fasttrap.c revision 179198
1/*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21
22/*
23 * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24 * Use is subject to license terms.
25 */
26
27#pragma ident	"%Z%%M%	%I%	%E% SMI"
28
29#include <sys/atomic.h>
30#include <sys/errno.h>
31#include <sys/stat.h>
32#include <sys/modctl.h>
33#include <sys/conf.h>
34#include <sys/systm.h>
35#include <sys/ddi.h>
36#include <sys/sunddi.h>
37#include <sys/cpuvar.h>
38#include <sys/kmem.h>
39#include <sys/strsubr.h>
40#include <sys/fasttrap.h>
41#include <sys/fasttrap_impl.h>
42#include <sys/fasttrap_isa.h>
43#include <sys/dtrace.h>
44#include <sys/dtrace_impl.h>
45#include <sys/sysmacros.h>
46#include <sys/proc.h>
47#include <sys/priv.h>
48#include <sys/policy.h>
49#include <util/qsort.h>
50
51/*
52 * User-Land Trap-Based Tracing
53 * ----------------------------
54 *
55 * The fasttrap provider allows DTrace consumers to instrument any user-level
56 * instruction to gather data; this includes probes with semantic
57 * signifigance like entry and return as well as simple offsets into the
58 * function. While the specific techniques used are very ISA specific, the
59 * methodology is generalizable to any architecture.
60 *
61 *
62 * The General Methodology
63 * -----------------------
64 *
65 * With the primary goal of tracing every user-land instruction and the
66 * limitation that we can't trust user space so don't want to rely on much
67 * information there, we begin by replacing the instructions we want to trace
68 * with trap instructions. Each instruction we overwrite is saved into a hash
69 * table keyed by process ID and pc address. When we enter the kernel due to
70 * this trap instruction, we need the effects of the replaced instruction to
71 * appear to have occurred before we proceed with the user thread's
72 * execution.
73 *
74 * Each user level thread is represented by a ulwp_t structure which is
75 * always easily accessible through a register. The most basic way to produce
76 * the effects of the instruction we replaced is to copy that instruction out
77 * to a bit of scratch space reserved in the user thread's ulwp_t structure
78 * (a sort of kernel-private thread local storage), set the PC to that
79 * scratch space and single step. When we reenter the kernel after single
80 * stepping the instruction we must then adjust the PC to point to what would
81 * normally be the next instruction. Of course, special care must be taken
82 * for branches and jumps, but these represent such a small fraction of any
83 * instruction set that writing the code to emulate these in the kernel is
84 * not too difficult.
85 *
86 * Return probes may require several tracepoints to trace every return site,
87 * and, conversely, each tracepoint may activate several probes (the entry
88 * and offset 0 probes, for example). To solve this muliplexing problem,
89 * tracepoints contain lists of probes to activate and probes contain lists
90 * of tracepoints to enable. If a probe is activated, it adds its ID to
91 * existing tracepoints or creates new ones as necessary.
92 *
93 * Most probes are activated _before_ the instruction is executed, but return
94 * probes are activated _after_ the effects of the last instruction of the
95 * function are visible. Return probes must be fired _after_ we have
96 * single-stepped the instruction whereas all other probes are fired
97 * beforehand.
98 *
99 *
100 * Lock Ordering
101 * -------------
102 *
103 * The lock ordering below -- both internally and with respect to the DTrace
104 * framework -- is a little tricky and bears some explanation. Each provider
105 * has a lock (ftp_mtx) that protects its members including reference counts
106 * for enabled probes (ftp_rcount), consumers actively creating probes
107 * (ftp_ccount) and USDT consumers (ftp_mcount); all three prevent a provider
108 * from being freed. A provider is looked up by taking the bucket lock for the
109 * provider hash table, and is returned with its lock held. The provider lock
110 * may be taken in functions invoked by the DTrace framework, but may not be
111 * held while calling functions in the DTrace framework.
112 *
113 * To ensure consistency over multiple calls to the DTrace framework, the
114 * creation lock (ftp_cmtx) should be held. Naturally, the creation lock may
115 * not be taken when holding the provider lock as that would create a cyclic
116 * lock ordering. In situations where one would naturally take the provider
117 * lock and then the creation lock, we instead up a reference count to prevent
118 * the provider from disappearing, drop the provider lock, and acquire the
119 * creation lock.
120 *
121 * Briefly:
122 * 	bucket lock before provider lock
123 *	DTrace before provider lock
124 *	creation lock before DTrace
125 *	never hold the provider lock and creation lock simultaneously
126 */
127
128static dev_info_t *fasttrap_devi;
129static dtrace_meta_provider_id_t fasttrap_meta_id;
130
131static timeout_id_t fasttrap_timeout;
132static kmutex_t fasttrap_cleanup_mtx;
133static uint_t fasttrap_cleanup_work;
134
135/*
136 * Generation count on modifications to the global tracepoint lookup table.
137 */
138static volatile uint64_t fasttrap_mod_gen;
139
140/*
141 * When the fasttrap provider is loaded, fasttrap_max is set to either
142 * FASTTRAP_MAX_DEFAULT or the value for fasttrap-max-probes in the
143 * fasttrap.conf file. Each time a probe is created, fasttrap_total is
144 * incremented by the number of tracepoints that may be associated with that
145 * probe; fasttrap_total is capped at fasttrap_max.
146 */
147#define	FASTTRAP_MAX_DEFAULT		250000
148static uint32_t fasttrap_max;
149static uint32_t fasttrap_total;
150
151
152#define	FASTTRAP_TPOINTS_DEFAULT_SIZE	0x4000
153#define	FASTTRAP_PROVIDERS_DEFAULT_SIZE	0x100
154#define	FASTTRAP_PROCS_DEFAULT_SIZE	0x100
155
156#define	FASTTRAP_PID_NAME		"pid"
157
158fasttrap_hash_t			fasttrap_tpoints;
159static fasttrap_hash_t		fasttrap_provs;
160static fasttrap_hash_t		fasttrap_procs;
161
162static uint64_t			fasttrap_pid_count;	/* pid ref count */
163static kmutex_t			fasttrap_count_mtx;	/* lock on ref count */
164
165#define	FASTTRAP_ENABLE_FAIL	1
166#define	FASTTRAP_ENABLE_PARTIAL	2
167
168static int fasttrap_tracepoint_enable(proc_t *, fasttrap_probe_t *, uint_t);
169static void fasttrap_tracepoint_disable(proc_t *, fasttrap_probe_t *, uint_t);
170
171static fasttrap_provider_t *fasttrap_provider_lookup(pid_t, const char *,
172    const dtrace_pattr_t *);
173static void fasttrap_provider_retire(pid_t, const char *, int);
174static void fasttrap_provider_free(fasttrap_provider_t *);
175
176static fasttrap_proc_t *fasttrap_proc_lookup(pid_t);
177static void fasttrap_proc_release(fasttrap_proc_t *);
178
179#define	FASTTRAP_PROVS_INDEX(pid, name) \
180	((fasttrap_hash_str(name) + (pid)) & fasttrap_provs.fth_mask)
181
182#define	FASTTRAP_PROCS_INDEX(pid) ((pid) & fasttrap_procs.fth_mask)
183
184static int
185fasttrap_highbit(ulong_t i)
186{
187	int h = 1;
188
189	if (i == 0)
190		return (0);
191#ifdef _LP64
192	if (i & 0xffffffff00000000ul) {
193		h += 32; i >>= 32;
194	}
195#endif
196	if (i & 0xffff0000) {
197		h += 16; i >>= 16;
198	}
199	if (i & 0xff00) {
200		h += 8; i >>= 8;
201	}
202	if (i & 0xf0) {
203		h += 4; i >>= 4;
204	}
205	if (i & 0xc) {
206		h += 2; i >>= 2;
207	}
208	if (i & 0x2) {
209		h += 1;
210	}
211	return (h);
212}
213
214static uint_t
215fasttrap_hash_str(const char *p)
216{
217	unsigned int g;
218	uint_t hval = 0;
219
220	while (*p) {
221		hval = (hval << 4) + *p++;
222		if ((g = (hval & 0xf0000000)) != 0)
223			hval ^= g >> 24;
224		hval &= ~g;
225	}
226	return (hval);
227}
228
229void
230fasttrap_sigtrap(proc_t *p, kthread_t *t, uintptr_t pc)
231{
232	sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP);
233
234	sqp->sq_info.si_signo = SIGTRAP;
235	sqp->sq_info.si_code = TRAP_DTRACE;
236	sqp->sq_info.si_addr = (caddr_t)pc;
237
238	mutex_enter(&p->p_lock);
239	sigaddqa(p, t, sqp);
240	mutex_exit(&p->p_lock);
241
242	if (t != NULL)
243		aston(t);
244}
245
246/*
247 * This function ensures that no threads are actively using the memory
248 * associated with probes that were formerly live.
249 */
250static void
251fasttrap_mod_barrier(uint64_t gen)
252{
253	int i;
254
255	if (gen < fasttrap_mod_gen)
256		return;
257
258	fasttrap_mod_gen++;
259
260	for (i = 0; i < NCPU; i++) {
261		mutex_enter(&cpu_core[i].cpuc_pid_lock);
262		mutex_exit(&cpu_core[i].cpuc_pid_lock);
263	}
264}
265
266/*
267 * This is the timeout's callback for cleaning up the providers and their
268 * probes.
269 */
270/*ARGSUSED*/
271static void
272fasttrap_pid_cleanup_cb(void *data)
273{
274	fasttrap_provider_t **fpp, *fp;
275	fasttrap_bucket_t *bucket;
276	dtrace_provider_id_t provid;
277	int i, later;
278
279	static volatile int in = 0;
280	ASSERT(in == 0);
281	in = 1;
282
283	mutex_enter(&fasttrap_cleanup_mtx);
284	while (fasttrap_cleanup_work) {
285		fasttrap_cleanup_work = 0;
286		mutex_exit(&fasttrap_cleanup_mtx);
287
288		later = 0;
289
290		/*
291		 * Iterate over all the providers trying to remove the marked
292		 * ones. If a provider is marked but not retired, we just
293		 * have to take a crack at removing it -- it's no big deal if
294		 * we can't.
295		 */
296		for (i = 0; i < fasttrap_provs.fth_nent; i++) {
297			bucket = &fasttrap_provs.fth_table[i];
298			mutex_enter(&bucket->ftb_mtx);
299			fpp = (fasttrap_provider_t **)&bucket->ftb_data;
300
301			while ((fp = *fpp) != NULL) {
302				if (!fp->ftp_marked) {
303					fpp = &fp->ftp_next;
304					continue;
305				}
306
307				mutex_enter(&fp->ftp_mtx);
308
309				/*
310				 * If this provider has consumers actively
311				 * creating probes (ftp_ccount) or is a USDT
312				 * provider (ftp_mcount), we can't unregister
313				 * or even condense.
314				 */
315				if (fp->ftp_ccount != 0 ||
316				    fp->ftp_mcount != 0) {
317					mutex_exit(&fp->ftp_mtx);
318					fp->ftp_marked = 0;
319					continue;
320				}
321
322				if (!fp->ftp_retired || fp->ftp_rcount != 0)
323					fp->ftp_marked = 0;
324
325				mutex_exit(&fp->ftp_mtx);
326
327				/*
328				 * If we successfully unregister this
329				 * provider we can remove it from the hash
330				 * chain and free the memory. If our attempt
331				 * to unregister fails and this is a retired
332				 * provider, increment our flag to try again
333				 * pretty soon. If we've consumed more than
334				 * half of our total permitted number of
335				 * probes call dtrace_condense() to try to
336				 * clean out the unenabled probes.
337				 */
338				provid = fp->ftp_provid;
339				if (dtrace_unregister(provid) != 0) {
340					if (fasttrap_total > fasttrap_max / 2)
341						(void) dtrace_condense(provid);
342					later += fp->ftp_marked;
343					fpp = &fp->ftp_next;
344				} else {
345					*fpp = fp->ftp_next;
346					fasttrap_provider_free(fp);
347				}
348			}
349			mutex_exit(&bucket->ftb_mtx);
350		}
351
352		mutex_enter(&fasttrap_cleanup_mtx);
353	}
354
355	ASSERT(fasttrap_timeout != 0);
356
357	/*
358	 * If we were unable to remove a retired provider, try again after
359	 * a second. This situation can occur in certain circumstances where
360	 * providers cannot be unregistered even though they have no probes
361	 * enabled because of an execution of dtrace -l or something similar.
362	 * If the timeout has been disabled (set to 1 because we're trying
363	 * to detach), we set fasttrap_cleanup_work to ensure that we'll
364	 * get a chance to do that work if and when the timeout is reenabled
365	 * (if detach fails).
366	 */
367	if (later > 0 && fasttrap_timeout != (timeout_id_t)1)
368		fasttrap_timeout = timeout(&fasttrap_pid_cleanup_cb, NULL, hz);
369	else if (later > 0)
370		fasttrap_cleanup_work = 1;
371	else
372		fasttrap_timeout = 0;
373
374	mutex_exit(&fasttrap_cleanup_mtx);
375	in = 0;
376}
377
378/*
379 * Activates the asynchronous cleanup mechanism.
380 */
381static void
382fasttrap_pid_cleanup(void)
383{
384	mutex_enter(&fasttrap_cleanup_mtx);
385	fasttrap_cleanup_work = 1;
386	if (fasttrap_timeout == 0)
387		fasttrap_timeout = timeout(&fasttrap_pid_cleanup_cb, NULL, 1);
388	mutex_exit(&fasttrap_cleanup_mtx);
389}
390
391/*
392 * This is called from cfork() via dtrace_fasttrap_fork(). The child
393 * process's address space is (roughly) a copy of the parent process's so
394 * we have to remove all the instrumentation we had previously enabled in the
395 * parent.
396 */
397static void
398fasttrap_fork(proc_t *p, proc_t *cp)
399{
400	pid_t ppid = p->p_pid;
401	int i;
402
403	ASSERT(curproc == p);
404	ASSERT(p->p_proc_flag & P_PR_LOCK);
405	ASSERT(p->p_dtrace_count > 0);
406	ASSERT(cp->p_dtrace_count == 0);
407
408	/*
409	 * This would be simpler and faster if we maintained per-process
410	 * hash tables of enabled tracepoints. It could, however, potentially
411	 * slow down execution of a tracepoint since we'd need to go
412	 * through two levels of indirection. In the future, we should
413	 * consider either maintaining per-process ancillary lists of
414	 * enabled tracepoints or hanging a pointer to a per-process hash
415	 * table of enabled tracepoints off the proc structure.
416	 */
417
418	/*
419	 * We don't have to worry about the child process disappearing
420	 * because we're in fork().
421	 */
422	mutex_enter(&cp->p_lock);
423	sprlock_proc(cp);
424	mutex_exit(&cp->p_lock);
425
426	/*
427	 * Iterate over every tracepoint looking for ones that belong to the
428	 * parent process, and remove each from the child process.
429	 */
430	for (i = 0; i < fasttrap_tpoints.fth_nent; i++) {
431		fasttrap_tracepoint_t *tp;
432		fasttrap_bucket_t *bucket = &fasttrap_tpoints.fth_table[i];
433
434		mutex_enter(&bucket->ftb_mtx);
435		for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
436			if (tp->ftt_pid == ppid &&
437			    tp->ftt_proc->ftpc_acount != 0) {
438				int ret = fasttrap_tracepoint_remove(cp, tp);
439				ASSERT(ret == 0);
440
441				/*
442				 * The count of active providers can only be
443				 * decremented (i.e. to zero) during exec,
444				 * exit, and removal of a meta provider so it
445				 * should be impossible to drop the count
446				 * mid-fork.
447				 */
448				ASSERT(tp->ftt_proc->ftpc_acount != 0);
449			}
450		}
451		mutex_exit(&bucket->ftb_mtx);
452	}
453
454	mutex_enter(&cp->p_lock);
455	sprunlock(cp);
456}
457
458/*
459 * This is called from proc_exit() or from exec_common() if p_dtrace_probes
460 * is set on the proc structure to indicate that there is a pid provider
461 * associated with this process.
462 */
463static void
464fasttrap_exec_exit(proc_t *p)
465{
466	ASSERT(p == curproc);
467	ASSERT(MUTEX_HELD(&p->p_lock));
468
469	mutex_exit(&p->p_lock);
470
471	/*
472	 * We clean up the pid provider for this process here; user-land
473	 * static probes are handled by the meta-provider remove entry point.
474	 */
475	fasttrap_provider_retire(p->p_pid, FASTTRAP_PID_NAME, 0);
476
477	mutex_enter(&p->p_lock);
478}
479
480
481/*ARGSUSED*/
482static void
483fasttrap_pid_provide(void *arg, const dtrace_probedesc_t *desc)
484{
485	/*
486	 * There are no "default" pid probes.
487	 */
488}
489
490static int
491fasttrap_tracepoint_enable(proc_t *p, fasttrap_probe_t *probe, uint_t index)
492{
493	fasttrap_tracepoint_t *tp, *new_tp = NULL;
494	fasttrap_bucket_t *bucket;
495	fasttrap_id_t *id;
496	pid_t pid;
497	uintptr_t pc;
498
499	ASSERT(index < probe->ftp_ntps);
500
501	pid = probe->ftp_pid;
502	pc = probe->ftp_tps[index].fit_tp->ftt_pc;
503	id = &probe->ftp_tps[index].fit_id;
504
505	ASSERT(probe->ftp_tps[index].fit_tp->ftt_pid == pid);
506
507	ASSERT(!(p->p_flag & SVFORK));
508
509	/*
510	 * Before we make any modifications, make sure we've imposed a barrier
511	 * on the generation in which this probe was last modified.
512	 */
513	fasttrap_mod_barrier(probe->ftp_gen);
514
515	bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
516
517	/*
518	 * If the tracepoint has already been enabled, just add our id to the
519	 * list of interested probes. This may be our second time through
520	 * this path in which case we'll have constructed the tracepoint we'd
521	 * like to install. If we can't find a match, and have an allocated
522	 * tracepoint ready to go, enable that one now.
523	 *
524	 * A tracepoint whose process is defunct is also considered defunct.
525	 */
526again:
527	mutex_enter(&bucket->ftb_mtx);
528	for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
529		/*
530		 * Note that it's safe to access the active count on the
531		 * associated proc structure because we know that at least one
532		 * provider (this one) will still be around throughout this
533		 * operation.
534		 */
535		if (tp->ftt_pid != pid || tp->ftt_pc != pc ||
536		    tp->ftt_proc->ftpc_acount == 0)
537			continue;
538
539		/*
540		 * Now that we've found a matching tracepoint, it would be
541		 * a decent idea to confirm that the tracepoint is still
542		 * enabled and the trap instruction hasn't been overwritten.
543		 * Since this is a little hairy, we'll punt for now.
544		 */
545
546		/*
547		 * This can't be the first interested probe. We don't have
548		 * to worry about another thread being in the midst of
549		 * deleting this tracepoint (which would be the only valid
550		 * reason for a tracepoint to have no interested probes)
551		 * since we're holding P_PR_LOCK for this process.
552		 */
553		ASSERT(tp->ftt_ids != NULL || tp->ftt_retids != NULL);
554
555		switch (id->fti_ptype) {
556		case DTFTP_ENTRY:
557		case DTFTP_OFFSETS:
558		case DTFTP_IS_ENABLED:
559			id->fti_next = tp->ftt_ids;
560			membar_producer();
561			tp->ftt_ids = id;
562			membar_producer();
563			break;
564
565		case DTFTP_RETURN:
566		case DTFTP_POST_OFFSETS:
567			id->fti_next = tp->ftt_retids;
568			membar_producer();
569			tp->ftt_retids = id;
570			membar_producer();
571			break;
572
573		default:
574			ASSERT(0);
575		}
576
577		mutex_exit(&bucket->ftb_mtx);
578
579		if (new_tp != NULL) {
580			new_tp->ftt_ids = NULL;
581			new_tp->ftt_retids = NULL;
582		}
583
584		return (0);
585	}
586
587	/*
588	 * If we have a good tracepoint ready to go, install it now while
589	 * we have the lock held and no one can screw with us.
590	 */
591	if (new_tp != NULL) {
592		int rc = 0;
593
594		new_tp->ftt_next = bucket->ftb_data;
595		membar_producer();
596		bucket->ftb_data = new_tp;
597		membar_producer();
598		mutex_exit(&bucket->ftb_mtx);
599
600		/*
601		 * Activate the tracepoint in the ISA-specific manner.
602		 * If this fails, we need to report the failure, but
603		 * indicate that this tracepoint must still be disabled
604		 * by calling fasttrap_tracepoint_disable().
605		 */
606		if (fasttrap_tracepoint_install(p, new_tp) != 0)
607			rc = FASTTRAP_ENABLE_PARTIAL;
608
609		/*
610		 * Increment the count of the number of tracepoints active in
611		 * the victim process.
612		 */
613		ASSERT(p->p_proc_flag & P_PR_LOCK);
614		p->p_dtrace_count++;
615
616		return (rc);
617	}
618
619	mutex_exit(&bucket->ftb_mtx);
620
621	/*
622	 * Initialize the tracepoint that's been preallocated with the probe.
