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; | |
4d37720a | 122 | |
fc52e449 DLM |
123 | if (!td->o.zone_size) { |
124 | td->o.zone_size = f->zbd_info->zone_size; | |
125 | if (!td->o.zone_size) { | |
126 | log_err("%s: invalid 0 zone size\n", | |
127 | f->file_name); | |
128 | return false; | |
129 | } | |
130 | } | |
131 | ||
4d37720a DLM |
132 | if (td->o.zone_skip && |
133 | (td->o.zone_skip < td->o.zone_size || | |
134 | td->o.zone_skip % td->o.zone_size)) { | |
135 | log_err("%s: zoneskip %llu is not a multiple of the device zone size %llu.\n", | |
136 | f->file_name, (unsigned long long) td->o.zone_skip, | |
137 | (unsigned long long) td->o.zone_size); | |
138 | return false; | |
139 | } | |
140 | ||
bfbdd35b BVA |
141 | zone_idx = zbd_zone_idx(f, f->file_offset); |
142 | z = &f->zbd_info->zone_info[zone_idx]; | |
ee3696bd DLM |
143 | if (f->file_offset != z->start) { |
144 | new_offset = (z+1)->start; | |
bfbdd35b BVA |
145 | if (new_offset >= f->file_offset + f->io_size) { |
146 | log_info("%s: io_size must be at least one zone\n", | |
147 | f->file_name); | |
148 | return false; | |
149 | } | |
a0c84dd4 JA |
150 | log_info("%s: rounded up offset from %llu to %llu\n", |
151 | f->file_name, (unsigned long long) f->file_offset, | |
152 | (unsigned long long) new_offset); | |
bfbdd35b BVA |
153 | f->io_size -= (new_offset - f->file_offset); |
154 | f->file_offset = new_offset; | |
155 | } | |
156 | zone_idx = zbd_zone_idx(f, f->file_offset + f->io_size); | |
157 | z = &f->zbd_info->zone_info[zone_idx]; | |
ee3696bd | 158 | new_end = z->start; |
bfbdd35b BVA |
159 | if (f->file_offset + f->io_size != new_end) { |
160 | if (new_end <= f->file_offset) { | |
161 | log_info("%s: io_size must be at least one zone\n", | |
162 | f->file_name); | |
163 | return false; | |
164 | } | |
a0c84dd4 JA |
165 | log_info("%s: rounded down io_size from %llu to %llu\n", |
166 | f->file_name, (unsigned long long) f->io_size, | |
167 | (unsigned long long) new_end - f->file_offset); | |
bfbdd35b BVA |
168 | f->io_size = new_end - f->file_offset; |
169 | } | |
170 | } | |
171 | } | |
172 | ||
173 | return true; | |
174 | } | |
175 | ||
176 | static bool zbd_verify_bs(void) | |
177 | { | |
178 | struct thread_data *td; | |
179 | struct fio_file *f; | |
180 | uint32_t zone_size; | |
181 | int i, j, k; | |
182 | ||
183 | for_each_td(td, i) { | |
184 | for_each_file(td, f, j) { | |
185 | if (!f->zbd_info) | |
186 | continue; | |
187 | zone_size = f->zbd_info->zone_size; | |
188 | for (k = 0; k < ARRAY_SIZE(td->o.bs); k++) { | |
189 | if (td->o.verify != VERIFY_NONE && | |
ee3696bd | 190 | zone_size % td->o.bs[k] != 0) { |
bfbdd35b BVA |
191 | log_info("%s: block size %llu is not a divisor of the zone size %d\n", |
192 | f->file_name, td->o.bs[k], | |
ee3696bd | 193 | zone_size); |
bfbdd35b BVA |
194 | return false; |
195 | } | |
196 | } | |
197 | } | |
198 | } | |
199 | return true; | |
200 | } | |
201 | ||
202 | /* | |
203 | * Read zone information into @buf starting from sector @start_sector. | |
204 | * @fd is a file descriptor that refers to a block device and @bufsz is the | |
205 | * size of @buf. | |
206 | * | |
