1 // SPDX-License-Identifier: CDDL-1.0
2 /*
3 * CDDL HEADER START
4 *
5 * The contents of this file are subject to the terms of the
6 * Common Development and Distribution License (the "License").
7 * You may not use this file except in compliance with the License.
8 *
9 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10 * or https://opensource.org/licenses/CDDL-1.0.
11 * See the License for the specific language governing permissions
12 * and limitations under the License.
13 *
14 * When distributing Covered Code, include this CDDL HEADER in each
15 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16 * If applicable, add the following below this CDDL HEADER, with the
17 * fields enclosed by brackets "[]" replaced with your own identifying
18 * information: Portions Copyright [yyyy] [name of copyright owner]
19 *
20 * CDDL HEADER END
21 */
22 /*
23 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Copyright (c) 2018, Joyent, Inc.
25 * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
26 * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
27 * Copyright 2017 Nexenta Systems, Inc. All rights reserved.
28 */
29
30 #include <sys/spa.h>
31 #include <sys/zio.h>
32 #include <sys/spa_impl.h>
33 #include <sys/zio_compress.h>
34 #include <sys/zio_checksum.h>
35 #include <sys/zfs_context.h>
36 #include <sys/arc.h>
37 #include <sys/zfs_refcount.h>
38 #include <sys/vdev.h>
39 #include <sys/vdev_trim.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dsl_pool.h>
42 #include <sys/multilist.h>
43 #include <sys/abd.h>
44 #include <sys/zil.h>
45 #include <sys/fm/fs/zfs.h>
46 #include <sys/shrinker.h>
47 #include <sys/vmsystm.h>
48 #include <sys/zpl.h>
49 #include <linux/page_compat.h>
50 #include <linux/notifier.h>
51 #include <linux/memory.h>
52 #include <linux/version.h>
53 #include <sys/callb.h>
54 #include <sys/kstat.h>
55 #include <sys/zthr.h>
56 #include <zfs_fletcher.h>
57 #include <sys/arc_impl.h>
58 #include <sys/trace_zfs.h>
59 #include <sys/aggsum.h>
60
61 /*
62 * This is a limit on how many pages the ARC shrinker makes available for
63 * eviction in response to one page allocation attempt. Note that in
64 * practice, the kernel's shrinker can ask us to evict up to about 4x this
65 * for one allocation attempt.
66 *
67 * For example a value of 10,000 (in practice, 160MB per allocation attempt
68 * with 4K pages) limits the amount of time spent attempting to reclaim ARC
69 * memory to less than 100ms per allocation attempt, even with a small
70 * average compressed block size of ~8KB.
71 *
72 * See also the comment in arc_shrinker_count().
73 * Set to 0 to disable limit.
74 */
75 static int zfs_arc_shrinker_limit = 0;
76
77 /*
78 * Relative cost of ARC eviction, AKA number of seeks needed to restore evicted
79 * page. Bigger values make ARC more precious and evictions smaller comparing
80 * to other kernel subsystems. Value of 4 means parity with page cache,
81 * according to my reading of kernel's do_shrink_slab() and other code.
82 */
83 static int zfs_arc_shrinker_seeks = DEFAULT_SEEKS;
84
85 #ifdef CONFIG_MEMORY_HOTPLUG
86 static struct notifier_block arc_hotplug_callback_mem_nb;
87 #endif
88
89 /*
90 * Return a default max arc size based on the amount of physical memory.
91 * This may be overridden by tuning the zfs_arc_max module parameter.
92 */
93 uint64_t
arc_default_max(uint64_t min,uint64_t allmem)94 arc_default_max(uint64_t min, uint64_t allmem)
95 {
96 uint64_t size;
97
98 if (allmem >= 1 << 30)
99 size = allmem - (1 << 30);
100 else
101 size = min;
102 return (MAX(allmem * 5 / 8, size));
103 }
104
105 /*
106 * Return maximum amount of memory that we could possibly use. Reduced
107 * to half of all memory in user space which is primarily used for testing.
108 */
109 uint64_t
arc_all_memory(void)110 arc_all_memory(void)
111 {
112 #ifdef CONFIG_HIGHMEM
113 return (ptob(zfs_totalram_pages - zfs_totalhigh_pages));
114 #else
115 return (ptob(zfs_totalram_pages));
116 #endif /* CONFIG_HIGHMEM */
117 }
118
119 /*
120 * Return the amount of memory that is considered free. In user space
121 * which is primarily used for testing we pretend that free memory ranges
122 * from 0-20% of all memory.
