2 * Copyright (C) 2007-2010 Lawrence Livermore National Security, LLC.
3 * Copyright (C) 2007 The Regents of the University of California.
4 * Produced at Lawrence Livermore National Laboratory (cf, DISCLAIMER).
5 * Written by Brian Behlendorf <behlendorf1@llnl.gov>.
8 * This file is part of the SPL, Solaris Porting Layer.
9 * For details, see <http://zfsonlinux.org/>.
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.
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
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/>.
25 #include <sys/debug.h>
26 #include <sys/sysmacros.h>
30 #include <linux/ratelimit.h>
33 * As a general rule kmem_alloc() allocations should be small, preferably
34 * just a few pages since they must by physically contiguous. Therefore, a
35 * rate limited warning will be printed to the console for any kmem_alloc()
36 * which exceeds a reasonable threshold.
38 * The default warning threshold is set to eight pages but capped at 32K to
39 * accommodate systems using large pages. This value was selected to be small
40 * enough to ensure the largest allocations are quickly noticed and fixed.
41 * But large enough to avoid logging any warnings when a allocation size is
42 * larger than optimal but not a serious concern. Since this value is tunable,
43 * developers are encouraged to set it lower when testing so any new largish
44 * allocations are quickly caught. These warnings may be disabled by setting
45 * the threshold to zero.
47 unsigned int spl_kmem_alloc_warn
= MAX(8 * PAGE_SIZE
, 32 * 1024);
48 module_param(spl_kmem_alloc_warn
, uint
, 0644);
49 MODULE_PARM_DESC(spl_kmem_alloc_warn
,
50 "Warning threshold in bytes for a kmem_alloc()");
51 EXPORT_SYMBOL(spl_kmem_alloc_warn
);
54 * Large kmem_alloc() allocations will fail if they exceed KMALLOC_MAX_SIZE.
55 * Allocations which are marginally smaller than this limit may succeed but
56 * should still be avoided due to the expense of locating a contiguous range
57 * of free pages. Therefore, a maximum kmem size with reasonable safely
58 * margin of 4x is set. Kmem_alloc() allocations larger than this maximum
59 * will quickly fail. Vmem_alloc() allocations less than or equal to this
60 * value will use kmalloc(), but shift to vmalloc() when exceeding this value.
62 unsigned int spl_kmem_alloc_max
= (KMALLOC_MAX_SIZE
>> 2);
63 module_param(spl_kmem_alloc_max
, uint
, 0644);
64 MODULE_PARM_DESC(spl_kmem_alloc_max
,
65 "Maximum size in bytes for a kmem_alloc()");
66 EXPORT_SYMBOL(spl_kmem_alloc_max
);
73 EXPORT_SYMBOL(kmem_debugging
);
76 kmem_vasprintf(const char *fmt
, va_list ap
)
83 ptr
= kvasprintf(kmem_flags_convert(KM_SLEEP
), fmt
, aq
);
85 } while (ptr
== NULL
);
89 EXPORT_SYMBOL(kmem_vasprintf
);
92 kmem_asprintf(const char *fmt
, ...)
99 ptr
= kvasprintf(kmem_flags_convert(KM_SLEEP
), fmt
, ap
);
101 } while (ptr
== NULL
);
105 EXPORT_SYMBOL(kmem_asprintf
);
108 __strdup(const char *str
, int flags
)
114 ptr
= kmalloc(n
+ 1, kmem_flags_convert(flags
));
116 memcpy(ptr
, str
, n
+ 1);
122 strdup(const char *str
)
124 return (__strdup(str
, KM_SLEEP
));
126 EXPORT_SYMBOL(strdup
);
133 EXPORT_SYMBOL(strfree
);
136 * Limit the number of large allocation stack traces dumped to not more than
137 * 5 every 60 seconds to prevent denial-of-service attacks from debug code.
139 DEFINE_RATELIMIT_STATE(kmem_alloc_ratelimit_state
, 60 * HZ
, 5);
142 * General purpose unified implementation of kmem_alloc(). It is an
143 * amalgamation of Linux and Illumos allocator design. It should never be
144 * exported to ensure that code using kmem_alloc()/kmem_zalloc() remains
145 * relatively portable. Consumers may only access this function through
146 * wrappers that enforce the common flags to ensure portability.