207 | * Returns 0 upon success and a negative error code upon failure. | |
7f125e7f DLM |
208 | * If the zone report is empty, always assume an error (device problem) and |
209 | * return -EIO. | |
bfbdd35b BVA |
210 | */ |
211 | static int read_zone_info(int fd, uint64_t start_sector, | |
212 | void *buf, unsigned int bufsz) | |
213 | { | |
214 | struct blk_zone_report *hdr = buf; | |
7f125e7f | 215 | int ret; |
bfbdd35b BVA |
216 | |
217 | if (bufsz < sizeof(*hdr)) | |
218 | return -EINVAL; | |
219 | ||
220 | memset(hdr, 0, sizeof(*hdr)); | |
221 | ||
222 | hdr->nr_zones = (bufsz - sizeof(*hdr)) / sizeof(struct blk_zone); | |
223 | hdr->sector = start_sector; | |
7f125e7f DLM |
224 | ret = ioctl(fd, BLKREPORTZONE, hdr); |
225 | if (ret) | |
226 | return -errno; | |
227 | if (!hdr->nr_zones) | |
228 | return -EIO; | |
229 | return 0; | |
bfbdd35b BVA |
230 | } |
231 | ||
232 | /* | |
233 | * Read up to 255 characters from the first line of a file. Strip the trailing | |
234 | * newline. | |
235 | */ | |
236 | static char *read_file(const char *path) | |
237 | { | |
238 | char line[256], *p = line; | |
239 | FILE *f; | |
240 | ||
241 | f = fopen(path, "rb"); | |
242 | if (!f) | |
243 | return NULL; | |
244 | if (!fgets(line, sizeof(line), f)) | |
245 | line[0] = '\0'; | |
246 | strsep(&p, "\n"); | |
247 | fclose(f); | |
248 | ||
249 | return strdup(line); | |
250 | } | |
251 | ||
252 | static enum blk_zoned_model get_zbd_model(const char *file_name) | |
253 | { | |
254 | enum blk_zoned_model model = ZBD_DM_NONE; | |
255 | char *zoned_attr_path = NULL; | |
256 | char *model_str = NULL; | |
257 | struct stat statbuf; | |
63e3e40d SK |
258 | char *sys_devno_path = NULL; |
259 | char *part_attr_path = NULL; | |
260 | char *part_str = NULL; | |
261 | char sys_path[PATH_MAX]; | |
262 | ssize_t sz; | |
263 | char *delim = NULL; | |
bfbdd35b BVA |
264 | |
265 | if (stat(file_name, &statbuf) < 0) | |
266 | goto out; | |
63e3e40d SK |
267 | |
268 | if (asprintf(&sys_devno_path, "/sys/dev/block/%d:%d", | |
bfbdd35b BVA |
269 | major(statbuf.st_rdev), minor(statbuf.st_rdev)) < 0) |
270 | goto out; | |
63e3e40d SK |
271 | |
272 | sz = readlink(sys_devno_path, sys_path, sizeof(sys_path) - 1); | |
273 | if (sz < 0) | |
274 | goto out; | |
275 | sys_path[sz] = '\0'; | |
276 | ||
277 | /* | |
278 | * If the device is a partition device, cut the device name in the | |
279 | * canonical sysfs path to obtain the sysfs path of the holder device. | |
280 | * e.g.: /sys/devices/.../sda/sda1 -> /sys/devices/.../sda | |
281 | */ | |
282 | if (asprintf(&part_attr_path, "/sys/dev/block/%s/partition", | |
283 | sys_path) < 0) | |
284 | goto out; | |
285 | part_str = read_file(part_attr_path); | |
286 | if (part_str && *part_str == '1') { | |
287 | delim = strrchr(sys_path, '/'); | |
288 | if (!delim) | |
289 | goto out; | |
290 | *delim = '\0'; | |
291 | } | |
292 | ||
293 | if (asprintf(&zoned_attr_path, | |
294 | "/sys/dev/block/%s/queue/zoned", sys_path) < 0) | |
295 | goto out; | |
296 | ||
bfbdd35b BVA |
297 | model_str = read_file(zoned_attr_path); |
298 | if (!model_str) | |
299 | goto out; | |
300 | dprint(FD_ZBD, "%s: zbd model string: %s\n", file_name, model_str); | |
