xref: /titanic_41/usr/src/cmd/mdb/common/modules/libumem/umem.c (revision 450396635f70344c58b6b1e4db38cf17ff34445c)
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  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include "umem.h"
27 
28 #include <sys/vmem_impl_user.h>
29 #include <umem_impl.h>
30 
31 #include <alloca.h>
32 #include <limits.h>
33 #include <mdb/mdb_whatis.h>
34 
35 #include "misc.h"
36 #include "leaky.h"
37 #include "dist.h"
38 
39 #include "umem_pagesize.h"
40 
41 #define	UM_ALLOCATED		0x1
42 #define	UM_FREE			0x2
43 #define	UM_BUFCTL		0x4
44 #define	UM_HASH			0x8
45 
46 int umem_ready;
47 
48 static int umem_stack_depth_warned;
49 static uint32_t umem_max_ncpus;
50 uint32_t umem_stack_depth;
51 
52 size_t umem_pagesize;
53 
54 #define	UMEM_READVAR(var)				\
55 	(umem_readvar(&(var), #var) == -1 &&		\
56 	    (mdb_warn("failed to read "#var), 1))
57 
58 int
59 umem_update_variables(void)
60 {
61 	size_t pagesize;
62 
63 	/*
64 	 * Figure out which type of umem is being used; if it's not there
65 	 * yet, succeed quietly.
66 	 */
67 	if (umem_set_standalone() == -1) {
68 		umem_ready = 0;
69 		return (0);		/* umem not there yet */
70 	}
71 
72 	/*
73 	 * Solaris 9 used a different name for umem_max_ncpus.  It's
74 	 * cheap backwards compatibility to check for both names.
75 	 */
76 	if (umem_readvar(&umem_max_ncpus, "umem_max_ncpus") == -1 &&
77 	    umem_readvar(&umem_max_ncpus, "max_ncpus") == -1) {
78 		mdb_warn("unable to read umem_max_ncpus or max_ncpus");
79 		return (-1);
80 	}
81 	if (UMEM_READVAR(umem_ready))
82 		return (-1);
83 	if (UMEM_READVAR(umem_stack_depth))
84 		return (-1);
85 	if (UMEM_READVAR(pagesize))
86 		return (-1);
87 
88 	if (umem_stack_depth > UMEM_MAX_STACK_DEPTH) {
89 		if (umem_stack_depth_warned == 0) {
90 			mdb_warn("umem_stack_depth corrupted (%d > %d)\n",
91 			    umem_stack_depth, UMEM_MAX_STACK_DEPTH);
92 			umem_stack_depth_warned = 1;
93 		}
94 		umem_stack_depth = 0;
95 	}
96 
97 	umem_pagesize = pagesize;
98 
99 	return (0);
100 }
101 
102 /*ARGSUSED*/
103 static int
104 umem_init_walkers(uintptr_t addr, const umem_cache_t *c, void *ignored)
105 {
106 	mdb_walker_t w;
107 	char descr[64];
108 
109 	(void) mdb_snprintf(descr, sizeof (descr),
110 	    "walk the %s cache", c->cache_name);
111 
112 	w.walk_name = c->cache_name;
113 	w.walk_descr = descr;
114 	w.walk_init = umem_walk_init;
115 	w.walk_step = umem_walk_step;
116 	w.walk_fini = umem_walk_fini;
117 	w.walk_init_arg = (void *)addr;
118 
119 	if (mdb_add_walker(&w) == -1)
120 		mdb_warn("failed to add %s walker", c->cache_name);
121 
122 	return (WALK_NEXT);
123 }
124 
125 /*ARGSUSED*/
126 static void
127 umem_statechange_cb(void *arg)
128 {
129 	static int been_ready = 0;
130 
131 #ifndef _KMDB
132 	leaky_cleanup(1);	/* state changes invalidate leaky state */
133 #endif
134 
135 	if (umem_update_variables() == -1)
136 		return;
137 
138 	if (been_ready)
139 		return;
140 
141 	if (umem_ready != UMEM_READY)
142 		return;
143 
144 	been_ready = 1;
145 	(void) mdb_walk("umem_cache", (mdb_walk_cb_t)umem_init_walkers, NULL);
146 }
147 
148 int
149 umem_abort_messages(void)
150 {
151 	char *umem_error_buffer;
152 	uint_t umem_error_begin;
153 	GElf_Sym sym;
154 	size_t bufsize;
155 
156 	if (UMEM_READVAR(umem_error_begin))
157 		return (DCMD_ERR);
158 
159 	if (umem_lookup_by_name("umem_error_buffer", &sym) == -1) {
160 		mdb_warn("unable to look up umem_error_buffer");
161 		return (DCMD_ERR);
162 	}
163 
164 	bufsize = (size_t)sym.st_size;
165 
166 	umem_error_buffer = mdb_alloc(bufsize+1, UM_SLEEP | UM_GC);
167 
168 	if (mdb_vread(umem_error_buffer, bufsize, (uintptr_t)sym.st_value)
169 	    != bufsize) {
170 		mdb_warn("unable to read umem_error_buffer");
171 		return (DCMD_ERR);
172 	}
173 	/* put a zero after the end of the buffer to simplify printing */
174 	umem_error_buffer[bufsize] = 0;
175 
176 	if ((umem_error_begin % bufsize) == 0)
177 		mdb_printf("%s\n", umem_error_buffer);
178 	else {
179 		umem_error_buffer[(umem_error_begin % bufsize) - 1] = 0;
180 		mdb_printf("%s%s\n",
181 		    &umem_error_buffer[umem_error_begin % bufsize],
182 		    umem_error_buffer);
183 	}
184 
185 	return (DCMD_OK);
186 }
187 
188 static void
189 umem_log_status(const char *name, umem_log_header_t *val)
190 {
191 	umem_log_header_t my_lh;
192 	uintptr_t pos = (uintptr_t)val;
193 	size_t size;
194 
195 	if (pos == NULL)
196 		return;
197 
198 	if (mdb_vread(&my_lh, sizeof (umem_log_header_t), pos) == -1) {
199 		mdb_warn("\nunable to read umem_%s_log pointer %p",
200 		    name, pos);
201 		return;
202 	}
203 
204 	size = my_lh.lh_chunksize * my_lh.lh_nchunks;
205 
206 	if (size % (1024 * 1024) == 0)
207 		mdb_printf("%s=%dm ", name, size / (1024 * 1024));
208 	else if (size % 1024 == 0)
209 		mdb_printf("%s=%dk ", name, size / 1024);
210 	else
211 		mdb_printf("%s=%d ", name, size);
212 }
213 
214 typedef struct umem_debug_flags {
215 	const char	*udf_name;
216 	uint_t		udf_flags;
217 	uint_t		udf_clear;	/* if 0, uses udf_flags */
218 } umem_debug_flags_t;
219 
220 umem_debug_flags_t umem_status_flags[] = {
221 	{ "random",	UMF_RANDOMIZE,	UMF_RANDOM },
222 	{ "default",	UMF_AUDIT | UMF_DEADBEEF | UMF_REDZONE | UMF_CONTENTS },
223 	{ "audit",	UMF_AUDIT },
224 	{ "guards",	UMF_DEADBEEF | UMF_REDZONE },
225 	{ "nosignal",	UMF_CHECKSIGNAL },
226 	{ "firewall",	UMF_FIREWALL },
227 	{ "lite",	UMF_LITE },
228 	{ NULL }
229 };
230 
231 /*ARGSUSED*/
232 int
233 umem_status(uintptr_t addr, uint_t flags, int ac, const mdb_arg_t *argv)
234 {
235 	int umem_logging;
236 
237 	umem_log_header_t *umem_transaction_log;
238 	umem_log_header_t *umem_content_log;
239 	umem_log_header_t *umem_failure_log;
240 	umem_log_header_t *umem_slab_log;
241 
242 	mdb_printf("Status:\t\t%s\n",
243 	    umem_ready == UMEM_READY_INIT_FAILED ? "initialization failed" :
244 	    umem_ready == UMEM_READY_STARTUP ? "uninitialized" :
245 	    umem_ready == UMEM_READY_INITING ? "initialization in process" :
246 	    umem_ready == UMEM_READY ? "ready and active" :
247 	    umem_ready == 0 ? "not loaded into address space" :
248 	    "unknown (umem_ready invalid)");
249 
250 	if (umem_ready == 0)
251 		return (DCMD_OK);
252 
253 	mdb_printf("Concurrency:\t%d\n", umem_max_ncpus);
254 
255 	if (UMEM_READVAR(umem_logging))
256 		goto err;
257 	if (UMEM_READVAR(umem_transaction_log))
258 		goto err;
259 	if (UMEM_READVAR(umem_content_log))
260 		goto err;
261 	if (UMEM_READVAR(umem_failure_log))
262 		goto err;
263 	if (UMEM_READVAR(umem_slab_log))
264 		goto err;
265 
266 	mdb_printf("Logs:\t\t");
267 	umem_log_status("transaction", umem_transaction_log);
268 	umem_log_status("content", umem_content_log);
269 	umem_log_status("fail", umem_failure_log);
270 	umem_log_status("slab", umem_slab_log);
271 	if (!umem_logging)
272 		mdb_printf("(inactive)");
273 	mdb_printf("\n");
274 
275 	mdb_printf("Message buffer:\n");
276 	return (umem_abort_messages());
277 
278 err:
279 	mdb_printf("Message buffer:\n");
280 	(void) umem_abort_messages();
281 	return (DCMD_ERR);
282 }
283 
284 typedef struct {
285 	uintptr_t ucw_first;
286 	uintptr_t ucw_current;
287 } umem_cache_walk_t;
288 
289 int
290 umem_cache_walk_init(mdb_walk_state_t *wsp)
291 {
292 	umem_cache_walk_t *ucw;
293 	umem_cache_t c;
294 	uintptr_t cp;
295 	GElf_Sym sym;
296 
297 	if (umem_lookup_by_name("umem_null_cache", &sym) == -1) {
298 		mdb_warn("couldn't find umem_null_cache");
299 		return (WALK_ERR);
300 	}
301 
302 	cp = (uintptr_t)sym.st_value;
303 
304 	if (mdb_vread(&c, sizeof (umem_cache_t), cp) == -1) {
305 		mdb_warn("couldn't read cache at %p", cp);
306 		return (WALK_ERR);
307 	}
308 
309 	ucw = mdb_alloc(sizeof (umem_cache_walk_t), UM_SLEEP);
310 
311 	ucw->ucw_first = cp;
312 	ucw->ucw_current = (uintptr_t)c.cache_next;
313 	wsp->walk_data = ucw;
314 
315 	return (WALK_NEXT);
316 }
317 
318 int
319 umem_cache_walk_step(mdb_walk_state_t *wsp)
320 {
321 	umem_cache_walk_t *ucw = wsp->walk_data;
322 	umem_cache_t c;
323 	int status;
324 
325 	if (mdb_vread(&c, sizeof (umem_cache_t), ucw->ucw_current) == -1) {
326 		mdb_warn("couldn't read cache at %p", ucw->ucw_current);
327 		return (WALK_DONE);
328 	}
329 
330 	status = wsp->walk_callback(ucw->ucw_current, &c, wsp->walk_cbdata);
331 
332 	if ((ucw->ucw_current = (uintptr_t)c.cache_next) == ucw->ucw_first)
333 		return (WALK_DONE);
334 
335 	return (status);
336 }
337 
338 void
339 umem_cache_walk_fini(mdb_walk_state_t *wsp)
340 {
341 	umem_cache_walk_t *ucw = wsp->walk_data;
342 	mdb_free(ucw, sizeof (umem_cache_walk_t));
343 }
344 
345 typedef struct {
346 	umem_cpu_t *ucw_cpus;
347 	uint32_t ucw_current;
348 	uint32_t ucw_max;
349 } umem_cpu_walk_state_t;
350 
351 int
352 umem_cpu_walk_init(mdb_walk_state_t *wsp)
353 {
354 	umem_cpu_t *umem_cpus;
355 
356 	umem_cpu_walk_state_t *ucw;
357 
358 	if (umem_readvar(&umem_cpus, "umem_cpus") == -1) {
359 		mdb_warn("failed to read 'umem_cpus'");
360 		return (WALK_ERR);
361 	}
362 
363 	ucw = mdb_alloc(sizeof (*ucw), UM_SLEEP);
364 
365 	ucw->ucw_cpus = umem_cpus;
366 	ucw->ucw_current = 0;
367 	ucw->ucw_max = umem_max_ncpus;
368 
369 	wsp->walk_data = ucw;
370 	return (WALK_NEXT);
371 }
372 
373 int
374 umem_cpu_walk_step(mdb_walk_state_t *wsp)
375 {
376 	umem_cpu_t cpu;
377 	umem_cpu_walk_state_t *ucw = wsp->walk_data;
378 
379 	uintptr_t caddr;
380 
381 	if (ucw->ucw_current >= ucw->ucw_max)
382 		return (WALK_DONE);
383 
384 	caddr = (uintptr_t)&(ucw->ucw_cpus[ucw->ucw_current]);
385 
386 	if (mdb_vread(&cpu, sizeof (umem_cpu_t), caddr) == -1) {
387 		mdb_warn("failed to read cpu %d", ucw->ucw_current);
388 		return (WALK_ERR);
389 	}
390 
391 	ucw->ucw_current++;
392 
393 	return (wsp->walk_callback(caddr, &cpu, wsp->walk_cbdata));
394 }
395 
396 void
397 umem_cpu_walk_fini(mdb_walk_state_t *wsp)
398 {
399 	umem_cpu_walk_state_t *ucw = wsp->walk_data;
400 
401 	mdb_free(ucw, sizeof (*ucw));
402 }
403 
404 int
405 umem_cpu_cache_walk_init(mdb_walk_state_t *wsp)
406 {
407 	if (wsp->walk_addr == NULL) {
408 		mdb_warn("umem_cpu_cache doesn't support global walks");
409 		return (WALK_ERR);
410 	}
411 
412 	if (mdb_layered_walk("umem_cpu", wsp) == -1) {
413 		mdb_warn("couldn't walk 'umem_cpu'");
414 		return (WALK_ERR);
415 	}
416 
417 	wsp->walk_data = (void *)wsp->walk_addr;
418 
419 	return (WALK_NEXT);
420 }
421 
422 int
423 umem_cpu_cache_walk_step(mdb_walk_state_t *wsp)
424 {
425 	uintptr_t caddr = (uintptr_t)wsp->walk_data;
426 	const umem_cpu_t *cpu = wsp->walk_layer;
427 	umem_cpu_cache_t cc;
428 
429 	caddr += cpu->cpu_cache_offset;
430 
431 	if (mdb_vread(&cc, sizeof (umem_cpu_cache_t), caddr) == -1) {
432 		mdb_warn("couldn't read umem_cpu_cache at %p", caddr);
433 		return (WALK_ERR);
434 	}
435 
436 	return (wsp->walk_callback(caddr, &cc, wsp->walk_cbdata));
437 }
438 
439 int
440 umem_slab_walk_init(mdb_walk_state_t *wsp)
441 {
442 	uintptr_t caddr = wsp->walk_addr;
443 	umem_cache_t c;
444 
445 	if (caddr == NULL) {
446 		mdb_warn("umem_slab doesn't support global walks\n");
447 		return (WALK_ERR);
448 	}
449 
450 	if (mdb_vread(&c, sizeof (c), caddr) == -1) {
451 		mdb_warn("couldn't read umem_cache at %p", caddr);
452 		return (WALK_ERR);
453 	}
454 
455 	wsp->walk_data =
456 	    (void *)(caddr + offsetof(umem_cache_t, cache_nullslab));
457 	wsp->walk_addr = (uintptr_t)c.cache_nullslab.slab_next;
458 
459 	return (WALK_NEXT);
460 }
461 
462 int
463 umem_slab_walk_partial_init(mdb_walk_state_t *wsp)
464 {
465 	uintptr_t caddr = wsp->walk_addr;
466 	umem_cache_t c;
467 
468 	if (caddr == NULL) {
469 		mdb_warn("umem_slab_partial doesn't support global walks\n");
470 		return (WALK_ERR);
471 	}
472 
473 	if (mdb_vread(&c, sizeof (c), caddr) == -1) {
474 		mdb_warn("couldn't read umem_cache at %p", caddr);
475 		return (WALK_ERR);
476 	}
477 
478 	wsp->walk_data =
479 	    (void *)(caddr + offsetof(umem_cache_t, cache_nullslab));
480 	wsp->walk_addr = (uintptr_t)c.cache_freelist;
481 
482 	/*
483 	 * Some consumers (umem_walk_step(), in particular) require at
484 	 * least one callback if there are any buffers in the cache.  So
485 	 * if there are *no* partial slabs, report the last full slab, if
486 	 * any.
487 	 *
488 	 * Yes, this is ugly, but it's cleaner than the other possibilities.
489 	 */
490 	if ((uintptr_t)wsp->walk_data == wsp->walk_addr)
491 		wsp->walk_addr = (uintptr_t)c.cache_nullslab.slab_prev;
492 
493 	return (WALK_NEXT);
494 }
495 
496 int
497 umem_slab_walk_step(mdb_walk_state_t *wsp)
498 {
499 	umem_slab_t s;
500 	uintptr_t addr = wsp->walk_addr;
501 	uintptr_t saddr = (uintptr_t)wsp->walk_data;
502 	uintptr_t caddr = saddr - offsetof(umem_cache_t, cache_nullslab);
503 
504 	if (addr == saddr)
505 		return (WALK_DONE);
506 
507 	if (mdb_vread(&s, sizeof (s), addr) == -1) {
508 		mdb_warn("failed to read slab at %p", wsp->walk_addr);
509 		return (WALK_ERR);
510 	}
511 
512 	if ((uintptr_t)s.slab_cache != caddr) {
513 		mdb_warn("slab %p isn't in cache %p (in cache %p)\n",
514 		    addr, caddr, s.slab_cache);
515 		return (WALK_ERR);
516 	}
517 
518 	wsp->walk_addr = (uintptr_t)s.slab_next;
519 
520 	return (wsp->walk_callback(addr, &s, wsp->walk_cbdata));
521 }
522 
523 int
524 umem_cache(uintptr_t addr, uint_t flags, int ac, const mdb_arg_t *argv)
525 {
526 	umem_cache_t c;
527 
528 	if (!(flags & DCMD_ADDRSPEC)) {
529 		if (mdb_walk_dcmd("umem_cache", "umem_cache", ac, argv) == -1) {
530 			mdb_warn("can't walk umem_cache");
531 			return (DCMD_ERR);
532 		}
533 		return (DCMD_OK);
534 	}
535 
536 	if (DCMD_HDRSPEC(flags))
537 		mdb_printf("%-?s %-25s %4s %8s %8s %8s\n", "ADDR", "NAME",
538 		    "FLAG", "CFLAG", "BUFSIZE", "BUFTOTL");
539 
540 	if (mdb_vread(&c, sizeof (c), addr) == -1) {
541 		mdb_warn("couldn't read umem_cache at %p", addr);
542 		return (DCMD_ERR);
543 	}
544 
545 	mdb_printf("%0?p %-25s %04x %08x %8ld %8lld\n", addr, c.cache_name,
546 	    c.cache_flags, c.cache_cflags, c.cache_bufsize, c.cache_buftotal);
547 
548 	return (DCMD_OK);
549 }
550 
551 static int
552 addrcmp(const void *lhs, const void *rhs)
553 {
554 	uintptr_t p1 = *((uintptr_t *)lhs);
555 	uintptr_t p2 = *((uintptr_t *)rhs);
556 
557 	if (p1 < p2)
558 		return (-1);
559 	if (p1 > p2)
560 		return (1);
561 	return (0);
562 }
563 
564 static int
565 bufctlcmp(const umem_bufctl_audit_t **lhs, const umem_bufctl_audit_t **rhs)
566 {
567 	const umem_bufctl_audit_t *bcp1 = *lhs;
568 	const umem_bufctl_audit_t *bcp2 = *rhs;
569 
570 	if (bcp1->bc_timestamp > bcp2->bc_timestamp)
571 		return (-1);
572 
573 	if (bcp1->bc_timestamp < bcp2->bc_timestamp)
574 		return (1);
575 
576 	return (0);
577 }
578 
579 typedef struct umem_hash_walk {
580 	uintptr_t *umhw_table;
581 	size_t umhw_nelems;
582 	size_t umhw_pos;
583 	umem_bufctl_t umhw_cur;
584 } umem_hash_walk_t;
585 
586 int
587 umem_hash_walk_init(mdb_walk_state_t *wsp)
588 {
589 	umem_hash_walk_t *umhw;
590 	uintptr_t *hash;
591 	umem_cache_t c;
592 	uintptr_t haddr, addr = wsp->walk_addr;
593 	size_t nelems;
594 	size_t hsize;
595 
596 	if (addr == NULL) {
597 		mdb_warn("umem_hash doesn't support global walks\n");
598 		return (WALK_ERR);
599 	}
600 
601 	if (mdb_vread(&c, sizeof (c), addr) == -1) {
602 		mdb_warn("couldn't read cache at addr %p", addr);
603 		return (WALK_ERR);
604 	}
605 
606 	if (!(c.cache_flags & UMF_HASH)) {
607 		mdb_warn("cache %p doesn't have a hash table\n", addr);
608 		return (WALK_DONE);		/* nothing to do */
609 	}
610 
611 	umhw = mdb_zalloc(sizeof (umem_hash_walk_t), UM_SLEEP);
612 	umhw->umhw_cur.bc_next = NULL;
613 	umhw->umhw_pos = 0;
614 
615 	umhw->umhw_nelems = nelems = c.cache_hash_mask + 1;
616 	hsize = nelems * sizeof (uintptr_t);
617 	haddr = (uintptr_t)c.cache_hash_table;
618 
619 	umhw->umhw_table = hash = mdb_alloc(hsize, UM_SLEEP);
620 	if (mdb_vread(hash, hsize, haddr) == -1) {
621 		mdb_warn("failed to read hash table at %p", haddr);
622 		mdb_free(hash, hsize);
623 		mdb_free(umhw, sizeof (umem_hash_walk_t));
624 		return (WALK_ERR);
625 	}
626 
627 	wsp->walk_data = umhw;
628 
629 	return (WALK_NEXT);
630 }
631 
632 int
633 umem_hash_walk_step(mdb_walk_state_t *wsp)
634 {
635 	umem_hash_walk_t *umhw = wsp->walk_data;
636 	uintptr_t addr = NULL;
637 
638 	if ((addr = (uintptr_t)umhw->umhw_cur.bc_next) == NULL) {
639 		while (umhw->umhw_pos < umhw->umhw_nelems) {
640 			if ((addr = umhw->umhw_table[umhw->umhw_pos++]) != NULL)
641 				break;
642 		}
643 	}
644 	if (addr == NULL)
645 		return (WALK_DONE);
646 
647 	if (mdb_vread(&umhw->umhw_cur, sizeof (umem_bufctl_t), addr) == -1) {
648 		mdb_warn("couldn't read umem_bufctl_t at addr %p", addr);
649 		return (WALK_ERR);
650 	}
651 
652 	return (wsp->walk_callback(addr, &umhw->umhw_cur, wsp->walk_cbdata));
653 }
654 
655 void
656 umem_hash_walk_fini(mdb_walk_state_t *wsp)
657 {
658 	umem_hash_walk_t *umhw = wsp->walk_data;
659 
660 	if (umhw == NULL)
661 		return;
662 
663 	mdb_free(umhw->umhw_table, umhw->umhw_nelems * sizeof (uintptr_t));
664 	mdb_free(umhw, sizeof (umem_hash_walk_t));
665 }
666 
667 /*
668  * Find the address of the bufctl structure for the address 'buf' in cache
669  * 'cp', which is at address caddr, and place it in *out.
