1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
4 * Copyright (C) 2007 The Regents of the University of California.
5 * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
6 * Written by Brian Behlendorf <behlendorf1@llnl.gov>.
7 * UCRL-CODE-235197
8 *
9 * This file is part of the SPL, Solaris Porting Layer.
10 *
11 * The SPL is free software; you can redistribute it and/or modify it
12 * under the terms of the GNU General Public License as published by the
13 * Free Software Foundation; either version 2 of the License, or (at your
14 * option) any later version.
15 *
16 * The SPL is distributed in the hope that it will be useful, but WITHOUT
17 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
18 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 * for more details.
20 *
21 * You should have received a copy of the GNU General Public License along
22 * with the SPL. If not, see <http://www.gnu.org/licenses/>.
23 */
24
25 #include <sys/debug.h>
26 #include <sys/sysmacros.h>
27 #include <sys/kmem.h>
28 #include <sys/vmem.h>
29
30 /*
31 * As a general rule kmem_alloc() allocations should be small, preferably
32 * just a few pages since they must by physically contiguous. Therefore, a
33 * rate limited warning will be printed to the console for any kmem_alloc()
34 * which exceeds a reasonable threshold.
35 *
36 * The default warning threshold is set to sixteen pages but capped at 64K to
37 * accommodate systems using large pages. This value was selected to be small
38 * enough to ensure the largest allocations are quickly noticed and fixed.
39 * But large enough to avoid logging any warnings when a allocation size is
40 * larger than optimal but not a serious concern. Since this value is tunable,
41 * developers are encouraged to set it lower when testing so any new largish
42 * allocations are quickly caught. These warnings may be disabled by setting
43 * the threshold to zero.
44 */
45 unsigned int spl_kmem_alloc_warn = MIN(16 * PAGE_SIZE, 64 * 1024);
46 module_param(spl_kmem_alloc_warn, uint, 0644);
47 MODULE_PARM_DESC(spl_kmem_alloc_warn,
48 "Warning threshold in bytes for a kmem_alloc()");
49 EXPORT_SYMBOL(spl_kmem_alloc_warn);
50
51 /*
52 * Large kmem_alloc() allocations will fail if they exceed KMALLOC_MAX_SIZE.
53 * Allocations which are marginally smaller than this limit may succeed but
54 * should still be avoided due to the expense of locating a contiguous range
55 * of free pages. Therefore, a maximum kmem size with reasonable safely
56 * margin of 4x is set. Kmem_alloc() allocations larger than this maximum
57 * will quickly fail. Vmem_alloc() allocations less than or equal to this
58 * value will use kmalloc(), but shift to vmalloc() when exceeding this value.
59 */
60 unsigned int spl_kmem_alloc_max = (KMALLOC_MAX_SIZE >> 2);
61 module_param(spl_kmem_alloc_max, uint, 0644);
62 MODULE_PARM_DESC(spl_kmem_alloc_max,
63 "Maximum size in bytes for a kmem_alloc()");
64 EXPORT_SYMBOL(spl_kmem_alloc_max);
65
66 int
kmem_debugging(void)67 kmem_debugging(void)
68 {
69 return (0);
70 }
71 EXPORT_SYMBOL(kmem_debugging);
72
73 char *
kmem_vasprintf(const char * fmt,va_list ap)74 kmem_vasprintf(const char *fmt, va_list ap)
75 {
76 va_list aq;
77 char *ptr;
78
79 do {
80 va_copy(aq, ap);
81 ptr = kvasprintf(kmem_flags_convert(KM_SLEEP), fmt, aq);
82 va_end(aq);
83 } while (ptr == NULL);
84
85 return (ptr);
86 }
87 EXPORT_SYMBOL(kmem_vasprintf);
88
89 char *
kmem_asprintf(const char * fmt,...)90 kmem_asprintf(const char *fmt, ...)
