xref: /freebsd/sys/contrib/openzfs/module/os/linux/zfs/arc_os.c (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
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 (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2018, Joyent, Inc.
24  * Copyright (c) 2011, 2019 by Delphix. All rights reserved.
25  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
26  * Copyright 2017 Nexenta Systems, Inc.  All rights reserved.
27  */
28 
29 #include <sys/spa.h>
30 #include <sys/zio.h>
31 #include <sys/spa_impl.h>
32 #include <sys/zio_compress.h>
33 #include <sys/zio_checksum.h>
34 #include <sys/zfs_context.h>
35 #include <sys/arc.h>
36 #include <sys/zfs_refcount.h>
37 #include <sys/vdev.h>
38 #include <sys/vdev_trim.h>
39 #include <sys/vdev_impl.h>
40 #include <sys/dsl_pool.h>
41 #include <sys/zio_checksum.h>
42 #include <sys/multilist.h>
43 #include <sys/abd.h>
44 #include <sys/zil.h>
45 #include <sys/fm/fs/zfs.h>
46 #ifdef _KERNEL
47 #include <sys/shrinker.h>
48 #include <sys/vmsystm.h>
49 #include <sys/zpl.h>
50 #include <linux/page_compat.h>
51 #include <linux/notifier.h>
52 #include <linux/memory.h>
53 #endif
54 #include <sys/callb.h>
55 #include <sys/kstat.h>
56 #include <sys/zthr.h>
57 #include <zfs_fletcher.h>
58 #include <sys/arc_impl.h>
59 #include <sys/trace_zfs.h>
60 #include <sys/aggsum.h>
61 
62 /*
63  * This is a limit on how many pages the ARC shrinker makes available for
64  * eviction in response to one page allocation attempt.  Note that in
65  * practice, the kernel's shrinker can ask us to evict up to about 4x this
66  * for one allocation attempt.
67  *
68  * The default limit of 10,000 (in practice, 160MB per allocation attempt
69  * with 4K pages) limits the amount of time spent attempting to reclaim ARC
70  * memory to less than 100ms per allocation attempt, even with a small
71  * average compressed block size of ~8KB.
72  *
73  * See also the comment in arc_shrinker_count().
74  * Set to 0 to disable limit.
75  */
76 int zfs_arc_shrinker_limit = 10000;
77 
78 #ifdef CONFIG_MEMORY_HOTPLUG
79 static struct notifier_block arc_hotplug_callback_mem_nb;
80 #endif
81 
82 /*
83  * Return a default max arc size based on the amount of physical memory.
84  */
85 uint64_t
86 arc_default_max(uint64_t min, uint64_t allmem)
87 {
88 	/* Default to 1/2 of all memory. */
89 	return (MAX(allmem / 2, min));
90 }
91 
92 #ifdef _KERNEL
93 /*
94  * Return maximum amount of memory that we could possibly use.  Reduced
95  * to half of all memory in user space which is primarily used for testing.
96  */
97 uint64_t
98 arc_all_memory(void)
99 {
100 #ifdef CONFIG_HIGHMEM
101 	return (ptob(zfs_totalram_pages - zfs_totalhigh_pages));
102 #else
103 	return (ptob(zfs_totalram_pages));
104 #endif /* CONFIG_HIGHMEM */
105 }
106 
107 /*
108  * Return the amount of memory that is considered free.  In user space
109  * which is primarily used for testing we pretend that free memory ranges
110  * from 0-20% of all memory.
111  */
112 uint64_t
113 arc_free_memory(void)
114 {
115 #ifdef CONFIG_HIGHMEM
116 	struct sysinfo si;
117 	si_meminfo(&si);
118 	return (ptob(si.freeram - si.freehigh));
119 #else
120 	return (ptob(nr_free_pages() +
121 	    nr_inactive_file_pages()));
122 #endif /* CONFIG_HIGHMEM */
123 }
124 
125 /*
126  * Return the amount of memory that can be consumed before reclaim will be
127  * needed.  Positive if there is sufficient free memory, negative indicates
128  * the amount of memory that needs to be freed up.