670  */
671 static int
672 umem_hash_lookup(umem_cache_t *cp, uintptr_t caddr, void *buf, uintptr_t *out)
673 {
674 	uintptr_t bucket = (uintptr_t)UMEM_HASH(cp, buf);
675 	umem_bufctl_t *bcp;
676 	umem_bufctl_t bc;
677 
678 	if (mdb_vread(&bcp, sizeof (umem_bufctl_t *), bucket) == -1) {
679 		mdb_warn("unable to read hash bucket for %p in cache %p",
680 		    buf, caddr);
681 		return (-1);
682 	}
683 
684 	while (bcp != NULL) {
685 		if (mdb_vread(&bc, sizeof (umem_bufctl_t),
686 		    (uintptr_t)bcp) == -1) {
687 			mdb_warn("unable to read bufctl at %p", bcp);
688 			return (-1);
689 		}
690 		if (bc.bc_addr == buf) {
691 			*out = (uintptr_t)bcp;
692 			return (0);
693 		}
694 		bcp = bc.bc_next;
695 	}
696 
697 	mdb_warn("unable to find bufctl for %p in cache %p\n", buf, caddr);
698 	return (-1);
699 }
700 
701 int
702 umem_get_magsize(const umem_cache_t *cp)
703 {
704 	uintptr_t addr = (uintptr_t)cp->cache_magtype;
705 	GElf_Sym mt_sym;
706 	umem_magtype_t mt;
707 	int res;
708 
709 	/*
710 	 * if cpu 0 has a non-zero magsize, it must be correct.  caches
711 	 * with UMF_NOMAGAZINE have disabled their magazine layers, so
712 	 * it is okay to return 0 for them.
713 	 */
714 	if ((res = cp->cache_cpu[0].cc_magsize) != 0 ||
715 	    (cp->cache_flags & UMF_NOMAGAZINE))
716 		return (res);
717 
718 	if (umem_lookup_by_name("umem_magtype", &mt_sym) == -1) {
719 		mdb_warn("unable to read 'umem_magtype'");
720 	} else if (addr < mt_sym.st_value ||
721 	    addr + sizeof (mt) - 1 > mt_sym.st_value + mt_sym.st_size - 1 ||
722 	    ((addr - mt_sym.st_value) % sizeof (mt)) != 0) {
723 		mdb_warn("cache '%s' has invalid magtype pointer (%p)\n",
724 		    cp->cache_name, addr);
725 		return (0);
726 	}
727 	if (mdb_vread(&mt, sizeof (mt), addr) == -1) {
728 		mdb_warn("unable to read magtype at %a", addr);
729 		return (0);
730 	}
731 	return (mt.mt_magsize);
732 }
733 
734 /*ARGSUSED*/
735 static int
736 umem_estimate_slab(uintptr_t addr, const umem_slab_t *sp, size_t *est)
737 {
738 	*est -= (sp->slab_chunks - sp->slab_refcnt);
739 
740 	return (WALK_NEXT);
741 }
742 
743 /*
744  * Returns an upper bound on the number of allocated buffers in a given
745  * cache.
746  */
747 size_t
748 umem_estimate_allocated(uintptr_t addr, const umem_cache_t *cp)
749 {
750 	int magsize;
751 	size_t cache_est;
752 
753 	cache_est = cp->cache_buftotal;
754 
755 	(void) mdb_pwalk("umem_slab_partial",
756 	    (mdb_walk_cb_t)umem_estimate_slab, &cache_est, addr);
757 
758 	if ((magsize = umem_get_magsize(cp)) != 0) {
759 		size_t mag_est = cp->cache_full.ml_total * magsize;
760 
761 		if (cache_est >= mag_est) {
762 			cache_est -= mag_est;
763 		} else {
764 			mdb_warn("cache %p's magazine layer holds more buffers "
765 			    "than the slab layer.\n", addr);
766 		}
767 	}
768 	return (cache_est);
769 }
770 
771 #define	READMAG_ROUNDS(rounds) { \
772 	if (mdb_vread(mp, magbsize, (uintptr_t)ump) == -1) { \
773 		mdb_warn("couldn't read magazine at %p", ump); \
774 		goto fail; \
775 	} \
776 	for (i = 0; i < rounds; i++) { \
777 		maglist[magcnt++] = mp->mag_round[i]; \
778 		if (magcnt == magmax) { \
779 			mdb_warn("%d magazines exceeds fudge factor\n", \
780 			    magcnt); \
781 			goto fail; \
782 		} \
783 	} \
784 }
785 
786 int
787 umem_read_magazines(umem_cache_t *cp, uintptr_t addr,
788     void ***maglistp, size_t *magcntp, size_t *magmaxp, int alloc_flags)
789 {
790 	umem_magazine_t *ump, *mp;
791 	void **maglist = NULL;
792 	int i, cpu;
793 	size_t magsize, magmax, magbsize;
794 	size_t magcnt = 0;
795 
796 	/*
797 	 * Read the magtype out of the cache, after verifying the pointer's
798 	 * correctness.
799 	 */
800 	magsize = umem_get_magsize(cp);
801 	if (magsize == 0) {
802 		*maglistp = NULL;
803 		*magcntp = 0;
804 		*magmaxp = 0;
805 		return (WALK_NEXT);
806 	}
807 
808 	/*
809 	 * There are several places where we need to go buffer hunting:
810 	 * the per-CPU loaded magazine, the per-CPU spare full magazine,
811 	 * and the full magazine list in the depot.
812 	 *
813 	 * For an upper bound on the number of buffers in the magazine
814 	 * layer, we have the number of magazines on the cache_full
815 	 * list plus at most two magazines per CPU (the loaded and the
816 	 * spare).  Toss in 100 magazines as a fudge factor in case this
817 	 * is live (the number "100" comes from the same fudge factor in
818 	 * crash(1M)).
819 	 */
820 	magmax = (cp->cache_full.ml_total + 2 * umem_max_ncpus + 100) * magsize;
821 	magbsize = offsetof(umem_magazine_t, mag_round[magsize]);
822 
823 	if (magbsize >= PAGESIZE / 2) {
824 		mdb_warn("magazine size for cache %p unreasonable (%x)\n",
825 		    addr, magbsize);
826 		return (WALK_ERR);
827 	}
828 
829 	maglist = mdb_alloc(magmax * sizeof (void *), alloc_flags);
830 	mp = mdb_alloc(magbsize, alloc_flags);
831 	if (mp == NULL || maglist == NULL)
832 		goto fail;
833 
834 	/*
835 	 * First up: the magazines in the depot (i.e. on the cache_full list).
836 	 */
837 	for (ump = cp->cache_full.ml_list; ump != NULL; ) {
838 		READMAG_ROUNDS(magsize);
839 		ump = mp->mag_next;
840 
841 		if (ump == cp->cache_full.ml_list)
842 			break; /* cache_full list loop detected */
843 	}
844 
845 	dprintf(("cache_full list done\n"));
846 
847 	/*
848 	 * Now whip through the CPUs, snagging the loaded magazines
849 	 * and full spares.
850 	 */
851 	for (cpu = 0; cpu < umem_max_ncpus; cpu++) {
852 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu];
853 
854 		dprintf(("reading cpu cache %p\n",
855 		    (uintptr_t)ccp - (uintptr_t)cp + addr));
856 
857 		if (ccp->cc_rounds > 0 &&
858 		    (ump = ccp->cc_loaded) != NULL) {
859 			dprintf(("reading %d loaded rounds\n", ccp->cc_rounds));
860 			READMAG_ROUNDS(ccp->cc_rounds);
861 		}
862 
863 		if (ccp->cc_prounds > 0 &&
864 		    (ump = ccp->cc_ploaded) != NULL) {
865 			dprintf(("reading %d previously loaded rounds\n",
866 			    ccp->cc_prounds));
867 			READMAG_ROUNDS(ccp->cc_prounds);
868 		}
869 	}
870 
871 	dprintf(("magazine layer: %d buffers\n", magcnt));
872 
873 	if (!(alloc_flags & UM_GC))
874 		mdb_free(mp, magbsize);
875 
876 	*maglistp = maglist;
877 	*magcntp = magcnt;
878 	*magmaxp = magmax;
879 
880 	return (WALK_NEXT);
881 
882 fail:
883 	if (!(alloc_flags & UM_GC)) {
884 		if (mp)
885 			mdb_free(mp, magbsize);
886 		if (maglist)
887 			mdb_free(maglist, magmax * sizeof (void *));
888 	}
889 	return (WALK_ERR);
890 }
891 
892 static int
893 umem_walk_callback(mdb_walk_state_t *wsp, uintptr_t buf)
894 {
895 	return (wsp->walk_callback(buf, NULL, wsp->walk_cbdata));
896 }
897 
898 static int
899 bufctl_walk_callback(umem_cache_t *cp, mdb_walk_state_t *wsp, uintptr_t buf)
900 {
901 	umem_bufctl_audit_t *b;
902 	UMEM_LOCAL_BUFCTL_AUDIT(&b);
903 
904 	/*
905 	 * if UMF_AUDIT is not set, we know that we're looking at a
906 	 * umem_bufctl_t.
907 	 */
908 	if (!(cp->cache_flags & UMF_AUDIT) ||
909 	    mdb_vread(b, UMEM_BUFCTL_AUDIT_SIZE, buf) == -1) {
910 		(void) memset(b, 0, UMEM_BUFCTL_AUDIT_SIZE);
911 		if (mdb_vread(b, sizeof (umem_bufctl_t), buf) == -1) {
912 			mdb_warn("unable to read bufctl at %p", buf);
913 			return (WALK_ERR);
914 		}
915 	}
916 
917 	return (wsp->walk_callback(buf, b, wsp->walk_cbdata));
918 }
919 
920 typedef struct umem_walk {
921 	int umw_type;
922 
923 	int umw_addr;			/* cache address */
924 	umem_cache_t *umw_cp;
925 	size_t umw_csize;
926 
927 	/*
928 	 * magazine layer
929 	 */
930 	void **umw_maglist;
931 	size_t umw_max;
932 	size_t umw_count;
933 	size_t umw_pos;
934 
935 	/*
936 	 * slab layer
937 	 */
938 	char *umw_valid;	/* to keep track of freed buffers */
939 	char *umw_ubase;	/* buffer for slab data */
940 } umem_walk_t;
941 
942 static int
943 umem_walk_init_common(mdb_walk_state_t *wsp, int type)
944 {
945 	umem_walk_t *umw;
946 	int csize;
947 	umem_cache_t *cp;
948 	size_t vm_quantum;
949 
950 	size_t magmax, magcnt;
951 	void **maglist = NULL;
952 	uint_t chunksize, slabsize;
953 	int status = WALK_ERR;
954 	uintptr_t addr = wsp->walk_addr;
955 	const char *layered;
956 
957 	type &= ~UM_HASH;
958 
959 	if (addr == NULL) {
960 		mdb_warn("umem walk doesn't support global walks\n");
961 		return (WALK_ERR);
962 	}
963 
964 	dprintf(("walking %p\n", addr));
965 
966 	/*
967 	 * The number of "cpus" determines how large the cache is.
968 	 */
969 	csize = UMEM_CACHE_SIZE(umem_max_ncpus);
970 	cp = mdb_alloc(csize, UM_SLEEP);
971 
972 	if (mdb_vread(cp, csize, addr) == -1) {
973 		mdb_warn("couldn't read cache at addr %p", addr);
974 		goto out2;
975 	}
976 
977 	/*
978 	 * It's easy for someone to hand us an invalid cache address.
979 	 * Unfortunately, it is hard for this walker to survive an
980 	 * invalid cache cleanly.  So we make sure that:
981 	 *
982 	 *	1. the vmem arena for the cache is readable,
983 	 *	2. the vmem arena's quantum is a power of 2,
984 	 *	3. our slabsize is a multiple of the quantum, and
985 	 *	4. our chunksize is >0 and less than our slabsize.
986 	 */
987 	if (mdb_vread(&vm_quantum, sizeof (vm_quantum),
988 	    (uintptr_t)&cp->cache_arena->vm_quantum) == -1 ||
989 	    vm_quantum == 0 ||
990 	    (vm_quantum & (vm_quantum - 1)) != 0 ||
991 	    cp->cache_slabsize < vm_quantum ||
992 	    P2PHASE(cp->cache_slabsize, vm_quantum) != 0 ||
993 	    cp->cache_chunksize == 0 ||
994 	    cp->cache_chunksize > cp->cache_slabsize) {
995 		mdb_warn("%p is not a valid umem_cache_t\n", addr);
996 		goto out2;
997 	}
998 
999 	dprintf(("buf total is %d\n", cp->cache_buftotal));
1000 
1001 	if (cp->cache_buftotal == 0) {
1002 		mdb_free(cp, csize);
1003 		return (WALK_DONE);
1004 	}
1005 
1006 	/*
1007 	 * If they ask for bufctls, but it's a small-slab cache,
1008 	 * there is nothing to report.
1009 	 */
1010 	if ((type & UM_BUFCTL) && !(cp->cache_flags & UMF_HASH)) {
1011 		dprintf(("bufctl requested, not UMF_HASH (flags: %p)\n",
1012 		    cp->cache_flags));
1013 		mdb_free(cp, csize);
1014 		return (WALK_DONE);
1015 	}
1016 
1017 	/*
1018 	 * Read in the contents of the magazine layer
1019 	 */
1020 	if (umem_read_magazines(cp, addr, &maglist, &magcnt, &magmax,
1021 	    UM_SLEEP) == WALK_ERR)
1022 		goto out2;
1023 
1024 	/*
1025 	 * We have all of the buffers from the magazines;  if we are walking
1026 	 * allocated buffers, sort them so we can bsearch them later.
1027 	 */
1028 	if (type & UM_ALLOCATED)
1029 		qsort(maglist, magcnt, sizeof (void *), addrcmp);
1030 
1031 	wsp->walk_data = umw = mdb_zalloc(sizeof (umem_walk_t), UM_SLEEP);
1032 
1033 	umw->umw_type = type;
1034 	umw->umw_addr = addr;
1035 	umw->umw_cp = cp;
1036 	umw->umw_csize = csize;
1037 	umw->umw_maglist = maglist;
1038 	umw->umw_max = magmax;
1039 	umw->umw_count = magcnt;
1040 	umw->umw_pos = 0;
1041 
1042 	/*
1043 	 * When walking allocated buffers in a UMF_HASH cache, we walk the
1044 	 * hash table instead of the slab layer.
1045 	 */
1046 	if ((cp->cache_flags & UMF_HASH) && (type & UM_ALLOCATED)) {
1047 		layered = "umem_hash";
1048 
1049 		umw->umw_type |= UM_HASH;
1050 	} else {
1051 		/*
1052 		 * If we are walking freed buffers, we only need the
1053 		 * magazine layer plus the partially allocated slabs.
1054 		 * To walk allocated buffers, we need all of the slabs.
1055 		 */
1056 		if (type & UM_ALLOCATED)
1057 			layered = "umem_slab";
1058 		else
1059 			layered = "umem_slab_partial";
1060 
1061 		/*
1062 		 * for small-slab caches, we read in the entire slab.  For
1063 		 * freed buffers, we can just walk the freelist.  For
1064 		 * allocated buffers, we use a 'valid' array to track
1065 		 * the freed buffers.
1066 		 */
1067 		if (!(cp->cache_flags & UMF_HASH)) {
1068 			chunksize = cp->cache_chunksize;
1069 			slabsize = cp->cache_slabsize;
1070 
1071 			umw->umw_ubase = mdb_alloc(slabsize +
1072 			    sizeof (umem_bufctl_t), UM_SLEEP);
1073 
1074 			if (type & UM_ALLOCATED)
1075 				umw->umw_valid =
1076 				    mdb_alloc(slabsize / chunksize, UM_SLEEP);
1077 		}
1078 	}
1079 
1080 	status = WALK_NEXT;
1081 
1082 	if (mdb_layered_walk(layered, wsp) == -1) {
1083 		mdb_warn("unable to start layered '%s' walk", layered);
1084 		status = WALK_ERR;
1085 	}
1086 
1087 out1:
1088 	if (status == WALK_ERR) {
1089 		if (umw->umw_valid)
1090 			mdb_free(umw->umw_valid, slabsize / chunksize);
1091 
1092 		if (umw->umw_ubase)
1093 			mdb_free(umw->umw_ubase, slabsize +
1094 			    sizeof (umem_bufctl_t));
1095 
1096 		if (umw->umw_maglist)
1097 			mdb_free(umw->umw_maglist, umw->umw_max *
1098 			    sizeof (uintptr_t));
1099 
1100 		mdb_free(umw, sizeof (umem_walk_t));
1101 		wsp->walk_data = NULL;
1102 	}
1103 
1104 out2:
1105 	if (status == WALK_ERR)
1106 		mdb_free(cp, csize);
1107 
1108 	return (status);
1109 }
1110 
1111 int
1112 umem_walk_step(mdb_walk_state_t *wsp)
1113 {
1114 	umem_walk_t *umw = wsp->walk_data;
1115 	int type = umw->umw_type;
1116 	umem_cache_t *cp = umw->umw_cp;
1117 
1118 	void **maglist = umw->umw_maglist;
1119 	int magcnt = umw->umw_count;
1120 
1121 	uintptr_t chunksize, slabsize;
1122 	uintptr_t addr;
1123 	const umem_slab_t *sp;
1124 	const umem_bufctl_t *bcp;
1125 	umem_bufctl_t bc;
1126 
1127 	int chunks;
1128 	char *kbase;
1129 	void *buf;
1130 	int i, ret;
1131 
1132 	char *valid, *ubase;
1133 
1134 	/*
1135 	 * first, handle the 'umem_hash' layered walk case
1136 	 */
1137 	if (type & UM_HASH) {
1138 		/*
1139 		 * We have a buffer which has been allocated out of the
1140 		 * global layer. We need to make sure that it's not
1141 		 * actually sitting in a magazine before we report it as
1142 		 * an allocated buffer.