123 */
124 uint64_t
arc_free_memory(void)125 arc_free_memory(void)
126 {
127 #ifdef CONFIG_HIGHMEM
128 struct sysinfo si;
129 si_meminfo(&si);
130 return (ptob(si.freeram - si.freehigh));
131 #else
132 return (ptob(nr_free_pages() +
133 nr_inactive_file_pages()));
134 #endif /* CONFIG_HIGHMEM */
135 }
136
137 /*
138 * Return the amount of memory that can be consumed before reclaim will be
139 * needed. Positive if there is sufficient free memory, negative indicates
140 * the amount of memory that needs to be freed up.
141 */
142 int64_t
arc_available_memory(void)143 arc_available_memory(void)
144 {
145 return (arc_free_memory() - arc_sys_free);
146 }
147
148 static uint64_t
arc_evictable_memory(void)149 arc_evictable_memory(void)
150 {
151 int64_t asize = aggsum_value(&arc_sums.arcstat_size);
152 uint64_t arc_clean =
153 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) +
154 zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) +
155 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) +
156 zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
157 uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0);
158
159 /*
160 * Scale reported evictable memory in proportion to page cache, cap
161 * at specified min/max.
162 */
163 uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent;
164 min = MAX(arc_c_min, MIN(arc_c_max, min));
165
166 if (arc_dirty >= min)
167 return (arc_clean);
168
169 return (MAX((int64_t)asize - (int64_t)min, 0));
170 }
171
172 /*
173 * The _count() function returns the number of free-able objects.
174 * The _scan() function returns the number of objects that were freed.
175 */
176 static unsigned long
arc_shrinker_count(struct shrinker * shrink,struct shrink_control * sc)177 arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc)
178 {
179 /*
180 * The kernel's shrinker code may not understand how many pages the
181 * ARC's callback actually frees, so it may ask the ARC to shrink a
182 * lot for one page allocation. This is problematic because it may
183 * take a long time, thus delaying the page allocation, and because
184 * it may force the ARC to unnecessarily shrink very small.
185 *
186 * Therefore, we limit the amount of data that we say is evictable,
187 * which limits the amount that the shrinker will ask us to evict for
188 * one page allocation attempt.
189 *
190 * In practice, we may be asked to shrink 4x the limit to satisfy one
191 * page allocation, before the kernel's shrinker code gives up on us.
192 * When that happens, we rely on the kernel code to find the pages
193 * that we freed before invoking the OOM killer. This happens in
194 * __alloc_pages_slowpath(), which retries and finds the pages we
195 * freed when it calls get_page_from_freelist().
196 *
197 * See also the comment above zfs_arc_shrinker_limit.
198 */
199 int64_t can_free = btop(arc_evictable_memory());
200 if (current_is_kswapd() && zfs_arc_shrinker_limit)
201 can_free = MIN(can_free, zfs_arc_shrinker_limit);
202 return (can_free);
203 }
204
205 static unsigned long
arc_shrinker_scan(struct shrinker * shrink,struct shrink_control * sc)206 arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc)
207 {
208 /* The arc is considered warm once reclaim has occurred */
209 if (unlikely(arc_warm == B_FALSE))
210 arc_warm = B_TRUE;
211
212 /*
213 * We are experiencing memory pressure which the arc_evict_zthr was
214 * unable to keep up with. Set arc_no_grow to briefly pause ARC
215 * growth to avoid compounding the memory pressure.
216 */
217 arc_no_grow = B_TRUE;
218
219 /*
220 * Evict the requested number of pages by reducing arc_c and waiting
221 * for the requested amount of data to be evicted. To avoid deadlock
222 * do not wait for eviction if we may be called from ZFS itself (see
223 * kmem_flags_convert() removing __GFP_FS). It may cause excessive
224 * eviction later if many evictions are accumulated, but just skipping
225 * the eviction is not good either if most of memory is used by ARC.