129  */
130 int64_t
131 arc_available_memory(void)
132 {
133 	return (arc_free_memory() - arc_sys_free);
134 }
135 
136 static uint64_t
137 arc_evictable_memory(void)
138 {
139 	int64_t asize = aggsum_value(&arc_size);
140 	uint64_t arc_clean =
141 	    zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_DATA]) +
142 	    zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA]) +
143 	    zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_DATA]) +
144 	    zfs_refcount_count(&arc_mfu->arcs_esize[ARC_BUFC_METADATA]);
145 	uint64_t arc_dirty = MAX((int64_t)asize - (int64_t)arc_clean, 0);
146 
147 	/*
148 	 * Scale reported evictable memory in proportion to page cache, cap
149 	 * at specified min/max.
150 	 */
151 	uint64_t min = (ptob(nr_file_pages()) / 100) * zfs_arc_pc_percent;
152 	min = MAX(arc_c_min, MIN(arc_c_max, min));
153 
154 	if (arc_dirty >= min)
155 		return (arc_clean);
156 
157 	return (MAX((int64_t)asize - (int64_t)min, 0));
158 }
159 
160 /*
161  * The _count() function returns the number of free-able objects.
162  * The _scan() function returns the number of objects that were freed.
163  */
164 static unsigned long
165 arc_shrinker_count(struct shrinker *shrink, struct shrink_control *sc)
166 {
167 	/*
168 	 * __GFP_FS won't be set if we are called from ZFS code (see
169 	 * kmem_flags_convert(), which removes it).  To avoid a deadlock, we
170 	 * don't allow evicting in this case.  We return 0 rather than
171 	 * SHRINK_STOP so that the shrinker logic doesn't accumulate a
172 	 * deficit against us.
173 	 */
174 	if (!(sc->gfp_mask & __GFP_FS)) {
175 		return (0);
176 	}
177 
178 	/*
179 	 * This code is reached in the "direct reclaim" case, where the
180 	 * kernel (outside ZFS) is trying to allocate a page, and the system
181 	 * is low on memory.
182 	 *
183 	 * The kernel's shrinker code doesn't understand how many pages the
184 	 * ARC's callback actually frees, so it may ask the ARC to shrink a
185 	 * lot for one page allocation. This is problematic because it may
186 	 * take a long time, thus delaying the page allocation, and because
187 	 * it may force the ARC to unnecessarily shrink very small.
188 	 *
189 	 * Therefore, we limit the amount of data that we say is evictable,
190 	 * which limits the amount that the shrinker will ask us to evict for
191 	 * one page allocation attempt.
192 	 *
193 	 * In practice, we may be asked to shrink 4x the limit to satisfy one
194 	 * page allocation, before the kernel's shrinker code gives up on us.
195 	 * When that happens, we rely on the kernel code to find the pages
196 	 * that we freed before invoking the OOM killer.  This happens in
197 	 * __alloc_pages_slowpath(), which retries and finds the pages we
198 	 * freed when it calls get_page_from_freelist().
199 	 *
200 	 * See also the comment above zfs_arc_shrinker_limit.
201 	 */
202 	int64_t limit = zfs_arc_shrinker_limit != 0 ?
203 	    zfs_arc_shrinker_limit : INT64_MAX;
204 	return (MIN(limit, btop((int64_t)arc_evictable_memory())));
205 }
206 
207 static unsigned long
208 arc_shrinker_scan(struct shrinker *shrink, struct shrink_control *sc)
209 {
210 	ASSERT((sc->gfp_mask & __GFP_FS) != 0);
211 
212 	/* The arc is considered warm once reclaim has occurred */
213 	if (unlikely(arc_warm == B_FALSE))
214 		arc_warm = B_TRUE;
215 
216 	/*
217 	 * Evict the requested number of pages by reducing arc_c and waiting
218 	 * for the requested amount of data to be evicted.
