xref: /freebsd/lib/libmemstat/memstat_uma.c (revision bdd1243d) 
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2005-2006 Robert N. M. Watson
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * $FreeBSD$
29  */
30 
31 #define _WANT_FREEBSD_BITSET
32 
33 #include <sys/param.h>
34 #include <sys/counter.h>
35 #include <sys/cpuset.h>
36 #include <sys/sysctl.h>
37 
38 #include <vm/uma.h>
39 #include <vm/uma_int.h>
40 
41 #include <err.h>
42 #include <errno.h>
43 #include <kvm.h>
44 #include <nlist.h>
45 #include <stddef.h>
46 #include <stdio.h>
47 #include <stdlib.h>
48 #include <string.h>
49 #include <unistd.h>
50 
51 #include "memstat.h"
52 #include "memstat_internal.h"
53 
54 static struct nlist namelist[] = {
55 #define	X_UMA_KEGS	0
56 	{ .n_name = "_uma_kegs" },
57 #define	X_MP_MAXID	1
58 	{ .n_name = "_mp_maxid" },
59 #define	X_ALL_CPUS	2
60 	{ .n_name = "_all_cpus" },
61 #define	X_VM_NDOMAINS	3
62 	{ .n_name = "_vm_ndomains" },
63 	{ .n_name = "" },
64 };
65 
66 /*
67  * Extract uma(9) statistics from the running kernel, and store all memory
68  * type information in the passed list.  For each type, check the list for an
69  * existing entry with the right name/allocator -- if present, update that
70  * entry.  Otherwise, add a new entry.  On error, the entire list will be
71  * cleared, as entries will be in an inconsistent state.
72  *
73  * To reduce the level of work for a list that starts empty, we keep around a
74  * hint as to whether it was empty when we began, so we can avoid searching
75  * the list for entries to update.  Updates are O(n^2) due to searching for
76  * each entry before adding it.
77  */
78 int
79 memstat_sysctl_uma(struct memory_type_list *list, int flags)
80 {
81 	struct uma_stream_header *ushp;
82 	struct uma_type_header *uthp;
83 	struct uma_percpu_stat *upsp;
84 	struct memory_type *mtp;
85 	int count, hint_dontsearch, i, j, maxcpus, maxid;
86 	char *buffer, *p;
87 	size_t size;
88 
89 	hint_dontsearch = LIST_EMPTY(&list->mtl_list);
90 
91 	/*
92 	 * Query the number of CPUs, number of malloc types so that we can
93 	 * guess an initial buffer size.  We loop until we succeed or really
94 	 * fail.  Note that the value of maxcpus we query using sysctl is not
95 	 * the version we use when processing the real data -- that is read
96 	 * from the header.
97 	 */
98 retry:
99 	size = sizeof(maxid);
100 	if (sysctlbyname("kern.smp.maxid", &maxid, &size, NULL, 0) < 0) {
101 		if (errno == EACCES || errno == EPERM)
102 			list->mtl_error = MEMSTAT_ERROR_PERMISSION;
103 		else
104 			list->mtl_error = MEMSTAT_ERROR_DATAERROR;
105 		return (-1);
106 	}
107 	if (size != sizeof(maxid)) {
108 		list->mtl_error = MEMSTAT_ERROR_DATAERROR;
109 		return (-1);
110 	}
111 
112 	size = sizeof(count);
113 	if (sysctlbyname("vm.zone_count", &count, &size, NULL, 0) < 0) {
114 		if (errno == EACCES || errno == EPERM)
115 			list->mtl_error = MEMSTAT_ERROR_PERMISSION;
116 		else
117 			list->mtl_error = MEMSTAT_ERROR_VERSION;
118 		return (-1);
119 	}
120 	if (size != sizeof(count)) {
121 		list->mtl_error = MEMSTAT_ERROR_DATAERROR;
122 		return (-1);
123 	}
124 
125 	size = sizeof(*uthp) + count * (sizeof(*uthp) + sizeof(*upsp) *
126 	    (maxid + 1));
127 
128 	buffer = malloc(size);
129 	if (buffer == NULL) {
130 		list->mtl_error = MEMSTAT_ERROR_NOMEMORY;
131 		return (-1);
132 	}
133 
134 	if (sysctlbyname("vm.zone_stats", buffer, &size, NULL, 0) < 0) {
135 		/*
136 		 * XXXRW: ENOMEM is an ambiguous return, we should bound the
137 		 * number of loops, perhaps.
