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 2009 Sun Microsystems, Inc.  All rights reserved.
24 * Use is subject to license terms.
25 */
26
27
28#include <sys/atomic.h>
29#include <sys/errno.h>
30#include <sys/stat.h>
31#include <sys/modctl.h>
32#include <sys/conf.h>
33#include <sys/systm.h>
34#include <sys/ddi.h>
35#include <sys/sunddi.h>
36#include <sys/cpuvar.h>
37#include <sys/kmem.h>
38#include <sys/strsubr.h>
39#include <sys/fasttrap.h>
40#include <sys/fasttrap_impl.h>
41#include <sys/fasttrap_isa.h>
42#include <sys/dtrace.h>
43#include <sys/dtrace_impl.h>
44#include <sys/sysmacros.h>
45#include <sys/proc.h>
46#include <sys/priv.h>
47#include <sys/policy.h>
48#include <util/qsort.h>
49
50/*
51 * User-Land Trap-Based Tracing
52 * ----------------------------
53 *
54 * The fasttrap provider allows DTrace consumers to instrument any user-level
55 * instruction to gather data; this includes probes with semantic
56 * signifigance like entry and return as well as simple offsets into the
57 * function. While the specific techniques used are very ISA specific, the
58 * methodology is generalizable to any architecture.
59 *
60 *
61 * The General Methodology
62 * -----------------------
63 *
64 * With the primary goal of tracing every user-land instruction and the
65 * limitation that we can't trust user space so don't want to rely on much
66 * information there, we begin by replacing the instructions we want to trace
67 * with trap instructions. Each instruction we overwrite is saved into a hash
68 * table keyed by process ID and pc address. When we enter the kernel due to
69 * this trap instruction, we need the effects of the replaced instruction to
70 * appear to have occurred before we proceed with the user thread's
71 * execution.
72 *
73 * Each user level thread is represented by a ulwp_t structure which is
74 * always easily accessible through a register. The most basic way to produce
75 * the effects of the instruction we replaced is to copy that instruction out
76 * to a bit of scratch space reserved in the user thread's ulwp_t structure
77 * (a sort of kernel-private thread local storage), set the PC to that
78 * scratch space and single step. When we reenter the kernel after single
79 * stepping the instruction we must then adjust the PC to point to what would
80 * normally be the next instruction. Of course, special care must be taken
81 * for branches and jumps, but these represent such a small fraction of any
82 * instruction set that writing the code to emulate these in the kernel is
83 * not too difficult.
84 *
85 * Return probes may require several tracepoints to trace every return site,
86 * and, conversely, each tracepoint may activate several probes (the entry
87 * and offset 0 probes, for example). To solve this muliplexing problem,
88 * tracepoints contain lists of probes to activate and probes contain lists
89 * of tracepoints to enable. If a probe is activated, it adds its ID to
90 * existing tracepoints or creates new ones as necessary.
91 *
92 * Most probes are activated _before_ the instruction is executed, but return
93 * probes are activated _after_ the effects of the last instruction of the
94 * function are visible. Return probes must be fired _after_ we have
95 * single-stepped the instruction whereas all other probes are fired
96 * beforehand.
97 *
98 *
99 * Lock Ordering
100 * -------------
101 *
102 * The lock ordering below -- both internally and with respect to the DTrace
103 * framework -- is a little tricky and bears some explanation. Each provider
104 * has a lock (ftp_mtx) that protects its members including reference counts
105 * for enabled probes (ftp_rcount), consumers actively creating probes
106 * (ftp_ccount) and USDT consumers (ftp_mcount); all three prevent a provider
107 * from being freed. A provider is looked up by taking the bucket lock for the
108 * provider hash table, and is returned with its lock held. The provider lock
109 * may be taken in functions invoked by the DTrace framework, but may not be
110 * held while calling functions in the DTrace framework.
111 *
112 * To ensure consistency over multiple calls to the DTrace framework, the
113 * creation lock (ftp_cmtx) should be held. Naturally, the creation lock may
114 * not be taken when holding the provider lock as that would create a cyclic
115 * lock ordering. In situations where one would naturally take the provider
116 * lock and then the creation lock, we instead up a reference count to prevent
117 * the provider from disappearing, drop the provider lock, and acquire the
118 * creation lock.
119 *
120 * Briefly:
121 * 	bucket lock before provider lock
122 *	DTrace before provider lock
123 *	creation lock before DTrace
124 *	never hold the provider lock and creation lock simultaneously
125 */
126
127static dev_info_t *fasttrap_devi;
128static dtrace_meta_provider_id_t fasttrap_meta_id;
129
130static timeout_id_t fasttrap_timeout;
131static kmutex_t fasttrap_cleanup_mtx;
132static uint_t fasttrap_cleanup_work;
133
134/*
135 * Generation count on modifications to the global tracepoint lookup table.
136 */
137static volatile uint64_t fasttrap_mod_gen;
138
139/*
140 * When the fasttrap provider is loaded, fasttrap_max is set to either
141 * FASTTRAP_MAX_DEFAULT or the value for fasttrap-max-probes in the
142 * fasttrap.conf file. Each time a probe is created, fasttrap_total is
143 * incremented by the number of tracepoints that may be associated with that
144 * probe; fasttrap_total is capped at fasttrap_max.