149 spl_kmem_alloc_impl(size_t size
, int flags
, int node
)
151 gfp_t lflags
= kmem_flags_convert(flags
);
156 * Log abnormally large allocations and rate limit the console output.
157 * Allocations larger than spl_kmem_alloc_warn should be performed
158 * through the vmem_alloc()/vmem_zalloc() interfaces.
160 if ((spl_kmem_alloc_warn
> 0) && (size
> spl_kmem_alloc_warn
) &&
161 !(flags
& KM_VMEM
) && __ratelimit(&kmem_alloc_ratelimit_state
)) {
163 "Large kmem_alloc(%lu, 0x%x), please file an issue at:\n"
164 "https://github.com/zfsonlinux/zfs/issues/new\n",
165 (unsigned long)size
, flags
);
170 * Use a loop because kmalloc_node() can fail when GFP_KERNEL is used
171 * unlike kmem_alloc() with KM_SLEEP on Illumos.
175 * Calling kmalloc_node() when the size >= spl_kmem_alloc_max
176 * is unsafe. This must fail for all for kmem_alloc() and
177 * kmem_zalloc() callers.
179 * For vmem_alloc() and vmem_zalloc() callers it is permissible
180 * to use __vmalloc(). However, in general use of __vmalloc()
181 * is strongly discouraged because a global lock must be
182 * acquired. Contention on this lock can significantly
183 * impact performance so frequently manipulating the virtual
184 * address space is strongly discouraged.
186 if ((size
> spl_kmem_alloc_max
) || use_vmem
) {
187 if (flags
& KM_VMEM
) {
188 ptr
= spl_vmalloc(size
, lflags
, PAGE_KERNEL
);
193 ptr
= kmalloc_node(size
, lflags
, node
);
196 if (likely(ptr
) || (flags
& KM_NOSLEEP
))
200 * For vmem_alloc() and vmem_zalloc() callers retry immediately
201 * using spl_vmalloc() which is unlikely to fail.
203 if ((flags
& KM_VMEM
) && (use_vmem
== 0)) {
208 if (unlikely(__ratelimit(&kmem_alloc_ratelimit_state
))) {
210 "Possible memory allocation deadlock: "
211 "size=%lu lflags=0x%x",
212 (unsigned long)size
, lflags
);
217 * Use cond_resched() instead of congestion_wait() to avoid
218 * deadlocking systems where there are no block devices.
227 spl_kmem_free_impl(const void *buf
, size_t size
)
229 if (is_vmalloc_addr(buf
))
236 * Memory allocation and accounting for kmem_* * style allocations. When
237 * DEBUG_KMEM is enabled the total memory allocated will be tracked and
238 * any memory leaked will be reported during module unload.
240 * ./configure --enable-debug-kmem
244 /* Shim layer memory accounting */
245 #ifdef HAVE_ATOMIC64_T
246 atomic64_t kmem_alloc_used
= ATOMIC64_INIT(0);
247 unsigned long long kmem_alloc_max
= 0;
248 #else /* HAVE_ATOMIC64_T */
249 atomic_t kmem_alloc_used
= ATOMIC_INIT(0);
250 unsigned long long kmem_alloc_max
= 0;
251 #endif /* HAVE_ATOMIC64_T */
253 EXPORT_SYMBOL(kmem_alloc_used
);
254 EXPORT_SYMBOL(kmem_alloc_max
);
257 spl_kmem_alloc_debug(size_t size
, int flags
, int node
)
261 ptr
= spl_kmem_alloc_impl(size
, flags
, node
);
263 kmem_alloc_used_add(size
);
264 if (unlikely(kmem_alloc_used_read() > kmem_alloc_max
))
265 kmem_alloc_max
= kmem_alloc_used_read();
272 spl_kmem_free_debug(const void *ptr
, size_t size
)
274 kmem_alloc_used_sub(size
);
275 spl_kmem_free_impl(ptr
, size
);
279 * When DEBUG_KMEM_TRACKING is enabled not only will total bytes be tracked
280 * but also the location of every alloc and free. When the SPL module is
281 * unloaded a list of all leaked addresses and where they were allocated
282 * will be dumped to the console. Enabling this feature has a significant
283 * impact on performance but it makes finding memory leaks straight forward.
285 * Not surprisingly with debugging enabled the xmem_locks are very highly
286 * contended particularly on xfree(). If we want to run with this detailed
287 * debugging enabled for anything other than debugging we need to minimize
288 * the contention by moving to a lock per xmem_table entry model.