91 {
92 va_list ap;
93 char *ptr;
94
95 do {
96 va_start(ap, fmt);
97 ptr = kvasprintf(kmem_flags_convert(KM_SLEEP), fmt, ap);
98 va_end(ap);
99 } while (ptr == NULL);
100
101 return (ptr);
102 }
103 EXPORT_SYMBOL(kmem_asprintf);
104
105 static char *
__strdup(const char * str,int flags)106 __strdup(const char *str, int flags)
107 {
108 char *ptr;
109 int n;
110
111 n = strlen(str);
112 ptr = kmalloc(n + 1, kmem_flags_convert(flags));
113 if (ptr)
114 memcpy(ptr, str, n + 1);
115
116 return (ptr);
117 }
118
119 char *
kmem_strdup(const char * str)120 kmem_strdup(const char *str)
121 {
122 return (__strdup(str, KM_SLEEP));
123 }
124 EXPORT_SYMBOL(kmem_strdup);
125
126 void
kmem_strfree(char * str)127 kmem_strfree(char *str)
128 {
129 kfree(str);
130 }
131 EXPORT_SYMBOL(kmem_strfree);
132
133 void *
spl_kvmalloc(size_t size,gfp_t lflags)134 spl_kvmalloc(size_t size, gfp_t lflags)
135 {
136 /*
137 * GFP_KERNEL allocations can safely use kvmalloc which may
138 * improve performance by avoiding a) high latency caused by
139 * vmalloc's on-access allocation, b) performance loss due to
140 * MMU memory address mapping and c) vmalloc locking overhead.
141 * This has the side-effect that the slab statistics will
142 * incorrectly report this as a vmem allocation, but that is
143 * purely cosmetic.
144 */
145 if ((lflags & GFP_KERNEL) == GFP_KERNEL)
146 return (kvmalloc(size, lflags));
147
148 gfp_t kmalloc_lflags = lflags;
149
150 if (size > PAGE_SIZE) {
151 /*
152 * We need to set __GFP_NOWARN here since spl_kvmalloc is not
153 * only called by spl_kmem_alloc_impl but can be called
154 * directly with custom lflags, too. In that case
155 * kmem_flags_convert does not get called, which would
156 * implicitly set __GFP_NOWARN.
157 */
158 kmalloc_lflags |= __GFP_NOWARN;
159
160 /*
161 * N.B. __GFP_RETRY_MAYFAIL is supported only for large
162 * e (>32kB) allocations.
163 *
164 * We have to override __GFP_RETRY_MAYFAIL by __GFP_NORETRY
165 * for !costly requests because there is no other way to tell
166 * the allocator that we want to fail rather than retry
167 * endlessly.
168 */
169 if (!(kmalloc_lflags & __GFP_RETRY_MAYFAIL) ||
170 (size <= PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER)) {
171 kmalloc_lflags |= __GFP_NORETRY;
172 }
173 }
174
175 /*
176 * We first try kmalloc - even for big sizes - and fall back to
177 * spl_vmalloc if that fails.
178 *
179 * For non-__GFP-RECLAIM allocations we always stick to
180 * kmalloc_node, and fail when kmalloc is not successful (returns
181 * NULL).
182 * We cannot fall back to spl_vmalloc in this case because spl_vmalloc
183 * internally uses GPF_KERNEL allocations.
184 */
185 void *ptr = kmalloc_node(size, kmalloc_lflags, NUMA_NO_NODE);
186 if (ptr || size <= PAGE_SIZE ||
187 (lflags & __GFP_RECLAIM) != __GFP_RECLAIM) {
188 return (ptr);
189 }
190
191 return (spl_vmalloc(size, lflags | __GFP_HIGHMEM));
192 }
193
194 /*
195 * General purpose unified implementation of kmem_alloc(). It is an
196 * amalgamation of Linux and Illumos allocator design. It should never be
197 * exported to ensure that code using kmem_alloc()/kmem_zalloc() remains
198 * relatively portable. Consumers may only access this function through
199 * wrappers that enforce the common flags to ensure portability.
