Print this page
10703 smatch unreachable code checking needs reworking
Reviewed by: Toomas Soome <tsoome@me.com>
Reviewed by: Yuri Pankov <yuri.pankov@nexenta.com>
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/common/fs/zfs/vdev_queue.c
+++ new/usr/src/uts/common/fs/zfs/vdev_queue.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 *
|
↓ open down ↓ |
18 lines elided |
↑ open up ↑ |
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 /*
27 27 * Copyright (c) 2012, 2018 by Delphix. All rights reserved.
28 28 * Copyright (c) 2014 Integros [integros.com]
29 + * Copyright 2019 Joyent, Inc.
29 30 */
30 31
31 32 #include <sys/zfs_context.h>
32 33 #include <sys/vdev_impl.h>
33 34 #include <sys/spa_impl.h>
34 35 #include <sys/zio.h>
35 36 #include <sys/avl.h>
36 37 #include <sys/dsl_pool.h>
37 38 #include <sys/metaslab_impl.h>
38 39 #include <sys/abd.h>
39 40
40 41 /*
41 42 * ZFS I/O Scheduler
42 43 * ---------------
43 44 *
44 45 * ZFS issues I/O operations to leaf vdevs to satisfy and complete zios. The
45 46 * I/O scheduler determines when and in what order those operations are
46 47 * issued. The I/O scheduler divides operations into five I/O classes
47 48 * prioritized in the following order: sync read, sync write, async read,
48 49 * async write, and scrub/resilver. Each queue defines the minimum and
49 50 * maximum number of concurrent operations that may be issued to the device.
50 51 * In addition, the device has an aggregate maximum. Note that the sum of the
51 52 * per-queue minimums must not exceed the aggregate maximum, and if the
52 53 * aggregate maximum is equal to or greater than the sum of the per-queue
53 54 * maximums, the per-queue minimum has no effect.
54 55 *
55 56 * For many physical devices, throughput increases with the number of
56 57 * concurrent operations, but latency typically suffers. Further, physical
57 58 * devices typically have a limit at which more concurrent operations have no
58 59 * effect on throughput or can actually cause it to decrease.
59 60 *
60 61 * The scheduler selects the next operation to issue by first looking for an
61 62 * I/O class whose minimum has not been satisfied. Once all are satisfied and
62 63 * the aggregate maximum has not been hit, the scheduler looks for classes
63 64 * whose maximum has not been satisfied. Iteration through the I/O classes is
64 65 * done in the order specified above. No further operations are issued if the
65 66 * aggregate maximum number of concurrent operations has been hit or if there
66 67 * are no operations queued for an I/O class that has not hit its maximum.
67 68 * Every time an i/o is queued or an operation completes, the I/O scheduler
68 69 * looks for new operations to issue.
69 70 *
70 71 * All I/O classes have a fixed maximum number of outstanding operations
71 72 * except for the async write class. Asynchronous writes represent the data
72 73 * that is committed to stable storage during the syncing stage for
73 74 * transaction groups (see txg.c). Transaction groups enter the syncing state
74 75 * periodically so the number of queued async writes will quickly burst up and
75 76 * then bleed down to zero. Rather than servicing them as quickly as possible,
76 77 * the I/O scheduler changes the maximum number of active async write i/os
77 78 * according to the amount of dirty data in the pool (see dsl_pool.c). Since
78 79 * both throughput and latency typically increase with the number of
79 80 * concurrent operations issued to physical devices, reducing the burstiness
80 81 * in the number of concurrent operations also stabilizes the response time of
81 82 * operations from other -- and in particular synchronous -- queues. In broad
82 83 * strokes, the I/O scheduler will issue more concurrent operations from the
83 84 * async write queue as there's more dirty data in the pool.
84 85 *
85 86 * Async Writes
86 87 *
87 88 * The number of concurrent operations issued for the async write I/O class
88 89 * follows a piece-wise linear function defined by a few adjustable points.
89 90 *
90 91 * | o---------| <-- zfs_vdev_async_write_max_active
91 92 * ^ | /^ |
92 93 * | | / | |
93 94 * active | / | |
94 95 * I/O | / | |
95 96 * count | / | |
96 97 * | / | |
97 98 * |------------o | | <-- zfs_vdev_async_write_min_active
98 99 * 0|____________^______|_________|
99 100 * 0% | | 100% of zfs_dirty_data_max
100 101 * | |
101 102 * | `-- zfs_vdev_async_write_active_max_dirty_percent
102 103 * `--------- zfs_vdev_async_write_active_min_dirty_percent
103 104 *
104 105 * Until the amount of dirty data exceeds a minimum percentage of the dirty
105 106 * data allowed in the pool, the I/O scheduler will limit the number of
106 107 * concurrent operations to the minimum. As that threshold is crossed, the
107 108 * number of concurrent operations issued increases linearly to the maximum at
108 109 * the specified maximum percentage of the dirty data allowed in the pool.