219 	 */
220 	arc_reduce_target_size(ptob(sc->nr_to_scan));
221 	arc_wait_for_eviction(ptob(sc->nr_to_scan));
222 	if (current->reclaim_state != NULL)
223 		current->reclaim_state->reclaimed_slab += sc->nr_to_scan;
224 
225 	/*
226 	 * We are experiencing memory pressure which the arc_evict_zthr was
227 	 * unable to keep up with. Set arc_no_grow to briefly pause arc
228 	 * growth to avoid compounding the memory pressure.
229 	 */
230 	arc_no_grow = B_TRUE;
231 
232 	/*
233 	 * When direct reclaim is observed it usually indicates a rapid
234 	 * increase in memory pressure.  This occurs because the kswapd
235 	 * threads were unable to asynchronously keep enough free memory
236 	 * available.
237 	 */
238 	if (current_is_kswapd()) {
239 		ARCSTAT_BUMP(arcstat_memory_indirect_count);
240 	} else {
241 		ARCSTAT_BUMP(arcstat_memory_direct_count);
242 	}
243 
244 	return (sc->nr_to_scan);
245 }
246 
247 SPL_SHRINKER_DECLARE(arc_shrinker,
248     arc_shrinker_count, arc_shrinker_scan, DEFAULT_SEEKS);
249 
250 int
251 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
252 {
253 	uint64_t free_memory = arc_free_memory();
254 
255 	if (free_memory > arc_all_memory() * arc_lotsfree_percent / 100)
256 		return (0);
257 
258 	if (txg > spa->spa_lowmem_last_txg) {
259 		spa->spa_lowmem_last_txg = txg;
260 		spa->spa_lowmem_page_load = 0;
261 	}
262 	/*
263 	 * If we are in pageout, we know that memory is already tight,
264 	 * the arc is already going to be evicting, so we just want to
265 	 * continue to let page writes occur as quickly as possible.
266 	 */
267 	if (current_is_kswapd()) {
268 		if (spa->spa_lowmem_page_load >
269 		    MAX(arc_sys_free / 4, free_memory) / 4) {
270 			DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
271 			return (SET_ERROR(ERESTART));
272 		}
273 		/* Note: reserve is inflated, so we deflate */
274 		atomic_add_64(&spa->spa_lowmem_page_load, reserve / 8);
275 		return (0);
276 	} else if (spa->spa_lowmem_page_load > 0 && arc_reclaim_needed()) {
277 		/* memory is low, delay before restarting */
278 		ARCSTAT_INCR(arcstat_memory_throttle_count, 1);
279 		DMU_TX_STAT_BUMP(dmu_tx_memory_reclaim);
280 		return (SET_ERROR(EAGAIN));
281 	}
282 	spa->spa_lowmem_page_load = 0;
283 	return (0);
284 }
285 
286 static void
287 arc_set_sys_free(uint64_t allmem)
288 {
289 	/*
290 	 * The ARC tries to keep at least this much memory available for the
291 	 * system.  This gives the ARC time to shrink in response to memory
292 	 * pressure, before running completely out of memory and invoking the
293 	 * direct-reclaim ARC shrinker.
294 	 *
295 	 * This should be more than twice high_wmark_pages(), so that
296 	 * arc_wait_for_eviction() will wait until at least the
297 	 * high_wmark_pages() are free (see arc_evict_state_impl()).
298 	 *
299 	 * Note: Even when the system is very low on memory, the kernel's
300 	 * shrinker code may only ask for one "batch" of pages (512KB) to be
301 	 * evicted.  If concurrent allocations consume these pages, there may
302 	 * still be insufficient free pages, and the OOM killer takes action.
