xref: /titanic_44/usr/src/uts/common/dtrace/fasttrap.c (revision 71269a2275bf5a143dad6461eee2710a344e7261)
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 
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 
127 static dev_info_t *fasttrap_devi;
128 static dtrace_meta_provider_id_t fasttrap_meta_id;
129 
130 static timeout_id_t fasttrap_timeout;
131 static kmutex_t fasttrap_cleanup_mtx;
132 static uint_t fasttrap_cleanup_work;
133 
134 /*
135  * Generation count on modifications to the global tracepoint lookup table.
136  */
137 static 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
147 static uint32_t fasttrap_max;
148 static 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 
157 fasttrap_hash_t			fasttrap_tpoints;
158 static fasttrap_hash_t		fasttrap_provs;
159 static fasttrap_hash_t		fasttrap_procs;
160 
161 static uint64_t			fasttrap_pid_count;	/* pid ref count */
162 static kmutex_t			fasttrap_count_mtx;	/* lock on ref count */
163 
164 #define	FASTTRAP_ENABLE_FAIL	1
165 #define	FASTTRAP_ENABLE_PARTIAL	2
166 
167 static int fasttrap_tracepoint_enable(proc_t *, fasttrap_probe_t *, uint_t);
168 static void fasttrap_tracepoint_disable(proc_t *, fasttrap_probe_t *, uint_t);
169 
170 static fasttrap_provider_t *fasttrap_provider_lookup(pid_t, const char *,
171     const dtrace_pattr_t *);
172 static void fasttrap_provider_retire(pid_t, const char *, int);
173 static void fasttrap_provider_free(fasttrap_provider_t *);
174 
175 static fasttrap_proc_t *fasttrap_proc_lookup(pid_t);
176 static 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 
183 static int
184 fasttrap_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 
213 static uint_t
214 fasttrap_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 
228 void
229 fasttrap_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  */
249 static void
250 fasttrap_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*/
270 static void
271 fasttrap_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  */
380 static void
381 fasttrap_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  */
396 static void
397 fasttrap_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  */
462 static void
463 fasttrap_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*/
481 static void
482 fasttrap_pid_provide(void *arg, const dtrace_probedesc_t *desc)
483 {
484 	/*
485 	 * There are no "default" pid probes.
486 	 */
487 }
488 
489 static int
490 fasttrap_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 	 */
525 again:
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 
662 static void
663 fasttrap_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 
828 static void
829 fasttrap_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 
850 static void
851 fasttrap_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*/
878 static void
879 fasttrap_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;
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;
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;
978 		}
979 	}
980 
981 	mutex_enter(&p->p_lock);
982 	sprunlock(p);
983 
984 	probe->ftp_enabled = 1;
985 }
986 
987 /*ARGSUSED*/
988 static void
989 fasttrap_pid_disable(void *arg, dtrace_id_t id, void *parg)
990 {
991 	fasttrap_probe_t *probe = parg;
992 	fasttrap_provider_t *provider = probe->ftp_prov;
993 	proc_t *p;
994 	int i, whack = 0;
995 
996 	ASSERT(id == probe->ftp_id);
997 
998 	/*
999 	 * We won't be able to acquire a /proc-esque lock on the process
1000 	 * iff the process is dead and gone. In this case, we rely on the
1001 	 * provider lock as a point of mutual exclusion to prevent other
1002 	 * DTrace consumers from disabling this probe.
1003 	 */
1004 	if ((p = sprlock(probe->ftp_pid)) != NULL) {
1005 		ASSERT(!(p->p_flag & SVFORK));
1006 		mutex_exit(&p->p_lock);
1007 	}
1008 
1009 	mutex_enter(&provider->ftp_mtx);
1010 
1011 	/*
1012 	 * Disable all the associated tracepoints (for fully enabled probes).
1013 	 */
1014 	if (probe->ftp_enabled) {
1015 		for (i = 0; i < probe->ftp_ntps; i++) {
1016 			fasttrap_tracepoint_disable(p, probe, i);
1017 		}
1018 	}
1019 
1020 	ASSERT(provider->ftp_rcount > 0);
1021 	provider->ftp_rcount--;
1022 
1023 	if (p != NULL) {
1024 		/*
1025 		 * Even though we may not be able to remove it entirely, we
1026 		 * mark this retired provider to get a chance to remove some
1027 		 * of the associated probes.
1028 		 */
1029 		if (provider->ftp_retired && !provider->ftp_marked)
1030 			whack = provider->ftp_marked = 1;
1031 		mutex_exit(&provider->ftp_mtx);
1032 
1033 		mutex_enter(&p->p_lock);
1034 		sprunlock(p);
1035 	} else {
1036 		/*
1037 		 * If the process is dead, we're just waiting for the
1038 		 * last probe to be disabled to be able to free it.
1039 		 */
1040 		if (provider->ftp_rcount == 0 && !provider->ftp_marked)
1041 			whack = provider->ftp_marked = 1;
1042 		mutex_exit(&provider->ftp_mtx);
1043 	}
1044 
1045 	if (whack)
1046 		fasttrap_pid_cleanup();
1047 
1048 	if (!probe->ftp_enabled)
1049 		return;
1050 
1051 	probe->ftp_enabled = 0;
1052 
1053 	ASSERT(MUTEX_HELD(&cpu_lock));
1054 	fasttrap_disable_callbacks();
1055 }
1056 
1057 /*ARGSUSED*/
1058 static void
1059 fasttrap_pid_getargdesc(void *arg, dtrace_id_t id, void *parg,
1060     dtrace_argdesc_t *desc)
1061 {
1062 	fasttrap_probe_t *probe = parg;
1063 	char *str;
1064 	int i, ndx;
1065 
1066 	desc->dtargd_native[0] = '\0';
1067 	desc->dtargd_xlate[0] = '\0';
1068 
1069 	if (probe->ftp_prov->ftp_retired != 0 ||
1070 	    desc->dtargd_ndx >= probe->ftp_nargs) {
1071 		desc->dtargd_ndx = DTRACE_ARGNONE;
1072 		return;
1073 	}
1074 
1075 	ndx = (probe->ftp_argmap != NULL) ?
1076 	    probe->ftp_argmap[desc->dtargd_ndx] : desc->dtargd_ndx;
1077 
1078 	str = probe->ftp_ntypes;
1079 	for (i = 0; i < ndx; i++) {
1080 		str += strlen(str) + 1;
1081 	}
1082 
1083 	ASSERT(strlen(str + 1) < sizeof (desc->dtargd_native));
1084 	(void) strcpy(desc->dtargd_native, str);
1085 
1086 	if (probe->ftp_xtypes == NULL)
1087 		return;
1088 
1089 	str = probe->ftp_xtypes;
1090 	for (i = 0; i < desc->dtargd_ndx; i++) {
1091 		str += strlen(str) + 1;
1092 	}
1093 
1094 	ASSERT(strlen(str + 1) < sizeof (desc->dtargd_xlate));
1095 	(void) strcpy(desc->dtargd_xlate, str);
1096 }
1097 
1098 /*ARGSUSED*/
1099 static void
1100 fasttrap_pid_destroy(void *arg, dtrace_id_t id, void *parg)
1101 {
1102 	fasttrap_probe_t *probe = parg;
1103 	int i;
1104 	size_t size;
1105 
1106 	ASSERT(probe != NULL);
1107 	ASSERT(!probe->ftp_enabled);
1108 	ASSERT(fasttrap_total >= probe->ftp_ntps);
1109 
1110 	atomic_add_32(&fasttrap_total, -probe->ftp_ntps);
1111 	size = offsetof(fasttrap_probe_t, ftp_tps[probe->ftp_ntps]);
1112 
1113 	if (probe->ftp_gen + 1 >= fasttrap_mod_gen)
1114 		fasttrap_mod_barrier(probe->ftp_gen);
1115 
1116 	for (i = 0; i < probe->ftp_ntps; i++) {
1117 		kmem_free(probe->ftp_tps[i].fit_tp,
1118 		    sizeof (fasttrap_tracepoint_t));
1119 	}
1120 
1121 	kmem_free(probe, size);
1122 }
1123 
1124 
1125 static const dtrace_pattr_t pid_attr = {
1126 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1127 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1128 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1129 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1130 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1131 };
1132 
1133 static dtrace_pops_t pid_pops = {
1134 	fasttrap_pid_provide,
1135 	NULL,
1136 	fasttrap_pid_enable,
1137 	fasttrap_pid_disable,
1138 	NULL,
1139 	NULL,
1140 	fasttrap_pid_getargdesc,
1141 	fasttrap_pid_getarg,
1142 	NULL,
1143 	fasttrap_pid_destroy
1144 };
1145 
1146 static dtrace_pops_t usdt_pops = {
1147 	fasttrap_pid_provide,
1148 	NULL,
1149 	fasttrap_pid_enable,
1150 	fasttrap_pid_disable,
1151 	NULL,
1152 	NULL,
1153 	fasttrap_pid_getargdesc,
1154 	fasttrap_usdt_getarg,
1155 	NULL,
1156 	fasttrap_pid_destroy
1157 };
1158 
1159 static fasttrap_proc_t *
1160 fasttrap_proc_lookup(pid_t pid)
1161 {
1162 	fasttrap_bucket_t *bucket;
1163 	fasttrap_proc_t *fprc, *new_fprc;
1164 
1165 	bucket = &fasttrap_procs.fth_table[FASTTRAP_PROCS_INDEX(pid)];
1166 	mutex_enter(&bucket->ftb_mtx);
1167 
1168 	for (fprc = bucket->ftb_data; fprc != NULL; fprc = fprc->ftpc_next) {
1169 		if (fprc->ftpc_pid == pid && fprc->ftpc_acount != 0) {
1170 			mutex_enter(&fprc->ftpc_mtx);
1171 			mutex_exit(&bucket->ftb_mtx);
1172 			fprc->ftpc_rcount++;
1173 			atomic_add_64(&fprc->ftpc_acount, 1);
1174 			ASSERT(fprc->ftpc_acount <= fprc->ftpc_rcount);
1175 			mutex_exit(&fprc->ftpc_mtx);
1176 
1177 			return (fprc);
1178 		}
1179 	}
1180 
1181 	/*
1182 	 * Drop the bucket lock so we don't try to perform a sleeping
1183 	 * allocation under it.
1184 	 */
1185 	mutex_exit(&bucket->ftb_mtx);
1186 
1187 	new_fprc = kmem_zalloc(sizeof (fasttrap_proc_t), KM_SLEEP);
1188 	new_fprc->ftpc_pid = pid;
1189 	new_fprc->ftpc_rcount = 1;
1190 	new_fprc->ftpc_acount = 1;
1191 
1192 	mutex_enter(&bucket->ftb_mtx);
1193 
1194 	/*
1195 	 * Take another lap through the list to make sure a proc hasn't
1196 	 * been created for this pid while we weren't under the bucket lock.
1197 	 */
1198 	for (fprc = bucket->ftb_data; fprc != NULL; fprc = fprc->ftpc_next) {
1199 		if (fprc->ftpc_pid == pid && fprc->ftpc_acount != 0) {
1200 			mutex_enter(&fprc->ftpc_mtx);
1201 			mutex_exit(&bucket->ftb_mtx);
1202 			fprc->ftpc_rcount++;
1203 			atomic_add_64(&fprc->ftpc_acount, 1);
1204 			ASSERT(fprc->ftpc_acount <= fprc->ftpc_rcount);
1205 			mutex_exit(&fprc->ftpc_mtx);
1206 
1207 			kmem_free(new_fprc, sizeof (fasttrap_proc_t));
1208 
1209 			return (fprc);
1210 		}
1211 	}
1212 
1213 	new_fprc->ftpc_next = bucket->ftb_data;
1214 	bucket->ftb_data = new_fprc;
1215 
1216 	mutex_exit(&bucket->ftb_mtx);
1217 
1218 	return (new_fprc);
1219 }
1220 
1221 static void
1222 fasttrap_proc_release(fasttrap_proc_t *proc)
1223 {
1224 	fasttrap_bucket_t *bucket;
1225 	fasttrap_proc_t *fprc, **fprcp;
1226 	pid_t pid = proc->ftpc_pid;
1227 
1228 	mutex_enter(&proc->ftpc_mtx);
1229 
1230 	ASSERT(proc->ftpc_rcount != 0);
1231 	ASSERT(proc->ftpc_acount <= proc->ftpc_rcount);
1232 
1233 	if (--proc->ftpc_rcount != 0) {
1234 		mutex_exit(&proc->ftpc_mtx);
1235 		return;
1236 	}
1237 
1238 	mutex_exit(&proc->ftpc_mtx);
1239 
1240 	/*
1241 	 * There should definitely be no live providers associated with this
1242 	 * process at this point.
1243 	 */
1244 	ASSERT(proc->ftpc_acount == 0);
1245 
1246 	bucket = &fasttrap_procs.fth_table[FASTTRAP_PROCS_INDEX(pid)];
1247 	mutex_enter(&bucket->ftb_mtx);
1248 
1249 	fprcp = (fasttrap_proc_t **)&bucket->ftb_data;
1250 	while ((fprc = *fprcp) != NULL) {
1251 		if (fprc == proc)
1252 			break;
1253 
1254 		fprcp = &fprc->ftpc_next;
1255 	}
1256 
1257 	/*
1258 	 * Something strange has happened if we can't find the proc.
1259 	 */
1260 	ASSERT(fprc != NULL);
1261 
1262 	*fprcp = fprc->ftpc_next;
1263 
1264 	mutex_exit(&bucket->ftb_mtx);
1265 
1266 	kmem_free(fprc, sizeof (fasttrap_proc_t));
1267 }
1268 
1269 /*
1270  * Lookup a fasttrap-managed provider based on its name and associated pid.
1271  * If the pattr argument is non-NULL, this function instantiates the provider
1272  * if it doesn't exist otherwise it returns NULL. The provider is returned
1273  * with its lock held.
1274  */
1275 static fasttrap_provider_t *
1276 fasttrap_provider_lookup(pid_t pid, const char *name,
1277     const dtrace_pattr_t *pattr)
1278 {
1279 	fasttrap_provider_t *fp, *new_fp = NULL;
1280 	fasttrap_bucket_t *bucket;
1281 	char provname[DTRACE_PROVNAMELEN];
1282 	proc_t *p;
1283 	cred_t *cred;
1284 
1285 	ASSERT(strlen(name) < sizeof (fp->ftp_name));
1286 	ASSERT(pattr != NULL);
1287 
1288 	bucket = &fasttrap_provs.fth_table[FASTTRAP_PROVS_INDEX(pid, name)];
1289 	mutex_enter(&bucket->ftb_mtx);
1290 
1291 	/*
1292 	 * Take a lap through the list and return the match if we find it.
1293 	 */
1294 	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1295 		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1296 		    !fp->ftp_retired) {
1297 			mutex_enter(&fp->ftp_mtx);
1298 			mutex_exit(&bucket->ftb_mtx);
1299 			return (fp);
1300 		}
1301 	}
1302 
1303 	/*
1304 	 * Drop the bucket lock so we don't try to perform a sleeping
1305 	 * allocation under it.
1306 	 */
1307 	mutex_exit(&bucket->ftb_mtx);
1308 
1309 	/*
1310 	 * Make sure the process exists, isn't a child created as the result
1311 	 * of a vfork(2), and isn't a zombie (but may be in fork).
1312 	 */
1313 	mutex_enter(&pidlock);
1314 	if ((p = prfind(pid)) == NULL) {
1315 		mutex_exit(&pidlock);
1316 		return (NULL);
1317 	}
1318 	mutex_enter(&p->p_lock);
1319 	mutex_exit(&pidlock);
1320 	if (p->p_flag & (SVFORK | SEXITING)) {
1321 		mutex_exit(&p->p_lock);
1322 		return (NULL);
1323 	}
1324 
1325 	/*
1326 	 * Increment p_dtrace_probes so that the process knows to inform us
1327 	 * when it exits or execs. fasttrap_provider_free() decrements this
1328 	 * when we're done with this provider.
1329 	 */
1330 	p->p_dtrace_probes++;
1331 
1332 	/*
1333 	 * Grab the credentials for this process so we have
1334 	 * something to pass to dtrace_register().
1335 	 */
1336 	mutex_enter(&p->p_crlock);
1337 	crhold(p->p_cred);
1338 	cred = p->p_cred;
1339 	mutex_exit(&p->p_crlock);
1340 	mutex_exit(&p->p_lock);
1341 
1342 	new_fp = kmem_zalloc(sizeof (fasttrap_provider_t), KM_SLEEP);
1343 	new_fp->ftp_pid = pid;
1344 	new_fp->ftp_proc = fasttrap_proc_lookup(pid);
1345 
1346 	ASSERT(new_fp->ftp_proc != NULL);
1347 
1348 	mutex_enter(&bucket->ftb_mtx);
1349 
1350 	/*
1351 	 * Take another lap through the list to make sure a provider hasn't
1352 	 * been created for this pid while we weren't under the bucket lock.
1353 	 */
1354 	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1355 		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1356 		    !fp->ftp_retired) {
1357 			mutex_enter(&fp->ftp_mtx);
1358 			mutex_exit(&bucket->ftb_mtx);
1359 			fasttrap_provider_free(new_fp);
1360 			crfree(cred);
1361 			return (fp);
1362 		}
1363 	}
1364 
1365 	(void) strcpy(new_fp->ftp_name, name);
1366 
1367 	/*
1368 	 * Fail and return NULL if either the provider name is too long
1369 	 * or we fail to register this new provider with the DTrace
1370 	 * framework. Note that this is the only place we ever construct
1371 	 * the full provider name -- we keep it in pieces in the provider
1372 	 * structure.
1373 	 */
1374 	if (snprintf(provname, sizeof (provname), "%s%u", name, (uint_t)pid) >=
1375 	    sizeof (provname) ||
1376 	    dtrace_register(provname, pattr,
1377 	    DTRACE_PRIV_PROC | DTRACE_PRIV_OWNER | DTRACE_PRIV_ZONEOWNER, cred,
1378 	    pattr == &pid_attr ? &pid_pops : &usdt_pops, new_fp,
1379 	    &new_fp->ftp_provid) != 0) {
1380 		mutex_exit(&bucket->ftb_mtx);
1381 		fasttrap_provider_free(new_fp);
1382 		crfree(cred);
1383 		return (NULL);
1384 	}
1385 
1386 	new_fp->ftp_next = bucket->ftb_data;
1387 	bucket->ftb_data = new_fp;
1388 
1389 	mutex_enter(&new_fp->ftp_mtx);
1390 	mutex_exit(&bucket->ftb_mtx);
1391 
1392 	crfree(cred);
1393 	return (new_fp);
1394 }
1395 
1396 static void
1397 fasttrap_provider_free(fasttrap_provider_t *provider)
1398 {
1399 	pid_t pid = provider->ftp_pid;
1400 	proc_t *p;
1401 
1402 	/*
1403 	 * There need to be no associated enabled probes, no consumers
1404 	 * creating probes, and no meta providers referencing this provider.
1405 	 */
1406 	ASSERT(provider->ftp_rcount == 0);
1407 	ASSERT(provider->ftp_ccount == 0);
1408 	ASSERT(provider->ftp_mcount == 0);
1409 
1410 	/*
1411 	 * If this provider hasn't been retired, we need to explicitly drop the
1412 	 * count of active providers on the associated process structure.
1413 	 */
1414 	if (!provider->ftp_retired) {
1415 		atomic_add_64(&provider->ftp_proc->ftpc_acount, -1);
1416 		ASSERT(provider->ftp_proc->ftpc_acount <
1417 		    provider->ftp_proc->ftpc_rcount);
1418 	}
1419 
1420 	fasttrap_proc_release(provider->ftp_proc);
1421 
1422 	kmem_free(provider, sizeof (fasttrap_provider_t));
1423 
1424 	/*
1425 	 * Decrement p_dtrace_probes on the process whose provider we're
1426 	 * freeing. We don't have to worry about clobbering somone else's
1427 	 * modifications to it because we have locked the bucket that
1428 	 * corresponds to this process's hash chain in the provider hash
1429 	 * table. Don't sweat it if we can't find the process.
1430 	 */
1431 	mutex_enter(&pidlock);
1432 	if ((p = prfind(pid)) == NULL) {
1433 		mutex_exit(&pidlock);
1434 		return;
1435 	}
1436 
1437 	mutex_enter(&p->p_lock);
1438 	mutex_exit(&pidlock);
1439 
1440 	p->p_dtrace_probes--;
1441 	mutex_exit(&p->p_lock);
1442 }
1443 
1444 static void
1445 fasttrap_provider_retire(pid_t pid, const char *name, int mprov)
1446 {
1447 	fasttrap_provider_t *fp;
1448 	fasttrap_bucket_t *bucket;
1449 	dtrace_provider_id_t provid;
1450 
1451 	ASSERT(strlen(name) < sizeof (fp->ftp_name));
1452 
1453 	bucket = &fasttrap_provs.fth_table[FASTTRAP_PROVS_INDEX(pid, name)];
1454 	mutex_enter(&bucket->ftb_mtx);
1455 
1456 	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1457 		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1458 		    !fp->ftp_retired)
1459 			break;
1460 	}
1461 
1462 	if (fp == NULL) {
1463 		mutex_exit(&bucket->ftb_mtx);
1464 		return;
1465 	}
1466 
1467 	mutex_enter(&fp->ftp_mtx);
1468 	ASSERT(!mprov || fp->ftp_mcount > 0);
1469 	if (mprov && --fp->ftp_mcount != 0)  {
1470 		mutex_exit(&fp->ftp_mtx);
1471 		mutex_exit(&bucket->ftb_mtx);
1472 		return;
1473 	}
1474 
1475 	/*
1476 	 * Mark the provider to be removed in our post-processing step, mark it
1477 	 * retired, and drop the active count on its proc. Marking it indicates
1478 	 * that we should try to remove it; setting the retired flag indicates
1479 	 * that we're done with this provider; dropping the active the proc
1480 	 * releases our hold, and when this reaches zero (as it will during
1481 	 * exit or exec) the proc and associated providers become defunct.
1482 	 *
1483 	 * We obviously need to take the bucket lock before the provider lock
1484 	 * to perform the lookup, but we need to drop the provider lock
1485 	 * before calling into the DTrace framework since we acquire the
1486 	 * provider lock in callbacks invoked from the DTrace framework. The
1487 	 * bucket lock therefore protects the integrity of the provider hash
1488 	 * table.
1489 	 */
1490 	atomic_add_64(&fp->ftp_proc->ftpc_acount, -1);
1491 	ASSERT(fp->ftp_proc->ftpc_acount < fp->ftp_proc->ftpc_rcount);
1492 
1493 	fp->ftp_retired = 1;
1494 	fp->ftp_marked = 1;
1495 	provid = fp->ftp_provid;
1496 	mutex_exit(&fp->ftp_mtx);
1497 
1498 	/*
1499 	 * We don't have to worry about invalidating the same provider twice
1500 	 * since fasttrap_provider_lookup() will ignore provider that have
1501 	 * been marked as retired.
1502 	 */
1503 	dtrace_invalidate(provid);
1504 
1505 	mutex_exit(&bucket->ftb_mtx);
1506 
1507 	fasttrap_pid_cleanup();
1508 }
1509 
1510 static int
1511 fasttrap_uint32_cmp(const void *ap, const void *bp)
1512 {
1513 	return (*(const uint32_t *)ap - *(const uint32_t *)bp);
1514 }
1515 
1516 static int
1517 fasttrap_uint64_cmp(const void *ap, const void *bp)
1518 {
1519 	return (*(const uint64_t *)ap - *(const uint64_t *)bp);
1520 }
1521 
1522 static int
1523 fasttrap_add_probe(fasttrap_probe_spec_t *pdata)
1524 {
1525 	fasttrap_provider_t *provider;
1526 	fasttrap_probe_t *pp;
1527 	fasttrap_tracepoint_t *tp;
1528 	char *name;
1529 	int i, aframes, whack;
1530 
1531 	/*
1532 	 * There needs to be at least one desired trace point.
1533 	 */
1534 	if (pdata->ftps_noffs == 0)
1535 		return (EINVAL);
1536 
1537 	switch (pdata->ftps_type) {
1538 	case DTFTP_ENTRY:
1539 		name = "entry";
1540 		aframes = FASTTRAP_ENTRY_AFRAMES;
1541 		break;
1542 	case DTFTP_RETURN:
1543 		name = "return";
1544 		aframes = FASTTRAP_RETURN_AFRAMES;
1545 		break;
1546 	case DTFTP_OFFSETS:
1547 		name = NULL;
1548 		break;
1549 	default:
1550 		return (EINVAL);
1551 	}
1552 
1553 	if ((provider = fasttrap_provider_lookup(pdata->ftps_pid,
1554 	    FASTTRAP_PID_NAME, &pid_attr)) == NULL)
1555 		return (ESRCH);
1556 
1557 	/*
1558 	 * Increment this reference count to indicate that a consumer is
1559 	 * actively adding a new probe associated with this provider. This
1560 	 * prevents the provider from being deleted -- we'll need to check
1561 	 * for pending deletions when we drop this reference count.
1562 	 */
1563 	provider->ftp_ccount++;
1564 	mutex_exit(&provider->ftp_mtx);
1565 
1566 	/*
1567 	 * Grab the creation lock to ensure consistency between calls to
1568 	 * dtrace_probe_lookup() and dtrace_probe_create() in the face of
1569 	 * other threads creating probes. We must drop the provider lock
1570 	 * before taking this lock to avoid a three-way deadlock with the
1571 	 * DTrace framework.
1572 	 */
1573 	mutex_enter(&provider->ftp_cmtx);
1574 
1575 	if (name == NULL) {
1576 		for (i = 0; i < pdata->ftps_noffs; i++) {
1577 			char name_str[17];
1578 
1579 			(void) sprintf(name_str, "%llx",
1580 			    (unsigned long long)pdata->ftps_offs[i]);
1581 
1582 			if (dtrace_probe_lookup(provider->ftp_provid,
1583 			    pdata->ftps_mod, pdata->ftps_func, name_str) != 0)
1584 				continue;
1585 
1586 			atomic_add_32(&fasttrap_total, 1);
1587 
1588 			if (fasttrap_total > fasttrap_max) {
1589 				atomic_add_32(&fasttrap_total, -1);
1590 				goto no_mem;
1591 			}
1592 
1593 			pp = kmem_zalloc(sizeof (fasttrap_probe_t), KM_SLEEP);
1594 
1595 			pp->ftp_prov = provider;
1596 			pp->ftp_faddr = pdata->ftps_pc;
1597 			pp->ftp_fsize = pdata->ftps_size;
1598 			pp->ftp_pid = pdata->ftps_pid;
1599 			pp->ftp_ntps = 1;
1600 
1601 			tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t),
1602 			    KM_SLEEP);
1603 
1604 			tp->ftt_proc = provider->ftp_proc;
1605 			tp->ftt_pc = pdata->ftps_offs[i] + pdata->ftps_pc;
1606 			tp->ftt_pid = pdata->ftps_pid;
1607 
1608 			pp->ftp_tps[0].fit_tp = tp;
1609 			pp->ftp_tps[0].fit_id.fti_probe = pp;
1610 			pp->ftp_tps[0].fit_id.fti_ptype = pdata->ftps_type;
1611 
1612 			pp->ftp_id = dtrace_probe_create(provider->ftp_provid,
1613 			    pdata->ftps_mod, pdata->ftps_func, name_str,
1614 			    FASTTRAP_OFFSET_AFRAMES, pp);
1615 		}
1616 
1617 	} else if (dtrace_probe_lookup(provider->ftp_provid, pdata->ftps_mod,
1618 	    pdata->ftps_func, name) == 0) {
1619 		atomic_add_32(&fasttrap_total, pdata->ftps_noffs);
1620 
1621 		if (fasttrap_total > fasttrap_max) {
1622 			atomic_add_32(&fasttrap_total, -pdata->ftps_noffs);
1623 			goto no_mem;
1624 		}
1625 
1626 		/*
1627 		 * Make sure all tracepoint program counter values are unique.
1628 		 * We later assume that each probe has exactly one tracepoint
1629 		 * for a given pc.
1630 		 */
1631 		qsort(pdata->ftps_offs, pdata->ftps_noffs,
1632 		    sizeof (uint64_t), fasttrap_uint64_cmp);
1633 		for (i = 1; i < pdata->ftps_noffs; i++) {
1634 			if (pdata->ftps_offs[i] > pdata->ftps_offs[i - 1])
1635 				continue;
1636 
1637 			atomic_add_32(&fasttrap_total, -pdata->ftps_noffs);
1638 			goto no_mem;
1639 		}
1640 
1641 		ASSERT(pdata->ftps_noffs > 0);
1642 		pp = kmem_zalloc(offsetof(fasttrap_probe_t,
1643 		    ftp_tps[pdata->ftps_noffs]), KM_SLEEP);
1644 
1645 		pp->ftp_prov = provider;
1646 		pp->ftp_faddr = pdata->ftps_pc;
1647 		pp->ftp_fsize = pdata->ftps_size;
1648 		pp->ftp_pid = pdata->ftps_pid;
1649 		pp->ftp_ntps = pdata->ftps_noffs;
1650 
1651 		for (i = 0; i < pdata->ftps_noffs; i++) {
1652 			tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t),
1653 			    KM_SLEEP);
1654 
1655 			tp->ftt_proc = provider->ftp_proc;
1656 			tp->ftt_pc = pdata->ftps_offs[i] + pdata->ftps_pc;
1657 			tp->ftt_pid = pdata->ftps_pid;
1658 
1659 			pp->ftp_tps[i].fit_tp = tp;
1660 			pp->ftp_tps[i].fit_id.fti_probe = pp;
1661 			pp->ftp_tps[i].fit_id.fti_ptype = pdata->ftps_type;
1662 		}
1663 
1664 		pp->ftp_id = dtrace_probe_create(provider->ftp_provid,
1665 		    pdata->ftps_mod, pdata->ftps_func, name, aframes, pp);
1666 	}
1667 
1668 	mutex_exit(&provider->ftp_cmtx);
1669 
1670 	/*
1671 	 * We know that the provider is still valid since we incremented the
1672 	 * creation reference count. If someone tried to clean up this provider
1673 	 * while we were using it (e.g. because the process called exec(2) or
1674 	 * exit(2)), take note of that and try to clean it up now.
1675 	 */
1676 	mutex_enter(&provider->ftp_mtx);
1677 	provider->ftp_ccount--;
1678 	whack = provider->ftp_retired;
1679 	mutex_exit(&provider->ftp_mtx);
1680 
1681 	if (whack)
1682 		fasttrap_pid_cleanup();
1683 
1684 	return (0);
1685 
1686 no_mem:
1687 	/*
1688 	 * If we've exhausted the allowable resources, we'll try to remove
1689 	 * this provider to free some up. This is to cover the case where
1690 	 * the user has accidentally created many more probes than was
1691 	 * intended (e.g. pid123:::).
1692 	 */
1693 	mutex_exit(&provider->ftp_cmtx);
1694 	mutex_enter(&provider->ftp_mtx);
1695 	provider->ftp_ccount--;
1696 	provider->ftp_marked = 1;
1697 	mutex_exit(&provider->ftp_mtx);
1698 
1699 	fasttrap_pid_cleanup();
1700 
1701 	return (ENOMEM);
1702 }
1703 
1704 /*ARGSUSED*/
1705 static void *
1706 fasttrap_meta_provide(void *arg, dtrace_helper_provdesc_t *dhpv, pid_t pid)
1707 {
1708 	fasttrap_provider_t *provider;
1709 
1710 	/*
1711 	 * A 32-bit unsigned integer (like a pid for example) can be
1712 	 * expressed in 10 or fewer decimal digits. Make sure that we'll
1713 	 * have enough space for the provider name.
1714 	 */
1715 	if (strlen(dhpv->dthpv_provname) + 10 >=
1716 	    sizeof (provider->ftp_name)) {
1717 		cmn_err(CE_WARN, "failed to instantiate provider %s: "
1718 		    "name too long to accomodate pid", dhpv->dthpv_provname);
1719 		return (NULL);
1720 	}
1721 
1722 	/*
1723 	 * Don't let folks spoof the true pid provider.
1724 	 */
1725 	if (strcmp(dhpv->dthpv_provname, FASTTRAP_PID_NAME) == 0) {
1726 		cmn_err(CE_WARN, "failed to instantiate provider %s: "
1727 		    "%s is an invalid name", dhpv->dthpv_provname,
1728 		    FASTTRAP_PID_NAME);
1729 		return (NULL);
1730 	}
1731 
1732 	/*
1733 	 * The highest stability class that fasttrap supports is ISA; cap
1734 	 * the stability of the new provider accordingly.
1735 	 */
1736 	if (dhpv->dthpv_pattr.dtpa_provider.dtat_class > DTRACE_CLASS_ISA)
1737 		dhpv->dthpv_pattr.dtpa_provider.dtat_class = DTRACE_CLASS_ISA;
1738 	if (dhpv->dthpv_pattr.dtpa_mod.dtat_class > DTRACE_CLASS_ISA)
1739 		dhpv->dthpv_pattr.dtpa_mod.dtat_class = DTRACE_CLASS_ISA;
1740 	if (dhpv->dthpv_pattr.dtpa_func.dtat_class > DTRACE_CLASS_ISA)
1741 		dhpv->dthpv_pattr.dtpa_func.dtat_class = DTRACE_CLASS_ISA;
1742 	if (dhpv->dthpv_pattr.dtpa_name.dtat_class > DTRACE_CLASS_ISA)
1743 		dhpv->dthpv_pattr.dtpa_name.dtat_class = DTRACE_CLASS_ISA;
1744 	if (dhpv->dthpv_pattr.dtpa_args.dtat_class > DTRACE_CLASS_ISA)
1745 		dhpv->dthpv_pattr.dtpa_args.dtat_class = DTRACE_CLASS_ISA;
1746 
1747 	if ((provider = fasttrap_provider_lookup(pid, dhpv->dthpv_provname,
1748 	    &dhpv->dthpv_pattr)) == NULL) {
1749 		cmn_err(CE_WARN, "failed to instantiate provider %s for "
1750 		    "process %u",  dhpv->dthpv_provname, (uint_t)pid);
1751 		return (NULL);
1752 	}
1753 
1754 	/*
1755 	 * Up the meta provider count so this provider isn't removed until
1756 	 * the meta provider has been told to remove it.
1757 	 */
1758 	provider->ftp_mcount++;
1759 
1760 	mutex_exit(&provider->ftp_mtx);
1761 
1762 	return (provider);
1763 }
1764 
1765 /*ARGSUSED*/
1766 static void
1767 fasttrap_meta_create_probe(void *arg, void *parg,
1768     dtrace_helper_probedesc_t *dhpb)
1769 {
1770 	fasttrap_provider_t *provider = parg;
1771 	fasttrap_probe_t *pp;
1772 	fasttrap_tracepoint_t *tp;
1773 	int i, j;
1774 	uint32_t ntps;
1775 
1776 	/*
1777 	 * Since the meta provider count is non-zero we don't have to worry
1778 	 * about this provider disappearing.
1779 	 */
1780 	ASSERT(provider->ftp_mcount > 0);
1781 
1782 	/*
1783 	 * The offsets must be unique.
1784 	 */
1785 	qsort(dhpb->dthpb_offs, dhpb->dthpb_noffs, sizeof (uint32_t),
1786 	    fasttrap_uint32_cmp);
1787 	for (i = 1; i < dhpb->dthpb_noffs; i++) {
1788 		if (dhpb->dthpb_base + dhpb->dthpb_offs[i] <=
1789 		    dhpb->dthpb_base + dhpb->dthpb_offs[i - 1])
1790 			return;
1791 	}
1792 
1793 	qsort(dhpb->dthpb_enoffs, dhpb->dthpb_nenoffs, sizeof (uint32_t),
1794 	    fasttrap_uint32_cmp);
1795 	for (i = 1; i < dhpb->dthpb_nenoffs; i++) {
1796 		if (dhpb->dthpb_base + dhpb->dthpb_enoffs[i] <=
1797 		    dhpb->dthpb_base + dhpb->dthpb_enoffs[i - 1])
1798 			return;
1799 	}
1800 
1801 	/*
1802 	 * Grab the creation lock to ensure consistency between calls to
1803 	 * dtrace_probe_lookup() and dtrace_probe_create() in the face of
1804 	 * other threads creating probes.
1805 	 */
1806 	mutex_enter(&provider->ftp_cmtx);
1807 
1808 	if (dtrace_probe_lookup(provider->ftp_provid, dhpb->dthpb_mod,
1809 	    dhpb->dthpb_func, dhpb->dthpb_name) != 0) {
1810 		mutex_exit(&provider->ftp_cmtx);
1811 		return;
1812 	}
1813 
1814 	ntps = dhpb->dthpb_noffs + dhpb->dthpb_nenoffs;
1815 	ASSERT(ntps > 0);
1816 
1817 	atomic_add_32(&fasttrap_total, ntps);
1818 
1819 	if (fasttrap_total > fasttrap_max) {
1820 		atomic_add_32(&fasttrap_total, -ntps);
1821 		mutex_exit(&provider->ftp_cmtx);
1822 		return;
1823 	}
1824 
1825 	pp = kmem_zalloc(offsetof(fasttrap_probe_t, ftp_tps[ntps]), KM_SLEEP);
1826 
1827 	pp->ftp_prov = provider;
1828 	pp->ftp_pid = provider->ftp_pid;
1829 	pp->ftp_ntps = ntps;
1830 	pp->ftp_nargs = dhpb->dthpb_xargc;
1831 	pp->ftp_xtypes = dhpb->dthpb_xtypes;
1832 	pp->ftp_ntypes = dhpb->dthpb_ntypes;
1833 
1834 	/*
1835 	 * First create a tracepoint for each actual point of interest.
1836 	 */
1837 	for (i = 0; i < dhpb->dthpb_noffs; i++) {
1838 		tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t), KM_SLEEP);
1839 
1840 		tp->ftt_proc = provider->ftp_proc;
1841 		tp->ftt_pc = dhpb->dthpb_base + dhpb->dthpb_offs[i];
1842 		tp->ftt_pid = provider->ftp_pid;
1843 
1844 		pp->ftp_tps[i].fit_tp = tp;
1845 		pp->ftp_tps[i].fit_id.fti_probe = pp;
1846 #ifdef __sparc
1847 		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_POST_OFFSETS;
1848 #else
1849 		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_OFFSETS;
1850 #endif
1851 	}
1852 
1853 	/*
1854 	 * Then create a tracepoint for each is-enabled point.
1855 	 */
1856 	for (j = 0; i < ntps; i++, j++) {
1857 		tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t), KM_SLEEP);
1858 
1859 		tp->ftt_proc = provider->ftp_proc;
1860 		tp->ftt_pc = dhpb->dthpb_base + dhpb->dthpb_enoffs[j];
1861 		tp->ftt_pid = provider->ftp_pid;
1862 
1863 		pp->ftp_tps[i].fit_tp = tp;
1864 		pp->ftp_tps[i].fit_id.fti_probe = pp;
1865 		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_IS_ENABLED;
1866 	}
1867 
1868 	/*
1869 	 * If the arguments are shuffled around we set the argument remapping
1870 	 * table. Later, when the probe fires, we only remap the arguments
1871 	 * if the table is non-NULL.
1872 	 */
1873 	for (i = 0; i < dhpb->dthpb_xargc; i++) {
1874 		if (dhpb->dthpb_args[i] != i) {
1875 			pp->ftp_argmap = dhpb->dthpb_args;
1876 			break;
1877 		}
1878 	}
1879 
1880 	/*
1881 	 * The probe is fully constructed -- register it with DTrace.
1882 	 */
1883 	pp->ftp_id = dtrace_probe_create(provider->ftp_provid, dhpb->dthpb_mod,
1884 	    dhpb->dthpb_func, dhpb->dthpb_name, FASTTRAP_OFFSET_AFRAMES, pp);
1885 
1886 	mutex_exit(&provider->ftp_cmtx);
1887 }
1888 
1889 /*ARGSUSED*/
1890 static void
1891 fasttrap_meta_remove(void *arg, dtrace_helper_provdesc_t *dhpv, pid_t pid)
1892 {
1893 	/*
1894 	 * Clean up the USDT provider. There may be active consumers of the
1895 	 * provider busy adding probes, no damage will actually befall the
1896 	 * provider until that count has dropped to zero. This just puts
1897 	 * the provider on death row.
1898 	 */
1899 	fasttrap_provider_retire(pid, dhpv->dthpv_provname, 1);
1900 }
1901 
1902 static dtrace_mops_t fasttrap_mops = {
1903 	fasttrap_meta_create_probe,
1904 	fasttrap_meta_provide,
1905 	fasttrap_meta_remove
1906 };
1907 
1908 /*ARGSUSED*/
1909 static int
1910 fasttrap_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
1911 {
1912 	return (0);
1913 }
1914 
1915 /*ARGSUSED*/
1916 static int
1917 fasttrap_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
1918 {
1919 	if (!dtrace_attached())
1920 		return (EAGAIN);
1921 
1922 	if (cmd == FASTTRAPIOC_MAKEPROBE) {
1923 		fasttrap_probe_spec_t *uprobe = (void *)arg;
1924 		fasttrap_probe_spec_t *probe;
1925 		uint64_t noffs;
1926 		size_t size;
1927 		int ret;
1928 		char *c;
1929 
1930 		if (copyin(&uprobe->ftps_noffs, &noffs,
1931 		    sizeof (uprobe->ftps_noffs)))
1932 			return (EFAULT);
1933 
1934 		/*
1935 		 * Probes must have at least one tracepoint.
1936 		 */
1937 		if (noffs == 0)
1938 			return (EINVAL);
1939 
1940 		size = sizeof (fasttrap_probe_spec_t) +
1941 		    sizeof (probe->ftps_offs[0]) * (noffs - 1);
1942 
1943 		if (size > 1024 * 1024)
1944 			return (ENOMEM);
1945 
1946 		probe = kmem_alloc(size, KM_SLEEP);
1947 
1948 		if (copyin(uprobe, probe, size) != 0) {
1949 			kmem_free(probe, size);
1950 			return (EFAULT);
1951 		}
1952 
1953 		/*
1954 		 * Verify that the function and module strings contain no
1955 		 * funny characters.
1956 		 */
1957 		for (c = &probe->ftps_func[0]; *c != '\0'; c++) {
1958 			if (*c < 0x20 || 0x7f <= *c) {
1959 				ret = EINVAL;
1960 				goto err;
1961 			}
1962 		}
1963 
1964 		for (c = &probe->ftps_mod[0]; *c != '\0'; c++) {
1965 			if (*c < 0x20 || 0x7f <= *c) {
1966 				ret = EINVAL;
1967 				goto err;
1968 			}
1969 		}
1970 
1971 		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
1972 			proc_t *p;
1973 			pid_t pid = probe->ftps_pid;
1974 
1975 			mutex_enter(&pidlock);
1976 			/*
1977 			 * Report an error if the process doesn't exist
1978 			 * or is actively being birthed.
1979 			 */
1980 			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
1981 				mutex_exit(&pidlock);
1982 				return (ESRCH);
1983 			}
1984 			mutex_enter(&p->p_lock);
1985 			mutex_exit(&pidlock);
1986 
1987 			if ((ret = priv_proc_cred_perm(cr, p, NULL,
1988 			    VREAD | VWRITE)) != 0) {
1989 				mutex_exit(&p->p_lock);
1990 				return (ret);
1991 			}
1992 
1993 			mutex_exit(&p->p_lock);
1994 		}
1995 
1996 		ret = fasttrap_add_probe(probe);
1997 err:
1998 		kmem_free(probe, size);
1999 
2000 		return (ret);
2001 
2002 	} else if (cmd == FASTTRAPIOC_GETINSTR) {
2003 		fasttrap_instr_query_t instr;
2004 		fasttrap_tracepoint_t *tp;
2005 		uint_t index;
2006 		int ret;
2007 
2008 		if (copyin((void *)arg, &instr, sizeof (instr)) != 0)
2009 			return (EFAULT);
2010 
2011 		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
2012 			proc_t *p;
2013 			pid_t pid = instr.ftiq_pid;
2014 
2015 			mutex_enter(&pidlock);
2016 			/*
2017 			 * Report an error if the process doesn't exist
2018 			 * or is actively being birthed.
2019 			 */
2020 			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
2021 				mutex_exit(&pidlock);
2022 				return (ESRCH);
2023 			}
2024 			mutex_enter(&p->p_lock);
2025 			mutex_exit(&pidlock);
2026 
2027 			if ((ret = priv_proc_cred_perm(cr, p, NULL,
2028 			    VREAD)) != 0) {
2029 				mutex_exit(&p->p_lock);
2030 				return (ret);
2031 			}
2032 
2033 			mutex_exit(&p->p_lock);
2034 		}
2035 
2036 		index = FASTTRAP_TPOINTS_INDEX(instr.ftiq_pid, instr.ftiq_pc);
2037 
2038 		mutex_enter(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2039 		tp = fasttrap_tpoints.fth_table[index].ftb_data;
2040 		while (tp != NULL) {
2041 			if (instr.ftiq_pid == tp->ftt_pid &&
2042 			    instr.ftiq_pc == tp->ftt_pc &&
2043 			    tp->ftt_proc->ftpc_acount != 0)
2044 				break;
2045 
2046 			tp = tp->ftt_next;
2047 		}
2048 
2049 		if (tp == NULL) {
2050 			mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2051 			return (ENOENT);
2052 		}
2053 
2054 		bcopy(&tp->ftt_instr, &instr.ftiq_instr,
2055 		    sizeof (instr.ftiq_instr));
2056 		mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2057 
2058 		if (copyout(&instr, (void *)arg, sizeof (instr)) != 0)
2059 			return (EFAULT);
2060 
2061 		return (0);
2062 	}
2063 
2064 	return (EINVAL);
2065 }
2066 
2067 static struct cb_ops fasttrap_cb_ops = {
2068 	fasttrap_open,		/* open */
2069 	nodev,			/* close */
2070 	nulldev,		/* strategy */
2071 	nulldev,		/* print */
2072 	nodev,			/* dump */
2073 	nodev,			/* read */
2074 	nodev,			/* write */
2075 	fasttrap_ioctl,		/* ioctl */
2076 	nodev,			/* devmap */
2077 	nodev,			/* mmap */
2078 	nodev,			/* segmap */
2079 	nochpoll,		/* poll */
2080 	ddi_prop_op,		/* cb_prop_op */
2081 	0,			/* streamtab  */
2082 	D_NEW | D_MP		/* Driver compatibility flag */
2083 };
2084 
2085 /*ARGSUSED*/
2086 static int
2087 fasttrap_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
2088 {
2089 	int error;
2090 
2091 	switch (infocmd) {
2092 	case DDI_INFO_DEVT2DEVINFO:
2093 		*result = (void *)fasttrap_devi;
2094 		error = DDI_SUCCESS;
2095 		break;
2096 	case DDI_INFO_DEVT2INSTANCE:
2097 		*result = (void *)0;
2098 		error = DDI_SUCCESS;
2099 		break;
2100 	default:
2101 		error = DDI_FAILURE;
2102 	}
2103 	return (error);
2104 }
2105 
2106 static int
2107 fasttrap_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
2108 {
2109 	ulong_t nent;
2110 
2111 	switch (cmd) {
2112 	case DDI_ATTACH:
2113 		break;
2114 	case DDI_RESUME:
2115 		return (DDI_SUCCESS);
2116 	default:
2117 		return (DDI_FAILURE);
2118 	}
2119 
2120 	if (ddi_create_minor_node(devi, "fasttrap", S_IFCHR, 0,
2121 	    DDI_PSEUDO, NULL) == DDI_FAILURE) {
2122 		ddi_remove_minor_node(devi, NULL);
2123 		return (DDI_FAILURE);
2124 	}
2125 
2126 	ddi_report_dev(devi);
2127 	fasttrap_devi = devi;
2128 
2129 	/*
2130 	 * Install our hooks into fork(2), exec(2), and exit(2).
2131 	 */
2132 	dtrace_fasttrap_fork_ptr = &fasttrap_fork;
2133 	dtrace_fasttrap_exit_ptr = &fasttrap_exec_exit;
2134 	dtrace_fasttrap_exec_ptr = &fasttrap_exec_exit;
2135 
2136 	fasttrap_max = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2137 	    "fasttrap-max-probes", FASTTRAP_MAX_DEFAULT);
2138 	fasttrap_total = 0;
2139 
2140 	/*
2141 	 * Conjure up the tracepoints hashtable...
2142 	 */
2143 	nent = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2144 	    "fasttrap-hash-size", FASTTRAP_TPOINTS_DEFAULT_SIZE);
2145 
2146 	if (nent == 0 || nent > 0x1000000)
2147 		nent = FASTTRAP_TPOINTS_DEFAULT_SIZE;
2148 
2149 	if ((nent & (nent - 1)) == 0)
2150 		fasttrap_tpoints.fth_nent = nent;
2151 	else
2152 		fasttrap_tpoints.fth_nent = 1 << fasttrap_highbit(nent);
2153 	ASSERT(fasttrap_tpoints.fth_nent > 0);
2154 	fasttrap_tpoints.fth_mask = fasttrap_tpoints.fth_nent - 1;
2155 	fasttrap_tpoints.fth_table = kmem_zalloc(fasttrap_tpoints.fth_nent *
2156 	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2157 
2158 	/*
2159 	 * ... and the providers hash table...
2160 	 */
2161 	nent = FASTTRAP_PROVIDERS_DEFAULT_SIZE;
2162 	if ((nent & (nent - 1)) == 0)
2163 		fasttrap_provs.fth_nent = nent;
2164 	else
2165 		fasttrap_provs.fth_nent = 1 << fasttrap_highbit(nent);
2166 	ASSERT(fasttrap_provs.fth_nent > 0);
2167 	fasttrap_provs.fth_mask = fasttrap_provs.fth_nent - 1;
2168 	fasttrap_provs.fth_table = kmem_zalloc(fasttrap_provs.fth_nent *
2169 	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2170 
2171 	/*
2172 	 * ... and the procs hash table.
2173 	 */
2174 	nent = FASTTRAP_PROCS_DEFAULT_SIZE;
2175 	if ((nent & (nent - 1)) == 0)
2176 		fasttrap_procs.fth_nent = nent;
2177 	else
2178 		fasttrap_procs.fth_nent = 1 << fasttrap_highbit(nent);
2179 	ASSERT(fasttrap_procs.fth_nent > 0);
2180 	fasttrap_procs.fth_mask = fasttrap_procs.fth_nent - 1;
2181 	fasttrap_procs.fth_table = kmem_zalloc(fasttrap_procs.fth_nent *
2182 	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2183 
2184 	(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2185 	    &fasttrap_meta_id);
2186 
2187 	return (DDI_SUCCESS);
2188 }
2189 
2190 static int
2191 fasttrap_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
2192 {
2193 	int i, fail = 0;
2194 	timeout_id_t tmp;
2195 
2196 	switch (cmd) {
2197 	case DDI_DETACH:
2198 		break;
2199 	case DDI_SUSPEND:
2200 		return (DDI_SUCCESS);
2201 	default:
2202 		return (DDI_FAILURE);
2203 	}
2204 
2205 	/*
2206 	 * Unregister the meta-provider to make sure no new fasttrap-
2207 	 * managed providers come along while we're trying to close up
2208 	 * shop. If we fail to detach, we'll need to re-register as a
2209 	 * meta-provider. We can fail to unregister as a meta-provider
2210 	 * if providers we manage still exist.
2211 	 */
2212 	if (fasttrap_meta_id != DTRACE_METAPROVNONE &&
2213 	    dtrace_meta_unregister(fasttrap_meta_id) != 0)
2214 		return (DDI_FAILURE);
2215 
2216 	/*
2217 	 * Prevent any new timeouts from running by setting fasttrap_timeout
2218 	 * to a non-zero value, and wait for the current timeout to complete.
2219 	 */
2220 	mutex_enter(&fasttrap_cleanup_mtx);
2221 	fasttrap_cleanup_work = 0;
2222 
2223 	while (fasttrap_timeout != (timeout_id_t)1) {
2224 		tmp = fasttrap_timeout;
2225 		fasttrap_timeout = (timeout_id_t)1;
2226 
2227 		if (tmp != 0) {
2228 			mutex_exit(&fasttrap_cleanup_mtx);
2229 			(void) untimeout(tmp);
2230 			mutex_enter(&fasttrap_cleanup_mtx);
2231 		}
2232 	}
2233 
2234 	fasttrap_cleanup_work = 0;
2235 	mutex_exit(&fasttrap_cleanup_mtx);
2236 
2237 	/*
2238 	 * Iterate over all of our providers. If there's still a process
2239 	 * that corresponds to that pid, fail to detach.
2240 	 */
2241 	for (i = 0; i < fasttrap_provs.fth_nent; i++) {
2242 		fasttrap_provider_t **fpp, *fp;
2243 		fasttrap_bucket_t *bucket = &fasttrap_provs.fth_table[i];
2244 
2245 		mutex_enter(&bucket->ftb_mtx);
2246 		fpp = (fasttrap_provider_t **)&bucket->ftb_data;
2247 		while ((fp = *fpp) != NULL) {
2248 			/*
2249 			 * Acquire and release the lock as a simple way of
2250 			 * waiting for any other consumer to finish with
2251 			 * this provider. A thread must first acquire the
2252 			 * bucket lock so there's no chance of another thread
2253 			 * blocking on the provider's lock.
2254 			 */
2255 			mutex_enter(&fp->ftp_mtx);
2256 			mutex_exit(&fp->ftp_mtx);
2257 
2258 			if (dtrace_unregister(fp->ftp_provid) != 0) {
2259 				fail = 1;
2260 				fpp = &fp->ftp_next;
2261 			} else {
2262 				*fpp = fp->ftp_next;
2263 				fasttrap_provider_free(fp);
2264 			}
2265 		}
2266 
2267 		mutex_exit(&bucket->ftb_mtx);
2268 	}
2269 
2270 	if (fail) {
2271 		uint_t work;
2272 		/*
2273 		 * If we're failing to detach, we need to unblock timeouts
2274 		 * and start a new timeout if any work has accumulated while
2275 		 * we've been unsuccessfully trying to detach.
2276 		 */
2277 		mutex_enter(&fasttrap_cleanup_mtx);
2278 		fasttrap_timeout = 0;
2279 		work = fasttrap_cleanup_work;
2280 		mutex_exit(&fasttrap_cleanup_mtx);
2281 
2282 		if (work)
2283 			fasttrap_pid_cleanup();
2284 
2285 		(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2286 		    &fasttrap_meta_id);
2287 
2288 		return (DDI_FAILURE);
2289 	}
2290 
2291 #ifdef DEBUG
2292 	mutex_enter(&fasttrap_count_mtx);
2293 	ASSERT(fasttrap_pid_count == 0);
2294 	mutex_exit(&fasttrap_count_mtx);
2295 #endif
2296 
2297 	kmem_free(fasttrap_tpoints.fth_table,
2298 	    fasttrap_tpoints.fth_nent * sizeof (fasttrap_bucket_t));
2299 	fasttrap_tpoints.fth_nent = 0;
2300 
2301 	kmem_free(fasttrap_provs.fth_table,
2302 	    fasttrap_provs.fth_nent * sizeof (fasttrap_bucket_t));
2303 	fasttrap_provs.fth_nent = 0;
2304 
2305 	kmem_free(fasttrap_procs.fth_table,
2306 	    fasttrap_procs.fth_nent * sizeof (fasttrap_bucket_t));
2307 	fasttrap_procs.fth_nent = 0;
2308 
2309 	/*
2310 	 * We know there are no tracepoints in any process anywhere in
2311 	 * the system so there is no process which has its p_dtrace_count
2312 	 * greater than zero, therefore we know that no thread can actively
2313 	 * be executing code in fasttrap_fork(). Similarly for p_dtrace_probes
2314 	 * and fasttrap_exec() and fasttrap_exit().
2315 	 */
2316 	ASSERT(dtrace_fasttrap_fork_ptr == &fasttrap_fork);
2317 	dtrace_fasttrap_fork_ptr = NULL;
2318 
2319 	ASSERT(dtrace_fasttrap_exec_ptr == &fasttrap_exec_exit);
2320 	dtrace_fasttrap_exec_ptr = NULL;
2321 
2322 	ASSERT(dtrace_fasttrap_exit_ptr == &fasttrap_exec_exit);
2323 	dtrace_fasttrap_exit_ptr = NULL;
2324 
2325 	ddi_remove_minor_node(devi, NULL);
2326 
2327 	return (DDI_SUCCESS);
2328 }
2329 
2330 static struct dev_ops fasttrap_ops = {
2331 	DEVO_REV,		/* devo_rev */
2332 	0,			/* refcnt */
2333 	fasttrap_info,		/* get_dev_info */
2334 	nulldev,		/* identify */
2335 	nulldev,		/* probe */
2336 	fasttrap_attach,	/* attach */
2337 	fasttrap_detach,	/* detach */
2338 	nodev,			/* reset */
2339 	&fasttrap_cb_ops,	/* driver operations */
2340 	NULL,			/* bus operations */
2341 	nodev,			/* dev power */
2342 	ddi_quiesce_not_needed,		/* quiesce */
2343 };
2344 
2345 /*
2346  * Module linkage information for the kernel.
2347  */
2348 static struct modldrv modldrv = {
2349 	&mod_driverops,		/* module type (this is a pseudo driver) */
2350 	"Fasttrap Tracing",	/* name of module */
2351 	&fasttrap_ops,		/* driver ops */
2352 };
2353 
2354 static struct modlinkage modlinkage = {
2355 	MODREV_1,
2356 	(void *)&modldrv,
2357 	NULL
2358 };
2359 
2360 int
2361 _init(void)
2362 {
2363 	return (mod_install(&modlinkage));
2364 }
2365 
2366 int
2367 _info(struct modinfo *modinfop)
2368 {
2369 	return (mod_info(&modlinkage, modinfop));
2370 }
2371 
2372 int
2373 _fini(void)
2374 {
2375 	return (mod_remove(&modlinkage));
2376 }
2377