109 110 *
110 111 * Ideally, the amount of dirty data on a busy pool will stay in the sloped
111 112 * part of the function between zfs_vdev_async_write_active_min_dirty_percent
112 113 * and zfs_vdev_async_write_active_max_dirty_percent. If it exceeds the
113 114 * maximum percentage, this indicates that the rate of incoming data is
114 115 * greater than the rate that the backend storage can handle. In this case, we
115 116 * must further throttle incoming writes (see dmu_tx_delay() for details).
116 117 */
117 118
118 119 /*
119 120 * The maximum number of i/os active to each device. Ideally, this will be >=
120 121 * the sum of each queue's max_active. It must be at least the sum of each
121 122 * queue's min_active.
122 123 */
123 124 uint32_t zfs_vdev_max_active = 1000;
124 125
125 126 /*
126 127 * Per-queue limits on the number of i/os active to each device. If the
127 128 * sum of the queue's max_active is < zfs_vdev_max_active, then the
128 129 * min_active comes into play. We will send min_active from each queue,
129 130 * and then select from queues in the order defined by zio_priority_t.
130 131 *
131 132 * In general, smaller max_active's will lead to lower latency of synchronous
132 133 * operations. Larger max_active's may lead to higher overall throughput,
133 134 * depending on underlying storage.
134 135 *
135 136 * The ratio of the queues' max_actives determines the balance of performance
136 137 * between reads, writes, and scrubs. E.g., increasing
137 138 * zfs_vdev_scrub_max_active will cause the scrub or resilver to complete
138 139 * more quickly, but reads and writes to have higher latency and lower
139 140 * throughput.
140 141 */
141 142 uint32_t zfs_vdev_sync_read_min_active = 10;
142 143 uint32_t zfs_vdev_sync_read_max_active = 10;
143 144 uint32_t zfs_vdev_sync_write_min_active = 10;
144 145 uint32_t zfs_vdev_sync_write_max_active = 10;
145 146 uint32_t zfs_vdev_async_read_min_active = 1;
146 147 uint32_t zfs_vdev_async_read_max_active = 3;
147 148 uint32_t zfs_vdev_async_write_min_active = 1;
148 149 uint32_t zfs_vdev_async_write_max_active = 10;
149 150 uint32_t zfs_vdev_scrub_min_active = 1;
150 151 uint32_t zfs_vdev_scrub_max_active = 2;
151 152 uint32_t zfs_vdev_removal_min_active = 1;
152 153 uint32_t zfs_vdev_removal_max_active = 2;
153 154 uint32_t zfs_vdev_initializing_min_active = 1;
154 155 uint32_t zfs_vdev_initializing_max_active = 1;
155 156
156 157 /*
157 158 * When the pool has less than zfs_vdev_async_write_active_min_dirty_percent
158 159 * dirty data, use zfs_vdev_async_write_min_active. When it has more than
159 160 * zfs_vdev_async_write_active_max_dirty_percent, use
160 161 * zfs_vdev_async_write_max_active. The value is linearly interpolated
161 162 * between min and max.
162 163 */
163 164 int zfs_vdev_async_write_active_min_dirty_percent = 30;
164 165 int zfs_vdev_async_write_active_max_dirty_percent = 60;
165 166
166 167 /*
167 168 * To reduce IOPs, we aggregate small adjacent I/Os into one large I/O.
168 169 * For read I/Os, we also aggregate across small adjacency gaps; for writes
169 170 * we include spans of optional I/Os to aid aggregation at the disk even when
170 171 * they aren't able to help us aggregate at this level.
171 172 */
172 173 int zfs_vdev_aggregation_limit = 1 << 20;
173 174 int zfs_vdev_read_gap_limit = 32 << 10;
174 175 int zfs_vdev_write_gap_limit = 4 << 10;
175 176
176 177 /*
177 178 * Define the queue depth percentage for each top-level. This percentage is
178 179 * used in conjunction with zfs_vdev_async_max_active to determine how many
179 180 * allocations a specific top-level vdev should handle. Once the queue depth
180 181 * reaches zfs_vdev_queue_depth_pct * zfs_vdev_async_write_max_active / 100
181 182 * then allocator will stop allocating blocks on that top-level device.
182 183 * The default kernel setting is 1000% which will yield 100 allocations per
183 184 * device. For userland testing, the default setting is 300% which equates
184 185 * to 30 allocations per device.
185 186 */
186 187 #ifdef _KERNEL
187 188 int zfs_vdev_queue_depth_pct = 1000;
188 189 #else
189 190 int zfs_vdev_queue_depth_pct = 300;
190 191 #endif
191 192
192 193 /*
193 194 * When performing allocations for a given metaslab, we want to make sure that
194 195 * there are enough IOs to aggregate together to improve throughput. We want to
195 196 * ensure that there are at least 128k worth of IOs that can be aggregated, and
196 197 * we assume that the average allocation size is 4k, so we need the queue depth
197 198 * to be 32 per allocator to get good aggregation of sequential writes.