226 */
227 uint64_t to_free = arc_reduce_target_size(ptob(sc->nr_to_scan));
228 if (sc->gfp_mask & __GFP_FS)
229 arc_wait_for_eviction(to_free, B_FALSE, B_FALSE);
230 if (current->reclaim_state != NULL)
231 #ifdef HAVE_RECLAIM_STATE_RECLAIMED
232 current->reclaim_state->reclaimed += btop(to_free);
233 #else
234 current->reclaim_state->reclaimed_slab += btop(to_free);
235 #endif
236
237 /*
238 * When direct reclaim is observed it usually indicates a rapid
239 * increase in memory pressure. This occurs because the kswapd
240 * threads were unable to asynchronously keep enough free memory
241 * available.
242 */
243 if (current_is_kswapd()) {
244 ARCSTAT_BUMP(arcstat_memory_indirect_count);
245 } else {
246 ARCSTAT_BUMP(arcstat_memory_direct_count);
247 }
248
249 return (btop(to_free));
250 }
251
252 static struct shrinker *arc_shrinker = NULL;
253
254 int
arc_memory_throttle(spa_t * spa,uint64_t reserve,uint64_t txg)255 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
256 {
257 uint64_t free_memory = arc_free_memory();
258
259 if (free_memory > arc_all_memory() * arc_lotsfree_percent / 100)
260 return (0);
261
262 if (txg > spa->spa_lowmem_last_txg) {
263 spa->spa_lowmem_last_txg = txg;
264 spa->spa_lowmem_page_load = 0;
265 }
266 /*
267 * If we are in pageout, we know that memory is already tight,
268 * the arc is already going to be evicting, so we just want to
269 * continue to let page writes occur as quickly as possible.
270 */
271 if (current_is_kswapd()) {
272 if (spa->spa_lowmem_page_load >
273 MAX(arc_sys_free / 4, free_memory) / 4) {
274 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
275 return (SET_ERROR(ERESTART));
276 }
277 /* Note: reserve is inflated, so we deflate */
278 atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
279 return (0);
280 } else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
281 /* memory is low, delay before restarting */
282 ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
283 DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
284 return (SET_ERROR(EAGAIN));
285 }
286 spa->spa_lowmem_page_load = 0;
287 return (0);
288 }
289
290 static void
arc_set_sys_free(uint64_t allmem)291 arc_set_sys_free(uint64_t allmem)
292 {
293 /*
294 * The ARC tries to keep at least this much memory available for the
295 * system. This gives the ARC time to shrink in response to memory
296 * pressure, before running completely out of memory and invoking the
297 * direct-reclaim ARC shrinker.
298 *
299 * This should be more than twice high_wmark_pages(), so that
300 * arc_wait_for_eviction() will wait until at least the
301 * high_wmark_pages() are free (see arc_evict_state_impl()).
302 *
303 * Note: If concurrent allocations consume these pages, there may
304 * still be insufficient free pages, and the OOM killer takes action.
305 *
306 * By setting arc_sys_free large enough, and having
307 * arc_wait_for_eviction() wait until there is at least arc_sys_free/2
308 * free memory, it is much less likely that concurrent allocations can
309 * consume all the memory that was evicted before checking for
310 * OOM.
311 *
312 * It's hard to iterate the zones from a linux kernel module, which
313 * makes it difficult to determine the watermark dynamically. Instead
314 * we compute the maximum high watermark for this system, based
315 * on the amount of memory, using the same method as the kernel uses
316 * to calculate its internal `min_free_kbytes` variable. See
317 * torvalds/linux@ee8eb9a5fe86 for the change in the upper clamp value
318 * from 64M to 256M.
319 */
320
321 /*
322 * Base wmark_low is 4 * the square root of Kbytes of RAM.
323 */
324 long wmark = int_sqrt(allmem / 1024 * 16) * 1024;
325
326 /*
327 * Clamp to between 128K and 256/64MB.
328 */
329 wmark = MAX(wmark, 128 * 1024);
330 #if LINUX_VERSION_CODE >= KERNEL_VERSION(5, 7, 0)
331 wmark = MIN(wmark, 256 * 1024 * 1024);
332 #else
333 wmark = MIN(wmark, 64 * 1024 * 1024);
334 #endif
335
336 /*
337 * watermark_boost can increase the wmark by up to 150%.
338 */
339 wmark += wmark * 150 / 100;
340
341 /*
342 * arc_sys_free needs to be more than 2x the watermark, because
343 * arc_wait_for_eviction() waits for half of arc_sys_free. Bump this up
344 * to 3x to ensure we're above it.
345 */
346 arc_sys_free = wmark * 3 + allmem / 32;
347 }
348
349 void
arc_lowmem_init(void)350 arc_lowmem_init(void)
351 {
352 uint64_t allmem = arc_all_memory();
353
354 /*
355 * Register a shrinker to support synchronous (direct) memory
356 * reclaim from the arc. This is done to prevent kswapd from
357 * swapping out pages when it is preferable to shrink the arc.
358 */
359 arc_shrinker = spl_register_shrinker("zfs-arc-shrinker",
360 arc_shrinker_count, arc_shrinker_scan, zfs_arc_shrinker_seeks);
361 VERIFY(arc_shrinker);
362
363 arc_set_sys_free(allmem);
364 }
365
366 void
arc_lowmem_fini(void)367 arc_lowmem_fini(void)
368 {
369 spl_unregister_shrinker(arc_shrinker);
370 arc_shrinker = NULL;
371 }
372
373 int
param_set_arc_u64(const char * buf,zfs_kernel_param_t * kp)374 param_set_arc_u64(const char *buf, zfs_kernel_param_t *kp)
375 {
376 int error;
377
378 error = spl_param_set_u64(buf, kp);
379 if (error < 0)
380 return (SET_ERROR(error));
381
382 arc_tuning_update(B_TRUE);
383
384 return (0);
385 }
386
387 int
param_set_arc_min(const char * buf,zfs_kernel_param_t * kp)388 param_set_arc_min(const char *buf, zfs_kernel_param_t *kp)
389 {
390 return (param_set_arc_u64(buf, kp));
391 }
392
393 int
param_set_arc_max(const char * buf,zfs_kernel_param_t * kp)394 param_set_arc_max(const char *buf, zfs_kernel_param_t *kp)
395 {
396 return (param_set_arc_u64(buf, kp));
397 }
398
399 int
param_set_arc_int(const char * buf,zfs_kernel_param_t * kp)400 param_set_arc_int(const char *buf, zfs_kernel_param_t *kp)
401 {
402 int error;
403
404 error = param_set_int(buf, kp);
405 if (error < 0)
406 return (SET_ERROR(error));
407
408 arc_tuning_update(B_TRUE);
409
410 return (0);
411 }
412
413 #ifdef CONFIG_MEMORY_HOTPLUG
414 static int
arc_hotplug_callback(struct notifier_block * self,unsigned long action,void * arg)415 arc_hotplug_callback(struct notifier_block *self, unsigned long action,
416 void *arg)
417 {
418 (void) self, (void) arg;
419 uint64_t allmem = arc_all_memory();
420 if (action != MEM_ONLINE)
421 return (NOTIFY_OK);
422
423 arc_set_limits(allmem);
424
425 #ifdef __LP64__
426 if (zfs_dirty_data_max_max == 0)
427 zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
428 allmem * zfs_dirty_data_max_max_percent / 100);
429 #else
430 if (zfs_dirty_data_max_max == 0)
431 zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
432 allmem * zfs_dirty_data_max_max_percent / 100);
433 #endif
434
435 arc_set_sys_free(allmem);
436 return (NOTIFY_OK);
437 }
438 #endif
439
440 void
arc_register_hotplug(void)441 arc_register_hotplug(void)
442 {
443 #ifdef CONFIG_MEMORY_HOTPLUG
444 arc_hotplug_callback_mem_nb.notifier_call = arc_hotplug_callback;
445 /* There is no significance to the value 100 */
446 arc_hotplug_callback_mem_nb.priority = 100;
447 register_memory_notifier(&arc_hotplug_callback_mem_nb);
448 #endif
449 }
450
451 void
arc_unregister_hotplug(void)452 arc_unregister_hotplug(void)
453 {
454 #ifdef CONFIG_MEMORY_HOTPLUG
455 unregister_memory_notifier(&arc_hotplug_callback_mem_nb);
456 #endif
457 }
458
459 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_limit, INT, ZMOD_RW,
460 "Limit on number of pages that ARC shrinker can reclaim at once");
461 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_seeks, INT, ZMOD_RD,
462 "Relative cost of ARC eviction vs other kernel subsystems");
463