301 | if (strcmp(model_str, "host-aware") == 0) | |
302 | model = ZBD_DM_HOST_AWARE; | |
303 | else if (strcmp(model_str, "host-managed") == 0) | |
304 | model = ZBD_DM_HOST_MANAGED; | |
305 | ||
306 | out: | |
307 | free(model_str); | |
308 | free(zoned_attr_path); | |
63e3e40d SK |
309 | free(part_str); |
310 | free(part_attr_path); | |
311 | free(sys_devno_path); | |
bfbdd35b BVA |
312 | return model; |
313 | } | |
314 | ||
315 | static int ilog2(uint64_t i) | |
316 | { | |
317 | int log = -1; | |
318 | ||
319 | while (i) { | |
320 | i >>= 1; | |
321 | log++; | |
322 | } | |
323 | return log; | |
324 | } | |
325 | ||
326 | /* | |
327 | * Initialize f->zbd_info for devices that are not zoned block devices. This | |
328 | * allows to execute a ZBD workload against a non-ZBD device. | |
329 | */ | |
330 | static int init_zone_info(struct thread_data *td, struct fio_file *f) | |
331 | { | |
332 | uint32_t nr_zones; | |
333 | struct fio_zone_info *p; | |
a4b7f12b | 334 | uint64_t zone_size = td->o.zone_size; |
bfbdd35b BVA |
335 | struct zoned_block_device_info *zbd_info = NULL; |
336 | pthread_mutexattr_t attr; | |
337 | int i; | |
338 | ||
a4b7f12b DLM |
339 | if (zone_size == 0) { |
340 | log_err("%s: Specifying the zone size is mandatory for regular block devices with --zonemode=zbd\n\n", | |
341 | f->file_name); | |
342 | return 1; | |
343 | } | |
344 | ||
345 | if (zone_size < 512) { | |
346 | log_err("%s: zone size must be at least 512 bytes for --zonemode=zbd\n\n", | |
347 | f->file_name); | |
348 | return 1; | |
349 | } | |
350 | ||
ee3696bd | 351 | nr_zones = (f->real_file_size + zone_size - 1) / zone_size; |
bfbdd35b BVA |
352 | zbd_info = scalloc(1, sizeof(*zbd_info) + |
353 | (nr_zones + 1) * sizeof(zbd_info->zone_info[0])); | |
354 | if (!zbd_info) | |
355 | return -ENOMEM; | |
356 | ||
357 | pthread_mutexattr_init(&attr); | |
358 | pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE); | |
359 | pthread_mutexattr_setpshared(&attr, true); | |
360 | pthread_mutex_init(&zbd_info->mutex, &attr); | |
361 | zbd_info->refcount = 1; | |
362 | p = &zbd_info->zone_info[0]; | |
363 | for (i = 0; i < nr_zones; i++, p++) { | |
364 | pthread_mutex_init(&p->mutex, &attr); | |
365 | p->start = i * zone_size; | |
366 | p->wp = p->start + zone_size; | |
367 | p->type = BLK_ZONE_TYPE_SEQWRITE_REQ; | |
368 | p->cond = BLK_ZONE_COND_EMPTY; | |
369 | } | |
370 | /* a sentinel */ | |
371 | p->start = nr_zones * zone_size; | |
372 | ||
373 | f->zbd_info = zbd_info; | |
374 | f->zbd_info->zone_size = zone_size; | |
375 | f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ? | |
ee3696bd | 376 | ilog2(zone_size) : -1; |
bfbdd35b BVA |
377 | f->zbd_info->nr_zones = nr_zones; |
378 | pthread_mutexattr_destroy(&attr); | |
379 | return 0; | |
380 | } | |
381 | ||
382 | /* | |
383 | * Parse the BLKREPORTZONE output and store it in f->zbd_info. Must be called | |
384 | * only for devices that support this ioctl, namely zoned block devices. | |
385 | */ | |
386 | static int parse_zone_info(struct thread_data *td, struct fio_file *f) | |
387 | { | |
388 | const unsigned int bufsz = sizeof(struct blk_zone_report) + | |