145 */
146#define	FASTTRAP_MAX_DEFAULT		250000
147static uint32_t fasttrap_max;
148static uint32_t fasttrap_total;
149
150
151#define	FASTTRAP_TPOINTS_DEFAULT_SIZE	0x4000
152#define	FASTTRAP_PROVIDERS_DEFAULT_SIZE	0x100
153#define	FASTTRAP_PROCS_DEFAULT_SIZE	0x100
154
155#define	FASTTRAP_PID_NAME		"pid"
156
157fasttrap_hash_t			fasttrap_tpoints;
158static fasttrap_hash_t		fasttrap_provs;
159static fasttrap_hash_t		fasttrap_procs;
160
161static uint64_t			fasttrap_pid_count;	/* pid ref count */
162static kmutex_t			fasttrap_count_mtx;	/* lock on ref count */
163
164#define	FASTTRAP_ENABLE_FAIL	1
165#define	FASTTRAP_ENABLE_PARTIAL	2
166
167static int fasttrap_tracepoint_enable(proc_t *, fasttrap_probe_t *, uint_t);
168static void fasttrap_tracepoint_disable(proc_t *, fasttrap_probe_t *, uint_t);
169
170static fasttrap_provider_t *fasttrap_provider_lookup(pid_t, const char *,
171    const dtrace_pattr_t *);
172static void fasttrap_provider_retire(pid_t, const char *, int);
173static void fasttrap_provider_free(fasttrap_provider_t *);
174
175static fasttrap_proc_t *fasttrap_proc_lookup(pid_t);
176static void fasttrap_proc_release(fasttrap_proc_t *);
177
178#define	FASTTRAP_PROVS_INDEX(pid, name) \
179	((fasttrap_hash_str(name) + (pid)) & fasttrap_provs.fth_mask)
180
181#define	FASTTRAP_PROCS_INDEX(pid) ((pid) & fasttrap_procs.fth_mask)
182
183static int
184fasttrap_highbit(ulong_t i)
185{
186	int h = 1;
187
188	if (i == 0)
189		return (0);
190#ifdef _LP64
191	if (i & 0xffffffff00000000ul) {
192		h += 32; i >>= 32;
193	}
194#endif
195	if (i & 0xffff0000) {
196		h += 16; i >>= 16;
197	}
198	if (i & 0xff00) {
199		h += 8; i >>= 8;
200	}
201	if (i & 0xf0) {
202		h += 4; i >>= 4;
203	}
204	if (i & 0xc) {
205		h += 2; i >>= 2;
206	}
207	if (i & 0x2) {
208		h += 1;
209	}
210	return (h);
211}
212
213static uint_t
214fasttrap_hash_str(const char *p)
215{
216	unsigned int g;
217	uint_t hval = 0;
218
219	while (*p) {
220		hval = (hval << 4) + *p++;
221		if ((g = (hval & 0xf0000000)) != 0)
222			hval ^= g >> 24;
223		hval &= ~g;
224	}
225	return (hval);
226}
227
228void
229fasttrap_sigtrap(proc_t *p, kthread_t *t, uintptr_t pc)
230{
231	sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP);
232
233	sqp->sq_info.si_signo = SIGTRAP;
234	sqp->sq_info.si_code = TRAP_DTRACE;
235	sqp->sq_info.si_addr = (caddr_t)pc;
236
237	mutex_enter(&p->p_lock);
238	sigaddqa(p, t, sqp);
239	mutex_exit(&p->p_lock);
240
241	if (t != NULL)
242		aston(t);
243}
244
245/*
246 * This function ensures that no threads are actively using the memory
247 * associated with probes that were formerly live.
248 */
249static void
250fasttrap_mod_barrier(uint64_t gen)
251{
252	int i;
253
254	if (gen < fasttrap_mod_gen)
255		return;
256
257	fasttrap_mod_gen++;
258
259	for (i = 0; i < NCPU; i++) {
260		mutex_enter(&cpu_core[i].cpuc_pid_lock);
261		mutex_exit(&cpu_core[i].cpuc_pid_lock);
262	}
263}
264
265/*
266 * This is the timeout's callback for cleaning up the providers and their
267 * probes.
268 */
269/*ARGSUSED*/
270static void
271fasttrap_pid_cleanup_cb(void *data)
272{
273	fasttrap_provider_t **fpp, *fp;
274	fasttrap_bucket_t *bucket;
275	dtrace_provider_id_t provid;
276	int i, later;
277
278	static volatile int in = 0;
279	ASSERT(in == 0);
280	in = 1;
281
282	mutex_enter(&fasttrap_cleanup_mtx);
283	while (fasttrap_cleanup_work) {
284		fasttrap_cleanup_work = 0;
285		mutex_exit(&fasttrap_cleanup_mtx);
286
287		later = 0;
288
289		/*
290		 * Iterate over all the providers trying to remove the marked
291		 * ones. If a provider is marked but not retired, we just
292		 * have to take a crack at removing it -- it's no big deal if
293		 * we can't.
294		 */
295		for (i = 0; i < fasttrap_provs.fth_nent; i++) {
296			bucket = &fasttrap_provs.fth_table[i];
297			mutex_enter(&bucket->ftb_mtx);
298			fpp = (fasttrap_provider_t **)&bucket->ftb_data;
299
300			while ((fp = *fpp) != NULL) {
301				if (!fp->ftp_marked) {
302					fpp = &fp->ftp_next;
303					continue;
304				}
305
306				mutex_enter(&fp->ftp_mtx);
307
308				/*
309				 * If this provider has consumers actively
310				 * creating probes (ftp_ccount) or is a USDT
311				 * provider (ftp_mcount), we can't unregister
312				 * or even condense.
313				 */
314				if (fp->ftp_ccount != 0 ||
315				    fp->ftp_mcount != 0) {
316					mutex_exit(&fp->ftp_mtx);
317					fp->ftp_marked = 0;
318					continue;
319				}
320
321				if (!fp->ftp_retired || fp->ftp_rcount != 0)
322					fp->ftp_marked = 0;
323
324				mutex_exit(&fp->ftp_mtx);
325
326				/*
327				 * If we successfully unregister this
328				 * provider we can remove it from the hash
329				 * chain and free the memory. If our attempt
330				 * to unregister fails and this is a retired
331				 * provider, increment our flag to try again
332				 * pretty soon. If we've consumed more than
333				 * half of our total permitted number of
334				 * probes call dtrace_condense() to try to
335				 * clean out the unenabled probes.
336				 */
337				provid = fp->ftp_provid;
338				if (dtrace_unregister(provid) != 0) {
339					if (fasttrap_total > fasttrap_max / 2)
340						(void) dtrace_condense(provid);
341					later += fp->ftp_marked;
342					fpp = &fp->ftp_next;
343				} else {
344					*fpp = fp->ftp_next;
345					fasttrap_provider_free(fp);
346				}
347			}
348			mutex_exit(&bucket->ftb_mtx);
349		}
350
351		mutex_enter(&fasttrap_cleanup_mtx);
352	}
353
354	ASSERT(fasttrap_timeout != 0);
355
356	/*
357	 * If we were unable to remove a retired provider, try again after
358	 * a second. This situation can occur in certain circumstances where
359	 * providers cannot be unregistered even though they have no probes
360	 * enabled because of an execution of dtrace -l or something similar.
