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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 * Copyright (c) 2013 by Delphix. All rights reserved.
29 #include <sys/zfs_context.h>
31 #include <sys/vdev_impl.h>
33 #include <sys/kstat.h>
36 * Virtual device read-ahead caching.
38 * This file implements a simple LRU read-ahead cache. When the DMU reads
39 * a given block, it will often want other, nearby blocks soon thereafter.
40 * We take advantage of this by reading a larger disk region and caching
41 * the result. In the best case, this can turn 128 back-to-back 512-byte
42 * reads into a single 64k read followed by 127 cache hits; this reduces
43 * latency dramatically. In the worst case, it can turn an isolated 512-byte
44 * read into a 64k read, which doesn't affect latency all that much but is
45 * terribly wasteful of bandwidth. A more intelligent version of the cache
46 * could keep track of access patterns and not do read-ahead unless it sees
47 * at least two temporally close I/Os to the same region. Currently, only
48 * metadata I/O is inflated. A futher enhancement could take advantage of
49 * more semantic information about the I/O. And it could use something
50 * faster than an AVL tree; that was chosen solely for convenience.
52 * There are five cache operations: allocate, fill, read, write, evict.
54 * (1) Allocate. This reserves a cache entry for the specified region.
55 * We separate the allocate and fill operations so that multiple threads
56 * don't generate I/O for the same cache miss.
58 * (2) Fill. When the I/O for a cache miss completes, the fill routine
59 * places the data in the previously allocated cache entry.
61 * (3) Read. Read data from the cache.
63 * (4) Write. Update cache contents after write completion.
65 * (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry
66 * if the total cache size exceeds zfs_vdev_cache_size.
70 * These tunables are for performance analysis.
73 * All i/os smaller than zfs_vdev_cache_max will be turned into
74 * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
75 * track buffer). At most zfs_vdev_cache_size bytes will be kept in each
78 * TODO: Note that with the current ZFS code, it turns out that the
79 * vdev cache is not helpful, and in some cases actually harmful. It
80 * is better if we disable this. Once some time has passed, we should
81 * actually remove this to simplify the code. For now we just disable
82 * it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11
83 * has made these same changes.
85 int zfs_vdev_cache_max = 1<<14; /* 16KB */
86 int zfs_vdev_cache_size = 0;
87 int zfs_vdev_cache_bshift = 16;
89 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
91 kstat_t *vdc_ksp = NULL;
93 typedef struct vdc_stats {
94 kstat_named_t vdc_stat_delegations;
95 kstat_named_t vdc_stat_hits;
96 kstat_named_t vdc_stat_misses;
99 static vdc_stats_t vdc_stats = {
100 { "delegations", KSTAT_DATA_UINT64 },
101 { "hits", KSTAT_DATA_UINT64 },
102 { "misses", KSTAT_DATA_UINT64 }
105 #define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1);
108 vdev_cache_offset_compare(const void *a1, const void *a2)
110 const vdev_cache_entry_t *ve1 = a1;
111 const vdev_cache_entry_t *ve2 = a2;
113 if (ve1->ve_offset < ve2->ve_offset)
115 if (ve1->ve_offset > ve2->ve_offset)
121 vdev_cache_lastused_compare(const void *a1, const void *a2)
123 const vdev_cache_entry_t *ve1 = a1;
124 const vdev_cache_entry_t *ve2 = a2;
126 if (ve1->ve_lastused < ve2->ve_lastused)
128 if (ve1->ve_lastused > ve2->ve_lastused)
132 * Among equally old entries, sort by offset to ensure uniqueness.
134 return (vdev_cache_offset_compare(a1, a2));
138 * Evict the specified entry from the cache.