623	 */
624	new_tp = probe->ftp_tps[index].fit_tp;
625
626	ASSERT(new_tp->ftt_pid == pid);
627	ASSERT(new_tp->ftt_pc == pc);
628	ASSERT(new_tp->ftt_proc == probe->ftp_prov->ftp_proc);
629	ASSERT(new_tp->ftt_ids == NULL);
630	ASSERT(new_tp->ftt_retids == NULL);
631
632	switch (id->fti_ptype) {
633	case DTFTP_ENTRY:
634	case DTFTP_OFFSETS:
635	case DTFTP_IS_ENABLED:
636		id->fti_next = NULL;
637		new_tp->ftt_ids = id;
638		break;
639
640	case DTFTP_RETURN:
641	case DTFTP_POST_OFFSETS:
642		id->fti_next = NULL;
643		new_tp->ftt_retids = id;
644		break;
645
646	default:
647		ASSERT(0);
648	}
649
650	/*
651	 * If the ISA-dependent initialization goes to plan, go back to the
652	 * beginning and try to install this freshly made tracepoint.
653	 */
654	if (fasttrap_tracepoint_init(p, new_tp, pc, id->fti_ptype) == 0)
655		goto again;
656
657	new_tp->ftt_ids = NULL;
658	new_tp->ftt_retids = NULL;
659
660	return (FASTTRAP_ENABLE_FAIL);
661}
662
663static void
664fasttrap_tracepoint_disable(proc_t *p, fasttrap_probe_t *probe, uint_t index)
665{
666	fasttrap_bucket_t *bucket;
667	fasttrap_provider_t *provider = probe->ftp_prov;
668	fasttrap_tracepoint_t **pp, *tp;
669	fasttrap_id_t *id, **idp;
670	pid_t pid;
671	uintptr_t pc;
672
673	ASSERT(index < probe->ftp_ntps);
674
675	pid = probe->ftp_pid;
676	pc = probe->ftp_tps[index].fit_tp->ftt_pc;
677	id = &probe->ftp_tps[index].fit_id;
678
679	ASSERT(probe->ftp_tps[index].fit_tp->ftt_pid == pid);
680
681	/*
682	 * Find the tracepoint and make sure that our id is one of the
683	 * ones registered with it.
684	 */
685	bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
686	mutex_enter(&bucket->ftb_mtx);
687	for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
688		if (tp->ftt_pid == pid && tp->ftt_pc == pc &&
689		    tp->ftt_proc == provider->ftp_proc)
690			break;
691	}
692
693	/*
694	 * If we somehow lost this tracepoint, we're in a world of hurt.
695	 */
696	ASSERT(tp != NULL);
697
698	switch (id->fti_ptype) {
699	case DTFTP_ENTRY:
700	case DTFTP_OFFSETS:
701	case DTFTP_IS_ENABLED:
702		ASSERT(tp->ftt_ids != NULL);
703		idp = &tp->ftt_ids;
704		break;
705
706	case DTFTP_RETURN:
707	case DTFTP_POST_OFFSETS:
708		ASSERT(tp->ftt_retids != NULL);
709		idp = &tp->ftt_retids;
710		break;
711
712	default:
713		ASSERT(0);
714	}
715
716	while ((*idp)->fti_probe != probe) {
717		idp = &(*idp)->fti_next;
718		ASSERT(*idp != NULL);
719	}
720
721	id = *idp;
722	*idp = id->fti_next;
723	membar_producer();
724
725	ASSERT(id->fti_probe == probe);
726
727	/*
728	 * If there are other registered enablings of this tracepoint, we're
729	 * all done, but if this was the last probe assocated with this
730	 * this tracepoint, we need to remove and free it.
731	 */
732	if (tp->ftt_ids != NULL || tp->ftt_retids != NULL) {
733
734		/*
735		 * If the current probe's tracepoint is in use, swap it
736		 * for an unused tracepoint.
737		 */
738		if (tp == probe->ftp_tps[index].fit_tp) {
739			fasttrap_probe_t *tmp_probe;
740			fasttrap_tracepoint_t **tmp_tp;
741			uint_t tmp_index;
742
743			if (tp->ftt_ids != NULL) {
744				tmp_probe = tp->ftt_ids->fti_probe;
745				/* LINTED - alignment */
746				tmp_index = FASTTRAP_ID_INDEX(tp->ftt_ids);
747				tmp_tp = &tmp_probe->ftp_tps[tmp_index].fit_tp;
748			} else {
749				tmp_probe = tp->ftt_retids->fti_probe;
750				/* LINTED - alignment */
751				tmp_index = FASTTRAP_ID_INDEX(tp->ftt_retids);
752				tmp_tp = &tmp_probe->ftp_tps[tmp_index].fit_tp;
753			}
754
755			ASSERT(*tmp_tp != NULL);
756			ASSERT(*tmp_tp != probe->ftp_tps[index].fit_tp);
757			ASSERT((*tmp_tp)->ftt_ids == NULL);
758			ASSERT((*tmp_tp)->ftt_retids == NULL);
759
760			probe->ftp_tps[index].fit_tp = *tmp_tp;
761			*tmp_tp = tp;
762		}
763
764		mutex_exit(&bucket->ftb_mtx);
765
766		/*
767		 * Tag the modified probe with the generation in which it was
768		 * changed.
769		 */
770		probe->ftp_gen = fasttrap_mod_gen;
771		return;
772	}
773
774	mutex_exit(&bucket->ftb_mtx);
775
776	/*
777	 * We can't safely remove the tracepoint from the set of active
778	 * tracepoints until we've actually removed the fasttrap instruction
779	 * from the process's text. We can, however, operate on this
780	 * tracepoint secure in the knowledge that no other thread is going to
781	 * be looking at it since we hold P_PR_LOCK on the process if it's
782	 * live or we hold the provider lock on the process if it's dead and
783	 * gone.
784	 */
785
786	/*
787	 * We only need to remove the actual instruction if we're looking
788	 * at an existing process
789	 */
790	if (p != NULL) {
791		/*
792		 * If we fail to restore the instruction we need to kill
793		 * this process since it's in a completely unrecoverable
794		 * state.
795		 */
796		if (fasttrap_tracepoint_remove(p, tp) != 0)
797			fasttrap_sigtrap(p, NULL, pc);
798
799		/*
800		 * Decrement the count of the number of tracepoints active
801		 * in the victim process.
802		 */
803		ASSERT(p->p_proc_flag & P_PR_LOCK);
804		p->p_dtrace_count--;
805	}
806
807	/*
808	 * Remove the probe from the hash table of active tracepoints.
809	 */
810	mutex_enter(&bucket->ftb_mtx);
811	pp = (fasttrap_tracepoint_t **)&bucket->ftb_data;
812	ASSERT(*pp != NULL);
813	while (*pp != tp) {
814		pp = &(*pp)->ftt_next;
815		ASSERT(*pp != NULL);
816	}
817
818	*pp = tp->ftt_next;
819	membar_producer();
820
821	mutex_exit(&bucket->ftb_mtx);
822
823	/*
824	 * Tag the modified probe with the generation in which it was changed.
825	 */
826	probe->ftp_gen = fasttrap_mod_gen;
827}
828
829static void
830fasttrap_enable_callbacks(void)
831{
832	/*
833	 * We don't have to play the rw lock game here because we're
834	 * providing something rather than taking something away --
835	 * we can be sure that no threads have tried to follow this
836	 * function pointer yet.
837	 */
838	mutex_enter(&fasttrap_count_mtx);
839	if (fasttrap_pid_count == 0) {
840		ASSERT(dtrace_pid_probe_ptr == NULL);
841		ASSERT(dtrace_return_probe_ptr == NULL);
842		dtrace_pid_probe_ptr = &fasttrap_pid_probe;
843		dtrace_return_probe_ptr = &fasttrap_return_probe;
844	}
845	ASSERT(dtrace_pid_probe_ptr == &fasttrap_pid_probe);
846	ASSERT(dtrace_return_probe_ptr == &fasttrap_return_probe);
847	fasttrap_pid_count++;
848	mutex_exit(&fasttrap_count_mtx);
849}
850
851static void
852fasttrap_disable_callbacks(void)
853{
854	ASSERT(MUTEX_HELD(&cpu_lock));
855
856	mutex_enter(&fasttrap_count_mtx);
857	ASSERT(fasttrap_pid_count > 0);
858	fasttrap_pid_count--;
859	if (fasttrap_pid_count == 0) {
860		cpu_t *cur, *cpu = CPU;
861
862		for (cur = cpu->cpu_next_onln; cur != cpu;
863		    cur = cur->cpu_next_onln) {
864			rw_enter(&cur->cpu_ft_lock, RW_WRITER);
865		}
866
867		dtrace_pid_probe_ptr = NULL;
868		dtrace_return_probe_ptr = NULL;
869
870		for (cur = cpu->cpu_next_onln; cur != cpu;
871		    cur = cur->cpu_next_onln) {
872			rw_exit(&cur->cpu_ft_lock);
873		}
874	}
875	mutex_exit(&fasttrap_count_mtx);
876}
877
878/*ARGSUSED*/
879static void
880fasttrap_pid_enable(void *arg, dtrace_id_t id, void *parg)
881{
882	fasttrap_probe_t *probe = parg;
883	proc_t *p;
884	int i, rc;
885
886	ASSERT(probe != NULL);
887	ASSERT(!probe->ftp_enabled);
888	ASSERT(id == probe->ftp_id);
889	ASSERT(MUTEX_HELD(&cpu_lock));
890
891	/*
892	 * Increment the count of enabled probes on this probe's provider;
893	 * the provider can't go away while the probe still exists. We
894	 * must increment this even if we aren't able to properly enable
895	 * this probe.
896	 */
897	mutex_enter(&probe->ftp_prov->ftp_mtx);
898	probe->ftp_prov->ftp_rcount++;
899	mutex_exit(&probe->ftp_prov->ftp_mtx);
900
901	/*
902	 * If this probe's provider is retired (meaning it was valid in a
903	 * previously exec'ed incarnation of this address space), bail out. The
904	 * provider can't go away while we're in this code path.
905	 */
906	if (probe->ftp_prov->ftp_retired)
907		return;
908
909	/*
910	 * If we can't find the process, it may be that we're in the context of
911	 * a fork in which the traced process is being born and we're copying
912	 * USDT probes. Otherwise, the process is gone so bail.
913	 */
914	if ((p = sprlock(probe->ftp_pid)) == NULL) {
915		if ((curproc->p_flag & SFORKING) == 0)
916			return;
917
918		mutex_enter(&pidlock);
919		p = prfind(probe->ftp_pid);
920
921		/*
922		 * Confirm that curproc is indeed forking the process in which
923		 * we're trying to enable probes.
924		 */
925		ASSERT(p != NULL);
926		ASSERT(p->p_parent == curproc);
927		ASSERT(p->p_stat == SIDL);
928
929		mutex_enter(&p->p_lock);
930		mutex_exit(&pidlock);
931
932		sprlock_proc(p);
933	}
934
935	ASSERT(!(p->p_flag & SVFORK));
936	mutex_exit(&p->p_lock);
937
938	/*
939	 * We have to enable the trap entry point before any user threads have
940	 * the chance to execute the trap instruction we're about to place
941	 * in their process's text.
942	 */
943	fasttrap_enable_callbacks();
944
945	/*
946	 * Enable all the tracepoints and add this probe's id to each
947	 * tracepoint's list of active probes.