198 199 */
199 200 int zfs_vdev_def_queue_depth = 32;
200 201
201 202
202 203 int
203 204 vdev_queue_offset_compare(const void *x1, const void *x2)
204 205 {
205 206 const zio_t *z1 = (const zio_t *)x1;
206 207 const zio_t *z2 = (const zio_t *)x2;
207 208
208 209 int cmp = AVL_CMP(z1->io_offset, z2->io_offset);
209 210
210 211 if (likely(cmp))
211 212 return (cmp);
212 213
213 214 return (AVL_PCMP(z1, z2));
214 215 }
215 216
216 217 static inline avl_tree_t *
217 218 vdev_queue_class_tree(vdev_queue_t *vq, zio_priority_t p)
218 219 {
219 220 return (&vq->vq_class[p].vqc_queued_tree);
220 221 }
221 222
222 223 static inline avl_tree_t *
223 224 vdev_queue_type_tree(vdev_queue_t *vq, zio_type_t t)
224 225 {
225 226 ASSERT(t == ZIO_TYPE_READ || t == ZIO_TYPE_WRITE);
226 227 if (t == ZIO_TYPE_READ)
227 228 return (&vq->vq_read_offset_tree);
228 229 else
229 230 return (&vq->vq_write_offset_tree);
230 231 }
231 232
232 233 int
233 234 vdev_queue_timestamp_compare(const void *x1, const void *x2)
234 235 {
235 236 const zio_t *z1 = (const zio_t *)x1;
236 237 const zio_t *z2 = (const zio_t *)x2;
237 238
238 239 int cmp = AVL_CMP(z1->io_timestamp, z2->io_timestamp);
239 240
240 241 if (likely(cmp))
241 242 return (cmp);
242 243
243 244 return (AVL_PCMP(z1, z2));
244 245 }
245 246
246 247 void
247 248 vdev_queue_init(vdev_t *vd)
248 249 {
249 250 vdev_queue_t *vq = &vd->vdev_queue;
250 251
251 252 mutex_init(&vq->vq_lock, NULL, MUTEX_DEFAULT, NULL);
252 253 vq->vq_vdev = vd;
253 254
254 255 avl_create(&vq->vq_active_tree, vdev_queue_offset_compare,
255 256 sizeof (zio_t), offsetof(struct zio, io_queue_node));
256 257 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_READ),
257 258 vdev_queue_offset_compare, sizeof (zio_t),
258 259 offsetof(struct zio, io_offset_node));
259 260 avl_create(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE),
260 261 vdev_queue_offset_compare, sizeof (zio_t),
261 262 offsetof(struct zio, io_offset_node));
262 263
263 264 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
264 265 int (*compfn) (const void *, const void *);
265 266
266 267 /*
267 268 * The synchronous i/o queues are dispatched in FIFO rather
268 269 * than LBA order. This provides more consistent latency for
269 270 * these i/os.
270 271 */
271 272 if (p == ZIO_PRIORITY_SYNC_READ || p == ZIO_PRIORITY_SYNC_WRITE)
272 273 compfn = vdev_queue_timestamp_compare;
273 274 else
274 275 compfn = vdev_queue_offset_compare;
275 276
276 277 avl_create(vdev_queue_class_tree(vq, p), compfn,
277 278 sizeof (zio_t), offsetof(struct zio, io_queue_node));
278 279 }
279 280
280 281 vq->vq_last_offset = 0;
281 282 }
282 283
283 284 void
284 285 vdev_queue_fini(vdev_t *vd)
285 286 {
286 287 vdev_queue_t *vq = &vd->vdev_queue;
287 288
288 289 for (zio_priority_t p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++)
289 290 avl_destroy(vdev_queue_class_tree(vq, p));
290 291 avl_destroy(&vq->vq_active_tree);
291 292 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_READ));
292 293 avl_destroy(vdev_queue_type_tree(vq, ZIO_TYPE_WRITE));
293 294
294 295 mutex_destroy(&vq->vq_lock);
295 296 }
296 297
297 298 static void
298 299 vdev_queue_io_add(vdev_queue_t *vq, zio_t *zio)
299 300 {
300 301 spa_t *spa = zio->io_spa;
301 302
302 303 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
303 304 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
304 305 avl_add(vdev_queue_type_tree(vq, zio->io_type), zio);
305 306
306 307 mutex_enter(&spa->spa_iokstat_lock);
307 308 spa->spa_queue_stats[zio->io_priority].spa_queued++;
308 309 if (spa->spa_iokstat != NULL)
309 310 kstat_waitq_enter(spa->spa_iokstat->ks_data);
310 311 mutex_exit(&spa->spa_iokstat_lock);
311 312 }
312 313
313 314 static void
314 315 vdev_queue_io_remove(vdev_queue_t *vq, zio_t *zio)
315 316 {
316 317 spa_t *spa = zio->io_spa;
317 318
318 319 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