1143 		 */
1144 		buf = ((const umem_bufctl_t *)wsp->walk_layer)->bc_addr;
1145 
1146 		if (magcnt > 0 &&
1147 		    bsearch(&buf, maglist, magcnt, sizeof (void *),
1148 		    addrcmp) != NULL)
1149 			return (WALK_NEXT);
1150 
1151 		if (type & UM_BUFCTL)
1152 			return (bufctl_walk_callback(cp, wsp, wsp->walk_addr));
1153 
1154 		return (umem_walk_callback(wsp, (uintptr_t)buf));
1155 	}
1156 
1157 	ret = WALK_NEXT;
1158 
1159 	addr = umw->umw_addr;
1160 
1161 	/*
1162 	 * If we're walking freed buffers, report everything in the
1163 	 * magazine layer before processing the first slab.
1164 	 */
1165 	if ((type & UM_FREE) && magcnt != 0) {
1166 		umw->umw_count = 0;		/* only do this once */
1167 		for (i = 0; i < magcnt; i++) {
1168 			buf = maglist[i];
1169 
1170 			if (type & UM_BUFCTL) {
1171 				uintptr_t out;
1172 
1173 				if (cp->cache_flags & UMF_BUFTAG) {
1174 					umem_buftag_t *btp;
1175 					umem_buftag_t tag;
1176 
1177 					/* LINTED - alignment */
1178 					btp = UMEM_BUFTAG(cp, buf);
1179 					if (mdb_vread(&tag, sizeof (tag),
1180 					    (uintptr_t)btp) == -1) {
1181 						mdb_warn("reading buftag for "
1182 						    "%p at %p", buf, btp);
1183 						continue;
1184 					}
1185 					out = (uintptr_t)tag.bt_bufctl;
1186 				} else {
1187 					if (umem_hash_lookup(cp, addr, buf,
1188 					    &out) == -1)
1189 						continue;
1190 				}
1191 				ret = bufctl_walk_callback(cp, wsp, out);
1192 			} else {
1193 				ret = umem_walk_callback(wsp, (uintptr_t)buf);
1194 			}
1195 
1196 			if (ret != WALK_NEXT)
1197 				return (ret);
1198 		}
1199 	}
1200 
1201 	/*
1202 	 * Handle the buffers in the current slab
1203 	 */
1204 	chunksize = cp->cache_chunksize;
1205 	slabsize = cp->cache_slabsize;
1206 
1207 	sp = wsp->walk_layer;
1208 	chunks = sp->slab_chunks;
1209 	kbase = sp->slab_base;
1210 
1211 	dprintf(("kbase is %p\n", kbase));
1212 
1213 	if (!(cp->cache_flags & UMF_HASH)) {
1214 		valid = umw->umw_valid;
1215 		ubase = umw->umw_ubase;
1216 
1217 		if (mdb_vread(ubase, chunks * chunksize,
1218 		    (uintptr_t)kbase) == -1) {
1219 			mdb_warn("failed to read slab contents at %p", kbase);
1220 			return (WALK_ERR);
1221 		}
1222 
1223 		/*
1224 		 * Set up the valid map as fully allocated -- we'll punch
1225 		 * out the freelist.
1226 		 */
1227 		if (type & UM_ALLOCATED)
1228 			(void) memset(valid, 1, chunks);
1229 	} else {
1230 		valid = NULL;
1231 		ubase = NULL;
1232 	}
1233 
1234 	/*
1235 	 * walk the slab's freelist
1236 	 */
1237 	bcp = sp->slab_head;
1238 
1239 	dprintf(("refcnt is %d; chunks is %d\n", sp->slab_refcnt, chunks));
1240 
1241 	/*
1242 	 * since we could be in the middle of allocating a buffer,
1243 	 * our refcnt could be one higher than it aught.  So we
1244 	 * check one further on the freelist than the count allows.
1245 	 */
1246 	for (i = sp->slab_refcnt; i <= chunks; i++) {
1247 		uint_t ndx;
1248 
1249 		dprintf(("bcp is %p\n", bcp));
1250 
1251 		if (bcp == NULL) {
1252 			if (i == chunks)
1253 				break;
1254 			mdb_warn(
1255 			    "slab %p in cache %p freelist too short by %d\n",
1256 			    sp, addr, chunks - i);
1257 			break;
1258 		}
1259 
1260 		if (cp->cache_flags & UMF_HASH) {
1261 			if (mdb_vread(&bc, sizeof (bc), (uintptr_t)bcp) == -1) {
1262 				mdb_warn("failed to read bufctl ptr at %p",
1263 				    bcp);
1264 				break;
1265 			}
1266 			buf = bc.bc_addr;
1267 		} else {
1268 			/*
1269 			 * Otherwise the buffer is in the slab which
1270 			 * we've read in;  we just need to determine
1271 			 * its offset in the slab to find the
1272 			 * umem_bufctl_t.
1273 			 */
1274 			bc = *((umem_bufctl_t *)
1275 			    ((uintptr_t)bcp - (uintptr_t)kbase +
1276 			    (uintptr_t)ubase));
1277 
1278 			buf = UMEM_BUF(cp, bcp);
1279 		}
1280 
1281 		ndx = ((uintptr_t)buf - (uintptr_t)kbase) / chunksize;
1282 
1283 		if (ndx > slabsize / cp->cache_bufsize) {
1284 			/*
1285 			 * This is very wrong; we have managed to find
1286 			 * a buffer in the slab which shouldn't
1287 			 * actually be here.  Emit a warning, and
1288 			 * try to continue.
1289 			 */
1290 			mdb_warn("buf %p is out of range for "
1291 			    "slab %p, cache %p\n", buf, sp, addr);
1292 		} else if (type & UM_ALLOCATED) {
1293 			/*
1294 			 * we have found a buffer on the slab's freelist;
1295 			 * clear its entry
1296 			 */
1297 			valid[ndx] = 0;
1298 		} else {
1299 			/*
1300 			 * Report this freed buffer
1301 			 */
1302 			if (type & UM_BUFCTL) {
1303 				ret = bufctl_walk_callback(cp, wsp,
1304 				    (uintptr_t)bcp);
1305 			} else {
1306 				ret = umem_walk_callback(wsp, (uintptr_t)buf);
1307 			}
1308 			if (ret != WALK_NEXT)
1309 				return (ret);
1310 		}
1311 
1312 		bcp = bc.bc_next;
1313 	}
1314 
1315 	if (bcp != NULL) {
1316 		dprintf(("slab %p in cache %p freelist too long (%p)\n",
1317 		    sp, addr, bcp));
1318 	}
1319 
1320 	/*
1321 	 * If we are walking freed buffers, the loop above handled reporting
1322 	 * them.
1323 	 */
1324 	if (type & UM_FREE)
1325 		return (WALK_NEXT);
1326 
1327 	if (type & UM_BUFCTL) {
1328 		mdb_warn("impossible situation: small-slab UM_BUFCTL walk for "
1329 		    "cache %p\n", addr);
1330 		return (WALK_ERR);
1331 	}
1332 
1333 	/*
1334 	 * Report allocated buffers, skipping buffers in the magazine layer.
1335 	 * We only get this far for small-slab caches.
1336 	 */
1337 	for (i = 0; ret == WALK_NEXT && i < chunks; i++) {
1338 		buf = (char *)kbase + i * chunksize;
1339 
1340 		if (!valid[i])
1341 			continue;		/* on slab freelist */
1342 
1343 		if (magcnt > 0 &&
1344 		    bsearch(&buf, maglist, magcnt, sizeof (void *),
1345 		    addrcmp) != NULL)
1346 			continue;		/* in magazine layer */
1347 
1348 		ret = umem_walk_callback(wsp, (uintptr_t)buf);
1349 	}
1350 	return (ret);
1351 }
1352 
1353 void
1354 umem_walk_fini(mdb_walk_state_t *wsp)
1355 {
1356 	umem_walk_t *umw = wsp->walk_data;
1357 	uintptr_t chunksize;
1358 	uintptr_t slabsize;
1359 
1360 	if (umw == NULL)
1361 		return;
1362 
1363 	if (umw->umw_maglist != NULL)
1364 		mdb_free(umw->umw_maglist, umw->umw_max * sizeof (void *));
1365 
1366 	chunksize = umw->umw_cp->cache_chunksize;
1367 	slabsize = umw->umw_cp->cache_slabsize;
1368 
1369 	if (umw->umw_valid != NULL)
1370 		mdb_free(umw->umw_valid, slabsize / chunksize);
1371 	if (umw->umw_ubase != NULL)
1372 		mdb_free(umw->umw_ubase, slabsize + sizeof (umem_bufctl_t));
1373 
1374 	mdb_free(umw->umw_cp, umw->umw_csize);
1375 	mdb_free(umw, sizeof (umem_walk_t));
1376 }
1377 
1378 /*ARGSUSED*/
1379 static int
1380 umem_walk_all(uintptr_t addr, const umem_cache_t *c, mdb_walk_state_t *wsp)
1381 {
1382 	/*
1383 	 * Buffers allocated from NOTOUCH caches can also show up as freed
1384 	 * memory in other caches.  This can be a little confusing, so we
1385 	 * don't walk NOTOUCH caches when walking all caches (thereby assuring
1386 	 * that "::walk umem" and "::walk freemem" yield disjoint output).
1387 	 */
1388 	if (c->cache_cflags & UMC_NOTOUCH)
1389 		return (WALK_NEXT);
1390 
1391 	if (mdb_pwalk(wsp->walk_data, wsp->walk_callback,
1392 	    wsp->walk_cbdata, addr) == -1)
1393 		return (WALK_DONE);
1394 
1395 	return (WALK_NEXT);
1396 }
1397 
1398 #define	UMEM_WALK_ALL(name, wsp) { \
1399 	wsp->walk_data = (name); \
1400 	if (mdb_walk("umem_cache", (mdb_walk_cb_t)umem_walk_all, wsp) == -1) \
1401 		return (WALK_ERR); \
1402 	return (WALK_DONE); \
1403 }
1404 
1405 int
1406 umem_walk_init(mdb_walk_state_t *wsp)
1407 {
1408 	if (wsp->walk_arg != NULL)
1409 		wsp->walk_addr = (uintptr_t)wsp->walk_arg;
1410 
1411 	if (wsp->walk_addr == NULL)
1412 		UMEM_WALK_ALL("umem", wsp);
1413 	return (umem_walk_init_common(wsp, UM_ALLOCATED));
1414 }
1415 
1416 int
1417 bufctl_walk_init(mdb_walk_state_t *wsp)
1418 {
1419 	if (wsp->walk_addr == NULL)
1420 		UMEM_WALK_ALL("bufctl", wsp);
1421 	return (umem_walk_init_common(wsp, UM_ALLOCATED | UM_BUFCTL));
1422 }
1423 
1424 int
1425 freemem_walk_init(mdb_walk_state_t *wsp)
1426 {
1427 	if (wsp->walk_addr == NULL)
1428 		UMEM_WALK_ALL("freemem", wsp);
1429 	return (umem_walk_init_common(wsp, UM_FREE));
1430 }
1431 
1432 int
1433 freectl_walk_init(mdb_walk_state_t *wsp)
1434 {
1435 	if (wsp->walk_addr == NULL)
1436 		UMEM_WALK_ALL("freectl", wsp);
1437 	return (umem_walk_init_common(wsp, UM_FREE | UM_BUFCTL));
1438 }
1439 
1440 typedef struct bufctl_history_walk {
1441 	void		*bhw_next;
1442 	umem_cache_t	*bhw_cache;
1443 	umem_slab_t	*bhw_slab;
1444 	hrtime_t	bhw_timestamp;
1445 } bufctl_history_walk_t;
1446 
1447 int
1448 bufctl_history_walk_init(mdb_walk_state_t *wsp)
1449 {
1450 	bufctl_history_walk_t *bhw;
1451 	umem_bufctl_audit_t bc;
1452 	umem_bufctl_audit_t bcn;
1453 
1454 	if (wsp->walk_addr == NULL) {
1455 		mdb_warn("bufctl_history walk doesn't support global walks\n");
1456 		return (WALK_ERR);
1457 	}
1458 
1459 	if (mdb_vread(&bc, sizeof (bc), wsp->walk_addr) == -1) {
1460 		mdb_warn("unable to read bufctl at %p", wsp->walk_addr);
1461 		return (WALK_ERR);
1462 	}
1463 
1464 	bhw = mdb_zalloc(sizeof (*bhw), UM_SLEEP);
1465 	bhw->bhw_timestamp = 0;
1466 	bhw->bhw_cache = bc.bc_cache;
1467 	bhw->bhw_slab = bc.bc_slab;
1468 
1469 	/*
1470 	 * sometimes the first log entry matches the base bufctl;  in that
1471 	 * case, skip the base bufctl.
1472 	 */
1473 	if (bc.bc_lastlog != NULL &&
1474 	    mdb_vread(&bcn, sizeof (bcn), (uintptr_t)bc.bc_lastlog) != -1 &&
1475 	    bc.bc_addr == bcn.bc_addr &&
1476 	    bc.bc_cache == bcn.bc_cache &&
1477 	    bc.bc_slab == bcn.bc_slab &&
1478 	    bc.bc_timestamp == bcn.bc_timestamp &&
1479 	    bc.bc_thread == bcn.bc_thread)
1480 		bhw->bhw_next = bc.bc_lastlog;
1481 	else
1482 		bhw->bhw_next = (void *)wsp->walk_addr;
1483 
1484 	wsp->walk_addr = (uintptr_t)bc.bc_addr;
1485 	wsp->walk_data = bhw;
1486 
1487 	return (WALK_NEXT);
1488 }
1489 
1490 int
1491 bufctl_history_walk_step(mdb_walk_state_t *wsp)
1492 {
1493 	bufctl_history_walk_t *bhw = wsp->walk_data;
1494 	uintptr_t addr = (uintptr_t)bhw->bhw_next;
1495 	uintptr_t baseaddr = wsp->walk_addr;
1496 	umem_bufctl_audit_t *b;
1497 	UMEM_LOCAL_BUFCTL_AUDIT(&b);
1498 
1499 	if (addr == NULL)
1500 		return (WALK_DONE);
1501 
1502 	if (mdb_vread(b, UMEM_BUFCTL_AUDIT_SIZE, addr) == -1) {
1503 		mdb_warn("unable to read bufctl at %p", bhw->bhw_next);
1504 		return (WALK_ERR);
1505 	}
1506 
1507 	/*
1508 	 * The bufctl is only valid if the address, cache, and slab are
1509 	 * correct.  We also check that the timestamp is decreasing, to
1510 	 * prevent infinite loops.
1511 	 */
1512 	if ((uintptr_t)b->bc_addr != baseaddr ||
1513 	    b->bc_cache != bhw->bhw_cache ||
1514 	    b->bc_slab != bhw->bhw_slab ||
1515 	    (bhw->bhw_timestamp != 0 && b->bc_timestamp >= bhw->bhw_timestamp))
1516 		return (WALK_DONE);
1517 
1518 	bhw->bhw_next = b->bc_lastlog;
1519 	bhw->bhw_timestamp = b->bc_timestamp;
1520 
1521 	return (wsp->walk_callback(addr, b, wsp->walk_cbdata));
1522 }
1523 
1524 void
1525 bufctl_history_walk_fini(mdb_walk_state_t *wsp)
1526 {
1527 	bufctl_history_walk_t *bhw = wsp->walk_data;
1528 
1529 	mdb_free(bhw, sizeof (*bhw));
1530 }
1531 
1532 typedef struct umem_log_walk {
1533 	umem_bufctl_audit_t *ulw_base;
1534 	umem_bufctl_audit_t **ulw_sorted;
1535 	umem_log_header_t ulw_lh;
1536 	size_t ulw_size;
1537 	size_t ulw_maxndx;
1538 	size_t ulw_ndx;
1539 } umem_log_walk_t;
1540 
1541 int
1542 umem_log_walk_init(mdb_walk_state_t *wsp)
1543 {
1544 	uintptr_t lp = wsp->walk_addr;
1545 	umem_log_walk_t *ulw;
1546 	umem_log_header_t *lhp;
1547 	int maxndx, i, j, k;
1548 
1549 	/*
1550 	 * By default (global walk), walk the umem_transaction_log.  Otherwise
1551 	 * read the log whose umem_log_header_t is stored at walk_addr.
1552 	 */
1553 	if (lp == NULL && umem_readvar(&lp, "umem_transaction_log") == -1) {
1554 		mdb_warn("failed to read 'umem_transaction_log'");
1555 		return (WALK_ERR);
1556 	}
1557 
1558 	if (lp == NULL) {
1559 		mdb_warn("log is disabled\n");
1560 		return (WALK_ERR);
1561 	}
1562 
1563 	ulw = mdb_zalloc(sizeof (umem_log_walk_t), UM_SLEEP);
1564 	lhp = &ulw->ulw_lh;
1565 
1566 	if (mdb_vread(lhp, sizeof (umem_log_header_t), lp) == -1) {
1567 		mdb_warn("failed to read log header at %p", lp);
1568 		mdb_free(ulw, sizeof (umem_log_walk_t));
1569 		return (WALK_ERR);
1570 	}
1571 
1572 	ulw->ulw_size = lhp->lh_chunksize * lhp->lh_nchunks;
1573 	ulw->ulw_base = mdb_alloc(ulw->ulw_size, UM_SLEEP);
1574 	maxndx = lhp->lh_chunksize / UMEM_BUFCTL_AUDIT_SIZE - 1;
1575 
1576 	if (mdb_vread(ulw->ulw_base, ulw->ulw_size,
1577 	    (uintptr_t)lhp->lh_base) == -1) {
1578 		mdb_warn("failed to read log at base %p", lhp->lh_base);
1579 		mdb_free(ulw->ulw_base, ulw->ulw_size);
1580 		mdb_free(ulw, sizeof (umem_log_walk_t));
1581 		return (WALK_ERR);
1582 	}
1583 
1584 	ulw->ulw_sorted = mdb_alloc(maxndx * lhp->lh_nchunks *
1585 	    sizeof (umem_bufctl_audit_t *), UM_SLEEP);
1586 
1587 	for (i = 0, k = 0; i < lhp->lh_nchunks; i++) {
1588 		caddr_t chunk = (caddr_t)
1589 		    ((uintptr_t)ulw->ulw_base + i * lhp->lh_chunksize);
1590 
1591 		for (j = 0; j < maxndx; j++) {
1592 			/* LINTED align */
1593 			ulw->ulw_sorted[k++] = (umem_bufctl_audit_t *)chunk;
1594 			chunk += UMEM_BUFCTL_AUDIT_SIZE;
1595 		}
1596 	}
1597 
1598 	qsort(ulw->ulw_sorted, k, sizeof (umem_bufctl_audit_t *),
1599 	    (int(*)(const void *, const void *))bufctlcmp);
1600 
1601 	ulw->ulw_maxndx = k;
1602 	wsp->walk_data = ulw;
1603 
1604 	return (WALK_NEXT);
1605 }
1606 
1607 int
1608 umem_log_walk_step(mdb_walk_state_t *wsp)
1609 {
1610 	umem_log_walk_t *ulw = wsp->walk_data;
1611 	umem_bufctl_audit_t *bcp;
1612 
1613 	if (ulw->ulw_ndx == ulw->ulw_maxndx)
1614 		return (WALK_DONE);
1615 
1616 	bcp = ulw->ulw_sorted[ulw->ulw_ndx++];
1617 
1618 	return (wsp->walk_callback((uintptr_t)bcp - (uintptr_t)ulw->ulw_base +
1619 	    (uintptr_t)ulw->ulw_lh.lh_base, bcp, wsp->walk_cbdata));
1620 }
1621 
1622 void
1623 umem_log_walk_fini(mdb_walk_state_t *wsp)
1624 {
1625 	umem_log_walk_t *ulw = wsp->walk_data;
1626 
1627 	mdb_free(ulw->ulw_base, ulw->ulw_size);
1628 	mdb_free(ulw->ulw_sorted, ulw->ulw_maxndx *
1629 	    sizeof (umem_bufctl_audit_t *));
1630 	mdb_free(ulw, sizeof (umem_log_walk_t));
1631 }
1632 
1633 typedef struct allocdby_bufctl {
1634 	uintptr_t abb_addr;
1635 	hrtime_t abb_ts;
1636 } allocdby_bufctl_t;
1637 
1638 typedef struct allocdby_walk {
1639 	const char *abw_walk;
1640 	uintptr_t abw_thread;
1641 	size_t abw_nbufs;
1642 	size_t abw_size;
1643 	allocdby_bufctl_t *abw_buf;
1644 	size_t abw_ndx;
1645 } allocdby_walk_t;
1646 
1647 int
1648 allocdby_walk_bufctl(uintptr_t addr, const umem_bufctl_audit_t *bcp,
1649     allocdby_walk_t *abw)
1650 {
1651 	if ((uintptr_t)bcp->bc_thread != abw->abw_thread)
1652 		return (WALK_NEXT);
1653 
1654 	if (abw->abw_nbufs == abw->abw_size) {
1655 		allocdby_bufctl_t *buf;
1656 		size_t oldsize = sizeof (allocdby_bufctl_t) * abw->abw_size;
1657 
1658 		buf = mdb_zalloc(oldsize << 1, UM_SLEEP);
1659 
1660 		bcopy(abw->abw_buf, buf, oldsize);
1661 		mdb_free(abw->abw_buf, oldsize);
1662 
1663 		abw->abw_size <<= 1;
1664 		abw->abw_buf = buf;
1665 	}
1666 
1667 	abw->abw_buf[abw->abw_nbufs].abb_addr = addr;
1668 	abw->abw_buf[abw->abw_nbufs].abb_ts = bcp->bc_timestamp;
1669 	abw->abw_nbufs++;
1670 
1671 	return (WALK_NEXT);
1672 }
1673 
1674 /*ARGSUSED*/
1675 int
1676 allocdby_walk_cache(uintptr_t addr, const umem_cache_t *c, allocdby_walk_t *abw)
1677 {
1678 	if (mdb_pwalk(abw->abw_walk, (mdb_walk_cb_t)allocdby_walk_bufctl,
1679 	    abw, addr) == -1) {
1680 		mdb_warn("couldn't walk bufctl for cache %p", addr);
1681 		return (WALK_DONE);
1682 	}
1683 
1684 	return (WALK_NEXT);
1685 }
1686 
1687 static int
1688 allocdby_cmp(const allocdby_bufctl_t *lhs, const allocdby_bufctl_t *rhs)
1689 {
1690 	if (lhs->abb_ts < rhs->abb_ts)
1691 		return (1);
1692 	if (lhs->abb_ts > rhs->abb_ts)
1693 		return (-1);
1694 	return (0);
1695 }
1696 
1697 static int
1698 allocdby_walk_init_common(mdb_walk_state_t *wsp, const char *walk)
1699 {
1700 	allocdby_walk_t *abw;
1701 
1702 	if (wsp->walk_addr == NULL) {
1703 		mdb_warn("allocdby walk doesn't support global walks\n");
1704 		return (WALK_ERR);
1705 	}
1706 
1707 	abw = mdb_zalloc(sizeof (allocdby_walk_t), UM_SLEEP);
1708 
1709 	abw->abw_thread = wsp->walk_addr;
1710 	abw->abw_walk = walk;
1711 	abw->abw_size = 128;	/* something reasonable */
1712 	abw->abw_buf =
1713 	    mdb_zalloc(abw->abw_size * sizeof (allocdby_bufctl_t), UM_SLEEP);
1714 
1715 	wsp->walk_data = abw;
1716 
1717 	if (mdb_walk("umem_cache",
1718 	    (mdb_walk_cb_t)allocdby_walk_cache, abw) == -1) {
1719 		mdb_warn("couldn't walk umem_cache");
1720 		allocdby_walk_fini(wsp);
1721 		return (WALK_ERR);
1722 	}
1723 
1724 	qsort(abw->abw_buf, abw->abw_nbufs, sizeof (allocdby_bufctl_t),
1725 	    (int(*)(const void *, const void *))allocdby_cmp);
1726 
1727 	return (WALK_NEXT);
1728 }
1729 
1730 int
1731 allocdby_walk_init(mdb_walk_state_t *wsp)
1732 {
1733 	return (allocdby_walk_init_common(wsp, "bufctl"));
1734 }
1735 
1736 int
1737 freedby_walk_init(mdb_walk_state_t *wsp)
1738 {
1739 	return (allocdby_walk_init_common(wsp, "freectl"));
1740 }
1741 
1742 int
1743 allocdby_walk_step(mdb_walk_state_t *wsp)
1744 {
1745 	allocdby_walk_t *abw = wsp->walk_data;
1746 	uintptr_t addr;
1747 	umem_bufctl_audit_t *bcp;
1748 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
1749 
1750 	if (abw->abw_ndx == abw->abw_nbufs)
1751 		return (WALK_DONE);
1752 
1753 	addr = abw->abw_buf[abw->abw_ndx++].abb_addr;
1754 
1755 	if (mdb_vread(bcp, UMEM_BUFCTL_AUDIT_SIZE, addr) == -1) {
1756 		mdb_warn("couldn't read bufctl at %p", addr);
1757 		return (WALK_DONE);
1758 	}
1759 
1760 	return (wsp->walk_callback(addr, bcp, wsp->walk_cbdata));
1761 }
1762 
1763 void
1764 allocdby_walk_fini(mdb_walk_state_t *wsp)
1765 {
1766 	allocdby_walk_t *abw = wsp->walk_data;
1767 
1768 	mdb_free(abw->abw_buf, sizeof (allocdby_bufctl_t) * abw->abw_size);
1769 	mdb_free(abw, sizeof (allocdby_walk_t));
1770 }
1771 
1772 /*ARGSUSED*/
1773 int
1774 allocdby_walk(uintptr_t addr, const umem_bufctl_audit_t *bcp, void *ignored)
1775 {
1776 	char c[MDB_SYM_NAMLEN];
1777 	GElf_Sym sym;
1778 	int i;
1779 
1780 	mdb_printf("%0?p %12llx ", addr, bcp->bc_timestamp);
1781 	for (i = 0; i < bcp->bc_depth; i++) {
1782 		if (mdb_lookup_by_addr(bcp->bc_stack[i],
1783 		    MDB_SYM_FUZZY, c, sizeof (c), &sym) == -1)
1784 			continue;
1785 		if (is_umem_sym(c, "umem_"))
1786 			continue;
1787 		mdb_printf("%s+0x%lx",
1788 		    c, bcp->bc_stack[i] - (uintptr_t)sym.st_value);
1789 		break;
1790 	}
1791 	mdb_printf("\n");
1792 
1793 	return (WALK_NEXT);
1794 }
1795 
1796 static int
1797 allocdby_common(uintptr_t addr, uint_t flags, const char *w)
1798 {
1799 	if (!(flags & DCMD_ADDRSPEC))
1800 		return (DCMD_USAGE);
1801 
1802 	mdb_printf("%-?s %12s %s\n", "BUFCTL", "TIMESTAMP", "CALLER");
1803 
1804 	if (mdb_pwalk(w, (mdb_walk_cb_t)allocdby_walk, NULL, addr) == -1) {
1805 		mdb_warn("can't walk '%s' for %p", w, addr);
1806 		return (DCMD_ERR);
1807 	}
1808 
1809 	return (DCMD_OK);
1810 }
1811 
1812 /*ARGSUSED*/
1813 int
1814 allocdby(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
1815 {
1816 	return (allocdby_common(addr, flags, "allocdby"));
1817 }
1818 
1819 /*ARGSUSED*/
1820 int
1821 freedby(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
1822 {
1823 	return (allocdby_common(addr, flags, "freedby"));
1824 }
1825 
1826 typedef struct whatis_info {
1827 	mdb_whatis_t *wi_w;
1828 	const umem_cache_t *wi_cache;
1829 	const vmem_t *wi_vmem;
1830 	vmem_t *wi_msb_arena;
1831 	size_t wi_slab_size;
1832 	int wi_slab_found;
1833 	uint_t wi_freemem;
1834 } whatis_info_t;
1835 
1836 /* call one of our dcmd functions with "-v" and the provided address */
1837 static void
1838 whatis_call_printer(mdb_dcmd_f *dcmd, uintptr_t addr)
1839 {
1840 	mdb_arg_t a;
1841 	a.a_type = MDB_TYPE_STRING;
1842 	a.a_un.a_str = "-v";
1843 
1844 	mdb_printf(":\n");
1845 	(void) (*dcmd)(addr, DCMD_ADDRSPEC, 1, &a);
1846 }
1847 
1848 static void
1849 whatis_print_umem(whatis_info_t *wi, uintptr_t maddr, uintptr_t addr,
1850     uintptr_t baddr)
1851 {
1852 	mdb_whatis_t *w = wi->wi_w;
1853 	const umem_cache_t *cp = wi->wi_cache;
1854 	int quiet = (mdb_whatis_flags(w) & WHATIS_QUIET);
1855 
1856 	int call_printer = (!quiet && (cp->cache_flags & UMF_AUDIT));
1857 
1858 	mdb_whatis_report_object(w, maddr, addr, "");
1859 
1860 	if (baddr != 0 && !call_printer)
1861 		mdb_printf("bufctl %p ", baddr);
1862 
1863 	mdb_printf("%s from %s",
1864 	    (wi->wi_freemem == FALSE) ? "allocated" : "freed", cp->cache_name);
1865 
1866 	if (call_printer && baddr != 0) {
1867 		whatis_call_printer(bufctl, baddr);
1868 		return;
1869 	}
1870 	mdb_printf("\n");
1871 }
1872 
1873 /*ARGSUSED*/
1874 static int
1875 whatis_walk_umem(uintptr_t addr, void *ignored, whatis_info_t *wi)
1876 {
1877 	mdb_whatis_t *w = wi->wi_w;
1878 
1879 	uintptr_t cur;
1880 	size_t size = wi->wi_cache->cache_bufsize;
1881 
1882 	while (mdb_whatis_match(w, addr, size, &cur))
1883 		whatis_print_umem(wi, cur, addr, NULL);
1884 
1885 	return (WHATIS_WALKRET(w));
1886 }
1887 
1888 /*ARGSUSED*/
1889 static int
1890 whatis_walk_bufctl(uintptr_t baddr, const umem_bufctl_t *bcp, whatis_info_t *wi)
1891 {
1892 	mdb_whatis_t *w = wi->wi_w;
1893 
1894 	uintptr_t cur;
1895 	uintptr_t addr = (uintptr_t)bcp->bc_addr;
1896 	size_t size = wi->wi_cache->cache_bufsize;
1897 
1898 	while (mdb_whatis_match(w, addr, size, &cur))
1899 		whatis_print_umem(wi, cur, addr, baddr);
1900 
1901 	return (WHATIS_WALKRET(w));
1902 }
1903 
1904 
1905 static int
1906 whatis_walk_seg(uintptr_t addr, const vmem_seg_t *vs, whatis_info_t *wi)
1907 {
1908 	mdb_whatis_t *w = wi->wi_w;
1909 
1910 	size_t size = vs->vs_end - vs->vs_start;
1911 	uintptr_t cur;
1912 
1913 	/* We're not interested in anything but alloc and free segments */
1914 	if (vs->vs_type != VMEM_ALLOC && vs->vs_type != VMEM_FREE)
1915 		return (WALK_NEXT);
1916 
1917 	while (mdb_whatis_match(w, vs->vs_start, size, &cur)) {
1918 		mdb_whatis_report_object(w, cur, vs->vs_start, "");
1919 
1920 		/*
1921 		 * If we're not printing it seperately, provide the vmem_seg
1922 		 * pointer if it has a stack trace.
1923 		 */
1924 		if ((mdb_whatis_flags(w) & WHATIS_QUIET) &&
1925 		    ((mdb_whatis_flags(w) & WHATIS_BUFCTL) != 0 ||
1926 		    (vs->vs_type == VMEM_ALLOC && vs->vs_depth != 0))) {
1927 			mdb_printf("vmem_seg %p ", addr);
1928 		}
1929 
1930 		mdb_printf("%s from %s vmem arena",
1931 		    (vs->vs_type == VMEM_ALLOC) ? "allocated" : "freed",
1932 		    wi->wi_vmem->vm_name);
1933 
1934 		if (!mdb_whatis_flags(w) & WHATIS_QUIET)
1935 			whatis_call_printer(vmem_seg, addr);
1936 		else
1937 			mdb_printf("\n");
1938 	}
1939 
1940 	return (WHATIS_WALKRET(w));
1941 }
1942 
1943 static int
1944 whatis_walk_vmem(uintptr_t addr, const vmem_t *vmem, whatis_info_t *wi)
1945 {
1946 	mdb_whatis_t *w = wi->wi_w;
1947 	const char *nm = vmem->vm_name;
1948 	wi->wi_vmem = vmem;
1949 
1950 	if (mdb_whatis_flags(w) & WHATIS_VERBOSE)
1951 		mdb_printf("Searching vmem arena %s...\n", nm);
1952 
1953 	if (mdb_pwalk("vmem_seg",
1954 	    (mdb_walk_cb_t)whatis_walk_seg, wi, addr) == -1) {
1955 		mdb_warn("can't walk vmem seg for %p", addr);
1956 		return (WALK_NEXT);
1957 	}
1958 
1959 	return (WHATIS_WALKRET(w));
1960 }
1961 
1962 /*ARGSUSED*/
1963 static int
1964 whatis_walk_slab(uintptr_t saddr, const umem_slab_t *sp, whatis_info_t *wi)
1965 {
1966 	mdb_whatis_t *w = wi->wi_w;
1967 
1968 	/* It must overlap with the slab data, or it's not interesting */
1969 	if (mdb_whatis_overlaps(w,
1970 	    (uintptr_t)sp->slab_base, wi->wi_slab_size)) {
1971 		wi->wi_slab_found++;
1972 		return (WALK_DONE);
1973 	}
1974 	return (WALK_NEXT);
1975 }
1976 
1977 static int
1978 whatis_walk_cache(uintptr_t addr, const umem_cache_t *c, whatis_info_t *wi)
1979 {
1980 	mdb_whatis_t *w = wi->wi_w;
1981 	char *walk, *freewalk;
1982 	mdb_walk_cb_t func;
1983 	int do_bufctl;
1984 
1985 	/* Override the '-b' flag as necessary */
1986 	if (!(c->cache_flags & UMF_HASH))
1987 		do_bufctl = FALSE;	/* no bufctls to walk */
1988 	else if (c->cache_flags & UMF_AUDIT)
1989 		do_bufctl = TRUE;	/* we always want debugging info */
1990 	else
1991 		do_bufctl = ((mdb_whatis_flags(w) & WHATIS_BUFCTL) != 0);
1992 
1993 	if (do_bufctl) {
1994 		walk = "bufctl";
1995 		freewalk = "freectl";
1996 		func = (mdb_walk_cb_t)whatis_walk_bufctl;
1997 	} else {
1998 		walk = "umem";
1999 		freewalk = "freemem";
2000 		func = (mdb_walk_cb_t)whatis_walk_umem;
2001 	}
2002 
2003 	wi->wi_cache = c;
2004 
2005 	if (mdb_whatis_flags(w) & WHATIS_VERBOSE)
2006 		mdb_printf("Searching %s...\n", c->cache_name);
2007 
2008 	/*
2009 	 * If more then two buffers live on each slab, figure out if we're
2010 	 * interested in anything in any slab before doing the more expensive
2011 	 * umem/freemem (bufctl/freectl) walkers.
2012 	 */
2013 	wi->wi_slab_size = c->cache_slabsize - c->cache_maxcolor;
2014 	if (!(c->cache_flags & UMF_HASH))
2015 		wi->wi_slab_size -= sizeof (umem_slab_t);
2016 
2017 	if ((wi->wi_slab_size / c->cache_chunksize) > 2) {
2018 		wi->wi_slab_found = 0;
2019 		if (mdb_pwalk("umem_slab", (mdb_walk_cb_t)whatis_walk_slab, wi,
2020 		    addr) == -1) {
2021 			mdb_warn("can't find umem_slab walker");
2022 			return (WALK_DONE);
2023 		}
2024 		if (wi->wi_slab_found == 0)
2025 			return (WALK_NEXT);
2026 	}
2027 
2028 	wi->wi_freemem = FALSE;
2029 	if (mdb_pwalk(walk, func, wi, addr) == -1) {
2030 		mdb_warn("can't find %s walker", walk);
2031 		return (WALK_DONE);
2032 	}
2033 
2034 	if (mdb_whatis_done(w))
2035 		return (WALK_DONE);
2036 
2037 	/*
2038 	 * We have searched for allocated memory; now search for freed memory.
2039 	 */
2040 	if (mdb_whatis_flags(w) & WHATIS_VERBOSE)
2041 		mdb_printf("Searching %s for free memory...\n", c->cache_name);
2042 
2043 	wi->wi_freemem = TRUE;
2044 
2045 	if (mdb_pwalk(freewalk, func, wi, addr) == -1) {
2046 		mdb_warn("can't find %s walker", freewalk);
2047 		return (WALK_DONE);
2048 	}
2049 
2050 	return (WHATIS_WALKRET(w));
2051 }
2052 
2053 static int
2054 whatis_walk_touch(uintptr_t addr, const umem_cache_t *c, whatis_info_t *wi)
2055 {
2056 	if (c->cache_arena == wi->wi_msb_arena ||
2057 	    (c->cache_cflags & UMC_NOTOUCH))
2058 		return (WALK_NEXT);
2059 
2060 	return (whatis_walk_cache(addr, c, wi));
2061 }
2062 
2063 static int
2064 whatis_walk_metadata(uintptr_t addr, const umem_cache_t *c, whatis_info_t *wi)
2065 {
2066 	if (c->cache_arena != wi->wi_msb_arena)
2067 		return (WALK_NEXT);
2068 
2069 	return (whatis_walk_cache(addr, c, wi));
2070 }
2071 
2072 static int
2073 whatis_walk_notouch(uintptr_t addr, const umem_cache_t *c, whatis_info_t *wi)
2074 {
2075 	if (c->cache_arena == wi->wi_msb_arena ||
2076 	    !(c->cache_cflags & UMC_NOTOUCH))
2077 		return (WALK_NEXT);
2078 
2079 	return (whatis_walk_cache(addr, c, wi));
2080 }
2081 
2082 /*ARGSUSED*/
2083 static int
2084 whatis_run_umem(mdb_whatis_t *w, void *ignored)
2085 {
2086 	whatis_info_t wi;
2087 
2088 	bzero(&wi, sizeof (wi));
2089 	wi.wi_w = w;
2090 
2091 	/* umem's metadata is allocated from the umem_internal_arena */
2092 	if (mdb_readvar(&wi.wi_msb_arena, "umem_internal_arena") == -1)
2093 		mdb_warn("unable to readvar \"umem_internal_arena\"");
2094 
2095 	/*
2096 	 * We process umem caches in the following order:
2097 	 *
2098 	 *	non-UMC_NOTOUCH, non-metadata	(typically the most interesting)
2099 	 *	metadata			(can be huge with UMF_AUDIT)
2100 	 *	UMC_NOTOUCH, non-metadata	(see umem_walk_all())
2101 	 */
2102 	if (mdb_walk("umem_cache", (mdb_walk_cb_t)whatis_walk_touch,
2103 	    &wi) == -1 ||
2104 	    mdb_walk("umem_cache", (mdb_walk_cb_t)whatis_walk_metadata,
2105 	    &wi) == -1 ||
2106 	    mdb_walk("umem_cache", (mdb_walk_cb_t)whatis_walk_notouch,
2107 	    &wi) == -1) {
2108 		mdb_warn("couldn't find umem_cache walker");
2109 		return (1);
2110 	}
2111 	return (0);
2112 }
2113 
2114 /*ARGSUSED*/
2115 static int
2116 whatis_run_vmem(mdb_whatis_t *w, void *ignored)
2117 {
2118 	whatis_info_t wi;
2119 
2120 	bzero(&wi, sizeof (wi));
2121 	wi.wi_w = w;
2122 
2123 	if (mdb_walk("vmem_postfix",
2124 	    (mdb_walk_cb_t)whatis_walk_vmem, &wi) == -1) {
2125 		mdb_warn("couldn't find vmem_postfix walker");
2126 		return (1);
2127 	}
2128 	return (0);
2129 }
2130 
2131 int
2132 umem_init(void)
2133 {
2134 	mdb_walker_t w = {
2135 		"umem_cache", "walk list of umem caches", umem_cache_walk_init,
2136 		umem_cache_walk_step, umem_cache_walk_fini
2137 	};
2138 
2139 	if (mdb_add_walker(&w) == -1) {
2140 		mdb_warn("failed to add umem_cache walker");
2141 		return (-1);
2142 	}
2143 
2144 	if (umem_update_variables() == -1)
2145 		return (-1);
2146 
2147 	/* install a callback so that our variables are always up-to-date */
2148 	(void) mdb_callback_add(MDB_CALLBACK_STCHG, umem_statechange_cb, NULL);
2149 	umem_statechange_cb(NULL);
2150 
2151 	/*
2152 	 * Register our ::whatis callbacks.