948	 */
949	for (i = 0; i < probe->ftp_ntps; i++) {
950		if ((rc = fasttrap_tracepoint_enable(p, probe, i)) != 0) {
951			/*
952			 * If enabling the tracepoint failed completely,
953			 * we don't have to disable it; if the failure
954			 * was only partial we must disable it.
955			 */
956			if (rc == FASTTRAP_ENABLE_FAIL)
957				i--;
958			else
959				ASSERT(rc == FASTTRAP_ENABLE_PARTIAL);
960
961			/*
962			 * Back up and pull out all the tracepoints we've
963			 * created so far for this probe.
964			 */
965			while (i >= 0) {
966				fasttrap_tracepoint_disable(p, probe, i);
967				i--;
968			}
969
970			mutex_enter(&p->p_lock);
971			sprunlock(p);
972
973			/*
974			 * Since we're not actually enabling this probe,
975			 * drop our reference on the trap table entry.
976			 */
977			fasttrap_disable_callbacks();
978			return;
979		}
980	}
981
982	mutex_enter(&p->p_lock);
983	sprunlock(p);
984
985	probe->ftp_enabled = 1;
986}
987
988/*ARGSUSED*/
989static void
990fasttrap_pid_disable(void *arg, dtrace_id_t id, void *parg)
991{
992	fasttrap_probe_t *probe = parg;
993	fasttrap_provider_t *provider = probe->ftp_prov;
994	proc_t *p;
995	int i, whack = 0;
996
997	ASSERT(id == probe->ftp_id);
998
999	/*
1000	 * We won't be able to acquire a /proc-esque lock on the process
1001	 * iff the process is dead and gone. In this case, we rely on the
1002	 * provider lock as a point of mutual exclusion to prevent other
1003	 * DTrace consumers from disabling this probe.
1004	 */
1005	if ((p = sprlock(probe->ftp_pid)) != NULL) {
1006		ASSERT(!(p->p_flag & SVFORK));
1007		mutex_exit(&p->p_lock);
1008	}
1009
1010	mutex_enter(&provider->ftp_mtx);
1011
1012	/*
1013	 * Disable all the associated tracepoints (for fully enabled probes).
1014	 */
1015	if (probe->ftp_enabled) {
1016		for (i = 0; i < probe->ftp_ntps; i++) {
1017			fasttrap_tracepoint_disable(p, probe, i);
1018		}
1019	}
1020
1021	ASSERT(provider->ftp_rcount > 0);
1022	provider->ftp_rcount--;
1023
1024	if (p != NULL) {
1025		/*
1026		 * Even though we may not be able to remove it entirely, we
1027		 * mark this retired provider to get a chance to remove some
1028		 * of the associated probes.
1029		 */
1030		if (provider->ftp_retired && !provider->ftp_marked)
1031			whack = provider->ftp_marked = 1;
1032		mutex_exit(&provider->ftp_mtx);
1033
1034		mutex_enter(&p->p_lock);
1035		sprunlock(p);
1036	} else {
1037		/*
1038		 * If the process is dead, we're just waiting for the
1039		 * last probe to be disabled to be able to free it.
1040		 */
1041		if (provider->ftp_rcount == 0 && !provider->ftp_marked)
1042			whack = provider->ftp_marked = 1;
1043		mutex_exit(&provider->ftp_mtx);
1044	}
1045
1046	if (whack)
1047		fasttrap_pid_cleanup();
1048
1049	if (!probe->ftp_enabled)
1050		return;
1051
1052	probe->ftp_enabled = 0;
1053
1054	ASSERT(MUTEX_HELD(&cpu_lock));
1055	fasttrap_disable_callbacks();
1056}
1057
1058/*ARGSUSED*/
1059static void
1060fasttrap_pid_getargdesc(void *arg, dtrace_id_t id, void *parg,
1061    dtrace_argdesc_t *desc)
1062{
1063	fasttrap_probe_t *probe = parg;
1064	char *str;
1065	int i, ndx;
1066
1067	desc->dtargd_native[0] = '\0';
1068	desc->dtargd_xlate[0] = '\0';
1069
1070	if (probe->ftp_prov->ftp_retired != 0 ||
1071	    desc->dtargd_ndx >= probe->ftp_nargs) {
1072		desc->dtargd_ndx = DTRACE_ARGNONE;
1073		return;
1074	}
1075
1076	ndx = (probe->ftp_argmap != NULL) ?
1077	    probe->ftp_argmap[desc->dtargd_ndx] : desc->dtargd_ndx;
1078
1079	str = probe->ftp_ntypes;
1080	for (i = 0; i < ndx; i++) {
1081		str += strlen(str) + 1;
1082	}
1083
1084	ASSERT(strlen(str + 1) < sizeof (desc->dtargd_native));
1085	(void) strcpy(desc->dtargd_native, str);
1086
1087	if (probe->ftp_xtypes == NULL)
1088		return;
1089
1090	str = probe->ftp_xtypes;
1091	for (i = 0; i < desc->dtargd_ndx; i++) {
1092		str += strlen(str) + 1;
1093	}
1094
1095	ASSERT(strlen(str + 1) < sizeof (desc->dtargd_xlate));
1096	(void) strcpy(desc->dtargd_xlate, str);
1097}
1098
1099/*ARGSUSED*/
1100static void
1101fasttrap_pid_destroy(void *arg, dtrace_id_t id, void *parg)
1102{
1103	fasttrap_probe_t *probe = parg;
1104	int i;
1105	size_t size;
1106
1107	ASSERT(probe != NULL);
1108	ASSERT(!probe->ftp_enabled);
1109	ASSERT(fasttrap_total >= probe->ftp_ntps);
1110
1111	atomic_add_32(&fasttrap_total, -probe->ftp_ntps);
1112	size = offsetof(fasttrap_probe_t, ftp_tps[probe->ftp_ntps]);
1113
1114	if (probe->ftp_gen + 1 >= fasttrap_mod_gen)
1115		fasttrap_mod_barrier(probe->ftp_gen);
1116
1117	for (i = 0; i < probe->ftp_ntps; i++) {
1118		kmem_free(probe->ftp_tps[i].fit_tp,
1119		    sizeof (fasttrap_tracepoint_t));
1120	}
1121
1122	kmem_free(probe, size);
1123}
1124
1125
1126static const dtrace_pattr_t pid_attr = {
1127{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1128{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1129{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1130{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1131{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1132};
1133
1134static dtrace_pops_t pid_pops = {
1135	fasttrap_pid_provide,
1136	NULL,
1137	fasttrap_pid_enable,
1138	fasttrap_pid_disable,
1139	NULL,
1140	NULL,
1141	fasttrap_pid_getargdesc,
1142	fasttrap_pid_getarg,
1143	NULL,
1144	fasttrap_pid_destroy
1145};
1146
1147static dtrace_pops_t usdt_pops = {
1148	fasttrap_pid_provide,
1149	NULL,
1150	fasttrap_pid_enable,
1151	fasttrap_pid_disable,
1152	NULL,
1153	NULL,
1154	fasttrap_pid_getargdesc,
1155	fasttrap_usdt_getarg,
1156	NULL,
1157	fasttrap_pid_destroy
1158};
1159
1160static fasttrap_proc_t *
1161fasttrap_proc_lookup(pid_t pid)
1162{
1163	fasttrap_bucket_t *bucket;
1164	fasttrap_proc_t *fprc, *new_fprc;
1165
1166	bucket = &fasttrap_procs.fth_table[FASTTRAP_PROCS_INDEX(pid)];
1167	mutex_enter(&bucket->ftb_mtx);
1168
1169	for (fprc = bucket->ftb_data; fprc != NULL; fprc = fprc->ftpc_next) {
1170		if (fprc->ftpc_pid == pid && fprc->ftpc_acount != 0) {
1171			mutex_enter(&fprc->ftpc_mtx);
1172			mutex_exit(&bucket->ftb_mtx);
1173			fprc->ftpc_rcount++;
1174			atomic_add_64(&fprc->ftpc_acount, 1);
1175			ASSERT(fprc->ftpc_acount <= fprc->ftpc_rcount);
1176			mutex_exit(&fprc->ftpc_mtx);
1177
1178			return (fprc);
1179		}
1180	}
1181
1182	/*
1183	 * Drop the bucket lock so we don't try to perform a sleeping
1184	 * allocation under it.
1185	 */
1186	mutex_exit(&bucket->ftb_mtx);
1187
1188	new_fprc = kmem_zalloc(sizeof (fasttrap_proc_t), KM_SLEEP);
1189	new_fprc->ftpc_pid = pid;
1190	new_fprc->ftpc_rcount = 1;
1191	new_fprc->ftpc_acount = 1;
1192
1193	mutex_enter(&bucket->ftb_mtx);
1194
1195	/*
1196	 * Take another lap through the list to make sure a proc hasn't
1197	 * been created for this pid while we weren't under the bucket lock.
1198	 */
1199	for (fprc = bucket->ftb_data; fprc != NULL; fprc = fprc->ftpc_next) {
1200		if (fprc->ftpc_pid == pid && fprc->ftpc_acount != 0) {
1201			mutex_enter(&fprc->ftpc_mtx);
1202			mutex_exit(&bucket->ftb_mtx);
1203			fprc->ftpc_rcount++;
1204			atomic_add_64(&fprc->ftpc_acount, 1);
1205			ASSERT(fprc->ftpc_acount <= fprc->ftpc_rcount);
1206			mutex_exit(&fprc->ftpc_mtx);
1207
1208			kmem_free(new_fprc, sizeof (fasttrap_proc_t));
1209
1210			return (fprc);
1211		}
1212	}
1213
1214	new_fprc->ftpc_next = bucket->ftb_data;
1215	bucket->ftb_data = new_fprc;
1216
1217	mutex_exit(&bucket->ftb_mtx);
1218
1219	return (new_fprc);
1220}
1221
1222static void
1223fasttrap_proc_release(fasttrap_proc_t *proc)
1224{
1225	fasttrap_bucket_t *bucket;
1226	fasttrap_proc_t *fprc, **fprcp;
1227	pid_t pid = proc->ftpc_pid;
1228
1229	mutex_enter(&proc->ftpc_mtx);
1230
1231	ASSERT(proc->ftpc_rcount != 0);
1232	ASSERT(proc->ftpc_acount <= proc->ftpc_rcount);
1233
1234	if (--proc->ftpc_rcount != 0) {
1235		mutex_exit(&proc->ftpc_mtx);
1236		return;
1237	}
1238
1239	mutex_exit(&proc->ftpc_mtx);
1240
1241	/*
1242	 * There should definitely be no live providers associated with this
1243	 * process at this point.
1244	 */
1245	ASSERT(proc->ftpc_acount == 0);
1246
1247	bucket = &fasttrap_procs.fth_table[FASTTRAP_PROCS_INDEX(pid)];
1248	mutex_enter(&bucket->ftb_mtx);
1249
1250	fprcp = (fasttrap_proc_t **)&bucket->ftb_data;
1251	while ((fprc = *fprcp) != NULL) {
1252		if (fprc == proc)
1253			break;
1254
1255		fprcp = &fprc->ftpc_next;
1256	}
1257
1258	/*
1259	 * Something strange has happened if we can't find the proc.
1260	 */
1261	ASSERT(fprc != NULL);
1262
1263	*fprcp = fprc->ftpc_next;
1264
1265	mutex_exit(&bucket->ftb_mtx);
1266
1267	kmem_free(fprc, sizeof (fasttrap_proc_t));
1268}
1269
1270/*
1271 * Lookup a fasttrap-managed provider based on its name and associated pid.
1272 * If the pattr argument is non-NULL, this function instantiates the provider
1273 * if it doesn't exist otherwise it returns NULL. The provider is returned
1274 * with its lock held.
1275 */
1276static fasttrap_provider_t *
1277fasttrap_provider_lookup(pid_t pid, const char *name,
1278    const dtrace_pattr_t *pattr)
1279{
1280	fasttrap_provider_t *fp, *new_fp = NULL;
1281	fasttrap_bucket_t *bucket;
1282	char provname[DTRACE_PROVNAMELEN];
1283	proc_t *p;
1284	cred_t *cred;
1285
1286	ASSERT(strlen(name) < sizeof (fp->ftp_name));
1287	ASSERT(pattr != NULL);
1288
1289	bucket = &fasttrap_provs.fth_table[FASTTRAP_PROVS_INDEX(pid, name)];
1290	mutex_enter(&bucket->ftb_mtx);
1291
1292	/*
1293	 * Take a lap through the list and return the match if we find it.
1294	 */
1295	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1296		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1297		    !fp->ftp_retired) {
1298			mutex_enter(&fp->ftp_mtx);
1299			mutex_exit(&bucket->ftb_mtx);
1300			return (fp);
1301		}
1302	}
1303
1304	/*
1305	 * Drop the bucket lock so we don't try to perform a sleeping
1306	 * allocation under it.
1307	 */
1308	mutex_exit(&bucket->ftb_mtx);
1309
1310	/*
1311	 * Make sure the process exists, isn't a child created as the result
1312	 * of a vfork(2), and isn't a zombie (but may be in fork).
1313	 */
1314	mutex_enter(&pidlock);
1315	if ((p = prfind(pid)) == NULL) {
1316		mutex_exit(&pidlock);
1317		return (NULL);
1318	}
1319	mutex_enter(&p->p_lock);
1320	mutex_exit(&pidlock);
1321	if (p->p_flag & (SVFORK | SEXITING)) {
1322		mutex_exit(&p->p_lock);
1323		return (NULL);
1324	}
1325
1326	/*
1327	 * Increment p_dtrace_probes so that the process knows to inform us
1328	 * when it exits or execs. fasttrap_provider_free() decrements this
1329	 * when we're done with this provider.
1330	 */
1331	p->p_dtrace_probes++;
1332
1333	/*
1334	 * Grab the credentials for this process so we have
1335	 * something to pass to dtrace_register().
1336	 */
1337	mutex_enter(&p->p_crlock);
1338	crhold(p->p_cred);
1339	cred = p->p_cred;
1340	mutex_exit(&p->p_crlock);
1341	mutex_exit(&p->p_lock);
1342
1343	new_fp = kmem_zalloc(sizeof (fasttrap_provider_t), KM_SLEEP);
1344	new_fp->ftp_pid = pid;
1345	new_fp->ftp_proc = fasttrap_proc_lookup(pid);
1346
1347	ASSERT(new_fp->ftp_proc != NULL);
1348
1349	mutex_enter(&bucket->ftb_mtx);
1350
1351	/*
1352	 * Take another lap through the list to make sure a provider hasn't
1353	 * been created for this pid while we weren't under the bucket lock.
1354	 */
1355	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1356		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1357		    !fp->ftp_retired) {
1358			mutex_enter(&fp->ftp_mtx);
1359			mutex_exit(&bucket->ftb_mtx);
1360			fasttrap_provider_free(new_fp);
1361			crfree(cred);
1362			return (fp);
1363		}
1364	}
1365
1366	(void) strcpy(new_fp->ftp_name, name);
1367
1368	/*
1369	 * Fail and return NULL if either the provider name is too long
1370	 * or we fail to register this new provider with the DTrace
1371	 * framework. Note that this is the only place we ever construct
1372	 * the full provider name -- we keep it in pieces in the provider
1373	 * structure.
1374	 */
1375	if (snprintf(provname, sizeof (provname), "%s%u", name, (uint_t)pid) >=
1376	    sizeof (provname) ||
1377	    dtrace_register(provname, pattr,
1378	    DTRACE_PRIV_PROC | DTRACE_PRIV_OWNER | DTRACE_PRIV_ZONEOWNER, cred,
1379	    pattr == &pid_attr ? &pid_pops : &usdt_pops, new_fp,
1380	    &new_fp->ftp_provid) != 0) {
1381		mutex_exit(&bucket->ftb_mtx);
1382		fasttrap_provider_free(new_fp);
1383		crfree(cred);
1384		return (NULL);
1385	}
1386
1387	new_fp->ftp_next = bucket->ftb_data;
1388	bucket->ftb_data = new_fp;
1389
1390	mutex_enter(&new_fp->ftp_mtx);
1391	mutex_exit(&bucket->ftb_mtx);
1392
1393	crfree(cred);
1394	return (new_fp);
1395}
1396
1397static void
1398fasttrap_provider_free(fasttrap_provider_t *provider)
1399{
1400	pid_t pid = provider->ftp_pid;
1401	proc_t *p;
1402
1403	/*
1404	 * There need to be no associated enabled probes, no consumers
1405	 * creating probes, and no meta providers referencing this provider.
1406	 */
1407	ASSERT(provider->ftp_rcount == 0);
1408	ASSERT(provider->ftp_ccount == 0);
1409	ASSERT(provider->ftp_mcount == 0);
1410
1411	/*
1412	 * If this provider hasn't been retired, we need to explicitly drop the
1413	 * count of active providers on the associated process structure.
1414	 */
1415	if (!provider->ftp_retired) {
1416		atomic_add_64(&provider->ftp_proc->ftpc_acount, -1);
1417		ASSERT(provider->ftp_proc->ftpc_acount <
1418		    provider->ftp_proc->ftpc_rcount);
1419	}
1420
1421	fasttrap_proc_release(provider->ftp_proc);
1422
1423	kmem_free(provider, sizeof (fasttrap_provider_t));
1424
1425	/*
1426	 * Decrement p_dtrace_probes on the process whose provider we're
1427	 * freeing. We don't have to worry about clobbering somone else's
1428	 * modifications to it because we have locked the bucket that
1429	 * corresponds to this process's hash chain in the provider hash
1430	 * table. Don't sweat it if we can't find the process.
1431	 */
1432	mutex_enter(&pidlock);
1433	if ((p = prfind(pid)) == NULL) {
1434		mutex_exit(&pidlock);
1435		return;
1436	}
1437
1438	mutex_enter(&p->p_lock);
1439	mutex_exit(&pidlock);
1440
1441	p->p_dtrace_probes--;
1442	mutex_exit(&p->p_lock);
1443}
1444
1445static void
1446fasttrap_provider_retire(pid_t pid, const char *name, int mprov)
1447{
1448	fasttrap_provider_t *fp;
1449	fasttrap_bucket_t *bucket;
1450	dtrace_provider_id_t provid;
1451
1452	ASSERT(strlen(name) < sizeof (fp->ftp_name));
1453
1454	bucket = &fasttrap_provs.fth_table[FASTTRAP_PROVS_INDEX(pid, name)];
1455	mutex_enter(&bucket->ftb_mtx);
1456
1457	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1458		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1459		    !fp->ftp_retired)
1460			break;
1461	}
1462
1463	if (fp == NULL) {
1464		mutex_exit(&bucket->ftb_mtx);
1465		return;
1466	}
1467
1468	mutex_enter(&fp->ftp_mtx);
1469	ASSERT(!mprov || fp->ftp_mcount > 0);
1470	if (mprov && --fp->ftp_mcount != 0)  {
1471		mutex_exit(&fp->ftp_mtx);
1472		mutex_exit(&bucket->ftb_mtx);
1473		return;
1474	}
1475
1476	/*
1477	 * Mark the provider to be removed in our post-processing step, mark it
1478	 * retired, and drop the active count on its proc. Marking it indicates
1479	 * that we should try to remove it; setting the retired flag indicates
1480	 * that we're done with this provider; dropping the active the proc
1481	 * releases our hold, and when this reaches zero (as it will during
1482	 * exit or exec) the proc and associated providers become defunct.
1483	 *
1484	 * We obviously need to take the bucket lock before the provider lock
1485	 * to perform the lookup, but we need to drop the provider lock
1486	 * before calling into the DTrace framework since we acquire the
1487	 * provider lock in callbacks invoked from the DTrace framework. The
1488	 * bucket lock therefore protects the integrity of the provider hash
1489	 * table.
1490	 */
1491	atomic_add_64(&fp->ftp_proc->ftpc_acount, -1);
1492	ASSERT(fp->ftp_proc->ftpc_acount < fp->ftp_proc->ftpc_rcount);
1493
1494	fp->ftp_retired = 1;
1495	fp->ftp_marked = 1;
1496	provid = fp->ftp_provid;
1497	mutex_exit(&fp->ftp_mtx);
1498
1499	/*
1500	 * We don't have to worry about invalidating the same provider twice
1501	 * since fasttrap_provider_lookup() will ignore provider that have
1502	 * been marked as retired.
1503	 */
1504	dtrace_invalidate(provid);
1505
1506	mutex_exit(&bucket->ftb_mtx);
1507
1508	fasttrap_pid_cleanup();
1509}
1510
1511static int
1512fasttrap_uint32_cmp(const void *ap, const void *bp)
1513{
1514	return (*(const uint32_t *)ap - *(const uint32_t *)bp);
1515}
1516
1517static int
1518fasttrap_uint64_cmp(const void *ap, const void *bp)
1519{
1520	return (*(const uint64_t *)ap - *(const uint64_t *)bp);
1521}
1522
1523static int
1524fasttrap_add_probe(fasttrap_probe_spec_t *pdata)
1525{
1526	fasttrap_provider_t *provider;
1527	fasttrap_probe_t *pp;
1528	fasttrap_tracepoint_t *tp;
1529	char *name;
1530	int i, aframes, whack;
1531
1532	/*
1533	 * There needs to be at least one desired trace point.
