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 2009 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 /*
26  * Copyright (c) 2013, 2015 by Delphix. All rights reserved.
27  */
28 
29 #include <sys/zfs_context.h>
30 #include <sys/spa.h>
31 #include <sys/vdev_impl.h>
32 #include <sys/zio.h>
33 #include <sys/kstat.h>
34 
35 /*
36  * Virtual device read-ahead caching.
37  *
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.
51  *
52  * There are five cache operations: allocate, fill, read, write, evict.
53  *
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.
57  *
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.
60  *
61  * (3) Read.  Read data from the cache.
62  *
63  * (4) Write.  Update cache contents after write completion.
64  *
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.
67  */
68 
69 /*
70  * These tunables are for performance analysis.
71  */
72 /*
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
76  * vdev's vdev_cache.
77  *
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.
84  */
85 int zfs_vdev_cache_max = 1<<14;			/* 16KB */
86 int zfs_vdev_cache_size = 0;
87 int zfs_vdev_cache_bshift = 16;
88 
89 #define	VCBS (1 << zfs_vdev_cache_bshift)	/* 64KB */
90 
91 SYSCTL_DECL(_vfs_zfs_vdev);
92 SYSCTL_NODE(_vfs_zfs_vdev, OID_AUTO, cache, CTLFLAG_RW, 0, "ZFS VDEV Cache");
93 TUNABLE_INT("vfs.zfs.vdev.cache.max", &zfs_vdev_cache_max);
94 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, max, CTLFLAG_RDTUN,
95     &zfs_vdev_cache_max, 0, "Maximum I/O request size that increase read size");
96 TUNABLE_INT("vfs.zfs.vdev.cache.size", &zfs_vdev_cache_size);
97 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, size, CTLFLAG_RDTUN,
98     &zfs_vdev_cache_size, 0, "Size of VDEV cache");
99 TUNABLE_INT("vfs.zfs.vdev.cache.bshift", &zfs_vdev_cache_bshift);
100 SYSCTL_INT(_vfs_zfs_vdev_cache, OID_AUTO, bshift, CTLFLAG_RDTUN,
101     &zfs_vdev_cache_bshift, 0, "Turn too small requests into 1 << this value");
102 
103 kstat_t	*vdc_ksp = NULL;
104 
105 typedef struct vdc_stats {
106 	kstat_named_t vdc_stat_delegations;
107 	kstat_named_t vdc_stat_hits;
108 	kstat_named_t vdc_stat_misses;
109 } vdc_stats_t;
110 
111 static vdc_stats_t vdc_stats = {
112 	{ "delegations",	KSTAT_DATA_UINT64 },
113 	{ "hits",		KSTAT_DATA_UINT64 },
114 	{ "misses",		KSTAT_DATA_UINT64 }
115 };
116 
117 #define	VDCSTAT_BUMP(stat)	atomic_inc_64(&vdc_stats.stat.value.ui64);
118 
119 static int
vdev_cache_offset_compare(const void * a1,const void * a2)120 vdev_cache_offset_compare(const void *a1, const void *a2)
121 {
122 	const vdev_cache_entry_t *ve1 = a1;
123 	const vdev_cache_entry_t *ve2 = a2;
124 
125 	if (ve1->ve_offset < ve2->ve_offset)
126 		return (-1);
127 	if (ve1->ve_offset > ve2->ve_offset)
128 		return (1);
129 	return (0);
130 }
131 
132 static int
vdev_cache_lastused_compare(const void * a1,const void * a2)133 vdev_cache_lastused_compare(const void *a1, const void *a2)
134 {
135 	const vdev_cache_entry_t *ve1 = a1;
136 	const vdev_cache_entry_t *ve2 = a2;
137 
138 	if (ve1->ve_lastused < ve2->ve_lastused)
139 		return (-1);
140 	if (ve1->ve_lastused > ve2->ve_lastused)
141 		return (1);
142 
143 	/*
144 	 * Among equally old entries, sort by offset to ensure uniqueness.
145 	 */
146 	return (vdev_cache_offset_compare(a1, a2));
147 }
148 
149 /*
150  * Evict the specified entry from the cache.
151  */
152 static void
vdev_cache_evict(vdev_cache_t * vc,vdev_cache_entry_t * ve)153 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
154 {
155 	ASSERT(MUTEX_HELD(&vc->vc_lock));
156 	ASSERT(ve->ve_fill_io == NULL);
157 	ASSERT(ve->ve_data != NULL);
158 
159 	avl_remove(&vc->vc_lastused_tree, ve);
160 	avl_remove(&vc->vc_offset_tree, ve);
161 	zio_buf_free(ve->ve_data, VCBS);
162 	kmem_free(ve, sizeof (vdev_cache_entry_t));
163 }
164 
165 /*
166  * Allocate an entry in the cache.  At the point we don't have the data,
167  * we're just creating a placeholder so that multiple threads don't all
168  * go off and read the same blocks.