200 */
201 inline void *
spl_kmem_alloc_impl(size_t size,int flags,int node)202 spl_kmem_alloc_impl(size_t size, int flags, int node)
203 {
204 gfp_t lflags = kmem_flags_convert(flags);
205 void *ptr;
206
207 /*
208 * Log abnormally large allocations and rate limit the console output.
209 * Allocations larger than spl_kmem_alloc_warn should be performed
210 * through the vmem_alloc()/vmem_zalloc() interfaces.
211 */
212 if ((spl_kmem_alloc_warn > 0) && (size > spl_kmem_alloc_warn) &&
213 !(flags & KM_VMEM)) {
214 printk(KERN_WARNING
215 "Large kmem_alloc(%lu, 0x%x), please file an issue at:\n"
216 "https://github.com/openzfs/zfs/issues/new\n",
217 (unsigned long)size, flags);
218 dump_stack();
219 }
220
221 /*
222 * Use a loop because kmalloc_node() can fail when GFP_KERNEL is used
223 * unlike kmem_alloc() with KM_SLEEP on Illumos.
224 */
225 do {
226 /*
227 * Calling kmalloc_node() when the size >= spl_kmem_alloc_max
228 * is unsafe. This must fail for all for kmem_alloc() and
229 * kmem_zalloc() callers.
230 *
231 * For vmem_alloc() and vmem_zalloc() callers it is permissible
232 * to use spl_vmalloc(). However, in general use of
233 * spl_vmalloc() is strongly discouraged because a global lock
234 * must be acquired. Contention on this lock can significantly
235 * impact performance so frequently manipulating the virtual
236 * address space is strongly discouraged.
237 */
238 if (size > spl_kmem_alloc_max) {
239 if (flags & KM_VMEM) {
240 ptr = spl_vmalloc(size, lflags | __GFP_HIGHMEM);
241 } else {
242 return (NULL);
243 }
244 } else {
245 /*
246 * We use kmalloc when doing kmem_alloc(KM_NOSLEEP),
247 * because kvmalloc/vmalloc may sleep. We also use
248 * kmalloc on systems with limited kernel VA space (e.g.
249 * 32-bit), which have HIGHMEM. Otherwise we use
250 * kvmalloc, which tries to get contiguous physical
251 * memory (fast, like kmalloc) and falls back on using
252 * virtual memory to stitch together pages (slow, like
253 * vmalloc).
254 */
255 #ifdef CONFIG_HIGHMEM
256 if (flags & KM_VMEM) {
257 #else
258 if ((flags & KM_VMEM) || !(flags & KM_NOSLEEP)) {
259 #endif
260 ptr = spl_kvmalloc(size, lflags);
261 } else {
262 ptr = kmalloc_node(size, lflags, node);
263 }
264 }
265
266 if (likely(ptr) || (flags & KM_NOSLEEP))
267 return (ptr);
268
269 /*
270 * Try hard to satisfy the allocation. However, when progress
271 * cannot be made, the allocation is allowed to fail.
272 */
273 if ((lflags & GFP_KERNEL) == GFP_KERNEL)
274 lflags |= __GFP_RETRY_MAYFAIL;
275
276 /*
277 * Use cond_resched() instead of congestion_wait() to avoid
278 * deadlocking systems where there are no block devices.
279 */
280 cond_resched();
281 } while (1);
282
283 return (NULL);
284 }
285
286 inline void
287 spl_kmem_free_impl(const void *buf, size_t size)
288 {
289 if (is_vmalloc_addr(buf))
290 vfree(buf);
291 else
292 kfree(buf);
293 }
294
295 /*
296 * Memory allocation and accounting for kmem_* * style allocations. When
297 * DEBUG_KMEM is enabled the total memory allocated will be tracked and
298 * any memory leaked will be reported during module unload.