2153 	 */
2154 	mdb_whatis_register("umem", whatis_run_umem, NULL,
2155 	    WHATIS_PRIO_ALLOCATOR, WHATIS_REG_NO_ID);
2156 	mdb_whatis_register("vmem", whatis_run_vmem, NULL,
2157 	    WHATIS_PRIO_ALLOCATOR, WHATIS_REG_NO_ID);
2158 
2159 	return (0);
2160 }
2161 
2162 typedef struct umem_log_cpu {
2163 	uintptr_t umc_low;
2164 	uintptr_t umc_high;
2165 } umem_log_cpu_t;
2166 
2167 int
2168 umem_log_walk(uintptr_t addr, const umem_bufctl_audit_t *b, umem_log_cpu_t *umc)
2169 {
2170 	int i;
2171 
2172 	for (i = 0; i < umem_max_ncpus; i++) {
2173 		if (addr >= umc[i].umc_low && addr < umc[i].umc_high)
2174 			break;
2175 	}
2176 
2177 	if (i == umem_max_ncpus)
2178 		mdb_printf("   ");
2179 	else
2180 		mdb_printf("%3d", i);
2181 
2182 	mdb_printf(" %0?p %0?p %16llx %0?p\n", addr, b->bc_addr,
2183 	    b->bc_timestamp, b->bc_thread);
2184 
2185 	return (WALK_NEXT);
2186 }
2187 
2188 /*ARGSUSED*/
2189 int
2190 umem_log(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
2191 {
2192 	umem_log_header_t lh;
2193 	umem_cpu_log_header_t clh;
2194 	uintptr_t lhp, clhp;
2195 	umem_log_cpu_t *umc;
2196 	int i;
2197 
2198 	if (umem_readvar(&lhp, "umem_transaction_log") == -1) {
2199 		mdb_warn("failed to read 'umem_transaction_log'");
2200 		return (DCMD_ERR);
2201 	}
2202 
2203 	if (lhp == NULL) {
2204 		mdb_warn("no umem transaction log\n");
2205 		return (DCMD_ERR);
2206 	}
2207 
2208 	if (mdb_vread(&lh, sizeof (umem_log_header_t), lhp) == -1) {
2209 		mdb_warn("failed to read log header at %p", lhp);
2210 		return (DCMD_ERR);
2211 	}
2212 
2213 	clhp = lhp + ((uintptr_t)&lh.lh_cpu[0] - (uintptr_t)&lh);
2214 
2215 	umc = mdb_zalloc(sizeof (umem_log_cpu_t) * umem_max_ncpus,
2216 	    UM_SLEEP | UM_GC);
2217 
2218 	for (i = 0; i < umem_max_ncpus; i++) {
2219 		if (mdb_vread(&clh, sizeof (clh), clhp) == -1) {
2220 			mdb_warn("cannot read cpu %d's log header at %p",
2221 			    i, clhp);
2222 			return (DCMD_ERR);
2223 		}
2224 
2225 		umc[i].umc_low = clh.clh_chunk * lh.lh_chunksize +
2226 		    (uintptr_t)lh.lh_base;
2227 		umc[i].umc_high = (uintptr_t)clh.clh_current;
2228 
2229 		clhp += sizeof (umem_cpu_log_header_t);
2230 	}
2231 
2232 	if (DCMD_HDRSPEC(flags)) {
2233 		mdb_printf("%3s %-?s %-?s %16s %-?s\n", "CPU", "ADDR",
2234 		    "BUFADDR", "TIMESTAMP", "THREAD");
2235 	}
2236 
2237 	/*
2238 	 * If we have been passed an address, we'll just print out that
2239 	 * log entry.
2240 	 */
2241 	if (flags & DCMD_ADDRSPEC) {
2242 		umem_bufctl_audit_t *bp;
2243 		UMEM_LOCAL_BUFCTL_AUDIT(&bp);
2244 
2245 		if (mdb_vread(bp, UMEM_BUFCTL_AUDIT_SIZE, addr) == -1) {
2246 			mdb_warn("failed to read bufctl at %p", addr);
2247 			return (DCMD_ERR);
2248 		}
2249 
2250 		(void) umem_log_walk(addr, bp, umc);
2251 
2252 		return (DCMD_OK);
2253 	}
2254 
2255 	if (mdb_walk("umem_log", (mdb_walk_cb_t)umem_log_walk, umc) == -1) {
2256 		mdb_warn("can't find umem log walker");
2257 		return (DCMD_ERR);
2258 	}
2259 
2260 	return (DCMD_OK);
2261 }
2262 
2263 typedef struct bufctl_history_cb {
2264 	int		bhc_flags;
2265 	int		bhc_argc;
2266 	const mdb_arg_t	*bhc_argv;
2267 	int		bhc_ret;
2268 } bufctl_history_cb_t;
2269 
2270 /*ARGSUSED*/
2271 static int
2272 bufctl_history_callback(uintptr_t addr, const void *ign, void *arg)
2273 {
2274 	bufctl_history_cb_t *bhc = arg;
2275 
2276 	bhc->bhc_ret =
2277 	    bufctl(addr, bhc->bhc_flags, bhc->bhc_argc, bhc->bhc_argv);
2278 
2279 	bhc->bhc_flags &= ~DCMD_LOOPFIRST;
2280 
2281 	return ((bhc->bhc_ret == DCMD_OK)? WALK_NEXT : WALK_DONE);
2282 }
2283 
2284 void
2285 bufctl_help(void)
2286 {
2287 	mdb_printf("%s\n",
2288 "Display the contents of umem_bufctl_audit_ts, with optional filtering.\n");
2289 	mdb_dec_indent(2);
2290 	mdb_printf("%<b>OPTIONS%</b>\n");
2291 	mdb_inc_indent(2);
2292 	mdb_printf("%s",
2293 "  -v    Display the full content of the bufctl, including its stack trace\n"
2294 "  -h    retrieve the bufctl's transaction history, if available\n"
2295 "  -a addr\n"
2296 "        filter out bufctls not involving the buffer at addr\n"
2297 "  -c caller\n"
2298 "        filter out bufctls without the function/PC in their stack trace\n"
2299 "  -e earliest\n"
2300 "        filter out bufctls timestamped before earliest\n"
2301 "  -l latest\n"
2302 "        filter out bufctls timestamped after latest\n"
2303 "  -t thread\n"
2304 "        filter out bufctls not involving thread\n");
2305 }
2306 
2307 int
2308 bufctl(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
2309 {
2310 	uint_t verbose = FALSE;
2311 	uint_t history = FALSE;
2312 	uint_t in_history = FALSE;
2313 	uintptr_t caller = NULL, thread = NULL;
2314 	uintptr_t laddr, haddr, baddr = NULL;
2315 	hrtime_t earliest = 0, latest = 0;
2316 	int i, depth;
2317 	char c[MDB_SYM_NAMLEN];
2318 	GElf_Sym sym;
2319 	umem_bufctl_audit_t *bcp;
2320 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
2321 
2322 	if (mdb_getopts(argc, argv,
2323 	    'v', MDB_OPT_SETBITS, TRUE, &verbose,
2324 	    'h', MDB_OPT_SETBITS, TRUE, &history,
2325 	    'H', MDB_OPT_SETBITS, TRUE, &in_history,		/* internal */
2326 	    'c', MDB_OPT_UINTPTR, &caller,
2327 	    't', MDB_OPT_UINTPTR, &thread,
2328 	    'e', MDB_OPT_UINT64, &earliest,
2329 	    'l', MDB_OPT_UINT64, &latest,
2330 	    'a', MDB_OPT_UINTPTR, &baddr, NULL) != argc)
2331 		return (DCMD_USAGE);
2332 
2333 	if (!(flags & DCMD_ADDRSPEC))
2334 		return (DCMD_USAGE);
2335 
2336 	if (in_history && !history)
2337 		return (DCMD_USAGE);
2338 
2339 	if (history && !in_history) {
2340 		mdb_arg_t *nargv = mdb_zalloc(sizeof (*nargv) * (argc + 1),
2341 		    UM_SLEEP | UM_GC);
2342 		bufctl_history_cb_t bhc;
2343 
2344 		nargv[0].a_type = MDB_TYPE_STRING;
2345 		nargv[0].a_un.a_str = "-H";		/* prevent recursion */
2346 
2347 		for (i = 0; i < argc; i++)
2348 			nargv[i + 1] = argv[i];
2349 
2350 		/*
2351 		 * When in history mode, we treat each element as if it
2352 		 * were in a seperate loop, so that the headers group
2353 		 * bufctls with similar histories.
2354 		 */
2355 		bhc.bhc_flags = flags | DCMD_LOOP | DCMD_LOOPFIRST;
2356 		bhc.bhc_argc = argc + 1;
2357 		bhc.bhc_argv = nargv;
2358 		bhc.bhc_ret = DCMD_OK;
2359 
2360 		if (mdb_pwalk("bufctl_history", bufctl_history_callback, &bhc,
2361 		    addr) == -1) {
2362 			mdb_warn("unable to walk bufctl_history");
2363 			return (DCMD_ERR);
2364 		}
2365 
2366 		if (bhc.bhc_ret == DCMD_OK && !(flags & DCMD_PIPE_OUT))
2367 			mdb_printf("\n");
2368 
2369 		return (bhc.bhc_ret);
2370 	}
2371 
2372 	if (DCMD_HDRSPEC(flags) && !(flags & DCMD_PIPE_OUT)) {
2373 		if (verbose) {
2374 			mdb_printf("%16s %16s %16s %16s\n"
2375 			    "%<u>%16s %16s %16s %16s%</u>\n",
2376 			    "ADDR", "BUFADDR", "TIMESTAMP", "THREAD",
2377 			    "", "CACHE", "LASTLOG", "CONTENTS");
2378 		} else {
2379 			mdb_printf("%<u>%-?s %-?s %-12s %5s %s%</u>\n",
2380 			    "ADDR", "BUFADDR", "TIMESTAMP", "THRD", "CALLER");
2381 		}
2382 	}
2383 
2384 	if (mdb_vread(bcp, UMEM_BUFCTL_AUDIT_SIZE, addr) == -1) {
2385 		mdb_warn("couldn't read bufctl at %p", addr);
2386 		return (DCMD_ERR);
2387 	}
2388 
2389 	/*
2390 	 * Guard against bogus bc_depth in case the bufctl is corrupt or
2391 	 * the address does not really refer to a bufctl.
2392 	 */
2393 	depth = MIN(bcp->bc_depth, umem_stack_depth);
2394 
2395 	if (caller != NULL) {
2396 		laddr = caller;
2397 		haddr = caller + sizeof (caller);
2398 
2399 		if (mdb_lookup_by_addr(caller, MDB_SYM_FUZZY, c, sizeof (c),
2400 		    &sym) != -1 && caller == (uintptr_t)sym.st_value) {
2401 			/*
2402 			 * We were provided an exact symbol value; any
2403 			 * address in the function is valid.
2404 			 */
2405 			laddr = (uintptr_t)sym.st_value;
2406 			haddr = (uintptr_t)sym.st_value + sym.st_size;
2407 		}
2408 
2409 		for (i = 0; i < depth; i++)
2410 			if (bcp->bc_stack[i] >= laddr &&
2411 			    bcp->bc_stack[i] < haddr)
2412 				break;
2413 
2414 		if (i == depth)
2415 			return (DCMD_OK);
2416 	}
2417 
2418 	if (thread != NULL && (uintptr_t)bcp->bc_thread != thread)
2419 		return (DCMD_OK);
2420 
2421 	if (earliest != 0 && bcp->bc_timestamp < earliest)
2422 		return (DCMD_OK);
2423 
2424 	if (latest != 0 && bcp->bc_timestamp > latest)
2425 		return (DCMD_OK);
2426 
2427 	if (baddr != 0 && (uintptr_t)bcp->bc_addr != baddr)
2428 		return (DCMD_OK);
2429 
2430 	if (flags & DCMD_PIPE_OUT) {
2431 		mdb_printf("%#r\n", addr);
2432 		return (DCMD_OK);
2433 	}
2434 
2435 	if (verbose) {
2436 		mdb_printf(
2437 		    "%<b>%16p%</b> %16p %16llx %16d\n"
2438 		    "%16s %16p %16p %16p\n",
2439 		    addr, bcp->bc_addr, bcp->bc_timestamp, bcp->bc_thread,
2440 		    "", bcp->bc_cache, bcp->bc_lastlog, bcp->bc_contents);
2441 
2442 		mdb_inc_indent(17);
2443 		for (i = 0; i < depth; i++)
2444 			mdb_printf("%a\n", bcp->bc_stack[i]);
2445 		mdb_dec_indent(17);
2446 		mdb_printf("\n");
2447 	} else {
2448 		mdb_printf("%0?p %0?p %12llx %5d", addr, bcp->bc_addr,
2449 		    bcp->bc_timestamp, bcp->bc_thread);
2450 
2451 		for (i = 0; i < depth; i++) {
2452 			if (mdb_lookup_by_addr(bcp->bc_stack[i],
2453 			    MDB_SYM_FUZZY, c, sizeof (c), &sym) == -1)
2454 				continue;
2455 			if (is_umem_sym(c, "umem_"))
2456 				continue;
2457 			mdb_printf(" %a\n", bcp->bc_stack[i]);
2458 			break;
2459 		}
2460 
2461 		if (i >= depth)
2462 			mdb_printf("\n");
2463 	}
2464 
2465 	return (DCMD_OK);
2466 }
2467 
2468 /*ARGSUSED*/
2469 int
2470 bufctl_audit(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
2471 {
2472 	mdb_arg_t a;
2473 
2474 	if (!(flags & DCMD_ADDRSPEC))
2475 		return (DCMD_USAGE);
2476 
2477 	if (argc != 0)
2478 		return (DCMD_USAGE);
2479 
2480 	a.a_type = MDB_TYPE_STRING;
2481 	a.a_un.a_str = "-v";
2482 
2483 	return (bufctl(addr, flags, 1, &a));
2484 }
2485 
2486 typedef struct umem_verify {
2487 	uint64_t *umv_buf;		/* buffer to read cache contents into */
2488 	size_t umv_size;		/* number of bytes in umv_buf */
2489 	int umv_corruption;		/* > 0 if corruption found. */
2490 	int umv_besilent;		/* report actual corruption sites */
2491 	struct umem_cache umv_cache;	/* the cache we're operating on */
2492 } umem_verify_t;
2493 
2494 /*
2495  * verify_pattern()
2496  *	verify that buf is filled with the pattern pat.
2497  */
2498 static int64_t
2499 verify_pattern(uint64_t *buf_arg, size_t size, uint64_t pat)
2500 {
2501 	/*LINTED*/
2502 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
2503 	uint64_t *buf;
2504 
2505 	for (buf = buf_arg; buf < bufend; buf++)
2506 		if (*buf != pat)
2507 			return ((uintptr_t)buf - (uintptr_t)buf_arg);
2508 	return (-1);
2509 }
2510 
2511 /*
2512  * verify_buftag()
2513  *	verify that btp->bt_bxstat == (bcp ^ pat)
2514  */
2515 static int
2516 verify_buftag(umem_buftag_t *btp, uintptr_t pat)
2517 {
2518 	return (btp->bt_bxstat == ((intptr_t)btp->bt_bufctl ^ pat) ? 0 : -1);
2519 }
2520 
2521 /*
2522  * verify_free()
2523  *	verify the integrity of a free block of memory by checking
2524  *	that it is filled with 0xdeadbeef and that its buftag is sane.
2525  */
2526 /*ARGSUSED1*/
2527 static int
2528 verify_free(uintptr_t addr, const void *data, void *private)
2529 {
2530 	umem_verify_t *umv = (umem_verify_t *)private;
2531 	uint64_t *buf = umv->umv_buf;	/* buf to validate */
2532 	int64_t corrupt;		/* corruption offset */
2533 	umem_buftag_t *buftagp;		/* ptr to buftag */
2534 	umem_cache_t *cp = &umv->umv_cache;
2535 	int besilent = umv->umv_besilent;
2536 
2537 	/*LINTED*/
2538 	buftagp = UMEM_BUFTAG(cp, buf);
2539 
2540 	/*
2541 	 * Read the buffer to check.
2542 	 */
2543 	if (mdb_vread(buf, umv->umv_size, addr) == -1) {
2544 		if (!besilent)
2545 			mdb_warn("couldn't read %p", addr);
2546 		return (WALK_NEXT);
2547 	}
2548 
2549 	if ((corrupt = verify_pattern(buf, cp->cache_verify,
2550 	    UMEM_FREE_PATTERN)) >= 0) {
2551 		if (!besilent)
2552 			mdb_printf("buffer %p (free) seems corrupted, at %p\n",
2553 			    addr, (uintptr_t)addr + corrupt);
2554 		goto corrupt;
2555 	}
2556 
2557 	if ((cp->cache_flags & UMF_HASH) &&
2558 	    buftagp->bt_redzone != UMEM_REDZONE_PATTERN) {
2559 		if (!besilent)
2560 			mdb_printf("buffer %p (free) seems to "
2561 			    "have a corrupt redzone pattern\n", addr);
2562 		goto corrupt;
2563 	}
2564 
2565 	/*
2566 	 * confirm bufctl pointer integrity.
2567 	 */
2568 	if (verify_buftag(buftagp, UMEM_BUFTAG_FREE) == -1) {
2569 		if (!besilent)
2570 			mdb_printf("buffer %p (free) has a corrupt "
2571 			    "buftag\n", addr);
2572 		goto corrupt;
2573 	}
2574 
2575 	return (WALK_NEXT);
2576 corrupt:
2577 	umv->umv_corruption++;
2578 	return (WALK_NEXT);
2579 }
2580 
2581 /*
2582  * verify_alloc()
2583  *	Verify that the buftag of an allocated buffer makes sense with respect
2584  *	to the buffer.
2585  */
2586 /*ARGSUSED1*/
2587 static int
2588 verify_alloc(uintptr_t addr, const void *data, void *private)
2589 {
2590 	umem_verify_t *umv = (umem_verify_t *)private;
2591 	umem_cache_t *cp = &umv->umv_cache;
2592 	uint64_t *buf = umv->umv_buf;	/* buf to validate */
2593 	/*LINTED*/
2594 	umem_buftag_t *buftagp = UMEM_BUFTAG(cp, buf);
2595 	uint32_t *ip = (uint32_t *)buftagp;
2596 	uint8_t *bp = (uint8_t *)buf;
2597 	int looks_ok = 0, size_ok = 1;	/* flags for finding corruption */
2598 	int besilent = umv->umv_besilent;
2599 
2600 	/*
2601 	 * Read the buffer to check.
2602 	 */
2603 	if (mdb_vread(buf, umv->umv_size, addr) == -1) {
2604 		if (!besilent)
2605 			mdb_warn("couldn't read %p", addr);
2606 		return (WALK_NEXT);
2607 	}
2608 
2609 	/*
2610 	 * There are two cases to handle:
2611 	 * 1. If the buf was alloc'd using umem_cache_alloc, it will have
2612 	 *    0xfeedfacefeedface at the end of it
2613 	 * 2. If the buf was alloc'd using umem_alloc, it will have
2614 	 *    0xbb just past the end of the region in use.  At the buftag,
2615 	 *    it will have 0xfeedface (or, if the whole buffer is in use,
2616 	 *    0xfeedface & bb000000 or 0xfeedfacf & 000000bb depending on
2617 	 *    endianness), followed by 32 bits containing the offset of the
2618 	 *    0xbb byte in the buffer.
