Print this page
4045 zfs write throttle & i/o scheduler performance work
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/vdev_cache.c
+++ new/usr/src/uts/common/fs/zfs/vdev_cache.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21 /*
22 22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 23 * Use is subject to license terms.
24 24 */
25 25 /*
26 26 * Copyright (c) 2013 by Delphix. All rights reserved.
27 27 */
28 28
29 29 #include <sys/zfs_context.h>
30 30 #include <sys/spa.h>
31 31 #include <sys/vdev_impl.h>
32 32 #include <sys/zio.h>
33 33 #include <sys/kstat.h>
34 34
35 35 /*
36 36 * Virtual device read-ahead caching.
37 37 *
38 38 * This file implements a simple LRU read-ahead cache. When the DMU reads
39 39 * a given block, it will often want other, nearby blocks soon thereafter.
40 40 * We take advantage of this by reading a larger disk region and caching
41 41 * the result. In the best case, this can turn 128 back-to-back 512-byte
42 42 * reads into a single 64k read followed by 127 cache hits; this reduces
43 43 * latency dramatically. In the worst case, it can turn an isolated 512-byte
44 44 * read into a 64k read, which doesn't affect latency all that much but is
45 45 * terribly wasteful of bandwidth. A more intelligent version of the cache
46 46 * could keep track of access patterns and not do read-ahead unless it sees
47 47 * at least two temporally close I/Os to the same region. Currently, only
48 48 * metadata I/O is inflated. A futher enhancement could take advantage of
49 49 * more semantic information about the I/O. And it could use something
50 50 * faster than an AVL tree; that was chosen solely for convenience.
51 51 *
52 52 * There are five cache operations: allocate, fill, read, write, evict.
53 53 *
54 54 * (1) Allocate. This reserves a cache entry for the specified region.
55 55 * We separate the allocate and fill operations so that multiple threads
56 56 * don't generate I/O for the same cache miss.
57 57 *
58 58 * (2) Fill. When the I/O for a cache miss completes, the fill routine
59 59 * places the data in the previously allocated cache entry.
60 60 *
61 61 * (3) Read. Read data from the cache.
62 62 *
63 63 * (4) Write. Update cache contents after write completion.
64 64 *
65 65 * (5) Evict. When allocating a new entry, we evict the oldest (LRU) entry
66 66 * if the total cache size exceeds zfs_vdev_cache_size.
67 67 */
68 68
69 69 /*
70 70 * These tunables are for performance analysis.
71 71 */
72 72 /*
73 73 * All i/os smaller than zfs_vdev_cache_max will be turned into
74 74 * 1<<zfs_vdev_cache_bshift byte reads by the vdev_cache (aka software
75 75 * track buffer). At most zfs_vdev_cache_size bytes will be kept in each
76 76 * vdev's vdev_cache.
77 77 *
78 78 * TODO: Note that with the current ZFS code, it turns out that the
79 79 * vdev cache is not helpful, and in some cases actually harmful. It
80 80 * is better if we disable this. Once some time has passed, we should
81 81 * actually remove this to simplify the code. For now we just disable
82 82 * it by setting the zfs_vdev_cache_size to zero. Note that Solaris 11
83 83 * has made these same changes.
84 84 */
85 85 int zfs_vdev_cache_max = 1<<14; /* 16KB */
86 86 int zfs_vdev_cache_size = 0;
87 87 int zfs_vdev_cache_bshift = 16;
88 88
89 89 #define VCBS (1 << zfs_vdev_cache_bshift) /* 64KB */
90 90
91 91 kstat_t *vdc_ksp = NULL;
92 92
93 93 typedef struct vdc_stats {
94 94 kstat_named_t vdc_stat_delegations;
95 95 kstat_named_t vdc_stat_hits;
96 96 kstat_named_t vdc_stat_misses;
97 97 } vdc_stats_t;
98 98
99 99 static vdc_stats_t vdc_stats = {
100 100 { "delegations", KSTAT_DATA_UINT64 },
101 101 { "hits", KSTAT_DATA_UINT64 },
102 102 { "misses", KSTAT_DATA_UINT64 }
103 103 };
104 104
105 105 #define VDCSTAT_BUMP(stat) atomic_add_64(&vdc_stats.stat.value.ui64, 1);
106 106
107 107 static int
108 108 vdev_cache_offset_compare(const void *a1, const void *a2)
109 109 {
110 110 const vdev_cache_entry_t *ve1 = a1;
111 111 const vdev_cache_entry_t *ve2 = a2;
112 112
113 113 if (ve1->ve_offset < ve2->ve_offset)
114 114 return (-1);
115 115 if (ve1->ve_offset > ve2->ve_offset)
116 116 return (1);
117 117 return (0);
118 118 }
119 119
120 120 static int
121 121 vdev_cache_lastused_compare(const void *a1, const void *a2)
122 122 {
123 123 const vdev_cache_entry_t *ve1 = a1;
124 124 const vdev_cache_entry_t *ve2 = a2;
125 125
126 126 if (ve1->ve_lastused < ve2->ve_lastused)
127 127 return (-1);
128 128 if (ve1->ve_lastused > ve2->ve_lastused)
129 129 return (1);
130 130
131 131 /*
132 132 * Among equally old entries, sort by offset to ensure uniqueness.