361	 * If the timeout has been disabled (set to 1 because we're trying
362	 * to detach), we set fasttrap_cleanup_work to ensure that we'll
363	 * get a chance to do that work if and when the timeout is reenabled
364	 * (if detach fails).
365	 */
366	if (later > 0 && fasttrap_timeout != (timeout_id_t)1)
367		fasttrap_timeout = timeout(&fasttrap_pid_cleanup_cb, NULL, hz);
368	else if (later > 0)
369		fasttrap_cleanup_work = 1;
370	else
371		fasttrap_timeout = 0;
372
373	mutex_exit(&fasttrap_cleanup_mtx);
374	in = 0;
375}
376
377/*
378 * Activates the asynchronous cleanup mechanism.
379 */
380static void
381fasttrap_pid_cleanup(void)
382{
383	mutex_enter(&fasttrap_cleanup_mtx);
384	fasttrap_cleanup_work = 1;
385	if (fasttrap_timeout == 0)
386		fasttrap_timeout = timeout(&fasttrap_pid_cleanup_cb, NULL, 1);
387	mutex_exit(&fasttrap_cleanup_mtx);
388}
389
390/*
391 * This is called from cfork() via dtrace_fasttrap_fork(). The child
392 * process's address space is (roughly) a copy of the parent process's so
393 * we have to remove all the instrumentation we had previously enabled in the
394 * parent.
395 */
396static void
397fasttrap_fork(proc_t *p, proc_t *cp)
398{
399	pid_t ppid = p->p_pid;
400	int i;
401
402	ASSERT(curproc == p);
403	ASSERT(p->p_proc_flag & P_PR_LOCK);
404	ASSERT(p->p_dtrace_count > 0);
405	ASSERT(cp->p_dtrace_count == 0);
406
407	/*
408	 * This would be simpler and faster if we maintained per-process
409	 * hash tables of enabled tracepoints. It could, however, potentially
410	 * slow down execution of a tracepoint since we'd need to go
411	 * through two levels of indirection. In the future, we should
412	 * consider either maintaining per-process ancillary lists of
413	 * enabled tracepoints or hanging a pointer to a per-process hash
414	 * table of enabled tracepoints off the proc structure.
415	 */
416
417	/*
418	 * We don't have to worry about the child process disappearing
419	 * because we're in fork().
420	 */
421	mutex_enter(&cp->p_lock);
422	sprlock_proc(cp);
423	mutex_exit(&cp->p_lock);
424
425	/*
426	 * Iterate over every tracepoint looking for ones that belong to the
427	 * parent process, and remove each from the child process.
428	 */
429	for (i = 0; i < fasttrap_tpoints.fth_nent; i++) {
430		fasttrap_tracepoint_t *tp;
431		fasttrap_bucket_t *bucket = &fasttrap_tpoints.fth_table[i];
432
433		mutex_enter(&bucket->ftb_mtx);
434		for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
435			if (tp->ftt_pid == ppid &&
436			    tp->ftt_proc->ftpc_acount != 0) {
437				int ret = fasttrap_tracepoint_remove(cp, tp);
438				ASSERT(ret == 0);
439
440				/*
441				 * The count of active providers can only be
442				 * decremented (i.e. to zero) during exec,
443				 * exit, and removal of a meta provider so it
444				 * should be impossible to drop the count
445				 * mid-fork.
446				 */
447				ASSERT(tp->ftt_proc->ftpc_acount != 0);
448			}
449		}
450		mutex_exit(&bucket->ftb_mtx);
451	}
452
453	mutex_enter(&cp->p_lock);
454	sprunlock(cp);
455}
456
457/*
458 * This is called from proc_exit() or from exec_common() if p_dtrace_probes
459 * is set on the proc structure to indicate that there is a pid provider
460 * associated with this process.
461 */
462static void
463fasttrap_exec_exit(proc_t *p)
464{
465	ASSERT(p == curproc);
466	ASSERT(MUTEX_HELD(&p->p_lock));
467
468	mutex_exit(&p->p_lock);
469
470	/*
471	 * We clean up the pid provider for this process here; user-land
472	 * static probes are handled by the meta-provider remove entry point.
473	 */
474	fasttrap_provider_retire(p->p_pid, FASTTRAP_PID_NAME, 0);
475
476	mutex_enter(&p->p_lock);
477}
478
479
480/*ARGSUSED*/
481static void
482fasttrap_pid_provide(void *arg, const dtrace_probedesc_t *desc)
483{
484	/*
485	 * There are no "default" pid probes.
486	 */
487}
488
489static int
490fasttrap_tracepoint_enable(proc_t *p, fasttrap_probe_t *probe, uint_t index)
491{
492	fasttrap_tracepoint_t *tp, *new_tp = NULL;
493	fasttrap_bucket_t *bucket;
494	fasttrap_id_t *id;
495	pid_t pid;
496	uintptr_t pc;
497
498	ASSERT(index < probe->ftp_ntps);
499
500	pid = probe->ftp_pid;
501	pc = probe->ftp_tps[index].fit_tp->ftt_pc;
502	id = &probe->ftp_tps[index].fit_id;
503
504	ASSERT(probe->ftp_tps[index].fit_tp->ftt_pid == pid);
505
506	ASSERT(!(p->p_flag & SVFORK));
507
508	/*
509	 * Before we make any modifications, make sure we've imposed a barrier
510	 * on the generation in which this probe was last modified.
511	 */
512	fasttrap_mod_barrier(probe->ftp_gen);
513
514	bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
515
516	/*
517	 * If the tracepoint has already been enabled, just add our id to the
518	 * list of interested probes. This may be our second time through
519	 * this path in which case we'll have constructed the tracepoint we'd
520	 * like to install. If we can't find a match, and have an allocated
521	 * tracepoint ready to go, enable that one now.
522	 *
523	 * A tracepoint whose process is defunct is also considered defunct.
524	 */
525again:
526	mutex_enter(&bucket->ftb_mtx);
527	for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
528		/*
529		 * Note that it's safe to access the active count on the
530		 * associated proc structure because we know that at least one
531		 * provider (this one) will still be around throughout this
532		 * operation.