169  */
170 static vdev_cache_entry_t *
vdev_cache_allocate(zio_t * zio)171 vdev_cache_allocate(zio_t *zio)
172 {
173 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
174 	uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
175 	vdev_cache_entry_t *ve;
176 
177 	ASSERT(MUTEX_HELD(&vc->vc_lock));
178 
179 	if (zfs_vdev_cache_size == 0)
180 		return (NULL);
181 
182 	/*
183 	 * If adding a new entry would exceed the cache size,
184 	 * evict the oldest entry (LRU).
185 	 */
186 	if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
187 	    zfs_vdev_cache_size) {
188 		ve = avl_first(&vc->vc_lastused_tree);
189 		if (ve->ve_fill_io != NULL)
190 			return (NULL);
191 		ASSERT(ve->ve_hits != 0);
192 		vdev_cache_evict(vc, ve);
193 	}
194 
195 	ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
196 	ve->ve_offset = offset;
197 	ve->ve_lastused = ddi_get_lbolt();
198 	ve->ve_data = zio_buf_alloc(VCBS);
199 
200 	avl_add(&vc->vc_offset_tree, ve);
201 	avl_add(&vc->vc_lastused_tree, ve);
202 
203 	return (ve);
204 }
205 
206 static void
vdev_cache_hit(vdev_cache_t * vc,vdev_cache_entry_t * ve,zio_t * zio)207 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
208 {
209 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
210 
211 	ASSERT(MUTEX_HELD(&vc->vc_lock));
212 	ASSERT(ve->ve_fill_io == NULL);
213 
214 	if (ve->ve_lastused != ddi_get_lbolt()) {
215 		avl_remove(&vc->vc_lastused_tree, ve);
216 		ve->ve_lastused = ddi_get_lbolt();
217 		avl_add(&vc->vc_lastused_tree, ve);
218 	}
219 
220 	ve->ve_hits++;
221 	bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
222 }
223 
224 /*
225  * Fill a previously allocated cache entry with data.
226  */
227 static void
vdev_cache_fill(zio_t * fio)228 vdev_cache_fill(zio_t *fio)
229 {
230 	vdev_t *vd = fio->io_vd;
231 	vdev_cache_t *vc = &vd->vdev_cache;
232 	vdev_cache_entry_t *ve = fio->io_private;
233 	zio_t *pio;
234 
235 	ASSERT(fio->io_size == VCBS);
236 
237 	/*
238 	 * Add data to the cache.
239 	 */
240 	mutex_enter(&vc->vc_lock);
241 
242 	ASSERT(ve->ve_fill_io == fio);
243 	ASSERT(ve->ve_offset == fio->io_offset);
244 	ASSERT(ve->ve_data == fio->io_data);
245 
246 	ve->ve_fill_io = NULL;
247 
248 	/*
249 	 * Even if this cache line was invalidated by a missed write update,
250 	 * any reads that were queued up before the missed update are still
251 	 * valid, so we can satisfy them from this line before we evict it.
252 	 */
253 	zio_link_t *zl = NULL;
254 	while ((pio = zio_walk_parents(fio, &zl)) != NULL)
255 		vdev_cache_hit(vc, ve, pio);
256 
257 	if (fio->io_error || ve->ve_missed_update)
258 		vdev_cache_evict(vc, ve);
259 
260 	mutex_exit(&vc->vc_lock);
261 }
262 
263 /*
264  * Read data from the cache.  Returns B_TRUE cache hit, B_FALSE on miss.
265  */
266 boolean_t
vdev_cache_read(zio_t * zio)267 vdev_cache_read(zio_t *zio)
268 {
269 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
270 	vdev_cache_entry_t *ve, ve_search;
271 	uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
272 	uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
273 	zio_t *fio;
274 
275 	ASSERT(zio->io_type == ZIO_TYPE_READ);
276 
277 	if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
278 		return (B_FALSE);
279 
280 	if (zio->io_size > zfs_vdev_cache_max)
281 		return (B_FALSE);
282 
283 	/*
284 	 * If the I/O straddles two or more cache blocks, don't cache it.