133 133 */
134 134 return (vdev_cache_offset_compare(a1, a2));
135 135 }
136 136
137 137 /*
138 138 * Evict the specified entry from the cache.
139 139 */
140 140 static void
141 141 vdev_cache_evict(vdev_cache_t *vc, vdev_cache_entry_t *ve)
142 142 {
143 143 ASSERT(MUTEX_HELD(&vc->vc_lock));
144 144 ASSERT(ve->ve_fill_io == NULL);
145 145 ASSERT(ve->ve_data != NULL);
146 146
147 147 avl_remove(&vc->vc_lastused_tree, ve);
148 148 avl_remove(&vc->vc_offset_tree, ve);
149 149 zio_buf_free(ve->ve_data, VCBS);
150 150 kmem_free(ve, sizeof (vdev_cache_entry_t));
151 151 }
152 152
153 153 /*
154 154 * Allocate an entry in the cache. At the point we don't have the data,
155 155 * we're just creating a placeholder so that multiple threads don't all
156 156 * go off and read the same blocks.
157 157 */
158 158 static vdev_cache_entry_t *
159 159 vdev_cache_allocate(zio_t *zio)
160 160 {
161 161 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
162 162 uint64_t offset = P2ALIGN(zio->io_offset, VCBS);
163 163 vdev_cache_entry_t *ve;
164 164
165 165 ASSERT(MUTEX_HELD(&vc->vc_lock));
166 166
167 167 if (zfs_vdev_cache_size == 0)
168 168 return (NULL);
169 169
170 170 /*
171 171 * If adding a new entry would exceed the cache size,
172 172 * evict the oldest entry (LRU).
173 173 */
174 174 if ((avl_numnodes(&vc->vc_lastused_tree) << zfs_vdev_cache_bshift) >
175 175 zfs_vdev_cache_size) {
176 176 ve = avl_first(&vc->vc_lastused_tree);
177 177 if (ve->ve_fill_io != NULL)
178 178 return (NULL);
179 179 ASSERT(ve->ve_hits != 0);
180 180 vdev_cache_evict(vc, ve);
181 181 }
182 182
183 183 ve = kmem_zalloc(sizeof (vdev_cache_entry_t), KM_SLEEP);
184 184 ve->ve_offset = offset;
185 185 ve->ve_lastused = ddi_get_lbolt();
186 186 ve->ve_data = zio_buf_alloc(VCBS);
187 187
188 188 avl_add(&vc->vc_offset_tree, ve);
189 189 avl_add(&vc->vc_lastused_tree, ve);
190 190
191 191 return (ve);
192 192 }
193 193
194 194 static void
195 195 vdev_cache_hit(vdev_cache_t *vc, vdev_cache_entry_t *ve, zio_t *zio)
196 196 {
197 197 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
198 198
199 199 ASSERT(MUTEX_HELD(&vc->vc_lock));
200 200 ASSERT(ve->ve_fill_io == NULL);
201 201
202 202 if (ve->ve_lastused != ddi_get_lbolt()) {
203 203 avl_remove(&vc->vc_lastused_tree, ve);
204 204 ve->ve_lastused = ddi_get_lbolt();
205 205 avl_add(&vc->vc_lastused_tree, ve);
206 206 }
207 207
208 208 ve->ve_hits++;
209 209 bcopy(ve->ve_data + cache_phase, zio->io_data, zio->io_size);
210 210 }
211 211
212 212 /*
213 213 * Fill a previously allocated cache entry with data.