2619 	 *
2620 	 * Finally, the two 32-bit words that comprise the second half of the
2621 	 * buftag should xor to UMEM_BUFTAG_ALLOC
2622 	 */
2623 
2624 	if (buftagp->bt_redzone == UMEM_REDZONE_PATTERN)
2625 		looks_ok = 1;
2626 	else if (!UMEM_SIZE_VALID(ip[1]))
2627 		size_ok = 0;
2628 	else if (bp[UMEM_SIZE_DECODE(ip[1])] == UMEM_REDZONE_BYTE)
2629 		looks_ok = 1;
2630 	else
2631 		size_ok = 0;
2632 
2633 	if (!size_ok) {
2634 		if (!besilent)
2635 			mdb_printf("buffer %p (allocated) has a corrupt "
2636 			    "redzone size encoding\n", addr);
2637 		goto corrupt;
2638 	}
2639 
2640 	if (!looks_ok) {
2641 		if (!besilent)
2642 			mdb_printf("buffer %p (allocated) has a corrupt "
2643 			    "redzone signature\n", addr);
2644 		goto corrupt;
2645 	}
2646 
2647 	if (verify_buftag(buftagp, UMEM_BUFTAG_ALLOC) == -1) {
2648 		if (!besilent)
2649 			mdb_printf("buffer %p (allocated) has a "
2650 			    "corrupt buftag\n", addr);
2651 		goto corrupt;
2652 	}
2653 
2654 	return (WALK_NEXT);
2655 corrupt:
2656 	umv->umv_corruption++;
2657 	return (WALK_NEXT);
2658 }
2659 
2660 /*ARGSUSED2*/
2661 int
2662 umem_verify(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
2663 {
2664 	if (flags & DCMD_ADDRSPEC) {
2665 		int check_alloc = 0, check_free = 0;
2666 		umem_verify_t umv;
2667 
2668 		if (mdb_vread(&umv.umv_cache, sizeof (umv.umv_cache),
2669 		    addr) == -1) {
2670 			mdb_warn("couldn't read umem_cache %p", addr);
2671 			return (DCMD_ERR);
2672 		}
2673 
2674 		umv.umv_size = umv.umv_cache.cache_buftag +
2675 		    sizeof (umem_buftag_t);
2676 		umv.umv_buf = mdb_alloc(umv.umv_size, UM_SLEEP | UM_GC);
2677 		umv.umv_corruption = 0;
2678 
2679 		if ((umv.umv_cache.cache_flags & UMF_REDZONE)) {
2680 			check_alloc = 1;
2681 			if (umv.umv_cache.cache_flags & UMF_DEADBEEF)
2682 				check_free = 1;
2683 		} else {
2684 			if (!(flags & DCMD_LOOP)) {
2685 				mdb_warn("cache %p (%s) does not have "
2686 				    "redzone checking enabled\n", addr,
2687 				    umv.umv_cache.cache_name);
2688 			}
2689 			return (DCMD_ERR);
2690 		}
2691 
2692 		if (flags & DCMD_LOOP) {
2693 			/*
2694 			 * table mode, don't print out every corrupt buffer
2695 			 */
2696 			umv.umv_besilent = 1;
2697 		} else {
2698 			mdb_printf("Summary for cache '%s'\n",
2699 			    umv.umv_cache.cache_name);
2700 			mdb_inc_indent(2);
2701 			umv.umv_besilent = 0;
2702 		}
2703 
2704 		if (check_alloc)
2705 			(void) mdb_pwalk("umem", verify_alloc, &umv, addr);
2706 		if (check_free)
2707 			(void) mdb_pwalk("freemem", verify_free, &umv, addr);
2708 
2709 		if (flags & DCMD_LOOP) {
2710 			if (umv.umv_corruption == 0) {
2711 				mdb_printf("%-*s %?p clean\n",
2712 				    UMEM_CACHE_NAMELEN,
2713 				    umv.umv_cache.cache_name, addr);
2714 			} else {
2715 				char *s = "";	/* optional s in "buffer[s]" */
2716 				if (umv.umv_corruption > 1)
2717 					s = "s";
2718 
2719 				mdb_printf("%-*s %?p %d corrupt buffer%s\n",
2720 				    UMEM_CACHE_NAMELEN,
2721 				    umv.umv_cache.cache_name, addr,
2722 				    umv.umv_corruption, s);
2723 			}
2724 		} else {
2725 			/*
2726 			 * This is the more verbose mode, when the user has
2727 			 * type addr::umem_verify.  If the cache was clean,
2728 			 * nothing will have yet been printed. So say something.
2729 			 */
2730 			if (umv.umv_corruption == 0)
2731 				mdb_printf("clean\n");
2732 
2733 			mdb_dec_indent(2);
2734 		}
2735 	} else {
2736 		/*
2737 		 * If the user didn't specify a cache to verify, we'll walk all
2738 		 * umem_cache's, specifying ourself as a callback for each...
2739 		 * this is the equivalent of '::walk umem_cache .::umem_verify'
2740 		 */
2741 		mdb_printf("%<u>%-*s %-?s %-20s%</b>\n", UMEM_CACHE_NAMELEN,
2742 		    "Cache Name", "Addr", "Cache Integrity");
2743 		(void) (mdb_walk_dcmd("umem_cache", "umem_verify", 0, NULL));
2744 	}
2745 
2746 	return (DCMD_OK);
2747 }
2748 
2749 typedef struct vmem_node {
2750 	struct vmem_node *vn_next;
2751 	struct vmem_node *vn_parent;
2752 	struct vmem_node *vn_sibling;
2753 	struct vmem_node *vn_children;
2754 	uintptr_t vn_addr;
2755 	int vn_marked;
2756 	vmem_t vn_vmem;
2757 } vmem_node_t;
2758 
2759 typedef struct vmem_walk {
2760 	vmem_node_t *vw_root;
2761 	vmem_node_t *vw_current;
2762 } vmem_walk_t;
2763 
2764 int
2765 vmem_walk_init(mdb_walk_state_t *wsp)
2766 {
2767 	uintptr_t vaddr, paddr;
2768 	vmem_node_t *head = NULL, *root = NULL, *current = NULL, *parent, *vp;
2769 	vmem_walk_t *vw;
2770 
2771 	if (umem_readvar(&vaddr, "vmem_list") == -1) {
2772 		mdb_warn("couldn't read 'vmem_list'");
2773 		return (WALK_ERR);
2774 	}
2775 
2776 	while (vaddr != NULL) {
2777 		vp = mdb_zalloc(sizeof (vmem_node_t), UM_SLEEP);
2778 		vp->vn_addr = vaddr;
2779 		vp->vn_next = head;
2780 		head = vp;
2781 
2782 		if (vaddr == wsp->walk_addr)
2783 			current = vp;
2784 
2785 		if (mdb_vread(&vp->vn_vmem, sizeof (vmem_t), vaddr) == -1) {
2786 			mdb_warn("couldn't read vmem_t at %p", vaddr);
2787 			goto err;
2788 		}
2789 
2790 		vaddr = (uintptr_t)vp->vn_vmem.vm_next;
2791 	}
2792 
2793 	for (vp = head; vp != NULL; vp = vp->vn_next) {
2794 
2795 		if ((paddr = (uintptr_t)vp->vn_vmem.vm_source) == NULL) {
2796 			vp->vn_sibling = root;
2797 			root = vp;
2798 			continue;
2799 		}
2800 
2801 		for (parent = head; parent != NULL; parent = parent->vn_next) {
2802 			if (parent->vn_addr != paddr)
2803 				continue;
2804 			vp->vn_sibling = parent->vn_children;
2805 			parent->vn_children = vp;
2806 			vp->vn_parent = parent;
2807 			break;
2808 		}
2809 
2810 		if (parent == NULL) {
2811 			mdb_warn("couldn't find %p's parent (%p)\n",
2812 			    vp->vn_addr, paddr);
2813 			goto err;
2814 		}
2815 	}
2816 
2817 	vw = mdb_zalloc(sizeof (vmem_walk_t), UM_SLEEP);
2818 	vw->vw_root = root;
2819 
2820 	if (current != NULL)
2821 		vw->vw_current = current;
2822 	else
2823 		vw->vw_current = root;
2824 
2825 	wsp->walk_data = vw;
2826 	return (WALK_NEXT);
2827 err:
2828 	for (vp = head; head != NULL; vp = head) {
2829 		head = vp->vn_next;
2830 		mdb_free(vp, sizeof (vmem_node_t));
2831 	}
2832 
2833 	return (WALK_ERR);
2834 }
2835 
2836 int
2837 vmem_walk_step(mdb_walk_state_t *wsp)
2838 {
2839 	vmem_walk_t *vw = wsp->walk_data;
2840 	vmem_node_t *vp;
2841 	int rval;
2842 
2843 	if ((vp = vw->vw_current) == NULL)
2844 		return (WALK_DONE);
2845 
2846 	rval = wsp->walk_callback(vp->vn_addr, &vp->vn_vmem, wsp->walk_cbdata);
2847 
2848 	if (vp->vn_children != NULL) {
2849 		vw->vw_current = vp->vn_children;
2850 		return (rval);
2851 	}
2852 
2853 	do {
2854 		vw->vw_current = vp->vn_sibling;
2855 		vp = vp->vn_parent;
2856 	} while (vw->vw_current == NULL && vp != NULL);
2857 
2858 	return (rval);
2859 }
2860 
2861 /*
2862  * The "vmem_postfix" walk walks the vmem arenas in post-fix order; all
2863  * children are visited before their parent.  We perform the postfix walk
2864  * iteratively (rather than recursively) to allow mdb to regain control
2865  * after each callback.
2866  */
2867 int
2868 vmem_postfix_walk_step(mdb_walk_state_t *wsp)
2869 {
2870 	vmem_walk_t *vw = wsp->walk_data;
2871 	vmem_node_t *vp = vw->vw_current;
2872 	int rval;
2873 
2874 	/*
2875 	 * If this node is marked, then we know that we have already visited
2876 	 * all of its children.  If the node has any siblings, they need to
2877 	 * be visited next; otherwise, we need to visit the parent.  Note
2878 	 * that vp->vn_marked will only be zero on the first invocation of
2879 	 * the step function.
2880 	 */
2881 	if (vp->vn_marked) {
2882 		if (vp->vn_sibling != NULL)
2883 			vp = vp->vn_sibling;
2884 		else if (vp->vn_parent != NULL)
2885 			vp = vp->vn_parent;
2886 		else {
2887 			/*
2888 			 * We have neither a parent, nor a sibling, and we
2889 			 * have already been visited; we're done.
2890 			 */
2891 			return (WALK_DONE);
2892 		}
2893 	}
2894 
2895 	/*
2896 	 * Before we visit this node, visit its children.
2897 	 */
2898 	while (vp->vn_children != NULL && !vp->vn_children->vn_marked)
2899 		vp = vp->vn_children;
2900 
2901 	vp->vn_marked = 1;
2902 	vw->vw_current = vp;
2903 	rval = wsp->walk_callback(vp->vn_addr, &vp->vn_vmem, wsp->walk_cbdata);
2904 
2905 	return (rval);
2906 }
2907 
2908 void
2909 vmem_walk_fini(mdb_walk_state_t *wsp)
2910 {
2911 	vmem_walk_t *vw = wsp->walk_data;
2912 	vmem_node_t *root = vw->vw_root;
2913 	int done;
2914 
2915 	if (root == NULL)
2916 		return;
2917 
2918 	if ((vw->vw_root = root->vn_children) != NULL)
2919 		vmem_walk_fini(wsp);
2920 
2921 	vw->vw_root = root->vn_sibling;
2922 	done = (root->vn_sibling == NULL && root->vn_parent == NULL);
2923 	mdb_free(root, sizeof (vmem_node_t));
2924 
2925 	if (done) {
2926 		mdb_free(vw, sizeof (vmem_walk_t));
2927 	} else {
2928 		vmem_walk_fini(wsp);
2929 	}
2930 }
2931 
2932 typedef struct vmem_seg_walk {
2933 	uint8_t vsw_type;
2934 	uintptr_t vsw_start;
2935 	uintptr_t vsw_current;
2936 } vmem_seg_walk_t;
2937 
2938 /*ARGSUSED*/
2939 int
2940 vmem_seg_walk_common_init(mdb_walk_state_t *wsp, uint8_t type, char *name)
2941 {
2942 	vmem_seg_walk_t *vsw;
2943 
2944 	if (wsp->walk_addr == NULL) {
2945 		mdb_warn("vmem_%s does not support global walks\n", name);
2946 		return (WALK_ERR);
2947 	}
2948 
2949 	wsp->walk_data = vsw = mdb_alloc(sizeof (vmem_seg_walk_t), UM_SLEEP);
2950 
2951 	vsw->vsw_type = type;
2952 	vsw->vsw_start = wsp->walk_addr + OFFSETOF(vmem_t, vm_seg0);
2953 	vsw->vsw_current = vsw->vsw_start;
2954 
2955 	return (WALK_NEXT);
2956 }
2957 
2958 /*
2959  * vmem segments can't have type 0 (this should be added to vmem_impl.h).
2960  */
2961 #define	VMEM_NONE	0
2962 
2963 int
2964 vmem_alloc_walk_init(mdb_walk_state_t *wsp)
2965 {
2966 	return (vmem_seg_walk_common_init(wsp, VMEM_ALLOC, "alloc"));
2967 }
2968 
2969 int
2970 vmem_free_walk_init(mdb_walk_state_t *wsp)
2971 {
2972 	return (vmem_seg_walk_common_init(wsp, VMEM_FREE, "free"));
2973 }
2974 
2975 int
2976 vmem_span_walk_init(mdb_walk_state_t *wsp)
2977 {
2978 	return (vmem_seg_walk_common_init(wsp, VMEM_SPAN, "span"));
2979 }
2980 
2981 int
2982 vmem_seg_walk_init(mdb_walk_state_t *wsp)
2983 {
2984 	return (vmem_seg_walk_common_init(wsp, VMEM_NONE, "seg"));
2985 }
2986 
2987 int
2988 vmem_seg_walk_step(mdb_walk_state_t *wsp)
2989 {
2990 	vmem_seg_t seg;
2991 	vmem_seg_walk_t *vsw = wsp->walk_data;
2992 	uintptr_t addr = vsw->vsw_current;
2993 	static size_t seg_size = 0;
2994 	int rval;
2995 
2996 	if (!seg_size) {
2997 		if (umem_readvar(&seg_size, "vmem_seg_size") == -1) {
2998 			mdb_warn("failed to read 'vmem_seg_size'");
2999 			seg_size = sizeof (vmem_seg_t);
3000 		}
3001 	}
3002 
3003 	if (seg_size < sizeof (seg))
3004 		bzero((caddr_t)&seg + seg_size, sizeof (seg) - seg_size);
3005 
3006 	if (mdb_vread(&seg, seg_size, addr) == -1) {
3007 		mdb_warn("couldn't read vmem_seg at %p", addr);
3008 		return (WALK_ERR);
3009 	}
3010 
3011 	vsw->vsw_current = (uintptr_t)seg.vs_anext;
3012 	if (vsw->vsw_type != VMEM_NONE && seg.vs_type != vsw->vsw_type) {
3013 		rval = WALK_NEXT;
3014 	} else {
3015 		rval = wsp->walk_callback(addr, &seg, wsp->walk_cbdata);
3016 	}
3017 
3018 	if (vsw->vsw_current == vsw->vsw_start)
3019 		return (WALK_DONE);
3020 
3021 	return (rval);
3022 }
3023 
3024 void
3025 vmem_seg_walk_fini(mdb_walk_state_t *wsp)
3026 {
3027 	vmem_seg_walk_t *vsw = wsp->walk_data;
3028 
3029 	mdb_free(vsw, sizeof (vmem_seg_walk_t));
3030 }
3031 
3032 #define	VMEM_NAMEWIDTH	22
3033 
3034 int
3035 vmem(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
3036 {
3037 	vmem_t v, parent;
3038 	uintptr_t paddr;
3039 	int ident = 0;
3040 	char c[VMEM_NAMEWIDTH];
3041 
3042 	if (!(flags & DCMD_ADDRSPEC)) {
3043 		if (mdb_walk_dcmd("vmem", "vmem", argc, argv) == -1) {
3044 			mdb_warn("can't walk vmem");
3045 			return (DCMD_ERR);
3046 		}
3047 		return (DCMD_OK);
3048 	}
3049 
3050 	if (DCMD_HDRSPEC(flags))
3051 		mdb_printf("%-?s %-*s %10s %12s %9s %5s\n",
3052 		    "ADDR", VMEM_NAMEWIDTH, "NAME", "INUSE",
3053 		    "TOTAL", "SUCCEED", "FAIL");
3054 
3055 	if (mdb_vread(&v, sizeof (v), addr) == -1) {
3056 		mdb_warn("couldn't read vmem at %p", addr);
3057 		return (DCMD_ERR);
3058 	}
3059 
3060 	for (paddr = (uintptr_t)v.vm_source; paddr != NULL; ident += 2) {
3061 		if (mdb_vread(&parent, sizeof (parent), paddr) == -1) {
3062 			mdb_warn("couldn't trace %p's ancestry", addr);
3063 			ident = 0;
3064 			break;
3065 		}
3066 		paddr = (uintptr_t)parent.vm_source;
3067 	}
3068 
3069 	(void) mdb_snprintf(c, VMEM_NAMEWIDTH, "%*s%s", ident, "", v.vm_name);
3070 
3071 	mdb_printf("%0?p %-*s %10llu %12llu %9llu %5llu\n",
3072 	    addr, VMEM_NAMEWIDTH, c,
3073 	    v.vm_kstat.vk_mem_inuse, v.vm_kstat.vk_mem_total,
3074 	    v.vm_kstat.vk_alloc, v.vm_kstat.vk_fail);
3075 
3076 	return (DCMD_OK);
3077 }
3078 
3079 void
3080 vmem_seg_help(void)
3081 {
3082 	mdb_printf("%s\n",
3083 "Display the contents of vmem_seg_ts, with optional filtering.\n"
3084 "\n"
3085 "A vmem_seg_t represents a range of addresses (or arbitrary numbers),\n"
3086 "representing a single chunk of data.  Only ALLOC segments have debugging\n"
3087 "information.\n");
3088 	mdb_dec_indent(2);
3089 	mdb_printf("%<b>OPTIONS%</b>\n");
3090 	mdb_inc_indent(2);
3091 	mdb_printf("%s",
3092 "  -v    Display the full content of the vmem_seg, including its stack trace\n"
3093 "  -s    report the size of the segment, instead of the end address\n"
3094 "  -c caller\n"
3095 "        filter out segments without the function/PC in their stack trace\n"
3096 "  -e earliest\n"
3097 "        filter out segments timestamped before earliest\n"
3098 "  -l latest\n"
3099 "        filter out segments timestamped after latest\n"
3100 "  -m minsize\n"
3101 "        filer out segments smaller than minsize\n"
3102 "  -M maxsize\n"
3103 "        filer out segments larger than maxsize\n"
3104 "  -t thread\n"
3105 "        filter out segments not involving thread\n"
3106 "  -T type\n"
3107 "        filter out segments not of type 'type'\n"
3108 "        type is one of: ALLOC/FREE/SPAN/ROTOR/WALKER\n");
3109 }
3110 
3111 
3112 /*ARGSUSED*/
3113 int
3114 vmem_seg(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
3115 {
3116 	vmem_seg_t vs;
3117 	uintptr_t *stk = vs.vs_stack;
3118 	uintptr_t sz;
3119 	uint8_t t;
3120 	const char *type = NULL;
3121 	GElf_Sym sym;
3122 	char c[MDB_SYM_NAMLEN];
3123 	int no_debug;
3124 	int i;
3125 	int depth;
3126 	uintptr_t laddr, haddr;
3127 
3128 	uintptr_t caller = NULL, thread = NULL;
3129 	uintptr_t minsize = 0, maxsize = 0;
3130 
3131 	hrtime_t earliest = 0, latest = 0;
3132 
3133 	uint_t size = 0;
3134 	uint_t verbose = 0;
3135 
3136 	if (!(flags & DCMD_ADDRSPEC))
3137 		return (DCMD_USAGE);
3138 
3139 	if (mdb_getopts(argc, argv,
3140 	    'c', MDB_OPT_UINTPTR, &caller,
3141 	    'e', MDB_OPT_UINT64, &earliest,
3142 	    'l', MDB_OPT_UINT64, &latest,
3143 	    's', MDB_OPT_SETBITS, TRUE, &size,
3144 	    'm', MDB_OPT_UINTPTR, &minsize,
3145 	    'M', MDB_OPT_UINTPTR, &maxsize,
3146 	    't', MDB_OPT_UINTPTR, &thread,
3147 	    'T', MDB_OPT_STR, &type,
3148 	    'v', MDB_OPT_SETBITS, TRUE, &verbose,
3149 	    NULL) != argc)
3150 		return (DCMD_USAGE);
3151 
3152 	if (DCMD_HDRSPEC(flags) && !(flags & DCMD_PIPE_OUT)) {
3153 		if (verbose) {
3154 			mdb_printf("%16s %4s %16s %16s %16s\n"
3155 			    "%<u>%16s %4s %16s %16s %16s%</u>\n",
3156 			    "ADDR", "TYPE", "START", "END", "SIZE",
3157 			    "", "", "THREAD", "TIMESTAMP", "");
3158 		} else {
3159 			mdb_printf("%?s %4s %?s %?s %s\n", "ADDR", "TYPE",
3160 			    "START", size? "SIZE" : "END", "WHO");
3161 		}
3162 	}
3163 
3164 	if (mdb_vread(&vs, sizeof (vs), addr) == -1) {
3165 		mdb_warn("couldn't read vmem_seg at %p", addr);
3166 		return (DCMD_ERR);
3167 	}
3168 
3169 	if (type != NULL) {
3170 		if (strcmp(type, "ALLC") == 0 || strcmp(type, "ALLOC") == 0)
3171 			t = VMEM_ALLOC;
3172 		else if (strcmp(type, "FREE") == 0)
3173 			t = VMEM_FREE;
3174 		else if (strcmp(type, "SPAN") == 0)
3175 			t = VMEM_SPAN;
3176 		else if (strcmp(type, "ROTR") == 0 ||
3177 		    strcmp(type, "ROTOR") == 0)
3178 			t = VMEM_ROTOR;
3179 		else if (strcmp(type, "WLKR") == 0 ||
3180 		    strcmp(type, "WALKER") == 0)
3181 			t = VMEM_WALKER;
3182 		else {
3183 			mdb_warn("\"%s\" is not a recognized vmem_seg type\n",
3184 			    type);
3185 			return (DCMD_ERR);
3186 		}
3187 
3188 		if (vs.vs_type != t)
3189 			return (DCMD_OK);
3190 	}
3191 
3192 	sz = vs.vs_end - vs.vs_start;
3193 
3194 	if (minsize != 0 && sz < minsize)
3195 		return (DCMD_OK);
3196 
3197 	if (maxsize != 0 && sz > maxsize)
3198 		return (DCMD_OK);
3199 
3200 	t = vs.vs_type;
3201 	depth = vs.vs_depth;
3202 
3203 	/*
3204 	 * debug info, when present, is only accurate for VMEM_ALLOC segments
3205 	 */
3206 	no_debug = (t != VMEM_ALLOC) ||
3207 	    (depth == 0 || depth > VMEM_STACK_DEPTH);
3208 
3209 	if (no_debug) {
3210 		if (caller != NULL || thread != NULL || earliest != 0 ||
3211 		    latest != 0)
3212 			return (DCMD_OK);		/* not enough info */
3213 	} else {
3214 		if (caller != NULL) {
3215 			laddr = caller;
3216 			haddr = caller + sizeof (caller);
3217 
3218 			if (mdb_lookup_by_addr(caller, MDB_SYM_FUZZY, c,
3219 			    sizeof (c), &sym) != -1 &&
3220 			    caller == (uintptr_t)sym.st_value) {
3221 				/*
3222 				 * We were provided an exact symbol value; any
3223 				 * address in the function is valid.
3224 				 */
3225 				laddr = (uintptr_t)sym.st_value;
3226 				haddr = (uintptr_t)sym.st_value + sym.st_size;
3227 			}
3228 
3229 			for (i = 0; i < depth; i++)
3230 				if (vs.vs_stack[i] >= laddr &&
3231 				    vs.vs_stack[i] < haddr)
3232 					break;
3233 
3234 			if (i == depth)
3235 				return (DCMD_OK);
3236 		}
3237 
3238 		if (thread != NULL && (uintptr_t)vs.vs_thread != thread)
3239 			return (DCMD_OK);
3240 
3241 		if (earliest != 0 && vs.vs_timestamp < earliest)
3242 			return (DCMD_OK);
3243 
3244 		if (latest != 0 && vs.vs_timestamp > latest)
3245 			return (DCMD_OK);
3246 	}
3247 
3248 	type = (t == VMEM_ALLOC ? "ALLC" :
3249 	    t == VMEM_FREE ? "FREE" :
3250 	    t == VMEM_SPAN ? "SPAN" :
3251 	    t == VMEM_ROTOR ? "ROTR" :
3252 	    t == VMEM_WALKER ? "WLKR" :
3253 	    "????");
3254 
3255 	if (flags & DCMD_PIPE_OUT) {
3256 		mdb_printf("%#r\n", addr);
3257 		return (DCMD_OK);
3258 	}
3259 
3260 	if (verbose) {
3261 		mdb_printf("%<b>%16p%</b> %4s %16p %16p %16d\n",
3262 		    addr, type, vs.vs_start, vs.vs_end, sz);
3263 
3264 		if (no_debug)
3265 			return (DCMD_OK);
3266 
3267 		mdb_printf("%16s %4s %16d %16llx\n",
3268 		    "", "", vs.vs_thread, vs.vs_timestamp);
3269 
3270 		mdb_inc_indent(17);
3271 		for (i = 0; i < depth; i++) {
3272 			mdb_printf("%a\n", stk[i]);
3273 		}
3274 		mdb_dec_indent(17);
3275 		mdb_printf("\n");
3276 	} else {
3277 		mdb_printf("%0?p %4s %0?p %0?p", addr, type,
3278 		    vs.vs_start, size? sz : vs.vs_end);
3279 
3280 		if (no_debug) {
3281 			mdb_printf("\n");
3282 			return (DCMD_OK);
3283 		}
3284 
3285 		for (i = 0; i < depth; i++) {
3286 			if (mdb_lookup_by_addr(stk[i], MDB_SYM_FUZZY,
3287 			    c, sizeof (c), &sym) == -1)
3288 				continue;
3289 			if (is_umem_sym(c, "vmem_"))
3290 				continue;
3291 			break;
3292 		}
3293 		mdb_printf(" %a\n", stk[i]);
3294 	}
3295 	return (DCMD_OK);
3296 }
3297 
3298 /*ARGSUSED*/
3299 static int
3300 showbc(uintptr_t addr, const umem_bufctl_audit_t *bcp, hrtime_t *newest)
3301 {
3302 	char name[UMEM_CACHE_NAMELEN + 1];
3303 	hrtime_t delta;
3304 	int i, depth;
3305 
3306 	if (bcp->bc_timestamp == 0)
3307 		return (WALK_DONE);
3308 
3309 	if (*newest == 0)
3310 		*newest = bcp->bc_timestamp;
3311 
3312 	delta = *newest - bcp->bc_timestamp;
3313 	depth = MIN(bcp->bc_depth, umem_stack_depth);
3314 
3315 	if (mdb_readstr(name, sizeof (name), (uintptr_t)
3316 	    &bcp->bc_cache->cache_name) <= 0)
3317 		(void) mdb_snprintf(name, sizeof (name), "%a", bcp->bc_cache);
3318 
3319 	mdb_printf("\nT-%lld.%09lld  addr=%p  %s\n",
3320 	    delta / NANOSEC, delta % NANOSEC, bcp->bc_addr, name);
3321 
3322 	for (i = 0; i < depth; i++)
3323 		mdb_printf("\t %a\n", bcp->bc_stack[i]);
3324 
3325 	return (WALK_NEXT);
3326 }
3327 
3328 int
3329 umalog(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
3330 {
3331 	const char *logname = "umem_transaction_log";
3332 	hrtime_t newest = 0;
3333 
3334 	if ((flags & DCMD_ADDRSPEC) || argc > 1)
3335 		return (DCMD_USAGE);
3336 
3337 	if (argc > 0) {
3338 		if (argv->a_type != MDB_TYPE_STRING)
3339 			return (DCMD_USAGE);
3340 		if (strcmp(argv->a_un.a_str, "fail") == 0)
3341 			logname = "umem_failure_log";
3342 		else if (strcmp(argv->a_un.a_str, "slab") == 0)
3343 			logname = "umem_slab_log";
3344 		else
3345 			return (DCMD_USAGE);
3346 	}
3347 
3348 	if (umem_readvar(&addr, logname) == -1) {
3349 		mdb_warn("failed to read %s log header pointer");
3350 		return (DCMD_ERR);
3351 	}
3352 
3353 	if (mdb_pwalk("umem_log", (mdb_walk_cb_t)showbc, &newest, addr) == -1) {
3354 		mdb_warn("failed to walk umem log");
3355 		return (DCMD_ERR);
3356 	}
3357 
3358 	return (DCMD_OK);
3359 }
3360 
3361 /*
3362  * As the final lure for die-hard crash(1M) users, we provide ::umausers here.
3363  * The first piece is a structure which we use to accumulate umem_cache_t
3364  * addresses of interest.  The umc_add is used as a callback for the umem_cache
3365  * walker; we either add all caches, or ones named explicitly as arguments.
3366  */
3367 
3368 typedef struct umclist {
3369 	const char *umc_name;			/* Name to match (or NULL) */
3370 	uintptr_t *umc_caches;			/* List of umem_cache_t addrs */
3371 	int umc_nelems;				/* Num entries in umc_caches */
3372 	int umc_size;				/* Size of umc_caches array */
3373 } umclist_t;
3374 
3375 static int
3376 umc_add(uintptr_t addr, const umem_cache_t *cp, umclist_t *umc)
3377 {
3378 	void *p;
3379 	int s;
3380 
3381 	if (umc->umc_name == NULL ||
3382 	    strcmp(cp->cache_name, umc->umc_name) == 0) {
3383 		/*
3384 		 * If we have a match, grow our array (if necessary), and then
3385 		 * add the virtual address of the matching cache to our list.
3386 		 */
3387 		if (umc->umc_nelems >= umc->umc_size) {
3388 			s = umc->umc_size ? umc->umc_size * 2 : 256;
3389 			p = mdb_alloc(sizeof (uintptr_t) * s, UM_SLEEP | UM_GC);
3390 
3391 			bcopy(umc->umc_caches, p,
3392 			    sizeof (uintptr_t) * umc->umc_size);
3393 
3394 			umc->umc_caches = p;
3395 			umc->umc_size = s;
3396 		}
3397 
3398 		umc->umc_caches[umc->umc_nelems++] = addr;
3399 		return (umc->umc_name ? WALK_DONE : WALK_NEXT);
3400 	}
3401 
3402 	return (WALK_NEXT);
3403 }
3404 
3405 /*
3406  * The second piece of ::umausers is a hash table of allocations.  Each
3407  * allocation owner is identified by its stack trace and data_size.  We then
3408  * track the total bytes of all such allocations, and the number of allocations
3409  * to report at the end.  Once we have a list of caches, we walk through the
3410  * allocated bufctls of each, and update our hash table accordingly.
3411  */
3412 
3413 typedef struct umowner {
3414 	struct umowner *umo_head;		/* First hash elt in bucket */
3415 	struct umowner *umo_next;		/* Next hash elt in chain */
3416 	size_t umo_signature;			/* Hash table signature */
3417 	uint_t umo_num;				/* Number of allocations */
3418 	size_t umo_data_size;			/* Size of each allocation */
3419 	size_t umo_total_size;			/* Total bytes of allocation */
3420 	int umo_depth;				/* Depth of stack trace */
3421 	uintptr_t *umo_stack;			/* Stack trace */
3422 } umowner_t;
3423 
3424 typedef struct umusers {
3425 	const umem_cache_t *umu_cache;		/* Current umem cache */
3426 	umowner_t *umu_hash;			/* Hash table of owners */
3427 	uintptr_t *umu_stacks;			/* stacks for owners */
3428 	int umu_nelems;				/* Number of entries in use */
3429 	int umu_size;				/* Total number of entries */
3430 } umusers_t;
3431 
3432 static void
3433 umu_add(umusers_t *umu, const umem_bufctl_audit_t *bcp,
3434     size_t size, size_t data_size)
3435 {
3436 	int i, depth = MIN(bcp->bc_depth, umem_stack_depth);
3437 	size_t bucket, signature = data_size;
3438 	umowner_t *umo, *umoend;
3439 
3440 	/*
3441 	 * If the hash table is full, double its size and rehash everything.
3442 	 */
3443 	if (umu->umu_nelems >= umu->umu_size) {
3444 		int s = umu->umu_size ? umu->umu_size * 2 : 1024;
3445 		size_t umowner_size = sizeof (umowner_t);
3446 		size_t trace_size = umem_stack_depth * sizeof (uintptr_t);
3447 		uintptr_t *new_stacks;
3448 
3449 		umo = mdb_alloc(umowner_size * s, UM_SLEEP | UM_GC);
3450 		new_stacks = mdb_alloc(trace_size * s, UM_SLEEP | UM_GC);
3451 
3452 		bcopy(umu->umu_hash, umo, umowner_size * umu->umu_size);
3453 		bcopy(umu->umu_stacks, new_stacks, trace_size * umu->umu_size);
3454 		umu->umu_hash = umo;
3455 		umu->umu_stacks = new_stacks;
3456 		umu->umu_size = s;
3457 
3458 		umoend = umu->umu_hash + umu->umu_size;
3459 		for (umo = umu->umu_hash; umo < umoend; umo++) {
3460 			umo->umo_head = NULL;
3461 			umo->umo_stack = &umu->umu_stacks[
3462 			    umem_stack_depth * (umo - umu->umu_hash)];
3463 		}
3464 
3465 		umoend = umu->umu_hash + umu->umu_nelems;
3466 		for (umo = umu->umu_hash; umo < umoend; umo++) {
3467 			bucket = umo->umo_signature & (umu->umu_size - 1);
3468 			umo->umo_next = umu->umu_hash[bucket].umo_head;
3469 			umu->umu_hash[bucket].umo_head = umo;
3470 		}
3471 	}
3472 
3473 	/*
3474 	 * Finish computing the hash signature from the stack trace, and then
3475 	 * see if the owner is in the hash table.  If so, update our stats.
3476 	 */
3477 	for (i = 0; i < depth; i++)
3478 		signature += bcp->bc_stack[i];
3479 
3480 	bucket = signature & (umu->umu_size - 1);
3481 
3482 	for (umo = umu->umu_hash[bucket].umo_head; umo; umo = umo->umo_next) {
3483 		if (umo->umo_signature == signature) {
3484 			size_t difference = 0;
3485 
3486 			difference |= umo->umo_data_size - data_size;
3487 			difference |= umo->umo_depth - depth;
3488 
3489 			for (i = 0; i < depth; i++) {
3490 				difference |= umo->umo_stack[i] -
3491 				    bcp->bc_stack[i];
3492 			}
3493 
3494 			if (difference == 0) {
3495 				umo->umo_total_size += size;
3496 				umo->umo_num++;
3497 				return;
3498 			}
3499 		}
3500 	}
3501 
3502 	/*
3503 	 * If the owner is not yet hashed, grab the next element and fill it
3504 	 * in based on the allocation information.
3505 	 */
3506 	umo = &umu->umu_hash[umu->umu_nelems++];
3507 	umo->umo_next = umu->umu_hash[bucket].umo_head;
3508 	umu->umu_hash[bucket].umo_head = umo;
3509 
3510 	umo->umo_signature = signature;
3511 	umo->umo_num = 1;
3512 	umo->umo_data_size = data_size;
3513 	umo->umo_total_size = size;
3514 	umo->umo_depth = depth;
3515 
3516 	for (i = 0; i < depth; i++)
3517 		umo->umo_stack[i] = bcp->bc_stack[i];
3518 }
3519 
3520 /*
3521  * When ::umausers is invoked without the -f flag, we simply update our hash
3522  * table with the information from each allocated bufctl.
3523  */
3524 /*ARGSUSED*/
3525 static int
3526 umause1(uintptr_t addr, const umem_bufctl_audit_t *bcp, umusers_t *umu)
3527 {
3528 	const umem_cache_t *cp = umu->umu_cache;
3529 
3530 	umu_add(umu, bcp, cp->cache_bufsize, cp->cache_bufsize);
3531 	return (WALK_NEXT);
3532 }
3533 
3534 /*
3535  * When ::umausers is invoked with the -f flag, we print out the information
3536  * for each bufctl as well as updating the hash table.
3537  */
3538 static int
3539 umause2(uintptr_t addr, const umem_bufctl_audit_t *bcp, umusers_t *umu)
3540 {
3541 	int i, depth = MIN(bcp->bc_depth, umem_stack_depth);
3542 	const umem_cache_t *cp = umu->umu_cache;
3543 
3544 	mdb_printf("size %d, addr %p, thread %p, cache %s\n",
3545 	    cp->cache_bufsize, addr, bcp->bc_thread, cp->cache_name);
3546 
3547 	for (i = 0; i < depth; i++)
3548 		mdb_printf("\t %a\n", bcp->bc_stack[i]);
3549 
3550 	umu_add(umu, bcp, cp->cache_bufsize, cp->cache_bufsize);
3551 	return (WALK_NEXT);
3552 }
3553 
3554 /*
3555  * We sort our results by allocation size before printing them.
3556  */
3557 static int
3558 umownercmp(const void *lp, const void *rp)
3559 {
3560 	const umowner_t *lhs = lp;
3561 	const umowner_t *rhs = rp;
3562 
3563 	return (rhs->umo_total_size - lhs->umo_total_size);
3564 }
3565 
3566 /*
3567  * The main engine of ::umausers is relatively straightforward: First we
3568  * accumulate our list of umem_cache_t addresses into the umclist_t. Next we
3569  * iterate over the allocated bufctls of each cache in the list.  Finally,
3570  * we sort and print our results.
3571  */
3572 /*ARGSUSED*/
3573 int
3574 umausers(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
3575 {
3576 	int mem_threshold = 8192;	/* Minimum # bytes for printing */
3577 	int cnt_threshold = 100;	/* Minimum # blocks for printing */
3578 	int audited_caches = 0;		/* Number of UMF_AUDIT caches found */
3579 	int do_all_caches = 1;		/* Do all caches (no arguments) */
3580 	int opt_e = FALSE;		/* Include "small" users */
3581 	int opt_f = FALSE;		/* Print stack traces */
3582 
3583 	mdb_walk_cb_t callback = (mdb_walk_cb_t)umause1;
3584 	umowner_t *umo, *umoend;
3585 	int i, oelems;
3586 
3587 	umclist_t umc;
3588 	umusers_t umu;
3589 
3590 	if (flags & DCMD_ADDRSPEC)
3591 		return (DCMD_USAGE);
3592 
3593 	bzero(&umc, sizeof (umc));
3594 	bzero(&umu, sizeof (umu));
3595 
3596 	while ((i = mdb_getopts(argc, argv,
3597 	    'e', MDB_OPT_SETBITS, TRUE, &opt_e,
3598 	    'f', MDB_OPT_SETBITS, TRUE, &opt_f, NULL)) != argc) {
3599 
3600 		argv += i;	/* skip past options we just processed */
3601 		argc -= i;	/* adjust argc */
3602 
3603 		if (argv->a_type != MDB_TYPE_STRING || *argv->a_un.a_str == '-')
3604 			return (DCMD_USAGE);
3605 
3606 		oelems = umc.umc_nelems;
3607 		umc.umc_name = argv->a_un.a_str;
3608 		(void) mdb_walk("umem_cache", (mdb_walk_cb_t)umc_add, &umc);
3609 
3610 		if (umc.umc_nelems == oelems) {
3611 			mdb_warn("unknown umem cache: %s\n", umc.umc_name);
3612 			return (DCMD_ERR);
3613 		}
3614 
3615 		do_all_caches = 0;
3616 		argv++;
3617 		argc--;
3618 	}
3619 
3620 	if (opt_e)
3621 		mem_threshold = cnt_threshold = 0;
3622 
3623 	if (opt_f)
3624 		callback = (mdb_walk_cb_t)umause2;
3625 
3626 	if (do_all_caches) {
3627 		umc.umc_name = NULL; /* match all cache names */
3628 		(void) mdb_walk("umem_cache", (mdb_walk_cb_t)umc_add, &umc);
3629 	}
3630 
3631 	for (i = 0; i < umc.umc_nelems; i++) {
3632 		uintptr_t cp = umc.umc_caches[i];
3633 		umem_cache_t c;
3634 
3635 		if (mdb_vread(&c, sizeof (c), cp) == -1) {
3636 			mdb_warn("failed to read cache at %p", cp);
3637 			continue;
3638 		}
3639 
3640 		if (!(c.cache_flags & UMF_AUDIT)) {
3641 			if (!do_all_caches) {
3642 				mdb_warn("UMF_AUDIT is not enabled for %s\n",
3643 				    c.cache_name);
3644 			}
3645 			continue;
3646 		}
3647 
3648 		umu.umu_cache = &c;
3649 		(void) mdb_pwalk("bufctl", callback, &umu, cp);
3650 		audited_caches++;
3651 	}
3652 
3653 	if (audited_caches == 0 && do_all_caches) {
3654 		mdb_warn("UMF_AUDIT is not enabled for any caches\n");
3655 		return (DCMD_ERR);
3656 	}
3657 
3658 	qsort(umu.umu_hash, umu.umu_nelems, sizeof (umowner_t), umownercmp);
3659 	umoend = umu.umu_hash + umu.umu_nelems;
3660 
3661 	for (umo = umu.umu_hash; umo < umoend; umo++) {
3662 		if (umo->umo_total_size < mem_threshold &&
3663 		    umo->umo_num < cnt_threshold)
3664 			continue;
3665 		mdb_printf("%lu bytes for %u allocations with data size %lu:\n",
3666 		    umo->umo_total_size, umo->umo_num, umo->umo_data_size);
3667 		for (i = 0; i < umo->umo_depth; i++)
3668 			mdb_printf("\t %a\n", umo->umo_stack[i]);
3669 	}
3670 
3671 	return (DCMD_OK);
3672 }
3673 
3674 struct malloc_data {
3675 	uint32_t malloc_size;
3676 	uint32_t malloc_stat; /* == UMEM_MALLOC_ENCODE(state, malloc_size) */
3677 };
3678 
3679 #ifdef _LP64
3680 #define	UMI_MAX_BUCKET		(UMEM_MAXBUF - 2*sizeof (struct malloc_data))
3681 #else
3682 #define	UMI_MAX_BUCKET		(UMEM_MAXBUF - sizeof (struct malloc_data))
3683 #endif
3684 
3685 typedef struct umem_malloc_info {
3686 	size_t um_total;	/* total allocated buffers */
3687 	size_t um_malloc;	/* malloc buffers */
3688 	size_t um_malloc_size;	/* sum of malloc buffer sizes */
3689 	size_t um_malloc_overhead; /* sum of in-chunk overheads */
3690 
3691 	umem_cache_t *um_cp;
3692 
3693 	uint_t *um_bucket;
3694 } umem_malloc_info_t;
3695 
3696 static void
3697 umem_malloc_print_dist(uint_t *um_bucket, size_t minmalloc, size_t maxmalloc,
3698     size_t maxbuckets, size_t minbucketsize, int geometric)
3699 {
3700 	uint64_t um_malloc;
3701 	int minb = -1;
3702 	int maxb = -1;
3703 	int buckets;
3704 	int nbucks;
3705 	int i;
3706 	int b;
3707 	const int *distarray;
3708 
3709 	minb = (int)minmalloc;
3710 	maxb = (int)maxmalloc;
3711 
3712 	nbucks = buckets = maxb - minb + 1;
3713 
3714 	um_malloc = 0;
3715 	for (b = minb; b <= maxb; b++)
3716 		um_malloc += um_bucket[b];
3717 
3718 	if (maxbuckets != 0)
3719 		buckets = MIN(buckets, maxbuckets);
3720 
3721 	if (minbucketsize > 1) {
3722 		buckets = MIN(buckets, nbucks/minbucketsize);
3723 		if (buckets == 0) {
3724 			buckets = 1;
3725 			minbucketsize = nbucks;
3726 		}
3727 	}
3728 
3729 	if (geometric)
3730 		distarray = dist_geometric(buckets, minb, maxb, minbucketsize);
3731 	else
3732 		distarray = dist_linear(buckets, minb, maxb);
3733 
3734 	dist_print_header("malloc size", 11, "count");
3735 	for (i = 0; i < buckets; i++) {
3736 		dist_print_bucket(distarray, i, um_bucket, um_malloc, 11);
3737 	}
3738 	mdb_printf("\n");
3739 }
3740 
3741 /*
3742  * A malloc()ed buffer looks like:
3743  *
3744  *	<----------- mi.malloc_size --->
3745  *	<----------- cp.cache_bufsize ------------------>
3746  *	<----------- cp.cache_chunksize -------------------------------->
3747  *	+-------+-----------------------+---------------+---------------+
3748  *	|/tag///| mallocsz		|/round-off/////|/debug info////|
3749  *	+-------+---------------------------------------+---------------+
3750  *		<-- usable space ------>
3751  *
3752  * mallocsz is the argument to malloc(3C).
3753  * mi.malloc_size is the actual size passed to umem_alloc(), which
3754  * is rounded up to the smallest available cache size, which is
3755  * cache_bufsize.  If there is debugging or alignment overhead in
3756  * the cache, that is reflected in a larger cache_chunksize.
3757  *
3758  * The tag at the beginning of the buffer is either 8-bytes or 16-bytes,
3759  * depending upon the ISA's alignment requirements.  For 32-bit allocations,
3760  * it is always a 8-byte tag.  For 64-bit allocations larger than 8 bytes,
3761  * the tag has 8 bytes of padding before it.
3762  *
3763  * 32-byte, 64-byte buffers <= 8 bytes:
3764  *	+-------+-------+--------- ...
3765  *	|/size//|/stat//| mallocsz ...
3766  *	+-------+-------+--------- ...
3767  *			^
3768  *			pointer returned from malloc(3C)
3769  *
3770  * 64-byte buffers > 8 bytes:
3771  *	+---------------+-------+-------+--------- ...
3772  *	|/padding///////|/size//|/stat//| mallocsz ...
3773  *	+---------------+-------+-------+--------- ...
3774  *					^
3775  *					pointer returned from malloc(3C)
3776  *
3777  * The "size" field is "malloc_size", which is mallocsz + the padding.
3778  * The "stat" field is derived from malloc_size, and functions as a
3779  * validation that this buffer is actually from malloc(3C).
3780  */
3781 /*ARGSUSED*/
3782 static int
3783 um_umem_buffer_cb(uintptr_t addr, void *buf, umem_malloc_info_t *ump)
3784 {
3785 	struct malloc_data md;
3786 	size_t m_addr = addr;
3787 	size_t overhead = sizeof (md);
3788 	size_t mallocsz;
3789 
3790 	ump->um_total++;
3791 
3792 #ifdef _LP64
3793 	if (ump->um_cp->cache_bufsize > UMEM_SECOND_ALIGN) {
3794 		m_addr += overhead;
3795 		overhead += sizeof (md);
3796 	}
3797 #endif
3798 
3799 	if (mdb_vread(&md, sizeof (md), m_addr) == -1) {
3800 		mdb_warn("unable to read malloc header at %p", m_addr);
3801 		return (WALK_NEXT);
3802 	}
3803 
3804 	switch (UMEM_MALLOC_DECODE(md.malloc_stat, md.malloc_size)) {
3805 	case MALLOC_MAGIC:
3806 #ifdef _LP64
3807 	case MALLOC_SECOND_MAGIC:
3808 #endif
3809 		mallocsz = md.malloc_size - overhead;
3810 
3811 		ump->um_malloc++;
3812 		ump->um_malloc_size += mallocsz;
3813 		ump->um_malloc_overhead += overhead;
3814 
3815 		/* include round-off and debug overhead */
3816 		ump->um_malloc_overhead +=
3817 		    ump->um_cp->cache_chunksize - md.malloc_size;
3818 
3819 		if (ump->um_bucket != NULL && mallocsz <= UMI_MAX_BUCKET)
3820 			ump->um_bucket[mallocsz]++;
3821 
3822 		break;
3823 	default:
3824 		break;
3825 	}
3826 
3827 	return (WALK_NEXT);
3828 }
3829 
3830 int
3831 get_umem_alloc_sizes(int **out, size_t *out_num)
3832 {
3833 	GElf_Sym sym;
3834 
3835 	if (umem_lookup_by_name("umem_alloc_sizes", &sym) == -1) {
3836 		mdb_warn("unable to look up umem_alloc_sizes");
3837 		return (-1);
3838 	}
3839 
3840 	*out = mdb_alloc(sym.st_size, UM_SLEEP | UM_GC);
3841 	*out_num = sym.st_size / sizeof (int);
3842 
3843 	if (mdb_vread(*out, sym.st_size, sym.st_value) == -1) {
3844 		mdb_warn("unable to read umem_alloc_sizes (%p)", sym.st_value);
3845 		*out = NULL;
3846 		return (-1);
3847 	}
3848 
3849 	return (0);
3850 }
3851 
3852 
3853 static int
3854 um_umem_cache_cb(uintptr_t addr, umem_cache_t *cp, umem_malloc_info_t *ump)
3855 {
3856 	if (strncmp(cp->cache_name, "umem_alloc_", strlen("umem_alloc_")) != 0)
3857 		return (WALK_NEXT);
3858 
3859 	ump->um_cp = cp;
3860 
3861 	if (mdb_pwalk("umem", (mdb_walk_cb_t)um_umem_buffer_cb, ump, addr) ==
3862 	    -1) {
3863 		mdb_warn("can't walk 'umem' for cache %p", addr);
3864 		return (WALK_ERR);
3865 	}
3866 
3867 	return (WALK_NEXT);
3868 }
3869 
3870 void
3871 umem_malloc_dist_help(void)
3872 {
3873 	mdb_printf("%s\n",
3874 	    "report distribution of outstanding malloc()s");
3875 	mdb_dec_indent(2);
3876 	mdb_printf("%<b>OPTIONS%</b>\n");
3877 	mdb_inc_indent(2);
3878 	mdb_printf("%s",
3879 "  -b maxbins\n"
3880 "        Use at most maxbins bins for the data\n"
3881 "  -B minbinsize\n"
3882 "        Make the bins at least minbinsize bytes apart\n"
3883 "  -d    dump the raw data out, without binning\n"
3884 "  -g    use geometric binning instead of linear binning\n");
3885 }
3886 
3887 /*ARGSUSED*/
3888 int
3889 umem_malloc_dist(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
3890 {
3891 	umem_malloc_info_t mi;
3892 	uint_t geometric = 0;
3893 	uint_t dump = 0;
3894 	size_t maxbuckets = 0;
3895 	size_t minbucketsize = 0;
3896 
3897 	size_t minalloc = 0;
3898 	size_t maxalloc = UMI_MAX_BUCKET;
3899 
3900 	if (flags & DCMD_ADDRSPEC)
3901 		return (DCMD_USAGE);
3902 
3903 	if (mdb_getopts(argc, argv,
3904 	    'd', MDB_OPT_SETBITS, TRUE, &dump,
3905 	    'g', MDB_OPT_SETBITS, TRUE, &geometric,
3906 	    'b', MDB_OPT_UINTPTR, &maxbuckets,
3907 	    'B', MDB_OPT_UINTPTR, &minbucketsize,
3908 	    0) != argc)
3909 		return (DCMD_USAGE);
3910 
3911 	bzero(&mi, sizeof (mi));
3912 	mi.um_bucket = mdb_zalloc((UMI_MAX_BUCKET + 1) * sizeof (*mi.um_bucket),
3913 	    UM_SLEEP | UM_GC);
3914 
3915 	if (mdb_walk("umem_cache", (mdb_walk_cb_t)um_umem_cache_cb,
3916 	    &mi) == -1) {
3917 		mdb_warn("unable to walk 'umem_cache'");
3918 		return (DCMD_ERR);
3919 	}
3920 
3921 	if (dump) {
3922 		int i;
3923 		for (i = minalloc; i <= maxalloc; i++)
3924 			mdb_printf("%d\t%d\n", i, mi.um_bucket[i]);
3925 
3926 		return (DCMD_OK);
3927 	}
3928 
3929 	umem_malloc_print_dist(mi.um_bucket, minalloc, maxalloc,
3930 	    maxbuckets, minbucketsize, geometric);
3931 
3932 	return (DCMD_OK);
3933 }
3934 
3935 void
3936 umem_malloc_info_help(void)
3937 {
3938 	mdb_printf("%s\n",
3939 	    "report information about malloc()s by cache.  ");
3940 	mdb_dec_indent(2);
3941 	mdb_printf("%<b>OPTIONS%</b>\n");
3942 	mdb_inc_indent(2);
3943 	mdb_printf("%s",
3944 "  -b maxbins\n"
3945 "        Use at most maxbins bins for the data\n"
3946 "  -B minbinsize\n"
3947 "        Make the bins at least minbinsize bytes apart\n"
3948 "  -d    dump the raw distribution data without binning\n"
3949 #ifndef _KMDB
3950 "  -g    use geometric binning instead of linear binning\n"
3951 #endif
3952 	    "");
3953 }
3954 int
3955 umem_malloc_info(uintptr_t addr, uint_t flags, int argc, const mdb_arg_t *argv)
3956 {
3957 	umem_cache_t c;
3958 	umem_malloc_info_t mi;
3959 
3960 	int skip = 0;
3961 
3962 	size_t maxmalloc;
3963 	size_t overhead;
3964 	size_t allocated;
3965 	size_t avg_malloc;
3966 	size_t overhead_pct;	/* 1000 * overhead_percent */
3967 
3968 	uint_t verbose = 0;
3969 	uint_t dump = 0;
3970 	uint_t geometric = 0;
3971 	size_t maxbuckets = 0;
3972 	size_t minbucketsize = 0;
3973 
3974 	int *alloc_sizes;
3975 	int idx;
3976 	size_t num;
3977 	size_t minmalloc;
3978 
3979 	if (mdb_getopts(argc, argv,
3980 	    'd', MDB_OPT_SETBITS, TRUE, &dump,
3981 	    'g', MDB_OPT_SETBITS, TRUE, &geometric,
3982 	    'b', MDB_OPT_UINTPTR, &maxbuckets,
3983 	    'B', MDB_OPT_UINTPTR, &minbucketsize,
3984 	    0) != argc)
3985 		return (DCMD_USAGE);
3986 
3987 	if (dump || geometric || (maxbuckets != 0) || (minbucketsize != 0))
3988 		verbose = 1;
3989 
3990 	if (!(flags & DCMD_ADDRSPEC)) {
3991 		if (mdb_walk_dcmd("umem_cache", "umem_malloc_info",
3992 		    argc, argv) == -1) {
3993 			mdb_warn("can't walk umem_cache");
3994 			return (DCMD_ERR);
3995 		}
3996 		return (DCMD_OK);
3997 	}
3998 
3999 	if (!mdb_vread(&c, sizeof (c), addr)) {
4000 		mdb_warn("unable to read cache at %p", addr);
4001 		return (DCMD_ERR);
4002 	}
4003 
4004 	if (strncmp(c.cache_name, "umem_alloc_", strlen("umem_alloc_")) != 0) {
4005 		if (!(flags & DCMD_LOOP))
4006 			mdb_warn("umem_malloc_info: cache \"%s\" is not used "
4007 			    "by malloc()\n", c.cache_name);
4008 		skip = 1;
4009 	}
4010 
4011 	/*
4012 	 * normally, print the header only the first time.  In verbose mode,
4013 	 * print the header on every non-skipped buffer
4014 	 */
4015 	if ((!verbose && DCMD_HDRSPEC(flags)) || (verbose && !skip))
4016 		mdb_printf("%<ul>%-?s %6s %6s %8s %8s %10s %10s %6s%</ul>\n",
4017 		    "CACHE", "BUFSZ", "MAXMAL",
4018 		    "BUFMALLC", "AVG_MAL", "MALLOCED", "OVERHEAD", "%OVER");
4019 
4020 	if (skip)
4021 		return (DCMD_OK);
4022 
4023 	maxmalloc = c.cache_bufsize - sizeof (struct malloc_data);
4024 #ifdef _LP64
4025 	if (c.cache_bufsize > UMEM_SECOND_ALIGN)
4026 		maxmalloc -= sizeof (struct malloc_data);
4027 #endif
4028 
4029 	bzero(&mi, sizeof (mi));
4030 	mi.um_cp = &c;
4031 	if (verbose)
4032 		mi.um_bucket =
4033 		    mdb_zalloc((UMI_MAX_BUCKET + 1) * sizeof (*mi.um_bucket),
4034 		    UM_SLEEP | UM_GC);
4035 
4036 	if (mdb_pwalk("umem", (mdb_walk_cb_t)um_umem_buffer_cb, &mi, addr) ==
4037 	    -1) {
4038 		mdb_warn("can't walk 'umem'");
4039 		return (DCMD_ERR);
4040 	}
4041 
4042 	overhead = mi.um_malloc_overhead;
4043 	allocated = mi.um_malloc_size;
4044 
4045 	/* do integer round off for the average */
4046 	if (mi.um_malloc != 0)
4047 		avg_malloc = (allocated + (mi.um_malloc - 1)/2) / mi.um_malloc;
4048 	else
4049 		avg_malloc = 0;
4050 
4051 	/*
4052 	 * include per-slab overhead
4053 	 *
4054 	 * Each slab in a given cache is the same size, and has the same
4055 	 * number of chunks in it;  we read in the first slab on the
4056 	 * slab list to get the number of chunks for all slabs.  To
4057 	 * compute the per-slab overhead, we just subtract the chunk usage
4058 	 * from the slabsize:
4059 	 *
4060 	 * +------------+-------+-------+ ... --+-------+-------+-------+
4061 	 * |////////////|	|	| ...	|	|///////|///////|
4062 	 * |////color///| chunk	| chunk	| ...	| chunk	|/color/|/slab//|
4063 	 * |////////////|	|	| ...	|	|///////|///////|
4064 	 * +------------+-------+-------+ ... --+-------+-------+-------+
4065 	 * |		\_______chunksize * chunks_____/		|
4066 	 * \__________________________slabsize__________________________/
4067 	 *
4068 	 * For UMF_HASH caches, there is an additional source of overhead;
4069 	 * the external umem_slab_t and per-chunk bufctl structures.  We
4070 	 * include those in our per-slab overhead.
4071 	 *
4072 	 * Once we have a number for the per-slab overhead, we estimate
4073 	 * the actual overhead by treating the malloc()ed buffers as if
4074 	 * they were densely packed:
4075 	 *
4076 	 *	additional overhead = (# mallocs) * (per-slab) / (chunks);
4077 	 *
4078 	 * carefully ordering the multiply before the divide, to avoid
4079 	 * round-off error.
4080 	 */
4081 	if (mi.um_malloc != 0) {
4082 		umem_slab_t slab;
4083 		uintptr_t saddr = (uintptr_t)c.cache_nullslab.slab_next;
4084 
4085 		if (mdb_vread(&slab, sizeof (slab), saddr) == -1) {
4086 			mdb_warn("unable to read slab at %p\n", saddr);
4087 		} else {
4088 			long chunks = slab.slab_chunks;
4089 			if (chunks != 0 && c.cache_chunksize != 0 &&
4090 			    chunks <= c.cache_slabsize / c.cache_chunksize) {
4091 				uintmax_t perslab =
4092 				    c.cache_slabsize -
4093 				    (c.cache_chunksize * chunks);
4094 
4095 				if (c.cache_flags & UMF_HASH) {
4096 					perslab += sizeof (umem_slab_t) +
4097 					    chunks *
4098 					    ((c.cache_flags & UMF_AUDIT) ?
4099 					    sizeof (umem_bufctl_audit_t) :
4100 					    sizeof (umem_bufctl_t));
4101 				}
4102 				overhead +=
4103 				    (perslab * (uintmax_t)mi.um_malloc)/chunks;
4104 			} else {
4105 				mdb_warn("invalid #chunks (%d) in slab %p\n",
4106 				    chunks, saddr);
4107 			}
4108 		}
4109 	}
4110 
4111 	if (allocated != 0)
4112 		overhead_pct = (1000ULL * overhead) / allocated;
4113 	else
4114 		overhead_pct = 0;
4115 
4116 	mdb_printf("%0?p %6ld %6ld %8ld %8ld %10ld %10ld %3ld.%01ld%%\n",
4117 	    addr, c.cache_bufsize, maxmalloc,
4118 	    mi.um_malloc, avg_malloc, allocated, overhead,
4119 	    overhead_pct / 10, overhead_pct % 10);
4120 
4121 	if (!verbose)
4122 		return (DCMD_OK);
4123 
4124 	if (!dump)
4125 		mdb_printf("\n");
4126 
4127 	if (get_umem_alloc_sizes(&alloc_sizes, &num) == -1)
4128 		return (DCMD_ERR);
4129 
4130 	for (idx = 0; idx < num; idx++) {
4131 		if (alloc_sizes[idx] == c.cache_bufsize)
4132 			break;
4133 		if (alloc_sizes[idx] == 0) {
4134 			idx = num;	/* 0-terminated array */
4135 			break;
4136 		}
4137 	}
4138 	if (idx == num) {
4139 		mdb_warn(
4140 		    "cache %p's size (%d) not in umem_alloc_sizes\n",
4141 		    addr, c.cache_bufsize);
4142 		return (DCMD_ERR);
4143 	}
4144 
4145 	minmalloc = (idx == 0)? 0 : alloc_sizes[idx - 1];
4146 	if (minmalloc > 0) {
4147 #ifdef _LP64
4148 		if (minmalloc > UMEM_SECOND_ALIGN)
4149 			minmalloc -= sizeof (struct malloc_data);
4150 #endif
4151 		minmalloc -= sizeof (struct malloc_data);
4152 		minmalloc += 1;
4153 	}
4154 
4155 	if (dump) {
4156 		for (idx = minmalloc; idx <= maxmalloc; idx++)
4157 			mdb_printf("%d\t%d\n", idx, mi.um_bucket[idx]);
4158 		mdb_printf("\n");
4159 	} else {
4160 		umem_malloc_print_dist(mi.um_bucket, minmalloc, maxmalloc,
4161 		    maxbuckets, minbucketsize, geometric);
4162 	}
4163 
4164 	return (DCMD_OK);
4165 }
4166