319 320 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
320 321 avl_remove(vdev_queue_type_tree(vq, zio->io_type), zio);
321 322
322 323 mutex_enter(&spa->spa_iokstat_lock);
323 324 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_queued, >, 0);
324 325 spa->spa_queue_stats[zio->io_priority].spa_queued--;
325 326 if (spa->spa_iokstat != NULL)
326 327 kstat_waitq_exit(spa->spa_iokstat->ks_data);
327 328 mutex_exit(&spa->spa_iokstat_lock);
328 329 }
329 330
330 331 static void
331 332 vdev_queue_pending_add(vdev_queue_t *vq, zio_t *zio)
332 333 {
333 334 spa_t *spa = zio->io_spa;
334 335 ASSERT(MUTEX_HELD(&vq->vq_lock));
335 336 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
336 337 vq->vq_class[zio->io_priority].vqc_active++;
337 338 avl_add(&vq->vq_active_tree, zio);
338 339
339 340 mutex_enter(&spa->spa_iokstat_lock);
340 341 spa->spa_queue_stats[zio->io_priority].spa_active++;
341 342 if (spa->spa_iokstat != NULL)
342 343 kstat_runq_enter(spa->spa_iokstat->ks_data);
343 344 mutex_exit(&spa->spa_iokstat_lock);
344 345 }
345 346
346 347 static void
347 348 vdev_queue_pending_remove(vdev_queue_t *vq, zio_t *zio)
348 349 {
349 350 spa_t *spa = zio->io_spa;
350 351 ASSERT(MUTEX_HELD(&vq->vq_lock));
351 352 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
352 353 vq->vq_class[zio->io_priority].vqc_active--;
353 354 avl_remove(&vq->vq_active_tree, zio);
354 355
355 356 mutex_enter(&spa->spa_iokstat_lock);
356 357 ASSERT3U(spa->spa_queue_stats[zio->io_priority].spa_active, >, 0);
357 358 spa->spa_queue_stats[zio->io_priority].spa_active--;
358 359 if (spa->spa_iokstat != NULL) {
359 360 kstat_io_t *ksio = spa->spa_iokstat->ks_data;
360 361
361 362 kstat_runq_exit(spa->spa_iokstat->ks_data);
362 363 if (zio->io_type == ZIO_TYPE_READ) {
363 364 ksio->reads++;
364 365 ksio->nread += zio->io_size;
365 366 } else if (zio->io_type == ZIO_TYPE_WRITE) {
366 367 ksio->writes++;
367 368 ksio->nwritten += zio->io_size;
368 369 }
369 370 }
370 371 mutex_exit(&spa->spa_iokstat_lock);
371 372 }
372 373
373 374 static void
374 375 vdev_queue_agg_io_done(zio_t *aio)
375 376 {
376 377 if (aio->io_type == ZIO_TYPE_READ) {
377 378 zio_t *pio;
378 379 zio_link_t *zl = NULL;
379 380 while ((pio = zio_walk_parents(aio, &zl)) != NULL) {
380 381 abd_copy_off(pio->io_abd, aio->io_abd,
381 382 0, pio->io_offset - aio->io_offset, pio->io_size);
382 383 }
383 384 }
384 385
385 386 abd_free(aio->io_abd);
386 387 }
387 388
388 389 static int
389 390 vdev_queue_class_min_active(zio_priority_t p)
390 391 {
391 392 switch (p) {
392 393 case ZIO_PRIORITY_SYNC_READ:
393 394 return (zfs_vdev_sync_read_min_active);
394 395 case ZIO_PRIORITY_SYNC_WRITE:
395 396 return (zfs_vdev_sync_write_min_active);
396 397 case ZIO_PRIORITY_ASYNC_READ:
397 398 return (zfs_vdev_async_read_min_active);
|
↓ open down ↓ |
359 lines elided |
↑ open up ↑ |
398 399 case ZIO_PRIORITY_ASYNC_WRITE:
399 400 return (zfs_vdev_async_write_min_active);
400 401 case ZIO_PRIORITY_SCRUB:
401 402 return (zfs_vdev_scrub_min_active);
402 403 case ZIO_PRIORITY_REMOVAL:
403 404 return (zfs_vdev_removal_min_active);
404 405 case ZIO_PRIORITY_INITIALIZING:
405 406 return (zfs_vdev_initializing_min_active);
406 407 default:
407 408 panic("invalid priority %u", p);
408 - return (0);
409 409 }
410 410 }
411 411
412 412 static int
413 413 vdev_queue_max_async_writes(spa_t *spa)
414 414 {
415 415 int writes;
416 416 uint64_t dirty = spa->spa_dsl_pool->dp_dirty_total;
417 417 uint64_t min_bytes = zfs_dirty_data_max *
418 418 zfs_vdev_async_write_active_min_dirty_percent / 100;
419 419 uint64_t max_bytes = zfs_dirty_data_max *
420 420 zfs_vdev_async_write_active_max_dirty_percent / 100;
421 421
422 422 /*
423 423 * Sync tasks correspond to interactive user actions. To reduce the
424 424 * execution time of those actions we push data out as fast as possible.
425 425 */
426 426 if (spa_has_pending_synctask(spa)) {
427 427 return (zfs_vdev_async_write_max_active);
428 428 }
429 429
430 430 if (dirty < min_bytes)
431 431 return (zfs_vdev_async_write_min_active);
432 432 if (dirty > max_bytes)
433 433 return (zfs_vdev_async_write_max_active);
434 434
435 435 /*
436 436 * linear interpolation:
437 437 * slope = (max_writes - min_writes) / (max_bytes - min_bytes)
438 438 * move right by min_bytes
439 439 * move up by min_writes
440 440 */
441 441 writes = (dirty - min_bytes) *
442 442 (zfs_vdev_async_write_max_active -
443 443 zfs_vdev_async_write_min_active) /
444 444 (max_bytes - min_bytes) +
445 445 zfs_vdev_async_write_min_active;
446 446 ASSERT3U(writes, >=, zfs_vdev_async_write_min_active);
447 447 ASSERT3U(writes, <=, zfs_vdev_async_write_max_active);
448 448 return (writes);
449 449 }
450 450
451 451 static int
452 452 vdev_queue_class_max_active(spa_t *spa, zio_priority_t p)
453 453 {
454 454 switch (p) {
455 455 case ZIO_PRIORITY_SYNC_READ:
456 456 return (zfs_vdev_sync_read_max_active);
457 457 case ZIO_PRIORITY_SYNC_WRITE:
458 458 return (zfs_vdev_sync_write_max_active);
459 459 case ZIO_PRIORITY_ASYNC_READ:
460 460 return (zfs_vdev_async_read_max_active);
|
↓ open down ↓ |
42 lines elided |
↑ open up ↑ |
461 461 case ZIO_PRIORITY_ASYNC_WRITE:
462 462 return (vdev_queue_max_async_writes(spa));
463 463 case ZIO_PRIORITY_SCRUB:
464 464 return (zfs_vdev_scrub_max_active);
465 465 case ZIO_PRIORITY_REMOVAL:
466 466 return (zfs_vdev_removal_max_active);
467 467 case ZIO_PRIORITY_INITIALIZING:
468 468 return (zfs_vdev_initializing_max_active);
469 469 default:
470 470 panic("invalid priority %u", p);
471 - return (0);
472 471 }
473 472 }
474 473
475 474 /*
476 475 * Return the i/o class to issue from, or ZIO_PRIORITY_MAX_QUEUEABLE if
477 476 * there is no eligible class.
478 477 */
479 478 static zio_priority_t
480 479 vdev_queue_class_to_issue(vdev_queue_t *vq)
481 480 {
482 481 spa_t *spa = vq->vq_vdev->vdev_spa;
483 482 zio_priority_t p;
484 483
485 484 if (avl_numnodes(&vq->vq_active_tree) >= zfs_vdev_max_active)
486 485 return (ZIO_PRIORITY_NUM_QUEUEABLE);
487 486
488 487 /* find a queue that has not reached its minimum # outstanding i/os */
489 488 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
490 489 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
491 490 vq->vq_class[p].vqc_active <
492 491 vdev_queue_class_min_active(p))
493 492 return (p);
494 493 }
495 494
496 495 /*
497 496 * If we haven't found a queue, look for one that hasn't reached its
498 497 * maximum # outstanding i/os.
499 498 */
500 499 for (p = 0; p < ZIO_PRIORITY_NUM_QUEUEABLE; p++) {
501 500 if (avl_numnodes(vdev_queue_class_tree(vq, p)) > 0 &&
502 501 vq->vq_class[p].vqc_active <
503 502 vdev_queue_class_max_active(spa, p))
504 503 return (p);
505 504 }
506 505
507 506 /* No eligible queued i/os */
508 507 return (ZIO_PRIORITY_NUM_QUEUEABLE);
509 508 }
510 509
511 510 /*
512 511 * Compute the range spanned by two i/os, which is the endpoint of the last
513 512 * (lio->io_offset + lio->io_size) minus start of the first (fio->io_offset).
514 513 * Conveniently, the gap between fio and lio is given by -IO_SPAN(lio, fio);
515 514 * thus fio and lio are adjacent if and only if IO_SPAN(lio, fio) == 0.
516 515 */
517 516 #define IO_SPAN(fio, lio) ((lio)->io_offset + (lio)->io_size - (fio)->io_offset)
518 517 #define IO_GAP(fio, lio) (-IO_SPAN(lio, fio))
519 518
520 519 static zio_t *
521 520 vdev_queue_aggregate(vdev_queue_t *vq, zio_t *zio)
522 521 {
523 522 zio_t *first, *last, *aio, *dio, *mandatory, *nio;
524 523 zio_link_t *zl = NULL;
525 524 uint64_t maxgap = 0;
526 525 uint64_t size;
527 526 boolean_t stretch = B_FALSE;
528 527 avl_tree_t *t = vdev_queue_type_tree(vq, zio->io_type);
529 528 enum zio_flag flags = zio->io_flags & ZIO_FLAG_AGG_INHERIT;
530 529
531 530 if (zio->io_flags & ZIO_FLAG_DONT_AGGREGATE)
532 531 return (NULL);
533 532
534 533 first = last = zio;
535 534
536 535 if (zio->io_type == ZIO_TYPE_READ)
537 536 maxgap = zfs_vdev_read_gap_limit;
538 537
539 538 /*
540 539 * We can aggregate I/Os that are sufficiently adjacent and of
541 540 * the same flavor, as expressed by the AGG_INHERIT flags.
542 541 * The latter requirement is necessary so that certain
543 542 * attributes of the I/O, such as whether it's a normal I/O
544 543 * or a scrub/resilver, can be preserved in the aggregate.
545 544 * We can include optional I/Os, but don't allow them
546 545 * to begin a range as they add no benefit in that situation.
547 546 */
548 547
549 548 /*
550 549 * We keep track of the last non-optional I/O.
551 550 */
552 551 mandatory = (first->io_flags & ZIO_FLAG_OPTIONAL) ? NULL : first;
553 552
554 553 /*
555 554 * Walk backwards through sufficiently contiguous I/Os
556 555 * recording the last non-optional I/O.
557 556 */
558 557 while ((dio = AVL_PREV(t, first)) != NULL &&
559 558 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
560 559 IO_SPAN(dio, last) <= zfs_vdev_aggregation_limit &&
561 560 IO_GAP(dio, first) <= maxgap &&
562 561 dio->io_type == zio->io_type) {
563 562 first = dio;
564 563 if (mandatory == NULL && !(first->io_flags & ZIO_FLAG_OPTIONAL))
565 564 mandatory = first;
566 565 }
567 566
568 567 /*
569 568 * Skip any initial optional I/Os.
570 569 */
571 570 while ((first->io_flags & ZIO_FLAG_OPTIONAL) && first != last) {
572 571 first = AVL_NEXT(t, first);
573 572 ASSERT(first != NULL);
574 573 }
575 574
576 575 /*
577 576 * Walk forward through sufficiently contiguous I/Os.
578 577 * The aggregation limit does not apply to optional i/os, so that
579 578 * we can issue contiguous writes even if they are larger than the
580 579 * aggregation limit.
581 580 */
582 581 while ((dio = AVL_NEXT(t, last)) != NULL &&
583 582 (dio->io_flags & ZIO_FLAG_AGG_INHERIT) == flags &&
584 583 (IO_SPAN(first, dio) <= zfs_vdev_aggregation_limit ||
585 584 (dio->io_flags & ZIO_FLAG_OPTIONAL)) &&
586 585 IO_GAP(last, dio) <= maxgap &&
587 586 dio->io_type == zio->io_type) {
588 587 last = dio;
589 588 if (!(last->io_flags & ZIO_FLAG_OPTIONAL))
590 589 mandatory = last;
591 590 }
592 591
593 592 /*
594 593 * Now that we've established the range of the I/O aggregation
595 594 * we must decide what to do with trailing optional I/Os.
596 595 * For reads, there's nothing to do. While we are unable to
597 596 * aggregate further, it's possible that a trailing optional
598 597 * I/O would allow the underlying device to aggregate with
599 598 * subsequent I/Os. We must therefore determine if the next
600 599 * non-optional I/O is close enough to make aggregation
601 600 * worthwhile.
602 601 */
603 602 if (zio->io_type == ZIO_TYPE_WRITE && mandatory != NULL) {
604 603 zio_t *nio = last;
605 604 while ((dio = AVL_NEXT(t, nio)) != NULL &&
606 605 IO_GAP(nio, dio) == 0 &&
607 606 IO_GAP(mandatory, dio) <= zfs_vdev_write_gap_limit) {
608 607 nio = dio;
609 608 if (!(nio->io_flags & ZIO_FLAG_OPTIONAL)) {
610 609 stretch = B_TRUE;
611 610 break;
612 611 }
613 612 }
614 613 }
615 614
616 615 if (stretch) {
617 616 /*
618 617 * We are going to include an optional io in our aggregated
619 618 * span, thus closing the write gap. Only mandatory i/os can
620 619 * start aggregated spans, so make sure that the next i/o
621 620 * after our span is mandatory.
622 621 */
623 622 dio = AVL_NEXT(t, last);
624 623 dio->io_flags &= ~ZIO_FLAG_OPTIONAL;
625 624 } else {
626 625 /* do not include the optional i/o */
627 626 while (last != mandatory && last != first) {
628 627 ASSERT(last->io_flags & ZIO_FLAG_OPTIONAL);
629 628 last = AVL_PREV(t, last);
630 629 ASSERT(last != NULL);
631 630 }
632 631 }
633 632
634 633 if (first == last)
635 634 return (NULL);
636 635
637 636 size = IO_SPAN(first, last);
638 637 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
639 638
640 639 aio = zio_vdev_delegated_io(first->io_vd, first->io_offset,
641 640 abd_alloc_for_io(size, B_TRUE), size, first->io_type,
642 641 zio->io_priority, flags | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE,
643 642 vdev_queue_agg_io_done, NULL);
644 643 aio->io_timestamp = first->io_timestamp;
645 644
646 645 nio = first;
647 646 do {
648 647 dio = nio;
649 648 nio = AVL_NEXT(t, dio);
650 649 ASSERT3U(dio->io_type, ==, aio->io_type);
651 650
652 651 if (dio->io_flags & ZIO_FLAG_NODATA) {
653 652 ASSERT3U(dio->io_type, ==, ZIO_TYPE_WRITE);
654 653 abd_zero_off(aio->io_abd,
655 654 dio->io_offset - aio->io_offset, dio->io_size);
656 655 } else if (dio->io_type == ZIO_TYPE_WRITE) {
657 656 abd_copy_off(aio->io_abd, dio->io_abd,
658 657 dio->io_offset - aio->io_offset, 0, dio->io_size);
659 658 }
660 659
661 660 zio_add_child(dio, aio);
662 661 vdev_queue_io_remove(vq, dio);
663 662 } while (dio != last);
664 663
665 664 /*
666 665 * We need to drop the vdev queue's lock to avoid a deadlock that we
667 666 * could encounter since this I/O will complete immediately.
668 667 */
669 668 mutex_exit(&vq->vq_lock);
670 669 while ((dio = zio_walk_parents(aio, &zl)) != NULL) {
671 670 zio_vdev_io_bypass(dio);
672 671 zio_execute(dio);
673 672 }
674 673 mutex_enter(&vq->vq_lock);
675 674
676 675 return (aio);
677 676 }
678 677
679 678 static zio_t *
680 679 vdev_queue_io_to_issue(vdev_queue_t *vq)
681 680 {
682 681 zio_t *zio, *aio;
683 682 zio_priority_t p;
684 683 avl_index_t idx;
685 684 avl_tree_t *tree;
686 685 zio_t search;
687 686
688 687 again:
689 688 ASSERT(MUTEX_HELD(&vq->vq_lock));
690 689
691 690 p = vdev_queue_class_to_issue(vq);
692 691
693 692 if (p == ZIO_PRIORITY_NUM_QUEUEABLE) {
694 693 /* No eligible queued i/os */
695 694 return (NULL);
696 695 }
697 696
698 697 /*
699 698 * For LBA-ordered queues (async / scrub / initializing), issue the
700 699 * i/o which follows the most recently issued i/o in LBA (offset) order.
701 700 *
702 701 * For FIFO queues (sync), issue the i/o with the lowest timestamp.
703 702 */
704 703 tree = vdev_queue_class_tree(vq, p);
705 704 search.io_timestamp = 0;
706 705 search.io_offset = vq->vq_last_offset - 1;
707 706 VERIFY3P(avl_find(tree, &search, &idx), ==, NULL);
708 707 zio = avl_nearest(tree, idx, AVL_AFTER);
709 708 if (zio == NULL)
710 709 zio = avl_first(tree);
711 710 ASSERT3U(zio->io_priority, ==, p);
712 711
713 712 aio = vdev_queue_aggregate(vq, zio);
714 713 if (aio != NULL)
715 714 zio = aio;
716 715 else
717 716 vdev_queue_io_remove(vq, zio);
718 717
719 718 /*
720 719 * If the I/O is or was optional and therefore has no data, we need to
721 720 * simply discard it. We need to drop the vdev queue's lock to avoid a
722 721 * deadlock that we could encounter since this I/O will complete
723 722 * immediately.
724 723 */
725 724 if (zio->io_flags & ZIO_FLAG_NODATA) {
726 725 mutex_exit(&vq->vq_lock);
727 726 zio_vdev_io_bypass(zio);
728 727 zio_execute(zio);
729 728 mutex_enter(&vq->vq_lock);
730 729 goto again;
731 730 }
732 731
733 732 vdev_queue_pending_add(vq, zio);
734 733 vq->vq_last_offset = zio->io_offset + zio->io_size;
735 734
736 735 return (zio);
737 736 }
738 737
739 738 zio_t *
740 739 vdev_queue_io(zio_t *zio)
741 740 {
742 741 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
743 742 zio_t *nio;
744 743
745 744 if (zio->io_flags & ZIO_FLAG_DONT_QUEUE)
746 745 return (zio);
747 746
748 747 /*
749 748 * Children i/os inherent their parent's priority, which might
750 749 * not match the child's i/o type. Fix it up here.
751 750 */
752 751 if (zio->io_type == ZIO_TYPE_READ) {
753 752 if (zio->io_priority != ZIO_PRIORITY_SYNC_READ &&
754 753 zio->io_priority != ZIO_PRIORITY_ASYNC_READ &&
755 754 zio->io_priority != ZIO_PRIORITY_SCRUB &&
756 755 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
757 756 zio->io_priority != ZIO_PRIORITY_INITIALIZING)
758 757 zio->io_priority = ZIO_PRIORITY_ASYNC_READ;
759 758 } else {
760 759 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
761 760 if (zio->io_priority != ZIO_PRIORITY_SYNC_WRITE &&
762 761 zio->io_priority != ZIO_PRIORITY_ASYNC_WRITE &&
763 762 zio->io_priority != ZIO_PRIORITY_REMOVAL &&
764 763 zio->io_priority != ZIO_PRIORITY_INITIALIZING)
765 764 zio->io_priority = ZIO_PRIORITY_ASYNC_WRITE;
766 765 }
767 766
768 767 zio->io_flags |= ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_QUEUE;
769 768
770 769 mutex_enter(&vq->vq_lock);
771 770 zio->io_timestamp = gethrtime();
772 771 vdev_queue_io_add(vq, zio);
773 772 nio = vdev_queue_io_to_issue(vq);
774 773 mutex_exit(&vq->vq_lock);
775 774
776 775 if (nio == NULL)
777 776 return (NULL);
778 777
779 778 if (nio->io_done == vdev_queue_agg_io_done) {
780 779 zio_nowait(nio);
781 780 return (NULL);
782 781 }
783 782
784 783 return (nio);
785 784 }
786 785
787 786 void
788 787 vdev_queue_io_done(zio_t *zio)
789 788 {
790 789 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
791 790 zio_t *nio;
792 791
793 792 mutex_enter(&vq->vq_lock);
794 793
795 794 vdev_queue_pending_remove(vq, zio);
796 795
797 796 vq->vq_io_complete_ts = gethrtime();
798 797
799 798 while ((nio = vdev_queue_io_to_issue(vq)) != NULL) {
800 799 mutex_exit(&vq->vq_lock);
801 800 if (nio->io_done == vdev_queue_agg_io_done) {
802 801 zio_nowait(nio);
803 802 } else {
804 803 zio_vdev_io_reissue(nio);
805 804 zio_execute(nio);
806 805 }
807 806 mutex_enter(&vq->vq_lock);
808 807 }
809 808
810 809 mutex_exit(&vq->vq_lock);
811 810 }
812 811
813 812 void
814 813 vdev_queue_change_io_priority(zio_t *zio, zio_priority_t priority)
815 814 {
816 815 vdev_queue_t *vq = &zio->io_vd->vdev_queue;
817 816 avl_tree_t *tree;
818 817
819 818 /*
820 819 * ZIO_PRIORITY_NOW is used by the vdev cache code and the aggregate zio
821 820 * code to issue IOs without adding them to the vdev queue. In this
822 821 * case, the zio is already going to be issued as quickly as possible
823 822 * and so it doesn't need any reprioitization to help.
824 823 */
825 824 if (zio->io_priority == ZIO_PRIORITY_NOW)
826 825 return;
827 826
828 827 ASSERT3U(zio->io_priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
829 828 ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE);
830 829
831 830 if (zio->io_type == ZIO_TYPE_READ) {
832 831 if (priority != ZIO_PRIORITY_SYNC_READ &&
833 832 priority != ZIO_PRIORITY_ASYNC_READ &&
834 833 priority != ZIO_PRIORITY_SCRUB)
835 834 priority = ZIO_PRIORITY_ASYNC_READ;
836 835 } else {
837 836 ASSERT(zio->io_type == ZIO_TYPE_WRITE);
838 837 if (priority != ZIO_PRIORITY_SYNC_WRITE &&
839 838 priority != ZIO_PRIORITY_ASYNC_WRITE)
840 839 priority = ZIO_PRIORITY_ASYNC_WRITE;
841 840 }
842 841
843 842 mutex_enter(&vq->vq_lock);
844 843
845 844 /*
846 845 * If the zio is in none of the queues we can simply change
847 846 * the priority. If the zio is waiting to be submitted we must
848 847 * remove it from the queue and re-insert it with the new priority.
849 848 * Otherwise, the zio is currently active and we cannot change its
850 849 * priority.
851 850 */
852 851 tree = vdev_queue_class_tree(vq, zio->io_priority);
853 852 if (avl_find(tree, zio, NULL) == zio) {
854 853 spa_t *spa = zio->io_spa;
855 854 zio_priority_t oldpri = zio->io_priority;
856 855
857 856 avl_remove(vdev_queue_class_tree(vq, zio->io_priority), zio);
858 857 zio->io_priority = priority;
859 858 avl_add(vdev_queue_class_tree(vq, zio->io_priority), zio);
860 859
861 860 mutex_enter(&spa->spa_iokstat_lock);
862 861 ASSERT3U(spa->spa_queue_stats[oldpri].spa_queued, >, 0);
863 862 spa->spa_queue_stats[oldpri].spa_queued--;
864 863 spa->spa_queue_stats[zio->io_priority].spa_queued++;
865 864 mutex_exit(&spa->spa_iokstat_lock);
866 865 } else if (avl_find(&vq->vq_active_tree, zio, NULL) != zio) {
867 866 zio->io_priority = priority;
868 867 }
869 868
870 869 mutex_exit(&vq->vq_lock);
871 870 }
872 871
873 872 /*
874 873 * As these two methods are only used for load calculations we're not
875 874 * concerned if we get an incorrect value on 32bit platforms due to lack of
876 875 * vq_lock mutex use here, instead we prefer to keep it lock free for
877 876 * performance.
878 877 */
879 878 int
880 879 vdev_queue_length(vdev_t *vd)
881 880 {
882 881 return (avl_numnodes(&vd->vdev_queue.vq_active_tree));
883 882 }
884 883
885 884 uint64_t
886 885 vdev_queue_last_offset(vdev_t *vd)
887 886 {
888 887 return (vd->vdev_queue.vq_last_offset);
889 888 }
|
↓ open down ↓ |
408 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX