * buffer, then copied by the network device off the system). Using 6790 * the VIS copy saves more L2 cache state. Network copies are 6791 * around 1.3K to 1.5K in size for historical reasons. 6792 * 6793 * Therefore, a limit of 1K bytes will be used for the 8 byte 6794 * aligned copy even for large caches and 8 MB ecache. The 6795 * infrastructure to allow different limits for different sized 6796 * caches is kept to allow further tuning in later releases. 6797 */ 6798 6799 if (min_ecache_size == 0 && use_hw_bcopy) { 6800 /* 6801 * First time through - should be before /etc/system 6802 * is read. 6803 * Could skip the checks for zero but this lets us 6804 * preserve any debugger rewrites. 6805 */ 6806 if (hw_copy_limit_1 == 0) { 6807 hw_copy_limit_1 = VIS_COPY_THRESHOLD; 6808 priv_hcl_1 = hw_copy_limit_1; 6809 } 6810 if (hw_copy_limit_2 == 0) { 6811 hw_copy_limit_2 = 2 * VIS_COPY_THRESHOLD; 6812 priv_hcl_2 = hw_copy_limit_2; 6813 } 6814 if (hw_copy_limit_4 == 0) { 6815 hw_copy_limit_4 = 4 * VIS_COPY_THRESHOLD; 6816 priv_hcl_4 = hw_copy_limit_4; 6817 } 6818 if (hw_copy_limit_8 == 0) { 6819 hw_copy_limit_8 = 4 * VIS_COPY_THRESHOLD; 6820 priv_hcl_8 = hw_copy_limit_8; 6821 } 6822 min_ecache_size = ecache_size; 6823 } else { 6824 /* 6825 * MP initialization. Called *after* /etc/system has 6826 * been parsed. One CPU has already been initialized. 6827 * Need to cater for /etc/system having scragged one 6828 * of our values. 6829 */ 6830 if (ecache_size == min_ecache_size) { 6831 /* 6832 * Same size ecache. We do nothing unless we 6833 * have a pessimistic ecache setting. In that 6834 * case we become more optimistic (if the cache is 6835 * large enough). 6836 */ 6837 if (hw_copy_limit_8 == 4 * VIS_COPY_THRESHOLD) { 6838 /* 6839 * Need to adjust hw_copy_limit* from our 6840 * pessimistic uniprocessor value to a more 6841 * optimistic UP value *iff* it hasn't been 6842 * reset. 6843 */ 6844 if ((ecache_size > 1048576) && 6845 (priv_hcl_8 == hw_copy_limit_8)) { 6846 if (ecache_size <= 2097152) 6847 hw_copy_limit_8 = 4 * 6848 VIS_COPY_THRESHOLD; 6849 else if (ecache_size <= 4194304) 6850 hw_copy_limit_8 = 4 * 6851 VIS_COPY_THRESHOLD; 6852 else 6853 hw_copy_limit_8 = 4 * 6854 VIS_COPY_THRESHOLD; 6855 priv_hcl_8 = hw_copy_limit_8; 6856 } 6857 } 6858 } else if (ecache_size < min_ecache_size) { 6859 /* 6860 * A different ecache size. Can this even happen? 6861 */ 6862 if (priv_hcl_8 == hw_copy_limit_8) { 6863 /* 6864 * The previous value that we set 6865 * is unchanged (i.e., it hasn't been 6866 * scragged by /etc/system). Rewrite it. 6867 */ 6868 if (ecache_size <= 1048576) 6869 hw_copy_limit_8 = 8 * 6870 VIS_COPY_THRESHOLD; 6871 else if (ecache_size <= 2097152) 6872 hw_copy_limit_8 = 8 * 6873 VIS_COPY_THRESHOLD; 6874 else if (ecache_size <= 4194304) 6875 hw_copy_limit_8 = 8 * 6876 VIS_COPY_THRESHOLD; 6877 else 6878 hw_copy_limit_8 = 10 * 6879 VIS_COPY_THRESHOLD; 6880 priv_hcl_8 = hw_copy_limit_8; 6881 min_ecache_size = ecache_size; 6882 } 6883 } 6884 } 6885 } 6886 6887 /* 6888 * Called from illegal instruction trap handler to see if we can attribute 6889 * the trap to a fpras check. 6890 */ 6891 int 6892 fpras_chktrap(struct regs *rp) 6893 { 6894 int op; 6895 struct fpras_chkfngrp *cgp; 6896 uintptr_t tpc = (uintptr_t)rp->r_pc; 6897 6898 if (fpras_chkfngrps == NULL) 6899 return (0); 6900 6901 cgp = &fpras_chkfngrps[CPU->cpu_id]; 6902 for (op = 0; op < FPRAS_NCOPYOPS; ++op) { 6903 if (tpc >= (uintptr_t)&cgp->fpras_fn[op].fpras_blk0 && 6904 tpc < (uintptr_t)&cgp->fpras_fn[op].fpras_chkresult) 6905 break; 6906 } 6907 if (op == FPRAS_NCOPYOPS) 6908 return (0); 6909 6910 /* 6911 * This is an fpRAS failure caught through an illegal 6912 * instruction - trampoline. 6913 */ 6914 rp->r_pc = (uintptr_t)&cgp->fpras_fn[op].fpras_trampoline; 6915 rp->r_npc = rp->r_pc + 4; 6916 return (1); 6917 } 6918 6919 /* 6920 * fpras_failure is called when a fpras check detects a bad calculation 6921 * result or an illegal instruction trap is attributed to an fpras 6922 * check. In all cases we are still bound to CPU. 6923 */ 6924 int 6925 fpras_failure(int op, int how) 6926 { 6927 int use_hw_bcopy_orig, use_hw_bzero_orig; 6928 uint_t hcl1_orig, hcl2_orig, hcl4_orig, hcl8_orig; 6929 ch_async_flt_t ch_flt; 6930 struct async_flt *aflt = (struct async_flt *)&ch_flt; 6931 struct fpras_chkfn *sfp, *cfp; 6932 uint32_t *sip, *cip; 6933 int i; 6934 6935 /* 6936 * We're running on a sick CPU. Avoid further FPU use at least for 6937 * the time in which we dispatch an ereport and (if applicable) panic. 6938 */ 6939 use_hw_bcopy_orig = use_hw_bcopy; 6940 use_hw_bzero_orig = use_hw_bzero; 6941 hcl1_orig = hw_copy_limit_1; 6942 hcl2_orig = hw_copy_limit_2; 6943 hcl4_orig = hw_copy_limit_4; 6944 hcl8_orig = hw_copy_limit_8; 6945 use_hw_bcopy = use_hw_bzero = 0; 6946 hw_copy_limit_1 = hw_copy_limit_2 = hw_copy_limit_4 = 6947 hw_copy_limit_8 = 0; 6948 6949 bzero(&ch_flt, sizeof (ch_async_flt_t)); 6950 aflt->flt_id = gethrtime_waitfree(); 6951 aflt->flt_class = CPU_FAULT; 6952 aflt->flt_inst = CPU->cpu_id; 6953 aflt->flt_status = (how << 8) | op; 6954 aflt->flt_payload = FM_EREPORT_PAYLOAD_FPU_HWCOPY; 6955 ch_flt.flt_type = CPU_FPUERR; 6956 6957 /* 6958 * We must panic if the copy operation had no lofault protection - 6959 * ie, don't panic for copyin, copyout, kcopy and bcopy called 6960 * under on_fault and do panic for unprotected bcopy and hwblkpagecopy. 6961 */ 6962 aflt->flt_panic = (curthread->t_lofault == NULL); 6963 6964 /* 6965 * XOR the source instruction block with the copied instruction 6966 * block - this will show us which bit(s) are corrupted. 6967 */ 6968 sfp = (struct fpras_chkfn *)fpras_chkfn_type1; 6969 cfp = &fpras_chkfngrps[CPU->cpu_id].fpras_fn[op]; 6970 if (op == FPRAS_BCOPY || op == FPRAS_COPYOUT) { 6971 sip = &sfp->fpras_blk0[0]; 6972 cip = &cfp->fpras_blk0[0]; 6973 } else { 6974 sip = &sfp->fpras_blk1[0]; 6975 cip = &cfp->fpras_blk1[0]; 6976 } 6977 for (i = 0; i < 16; ++i, ++sip, ++cip) 6978 ch_flt.flt_fpdata[i] = *sip ^ *cip; 6979 6980 cpu_errorq_dispatch(FM_EREPORT_CPU_USIII_FPU_HWCOPY, (void *)&ch_flt, 6981 sizeof (ch_async_flt_t), ue_queue, aflt->flt_panic); 6982 6983 if (aflt->flt_panic) 6984 fm_panic("FPU failure on CPU %d", CPU->cpu_id); 6985 6986 /* 6987 * We get here for copyin/copyout and kcopy or bcopy where the 6988 * caller has used on_fault. We will flag the error so that 6989 * the process may be killed The trap_async_hwerr mechanism will 6990 * take appropriate further action (such as a reboot, contract 6991 * notification etc). Since we may be continuing we will 6992 * restore the global hardware copy acceleration switches. 6993 * 6994 * When we return from this function to the copy function we want to 6995 * avoid potentially bad data being used, ie we want the affected 6996 * copy function to return an error. The caller should therefore 6997 * invoke its lofault handler (which always exists for these functions) 6998 * which will return the appropriate error. 6999 */ 7000 ttolwp(curthread)->lwp_pcb.pcb_flags |= ASYNC_HWERR; 7001 aston(curthread); 7002 7003 use_hw_bcopy = use_hw_bcopy_orig; 7004 use_hw_bzero = use_hw_bzero_orig; 7005 hw_copy_limit_1 = hcl1_orig; 7006 hw_copy_limit_2 = hcl2_orig; 7007 hw_copy_limit_4 = hcl4_orig; 7008 hw_copy_limit_8 = hcl8_orig; 7009 7010 return (1); 7011 } 7012 7013 #define VIS_BLOCKSIZE 64 7014 7015 int 7016 dtrace_blksuword32_err(uintptr_t addr, uint32_t *data) 7017 { 7018 int ret, watched; 7019 7020 watched = watch_disable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE); 7021 ret = dtrace_blksuword32(addr, data, 0); 7022 if (watched) 7023 watch_enable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE); 7024 7025 return (ret); 7026 } 7027 7028 /* 7029 * Called when a cpu enters the CPU_FAULTED state (by the cpu placing the 7030 * faulted cpu into that state). Cross-trap to the faulted cpu to clear 7031 * CEEN from the EER to disable traps for further disrupting error types 7032 * on that cpu. We could cross-call instead, but that has a larger 7033 * instruction and data footprint than cross-trapping, and the cpu is known 7034 * to be faulted. 7035 */ 7036 7037 void 7038 cpu_faulted_enter(struct cpu *cp) 7039 { 7040 xt_one(cp->cpu_id, set_error_enable_tl1, EN_REG_CEEN, EER_SET_CLRBITS); 7041 } 7042 7043 /* 7044 * Called when a cpu leaves the CPU_FAULTED state to return to one of 7045 * offline, spare, or online (by the cpu requesting this state change). 7046 * First we cross-call to clear the AFSR (and AFSR_EXT on Panther) of 7047 * disrupting error bits that have accumulated without trapping, then 7048 * we cross-trap to re-enable CEEN controlled traps. 7049 */ 7050 void 7051 cpu_faulted_exit(struct cpu *cp) 7052 { 7053 ch_cpu_errors_t cpu_error_regs; 7054 7055 cpu_error_regs.afsr = C_AFSR_CECC_ERRS; 7056 if (IS_PANTHER(cpunodes[cp->cpu_id].implementation)) 7057 cpu_error_regs.afsr_ext &= C_AFSR_EXT_CECC_ERRS; 7058 xc_one(cp->cpu_id, (xcfunc_t *)set_cpu_error_state, 7059 (uint64_t)&cpu_error_regs, 0); 7060 7061 xt_one(cp->cpu_id, set_error_enable_tl1, EN_REG_CEEN, EER_SET_SETBITS); 7062 } 7063 7064 /* 7065 * Return 1 if the errors in ch_flt's AFSR are secondary errors caused by 7066 * the errors in the original AFSR, 0 otherwise. 7067 * 7068 * For all procs if the initial error was a BERR or TO, then it is possible 7069 * that we may have caused a secondary BERR or TO in the process of logging the 7070 * inital error via cpu_run_bus_error_handlers(). If this is the case then 7071 * if the request was protected then a panic is still not necessary, if not 7072 * protected then aft_panic is already set - so either way there's no need 7073 * to set aft_panic for the secondary error. 7074 * 7075 * For Cheetah and Jalapeno, if the original error was a UE which occurred on 7076 * a store merge, then the error handling code will call cpu_deferred_error(). 7077 * When clear_errors() is called, it will determine that secondary errors have 7078 * occurred - in particular, the store merge also caused a EDU and WDU that 7079 * weren't discovered until this point. 7080 * 7081 * We do three checks to verify that we are in this case. If we pass all three 7082 * checks, we return 1 to indicate that we should not panic. If any unexpected 7083 * errors occur, we return 0. 7084 * 7085 * For Cheetah+ and derivative procs, the store merge causes a DUE, which is 7086 * handled in cpu_disrupting_errors(). Since this function is not even called 7087 * in the case we are interested in, we just return 0 for these processors. 7088 */ 7089 /*ARGSUSED*/ 7090 static int 7091 cpu_check_secondary_errors(ch_async_flt_t *ch_flt, uint64_t t_afsr_errs, 7092 uint64_t t_afar) 7093 { 7094 #if defined(CHEETAH_PLUS) 7095 #else /* CHEETAH_PLUS */ 7096 struct async_flt *aflt = (struct async_flt *)ch_flt; 7097 #endif /* CHEETAH_PLUS */ 7098 7099 /* 7100 * Was the original error a BERR or TO and only a BERR or TO 7101 * (multiple errors are also OK) 7102 */ 7103 if ((t_afsr_errs & ~(C_AFSR_BERR | C_AFSR_TO | C_AFSR_ME)) == 0) { 7104 /* 7105 * Is the new error a BERR or TO and only a BERR or TO 7106 * (multiple errors are also OK) 7107 */ 7108 if ((ch_flt->afsr_errs & 7109 ~(C_AFSR_BERR | C_AFSR_TO | C_AFSR_ME)) == 0) 7110 return (1); 7111 } 7112 7113 #if defined(CHEETAH_PLUS) 7114 return (0); 7115 #else /* CHEETAH_PLUS */ 7116 /* 7117 * Now look for secondary effects of a UE on cheetah/jalapeno 7118 * 7119 * Check the original error was a UE, and only a UE. Note that 7120 * the ME bit will cause us to fail this check. 7121 */ 7122 if (t_afsr_errs != C_AFSR_UE) 7123 return (0); 7124 7125 /* 7126 * Check the secondary errors were exclusively an EDU and/or WDU. 7127 */ 7128 if ((ch_flt->afsr_errs & ~(C_AFSR_EDU|C_AFSR_WDU)) != 0) 7129 return (0); 7130 7131 /* 7132 * Check the AFAR of the original error and secondary errors 7133 * match to the 64-byte boundary 7134 */ 7135 if (P2ALIGN(aflt->flt_addr, 64) != P2ALIGN(t_afar, 64)) 7136 return (0); 7137 7138 /* 7139 * We've passed all the checks, so it's a secondary error! 7140 */ 7141 return (1); 7142 #endif /* CHEETAH_PLUS */ 7143 } 7144 7145 /* 7146 * Translate the flt_bit or flt_type into an error type. First, flt_bit 7147 * is checked for any valid errors. If found, the error type is 7148 * returned. If not found, the flt_type is checked for L1$ parity errors. 7149 */ 7150 /*ARGSUSED*/ 7151 static uint8_t 7152 cpu_flt_bit_to_plat_error(struct async_flt *aflt) 7153 { 7154 #if defined(JALAPENO) 7155 /* 7156 * Currently, logging errors to the SC is not supported on Jalapeno 7157 */ 7158 return (PLAT_ECC_ERROR2_NONE); 7159 #else 7160 ch_async_flt_t *ch_flt = (ch_async_flt_t *)aflt; 7161 7162 switch (ch_flt->flt_bit) { 7163 case C_AFSR_CE: 7164 return (PLAT_ECC_ERROR2_CE); 7165 case C_AFSR_UCC: 7166 case C_AFSR_EDC: 7167 case C_AFSR_WDC: 7168 case C_AFSR_CPC: 7169 return (PLAT_ECC_ERROR2_L2_CE); 7170 case C_AFSR_EMC: 7171 return (PLAT_ECC_ERROR2_EMC); 7172 case C_AFSR_IVC: 7173 return (PLAT_ECC_ERROR2_IVC); 7174 case C_AFSR_UE: 7175 return (PLAT_ECC_ERROR2_UE); 7176 case C_AFSR_UCU: 7177 case C_AFSR_EDU: 7178 case C_AFSR_WDU: 7179 case C_AFSR_CPU: 7180 return (PLAT_ECC_ERROR2_L2_UE); 7181 case C_AFSR_IVU: 7182 return (PLAT_ECC_ERROR2_IVU); 7183 case C_AFSR_TO: 7184 return (PLAT_ECC_ERROR2_TO); 7185 case C_AFSR_BERR: 7186 return (PLAT_ECC_ERROR2_BERR); 7187 #if defined(CHEETAH_PLUS) 7188 case C_AFSR_L3_EDC: 7189 case C_AFSR_L3_UCC: 7190 case C_AFSR_L3_CPC: 7191 case C_AFSR_L3_WDC: 7192 return (PLAT_ECC_ERROR2_L3_CE); 7193 case C_AFSR_IMC: 7194 return (PLAT_ECC_ERROR2_IMC); 7195 case C_AFSR_TSCE: 7196 return (PLAT_ECC_ERROR2_L2_TSCE); 7197 case C_AFSR_THCE: 7198 return (PLAT_ECC_ERROR2_L2_THCE); 7199 case C_AFSR_L3_MECC: 7200 return (PLAT_ECC_ERROR2_L3_MECC); 7201 case C_AFSR_L3_THCE: 7202 return (PLAT_ECC_ERROR2_L3_THCE); 7203 case C_AFSR_L3_CPU: 7204 case C_AFSR_L3_EDU: 7205 case C_AFSR_L3_UCU: 7206 case C_AFSR_L3_WDU: 7207 return (PLAT_ECC_ERROR2_L3_UE); 7208 case C_AFSR_DUE: 7209 return (PLAT_ECC_ERROR2_DUE); 7210 case C_AFSR_DTO: 7211 return (PLAT_ECC_ERROR2_DTO); 7212 case C_AFSR_DBERR: 7213 return (PLAT_ECC_ERROR2_DBERR); 7214 #endif /* CHEETAH_PLUS */ 7215 default: 7216 switch (ch_flt->flt_type) { 7217 #if defined(CPU_IMP_L1_CACHE_PARITY) 7218 case CPU_IC_PARITY: 7219 return (PLAT_ECC_ERROR2_IPE); 7220 case CPU_DC_PARITY: 7221 if (IS_PANTHER(cpunodes[CPU->cpu_id].implementation)) { 7222 if (ch_flt->parity_data.dpe.cpl_cache == 7223 CPU_PC_PARITY) { 7224 return (PLAT_ECC_ERROR2_PCACHE); 7225 } 7226 } 7227 return (PLAT_ECC_ERROR2_DPE); 7228 #endif /* CPU_IMP_L1_CACHE_PARITY */ 7229 case CPU_ITLB_PARITY: 7230 return (PLAT_ECC_ERROR2_ITLB); 7231 case CPU_DTLB_PARITY: 7232 return (PLAT_ECC_ERROR2_DTLB); 7233 default: 7234 return (PLAT_ECC_ERROR2_NONE); 7235 } 7236 } 7237 #endif /* JALAPENO */ 7238 } 7239
Indexes created Sun Nov 22 02:52:53 UTC 2009
Terms of Use |
Privacy |
Trademarks |
Copyright Policy |
Site Guidelines |
Help
Your use of this web site or any of its content or software indicates your agreement to be bound by these Terms of Use.
Copyright © 1995-2008 Sun Microsystems, Inc.