389 | 4096 * sizeof(struct blk_zone); | |
390 | uint32_t nr_zones; | |
391 | struct blk_zone_report *hdr; | |
392 | const struct blk_zone *z; | |
393 | struct fio_zone_info *p; | |
394 | uint64_t zone_size, start_sector; | |
395 | struct zoned_block_device_info *zbd_info = NULL; | |
396 | pthread_mutexattr_t attr; | |
397 | void *buf; | |
398 | int fd, i, j, ret = 0; | |
399 | ||
400 | pthread_mutexattr_init(&attr); | |
401 | pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE); | |
402 | pthread_mutexattr_setpshared(&attr, true); | |
403 | ||
404 | buf = malloc(bufsz); | |
405 | if (!buf) | |
406 | goto out; | |
407 | ||
408 | fd = open(f->file_name, O_RDONLY | O_LARGEFILE); | |
409 | if (fd < 0) { | |
410 | ret = -errno; | |
411 | goto free; | |
412 | } | |
413 | ||
414 | ret = read_zone_info(fd, 0, buf, bufsz); | |
415 | if (ret < 0) { | |
416 | log_info("fio: BLKREPORTZONE(%lu) failed for %s (%d).\n", | |
417 | 0UL, f->file_name, -ret); | |
418 | goto close; | |
419 | } | |
420 | hdr = buf; | |
421 | if (hdr->nr_zones < 1) { | |
422 | log_info("fio: %s has invalid zone information.\n", | |
423 | f->file_name); | |
424 | goto close; | |
425 | } | |
426 | z = (void *)(hdr + 1); | |
ee3696bd DLM |
427 | zone_size = z->len << 9; |
428 | nr_zones = (f->real_file_size + zone_size - 1) / zone_size; | |
bfbdd35b BVA |
429 | |
430 | if (td->o.zone_size == 0) { | |
ee3696bd DLM |
431 | td->o.zone_size = zone_size; |
432 | } else if (td->o.zone_size != zone_size) { | |
4580ff92 DLM |
433 | log_err("fio: %s job parameter zonesize %llu does not match disk zone size %llu.\n", |
434 | f->file_name, (unsigned long long) td->o.zone_size, | |
a0c84dd4 | 435 | (unsigned long long) zone_size); |
bfbdd35b BVA |
436 | ret = -EINVAL; |
437 | goto close; | |
438 | } | |
439 | ||
a0c84dd4 JA |
440 | dprint(FD_ZBD, "Device %s has %d zones of size %llu KB\n", f->file_name, |
441 | nr_zones, (unsigned long long) zone_size / 1024); | |
bfbdd35b BVA |
442 | |
443 | zbd_info = scalloc(1, sizeof(*zbd_info) + | |
444 | (nr_zones + 1) * sizeof(zbd_info->zone_info[0])); | |
445 | ret = -ENOMEM; | |
446 | if (!zbd_info) | |
447 | goto close; | |
448 | pthread_mutex_init(&zbd_info->mutex, &attr); | |
449 | zbd_info->refcount = 1; | |
450 | p = &zbd_info->zone_info[0]; | |
451 | for (start_sector = 0, j = 0; j < nr_zones;) { | |
452 | z = (void *)(hdr + 1); | |
453 | for (i = 0; i < hdr->nr_zones; i++, j++, z++, p++) { | |
454 | pthread_mutex_init(&p->mutex, &attr); | |
ee3696bd | 455 | p->start = z->start << 9; |
bfbdd35b BVA |
456 | switch (z->cond) { |
457 | case BLK_ZONE_COND_NOT_WP: | |
bfbdd35b | 458 | case BLK_ZONE_COND_FULL: |
ee3696bd | 459 | p->wp = p->start + zone_size; |
bfbdd35b BVA |
460 | break; |
461 | default: | |
462 | assert(z->start <= z->wp); | |
ee3696bd DLM |
463 | assert(z->wp <= z->start + (zone_size >> 9)); |
464 | p->wp = z->wp << 9; | |
bfbdd35b BVA |
465 | break; |
466 | } | |
467 | p->type = z->type; | |
468 | p->cond = z->cond; | |
469 | if (j > 0 && p->start != p[-1].start + zone_size) { | |
470 | log_info("%s: invalid zone data\n", | |
471 | f->file_name); | |
472 | ret = -EINVAL; | |
473 | goto close; | |
474 | } | |
475 | } | |
476 | z--; | |