138 		 */
139 		if (errno == ENOMEM) {
140 			free(buffer);
141 			goto retry;
142 		}
143 		if (errno == EACCES || errno == EPERM)
144 			list->mtl_error = MEMSTAT_ERROR_PERMISSION;
145 		else
146 			list->mtl_error = MEMSTAT_ERROR_VERSION;
147 		free(buffer);
148 		return (-1);
149 	}
150 
151 	if (size == 0) {
152 		free(buffer);
153 		return (0);
154 	}
155 
156 	if (size < sizeof(*ushp)) {
157 		list->mtl_error = MEMSTAT_ERROR_VERSION;
158 		free(buffer);
159 		return (-1);
160 	}
161 	p = buffer;
162 	ushp = (struct uma_stream_header *)p;
163 	p += sizeof(*ushp);
164 
165 	if (ushp->ush_version != UMA_STREAM_VERSION) {
166 		list->mtl_error = MEMSTAT_ERROR_VERSION;
167 		free(buffer);
168 		return (-1);
169 	}
170 
171 	/*
172 	 * For the remainder of this function, we are quite trusting about
173 	 * the layout of structures and sizes, since we've determined we have
174 	 * a matching version and acceptable CPU count.
175 	 */
176 	maxcpus = ushp->ush_maxcpus;
177 	count = ushp->ush_count;
178 	for (i = 0; i < count; i++) {
179 		uthp = (struct uma_type_header *)p;
180 		p += sizeof(*uthp);
181 
182 		if (hint_dontsearch == 0) {
183 			mtp = memstat_mtl_find(list, ALLOCATOR_UMA,
184 			    uthp->uth_name);
185 		} else
186 			mtp = NULL;
187 		if (mtp == NULL)
188 			mtp = _memstat_mt_allocate(list, ALLOCATOR_UMA,
189 			    uthp->uth_name, maxid + 1);
190 		if (mtp == NULL) {
191 			_memstat_mtl_empty(list);
192 			free(buffer);
193 			list->mtl_error = MEMSTAT_ERROR_NOMEMORY;
194 			return (-1);
195 		}
196 
197 		/*
198 		 * Reset the statistics on a current node.
199 		 */
200 		_memstat_mt_reset_stats(mtp, maxid + 1);
201 
202 		mtp->mt_numallocs = uthp->uth_allocs;
203 		mtp->mt_numfrees = uthp->uth_frees;
204 		mtp->mt_failures = uthp->uth_fails;
205 		mtp->mt_sleeps = uthp->uth_sleeps;
206 		mtp->mt_xdomain = uthp->uth_xdomain;
207 
208 		for (j = 0; j < maxcpus; j++) {
209 			upsp = (struct uma_percpu_stat *)p;
210 			p += sizeof(*upsp);
211 
212 			mtp->mt_percpu_cache[j].mtp_free =
213 			    upsp->ups_cache_free;
214 			mtp->mt_free += upsp->ups_cache_free;
215 			mtp->mt_numallocs += upsp->ups_allocs;
216 			mtp->mt_numfrees += upsp->ups_frees;
217 		}
218 
219 		/*
220 		 * Values for uth_allocs and uth_frees frees are snap.
221 		 * It may happen that kernel reports that number of frees
222 		 * is greater than number of allocs. See counter(9) for
223 		 * details.
