xref: /illumos-gate/usr/src/uts/common/dtrace/fasttrap.c (revision 598f4ceed9327d2d6c2325dd67cae3aa06f7fea6)
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  * Copyright (c) 2013, Joyent, Inc. All rights reserved.
29  */
30 
31 #include <sys/atomic.h>
32 #include <sys/errno.h>
33 #include <sys/stat.h>
34 #include <sys/modctl.h>
35 #include <sys/conf.h>
36 #include <sys/systm.h>
37 #include <sys/ddi.h>
38 #include <sys/sunddi.h>
39 #include <sys/cpuvar.h>
40 #include <sys/kmem.h>
41 #include <sys/strsubr.h>
42 #include <sys/fasttrap.h>
43 #include <sys/fasttrap_impl.h>
44 #include <sys/fasttrap_isa.h>
45 #include <sys/dtrace.h>
46 #include <sys/dtrace_impl.h>
47 #include <sys/sysmacros.h>
48 #include <sys/proc.h>
49 #include <sys/priv.h>
50 #include <sys/policy.h>
51 #include <util/qsort.h>
52 
53 /*
54  * User-Land Trap-Based Tracing
55  * ----------------------------
56  *
57  * The fasttrap provider allows DTrace consumers to instrument any user-level
58  * instruction to gather data; this includes probes with semantic
59  * signifigance like entry and return as well as simple offsets into the
60  * function. While the specific techniques used are very ISA specific, the
61  * methodology is generalizable to any architecture.
62  *
63  *
64  * The General Methodology
65  * -----------------------
66  *
67  * With the primary goal of tracing every user-land instruction and the
68  * limitation that we can't trust user space so don't want to rely on much
69  * information there, we begin by replacing the instructions we want to trace
70  * with trap instructions. Each instruction we overwrite is saved into a hash
71  * table keyed by process ID and pc address. When we enter the kernel due to
72  * this trap instruction, we need the effects of the replaced instruction to
73  * appear to have occurred before we proceed with the user thread's
74  * execution.
75  *
76  * Each user level thread is represented by a ulwp_t structure which is
77  * always easily accessible through a register. The most basic way to produce
78  * the effects of the instruction we replaced is to copy that instruction out
79  * to a bit of scratch space reserved in the user thread's ulwp_t structure
80  * (a sort of kernel-private thread local storage), set the PC to that
81  * scratch space and single step. When we reenter the kernel after single
82  * stepping the instruction we must then adjust the PC to point to what would
83  * normally be the next instruction. Of course, special care must be taken
84  * for branches and jumps, but these represent such a small fraction of any
85  * instruction set that writing the code to emulate these in the kernel is
86  * not too difficult.
87  *
88  * Return probes may require several tracepoints to trace every return site,
89  * and, conversely, each tracepoint may activate several probes (the entry
90  * and offset 0 probes, for example). To solve this muliplexing problem,
91  * tracepoints contain lists of probes to activate and probes contain lists
92  * of tracepoints to enable. If a probe is activated, it adds its ID to
93  * existing tracepoints or creates new ones as necessary.
94  *
95  * Most probes are activated _before_ the instruction is executed, but return
96  * probes are activated _after_ the effects of the last instruction of the
97  * function are visible. Return probes must be fired _after_ we have
98  * single-stepped the instruction whereas all other probes are fired
99  * beforehand.
100  *
101  *
102  * Lock Ordering
103  * -------------
104  *
105  * The lock ordering below -- both internally and with respect to the DTrace
106  * framework -- is a little tricky and bears some explanation. Each provider
107  * has a lock (ftp_mtx) that protects its members including reference counts
108  * for enabled probes (ftp_rcount), consumers actively creating probes
109  * (ftp_ccount) and USDT consumers (ftp_mcount); all three prevent a provider
110  * from being freed. A provider is looked up by taking the bucket lock for the
111  * provider hash table, and is returned with its lock held. The provider lock
112  * may be taken in functions invoked by the DTrace framework, but may not be
113  * held while calling functions in the DTrace framework.
114  *
115  * To ensure consistency over multiple calls to the DTrace framework, the
116  * creation lock (ftp_cmtx) should be held. Naturally, the creation lock may
117  * not be taken when holding the provider lock as that would create a cyclic
118  * lock ordering. In situations where one would naturally take the provider
119  * lock and then the creation lock, we instead up a reference count to prevent
120  * the provider from disappearing, drop the provider lock, and acquire the
121  * creation lock.
122  *
123  * Briefly:
124  * 	bucket lock before provider lock
125  *	DTrace before provider lock
126  *	creation lock before DTrace
127  *	never hold the provider lock and creation lock simultaneously
128  */
129 
130 static dev_info_t *fasttrap_devi;
131 static dtrace_meta_provider_id_t fasttrap_meta_id;
132 
133 static timeout_id_t fasttrap_timeout;
134 static kmutex_t fasttrap_cleanup_mtx;
135 static uint_t fasttrap_cleanup_work;
136 
137 /*
138  * Generation count on modifications to the global tracepoint lookup table.
139  */
140 static volatile uint64_t fasttrap_mod_gen;
141 
142 /*
143  * When the fasttrap provider is loaded, fasttrap_max is set to either
144  * FASTTRAP_MAX_DEFAULT or the value for fasttrap-max-probes in the
145  * fasttrap.conf file. Each time a probe is created, fasttrap_total is
146  * incremented by the number of tracepoints that may be associated with that
147  * probe; fasttrap_total is capped at fasttrap_max.
148  */
149 #define	FASTTRAP_MAX_DEFAULT		250000
150 static uint32_t fasttrap_max;
151 static uint32_t fasttrap_total;
152 
153 
154 #define	FASTTRAP_TPOINTS_DEFAULT_SIZE	0x4000
155 #define	FASTTRAP_PROVIDERS_DEFAULT_SIZE	0x100
156 #define	FASTTRAP_PROCS_DEFAULT_SIZE	0x100
157 
158 #define	FASTTRAP_PID_NAME		"pid"
159 
160 fasttrap_hash_t			fasttrap_tpoints;
161 static fasttrap_hash_t		fasttrap_provs;
162 static fasttrap_hash_t		fasttrap_procs;
163 
164 static uint64_t			fasttrap_pid_count;	/* pid ref count */
165 static kmutex_t			fasttrap_count_mtx;	/* lock on ref count */
166 
167 #define	FASTTRAP_ENABLE_FAIL	1
168 #define	FASTTRAP_ENABLE_PARTIAL	2
169 
170 static int fasttrap_tracepoint_enable(proc_t *, fasttrap_probe_t *, uint_t);
171 static void fasttrap_tracepoint_disable(proc_t *, fasttrap_probe_t *, uint_t);
172 
173 static fasttrap_provider_t *fasttrap_provider_lookup(pid_t, const char *,
174     const dtrace_pattr_t *);
175 static void fasttrap_provider_retire(pid_t, const char *, int);
176 static void fasttrap_provider_free(fasttrap_provider_t *);
177 
178 static fasttrap_proc_t *fasttrap_proc_lookup(pid_t);
179 static void fasttrap_proc_release(fasttrap_proc_t *);
180 
181 #define	FASTTRAP_PROVS_INDEX(pid, name) \
182 	((fasttrap_hash_str(name) + (pid)) & fasttrap_provs.fth_mask)
183 
184 #define	FASTTRAP_PROCS_INDEX(pid) ((pid) & fasttrap_procs.fth_mask)
185 
186 static int
187 fasttrap_highbit(ulong_t i)
188 {
189 	int h = 1;
190 
191 	if (i == 0)
192 		return (0);
193 #ifdef _LP64
194 	if (i & 0xffffffff00000000ul) {
195 		h += 32; i >>= 32;
196 	}
197 #endif
198 	if (i & 0xffff0000) {
199 		h += 16; i >>= 16;
200 	}
201 	if (i & 0xff00) {
202 		h += 8; i >>= 8;
203 	}
204 	if (i & 0xf0) {
205 		h += 4; i >>= 4;
206 	}
207 	if (i & 0xc) {
208 		h += 2; i >>= 2;
209 	}
210 	if (i & 0x2) {
211 		h += 1;
212 	}
213 	return (h);
214 }
215 
216 static uint_t
217 fasttrap_hash_str(const char *p)
218 {
219 	unsigned int g;
220 	uint_t hval = 0;
221 
222 	while (*p) {
223 		hval = (hval << 4) + *p++;
224 		if ((g = (hval & 0xf0000000)) != 0)
225 			hval ^= g >> 24;
226 		hval &= ~g;
227 	}
228 	return (hval);
229 }
230 
231 void
232 fasttrap_sigtrap(proc_t *p, kthread_t *t, uintptr_t pc)
233 {
234 	sigqueue_t *sqp = kmem_zalloc(sizeof (sigqueue_t), KM_SLEEP);
235 
236 	sqp->sq_info.si_signo = SIGTRAP;
237 	sqp->sq_info.si_code = TRAP_DTRACE;
238 	sqp->sq_info.si_addr = (caddr_t)pc;
239 
240 	mutex_enter(&p->p_lock);
241 	sigaddqa(p, t, sqp);
242 	mutex_exit(&p->p_lock);
243 
244 	if (t != NULL)
245 		aston(t);
246 }
247 
248 /*
249  * This function ensures that no threads are actively using the memory
250  * associated with probes that were formerly live.
251  */
252 static void
253 fasttrap_mod_barrier(uint64_t gen)
254 {
255 	int i;
256 
257 	if (gen < fasttrap_mod_gen)
258 		return;
259 
260 	fasttrap_mod_gen++;
261 
262 	for (i = 0; i < NCPU; i++) {
263 		mutex_enter(&cpu_core[i].cpuc_pid_lock);
264 		mutex_exit(&cpu_core[i].cpuc_pid_lock);
265 	}
266 }
267 
268 /*
269  * This is the timeout's callback for cleaning up the providers and their
270  * probes.
271  */
272 /*ARGSUSED*/
273 static void
274 fasttrap_pid_cleanup_cb(void *data)
275 {
276 	fasttrap_provider_t **fpp, *fp;
277 	fasttrap_bucket_t *bucket;
278 	dtrace_provider_id_t provid;
279 	int i, later, rval;
280 
281 	static volatile int in = 0;
282 	ASSERT(in == 0);
283 	in = 1;
284 
285 	mutex_enter(&fasttrap_cleanup_mtx);
286 	while (fasttrap_cleanup_work) {
287 		fasttrap_cleanup_work = 0;
288 		mutex_exit(&fasttrap_cleanup_mtx);
289 
290 		later = 0;
291 
292 		/*
293 		 * Iterate over all the providers trying to remove the marked
294 		 * ones. If a provider is marked but not retired, we just
295 		 * have to take a crack at removing it -- it's no big deal if
296 		 * we can't.
297 		 */
298 		for (i = 0; i < fasttrap_provs.fth_nent; i++) {
299 			bucket = &fasttrap_provs.fth_table[i];
300 			mutex_enter(&bucket->ftb_mtx);
301 			fpp = (fasttrap_provider_t **)&bucket->ftb_data;
302 
303 			while ((fp = *fpp) != NULL) {
304 				if (!fp->ftp_marked) {
305 					fpp = &fp->ftp_next;
306 					continue;
307 				}
308 
309 				mutex_enter(&fp->ftp_mtx);
310 
311 				/*
312 				 * If this provider has consumers actively
313 				 * creating probes (ftp_ccount) or is a USDT
314 				 * provider (ftp_mcount), we can't unregister
315 				 * or even condense.