477 | start_sector = z->start + z->len; | |
478 | if (j >= nr_zones) | |
479 | break; | |
480 | ret = read_zone_info(fd, start_sector, buf, bufsz); | |
481 | if (ret < 0) { | |
a0c84dd4 JA |
482 | log_info("fio: BLKREPORTZONE(%llu) failed for %s (%d).\n", |
483 | (unsigned long long) start_sector, f->file_name, -ret); | |
bfbdd35b BVA |
484 | goto close; |
485 | } | |
486 | } | |
487 | /* a sentinel */ | |
ee3696bd | 488 | zbd_info->zone_info[nr_zones].start = start_sector << 9; |
bfbdd35b BVA |
489 | |
490 | f->zbd_info = zbd_info; | |
491 | f->zbd_info->zone_size = zone_size; | |
492 | f->zbd_info->zone_size_log2 = is_power_of_2(zone_size) ? | |
ee3696bd | 493 | ilog2(zone_size) : -1; |
bfbdd35b BVA |
494 | f->zbd_info->nr_zones = nr_zones; |
495 | zbd_info = NULL; | |
496 | ret = 0; | |
497 | ||
498 | close: | |
499 | sfree(zbd_info); | |
500 | close(fd); | |
501 | free: | |
502 | free(buf); | |
503 | out: | |
504 | pthread_mutexattr_destroy(&attr); | |
505 | return ret; | |
506 | } | |
507 | ||
508 | /* | |
509 | * Allocate zone information and store it into f->zbd_info if zonemode=zbd. | |
510 | * | |
511 | * Returns 0 upon success and a negative error code upon failure. | |
512 | */ | |
379e5f09 | 513 | static int zbd_create_zone_info(struct thread_data *td, struct fio_file *f) |
bfbdd35b BVA |
514 | { |
515 | enum blk_zoned_model zbd_model; | |
516 | int ret = 0; | |
517 | ||
518 | assert(td->o.zone_mode == ZONE_MODE_ZBD); | |
519 | ||
520 | zbd_model = get_zbd_model(f->file_name); | |
521 | switch (zbd_model) { | |
522 | case ZBD_DM_HOST_AWARE: | |
523 | case ZBD_DM_HOST_MANAGED: | |
524 | ret = parse_zone_info(td, f); | |
525 | break; | |
526 | case ZBD_DM_NONE: | |
527 | ret = init_zone_info(td, f); | |
528 | break; | |
529 | } | |
530 | if (ret == 0) | |
531 | f->zbd_info->model = zbd_model; | |
532 | return ret; | |
533 | } | |
534 | ||
535 | void zbd_free_zone_info(struct fio_file *f) | |
536 | { | |
537 | uint32_t refcount; | |
538 | ||
539 | if (!f->zbd_info) | |
540 | return; | |
541 | ||
542 | pthread_mutex_lock(&f->zbd_info->mutex); | |
543 | refcount = --f->zbd_info->refcount; | |
544 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
545 | ||
546 | assert((int32_t)refcount >= 0); | |
547 | if (refcount == 0) | |
548 | sfree(f->zbd_info); | |
549 | f->zbd_info = NULL; | |
550 | } | |
551 | ||
552 | /* | |
553 | * Initialize f->zbd_info. | |
554 | * | |
555 | * Returns 0 upon success and a negative error code upon failure. | |
556 | * | |
557 | * Note: this function can only work correctly if it is called before the first | |
558 | * fio fork() call. | |
559 | */ | |
560 | static int zbd_init_zone_info(struct thread_data *td, struct fio_file *file) | |
561 | { | |
562 | struct thread_data *td2; | |
563 | struct fio_file *f2; | |
564 | int i, j, ret; | |
565 | ||
566 | for_each_td(td2, i) { | |
567 | for_each_file(td2, f2, j) { | |
568 | if (td2 == td && f2 == file) | |
569 | continue; | |
570 | if (!f2->zbd_info || | |
571 | strcmp(f2->file_name, file->file_name) != 0) | |
572 | continue; | |
573 | file->zbd_info = f2->zbd_info; | |
574 | file->zbd_info->refcount++; | |
575 | return 0; | |
576 | } | |
577 | } | |
578 | ||
579 | ret = zbd_create_zone_info(td, file); | |
580 | if (ret < 0) | |
c5837eec | 581 | td_verror(td, -ret, "zbd_create_zone_info() failed"); |
bfbdd35b BVA |
582 | return ret; |
583 | } | |
584 | ||
585 | int zbd_init(struct thread_data *td) | |
586 | { | |
587 | struct fio_file *f; | |
588 | int i; | |
589 | ||
590 | for_each_file(td, f, i) { | |
591 | if (f->filetype != FIO_TYPE_BLOCK) | |
592 | continue; | |
a4b7f12b | 593 | if (zbd_init_zone_info(td, f)) |
bfbdd35b | 594 | return 1; |
bfbdd35b BVA |
595 | } |
596 | ||
597 | if (!zbd_using_direct_io()) { | |
598 | log_err("Using direct I/O is mandatory for writing to ZBD drives\n\n"); | |
599 | return 1; | |
600 | } | |
601 | ||
602 | if (!zbd_verify_sizes()) | |
603 | return 1; | |
604 | ||
605 | if (!zbd_verify_bs()) | |
606 | return 1; | |
607 | ||
608 | return 0; | |
609 | } | |
610 | ||
611 | /** | |
612 | * zbd_reset_range - reset zones for a range of sectors | |
613 | * @td: FIO thread data. | |
614 | * @f: Fio file for which to reset zones | |
615 | * @sector: Starting sector in units of 512 bytes | |
616 | * @nr_sectors: Number of sectors in units of 512 bytes | |
617 | * | |
618 | * Returns 0 upon success and a negative error code upon failure. | |
619 | */ | |
620 | static int zbd_reset_range(struct thread_data *td, const struct fio_file *f, | |
ee3696bd | 621 | uint64_t offset, uint64_t length) |
bfbdd35b BVA |
622 | { |
623 | struct blk_zone_range zr = { | |
ee3696bd DLM |
624 | .sector = offset >> 9, |
625 | .nr_sectors = length >> 9, | |
bfbdd35b BVA |
626 | }; |
627 | uint32_t zone_idx_b, zone_idx_e; | |
628 | struct fio_zone_info *zb, *ze, *z; | |
629 | int ret = 0; | |
630 | ||
631 | assert(f->fd != -1); | |
ee3696bd | 632 | assert(is_valid_offset(f, offset + length - 1)); |
bfbdd35b BVA |
633 | switch (f->zbd_info->model) { |
634 | case ZBD_DM_HOST_AWARE: | |
635 | case ZBD_DM_HOST_MANAGED: | |
636 | ret = ioctl(f->fd, BLKRESETZONE, &zr); | |
637 | if (ret < 0) { | |
638 | td_verror(td, errno, "resetting wp failed"); | |
639 | log_err("%s: resetting wp for %llu sectors at sector %llu failed (%d).\n", | |
640 | f->file_name, zr.nr_sectors, zr.sector, errno); | |
641 | return ret; | |
642 | } | |
643 | break; | |
644 | case ZBD_DM_NONE: | |
645 | break; | |
646 | } | |
647 | ||
ee3696bd | 648 | zone_idx_b = zbd_zone_idx(f, offset); |
bfbdd35b | 649 | zb = &f->zbd_info->zone_info[zone_idx_b]; |
ee3696bd | 650 | zone_idx_e = zbd_zone_idx(f, offset + length); |
bfbdd35b BVA |
651 | ze = &f->zbd_info->zone_info[zone_idx_e]; |
652 | for (z = zb; z < ze; z++) { | |
653 | pthread_mutex_lock(&z->mutex); | |
a7c2b6fc BVA |
654 | pthread_mutex_lock(&f->zbd_info->mutex); |
655 | f->zbd_info->sectors_with_data -= z->wp - z->start; | |
656 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
bfbdd35b BVA |
657 | z->wp = z->start; |
658 | z->verify_block = 0; | |
659 | pthread_mutex_unlock(&z->mutex); | |
660 | } | |
661 | ||
fd5d733f BVA |
662 | td->ts.nr_zone_resets += ze - zb; |
663 | ||
bfbdd35b BVA |
664 | return ret; |
665 | } | |
666 | ||
a0c84dd4 JA |
667 | static unsigned int zbd_zone_nr(struct zoned_block_device_info *zbd_info, |
668 | struct fio_zone_info *zone) | |
669 | { | |
089ddd95 | 670 | return zone - zbd_info->zone_info; |
a0c84dd4 JA |
671 | } |
672 | ||
bfbdd35b BVA |
673 | /** |
674 | * zbd_reset_zone - reset the write pointer of a single zone | |
675 | * @td: FIO thread data. | |
676 | * @f: FIO file associated with the disk for which to reset a write pointer. | |
677 | * @z: Zone to reset. | |
678 | * | |
679 | * Returns 0 upon success and a negative error code upon failure. | |
680 | */ | |
681 | static int zbd_reset_zone(struct thread_data *td, const struct fio_file *f, | |
682 | struct fio_zone_info *z) | |
683 | { | |
a0c84dd4 JA |
684 | dprint(FD_ZBD, "%s: resetting wp of zone %u.\n", f->file_name, |
685 | zbd_zone_nr(f->zbd_info, z)); | |
d60be7d5 DLM |
686 | |
687 | return zbd_reset_range(td, f, z->start, (z+1)->start - z->start); | |
bfbdd35b BVA |
688 | } |
689 | ||
690 | /* | |
691 | * Reset a range of zones. Returns 0 upon success and 1 upon failure. | |
692 | * @td: fio thread data. | |
693 | * @f: fio file for which to reset zones | |
694 | * @zb: first zone to reset. | |
695 | * @ze: first zone not to reset. | |
696 | * @all_zones: whether to reset all zones or only those zones for which the | |
697 | * write pointer is not a multiple of td->o.min_bs[DDIR_WRITE]. | |
698 | */ | |
699 | static int zbd_reset_zones(struct thread_data *td, struct fio_file *f, | |
700 | struct fio_zone_info *const zb, | |
701 | struct fio_zone_info *const ze, bool all_zones) | |
702 | { | |
703 | struct fio_zone_info *z, *start_z = ze; | |
ee3696bd | 704 | const uint32_t min_bs = td->o.min_bs[DDIR_WRITE]; |
bfbdd35b BVA |
705 | bool reset_wp; |
706 | int res = 0; | |
707 | ||
a0c84dd4 JA |
708 | dprint(FD_ZBD, "%s: examining zones %u .. %u\n", f->file_name, |
709 | zbd_zone_nr(f->zbd_info, zb), zbd_zone_nr(f->zbd_info, ze)); | |
bfbdd35b BVA |
710 | assert(f->fd != -1); |
711 | for (z = zb; z < ze; z++) { | |
712 | pthread_mutex_lock(&z->mutex); | |
713 | switch (z->type) { | |
714 | case BLK_ZONE_TYPE_SEQWRITE_REQ: | |
715 | reset_wp = all_zones ? z->wp != z->start : | |
716 | (td->o.td_ddir & TD_DDIR_WRITE) && | |
717 | z->wp % min_bs != 0; | |
718 | if (start_z == ze && reset_wp) { | |
719 | start_z = z; | |
720 | } else if (start_z < ze && !reset_wp) { | |
721 | dprint(FD_ZBD, | |
a0c84dd4 | 722 | "%s: resetting zones %u .. %u\n", |
bfbdd35b | 723 | f->file_name, |
a0c84dd4 JA |
724 | zbd_zone_nr(f->zbd_info, start_z), |
725 | zbd_zone_nr(f->zbd_info, z)); | |
bfbdd35b BVA |
726 | if (zbd_reset_range(td, f, start_z->start, |
727 | z->start - start_z->start) < 0) | |
728 | res = 1; | |
729 | start_z = ze; | |
730 | } | |
731 | break; | |
732 | default: | |
733 | if (start_z == ze) | |
734 | break; | |
a0c84dd4 JA |
735 | dprint(FD_ZBD, "%s: resetting zones %u .. %u\n", |
736 | f->file_name, zbd_zone_nr(f->zbd_info, start_z), | |
737 | zbd_zone_nr(f->zbd_info, z)); | |
bfbdd35b BVA |
738 | if (zbd_reset_range(td, f, start_z->start, |
739 | z->start - start_z->start) < 0) | |
740 | res = 1; | |
741 | start_z = ze; | |
742 | break; | |
743 | } | |
744 | } | |
745 | if (start_z < ze) { | |
a0c84dd4 JA |
746 | dprint(FD_ZBD, "%s: resetting zones %u .. %u\n", f->file_name, |