290 * ./configure --enable-debug-kmem-tracking
292 #ifdef DEBUG_KMEM_TRACKING
294 #include <linux/hash.h>
295 #include <linux/ctype.h>
297 #define KMEM_HASH_BITS 10
298 #define KMEM_TABLE_SIZE (1 << KMEM_HASH_BITS)
300 typedef struct kmem_debug
{
301 struct hlist_node kd_hlist
; /* Hash node linkage */
302 struct list_head kd_list
; /* List of all allocations */
303 void *kd_addr
; /* Allocation pointer */
304 size_t kd_size
; /* Allocation size */
305 const char *kd_func
; /* Allocation function */
306 int kd_line
; /* Allocation line */
309 static spinlock_t kmem_lock
;
310 static struct hlist_head kmem_table
[KMEM_TABLE_SIZE
];
311 static struct list_head kmem_list
;
313 static kmem_debug_t
*
314 kmem_del_init(spinlock_t
*lock
, struct hlist_head
*table
,
315 int bits
, const void *addr
)
317 struct hlist_head
*head
;
318 struct hlist_node
*node
;
319 struct kmem_debug
*p
;
322 spin_lock_irqsave(lock
, flags
);
324 head
= &table
[hash_ptr((void *)addr
, bits
)];
325 hlist_for_each(node
, head
) {
326 p
= list_entry(node
, struct kmem_debug
, kd_hlist
);
327 if (p
->kd_addr
== addr
) {
328 hlist_del_init(&p
->kd_hlist
);
329 list_del_init(&p
->kd_list
);
330 spin_unlock_irqrestore(lock
, flags
);
335 spin_unlock_irqrestore(lock
, flags
);
341 spl_kmem_alloc_track(size_t size
, int flags
,
342 const char *func
, int line
, int node
)
346 unsigned long irq_flags
;
348 dptr
= kmalloc(sizeof (kmem_debug_t
), kmem_flags_convert(flags
));
352 dptr
->kd_func
= __strdup(func
, flags
);
353 if (dptr
->kd_func
== NULL
) {
358 ptr
= spl_kmem_alloc_debug(size
, flags
, node
);
360 kfree(dptr
->kd_func
);
365 INIT_HLIST_NODE(&dptr
->kd_hlist
);
366 INIT_LIST_HEAD(&dptr
->kd_list
);
369 dptr
->kd_size
= size
;
370 dptr
->kd_line
= line
;
372 spin_lock_irqsave(&kmem_lock
, irq_flags
);
373 hlist_add_head(&dptr
->kd_hlist
,
374 &kmem_table
[hash_ptr(ptr
, KMEM_HASH_BITS
)]);
375 list_add_tail(&dptr
->kd_list
, &kmem_list
);
376 spin_unlock_irqrestore(&kmem_lock
, irq_flags
);
382 spl_kmem_free_track(const void *ptr
, size_t size
)
386 /* Must exist in hash due to kmem_alloc() */
387 dptr
= kmem_del_init(&kmem_lock
, kmem_table
, KMEM_HASH_BITS
, ptr
);
388 ASSERT3P(dptr
, !=, NULL
);
389 ASSERT3S(dptr
->kd_size
, ==, size
);
391 kfree(dptr
->kd_func
);
394 spl_kmem_free_debug(ptr
, size
);
396 #endif /* DEBUG_KMEM_TRACKING */
397 #endif /* DEBUG_KMEM */
400 * Public kmem_alloc(), kmem_zalloc() and kmem_free() interfaces.
