Commit | Line | Data |
---|---|---|
bfbdd35b BVA |
1 | /* |
2 | * Copyright (C) 2018 Western Digital Corporation or its affiliates. | |
3 | * | |
4 | * This file is released under the GPL. | |
5 | */ | |
6 | ||
7 | #include <errno.h> | |
8 | #include <string.h> | |
9 | #include <stdlib.h> | |
10 | #include <dirent.h> | |
11 | #include <fcntl.h> | |
12 | #include <sys/ioctl.h> | |
13 | #include <sys/stat.h> | |
14 | #include <unistd.h> | |
15 | #include <linux/blkzoned.h> | |
16 | #include "file.h" | |
17 | #include "fio.h" | |
18 | #include "lib/pow2.h" | |
19 | #include "log.h" | |
20 | #include "smalloc.h" | |
21 | #include "verify.h" | |
22 | #include "zbd.h" | |
23 | ||
24 | /** | |
25 | * zbd_zone_idx - convert an offset into a zone number | |
26 | * @f: file pointer. | |
27 | * @offset: offset in bytes. If this offset is in the first zone_size bytes | |
28 | * past the disk size then the index of the sentinel is returned. | |
29 | */ | |
30 | static uint32_t zbd_zone_idx(const struct fio_file *f, uint64_t offset) | |
31 | { | |
32 | uint32_t zone_idx; | |
33 | ||
cd775e06 | 34 | if (f->zbd_info->zone_size_log2 > 0) |
bfbdd35b BVA |
35 | zone_idx = offset >> f->zbd_info->zone_size_log2; |
36 | else | |
ee3696bd | 37 | zone_idx = offset / f->zbd_info->zone_size; |
bfbdd35b BVA |
38 | |
39 | return min(zone_idx, f->zbd_info->nr_zones); | |
40 | } | |
41 | ||
42 | /** | |
43 | * zbd_zone_full - verify whether a minimum number of bytes remain in a zone | |
44 | * @f: file pointer. | |
45 | * @z: zone info pointer. | |
46 | * @required: minimum number of bytes that must remain in a zone. | |
47 | * | |
48 | * The caller must hold z->mutex. | |
49 | */ | |
50 | static bool zbd_zone_full(const struct fio_file *f, struct fio_zone_info *z, | |
51 | uint64_t required) | |
52 | { | |
53 | assert((required & 511) == 0); | |
54 | ||
55 | return z->type == BLK_ZONE_TYPE_SEQWRITE_REQ && | |
ee3696bd | 56 | z->wp + required > z->start + f->zbd_info->zone_size; |
bfbdd35b BVA |
57 | } |
58 | ||
59 | static bool is_valid_offset(const struct fio_file *f, uint64_t offset) | |
60 | { | |
61 | return (uint64_t)(offset - f->file_offset) < f->io_size; | |
62 | } | |
63 | ||
64 | /* Verify whether direct I/O is used for all host-managed zoned drives. */ | |
65 | static bool zbd_using_direct_io(void) | |
66 | { | |
67 | struct thread_data *td; | |
68 | struct fio_file *f; | |
69 | int i, j; | |
70 | ||
71 | for_each_td(td, i) { | |
72 | if (td->o.odirect || !(td->o.td_ddir & TD_DDIR_WRITE)) | |
73 | continue; | |
74 | for_each_file(td, f, j) { | |
75 | if (f->zbd_info && | |
76 | f->zbd_info->model == ZBD_DM_HOST_MANAGED) | |
77 | return false; | |
78 | } | |
79 | } | |
80 | ||
81 | return true; | |
82 | } | |
83 | ||
84 | /* Whether or not the I/O range for f includes one or more sequential zones */ | |
85 | static bool zbd_is_seq_job(struct fio_file *f) | |
86 | { | |
87 | uint32_t zone_idx, zone_idx_b, zone_idx_e; | |
88 | ||
89 | assert(f->zbd_info); | |
90 | if (f->io_size == 0) | |
91 | return false; | |
92 | zone_idx_b = zbd_zone_idx(f, f->file_offset); | |
93 | zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size - 1); | |
94 | for (zone_idx = zone_idx_b; zone_idx <= zone_idx_e; zone_idx++) | |
95 | if (f->zbd_info->zone_info[zone_idx].type == | |
96 | BLK_ZONE_TYPE_SEQWRITE_REQ) | |
97 | return true; | |
98 | ||
99 | return false; | |
100 | } | |
101 | ||
102 | /* | |
103 | * Verify whether offset and size parameters are aligned with zone boundaries. | |
104 | */ | |
105 | static bool zbd_verify_sizes(void) | |
106 | { | |
107 | const struct fio_zone_info *z; | |
108 | struct thread_data *td; | |
109 | struct fio_file *f; | |
110 | uint64_t new_offset, new_end; | |
111 | uint32_t zone_idx; | |
112 | int i, j; | |
113 | ||
114 | for_each_td(td, i) { | |
115 | for_each_file(td, f, j) { | |
116 | if (!f->zbd_info) | |
117 | continue; | |
118 | if (f->file_offset >= f->real_file_size) | |
119 | continue; | |
120 | if (!zbd_is_seq_job(f)) | |
121 | continue; | |
122 | zone_idx = zbd_zone_idx(f, f->file_offset); | |
123 | z = &f->zbd_info->zone_info[zone_idx]; | |
ee3696bd DLM |
124 | if (f->file_offset != z->start) { |
125 | new_offset = (z+1)->start; | |
bfbdd35b BVA |
126 | if (new_offset >= f->file_offset + f->io_size) { |
127 | log_info("%s: io_size must be at least one zone\n", | |
128 | f->file_name); | |
129 | return false; | |
130 | } | |
a0c84dd4 JA |
131 | log_info("%s: rounded up offset from %llu to %llu\n", |
132 | f->file_name, (unsigned long long) f->file_offset, | |
133 | (unsigned long long) new_offset); | |
bfbdd35b BVA |
134 | f->io_size -= (new_offset - f->file_offset); |
135 | f->file_offset = new_offset; | |
136 | } | |
137 | zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size); | |
138 | z = &f->zbd_info->zone_info[zone_idx]; | |
ee3696bd | 139 | new_end = z->start; |
bfbdd35b BVA |
140 | if (f->file_offset + f->io_size != new_end) { |
141 | if (new_end <= f->file_offset) { | |
142 | log_info("%s: io_size must be at least one zone\n", | |
143 | f->file_name); | |
144 | return false; | |
145 | } | |
a0c84dd4 JA |
146 | log_info("%s: rounded down io_size from %llu to %llu\n", |
147 | f->file_name, (unsigned long long) f->io_size, | |
148 | (unsigned long long) new_end - f->file_offset); | |
bfbdd35b BVA |
149 | f->io_size = new_end - f->file_offset; |
150 | } | |
151 | } | |
152 | } | |
153 | ||
154 | return true; | |
155 | } | |
156 | ||
157 | static bool zbd_verify_bs(void) | |
158 | { | |
159 | struct thread_data *td; | |
160 | struct fio_file *f; | |
161 | uint32_t zone_size; | |
162 | int i, j, k; | |
163 | ||
164 | for_each_td(td, i) { | |
165 | for_each_file(td, f, j) { | |
166 | if (!f->zbd_info) | |
167 | continue; | |
168 | zone_size = f->zbd_info->zone_size; | |
169 | for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) { | |
170 | if (td->o.verify != VERIFY_NONE && | |
ee3696bd | 171 | zone_size % td->o.bs[k] != 0) { |
bfbdd35b BVA |
172 | log_info("%s: block size %llu is not a divisor of the zone size %d\n", |
173 | f->file_name, td->o.bs[k], | |
ee3696bd | 174 | zone_size); |
bfbdd35b BVA |
175 | return false; |
176 | } | |
177 | } | |
178 | } | |
179 | } | |
180 | return true; | |
181 | } | |
182 | ||
183 | /* | |
184 | * Read zone information into @buf starting from sector @start_sector. | |
185 | * @fd is a file descriptor that refers to a block device and @bufsz is the | |
186 | * size of @buf. | |
187 | * | |
188 | * Returns 0 upon success and a negative error code upon failure. | |
189 | */ | |
190 | static int read_zone_info(int fd, uint64_t start_sector, | |
191 | void *buf, unsigned int bufsz) | |
192 | { | |
193 | struct blk_zone_report *hdr = buf; | |
194 | ||
195 | if (bufsz < sizeof(*hdr)) | |
196 | return -EINVAL; | |
197 | ||
198 | memset(hdr, 0, sizeof(*hdr)); | |
199 | ||
200 | hdr->nr_zones = (bufsz - sizeof(*hdr)) / sizeof(struct blk_zone); | |
201 | hdr->sector = start_sector; | |
202 | return ioctl(fd, BLKREPORTZONE, hdr) >= 0 ? 0 : -errno; | |
203 | } | |
204 | ||
205 | /* | |
206 | * Read up to 255 characters from the first line of a file. Strip the trailing | |
207 | * newline. | |
208 | */ | |
209 | static char *read_file(const char *path) | |
210 | { | |
211 | char line[256], *p = line; | |
212 | FILE *f; | |
213 | ||
214 | f = fopen(path, "rb"); | |
215 | if (!f) | |
216 | return NULL; | |
217 | if (!fgets(line, sizeof(line), f)) | |
218 | line[0] = '\0'; | |
219 | strsep(&p, "\n"); | |
220 | fclose(f); | |
221 | ||
222 | return strdup(line); | |
223 | } | |
224 | ||
225 | static enum blk_zoned_model get_zbd_model(const char *file_name) | |
226 | { | |
227 | enum blk_zoned_model model = ZBD_DM_NONE; | |
228 | char *zoned_attr_path = NULL; | |
229 | char *model_str = NULL; | |
230 | struct stat statbuf; | |
231 | ||
232 | if (stat(file_name, &statbuf) < 0) | |
233 | goto out; | |
234 | if (asprintf(&zoned_attr_path, "/sys/dev/block/%d:%d/queue/zoned", | |
235 | major(statbuf.st_rdev), minor(statbuf.st_rdev)) < 0) | |
236 | goto out; | |
237 | model_str = read_file(zoned_attr_path); | |
238 | if (!model_str) | |
239 | goto out; | |
240 | dprint(FD_ZBD, "%s: zbd model string: %s\n", file_name, model_str); | |
241 | if (strcmp(model_str, "host-aware") == 0) | |
242 | model = ZBD_DM_HOST_AWARE; | |
243 | else if (strcmp(model_str, "host-managed") == 0) | |
244 | model = ZBD_DM_HOST_MANAGED; | |
245 | ||
246 | out: | |
247 | free(model_str); | |
248 | free(zoned_attr_path); | |
249 | return model; | |
250 | } | |
251 | ||
252 | static int ilog2(uint64_t i) | |
253 | { | |
254 | int log = -1; | |
255 | ||
256 | while (i) { | |
257 | i >>= 1; | |
258 | log++; | |
259 | } | |
260 | return log; | |
261 | } | |
262 | ||
263 | /* | |
264 | * Initialize f->zbd_info for devices that are not zoned block devices. This | |
265 | * allows to execute a ZBD workload against a non-ZBD device. | |
266 | */ | |
267 | static int init_zone_info(struct thread_data *td, struct fio_file *f) | |
268 | { | |
269 | uint32_t nr_zones; | |
270 | struct fio_zone_info *p; | |
271 | uint64_t zone_size; | |
272 | struct zoned_block_device_info *zbd_info = NULL; | |
273 | pthread_mutexattr_t attr; | |
274 | int i; | |
275 | ||
ee3696bd | 276 | zone_size = td->o.zone_size; |
bfbdd35b | 277 | assert(zone_size); |
ee3696bd | 278 | nr_zones = (f->real_file_size + zone_size - 1) / zone_size; |
bfbdd35b BVA |
279 | zbd_info = scalloc(1, sizeof(*zbd_info) + |
280 | (nr_zones + 1) * sizeof(zbd_info->zone_info[0])); | |
281 | if (!zbd_info) | |
282 | return -ENOMEM; | |
283 | ||
284 | pthread_mutexattr_init(&attr); | |
285 | pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE); | |
286 | pthread_mutexattr_setpshared(&attr, true); | |
287 | pthread_mutex_init(&zbd_info->mutex, &attr); | |
288 | zbd_info->refcount = 1; | |
289 | p = &zbd_info->zone_info[0]; | |
290 | for (i = 0; i < nr_zones; i++, p++) { | |
291 | pthread_mutex_init(&p->mutex, &attr); | |
292 | p->start = i * zone_size; | |
293 | p->wp = p->start + zone_size; | |
294 | p->type = BLK_ZONE_TYPE_SEQWRITE_REQ; | |
295 | p->cond = BLK_ZONE_COND_EMPTY; | |
296 | } | |
297 | /* a sentinel */ | |
298 | p->start = nr_zones * zone_size; | |
299 | ||
300 | f->zbd_info = zbd_info; | |
301 | f->zbd_info->zone_size = zone_size; | |
302 | f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ? | |
ee3696bd | 303 | ilog2(zone_size) : -1; |
bfbdd35b BVA |
304 | f->zbd_info->nr_zones = nr_zones; |
305 | pthread_mutexattr_destroy(&attr); | |
306 | return 0; | |
307 | } | |
308 | ||
309 | /* | |
310 | * Parse the BLKREPORTZONE output and store it in f->zbd_info. Must be called | |
311 | * only for devices that support this ioctl, namely zoned block devices. | |
312 | */ | |
313 | static int parse_zone_info(struct thread_data *td, struct fio_file *f) | |
314 | { | |
315 | const unsigned int bufsz = sizeof(struct blk_zone_report) + | |
316 | 4096 * sizeof(struct blk_zone); | |
317 | uint32_t nr_zones; | |
318 | struct blk_zone_report *hdr; | |
319 | const struct blk_zone *z; | |
320 | struct fio_zone_info *p; | |
321 | uint64_t zone_size, start_sector; | |
322 | struct zoned_block_device_info *zbd_info = NULL; | |
323 | pthread_mutexattr_t attr; | |
324 | void *buf; | |
325 | int fd, i, j, ret = 0; | |
326 | ||
327 | pthread_mutexattr_init(&attr); | |
328 | pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE); | |
329 | pthread_mutexattr_setpshared(&attr, true); | |
330 | ||
331 | buf = malloc(bufsz); | |
332 | if (!buf) | |
333 | goto out; | |
334 | ||
335 | fd = open(f->file_name, O_RDONLY | O_LARGEFILE); | |
336 | if (fd < 0) { | |
337 | ret = -errno; | |
338 | goto free; | |
339 | } | |
340 | ||
341 | ret = read_zone_info(fd, 0, buf, bufsz); | |
342 | if (ret < 0) { | |
343 | log_info("fio: BLKREPORTZONE(%lu) failed for %s (%d).\n", | |
344 | 0UL, f->file_name, -ret); | |
345 | goto close; | |
346 | } | |
347 | hdr = buf; | |
348 | if (hdr->nr_zones < 1) { | |
349 | log_info("fio: %s has invalid zone information.\n", | |
350 | f->file_name); | |
351 | goto close; | |
352 | } | |
353 | z = (void *)(hdr + 1); | |
ee3696bd DLM |
354 | zone_size = z->len << 9; |
355 | nr_zones = (f->real_file_size + zone_size - 1) / zone_size; | |
bfbdd35b BVA |
356 | |
357 | if (td->o.zone_size == 0) { | |
ee3696bd DLM |
358 | td->o.zone_size = zone_size; |
359 | } else if (td->o.zone_size != zone_size) { | |
a0c84dd4 JA |
360 | log_info("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n", |
361 | f->file_name, (unsigned long long) td->o.zone_size, | |
362 | (unsigned long long) zone_size); | |
bfbdd35b BVA |
363 | ret = -EINVAL; |
364 | goto close; | |
365 | } | |
366 | ||
a0c84dd4 JA |
367 | dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name, |
368 | nr_zones, (unsigned long long) zone_size / 1024); | |
bfbdd35b BVA |
369 | |
370 | zbd_info = scalloc(1, sizeof(*zbd_info) + | |
371 | (nr_zones + 1) * sizeof(zbd_info->zone_info[0])); | |
372 | ret = -ENOMEM; | |
373 | if (!zbd_info) | |
374 | goto close; | |
375 | pthread_mutex_init(&zbd_info->mutex, &attr); | |
376 | zbd_info->refcount = 1; | |
377 | p = &zbd_info->zone_info[0]; | |
378 | for (start_sector = 0, j = 0; j < nr_zones;) { | |
379 | z = (void *)(hdr + 1); | |
380 | for (i = 0; i < hdr->nr_zones; i++, j++, z++, p++) { | |
381 | pthread_mutex_init(&p->mutex, &attr); | |
ee3696bd | 382 | p->start = z->start << 9; |
bfbdd35b BVA |
383 | switch (z->cond) { |
384 | case BLK_ZONE_COND_NOT_WP: | |
ee3696bd | 385 | p->wp = p->start; |
bfbdd35b BVA |
386 | break; |
387 | case BLK_ZONE_COND_FULL: | |
ee3696bd | 388 | p->wp = p->start + zone_size; |
bfbdd35b BVA |
389 | break; |
390 | default: | |
391 | assert(z->start <= z->wp); | |
ee3696bd DLM |
392 | assert(z->wp <= z->start + (zone_size >> 9)); |
393 | p->wp = z->wp << 9; | |
bfbdd35b BVA |
394 | break; |
395 | } | |
396 | p->type = z->type; | |
397 | p->cond = z->cond; | |
398 | if (j > 0 && p->start != p[-1].start + zone_size) { | |
399 | log_info("%s: invalid zone data\n", | |
400 | f->file_name); | |
401 | ret = -EINVAL; | |
402 | goto close; | |
403 | } | |
404 | } | |
405 | z--; | |
406 | start_sector = z->start + z->len; | |
407 | if (j >= nr_zones) | |
408 | break; | |
409 | ret = read_zone_info(fd, start_sector, buf, bufsz); | |
410 | if (ret < 0) { | |
a0c84dd4 JA |
411 | log_info("fio: BLKREPORTZONE(%llu) failed for %s (%d).\n", |
412 | (unsigned long long) start_sector, f->file_name, -ret); | |
bfbdd35b BVA |
413 | goto close; |
414 | } | |
415 | } | |
416 | /* a sentinel */ | |
ee3696bd | 417 | zbd_info->zone_info[nr_zones].start = start_sector << 9; |
bfbdd35b BVA |
418 | |
419 | f->zbd_info = zbd_info; | |
420 | f->zbd_info->zone_size = zone_size; | |
421 | f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ? | |
ee3696bd | 422 | ilog2(zone_size) : -1; |
bfbdd35b BVA |
423 | f->zbd_info->nr_zones = nr_zones; |
424 | zbd_info = NULL; | |
425 | ret = 0; | |
426 | ||
427 | close: | |
428 | sfree(zbd_info); | |
429 | close(fd); | |
430 | free: | |
431 | free(buf); | |
432 | out: | |
433 | pthread_mutexattr_destroy(&attr); | |
434 | return ret; | |
435 | } | |
436 | ||
437 | /* | |
438 | * Allocate zone information and store it into f->zbd_info if zonemode=zbd. | |
439 | * | |
440 | * Returns 0 upon success and a negative error code upon failure. | |
441 | */ | |
442 | int zbd_create_zone_info(struct thread_data *td, struct fio_file *f) | |
443 | { | |
444 | enum blk_zoned_model zbd_model; | |
445 | int ret = 0; | |
446 | ||
447 | assert(td->o.zone_mode == ZONE_MODE_ZBD); | |
448 | ||
449 | zbd_model = get_zbd_model(f->file_name); | |
450 | switch (zbd_model) { | |
451 | case ZBD_DM_HOST_AWARE: | |
452 | case ZBD_DM_HOST_MANAGED: | |
453 | ret = parse_zone_info(td, f); | |
454 | break; | |
455 | case ZBD_DM_NONE: | |
456 | ret = init_zone_info(td, f); | |
457 | break; | |
458 | } | |
459 | if (ret == 0) | |
460 | f->zbd_info->model = zbd_model; | |
461 | return ret; | |
462 | } | |
463 | ||
464 | void zbd_free_zone_info(struct fio_file *f) | |
465 | { | |
466 | uint32_t refcount; | |
467 | ||
468 | if (!f->zbd_info) | |
469 | return; | |
470 | ||
471 | pthread_mutex_lock(&f->zbd_info->mutex); | |
472 | refcount = --f->zbd_info->refcount; | |
473 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
474 | ||
475 | assert((int32_t)refcount >= 0); | |
476 | if (refcount == 0) | |
477 | sfree(f->zbd_info); | |
478 | f->zbd_info = NULL; | |
479 | } | |
480 | ||
481 | /* | |
482 | * Initialize f->zbd_info. | |
483 | * | |
484 | * Returns 0 upon success and a negative error code upon failure. | |
485 | * | |
486 | * Note: this function can only work correctly if it is called before the first | |
487 | * fio fork() call. | |
488 | */ | |
489 | static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file) | |
490 | { | |
491 | struct thread_data *td2; | |
492 | struct fio_file *f2; | |
493 | int i, j, ret; | |
494 | ||
495 | for_each_td(td2, i) { | |
496 | for_each_file(td2, f2, j) { | |
497 | if (td2 == td && f2 == file) | |
498 | continue; | |
499 | if (!f2->zbd_info || | |
500 | strcmp(f2->file_name, file->file_name) != 0) | |
501 | continue; | |
502 | file->zbd_info = f2->zbd_info; | |
503 | file->zbd_info->refcount++; | |
504 | return 0; | |
505 | } | |
506 | } | |
507 | ||
508 | ret = zbd_create_zone_info(td, file); | |
509 | if (ret < 0) | |
510 | td_verror(td, -ret, "BLKREPORTZONE failed"); | |
511 | return ret; | |
512 | } | |
513 | ||
514 | int zbd_init(struct thread_data *td) | |
515 | { | |
516 | struct fio_file *f; | |
517 | int i; | |
518 | ||
519 | for_each_file(td, f, i) { | |
520 | if (f->filetype != FIO_TYPE_BLOCK) | |
521 | continue; | |
522 | if (td->o.zone_size && td->o.zone_size < 512) { | |
523 | log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n", | |
524 | f->file_name); | |
525 | return 1; | |
526 | } | |
527 | if (td->o.zone_size == 0 && | |
528 | get_zbd_model(f->file_name) == ZBD_DM_NONE) { | |
529 | log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n", | |
530 | f->file_name); | |
531 | return 1; | |
532 | } | |
533 | zbd_init_zone_info(td, f); | |
534 | } | |
535 | ||
536 | if (!zbd_using_direct_io()) { | |
537 | log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n"); | |
538 | return 1; | |
539 | } | |
540 | ||
541 | if (!zbd_verify_sizes()) | |
542 | return 1; | |
543 | ||
544 | if (!zbd_verify_bs()) | |
545 | return 1; | |
546 | ||
547 | return 0; | |
548 | } | |
549 | ||
550 | /** | |
551 | * zbd_reset_range - reset zones for a range of sectors | |
552 | * @td: FIO thread data. | |
553 | * @f: Fio file for which to reset zones | |
554 | * @sector: Starting sector in units of 512 bytes | |
555 | * @nr_sectors: Number of sectors in units of 512 bytes | |
556 | * | |
557 | * Returns 0 upon success and a negative error code upon failure. | |
558 | */ | |
559 | static int zbd_reset_range(struct thread_data *td, const struct fio_file *f, | |
ee3696bd | 560 | uint64_t offset, uint64_t length) |
bfbdd35b BVA |
561 | { |
562 | struct blk_zone_range zr = { | |
ee3696bd DLM |
563 | .sector = offset >> 9, |
564 | .nr_sectors = length >> 9, | |
bfbdd35b BVA |
565 | }; |
566 | uint32_t zone_idx_b, zone_idx_e; | |
567 | struct fio_zone_info *zb, *ze, *z; | |
568 | int ret = 0; | |
569 | ||
570 | assert(f->fd != -1); | |
ee3696bd | 571 | assert(is_valid_offset(f, offset + length - 1)); |
bfbdd35b BVA |
572 | switch (f->zbd_info->model) { |
573 | case ZBD_DM_HOST_AWARE: | |
574 | case ZBD_DM_HOST_MANAGED: | |
575 | ret = ioctl(f->fd, BLKRESETZONE, &zr); | |
576 | if (ret < 0) { | |
577 | td_verror(td, errno, "resetting wp failed"); | |
578 | log_err("%s: resetting wp for %llu sectors at sector %llu failed (%d).\n", | |
579 | f->file_name, zr.nr_sectors, zr.sector, errno); | |
580 | return ret; | |
581 | } | |
582 | break; | |
583 | case ZBD_DM_NONE: | |
584 | break; | |
585 | } | |
586 | ||
ee3696bd | 587 | zone_idx_b = zbd_zone_idx(f, offset); |
bfbdd35b | 588 | zb = &f->zbd_info->zone_info[zone_idx_b]; |
ee3696bd | 589 | zone_idx_e = zbd_zone_idx(f, offset + length); |
bfbdd35b BVA |
590 | ze = &f->zbd_info->zone_info[zone_idx_e]; |
591 | for (z = zb; z < ze; z++) { | |
592 | pthread_mutex_lock(&z->mutex); | |
a7c2b6fc BVA |
593 | pthread_mutex_lock(&f->zbd_info->mutex); |
594 | f->zbd_info->sectors_with_data -= z->wp - z->start; | |
595 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
bfbdd35b BVA |
596 | z->wp = z->start; |
597 | z->verify_block = 0; | |
598 | pthread_mutex_unlock(&z->mutex); | |
599 | } | |
600 | ||
fd5d733f BVA |
601 | td->ts.nr_zone_resets += ze - zb; |
602 | ||
bfbdd35b BVA |
603 | return ret; |
604 | } | |
605 | ||
a0c84dd4 JA |
606 | static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info, |
607 | struct fio_zone_info *zone) | |
608 | { | |
089ddd95 | 609 | return zone - zbd_info->zone_info; |
a0c84dd4 JA |
610 | } |
611 | ||
bfbdd35b BVA |
612 | /** |
613 | * zbd_reset_zone - reset the write pointer of a single zone | |
614 | * @td: FIO thread data. | |
615 | * @f: FIO file associated with the disk for which to reset a write pointer. | |
616 | * @z: Zone to reset. | |
617 | * | |
618 | * Returns 0 upon success and a negative error code upon failure. | |
619 | */ | |
620 | static int zbd_reset_zone(struct thread_data *td, const struct fio_file *f, | |
621 | struct fio_zone_info *z) | |
622 | { | |
a0c84dd4 JA |
623 | dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name, |
624 | zbd_zone_nr(f->zbd_info, z)); | |
d60be7d5 DLM |
625 | |
626 | return zbd_reset_range(td, f, z->start, (z+1)->start - z->start); | |
bfbdd35b BVA |
627 | } |
628 | ||
629 | /* | |
630 | * Reset a range of zones. Returns 0 upon success and 1 upon failure. | |
631 | * @td: fio thread data. | |
632 | * @f: fio file for which to reset zones | |
633 | * @zb: first zone to reset. | |
634 | * @ze: first zone not to reset. | |
635 | * @all_zones: whether to reset all zones or only those zones for which the | |
636 | * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE]. | |
637 | */ | |
638 | static int zbd_reset_zones(struct thread_data *td, struct fio_file *f, | |
639 | struct fio_zone_info *const zb, | |
640 | struct fio_zone_info *const ze, bool all_zones) | |
641 | { | |
642 | struct fio_zone_info *z, *start_z = ze; | |
ee3696bd | 643 | const uint32_t min_bs = td->o.min_bs[DDIR_WRITE]; |
bfbdd35b BVA |
644 | bool reset_wp; |
645 | int res = 0; | |
646 | ||
a0c84dd4 JA |
647 | dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name, |
648 | zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze)); | |
bfbdd35b BVA |
649 | assert(f->fd != -1); |
650 | for (z = zb; z < ze; z++) { | |
651 | pthread_mutex_lock(&z->mutex); | |
652 | switch (z->type) { | |
653 | case BLK_ZONE_TYPE_SEQWRITE_REQ: | |
654 | reset_wp = all_zones ? z->wp != z->start : | |
655 | (td->o.td_ddir & TD_DDIR_WRITE) && | |
656 | z->wp % min_bs != 0; | |
657 | if (start_z == ze && reset_wp) { | |
658 | start_z = z; | |
659 | } else if (start_z < ze && !reset_wp) { | |
660 | dprint(FD_ZBD, | |
a0c84dd4 | 661 | "%s: resetting zones %u .. %u\n", |
bfbdd35b | 662 | f->file_name, |
a0c84dd4 JA |
663 | zbd_zone_nr(f->zbd_info, start_z), |
664 | zbd_zone_nr(f->zbd_info, z)); | |
bfbdd35b BVA |
665 | if (zbd_reset_range(td, f, start_z->start, |
666 | z->start - start_z->start) < 0) | |
667 | res = 1; | |
668 | start_z = ze; | |
669 | } | |
670 | break; | |
671 | default: | |
672 | if (start_z == ze) | |
673 | break; | |
a0c84dd4 JA |
674 | dprint(FD_ZBD, "%s: resetting zones %u .. %u\n", |
675 | f->file_name, zbd_zone_nr(f->zbd_info, start_z), | |
676 | zbd_zone_nr(f->zbd_info, z)); | |
bfbdd35b BVA |
677 | if (zbd_reset_range(td, f, start_z->start, |
678 | z->start - start_z->start) < 0) | |
679 | res = 1; | |
680 | start_z = ze; | |
681 | break; | |
682 | } | |
683 | } | |
684 | if (start_z < ze) { | |
a0c84dd4 JA |
685 | dprint(FD_ZBD, "%s: resetting zones %u .. %u\n", f->file_name, |
686 | zbd_zone_nr(f->zbd_info, start_z), | |
687 | zbd_zone_nr(f->zbd_info, z)); | |
bfbdd35b BVA |
688 | if (zbd_reset_range(td, f, start_z->start, |
689 | z->start - start_z->start) < 0) | |
690 | res = 1; | |
691 | } | |
692 | for (z = zb; z < ze; z++) | |
693 | pthread_mutex_unlock(&z->mutex); | |
694 | ||
695 | return res; | |
696 | } | |
697 | ||
a7c2b6fc BVA |
698 | /* |
699 | * Reset zbd_info.write_cnt, the counter that counts down towards the next | |
700 | * zone reset. | |
701 | */ | |
702 | static void zbd_reset_write_cnt(const struct thread_data *td, | |
703 | const struct fio_file *f) | |
704 | { | |
705 | assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1); | |
706 | ||
707 | pthread_mutex_lock(&f->zbd_info->mutex); | |
708 | f->zbd_info->write_cnt = td->o.zrf.u.f ? | |
709 | min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX; | |
710 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
711 | } | |
712 | ||
713 | static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td, | |
714 | const struct fio_file *f) | |
715 | { | |
716 | uint32_t write_cnt = 0; | |
717 | ||
718 | pthread_mutex_lock(&f->zbd_info->mutex); | |
719 | assert(f->zbd_info->write_cnt); | |
720 | if (f->zbd_info->write_cnt) | |
721 | write_cnt = --f->zbd_info->write_cnt; | |
722 | if (write_cnt == 0) | |
723 | zbd_reset_write_cnt(td, f); | |
724 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
725 | ||
726 | return write_cnt == 0; | |
727 | } | |
728 | ||
91d25131 BVA |
729 | enum swd_action { |
730 | CHECK_SWD, | |
731 | SET_SWD, | |
732 | }; | |
733 | ||
734 | /* Calculate the number of sectors with data (swd) and perform action 'a' */ | |
735 | static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a) | |
615555bb | 736 | { |
615555bb BVA |
737 | struct fio_zone_info *zb, *ze, *z; |
738 | uint64_t swd = 0; | |
739 | ||
740 | zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)]; | |
741 | ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset + | |
742 | f->io_size)]; | |
743 | for (z = zb; z < ze; z++) { | |
744 | pthread_mutex_lock(&z->mutex); | |
745 | swd += z->wp - z->start; | |
746 | } | |
747 | pthread_mutex_lock(&f->zbd_info->mutex); | |
91d25131 BVA |
748 | switch (a) { |
749 | case CHECK_SWD: | |
750 | assert(f->zbd_info->sectors_with_data == swd); | |
751 | break; | |
752 | case SET_SWD: | |
753 | f->zbd_info->sectors_with_data = swd; | |
754 | break; | |
755 | } | |
615555bb BVA |
756 | pthread_mutex_unlock(&f->zbd_info->mutex); |
757 | for (z = zb; z < ze; z++) | |
758 | pthread_mutex_unlock(&z->mutex); | |
91d25131 BVA |
759 | |
760 | return swd; | |
761 | } | |
762 | ||
763 | /* | |
764 | * The swd check is useful for debugging but takes too much time to leave | |
765 | * it enabled all the time. Hence it is disabled by default. | |
766 | */ | |
767 | static const bool enable_check_swd = false; | |
768 | ||
769 | /* Check whether the value of zbd_info.sectors_with_data is correct. */ | |
770 | static void zbd_check_swd(const struct fio_file *f) | |
771 | { | |
772 | if (!enable_check_swd) | |
773 | return; | |
774 | ||
775 | zbd_process_swd(f, CHECK_SWD); | |
776 | } | |
777 | ||
778 | static void zbd_init_swd(struct fio_file *f) | |
779 | { | |
780 | uint64_t swd; | |
781 | ||
782 | swd = zbd_process_swd(f, SET_SWD); | |
783 | dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name, | |
784 | swd); | |
615555bb BVA |
785 | } |
786 | ||
bfbdd35b BVA |
787 | void zbd_file_reset(struct thread_data *td, struct fio_file *f) |
788 | { | |
91d25131 | 789 | struct fio_zone_info *zb, *ze; |
bfbdd35b BVA |
790 | uint32_t zone_idx_e; |
791 | ||
792 | if (!f->zbd_info) | |
793 | return; | |
794 | ||
795 | zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)]; | |
796 | zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size); | |
797 | ze = &f->zbd_info->zone_info[zone_idx_e]; | |
91d25131 | 798 | zbd_init_swd(f); |
bfbdd35b BVA |
799 | /* |
800 | * If data verification is enabled reset the affected zones before | |
801 | * writing any data to avoid that a zone reset has to be issued while | |
802 | * writing data, which causes data loss. | |
803 | */ | |
804 | zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE && | |
805 | (td->o.td_ddir & TD_DDIR_WRITE) && | |
806 | td->runstate != TD_VERIFYING); | |
a7c2b6fc | 807 | zbd_reset_write_cnt(td, f); |
bfbdd35b BVA |
808 | } |
809 | ||
59b07544 BVA |
810 | /* The caller must hold f->zbd_info->mutex. */ |
811 | static bool is_zone_open(const struct thread_data *td, const struct fio_file *f, | |
812 | unsigned int zone_idx) | |
813 | { | |
814 | struct zoned_block_device_info *zbdi = f->zbd_info; | |
815 | int i; | |
816 | ||
817 | assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones)); | |
818 | assert(zbdi->num_open_zones <= td->o.max_open_zones); | |
819 | ||
820 | for (i = 0; i < zbdi->num_open_zones; i++) | |
821 | if (zbdi->open_zones[i] == zone_idx) | |
822 | return true; | |
823 | ||
824 | return false; | |
825 | } | |
826 | ||
827 | /* | |
828 | * Open a ZBD zone if it was not yet open. Returns true if either the zone was | |
829 | * already open or if opening a new zone is allowed. Returns false if the zone | |
830 | * was not yet open and opening a new zone would cause the zone limit to be | |
831 | * exceeded. | |
832 | */ | |
833 | static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u, | |
834 | uint32_t zone_idx) | |
835 | { | |
836 | const uint32_t min_bs = td->o.min_bs[DDIR_WRITE]; | |
837 | const struct fio_file *f = io_u->file; | |
838 | struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx]; | |
839 | bool res = true; | |
840 | ||
841 | if (z->cond == BLK_ZONE_COND_OFFLINE) | |
842 | return false; | |
843 | ||
844 | /* | |
845 | * Skip full zones with data verification enabled because resetting a | |
846 | * zone causes data loss and hence causes verification to fail. | |
847 | */ | |
848 | if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs)) | |
849 | return false; | |
850 | ||
851 | /* Zero means no limit */ | |
852 | if (!td->o.max_open_zones) | |
853 | return true; | |
854 | ||
855 | pthread_mutex_lock(&f->zbd_info->mutex); | |
856 | if (is_zone_open(td, f, zone_idx)) | |
857 | goto out; | |
858 | res = false; | |
859 | if (f->zbd_info->num_open_zones >= td->o.max_open_zones) | |
860 | goto out; | |
861 | dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx); | |
862 | f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx; | |
863 | z->open = 1; | |
864 | res = true; | |
865 | ||
866 | out: | |
867 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
868 | return res; | |
869 | } | |
870 | ||
871 | /* The caller must hold f->zbd_info->mutex */ | |
872 | static void zbd_close_zone(struct thread_data *td, const struct fio_file *f, | |
873 | unsigned int open_zone_idx) | |
874 | { | |
875 | uint32_t zone_idx; | |
876 | ||
877 | assert(open_zone_idx < f->zbd_info->num_open_zones); | |
878 | zone_idx = f->zbd_info->open_zones[open_zone_idx]; | |
879 | memmove(f->zbd_info->open_zones + open_zone_idx, | |
880 | f->zbd_info->open_zones + open_zone_idx + 1, | |
881 | (FIO_MAX_OPEN_ZBD_ZONES - (open_zone_idx + 1)) * | |
882 | sizeof(f->zbd_info->open_zones[0])); | |
883 | f->zbd_info->num_open_zones--; | |
884 | f->zbd_info->zone_info[zone_idx].open = 0; | |
885 | } | |
886 | ||
887 | /* | |
888 | * Modify the offset of an I/O unit that does not refer to an open zone such | |
889 | * that it refers to an open zone. Close an open zone and open a new zone if | |
890 | * necessary. This algorithm can only work correctly if all write pointers are | |
891 | * a multiple of the fio block size. The caller must neither hold z->mutex | |
892 | * nor f->zbd_info->mutex. Returns with z->mutex held upon success. | |
893 | */ | |
894 | struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td, | |
895 | struct io_u *io_u) | |
896 | { | |
897 | const uint32_t min_bs = td->o.min_bs[io_u->ddir]; | |
898 | const struct fio_file *f = io_u->file; | |
899 | struct fio_zone_info *z; | |
900 | unsigned int open_zone_idx = -1; | |
901 | uint32_t zone_idx, new_zone_idx; | |
902 | int i; | |
903 | ||
904 | assert(is_valid_offset(f, io_u->offset)); | |
905 | ||
906 | if (td->o.max_open_zones) { | |
907 | /* | |
908 | * This statement accesses f->zbd_info->open_zones[] on purpose | |
909 | * without locking. | |
910 | */ | |
911 | zone_idx = f->zbd_info->open_zones[(io_u->offset - | |
912 | f->file_offset) * | |
913 | f->zbd_info->num_open_zones / f->io_size]; | |
914 | } else { | |
915 | zone_idx = zbd_zone_idx(f, io_u->offset); | |
916 | } | |
917 | dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n", | |
918 | __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen); | |
919 | ||
920 | /* | |
921 | * Since z->mutex is the outer lock and f->zbd_info->mutex the inner | |
922 | * lock it can happen that the state of the zone with index zone_idx | |
923 | * has changed after 'z' has been assigned and before f->zbd_info->mutex | |
924 | * has been obtained. Hence the loop. | |
925 | */ | |
926 | for (;;) { | |
927 | z = &f->zbd_info->zone_info[zone_idx]; | |
928 | ||
929 | pthread_mutex_lock(&z->mutex); | |
930 | pthread_mutex_lock(&f->zbd_info->mutex); | |
931 | if (td->o.max_open_zones == 0) | |
932 | goto examine_zone; | |
933 | if (f->zbd_info->num_open_zones == 0) { | |
934 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
935 | pthread_mutex_unlock(&z->mutex); | |
936 | dprint(FD_ZBD, "%s(%s): no zones are open\n", | |
937 | __func__, f->file_name); | |
938 | return NULL; | |
939 | } | |
940 | open_zone_idx = (io_u->offset - f->file_offset) * | |
941 | f->zbd_info->num_open_zones / f->io_size; | |
942 | assert(open_zone_idx < f->zbd_info->num_open_zones); | |
943 | new_zone_idx = f->zbd_info->open_zones[open_zone_idx]; | |
944 | if (new_zone_idx == zone_idx) | |
945 | break; | |
946 | zone_idx = new_zone_idx; | |
947 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
948 | pthread_mutex_unlock(&z->mutex); | |
949 | } | |
950 | ||
951 | /* Both z->mutex and f->zbd_info->mutex are held. */ | |
952 | ||
953 | examine_zone: | |
ee3696bd | 954 | if (z->wp + min_bs <= (z+1)->start) { |
59b07544 BVA |
955 | pthread_mutex_unlock(&f->zbd_info->mutex); |
956 | goto out; | |
957 | } | |
958 | dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name, | |
959 | zone_idx); | |
960 | if (td->o.max_open_zones) | |
961 | zbd_close_zone(td, f, open_zone_idx); | |
962 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
963 | ||
964 | /* Only z->mutex is held. */ | |
965 | ||
966 | /* Zone 'z' is full, so try to open a new zone. */ | |
967 | for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) { | |
968 | zone_idx++; | |
969 | pthread_mutex_unlock(&z->mutex); | |
970 | z++; | |
ee3696bd | 971 | if (!is_valid_offset(f, z->start)) { |
59b07544 BVA |
972 | /* Wrap-around. */ |
973 | zone_idx = zbd_zone_idx(f, f->file_offset); | |
974 | z = &f->zbd_info->zone_info[zone_idx]; | |
975 | } | |
ee3696bd | 976 | assert(is_valid_offset(f, z->start)); |
59b07544 BVA |
977 | pthread_mutex_lock(&z->mutex); |
978 | if (z->open) | |
979 | continue; | |
980 | if (zbd_open_zone(td, io_u, zone_idx)) | |
981 | goto out; | |
982 | } | |
983 | ||
984 | /* Only z->mutex is held. */ | |
985 | ||
986 | /* Check whether the write fits in any of the already opened zones. */ | |
987 | pthread_mutex_lock(&f->zbd_info->mutex); | |
988 | for (i = 0; i < f->zbd_info->num_open_zones; i++) { | |
989 | zone_idx = f->zbd_info->open_zones[i]; | |
990 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
991 | pthread_mutex_unlock(&z->mutex); | |
992 | ||
993 | z = &f->zbd_info->zone_info[zone_idx]; | |
994 | ||
995 | pthread_mutex_lock(&z->mutex); | |
ee3696bd | 996 | if (z->wp + min_bs <= (z+1)->start) |
59b07544 BVA |
997 | goto out; |
998 | pthread_mutex_lock(&f->zbd_info->mutex); | |
999 | } | |
1000 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
1001 | pthread_mutex_unlock(&z->mutex); | |
1002 | dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__, | |
1003 | f->file_name); | |
1004 | return NULL; | |
1005 | ||
1006 | out: | |
1007 | dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name, | |
1008 | zone_idx); | |
ee3696bd | 1009 | io_u->offset = z->start; |
59b07544 BVA |
1010 | return z; |
1011 | } | |
1012 | ||
bfbdd35b | 1013 | /* The caller must hold z->mutex. */ |
59b07544 BVA |
1014 | static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td, |
1015 | struct io_u *io_u, | |
1016 | struct fio_zone_info *z) | |
bfbdd35b BVA |
1017 | { |
1018 | const struct fio_file *f = io_u->file; | |
1019 | const uint32_t min_bs = td->o.min_bs[DDIR_WRITE]; | |
1020 | ||
59b07544 BVA |
1021 | if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) { |
1022 | pthread_mutex_unlock(&z->mutex); | |
1023 | z = zbd_convert_to_open_zone(td, io_u); | |
1024 | assert(z); | |
1025 | } | |
1026 | ||
bfbdd35b | 1027 | if (z->verify_block * min_bs >= f->zbd_info->zone_size) |
a0c84dd4 JA |
1028 | log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block, |
1029 | min_bs, (unsigned long long) f->zbd_info->zone_size); | |
ee3696bd | 1030 | io_u->offset = z->start + z->verify_block++ * min_bs; |
59b07544 | 1031 | return z; |
bfbdd35b BVA |
1032 | } |
1033 | ||
1034 | /* | |
1035 | * Find another zone for which @io_u fits below the write pointer. Start | |
1036 | * searching in zones @zb + 1 .. @zl and continue searching in zones | |
1037 | * @zf .. @zb - 1. | |
1038 | * | |
1039 | * Either returns NULL or returns a zone pointer and holds the mutex for that | |
1040 | * zone. | |
1041 | */ | |
1042 | static struct fio_zone_info * | |
1043 | zbd_find_zone(struct thread_data *td, struct io_u *io_u, | |
1044 | struct fio_zone_info *zb, struct fio_zone_info *zl) | |
1045 | { | |
1046 | const uint32_t min_bs = td->o.min_bs[io_u->ddir]; | |
1047 | const struct fio_file *f = io_u->file; | |
1048 | struct fio_zone_info *z1, *z2; | |
1049 | const struct fio_zone_info *const zf = | |
1050 | &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)]; | |
1051 | ||
1052 | /* | |
1053 | * Skip to the next non-empty zone in case of sequential I/O and to | |
1054 | * the nearest non-empty zone in case of random I/O. | |
1055 | */ | |
1056 | for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) { | |
1057 | if (z1 < zl && z1->cond != BLK_ZONE_COND_OFFLINE) { | |
1058 | pthread_mutex_lock(&z1->mutex); | |
ee3696bd | 1059 | if (z1->start + min_bs <= z1->wp) |
bfbdd35b BVA |
1060 | return z1; |
1061 | pthread_mutex_unlock(&z1->mutex); | |
1062 | } else if (!td_random(td)) { | |
1063 | break; | |
1064 | } | |
1065 | if (td_random(td) && z2 >= zf && | |
1066 | z2->cond != BLK_ZONE_COND_OFFLINE) { | |
1067 | pthread_mutex_lock(&z2->mutex); | |
ee3696bd | 1068 | if (z2->start + min_bs <= z2->wp) |
bfbdd35b BVA |
1069 | return z2; |
1070 | pthread_mutex_unlock(&z2->mutex); | |
1071 | } | |
1072 | } | |
1073 | dprint(FD_ZBD, "%s: adjusting random read offset failed\n", | |
1074 | f->file_name); | |
1075 | return NULL; | |
1076 | } | |
1077 | ||
1078 | ||
1079 | /** | |
1080 | * zbd_post_submit - update the write pointer and unlock the zone lock | |
1081 | * @io_u: I/O unit | |
1082 | * @success: Whether or not the I/O unit has been executed successfully | |
1083 | * | |
1084 | * For write and trim operations, update the write pointer of all affected | |
1085 | * zones. | |
1086 | */ | |
1087 | static void zbd_post_submit(const struct io_u *io_u, bool success) | |
1088 | { | |
1089 | struct zoned_block_device_info *zbd_info; | |
1090 | struct fio_zone_info *z; | |
1091 | uint32_t zone_idx; | |
1092 | uint64_t end, zone_end; | |
1093 | ||
1094 | zbd_info = io_u->file->zbd_info; | |
1095 | if (!zbd_info) | |
1096 | return; | |
1097 | ||
1098 | zone_idx = zbd_zone_idx(io_u->file, io_u->offset); | |
ee3696bd | 1099 | end = io_u->offset + io_u->buflen; |
bfbdd35b BVA |
1100 | z = &zbd_info->zone_info[zone_idx]; |
1101 | assert(zone_idx < zbd_info->nr_zones); | |
1102 | if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ) | |
1103 | return; | |
1104 | if (!success) | |
1105 | goto unlock; | |
1106 | switch (io_u->ddir) { | |
1107 | case DDIR_WRITE: | |
1108 | zone_end = min(end, (z + 1)->start); | |
a7c2b6fc BVA |
1109 | pthread_mutex_lock(&zbd_info->mutex); |
1110 | /* | |
1111 | * z->wp > zone_end means that one or more I/O errors | |
1112 | * have occurred. | |
1113 | */ | |
1114 | if (z->wp <= zone_end) | |
1115 | zbd_info->sectors_with_data += zone_end - z->wp; | |
1116 | pthread_mutex_unlock(&zbd_info->mutex); | |
bfbdd35b BVA |
1117 | z->wp = zone_end; |
1118 | break; | |
1119 | case DDIR_TRIM: | |
1120 | assert(z->wp == z->start); | |
1121 | break; | |
1122 | default: | |
1123 | break; | |
1124 | } | |
1125 | unlock: | |
1126 | pthread_mutex_unlock(&z->mutex); | |
615555bb BVA |
1127 | |
1128 | zbd_check_swd(io_u->file); | |
bfbdd35b BVA |
1129 | } |
1130 | ||
1131 | bool zbd_unaligned_write(int error_code) | |
1132 | { | |
1133 | switch (error_code) { | |
1134 | case EIO: | |
1135 | case EREMOTEIO: | |
1136 | return true; | |
1137 | } | |
1138 | return false; | |
1139 | } | |
1140 | ||
1141 | /** | |
1142 | * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives | |
1143 | * @td: FIO thread data. | |
1144 | * @io_u: FIO I/O unit. | |
1145 | * | |
1146 | * Locking strategy: returns with z->mutex locked if and only if z refers | |
1147 | * to a sequential zone and if io_u_accept is returned. z is the zone that | |
1148 | * corresponds to io_u->offset at the end of this function. | |
1149 | */ | |
1150 | enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u) | |
1151 | { | |
1152 | const struct fio_file *f = io_u->file; | |
1153 | uint32_t zone_idx_b; | |
de65f7b7 | 1154 | struct fio_zone_info *zb, *zl, *orig_zb; |
bfbdd35b BVA |
1155 | uint32_t orig_len = io_u->buflen; |
1156 | uint32_t min_bs = td->o.min_bs[io_u->ddir]; | |
1157 | uint64_t new_len; | |
1158 | int64_t range; | |
1159 | ||
1160 | if (!f->zbd_info) | |
1161 | return io_u_accept; | |
1162 | ||
1163 | assert(is_valid_offset(f, io_u->offset)); | |
1164 | assert(io_u->buflen); | |
1165 | zone_idx_b = zbd_zone_idx(f, io_u->offset); | |
1166 | zb = &f->zbd_info->zone_info[zone_idx_b]; | |
de65f7b7 | 1167 | orig_zb = zb; |
bfbdd35b BVA |
1168 | |
1169 | /* Accept the I/O offset for conventional zones. */ | |
1170 | if (zb->type == BLK_ZONE_TYPE_CONVENTIONAL) | |
1171 | return io_u_accept; | |
1172 | ||
1173 | /* | |
1174 | * Accept the I/O offset for reads if reading beyond the write pointer | |
1175 | * is enabled. | |
1176 | */ | |
1177 | if (zb->cond != BLK_ZONE_COND_OFFLINE && | |
1178 | io_u->ddir == DDIR_READ && td->o.read_beyond_wp) | |
1179 | return io_u_accept; | |
1180 | ||
615555bb BVA |
1181 | zbd_check_swd(f); |
1182 | ||
bfbdd35b BVA |
1183 | pthread_mutex_lock(&zb->mutex); |
1184 | switch (io_u->ddir) { | |
1185 | case DDIR_READ: | |
1186 | if (td->runstate == TD_VERIFYING) { | |
59b07544 | 1187 | zb = zbd_replay_write_order(td, io_u, zb); |
bfbdd35b BVA |
1188 | goto accept; |
1189 | } | |
1190 | /* | |
de65f7b7 DLM |
1191 | * Check that there is enough written data in the zone to do an |
1192 | * I/O of at least min_bs B. If there isn't, find a new zone for | |
1193 | * the I/O. | |
bfbdd35b BVA |
1194 | */ |
1195 | range = zb->cond != BLK_ZONE_COND_OFFLINE ? | |
ee3696bd | 1196 | zb->wp - zb->start : 0; |
de65f7b7 | 1197 | if (range < min_bs || |
ee3696bd | 1198 | ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) { |
bfbdd35b BVA |
1199 | pthread_mutex_unlock(&zb->mutex); |
1200 | zl = &f->zbd_info->zone_info[zbd_zone_idx(f, | |
1201 | f->file_offset + f->io_size)]; | |
1202 | zb = zbd_find_zone(td, io_u, zb, zl); | |
1203 | if (!zb) { | |
1204 | dprint(FD_ZBD, | |
1205 | "%s: zbd_find_zone(%lld, %llu) failed\n", | |
1206 | f->file_name, io_u->offset, | |
1207 | io_u->buflen); | |
1208 | goto eof; | |
1209 | } | |
de65f7b7 DLM |
1210 | /* |
1211 | * zbd_find_zone() returned a zone with a range of at | |
1212 | * least min_bs. | |
1213 | */ | |
ee3696bd | 1214 | range = zb->wp - zb->start; |
de65f7b7 DLM |
1215 | assert(range >= min_bs); |
1216 | ||
1217 | if (!td_random(td)) | |
ee3696bd | 1218 | io_u->offset = zb->start; |
bfbdd35b | 1219 | } |
de65f7b7 DLM |
1220 | /* |
1221 | * Make sure the I/O is within the zone valid data range while | |
1222 | * maximizing the I/O size and preserving randomness. | |
1223 | */ | |
1224 | if (range <= io_u->buflen) | |
ee3696bd | 1225 | io_u->offset = zb->start; |
de65f7b7 | 1226 | else if (td_random(td)) |
ee3696bd DLM |
1227 | io_u->offset = zb->start + |
1228 | ((io_u->offset - orig_zb->start) % | |
de65f7b7 DLM |
1229 | (range - io_u->buflen)) / min_bs * min_bs; |
1230 | /* | |
1231 | * Make sure the I/O does not cross over the zone wp position. | |
1232 | */ | |
1233 | new_len = min((unsigned long long)io_u->buflen, | |
ee3696bd | 1234 | (unsigned long long)(zb->wp - io_u->offset)); |
de65f7b7 DLM |
1235 | new_len = new_len / min_bs * min_bs; |
1236 | if (new_len < io_u->buflen) { | |
1237 | io_u->buflen = new_len; | |
1238 | dprint(FD_IO, "Changed length from %u into %llu\n", | |
1239 | orig_len, io_u->buflen); | |
bfbdd35b | 1240 | } |
ee3696bd DLM |
1241 | assert(zb->start <= io_u->offset); |
1242 | assert(io_u->offset + io_u->buflen <= zb->wp); | |
bfbdd35b BVA |
1243 | goto accept; |
1244 | case DDIR_WRITE: | |
ee3696bd | 1245 | if (io_u->buflen > f->zbd_info->zone_size) |
bfbdd35b | 1246 | goto eof; |
59b07544 BVA |
1247 | if (!zbd_open_zone(td, io_u, zone_idx_b)) { |
1248 | pthread_mutex_unlock(&zb->mutex); | |
1249 | zb = zbd_convert_to_open_zone(td, io_u); | |
1250 | if (!zb) | |
1251 | goto eof; | |
1252 | zone_idx_b = zb - f->zbd_info->zone_info; | |
1253 | } | |
a7c2b6fc BVA |
1254 | /* Check whether the zone reset threshold has been exceeded */ |
1255 | if (td->o.zrf.u.f) { | |
ee3696bd | 1256 | if (f->zbd_info->sectors_with_data >= |
a7c2b6fc BVA |
1257 | f->io_size * td->o.zrt.u.f && |
1258 | zbd_dec_and_reset_write_cnt(td, f)) { | |
1259 | zb->reset_zone = 1; | |
1260 | } | |
1261 | } | |
bfbdd35b BVA |
1262 | /* Reset the zone pointer if necessary */ |
1263 | if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) { | |
1264 | assert(td->o.verify == VERIFY_NONE); | |
1265 | /* | |
1266 | * Since previous write requests may have been submitted | |
1267 | * asynchronously and since we will submit the zone | |
1268 | * reset synchronously, wait until previously submitted | |
1269 | * write requests have completed before issuing a | |
1270 | * zone reset. | |
1271 | */ | |
1272 | io_u_quiesce(td); | |
1273 | zb->reset_zone = 0; | |
1274 | if (zbd_reset_zone(td, f, zb) < 0) | |
1275 | goto eof; | |
1276 | } | |
1277 | /* Make writes occur at the write pointer */ | |
1278 | assert(!zbd_zone_full(f, zb, min_bs)); | |
ee3696bd | 1279 | io_u->offset = zb->wp; |
bfbdd35b BVA |
1280 | if (!is_valid_offset(f, io_u->offset)) { |
1281 | dprint(FD_ZBD, "Dropped request with offset %llu\n", | |
1282 | io_u->offset); | |
1283 | goto eof; | |
1284 | } | |
1285 | /* | |
1286 | * Make sure that the buflen is a multiple of the minimal | |
1287 | * block size. Give up if shrinking would make the request too | |
1288 | * small. | |
1289 | */ | |
1290 | new_len = min((unsigned long long)io_u->buflen, | |
ee3696bd | 1291 | (zb + 1)->start - io_u->offset); |
bfbdd35b BVA |
1292 | new_len = new_len / min_bs * min_bs; |
1293 | if (new_len == io_u->buflen) | |
1294 | goto accept; | |
1295 | if (new_len >= min_bs) { | |
1296 | io_u->buflen = new_len; | |
1297 | dprint(FD_IO, "Changed length from %u into %llu\n", | |
1298 | orig_len, io_u->buflen); | |
1299 | goto accept; | |
1300 | } | |
1301 | log_err("Zone remainder %lld smaller than minimum block size %d\n", | |
ee3696bd | 1302 | ((zb + 1)->start - io_u->offset), |
bfbdd35b BVA |
1303 | min_bs); |
1304 | goto eof; | |
1305 | case DDIR_TRIM: | |
1306 | /* fall-through */ | |
1307 | case DDIR_SYNC: | |
1308 | case DDIR_DATASYNC: | |
1309 | case DDIR_SYNC_FILE_RANGE: | |
1310 | case DDIR_WAIT: | |
1311 | case DDIR_LAST: | |
1312 | case DDIR_INVAL: | |
1313 | goto accept; | |
1314 | } | |
1315 | ||
1316 | assert(false); | |
1317 | ||
1318 | accept: | |
1319 | assert(zb); | |
1320 | assert(zb->cond != BLK_ZONE_COND_OFFLINE); | |
1321 | assert(!io_u->post_submit); | |
1322 | io_u->post_submit = zbd_post_submit; | |
1323 | return io_u_accept; | |
1324 | ||
1325 | eof: | |
1326 | if (zb) | |
1327 | pthread_mutex_unlock(&zb->mutex); | |
1328 | return io_u_eof; | |
1329 | } | |
fd5d733f BVA |
1330 | |
1331 | /* Return a string with ZBD statistics */ | |
1332 | char *zbd_write_status(const struct thread_stat *ts) | |
1333 | { | |
1334 | char *res; | |
1335 | ||
a0c84dd4 | 1336 | if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0) |
fd5d733f BVA |
1337 | return NULL; |
1338 | return res; | |
1339 | } |