224 		 */
225 		if (mtp->mt_numallocs < mtp->mt_numfrees)
226 			mtp->mt_numallocs = mtp->mt_numfrees;
227 
228 		mtp->mt_size = uthp->uth_size;
229 		mtp->mt_rsize = uthp->uth_rsize;
230 		mtp->mt_memalloced = mtp->mt_numallocs * uthp->uth_size;
231 		mtp->mt_memfreed = mtp->mt_numfrees * uthp->uth_size;
232 		mtp->mt_bytes = mtp->mt_memalloced - mtp->mt_memfreed;
233 		mtp->mt_countlimit = uthp->uth_limit;
234 		mtp->mt_byteslimit = uthp->uth_limit * uthp->uth_size;
235 
236 		mtp->mt_count = mtp->mt_numallocs - mtp->mt_numfrees;
237 		mtp->mt_zonefree = uthp->uth_zone_free;
238 
239 		/*
240 		 * UMA secondary zones share a keg with the primary zone.  To
241 		 * avoid double-reporting of free items, report keg free
242 		 * items only in the primary zone.
243 		 */
244 		if (!(uthp->uth_zone_flags & UTH_ZONE_SECONDARY)) {
245 			mtp->mt_kegfree = uthp->uth_keg_free;
246 			mtp->mt_free += mtp->mt_kegfree;
247 		}
248 		mtp->mt_free += mtp->mt_zonefree;
249 	}
250 
251 	free(buffer);
252 
253 	return (0);
254 }
255 
256 static int
257 kread(kvm_t *kvm, void *kvm_pointer, void *address, size_t size,
258     size_t offset)
259 {
260 	ssize_t ret;
261 
262 	ret = kvm_read(kvm, (unsigned long)kvm_pointer + offset, address,
263 	    size);
264 	if (ret < 0)
265 		return (MEMSTAT_ERROR_KVM);
266 	if ((size_t)ret != size)
267 		return (MEMSTAT_ERROR_KVM_SHORTREAD);
268 	return (0);
269 }
270 
271 static int
272 kread_string(kvm_t *kvm, const void *kvm_pointer, char *buffer, int buflen)
273 {
274 	ssize_t ret;
275 	int i;
276 
277 	for (i = 0; i < buflen; i++) {
278 		ret = kvm_read(kvm, (unsigned long)kvm_pointer + i,
279 		    &(buffer[i]), sizeof(char));
280 		if (ret < 0)
281 			return (MEMSTAT_ERROR_KVM);
282 		if ((size_t)ret != sizeof(char))
283 			return (MEMSTAT_ERROR_KVM_SHORTREAD);
284 		if (buffer[i] == '\0')
285 			return (0);
286 	}
287 	/* Truncate. */
288 	buffer[i-1] = '\0';
289 	return (0);
290 }
291 
292 static int
293 kread_symbol(kvm_t *kvm, int index, void *address, size_t size,
294     size_t offset)
295 {
296 	ssize_t ret;
297 
298 	ret = kvm_read(kvm, namelist[index].n_value + offset, address, size);
299 	if (ret < 0)
300 		return (MEMSTAT_ERROR_KVM);
301 	if ((size_t)ret != size)
302 		return (MEMSTAT_ERROR_KVM_SHORTREAD);
303 	return (0);
304 }
305 
306 /*
307  * memstat_kvm_uma() is similar to memstat_sysctl_uma(), only it extracts
308  * UMA(9) statistics from a kernel core/memory file.