533		 */
534		if (tp->ftt_pid != pid || tp->ftt_pc != pc ||
535		    tp->ftt_proc->ftpc_acount == 0)
536			continue;
537
538		/*
539		 * Now that we've found a matching tracepoint, it would be
540		 * a decent idea to confirm that the tracepoint is still
541		 * enabled and the trap instruction hasn't been overwritten.
542		 * Since this is a little hairy, we'll punt for now.
543		 */
544
545		/*
546		 * This can't be the first interested probe. We don't have
547		 * to worry about another thread being in the midst of
548		 * deleting this tracepoint (which would be the only valid
549		 * reason for a tracepoint to have no interested probes)
550		 * since we're holding P_PR_LOCK for this process.
551		 */
552		ASSERT(tp->ftt_ids != NULL || tp->ftt_retids != NULL);
553
554		switch (id->fti_ptype) {
555		case DTFTP_ENTRY:
556		case DTFTP_OFFSETS:
557		case DTFTP_IS_ENABLED:
558			id->fti_next = tp->ftt_ids;
559			membar_producer();
560			tp->ftt_ids = id;
561			membar_producer();
562			break;
563
564		case DTFTP_RETURN:
565		case DTFTP_POST_OFFSETS:
566			id->fti_next = tp->ftt_retids;
567			membar_producer();
568			tp->ftt_retids = id;
569			membar_producer();
570			break;
571
572		default:
573			ASSERT(0);
574		}
575
576		mutex_exit(&bucket->ftb_mtx);
577
578		if (new_tp != NULL) {
579			new_tp->ftt_ids = NULL;
580			new_tp->ftt_retids = NULL;
581		}
582
583		return (0);
584	}
585
586	/*
587	 * If we have a good tracepoint ready to go, install it now while
588	 * we have the lock held and no one can screw with us.
589	 */
590	if (new_tp != NULL) {
591		int rc = 0;
592
593		new_tp->ftt_next = bucket->ftb_data;
594		membar_producer();
595		bucket->ftb_data = new_tp;
596		membar_producer();
597		mutex_exit(&bucket->ftb_mtx);
598
599		/*
600		 * Activate the tracepoint in the ISA-specific manner.
601		 * If this fails, we need to report the failure, but
602		 * indicate that this tracepoint must still be disabled
603		 * by calling fasttrap_tracepoint_disable().
604		 */
605		if (fasttrap_tracepoint_install(p, new_tp) != 0)
606			rc = FASTTRAP_ENABLE_PARTIAL;
607
608		/*
609		 * Increment the count of the number of tracepoints active in
610		 * the victim process.
611		 */
612		ASSERT(p->p_proc_flag & P_PR_LOCK);
613		p->p_dtrace_count++;
614
615		return (rc);
616	}
617
618	mutex_exit(&bucket->ftb_mtx);
619
620	/*
621	 * Initialize the tracepoint that's been preallocated with the probe.
622	 */
623	new_tp = probe->ftp_tps[index].fit_tp;
624
625	ASSERT(new_tp->ftt_pid == pid);
626	ASSERT(new_tp->ftt_pc == pc);
627	ASSERT(new_tp->ftt_proc == probe->ftp_prov->ftp_proc);
628	ASSERT(new_tp->ftt_ids == NULL);
629	ASSERT(new_tp->ftt_retids == NULL);
630
631	switch (id->fti_ptype) {
632	case DTFTP_ENTRY:
633	case DTFTP_OFFSETS:
634	case DTFTP_IS_ENABLED:
635		id->fti_next = NULL;
636		new_tp->ftt_ids = id;
637		break;
638
639	case DTFTP_RETURN:
640	case DTFTP_POST_OFFSETS:
641		id->fti_next = NULL;
642		new_tp->ftt_retids = id;
643		break;
644
645	default:
646		ASSERT(0);
647	}
648
649	/*
650	 * If the ISA-dependent initialization goes to plan, go back to the
651	 * beginning and try to install this freshly made tracepoint.
652	 */
653	if (fasttrap_tracepoint_init(p, new_tp, pc, id->fti_ptype) == 0)
654		goto again;
655
656	new_tp->ftt_ids = NULL;
657	new_tp->ftt_retids = NULL;
658
659	return (FASTTRAP_ENABLE_FAIL);
660}
661
662static void
663fasttrap_tracepoint_disable(proc_t *p, fasttrap_probe_t *probe, uint_t index)
664{
665	fasttrap_bucket_t *bucket;
666	fasttrap_provider_t *provider = probe->ftp_prov;
667	fasttrap_tracepoint_t **pp, *tp;
668	fasttrap_id_t *id, **idp;
669	pid_t pid;
670	uintptr_t pc;
671
672	ASSERT(index < probe->ftp_ntps);
673
674	pid = probe->ftp_pid;
675	pc = probe->ftp_tps[index].fit_tp->ftt_pc;
676	id = &probe->ftp_tps[index].fit_id;
677
678	ASSERT(probe->ftp_tps[index].fit_tp->ftt_pid == pid);
679
680	/*
681	 * Find the tracepoint and make sure that our id is one of the
682	 * ones registered with it.
683	 */
684	bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
685	mutex_enter(&bucket->ftb_mtx);
686	for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
687		if (tp->ftt_pid == pid && tp->ftt_pc == pc &&
688		    tp->ftt_proc == provider->ftp_proc)
689			break;
690	}
691
692	/*
693	 * If we somehow lost this tracepoint, we're in a world of hurt.
694	 */
695	ASSERT(tp != NULL);
696
697	switch (id->fti_ptype) {
698	case DTFTP_ENTRY:
699	case DTFTP_OFFSETS:
700	case DTFTP_IS_ENABLED:
701		ASSERT(tp->ftt_ids != NULL);
702		idp = &tp->ftt_ids;
703		break;
704
705	case DTFTP_RETURN:
706	case DTFTP_POST_OFFSETS:
707		ASSERT(tp->ftt_retids != NULL);
708		idp = &tp->ftt_retids;
709		break;
710
711	default:
712		ASSERT(0);
713	}
714
715	while ((*idp)->fti_probe != probe) {
716		idp = &(*idp)->fti_next;
717		ASSERT(*idp != NULL);
718	}
719
720	id = *idp;
721	*idp = id->fti_next;
722	membar_producer();
723
724	ASSERT(id->fti_probe == probe);
725
726	/*
727	 * If there are other registered enablings of this tracepoint, we're
728	 * all done, but if this was the last probe assocated with this
729	 * this tracepoint, we need to remove and free it.
730	 */
731	if (tp->ftt_ids != NULL || tp->ftt_retids != NULL) {
732
733		/*
734		 * If the current probe's tracepoint is in use, swap it
735		 * for an unused tracepoint.
736		 */
737		if (tp == probe->ftp_tps[index].fit_tp) {
738			fasttrap_probe_t *tmp_probe;
739			fasttrap_tracepoint_t **tmp_tp;
740			uint_t tmp_index;
741
742			if (tp->ftt_ids != NULL) {
743				tmp_probe = tp->ftt_ids->fti_probe;
744				/* LINTED - alignment */
745				tmp_index = FASTTRAP_ID_INDEX(tp->ftt_ids);
746				tmp_tp = &tmp_probe->ftp_tps[tmp_index].fit_tp;
747			} else {
748				tmp_probe = tp->ftt_retids->fti_probe;
749				/* LINTED - alignment */
750				tmp_index = FASTTRAP_ID_INDEX(tp->ftt_retids);
751				tmp_tp = &tmp_probe->ftp_tps[tmp_index].fit_tp;
752			}
753
754			ASSERT(*tmp_tp != NULL);
755			ASSERT(*tmp_tp != probe->ftp_tps[index].fit_tp);
756			ASSERT((*tmp_tp)->ftt_ids == NULL);
757			ASSERT((*tmp_tp)->ftt_retids == NULL);
758
759			probe->ftp_tps[index].fit_tp = *tmp_tp;
760			*tmp_tp = tp;
761		}
762
763		mutex_exit(&bucket->ftb_mtx);
764
765		/*
766		 * Tag the modified probe with the generation in which it was
767		 * changed.
768		 */
769		probe->ftp_gen = fasttrap_mod_gen;
770		return;
771	}
772
773	mutex_exit(&bucket->ftb_mtx);
774
775	/*
776	 * We can't safely remove the tracepoint from the set of active
777	 * tracepoints until we've actually removed the fasttrap instruction
778	 * from the process's text. We can, however, operate on this
779	 * tracepoint secure in the knowledge that no other thread is going to
780	 * be looking at it since we hold P_PR_LOCK on the process if it's
781	 * live or we hold the provider lock on the process if it's dead and
782	 * gone.
783	 */
784
785	/*
786	 * We only need to remove the actual instruction if we're looking
787	 * at an existing process
788	 */
789	if (p != NULL) {
790		/*
791		 * If we fail to restore the instruction we need to kill
792		 * this process since it's in a completely unrecoverable
793		 * state.
794		 */
795		if (fasttrap_tracepoint_remove(p, tp) != 0)
796			fasttrap_sigtrap(p, NULL, pc);
797
798		/*
799		 * Decrement the count of the number of tracepoints active
800		 * in the victim process.
801		 */
802		ASSERT(p->p_proc_flag & P_PR_LOCK);
803		p->p_dtrace_count--;
804	}
805
806	/*
807	 * Remove the probe from the hash table of active tracepoints.
808	 */
809	mutex_enter(&bucket->ftb_mtx);
810	pp = (fasttrap_tracepoint_t **)&bucket->ftb_data;
811	ASSERT(*pp != NULL);
812	while (*pp != tp) {
813		pp = &(*pp)->ftt_next;
814		ASSERT(*pp != NULL);
815	}
816
817	*pp = tp->ftt_next;
818	membar_producer();
819
820	mutex_exit(&bucket->ftb_mtx);
821
822	/*
823	 * Tag the modified probe with the generation in which it was changed.
824	 */
825	probe->ftp_gen = fasttrap_mod_gen;
826}
827
828static void
829fasttrap_enable_callbacks(void)
830{
831	/*
832	 * We don't have to play the rw lock game here because we're
833	 * providing something rather than taking something away --
834	 * we can be sure that no threads have tried to follow this
835	 * function pointer yet.
836	 */
837	mutex_enter(&fasttrap_count_mtx);
838	if (fasttrap_pid_count == 0) {
839		ASSERT(dtrace_pid_probe_ptr == NULL);
840		ASSERT(dtrace_return_probe_ptr == NULL);
841		dtrace_pid_probe_ptr = &fasttrap_pid_probe;
842		dtrace_return_probe_ptr = &fasttrap_return_probe;
843	}
844	ASSERT(dtrace_pid_probe_ptr == &fasttrap_pid_probe);
845	ASSERT(dtrace_return_probe_ptr == &fasttrap_return_probe);
846	fasttrap_pid_count++;
847	mutex_exit(&fasttrap_count_mtx);
848}
849
850static void
851fasttrap_disable_callbacks(void)
852{
853	ASSERT(MUTEX_HELD(&cpu_lock));
854
855	mutex_enter(&fasttrap_count_mtx);
856	ASSERT(fasttrap_pid_count > 0);
857	fasttrap_pid_count--;
858	if (fasttrap_pid_count == 0) {
859		cpu_t *cur, *cpu = CPU;
860
861		for (cur = cpu->cpu_next_onln; cur != cpu;
862		    cur = cur->cpu_next_onln) {
863			rw_enter(&cur->cpu_ft_lock, RW_WRITER);
864		}
865
866		dtrace_pid_probe_ptr = NULL;
867		dtrace_return_probe_ptr = NULL;
868
869		for (cur = cpu->cpu_next_onln; cur != cpu;
870		    cur = cur->cpu_next_onln) {
871			rw_exit(&cur->cpu_ft_lock);
872		}
873	}
874	mutex_exit(&fasttrap_count_mtx);
875}
876
877/*ARGSUSED*/
878static int
879fasttrap_pid_enable(void *arg, dtrace_id_t id, void *parg)
880{
881	fasttrap_probe_t *probe = parg;
882	proc_t *p;
883	int i, rc;
884
885	ASSERT(probe != NULL);
886	ASSERT(!probe->ftp_enabled);
887	ASSERT(id == probe->ftp_id);
888	ASSERT(MUTEX_HELD(&cpu_lock));
889
890	/*
891	 * Increment the count of enabled probes on this probe's provider;
892	 * the provider can't go away while the probe still exists. We
893	 * must increment this even if we aren't able to properly enable
894	 * this probe.
895	 */
896	mutex_enter(&probe->ftp_prov->ftp_mtx);
897	probe->ftp_prov->ftp_rcount++;
898	mutex_exit(&probe->ftp_prov->ftp_mtx);
899
900	/*
901	 * If this probe's provider is retired (meaning it was valid in a
902	 * previously exec'ed incarnation of this address space), bail out. The
903	 * provider can't go away while we're in this code path.
904	 */
905	if (probe->ftp_prov->ftp_retired)
906		return (0);
907
908	/*
909	 * If we can't find the process, it may be that we're in the context of
910	 * a fork in which the traced process is being born and we're copying
911	 * USDT probes. Otherwise, the process is gone so bail.
912	 */
913	if ((p = sprlock(probe->ftp_pid)) == NULL) {
914		if ((curproc->p_flag & SFORKING) == 0)
915			return (0);
916
917		mutex_enter(&pidlock);
918		p = prfind(probe->ftp_pid);
919
920		/*
921		 * Confirm that curproc is indeed forking the process in which
922		 * we're trying to enable probes.
923		 */
924		ASSERT(p != NULL);
925		ASSERT(p->p_parent == curproc);
926		ASSERT(p->p_stat == SIDL);
927
928		mutex_enter(&p->p_lock);
929		mutex_exit(&pidlock);
930
931		sprlock_proc(p);
932	}
933
934	ASSERT(!(p->p_flag & SVFORK));
935	mutex_exit(&p->p_lock);
936
937	/*
938	 * We have to enable the trap entry point before any user threads have
939	 * the chance to execute the trap instruction we're about to place
940	 * in their process's text.
941	 */
942	fasttrap_enable_callbacks();
943
944	/*
945	 * Enable all the tracepoints and add this probe's id to each
946	 * tracepoint's list of active probes.
947	 */
948	for (i = 0; i < probe->ftp_ntps; i++) {
949		if ((rc = fasttrap_tracepoint_enable(p, probe, i)) != 0) {
950			/*
951			 * If enabling the tracepoint failed completely,
952			 * we don't have to disable it; if the failure
953			 * was only partial we must disable it.
954			 */
955			if (rc == FASTTRAP_ENABLE_FAIL)
956				i--;
957			else
958				ASSERT(rc == FASTTRAP_ENABLE_PARTIAL);
959
960			/*
961			 * Back up and pull out all the tracepoints we've
962			 * created so far for this probe.
963			 */
964			while (i >= 0) {
965				fasttrap_tracepoint_disable(p, probe, i);
966				i--;
967			}
968
969			mutex_enter(&p->p_lock);
970			sprunlock(p);
971
972			/*
973			 * Since we're not actually enabling this probe,
974			 * drop our reference on the trap table entry.
975			 */
976			fasttrap_disable_callbacks();
977			return (0);
978		}
979	}
980
981	mutex_enter(&p->p_lock);
982	sprunlock(p);
983
984	probe->ftp_enabled = 1;
985	return (0);
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		    probe->ftps_noffs != noffs) {
1951			kmem_free(probe, size);
1952			return (EFAULT);
1953		}
1954
1955		/*
1956		 * Verify that the function and module strings contain no
1957		 * funny characters.
1958		 */
1959		for (c = &probe->ftps_func[0]; *c != '\0'; c++) {
1960			if (*c < 0x20 || 0x7f <= *c) {
1961				ret = EINVAL;
1962				goto err;
1963			}
1964		}
1965
1966		for (c = &probe->ftps_mod[0]; *c != '\0'; c++) {
1967			if (*c < 0x20 || 0x7f <= *c) {
1968				ret = EINVAL;
1969				goto err;
1970			}
1971		}
1972
1973		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
1974			proc_t *p;
1975			pid_t pid = probe->ftps_pid;
1976
1977			mutex_enter(&pidlock);
1978			/*
1979			 * Report an error if the process doesn't exist
1980			 * or is actively being birthed.
1981			 */
1982			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
1983				mutex_exit(&pidlock);
1984				return (ESRCH);
1985			}
1986			mutex_enter(&p->p_lock);
1987			mutex_exit(&pidlock);
1988
1989			if ((ret = priv_proc_cred_perm(cr, p, NULL,
1990			    VREAD | VWRITE)) != 0) {
1991				mutex_exit(&p->p_lock);
1992				return (ret);
1993			}
1994
1995			mutex_exit(&p->p_lock);
1996		}
1997
1998		ret = fasttrap_add_probe(probe);
1999err:
2000		kmem_free(probe, size);
2001
2002		return (ret);
2003
2004	} else if (cmd == FASTTRAPIOC_GETINSTR) {
2005		fasttrap_instr_query_t instr;
2006		fasttrap_tracepoint_t *tp;
2007		uint_t index;
2008		int ret;
2009
2010		if (copyin((void *)arg, &instr, sizeof (instr)) != 0)
2011			return (EFAULT);
2012
2013		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
2014			proc_t *p;
2015			pid_t pid = instr.ftiq_pid;
2016
2017			mutex_enter(&pidlock);
2018			/*
2019			 * Report an error if the process doesn't exist
2020			 * or is actively being birthed.
2021			 */
2022			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
2023				mutex_exit(&pidlock);
2024				return (ESRCH);
2025			}
2026			mutex_enter(&p->p_lock);
2027			mutex_exit(&pidlock);
2028
2029			if ((ret = priv_proc_cred_perm(cr, p, NULL,
2030			    VREAD)) != 0) {
2031				mutex_exit(&p->p_lock);
2032				return (ret);
2033			}
2034
2035			mutex_exit(&p->p_lock);
2036		}
2037
2038		index = FASTTRAP_TPOINTS_INDEX(instr.ftiq_pid, instr.ftiq_pc);
2039
2040		mutex_enter(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2041		tp = fasttrap_tpoints.fth_table[index].ftb_data;
2042		while (tp != NULL) {
2043			if (instr.ftiq_pid == tp->ftt_pid &&
2044			    instr.ftiq_pc == tp->ftt_pc &&
2045			    tp->ftt_proc->ftpc_acount != 0)
2046				break;
2047
2048			tp = tp->ftt_next;
2049		}
2050
2051		if (tp == NULL) {
2052			mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2053			return (ENOENT);
2054		}
2055
2056		bcopy(&tp->ftt_instr, &instr.ftiq_instr,
2057		    sizeof (instr.ftiq_instr));
2058		mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2059
2060		if (copyout(&instr, (void *)arg, sizeof (instr)) != 0)
2061			return (EFAULT);
2062
2063		return (0);
2064	}
2065
2066	return (EINVAL);
2067}
2068
2069static struct cb_ops fasttrap_cb_ops = {
2070	fasttrap_open,		/* open */
2071	nodev,			/* close */
2072	nulldev,		/* strategy */
2073	nulldev,		/* print */
2074	nodev,			/* dump */
2075	nodev,			/* read */
2076	nodev,			/* write */
2077	fasttrap_ioctl,		/* ioctl */
2078	nodev,			/* devmap */
2079	nodev,			/* mmap */
2080	nodev,			/* segmap */
2081	nochpoll,		/* poll */
2082	ddi_prop_op,		/* cb_prop_op */
2083	0,			/* streamtab  */
2084	D_NEW | D_MP		/* Driver compatibility flag */
2085};
2086
2087/*ARGSUSED*/
2088static int
2089fasttrap_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
2090{
2091	int error;
2092
2093	switch (infocmd) {
2094	case DDI_INFO_DEVT2DEVINFO:
2095		*result = (void *)fasttrap_devi;
2096		error = DDI_SUCCESS;
2097		break;
2098	case DDI_INFO_DEVT2INSTANCE:
2099		*result = (void *)0;
2100		error = DDI_SUCCESS;
2101		break;
2102	default:
2103		error = DDI_FAILURE;
2104	}
2105	return (error);
2106}
2107
2108static int
2109fasttrap_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
2110{
2111	ulong_t nent;
2112
2113	switch (cmd) {
2114	case DDI_ATTACH:
2115		break;
2116	case DDI_RESUME:
2117		return (DDI_SUCCESS);
2118	default:
2119		return (DDI_FAILURE);
2120	}
2121
2122	if (ddi_create_minor_node(devi, "fasttrap", S_IFCHR, 0,
2123	    DDI_PSEUDO, NULL) == DDI_FAILURE) {
2124		ddi_remove_minor_node(devi, NULL);
2125		return (DDI_FAILURE);
2126	}
2127
2128	ddi_report_dev(devi);
2129	fasttrap_devi = devi;
2130
2131	/*
2132	 * Install our hooks into fork(2), exec(2), and exit(2).
2133	 */
2134	dtrace_fasttrap_fork_ptr = &fasttrap_fork;
2135	dtrace_fasttrap_exit_ptr = &fasttrap_exec_exit;
2136	dtrace_fasttrap_exec_ptr = &fasttrap_exec_exit;
2137
2138	fasttrap_max = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2139	    "fasttrap-max-probes", FASTTRAP_MAX_DEFAULT);
2140	fasttrap_total = 0;
2141
2142	/*
2143	 * Conjure up the tracepoints hashtable...
2144	 */
2145	nent = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2146	    "fasttrap-hash-size", FASTTRAP_TPOINTS_DEFAULT_SIZE);
2147
2148	if (nent == 0 || nent > 0x1000000)
2149		nent = FASTTRAP_TPOINTS_DEFAULT_SIZE;
2150
2151	if ((nent & (nent - 1)) == 0)
2152		fasttrap_tpoints.fth_nent = nent;
2153	else
2154		fasttrap_tpoints.fth_nent = 1 << fasttrap_highbit(nent);
2155	ASSERT(fasttrap_tpoints.fth_nent > 0);
2156	fasttrap_tpoints.fth_mask = fasttrap_tpoints.fth_nent - 1;
2157	fasttrap_tpoints.fth_table = kmem_zalloc(fasttrap_tpoints.fth_nent *
2158	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2159
2160	/*
2161	 * ... and the providers hash table...
2162	 */
2163	nent = FASTTRAP_PROVIDERS_DEFAULT_SIZE;
2164	if ((nent & (nent - 1)) == 0)
2165		fasttrap_provs.fth_nent = nent;
2166	else
2167		fasttrap_provs.fth_nent = 1 << fasttrap_highbit(nent);
2168	ASSERT(fasttrap_provs.fth_nent > 0);
2169	fasttrap_provs.fth_mask = fasttrap_provs.fth_nent - 1;
2170	fasttrap_provs.fth_table = kmem_zalloc(fasttrap_provs.fth_nent *
2171	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2172
2173	/*
2174	 * ... and the procs hash table.
2175	 */
2176	nent = FASTTRAP_PROCS_DEFAULT_SIZE;
2177	if ((nent & (nent - 1)) == 0)
2178		fasttrap_procs.fth_nent = nent;
2179	else
2180		fasttrap_procs.fth_nent = 1 << fasttrap_highbit(nent);
2181	ASSERT(fasttrap_procs.fth_nent > 0);
2182	fasttrap_procs.fth_mask = fasttrap_procs.fth_nent - 1;
2183	fasttrap_procs.fth_table = kmem_zalloc(fasttrap_procs.fth_nent *
2184	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2185
2186	(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2187	    &fasttrap_meta_id);
2188
2189	return (DDI_SUCCESS);
2190}
2191
2192static int
2193fasttrap_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
2194{
2195	int i, fail = 0;
2196	timeout_id_t tmp;
2197
2198	switch (cmd) {
2199	case DDI_DETACH:
2200		break;
2201	case DDI_SUSPEND:
2202		return (DDI_SUCCESS);
2203	default:
2204		return (DDI_FAILURE);
2205	}
2206
2207	/*
2208	 * Unregister the meta-provider to make sure no new fasttrap-
2209	 * managed providers come along while we're trying to close up
2210	 * shop. If we fail to detach, we'll need to re-register as a
2211	 * meta-provider. We can fail to unregister as a meta-provider
2212	 * if providers we manage still exist.
2213	 */
2214	if (fasttrap_meta_id != DTRACE_METAPROVNONE &&
2215	    dtrace_meta_unregister(fasttrap_meta_id) != 0)
2216		return (DDI_FAILURE);
2217
2218	/*
2219	 * Prevent any new timeouts from running by setting fasttrap_timeout
2220	 * to a non-zero value, and wait for the current timeout to complete.
2221	 */
2222	mutex_enter(&fasttrap_cleanup_mtx);
2223	fasttrap_cleanup_work = 0;
2224
2225	while (fasttrap_timeout != (timeout_id_t)1) {
2226		tmp = fasttrap_timeout;
2227		fasttrap_timeout = (timeout_id_t)1;
2228
2229		if (tmp != 0) {
2230			mutex_exit(&fasttrap_cleanup_mtx);
2231			(void) untimeout(tmp);
2232			mutex_enter(&fasttrap_cleanup_mtx);
2233		}
2234	}
2235
2236	fasttrap_cleanup_work = 0;
2237	mutex_exit(&fasttrap_cleanup_mtx);
2238
2239	/*
2240	 * Iterate over all of our providers. If there's still a process
2241	 * that corresponds to that pid, fail to detach.
2242	 */
2243	for (i = 0; i < fasttrap_provs.fth_nent; i++) {
2244		fasttrap_provider_t **fpp, *fp;
2245		fasttrap_bucket_t *bucket = &fasttrap_provs.fth_table[i];
2246
2247		mutex_enter(&bucket->ftb_mtx);
2248		fpp = (fasttrap_provider_t **)&bucket->ftb_data;
2249		while ((fp = *fpp) != NULL) {
2250			/*
2251			 * Acquire and release the lock as a simple way of
2252			 * waiting for any other consumer to finish with
2253			 * this provider. A thread must first acquire the
2254			 * bucket lock so there's no chance of another thread
2255			 * blocking on the provider's lock.
2256			 */
2257			mutex_enter(&fp->ftp_mtx);
2258			mutex_exit(&fp->ftp_mtx);
2259
2260			if (dtrace_unregister(fp->ftp_provid) != 0) {
2261				fail = 1;
2262				fpp = &fp->ftp_next;
2263			} else {
2264				*fpp = fp->ftp_next;
2265				fasttrap_provider_free(fp);
2266			}
2267		}
2268
2269		mutex_exit(&bucket->ftb_mtx);
2270	}
2271
2272	if (fail) {
2273		uint_t work;
2274		/*
2275		 * If we're failing to detach, we need to unblock timeouts
2276		 * and start a new timeout if any work has accumulated while
2277		 * we've been unsuccessfully trying to detach.
2278		 */
2279		mutex_enter(&fasttrap_cleanup_mtx);
2280		fasttrap_timeout = 0;
2281		work = fasttrap_cleanup_work;
2282		mutex_exit(&fasttrap_cleanup_mtx);
2283
2284		if (work)
2285			fasttrap_pid_cleanup();
2286
2287		(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2288		    &fasttrap_meta_id);
2289
2290		return (DDI_FAILURE);
2291	}
2292
2293#ifdef DEBUG
2294	mutex_enter(&fasttrap_count_mtx);
2295	ASSERT(fasttrap_pid_count == 0);
2296	mutex_exit(&fasttrap_count_mtx);
2297#endif
2298
2299	kmem_free(fasttrap_tpoints.fth_table,
2300	    fasttrap_tpoints.fth_nent * sizeof (fasttrap_bucket_t));
2301	fasttrap_tpoints.fth_nent = 0;
2302
2303	kmem_free(fasttrap_provs.fth_table,
2304	    fasttrap_provs.fth_nent * sizeof (fasttrap_bucket_t));
2305	fasttrap_provs.fth_nent = 0;
2306
2307	kmem_free(fasttrap_procs.fth_table,
2308	    fasttrap_procs.fth_nent * sizeof (fasttrap_bucket_t));
2309	fasttrap_procs.fth_nent = 0;
2310
2311	/*
2312	 * We know there are no tracepoints in any process anywhere in
2313	 * the system so there is no process which has its p_dtrace_count
2314	 * greater than zero, therefore we know that no thread can actively
2315	 * be executing code in fasttrap_fork(). Similarly for p_dtrace_probes
2316	 * and fasttrap_exec() and fasttrap_exit().
2317	 */
2318	ASSERT(dtrace_fasttrap_fork_ptr == &fasttrap_fork);
2319	dtrace_fasttrap_fork_ptr = NULL;
2320
2321	ASSERT(dtrace_fasttrap_exec_ptr == &fasttrap_exec_exit);
2322	dtrace_fasttrap_exec_ptr = NULL;
2323
2324	ASSERT(dtrace_fasttrap_exit_ptr == &fasttrap_exec_exit);
2325	dtrace_fasttrap_exit_ptr = NULL;
2326
2327	ddi_remove_minor_node(devi, NULL);
2328
2329	return (DDI_SUCCESS);
2330}
2331
2332static struct dev_ops fasttrap_ops = {
2333	DEVO_REV,		/* devo_rev */
2334	0,			/* refcnt */
2335	fasttrap_info,		/* get_dev_info */
2336	nulldev,		/* identify */
2337	nulldev,		/* probe */
2338	fasttrap_attach,	/* attach */
2339	fasttrap_detach,	/* detach */
2340	nodev,			/* reset */
2341	&fasttrap_cb_ops,	/* driver operations */
2342	NULL,			/* bus operations */
2343	nodev,			/* dev power */
2344	ddi_quiesce_not_needed,		/* quiesce */
2345};
2346
2347/*
2348 * Module linkage information for the kernel.
2349 */
2350static struct modldrv modldrv = {
2351	&mod_driverops,		/* module type (this is a pseudo driver) */
2352	"Fasttrap Tracing",	/* name of module */
2353	&fasttrap_ops,		/* driver ops */
2354};
2355
2356static struct modlinkage modlinkage = {
2357	MODREV_1,
2358	(void *)&modldrv,
2359	NULL
2360};
2361
2362int
2363_init(void)
2364{
2365	return (mod_install(&modlinkage));
2366}
2367
2368int
2369_info(struct modinfo *modinfop)
2370{
2371	return (mod_info(&modlinkage, modinfop));
2372}
2373
2374int
2375_fini(void)
2376{
2377	return (mod_remove(&modlinkage));
2378}
2379