214 214 */
215 215 static void
216 216 vdev_cache_fill(zio_t *fio)
217 217 {
218 218 vdev_t *vd = fio->io_vd;
219 219 vdev_cache_t *vc = &vd->vdev_cache;
220 220 vdev_cache_entry_t *ve = fio->io_private;
221 221 zio_t *pio;
222 222
223 223 ASSERT(fio->io_size == VCBS);
224 224
225 225 /*
226 226 * Add data to the cache.
227 227 */
228 228 mutex_enter(&vc->vc_lock);
229 229
230 230 ASSERT(ve->ve_fill_io == fio);
231 231 ASSERT(ve->ve_offset == fio->io_offset);
232 232 ASSERT(ve->ve_data == fio->io_data);
233 233
234 234 ve->ve_fill_io = NULL;
235 235
236 236 /*
237 237 * Even if this cache line was invalidated by a missed write update,
238 238 * any reads that were queued up before the missed update are still
239 239 * valid, so we can satisfy them from this line before we evict it.
240 240 */
241 241 while ((pio = zio_walk_parents(fio)) != NULL)
242 242 vdev_cache_hit(vc, ve, pio);
243 243
244 244 if (fio->io_error || ve->ve_missed_update)
245 245 vdev_cache_evict(vc, ve);
246 246
247 247 mutex_exit(&vc->vc_lock);
248 248 }
249 249
250 250 /*
251 251 * Read data from the cache. Returns 0 on cache hit, errno on a miss.
252 252 */
253 253 int
254 254 vdev_cache_read(zio_t *zio)
255 255 {
256 256 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
257 257 vdev_cache_entry_t *ve, ve_search;
258 258 uint64_t cache_offset = P2ALIGN(zio->io_offset, VCBS);
259 259 uint64_t cache_phase = P2PHASE(zio->io_offset, VCBS);
260 260 zio_t *fio;
261 261
262 262 ASSERT(zio->io_type == ZIO_TYPE_READ);
263 263
264 264 if (zio->io_flags & ZIO_FLAG_DONT_CACHE)
265 265 return (SET_ERROR(EINVAL));
266 266
267 267 if (zio->io_size > zfs_vdev_cache_max)
268 268 return (SET_ERROR(EOVERFLOW));
269 269
270 270 /*
271 271 * If the I/O straddles two or more cache blocks, don't cache it.
272 272 */
273 273 if (P2BOUNDARY(zio->io_offset, zio->io_size, VCBS))
274 274 return (SET_ERROR(EXDEV));
275 275
276 276 ASSERT(cache_phase + zio->io_size <= VCBS);
277 277
278 278 mutex_enter(&vc->vc_lock);
279 279
280 280 ve_search.ve_offset = cache_offset;
281 281 ve = avl_find(&vc->vc_offset_tree, &ve_search, NULL);
282 282
283 283 if (ve != NULL) {
284 284 if (ve->ve_missed_update) {
285 285 mutex_exit(&vc->vc_lock);
286 286 return (SET_ERROR(ESTALE));
287 287 }
288 288
289 289 if ((fio = ve->ve_fill_io) != NULL) {
290 290 zio_vdev_io_bypass(zio);
291 291 zio_add_child(zio, fio);
292 292 mutex_exit(&vc->vc_lock);
293 293 VDCSTAT_BUMP(vdc_stat_delegations);
294 294 return (0);
295 295 }
296 296
297 297 vdev_cache_hit(vc, ve, zio);
298 298 zio_vdev_io_bypass(zio);
299 299
300 300 mutex_exit(&vc->vc_lock);
301 301 VDCSTAT_BUMP(vdc_stat_hits);
302 302 return (0);
|
↓ open down ↓ |
302 lines elided |
↑ open up ↑ |
303 303 }
304 304
305 305 ve = vdev_cache_allocate(zio);
306 306
307 307 if (ve == NULL) {
308 308 mutex_exit(&vc->vc_lock);
309 309 return (SET_ERROR(ENOMEM));
310 310 }
311 311
312 312 fio = zio_vdev_delegated_io(zio->io_vd, cache_offset,
313 - ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_CACHE_FILL,
313 + ve->ve_data, VCBS, ZIO_TYPE_READ, ZIO_PRIORITY_NOW,
314 314 ZIO_FLAG_DONT_CACHE, vdev_cache_fill, ve);
315 315
316 316 ve->ve_fill_io = fio;
317 317 zio_vdev_io_bypass(zio);
318 318 zio_add_child(zio, fio);
319 319
320 320 mutex_exit(&vc->vc_lock);
321 321 zio_nowait(fio);
322 322 VDCSTAT_BUMP(vdc_stat_misses);
323 323
324 324 return (0);
325 325 }
326 326
327 327 /*
328 328 * Update cache contents upon write completion.