316 				 */
317 				if (fp->ftp_ccount != 0 ||
318 				    fp->ftp_mcount != 0) {
319 					mutex_exit(&fp->ftp_mtx);
320 					fp->ftp_marked = 0;
321 					continue;
322 				}
323 
324 				if (!fp->ftp_retired || fp->ftp_rcount != 0)
325 					fp->ftp_marked = 0;
326 
327 				mutex_exit(&fp->ftp_mtx);
328 
329 				/*
330 				 * If we successfully unregister this
331 				 * provider we can remove it from the hash
332 				 * chain and free the memory. If our attempt
333 				 * to unregister fails and this is a retired
334 				 * provider, increment our flag to try again
335 				 * pretty soon. If we've consumed more than
336 				 * half of our total permitted number of
337 				 * probes call dtrace_condense() to try to
338 				 * clean out the unenabled probes.
339 				 */
340 				provid = fp->ftp_provid;
341 				if ((rval = dtrace_unregister(provid)) != 0) {
342 					if (fasttrap_total > fasttrap_max / 2)
343 						(void) dtrace_condense(provid);
344 
345 					if (rval == EAGAIN)
346 						fp->ftp_marked = 1;
347 
348 					later += fp->ftp_marked;
349 					fpp = &fp->ftp_next;
350 				} else {
351 					*fpp = fp->ftp_next;
352 					fasttrap_provider_free(fp);
353 				}
354 			}
355 			mutex_exit(&bucket->ftb_mtx);
356 		}
357 
358 		mutex_enter(&fasttrap_cleanup_mtx);
359 	}
360 
361 	ASSERT(fasttrap_timeout != 0);
362 
363 	/*
364 	 * If we were unable to remove a retired provider, try again after
365 	 * a second. This situation can occur in certain circumstances where
366 	 * providers cannot be unregistered even though they have no probes
367 	 * enabled because of an execution of dtrace -l or something similar.
368 	 * If the timeout has been disabled (set to 1 because we're trying
369 	 * to detach), we set fasttrap_cleanup_work to ensure that we'll
370 	 * get a chance to do that work if and when the timeout is reenabled
371 	 * (if detach fails).
372 	 */
373 	if (later > 0) {
374 		if (fasttrap_timeout != (timeout_id_t)1) {
375 			fasttrap_timeout =
376 			    timeout(&fasttrap_pid_cleanup_cb, NULL, hz);
377 		}
378 
379 		fasttrap_cleanup_work = 1;
380 	} else {
381 		fasttrap_timeout = 0;
382 	}
383 
384 	mutex_exit(&fasttrap_cleanup_mtx);
385 	in = 0;
386 }
387 
388 /*
389  * Activates the asynchronous cleanup mechanism.
390  */
391 static void
392 fasttrap_pid_cleanup(void)
393 {
394 	mutex_enter(&fasttrap_cleanup_mtx);
395 	fasttrap_cleanup_work = 1;
396 	if (fasttrap_timeout == 0)
397 		fasttrap_timeout = timeout(&fasttrap_pid_cleanup_cb, NULL, 1);
398 	mutex_exit(&fasttrap_cleanup_mtx);
399 }
400 
401 /*
402  * This is called from cfork() via dtrace_fasttrap_fork(). The child
403  * process's address space is (roughly) a copy of the parent process's so
404  * we have to remove all the instrumentation we had previously enabled in the
405  * parent.
406  */
407 static void
408 fasttrap_fork(proc_t *p, proc_t *cp)
409 {
410 	pid_t ppid = p->p_pid;
411 	int i;
412 
413 	ASSERT(curproc == p);
414 	ASSERT(p->p_proc_flag & P_PR_LOCK);
415 	ASSERT(p->p_dtrace_count > 0);
416 	ASSERT(cp->p_dtrace_count == 0);
417 
418 	/*
419 	 * This would be simpler and faster if we maintained per-process
420 	 * hash tables of enabled tracepoints. It could, however, potentially
421 	 * slow down execution of a tracepoint since we'd need to go
422 	 * through two levels of indirection. In the future, we should
423 	 * consider either maintaining per-process ancillary lists of
424 	 * enabled tracepoints or hanging a pointer to a per-process hash
425 	 * table of enabled tracepoints off the proc structure.
426 	 */
427 
428 	/*
429 	 * We don't have to worry about the child process disappearing
430 	 * because we're in fork().
431 	 */
432 	mutex_enter(&cp->p_lock);
433 	sprlock_proc(cp);
434 	mutex_exit(&cp->p_lock);
435 
436 	/*
437 	 * Iterate over every tracepoint looking for ones that belong to the
438 	 * parent process, and remove each from the child process.
439 	 */
440 	for (i = 0; i < fasttrap_tpoints.fth_nent; i++) {
441 		fasttrap_tracepoint_t *tp;
442 		fasttrap_bucket_t *bucket = &fasttrap_tpoints.fth_table[i];
443 
444 		mutex_enter(&bucket->ftb_mtx);
445 		for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
446 			if (tp->ftt_pid == ppid &&
447 			    tp->ftt_proc->ftpc_acount != 0) {
448 				int ret = fasttrap_tracepoint_remove(cp, tp);
449 				ASSERT(ret == 0);
450 
451 				/*
452 				 * The count of active providers can only be
453 				 * decremented (i.e. to zero) during exec,
454 				 * exit, and removal of a meta provider so it
455 				 * should be impossible to drop the count
456 				 * mid-fork.
457 				 */
458 				ASSERT(tp->ftt_proc->ftpc_acount != 0);
459 			}
460 		}
461 		mutex_exit(&bucket->ftb_mtx);
462 	}
463 
464 	mutex_enter(&cp->p_lock);
465 	sprunlock(cp);
466 }
467 
468 /*
469  * This is called from proc_exit() or from exec_common() if p_dtrace_probes
470  * is set on the proc structure to indicate that there is a pid provider
471  * associated with this process.
472  */
473 static void
474 fasttrap_exec_exit(proc_t *p)
475 {
476 	ASSERT(p == curproc);
477 	ASSERT(MUTEX_HELD(&p->p_lock));
478 
479 	mutex_exit(&p->p_lock);
480 
481 	/*
482 	 * We clean up the pid provider for this process here; user-land
483 	 * static probes are handled by the meta-provider remove entry point.
484 	 */
485 	fasttrap_provider_retire(p->p_pid, FASTTRAP_PID_NAME, 0);
486 
487 	mutex_enter(&p->p_lock);
488 }
489 
490 
491 /*ARGSUSED*/
492 static void
493 fasttrap_pid_provide(void *arg, const dtrace_probedesc_t *desc)
494 {
495 	/*
496 	 * There are no "default" pid probes.
497 	 */
498 }
499 
500 static int
501 fasttrap_tracepoint_enable(proc_t *p, fasttrap_probe_t *probe, uint_t index)
502 {
503 	fasttrap_tracepoint_t *tp, *new_tp = NULL;
504 	fasttrap_bucket_t *bucket;
505 	fasttrap_id_t *id;
506 	pid_t pid;
507 	uintptr_t pc;
508 
509 	ASSERT(index < probe->ftp_ntps);
510 
511 	pid = probe->ftp_pid;
512 	pc = probe->ftp_tps[index].fit_tp->ftt_pc;
513 	id = &probe->ftp_tps[index].fit_id;
514 
515 	ASSERT(probe->ftp_tps[index].fit_tp->ftt_pid == pid);
516 
517 	ASSERT(!(p->p_flag & SVFORK));
518 
519 	/*
520 	 * Before we make any modifications, make sure we've imposed a barrier
521 	 * on the generation in which this probe was last modified.
522 	 */
523 	fasttrap_mod_barrier(probe->ftp_gen);
524 
525 	bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
526 
527 	/*
528 	 * If the tracepoint has already been enabled, just add our id to the
529 	 * list of interested probes. This may be our second time through
530 	 * this path in which case we'll have constructed the tracepoint we'd
531 	 * like to install. If we can't find a match, and have an allocated
532 	 * tracepoint ready to go, enable that one now.
533 	 *
534 	 * A tracepoint whose process is defunct is also considered defunct.
535 	 */
536 again:
537 	mutex_enter(&bucket->ftb_mtx);
538 	for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
539 		/*
540 		 * Note that it's safe to access the active count on the
541 		 * associated proc structure because we know that at least one
542 		 * provider (this one) will still be around throughout this
543 		 * operation.
544 		 */
545 		if (tp->ftt_pid != pid || tp->ftt_pc != pc ||
546 		    tp->ftt_proc->ftpc_acount == 0)
547 			continue;
548 
549 		/*
550 		 * Now that we've found a matching tracepoint, it would be
551 		 * a decent idea to confirm that the tracepoint is still
552 		 * enabled and the trap instruction hasn't been overwritten.
553 		 * Since this is a little hairy, we'll punt for now.
554 		 */
555 
556 		/*
557 		 * This can't be the first interested probe. We don't have
558 		 * to worry about another thread being in the midst of
559 		 * deleting this tracepoint (which would be the only valid
560 		 * reason for a tracepoint to have no interested probes)
561 		 * since we're holding P_PR_LOCK for this process.
562 		 */
563 		ASSERT(tp->ftt_ids != NULL || tp->ftt_retids != NULL);
564 
565 		switch (id->fti_ptype) {
566 		case DTFTP_ENTRY:
567 		case DTFTP_OFFSETS:
568 		case DTFTP_IS_ENABLED:
569 			id->fti_next = tp->ftt_ids;
570 			membar_producer();
571 			tp->ftt_ids = id;
572 			membar_producer();
573 			break;
574 
575 		case DTFTP_RETURN:
576 		case DTFTP_POST_OFFSETS:
577 			id->fti_next = tp->ftt_retids;
578 			membar_producer();
579 			tp->ftt_retids = id;
580 			membar_producer();
581 			break;
582 
583 		default:
584 			ASSERT(0);
585 		}
586 
587 		mutex_exit(&bucket->ftb_mtx);
588 
589 		if (new_tp != NULL) {
590 			new_tp->ftt_ids = NULL;
591 			new_tp->ftt_retids = NULL;
592 		}
593 
594 		return (0);
595 	}
596 
597 	/*
598 	 * If we have a good tracepoint ready to go, install it now while
599 	 * we have the lock held and no one can screw with us.
600 	 */
601 	if (new_tp != NULL) {
602 		int rc = 0;
603 
604 		new_tp->ftt_next = bucket->ftb_data;
605 		membar_producer();
606 		bucket->ftb_data = new_tp;
607 		membar_producer();
608 		mutex_exit(&bucket->ftb_mtx);
609 
610 		/*
611 		 * Activate the tracepoint in the ISA-specific manner.
612 		 * If this fails, we need to report the failure, but
613 		 * indicate that this tracepoint must still be disabled
614 		 * by calling fasttrap_tracepoint_disable().
615 		 */
616 		if (fasttrap_tracepoint_install(p, new_tp) != 0)
617 			rc = FASTTRAP_ENABLE_PARTIAL;
618 
619 		/*
620 		 * Increment the count of the number of tracepoints active in
621 		 * the victim process.
622 		 */
623 		ASSERT(p->p_proc_flag & P_PR_LOCK);
624 		p->p_dtrace_count++;
625 
626 		return (rc);
627 	}
628 
629 	mutex_exit(&bucket->ftb_mtx);
630 
631 	/*
632 	 * Initialize the tracepoint that's been preallocated with the probe.
633 	 */
634 	new_tp = probe->ftp_tps[index].fit_tp;
635 
636 	ASSERT(new_tp->ftt_pid == pid);
637 	ASSERT(new_tp->ftt_pc == pc);
638 	ASSERT(new_tp->ftt_proc == probe->ftp_prov->ftp_proc);
639 	ASSERT(new_tp->ftt_ids == NULL);
640 	ASSERT(new_tp->ftt_retids == NULL);
641 
642 	switch (id->fti_ptype) {
643 	case DTFTP_ENTRY:
644 	case DTFTP_OFFSETS:
645 	case DTFTP_IS_ENABLED:
646 		id->fti_next = NULL;
647 		new_tp->ftt_ids = id;
648 		break;
649 
650 	case DTFTP_RETURN:
651 	case DTFTP_POST_OFFSETS:
652 		id->fti_next = NULL;
653 		new_tp->ftt_retids = id;
654 		break;
655 
656 	default:
657 		ASSERT(0);
658 	}
659 
660 	/*
661 	 * If the ISA-dependent initialization goes to plan, go back to the
662 	 * beginning and try to install this freshly made tracepoint.
663 	 */
664 	if (fasttrap_tracepoint_init(p, new_tp, pc, id->fti_ptype) == 0)
665 		goto again;
666 
667 	new_tp->ftt_ids = NULL;
668 	new_tp->ftt_retids = NULL;
669 
670 	return (FASTTRAP_ENABLE_FAIL);
671 }
672 
673 static void
674 fasttrap_tracepoint_disable(proc_t *p, fasttrap_probe_t *probe, uint_t index)
675 {
676 	fasttrap_bucket_t *bucket;
677 	fasttrap_provider_t *provider = probe->ftp_prov;
678 	fasttrap_tracepoint_t **pp, *tp;
679 	fasttrap_id_t *id, **idp;
680 	pid_t pid;
681 	uintptr_t pc;
682 
683 	ASSERT(index < probe->ftp_ntps);
684 
685 	pid = probe->ftp_pid;
686 	pc = probe->ftp_tps[index].fit_tp->ftt_pc;
687 	id = &probe->ftp_tps[index].fit_id;
688 
689 	ASSERT(probe->ftp_tps[index].fit_tp->ftt_pid == pid);
690 
691 	/*
692 	 * Find the tracepoint and make sure that our id is one of the
693 	 * ones registered with it.
694 	 */
695 	bucket = &fasttrap_tpoints.fth_table[FASTTRAP_TPOINTS_INDEX(pid, pc)];
696 	mutex_enter(&bucket->ftb_mtx);
697 	for (tp = bucket->ftb_data; tp != NULL; tp = tp->ftt_next) {
698 		if (tp->ftt_pid == pid && tp->ftt_pc == pc &&
699 		    tp->ftt_proc == provider->ftp_proc)
700 			break;
701 	}
702 
703 	/*
704 	 * If we somehow lost this tracepoint, we're in a world of hurt.
705 	 */
706 	ASSERT(tp != NULL);
707 
708 	switch (id->fti_ptype) {
709 	case DTFTP_ENTRY:
710 	case DTFTP_OFFSETS:
711 	case DTFTP_IS_ENABLED:
712 		ASSERT(tp->ftt_ids != NULL);
713 		idp = &tp->ftt_ids;
714 		break;
715 
716 	case DTFTP_RETURN:
717 	case DTFTP_POST_OFFSETS:
718 		ASSERT(tp->ftt_retids != NULL);
719 		idp = &tp->ftt_retids;
720 		break;
721 
722 	default:
723 		ASSERT(0);
724 	}
725 
726 	while ((*idp)->fti_probe != probe) {
727 		idp = &(*idp)->fti_next;
728 		ASSERT(*idp != NULL);
729 	}
730 
731 	id = *idp;
732 	*idp = id->fti_next;
733 	membar_producer();
734 
735 	ASSERT(id->fti_probe == probe);
736 
737 	/*
738 	 * If there are other registered enablings of this tracepoint, we're
739 	 * all done, but if this was the last probe assocated with this
740 	 * this tracepoint, we need to remove and free it.
741 	 */
742 	if (tp->ftt_ids != NULL || tp->ftt_retids != NULL) {
743 
744 		/*
745 		 * If the current probe's tracepoint is in use, swap it
746 		 * for an unused tracepoint.
747 		 */
748 		if (tp == probe->ftp_tps[index].fit_tp) {
749 			fasttrap_probe_t *tmp_probe;
750 			fasttrap_tracepoint_t **tmp_tp;
751 			uint_t tmp_index;
752 
753 			if (tp->ftt_ids != NULL) {
754 				tmp_probe = tp->ftt_ids->fti_probe;
755 				/* LINTED - alignment */
756 				tmp_index = FASTTRAP_ID_INDEX(tp->ftt_ids);
757 				tmp_tp = &tmp_probe->ftp_tps[tmp_index].fit_tp;
758 			} else {
759 				tmp_probe = tp->ftt_retids->fti_probe;
760 				/* LINTED - alignment */
761 				tmp_index = FASTTRAP_ID_INDEX(tp->ftt_retids);
762 				tmp_tp = &tmp_probe->ftp_tps[tmp_index].fit_tp;
763 			}
764 
765 			ASSERT(*tmp_tp != NULL);
766 			ASSERT(*tmp_tp != probe->ftp_tps[index].fit_tp);
767 			ASSERT((*tmp_tp)->ftt_ids == NULL);
768 			ASSERT((*tmp_tp)->ftt_retids == NULL);
769 
770 			probe->ftp_tps[index].fit_tp = *tmp_tp;
771 			*tmp_tp = tp;
772 		}
773 
774 		mutex_exit(&bucket->ftb_mtx);
775 
776 		/*
777 		 * Tag the modified probe with the generation in which it was
778 		 * changed.
779 		 */
780 		probe->ftp_gen = fasttrap_mod_gen;
781 		return;
782 	}
783 
784 	mutex_exit(&bucket->ftb_mtx);
785 
786 	/*
787 	 * We can't safely remove the tracepoint from the set of active
788 	 * tracepoints until we've actually removed the fasttrap instruction
789 	 * from the process's text. We can, however, operate on this
790 	 * tracepoint secure in the knowledge that no other thread is going to
791 	 * be looking at it since we hold P_PR_LOCK on the process if it's
792 	 * live or we hold the provider lock on the process if it's dead and
793 	 * gone.
794 	 */
795 
796 	/*
797 	 * We only need to remove the actual instruction if we're looking
798 	 * at an existing process
799 	 */
800 	if (p != NULL) {
801 		/*
802 		 * If we fail to restore the instruction we need to kill
803 		 * this process since it's in a completely unrecoverable
804 		 * state.
805 		 */
806 		if (fasttrap_tracepoint_remove(p, tp) != 0)
807 			fasttrap_sigtrap(p, NULL, pc);
808 
809 		/*
810 		 * Decrement the count of the number of tracepoints active
811 		 * in the victim process.
812 		 */
813 		ASSERT(p->p_proc_flag & P_PR_LOCK);
814 		p->p_dtrace_count--;
815 	}
816 
817 	/*
818 	 * Remove the probe from the hash table of active tracepoints.
819 	 */
820 	mutex_enter(&bucket->ftb_mtx);
821 	pp = (fasttrap_tracepoint_t **)&bucket->ftb_data;
822 	ASSERT(*pp != NULL);
823 	while (*pp != tp) {
824 		pp = &(*pp)->ftt_next;
825 		ASSERT(*pp != NULL);
826 	}
827 
828 	*pp = tp->ftt_next;
829 	membar_producer();
830 
831 	mutex_exit(&bucket->ftb_mtx);
832 
833 	/*
834 	 * Tag the modified probe with the generation in which it was changed.
835 	 */
836 	probe->ftp_gen = fasttrap_mod_gen;
837 }
838 
839 static void
840 fasttrap_enable_callbacks(void)
841 {
842 	/*
843 	 * We don't have to play the rw lock game here because we're
844 	 * providing something rather than taking something away --
845 	 * we can be sure that no threads have tried to follow this
846 	 * function pointer yet.
847 	 */
848 	mutex_enter(&fasttrap_count_mtx);
849 	if (fasttrap_pid_count == 0) {
850 		ASSERT(dtrace_pid_probe_ptr == NULL);
851 		ASSERT(dtrace_return_probe_ptr == NULL);
852 		dtrace_pid_probe_ptr = &fasttrap_pid_probe;
853 		dtrace_return_probe_ptr = &fasttrap_return_probe;
854 	}
855 	ASSERT(dtrace_pid_probe_ptr == &fasttrap_pid_probe);
856 	ASSERT(dtrace_return_probe_ptr == &fasttrap_return_probe);
857 	fasttrap_pid_count++;
858 	mutex_exit(&fasttrap_count_mtx);
859 }
860 
861 static void
862 fasttrap_disable_callbacks(void)
863 {
864 	ASSERT(MUTEX_HELD(&cpu_lock));
865 
866 	mutex_enter(&fasttrap_count_mtx);
867 	ASSERT(fasttrap_pid_count > 0);
868 	fasttrap_pid_count--;
869 	if (fasttrap_pid_count == 0) {
870 		cpu_t *cur, *cpu = CPU;
871 
872 		for (cur = cpu->cpu_next_onln; cur != cpu;
873 		    cur = cur->cpu_next_onln) {
874 			rw_enter(&cur->cpu_ft_lock, RW_WRITER);
875 		}
876 
877 		dtrace_pid_probe_ptr = NULL;
878 		dtrace_return_probe_ptr = NULL;
879 
880 		for (cur = cpu->cpu_next_onln; cur != cpu;
881 		    cur = cur->cpu_next_onln) {
882 			rw_exit(&cur->cpu_ft_lock);
883 		}
884 	}
885 	mutex_exit(&fasttrap_count_mtx);
886 }
887 
888 /*ARGSUSED*/
889 static int
890 fasttrap_pid_enable(void *arg, dtrace_id_t id, void *parg)
891 {
892 	fasttrap_probe_t *probe = parg;
893 	proc_t *p;
894 	int i, rc;
895 
896 	ASSERT(probe != NULL);
897 	ASSERT(!probe->ftp_enabled);
898 	ASSERT(id == probe->ftp_id);
899 	ASSERT(MUTEX_HELD(&cpu_lock));
900 
901 	/*
902 	 * Increment the count of enabled probes on this probe's provider;
903 	 * the provider can't go away while the probe still exists. We
904 	 * must increment this even if we aren't able to properly enable
905 	 * this probe.
906 	 */
907 	mutex_enter(&probe->ftp_prov->ftp_mtx);
908 	probe->ftp_prov->ftp_rcount++;
909 	mutex_exit(&probe->ftp_prov->ftp_mtx);
910 
911 	/*
912 	 * If this probe's provider is retired (meaning it was valid in a
913 	 * previously exec'ed incarnation of this address space), bail out. The
914 	 * provider can't go away while we're in this code path.
915 	 */
916 	if (probe->ftp_prov->ftp_retired)
917 		return (0);
918 
919 	/*
920 	 * If we can't find the process, it may be that we're in the context of
921 	 * a fork in which the traced process is being born and we're copying
922 	 * USDT probes. Otherwise, the process is gone so bail.
923 	 */
924 	if ((p = sprlock(probe->ftp_pid)) == NULL) {
925 		if ((curproc->p_flag & SFORKING) == 0)
926 			return (0);
927 
928 		mutex_enter(&pidlock);
929 		p = prfind(probe->ftp_pid);
930 
931 		/*
932 		 * Confirm that curproc is indeed forking the process in which
933 		 * we're trying to enable probes.
934 		 */
935 		ASSERT(p != NULL);
936 		ASSERT(p->p_parent == curproc);
937 		ASSERT(p->p_stat == SIDL);
938 
939 		mutex_enter(&p->p_lock);
940 		mutex_exit(&pidlock);
941 
942 		sprlock_proc(p);
943 	}
944 
945 	ASSERT(!(p->p_flag & SVFORK));
946 	mutex_exit(&p->p_lock);
947 
948 	/*
949 	 * We have to enable the trap entry point before any user threads have
950 	 * the chance to execute the trap instruction we're about to place
951 	 * in their process's text.