299 *
300 * ./configure --enable-debug-kmem
301 */
302 #ifdef DEBUG_KMEM
303
304 /* Shim layer memory accounting */
305 atomic64_t kmem_alloc_used = ATOMIC64_INIT(0);
306 uint64_t kmem_alloc_max = 0;
307
308 EXPORT_SYMBOL(kmem_alloc_used);
309 EXPORT_SYMBOL(kmem_alloc_max);
310
311 inline void *
312 spl_kmem_alloc_debug(size_t size, int flags, int node)
313 {
314 void *ptr;
315
316 ptr = spl_kmem_alloc_impl(size, flags, node);
317 if (ptr) {
318 atomic64_add(size, &kmem_alloc_used);
319 if (unlikely(atomic64_read(&kmem_alloc_used) > kmem_alloc_max))
320 kmem_alloc_max = atomic64_read(&kmem_alloc_used);
321 }
322
323 return (ptr);
324 }
325
326 inline void
327 spl_kmem_free_debug(const void *ptr, size_t size)
328 {
329 atomic64_sub(size, &kmem_alloc_used);
330 spl_kmem_free_impl(ptr, size);
331 }
332
333 /*
334 * When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked
335 * but also the location of every alloc and free. When the SPL module is
336 * unloaded a list of all leaked addresses and where they were allocated
337 * will be dumped to the console. Enabling this feature has a significant
338 * impact on performance but it makes finding memory leaks straight forward.
339 *
340 * Not surprisingly with debugging enabled the xmem_locks are very highly
341 * contended particularly on xfree(). If we want to run with this detailed
342 * debugging enabled for anything other than debugging we need to minimize
343 * the contention by moving to a lock per xmem_table entry model.
344 *
345 * ./configure --enable-debug-kmem-tracking
346 */
347 #ifdef DEBUG_KMEM_TRACKING
348
349 #include <linux/hash.h>
350 #include <linux/ctype.h>
351
352 #define KMEM_HASH_BITS 10
353 #define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS)
354
355 typedef struct kmem_debug {
356 struct hlist_node kd_hlist; /* Hash node linkage */
357 struct list_head kd_list; /* List of all allocations */
358 void *kd_addr; /* Allocation pointer */
359 size_t kd_size; /* Allocation size */
360 const char *kd_func; /* Allocation function */
361 int kd_line; /* Allocation line */
362 } kmem_debug_t;
363
364 static spinlock_t kmem_lock;
365 static struct hlist_head kmem_table[KMEM_TABLE_SIZE];
366 static struct list_head kmem_list;
367
368 static kmem_debug_t *
369 kmem_del_init(spinlock_t *lock, struct hlist_head *table,
370 int bits, const void *addr)
371 {
372 struct hlist_head *head;
373 struct hlist_node *node = NULL;
374 struct kmem_debug *p;
375 unsigned long flags;
376
377 spin_lock_irqsave(lock, flags);
378
379 head = &table[hash_ptr((void *)addr, bits)];
380 hlist_for_each(node, head) {
381 p = list_entry(node, struct kmem_debug, kd_hlist);
382 if (p->kd_addr == addr) {
383 hlist_del_init(&p->kd_hlist);
384 list_del_init(&p->kd_list);
385 spin_unlock_irqrestore(lock, flags);
386 return (p);
387 }
388 }
389
390 spin_unlock_irqrestore(lock, flags);
391
392 return (NULL);
393 }
394
395 inline void *
396 spl_kmem_alloc_track(size_t size, int flags,
397 const char *func, int line, int node)
398 {
399 void *ptr = NULL;
400 kmem_debug_t *dptr;
401 unsigned long irq_flags;
402
403 dptr = kmalloc(sizeof (kmem_debug_t), kmem_flags_convert(flags));