303 	 *
304 	 * By setting arc_sys_free large enough, and having
305 	 * arc_wait_for_eviction() wait until there is at least arc_sys_free/2
306 	 * free memory, it is much less likely that concurrent allocations can
307 	 * consume all the memory that was evicted before checking for
308 	 * OOM.
309 	 *
310 	 * It's hard to iterate the zones from a linux kernel module, which
311 	 * makes it difficult to determine the watermark dynamically. Instead
312 	 * we compute the maximum high watermark for this system, based
313 	 * on the amount of memory, assuming default parameters on Linux kernel
314 	 * 5.3.
315 	 */
316 
317 	/*
318 	 * Base wmark_low is 4 * the square root of Kbytes of RAM.
319 	 */
320 	long wmark = 4 * int_sqrt(allmem/1024) * 1024;
321 
322 	/*
323 	 * Clamp to between 128K and 64MB.
324 	 */
325 	wmark = MAX(wmark, 128 * 1024);
326 	wmark = MIN(wmark, 64 * 1024 * 1024);
327 
328 	/*
329 	 * watermark_boost can increase the wmark by up to 150%.
330 	 */
331 	wmark += wmark * 150 / 100;
332 
333 	/*
334 	 * arc_sys_free needs to be more than 2x the watermark, because
335 	 * arc_wait_for_eviction() waits for half of arc_sys_free.  Bump this up
336 	 * to 3x to ensure we're above it.
337 	 */
338 	arc_sys_free = wmark * 3 + allmem / 32;
339 }
340 
341 void
342 arc_lowmem_init(void)
343 {
344 	uint64_t allmem = arc_all_memory();
345 
346 	/*
347 	 * Register a shrinker to support synchronous (direct) memory
348 	 * reclaim from the arc.  This is done to prevent kswapd from
349 	 * swapping out pages when it is preferable to shrink the arc.
350 	 */
351 	spl_register_shrinker(&arc_shrinker);
352 	arc_set_sys_free(allmem);
353 }
354 
355 void
356 arc_lowmem_fini(void)
357 {
358 	spl_unregister_shrinker(&arc_shrinker);
359 }
360 
361 int
362 param_set_arc_long(const char *buf, zfs_kernel_param_t *kp)
363 {
364 	int error;
365 
366 	error = param_set_long(buf, kp);
367 	if (error < 0)
368 		return (SET_ERROR(error));
369 
370 	arc_tuning_update(B_TRUE);
371 
372 	return (0);
373 }
374 
375 int
376 param_set_arc_int(const char *buf, zfs_kernel_param_t *kp)
377 {
378 	int error;
379 
380 	error = param_set_int(buf, kp);
381 	if (error < 0)
382 		return (SET_ERROR(error));
383 
384 	arc_tuning_update(B_TRUE);
385 
386 	return (0);
387 }
388 
389 #ifdef CONFIG_MEMORY_HOTPLUG
390 /* ARGSUSED */
391 static int
392 arc_hotplug_callback(struct notifier_block *self, unsigned long action,
393     void *arg)
394 {
395 	uint64_t allmem = arc_all_memory();
396 	if (action != MEM_ONLINE)
397 		return (NOTIFY_OK);
398 
399 	arc_set_limits(allmem);
400 
401 #ifdef __LP64__
402 	if (zfs_dirty_data_max_max == 0)
403 		zfs_dirty_data_max_max = MIN(4ULL * 1024 * 1024 * 1024,
404 		    allmem * zfs_dirty_data_max_max_percent / 100);
405 #else
406 	if (zfs_dirty_data_max_max == 0)
407 		zfs_dirty_data_max_max = MIN(1ULL * 1024 * 1024 * 1024,
408 		    allmem * zfs_dirty_data_max_max_percent / 100);
409 #endif
410 
411 	arc_set_sys_free(allmem);
412 	return (NOTIFY_OK);
413 }
414 #endif
415 
416 void
417 arc_register_hotplug(void)
418 {
419 #ifdef CONFIG_MEMORY_HOTPLUG
420 	arc_hotplug_callback_mem_nb.notifier_call = arc_hotplug_callback;
421 	/* There is no significance to the value 100 */
422 	arc_hotplug_callback_mem_nb.priority = 100;
423 	register_memory_notifier(&arc_hotplug_callback_mem_nb);
424 #endif
425 }
426 
427 void
428 arc_unregister_hotplug(void)
429 {
430 #ifdef CONFIG_MEMORY_HOTPLUG
431 	unregister_memory_notifier(&arc_hotplug_callback_mem_nb);
432 #endif
433 }
434 #else /* _KERNEL */
435 int64_t
436 arc_available_memory(void)
437 {
438 	int64_t lowest = INT64_MAX;
439 
440 	/* Every 100 calls, free a small amount */
441 	if (spa_get_random(100) == 0)
442 		lowest = -1024;
443 
444 	return (lowest);
445 }
446 
447 int
448 arc_memory_throttle(spa_t *spa, uint64_t reserve, uint64_t txg)
449 {
450 	return (0);
451 }
452 
453 uint64_t
454 arc_all_memory(void)
455 {
456 	return (ptob(physmem) / 2);
457 }
458 
459 uint64_t
460 arc_free_memory(void)
461 {
462 	return (spa_get_random(arc_all_memory() * 20 / 100));
463 }
464 
465 void
466 arc_register_hotplug(void)
467 {
468 }
469 
470 void
471 arc_unregister_hotplug(void)
472 {
473 }
474 #endif /* _KERNEL */
475 
476 /*
477  * Helper function for arc_prune_async() it is responsible for safely
478  * handling the execution of a registered arc_prune_func_t.
479  */
480 static void
481 arc_prune_task(void *ptr)
482 {
483 	arc_prune_t *ap = (arc_prune_t *)ptr;
484 	arc_prune_func_t *func = ap->p_pfunc;
485 
486 	if (func != NULL)
487 		func(ap->p_adjust, ap->p_private);
488 
489 	zfs_refcount_remove(&ap->p_refcnt, func);
490 }
491 
492 /*
493  * Notify registered consumers they must drop holds on a portion of the ARC
494  * buffered they reference.  This provides a mechanism to ensure the ARC can
495  * honor the arc_meta_limit and reclaim otherwise pinned ARC buffers.  This
496  * is analogous to dnlc_reduce_cache() but more generic.
497  *
498  * This operation is performed asynchronously so it may be safely called
499  * in the context of the arc_reclaim_thread().  A reference is taken here
500  * for each registered arc_prune_t and the arc_prune_task() is responsible
501  * for releasing it once the registered arc_prune_func_t has completed.
502  */
503 void
504 arc_prune_async(int64_t adjust)
505 {
506 	arc_prune_t *ap;
507 
508 	mutex_enter(&arc_prune_mtx);
509 	for (ap = list_head(&arc_prune_list); ap != NULL;
510 	    ap = list_next(&arc_prune_list, ap)) {
511 
512 		if (zfs_refcount_count(&ap->p_refcnt) >= 2)
513 			continue;
514 
515 		zfs_refcount_add(&ap->p_refcnt, ap->p_pfunc);
516 		ap->p_adjust = adjust;
517 		if (taskq_dispatch(arc_prune_taskq, arc_prune_task,
518 		    ap, TQ_SLEEP) == TASKQID_INVALID) {
519 			zfs_refcount_remove(&ap->p_refcnt, ap->p_pfunc);
520 			continue;
521 		}
522 		ARCSTAT_BUMP(arcstat_prune);
523 	}
524 	mutex_exit(&arc_prune_mtx);
525 }
526 
527 /* BEGIN CSTYLED */
528 ZFS_MODULE_PARAM(zfs_arc, zfs_arc_, shrinker_limit, INT, ZMOD_RW,
529 	"Limit on number of pages that ARC shrinker can reclaim at once");
530 /* END CSTYLED */
531