329 329 */
330 330 void
331 331 vdev_cache_write(zio_t *zio)
332 332 {
333 333 vdev_cache_t *vc = &zio->io_vd->vdev_cache;
334 334 vdev_cache_entry_t *ve, ve_search;
335 335 uint64_t io_start = zio->io_offset;
336 336 uint64_t io_end = io_start + zio->io_size;
337 337 uint64_t min_offset = P2ALIGN(io_start, VCBS);
338 338 uint64_t max_offset = P2ROUNDUP(io_end, VCBS);
339 339 avl_index_t where;
340 340
341 341 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
342 342
343 343 mutex_enter(&vc->vc_lock);
344 344
345 345 ve_search.ve_offset = min_offset;
346 346 ve = avl_find(&vc->vc_offset_tree, &ve_search, &where);
347 347
348 348 if (ve == NULL)
349 349 ve = avl_nearest(&vc->vc_offset_tree, where, AVL_AFTER);
350 350
351 351 while (ve != NULL && ve->ve_offset < max_offset) {
352 352 uint64_t start = MAX(ve->ve_offset, io_start);
353 353 uint64_t end = MIN(ve->ve_offset + VCBS, io_end);
354 354
355 355 if (ve->ve_fill_io != NULL) {
356 356 ve->ve_missed_update = 1;
357 357 } else {
358 358 bcopy((char *)zio->io_data + start - io_start,
359 359 ve->ve_data + start - ve->ve_offset, end - start);
360 360 }
361 361 ve = AVL_NEXT(&vc->vc_offset_tree, ve);
362 362 }
363 363 mutex_exit(&vc->vc_lock);
364 364 }
365 365
366 366 void
367 367 vdev_cache_purge(vdev_t *vd)
368 368 {
369 369 vdev_cache_t *vc = &vd->vdev_cache;
370 370 vdev_cache_entry_t *ve;
371 371
372 372 mutex_enter(&vc->vc_lock);
373 373 while ((ve = avl_first(&vc->vc_offset_tree)) != NULL)
374 374 vdev_cache_evict(vc, ve);
375 375 mutex_exit(&vc->vc_lock);
376 376 }
377 377
378 378 void
379 379 vdev_cache_init(vdev_t *vd)
380 380 {
381 381 vdev_cache_t *vc = &vd->vdev_cache;
382 382
383 383 mutex_init(&vc->vc_lock, NULL, MUTEX_DEFAULT, NULL);
384 384
385 385 avl_create(&vc->vc_offset_tree, vdev_cache_offset_compare,
386 386 sizeof (vdev_cache_entry_t),
387 387 offsetof(struct vdev_cache_entry, ve_offset_node));
388 388
389 389 avl_create(&vc->vc_lastused_tree, vdev_cache_lastused_compare,
390 390 sizeof (vdev_cache_entry_t),
391 391 offsetof(struct vdev_cache_entry, ve_lastused_node));
392 392 }
393 393
394 394 void
395 395 vdev_cache_fini(vdev_t *vd)
396 396 {
397 397 vdev_cache_t *vc = &vd->vdev_cache;
398 398
399 399 vdev_cache_purge(vd);
400 400
401 401 avl_destroy(&vc->vc_offset_tree);
402 402 avl_destroy(&vc->vc_lastused_tree);
403 403
404 404 mutex_destroy(&vc->vc_lock);
405 405 }
406 406
407 407 void
408 408 vdev_cache_stat_init(void)
409 409 {
410 410 vdc_ksp = kstat_create("zfs", 0, "vdev_cache_stats", "misc",
411 411 KSTAT_TYPE_NAMED, sizeof (vdc_stats) / sizeof (kstat_named_t),
412 412 KSTAT_FLAG_VIRTUAL);
413 413 if (vdc_ksp != NULL) {
414 414 vdc_ksp->ks_data = &vdc_stats;
415 415 kstat_install(vdc_ksp);
416 416 }
417 417 }
418 418
419 419 void
420 420 vdev_cache_stat_fini(void)
421 421 {
422 422 if (vdc_ksp != NULL) {
423 423 kstat_delete(vdc_ksp);
424 424 vdc_ksp = NULL;
425 425 }
426 426 }
|
↓ open down ↓ |
103 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX