/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */

/*
 * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

/*
 * DTrace - Dynamic Tracing for Solaris
 *
 * This is the implementation of the Solaris Dynamic Tracing framework
 * (DTrace).  The user-visible interface to DTrace is described at length in
 * the "Solaris Dynamic Tracing Guide".  The interfaces between the libdtrace
 * library, the in-kernel DTrace framework, and the DTrace providers are
 * described in the block comments in the <sys/dtrace.h> header file.  The
 * internal architecture of DTrace is described in the block comments in the
 * <sys/dtrace_impl.h> header file.  The comments contained within the DTrace
 * implementation very much assume mastery of all of these sources; if one has
 * an unanswered question about the implementation, one should consult them
 * first.
 *
 * The functions here are ordered roughly as follows:
 *
 *   - Probe context functions
 *   - Probe hashing functions
 *   - Non-probe context utility functions
 *   - Matching functions
 *   - Provider-to-Framework API functions
 *   - Probe management functions
 *   - DIF object functions
 *   - Format functions
 *   - Predicate functions
 *   - ECB functions
 *   - Buffer functions
 *   - Enabling functions
 *   - DOF functions
 *   - Anonymous enabling functions
 *   - Consumer state functions
 *   - Helper functions
 *   - Hook functions
 *   - Driver cookbook functions
 *
 * Each group of functions begins with a block comment labelled the "DTrace
 * [Group] Functions", allowing one to find each block by searching forward
 * on capital-f functions.
 */
#include <sys/errno.h>
#include <sys/stat.h>
#include <sys/modctl.h>
#include <sys/conf.h>
#include <sys/systm.h>
#include <sys/ddi.h>
#include <sys/sunddi.h>
#include <sys/cpuvar.h>
#include <sys/kmem.h>
#include <sys/strsubr.h>
#include <sys/sysmacros.h>
#include <sys/dtrace_impl.h>
#include <sys/atomic.h>
#include <sys/cmn_err.h>
#include <sys/mutex_impl.h>
#include <sys/rwlock_impl.h>
#include <sys/ctf_api.h>
#include <sys/panic.h>
#include <sys/priv_impl.h>
#include <sys/policy.h>
#include <sys/cred_impl.h>
#include <sys/procfs_isa.h>
#include <sys/taskq.h>
#include <sys/mkdev.h>
#include <sys/kdi.h>
#include <sys/zone.h>
#include <sys/socket.h>
#include <netinet/in.h>

/*
 * DTrace Tunable Variables
 *
 * The following variables may be tuned by adding a line to /etc/system that
 * includes both the name of the DTrace module ("dtrace") and the name of the
 * variable.  For example:
 *
 *   set dtrace:dtrace_destructive_disallow = 1
 *
 * In general, the only variables that one should be tuning this way are those
 * that affect system-wide DTrace behavior, and for which the default behavior
 * is undesirable.  Most of these variables are tunable on a per-consumer
 * basis using DTrace options, and need not be tuned on a system-wide basis.
 * When tuning these variables, avoid pathological values; while some attempt
 * is made to verify the integrity of these variables, they are not considered
 * part of the supported interface to DTrace, and they are therefore not
 * checked comprehensively.  Further, these variables should not be tuned
 * dynamically via "mdb -kw" or other means; they should only be tuned via
 * /etc/system.
 */
int		dtrace_destructive_disallow = 0;
dtrace_optval_t	dtrace_nonroot_maxsize = (16 * 1024 * 1024);
size_t		dtrace_difo_maxsize = (256 * 1024);
dtrace_optval_t	dtrace_dof_maxsize = (256 * 1024);
size_t		dtrace_global_maxsize = (16 * 1024);
size_t		dtrace_actions_max = (16 * 1024);
size_t		dtrace_retain_max = 1024;
dtrace_optval_t	dtrace_helper_actions_max = 32;
dtrace_optval_t	dtrace_helper_providers_max = 32;
dtrace_optval_t	dtrace_dstate_defsize = (1 * 1024 * 1024);
size_t		dtrace_strsize_default = 256;
dtrace_optval_t	dtrace_cleanrate_default = 9900990;		/* 101 hz */
dtrace_optval_t	dtrace_cleanrate_min = 200000;			/* 5000 hz */
dtrace_optval_t	dtrace_cleanrate_max = (uint64_t)60 * NANOSEC;	/* 1/minute */
dtrace_optval_t	dtrace_aggrate_default = NANOSEC;		/* 1 hz */
dtrace_optval_t	dtrace_statusrate_default = NANOSEC;		/* 1 hz */
dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC;	 /* 6/minute */
dtrace_optval_t	dtrace_switchrate_default = NANOSEC;		/* 1 hz */
dtrace_optval_t	dtrace_nspec_default = 1;
dtrace_optval_t	dtrace_specsize_default = 32 * 1024;
dtrace_optval_t dtrace_stackframes_default = 20;
dtrace_optval_t dtrace_ustackframes_default = 20;
dtrace_optval_t dtrace_jstackframes_default = 50;
dtrace_optval_t dtrace_jstackstrsize_default = 512;
int		dtrace_msgdsize_max = 128;
hrtime_t	dtrace_chill_max = 500 * (NANOSEC / MILLISEC);	/* 500 ms */
hrtime_t	dtrace_chill_interval = NANOSEC;		/* 1000 ms */
int		dtrace_devdepth_max = 32;
int		dtrace_err_verbose;
hrtime_t	dtrace_deadman_interval = NANOSEC;
hrtime_t	dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
hrtime_t	dtrace_deadman_user = (hrtime_t)30 * NANOSEC;

/*
 * DTrace External Variables
 *
 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
 * available to DTrace consumers via the backtick (`) syntax.  One of these,
 * dtrace_zero, is made deliberately so:  it is provided as a source of
 * well-known, zero-filled memory.  While this variable is not documented,
 * it is used by some translators as an implementation detail.
 */
const char	dtrace_zero[256] = { 0 };	/* zero-filled memory */

/*
 * DTrace Internal Variables
 */
static dev_info_t	*dtrace_devi;		/* device info */
static vmem_t		*dtrace_arena;		/* probe ID arena */
static vmem_t		*dtrace_minor;		/* minor number arena */
static taskq_t		*dtrace_taskq;		/* task queue */
static dtrace_probe_t	**dtrace_probes;	/* array of all probes */
static int		dtrace_nprobes;		/* number of probes */
static dtrace_provider_t *dtrace_provider;	/* provider list */
static dtrace_meta_t	*dtrace_meta_pid;	/* user-land meta provider */
static int		dtrace_opens;		/* number of opens */
static int		dtrace_helpers;		/* number of helpers */
static void		*dtrace_softstate;	/* softstate pointer */
static dtrace_hash_t	*dtrace_bymod;		/* probes hashed by module */
static dtrace_hash_t	*dtrace_byfunc;		/* probes hashed by function */
static dtrace_hash_t	*dtrace_byname;		/* probes hashed by name */
static dtrace_toxrange_t *dtrace_toxrange;	/* toxic range array */
static int		dtrace_toxranges;	/* number of toxic ranges */
static int		dtrace_toxranges_max;	/* size of toxic range array */
static dtrace_anon_t	dtrace_anon;		/* anonymous enabling */
static kmem_cache_t	*dtrace_state_cache;	/* cache for dynamic state */
static uint64_t		dtrace_vtime_references; /* number of vtimestamp refs */
static kthread_t	*dtrace_panicked;	/* panicking thread */
static dtrace_ecb_t	*dtrace_ecb_create_cache; /* cached created ECB */
static dtrace_genid_t	dtrace_probegen;	/* current probe generation */
static dtrace_helpers_t *dtrace_deferred_pid;	/* deferred helper list */
static dtrace_enabling_t *dtrace_retained;	/* list of retained enablings */
static dtrace_genid_t	dtrace_retained_gen;	/* current retained enab gen */
static dtrace_dynvar_t	dtrace_dynhash_sink;	/* end of dynamic hash chains */

/*
 * DTrace Locking
 * DTrace is protected by three (relatively coarse-grained) locks:
 *
 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
 *     including enabling state, probes, ECBs, consumer state, helper state,
 *     etc.  Importantly, dtrace_lock is _not_ required when in probe context;
 *     probe context is lock-free -- synchronization is handled via the
 *     dtrace_sync() cross call mechanism.
 *
 * (2) dtrace_provider_lock is required when manipulating provider state, or
 *     when provider state must be held constant.
 *
 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
 *     when meta provider state must be held constant.
 *
 * The lock ordering between these three locks is dtrace_meta_lock before
 * dtrace_provider_lock before dtrace_lock.  (In particular, there are
 * several places where dtrace_provider_lock is held by the framework as it
 * calls into the providers -- which then call back into the framework,
 * grabbing dtrace_lock.)
 *
 * There are two other locks in the mix:  mod_lock and cpu_lock.  With respect
 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
 * role as a coarse-grained lock; it is acquired before both of these locks.
 * With respect to dtrace_meta_lock, its behavior is stranger:  cpu_lock must
 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
 * acquired _between_ dtrace_provider_lock and dtrace_lock.
 */
static kmutex_t		dtrace_lock;		/* probe state lock */
static kmutex_t		dtrace_provider_lock;	/* provider state lock */
static kmutex_t		dtrace_meta_lock;	/* meta-provider state lock */

/*
 * DTrace Provider Variables
 *
 * These are the variables relating to DTrace as a provider (that is, the
 * provider of the BEGIN, END, and ERROR probes).
 */
static dtrace_pattr_t	dtrace_provider_attr = {
{ DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
{ DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
{ DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
};

static void
dtrace_nullop(void)
{}

static int
dtrace_enable_nullop(void)
{
	return (0);
}

static dtrace_pops_t	dtrace_provider_ops = {
	(void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop,
	(void (*)(void *, struct modctl *))dtrace_nullop,
	(int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop,
	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
	NULL,
	NULL,
	NULL,
	(void (*)(void *, dtrace_id_t, void *))dtrace_nullop
};

static dtrace_id_t	dtrace_probeid_begin;	/* special BEGIN probe */
static dtrace_id_t	dtrace_probeid_end;	/* special END probe */
dtrace_id_t		dtrace_probeid_error;	/* special ERROR probe */

/*
 * DTrace Helper Tracing Variables
 */
uint32_t dtrace_helptrace_next = 0;
uint32_t dtrace_helptrace_nlocals;
char	*dtrace_helptrace_buffer;
int	dtrace_helptrace_bufsize = 512 * 1024;

#ifdef DEBUG
int	dtrace_helptrace_enabled = 1;
#else
int	dtrace_helptrace_enabled = 0;
#endif

/*
 * DTrace Error Hashing
 *
 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
 * table.  This is very useful for checking coverage of tests that are
 * expected to induce DIF or DOF processing errors, and may be useful for
 * debugging problems in the DIF code generator or in DOF generation .  The
 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
 */
#ifdef DEBUG
static dtrace_errhash_t	dtrace_errhash[DTRACE_ERRHASHSZ];
static const char *dtrace_errlast;
static kthread_t *dtrace_errthread;
static kmutex_t dtrace_errlock;
#endif

/*
 * DTrace Macros and Constants
 *
 * These are various macros that are useful in various spots in the
 * implementation, along with a few random constants that have no meaning
 * outside of the implementation.  There is no real structure to this cpp
 * mishmash -- but is there ever?
 */
#define	DTRACE_HASHSTR(hash, probe)	\
	dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))

#define	DTRACE_HASHNEXT(hash, probe)	\
	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)

#define	DTRACE_HASHPREV(hash, probe)	\
	(dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)

#define	DTRACE_HASHEQ(hash, lhs, rhs)	\
	(strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
	    *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)

#define	DTRACE_AGGHASHSIZE_SLEW		17

#define	DTRACE_V4MAPPED_OFFSET		(sizeof (uint32_t) * 3)

/*
 * The key for a thread-local variable consists of the lower 61 bits of the
 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
 * equal to a variable identifier.  This is necessary (but not sufficient) to
 * assure that global associative arrays never collide with thread-local
 * variables.  To guarantee that they cannot collide, we must also define the
 * order for keying dynamic variables.  That order is:
 *
 *   [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
 *
 * Because the variable-key and the tls-key are in orthogonal spaces, there is
 * no way for a global variable key signature to match a thread-local key
 * signature.
 */
#define	DTRACE_TLS_THRKEY(where) { \
	uint_t intr = 0; \
	uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
	for (; actv; actv >>= 1) \
		intr++; \
	ASSERT(intr < (1 << 3)); \
	(where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
	    (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
}

#define	DT_BSWAP_8(x)	((x) & 0xff)
#define	DT_BSWAP_16(x)	((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
#define	DT_BSWAP_32(x)	((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
#define	DT_BSWAP_64(x)	((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))

#define	DT_MASK_LO 0x00000000FFFFFFFFULL

#define	DTRACE_STORE(type, tomax, offset, what) \
	*((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);

#ifndef __i386
#define	DTRACE_ALIGNCHECK(addr, size, flags)				\
	if (addr & (size - 1)) {					\
		*flags |= CPU_DTRACE_BADALIGN;				\
		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
		return (0);						\
	}
#else
#define	DTRACE_ALIGNCHECK(addr, size, flags)
#endif

/*
 * Test whether a range of memory starting at testaddr of size testsz falls
 * within the range of memory described by addr, sz.  We take care to avoid
 * problems with overflow and underflow of the unsigned quantities, and
 * disallow all negative sizes.  Ranges of size 0 are allowed.
 */
#define	DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
	((testaddr) - (baseaddr) < (basesz) && \
	(testaddr) + (testsz) - (baseaddr) <= (basesz) && \
	(testaddr) + (testsz) >= (testaddr))

/*
 * Test whether alloc_sz bytes will fit in the scratch region.  We isolate
 * alloc_sz on the righthand side of the comparison in order to avoid overflow
 * or underflow in the comparison with it.  This is simpler than the INRANGE
 * check above, because we know that the dtms_scratch_ptr is valid in the
 * range.  Allocations of size zero are allowed.
 */
#define	DTRACE_INSCRATCH(mstate, alloc_sz) \
	((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
	(mstate)->dtms_scratch_ptr >= (alloc_sz))

#define	DTRACE_LOADFUNC(bits)						\
/*CSTYLED*/								\
uint##bits##_t								\
dtrace_load##bits(uintptr_t addr)					\
{									\
	size_t size = bits / NBBY;					\
	/*CSTYLED*/							\
	uint##bits##_t rval;						\
	int i;								\
	volatile uint16_t *flags = (volatile uint16_t *)		\
	    &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;			\
									\
	DTRACE_ALIGNCHECK(addr, size, flags);				\
									\
	for (i = 0; i < dtrace_toxranges; i++) {			\
		if (addr >= dtrace_toxrange[i].dtt_limit)		\
			continue;					\
									\
		if (addr + size <= dtrace_toxrange[i].dtt_base)		\
			continue;					\
									\
		/*							\
		 * This address falls within a toxic region; return 0.	\
		 */							\
		*flags |= CPU_DTRACE_BADADDR;				\
		cpu_core[CPU->cpu_id].cpuc_dtrace_illval = addr;	\
		return (0);						\
	}								\
									\
	*flags |= CPU_DTRACE_NOFAULT;					\
	/*CSTYLED*/							\
	rval = *((volatile uint##bits##_t *)addr);			\
	*flags &= ~CPU_DTRACE_NOFAULT;					\
									\
	return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0);		\
}

#ifdef _LP64
#define	dtrace_loadptr	dtrace_load64
#else
#define	dtrace_loadptr	dtrace_load32
#endif

#define	DTRACE_DYNHASH_FREE	0
#define	DTRACE_DYNHASH_SINK	1
#define	DTRACE_DYNHASH_VALID	2

#define	DTRACE_MATCH_FAIL	-1
#define	DTRACE_MATCH_NEXT	0
#define	DTRACE_MATCH_DONE	1
#define	DTRACE_ANCHORED(probe)	((probe)->dtpr_func[0] != '\0')
#define	DTRACE_STATE_ALIGN	64

#define	DTRACE_FLAGS2FLT(flags)						\
	(((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR :		\
	((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP :		\
	((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO :		\
	((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV :		\
	((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV :		\
	((flags) & CPU_DTRACE_TUPOFLOW) ?  DTRACEFLT_TUPOFLOW :		\
	((flags) & CPU_DTRACE_BADALIGN) ?  DTRACEFLT_BADALIGN :		\
	((flags) & CPU_DTRACE_NOSCRATCH) ?  DTRACEFLT_NOSCRATCH :	\
	((flags) & CPU_DTRACE_BADSTACK) ?  DTRACEFLT_BADSTACK :		\
	DTRACEFLT_UNKNOWN)

#define	DTRACEACT_ISSTRING(act)						\
	((act)->dta_kind == DTRACEACT_DIFEXPR &&			\
	(act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)

static size_t dtrace_strlen(const char *, size_t);
static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
static void dtrace_enabling_provide(dtrace_provider_t *);
static int dtrace_enabling_match(dtrace_enabling_t *, int *);
static void dtrace_enabling_matchall(void);
static dtrace_state_t *dtrace_anon_grab(void);
static uint64_t dtrace_helper(int, dtrace_mstate_t *,
    dtrace_state_t *, uint64_t, uint64_t);
static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
static void dtrace_buffer_drop(dtrace_buffer_t *);
static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
    dtrace_state_t *, dtrace_mstate_t *);
static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
    dtrace_optval_t);
static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);

/*
 * DTrace Probe Context Functions
 *
 * These functions are called from probe context.  Because probe context is
 * any context in which C may be called, arbitrarily locks may be held,
 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
 * As a result, functions called from probe context may only call other DTrace
 * support functions -- they may not interact at all with the system at large.
 * (Note that the ASSERT macro is made probe-context safe by redefining it in
 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
 * loads are to be performed from probe context, they _must_ be in terms of
 * the safe dtrace_load*() variants.
 *
 * Some functions in this block are not actually called from probe context;
 * for these functions, there will be a comment above the function reading
 * "Note:  not called from probe context."
 */
void
dtrace_panic(const char *format, ...)
{
	va_list alist;

	va_start(alist, format);
	dtrace_vpanic(format, alist);
	va_end(alist);
}

int
dtrace_assfail(const char *a, const char *f, int l)
{
	dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);

	/*
	 * We just need something here that even the most clever compiler
	 * cannot optimize away.
	 */
	return (a[(uintptr_t)f]);
}

/*
 * Atomically increment a specified error counter from probe context.
 */
static void
dtrace_error(uint32_t *counter)
{
	/*
	 * Most counters stored to in probe context are per-CPU counters.
	 * However, there are some error conditions that are sufficiently
	 * arcane that they don't merit per-CPU storage.  If these counters
	 * are incremented concurrently on different CPUs, scalability will be
	 * adversely affected -- but we don't expect them to be white-hot in a
	 * correctly constructed enabling...
	 */
	uint32_t oval, nval;

	do {
		oval = *counter;

		if ((nval = oval + 1) == 0) {
			/*
			 * If the counter would wrap, set it to 1 -- assuring
			 * that the counter is never zero when we have seen
			 * errors.  (The counter must be 32-bits because we
			 * aren't guaranteed a 64-bit compare&swap operation.)
			 * To save this code both the infamy of being fingered
			 * by a priggish news story and the indignity of being
			 * the target of a neo-puritan witch trial, we're
			 * carefully avoiding any colorful description of the
			 * likelihood of this condition -- but suffice it to
			 * say that it is only slightly more likely than the
			 * overflow of predicate cache IDs, as discussed in
			 * dtrace_predicate_create().
			 */
			nval = 1;
		}
	} while (dtrace_cas32(counter, oval, nval) != oval);
}

/*
 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
 */
DTRACE_LOADFUNC(8)
DTRACE_LOADFUNC(16)
DTRACE_LOADFUNC(32)
DTRACE_LOADFUNC(64)

static int
dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
{
	if (dest < mstate->dtms_scratch_base)
		return (0);

	if (dest + size < dest)
		return (0);

	if (dest + size > mstate->dtms_scratch_ptr)
		return (0);

	return (1);
}

static int
dtrace_canstore_statvar(uint64_t addr, size_t sz,
    dtrace_statvar_t **svars, int nsvars)
{
	int i;

	for (i = 0; i < nsvars; i++) {
		dtrace_statvar_t *svar = svars[i];

		if (svar == NULL || svar->dtsv_size == 0)
			continue;

		if (DTRACE_INRANGE(addr, sz, svar->dtsv_data, svar->dtsv_size))
			return (1);
	}

	return (0);
}

/*
 * Check to see if the address is within a memory region to which a store may
 * be issued.  This includes the DTrace scratch areas, and any DTrace variable
 * region.  The caller of dtrace_canstore() is responsible for performing any
 * alignment checks that are needed before stores are actually executed.
 */
static int
dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
    dtrace_vstate_t *vstate)
{
	/*
	 * First, check to see if the address is in scratch space...
	 */
	if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
	    mstate->dtms_scratch_size))
		return (1);

	/*
	 * Now check to see if it's a dynamic variable.  This check will pick
	 * up both thread-local variables and any global dynamically-allocated
	 * variables.
	 */
	if (DTRACE_INRANGE(addr, sz, (uintptr_t)vstate->dtvs_dynvars.dtds_base,
	    vstate->dtvs_dynvars.dtds_size)) {
		dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
		uintptr_t base = (uintptr_t)dstate->dtds_base +
		    (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
		uintptr_t chunkoffs;

		/*
		 * Before we assume that we can store here, we need to make
		 * sure that it isn't in our metadata -- storing to our
		 * dynamic variable metadata would corrupt our state.  For
		 * the range to not include any dynamic variable metadata,
		 * it must:
		 *
		 *	(1) Start above the hash table that is at the base of
		 *	the dynamic variable space
		 *
		 *	(2) Have a starting chunk offset that is beyond the
		 *	dtrace_dynvar_t that is at the base of every chunk
		 *
		 *	(3) Not span a chunk boundary
		 *
		 */
		if (addr < base)
			return (0);

		chunkoffs = (addr - base) % dstate->dtds_chunksize;

		if (chunkoffs < sizeof (dtrace_dynvar_t))
			return (0);

		if (chunkoffs + sz > dstate->dtds_chunksize)
			return (0);

		return (1);
	}

	/*
	 * Finally, check the static local and global variables.  These checks
	 * take the longest, so we perform them last.
	 */
	if (dtrace_canstore_statvar(addr, sz,
	    vstate->dtvs_locals, vstate->dtvs_nlocals))
		return (1);

	if (dtrace_canstore_statvar(addr, sz,
	    vstate->dtvs_globals, vstate->dtvs_nglobals))
		return (1);

	return (0);
}


/*
 * Convenience routine to check to see if the address is within a memory
 * region in which a load may be issued given the user's privilege level;
 * if not, it sets the appropriate error flags and loads 'addr' into the
 * illegal value slot.
 *
 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
 * appropriate memory access protection.
 */
static int
dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
    dtrace_vstate_t *vstate)
{
	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;

	/*
	 * If we hold the privilege to read from kernel memory, then
	 * everything is readable.
	 */
	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
		return (1);

	/*
	 * You can obviously read that which you can store.
	 */
	if (dtrace_canstore(addr, sz, mstate, vstate))
		return (1);

	/*
	 * We're allowed to read from our own string table.
	 */
	if (DTRACE_INRANGE(addr, sz, (uintptr_t)mstate->dtms_difo->dtdo_strtab,
	    mstate->dtms_difo->dtdo_strlen))
		return (1);

	DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
	*illval = addr;
	return (0);
}

/*
 * Convenience routine to check to see if a given string is within a memory
 * region in which a load may be issued given the user's privilege level;
 * this exists so that we don't need to issue unnecessary dtrace_strlen()
 * calls in the event that the user has all privileges.
 */
static int
dtrace_strcanload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
    dtrace_vstate_t *vstate)
{
	size_t strsz;

	/*
	 * If we hold the privilege to read from kernel memory, then
	 * everything is readable.
	 */
	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
		return (1);

	strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr, sz);
	if (dtrace_canload(addr, strsz, mstate, vstate))
		return (1);

	return (0);
}

/*
 * Convenience routine to check to see if a given variable is within a memory
 * region in which a load may be issued given the user's privilege level.
 */
static int
dtrace_vcanload(void *src, dtrace_diftype_t *type, dtrace_mstate_t *mstate,
    dtrace_vstate_t *vstate)
{
	size_t sz;
	ASSERT(type->dtdt_flags & DIF_TF_BYREF);

	/*
	 * If we hold the privilege to read from kernel memory, then
	 * everything is readable.
	 */
	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
		return (1);

	if (type->dtdt_kind == DIF_TYPE_STRING)
		sz = dtrace_strlen(src,
		    vstate->dtvs_state->dts_options[DTRACEOPT_STRSIZE]) + 1;
	else
		sz = type->dtdt_size;

	return (dtrace_canload((uintptr_t)src, sz, mstate, vstate));
}

/*
 * Compare two strings using safe loads.
 */
static int
dtrace_strncmp(char *s1, char *s2, size_t limit)
{
	uint8_t c1, c2;
	volatile uint16_t *flags;

	if (s1 == s2 || limit == 0)
		return (0);

	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;

	do {
		if (s1 == NULL) {
			c1 = '\0';
		} else {
			c1 = dtrace_load8((uintptr_t)s1++);
		}

		if (s2 == NULL) {
			c2 = '\0';
		} else {
			c2 = dtrace_load8((uintptr_t)s2++);
		}

		if (c1 != c2)
			return (c1 - c2);
	} while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));

	return (0);
}

/*
 * Compute strlen(s) for a string using safe memory accesses.  The additional
 * len parameter is used to specify a maximum length to ensure completion.
 */
static size_t
dtrace_strlen(const char *s, size_t lim)
{
	uint_t len;

	for (len = 0; len != lim; len++) {
		if (dtrace_load8((uintptr_t)s++) == '\0')
			break;
	}

	return (len);
}

/*
 * Check if an address falls within a toxic region.
 */
static int
dtrace_istoxic(uintptr_t kaddr, size_t size)
{
	uintptr_t taddr, tsize;
	int i;

	for (i = 0; i < dtrace_toxranges; i++) {
		taddr = dtrace_toxrange[i].dtt_base;
		tsize = dtrace_toxrange[i].dtt_limit - taddr;

		if (kaddr - taddr < tsize) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = kaddr;
			return (1);
		}

		if (taddr - kaddr < size) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = taddr;
			return (1);
		}
	}

	return (0);
}

/*
 * Copy src to dst using safe memory accesses.  The src is assumed to be unsafe
 * memory specified by the DIF program.  The dst is assumed to be safe memory
 * that we can store to directly because it is managed by DTrace.  As with
 * standard bcopy, overlapping copies are handled properly.
 */
static void
dtrace_bcopy(const void *src, void *dst, size_t len)
{
	if (len != 0) {
		uint8_t *s1 = dst;
		const uint8_t *s2 = src;

		if (s1 <= s2) {
			do {
				*s1++ = dtrace_load8((uintptr_t)s2++);
			} while (--len != 0);
		} else {
			s2 += len;
			s1 += len;

			do {
				*--s1 = dtrace_load8((uintptr_t)--s2);
			} while (--len != 0);
		}
	}
}

/*
 * Copy src to dst using safe memory accesses, up to either the specified
 * length, or the point that a nul byte is encountered.  The src is assumed to
 * be unsafe memory specified by the DIF program.  The dst is assumed to be
 * safe memory that we can store to directly because it is managed by DTrace.
 * Unlike dtrace_bcopy(), overlapping regions are not handled.
 */
static void
dtrace_strcpy(const void *src, void *dst, size_t len)
{
	if (len != 0) {
		uint8_t *s1 = dst, c;
		const uint8_t *s2 = src;

		do {
			*s1++ = c = dtrace_load8((uintptr_t)s2++);
		} while (--len != 0 && c != '\0');
	}
}

/*
 * Copy src to dst, deriving the size and type from the specified (BYREF)
 * variable type.  The src is assumed to be unsafe memory specified by the DIF
 * program.  The dst is assumed to be DTrace variable memory that is of the
 * specified type; we assume that we can store to directly.
 */
static void
dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type)
{
	ASSERT(type->dtdt_flags & DIF_TF_BYREF);

	if (type->dtdt_kind == DIF_TYPE_STRING) {
		dtrace_strcpy(src, dst, type->dtdt_size);
	} else {
		dtrace_bcopy(src, dst, type->dtdt_size);
	}
}

/*
 * Compare s1 to s2 using safe memory accesses.  The s1 data is assumed to be
 * unsafe memory specified by the DIF program.  The s2 data is assumed to be
 * safe memory that we can access directly because it is managed by DTrace.
 */
static int
dtrace_bcmp(const void *s1, const void *s2, size_t len)
{
	volatile uint16_t *flags;

	flags = (volatile uint16_t *)&cpu_core[CPU->cpu_id].cpuc_dtrace_flags;

	if (s1 == s2)
		return (0);

	if (s1 == NULL || s2 == NULL)
		return (1);

	if (s1 != s2 && len != 0) {
		const uint8_t *ps1 = s1;
		const uint8_t *ps2 = s2;

		do {
			if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
				return (1);
		} while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
	}
	return (0);
}

/*
 * Zero the specified region using a simple byte-by-byte loop.  Note that this
 * is for safe DTrace-managed memory only.
 */
static void
dtrace_bzero(void *dst, size_t len)
{
	uchar_t *cp;

	for (cp = dst; len != 0; len--)
		*cp++ = 0;
}

static void
dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
{
	uint64_t result[2];

	result[0] = addend1[0] + addend2[0];
	result[1] = addend1[1] + addend2[1] +
	    (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);

	sum[0] = result[0];
	sum[1] = result[1];
}

/*
 * Shift the 128-bit value in a by b. If b is positive, shift left.
 * If b is negative, shift right.
 */
static void
dtrace_shift_128(uint64_t *a, int b)
{
	uint64_t mask;

	if (b == 0)
		return;

	if (b < 0) {
		b = -b;
		if (b >= 64) {
			a[0] = a[1] >> (b - 64);
			a[1] = 0;
		} else {
			a[0] >>= b;
			mask = 1LL << (64 - b);
			mask -= 1;
			a[0] |= ((a[1] & mask) << (64 - b));
			a[1] >>= b;
		}
	} else {
		if (b >= 64) {
			a[1] = a[0] << (b - 64);
			a[0] = 0;
		} else {
			a[1] <<= b;
			mask = a[0] >> (64 - b);
			a[1] |= mask;
			a[0] <<= b;
		}
	}
}

/*
 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
 * use native multiplication on those, and then re-combine into the
 * resulting 128-bit value.
 *
 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
 *     hi1 * hi2 << 64 +
 *     hi1 * lo2 << 32 +
 *     hi2 * lo1 << 32 +
 *     lo1 * lo2
 */
static void
dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
{
	uint64_t hi1, hi2, lo1, lo2;
	uint64_t tmp[2];

	hi1 = factor1 >> 32;
	hi2 = factor2 >> 32;

	lo1 = factor1 & DT_MASK_LO;
	lo2 = factor2 & DT_MASK_LO;

	product[0] = lo1 * lo2;
	product[1] = hi1 * hi2;

	tmp[0] = hi1 * lo2;
	tmp[1] = 0;
	dtrace_shift_128(tmp, 32);
	dtrace_add_128(product, tmp, product);

	tmp[0] = hi2 * lo1;
	tmp[1] = 0;
	dtrace_shift_128(tmp, 32);
	dtrace_add_128(product, tmp, product);
}

/*
 * This privilege check should be used by actions and subroutines to
 * verify that the user credentials of the process that enabled the
 * invoking ECB match the target credentials
 */
static int
dtrace_priv_proc_common_user(dtrace_state_t *state)
{
	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;

	/*
	 * We should always have a non-NULL state cred here, since if cred
	 * is null (anonymous tracing), we fast-path bypass this routine.
	 */
	ASSERT(s_cr != NULL);

	if ((cr = CRED()) != NULL &&
	    s_cr->cr_uid == cr->cr_uid &&
	    s_cr->cr_uid == cr->cr_ruid &&
	    s_cr->cr_uid == cr->cr_suid &&
	    s_cr->cr_gid == cr->cr_gid &&
	    s_cr->cr_gid == cr->cr_rgid &&
	    s_cr->cr_gid == cr->cr_sgid)
		return (1);

	return (0);
}

/*
 * This privilege check should be used by actions and subroutines to
 * verify that the zone of the process that enabled the invoking ECB
 * matches the target credentials
 */
static int
dtrace_priv_proc_common_zone(dtrace_state_t *state)
{
	cred_t *cr, *s_cr = state->dts_cred.dcr_cred;

	/*
	 * We should always have a non-NULL state cred here, since if cred
	 * is null (anonymous tracing), we fast-path bypass this routine.
	 */
	ASSERT(s_cr != NULL);

	if ((cr = CRED()) != NULL &&
	    s_cr->cr_zone == cr->cr_zone)
		return (1);

	return (0);
}

/*
 * This privilege check should be used by actions and subroutines to
 * verify that the process has not setuid or changed credentials.
 */
static int
dtrace_priv_proc_common_nocd()
{
	proc_t *proc;

	if ((proc = ttoproc(curthread)) != NULL &&
	    !(proc->p_flag & SNOCD))
		return (1);

	return (0);
}

static int
dtrace_priv_proc_destructive(dtrace_state_t *state)
{
	int action = state->dts_cred.dcr_action;

	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
	    dtrace_priv_proc_common_zone(state) == 0)
		goto bad;

	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
	    dtrace_priv_proc_common_user(state) == 0)
		goto bad;

	if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
	    dtrace_priv_proc_common_nocd() == 0)
		goto bad;

	return (1);

bad:
	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;

	return (0);
}

static int
dtrace_priv_proc_control(dtrace_state_t *state)
{
	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
		return (1);

	if (dtrace_priv_proc_common_zone(state) &&
	    dtrace_priv_proc_common_user(state) &&
	    dtrace_priv_proc_common_nocd())
		return (1);

	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;

	return (0);
}

static int
dtrace_priv_proc(dtrace_state_t *state)
{
	if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
		return (1);

	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;

	return (0);
}

static int
dtrace_priv_kernel(dtrace_state_t *state)
{
	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
		return (1);

	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;

	return (0);
}

static int
dtrace_priv_kernel_destructive(dtrace_state_t *state)
{
	if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
		return (1);

	cpu_core[CPU->cpu_id].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;

	return (0);
}

/*
 * Note:  not called from probe context.  This function is called
 * asynchronously (and at a regular interval) from outside of probe context to
 * clean the dirty dynamic variable lists on all CPUs.  Dynamic variable
 * cleaning is explained in detail in <sys/dtrace_impl.h>.
 */
void
dtrace_dynvar_clean(dtrace_dstate_t *dstate)
{
	dtrace_dynvar_t *dirty;
	dtrace_dstate_percpu_t *dcpu;
	int i, work = 0;

	for (i = 0; i < NCPU; i++) {
		dcpu = &dstate->dtds_percpu[i];

		ASSERT(dcpu->dtdsc_rinsing == NULL);

		/*
		 * If the dirty list is NULL, there is no dirty work to do.
		 */
		if (dcpu->dtdsc_dirty == NULL)
			continue;

		/*
		 * If the clean list is non-NULL, then we're not going to do
		 * any work for this CPU -- it means that there has not been
		 * a dtrace_dynvar() allocation on this CPU (or from this CPU)
		 * since the last time we cleaned house.
		 */
		if (dcpu->dtdsc_clean != NULL)
			continue;

		work = 1;

		/*
		 * Atomically move the dirty list aside.
		 */
		do {
			dirty = dcpu->dtdsc_dirty;

			/*
			 * Before we zap the dirty list, set the rinsing list.
			 * (This allows for a potential assertion in
			 * dtrace_dynvar():  if a free dynamic variable appears
			 * on a hash chain, either the dirty list or the
			 * rinsing list for some CPU must be non-NULL.)
			 */
			dcpu->dtdsc_rinsing = dirty;
			dtrace_membar_producer();
		} while (dtrace_casptr(&dcpu->dtdsc_dirty,
		    dirty, NULL) != dirty);
	}

	if (!work) {
		/*
		 * We have no work to do; we can simply return.
		 */
		return;
	}

	dtrace_sync();

	for (i = 0; i < NCPU; i++) {
		dcpu = &dstate->dtds_percpu[i];

		if (dcpu->dtdsc_rinsing == NULL)
			continue;

		/*
		 * We are now guaranteed that no hash chain contains a pointer
		 * into this dirty list; we can make it clean.
		 */
		ASSERT(dcpu->dtdsc_clean == NULL);
		dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
		dcpu->dtdsc_rinsing = NULL;
	}

	/*
	 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
	 * sure that all CPUs have seen all of the dtdsc_clean pointers.
	 * This prevents a race whereby a CPU incorrectly decides that
	 * the state should be something other than DTRACE_DSTATE_CLEAN
	 * after dtrace_dynvar_clean() has completed.
	 */
	dtrace_sync();

	dstate->dtds_state = DTRACE_DSTATE_CLEAN;
}

/*
 * Depending on the value of the op parameter, this function looks-up,
 * allocates or deallocates an arbitrarily-keyed dynamic variable.  If an
 * allocation is requested, this function will return a pointer to a
 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
 * variable can be allocated.  If NULL is returned, the appropriate counter
 * will be incremented.
 */
dtrace_dynvar_t *
dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
    dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
    dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
{
	uint64_t hashval = DTRACE_DYNHASH_VALID;
	dtrace_dynhash_t *hash = dstate->dtds_hash;
	dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
	processorid_t me = CPU->cpu_id, cpu = me;
	dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
	size_t bucket, ksize;
	size_t chunksize = dstate->dtds_chunksize;
	uintptr_t kdata, lock, nstate;
	uint_t i;

	ASSERT(nkeys != 0);

	/*
	 * Hash the key.  As with aggregations, we use Jenkins' "One-at-a-time"
	 * algorithm.  For the by-value portions, we perform the algorithm in
	 * 16-bit chunks (as opposed to 8-bit chunks).  This speeds things up a
	 * bit, and seems to have only a minute effect on distribution.  For
	 * the by-reference data, we perform "One-at-a-time" iterating (safely)
	 * over each referenced byte.  It's painful to do this, but it's much
	 * better than pathological hash distribution.  The efficacy of the
	 * hashing algorithm (and a comparison with other algorithms) may be
	 * found by running the ::dtrace_dynstat MDB dcmd.
	 */
	for (i = 0; i < nkeys; i++) {
		if (key[i].dttk_size == 0) {
			uint64_t val = key[i].dttk_value;

			hashval += (val >> 48) & 0xffff;
			hashval += (hashval << 10);
			hashval ^= (hashval >> 6);

			hashval += (val >> 32) & 0xffff;
			hashval += (hashval << 10);
			hashval ^= (hashval >> 6);

			hashval += (val >> 16) & 0xffff;
			hashval += (hashval << 10);
			hashval ^= (hashval >> 6);

			hashval += val & 0xffff;
			hashval += (hashval << 10);
			hashval ^= (hashval >> 6);
		} else {
			/*
			 * This is incredibly painful, but it beats the hell
			 * out of the alternative.
			 */
			uint64_t j, size = key[i].dttk_size;
			uintptr_t base = (uintptr_t)key[i].dttk_value;

			if (!dtrace_canload(base, size, mstate, vstate))
				break;

			for (j = 0; j < size; j++) {
				hashval += dtrace_load8(base + j);
				hashval += (hashval << 10);
				hashval ^= (hashval >> 6);
			}
		}
	}

	if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
		return (NULL);

	hashval += (hashval << 3);
	hashval ^= (hashval >> 11);
	hashval += (hashval << 15);

	/*
	 * There is a remote chance (ideally, 1 in 2^31) that our hashval
	 * comes out to be one of our two sentinel hash values.  If this
	 * actually happens, we set the hashval to be a value known to be a
	 * non-sentinel value.
	 */
	if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
		hashval = DTRACE_DYNHASH_VALID;

	/*
	 * Yes, it's painful to do a divide here.  If the cycle count becomes
	 * important here, tricks can be pulled to reduce it.  (However, it's
	 * critical that hash collisions be kept to an absolute minimum;
	 * they're much more painful than a divide.)  It's better to have a
	 * solution that generates few collisions and still keeps things
	 * relatively simple.
	 */
	bucket = hashval % dstate->dtds_hashsize;

	if (op == DTRACE_DYNVAR_DEALLOC) {
		volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;

		for (;;) {
			while ((lock = *lockp) & 1)
				continue;

			if (dtrace_casptr((void *)lockp,
			    (void *)lock, (void *)(lock + 1)) == (void *)lock)
				break;
		}

		dtrace_membar_producer();
	}

top:
	prev = NULL;
	lock = hash[bucket].dtdh_lock;

	dtrace_membar_consumer();

	start = hash[bucket].dtdh_chain;
	ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
	    start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
	    op != DTRACE_DYNVAR_DEALLOC));

	for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
		dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
		dtrace_key_t *dkey = &dtuple->dtt_key[0];

		if (dvar->dtdv_hashval != hashval) {
			if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
				/*
				 * We've reached the sink, and therefore the
				 * end of the hash chain; we can kick out of
				 * the loop knowing that we have seen a valid
				 * snapshot of state.
				 */
				ASSERT(dvar->dtdv_next == NULL);
				ASSERT(dvar == &dtrace_dynhash_sink);
				break;
			}

			if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
				/*
				 * We've gone off the rails:  somewhere along
				 * the line, one of the members of this hash
				 * chain was deleted.  Note that we could also
				 * detect this by simply letting this loop run
				 * to completion, as we would eventually hit
				 * the end of the dirty list.  However, we
				 * want to avoid running the length of the
				 * dirty list unnecessarily (it might be quite
				 * long), so we catch this as early as
				 * possible by detecting the hash marker.  In
				 * this case, we simply set dvar to NULL and
				 * break; the conditional after the loop will
				 * send us back to top.
				 */
				dvar = NULL;
				break;
			}

			goto next;
		}

		if (dtuple->dtt_nkeys != nkeys)
			goto next;

		for (i = 0; i < nkeys; i++, dkey++) {
			if (dkey->dttk_size != key[i].dttk_size)
				goto next; /* size or type mismatch */

			if (dkey->dttk_size != 0) {
				if (dtrace_bcmp(
				    (void *)(uintptr_t)key[i].dttk_value,
				    (void *)(uintptr_t)dkey->dttk_value,
				    dkey->dttk_size))
					goto next;
			} else {
				if (dkey->dttk_value != key[i].dttk_value)
					goto next;
			}
		}

		if (op != DTRACE_DYNVAR_DEALLOC)
			return (dvar);

		ASSERT(dvar->dtdv_next == NULL ||
		    dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);

		if (prev != NULL) {
			ASSERT(hash[bucket].dtdh_chain != dvar);
			ASSERT(start != dvar);
			ASSERT(prev->dtdv_next == dvar);
			prev->dtdv_next = dvar->dtdv_next;
		} else {
			if (dtrace_casptr(&hash[bucket].dtdh_chain,
			    start, dvar->dtdv_next) != start) {
				/*
				 * We have failed to atomically swing the
				 * hash table head pointer, presumably because
				 * of a conflicting allocation on another CPU.
				 * We need to reread the hash chain and try
				 * again.
				 */
				goto top;
			}
		}

		dtrace_membar_producer();

		/*
		 * Now set the hash value to indicate that it's free.
		 */
		ASSERT(hash[bucket].dtdh_chain != dvar);
		dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;

		dtrace_membar_producer();

		/*
		 * Set the next pointer to point at the dirty list, and
		 * atomically swing the dirty pointer to the newly freed dvar.
		 */
		do {
			next = dcpu->dtdsc_dirty;
			dvar->dtdv_next = next;
		} while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);

		/*
		 * Finally, unlock this hash bucket.
		 */
		ASSERT(hash[bucket].dtdh_lock == lock);
		ASSERT(lock & 1);
		hash[bucket].dtdh_lock++;

		return (NULL);
next:
		prev = dvar;
		continue;
	}

	if (dvar == NULL) {
		/*
		 * If dvar is NULL, it is because we went off the rails:
		 * one of the elements that we traversed in the hash chain
		 * was deleted while we were traversing it.  In this case,
		 * we assert that we aren't doing a dealloc (deallocs lock
		 * the hash bucket to prevent themselves from racing with
		 * one another), and retry the hash chain traversal.
		 */
		ASSERT(op != DTRACE_DYNVAR_DEALLOC);
		goto top;
	}

	if (op != DTRACE_DYNVAR_ALLOC) {
		/*
		 * If we are not to allocate a new variable, we want to
		 * return NULL now.  Before we return, check that the value
		 * of the lock word hasn't changed.  If it has, we may have
		 * seen an inconsistent snapshot.
		 */
		if (op == DTRACE_DYNVAR_NOALLOC) {
			if (hash[bucket].dtdh_lock != lock)
				goto top;
		} else {
			ASSERT(op == DTRACE_DYNVAR_DEALLOC);
			ASSERT(hash[bucket].dtdh_lock == lock);
			ASSERT(lock & 1);
			hash[bucket].dtdh_lock++;
		}

		return (NULL);
	}

	/*
	 * We need to allocate a new dynamic variable.  The size we need is the
	 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
	 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
	 * the size of any referred-to data (dsize).  We then round the final
	 * size up to the chunksize for allocation.
	 */
	for (ksize = 0, i = 0; i < nkeys; i++)
		ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));

	/*
	 * This should be pretty much impossible, but could happen if, say,
	 * strange DIF specified the tuple.  Ideally, this should be an
	 * assertion and not an error condition -- but that requires that the
	 * chunksize calculation in dtrace_difo_chunksize() be absolutely
	 * bullet-proof.  (That is, it must not be able to be fooled by
	 * malicious DIF.)  Given the lack of backwards branches in DIF,
	 * solving this would presumably not amount to solving the Halting
	 * Problem -- but it still seems awfully hard.
	 */
	if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
	    ksize + dsize > chunksize) {
		dcpu->dtdsc_drops++;
		return (NULL);
	}

	nstate = DTRACE_DSTATE_EMPTY;

	do {
retry:
		free = dcpu->dtdsc_free;

		if (free == NULL) {
			dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
			void *rval;

			if (clean == NULL) {
				/*
				 * We're out of dynamic variable space on
				 * this CPU.  Unless we have tried all CPUs,
				 * we'll try to allocate from a different
				 * CPU.
				 */
				switch (dstate->dtds_state) {
				case DTRACE_DSTATE_CLEAN: {
					void *sp = &dstate->dtds_state;

					if (++cpu >= NCPU)
						cpu = 0;

					if (dcpu->dtdsc_dirty != NULL &&
					    nstate == DTRACE_DSTATE_EMPTY)
						nstate = DTRACE_DSTATE_DIRTY;

					if (dcpu->dtdsc_rinsing != NULL)
						nstate = DTRACE_DSTATE_RINSING;

					dcpu = &dstate->dtds_percpu[cpu];

					if (cpu != me)
						goto retry;

					(void) dtrace_cas32(sp,
					    DTRACE_DSTATE_CLEAN, nstate);

					/*
					 * To increment the correct bean
					 * counter, take another lap.
					 */
					goto retry;
				}

				case DTRACE_DSTATE_DIRTY:
					dcpu->dtdsc_dirty_drops++;
					break;

				case DTRACE_DSTATE_RINSING:
					dcpu->dtdsc_rinsing_drops++;
					break;

				case DTRACE_DSTATE_EMPTY:
					dcpu->dtdsc_drops++;
					break;
				}

				DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
				return (NULL);
			}

			/*
			 * The clean list appears to be non-empty.  We want to
			 * move the clean list to the free list; we start by
			 * moving the clean pointer aside.
			 */
			if (dtrace_casptr(&dcpu->dtdsc_clean,
			    clean, NULL) != clean) {
				/*
				 * We are in one of two situations:
				 *
				 *  (a)	The clean list was switched to the
				 *	free list by another CPU.
				 *
				 *  (b)	The clean list was added to by the
				 *	cleansing cyclic.
				 *
				 * In either of these situations, we can
				 * just reattempt the free list allocation.
				 */
				goto retry;
			}

			ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);

			/*
			 * Now we'll move the clean list to the free list.
			 * It's impossible for this to fail:  the only way
			 * the free list can be updated is through this
			 * code path, and only one CPU can own the clean list.
			 * Thus, it would only be possible for this to fail if
			 * this code were racing with dtrace_dynvar_clean().
			 * (That is, if dtrace_dynvar_clean() updated the clean
			 * list, and we ended up racing to update the free
			 * list.)  This race is prevented by the dtrace_sync()
			 * in dtrace_dynvar_clean() -- which flushes the
			 * owners of the clean lists out before resetting
			 * the clean lists.
			 */
			rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
			ASSERT(rval == NULL);
			goto retry;
		}

		dvar = free;
		new_free = dvar->dtdv_next;
	} while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);

	/*
	 * We have now allocated a new chunk.  We copy the tuple keys into the
	 * tuple array and copy any referenced key data into the data space
	 * following the tuple array.  As we do this, we relocate dttk_value
	 * in the final tuple to point to the key data address in the chunk.
	 */
	kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
	dvar->dtdv_data = (void *)(kdata + ksize);
	dvar->dtdv_tuple.dtt_nkeys = nkeys;

	for (i = 0; i < nkeys; i++) {
		dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
		size_t kesize = key[i].dttk_size;

		if (kesize != 0) {
			dtrace_bcopy(
			    (const void *)(uintptr_t)key[i].dttk_value,
			    (void *)kdata, kesize);
			dkey->dttk_value = kdata;
			kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
		} else {
			dkey->dttk_value = key[i].dttk_value;
		}

		dkey->dttk_size = kesize;
	}

	ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
	dvar->dtdv_hashval = hashval;
	dvar->dtdv_next = start;

	if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
		return (dvar);

	/*
	 * The cas has failed.  Either another CPU is adding an element to
	 * this hash chain, or another CPU is deleting an element from this
	 * hash chain.  The simplest way to deal with both of these cases
	 * (though not necessarily the most efficient) is to free our
	 * allocated block and tail-call ourselves.  Note that the free is
	 * to the dirty list and _not_ to the free list.  This is to prevent
	 * races with allocators, above.
	 */
	dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;

	dtrace_membar_producer();

	do {
		free = dcpu->dtdsc_dirty;
		dvar->dtdv_next = free;
	} while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);

	return (dtrace_dynvar(dstate, nkeys, key, dsize, op, mstate, vstate));
}

/*ARGSUSED*/
static void
dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
{
	if ((int64_t)nval < (int64_t)*oval)
		*oval = nval;
}

/*ARGSUSED*/
static void
dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
{
	if ((int64_t)nval > (int64_t)*oval)
		*oval = nval;
}

static void
dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
{
	int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
	int64_t val = (int64_t)nval;

	if (val < 0) {
		for (i = 0; i < zero; i++) {
			if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
				quanta[i] += incr;
				return;
			}
		}
	} else {
		for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
			if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
				quanta[i - 1] += incr;
				return;
			}
		}

		quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
		return;
	}

	ASSERT(0);
}

static void
dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
{
	uint64_t arg = *lquanta++;
	int32_t base = DTRACE_LQUANTIZE_BASE(arg);
	uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
	uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
	int32_t val = (int32_t)nval, level;

	ASSERT(step != 0);
	ASSERT(levels != 0);

	if (val < base) {
		/*
		 * This is an underflow.
		 */
		lquanta[0] += incr;
		return;
	}

	level = (val - base) / step;

	if (level < levels) {
		lquanta[level + 1] += incr;
		return;
	}

	/*
	 * This is an overflow.
	 */
	lquanta[levels + 1] += incr;
}

/*ARGSUSED*/
static void
dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
{
	data[0]++;
	data[1] += nval;
}

/*ARGSUSED*/
static void
dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
{
	int64_t snval = (int64_t)nval;
	uint64_t tmp[2];

	data[0]++;
	data[1] += nval;

	/*
	 * What we want to say here is:
	 *
	 * data[2] += nval * nval;
	 *
	 * But given that nval is 64-bit, we could easily overflow, so
	 * we do this as 128-bit arithmetic.
	 */
	if (snval < 0)
		snval = -snval;

	dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
	dtrace_add_128(data + 2, tmp, data + 2);
}

/*ARGSUSED*/
static void
dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
{
	*oval = *oval + 1;
}

/*ARGSUSED*/
static void
dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
{
	*oval += nval;
}

/*
 * Aggregate given the tuple in the principal data buffer, and the aggregating
 * action denoted by the specified dtrace_aggregation_t.  The aggregation
 * buffer is specified as the buf parameter.  This routine does not return
 * failure; if there is no space in the aggregation buffer, the data will be
 * dropped, and a corresponding counter incremented.
 */
static void
dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
    intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
{
	dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
	uint32_t i, ndx, size, fsize;
	uint32_t align = sizeof (uint64_t) - 1;
	dtrace_aggbuffer_t *agb;
	dtrace_aggkey_t *key;
	uint32_t hashval = 0, limit, isstr;
	caddr_t tomax, data, kdata;
	dtrace_actkind_t action;
	dtrace_action_t *act;
	uintptr_t offs;

	if (buf == NULL)
		return;

	if (!agg->dtag_hasarg) {
		/*
		 * Currently, only quantize() and lquantize() take additional
		 * arguments, and they have the same semantics:  an increment
		 * value that defaults to 1 when not present.  If additional
		 * aggregating actions take arguments, the setting of the
		 * default argument value will presumably have to become more
		 * sophisticated...
		 */
		arg = 1;
	}

	action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
	size = rec->dtrd_offset - agg->dtag_base;
	fsize = size + rec->dtrd_size;

	ASSERT(dbuf->dtb_tomax != NULL);
	data = dbuf->dtb_tomax + offset + agg->dtag_base;

	if ((tomax = buf->dtb_tomax) == NULL) {
		dtrace_buffer_drop(buf);
		return;
	}

	/*
	 * The metastructure is always at the bottom of the buffer.
	 */
	agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
	    sizeof (dtrace_aggbuffer_t));

	if (buf->dtb_offset == 0) {
		/*
		 * We just kludge up approximately 1/8th of the size to be
		 * buckets.  If this guess ends up being routinely
		 * off-the-mark, we may need to dynamically readjust this
		 * based on past performance.
		 */
		uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);

		if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
		    (uintptr_t)tomax || hashsize == 0) {
			/*
			 * We've been given a ludicrously small buffer;
			 * increment our drop count and leave.
			 */
			dtrace_buffer_drop(buf);
			return;
		}

		/*
		 * And now, a pathetic attempt to try to get a an odd (or
		 * perchance, a prime) hash size for better hash distribution.
		 */
		if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
			hashsize -= DTRACE_AGGHASHSIZE_SLEW;

		agb->dtagb_hashsize = hashsize;
		agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
		    agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
		agb->dtagb_free = (uintptr_t)agb->dtagb_hash;

		for (i = 0; i < agb->dtagb_hashsize; i++)
			agb->dtagb_hash[i] = NULL;
	}

	ASSERT(agg->dtag_first != NULL);
	ASSERT(agg->dtag_first->dta_intuple);

	/*
	 * Calculate the hash value based on the key.  Note that we _don't_
	 * include the aggid in the hashing (but we will store it as part of
	 * the key).  The hashing algorithm is Bob Jenkins' "One-at-a-time"
	 * algorithm: a simple, quick algorithm that has no known funnels, and
	 * gets good distribution in practice.  The efficacy of the hashing
	 * algorithm (and a comparison with other algorithms) may be found by
	 * running the ::dtrace_aggstat MDB dcmd.
	 */
	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
		i = act->dta_rec.dtrd_offset - agg->dtag_base;
		limit = i + act->dta_rec.dtrd_size;
		ASSERT(limit <= size);
		isstr = DTRACEACT_ISSTRING(act);

		for (; i < limit; i++) {
			hashval += data[i];
			hashval += (hashval << 10);
			hashval ^= (hashval >> 6);

			if (isstr && data[i] == '\0')
				break;
		}
	}

	hashval += (hashval << 3);
	hashval ^= (hashval >> 11);
	hashval += (hashval << 15);

	/*
	 * Yes, the divide here is expensive -- but it's generally the least
	 * of the performance issues given the amount of data that we iterate
	 * over to compute hash values, compare data, etc.
	 */
	ndx = hashval % agb->dtagb_hashsize;

	for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
		ASSERT((caddr_t)key >= tomax);
		ASSERT((caddr_t)key < tomax + buf->dtb_size);

		if (hashval != key->dtak_hashval || key->dtak_size != size)
			continue;

		kdata = key->dtak_data;
		ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);

		for (act = agg->dtag_first; act->dta_intuple;
		    act = act->dta_next) {
			i = act->dta_rec.dtrd_offset - agg->dtag_base;
			limit = i + act->dta_rec.dtrd_size;
			ASSERT(limit <= size);
			isstr = DTRACEACT_ISSTRING(act);

			for (; i < limit; i++) {
				if (kdata[i] != data[i])
					goto next;

				if (isstr && data[i] == '\0')
					break;
			}
		}

		if (action != key->dtak_action) {
			/*
			 * We are aggregating on the same value in the same
			 * aggregation with two different aggregating actions.
			 * (This should have been picked up in the compiler,
			 * so we may be dealing with errant or devious DIF.)
			 * This is an error condition; we indicate as much,
			 * and return.
			 */
			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
			return;
		}

		/*
		 * This is a hit:  we need to apply the aggregator to
		 * the value at this key.
		 */
		agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
		return;
next:
		continue;
	}

	/*
	 * We didn't find it.  We need to allocate some zero-filled space,
	 * link it into the hash table appropriately, and apply the aggregator
	 * to the (zero-filled) value.
	 */
	offs = buf->dtb_offset;
	while (offs & (align - 1))
		offs += sizeof (uint32_t);

	/*
	 * If we don't have enough room to both allocate a new key _and_
	 * its associated data, increment the drop count and return.
	 */
	if ((uintptr_t)tomax + offs + fsize >
	    agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
		dtrace_buffer_drop(buf);
		return;
	}

	/*CONSTCOND*/
	ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
	key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
	agb->dtagb_free -= sizeof (dtrace_aggkey_t);

	key->dtak_data = kdata = tomax + offs;
	buf->dtb_offset = offs + fsize;

	/*
	 * Now copy the data across.
	 */
	*((dtrace_aggid_t *)kdata) = agg->dtag_id;

	for (i = sizeof (dtrace_aggid_t); i < size; i++)
		kdata[i] = data[i];

	/*
	 * Because strings are not zeroed out by default, we need to iterate
	 * looking for actions that store strings, and we need to explicitly
	 * pad these strings out with zeroes.
	 */
	for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
		int nul;

		if (!DTRACEACT_ISSTRING(act))
			continue;

		i = act->dta_rec.dtrd_offset - agg->dtag_base;
		limit = i + act->dta_rec.dtrd_size;
		ASSERT(limit <= size);

		for (nul = 0; i < limit; i++) {
			if (nul) {
				kdata[i] = '\0';
				continue;
			}

			if (data[i] != '\0')
				continue;

			nul = 1;
		}
	}

	for (i = size; i < fsize; i++)
		kdata[i] = 0;

	key->dtak_hashval = hashval;
	key->dtak_size = size;
	key->dtak_action = action;
	key->dtak_next = agb->dtagb_hash[ndx];
	agb->dtagb_hash[ndx] = key;

	/*
	 * Finally, apply the aggregator.
	 */
	*((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
	agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
}

/*
 * Given consumer state, this routine finds a speculation in the INACTIVE
 * state and transitions it into the ACTIVE state.  If there is no speculation
 * in the INACTIVE state, 0 is returned.  In this case, no error counter is
 * incremented -- it is up to the caller to take appropriate action.
 */
static int
dtrace_speculation(dtrace_state_t *state)
{
	int i = 0;
	dtrace_speculation_state_t current;
	uint32_t *stat = &state->dts_speculations_unavail, count;

	while (i < state->dts_nspeculations) {
		dtrace_speculation_t *spec = &state->dts_speculations[i];

		current = spec->dtsp_state;

		if (current != DTRACESPEC_INACTIVE) {
			if (current == DTRACESPEC_COMMITTINGMANY ||
			    current == DTRACESPEC_COMMITTING ||
			    current == DTRACESPEC_DISCARDING)
				stat = &state->dts_speculations_busy;
			i++;
			continue;
		}

		if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
		    current, DTRACESPEC_ACTIVE) == current)
			return (i + 1);
	}

	/*
	 * We couldn't find a speculation.  If we found as much as a single
	 * busy speculation buffer, we'll attribute this failure as "busy"
	 * instead of "unavail".
	 */
	do {
		count = *stat;
	} while (dtrace_cas32(stat, count, count + 1) != count);

	return (0);
}

/*
 * This routine commits an active speculation.  If the specified speculation
 * is not in a valid state to perform a commit(), this routine will silently do
 * nothing.  The state of the specified speculation is transitioned according
 * to the state transition diagram outlined in <sys/dtrace_impl.h>
 */
static void
dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
    dtrace_specid_t which)
{
	dtrace_speculation_t *spec;
	dtrace_buffer_t *src, *dest;
	uintptr_t daddr, saddr, dlimit;
	dtrace_speculation_state_t current, new;
	intptr_t offs;

	if (which == 0)
		return;

	if (which > state->dts_nspeculations) {
		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
		return;
	}

	spec = &state->dts_speculations[which - 1];
	src = &spec->dtsp_buffer[cpu];
	dest = &state->dts_buffer[cpu];

	do {
		current = spec->dtsp_state;

		if (current == DTRACESPEC_COMMITTINGMANY)
			break;

		switch (current) {
		case DTRACESPEC_INACTIVE:
		case DTRACESPEC_DISCARDING:
			return;

		case DTRACESPEC_COMMITTING:
			/*
			 * This is only possible if we are (a) commit()'ing
			 * without having done a prior speculate() on this CPU
			 * and (b) racing with another commit() on a different
			 * CPU.  There's nothing to do -- we just assert that
			 * our offset is 0.
			 */
			ASSERT(src->dtb_offset == 0);
			return;

		case DTRACESPEC_ACTIVE:
			new = DTRACESPEC_COMMITTING;
			break;

		case DTRACESPEC_ACTIVEONE:
			/*
			 * This speculation is active on one CPU.  If our
			 * buffer offset is non-zero, we know that the one CPU
			 * must be us.  Otherwise, we are committing on a
			 * different CPU from the speculate(), and we must
			 * rely on being asynchronously cleaned.
			 */
			if (src->dtb_offset != 0) {
				new = DTRACESPEC_COMMITTING;
				break;
			}
			/*FALLTHROUGH*/

		case DTRACESPEC_ACTIVEMANY:
			new = DTRACESPEC_COMMITTINGMANY;
			break;

		default:
			ASSERT(0);
		}
	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
	    current, new) != current);

	/*
	 * We have set the state to indicate that we are committing this
	 * speculation.  Now reserve the necessary space in the destination
	 * buffer.
	 */
	if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
	    sizeof (uint64_t), state, NULL)) < 0) {
		dtrace_buffer_drop(dest);
		goto out;
	}

	/*
	 * We have the space; copy the buffer across.  (Note that this is a
	 * highly subobtimal bcopy(); in the unlikely event that this becomes
	 * a serious performance issue, a high-performance DTrace-specific
	 * bcopy() should obviously be invented.)
	 */
	daddr = (uintptr_t)dest->dtb_tomax + offs;
	dlimit = daddr + src->dtb_offset;
	saddr = (uintptr_t)src->dtb_tomax;

	/*
	 * First, the aligned portion.
	 */
	while (dlimit - daddr >= sizeof (uint64_t)) {
		*((uint64_t *)daddr) = *((uint64_t *)saddr);

		daddr += sizeof (uint64_t);
		saddr += sizeof (uint64_t);
	}

	/*
	 * Now any left-over bit...
	 */
	while (dlimit - daddr)
		*((uint8_t *)daddr++) = *((uint8_t *)saddr++);

	/*
	 * Finally, commit the reserved space in the destination buffer.
	 */
	dest->dtb_offset = offs + src->dtb_offset;

out:
	/*
	 * If we're lucky enough to be the only active CPU on this speculation
	 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
	 */
	if (current == DTRACESPEC_ACTIVE ||
	    (current == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
		uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
		    DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);

		ASSERT(rval == DTRACESPEC_COMMITTING);
	}

	src->dtb_offset = 0;
	src->dtb_xamot_drops += src->dtb_drops;
	src->dtb_drops = 0;
}

/*
 * This routine discards an active speculation.  If the specified speculation
 * is not in a valid state to perform a discard(), this routine will silently
 * do nothing.  The state of the specified speculation is transitioned
 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
 */
static void
dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
    dtrace_specid_t which)
{
	dtrace_speculation_t *spec;
	dtrace_speculation_state_t current, new;
	dtrace_buffer_t *buf;

	if (which == 0)
		return;

	if (which > state->dts_nspeculations) {
		cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
		return;
	}

	spec = &state->dts_speculations[which - 1];
	buf = &spec->dtsp_buffer[cpu];

	do {
		current = spec->dtsp_state;

		switch (current) {
		case DTRACESPEC_INACTIVE:
		case DTRACESPEC_COMMITTINGMANY:
		case DTRACESPEC_COMMITTING:
		case DTRACESPEC_DISCARDING:
			return;

		case DTRACESPEC_ACTIVE:
		case DTRACESPEC_ACTIVEMANY:
			new = DTRACESPEC_DISCARDING;
			break;

		case DTRACESPEC_ACTIVEONE:
			if (buf->dtb_offset != 0) {
				new = DTRACESPEC_INACTIVE;
			} else {
				new = DTRACESPEC_DISCARDING;
			}
			break;

		default:
			ASSERT(0);
		}
	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
	    current, new) != current);

	buf->dtb_offset = 0;
	buf->dtb_drops = 0;
}

/*
 * Note:  not called from probe context.  This function is called
 * asynchronously from cross call context to clean any speculations that are
 * in the COMMITTINGMANY or DISCARDING states.  These speculations may not be
 * transitioned back to the INACTIVE state until all CPUs have cleaned the
 * speculation.
 */
static void
dtrace_speculation_clean_here(dtrace_state_t *state)
{
	dtrace_icookie_t cookie;
	processorid_t cpu = CPU->cpu_id;
	dtrace_buffer_t *dest = &state->dts_buffer[cpu];
	dtrace_specid_t i;

	cookie = dtrace_interrupt_disable();

	if (dest->dtb_tomax == NULL) {
		dtrace_interrupt_enable(cookie);
		return;
	}

	for (i = 0; i < state->dts_nspeculations; i++) {
		dtrace_speculation_t *spec = &state->dts_speculations[i];
		dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];

		if (src->dtb_tomax == NULL)
			continue;

		if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
			src->dtb_offset = 0;
			continue;
		}

		if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
			continue;

		if (src->dtb_offset == 0)
			continue;

		dtrace_speculation_commit(state, cpu, i + 1);
	}

	dtrace_interrupt_enable(cookie);
}

/*
 * Note:  not called from probe context.  This function is called
 * asynchronously (and at a regular interval) to clean any speculations that
 * are in the COMMITTINGMANY or DISCARDING states.  If it discovers that there
 * is work to be done, it cross calls all CPUs to perform that work;
 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
 * INACTIVE state until they have been cleaned by all CPUs.
 */
static void
dtrace_speculation_clean(dtrace_state_t *state)
{
	int work = 0, rv;
	dtrace_specid_t i;

	for (i = 0; i < state->dts_nspeculations; i++) {
		dtrace_speculation_t *spec = &state->dts_speculations[i];

		ASSERT(!spec->dtsp_cleaning);

		if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
		    spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
			continue;

		work++;
		spec->dtsp_cleaning = 1;
	}

	if (!work)
		return;

	dtrace_xcall(DTRACE_CPUALL,
	    (dtrace_xcall_t)dtrace_speculation_clean_here, state);

	/*
	 * We now know that all CPUs have committed or discarded their
	 * speculation buffers, as appropriate.  We can now set the state
	 * to inactive.
	 */
	for (i = 0; i < state->dts_nspeculations; i++) {
		dtrace_speculation_t *spec = &state->dts_speculations[i];
		dtrace_speculation_state_t current, new;

		if (!spec->dtsp_cleaning)
			continue;

		current = spec->dtsp_state;
		ASSERT(current == DTRACESPEC_DISCARDING ||
		    current == DTRACESPEC_COMMITTINGMANY);

		new = DTRACESPEC_INACTIVE;

		rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, current, new);
		ASSERT(rv == current);
		spec->dtsp_cleaning = 0;
	}
}

/*
 * Called as part of a speculate() to get the speculative buffer associated
 * with a given speculation.  Returns NULL if the specified speculation is not
 * in an ACTIVE state.  If the speculation is in the ACTIVEONE state -- and
 * the active CPU is not the specified CPU -- the speculation will be
 * atomically transitioned into the ACTIVEMANY state.
 */
static dtrace_buffer_t *
dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
    dtrace_specid_t which)
{
	dtrace_speculation_t *spec;
	dtrace_speculation_state_t current, new;
	dtrace_buffer_t *buf;

	if (which == 0)
		return (NULL);

	if (which > state->dts_nspeculations) {
		cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
		return (NULL);
	}

	spec = &state->dts_speculations[which - 1];
	buf = &spec->dtsp_buffer[cpuid];

	do {
		current = spec->dtsp_state;

		switch (current) {
		case DTRACESPEC_INACTIVE:
		case DTRACESPEC_COMMITTINGMANY:
		case DTRACESPEC_DISCARDING:
			return (NULL);

		case DTRACESPEC_COMMITTING:
			ASSERT(buf->dtb_offset == 0);
			return (NULL);

		case DTRACESPEC_ACTIVEONE:
			/*
			 * This speculation is currently active on one CPU.
			 * Check the offset in the buffer; if it's non-zero,
			 * that CPU must be us (and we leave the state alone).
			 * If it's zero, assume that we're starting on a new
			 * CPU -- and change the state to indicate that the
			 * speculation is active on more than one CPU.
			 */
			if (buf->dtb_offset != 0)
				return (buf);

			new = DTRACESPEC_ACTIVEMANY;
			break;

		case DTRACESPEC_ACTIVEMANY:
			return (buf);

		case DTRACESPEC_ACTIVE:
			new = DTRACESPEC_ACTIVEONE;
			break;

		default:
			ASSERT(0);
		}
	} while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
	    current, new) != current);

	ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
	return (buf);
}

/*
 * Return a string.  In the event that the user lacks the privilege to access
 * arbitrary kernel memory, we copy the string out to scratch memory so that we
 * don't fail access checking.
 *
 * dtrace_dif_variable() uses this routine as a helper for various
 * builtin values such as 'execname' and 'probefunc.'
 */
uintptr_t
dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
    dtrace_mstate_t *mstate)
{
	uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
	uintptr_t ret;
	size_t strsz;

	/*
	 * The easy case: this probe is allowed to read all of memory, so
	 * we can just return this as a vanilla pointer.
	 */
	if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
		return (addr);

	/*
	 * This is the tougher case: we copy the string in question from
	 * kernel memory into scratch memory and return it that way: this
	 * ensures that we won't trip up when access checking tests the
	 * BYREF return value.
	 */
	strsz = dtrace_strlen((char *)addr, size) + 1;

	if (mstate->dtms_scratch_ptr + strsz >
	    mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
		return (NULL);
	}

	dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
	    strsz);
	ret = mstate->dtms_scratch_ptr;
	mstate->dtms_scratch_ptr += strsz;
	return (ret);
}

/*
 * This function implements the DIF emulator's variable lookups.  The emulator
 * passes a reserved variable identifier and optional built-in array index.
 */
static uint64_t
dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
    uint64_t ndx)
{
	/*
	 * If we're accessing one of the uncached arguments, we'll turn this
	 * into a reference in the args array.
	 */
	if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
		ndx = v - DIF_VAR_ARG0;
		v = DIF_VAR_ARGS;
	}

	switch (v) {
	case DIF_VAR_ARGS:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
		if (ndx >= sizeof (mstate->dtms_arg) /
		    sizeof (mstate->dtms_arg[0])) {
			int aframes = mstate->dtms_probe->dtpr_aframes + 2;
			dtrace_provider_t *pv;
			uint64_t val;

			pv = mstate->dtms_probe->dtpr_provider;
			if (pv->dtpv_pops.dtps_getargval != NULL)
				val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
				    mstate->dtms_probe->dtpr_id,
				    mstate->dtms_probe->dtpr_arg, ndx, aframes);
			else
				val = dtrace_getarg(ndx, aframes);

			/*
			 * This is regrettably required to keep the compiler
			 * from tail-optimizing the call to dtrace_getarg().
			 * The condition always evaluates to true, but the
			 * compiler has no way of figuring that out a priori.
			 * (None of this would be necessary if the compiler
			 * could be relied upon to _always_ tail-optimize
			 * the call to dtrace_getarg() -- but it can't.)
			 */
			if (mstate->dtms_probe != NULL)
				return (val);

			ASSERT(0);
		}

		return (mstate->dtms_arg[ndx]);

	case DIF_VAR_UREGS: {
		klwp_t *lwp;

		if (!dtrace_priv_proc(state))
			return (0);

		if ((lwp = curthread->t_lwp) == NULL) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
			cpu_core[CPU->cpu_id].cpuc_dtrace_illval = NULL;
			return (0);
		}

		return (dtrace_getreg(lwp->lwp_regs, ndx));
	}

	case DIF_VAR_CURTHREAD:
		if (!dtrace_priv_kernel(state))
			return (0);
		return ((uint64_t)(uintptr_t)curthread);

	case DIF_VAR_TIMESTAMP:
		if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
			mstate->dtms_timestamp = dtrace_gethrtime();
			mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
		}
		return (mstate->dtms_timestamp);

	case DIF_VAR_VTIMESTAMP:
		ASSERT(dtrace_vtime_references != 0);
		return (curthread->t_dtrace_vtime);

	case DIF_VAR_WALLTIMESTAMP:
		if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
			mstate->dtms_walltimestamp = dtrace_gethrestime();
			mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
		}
		return (mstate->dtms_walltimestamp);

	case DIF_VAR_IPL:
		if (!dtrace_priv_kernel(state))
			return (0);
		if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
			mstate->dtms_ipl = dtrace_getipl();
			mstate->dtms_present |= DTRACE_MSTATE_IPL;
		}
		return (mstate->dtms_ipl);

	case DIF_VAR_EPID:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
		return (mstate->dtms_epid);

	case DIF_VAR_ID:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
		return (mstate->dtms_probe->dtpr_id);

	case DIF_VAR_STACKDEPTH:
		if (!dtrace_priv_kernel(state))
			return (0);
		if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
			int aframes = mstate->dtms_probe->dtpr_aframes + 2;

			mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
			mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
		}
		return (mstate->dtms_stackdepth);

	case DIF_VAR_USTACKDEPTH:
		if (!dtrace_priv_proc(state))
			return (0);
		if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
			/*
			 * See comment in DIF_VAR_PID.
			 */
			if (DTRACE_ANCHORED(mstate->dtms_probe) &&
			    CPU_ON_INTR(CPU)) {
				mstate->dtms_ustackdepth = 0;
			} else {
				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
				mstate->dtms_ustackdepth =
				    dtrace_getustackdepth();
				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
			}
			mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
		}
		return (mstate->dtms_ustackdepth);

	case DIF_VAR_CALLER:
		if (!dtrace_priv_kernel(state))
			return (0);
		if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
			int aframes = mstate->dtms_probe->dtpr_aframes + 2;

			if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
				/*
				 * If this is an unanchored probe, we are
				 * required to go through the slow path:
				 * dtrace_caller() only guarantees correct
				 * results for anchored probes.
				 */
				pc_t caller[2];

				dtrace_getpcstack(caller, 2, aframes,
				    (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
				mstate->dtms_caller = caller[1];
			} else if ((mstate->dtms_caller =
			    dtrace_caller(aframes)) == -1) {
				/*
				 * We have failed to do this the quick way;
				 * we must resort to the slower approach of
				 * calling dtrace_getpcstack().
				 */
				pc_t caller;

				dtrace_getpcstack(&caller, 1, aframes, NULL);
				mstate->dtms_caller = caller;
			}

			mstate->dtms_present |= DTRACE_MSTATE_CALLER;
		}
		return (mstate->dtms_caller);

	case DIF_VAR_UCALLER:
		if (!dtrace_priv_proc(state))
			return (0);

		if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
			uint64_t ustack[3];

			/*
			 * dtrace_getupcstack() fills in the first uint64_t
			 * with the current PID.  The second uint64_t will
			 * be the program counter at user-level.  The third
			 * uint64_t will contain the caller, which is what
			 * we're after.
			 */
			ustack[2] = NULL;
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
			dtrace_getupcstack(ustack, 3);
			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
			mstate->dtms_ucaller = ustack[2];
			mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
		}

		return (mstate->dtms_ucaller);

	case DIF_VAR_PROBEPROV:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
		return (dtrace_dif_varstr(
		    (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
		    state, mstate));

	case DIF_VAR_PROBEMOD:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
		return (dtrace_dif_varstr(
		    (uintptr_t)mstate->dtms_probe->dtpr_mod,
		    state, mstate));

	case DIF_VAR_PROBEFUNC:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
		return (dtrace_dif_varstr(
		    (uintptr_t)mstate->dtms_probe->dtpr_func,
		    state, mstate));

	case DIF_VAR_PROBENAME:
		ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
		return (dtrace_dif_varstr(
		    (uintptr_t)mstate->dtms_probe->dtpr_name,
		    state, mstate));

	case DIF_VAR_PID:
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * Note that we are assuming that an unanchored probe is
		 * always due to a high-level interrupt.  (And we're assuming
		 * that there is only a single high level interrupt.)
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return (pid0.pid_id);

		/*
		 * It is always safe to dereference one's own t_procp pointer:
		 * it always points to a valid, allocated proc structure.
		 * Further, it is always safe to dereference the p_pidp member
		 * of one's own proc structure.  (These are truisms becuase
		 * threads and processes don't clean up their own state --
		 * they leave that task to whomever reaps them.)
		 */
		return ((uint64_t)curthread->t_procp->p_pidp->pid_id);

	case DIF_VAR_PPID:
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return (pid0.pid_id);

		/*
		 * It is always safe to dereference one's own t_procp pointer:
		 * it always points to a valid, allocated proc structure.
		 * (This is true because threads don't clean up their own
		 * state -- they leave that task to whomever reaps them.)
		 */
		return ((uint64_t)curthread->t_procp->p_ppid);

	case DIF_VAR_TID:
		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return (0);

		return ((uint64_t)curthread->t_tid);

	case DIF_VAR_EXECNAME:
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return ((uint64_t)(uintptr_t)p0.p_user.u_comm);

		/*
		 * It is always safe to dereference one's own t_procp pointer:
		 * it always points to a valid, allocated proc structure.
		 * (This is true because threads don't clean up their own
		 * state -- they leave that task to whomever reaps them.)
		 */
		return (dtrace_dif_varstr(
		    (uintptr_t)curthread->t_procp->p_user.u_comm,
		    state, mstate));

	case DIF_VAR_ZONENAME:
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);

		/*
		 * It is always safe to dereference one's own t_procp pointer:
		 * it always points to a valid, allocated proc structure.
		 * (This is true because threads don't clean up their own
		 * state -- they leave that task to whomever reaps them.)
		 */
		return (dtrace_dif_varstr(
		    (uintptr_t)curthread->t_procp->p_zone->zone_name,
		    state, mstate));

	case DIF_VAR_UID:
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return ((uint64_t)p0.p_cred->cr_uid);

		/*
		 * It is always safe to dereference one's own t_procp pointer:
		 * it always points to a valid, allocated proc structure.
		 * (This is true because threads don't clean up their own
		 * state -- they leave that task to whomever reaps them.)
		 *
		 * Additionally, it is safe to dereference one's own process
		 * credential, since this is never NULL after process birth.
		 */
		return ((uint64_t)curthread->t_procp->p_cred->cr_uid);

	case DIF_VAR_GID:
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return ((uint64_t)p0.p_cred->cr_gid);

		/*
		 * It is always safe to dereference one's own t_procp pointer:
		 * it always points to a valid, allocated proc structure.
		 * (This is true because threads don't clean up their own
		 * state -- they leave that task to whomever reaps them.)
		 *
		 * Additionally, it is safe to dereference one's own process
		 * credential, since this is never NULL after process birth.
		 */
		return ((uint64_t)curthread->t_procp->p_cred->cr_gid);

	case DIF_VAR_ERRNO: {
		klwp_t *lwp;
		if (!dtrace_priv_proc(state))
			return (0);

		/*
		 * See comment in DIF_VAR_PID.
		 */
		if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
			return (0);

		/*
		 * It is always safe to dereference one's own t_lwp pointer in
		 * the event that this pointer is non-NULL.  (This is true
		 * because threads and lwps don't clean up their own state --
		 * they leave that task to whomever reaps them.)
		 */
		if ((lwp = curthread->t_lwp) == NULL)
			return (0);

		return ((uint64_t)lwp->lwp_errno);
	}
	default:
		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
		return (0);
	}
}

/*
 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
 * Notice that we don't bother validating the proper number of arguments or
 * their types in the tuple stack.  This isn't needed because all argument
 * interpretation is safe because of our load safety -- the worst that can
 * happen is that a bogus program can obtain bogus results.
 */
static void
dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
    dtrace_key_t *tupregs, int nargs,
    dtrace_mstate_t *mstate, dtrace_state_t *state)
{
	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;
	dtrace_vstate_t *vstate = &state->dts_vstate;

	union {
		mutex_impl_t mi;
		uint64_t mx;
	} m;

	union {
		krwlock_t ri;
		uintptr_t rw;
	} r;

	switch (subr) {
	case DIF_SUBR_RAND:
		regs[rd] = (dtrace_gethrtime() * 2416 + 374441) % 1771875;
		break;

	case DIF_SUBR_MUTEX_OWNED:
		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		m.mx = dtrace_load64(tupregs[0].dttk_value);
		if (MUTEX_TYPE_ADAPTIVE(&m.mi))
			regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
		else
			regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
		break;

	case DIF_SUBR_MUTEX_OWNER:
		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		m.mx = dtrace_load64(tupregs[0].dttk_value);
		if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
		    MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
			regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
		else
			regs[rd] = 0;
		break;

	case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		m.mx = dtrace_load64(tupregs[0].dttk_value);
		regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
		break;

	case DIF_SUBR_MUTEX_TYPE_SPIN:
		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		m.mx = dtrace_load64(tupregs[0].dttk_value);
		regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
		break;

	case DIF_SUBR_RW_READ_HELD: {
		uintptr_t tmp;

		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
		regs[rd] = _RW_READ_HELD(&r.ri, tmp);
		break;
	}

	case DIF_SUBR_RW_WRITE_HELD:
		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
		regs[rd] = _RW_WRITE_HELD(&r.ri);
		break;

	case DIF_SUBR_RW_ISWRITER:
		if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
		    mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		r.rw = dtrace_loadptr(tupregs[0].dttk_value);
		regs[rd] = _RW_ISWRITER(&r.ri);
		break;

	case DIF_SUBR_BCOPY: {
		/*
		 * We need to be sure that the destination is in the scratch
		 * region -- no other region is allowed.
		 */
		uintptr_t src = tupregs[0].dttk_value;
		uintptr_t dest = tupregs[1].dttk_value;
		size_t size = tupregs[2].dttk_value;

		if (!dtrace_inscratch(dest, size, mstate)) {
			*flags |= CPU_DTRACE_BADADDR;
			*illval = regs[rd];
			break;
		}

		if (!dtrace_canload(src, size, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		dtrace_bcopy((void *)src, (void *)dest, size);
		break;
	}

	case DIF_SUBR_ALLOCA:
	case DIF_SUBR_COPYIN: {
		uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
		uint64_t size =
		    tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
		size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;

		/*
		 * This action doesn't require any credential checks since
		 * probes will not activate in user contexts to which the
		 * enabling user does not have permissions.
		 */

		/*
		 * Rounding up the user allocation size could have overflowed
		 * a large, bogus allocation (like -1ULL) to 0.
		 */
		if (scratch_size < size ||
		    !DTRACE_INSCRATCH(mstate, scratch_size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		if (subr == DIF_SUBR_COPYIN) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
			dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
		}

		mstate->dtms_scratch_ptr += scratch_size;
		regs[rd] = dest;
		break;
	}

	case DIF_SUBR_COPYINTO: {
		uint64_t size = tupregs[1].dttk_value;
		uintptr_t dest = tupregs[2].dttk_value;

		/*
		 * This action doesn't require any credential checks since
		 * probes will not activate in user contexts to which the
		 * enabling user does not have permissions.
		 */
		if (!dtrace_inscratch(dest, size, mstate)) {
			*flags |= CPU_DTRACE_BADADDR;
			*illval = regs[rd];
			break;
		}

		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
		dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
		break;
	}

	case DIF_SUBR_COPYINSTR: {
		uintptr_t dest = mstate->dtms_scratch_ptr;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];

		if (nargs > 1 && tupregs[1].dttk_value < size)
			size = tupregs[1].dttk_value + 1;

		/*
		 * This action doesn't require any credential checks since
		 * probes will not activate in user contexts to which the
		 * enabling user does not have permissions.
		 */
		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
		dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);

		((char *)dest)[size - 1] = '\0';
		mstate->dtms_scratch_ptr += size;
		regs[rd] = dest;
		break;
	}

	case DIF_SUBR_MSGSIZE:
	case DIF_SUBR_MSGDSIZE: {
		uintptr_t baddr = tupregs[0].dttk_value, daddr;
		uintptr_t wptr, rptr;
		size_t count = 0;
		int cont = 0;

		while (baddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {

			if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
			    vstate)) {
				regs[rd] = NULL;
				break;
			}

			wptr = dtrace_loadptr(baddr +
			    offsetof(mblk_t, b_wptr));

			rptr = dtrace_loadptr(baddr +
			    offsetof(mblk_t, b_rptr));

			if (wptr < rptr) {
				*flags |= CPU_DTRACE_BADADDR;
				*illval = tupregs[0].dttk_value;
				break;
			}

			daddr = dtrace_loadptr(baddr +
			    offsetof(mblk_t, b_datap));

			baddr = dtrace_loadptr(baddr +
			    offsetof(mblk_t, b_cont));

			/*
			 * We want to prevent against denial-of-service here,
			 * so we're only going to search the list for
			 * dtrace_msgdsize_max mblks.
			 */
			if (cont++ > dtrace_msgdsize_max) {
				*flags |= CPU_DTRACE_ILLOP;
				break;
			}

			if (subr == DIF_SUBR_MSGDSIZE) {
				if (dtrace_load8(daddr +
				    offsetof(dblk_t, db_type)) != M_DATA)
					continue;
			}

			count += wptr - rptr;
		}

		if (!(*flags & CPU_DTRACE_FAULT))
			regs[rd] = count;

		break;
	}

	case DIF_SUBR_PROGENYOF: {
		pid_t pid = tupregs[0].dttk_value;
		proc_t *p;
		int rval = 0;

		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);

		for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
			if (p->p_pidp->pid_id == pid) {
				rval = 1;
				break;
			}
		}

		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);

		regs[rd] = rval;
		break;
	}

	case DIF_SUBR_SPECULATION:
		regs[rd] = dtrace_speculation(state);
		break;

	case DIF_SUBR_COPYOUT: {
		uintptr_t kaddr = tupregs[0].dttk_value;
		uintptr_t uaddr = tupregs[1].dttk_value;
		uint64_t size = tupregs[2].dttk_value;

		if (!dtrace_destructive_disallow &&
		    dtrace_priv_proc_control(state) &&
		    !dtrace_istoxic(kaddr, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
			dtrace_copyout(kaddr, uaddr, size, flags);
			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
		}
		break;
	}

	case DIF_SUBR_COPYOUTSTR: {
		uintptr_t kaddr = tupregs[0].dttk_value;
		uintptr_t uaddr = tupregs[1].dttk_value;
		uint64_t size = tupregs[2].dttk_value;

		if (!dtrace_destructive_disallow &&
		    dtrace_priv_proc_control(state) &&
		    !dtrace_istoxic(kaddr, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
			dtrace_copyoutstr(kaddr, uaddr, size, flags);
			DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
		}
		break;
	}

	case DIF_SUBR_STRLEN: {
		size_t sz;
		uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
		sz = dtrace_strlen((char *)addr,
		    state->dts_options[DTRACEOPT_STRSIZE]);

		if (!dtrace_canload(addr, sz + 1, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		regs[rd] = sz;

		break;
	}

	case DIF_SUBR_STRCHR:
	case DIF_SUBR_STRRCHR: {
		/*
		 * We're going to iterate over the string looking for the
		 * specified character.  We will iterate until we have reached
		 * the string length or we have found the character.  If this
		 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
		 * of the specified character instead of the first.
		 */
		uintptr_t saddr = tupregs[0].dttk_value;
		uintptr_t addr = tupregs[0].dttk_value;
		uintptr_t limit = addr + state->dts_options[DTRACEOPT_STRSIZE];
		char c, target = (char)tupregs[1].dttk_value;

		for (regs[rd] = NULL; addr < limit; addr++) {
			if ((c = dtrace_load8(addr)) == target) {
				regs[rd] = addr;

				if (subr == DIF_SUBR_STRCHR)
					break;
			}

			if (c == '\0')
				break;
		}

		if (!dtrace_canload(saddr, addr - saddr, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		break;
	}

	case DIF_SUBR_STRSTR:
	case DIF_SUBR_INDEX:
	case DIF_SUBR_RINDEX: {
		/*
		 * We're going to iterate over the string looking for the
		 * specified string.  We will iterate until we have reached
		 * the string length or we have found the string.  (Yes, this
		 * is done in the most naive way possible -- but considering
		 * that the string we're searching for is likely to be
		 * relatively short, the complexity of Rabin-Karp or similar
		 * hardly seems merited.)
		 */
		char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
		char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		size_t len = dtrace_strlen(addr, size);
		size_t sublen = dtrace_strlen(substr, size);
		char *limit = addr + len, *orig = addr;
		int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
		int inc = 1;

		regs[rd] = notfound;

		if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
		    vstate)) {
			regs[rd] = NULL;
			break;
		}

		/*
		 * strstr() and index()/rindex() have similar semantics if
		 * both strings are the empty string: strstr() returns a
		 * pointer to the (empty) string, and index() and rindex()
		 * both return index 0 (regardless of any position argument).
		 */
		if (sublen == 0 && len == 0) {
			if (subr == DIF_SUBR_STRSTR)
				regs[rd] = (uintptr_t)addr;
			else
				regs[rd] = 0;
			break;
		}

		if (subr != DIF_SUBR_STRSTR) {
			if (subr == DIF_SUBR_RINDEX) {
				limit = orig - 1;
				addr += len;
				inc = -1;
			}

			/*
			 * Both index() and rindex() take an optional position
			 * argument that denotes the starting position.
			 */
			if (nargs == 3) {
				int64_t pos = (int64_t)tupregs[2].dttk_value;

				/*
				 * If the position argument to index() is
				 * negative, Perl implicitly clamps it at
				 * zero.  This semantic is a little surprising
				 * given the special meaning of negative
				 * positions to similar Perl functions like
				 * substr(), but it appears to reflect a
				 * notion that index() can start from a
				 * negative index and increment its way up to
				 * the string.  Given this notion, Perl's
				 * rindex() is at least self-consistent in
				 * that it implicitly clamps positions greater
				 * than the string length to be the string
				 * length.  Where Perl completely loses
				 * coherence, however, is when the specified
				 * substring is the empty string ("").  In
				 * this case, even if the position is
				 * negative, rindex() returns 0 -- and even if
				 * the position is greater than the length,
				 * index() returns the string length.  These
				 * semantics violate the notion that index()
				 * should never return a value less than the
				 * specified position and that rindex() should
				 * never return a value greater than the
				 * specified position.  (One assumes that
				 * these semantics are artifacts of Perl's
				 * implementation and not the results of
				 * deliberate design -- it beggars belief that
				 * even Larry Wall could desire such oddness.)
				 * While in the abstract one would wish for
				 * consistent position semantics across
				 * substr(), index() and rindex() -- or at the
				 * very least self-consistent position
				 * semantics for index() and rindex() -- we
				 * instead opt to keep with the extant Perl
				 * semantics, in all their broken glory.  (Do
				 * we have more desire to maintain Perl's
				 * semantics than Perl does?  Probably.)
				 */
				if (subr == DIF_SUBR_RINDEX) {
					if (pos < 0) {
						if (sublen == 0)
							regs[rd] = 0;
						break;
					}

					if (pos > len)
						pos = len;
				} else {
					if (pos < 0)
						pos = 0;

					if (pos >= len) {
						if (sublen == 0)
							regs[rd] = len;
						break;
					}
				}

				addr = orig + pos;
			}
		}

		for (regs[rd] = notfound; addr != limit; addr += inc) {
			if (dtrace_strncmp(addr, substr, sublen) == 0) {
				if (subr != DIF_SUBR_STRSTR) {
					/*
					 * As D index() and rindex() are
					 * modeled on Perl (and not on awk),
					 * we return a zero-based (and not a
					 * one-based) index.  (For you Perl
					 * weenies: no, we're not going to add
					 * $[ -- and shouldn't you be at a con
					 * or something?)
					 */
					regs[rd] = (uintptr_t)(addr - orig);
					break;
				}

				ASSERT(subr == DIF_SUBR_STRSTR);
				regs[rd] = (uintptr_t)addr;
				break;
			}
		}

		break;
	}

	case DIF_SUBR_STRTOK: {
		uintptr_t addr = tupregs[0].dttk_value;
		uintptr_t tokaddr = tupregs[1].dttk_value;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		uintptr_t limit, toklimit = tokaddr + size;
		uint8_t c, tokmap[32];	 /* 256 / 8 */
		char *dest = (char *)mstate->dtms_scratch_ptr;
		int i;

		/*
		 * Check both the token buffer and (later) the input buffer,
		 * since both could be non-scratch addresses.
		 */
		if (!dtrace_strcanload(tokaddr, size, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		if (addr == NULL) {
			/*
			 * If the address specified is NULL, we use our saved
			 * strtok pointer from the mstate.  Note that this
			 * means that the saved strtok pointer is _only_
			 * valid within multiple enablings of the same probe --
			 * it behaves like an implicit clause-local variable.
			 */
			addr = mstate->dtms_strtok;
		} else {
			/*
			 * If the user-specified address is non-NULL we must
			 * access check it.  This is the only time we have
			 * a chance to do so, since this address may reside
			 * in the string table of this clause-- future calls
			 * (when we fetch addr from mstate->dtms_strtok)
			 * would fail this access check.
			 */
			if (!dtrace_strcanload(addr, size, mstate, vstate)) {
				regs[rd] = NULL;
				break;
			}
		}

		/*
		 * First, zero the token map, and then process the token
		 * string -- setting a bit in the map for every character
		 * found in the token string.
		 */
		for (i = 0; i < sizeof (tokmap); i++)
			tokmap[i] = 0;

		for (; tokaddr < toklimit; tokaddr++) {
			if ((c = dtrace_load8(tokaddr)) == '\0')
				break;

			ASSERT((c >> 3) < sizeof (tokmap));
			tokmap[c >> 3] |= (1 << (c & 0x7));
		}

		for (limit = addr + size; addr < limit; addr++) {
			/*
			 * We're looking for a character that is _not_ contained
			 * in the token string.
			 */
			if ((c = dtrace_load8(addr)) == '\0')
				break;

			if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
				break;
		}

		if (c == '\0') {
			/*
			 * We reached the end of the string without finding
			 * any character that was not in the token string.
			 * We return NULL in this case, and we set the saved
			 * address to NULL as well.
			 */
			regs[rd] = NULL;
			mstate->dtms_strtok = NULL;
			break;
		}

		/*
		 * From here on, we're copying into the destination string.
		 */
		for (i = 0; addr < limit && i < size - 1; addr++) {
			if ((c = dtrace_load8(addr)) == '\0')
				break;

			if (tokmap[c >> 3] & (1 << (c & 0x7)))
				break;

			ASSERT(i < size);
			dest[i++] = c;
		}

		ASSERT(i < size);
		dest[i] = '\0';
		regs[rd] = (uintptr_t)dest;
		mstate->dtms_scratch_ptr += size;
		mstate->dtms_strtok = addr;
		break;
	}

	case DIF_SUBR_SUBSTR: {
		uintptr_t s = tupregs[0].dttk_value;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		char *d = (char *)mstate->dtms_scratch_ptr;
		int64_t index = (int64_t)tupregs[1].dttk_value;
		int64_t remaining = (int64_t)tupregs[2].dttk_value;
		size_t len = dtrace_strlen((char *)s, size);
		int64_t i = 0;

		if (!dtrace_canload(s, len + 1, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		if (nargs <= 2)
			remaining = (int64_t)size;

		if (index < 0) {
			index += len;

			if (index < 0 && index + remaining > 0) {
				remaining += index;
				index = 0;
			}
		}

		if (index >= len || index < 0) {
			remaining = 0;
		} else if (remaining < 0) {
			remaining += len - index;
		} else if (index + remaining > size) {
			remaining = size - index;
		}

		for (i = 0; i < remaining; i++) {
			if ((d[i] = dtrace_load8(s + index + i)) == '\0')
				break;
		}

		d[i] = '\0';

		mstate->dtms_scratch_ptr += size;
		regs[rd] = (uintptr_t)d;
		break;
	}

	case DIF_SUBR_GETMAJOR:
#ifdef _LP64
		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
#else
		regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
#endif
		break;

	case DIF_SUBR_GETMINOR:
#ifdef _LP64
		regs[rd] = tupregs[0].dttk_value & MAXMIN64;
#else
		regs[rd] = tupregs[0].dttk_value & MAXMIN;
#endif
		break;

	case DIF_SUBR_DDI_PATHNAME: {
		/*
		 * This one is a galactic mess.  We are going to roughly
		 * emulate ddi_pathname(), but it's made more complicated
		 * by the fact that we (a) want to include the minor name and
		 * (b) must proceed iteratively instead of recursively.
		 */
		uintptr_t dest = mstate->dtms_scratch_ptr;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		char *start = (char *)dest, *end = start + size - 1;
		uintptr_t daddr = tupregs[0].dttk_value;
		int64_t minor = (int64_t)tupregs[1].dttk_value;
		char *s;
		int i, len, depth = 0;

		/*
		 * Due to all the pointer jumping we do and context we must
		 * rely upon, we just mandate that the user must have kernel
		 * read privileges to use this routine.
		 */
		if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
			*flags |= CPU_DTRACE_KPRIV;
			*illval = daddr;
			regs[rd] = NULL;
		}

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		*end = '\0';

		/*
		 * We want to have a name for the minor.  In order to do this,
		 * we need to walk the minor list from the devinfo.  We want
		 * to be sure that we don't infinitely walk a circular list,
		 * so we check for circularity by sending a scout pointer
		 * ahead two elements for every element that we iterate over;
		 * if the list is circular, these will ultimately point to the
		 * same element.  You may recognize this little trick as the
		 * answer to a stupid interview question -- one that always
		 * seems to be asked by those who had to have it laboriously
		 * explained to them, and who can't even concisely describe
		 * the conditions under which one would be forced to resort to
		 * this technique.  Needless to say, those conditions are
		 * found here -- and probably only here.  Is this the only use
		 * of this infamous trick in shipping, production code?  If it
		 * isn't, it probably should be...
		 */
		if (minor != -1) {
			uintptr_t maddr = dtrace_loadptr(daddr +
			    offsetof(struct dev_info, devi_minor));

			uintptr_t next = offsetof(struct ddi_minor_data, next);
			uintptr_t name = offsetof(struct ddi_minor_data,
			    d_minor) + offsetof(struct ddi_minor, name);
			uintptr_t dev = offsetof(struct ddi_minor_data,
			    d_minor) + offsetof(struct ddi_minor, dev);
			uintptr_t scout;

			if (maddr != NULL)
				scout = dtrace_loadptr(maddr + next);

			while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
				uint64_t m;
#ifdef _LP64
				m = dtrace_load64(maddr + dev) & MAXMIN64;
#else
				m = dtrace_load32(maddr + dev) & MAXMIN;
#endif
				if (m != minor) {
					maddr = dtrace_loadptr(maddr + next);

					if (scout == NULL)
						continue;

					scout = dtrace_loadptr(scout + next);

					if (scout == NULL)
						continue;

					scout = dtrace_loadptr(scout + next);

					if (scout == NULL)
						continue;

					if (scout == maddr) {
						*flags |= CPU_DTRACE_ILLOP;
						break;
					}

					continue;
				}

				/*
				 * We have the minor data.  Now we need to
				 * copy the minor's name into the end of the
				 * pathname.
				 */
				s = (char *)dtrace_loadptr(maddr + name);
				len = dtrace_strlen(s, size);

				if (*flags & CPU_DTRACE_FAULT)
					break;

				if (len != 0) {
					if ((end -= (len + 1)) < start)
						break;

					*end = ':';
				}

				for (i = 1; i <= len; i++)
					end[i] = dtrace_load8((uintptr_t)s++);
				break;
			}
		}

		while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
			ddi_node_state_t devi_state;

			devi_state = dtrace_load32(daddr +
			    offsetof(struct dev_info, devi_node_state));

			if (*flags & CPU_DTRACE_FAULT)
				break;

			if (devi_state >= DS_INITIALIZED) {
				s = (char *)dtrace_loadptr(daddr +
				    offsetof(struct dev_info, devi_addr));
				len = dtrace_strlen(s, size);

				if (*flags & CPU_DTRACE_FAULT)
					break;

				if (len != 0) {
					if ((end -= (len + 1)) < start)
						break;

					*end = '@';
				}

				for (i = 1; i <= len; i++)
					end[i] = dtrace_load8((uintptr_t)s++);
			}

			/*
			 * Now for the node name...
			 */
			s = (char *)dtrace_loadptr(daddr +
			    offsetof(struct dev_info, devi_node_name));

			daddr = dtrace_loadptr(daddr +
			    offsetof(struct dev_info, devi_parent));

			/*
			 * If our parent is NULL (that is, if we're the root
			 * node), we're going to use the special path
			 * "devices".
			 */
			if (daddr == NULL)
				s = "devices";

			len = dtrace_strlen(s, size);
			if (*flags & CPU_DTRACE_FAULT)
				break;

			if ((end -= (len + 1)) < start)
				break;

			for (i = 1; i <= len; i++)
				end[i] = dtrace_load8((uintptr_t)s++);
			*end = '/';

			if (depth++ > dtrace_devdepth_max) {
				*flags |= CPU_DTRACE_ILLOP;
				break;
			}
		}

		if (end < start)
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);

		if (daddr == NULL) {
			regs[rd] = (uintptr_t)end;
			mstate->dtms_scratch_ptr += size;
		}

		break;
	}

	case DIF_SUBR_STRJOIN: {
		char *d = (char *)mstate->dtms_scratch_ptr;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		uintptr_t s1 = tupregs[0].dttk_value;
		uintptr_t s2 = tupregs[1].dttk_value;
		int i = 0;

		if (!dtrace_strcanload(s1, size, mstate, vstate) ||
		    !dtrace_strcanload(s2, size, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		for (;;) {
			if (i >= size) {
				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
				regs[rd] = NULL;
				break;
			}

			if ((d[i++] = dtrace_load8(s1++)) == '\0') {
				i--;
				break;
			}
		}

		for (;;) {
			if (i >= size) {
				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
				regs[rd] = NULL;
				break;
			}

			if ((d[i++] = dtrace_load8(s2++)) == '\0')
				break;
		}

		if (i < size) {
			mstate->dtms_scratch_ptr += i;
			regs[rd] = (uintptr_t)d;
		}

		break;
	}

	case DIF_SUBR_LLTOSTR: {
		int64_t i = (int64_t)tupregs[0].dttk_value;
		int64_t val = i < 0 ? i * -1 : i;
		uint64_t size = 22;	/* enough room for 2^64 in decimal */
		char *end = (char *)mstate->dtms_scratch_ptr + size - 1;

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		for (*end-- = '\0'; val; val /= 10)
			*end-- = '0' + (val % 10);

		if (i == 0)
			*end-- = '0';

		if (i < 0)
			*end-- = '-';

		regs[rd] = (uintptr_t)end + 1;
		mstate->dtms_scratch_ptr += size;
		break;
	}

	case DIF_SUBR_HTONS:
	case DIF_SUBR_NTOHS:
#ifdef _BIG_ENDIAN
		regs[rd] = (uint16_t)tupregs[0].dttk_value;
#else
		regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
#endif
		break;


	case DIF_SUBR_HTONL:
	case DIF_SUBR_NTOHL:
#ifdef _BIG_ENDIAN
		regs[rd] = (uint32_t)tupregs[0].dttk_value;
#else
		regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
#endif
		break;


	case DIF_SUBR_HTONLL:
	case DIF_SUBR_NTOHLL:
#ifdef _BIG_ENDIAN
		regs[rd] = (uint64_t)tupregs[0].dttk_value;
#else
		regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
#endif
		break;


	case DIF_SUBR_DIRNAME:
	case DIF_SUBR_BASENAME: {
		char *dest = (char *)mstate->dtms_scratch_ptr;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		uintptr_t src = tupregs[0].dttk_value;
		int i, j, len = dtrace_strlen((char *)src, size);
		int lastbase = -1, firstbase = -1, lastdir = -1;
		int start, end;

		if (!dtrace_canload(src, len + 1, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		/*
		 * The basename and dirname for a zero-length string is
		 * defined to be "."
		 */
		if (len == 0) {
			len = 1;
			src = (uintptr_t)".";
		}

		/*
		 * Start from the back of the string, moving back toward the
		 * front until we see a character that isn't a slash.  That
		 * character is the last character in the basename.
		 */
		for (i = len - 1; i >= 0; i--) {
			if (dtrace_load8(src + i) != '/')
				break;
		}

		if (i >= 0)
			lastbase = i;

		/*
		 * Starting from the last character in the basename, move
		 * towards the front until we find a slash.  The character
		 * that we processed immediately before that is the first
		 * character in the basename.
		 */
		for (; i >= 0; i--) {
			if (dtrace_load8(src + i) == '/')
				break;
		}

		if (i >= 0)
			firstbase = i + 1;

		/*
		 * Now keep going until we find a non-slash character.  That
		 * character is the last character in the dirname.
		 */
		for (; i >= 0; i--) {
			if (dtrace_load8(src + i) != '/')
				break;
		}

		if (i >= 0)
			lastdir = i;

		ASSERT(!(lastbase == -1 && firstbase != -1));
		ASSERT(!(firstbase == -1 && lastdir != -1));

		if (lastbase == -1) {
			/*
			 * We didn't find a non-slash character.  We know that
			 * the length is non-zero, so the whole string must be
			 * slashes.  In either the dirname or the basename
			 * case, we return '/'.
			 */
			ASSERT(firstbase == -1);
			firstbase = lastbase = lastdir = 0;
		}

		if (firstbase == -1) {
			/*
			 * The entire string consists only of a basename
			 * component.  If we're looking for dirname, we need
			 * to change our string to be just "."; if we're
			 * looking for a basename, we'll just set the first
			 * character of the basename to be 0.
			 */
			if (subr == DIF_SUBR_DIRNAME) {
				ASSERT(lastdir == -1);
				src = (uintptr_t)".";
				lastdir = 0;
			} else {
				firstbase = 0;
			}
		}

		if (subr == DIF_SUBR_DIRNAME) {
			if (lastdir == -1) {
				/*
				 * We know that we have a slash in the name --
				 * or lastdir would be set to 0, above.  And
				 * because lastdir is -1, we know that this
				 * slash must be the first character.  (That
				 * is, the full string must be of the form
				 * "/basename".)  In this case, the last
				 * character of the directory name is 0.
				 */
				lastdir = 0;
			}

			start = 0;
			end = lastdir;
		} else {
			ASSERT(subr == DIF_SUBR_BASENAME);
			ASSERT(firstbase != -1 && lastbase != -1);
			start = firstbase;
			end = lastbase;
		}

		for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
			dest[j] = dtrace_load8(src + i);

		dest[j] = '\0';
		regs[rd] = (uintptr_t)dest;
		mstate->dtms_scratch_ptr += size;
		break;
	}

	case DIF_SUBR_CLEANPATH: {
		char *dest = (char *)mstate->dtms_scratch_ptr, c;
		uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
		uintptr_t src = tupregs[0].dttk_value;
		int i = 0, j = 0;

		if (!dtrace_strcanload(src, size, mstate, vstate)) {
			regs[rd] = NULL;
			break;
		}

		if (!DTRACE_INSCRATCH(mstate, size)) {
			DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
			regs[rd] = NULL;
			break;
		}

		/*
		 * Move forward, loading each character.
		 */
		do {
			c = dtrace_load8(src + i++);
next:
			if (j + 5 >= size)	/* 5 = strlen("/..c\0") */
				break;

			if (c != '/') {
				dest[j++] = c;
				continue;
			}

			c = dtrace_load8(src + i++);

			if (c == '/') {
				/*
				 * We have two slashes -- we can just advance
				 * to the next character.
				 */
				goto next;
			}

			if (c != '.') {
				/*
				 * This is not "." and it's not ".." -- we can
				 * just store the "/" and this character and
				 * drive on.
				 */
				dest[j++] = '/';
				dest[j++] = c;
				continue;
			}

			c = dtrace_load8(src + i++);

			if (c == '/') {
				/*
				 * This is a "/./" component.  We're not going
				 * to store anything in the destination buffer;
				 * we're just going to go to the next component.
				 */
				goto next;
			}

			if (c != '.') {
				/*
				 * This is not ".." -- we can just store the
				 * "/." and this character and continue
				 * processing.
				 */
				dest[j++] = '/';
				dest[j++] = '.';
				dest[j++] = c;
				continue;
			}

			c = dtrace_load8(src + i++);

			if (c != '/' && c != '\0') {
				/*
				 * This is not ".." -- it's "..[mumble]".
				 * We'll store the "/.." and this character
				 * and continue processing.
				 */
				dest[j++] = '/';
				dest[j++] = '.';
				dest[j++] = '.';
				dest[j++] = c;
				continue;
			}

			/*
			 * This is "/../" or "/..\0".  We need to back up
			 * our destination pointer until we find a "/".
			 */
			i--;
			while (j != 0 && dest[--j] != '/')
				continue;

			if (c == '\0')
				dest[++j] = '/';
		} while (c != '\0');

		dest[j] = '\0';
		regs[rd] = (uintptr_t)dest;
		mstate->dtms_scratch_ptr += size;
		break;
	}

	case DIF_SUBR_INET_NTOA:
	case DIF_SUBR_INET_NTOA6:
	case DIF_SUBR_INET_NTOP: {
		size_t size;
		int af, argi, i;
		char *base, *end;

		if (subr == DIF_SUBR_INET_NTOP) {
			af = (int)tupregs[0].dttk_value;
			argi = 1;
		} else {
			af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
			argi = 0;
		}

		if (af == AF_INET) {
			ipaddr_t ip4;
			uint8_t *ptr8, val;

			/*
			 * Safely load the IPv4 address.
			 */
			ip4 = dtrace_load32(tupregs[argi].dttk_value);

			/*
			 * Check an IPv4 string will fit in scratch.
			 */
			size = INET_ADDRSTRLEN;
			if (!DTRACE_INSCRATCH(mstate, size)) {
				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
				regs[rd] = NULL;
				break;
			}
			base = (char *)mstate->dtms_scratch_ptr;
			end = (char *)mstate->dtms_scratch_ptr + size - 1;

			/*
			 * Stringify as a dotted decimal quad.
			 */
			*end-- = '\0';
			ptr8 = (uint8_t *)&ip4;
			for (i = 3; i >= 0; i--) {
				val = ptr8[i];

				if (val == 0) {
					*end-- = '0';
				} else {
					for (; val; val /= 10) {
						*end-- = '0' + (val % 10);
					}
				}

				if (i > 0)
					*end-- = '.';
			}
			ASSERT(end + 1 >= base);

		} else if (af == AF_INET6) {
			struct in6_addr ip6;
			int firstzero, tryzero, numzero, v6end;
			uint16_t val;
			const char digits[] = "0123456789abcdef";

			/*
			 * Stringify using RFC 1884 convention 2 - 16 bit
			 * hexadecimal values with a zero-run compression.
			 * Lower case hexadecimal digits are used.
			 * 	eg, fe80::214:4fff:fe0b:76c8.
			 * The IPv4 embedded form is returned for inet_ntop,
			 * just the IPv4 string is returned for inet_ntoa6.
			 */

			/*
			 * Safely load the IPv6 address.
			 */
			dtrace_bcopy(
			    (void *)(uintptr_t)tupregs[argi].dttk_value,
			    (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));

			/*
			 * Check an IPv6 string will fit in scratch.
			 */
			size = INET6_ADDRSTRLEN;
			if (!DTRACE_INSCRATCH(mstate, size)) {
				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
				regs[rd] = NULL;
				break;
			}
			base = (char *)mstate->dtms_scratch_ptr;
			end = (char *)mstate->dtms_scratch_ptr + size - 1;
			*end-- = '\0';

			/*
			 * Find the longest run of 16 bit zero values
			 * for the single allowed zero compression - "::".
			 */
			firstzero = -1;
			tryzero = -1;
			numzero = 1;
			for (i = 0; i < sizeof (struct in6_addr); i++) {
				if (ip6._S6_un._S6_u8[i] == 0 &&
				    tryzero == -1 && i % 2 == 0) {
					tryzero = i;
					continue;
				}

				if (tryzero != -1 &&
				    (ip6._S6_un._S6_u8[i] != 0 ||
				    i == sizeof (struct in6_addr) - 1)) {

					if (i - tryzero <= numzero) {
						tryzero = -1;
						continue;
					}

					firstzero = tryzero;
					numzero = i - i % 2 - tryzero;
					tryzero = -1;

					if (ip6._S6_un._S6_u8[i] == 0 &&
					    i == sizeof (struct in6_addr) - 1)
						numzero += 2;
				}
			}
			ASSERT(firstzero + numzero <= sizeof (struct in6_addr));

			/*
			 * Check for an IPv4 embedded address.
			 */
			v6end = sizeof (struct in6_addr) - 2;
			if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
			    IN6_IS_ADDR_V4COMPAT(&ip6)) {
				for (i = sizeof (struct in6_addr) - 1;
				    i >= DTRACE_V4MAPPED_OFFSET; i--) {
					ASSERT(end >= base);

					val = ip6._S6_un._S6_u8[i];

					if (val == 0) {
						*end-- = '0';
					} else {
						for (; val; val /= 10) {
							*end-- = '0' + val % 10;
						}
					}

					if (i > DTRACE_V4MAPPED_OFFSET)
						*end-- = '.';
				}

				if (subr == DIF_SUBR_INET_NTOA6)
					goto inetout;

				/*
				 * Set v6end to skip the IPv4 address that
				 * we have already stringified.
				 */
				v6end = 10;
			}

			/*
			 * Build the IPv6 string by working through the
			 * address in reverse.
			 */
			for (i = v6end; i >= 0; i -= 2) {
				ASSERT(end >= base);

				if (i == firstzero + numzero - 2) {
					*end-- = ':';
					*end-- = ':';
					i -= numzero - 2;
					continue;
				}

				if (i < 14 && i != firstzero - 2)
					*end-- = ':';

				val = (ip6._S6_un._S6_u8[i] << 8) +
				    ip6._S6_un._S6_u8[i + 1];

				if (val == 0) {
					*end-- = '0';
				} else {
					for (; val; val /= 16) {
						*end-- = digits[val % 16];
					}
				}
			}
			ASSERT(end + 1 >= base);

		} else {
			/*
			 * The user didn't use AH_INET or AH_INET6.
			 */
			DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
			regs[rd] = NULL;
			break;
		}

inetout:	regs[rd] = (uintptr_t)end + 1;
		mstate->dtms_scratch_ptr += size;
		break;
	}

	}
}

/*
 * Emulate the execution of DTrace IR instructions specified by the given
 * DIF object.  This function is deliberately void of assertions as all of
 * the necessary checks are handled by a call to dtrace_difo_validate().
 */
static uint64_t
dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
    dtrace_vstate_t *vstate, dtrace_state_t *state)
{
	const dif_instr_t *text = difo->dtdo_buf;
	const uint_t textlen = difo->dtdo_len;
	const char *strtab = difo->dtdo_strtab;
	const uint64_t *inttab = difo->dtdo_inttab;

	uint64_t rval = 0;
	dtrace_statvar_t *svar;
	dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
	dtrace_difv_t *v;
	volatile uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
	volatile uintptr_t *illval = &cpu_core[CPU->cpu_id].cpuc_dtrace_illval;

	dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
	uint64_t regs[DIF_DIR_NREGS];
	uint64_t *tmp;

	uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
	int64_t cc_r;
	uint_t pc = 0, id, opc;
	uint8_t ttop = 0;
	dif_instr_t instr;
	uint_t r1, r2, rd;

	/*
	 * We stash the current DIF object into the machine state: we need it
	 * for subsequent access checking.
	 */
	mstate->dtms_difo = difo;

	regs[DIF_REG_R0] = 0; 		/* %r0 is fixed at zero */

	while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
		opc = pc;

		instr = text[pc++];
		r1 = DIF_INSTR_R1(instr);
		r2 = DIF_INSTR_R2(instr);
		rd = DIF_INSTR_RD(instr);

		switch (DIF_INSTR_OP(instr)) {
		case DIF_OP_OR:
			regs[rd] = regs[r1] | regs[r2];
			break;
		case DIF_OP_XOR:
			regs[rd] = regs[r1] ^ regs[r2];
			break;
		case DIF_OP_AND:
			regs[rd] = regs[r1] & regs[r2];
			break;
		case DIF_OP_SLL:
			regs[rd] = regs[r1] << regs[r2];
			break;
		case DIF_OP_SRL:
			regs[rd] = regs[r1] >> regs[r2];
			break;
		case DIF_OP_SUB:
			regs[rd] = regs[r1] - regs[r2];
			break;
		case DIF_OP_ADD:
			regs[rd] = regs[r1] + regs[r2];
			break;
		case DIF_OP_MUL:
			regs[rd] = regs[r1] * regs[r2];
			break;
		case DIF_OP_SDIV:
			if (regs[r2] == 0) {
				regs[rd] = 0;
				*flags |= CPU_DTRACE_DIVZERO;
			} else {
				regs[rd] = (int64_t)regs[r1] /
				    (int64_t)regs[r2];
			}
			break;

		case DIF_OP_UDIV:
			if (regs[r2] == 0) {
				regs[rd] = 0;
				*flags |= CPU_DTRACE_DIVZERO;
			} else {
				regs[rd] = regs[r1] / regs[r2];
			}
			break;

		case DIF_OP_SREM:
			if (regs[r2] == 0) {
				regs[rd] = 0;
				*flags |= CPU_DTRACE_DIVZERO;
			} else {
				regs[rd] = (int64_t)regs[r1] %
				    (int64_t)regs[r2];
			}
			break;

		case DIF_OP_UREM:
			if (regs[r2] == 0) {
				regs[rd] = 0;
				*flags |= CPU_DTRACE_DIVZERO;
			} else {
				regs[rd] = regs[r1] % regs[r2];
			}
			break;

		case DIF_OP_NOT:
			regs[rd] = ~regs[r1];
			break;
		case DIF_OP_MOV:
			regs[rd] = regs[r1];
			break;
		case DIF_OP_CMP:
			cc_r = regs[r1] - regs[r2];
			cc_n = cc_r < 0;
			cc_z = cc_r == 0;
			cc_v = 0;
			cc_c = regs[r1] < regs[r2];
			break;
		case DIF_OP_TST:
			cc_n = cc_v = cc_c = 0;
			cc_z = regs[r1] == 0;
			break;
		case DIF_OP_BA:
			pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BE:
			if (cc_z)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BNE:
			if (cc_z == 0)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BG:
			if ((cc_z | (cc_n ^ cc_v)) == 0)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BGU:
			if ((cc_c | cc_z) == 0)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BGE:
			if ((cc_n ^ cc_v) == 0)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BGEU:
			if (cc_c == 0)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BL:
			if (cc_n ^ cc_v)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BLU:
			if (cc_c)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BLE:
			if (cc_z | (cc_n ^ cc_v))
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_BLEU:
			if (cc_c | cc_z)
				pc = DIF_INSTR_LABEL(instr);
			break;
		case DIF_OP_RLDSB:
			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDSB:
			regs[rd] = (int8_t)dtrace_load8(regs[r1]);
			break;
		case DIF_OP_RLDSH:
			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDSH:
			regs[rd] = (int16_t)dtrace_load16(regs[r1]);
			break;
		case DIF_OP_RLDSW:
			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDSW:
			regs[rd] = (int32_t)dtrace_load32(regs[r1]);
			break;
		case DIF_OP_RLDUB:
			if (!dtrace_canstore(regs[r1], 1, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDUB:
			regs[rd] = dtrace_load8(regs[r1]);
			break;
		case DIF_OP_RLDUH:
			if (!dtrace_canstore(regs[r1], 2, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDUH:
			regs[rd] = dtrace_load16(regs[r1]);
			break;
		case DIF_OP_RLDUW:
			if (!dtrace_canstore(regs[r1], 4, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDUW:
			regs[rd] = dtrace_load32(regs[r1]);
			break;
		case DIF_OP_RLDX:
			if (!dtrace_canstore(regs[r1], 8, mstate, vstate)) {
				*flags |= CPU_DTRACE_KPRIV;
				*illval = regs[r1];
				break;
			}
			/*FALLTHROUGH*/
		case DIF_OP_LDX:
			regs[rd] = dtrace_load64(regs[r1]);
			break;
		case DIF_OP_ULDSB:
			regs[rd] = (int8_t)
			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_ULDSH:
			regs[rd] = (int16_t)
			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_ULDSW:
			regs[rd] = (int32_t)
			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_ULDUB:
			regs[rd] =
			    dtrace_fuword8((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_ULDUH:
			regs[rd] =
			    dtrace_fuword16((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_ULDUW:
			regs[rd] =
			    dtrace_fuword32((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_ULDX:
			regs[rd] =
			    dtrace_fuword64((void *)(uintptr_t)regs[r1]);
			break;
		case DIF_OP_RET:
			rval = regs[rd];
			pc = textlen;
			break;
		case DIF_OP_NOP:
			break;
		case DIF_OP_SETX:
			regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
			break;
		case DIF_OP_SETS:
			regs[rd] = (uint64_t)(uintptr_t)
			    (strtab + DIF_INSTR_STRING(instr));
			break;
		case DIF_OP_SCMP: {
			size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
			uintptr_t s1 = regs[r1];
			uintptr_t s2 = regs[r2];

			if (s1 != NULL &&
			    !dtrace_strcanload(s1, sz, mstate, vstate))
				break;
			if (s2 != NULL &&
			    !dtrace_strcanload(s2, sz, mstate, vstate))
				break;

			cc_r = dtrace_strncmp((char *)s1, (char *)s2, sz);

			cc_n = cc_r < 0;
			cc_z = cc_r == 0;
			cc_v = cc_c = 0;
			break;
		}
		case DIF_OP_LDGA:
			regs[rd] = dtrace_dif_variable(mstate, state,
			    r1, regs[r2]);
			break;
		case DIF_OP_LDGS:
			id = DIF_INSTR_VAR(instr);

			if (id >= DIF_VAR_OTHER_UBASE) {
				uintptr_t a;

				id -= DIF_VAR_OTHER_UBASE;
				svar = vstate->dtvs_globals[id];
				ASSERT(svar != NULL);
				v = &svar->dtsv_var;

				if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
					regs[rd] = svar->dtsv_data;
					break;
				}

				a = (uintptr_t)svar->dtsv_data;

				if (*(uint8_t *)a == UINT8_MAX) {
					/*
					 * If the 0th byte is set to UINT8_MAX
					 * then this is to be treated as a
					 * reference to a NULL variable.
					 */
					regs[rd] = NULL;
				} else {
					regs[rd] = a + sizeof (uint64_t);
				}

				break;
			}

			regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
			break;

		case DIF_OP_STGS:
			id = DIF_INSTR_VAR(instr);

			ASSERT(id >= DIF_VAR_OTHER_UBASE);
			id -= DIF_VAR_OTHER_UBASE;

			svar = vstate->dtvs_globals[id];
			ASSERT(svar != NULL);
			v = &svar->dtsv_var;

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				uintptr_t a = (uintptr_t)svar->dtsv_data;

				ASSERT(a != NULL);
				ASSERT(svar->dtsv_size != 0);

				if (regs[rd] == NULL) {
					*(uint8_t *)a = UINT8_MAX;
					break;
				} else {
					*(uint8_t *)a = 0;
					a += sizeof (uint64_t);
				}
				if (!dtrace_vcanload(
				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
				    mstate, vstate))
					break;

				dtrace_vcopy((void *)(uintptr_t)regs[rd],
				    (void *)a, &v->dtdv_type);
				break;
			}

			svar->dtsv_data = regs[rd];
			break;

		case DIF_OP_LDTA:
			/*
			 * There are no DTrace built-in thread-local arrays at
			 * present.  This opcode is saved for future work.
			 */
			*flags |= CPU_DTRACE_ILLOP;
			regs[rd] = 0;
			break;

		case DIF_OP_LDLS:
			id = DIF_INSTR_VAR(instr);

			if (id < DIF_VAR_OTHER_UBASE) {
				/*
				 * For now, this has no meaning.
				 */
				regs[rd] = 0;
				break;
			}

			id -= DIF_VAR_OTHER_UBASE;

			ASSERT(id < vstate->dtvs_nlocals);
			ASSERT(vstate->dtvs_locals != NULL);

			svar = vstate->dtvs_locals[id];
			ASSERT(svar != NULL);
			v = &svar->dtsv_var;

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				uintptr_t a = (uintptr_t)svar->dtsv_data;
				size_t sz = v->dtdv_type.dtdt_size;

				sz += sizeof (uint64_t);
				ASSERT(svar->dtsv_size == NCPU * sz);
				a += CPU->cpu_id * sz;

				if (*(uint8_t *)a == UINT8_MAX) {
					/*
					 * If the 0th byte is set to UINT8_MAX
					 * then this is to be treated as a
					 * reference to a NULL variable.
					 */
					regs[rd] = NULL;
				} else {
					regs[rd] = a + sizeof (uint64_t);
				}

				break;
			}

			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
			regs[rd] = tmp[CPU->cpu_id];
			break;

		case DIF_OP_STLS:
			id = DIF_INSTR_VAR(instr);

			ASSERT(id >= DIF_VAR_OTHER_UBASE);
			id -= DIF_VAR_OTHER_UBASE;
			ASSERT(id < vstate->dtvs_nlocals);

			ASSERT(vstate->dtvs_locals != NULL);
			svar = vstate->dtvs_locals[id];
			ASSERT(svar != NULL);
			v = &svar->dtsv_var;

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				uintptr_t a = (uintptr_t)svar->dtsv_data;
				size_t sz = v->dtdv_type.dtdt_size;

				sz += sizeof (uint64_t);
				ASSERT(svar->dtsv_size == NCPU * sz);
				a += CPU->cpu_id * sz;

				if (regs[rd] == NULL) {
					*(uint8_t *)a = UINT8_MAX;
					break;
				} else {
					*(uint8_t *)a = 0;
					a += sizeof (uint64_t);
				}

				if (!dtrace_vcanload(
				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
				    mstate, vstate))
					break;

				dtrace_vcopy((void *)(uintptr_t)regs[rd],
				    (void *)a, &v->dtdv_type);
				break;
			}

			ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
			tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
			tmp[CPU->cpu_id] = regs[rd];
			break;

		case DIF_OP_LDTS: {
			dtrace_dynvar_t *dvar;
			dtrace_key_t *key;

			id = DIF_INSTR_VAR(instr);
			ASSERT(id >= DIF_VAR_OTHER_UBASE);
			id -= DIF_VAR_OTHER_UBASE;
			v = &vstate->dtvs_tlocals[id];

			key = &tupregs[DIF_DTR_NREGS];
			key[0].dttk_value = (uint64_t)id;
			key[0].dttk_size = 0;
			DTRACE_TLS_THRKEY(key[1].dttk_value);
			key[1].dttk_size = 0;

			dvar = dtrace_dynvar(dstate, 2, key,
			    sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
			    mstate, vstate);

			if (dvar == NULL) {
				regs[rd] = 0;
				break;
			}

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
			} else {
				regs[rd] = *((uint64_t *)dvar->dtdv_data);
			}

			break;
		}

		case DIF_OP_STTS: {
			dtrace_dynvar_t *dvar;
			dtrace_key_t *key;

			id = DIF_INSTR_VAR(instr);
			ASSERT(id >= DIF_VAR_OTHER_UBASE);
			id -= DIF_VAR_OTHER_UBASE;

			key = &tupregs[DIF_DTR_NREGS];
			key[0].dttk_value = (uint64_t)id;
			key[0].dttk_size = 0;
			DTRACE_TLS_THRKEY(key[1].dttk_value);
			key[1].dttk_size = 0;
			v = &vstate->dtvs_tlocals[id];

			dvar = dtrace_dynvar(dstate, 2, key,
			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);

			/*
			 * Given that we're storing to thread-local data,
			 * we need to flush our predicate cache.
			 */
			curthread->t_predcache = NULL;

			if (dvar == NULL)
				break;

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				if (!dtrace_vcanload(
				    (void *)(uintptr_t)regs[rd],
				    &v->dtdv_type, mstate, vstate))
					break;

				dtrace_vcopy((void *)(uintptr_t)regs[rd],
				    dvar->dtdv_data, &v->dtdv_type);
			} else {
				*((uint64_t *)dvar->dtdv_data) = regs[rd];
			}

			break;
		}

		case DIF_OP_SRA:
			regs[rd] = (int64_t)regs[r1] >> regs[r2];
			break;

		case DIF_OP_CALL:
			dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
			    regs, tupregs, ttop, mstate, state);
			break;

		case DIF_OP_PUSHTR:
			if (ttop == DIF_DTR_NREGS) {
				*flags |= CPU_DTRACE_TUPOFLOW;
				break;
			}

			if (r1 == DIF_TYPE_STRING) {
				/*
				 * If this is a string type and the size is 0,
				 * we'll use the system-wide default string
				 * size.  Note that we are _not_ looking at
				 * the value of the DTRACEOPT_STRSIZE option;
				 * had this been set, we would expect to have
				 * a non-zero size value in the "pushtr".
				 */
				tupregs[ttop].dttk_size =
				    dtrace_strlen((char *)(uintptr_t)regs[rd],
				    regs[r2] ? regs[r2] :
				    dtrace_strsize_default) + 1;
			} else {
				tupregs[ttop].dttk_size = regs[r2];
			}

			tupregs[ttop++].dttk_value = regs[rd];
			break;

		case DIF_OP_PUSHTV:
			if (ttop == DIF_DTR_NREGS) {
				*flags |= CPU_DTRACE_TUPOFLOW;
				break;
			}

			tupregs[ttop].dttk_value = regs[rd];
			tupregs[ttop++].dttk_size = 0;
			break;

		case DIF_OP_POPTS:
			if (ttop != 0)
				ttop--;
			break;

		case DIF_OP_FLUSHTS:
			ttop = 0;
			break;

		case DIF_OP_LDGAA:
		case DIF_OP_LDTAA: {
			dtrace_dynvar_t *dvar;
			dtrace_key_t *key = tupregs;
			uint_t nkeys = ttop;

			id = DIF_INSTR_VAR(instr);
			ASSERT(id >= DIF_VAR_OTHER_UBASE);
			id -= DIF_VAR_OTHER_UBASE;

			key[nkeys].dttk_value = (uint64_t)id;
			key[nkeys++].dttk_size = 0;

			if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
				key[nkeys++].dttk_size = 0;
				v = &vstate->dtvs_tlocals[id];
			} else {
				v = &vstate->dtvs_globals[id]->dtsv_var;
			}

			dvar = dtrace_dynvar(dstate, nkeys, key,
			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
			    DTRACE_DYNVAR_NOALLOC, mstate, vstate);

			if (dvar == NULL) {
				regs[rd] = 0;
				break;
			}

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
			} else {
				regs[rd] = *((uint64_t *)dvar->dtdv_data);
			}

			break;
		}

		case DIF_OP_STGAA:
		case DIF_OP_STTAA: {
			dtrace_dynvar_t *dvar;
			dtrace_key_t *key = tupregs;
			uint_t nkeys = ttop;

			id = DIF_INSTR_VAR(instr);
			ASSERT(id >= DIF_VAR_OTHER_UBASE);
			id -= DIF_VAR_OTHER_UBASE;

			key[nkeys].dttk_value = (uint64_t)id;
			key[nkeys++].dttk_size = 0;

			if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
				DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
				key[nkeys++].dttk_size = 0;
				v = &vstate->dtvs_tlocals[id];
			} else {
				v = &vstate->dtvs_globals[id]->dtsv_var;
			}

			dvar = dtrace_dynvar(dstate, nkeys, key,
			    v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
			    v->dtdv_type.dtdt_size : sizeof (uint64_t),
			    regs[rd] ? DTRACE_DYNVAR_ALLOC :
			    DTRACE_DYNVAR_DEALLOC, mstate, vstate);

			if (dvar == NULL)
				break;

			if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
				if (!dtrace_vcanload(
				    (void *)(uintptr_t)regs[rd], &v->dtdv_type,
				    mstate, vstate))
					break;

				dtrace_vcopy((void *)(uintptr_t)regs[rd],
				    dvar->dtdv_data, &v->dtdv_type);
			} else {
				*((uint64_t *)dvar->dtdv_data) = regs[rd];
			}

			break;
		}

		case DIF_OP_ALLOCS: {
			uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
			size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];

			/*
			 * Rounding up the user allocation size could have
			 * overflowed large, bogus allocations (like -1ULL) to
			 * 0.
			 */
			if (size < regs[r1] ||
			    !DTRACE_INSCRATCH(mstate, size)) {
				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
				regs[rd] = NULL;
				break;
			}

			dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
			mstate->dtms_scratch_ptr += size;
			regs[rd] = ptr;
			break;
		}

		case DIF_OP_COPYS:
			if (!dtrace_canstore(regs[rd], regs[r2],
			    mstate, vstate)) {
				*flags |= CPU_DTRACE_BADADDR;
				*illval = regs[rd];
				break;
			}

			if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
				break;

			dtrace_bcopy((void *)(uintptr_t)regs[r1],
			    (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
			break;

		case DIF_OP_STB:
			if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
				*flags |= CPU_DTRACE_BADADDR;
				*illval = regs[rd];
				break;
			}
			*((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
			break;

		case DIF_OP_STH:
			if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
				*flags |= CPU_DTRACE_BADADDR;
				*illval = regs[rd];
				break;
			}
			if (regs[rd] & 1) {
				*flags |= CPU_DTRACE_BADALIGN;
				*illval = regs[rd];
				break;
			}
			*((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
			break;

		case DIF_OP_STW:
			if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
				*flags |= CPU_DTRACE_BADADDR;
				*illval = regs[rd];
				break;
			}
			if (regs[rd] & 3) {
				*flags |= CPU_DTRACE_BADALIGN;
				*illval = regs[rd];
				break;
			}
			*((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
			break;

		case DIF_OP_STX:
			if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
				*flags |= CPU_DTRACE_BADADDR;
				*illval = regs[rd];
				break;
			}
			if (regs[rd] & 7) {
				*flags |= CPU_DTRACE_BADALIGN;
				*illval = regs[rd];
				break;
			}
			*((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
			break;
		}
	}

	if (!(*flags & CPU_DTRACE_FAULT))
		return (rval);

	mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
	mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;

	return (0);
}

static void
dtrace_action_breakpoint(dtrace_ecb_t *ecb)
{
	dtrace_probe_t *probe = ecb->dte_probe;
	dtrace_provider_t *prov = probe->dtpr_provider;
	char c[DTRACE_FULLNAMELEN + 80], *str;
	char *msg = "dtrace: breakpoint action at probe ";
	char *ecbmsg = " (ecb ";
	uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
	uintptr_t val = (uintptr_t)ecb;
	int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;

	if (dtrace_destructive_disallow)
		return;

	/*
	 * It's impossible to be taking action on the NULL probe.
	 */
	ASSERT(probe != NULL);

	/*
	 * This is a poor man's (destitute man's?) sprintf():  we want to
	 * print the provider name, module name, function name and name of
	 * the probe, along with the hex address of the ECB with the breakpoint
	 * action -- all of which we must place in the character buffer by
	 * hand.
	 */
	while (*msg != '\0')
		c[i++] = *msg++;

	for (str = prov->dtpv_name; *str != '\0'; str++)
		c[i++] = *str;
	c[i++] = ':';

	for (str = probe->dtpr_mod; *str != '\0'; str++)
		c[i++] = *str;
	c[i++] = ':';

	for (str = probe->dtpr_func; *str != '\0'; str++)
		c[i++] = *str;
	c[i++] = ':';

	for (str = probe->dtpr_name; *str != '\0'; str++)
		c[i++] = *str;

	while (*ecbmsg != '\0')
		c[i++] = *ecbmsg++;

	while (shift >= 0) {
		mask = (uintptr_t)0xf << shift;

		if (val >= ((uintptr_t)1 << shift))
			c[i++] = "0123456789abcdef"[(val & mask) >> shift];
		shift -= 4;
	}

	c[i++] = ')';
	c[i] = '\0';

	debug_enter(c);
}

static void
dtrace_action_panic(dtrace_ecb_t *ecb)
{
	dtrace_probe_t *probe = ecb->dte_probe;

	/*
	 * It's impossible to be taking action on the NULL probe.
	 */
	ASSERT(probe != NULL);

	if (dtrace_destructive_disallow)
		return;

	if (dtrace_panicked != NULL)
		return;

	if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
		return;

	/*
	 * We won the right to panic.  (We want to be sure that only one
	 * thread calls panic() from dtrace_probe(), and that panic() is
	 * called exactly once.)
	 */
	dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
	    probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
	    probe->dtpr_func, probe->dtpr_name, (void *)ecb);
}

static void
dtrace_action_raise(uint64_t sig)
{
	if (dtrace_destructive_disallow)
		return;

	if (sig >= NSIG) {
		DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
		return;
	}

	/*
	 * raise() has a queue depth of 1 -- we ignore all subsequent
	 * invocations of the raise() action.
	 */
	if (curthread->t_dtrace_sig == 0)
		curthread->t_dtrace_sig = (uint8_t)sig;

	curthread->t_sig_check = 1;
	aston(curthread);
}

static void
dtrace_action_stop(void)
{
	if (dtrace_destructive_disallow)
		return;

	if (!curthread->t_dtrace_stop) {
		curthread->t_dtrace_stop = 1;
		curthread->t_sig_check = 1;
		aston(curthread);
	}
}

static void
dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
{
	hrtime_t now;
	volatile uint16_t *flags;
	cpu_t *cpu = CPU;

	if (dtrace_destructive_disallow)
		return;

	flags = (volatile uint16_t *)&cpu_core[cpu->cpu_id].cpuc_dtrace_flags;

	now = dtrace_gethrtime();

	if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
		/*
		 * We need to advance the mark to the current time.
		 */
		cpu->cpu_dtrace_chillmark = now;
		cpu->cpu_dtrace_chilled = 0;
	}

	/*
	 * Now check to see if the requested chill time would take us over
	 * the maximum amount of time allowed in the chill interval.  (Or
	 * worse, if the calculation itself induces overflow.)
	 */
	if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
	    cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
		*flags |= CPU_DTRACE_ILLOP;
		return;
	}

	while (dtrace_gethrtime() - now < val)
		continue;

	/*
	 * Normally, we assure that the value of the variable "timestamp" does
	 * not change within an ECB.  The presence of chill() represents an
	 * exception to this rule, however.
	 */
	mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
	cpu->cpu_dtrace_chilled += val;
}

static void
dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
    uint64_t *buf, uint64_t arg)
{
	int nframes = DTRACE_USTACK_NFRAMES(arg);
	int strsize = DTRACE_USTACK_STRSIZE(arg);
	uint64_t *pcs = &buf[1], *fps;
	char *str = (char *)&pcs[nframes];
	int size, offs = 0, i, j;
	uintptr_t old = mstate->dtms_scratch_ptr, saved;
	uint16_t *flags = &cpu_core[CPU->cpu_id].cpuc_dtrace_flags;
	char *sym;

	/*
	 * Should be taking a faster path if string space has not been
	 * allocated.
	 */
	ASSERT(strsize != 0);

	/*
	 * We will first allocate some temporary space for the frame pointers.
	 */
	fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
	size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
	    (nframes * sizeof (uint64_t));

	if (!DTRACE_INSCRATCH(mstate, size)) {
		/*
		 * Not enough room for our frame pointers -- need to indicate
		 * that we ran out of scratch space.
		 */
		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
		return;
	}

	mstate->dtms_scratch_ptr += size;
	saved = mstate->dtms_scratch_ptr;

	/*
	 * Now get a stack with both program counters and frame pointers.
	 */
	DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
	dtrace_getufpstack(buf, fps, nframes + 1);
	DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);

	/*
	 * If that faulted, we're cooked.
	 */
	if (*flags & CPU_DTRACE_FAULT)
		goto out;

	/*
	 * Now we want to walk up the stack, calling the USTACK helper.  For
	 * each iteration, we restore the scratch pointer.
	 */
	for (i = 0; i < nframes; i++) {
		mstate->dtms_scratch_ptr = saved;

		if (offs >= strsize)
			break;

		sym = (char *)(uintptr_t)dtrace_helper(
		    DTRACE_HELPER_ACTION_USTACK,
		    mstate, state, pcs[i], fps[i]);

		/*
		 * If we faulted while running the helper, we're going to
		 * clear the fault and null out the corresponding string.
		 */
		if (*flags & CPU_DTRACE_FAULT) {
			*flags &= ~CPU_DTRACE_FAULT;
			str[offs++] = '\0';
			continue;
		}

		if (sym == NULL) {
			str[offs++] = '\0';
			continue;
		}

		DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);

		/*
		 * Now copy in the string that the helper returned to us.
		 */
		for (j = 0; offs + j < strsize; j++) {
			if ((str[offs + j] = sym[j]) == '\0')
				break;
		}

		DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);

		offs += j + 1;
	}

	if (offs >= strsize) {
		/*
		 * If we didn't have room for all of the strings, we don't
		 * abort processing -- this needn't be a fatal error -- but we
		 * still want to increment a counter (dts_stkstroverflows) to
		 * allow this condition to be warned about.  (If this is from
		 * a jstack() action, it is easily tuned via jstackstrsize.)
		 */
		dtrace_error(&state->dts_stkstroverflows);
	}

	while (offs < strsize)
		str[offs++] = '\0';

out:
	mstate->dtms_scratch_ptr = old;
}

/*
 * If you're looking for the epicenter of DTrace, you just found it.  This
 * is the function called by the provider to fire a probe -- from which all
 * subsequent probe-context DTrace activity emanates.
 */
void
dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
    uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
{
	processorid_t cpuid;
	dtrace_icookie_t cookie;
	dtrace_probe_t *probe;
	dtrace_mstate_t mstate;
	dtrace_ecb_t *ecb;
	dtrace_action_t *act;
	intptr_t offs;
	size_t size;
	int vtime, onintr;
	volatile uint16_t *flags;
	hrtime_t now;

	/*
	 * Kick out immediately if this CPU is still being born (in which case
	 * curthread will be set to -1) or the current thread can't allow
	 * probes in its current context.
	 */
	if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
		return;

	cookie = dtrace_interrupt_disable();
	probe = dtrace_probes[id - 1];
	cpuid = CPU->cpu_id;
	onintr = CPU_ON_INTR(CPU);

	if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
	    probe->dtpr_predcache == curthread->t_predcache) {
		/*
		 * We have hit in the predicate cache; we know that
		 * this predicate would evaluate to be false.
		 */
		dtrace_interrupt_enable(cookie);
		return;
	}

	if (panic_quiesce) {
		/*
		 * We don't trace anything if we're panicking.
		 */
		dtrace_interrupt_enable(cookie);
		return;
	}

	now = dtrace_gethrtime();
	vtime = dtrace_vtime_references != 0;

	if (vtime && curthread->t_dtrace_start)
		curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;

	mstate.dtms_difo = NULL;
	mstate.dtms_probe = probe;
	mstate.dtms_strtok = NULL;
	mstate.dtms_arg[0] = arg0;
	mstate.dtms_arg[1] = arg1;
	mstate.dtms_arg[2] = arg2;
	mstate.dtms_arg[3] = arg3;
	mstate.dtms_arg[4] = arg4;

	flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;

	for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
		dtrace_predicate_t *pred = ecb->dte_predicate;
		dtrace_state_t *state = ecb->dte_state;
		dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
		dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
		dtrace_vstate_t *vstate = &state->dts_vstate;
		dtrace_provider_t *prov = probe->dtpr_provider;
		int committed = 0;
		caddr_t tomax;

		/*
		 * A little subtlety with the following (seemingly innocuous)
		 * declaration of the automatic 'val':  by looking at the
		 * code, you might think that it could be declared in the
		 * action processing loop, below.  (That is, it's only used in
		 * the action processing loop.)  However, it must be declared
		 * out of that scope because in the case of DIF expression
		 * arguments to aggregating actions, one iteration of the
		 * action loop will use the last iteration's value.
		 */
#ifdef lint
		uint64_t val = 0;
#else
		uint64_t val;
#endif

		mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
		*flags &= ~CPU_DTRACE_ERROR;

		if (prov == dtrace_provider) {
			/*
			 * If dtrace itself is the provider of this probe,
			 * we're only going to continue processing the ECB if
			 * arg0 (the dtrace_state_t) is equal to the ECB's
			 * creating state.  (This prevents disjoint consumers
			 * from seeing one another's metaprobes.)
			 */
			if (arg0 != (uint64_t)(uintptr_t)state)
				continue;
		}

		if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
			/*
			 * We're not currently active.  If our provider isn't
			 * the dtrace pseudo provider, we're not interested.
			 */
			if (prov != dtrace_provider)
				continue;

			/*
			 * Now we must further check if we are in the BEGIN
			 * probe.  If we are, we will only continue processing
			 * if we're still in WARMUP -- if one BEGIN enabling
			 * has invoked the exit() action, we don't want to
			 * evaluate subsequent BEGIN enablings.
			 */
			if (probe->dtpr_id == dtrace_probeid_begin &&
			    state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
				ASSERT(state->dts_activity ==
				    DTRACE_ACTIVITY_DRAINING);
				continue;
			}
		}

		if (ecb->dte_cond) {
			/*
			 * If the dte_cond bits indicate that this
			 * consumer is only allowed to see user-mode firings
			 * of this probe, call the provider's dtps_usermode()
			 * entry point to check that the probe was fired
			 * while in a user context. Skip this ECB if that's
			 * not the case.
			 */
			if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
			    prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
			    probe->dtpr_id, probe->dtpr_arg) == 0)
				continue;

			/*
			 * This is more subtle than it looks. We have to be
			 * absolutely certain that CRED() isn't going to
			 * change out from under us so it's only legit to
			 * examine that structure if we're in constrained
			 * situations. Currently, the only times we'll this
			 * check is if a non-super-user has enabled the
			 * profile or syscall providers -- providers that
			 * allow visibility of all processes. For the
			 * profile case, the check above will ensure that
			 * we're examining a user context.
			 */
			if (ecb->dte_cond & DTRACE_COND_OWNER) {
				cred_t *cr;
				cred_t *s_cr =
				    ecb->dte_state->dts_cred.dcr_cred;
				proc_t *proc;

				ASSERT(s_cr != NULL);

				if ((cr = CRED()) == NULL ||
				    s_cr->cr_uid != cr->cr_uid ||
				    s_cr->cr_uid != cr->cr_ruid ||
				    s_cr->cr_uid != cr->cr_suid ||
				    s_cr->cr_gid != cr->cr_gid ||
				    s_cr->cr_gid != cr->cr_rgid ||
				    s_cr->cr_gid != cr->cr_sgid ||
				    (proc = ttoproc(curthread)) == NULL ||
				    (proc->p_flag & SNOCD))
					continue;
			}

			if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
				cred_t *cr;
				cred_t *s_cr =
				    ecb->dte_state->dts_cred.dcr_cred;

				ASSERT(s_cr != NULL);

				if ((cr = CRED()) == NULL ||
				    s_cr->cr_zone->zone_id !=
				    cr->cr_zone->zone_id)
					continue;
			}
		}

		if (now - state->dts_alive > dtrace_deadman_timeout) {
			/*
			 * We seem to be dead.  Unless we (a) have kernel
			 * destructive permissions (b) have expicitly enabled
			 * destructive actions and (c) destructive actions have
			 * not been disabled, we're going to transition into
			 * the KILLED state, from which no further processing
			 * on this state will be performed.
			 */
			if (!dtrace_priv_kernel_destructive(state) ||
			    !state->dts_cred.dcr_destructive ||
			    dtrace_destructive_disallow) {
				void *activity = &state->dts_activity;
				dtrace_activity_t current;

				do {
					current = state->dts_activity;
				} while (dtrace_cas32(activity, current,
				    DTRACE_ACTIVITY_KILLED) != current);

				continue;
			}
		}

		if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
		    ecb->dte_alignment, state, &mstate)) < 0)
			continue;

		tomax = buf->dtb_tomax;
		ASSERT(tomax != NULL);

		if (ecb->dte_size != 0)
			DTRACE_STORE(uint32_t, tomax, offs, ecb->dte_epid);

		mstate.dtms_epid = ecb->dte_epid;
		mstate.dtms_present |= DTRACE_MSTATE_EPID;

		if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
			mstate.dtms_access = DTRACE_ACCESS_KERNEL;
		else
			mstate.dtms_access = 0;

		if (pred != NULL) {
			dtrace_difo_t *dp = pred->dtp_difo;
			int rval;

			rval = dtrace_dif_emulate(dp, &mstate, vstate, state);

			if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
				dtrace_cacheid_t cid = probe->dtpr_predcache;

				if (cid != DTRACE_CACHEIDNONE && !onintr) {
					/*
					 * Update the predicate cache...
					 */
					ASSERT(cid == pred->dtp_cacheid);
					curthread->t_predcache = cid;
				}

				continue;
			}
		}

		for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
		    act != NULL; act = act->dta_next) {
			size_t valoffs;
			dtrace_difo_t *dp;
			dtrace_recdesc_t *rec = &act->dta_rec;

			size = rec->dtrd_size;
			valoffs = offs + rec->dtrd_offset;

			if (DTRACEACT_ISAGG(act->dta_kind)) {
				uint64_t v = 0xbad;
				dtrace_aggregation_t *agg;

				agg = (dtrace_aggregation_t *)act;

				if ((dp = act->dta_difo) != NULL)
					v = dtrace_dif_emulate(dp,
					    &mstate, vstate, state);

				if (*flags & CPU_DTRACE_ERROR)
					continue;

				/*
				 * Note that we always pass the expression
				 * value from the previous iteration of the
				 * action loop.  This value will only be used
				 * if there is an expression argument to the
				 * aggregating action, denoted by the
				 * dtag_hasarg field.
				 */
				dtrace_aggregate(agg, buf,
				    offs, aggbuf, v, val);
				continue;
			}

			switch (act->dta_kind) {
			case DTRACEACT_STOP:
				if (dtrace_priv_proc_destructive(state))
					dtrace_action_stop();
				continue;

			case DTRACEACT_BREAKPOINT:
				if (dtrace_priv_kernel_destructive(state))
					dtrace_action_breakpoint(ecb);
				continue;

			case DTRACEACT_PANIC:
				if (dtrace_priv_kernel_destructive(state))
					dtrace_action_panic(ecb);
				continue;

			case DTRACEACT_STACK:
				if (!dtrace_priv_kernel(state))
					continue;

				dtrace_getpcstack((pc_t *)(tomax + valoffs),
				    size / sizeof (pc_t), probe->dtpr_aframes,
				    DTRACE_ANCHORED(probe) ? NULL :
				    (uint32_t *)arg0);

				continue;

			case DTRACEACT_JSTACK:
			case DTRACEACT_USTACK:
				if (!dtrace_priv_proc(state))
					continue;

				/*
				 * See comment in DIF_VAR_PID.
				 */
				if (DTRACE_ANCHORED(mstate.dtms_probe) &&
				    CPU_ON_INTR(CPU)) {
					int depth = DTRACE_USTACK_NFRAMES(
					    rec->dtrd_arg) + 1;

					dtrace_bzero((void *)(tomax + valoffs),
					    DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
					    + depth * sizeof (uint64_t));

					continue;
				}

				if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
				    curproc->p_dtrace_helpers != NULL) {
					/*
					 * This is the slow path -- we have
					 * allocated string space, and we're
					 * getting the stack of a process that
					 * has helpers.  Call into a separate
					 * routine to perform this processing.
					 */
					dtrace_action_ustack(&mstate, state,
					    (uint64_t *)(tomax + valoffs),
					    rec->dtrd_arg);
					continue;
				}

				DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
				dtrace_getupcstack((uint64_t *)
				    (tomax + valoffs),
				    DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
				DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
				continue;

			default:
				break;
			}

			dp = act->dta_difo;
			ASSERT(dp != NULL);

			val = dtrace_dif_emulate(dp, &mstate, vstate, state);

			if (*flags & CPU_DTRACE_ERROR)
				continue;

			switch (act->dta_kind) {
			case DTRACEACT_SPECULATE:
				ASSERT(buf == &state->dts_buffer[cpuid]);
				buf = dtrace_speculation_buffer(state,
				    cpuid, val);

				if (buf == NULL) {
					*flags |= CPU_DTRACE_DROP;
					continue;
				}

				offs = dtrace_buffer_reserve(buf,
				    ecb->dte_needed, ecb->dte_alignment,
				    state, NULL);

				if (offs < 0) {
					*flags |= CPU_DTRACE_DROP;
					continue;
				}

				tomax = buf->dtb_tomax;
				ASSERT(tomax != NULL);

				if (ecb->dte_size != 0)
					DTRACE_STORE(uint32_t, tomax, offs,
					    ecb->dte_epid);
				continue;

			case DTRACEACT_CHILL:
				if (dtrace_priv_kernel_destructive(state))
					dtrace_action_chill(&mstate, val);
				continue;

			case DTRACEACT_RAISE:
				if (dtrace_priv_proc_destructive(state))
					dtrace_action_raise(val);
				continue;

			case DTRACEACT_COMMIT:
				ASSERT(!committed);

				/*
				 * We need to commit our buffer state.
				 */
				if (ecb->dte_size)
					buf->dtb_offset = offs + ecb->dte_size;
				buf = &state->dts_buffer[cpuid];
				dtrace_speculation_commit(state, cpuid, val);
				committed = 1;
				continue;

			case DTRACEACT_DISCARD:
				dtrace_speculation_discard(state, cpuid, val);
				continue;

			case DTRACEACT_DIFEXPR:
			case DTRACEACT_LIBACT:
			case DTRACEACT_PRINTF:
			case DTRACEACT_PRINTA:
			case DTRACEACT_SYSTEM:
			case DTRACEACT_FREOPEN:
				break;

			case DTRACEACT_SYM:
			case DTRACEACT_MOD:
				if (!dtrace_priv_kernel(state))
					continue;
				break;

			case DTRACEACT_USYM:
			case DTRACEACT_UMOD:
			case DTRACEACT_UADDR: {
				struct pid *pid = curthread->t_procp->p_pidp;

				if (!dtrace_priv_proc(state))
					continue;

				DTRACE_STORE(uint64_t, tomax,
				    valoffs, (uint64_t)pid->pid_id);
				DTRACE_STORE(uint64_t, tomax,
				    valoffs + sizeof (uint64_t), val);

				continue;
			}

			case DTRACEACT_EXIT: {
				/*
				 * For the exit action, we are going to attempt
				 * to atomically set our activity to be
				 * draining.  If this fails (either because
				 * another CPU has beat us to the exit action,
				 * or because our current activity is something
				 * other than ACTIVE or WARMUP), we will
				 * continue.  This assures that the exit action
				 * can be successfully recorded at most once
				 * when we're in the ACTIVE state.  If we're
				 * encountering the exit() action while in
				 * COOLDOWN, however, we want to honor the new
				 * status code.  (We know that we're the only
				 * thread in COOLDOWN, so there is no race.)
				 */
				void *activity = &state->dts_activity;
				dtrace_activity_t current = state->dts_activity;

				if (current == DTRACE_ACTIVITY_COOLDOWN)
					break;

				if (current != DTRACE_ACTIVITY_WARMUP)
					current = DTRACE_ACTIVITY_ACTIVE;

				if (dtrace_cas32(activity, current,
				    DTRACE_ACTIVITY_DRAINING) != current) {
					*flags |= CPU_DTRACE_DROP;
					continue;
				}

				break;
			}

			default:
				ASSERT(0);
			}

			if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF) {
				uintptr_t end = valoffs + size;

				if (!dtrace_vcanload((void *)(uintptr_t)val,
				    &dp->dtdo_rtype, &mstate, vstate))
					continue;

				/*
				 * If this is a string, we're going to only
				 * load until we find the zero byte -- after
				 * which we'll store zero bytes.
				 */
				if (dp->dtdo_rtype.dtdt_kind ==
				    DIF_TYPE_STRING) {
					char c = '\0' + 1;
					int intuple = act->dta_intuple;
					size_t s;

					for (s = 0; s < size; s++) {
						if (c != '\0')
							c = dtrace_load8(val++);

						DTRACE_STORE(uint8_t, tomax,
						    valoffs++, c);

						if (c == '\0' && intuple)
							break;
					}

					continue;
				}

				while (valoffs < end) {
					DTRACE_STORE(uint8_t, tomax, valoffs++,
					    dtrace_load8(val++));
				}

				continue;
			}

			switch (size) {
			case 0:
				break;

			case sizeof (uint8_t):
				DTRACE_STORE(uint8_t, tomax, valoffs, val);
				break;
			case sizeof (uint16_t):
				DTRACE_STORE(uint16_t, tomax, valoffs, val);
				break;
			case sizeof (uint32_t):
				DTRACE_STORE(uint32_t, tomax, valoffs, val);
				break;
			case sizeof (uint64_t):
				DTRACE_STORE(uint64_t, tomax, valoffs, val);
				break;
			default:
				/*
				 * Any other size should have been returned by
				 * reference, not by value.
				 */
				ASSERT(0);
				break;
			}
		}

		if (*flags & CPU_DTRACE_DROP)
			continue;

		if (*flags & CPU_DTRACE_FAULT) {
			int ndx;
			dtrace_action_t *err;

			buf->dtb_errors++;

			if (probe->dtpr_id == dtrace_probeid_error) {
				/*
				 * There's nothing we can do -- we had an
				 * error on the error probe.  We bump an
				 * error counter to at least indicate that
				 * this condition happened.
				 */
				dtrace_error(&state->dts_dblerrors);
				continue;
			}

			if (vtime) {
				/*
				 * Before recursing on dtrace_probe(), we
				 * need to explicitly clear out our start
				 * time to prevent it from being accumulated
				 * into t_dtrace_vtime.
				 */
				curthread->t_dtrace_start = 0;
			}

			/*
			 * Iterate over the actions to figure out which action
			 * we were processing when we experienced the error.
			 * Note that act points _past_ the faulting action; if
			 * act is ecb->dte_action, the fault was in the
			 * predicate, if it's ecb->dte_action->dta_next it's
			 * in action #1, and so on.
			 */
			for (err = ecb->dte_action, ndx = 0;
			    err != act; err = err->dta_next, ndx++)
				continue;

			dtrace_probe_error(state, ecb->dte_epid, ndx,
			    (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
			    mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
			    cpu_core[cpuid].cpuc_dtrace_illval);

			continue;
		}

		if (!committed)
			buf->dtb_offset = offs + ecb->dte_size;
	}

	if (vtime)
		curthread->t_dtrace_start = dtrace_gethrtime();

	dtrace_interrupt_enable(cookie);
}

/*
 * DTrace Probe Hashing Functions
 *
 * The functions in this section (and indeed, the functions in remaining
 * sections) are not _called_ from probe context.  (Any exceptions to this are
 * marked with a "Note:".)  Rather, they are called from elsewhere in the
 * DTrace framework to look-up probes in, add probes to and remove probes from
 * the DTrace probe hashes.  (Each probe is hashed by each element of the
 * probe tuple -- allowing for fast lookups, regardless of what was
 * specified.)
 */
static uint_t
dtrace_hash_str(char *p)
{
	unsigned int g;
	uint_t hval = 0;

	while (*p) {
		hval = (hval << 4) + *p++;
		if ((g = (hval & 0xf0000000)) != 0)
			hval ^= g >> 24;
		hval &= ~g;
	}
	return (hval);
}

static dtrace_hash_t *
dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
{
	dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);

	hash->dth_stroffs = stroffs;
	hash->dth_nextoffs = nextoffs;
	hash->dth_prevoffs = prevoffs;

	hash->dth_size = 1;
	hash->dth_mask = hash->dth_size - 1;

	hash->dth_tab = kmem_zalloc(hash->dth_size *
	    sizeof (dtrace_hashbucket_t *), KM_SLEEP);

	return (hash);
}

static void
dtrace_hash_destroy(dtrace_hash_t *hash)
{
#ifdef DEBUG
	int i;

	for (i = 0; i < hash->dth_size; i++)
		ASSERT(hash->dth_tab[i] == NULL);
#endif

	kmem_free(hash->dth_tab,
	    hash->dth_size * sizeof (dtrace_hashbucket_t *));
	kmem_free(hash, sizeof (dtrace_hash_t));
}

static void
dtrace_hash_resize(dtrace_hash_t *hash)
{
	int size = hash->dth_size, i, ndx;
	int new_size = hash->dth_size << 1;
	int new_mask = new_size - 1;
	dtrace_hashbucket_t **new_tab, *bucket, *next;

	ASSERT((new_size & new_mask) == 0);

	new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);

	for (i = 0; i < size; i++) {
		for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
			dtrace_probe_t *probe = bucket->dthb_chain;

			ASSERT(probe != NULL);
			ndx = DTRACE_HASHSTR(hash, probe) & new_mask;

			next = bucket->dthb_next;
			bucket->dthb_next = new_tab[ndx];
			new_tab[ndx] = bucket;
		}
	}

	kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
	hash->dth_tab = new_tab;
	hash->dth_size = new_size;
	hash->dth_mask = new_mask;
}

static void
dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
{
	int hashval = DTRACE_HASHSTR(hash, new);
	int ndx = hashval & hash->dth_mask;
	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
	dtrace_probe_t **nextp, **prevp;

	for (; bucket != NULL; bucket = bucket->dthb_next) {
		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
			goto add;
	}

	if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
		dtrace_hash_resize(hash);
		dtrace_hash_add(hash, new);
		return;
	}

	bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
	bucket->dthb_next = hash->dth_tab[ndx];
	hash->dth_tab[ndx] = bucket;
	hash->dth_nbuckets++;

add:
	nextp = DTRACE_HASHNEXT(hash, new);
	ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
	*nextp = bucket->dthb_chain;

	if (bucket->dthb_chain != NULL) {
		prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
		ASSERT(*prevp == NULL);
		*prevp = new;
	}

	bucket->dthb_chain = new;
	bucket->dthb_len++;
}

static dtrace_probe_t *
dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
{
	int hashval = DTRACE_HASHSTR(hash, template);
	int ndx = hashval & hash->dth_mask;
	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];

	for (; bucket != NULL; bucket = bucket->dthb_next) {
		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
			return (bucket->dthb_chain);
	}

	return (NULL);
}

static int
dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
{
	int hashval = DTRACE_HASHSTR(hash, template);
	int ndx = hashval & hash->dth_mask;
	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];

	for (; bucket != NULL; bucket = bucket->dthb_next) {
		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
			return (bucket->dthb_len);
	}

	return (NULL);
}

static void
dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
{
	int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
	dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];

	dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
	dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);

	/*
	 * Find the bucket that we're removing this probe from.
	 */
	for (; bucket != NULL; bucket = bucket->dthb_next) {
		if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
			break;
	}

	ASSERT(bucket != NULL);

	if (*prevp == NULL) {
		if (*nextp == NULL) {
			/*
			 * The removed probe was the only probe on this
			 * bucket; we need to remove the bucket.
			 */
			dtrace_hashbucket_t *b = hash->dth_tab[ndx];

			ASSERT(bucket->dthb_chain == probe);
			ASSERT(b != NULL);

			if (b == bucket) {
				hash->dth_tab[ndx] = bucket->dthb_next;
			} else {
				while (b->dthb_next != bucket)
					b = b->dthb_next;
				b->dthb_next = bucket->dthb_next;
			}

			ASSERT(hash->dth_nbuckets > 0);
			hash->dth_nbuckets--;
			kmem_free(bucket, sizeof (dtrace_hashbucket_t));
			return;
		}

		bucket->dthb_chain = *nextp;
	} else {
		*(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
	}

	if (*nextp != NULL)
		*(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
}

/*
 * DTrace Utility Functions
 *
 * These are random utility functions that are _not_ called from probe context.
 */
static int
dtrace_badattr(const dtrace_attribute_t *a)
{
	return (a->dtat_name > DTRACE_STABILITY_MAX ||
	    a->dtat_data > DTRACE_STABILITY_MAX ||
	    a->dtat_class > DTRACE_CLASS_MAX);
}

/*
 * Return a duplicate copy of a string.  If the specified string is NULL,
 * this function returns a zero-length string.
 */
static char *
dtrace_strdup(const char *str)
{
	char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);

	if (str != NULL)
		(void) strcpy(new, str);

	return (new);
}

#define	DTRACE_ISALPHA(c)	\
	(((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))

static int
dtrace_badname(const char *s)
{
	char c;

	if (s == NULL || (c = *s++) == '\0')
		return (0);

	if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
		return (1);

	while ((c = *s++) != '\0') {
		if (!DTRACE_ISALPHA(c) && (c < '0' || c > '9') &&
		    c != '-' && c != '_' && c != '.' && c != '`')
			return (1);
	}

	return (0);
}

static void
dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
{
	uint32_t priv;

	if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
		/*
		 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
		 */
		priv = DTRACE_PRIV_ALL;
	} else {
		*uidp = crgetuid(cr);
		*zoneidp = crgetzoneid(cr);

		priv = 0;
		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
			priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
		else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
			priv |= DTRACE_PRIV_USER;
		if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
			priv |= DTRACE_PRIV_PROC;
		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
			priv |= DTRACE_PRIV_OWNER;
		if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
			priv |= DTRACE_PRIV_ZONEOWNER;
	}

	*privp = priv;
}

#ifdef DTRACE_ERRDEBUG
static void
dtrace_errdebug(const char *str)
{
	int hval = dtrace_hash_str((char *)str) % DTRACE_ERRHASHSZ;
	int occupied = 0;

	mutex_enter(&dtrace_errlock);
	dtrace_errlast = str;
	dtrace_errthread = curthread;

	while (occupied++ < DTRACE_ERRHASHSZ) {
		if (dtrace_errhash[hval].dter_msg == str) {
			dtrace_errhash[hval].dter_count++;
			goto out;
		}

		if (dtrace_errhash[hval].dter_msg != NULL) {
			hval = (hval + 1) % DTRACE_ERRHASHSZ;
			continue;
		}

		dtrace_errhash[hval].dter_msg = str;
		dtrace_errhash[hval].dter_count = 1;
		goto out;
	}

	panic("dtrace: undersized error hash");
out:
	mutex_exit(&dtrace_errlock);
}
#endif

/*
 * DTrace Matching Functions
 *
 * These functions are used to match groups of probes, given some elements of
 * a probe tuple, or some globbed expressions for elements of a probe tuple.
 */
static int
dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
    zoneid_t zoneid)
{
	if (priv != DTRACE_PRIV_ALL) {
		uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
		uint32_t match = priv & ppriv;

		/*
		 * No PRIV_DTRACE_* privileges...
		 */
		if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
		    DTRACE_PRIV_KERNEL)) == 0)
			return (0);

		/*
		 * No matching bits, but there were bits to match...
		 */
		if (match == 0 && ppriv != 0)
			return (0);

		/*
		 * Need to have permissions to the process, but don't...
		 */
		if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
		    uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
			return (0);
		}

		/*
		 * Need to be in the same zone unless we possess the
		 * privilege to examine all zones.
		 */
		if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
		    zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
			return (0);
		}
	}

	return (1);
}

/*
 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
 * consists of input pattern strings and an ops-vector to evaluate them.
 * This function returns >0 for match, 0 for no match, and <0 for error.
 */
static int
dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
    uint32_t priv, uid_t uid, zoneid_t zoneid)
{
	dtrace_provider_t *pvp = prp->dtpr_provider;
	int rv;

	if (pvp->dtpv_defunct)
		return (0);

	if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
		return (rv);

	if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
		return (rv);

	if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
		return (rv);

	if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
		return (rv);

	if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
		return (0);

	return (rv);
}

/*
 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
 * interface for matching a glob pattern 'p' to an input string 's'.  Unlike
 * libc's version, the kernel version only applies to 8-bit ASCII strings.
 * In addition, all of the recursion cases except for '*' matching have been
 * unwound.  For '*', we still implement recursive evaluation, but a depth
 * counter is maintained and matching is aborted if we recurse too deep.
 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
 */
static int
dtrace_match_glob(const char *s, const char *p, int depth)
{
	const char *olds;
	char s1, c;
	int gs;

	if (depth > DTRACE_PROBEKEY_MAXDEPTH)
		return (-1);

	if (s == NULL)
		s = ""; /* treat NULL as empty string */

top:
	olds = s;
	s1 = *s++;

	if (p == NULL)
		return (0);

	if ((c = *p++) == '\0')
		return (s1 == '\0');

	switch (c) {
	case '[': {
		int ok = 0, notflag = 0;
		char lc = '\0';

		if (s1 == '\0')
			return (0);

		if (*p == '!') {
			notflag = 1;
			p++;
		}

		if ((c = *p++) == '\0')
			return (0);

		do {
			if (c == '-' && lc != '\0' && *p != ']') {
				if ((c = *p++) == '\0')
					return (0);
				if (c == '\\' && (c = *p++) == '\0')
					return (0);

				if (notflag) {
					if (s1 < lc || s1 > c)
						ok++;
					else
						return (0);
				} else if (lc <= s1 && s1 <= c)
					ok++;

			} else if (c == '\\' && (c = *p++) == '\0')
				return (0);

			lc = c; /* save left-hand 'c' for next iteration */

			if (notflag) {
				if (s1 != c)
					ok++;
				else
					return (0);
			} else if (s1 == c)
				ok++;

			if ((c = *p++) == '\0')
				return (0);

		} while (c != ']');

		if (ok)
			goto top;

		return (0);
	}

	case '\\':
		if ((c = *p++) == '\0')
			return (0);
		/*FALLTHRU*/

	default:
		if (c != s1)
			return (0);
		/*FALLTHRU*/

	case '?':
		if (s1 != '\0')
			goto top;
		return (0);

	case '*':
		while (*p == '*')
			p++; /* consecutive *'s are identical to a single one */

		if (*p == '\0')
			return (1);

		for (s = olds; *s != '\0'; s++) {
			if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
				return (gs);
		}

		return (0);
	}
}

/*ARGSUSED*/
static int
dtrace_match_string(const char *s, const char *p, int depth)
{
	return (s != NULL && strcmp(s, p) == 0);
}

/*ARGSUSED*/
static int
dtrace_match_nul(const char *s, const char *p, int depth)
{
	return (1); /* always match the empty pattern */
}

/*ARGSUSED*/
static int
dtrace_match_nonzero(const char *s, const char *p, int depth)
{
	return (s != NULL && s[0] != '\0');
}

static int
dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
    zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
{
	dtrace_probe_t template, *probe;
	dtrace_hash_t *hash = NULL;
	int len, rc, best = INT_MAX, nmatched = 0;
	dtrace_id_t i;

	ASSERT(MUTEX_HELD(&dtrace_lock));

	/*
	 * If the probe ID is specified in the key, just lookup by ID and
	 * invoke the match callback once if a matching probe is found.
	 */
	if (pkp->dtpk_id != DTRACE_IDNONE) {
		if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
		    dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
			if ((*matched)(probe, arg) == DTRACE_MATCH_FAIL)
				return (DTRACE_MATCH_FAIL);
			nmatched++;
		}
		return (nmatched);
	}

	template.dtpr_mod = (char *)pkp->dtpk_mod;
	template.dtpr_func = (char *)pkp->dtpk_func;
	template.dtpr_name = (char *)pkp->dtpk_name;

	/*
	 * We want to find the most distinct of the module name, function
	 * name, and name.  So for each one that is not a glob pattern or
	 * empty string, we perform a lookup in the corresponding hash and
	 * use the hash table with the fewest collisions to do our search.
	 */
	if (pkp->dtpk_mmatch == &dtrace_match_string &&
	    (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
		best = len;
		hash = dtrace_bymod;
	}

	if (pkp->dtpk_fmatch == &dtrace_match_string &&
	    (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
		best = len;
		hash = dtrace_byfunc;
	}

	if (pkp->dtpk_nmatch == &dtrace_match_string &&
	    (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
		best = len;
		hash = dtrace_byname;
	}

	/*
	 * If we did not select a hash table, iterate over every probe and
	 * invoke our callback for each one that matches our input probe key.
	 */
	if (hash == NULL) {
		for (i = 0; i < dtrace_nprobes; i++) {
			if ((probe = dtrace_probes[i]) == NULL ||
			    dtrace_match_probe(probe, pkp, priv, uid,
			    zoneid) <= 0)
				continue;

			nmatched++;

			if ((rc = (*matched)(probe, arg)) !=
			    DTRACE_MATCH_NEXT) {
				if (rc == DTRACE_MATCH_FAIL)
					return (DTRACE_MATCH_FAIL);
				break;
			}
		}

		return (nmatched);
	}

	/*
	 * If we selected a hash table, iterate over each probe of the same key
	 * name and invoke the callback for every probe that matches the other
	 * attributes of our input probe key.
	 */
	for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
	    probe = *(DTRACE_HASHNEXT(hash, probe))) {

		if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
			continue;

		nmatched++;

		if ((rc = (*matched)(probe, arg)) != DTRACE_MATCH_NEXT) {
			if (rc == DTRACE_MATCH_FAIL)
				return (DTRACE_MATCH_FAIL);
			break;
		}
	}

	return (nmatched);
}

/*
 * Return the function pointer dtrace_probecmp() should use to compare the
 * specified pattern with a string.  For NULL or empty patterns, we select
 * dtrace_match_nul().  For glob pattern strings, we use dtrace_match_glob().
 * For non-empty non-glob strings, we use dtrace_match_string().
 */
static dtrace_probekey_f *
dtrace_probekey_func(const char *p)
{
	char c;

	if (p == NULL || *p == '\0')
		return (&dtrace_match_nul);

	while ((c = *p++) != '\0') {
		if (c == '[' || c == '?' || c == '*' || c == '\\')
			return (&dtrace_match_glob);
	}

	return (&dtrace_match_string);
}

/*
 * Build a probe comparison key for use with dtrace_match_probe() from the
 * given probe description.  By convention, a null key only matches anchored
 * probes: if each field is the empty string, reset dtpk_fmatch to
 * dtrace_match_nonzero().
 */
static void
dtrace_probekey(const dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
{
	pkp->dtpk_prov = pdp->dtpd_provider;
	pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);

	pkp->dtpk_mod = pdp->dtpd_mod;
	pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);

	pkp->dtpk_func = pdp->dtpd_func;
	pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);

	pkp->dtpk_name = pdp->dtpd_name;
	pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);

	pkp->dtpk_id = pdp->dtpd_id;

	if (pkp->dtpk_id == DTRACE_IDNONE &&
	    pkp->dtpk_pmatch == &dtrace_match_nul &&
	    pkp->dtpk_mmatch == &dtrace_match_nul &&
	    pkp->dtpk_fmatch == &dtrace_match_nul &&
	    pkp->dtpk_nmatch == &dtrace_match_nul)
		pkp->dtpk_fmatch = &dtrace_match_nonzero;
}

/*
 * DTrace Provider-to-Framework API Functions
 *
 * These functions implement much of the Provider-to-Framework API, as
 * described in <sys/dtrace.h>.  The parts of the API not in this section are
 * the functions in the API for probe management (found below), and
 * dtrace_probe() itself (found above).
 */

/*
 * Register the calling provider with the DTrace framework.  This should
 * generally be called by DTrace providers in their attach(9E) entry point.
 */
int
dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
    cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
{
	dtrace_provider_t *provider;

	if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
		    "arguments", name ? name : "<NULL>");
		return (EINVAL);
	}

	if (name[0] == '\0' || dtrace_badname(name)) {
		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
		    "provider name", name);
		return (EINVAL);
	}

	if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
	    pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
	    pops->dtps_destroy == NULL ||
	    ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
		    "provider ops", name);
		return (EINVAL);
	}

	if (dtrace_badattr(&pap->dtpa_provider) ||
	    dtrace_badattr(&pap->dtpa_mod) ||
	    dtrace_badattr(&pap->dtpa_func) ||
	    dtrace_badattr(&pap->dtpa_name) ||
	    dtrace_badattr(&pap->dtpa_args)) {
		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
		    "provider attributes", name);
		return (EINVAL);
	}

	if (priv & ~DTRACE_PRIV_ALL) {
		cmn_err(CE_WARN, "failed to register provider '%s': invalid "
		    "privilege attributes", name);
		return (EINVAL);
	}

	if ((priv & DTRACE_PRIV_KERNEL) &&
	    (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
	    pops->dtps_usermode == NULL) {
		cmn_err(CE_WARN, "failed to register provider '%s': need "
		    "dtps_usermode() op for given privilege attributes", name);
		return (EINVAL);
	}

	provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
	provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
	(void) strcpy(provider->dtpv_name, name);

	provider->dtpv_attr = *pap;
	provider->dtpv_priv.dtpp_flags = priv;
	if (cr != NULL) {
		provider->dtpv_priv.dtpp_uid = crgetuid(cr);
		provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
	}
	provider->dtpv_pops = *pops;

	if (pops->dtps_provide == NULL) {
		ASSERT(pops->dtps_provide_module != NULL);
		provider->dtpv_pops.dtps_provide =
		    (void (*)(void *, const dtrace_probedesc_t *))dtrace_nullop;
	}

	if (pops->dtps_provide_module == NULL) {
		ASSERT(pops->dtps_provide != NULL);
		provider->dtpv_pops.dtps_provide_module =
		    (void (*)(void *, struct modctl *))dtrace_nullop;
	}

	if (pops->dtps_suspend == NULL) {
		ASSERT(pops->dtps_resume == NULL);
		provider->dtpv_pops.dtps_suspend =
		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
		provider->dtpv_pops.dtps_resume =
		    (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
	}

	provider->dtpv_arg = arg;
	*idp = (dtrace_provider_id_t)provider;

	if (pops == &dtrace_provider_ops) {
		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
		ASSERT(MUTEX_HELD(&dtrace_lock));
		ASSERT(dtrace_anon.dta_enabling == NULL);

		/*
		 * We make sure that the DTrace provider is at the head of
		 * the provider chain.
		 */
		provider->dtpv_next = dtrace_provider;
		dtrace_provider = provider;
		return (0);
	}

	mutex_enter(&dtrace_provider_lock);
	mutex_enter(&dtrace_lock);

	/*
	 * If there is at least one provider registered, we'll add this
	 * provider after the first provider.
	 */
	if (dtrace_provider != NULL) {
		provider->dtpv_next = dtrace_provider->dtpv_next;
		dtrace_provider->dtpv_next = provider;
	} else {
		dtrace_provider = provider;
	}

	if (dtrace_retained != NULL) {
		dtrace_enabling_provide(provider);

		/*
		 * Now we need to call dtrace_enabling_matchall() -- which
		 * will acquire cpu_lock and dtrace_lock.  We therefore need
		 * to drop all of our locks before calling into it...
		 */
		mutex_exit(&dtrace_lock);
		mutex_exit(&dtrace_provider_lock);
		dtrace_enabling_matchall();

		return (0);
	}

	mutex_exit(&dtrace_lock);
	mutex_exit(&dtrace_provider_lock);

	return (0);
}

/*
 * Unregister the specified provider from the DTrace framework.  This should
 * generally be called by DTrace providers in their detach(9E) entry point.
 */
int
dtrace_unregister(dtrace_provider_id_t id)
{
	dtrace_provider_t *old = (dtrace_provider_t *)id;
	dtrace_provider_t *prev = NULL;
	int i, self = 0;
	dtrace_probe_t *probe, *first = NULL;

	if (old->dtpv_pops.dtps_enable ==
	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop) {
		/*
		 * If DTrace itself is the provider, we're called with locks
		 * already held.
		 */
		ASSERT(old == dtrace_provider);
		ASSERT(dtrace_devi != NULL);
		ASSERT(MUTEX_HELD(&dtrace_provider_lock));
		ASSERT(MUTEX_HELD(&dtrace_lock));
		self = 1;

		if (dtrace_provider->dtpv_next != NULL) {
			/*
			 * There's another provider here; return failure.
			 */
			return (EBUSY);
		}
	} else {
		mutex_enter(&dtrace_provider_lock);
		mutex_enter(&mod_lock);
		mutex_enter(&dtrace_lock);
	}

	/*
	 * If anyone has /dev/dtrace open, or if there are anonymous enabled
	 * probes, we refuse to let providers slither away, unless this
	 * provider has already been explicitly invalidated.
	 */
	if (!old->dtpv_defunct &&
	    (dtrace_opens || (dtrace_anon.dta_state != NULL &&
	    dtrace_anon.dta_state->dts_necbs > 0))) {
		if (!self) {
			mutex_exit(&dtrace_lock);
			mutex_exit(&mod_lock);
			mutex_exit(&dtrace_provider_lock);
		}
		return (EBUSY);
	}

	/*
	 * Attempt to destroy the probes associated with this provider.
	 */
	for (i = 0; i < dtrace_nprobes; i++) {
		if ((probe = dtrace_probes[i]) == NULL)
			continue;

		if (probe->dtpr_provider != old)
			continue;

		if (probe->dtpr_ecb == NULL)
			continue;

		/*
		 * We have at least one ECB; we can't remove this provider.
		 */
		if (!self) {
			mutex_exit(&dtrace_lock);
			mutex_exit(&mod_lock);
			mutex_exit(&dtrace_provider_lock);
		}
		return (EBUSY);
	}

	/*
	 * All of the probes for this provider are disabled; we can safely
	 * remove all of them from their hash chains and from the probe array.
	 */
	for (i = 0; i < dtrace_nprobes; i++) {
		if ((probe = dtrace_probes[i]) == NULL)
			continue;

		if (probe->dtpr_provider != old)
			continue;

		dtrace_probes[i] = NULL;

		dtrace_hash_remove(dtrace_bymod, probe);
		dtrace_hash_remove(dtrace_byfunc, probe);
		dtrace_hash_remove(dtrace_byname, probe);

		if (first == NULL) {
			first = probe;
			probe->dtpr_nextmod = NULL;
		} else {
			probe->dtpr_nextmod = first;
			first = probe;
		}
	}

	/*
	 * The provider's probes have been removed from the hash chains and
	 * from the probe array.  Now issue a dtrace_sync() to be sure that
	 * everyone has cleared out from any probe array processing.
	 */
	dtrace_sync();

	for (probe = first; probe != NULL; probe = first) {
		first = probe->dtpr_nextmod;

		old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
		    probe->dtpr_arg);
		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
		vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
		kmem_free(probe, sizeof (dtrace_probe_t));
	}

	if ((prev = dtrace_provider) == old) {
		ASSERT(self || dtrace_devi == NULL);
		ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
		dtrace_provider = old->dtpv_next;
	} else {
		while (prev != NULL && prev->dtpv_next != old)
			prev = prev->dtpv_next;

		if (prev == NULL) {
			panic("attempt to unregister non-existent "
			    "dtrace provider %p\n", (void *)id);
		}

		prev->dtpv_next = old->dtpv_next;
	}

	if (!self) {
		mutex_exit(&dtrace_lock);
		mutex_exit(&mod_lock);
		mutex_exit(&dtrace_provider_lock);
	}

	kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
	kmem_free(old, sizeof (dtrace_provider_t));

	return (0);
}

/*
 * Invalidate the specified provider.  All subsequent probe lookups for the
 * specified provider will fail, but its probes will not be removed.
 */
void
dtrace_invalidate(dtrace_provider_id_t id)
{
	dtrace_provider_t *pvp = (dtrace_provider_t *)id;

	ASSERT(pvp->dtpv_pops.dtps_enable !=
	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);

	mutex_enter(&dtrace_provider_lock);
	mutex_enter(&dtrace_lock);

	pvp->dtpv_defunct = 1;

	mutex_exit(&dtrace_lock);
	mutex_exit(&dtrace_provider_lock);
}

/*
 * Indicate whether or not DTrace has attached.
 */
int
dtrace_attached(void)
{
	/*
	 * dtrace_provider will be non-NULL iff the DTrace driver has
	 * attached.  (It's non-NULL because DTrace is always itself a
	 * provider.)
	 */
	return (dtrace_provider != NULL);
}

/*
 * Remove all the unenabled probes for the given provider.  This function is
 * not unlike dtrace_unregister(), except that it doesn't remove the provider
 * -- just as many of its associated probes as it can.
 */
int
dtrace_condense(dtrace_provider_id_t id)
{
	dtrace_provider_t *prov = (dtrace_provider_t *)id;
	int i;
	dtrace_probe_t *probe;

	/*
	 * Make sure this isn't the dtrace provider itself.
	 */
	ASSERT(prov->dtpv_pops.dtps_enable !=
	    (int (*)(void *, dtrace_id_t, void *))dtrace_enable_nullop);

	mutex_enter(&dtrace_provider_lock);
	mutex_enter(&dtrace_lock);

	/*
	 * Attempt to destroy the probes associated with this provider.
	 */
	for (i = 0; i < dtrace_nprobes; i++) {
		if ((probe = dtrace_probes[i]) == NULL)
			continue;

		if (probe->dtpr_provider != prov)
			continue;

		if (probe->dtpr_ecb != NULL)
			continue;

		dtrace_probes[i] = NULL;

		dtrace_hash_remove(dtrace_bymod, probe);
		dtrace_hash_remove(dtrace_byfunc, probe);
		dtrace_hash_remove(dtrace_byname, probe);

		prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
		    probe->dtpr_arg);
		kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
		kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
		kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
		kmem_free(probe, sizeof (dtrace_probe_t));
		vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
	}

	mutex_exit(&dtrace_lock);
	mutex_exit(&dtrace_provider_lock);

	return (0);
}

/*
 * DTrace Probe Management Functions
 *
 * The functions in this section perform the DTrace probe management,
 * including functions to create probes, look-up probes, and call into the
 * providers to request that probes be provided.  Some of these functions are
 * in the Provider-to-Framework API; these functions can be identified by the
 * fact that they are not declared "static".
 */

/*
 * Create a probe with the specified module name, function name, and name.
 */
dtrace_id_t
dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
    const char *func, const char *name, int aframes, void *arg)
{
	dtrace_probe_t *probe, **probes;
	dtrace_provider_t *provider = (dtrace_provider_t *)prov;
	dtrace_id_t id;

	if (provider == dtrace_provider) {
		ASSERT(MUTEX_HELD(&dtrace_lock));
	} else {
		mutex_enter(&dtrace_lock);
	}

	id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
	    VM_BESTFIT | VM_SLEEP);
	probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);

	probe->dtpr_id = id;
	probe->dtpr_gen = dtrace_probegen++;
	probe->dtpr_mod = dtrace_strdup(mod);
	probe->dtpr_func = dtrace_strdup(func);
	probe->dtpr_name = dtrace_strdup(name);
	probe->dtpr_arg = arg;
	probe->dtpr_aframes = aframes;
	probe->dtpr_provider = provider;

	dtrace_hash_add(dtrace_bymod, probe);
	dtrace_hash_add(dtrace_byfunc, probe);
	dtrace_hash_add(dtrace_byname, probe);

	if (id - 1 >= dtrace_nprobes) {
		size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
		size_t nsize = osize << 1;

		if (nsize == 0) {
			ASSERT(osize == 0);
			ASSERT(dtrace_probes == NULL);
			nsize = sizeof (dtrace_probe_t *);
		}

		probes = kmem_zalloc(nsize, KM_SLEEP);

		if (dtrace_probes == NULL) {
			ASSERT(osize == 0);
			dtrace_probes = probes;
			dtrace_nprobes = 1;
		} else {
			dtrace_probe_t **oprobes = dtrace_probes;

			bcopy(oprobes, probes, osize);
			dtrace_membar_producer();
			dtrace_probes = probes;

			dtrace_sync();

			/*
			 * All CPUs are now seeing the new probes array; we can
			 * safely free the old array.
			 */
			kmem_free(oprobes, osize);
			dtrace_nprobes <<= 1;
		}

		ASSERT(id - 1 < dtrace_nprobes);
	}

	ASSERT(dtrace_probes[id - 1] == NULL);
	dtrace_probes[id - 1] = probe;

	if (provider != dtrace_provider)
		mutex_exit(&dtrace_lock);

	return (id);
}

static dtrace_probe_t *
dtrace_probe_lookup_id(dtrace_id_t id)
{
	ASSERT(MUTEX_HELD(&dtrace_lock));

	if (id == 0 || id > dtrace_nprobes)
		return (NULL);

	return (dtrace_probes[id - 1]);
}

static int
dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
{
	*((dtrace_id_t *)arg) = probe->dtpr_id;

	return (DTRACE_MATCH_DONE);
}

/*
 * Look up a probe based on provider and one or more of module name, function
 * name and probe name.
 */
dtrace_id_t
dtrace_probe_lookup(dtrace_provider_id_t prid, const char *mod,
    const char *func, const char *name)
{
	dtrace_probekey_t pkey;
	dtrace_id_t id;
	int match;

	pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
	pkey.dtpk_pmatch = &dtrace_match_string;
	pkey.dtpk_mod = mod;
	pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
	pkey.dtpk_func = func;
	pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
	pkey.dtpk_name = name;
	pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
	pkey.dtpk_id = DTRACE_IDNONE;

	mutex_enter(&dtrace_lock);
	match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
	    dtrace_probe_lookup_match, &id);
	mutex_exit(&dtrace_lock);

	ASSERT(match == 1 || match == 0);
	return (match ? id : 0);
}

/*
 * Returns the probe argument associated with the specified probe.
 */
void *
dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
{
	dtrace_probe_t *probe;
	void *rval = NULL;

	mutex_enter(&dtrace_lock);

	if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
	    probe->dtpr_provider == (dtrace_provider_t *)id)
		rval = probe->dtpr_arg;

	mutex_exit(&dtrace_lock);

	return (rval);
}

/*
 * Copy a probe into a probe description.
 */
static void
dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
{
	bzero(pdp, sizeof (dtrace_probedesc_t));
	pdp->dtpd_id = prp->dtpr_id;

	(void) strncpy(pdp->dtpd_provider,
	    prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);

	(void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
	(void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
	(void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
}

/*
 * Called to indicate that a probe -- or probes -- should be provided by a
 * specfied provider.  If the specified description is NULL, the provider will
 * be told to provide all of its probes.  (This is done whenever a new
 * consumer comes along, or whenever a retained enabling is to be matched.) If
 * the specified description is non-NULL, the provider is given the
 * opportunity to dynamically provide the specified probe, allowing providers
 * to support the creation of probes on-the-fly.  (So-called _autocreated_
 * probes.)  If the provider is NULL, the operations will be applied to all
 * providers; if the provider is non-NULL the operations will only be applied
 * to the specified provider.  The dtrace_provider_lock must be held, and the
 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
 * will need to grab the dtrace_lock when it reenters the framework through
 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
 */
static void
dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
{
	struct modctl *ctl;
	int all = 0;

	ASSERT(MUTEX_HELD(&dtrace_provider_lock));

	if (prv == NULL) {
		all = 1;
		prv = dtrace_provider;
	}

	do {
		/*
		 * First, call the blanket provide operation.
		 */
		prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);

		/*
		 * Now call the per-module provide operation.  We will grab
		 * mod_lock to prevent the list from being modified.  Note
		 * that this also prevents the mod_busy bits from changing.
		 * (mod_busy can only be changed with mod_lock held.)
		 */
		mutex_enter(&mod_lock);

		ctl = &modules;
		do {
			if (ctl->mod_busy || ctl->mod_mp == NULL)
				continue;

			prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);

		} while ((ctl = ctl->mod_next) != &modules);

		mutex_exit(&mod_lock);
	} while (all && (prv = prv->dtpv_next) != NULL);
}

/*
 * Iterate over each probe, and call the Framework-to-Provider API function
 * denoted by offs.
 */
static void
dtrace_probe_foreach(uintptr_t offs)
{
	dtrace_provider_t *prov;
	void (*func)(void *, dtrace_id_t, void *);
	dtrace_probe_t *probe;
	dtrace_icookie_t cookie;
	int i;

	/*
	 * We disable interrupts to walk through the probe array.  This is
	 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
	 * won't see stale data.
	 */
	cookie = dtrace_interrupt_disable();

	for (i = 0; i < dtrace_nprobes; i++) {
		if ((probe = dtrace_probes[i]) == NULL)
			continue;

		if (probe->dtpr_ecb == NULL) {
			/*
			 * This probe isn't enabled -- don't call the function.
			 */
			continue;
		}

		prov = probe->dtpr_provider;
		func = *((void(**)(void *, dtrace_id_t, void *))
		    ((uintptr_t)&prov->dtpv_pops + offs));

		func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
	}

	dtrace_interrupt_enable(cookie);
}

static int
dtrace_probe_enable(const dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
{
	dtrace_probekey_t pkey;
	uint32_t priv;
	uid_t uid;
	zoneid_t zoneid;

	ASSERT(MUTEX_HELD(&dtrace_lock));
	dtrace_ecb_create_cache = NULL;

	if (desc == NULL) {
		/*
		 * If we're passed a NULL description, we're being asked to
		 * create an ECB with a NULL probe.
		 */
		(void) dtrace_ecb_create_enable(NULL, enab);
		return (0);
	}

	dtrace_probekey(desc, &pkey);
	dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
	    &priv, &uid, &zoneid);

	return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
	    enab));
}

/*
 * DTrace Helper Provider Functions
 */
static void
dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
{
	attr->dtat_name = DOF_ATTR_NAME(dofattr);
	attr->dtat_data = DOF_ATTR_DATA(dofattr);
	attr->dtat_class = DOF_ATTR_CLASS(dofattr);
}

static void
dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
    const dof_provider_t *dofprov, char *strtab)
{
	hprov->dthpv_provname = strtab + dofprov->dofpv_name;
	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
	    dofprov->dofpv_provattr);
	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
	    dofprov->dofpv_modattr);
	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
	    dofprov->dofpv_funcattr);
	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
	    dofprov->dofpv_nameattr);
	dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
	    dofprov->dofpv_argsattr);
}

static void
dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
{
	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
	dof_hdr_t *dof = (dof_hdr_t *)daddr;
	dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
	dof_provider_t *provider;
	dof_probe_t *probe;
	uint32_t *off, *enoff;
	uint8_t *arg;
	char *strtab;
	uint_t i, nprobes;
	dtrace_helper_provdesc_t dhpv;
	dtrace_helper_probedesc_t dhpb;
	dtrace_meta_t *meta = dtrace_meta_pid;
	dtrace_mops_t *mops = &meta->dtm_mops;
	void *parg;

	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
	    provider->dofpv_strtab * dof->dofh_secsize);
	prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
	    provider->dofpv_probes * dof->dofh_secsize);
	arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
	    provider->dofpv_prargs * dof->dofh_secsize);
	off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
	    provider->dofpv_proffs * dof->dofh_secsize);

	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
	off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
	arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
	enoff = NULL;

	/*
	 * See dtrace_helper_provider_validate().
	 */
	if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
	    provider->dofpv_prenoffs != DOF_SECT_NONE) {
		enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
		    provider->dofpv_prenoffs * dof->dofh_secsize);
		enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
	}

	nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;

	/*
	 * Create the provider.
	 */
	dtrace_dofprov2hprov(&dhpv, provider, strtab);

	if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
		return;

	meta->dtm_count++;

	/*
	 * Create the probes.
	 */
	for (i = 0; i < nprobes; i++) {
		probe = (dof_probe_t *)(uintptr_t)(daddr +
		    prb_sec->dofs_offset + i * prb_sec->dofs_entsize);

		dhpb.dthpb_mod = dhp->dofhp_mod;
		dhpb.dthpb_func = strtab + probe->dofpr_func;
		dhpb.dthpb_name = strtab + probe->dofpr_name;
		dhpb.dthpb_base = probe->dofpr_addr;
		dhpb.dthpb_offs = off + probe->dofpr_offidx;
		dhpb.dthpb_noffs = probe->dofpr_noffs;
		if (enoff != NULL) {
			dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
			dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
		} else {
			dhpb.dthpb_enoffs = NULL;
			dhpb.dthpb_nenoffs = 0;
		}
		dhpb.dthpb_args = arg + probe->dofpr_argidx;
		dhpb.dthpb_nargc = probe->dofpr_nargc;
		dhpb.dthpb_xargc = probe->dofpr_xargc;
		dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
		dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;

		mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
	}
}

static void
dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
{
	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
	dof_hdr_t *dof = (dof_hdr_t *)daddr;
	int i;

	ASSERT(MUTEX_HELD(&dtrace_meta_lock));

	for (i = 0; i < dof->dofh_secnum; i++) {
		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
		    dof->dofh_secoff + i * dof->dofh_secsize);

		if (sec->dofs_type != DOF_SECT_PROVIDER)
			continue;

		dtrace_helper_provide_one(dhp, sec, pid);
	}

	/*
	 * We may have just created probes, so we must now rematch against
	 * any retained enablings.  Note that this call will acquire both
	 * cpu_lock and dtrace_lock; the fact that we are holding
	 * dtrace_meta_lock now is what defines the ordering with respect to
	 * these three locks.
	 */
	dtrace_enabling_matchall();
}

static void
dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
{
	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
	dof_hdr_t *dof = (dof_hdr_t *)daddr;
	dof_sec_t *str_sec;
	dof_provider_t *provider;
	char *strtab;
	dtrace_helper_provdesc_t dhpv;
	dtrace_meta_t *meta = dtrace_meta_pid;
	dtrace_mops_t *mops = &meta->dtm_mops;

	provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
	str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
	    provider->dofpv_strtab * dof->dofh_secsize);

	strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);

	/*
	 * Create the provider.
	 */
	dtrace_dofprov2hprov(&dhpv, provider, strtab);

	mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);

	meta->dtm_count--;
}

static void
dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
{
	uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
	dof_hdr_t *dof = (dof_hdr_t *)daddr;
	int i;

	ASSERT(MUTEX_HELD(&dtrace_meta_lock));

	for (i = 0; i < dof->dofh_secnum; i++) {
		dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
		    dof->dofh_secoff + i * dof->dofh_secsize);

		if (sec->dofs_type != DOF_SECT_PROVIDER)
			continue;

		dtrace_helper_provider_remove_one(dhp, sec, pid);
	}
}

/*
 * DTrace Meta Provider-to-Framework API Functions
 *
 * These functions implement the Meta Provider-to-Framework API, as described
 * in <sys/dtrace.h>.
 */
int
dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
    dtrace_meta_provider_id_t *idp)
{
	dtrace_meta_t *meta;
	dtrace_helpers_t *help, *next;
	int i;

	*idp = DTRACE_METAPROVNONE;

	/*
	 * We strictly don't need the name, but we hold onto it for
	 * debuggability. All hail error queues!
	 */
	if (name == NULL) {
		cmn_err(CE_WARN, "failed to register meta-provider: "
		    "invalid name");
		return (EINVAL);
	}

	if (mops == NULL ||
	    mops->dtms_create_probe == NULL ||
	    mops->dtms_provide_pid == NULL ||
	    mops->dtms_remove_pid == NULL) {
		cmn_err(CE_WARN, "failed to register meta-register %s: "
		    "invalid ops", name);
		return (EINVAL);
	}

	meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
	meta->dtm_mops = *mops;
	meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
	(void) strcpy(meta->dtm_name, name);
	meta->dtm_arg = arg;

	mutex_enter(&dtrace_meta_lock);
	mutex_enter(&dtrace_lock);

	if (dtrace_meta_pid != NULL) {
		mutex_exit(&dtrace_lock);
		mutex_exit(&dtrace_meta_lock);
		cmn_err(CE_WARN, "failed to register meta-register %s: "
		    "user-land meta-provider exists", name);
		kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
		kmem_free(meta, sizeof (dtrace_meta_t));
		return (EINVAL);
	}

	dtrace_meta_pid = meta;
	*idp = (dtrace_meta_provider_id_t)meta;

	/*
	 * If there are providers and probes ready to go, pass them
	 * off to the new meta provider now.
	 */

	help = dtrace_deferred_pid;
	dtrace_deferred_pid = NULL;

	mutex_exit(&dtrace_lock);

	while (help != NULL) {
		for (i = 0; i < help->dthps_nprovs; i++) {
			dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
			    help->dthps_pid);
		}

		next = help->dthps_next;
		help->dthps_next = NULL;
		help->dthps_prev = NULL;
		help->dthps_deferred = 0;
		help = next;
	}

	mutex_exit(&dtrace_meta_lock);

	return (0);
}

int
dtrace_meta_unregister(dtrace_meta_provider_id_t id)
{
	dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;

	mutex_enter(&dtrace_meta_lock);
	mutex_enter(&dtrace_lock);

	if (old == dtrace_meta_pid) {
		pp = &dtrace_meta_pid;
	} else {
		panic("attempt to unregister non-existent "
		    "dtrace meta-provider %p\n", (void *)old);
	}

	if (old->dtm_count != 0) {
		mutex_exit(&dtrace_lock);
		mutex_exit(&dtrace_meta_lock);
		return (EBUSY);
	}

	*pp = NULL;

	mutex_exit(&dtrace_lock);
	mutex_exit(&dtrace_meta_lock);

	kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
	kmem_free(old, sizeof (dtrace_meta_t));

	return (0);
}


/*
 * DTrace DIF Object Functions
 */
static int
dtrace_difo_err(uint_t pc, const char *format, ...)
{
	if (dtrace_err_verbose) {
		va_list alist;

		(void) uprintf("dtrace DIF object error: [%u]: ", pc);
		va_start(alist, format);
		(void) vuprintf(format, alist);
		va_end(alist);
	}

#ifdef DTRACE_ERRDEBUG
	dtrace_errdebug(format);
#endif
	return (1);
}

/*
 * Validate a DTrace DIF object by checking the IR instructions.  The following
 * rules are currently enforced by dtrace_difo_validate():
 *
 * 1. Each instruction must have a valid opcode
 * 2. Each register, string, variable, or subroutine reference must be valid
 * 3. No instruction can modify register %r0 (must be zero)
 * 4. All instruction reserved bits must be set to zero
 * 5. The last instruction must be a "ret" instruction
 * 6. All branch targets must reference a valid instruction _after_ the branch
 */
static int
dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
    cred_t *cr)
{
	int err = 0, i;
	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
	int kcheckload;
	uint_t pc;

	kcheckload = cr == NULL ||
	    (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;

	dp->dtdo_destructive = 0;

	for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
		dif_instr_t instr = dp->dtdo_buf[pc];

		uint_t r1 = DIF_INSTR_R1(instr);
		uint_t r2 = DIF_INSTR_R2(instr);
		uint_t rd = DIF_INSTR_RD(instr);
		uint_t rs = DIF_INSTR_RS(instr);
		uint_t label = DIF_INSTR_LABEL(instr);
		uint_t v = DIF_INSTR_VAR(instr);
		uint_t subr = DIF_INSTR_SUBR(instr);
		uint_t type = DIF_INSTR_TYPE(instr);
		uint_t op = DIF_INSTR_OP(instr);

		switch (op) {
		case DIF_OP_OR:
		case DIF_OP_XOR:
		case DIF_OP_AND:
		case DIF_OP_SLL:
		case DIF_OP_SRL:
		case DIF_OP_SRA:
		case DIF_OP_SUB:
		case DIF_OP_ADD:
		case DIF_OP_MUL:
		case DIF_OP_SDIV:
		case DIF_OP_UDIV:
		case DIF_OP_SREM:
		case DIF_OP_UREM:
		case DIF_OP_COPYS:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 >= nregs)
				err += efunc(pc, "invalid register %u\n", r2);
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_NOT:
		case DIF_OP_MOV:
		case DIF_OP_ALLOCS:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_LDSB:
		case DIF_OP_LDSH:
		case DIF_OP_LDSW:
		case DIF_OP_LDUB:
		case DIF_OP_LDUH:
		case DIF_OP_LDUW:
		case DIF_OP_LDX:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			if (kcheckload)
				dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
				    DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
			break;
		case DIF_OP_RLDSB:
		case DIF_OP_RLDSH:
		case DIF_OP_RLDSW:
		case DIF_OP_RLDUB:
		case DIF_OP_RLDUH:
		case DIF_OP_RLDUW:
		case DIF_OP_RLDX:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_ULDSB:
		case DIF_OP_ULDSH:
		case DIF_OP_ULDSW:
		case DIF_OP_ULDUB:
		case DIF_OP_ULDUH:
		case DIF_OP_ULDUW:
		case DIF_OP_ULDX:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_STB:
		case DIF_OP_STH:
		case DIF_OP_STW:
		case DIF_OP_STX:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to 0 address\n");
			break;
		case DIF_OP_CMP:
		case DIF_OP_SCMP:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 >= nregs)
				err += efunc(pc, "invalid register %u\n", r2);
			if (rd != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			break;
		case DIF_OP_TST:
			if (r1 >= nregs)
				err += efunc(pc, "invalid register %u\n", r1);
			if (r2 != 0 || rd != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			break;
		case DIF_OP_BA:
		case DIF_OP_BE:
		case DIF_OP_BNE:
		case DIF_OP_BG:
		case DIF_OP_BGU:
		case DIF_OP_BGE:
		case DIF_OP_BGEU:
		case DIF_OP_BL:
		case DIF_OP_BLU:
		case DIF_OP_BLE:
		case DIF_OP_BLEU:
			if (label >= dp->dtdo_len) {
				err += efunc(pc, "invalid branch target %u\n",
				    label);
			}
			if (label <= pc) {
				err += efunc(pc, "backward branch to %u\n",
				    label);
			}
			break;
		case DIF_OP_RET:
			if (r1 != 0 || r2 != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			break;
		case DIF_OP_NOP:
		case DIF_OP_POPTS:
		case DIF_OP_FLUSHTS:
			if (r1 != 0 || r2 != 0 || rd != 0)
				err += efunc(pc, "non-zero reserved bits\n");
			break;
		case DIF_OP_SETX:
			if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
				err += efunc(pc, "invalid integer ref %u\n",
				    DIF_INSTR_INTEGER(instr));
			}
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_SETS:
			if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
				err += efunc(pc, "invalid string ref %u\n",
				    DIF_INSTR_STRING(instr));
			}
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_LDGA:
		case DIF_OP_LDTA:
			if (r1 > DIF_VAR_ARRAY_MAX)
				err += efunc(pc, "invalid array %u\n", r1);
			if (r2 >= nregs)
				err += efunc(pc, "invalid register %u\n", r2);
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_LDGS:
		case DIF_OP_LDTS:
		case DIF_OP_LDLS:
		case DIF_OP_LDGAA:
		case DIF_OP_LDTAA:
			if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
				err += efunc(pc, "invalid variable %u\n", v);
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");
			break;
		case DIF_OP_STGS:
		case DIF_OP_STTS:
		case DIF_OP_STLS:
		case DIF_OP_STGAA:
		case DIF_OP_STTAA:
			if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
				err += efunc(pc, "invalid variable %u\n", v);
			if (rs >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			break;
		case DIF_OP_CALL:
			if (subr > DIF_SUBR_MAX)
				err += efunc(pc, "invalid subr %u\n", subr);
			if (rd >= nregs)
				err += efunc(pc, "invalid register %u\n", rd);
			if (rd == 0)
				err += efunc(pc, "cannot write to %r0\n");

			if (subr == DIF_SUBR_COPYOUT ||
			    subr == DIF_SUBR_COPYOUTSTR) {
				dp->dtdo_destructive = 1;
			}
			break;
		case DIF_OP_PUSHTR:
			if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
				err += efunc(pc, "invalid ref type %u\n", type);
			if (r2 >= nregs)
				err += efunc(pc, "invalid register %u\n", r2);
			if (rs >= nregs)
				err += efunc(pc, "invalid register %u\n", rs);
			break;
		case DIF_OP_PUSHTV:
			if (type != DIF_TYPE_CTF)
				err += efunc(pc, "invalid val type %u\n", type);
			if (r2 >= nregs)
				err += efunc(pc, "invalid register %u\n", r2);
			if (rs >= nregs)
				err += efunc(pc, "invalid register %u\n", rs);
			break;
		default:
			err += efunc(pc, "invalid opcode %u\n",
			    DIF_INSTR_OP(instr));
		}
	}

	if (dp->dtdo_len != 0 &&
	    DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
		err += efunc(dp->dtdo_len - 1,
		    "expected 'ret' as last DIF instruction\n");
	}

	if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF)) {
		/*
		 * If we're not returning by reference, the size must be either
		 * 0 or the size of one of the base types.
		 */
		switch (dp->dtdo_rtype.dtdt_size) {
		case 0:
		case sizeof (uint8_t):
		case sizeof (uint16_t):
		case sizeof (uint32_t):
		case sizeof (uint64_t):
			break;

		default:
			err += efunc(dp->dtdo_len - 1, "bad return size\n");
		}
	}

	for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
		dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
		dtrace_diftype_t *vt, *et;
		uint_t id, ndx;

		if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
		    v->dtdv_scope != DIFV_SCOPE_THREAD &&
		    v->dtdv_scope != DIFV_SCOPE_LOCAL) {
			err += efunc(i, "unrecognized variable scope %d\n",
			    v->dtdv_scope);
			break;
		}

		if (v->dtdv_kind != DIFV_KIND_ARRAY &&
		    v->dtdv_kind != DIFV_KIND_SCALAR) {
			err += efunc(i, "unrecognized variable type %d\n",
			    v->dtdv_kind);
			break;
		}

		if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
			err += efunc(i, "%d exceeds variable id limit\n", id);
			break;
		}

		if (id < DIF_VAR_OTHER_UBASE)
			continue;

		/*
		 * For user-defined variables, we need to check that this
		 * definition is identical to any previous definition that we
		 * encountered.
		 */
		ndx = id - DIF_VAR_OTHER_UBASE;

		switch (v->dtdv_scope) {
		case DIFV_SCOPE_GLOBAL:
			if (ndx < vstate->dtvs_nglobals) {
				dtrace_statvar_t *svar;

				if ((svar = vstate->dtvs_globals[ndx]) != NULL)
					existing = &svar->dtsv_var;
			}

			break;

		case DIFV_SCOPE_THREAD:
			if (ndx < vstate->dtvs_ntlocals)
				existing = &vstate->dtvs_tlocals[ndx];
			break;

		case DIFV_SCOPE_LOCAL:
			if (ndx < vstate->dtvs_nlocals) {
				dtrace_statvar_t *svar;

				if ((svar = vstate->dtvs_locals[ndx]) != NULL)
					existing = &svar->dtsv_var;
			}

			break;
		}

		vt = &v->dtdv_type;

		if (vt->dtdt_flags & DIF_TF_BYREF) {
			if (vt->dtdt_size == 0) {
				err += efunc(i, "zero-sized variable\n");
				break;
			}

			if (v->dtdv_scope == DIFV_SCOPE_GLOBAL &&
			    vt->dtdt_size > dtrace_global_maxsize) {
				err += efunc(i, "oversized by-ref global\n");
				break;
			}
		}

		if (existing == NULL || existing->dtdv_id == 0)
			continue;

		ASSERT(existing->dtdv_id == v->dtdv_id);
		ASSERT(existing->dtdv_scope == v->dtdv_scope);

		if (existing->dtdv_kind != v->dtdv_kind)
			err += efunc(i, "%d changed variable kind\n", id);

		et = &existing->dtdv_type;

		if (vt->dtdt_flags != et->dtdt_flags) {
			err += efunc(i, "%d changed variable type flags\n", id);
			break;
		}

		if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
			err += efunc(i, "%d changed variable type size\n", id);
			break;
		}
	}

	return (err);
}

/*
 * Validate a DTrace DIF object that it is to be used as a helper.  Helpers
 * are much more constrained than normal DIFOs.  Specifically, they may
 * not:
 *
 * 1. Make calls to subroutines other than copyin(), copyinstr() or
 *    miscellaneous string routines
 * 2. Access DTrace variables other than the args[] array, and the
 *    curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
 * 3. Have thread-local variables.
 * 4. Have dynamic variables.
 */
static int
dtrace_difo_validate_helper(dtrace_difo_t *dp)
{
	int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
	int err = 0;
	uint_t pc;

	for (pc = 0; pc < dp->dtdo_len; pc++) {
		dif_instr_t instr = dp->dtdo_buf[pc];

		uint_t v = DIF_INSTR_VAR(instr);
		uint_t subr = DIF_INSTR_SUBR(instr);
		uint_t op = DIF_INSTR_OP(instr);

		switch (op) {
		case DIF_OP_OR:
		case DIF_OP_XOR:
		case DIF_OP_AND:
		case DIF_OP_SLL:
		case DIF_OP_SRL:
		case DIF_OP_SRA:
		case DIF_OP_SUB:
		case DIF_OP_ADD:
		case DIF_OP_MUL:
		case DIF_OP_SDIV:
		case DIF_OP_UDIV:
		case DIF_OP_SREM:
		case DIF_OP_UREM:
		case DIF_OP_COPYS:
		case DIF_OP_NOT:
		case DIF_OP_MOV:
		case DIF_OP_RLDSB:
		case DIF_OP_RLDSH:
		case DIF_OP_RLDSW:
		case DIF_OP_RLDUB:
		case DIF_OP_RLDUH:
		case DIF_OP_RLDUW:
		case DIF_OP_RLDX:
		case DIF_OP_ULDSB:
		case DIF_OP_ULDSH:
		case DIF_OP_ULDSW:
		case DIF_OP_ULDUB:
		case DIF_OP_ULDUH:
		case DIF_OP_ULDUW:
		case DIF_OP_ULDX:
		case DIF_OP_STB:
		case DIF_OP_STH:
		case DIF_OP_STW:
		case DIF_OP_STX:
		case DIF_OP_ALLOCS:
		case DIF_OP_CMP:
		case DIF_OP_SCMP:
		case DIF_OP_TST:
		case DIF_OP_BA:
		case DIF_OP_BE:
		case DIF_OP_BNE:
		case DIF_OP_BG:
		case DIF_OP_BGU:
		case DIF_OP_BGE:
		case DIF_OP_BGEU:
		case DIF_OP_BL:
		case DIF_OP_BLU:
		case DIF_OP_BLE:
		case DIF_OP_BLEU:
		case DIF_OP_RET:
		case DIF_OP_NOP:
		case DIF_OP_POPTS:
		case DIF_OP_FLUSHTS:
		case DIF_OP_SETX:
		case DIF_OP_SETS:
		case DIF_OP_LDGA:
		case DIF_OP_LDLS:
		case DIF_OP_STGS:
		case DIF_OP_STLS:
		case DIF_OP_PUSHTR:
		case DIF_OP_PUSHTV:
			break;

		case DIF_OP_LDGS:
			if (v >= DIF_VAR_OTHER_UBASE)
				break;

			if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
				break;

			if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
			    v == DIF_VAR_PPID || v == DIF_VAR_TID ||
			    v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
			    v == DIF_VAR_UID || v == DIF_VAR_GID)
				break;

			err += efunc(pc, "illegal variable %u\n", v);
			break;

		case DIF_OP_LDTA:
		case DIF_OP_LDTS:
		case DIF_OP_LDGAA:
		case DIF_OP_LDTAA:
			err += efunc(pc, "illegal dynamic variable load\n");
			break;

		case DIF_OP_STTS:
		case DIF_OP_STGAA:
		case DIF_OP_STTAA:
			err += efunc(pc, "illegal dynamic variable store\n");
			break;

		case DIF_OP_CALL:
			if (subr == DIF_SUBR_ALLOCA ||
			    subr == DIF_SUBR_BCOPY ||
			    subr == DIF_SUBR_COPYIN ||
			    subr == DIF_SUBR_COPYINTO ||
			    subr == DIF_SUBR_COPYINSTR ||
			    subr == DIF_SUBR_INDEX ||
			    subr == DIF_SUBR_INET_NTOA ||
			    subr == DIF_SUBR_INET_NTOA6 ||
			    subr == DIF_SUBR_INET_NTOP ||
			    subr == DIF_SUBR_LLTOSTR ||
			    subr == DIF_SUBR_RINDEX ||
			    subr == DIF_SUBR_STRCHR ||
			    subr == DIF_SUBR_STRJOIN ||
			    subr == DIF_SUBR_STRRCHR ||
			    subr == DIF_SUBR_STRSTR ||
			    subr == DIF_SUBR_HTONS ||
			    subr == DIF_SUBR_HTONL ||
			    subr == DIF_SUBR_HTONLL</