309  */
310 int
311 memstat_kvm_uma(struct memory_type_list *list, void *kvm_handle)
312 {
313 	LIST_HEAD(, uma_keg) uma_kegs;
314 	struct memory_type *mtp;
315 	struct uma_zone_domain uzd;
316 	struct uma_domain ukd;
317 	struct uma_bucket *ubp, ub;
318 	struct uma_cache *ucp, *ucp_array;
319 	struct uma_zone *uzp, uz;
320 	struct uma_keg *kzp, kz;
321 	uint64_t kegfree;
322 	int hint_dontsearch, i, mp_maxid, ndomains, ret;
323 	char name[MEMTYPE_MAXNAME];
324 	cpuset_t all_cpus;
325 	long cpusetsize;
326 	kvm_t *kvm;
327 
328 	kvm = (kvm_t *)kvm_handle;
329 	hint_dontsearch = LIST_EMPTY(&list->mtl_list);
330 	if (kvm_nlist(kvm, namelist) != 0) {
331 		list->mtl_error = MEMSTAT_ERROR_KVM;
332 		return (-1);
333 	}
334 	if (namelist[X_UMA_KEGS].n_type == 0 ||
335 	    namelist[X_UMA_KEGS].n_value == 0) {
336 		list->mtl_error = MEMSTAT_ERROR_KVM_NOSYMBOL;
337 		return (-1);
338 	}
339 	ret = kread_symbol(kvm, X_MP_MAXID, &mp_maxid, sizeof(mp_maxid), 0);
340 	if (ret != 0) {
341 		list->mtl_error = ret;
342 		return (-1);
343 	}
344 	ret = kread_symbol(kvm, X_VM_NDOMAINS, &ndomains,
345 	    sizeof(ndomains), 0);
346 	if (ret != 0) {
347 		list->mtl_error = ret;
348 		return (-1);
349 	}
350 	ret = kread_symbol(kvm, X_UMA_KEGS, &uma_kegs, sizeof(uma_kegs), 0);
351 	if (ret != 0) {
352 		list->mtl_error = ret;
353 		return (-1);
354 	}
355 	cpusetsize = sysconf(_SC_CPUSET_SIZE);
356 	if (cpusetsize == -1 || (u_long)cpusetsize > sizeof(cpuset_t)) {
357 		list->mtl_error = MEMSTAT_ERROR_KVM_NOSYMBOL;
358 		return (-1);
359 	}
360 	CPU_ZERO(&all_cpus);
361 	ret = kread_symbol(kvm, X_ALL_CPUS, &all_cpus, cpusetsize, 0);
362 	if (ret != 0) {
363 		list->mtl_error = ret;
364 		return (-1);
365 	}
366 	ucp_array = malloc(sizeof(struct uma_cache) * (mp_maxid + 1));
367 	if (ucp_array == NULL) {
368 		list->mtl_error = MEMSTAT_ERROR_NOMEMORY;
369 		return (-1);
370 	}
371 	for (kzp = LIST_FIRST(&uma_kegs); kzp != NULL; kzp =
372 	    LIST_NEXT(&kz, uk_link)) {
373 		ret = kread(kvm, kzp, &kz, sizeof(kz), 0);
374 		if (ret != 0) {
375 			free(ucp_array);
376 			_memstat_mtl_empty(list);
377 			list->mtl_error = ret;
378 			return (-1);
379 		}
380 		for (uzp = LIST_FIRST(&kz.uk_zones); uzp != NULL; uzp =
381 		    LIST_NEXT(&uz, uz_link)) {
382 			ret = kread(kvm, uzp, &uz, sizeof(uz), 0);
383 			if (ret != 0) {
384 				free(ucp_array);
385 				_memstat_mtl_empty(list);
386 				list->mtl_error = ret;
387 				return (-1);
388 			}
389 			ret = kread(kvm, uzp, ucp_array,
390 			    sizeof(struct uma_cache) * (mp_maxid + 1),
391 			    offsetof(struct uma_zone, uz_cpu[0]));
392 			if (ret != 0) {
393 				free(ucp_array);
394 				_memstat_mtl_empty(list);
395 				list->mtl_error = ret;
396 				return (-1);
397 			}
398 			ret = kread_string(kvm, uz.uz_name, name,
399 			    MEMTYPE_MAXNAME);
400 			if (ret != 0) {
401 				free(ucp_array);
402 				_memstat_mtl_empty(list);
403 				list->mtl_error = ret;
404 				return (-1);
405 			}
406 			if (hint_dontsearch == 0) {
407 				mtp = memstat_mtl_find(list, ALLOCATOR_UMA,
408 				    name);
409 			} else
410 				mtp = NULL;
411 			if (mtp == NULL)
412 				mtp = _memstat_mt_allocate(list, ALLOCATOR_UMA,
413 				    name, mp_maxid + 1);
414 			if (mtp == NULL) {
415 				free(ucp_array);
416 				_memstat_mtl_empty(list);
417 				list->mtl_error = MEMSTAT_ERROR_NOMEMORY;
418 				return (-1);
419 			}
420 			/*
421 			 * Reset the statistics on a current node.