141 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
143 ASSERT(MUTEX_HELD(&vc->vc_lock));
144 ASSERT(ve->ve_fill_io == NULL);
145 ASSERT(ve->ve_data != NULL);
147 avl_remove(&vc->vc_lastused_tree, ve);
148 avl_remove(&vc->vc_offset_tree, ve);
149 zio_buf_free(ve->ve_data, VCBS);
150 kmem_free(ve, sizeof (vdev_cache_entry_t));
154 * Allocate an entry in the cache. At the point we don't have the data,
155 * we're just creating a placeholder so that multiple threads don't all
156 * go off and read the same blocks.
158 static vdev_cache_entry_t *
159 vdev_cache_allocate(zio_t *zio)
161 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
162 uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
163 vdev_cache_entry_t *ve;
165 ASSERT(MUTEX_HELD(&vc->vc_lock));
167 if (zfs_vdev_cache_size == 0)
171 * If adding a new entry would exceed the cache size,
172 * evict the oldest entry (LRU).
174 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
175 zfs_vdev_cache_size) {
176 ve = avl_first(&vc->vc_lastused_tree);
177 if (ve->ve_fill_io != NULL)
179 ASSERT(ve->ve_hits != 0);
180 vdev_cache_evict(vc, ve);
183 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_PUSHPAGE);
184 ve->ve_offset = offset;
185 ve->ve_lastused = ddi_get_lbolt();
186 ve->ve_data = zio_buf_alloc(VCBS);
188 avl_add(&vc->vc_offset_tree, ve);
189 avl_add(&vc->vc_lastused_tree, ve);
195 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
197 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
199 ASSERT(MUTEX_HELD(&vc->vc_lock));
200 ASSERT(ve->ve_fill_io == NULL);
202 if (ve->ve_lastused != ddi_get_lbolt()) {
203 avl_remove(&vc->vc_lastused_tree, ve);
204 ve->ve_lastused = ddi_get_lbolt();
205 avl_add(&vc->vc_lastused_tree, ve);
209 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
213 * Fill a previously allocated cache entry with data.
216 vdev_cache_fill(zio_t *fio)
218 vdev_t *vd = fio->io_vd;
219 vdev_cache_t *vc = &vd->vdev_cache;
220 vdev_cache_entry_t *ve = fio->io_private;
223 ASSERT(fio->io_size == VCBS);
226 * Add data to the cache.
228 mutex_enter(&vc->vc_lock);
230 ASSERT(ve->ve_fill_io == fio);
231 ASSERT(ve->ve_offset == fio->io_offset);
232 ASSERT(ve->ve_data == fio->io_data);
234 ve->ve_fill_io = NULL;
237 * Even if this cache line was invalidated by a missed write update,
238 * any reads that were queued up before the missed update are still
239 * valid, so we can satisfy them from this line before we evict it.
241 while ((pio = zio_walk_parents(fio)) != NULL)
242 vdev_cache_hit(vc, ve, pio);
244 if (fio->io_error || ve->ve_missed_update)
245 vdev_cache_evict(vc, ve);
247 mutex_exit(&vc->vc_lock);
251 * Read data from the cache. Returns 0 on cache hit, errno on a miss.
254 vdev_cache_read(zio_t *zio)
256 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
257 vdev_cache_entry_t *ve, *ve_search;
258 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
260 ASSERTV(uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS));
262 ASSERT(zio->io_type == ZIO_TYPE_READ);
264 if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
265 return (SET_ERROR(EINVAL));
267 if (zio->io_size > zfs_vdev_cache_max)
268 return (SET_ERROR(EOVERFLOW));
271 * If the I/O straddles two or more cache blocks, don't cache it.
273 if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
274 return (SET_ERROR(EXDEV));
276 ASSERT(cache_phase + zio->io_size <= VCBS);
278 mutex_enter(&vc->vc_lock);
280 ve_search = kmem_alloc(sizeof (vdev_cache_entry_t), KM_PUSHPAGE);
281 ve_search->ve_offset = cache_offset;
282 ve = avl_find(&vc->vc_offset_tree, ve_search, NULL);
283 kmem_free(ve_search, sizeof (vdev_cache_entry_t));
286 if (ve->ve_missed_update) {
287 mutex_exit(&vc->vc_lock);
288 return (SET_ERROR(ESTALE));
291 if ((fio = ve->ve_fill_io) != NULL) {
292 zio_vdev_io_bypass(zio);
293 zio_add_child(zio, fio);
294 mutex_exit(&vc->vc_lock);
295 VDCSTAT_BUMP(vdc_stat_delegations);
299 vdev_cache_hit(vc, ve, zio);
300 zio_vdev_io_bypass(zio);
302 mutex_exit(&vc->vc_lock);
303 VDCSTAT_BUMP(vdc_stat_hits);
307 ve = vdev_cache_allocate(zio);
310 mutex_exit(&vc->vc_lock);
311 return (SET_ERROR(ENOMEM));
314 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
315 ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
316 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
318 ve->ve_fill_io = fio;
319 zio_vdev_io_bypass(zio);
320 zio_add_child(zio, fio);
322 mutex_exit(&vc->vc_lock);
324 VDCSTAT_BUMP(vdc_stat_misses);
330 * Update cache contents upon write completion.
333 vdev_cache_write(zio_t *zio)
335 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
336 vdev_cache_entry_t *ve, ve_search;
337 uint64_t io_start = zio->io_offset;
338 uint64_t io_end = io_start + zio->io_size;
339 uint64_t min_offset = P2ALIGN(io_start, VCBS);
340 uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
343 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
345 mutex_enter(&vc->vc_lock);
347 ve_search.ve_offset = min_offset;
348 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
351 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
353 while (ve != NULL && ve->ve_offset < max_offset) {
354 uint64_t start = MAX(ve->ve_offset, io_start);
355 uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
357 if (ve->ve_fill_io != NULL) {
358 ve->ve_missed_update = 1;
360 bcopy((char *)zio->io_data + start - io_start,
361 ve->ve_data + start - ve->ve_offset, end - start);
363 ve = AVL_NEXT(&vc->vc_offset_tree, ve);
365 mutex_exit(&vc->vc_lock);
369 vdev_cache_purge(vdev_t *vd)
371 vdev_cache_t *vc = &vd->vdev_cache;
372 vdev_cache_entry_t *ve;
374 mutex_enter(&vc->vc_lock);
375 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
376 vdev_cache_evict(vc, ve);
377 mutex_exit(&vc->vc_lock);
381 vdev_cache_init(vdev_t *vd)
383 vdev_cache_t *vc = &vd->vdev_cache;
385 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
387 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
388 sizeof (vdev_cache_entry_t),
389 offsetof(struct vdev_cache_entry, ve_offset_node));
391 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
392 sizeof (vdev_cache_entry_t),
393 offsetof(struct vdev_cache_entry, ve_lastused_node));
397 vdev_cache_fini(vdev_t *vd)
399 vdev_cache_t *vc = &vd->vdev_cache;
401 vdev_cache_purge(vd);
403 avl_destroy(&vc->vc_offset_tree);
404 avl_destroy(&vc->vc_lastused_tree);
406 mutex_destroy(&vc->vc_lock);
410 vdev_cache_stat_init(void)
412 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
413 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
415 if (vdc_ksp != NULL) {
416 vdc_ksp->ks_data = &vdc_stats;
417 kstat_install(vdc_ksp);
422 vdev_cache_stat_fini(void)
424 if (vdc_ksp != NULL) {
425 kstat_delete(vdc_ksp);
430 #if defined(_KERNEL) && defined(HAVE_SPL)
431 module_param(zfs_vdev_cache_max, int, 0644);
432 MODULE_PARM_DESC(zfs_vdev_cache_max, "Inflate reads small than max");
434 module_param(zfs_vdev_cache_size, int, 0444);
435 MODULE_PARM_DESC(zfs_vdev_cache_size, "Total size of the per-disk cache");
437 module_param(zfs_vdev_cache_bshift, int, 0644);
438 MODULE_PARM_DESC(zfs_vdev_cache_bshift, "Shift size to inflate reads too");