952 	 */
953 	fasttrap_enable_callbacks();
954 
955 	/*
956 	 * Enable all the tracepoints and add this probe's id to each
957 	 * tracepoint's list of active probes.
958 	 */
959 	for (i = 0; i < probe->ftp_ntps; i++) {
960 		if ((rc = fasttrap_tracepoint_enable(p, probe, i)) != 0) {
961 			/*
962 			 * If enabling the tracepoint failed completely,
963 			 * we don't have to disable it; if the failure
964 			 * was only partial we must disable it.
965 			 */
966 			if (rc == FASTTRAP_ENABLE_FAIL)
967 				i--;
968 			else
969 				ASSERT(rc == FASTTRAP_ENABLE_PARTIAL);
970 
971 			/*
972 			 * Back up and pull out all the tracepoints we've
973 			 * created so far for this probe.
974 			 */
975 			while (i >= 0) {
976 				fasttrap_tracepoint_disable(p, probe, i);
977 				i--;
978 			}
979 
980 			mutex_enter(&p->p_lock);
981 			sprunlock(p);
982 
983 			/*
984 			 * Since we're not actually enabling this probe,
985 			 * drop our reference on the trap table entry.
986 			 */
987 			fasttrap_disable_callbacks();
988 			return (0);
989 		}
990 	}
991 
992 	mutex_enter(&p->p_lock);
993 	sprunlock(p);
994 
995 	probe->ftp_enabled = 1;
996 	return (0);
997 }
998 
999 /*ARGSUSED*/
1000 static void
1001 fasttrap_pid_disable(void *arg, dtrace_id_t id, void *parg)
1002 {
1003 	fasttrap_probe_t *probe = parg;
1004 	fasttrap_provider_t *provider = probe->ftp_prov;
1005 	proc_t *p;
1006 	int i, whack = 0;
1007 
1008 	ASSERT(id == probe->ftp_id);
1009 
1010 	/*
1011 	 * We won't be able to acquire a /proc-esque lock on the process
1012 	 * iff the process is dead and gone. In this case, we rely on the
1013 	 * provider lock as a point of mutual exclusion to prevent other
1014 	 * DTrace consumers from disabling this probe.
1015 	 */
1016 	if ((p = sprlock(probe->ftp_pid)) != NULL) {
1017 		ASSERT(!(p->p_flag & SVFORK));
1018 		mutex_exit(&p->p_lock);
1019 	}
1020 
1021 	mutex_enter(&provider->ftp_mtx);
1022 
1023 	/*
1024 	 * Disable all the associated tracepoints (for fully enabled probes).
1025 	 */
1026 	if (probe->ftp_enabled) {
1027 		for (i = 0; i < probe->ftp_ntps; i++) {
1028 			fasttrap_tracepoint_disable(p, probe, i);
1029 		}
1030 	}
1031 
1032 	ASSERT(provider->ftp_rcount > 0);
1033 	provider->ftp_rcount--;
1034 
1035 	if (p != NULL) {
1036 		/*
1037 		 * Even though we may not be able to remove it entirely, we
1038 		 * mark this retired provider to get a chance to remove some
1039 		 * of the associated probes.
1040 		 */
1041 		if (provider->ftp_retired && !provider->ftp_marked)
1042 			whack = provider->ftp_marked = 1;
1043 		mutex_exit(&provider->ftp_mtx);
1044 
1045 		mutex_enter(&p->p_lock);
1046 		sprunlock(p);
1047 	} else {
1048 		/*
1049 		 * If the process is dead, we're just waiting for the
1050 		 * last probe to be disabled to be able to free it.
1051 		 */
1052 		if (provider->ftp_rcount == 0 && !provider->ftp_marked)
1053 			whack = provider->ftp_marked = 1;
1054 		mutex_exit(&provider->ftp_mtx);
1055 	}
1056 
1057 	if (whack)
1058 		fasttrap_pid_cleanup();
1059 
1060 	if (!probe->ftp_enabled)
1061 		return;
1062 
1063 	probe->ftp_enabled = 0;
1064 
1065 	ASSERT(MUTEX_HELD(&cpu_lock));
1066 	fasttrap_disable_callbacks();
1067 }
1068 
1069 /*ARGSUSED*/
1070 static void
1071 fasttrap_pid_getargdesc(void *arg, dtrace_id_t id, void *parg,
1072     dtrace_argdesc_t *desc)
1073 {
1074 	fasttrap_probe_t *probe = parg;
1075 	char *str;
1076 	int i, ndx;
1077 
1078 	desc->dtargd_native[0] = '\0';
1079 	desc->dtargd_xlate[0] = '\0';
1080 
1081 	if (probe->ftp_prov->ftp_retired != 0 ||
1082 	    desc->dtargd_ndx >= probe->ftp_nargs) {
1083 		desc->dtargd_ndx = DTRACE_ARGNONE;
1084 		return;
1085 	}
1086 
1087 	ndx = (probe->ftp_argmap != NULL) ?
1088 	    probe->ftp_argmap[desc->dtargd_ndx] : desc->dtargd_ndx;
1089 
1090 	str = probe->ftp_ntypes;
1091 	for (i = 0; i < ndx; i++) {
1092 		str += strlen(str) + 1;
1093 	}
1094 
1095 	ASSERT(strlen(str + 1) < sizeof (desc->dtargd_native));
1096 	(void) strcpy(desc->dtargd_native, str);
1097 
1098 	if (probe->ftp_xtypes == NULL)
1099 		return;
1100 
1101 	str = probe->ftp_xtypes;
1102 	for (i = 0; i < desc->dtargd_ndx; i++) {
1103 		str += strlen(str) + 1;
1104 	}
1105 
1106 	ASSERT(strlen(str + 1) < sizeof (desc->dtargd_xlate));
1107 	(void) strcpy(desc->dtargd_xlate, str);
1108 }
1109 
1110 /*ARGSUSED*/
1111 static void
1112 fasttrap_pid_destroy(void *arg, dtrace_id_t id, void *parg)
1113 {
1114 	fasttrap_probe_t *probe = parg;
1115 	int i;
1116 	size_t size;
1117 
1118 	ASSERT(probe != NULL);
1119 	ASSERT(!probe->ftp_enabled);
1120 	ASSERT(fasttrap_total >= probe->ftp_ntps);
1121 
1122 	atomic_add_32(&fasttrap_total, -probe->ftp_ntps);
1123 	size = offsetof(fasttrap_probe_t, ftp_tps[probe->ftp_ntps]);
1124 
1125 	if (probe->ftp_gen + 1 >= fasttrap_mod_gen)
1126 		fasttrap_mod_barrier(probe->ftp_gen);
1127 
1128 	for (i = 0; i < probe->ftp_ntps; i++) {
1129 		kmem_free(probe->ftp_tps[i].fit_tp,
1130 		    sizeof (fasttrap_tracepoint_t));
1131 	}
1132 
1133 	kmem_free(probe, size);
1134 }
1135 
1136 
1137 static const dtrace_pattr_t pid_attr = {
1138 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1139 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1140 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1141 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_ISA },
1142 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
1143 };
1144 
1145 static dtrace_pops_t pid_pops = {
1146 	fasttrap_pid_provide,
1147 	NULL,
1148 	fasttrap_pid_enable,
1149 	fasttrap_pid_disable,
1150 	NULL,
1151 	NULL,
1152 	fasttrap_pid_getargdesc,
1153 	fasttrap_pid_getarg,
1154 	NULL,
1155 	fasttrap_pid_destroy
1156 };
1157 
1158 static dtrace_pops_t usdt_pops = {
1159 	fasttrap_pid_provide,
1160 	NULL,
1161 	fasttrap_pid_enable,
1162 	fasttrap_pid_disable,
1163 	NULL,
1164 	NULL,
1165 	fasttrap_pid_getargdesc,
1166 	fasttrap_usdt_getarg,
1167 	NULL,
1168 	fasttrap_pid_destroy
1169 };
1170 
1171 static fasttrap_proc_t *
1172 fasttrap_proc_lookup(pid_t pid)
1173 {
1174 	fasttrap_bucket_t *bucket;
1175 	fasttrap_proc_t *fprc, *new_fprc;
1176 
1177 	bucket = &fasttrap_procs.fth_table[FASTTRAP_PROCS_INDEX(pid)];
1178 	mutex_enter(&bucket->ftb_mtx);
1179 
1180 	for (fprc = bucket->ftb_data; fprc != NULL; fprc = fprc->ftpc_next) {
1181 		if (fprc->ftpc_pid == pid && fprc->ftpc_acount != 0) {
1182 			mutex_enter(&fprc->ftpc_mtx);
1183 			mutex_exit(&bucket->ftb_mtx);
1184 			fprc->ftpc_rcount++;
1185 			atomic_add_64(&fprc->ftpc_acount, 1);
1186 			ASSERT(fprc->ftpc_acount <= fprc->ftpc_rcount);
1187 			mutex_exit(&fprc->ftpc_mtx);
1188 
1189 			return (fprc);
1190 		}
1191 	}
1192 
1193 	/*
1194 	 * Drop the bucket lock so we don't try to perform a sleeping
1195 	 * allocation under it.
1196 	 */
1197 	mutex_exit(&bucket->ftb_mtx);
1198 
1199 	new_fprc = kmem_zalloc(sizeof (fasttrap_proc_t), KM_SLEEP);
1200 	new_fprc->ftpc_pid = pid;
1201 	new_fprc->ftpc_rcount = 1;
1202 	new_fprc->ftpc_acount = 1;
1203 
1204 	mutex_enter(&bucket->ftb_mtx);
1205 
1206 	/*
1207 	 * Take another lap through the list to make sure a proc hasn't
1208 	 * been created for this pid while we weren't under the bucket lock.
1209 	 */
1210 	for (fprc = bucket->ftb_data; fprc != NULL; fprc = fprc->ftpc_next) {
1211 		if (fprc->ftpc_pid == pid && fprc->ftpc_acount != 0) {
1212 			mutex_enter(&fprc->ftpc_mtx);
1213 			mutex_exit(&bucket->ftb_mtx);
1214 			fprc->ftpc_rcount++;
1215 			atomic_add_64(&fprc->ftpc_acount, 1);
1216 			ASSERT(fprc->ftpc_acount <= fprc->ftpc_rcount);
1217 			mutex_exit(&fprc->ftpc_mtx);
1218 
1219 			kmem_free(new_fprc, sizeof (fasttrap_proc_t));
1220 
1221 			return (fprc);
1222 		}
1223 	}
1224 
1225 	new_fprc->ftpc_next = bucket->ftb_data;
1226 	bucket->ftb_data = new_fprc;
1227 
1228 	mutex_exit(&bucket->ftb_mtx);
1229 
1230 	return (new_fprc);
1231 }
1232 
1233 static void
1234 fasttrap_proc_release(fasttrap_proc_t *proc)
1235 {
1236 	fasttrap_bucket_t *bucket;
1237 	fasttrap_proc_t *fprc, **fprcp;
1238 	pid_t pid = proc->ftpc_pid;
1239 
1240 	mutex_enter(&proc->ftpc_mtx);
1241 
1242 	ASSERT(proc->ftpc_rcount != 0);
1243 	ASSERT(proc->ftpc_acount <= proc->ftpc_rcount);
1244 
1245 	if (--proc->ftpc_rcount != 0) {
1246 		mutex_exit(&proc->ftpc_mtx);
1247 		return;
1248 	}
1249 
1250 	mutex_exit(&proc->ftpc_mtx);
1251 
1252 	/*
1253 	 * There should definitely be no live providers associated with this
1254 	 * process at this point.
1255 	 */
1256 	ASSERT(proc->ftpc_acount == 0);
1257 
1258 	bucket = &fasttrap_procs.fth_table[FASTTRAP_PROCS_INDEX(pid)];
1259 	mutex_enter(&bucket->ftb_mtx);
1260 
1261 	fprcp = (fasttrap_proc_t **)&bucket->ftb_data;
1262 	while ((fprc = *fprcp) != NULL) {
1263 		if (fprc == proc)
1264 			break;
1265 
1266 		fprcp = &fprc->ftpc_next;
1267 	}
1268 
1269 	/*
1270 	 * Something strange has happened if we can't find the proc.
1271 	 */
1272 	ASSERT(fprc != NULL);
1273 
1274 	*fprcp = fprc->ftpc_next;
1275 
1276 	mutex_exit(&bucket->ftb_mtx);
1277 
1278 	kmem_free(fprc, sizeof (fasttrap_proc_t));
1279 }
1280 
1281 /*
1282  * Lookup a fasttrap-managed provider based on its name and associated pid.
1283  * If the pattr argument is non-NULL, this function instantiates the provider
1284  * if it doesn't exist otherwise it returns NULL. The provider is returned
1285  * with its lock held.
1286  */
1287 static fasttrap_provider_t *
1288 fasttrap_provider_lookup(pid_t pid, const char *name,
1289     const dtrace_pattr_t *pattr)
1290 {
1291 	fasttrap_provider_t *fp, *new_fp = NULL;
1292 	fasttrap_bucket_t *bucket;
1293 	char provname[DTRACE_PROVNAMELEN];
1294 	proc_t *p;
1295 	cred_t *cred;
1296 
1297 	ASSERT(strlen(name) < sizeof (fp->ftp_name));
1298 	ASSERT(pattr != NULL);
1299 
1300 	bucket = &fasttrap_provs.fth_table[FASTTRAP_PROVS_INDEX(pid, name)];
1301 	mutex_enter(&bucket->ftb_mtx);
1302 
1303 	/*
1304 	 * Take a lap through the list and return the match if we find it.
1305 	 */
1306 	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1307 		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1308 		    !fp->ftp_retired) {
1309 			mutex_enter(&fp->ftp_mtx);
1310 			mutex_exit(&bucket->ftb_mtx);
1311 			return (fp);
1312 		}
1313 	}
1314 
1315 	/*
1316 	 * Drop the bucket lock so we don't try to perform a sleeping
1317 	 * allocation under it.
1318 	 */
1319 	mutex_exit(&bucket->ftb_mtx);
1320 
1321 	/*
1322 	 * Make sure the process exists, isn't a child created as the result
1323 	 * of a vfork(2), and isn't a zombie (but may be in fork).
1324 	 */
1325 	mutex_enter(&pidlock);
1326 	if ((p = prfind(pid)) == NULL) {
1327 		mutex_exit(&pidlock);
1328 		return (NULL);
1329 	}
1330 	mutex_enter(&p->p_lock);
1331 	mutex_exit(&pidlock);
1332 	if (p->p_flag & (SVFORK | SEXITING)) {
1333 		mutex_exit(&p->p_lock);
1334 		return (NULL);
1335 	}
1336 
1337 	/*
1338 	 * Increment p_dtrace_probes so that the process knows to inform us
1339 	 * when it exits or execs. fasttrap_provider_free() decrements this
1340 	 * when we're done with this provider.
1341 	 */
1342 	p->p_dtrace_probes++;
1343 
1344 	/*
1345 	 * Grab the credentials for this process so we have
1346 	 * something to pass to dtrace_register().
1347 	 */
1348 	mutex_enter(&p->p_crlock);
1349 	crhold(p->p_cred);
1350 	cred = p->p_cred;
1351 	mutex_exit(&p->p_crlock);
1352 	mutex_exit(&p->p_lock);
1353 
1354 	new_fp = kmem_zalloc(sizeof (fasttrap_provider_t), KM_SLEEP);
1355 	new_fp->ftp_pid = pid;
1356 	new_fp->ftp_proc = fasttrap_proc_lookup(pid);
1357 
1358 	ASSERT(new_fp->ftp_proc != NULL);
1359 
1360 	mutex_enter(&bucket->ftb_mtx);
1361 
1362 	/*
1363 	 * Take another lap through the list to make sure a provider hasn't
1364 	 * been created for this pid while we weren't under the bucket lock.
1365 	 */
1366 	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1367 		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1368 		    !fp->ftp_retired) {
1369 			mutex_enter(&fp->ftp_mtx);
1370 			mutex_exit(&bucket->ftb_mtx);
1371 			fasttrap_provider_free(new_fp);
1372 			crfree(cred);
1373 			return (fp);
1374 		}
1375 	}
1376 
1377 	(void) strcpy(new_fp->ftp_name, name);
1378 
1379 	/*
1380 	 * Fail and return NULL if either the provider name is too long
1381 	 * or we fail to register this new provider with the DTrace
1382 	 * framework. Note that this is the only place we ever construct
1383 	 * the full provider name -- we keep it in pieces in the provider
1384 	 * structure.
1385 	 */
1386 	if (snprintf(provname, sizeof (provname), "%s%u", name, (uint_t)pid) >=
1387 	    sizeof (provname) ||
1388 	    dtrace_register(provname, pattr,
1389 	    DTRACE_PRIV_PROC | DTRACE_PRIV_OWNER | DTRACE_PRIV_ZONEOWNER, cred,
1390 	    pattr == &pid_attr ? &pid_pops : &usdt_pops, new_fp,
1391 	    &new_fp->ftp_provid) != 0) {
1392 		mutex_exit(&bucket->ftb_mtx);
1393 		fasttrap_provider_free(new_fp);
1394 		crfree(cred);
1395 		return (NULL);
1396 	}
1397 
1398 	new_fp->ftp_next = bucket->ftb_data;
1399 	bucket->ftb_data = new_fp;
1400 
1401 	mutex_enter(&new_fp->ftp_mtx);
1402 	mutex_exit(&bucket->ftb_mtx);
1403 
1404 	crfree(cred);
1405 	return (new_fp);
1406 }
1407 
1408 static void
1409 fasttrap_provider_free(fasttrap_provider_t *provider)
1410 {
1411 	pid_t pid = provider->ftp_pid;
1412 	proc_t *p;
1413 
1414 	/*
1415 	 * There need to be no associated enabled probes, no consumers
1416 	 * creating probes, and no meta providers referencing this provider.
1417 	 */
1418 	ASSERT(provider->ftp_rcount == 0);
1419 	ASSERT(provider->ftp_ccount == 0);
1420 	ASSERT(provider->ftp_mcount == 0);
1421 
1422 	/*
1423 	 * If this provider hasn't been retired, we need to explicitly drop the
1424 	 * count of active providers on the associated process structure.
1425 	 */
1426 	if (!provider->ftp_retired) {
1427 		atomic_add_64(&provider->ftp_proc->ftpc_acount, -1);
1428 		ASSERT(provider->ftp_proc->ftpc_acount <
1429 		    provider->ftp_proc->ftpc_rcount);
1430 	}
1431 
1432 	fasttrap_proc_release(provider->ftp_proc);
1433 
1434 	kmem_free(provider, sizeof (fasttrap_provider_t));
1435 
1436 	/*
1437 	 * Decrement p_dtrace_probes on the process whose provider we're
1438 	 * freeing. We don't have to worry about clobbering somone else's
1439 	 * modifications to it because we have locked the bucket that
1440 	 * corresponds to this process's hash chain in the provider hash
1441 	 * table. Don't sweat it if we can't find the process.
1442 	 */
1443 	mutex_enter(&pidlock);
1444 	if ((p = prfind(pid)) == NULL) {
1445 		mutex_exit(&pidlock);
1446 		return;
1447 	}
1448 
1449 	mutex_enter(&p->p_lock);
1450 	mutex_exit(&pidlock);
1451 
1452 	p->p_dtrace_probes--;
1453 	mutex_exit(&p->p_lock);
1454 }
1455 
1456 static void
1457 fasttrap_provider_retire(pid_t pid, const char *name, int mprov)
1458 {
1459 	fasttrap_provider_t *fp;
1460 	fasttrap_bucket_t *bucket;
1461 	dtrace_provider_id_t provid;
1462 
1463 	ASSERT(strlen(name) < sizeof (fp->ftp_name));
1464 
1465 	bucket = &fasttrap_provs.fth_table[FASTTRAP_PROVS_INDEX(pid, name)];
1466 	mutex_enter(&bucket->ftb_mtx);
1467 
1468 	for (fp = bucket->ftb_data; fp != NULL; fp = fp->ftp_next) {
1469 		if (fp->ftp_pid == pid && strcmp(fp->ftp_name, name) == 0 &&
1470 		    !fp->ftp_retired)
1471 			break;
1472 	}
1473 
1474 	if (fp == NULL) {
1475 		mutex_exit(&bucket->ftb_mtx);
1476 		return;
1477 	}
1478 
1479 	mutex_enter(&fp->ftp_mtx);
1480 	ASSERT(!mprov || fp->ftp_mcount > 0);
1481 	if (mprov && --fp->ftp_mcount != 0)  {
1482 		mutex_exit(&fp->ftp_mtx);
1483 		mutex_exit(&bucket->ftb_mtx);
1484 		return;
1485 	}
1486 
1487 	/*
1488 	 * Mark the provider to be removed in our post-processing step, mark it
1489 	 * retired, and drop the active count on its proc. Marking it indicates
1490 	 * that we should try to remove it; setting the retired flag indicates
1491 	 * that we're done with this provider; dropping the active the proc
1492 	 * releases our hold, and when this reaches zero (as it will during
1493 	 * exit or exec) the proc and associated providers become defunct.
1494 	 *
1495 	 * We obviously need to take the bucket lock before the provider lock
1496 	 * to perform the lookup, but we need to drop the provider lock
1497 	 * before calling into the DTrace framework since we acquire the
1498 	 * provider lock in callbacks invoked from the DTrace framework. The
1499 	 * bucket lock therefore protects the integrity of the provider hash
1500 	 * table.
1501 	 */
1502 	atomic_add_64(&fp->ftp_proc->ftpc_acount, -1);
1503 	ASSERT(fp->ftp_proc->ftpc_acount < fp->ftp_proc->ftpc_rcount);
1504 
1505 	fp->ftp_retired = 1;
1506 	fp->ftp_marked = 1;
1507 	provid = fp->ftp_provid;
1508 	mutex_exit(&fp->ftp_mtx);
1509 
1510 	/*
1511 	 * We don't have to worry about invalidating the same provider twice
1512 	 * since fasttrap_provider_lookup() will ignore provider that have
1513 	 * been marked as retired.
1514 	 */
1515 	dtrace_invalidate(provid);
1516 
1517 	mutex_exit(&bucket->ftb_mtx);
1518 
1519 	fasttrap_pid_cleanup();
1520 }
1521 
1522 static int
1523 fasttrap_uint32_cmp(const void *ap, const void *bp)
1524 {
1525 	return (*(const uint32_t *)ap - *(const uint32_t *)bp);
1526 }
1527 
1528 static int
1529 fasttrap_uint64_cmp(const void *ap, const void *bp)
1530 {
1531 	return (*(const uint64_t *)ap - *(const uint64_t *)bp);
1532 }
1533 
1534 static int
1535 fasttrap_add_probe(fasttrap_probe_spec_t *pdata)
1536 {
1537 	fasttrap_provider_t *provider;
1538 	fasttrap_probe_t *pp;
1539 	fasttrap_tracepoint_t *tp;
1540 	char *name;
1541 	int i, aframes, whack;
1542 
1543 	/*
1544 	 * There needs to be at least one desired trace point.
1545 	 */
1546 	if (pdata->ftps_noffs == 0)
1547 		return (EINVAL);
1548 
1549 	switch (pdata->ftps_type) {
1550 	case DTFTP_ENTRY:
1551 		name = "entry";
1552 		aframes = FASTTRAP_ENTRY_AFRAMES;
1553 		break;
1554 	case DTFTP_RETURN:
1555 		name = "return";
1556 		aframes = FASTTRAP_RETURN_AFRAMES;
1557 		break;
1558 	case DTFTP_OFFSETS:
1559 		name = NULL;
1560 		break;
1561 	default:
1562 		return (EINVAL);
1563 	}
1564 
1565 	if ((provider = fasttrap_provider_lookup(pdata->ftps_pid,
1566 	    FASTTRAP_PID_NAME, &pid_attr)) == NULL)
1567 		return (ESRCH);
1568 
1569 	/*
1570 	 * Increment this reference count to indicate that a consumer is
1571 	 * actively adding a new probe associated with this provider. This
1572 	 * prevents the provider from being deleted -- we'll need to check
1573 	 * for pending deletions when we drop this reference count.
1574 	 */
1575 	provider->ftp_ccount++;
1576 	mutex_exit(&provider->ftp_mtx);
1577 
1578 	/*
1579 	 * Grab the creation lock to ensure consistency between calls to
1580 	 * dtrace_probe_lookup() and dtrace_probe_create() in the face of
1581 	 * other threads creating probes. We must drop the provider lock
1582 	 * before taking this lock to avoid a three-way deadlock with the
1583 	 * DTrace framework.
1584 	 */
1585 	mutex_enter(&provider->ftp_cmtx);
1586 
1587 	if (name == NULL) {
1588 		for (i = 0; i < pdata->ftps_noffs; i++) {
1589 			char name_str[17];
1590 
1591 			(void) sprintf(name_str, "%llx",
1592 			    (unsigned long long)pdata->ftps_offs[i]);
1593 
1594 			if (dtrace_probe_lookup(provider->ftp_provid,
1595 			    pdata->ftps_mod, pdata->ftps_func, name_str) != 0)
1596 				continue;
1597 
1598 			atomic_add_32(&fasttrap_total, 1);
1599 
1600 			if (fasttrap_total > fasttrap_max) {
1601 				atomic_add_32(&fasttrap_total, -1);
1602 				goto no_mem;
1603 			}
1604 
1605 			pp = kmem_zalloc(sizeof (fasttrap_probe_t), KM_SLEEP);
1606 
1607 			pp->ftp_prov = provider;
1608 			pp->ftp_faddr = pdata->ftps_pc;
1609 			pp->ftp_fsize = pdata->ftps_size;
1610 			pp->ftp_pid = pdata->ftps_pid;
1611 			pp->ftp_ntps = 1;
1612 
1613 			tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t),
1614 			    KM_SLEEP);
1615 
1616 			tp->ftt_proc = provider->ftp_proc;
1617 			tp->ftt_pc = pdata->ftps_offs[i] + pdata->ftps_pc;
1618 			tp->ftt_pid = pdata->ftps_pid;
1619 
1620 			pp->ftp_tps[0].fit_tp = tp;
1621 			pp->ftp_tps[0].fit_id.fti_probe = pp;
1622 			pp->ftp_tps[0].fit_id.fti_ptype = pdata->ftps_type;
1623 
1624 			pp->ftp_id = dtrace_probe_create(provider->ftp_provid,
1625 			    pdata->ftps_mod, pdata->ftps_func, name_str,
1626 			    FASTTRAP_OFFSET_AFRAMES, pp);
1627 		}
1628 
1629 	} else if (dtrace_probe_lookup(provider->ftp_provid, pdata->ftps_mod,
1630 	    pdata->ftps_func, name) == 0) {
1631 		atomic_add_32(&fasttrap_total, pdata->ftps_noffs);
1632 
1633 		if (fasttrap_total > fasttrap_max) {
1634 			atomic_add_32(&fasttrap_total, -pdata->ftps_noffs);
1635 			goto no_mem;
1636 		}
1637 
1638 		/*
1639 		 * Make sure all tracepoint program counter values are unique.
1640 		 * We later assume that each probe has exactly one tracepoint
1641 		 * for a given pc.
1642 		 */
1643 		qsort(pdata->ftps_offs, pdata->ftps_noffs,
1644 		    sizeof (uint64_t), fasttrap_uint64_cmp);
1645 		for (i = 1; i < pdata->ftps_noffs; i++) {
1646 			if (pdata->ftps_offs[i] > pdata->ftps_offs[i - 1])
1647 				continue;
1648 
1649 			atomic_add_32(&fasttrap_total, -pdata->ftps_noffs);
1650 			goto no_mem;
1651 		}
1652 
1653 		ASSERT(pdata->ftps_noffs > 0);
1654 		pp = kmem_zalloc(offsetof(fasttrap_probe_t,
1655 		    ftp_tps[pdata->ftps_noffs]), KM_SLEEP);
1656 
1657 		pp->ftp_prov = provider;
1658 		pp->ftp_faddr = pdata->ftps_pc;
1659 		pp->ftp_fsize = pdata->ftps_size;
1660 		pp->ftp_pid = pdata->ftps_pid;
1661 		pp->ftp_ntps = pdata->ftps_noffs;
1662 
1663 		for (i = 0; i < pdata->ftps_noffs; i++) {
1664 			tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t),
1665 			    KM_SLEEP);
1666 
1667 			tp->ftt_proc = provider->ftp_proc;
1668 			tp->ftt_pc = pdata->ftps_offs[i] + pdata->ftps_pc;
1669 			tp->ftt_pid = pdata->ftps_pid;
1670 
1671 			pp->ftp_tps[i].fit_tp = tp;
1672 			pp->ftp_tps[i].fit_id.fti_probe = pp;
1673 			pp->ftp_tps[i].fit_id.fti_ptype = pdata->ftps_type;
1674 		}
1675 
1676 		pp->ftp_id = dtrace_probe_create(provider->ftp_provid,
1677 		    pdata->ftps_mod, pdata->ftps_func, name, aframes, pp);
1678 	}
1679 
1680 	mutex_exit(&provider->ftp_cmtx);
1681 
1682 	/*
1683 	 * We know that the provider is still valid since we incremented the
1684 	 * creation reference count. If someone tried to clean up this provider
1685 	 * while we were using it (e.g. because the process called exec(2) or
1686 	 * exit(2)), take note of that and try to clean it up now.
1687 	 */
1688 	mutex_enter(&provider->ftp_mtx);
1689 	provider->ftp_ccount--;
1690 	whack = provider->ftp_retired;
1691 	mutex_exit(&provider->ftp_mtx);
1692 
1693 	if (whack)
1694 		fasttrap_pid_cleanup();
1695 
1696 	return (0);
1697 
1698 no_mem:
1699 	/*
1700 	 * If we've exhausted the allowable resources, we'll try to remove
1701 	 * this provider to free some up. This is to cover the case where
1702 	 * the user has accidentally created many more probes than was
1703 	 * intended (e.g. pid123:::).
1704 	 */
1705 	mutex_exit(&provider->ftp_cmtx);
1706 	mutex_enter(&provider->ftp_mtx);
1707 	provider->ftp_ccount--;
1708 	provider->ftp_marked = 1;
1709 	mutex_exit(&provider->ftp_mtx);
1710 
1711 	fasttrap_pid_cleanup();
1712 
1713 	return (ENOMEM);
1714 }
1715 
1716 /*ARGSUSED*/
1717 static void *
1718 fasttrap_meta_provide(void *arg, dtrace_helper_provdesc_t *dhpv, pid_t pid)
1719 {
1720 	fasttrap_provider_t *provider;
1721 
1722 	/*
1723 	 * A 32-bit unsigned integer (like a pid for example) can be
1724 	 * expressed in 10 or fewer decimal digits. Make sure that we'll
1725 	 * have enough space for the provider name.
1726 	 */
1727 	if (strlen(dhpv->dthpv_provname) + 10 >=
1728 	    sizeof (provider->ftp_name)) {
1729 		cmn_err(CE_WARN, "failed to instantiate provider %s: "
1730 		    "name too long to accomodate pid", dhpv->dthpv_provname);
1731 		return (NULL);
1732 	}
1733 
1734 	/*
1735 	 * Don't let folks spoof the true pid provider.
1736 	 */
1737 	if (strcmp(dhpv->dthpv_provname, FASTTRAP_PID_NAME) == 0) {
1738 		cmn_err(CE_WARN, "failed to instantiate provider %s: "
1739 		    "%s is an invalid name", dhpv->dthpv_provname,
1740 		    FASTTRAP_PID_NAME);
1741 		return (NULL);
1742 	}
1743 
1744 	/*
1745 	 * The highest stability class that fasttrap supports is ISA; cap
1746 	 * the stability of the new provider accordingly.
1747 	 */
1748 	if (dhpv->dthpv_pattr.dtpa_provider.dtat_class > DTRACE_CLASS_ISA)
1749 		dhpv->dthpv_pattr.dtpa_provider.dtat_class = DTRACE_CLASS_ISA;
1750 	if (dhpv->dthpv_pattr.dtpa_mod.dtat_class > DTRACE_CLASS_ISA)
1751 		dhpv->dthpv_pattr.dtpa_mod.dtat_class = DTRACE_CLASS_ISA;
1752 	if (dhpv->dthpv_pattr.dtpa_func.dtat_class > DTRACE_CLASS_ISA)
1753 		dhpv->dthpv_pattr.dtpa_func.dtat_class = DTRACE_CLASS_ISA;
1754 	if (dhpv->dthpv_pattr.dtpa_name.dtat_class > DTRACE_CLASS_ISA)
1755 		dhpv->dthpv_pattr.dtpa_name.dtat_class = DTRACE_CLASS_ISA;
1756 	if (dhpv->dthpv_pattr.dtpa_args.dtat_class > DTRACE_CLASS_ISA)
1757 		dhpv->dthpv_pattr.dtpa_args.dtat_class = DTRACE_CLASS_ISA;
1758 
1759 	if ((provider = fasttrap_provider_lookup(pid, dhpv->dthpv_provname,
1760 	    &dhpv->dthpv_pattr)) == NULL) {
1761 		cmn_err(CE_WARN, "failed to instantiate provider %s for "
1762 		    "process %u",  dhpv->dthpv_provname, (uint_t)pid);
1763 		return (NULL);
1764 	}
1765 
1766 	/*
1767 	 * Up the meta provider count so this provider isn't removed until
1768 	 * the meta provider has been told to remove it.
1769 	 */
1770 	provider->ftp_mcount++;
1771 
1772 	mutex_exit(&provider->ftp_mtx);
1773 
1774 	return (provider);
1775 }
1776 
1777 /*ARGSUSED*/
1778 static void
1779 fasttrap_meta_create_probe(void *arg, void *parg,
1780     dtrace_helper_probedesc_t *dhpb)
1781 {
1782 	fasttrap_provider_t *provider = parg;
1783 	fasttrap_probe_t *pp;
1784 	fasttrap_tracepoint_t *tp;
1785 	int i, j;
1786 	uint32_t ntps;
1787 
1788 	/*
1789 	 * Since the meta provider count is non-zero we don't have to worry
1790 	 * about this provider disappearing.
1791 	 */
1792 	ASSERT(provider->ftp_mcount > 0);
1793 
1794 	/*
1795 	 * The offsets must be unique.
1796 	 */
1797 	qsort(dhpb->dthpb_offs, dhpb->dthpb_noffs, sizeof (uint32_t),
1798 	    fasttrap_uint32_cmp);
1799 	for (i = 1; i < dhpb->dthpb_noffs; i++) {
1800 		if (dhpb->dthpb_base + dhpb->dthpb_offs[i] <=
1801 		    dhpb->dthpb_base + dhpb->dthpb_offs[i - 1])
1802 			return;
1803 	}
1804 
1805 	qsort(dhpb->dthpb_enoffs, dhpb->dthpb_nenoffs, sizeof (uint32_t),
1806 	    fasttrap_uint32_cmp);
1807 	for (i = 1; i < dhpb->dthpb_nenoffs; i++) {
1808 		if (dhpb->dthpb_base + dhpb->dthpb_enoffs[i] <=
1809 		    dhpb->dthpb_base + dhpb->dthpb_enoffs[i - 1])
1810 			return;
1811 	}
1812 
1813 	/*
1814 	 * Grab the creation lock to ensure consistency between calls to
1815 	 * dtrace_probe_lookup() and dtrace_probe_create() in the face of
1816 	 * other threads creating probes.
1817 	 */
1818 	mutex_enter(&provider->ftp_cmtx);
1819 
1820 	if (dtrace_probe_lookup(provider->ftp_provid, dhpb->dthpb_mod,
1821 	    dhpb->dthpb_func, dhpb->dthpb_name) != 0) {
1822 		mutex_exit(&provider->ftp_cmtx);
1823 		return;
1824 	}
1825 
1826 	ntps = dhpb->dthpb_noffs + dhpb->dthpb_nenoffs;
1827 	ASSERT(ntps > 0);
1828 
1829 	atomic_add_32(&fasttrap_total, ntps);
1830 
1831 	if (fasttrap_total > fasttrap_max) {
1832 		atomic_add_32(&fasttrap_total, -ntps);
1833 		mutex_exit(&provider->ftp_cmtx);
1834 		return;
1835 	}
1836 
1837 	pp = kmem_zalloc(offsetof(fasttrap_probe_t, ftp_tps[ntps]), KM_SLEEP);
1838 
1839 	pp->ftp_prov = provider;
1840 	pp->ftp_pid = provider->ftp_pid;
1841 	pp->ftp_ntps = ntps;
1842 	pp->ftp_nargs = dhpb->dthpb_xargc;
1843 	pp->ftp_xtypes = dhpb->dthpb_xtypes;
1844 	pp->ftp_ntypes = dhpb->dthpb_ntypes;
1845 
1846 	/*
1847 	 * First create a tracepoint for each actual point of interest.
1848 	 */
1849 	for (i = 0; i < dhpb->dthpb_noffs; i++) {
1850 		tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t), KM_SLEEP);
1851 
1852 		tp->ftt_proc = provider->ftp_proc;
1853 		tp->ftt_pc = dhpb->dthpb_base + dhpb->dthpb_offs[i];
1854 		tp->ftt_pid = provider->ftp_pid;
1855 
1856 		pp->ftp_tps[i].fit_tp = tp;
1857 		pp->ftp_tps[i].fit_id.fti_probe = pp;
1858 #ifdef __sparc
1859 		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_POST_OFFSETS;
1860 #else
1861 		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_OFFSETS;
1862 #endif
1863 	}
1864 
1865 	/*
1866 	 * Then create a tracepoint for each is-enabled point.
1867 	 */
1868 	for (j = 0; i < ntps; i++, j++) {
1869 		tp = kmem_zalloc(sizeof (fasttrap_tracepoint_t), KM_SLEEP);
1870 
1871 		tp->ftt_proc = provider->ftp_proc;
1872 		tp->ftt_pc = dhpb->dthpb_base + dhpb->dthpb_enoffs[j];
1873 		tp->ftt_pid = provider->ftp_pid;
1874 
1875 		pp->ftp_tps[i].fit_tp = tp;
1876 		pp->ftp_tps[i].fit_id.fti_probe = pp;
1877 		pp->ftp_tps[i].fit_id.fti_ptype = DTFTP_IS_ENABLED;
1878 	}
1879 
1880 	/*
1881 	 * If the arguments are shuffled around we set the argument remapping
1882 	 * table. Later, when the probe fires, we only remap the arguments
1883 	 * if the table is non-NULL.
1884 	 */
1885 	for (i = 0; i < dhpb->dthpb_xargc; i++) {
1886 		if (dhpb->dthpb_args[i] != i) {
1887 			pp->ftp_argmap = dhpb->dthpb_args;
1888 			break;
1889 		}
1890 	}
1891 
1892 	/*
1893 	 * The probe is fully constructed -- register it with DTrace.
1894 	 */
1895 	pp->ftp_id = dtrace_probe_create(provider->ftp_provid, dhpb->dthpb_mod,
1896 	    dhpb->dthpb_func, dhpb->dthpb_name, FASTTRAP_OFFSET_AFRAMES, pp);
1897 
1898 	mutex_exit(&provider->ftp_cmtx);
1899 }
1900 
1901 /*ARGSUSED*/
1902 static void
1903 fasttrap_meta_remove(void *arg, dtrace_helper_provdesc_t *dhpv, pid_t pid)
1904 {
1905 	/*
1906 	 * Clean up the USDT provider. There may be active consumers of the
1907 	 * provider busy adding probes, no damage will actually befall the
1908 	 * provider until that count has dropped to zero. This just puts
1909 	 * the provider on death row.
1910 	 */
1911 	fasttrap_provider_retire(pid, dhpv->dthpv_provname, 1);
1912 }
1913 
1914 static dtrace_mops_t fasttrap_mops = {
1915 	fasttrap_meta_create_probe,
1916 	fasttrap_meta_provide,
1917 	fasttrap_meta_remove
1918 };
1919 
1920 /*ARGSUSED*/
1921 static int
1922 fasttrap_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
1923 {
1924 	return (0);
1925 }
1926 
1927 /*ARGSUSED*/
1928 static int
1929 fasttrap_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
1930 {
1931 	if (!dtrace_attached())
1932 		return (EAGAIN);
1933 
1934 	if (cmd == FASTTRAPIOC_MAKEPROBE) {
1935 		fasttrap_probe_spec_t *uprobe = (void *)arg;
1936 		fasttrap_probe_spec_t *probe;
1937 		uint64_t noffs;
1938 		size_t size;
1939 		int ret, err;
1940 
1941 		if (copyin(&uprobe->ftps_noffs, &noffs,
1942 		    sizeof (uprobe->ftps_noffs)))
1943 			return (EFAULT);
1944 
1945 		/*
1946 		 * Probes must have at least one tracepoint.
1947 		 */
1948 		if (noffs == 0)
1949 			return (EINVAL);
1950 
1951 		size = sizeof (fasttrap_probe_spec_t) +
1952 		    sizeof (probe->ftps_offs[0]) * (noffs - 1);
1953 
1954 		if (size > 1024 * 1024)
1955 			return (ENOMEM);
1956 
1957 		probe = kmem_alloc(size, KM_SLEEP);
1958 
1959 		if (copyin(uprobe, probe, size) != 0 ||
1960 		    probe->ftps_noffs != noffs) {
1961 			kmem_free(probe, size);
1962 			return (EFAULT);
1963 		}
1964 
1965 		/*
1966 		 * Verify that the function and module strings contain no
1967 		 * funny characters.
1968 		 */
1969 		if (u8_validate(probe->ftps_func, strlen(probe->ftps_func),
1970 		    NULL, U8_VALIDATE_ENTIRE, &err) < 0) {
1971 			ret = EINVAL;
1972 			goto err;
1973 		}
1974 
1975 		if (u8_validate(probe->ftps_mod, strlen(probe->ftps_mod),
1976 		    NULL, U8_VALIDATE_ENTIRE, &err) < 0) {
1977 			ret = EINVAL;
1978 			goto err;
1979 		}
1980 
1981 		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
1982 			proc_t *p;
1983 			pid_t pid = probe->ftps_pid;
1984 
1985 			mutex_enter(&pidlock);
1986 			/*
1987 			 * Report an error if the process doesn't exist
1988 			 * or is actively being birthed.
1989 			 */
1990 			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
1991 				mutex_exit(&pidlock);
1992 				return (ESRCH);
1993 			}
1994 			mutex_enter(&p->p_lock);
1995 			mutex_exit(&pidlock);
1996 
1997 			if ((ret = priv_proc_cred_perm(cr, p, NULL,
1998 			    VREAD | VWRITE)) != 0) {
1999 				mutex_exit(&p->p_lock);
2000 				return (ret);
2001 			}
2002 
2003 			mutex_exit(&p->p_lock);
2004 		}
2005 
2006 		ret = fasttrap_add_probe(probe);
2007 err:
2008 		kmem_free(probe, size);
2009 
2010 		return (ret);
2011 
2012 	} else if (cmd == FASTTRAPIOC_GETINSTR) {
2013 		fasttrap_instr_query_t instr;
2014 		fasttrap_tracepoint_t *tp;
2015 		uint_t index;
2016 		int ret;
2017 
2018 		if (copyin((void *)arg, &instr, sizeof (instr)) != 0)
2019 			return (EFAULT);
2020 
2021 		if (!PRIV_POLICY_CHOICE(cr, PRIV_ALL, B_FALSE)) {
2022 			proc_t *p;
2023 			pid_t pid = instr.ftiq_pid;
2024 
2025 			mutex_enter(&pidlock);
2026 			/*
2027 			 * Report an error if the process doesn't exist
2028 			 * or is actively being birthed.
2029 			 */
2030 			if ((p = prfind(pid)) == NULL || p->p_stat == SIDL) {
2031 				mutex_exit(&pidlock);
2032 				return (ESRCH);
2033 			}
2034 			mutex_enter(&p->p_lock);
2035 			mutex_exit(&pidlock);
2036 
2037 			if ((ret = priv_proc_cred_perm(cr, p, NULL,
2038 			    VREAD)) != 0) {
2039 				mutex_exit(&p->p_lock);
2040 				return (ret);
2041 			}
2042 
2043 			mutex_exit(&p->p_lock);
2044 		}
2045 
2046 		index = FASTTRAP_TPOINTS_INDEX(instr.ftiq_pid, instr.ftiq_pc);
2047 
2048 		mutex_enter(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2049 		tp = fasttrap_tpoints.fth_table[index].ftb_data;
2050 		while (tp != NULL) {
2051 			if (instr.ftiq_pid == tp->ftt_pid &&
2052 			    instr.ftiq_pc == tp->ftt_pc &&
2053 			    tp->ftt_proc->ftpc_acount != 0)
2054 				break;
2055 
2056 			tp = tp->ftt_next;
2057 		}
2058 
2059 		if (tp == NULL) {
2060 			mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2061 			return (ENOENT);
2062 		}
2063 
2064 		bcopy(&tp->ftt_instr, &instr.ftiq_instr,
2065 		    sizeof (instr.ftiq_instr));
2066 		mutex_exit(&fasttrap_tpoints.fth_table[index].ftb_mtx);
2067 
2068 		if (copyout(&instr, (void *)arg, sizeof (instr)) != 0)
2069 			return (EFAULT);
2070 
2071 		return (0);
2072 	}
2073 
2074 	return (EINVAL);
2075 }
2076 
2077 static struct cb_ops fasttrap_cb_ops = {
2078 	fasttrap_open,		/* open */
2079 	nodev,			/* close */
2080 	nulldev,		/* strategy */
2081 	nulldev,		/* print */
2082 	nodev,			/* dump */
2083 	nodev,			/* read */
2084 	nodev,			/* write */
2085 	fasttrap_ioctl,		/* ioctl */
2086 	nodev,			/* devmap */
2087 	nodev,			/* mmap */
2088 	nodev,			/* segmap */
2089 	nochpoll,		/* poll */
2090 	ddi_prop_op,		/* cb_prop_op */
2091 	0,			/* streamtab  */
2092 	D_NEW | D_MP		/* Driver compatibility flag */
2093 };
2094 
2095 /*ARGSUSED*/
2096 static int
2097 fasttrap_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
2098 {
2099 	int error;
2100 
2101 	switch (infocmd) {
2102 	case DDI_INFO_DEVT2DEVINFO:
2103 		*result = (void *)fasttrap_devi;
2104 		error = DDI_SUCCESS;
2105 		break;
2106 	case DDI_INFO_DEVT2INSTANCE:
2107 		*result = (void *)0;
2108 		error = DDI_SUCCESS;
2109 		break;
2110 	default:
2111 		error = DDI_FAILURE;
2112 	}
2113 	return (error);
2114 }
2115 
2116 static int
2117 fasttrap_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
2118 {
2119 	ulong_t nent;
2120 
2121 	switch (cmd) {
2122 	case DDI_ATTACH:
2123 		break;
2124 	case DDI_RESUME:
2125 		return (DDI_SUCCESS);
2126 	default:
2127 		return (DDI_FAILURE);
2128 	}
2129 
2130 	if (ddi_create_minor_node(devi, "fasttrap", S_IFCHR, 0,
2131 	    DDI_PSEUDO, NULL) == DDI_FAILURE) {
2132 		ddi_remove_minor_node(devi, NULL);
2133 		return (DDI_FAILURE);
2134 	}
2135 
2136 	ddi_report_dev(devi);
2137 	fasttrap_devi = devi;
2138 
2139 	/*
2140 	 * Install our hooks into fork(2), exec(2), and exit(2).
2141 	 */
2142 	dtrace_fasttrap_fork_ptr = &fasttrap_fork;
2143 	dtrace_fasttrap_exit_ptr = &fasttrap_exec_exit;
2144 	dtrace_fasttrap_exec_ptr = &fasttrap_exec_exit;
2145 
2146 	fasttrap_max = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2147 	    "fasttrap-max-probes", FASTTRAP_MAX_DEFAULT);
2148 	fasttrap_total = 0;
2149 
2150 	/*
2151 	 * Conjure up the tracepoints hashtable...
2152 	 */
2153 	nent = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
2154 	    "fasttrap-hash-size", FASTTRAP_TPOINTS_DEFAULT_SIZE);
2155 
2156 	if (nent == 0 || nent > 0x1000000)
2157 		nent = FASTTRAP_TPOINTS_DEFAULT_SIZE;
2158 
2159 	if ((nent & (nent - 1)) == 0)
2160 		fasttrap_tpoints.fth_nent = nent;
2161 	else
2162 		fasttrap_tpoints.fth_nent = 1 << fasttrap_highbit(nent);
2163 	ASSERT(fasttrap_tpoints.fth_nent > 0);
2164 	fasttrap_tpoints.fth_mask = fasttrap_tpoints.fth_nent - 1;
2165 	fasttrap_tpoints.fth_table = kmem_zalloc(fasttrap_tpoints.fth_nent *
2166 	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2167 
2168 	/*
2169 	 * ... and the providers hash table...
2170 	 */
2171 	nent = FASTTRAP_PROVIDERS_DEFAULT_SIZE;
2172 	if ((nent & (nent - 1)) == 0)
2173 		fasttrap_provs.fth_nent = nent;
2174 	else
2175 		fasttrap_provs.fth_nent = 1 << fasttrap_highbit(nent);
2176 	ASSERT(fasttrap_provs.fth_nent > 0);
2177 	fasttrap_provs.fth_mask = fasttrap_provs.fth_nent - 1;
2178 	fasttrap_provs.fth_table = kmem_zalloc(fasttrap_provs.fth_nent *
2179 	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2180 
2181 	/*
2182 	 * ... and the procs hash table.
2183 	 */
2184 	nent = FASTTRAP_PROCS_DEFAULT_SIZE;
2185 	if ((nent & (nent - 1)) == 0)
2186 		fasttrap_procs.fth_nent = nent;
2187 	else
2188 		fasttrap_procs.fth_nent = 1 << fasttrap_highbit(nent);
2189 	ASSERT(fasttrap_procs.fth_nent > 0);
2190 	fasttrap_procs.fth_mask = fasttrap_procs.fth_nent - 1;
2191 	fasttrap_procs.fth_table = kmem_zalloc(fasttrap_procs.fth_nent *
2192 	    sizeof (fasttrap_bucket_t), KM_SLEEP);
2193 
2194 	(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2195 	    &fasttrap_meta_id);
2196 
2197 	return (DDI_SUCCESS);
2198 }
2199 
2200 static int
2201 fasttrap_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
2202 {
2203 	int i, fail = 0;
2204 	timeout_id_t tmp;
2205 
2206 	switch (cmd) {
2207 	case DDI_DETACH:
2208 		break;
2209 	case DDI_SUSPEND:
2210 		return (DDI_SUCCESS);
2211 	default:
2212 		return (DDI_FAILURE);
2213 	}
2214 
2215 	/*
2216 	 * Unregister the meta-provider to make sure no new fasttrap-
2217 	 * managed providers come along while we're trying to close up
2218 	 * shop. If we fail to detach, we'll need to re-register as a
2219 	 * meta-provider. We can fail to unregister as a meta-provider
2220 	 * if providers we manage still exist.
2221 	 */
2222 	if (fasttrap_meta_id != DTRACE_METAPROVNONE &&
2223 	    dtrace_meta_unregister(fasttrap_meta_id) != 0)
2224 		return (DDI_FAILURE);
2225 
2226 	/*
2227 	 * Prevent any new timeouts from running by setting fasttrap_timeout
2228 	 * to a non-zero value, and wait for the current timeout to complete.
2229 	 */
2230 	mutex_enter(&fasttrap_cleanup_mtx);
2231 	fasttrap_cleanup_work = 0;
2232 
2233 	while (fasttrap_timeout != (timeout_id_t)1) {
2234 		tmp = fasttrap_timeout;
2235 		fasttrap_timeout = (timeout_id_t)1;
2236 
2237 		if (tmp != 0) {
2238 			mutex_exit(&fasttrap_cleanup_mtx);
2239 			(void) untimeout(tmp);
2240 			mutex_enter(&fasttrap_cleanup_mtx);
2241 		}
2242 	}
2243 
2244 	fasttrap_cleanup_work = 0;
2245 	mutex_exit(&fasttrap_cleanup_mtx);
2246 
2247 	/*
2248 	 * Iterate over all of our providers. If there's still a process
2249 	 * that corresponds to that pid, fail to detach.
2250 	 */
2251 	for (i = 0; i < fasttrap_provs.fth_nent; i++) {
2252 		fasttrap_provider_t **fpp, *fp;
2253 		fasttrap_bucket_t *bucket = &fasttrap_provs.fth_table[i];
2254 
2255 		mutex_enter(&bucket->ftb_mtx);
2256 		fpp = (fasttrap_provider_t **)&bucket->ftb_data;
2257 		while ((fp = *fpp) != NULL) {
2258 			/*
2259 			 * Acquire and release the lock as a simple way of
2260 			 * waiting for any other consumer to finish with
2261 			 * this provider. A thread must first acquire the
2262 			 * bucket lock so there's no chance of another thread
2263 			 * blocking on the provider's lock.
2264 			 */
2265 			mutex_enter(&fp->ftp_mtx);
2266 			mutex_exit(&fp->ftp_mtx);
2267 
2268 			if (dtrace_unregister(fp->ftp_provid) != 0) {
2269 				fail = 1;
2270 				fpp = &fp->ftp_next;
2271 			} else {
2272 				*fpp = fp->ftp_next;
2273 				fasttrap_provider_free(fp);
2274 			}
2275 		}
2276 
2277 		mutex_exit(&bucket->ftb_mtx);
2278 	}
2279 
2280 	if (fail) {
2281 		uint_t work;
2282 		/*
2283 		 * If we're failing to detach, we need to unblock timeouts
2284 		 * and start a new timeout if any work has accumulated while
2285 		 * we've been unsuccessfully trying to detach.
2286 		 */
2287 		mutex_enter(&fasttrap_cleanup_mtx);
2288 		fasttrap_timeout = 0;
2289 		work = fasttrap_cleanup_work;
2290 		mutex_exit(&fasttrap_cleanup_mtx);
2291 
2292 		if (work)
2293 			fasttrap_pid_cleanup();
2294 
2295 		(void) dtrace_meta_register("fasttrap", &fasttrap_mops, NULL,
2296 		    &fasttrap_meta_id);
2297 
2298 		return (DDI_FAILURE);
2299 	}
2300 
2301 #ifdef DEBUG
2302 	mutex_enter(&fasttrap_count_mtx);
2303 	ASSERT(fasttrap_pid_count == 0);
2304 	mutex_exit(&fasttrap_count_mtx);
2305 #endif
2306 
2307 	kmem_free(fasttrap_tpoints.fth_table,
2308 	    fasttrap_tpoints.fth_nent * sizeof (fasttrap_bucket_t));
2309 	fasttrap_tpoints.fth_nent = 0;
2310 
2311 	kmem_free(fasttrap_provs.fth_table,
2312 	    fasttrap_provs.fth_nent * sizeof (fasttrap_bucket_t));
2313 	fasttrap_provs.fth_nent = 0;
2314 
2315 	kmem_free(fasttrap_procs.fth_table,
2316 	    fasttrap_procs.fth_nent * sizeof (fasttrap_bucket_t));
2317 	fasttrap_procs.fth_nent = 0;
2318 
2319 	/*
2320 	 * We know there are no tracepoints in any process anywhere in
2321 	 * the system so there is no process which has its p_dtrace_count
2322 	 * greater than zero, therefore we know that no thread can actively
2323 	 * be executing code in fasttrap_fork(). Similarly for p_dtrace_probes
2324 	 * and fasttrap_exec() and fasttrap_exit().
2325 	 */
2326 	ASSERT(dtrace_fasttrap_fork_ptr == &fasttrap_fork);
2327 	dtrace_fasttrap_fork_ptr = NULL;
2328 
2329 	ASSERT(dtrace_fasttrap_exec_ptr == &fasttrap_exec_exit);
2330 	dtrace_fasttrap_exec_ptr = NULL;
2331 
2332 	ASSERT(dtrace_fasttrap_exit_ptr == &fasttrap_exec_exit);
2333 	dtrace_fasttrap_exit_ptr = NULL;
2334 
2335 	ddi_remove_minor_node(devi, NULL);
2336 
2337 	return (DDI_SUCCESS);
2338 }
2339 
2340 static struct dev_ops fasttrap_ops = {
2341 	DEVO_REV,		/* devo_rev */
2342 	0,			/* refcnt */
2343 	fasttrap_info,		/* get_dev_info */
2344 	nulldev,		/* identify */
2345 	nulldev,		/* probe */
2346 	fasttrap_attach,	/* attach */
2347 	fasttrap_detach,	/* detach */
2348 	nodev,			/* reset */
2349 	&fasttrap_cb_ops,	/* driver operations */
2350 	NULL,			/* bus operations */
2351 	nodev,			/* dev power */
2352 	ddi_quiesce_not_needed,		/* quiesce */
2353 };
2354 
2355 /*
2356  * Module linkage information for the kernel.
2357  */
2358 static struct modldrv modldrv = {
2359 	&mod_driverops,		/* module type (this is a pseudo driver) */
2360 	"Fasttrap Tracing",	/* name of module */
2361 	&fasttrap_ops,		/* driver ops */
2362 };
2363 
2364 static struct modlinkage modlinkage = {
2365 	MODREV_1,
2366 	(void *)&modldrv,
2367 	NULL
2368 };
2369 
2370 int
2371 _init(void)
2372 {
2373 	return (mod_install(&modlinkage));
2374 }
2375 
2376 int
2377 _info(struct modinfo *modinfop)
2378 {
2379 	return (mod_info(&modlinkage, modinfop));
2380 }
2381 
2382 int
2383 _fini(void)
2384 {
2385 	return (mod_remove(&modlinkage));
2386 }
2387