1534	 */
1535	if (pdata->ftps_noffs == 0)
1536		return (EINVAL);
1537
1538	switch (pdata->ftps_type) {
1539	case DTFTP_ENTRY:
1540		name = "entry";
1541		aframes = FASTTRAP_ENTRY_AFRAMES;
1542		break;
1543	case DTFTP_RETURN:
1544		name = "return";
1545		aframes = FASTTRAP_RETURN_AFRAMES;
1546		break;
1547	case DTFTP_OFFSETS:
1548		name = NULL;
1549		break;
1550	default:
1551		return (EINVAL);
1552	}
1553
1554	if ((provider = fasttrap_provider_lookup(pdata->ftps_pid,
1555	    FASTTRAP_PID_NAME, &pid_attr)) == NULL)
1556		return (ESRCH);
1557
1558	/*
1559	 * Increment this reference count to indicate that a consumer is
1560	 * actively adding a new probe associated with this provider. This
1561	 * prevents the provider from being deleted -- we'll need to check
1562	 * for pending deletions when we drop this reference count.
1563	 */
1564	provider->ftp_ccount++;
1565	mutex_exit(&provider->ftp_mtx);
1566
1567	/*
1568	 * Grab the creation lock to ensure consistency between calls to
1569	 * dtrace_probe_lookup() and dtrace_probe_create() in the face of
1570	 * other threads creating probes. We must drop the provider lock
1571	 * before taking this lock to avoid a three-way deadlock with the
1572	 * DTrace framework.
1573	 */
1574	mutex_enter(&provider->ftp_cmtx);
1575
1576	if (name == NULL) {
1577		for (i = 0; i < pdata->ftps_noffs; i++) {
1578			char name_str[17];
1579
1580			(void) sprintf(name_str, "%llx",
1581			    (unsigned long long)pdata->ftps_offs[i]);
1582
1583			if (dtrace_probe_lookup(provider->ftp_provid,
1584			    pdata->ftps_mod, pdata->ftps_func, name_str) != 0)
1585				continue;
1586
1587			atomic_add_32(&fasttrap_total, 1);
1588
1589			if (fasttrap_total > fasttrap_max) {
1590				atomic_add_32(&fasttrap_total, -1);
1591				goto no_mem;
1592			}
1593
1594			pp = kmem_zalloc(sizeof (fasttrap_probe_t), KM_SLEEP);
1595
1596			pp->ftp_prov = provider;
1597			pp->ftp_faddr = pdata->ftps_pc;
1598			pp->ftp_fsize = pdata->ftps_size;
1599			pp->ftp_pid = pdata->ftps_pid;
1600			pp->ftp_ntps = 1;
1601
1602			tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t),
1603			    KM_SLEEP);
1604
1605			tp->ftt_proc = provider->ftp_proc;
1606			tp->ftt_pc = pdata->ftps_offs[i] + pdata->ftps_pc;
1607			tp->ftt_pid = pdata->ftps_pid;
1608
1609			pp->ftp_tps[0].fit_tp = tp;
1610			pp->ftp_tps[0].fit_id.fti_probe = pp;
1611			pp->ftp_tps[0].fit_id.fti_ptype = pdata->ftps_type;
1612
1613			pp->ftp_id = dtrace_probe_create(provider->ftp_provid,
1614			    pdata->ftps_mod, pdata->ftps_func, name_str,
1615			    FASTTRAP_OFFSET_AFRAMES, pp);
1616		}
1617
1618	} else if (dtrace_probe_lookup(provider->ftp_provid, pdata->ftps_mod,
1619	    pdata->ftps_func, name) == 0) {
1620		atomic_add_32(&fasttrap_total, pdata->ftps_noffs);
1621
1622		if (fasttrap_total > fasttrap_max) {
1623			atomic_add_32(&fasttrap_total, -pdata->ftps_noffs);
1624			goto no_mem;
1625		}
1626
1627		/*
1628		 * Make sure all tracepoint program counter values are unique.
1629		 * We later assume that each probe has exactly one tracepoint
1630		 * for a given pc.
1631		 */
1632		qsort(pdata->ftps_offs, pdata->ftps_noffs,
1633		    sizeof (uint64_t), fasttrap_uint64_cmp);
1634		for (i = 1; i < pdata->ftps_noffs; i++) {
1635			if (pdata->ftps_offs[i] > pdata->ftps_offs[i - 1])
1636				continue;
1637
1638			atomic_add_32(&fasttrap_total, -pdata->ftps_noffs);
1639			goto no_mem;
1640		}
1641
1642		ASSERT(pdata->ftps_noffs > 0);
1643		pp = kmem_zalloc(offsetof(fasttrap_probe_t,
1644		    ftp_tps[pdata->ftps_noffs]), KM_SLEEP);
1645
1646		pp->ftp_prov = provider;
1647		pp->ftp_faddr = pdata->ftps_pc;
1648		pp->ftp_fsize = pdata->ftps_size;
1649		pp->ftp_pid = pdata->ftps_pid;
1650		pp->ftp_ntps = pdata->ftps_noffs;
1651
1652		for (i = 0; i < pdata->ftps_noffs; i++) {
1653			tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t),
1654			    KM_SLEEP);
1655
1656			tp->ftt_proc = provider->ftp_proc;
1657			tp->ftt_pc = pdata->ftps_offs[i] + pdata->ftps_pc;
1658			tp->ftt_pid = pdata->ftps_pid;
1659
1660			pp->ftp_tps[i].fit_tp = tp;
1661			pp->ftp_tps[i].fit_id.fti_probe = pp;
1662			pp->ftp_tps[i].fit_id.fti_ptype = pdata->ftps_type;
1663		}
1664
1665		pp->ftp_id = dtrace_probe_create(provider->ftp_provid,
1666		    pdata->ftps_mod, pdata->ftps_func, name, aframes, pp);
1667	}
1668
1669	mutex_exit(&provider->ftp_cmtx);
1670
1671	/*
1672	 * We know that the provider is still valid since we incremented the
1673	 * creation reference count. If someone tried to clean up this provider
1674	 * while we were using it (e.g. because the process called exec(2) or
1675	 * exit(2)), take note of that and try to clean it up now.
1676	 */
1677	mutex_enter(&provider->ftp_mtx);
1678	provider->ftp_ccount--;
1679	whack = provider->ftp_retired;
1680	mutex_exit(&provider->ftp_mtx);
1681
1682	if (whack)
1683		fasttrap_pid_cleanup();
1684
1685	return (0);
1686
1687no_mem:
1688	/*
1689	 * If we've exhausted the allowable resources, we'll try to remove
1690	 * this provider to free some up. This is to cover the case where
1691	 * the user has accidentally created many more probes than was
1692	 * intended (e.g. pid123:::).
1693	 */
1694	mutex_exit(&provider->ftp_cmtx);
1695	mutex_enter(&provider->ftp_mtx);
1696	provider->ftp_ccount--;
1697	provider->ftp_marked = 1;
1698	mutex_exit(&provider->ftp_mtx);
1699
1700	fasttrap_pid_cleanup();
1701
1702	return (ENOMEM);
1703}
1704
1705/*ARGSUSED*/
1706static void *
1707fasttrap_meta_provide(void *arg, dtrace_helper_provdesc_t *dhpv, pid_t pid)
1708{
1709	fasttrap_provider_t *provider;
1710
1711	/*
1712	 * A 32-bit unsigned integer (like a pid for example) can be
1713	 * expressed in 10 or fewer decimal digits. Make sure that we'll
1714	 * have enough space for the provider name.
1715	 */
1716	if (strlen(dhpv->dthpv_provname) + 10 >=
1717	    sizeof (provider->ftp_name)) {
1718		cmn_err(CE_WARN, "failed to instantiate provider %s: "
1719		    "name too long to accomodate pid", dhpv->dthpv_provname);
1720		return (NULL);
1721	}
1722
1723	/*
1724	 * Don't let folks spoof the true pid provider.
1725	 */
1726	if (strcmp(dhpv->dthpv_provname, FASTTRAP_PID_NAME) == 0) {
1727		cmn_err(CE_WARN, "failed to instantiate provider %s: "
1728		    "%s is an invalid name", dhpv->dthpv_provname,
1729		    FASTTRAP_PID_NAME);
1730		return (NULL);
1731	}
1732
1733	/*
1734	 * The highest stability class that fasttrap supports is ISA; cap
1735	 * the stability of the new provider accordingly.
1736	 */
1737	if (dhpv->dthpv_pattr.dtpa_provider.dtat_class > DTRACE_CLASS_ISA)
1738		dhpv->dthpv_pattr.dtpa_provider.dtat_class = DTRACE_CLASS_ISA;
1739	if (dhpv->dthpv_pattr.dtpa_mod.dtat_class > DTRACE_CLASS_ISA)
1740		dhpv->dthpv_pattr.dtpa_mod.dtat_class = DTRACE_CLASS_ISA;
1741	if (dhpv->dthpv_pattr.dtpa_func.dtat_class > DTRACE_CLASS_ISA)
1742		dhpv->dthpv_pattr.dtpa_func.dtat_class = DTRACE_CLASS_ISA;
1743	if (dhpv->dthpv_pattr.dtpa_name.dtat_class > DTRACE_CLASS_ISA)
1744		dhpv->dthpv_pattr.dtpa_name.dtat_class = DTRACE_CLASS_ISA;
1745	if (dhpv->dthpv_pattr.dtpa_args.dtat_class > DTRACE_CLASS_ISA)
1746		dhpv->dthpv_pattr.dtpa_args.dtat_class = DTRACE_CLASS_ISA;
1747
1748	if ((provider = fasttrap_provider_lookup(pid, dhpv->dthpv_provname,
1749	    &dhpv->dthpv_pattr)) == NULL) {
1750		cmn_err(CE_WARN, "failed to instantiate provider %s for "
1751		    "process %u",  dhpv->dthpv_provname, (uint_t)pid);
1752		return (NULL);
1753	}
1754
1755	/*
1756	 * Up the meta provider count so this provider isn't removed until
1757	 * the meta provider has been told to remove it.
1758	 */
1759	provider->ftp_mcount++;
1760
1761	mutex_exit(&provider->ftp_mtx);
1762
1763	return (provider);
1764}
1765
1766/*ARGSUSED*/
1767static void
1768fasttrap_meta_create_probe(void *arg, void *parg,
1769    dtrace_helper_probedesc_t *dhpb)
1770{
1771	fasttrap_provider_t *provider = parg;
1772	fasttrap_probe_t *pp;
1773	fasttrap_tracepoint_t *tp;
1774	int i, j;
1775	uint32_t ntps;
1776
1777	/*
1778	 * Since the meta provider count is non-zero we don't have to worry
1779	 * about this provider disappearing.
1780	 */
1781	ASSERT(provider->ftp_mcount > 0);
1782
1783	/*
1784	 * The offsets must be unique.