285 	 */
286 	if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
287 		return (B_FALSE);
288 
289 	ASSERT(cache_phase + zio->io_size <= VCBS);
290 
291 	mutex_enter(&vc->vc_lock);
292 
293 	ve_search.ve_offset = cache_offset;
294 	ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
295 
296 	if (ve != NULL) {
297 		if (ve->ve_missed_update) {
298 			mutex_exit(&vc->vc_lock);
299 			return (B_FALSE);
300 		}
301 
302 		if ((fio = ve->ve_fill_io) != NULL) {
303 			zio_vdev_io_bypass(zio);
304 			zio_add_child(zio, fio);
305 			mutex_exit(&vc->vc_lock);
306 			VDCSTAT_BUMP(vdc_stat_delegations);
307 			return (B_TRUE);
308 		}
309 
310 		vdev_cache_hit(vc, ve, zio);
311 		zio_vdev_io_bypass(zio);
312 
313 		mutex_exit(&vc->vc_lock);
314 		VDCSTAT_BUMP(vdc_stat_hits);
315 		return (B_TRUE);
316 	}
317 
318 	ve = vdev_cache_allocate(zio);
319 
320 	if (ve == NULL) {
321 		mutex_exit(&vc->vc_lock);
322 		return (B_FALSE);
323 	}
324 
325 	fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
326 	    ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
327 	    ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
328 
329 	ve->ve_fill_io = fio;
330 	zio_vdev_io_bypass(zio);
331 	zio_add_child(zio, fio);
332 
333 	mutex_exit(&vc->vc_lock);
334 	zio_nowait(fio);
335 	VDCSTAT_BUMP(vdc_stat_misses);
336 
337 	return (B_TRUE);
338 }
339 
340 /*
341  * Update cache contents upon write completion.
342  */
343 void
vdev_cache_write(zio_t * zio)344 vdev_cache_write(zio_t *zio)
345 {
346 	vdev_cache_t *vc = &zio->io_vd->vdev_cache;
347 	vdev_cache_entry_t *ve, ve_search;
348 	uint64_t io_start = zio->io_offset;
349 	uint64_t io_end = io_start + zio->io_size;
350 	uint64_t min_offset = P2ALIGN(io_start, VCBS);
351 	uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
352 	avl_index_t where;
353 
354 	ASSERT(zio->io_type == ZIO_TYPE_WRITE);
355 
356 	mutex_enter(&vc->vc_lock);
357 
358 	ve_search.ve_offset = min_offset;
359 	ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
360 
361 	if (ve == NULL)
362 		ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
363 
364 	while (ve != NULL && ve->ve_offset < max_offset) {
365 		uint64_t start = MAX(ve->ve_offset, io_start);
366 		uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
367 
368 		if (ve->ve_fill_io != NULL) {
369 			ve->ve_missed_update = 1;
370 		} else {
371 			bcopy((char *)zio->io_data + start - io_start,
372 			    ve->ve_data + start - ve->ve_offset, end - start);
373 		}
374 		ve = AVL_NEXT(&vc->vc_offset_tree, ve);
375 	}
376 	mutex_exit(&vc->vc_lock);
377 }
378 
379 void
vdev_cache_purge(vdev_t * vd)380 vdev_cache_purge(vdev_t *vd)
381 {
382 	vdev_cache_t *vc = &vd->vdev_cache;
383 	vdev_cache_entry_t *ve;
384 
385 	mutex_enter(&vc->vc_lock);
386 	while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
387 		vdev_cache_evict(vc, ve);
388 	mutex_exit(&vc->vc_lock);
389 }
390 
391 void
vdev_cache_init(vdev_t * vd)392 vdev_cache_init(vdev_t *vd)
393 {
394 	vdev_cache_t *vc = &vd->vdev_cache;
395 
396 	mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
397 
398 	avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
399 	    sizeof (vdev_cache_entry_t),
400 	    offsetof(struct vdev_cache_entry, ve_offset_node));
401 
402 	avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
403 	    sizeof (vdev_cache_entry_t),
404 	    offsetof(struct vdev_cache_entry, ve_lastused_node));
405 }
406 
407 void
vdev_cache_fini(vdev_t * vd)408 vdev_cache_fini(vdev_t *vd)
409 {
410 	vdev_cache_t *vc = &vd->vdev_cache;
411 
412 	vdev_cache_purge(vd);
413 
414 	avl_destroy(&vc->vc_offset_tree);
415 	avl_destroy(&vc->vc_lastused_tree);
416 
417 	mutex_destroy(&vc->vc_lock);
418 }
419 
420 void
vdev_cache_stat_init(void)421 vdev_cache_stat_init(void)
422 {
423 	vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
424 	    KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
425 	    KSTAT_FLAG_VIRTUAL);
426 	if (vdc_ksp != NULL) {
427 		vdc_ksp->ks_data = &vdc_stats;
428 		kstat_install(vdc_ksp);
429 	}
430 }
431 
432 void
vdev_cache_stat_fini(void)433 vdev_cache_stat_fini(void)
434 {
435 	if (vdc_ksp != NULL) {
436 		kstat_delete(vdc_ksp);
437 		vdc_ksp = NULL;
438 	}
439 }
440