404 if (dptr == NULL)
405 return (NULL);
406
407 dptr->kd_func = __strdup(func, flags);
408 if (dptr->kd_func == NULL) {
409 kfree(dptr);
410 return (NULL);
411 }
412
413 ptr = spl_kmem_alloc_debug(size, flags, node);
414 if (ptr == NULL) {
415 kfree(dptr->kd_func);
416 kfree(dptr);
417 return (NULL);
418 }
419
420 INIT_HLIST_NODE(&dptr->kd_hlist);
421 INIT_LIST_HEAD(&dptr->kd_list);
422
423 dptr->kd_addr = ptr;
424 dptr->kd_size = size;
425 dptr->kd_line = line;
426
427 spin_lock_irqsave(&kmem_lock, irq_flags);
428 hlist_add_head(&dptr->kd_hlist,
429 &kmem_table[hash_ptr(ptr, KMEM_HASH_BITS)]);
430 list_add_tail(&dptr->kd_list, &kmem_list);
431 spin_unlock_irqrestore(&kmem_lock, irq_flags);
432
433 return (ptr);
434 }
435
436 inline void
437 spl_kmem_free_track(const void *ptr, size_t size)
438 {
439 kmem_debug_t *dptr;
440
441 /* Ignore NULL pointer since we haven't tracked it at all */
442 if (ptr == NULL)
443 return;
444
445 /* Must exist in hash due to kmem_alloc() */
446 dptr = kmem_del_init(&kmem_lock, kmem_table, KMEM_HASH_BITS, ptr);
447 ASSERT3P(dptr, !=, NULL);
448 ASSERT3S(dptr->kd_size, ==, size);
449
450 kfree(dptr->kd_func);
451 kfree(dptr);
452
453 spl_kmem_free_debug(ptr, size);
454 }
455 #endif /* DEBUG_KMEM_TRACKING */
456 #endif /* DEBUG_KMEM */
457
458 /*
459 * Public kmem_alloc(), kmem_zalloc() and kmem_free() interfaces.
460 */
461 void *
462 spl_kmem_alloc(size_t size, int flags, const char *func, int line)
463 {
464 ASSERT0(flags & ~KM_PUBLIC_MASK);
465
466 #if !defined(DEBUG_KMEM)
467 return (spl_kmem_alloc_impl(size, flags, NUMA_NO_NODE));
468 #elif !defined(DEBUG_KMEM_TRACKING)
469 return (spl_kmem_alloc_debug(size, flags, NUMA_NO_NODE));
470 #else
471 return (spl_kmem_alloc_track(size, flags, func, line, NUMA_NO_NODE));
472 #endif
473 }
474 EXPORT_SYMBOL(spl_kmem_alloc);
475
476 void *
477 spl_kmem_zalloc(size_t size, int flags, const char *func, int line)
478 {
479 ASSERT0(flags & ~KM_PUBLIC_MASK);
480
481 flags |= KM_ZERO;
482
483 #if !defined(DEBUG_KMEM)
484 return (spl_kmem_alloc_impl(size, flags, NUMA_NO_NODE));
485 #elif !defined(DEBUG_KMEM_TRACKING)
486 return (spl_kmem_alloc_debug(size, flags, NUMA_NO_NODE));
487 #else
488 return (spl_kmem_alloc_track(size, flags, func, line, NUMA_NO_NODE));
489 #endif
490 }
491 EXPORT_SYMBOL(spl_kmem_zalloc);
492
493 void
494 spl_kmem_free(const void *buf, size_t size)
495 {
496 #if !defined(DEBUG_KMEM)
497 return (spl_kmem_free_impl(buf, size));
498 #elif !defined(DEBUG_KMEM_TRACKING)
499 return (spl_kmem_free_debug(buf, size));
500 #else
501 return (spl_kmem_free_track(buf, size));
502 #endif
503 }
504 EXPORT_SYMBOL(spl_kmem_free);
505
506 #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING)
507 static char *
508 spl_sprintf_addr(kmem_debug_t *kd, char *str, int len, int min)
509 {
510 int size = ((len - 1) < kd->kd_size) ? (len - 1) : kd->kd_size;
511 int i, flag = 1;
512
513 ASSERT(str != NULL && len >= 17);
514 memset(str, 0, len);
515
516 /*
517 * Check for a fully printable string, and while we are at
518 * it place the printable characters in the passed buffer.
519 */
520 for (i = 0; i < size; i++) {
521 str[i] = ((char *)(kd->kd_addr))[i];
522 if (isprint(str[i])) {
523 continue;
524 } else {
525 /*
526 * Minimum number of printable characters found
527 * to make it worthwhile to print this as ascii.
528 */
529 if (i > min)
530 break;
531
532 flag = 0;
533 break;
534 }
535 }
536
537 if (!flag) {
538 sprintf(str, "%02x%02x%02x%02x%02x%02x%02x%02x",
539 *((uint8_t *)kd->kd_addr),
540 *((uint8_t *)kd->kd_addr + 2),
541 *((uint8_t *)kd->kd_addr + 4),
542 *((uint8_t *)kd->kd_addr + 6),
543 *((uint8_t *)kd->kd_addr + 8),
544 *((uint8_t *)kd->kd_addr + 10),
545 *((uint8_t *)kd->kd_addr + 12),
546 *((uint8_t *)kd->kd_addr + 14));
547 }
548
549 return (str);
550 }
551
552 static int
553 spl_kmem_init_tracking(struct list_head *list, spinlock_t *lock, int size)
554 {
555 int i;
556
557 spin_lock_init(lock);
558 INIT_LIST_HEAD(list);
559
560 for (i = 0; i < size; i++)
561 INIT_HLIST_HEAD(&kmem_table[i]);
562
563 return (0);
564 }
565
566 static void
567 spl_kmem_fini_tracking(struct list_head *list, spinlock_t *lock)
568 {
569 unsigned long flags;
570 kmem_debug_t *kd = NULL;
571 char str[17];
572
573 spin_lock_irqsave(lock, flags);
574 if (!list_empty(list))
575 printk(KERN_WARNING "%-16s %-5s %-16s %s:%s\n", "address",
576 "size", "data", "func", "line");
577
578 list_for_each_entry(kd, list, kd_list) {
579 printk(KERN_WARNING "%p %-5d %-16s %s:%d\n", kd->kd_addr,
580 (int)kd->kd_size, spl_sprintf_addr(kd, str, 17, 8),
581 kd->kd_func, kd->kd_line);
582 }
583
584 spin_unlock_irqrestore(lock, flags);
585 }
586 #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
587
588 int
589 spl_kmem_init(void)
590 {
591
592 #ifdef DEBUG_KMEM
593 atomic64_set(&kmem_alloc_used, 0);
594
595
596
597 #ifdef DEBUG_KMEM_TRACKING
598 spl_kmem_init_tracking(&kmem_list, &kmem_lock, KMEM_TABLE_SIZE);
599 #endif /* DEBUG_KMEM_TRACKING */
600 #endif /* DEBUG_KMEM */
601
602 return (0);
603 }
604
605 void
606 spl_kmem_fini(void)
607 {
608 #ifdef DEBUG_KMEM
609 /*
610 * Display all unreclaimed memory addresses, including the
611 * allocation size and the first few bytes of what's located
612 * at that address to aid in debugging. Performance is not
613 * a serious concern here since it is module unload time.
614 */
615 if (atomic64_read(&kmem_alloc_used) != 0)
616 printk(KERN_WARNING "kmem leaked %ld/%llu bytes\n",
617 (unsigned long)atomic64_read(&kmem_alloc_used),
618 kmem_alloc_max);
619
620 #ifdef DEBUG_KMEM_TRACKING
621 spl_kmem_fini_tracking(&kmem_list, &kmem_lock);
622 #endif /* DEBUG_KMEM_TRACKING */
623 #endif /* DEBUG_KMEM */
624 }
625