403 spl_kmem_alloc(size_t size
, int flags
, const char *func
, int line
)
405 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
407 #if !defined(DEBUG_KMEM)
408 return (spl_kmem_alloc_impl(size
, flags
, NUMA_NO_NODE
));
409 #elif !defined(DEBUG_KMEM_TRACKING)
410 return (spl_kmem_alloc_debug(size
, flags
, NUMA_NO_NODE
));
412 return (spl_kmem_alloc_track(size
, flags
, func
, line
, NUMA_NO_NODE
));
415 EXPORT_SYMBOL(spl_kmem_alloc
);
418 spl_kmem_zalloc(size_t size
, int flags
, const char *func
, int line
)
420 ASSERT0(flags
& ~KM_PUBLIC_MASK
);
424 #if !defined(DEBUG_KMEM)
425 return (spl_kmem_alloc_impl(size
, flags
, NUMA_NO_NODE
));
426 #elif !defined(DEBUG_KMEM_TRACKING)
427 return (spl_kmem_alloc_debug(size
, flags
, NUMA_NO_NODE
));
429 return (spl_kmem_alloc_track(size
, flags
, func
, line
, NUMA_NO_NODE
));
432 EXPORT_SYMBOL(spl_kmem_zalloc
);
435 spl_kmem_free(const void *buf
, size_t size
)
437 #if !defined(DEBUG_KMEM)
438 return (spl_kmem_free_impl(buf
, size
));
439 #elif !defined(DEBUG_KMEM_TRACKING)
440 return (spl_kmem_free_debug(buf
, size
));
442 return (spl_kmem_free_track(buf
, size
));
445 EXPORT_SYMBOL(spl_kmem_free
);
447 #if defined(DEBUG_KMEM) && defined(DEBUG_KMEM_TRACKING)
449 spl_sprintf_addr(kmem_debug_t
*kd
, char *str
, int len
, int min
)
451 int size
= ((len
- 1) < kd
->kd_size
) ? (len
- 1) : kd
->kd_size
;
454 ASSERT(str
!= NULL
&& len
>= 17);
458 * Check for a fully printable string, and while we are at
459 * it place the printable characters in the passed buffer.
461 for (i
= 0; i
< size
; i
++) {
462 str
[i
] = ((char *)(kd
->kd_addr
))[i
];
463 if (isprint(str
[i
])) {
467 * Minimum number of printable characters found
468 * to make it worthwhile to print this as ascii.
479 sprintf(str
, "%02x%02x%02x%02x%02x%02x%02x%02x",
480 *((uint8_t *)kd
->kd_addr
),
481 *((uint8_t *)kd
->kd_addr
+ 2),
482 *((uint8_t *)kd
->kd_addr
+ 4),
483 *((uint8_t *)kd
->kd_addr
+ 6),
484 *((uint8_t *)kd
->kd_addr
+ 8),
485 *((uint8_t *)kd
->kd_addr
+ 10),
486 *((uint8_t *)kd
->kd_addr
+ 12),
487 *((uint8_t *)kd
->kd_addr
+ 14));
494 spl_kmem_init_tracking(struct list_head
*list
, spinlock_t
*lock
, int size
)
498 spin_lock_init(lock
);
499 INIT_LIST_HEAD(list
);
501 for (i
= 0; i
< size
; i
++)
502 INIT_HLIST_HEAD(&kmem_table
[i
]);
508 spl_kmem_fini_tracking(struct list_head
*list
, spinlock_t
*lock
)
514 spin_lock_irqsave(lock
, flags
);
515 if (!list_empty(list
))
516 printk(KERN_WARNING
"%-16s %-5s %-16s %s:%s\n", "address",
517 "size", "data", "func", "line");
519 list_for_each_entry(kd
, list
, kd_list
)
520 printk(KERN_WARNING
"%p %-5d %-16s %s:%d\n", kd
->kd_addr
,
521 (int)kd
->kd_size
, spl_sprintf_addr(kd
, str
, 17, 8),
522 kd
->kd_func
, kd
->kd_line
);
524 spin_unlock_irqrestore(lock
, flags
);
526 #endif /* DEBUG_KMEM && DEBUG_KMEM_TRACKING */
532 kmem_alloc_used_set(0);
534 #ifdef DEBUG_KMEM_TRACKING
535 spl_kmem_init_tracking(&kmem_list
, &kmem_lock
, KMEM_TABLE_SIZE
);
536 #endif /* DEBUG_KMEM_TRACKING */
537 #endif /* DEBUG_KMEM */
547 * Display all unreclaimed memory addresses, including the
548 * allocation size and the first few bytes of what's located
549 * at that address to aid in debugging. Performance is not
550 * a serious concern here since it is module unload time.
552 if (kmem_alloc_used_read() != 0)
553 printk(KERN_WARNING
"kmem leaked %ld/%llu bytes\n",
554 (unsigned long)kmem_alloc_used_read(), kmem_alloc_max
);
556 #ifdef DEBUG_KMEM_TRACKING
557 spl_kmem_fini_tracking(&kmem_list
, &kmem_lock
);
558 #endif /* DEBUG_KMEM_TRACKING */
559 #endif /* DEBUG_KMEM */