1785	 */
1786	qsort(dhpb->dthpb_offs, dhpb->dthpb_noffs, sizeof (uint32_t),
1787	    fasttrap_uint32_cmp);
1788	for (i = 1; i < dhpb->dthpb_noffs; i++) {
1789		if (dhpb->dthpb_base + dhpb->dthpb_offs[i] <=
1790		    dhpb->dthpb_base + dhpb->dthpb_offs[i - 1])
1791			return;
1792	}
1793
1794	qsort(dhpb->dthpb_enoffs, dhpb->dthpb_nenoffs, sizeof (uint32_t),
1795	    fasttrap_uint32_cmp);
1796	for (i = 1; i < dhpb->dthpb_nenoffs; i++) {
1797		if (dhpb->dthpb_base + dhpb->dthpb_enoffs[i] <=
1798		    dhpb->dthpb_base + dhpb->dthpb_enoffs[i - 1])
1799			return;
1800	}
1801
1802	/*
1803	 * Grab the creation lock to ensure consistency between calls to
1804	 * dtrace_probe_lookup() and dtrace_probe_create() in the face of
1805	 * other threads creating probes.
1806	 */
1807	mutex_enter(&provider->ftp_cmtx);
1808
1809	if (dtrace_probe_lookup(provider->ftp_provid, dhpb->dthpb_mod,
1810	    dhpb->dthpb_func, dhpb->dthpb_name) != 0) {
1811		mutex_exit(&provider->ftp_cmtx);
1812		return;
1813	}
1814
1815	ntps = dhpb->dthpb_noffs + dhpb->dthpb_nenoffs;
1816	ASSERT(ntps > 0);
1817
1818	atomic_add_32(&fasttrap_total, ntps);
1819
1820	if (fasttrap_total > fasttrap_max) {
1821		atomic_add_32(&fasttrap_total, -ntps);
1822		mutex_exit(&provider->ftp_cmtx);
1823		return;
1824	}
1825
1826	pp = kmem_zalloc(offsetof(fasttrap_probe_t, ftp_tps[ntps]), KM_SLEEP);
1827
1828	pp->ftp_prov = provider;
1829	pp->ftp_pid = provider->ftp_pid;
1830	pp->ftp_ntps = ntps;
1831	pp->ftp_nargs = dhpb->dthpb_xargc;
1832	pp->ftp_xtypes = dhpb->dthpb_xtypes;
1833	pp->ftp_ntypes = dhpb->dthpb_ntypes;
1834
1835	/*
1836	 * First create a tracepoint for each actual point of interest.
1837	 */
1838	for (i = 0; i < dhpb->dthpb_noffs; i++) {
1839		tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t), KM_SLEEP);
1840
1841		tp->ftt_proc = provider->ftp_proc;
1842		tp->ftt_pc = dhpb->dthpb_base + dhpb->dthpb_offs[i];
1843		tp->ftt_pid = provider->ftp_pid;
1844
1845		pp->ftp_tps[i].fit_tp = tp;
1846		pp->ftp_tps[i].fit_id.fti_probe = pp;
1847#ifdef __sparc
1848		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_POST_OFFSETS;
1849#else
1850		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_OFFSETS;
1851#endif
1852	}
1853
1854	/*
1855	 * Then create a tracepoint for each is-enabled point.
1856	 */
1857	for (j = 0; i < ntps; i++, j++) {
1858		tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t), KM_SLEEP);
1859
1860		tp->ftt_proc = provider->ftp_proc;
1861		tp->ftt_pc = dhpb->dthpb_base + dhpb->dthpb_enoffs[j];
1862		tp->ftt_pid = provider->ftp_pid;
1863
1864		pp->ftp_tps[i].fit_tp = tp;
1865		pp->ftp_tps[i].fit_id.fti_probe = pp;
1866		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_IS_ENABLED;
1867	}
1868
1869	/*
1870	 * If the arguments are shuffled around we set the argument remapping
1871	 * table. Later, when the probe fires, we only remap the arguments
1872	 * if the table is non-NULL.
1873	 */
1874	for (i = 0; i < dhpb->dthpb_xargc; i++) {
1875		if (dhpb->dthpb_args[i] != i) {
1876			pp->ftp_argmap = dhpb->dthpb_args;
1877			break;
1878		}
1879	}
1880
1881	/*
1882	 * The probe is fully constructed -- register it with DTrace.
1883	 */
1884	pp->ftp_id = dtrace_probe_create(provider->ftp_provid, dhpb->dthpb_mod,
1885	    dhpb->dthpb_func, dhpb->dthpb_name, FASTTRAP_OFFSET_AFRAMES, pp);
1886
1887	mutex_exit(&provider->ftp_cmtx);
1888}
1889
1890/*ARGSUSED*/
1891static void
1892fasttrap_meta_remove(void *arg, dtrace_helper_provdesc_t *dhpv, pid_t pid)
1893{
1894	/*
1895	 * Clean up the USDT provider. There may be active consumers of the
1896	 * provider busy adding probes, no damage will actually befall the
1897	 * provider until that count has dropped to zero. This just puts
1898	 * the provider on death row.
1899	 */
1900	fasttrap_provider_retire(pid, dhpv->dthpv_provname, 1);
1901}
1902
1903static dtrace_mops_t fasttrap_mops = {
1904	fasttrap_meta_create_probe,
1905	fasttrap_meta_provide,
1906	fasttrap_meta_remove
1907};
1908
1909/*ARGSUSED*/
1910static int
1911fasttrap_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
1912{
1913	return (0);
1914}
1915
1916/*ARGSUSED*/
1917static int
1918fasttrap_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
1919{
1920	if (!dtrace_attached())
1921		return (EAGAIN);
1922
1923	if (cmd == FASTTRAPIOC_MAKEPROBE) {
1924		fasttrap_probe_spec_t *uprobe = (void *)arg;
1925		fasttrap_probe_spec_t *probe;
1926		uint64_t noffs;
1927		size_t size;
1928		int ret;
1929		char *c;
1930
1931		if (copyin(&uprobe->ftps_noffs, &noffs,
1932		    sizeof (uprobe->ftps_noffs)))
1933			return (EFAULT);
1934
1935		/*
1936		 * Probes must have at least one tracepoint.
1937		 */
1938		if (noffs == 0)
1939			return (EINVAL);
1940
1941		size = sizeof (fasttrap_probe_spec_t) +
1942		    sizeof (probe->ftps_offs[0]) * (noffs - 1);
1943
1944		if (size > 1024 * 1024)
1945			return (ENOMEM);
1946
1947		probe = kmem_alloc(size, KM_SLEEP);
1948
1949		if (copyin(uprobe, probe, size) != 0) {
1950			kmem_free(probe, size);
1951			return (EFAULT);
1952		}
1953
1954		/*
1955		 * Verify that the function and module strings contain no
1956		 * funny characters.
1957		 */
1958		for (c = &probe->ftps_func[0]; *c != '\0'; c++) {
1959			if (*c < 0x20 || 0x7f <= *c) {
1960				ret = EINVAL;
1961				goto err;
1962			}
1963		}
1964
1965		for (c = &probe->ftps_mod[0]; *c != '\0'; c++) {
1966			if (*c < 0x20 || 0x7f <= *c) {
1967				ret = EINVAL;
1968				goto err;
1969			}
1970		}
1971
1972		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
1973			proc_t *p;
1974			pid_t pid = probe->ftps_pid;
1975
1976			mutex_enter(&pidlock);
1977			/*
1978			 * Report an error if the process doesn't exist
1979			 * or is actively being birthed.
1980			 */
1981			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
1982				mutex_exit(&pidlock);
1983				return (ESRCH);
1984			}
1985			mutex_enter(&p->p_lock);
1986			mutex_exit(&pidlock);
1987
1988			if ((ret = priv_proc_cred_perm(cr, p, NULL,
1989			    VREAD | VWRITE)) != 0) {
1990				mutex_exit(&p->p_lock);
1991				return (ret);
1992			}
1993
1994			mutex_exit(&p->p_lock);
1995		}
1996
1997		ret = fasttrap_add_probe(probe);
1998err:
1999		kmem_free(probe, size);
2000
2001		return (ret);
2002
2003	} else if (cmd == FASTTRAPIOC_GETINSTR) {
2004		fasttrap_instr_query_t instr;
2005		fasttrap_tracepoint_t *tp;
2006		uint_t index;
2007		int ret;
2008
2009		if (copyin((void *)arg, &instr, sizeof (instr)) != 0)
2010			return (EFAULT);
2011
2012		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
2013			proc_t *p;
2014			pid_t pid = instr.ftiq_pid;
2015
2016			mutex_enter(&pidlock);
2017			/*
2018			 * Report an error if the process doesn't exist
2019			 * or is actively being birthed.
2020			 */
2021			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
2022				mutex_exit(&pidlock);
2023				return (ESRCH);
2024			}
2025			mutex_enter(&p->p_lock);
2026			mutex_exit(&pidlock);
2027
2028			if ((ret = priv_proc_cred_perm(cr, p, NULL,
2029			    VREAD)) != 0) {
2030				mutex_exit(&p->p_lock);
2031				return (ret);
2032			}
2033
2034			mutex_exit(&p->p_lock);
2035		}
2036
2037		index = FASTTRAP_TPOINTS_INDEX(instr.ftiq_pid, instr.ftiq_pc);
2038
2039		mutex_enter(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2040		tp = fasttrap_tpoints.fth_table[index].ftb_data;
2041		while (tp != NULL) {
2042			if (instr.ftiq_pid == tp->ftt_pid &&
2043			    instr.ftiq_pc == tp->ftt_pc &&
2044			    tp->ftt_proc->ftpc_acount != 0)
2045				break;
2046
2047			/*
2048			 * The count of active providers can only be
2049			 * decremented (i.e. to zero) during exec, exit, and
2050			 * removal of a meta provider so it should be
2051			 * impossible to drop the count during this operation().
2052			 */
2053			ASSERT(tp->ftt_proc->ftpc_acount != 0);
2054			tp = tp->ftt_next;
2055		}
2056
2057		if (tp == NULL) {
2058			mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2059			return (ENOENT);
2060		}
2061
2062		bcopy(&tp->ftt_instr, &instr.ftiq_instr,
2063		    sizeof (instr.ftiq_instr));
2064		mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2065
2066		if (copyout(&instr, (void *)arg, sizeof (instr)) != 0)
2067			return (EFAULT);
2068
2069		return (0);
2070	}
2071
2072	return (EINVAL);
2073}
2074
2075static struct cb_ops fasttrap_cb_ops = {
2076	fasttrap_open,		/* open */
2077	nodev,			/* close */
2078	nulldev,		/* strategy */
2079	nulldev,		/* print */
2080	nodev,			/* dump */
2081	nodev,			/* read */
2082	nodev,			/* write */
2083	fasttrap_ioctl,		/* ioctl */
2084	nodev,			/* devmap */
2085	nodev,			/* mmap */
2086	nodev,			/* segmap */
2087	nochpoll,		/* poll */
2088	ddi_prop_op,		/* cb_prop_op */
2089	0,			/* streamtab  */
2090	D_NEW | D_MP		/* Driver compatibility flag */
2091};
2092
2093/*ARGSUSED*/
2094static int
2095fasttrap_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
2096{
2097	int error;
2098
2099	switch (infocmd) {
2100	case DDI_INFO_DEVT2DEVINFO:
2101		*result = (void *)fasttrap_devi;
2102		error = DDI_SUCCESS;
2103		break;
2104	case DDI_INFO_DEVT2INSTANCE:
2105		*result = (void *)0;
2106		error = DDI_SUCCESS;
2107		break;
2108	default:
2109		error = DDI_FAILURE;
2110	}
2111	return (error);
2112}
2113
2114static int
2115fasttrap_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
2116{
2117	ulong_t nent;
2118
2119	switch (cmd) {
2120	case DDI_ATTACH:
2121		break;
2122	case DDI_RESUME:
2123		return (DDI_SUCCESS);
2124	default:
2125		return (DDI_FAILURE);
2126	}
2127
2128	if (ddi_create_minor_node(devi, "fasttrap", S_IFCHR, 0,
2129	    DDI_PSEUDO, NULL) == DDI_FAILURE) {
2130		ddi_remove_minor_node(devi, NULL);
2131		return (DDI_FAILURE);
2132	}
2133
2134	ddi_report_dev(devi);
2135	fasttrap_devi = devi;
2136
2137	/*
2138	 * Install our hooks into fork(2), exec(2), and exit(2).