422 			 */
423 			_memstat_mt_reset_stats(mtp, mp_maxid + 1);
424 			mtp->mt_numallocs = kvm_counter_u64_fetch(kvm,
425 			    (unsigned long )uz.uz_allocs);
426 			mtp->mt_numfrees = kvm_counter_u64_fetch(kvm,
427 			    (unsigned long )uz.uz_frees);
428 			mtp->mt_failures = kvm_counter_u64_fetch(kvm,
429 			    (unsigned long )uz.uz_fails);
430 			mtp->mt_xdomain = kvm_counter_u64_fetch(kvm,
431 			    (unsigned long )uz.uz_xdomain);
432 			mtp->mt_sleeps = uz.uz_sleeps;
433 			/* See comment above in memstat_sysctl_uma(). */
434 			if (mtp->mt_numallocs < mtp->mt_numfrees)
435 				mtp->mt_numallocs = mtp->mt_numfrees;
436 
437 			if (kz.uk_flags & UMA_ZFLAG_INTERNAL)
438 				goto skip_percpu;
439 			for (i = 0; i < mp_maxid + 1; i++) {
440 				if (!CPU_ISSET(i, &all_cpus))
441 					continue;
442 				ucp = &ucp_array[i];
443 				mtp->mt_numallocs += ucp->uc_allocs;
444 				mtp->mt_numfrees += ucp->uc_frees;
445 
446 				mtp->mt_free += ucp->uc_allocbucket.ucb_cnt;
447 				mtp->mt_free += ucp->uc_freebucket.ucb_cnt;
448 				mtp->mt_free += ucp->uc_crossbucket.ucb_cnt;
449 			}
450 skip_percpu:
451 			mtp->mt_size = kz.uk_size;
452 			mtp->mt_rsize = kz.uk_rsize;
453 			mtp->mt_memalloced = mtp->mt_numallocs * mtp->mt_size;
454 			mtp->mt_memfreed = mtp->mt_numfrees * mtp->mt_size;
455 			mtp->mt_bytes = mtp->mt_memalloced - mtp->mt_memfreed;
456 			mtp->mt_countlimit = uz.uz_max_items;
457 			mtp->mt_byteslimit = mtp->mt_countlimit * mtp->mt_size;
458 			mtp->mt_count = mtp->mt_numallocs - mtp->mt_numfrees;
459 			for (i = 0; i < ndomains; i++) {
460 				ret = kread(kvm, ZDOM_GET(uzp, i), &uzd,
461 				    sizeof(uzd), 0);
462 				if (ret != 0)
463 					continue;
464 				for (ubp =
465 				    STAILQ_FIRST(&uzd.uzd_buckets);
466 				    ubp != NULL;
467 				    ubp = STAILQ_NEXT(&ub, ub_link)) {
468 					ret = kread(kvm, ubp, &ub,
469 					   sizeof(ub), 0);
470 					if (ret != 0)
471 						continue;
472 					mtp->mt_zonefree += ub.ub_cnt;
473 				}
474 			}
475 			if (!((kz.uk_flags & UMA_ZONE_SECONDARY) &&
476 			    LIST_FIRST(&kz.uk_zones) != uzp)) {
477 				kegfree = 0;
478 				for (i = 0; i < ndomains; i++) {
479 					ret = kread(kvm, &kzp->uk_domain[i],
480 					    &ukd, sizeof(ukd), 0);
481 					if (ret != 0)
482 						kegfree += ukd.ud_free_items;
483 				}
484 				mtp->mt_kegfree = kegfree;
485 				mtp->mt_free += mtp->mt_kegfree;
486 			}
487 			mtp->mt_free += mtp->mt_zonefree;
488 		}
489 	}
490 	free(ucp_array);
491 	return (0);
492 }
493