2139	 */
2140	dtrace_fasttrap_fork_ptr = &fasttrap_fork;
2141	dtrace_fasttrap_exit_ptr = &fasttrap_exec_exit;
2142	dtrace_fasttrap_exec_ptr = &fasttrap_exec_exit;
2143
2144	fasttrap_max = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2145	    "fasttrap-max-probes", FASTTRAP_MAX_DEFAULT);
2146	fasttrap_total = 0;
2147
2148	/*
2149	 * Conjure up the tracepoints hashtable...
2150	 */
2151	nent = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2152	    "fasttrap-hash-size", FASTTRAP_TPOINTS_DEFAULT_SIZE);
2153
2154	if (nent == 0 || nent > 0x1000000)
2155		nent = FASTTRAP_TPOINTS_DEFAULT_SIZE;
2156
2157	if ((nent & (nent - 1)) == 0)
2158		fasttrap_tpoints.fth_nent = nent;
2159	else
2160		fasttrap_tpoints.fth_nent = 1 << fasttrap_highbit(nent);
2161	ASSERT(fasttrap_tpoints.fth_nent > 0);
2162	fasttrap_tpoints.fth_mask = fasttrap_tpoints.fth_nent - 1;
2163	fasttrap_tpoints.fth_table = kmem_zalloc(fasttrap_tpoints.fth_nent *
2164	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2165
2166	/*
2167	 * ... and the providers hash table...
2168	 */
2169	nent = FASTTRAP_PROVIDERS_DEFAULT_SIZE;
2170	if ((nent & (nent - 1)) == 0)
2171		fasttrap_provs.fth_nent = nent;
2172	else
2173		fasttrap_provs.fth_nent = 1 << fasttrap_highbit(nent);
2174	ASSERT(fasttrap_provs.fth_nent > 0);
2175	fasttrap_provs.fth_mask = fasttrap_provs.fth_nent - 1;
2176	fasttrap_provs.fth_table = kmem_zalloc(fasttrap_provs.fth_nent *
2177	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2178
2179	/*
2180	 * ... and the procs hash table.
2181	 */
2182	nent = FASTTRAP_PROCS_DEFAULT_SIZE;
2183	if ((nent & (nent - 1)) == 0)
2184		fasttrap_procs.fth_nent = nent;
2185	else
2186		fasttrap_procs.fth_nent = 1 << fasttrap_highbit(nent);
2187	ASSERT(fasttrap_procs.fth_nent > 0);
2188	fasttrap_procs.fth_mask = fasttrap_procs.fth_nent - 1;
2189	fasttrap_procs.fth_table = kmem_zalloc(fasttrap_procs.fth_nent *
2190	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2191
2192	(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2193	    &fasttrap_meta_id);
2194
2195	return (DDI_SUCCESS);
2196}
2197
2198static int
2199fasttrap_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
2200{
2201	int i, fail = 0;
2202	timeout_id_t tmp;
2203
2204	switch (cmd) {
2205	case DDI_DETACH:
2206		break;
2207	case DDI_SUSPEND:
2208		return (DDI_SUCCESS);
2209	default:
2210		return (DDI_FAILURE);
2211	}
2212
2213	/*
2214	 * Unregister the meta-provider to make sure no new fasttrap-
2215	 * managed providers come along while we're trying to close up
2216	 * shop. If we fail to detach, we'll need to re-register as a
2217	 * meta-provider. We can fail to unregister as a meta-provider
2218	 * if providers we manage still exist.
2219	 */
2220	if (fasttrap_meta_id != DTRACE_METAPROVNONE &&
2221	    dtrace_meta_unregister(fasttrap_meta_id) != 0)
2222		return (DDI_FAILURE);
2223
2224	/*
2225	 * Prevent any new timeouts from running by setting fasttrap_timeout
2226	 * to a non-zero value, and wait for the current timeout to complete.
2227	 */
2228	mutex_enter(&fasttrap_cleanup_mtx);
2229	fasttrap_cleanup_work = 0;
2230
2231	while (fasttrap_timeout != (timeout_id_t)1) {
2232		tmp = fasttrap_timeout;
2233		fasttrap_timeout = (timeout_id_t)1;
2234
2235		if (tmp != 0) {
2236			mutex_exit(&fasttrap_cleanup_mtx);
2237			(void) untimeout(tmp);
2238			mutex_enter(&fasttrap_cleanup_mtx);
2239		}
2240	}
2241
2242	fasttrap_cleanup_work = 0;
2243	mutex_exit(&fasttrap_cleanup_mtx);
2244
2245	/*
2246	 * Iterate over all of our providers. If there's still a process
2247	 * that corresponds to that pid, fail to detach.
2248	 */
2249	for (i = 0; i < fasttrap_provs.fth_nent; i++) {
2250		fasttrap_provider_t **fpp, *fp;
2251		fasttrap_bucket_t *bucket = &fasttrap_provs.fth_table[i];
2252
2253		mutex_enter(&bucket->ftb_mtx);
2254		fpp = (fasttrap_provider_t **)&bucket->ftb_data;
2255		while ((fp = *fpp) != NULL) {
2256			/*
2257			 * Acquire and release the lock as a simple way of
2258			 * waiting for any other consumer to finish with
2259			 * this provider. A thread must first acquire the
2260			 * bucket lock so there's no chance of another thread
2261			 * blocking on the provider's lock.
2262			 */
2263			mutex_enter(&fp->ftp_mtx);
2264			mutex_exit(&fp->ftp_mtx);
2265
2266			if (dtrace_unregister(fp->ftp_provid) != 0) {
2267				fail = 1;
2268				fpp = &fp->ftp_next;
2269			} else {
2270				*fpp = fp->ftp_next;
2271				fasttrap_provider_free(fp);
2272			}
2273		}
2274
2275		mutex_exit(&bucket->ftb_mtx);
2276	}
2277
2278	if (fail) {
2279		uint_t work;
2280		/*
2281		 * If we're failing to detach, we need to unblock timeouts
2282		 * and start a new timeout if any work has accumulated while
2283		 * we've been unsuccessfully trying to detach.
2284		 */
2285		mutex_enter(&fasttrap_cleanup_mtx);
2286		fasttrap_timeout = 0;
2287		work = fasttrap_cleanup_work;
2288		mutex_exit(&fasttrap_cleanup_mtx);
2289
2290		if (work)
2291			fasttrap_pid_cleanup();
2292
2293		(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2294		    &fasttrap_meta_id);
2295
2296		return (DDI_FAILURE);
2297	}
2298
2299#ifdef DEBUG
2300	mutex_enter(&fasttrap_count_mtx);
2301	ASSERT(fasttrap_pid_count == 0);
2302	mutex_exit(&fasttrap_count_mtx);
2303#endif
2304
2305	kmem_free(fasttrap_tpoints.fth_table,
2306	    fasttrap_tpoints.fth_nent * sizeof (fasttrap_bucket_t));
2307	fasttrap_tpoints.fth_nent = 0;
2308
2309	kmem_free(fasttrap_provs.fth_table,
2310	    fasttrap_provs.fth_nent * sizeof (fasttrap_bucket_t));
2311	fasttrap_provs.fth_nent = 0;
2312
2313	kmem_free(fasttrap_procs.fth_table,
2314	    fasttrap_procs.fth_nent * sizeof (fasttrap_bucket_t));
2315	fasttrap_procs.fth_nent = 0;
2316
2317	/*
2318	 * We know there are no tracepoints in any process anywhere in
2319	 * the system so there is no process which has its p_dtrace_count
2320	 * greater than zero, therefore we know that no thread can actively
2321	 * be executing code in fasttrap_fork(). Similarly for p_dtrace_probes
2322	 * and fasttrap_exec() and fasttrap_exit().
2323	 */
2324	ASSERT(dtrace_fasttrap_fork_ptr == &fasttrap_fork);
2325	dtrace_fasttrap_fork_ptr = NULL;
2326
2327	ASSERT(dtrace_fasttrap_exec_ptr == &fasttrap_exec_exit);
2328	dtrace_fasttrap_exec_ptr = NULL;
2329
2330	ASSERT(dtrace_fasttrap_exit_ptr == &fasttrap_exec_exit);
2331	dtrace_fasttrap_exit_ptr = NULL;
2332
2333	ddi_remove_minor_node(devi, NULL);
2334
2335	return (DDI_SUCCESS);
2336}
2337
2338static struct dev_ops fasttrap_ops = {
2339	DEVO_REV,		/* devo_rev */
2340	0,			/* refcnt */
2341	fasttrap_info,		/* get_dev_info */
2342	nulldev,		/* identify */
2343	nulldev,		/* probe */
2344	fasttrap_attach,	/* attach */
2345	fasttrap_detach,	/* detach */
2346	nodev,			/* reset */
2347	&fasttrap_cb_ops,	/* driver operations */
2348	NULL,			/* bus operations */
2349	nodev			/* dev power */
2350};
2351
2352/*
2353 * Module linkage information for the kernel.
2354 */
2355static struct modldrv modldrv = {
2356	&mod_driverops,		/* module type (this is a pseudo driver) */
2357	"Fasttrap Tracing",	/* name of module */
2358	&fasttrap_ops,		/* driver ops */
2359};
2360
2361static struct modlinkage modlinkage = {
2362	MODREV_1,
2363	(void *)&modldrv,
2364	NULL
2365};
2366
2367int
2368_init(void)
2369{
2370	return (mod_install(&modlinkage));
2371}
2372
2373int
2374_info(struct modinfo *modinfop)
2375{
2376	return (mod_info(&modlinkage, modinfop));
2377}
2378
2379int
2380_fini(void)
2381{
2382	return (mod_remove(&modlinkage));
2383}
2384