747 | zbd_zone_nr(f->zbd_info, start_z), | |
748 | zbd_zone_nr(f->zbd_info, z)); | |
bfbdd35b BVA |
749 | if (zbd_reset_range(td, f, start_z->start, |
750 | z->start - start_z->start) < 0) | |
751 | res = 1; | |
752 | } | |
753 | for (z = zb; z < ze; z++) | |
754 | pthread_mutex_unlock(&z->mutex); | |
755 | ||
756 | return res; | |
757 | } | |
758 | ||
a7c2b6fc BVA |
759 | /* |
760 | * Reset zbd_info.write_cnt, the counter that counts down towards the next | |
761 | * zone reset. | |
762 | */ | |
763 | static void zbd_reset_write_cnt(const struct thread_data *td, | |
764 | const struct fio_file *f) | |
765 | { | |
766 | assert(0 <= td->o.zrf.u.f && td->o.zrf.u.f <= 1); | |
767 | ||
768 | pthread_mutex_lock(&f->zbd_info->mutex); | |
769 | f->zbd_info->write_cnt = td->o.zrf.u.f ? | |
770 | min(1.0 / td->o.zrf.u.f, 0.0 + UINT_MAX) : UINT_MAX; | |
771 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
772 | } | |
773 | ||
774 | static bool zbd_dec_and_reset_write_cnt(const struct thread_data *td, | |
775 | const struct fio_file *f) | |
776 | { | |
777 | uint32_t write_cnt = 0; | |
778 | ||
779 | pthread_mutex_lock(&f->zbd_info->mutex); | |
780 | assert(f->zbd_info->write_cnt); | |
781 | if (f->zbd_info->write_cnt) | |
782 | write_cnt = --f->zbd_info->write_cnt; | |
783 | if (write_cnt == 0) | |
784 | zbd_reset_write_cnt(td, f); | |
785 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
786 | ||
787 | return write_cnt == 0; | |
788 | } | |
789 | ||
91d25131 BVA |
790 | enum swd_action { |
791 | CHECK_SWD, | |
792 | SET_SWD, | |
793 | }; | |
794 | ||
795 | /* Calculate the number of sectors with data (swd) and perform action 'a' */ | |
796 | static uint64_t zbd_process_swd(const struct fio_file *f, enum swd_action a) | |
615555bb | 797 | { |
615555bb BVA |
798 | struct fio_zone_info *zb, *ze, *z; |
799 | uint64_t swd = 0; | |
800 | ||
801 | zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)]; | |
802 | ze = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset + | |
803 | f->io_size)]; | |
804 | for (z = zb; z < ze; z++) { | |
805 | pthread_mutex_lock(&z->mutex); | |
806 | swd += z->wp - z->start; | |
807 | } | |
808 | pthread_mutex_lock(&f->zbd_info->mutex); | |
91d25131 BVA |
809 | switch (a) { |
810 | case CHECK_SWD: | |
811 | assert(f->zbd_info->sectors_with_data == swd); | |
812 | break; | |
813 | case SET_SWD: | |
814 | f->zbd_info->sectors_with_data = swd; | |
815 | break; | |
816 | } | |
615555bb BVA |
817 | pthread_mutex_unlock(&f->zbd_info->mutex); |
818 | for (z = zb; z < ze; z++) | |
819 | pthread_mutex_unlock(&z->mutex); | |
91d25131 BVA |
820 | |
821 | return swd; | |
822 | } | |
823 | ||
824 | /* | |
825 | * The swd check is useful for debugging but takes too much time to leave | |
826 | * it enabled all the time. Hence it is disabled by default. | |
827 | */ | |
828 | static const bool enable_check_swd = false; | |
829 | ||
830 | /* Check whether the value of zbd_info.sectors_with_data is correct. */ | |
831 | static void zbd_check_swd(const struct fio_file *f) | |
832 | { | |
833 | if (!enable_check_swd) | |
834 | return; | |
835 | ||
836 | zbd_process_swd(f, CHECK_SWD); | |
837 | } | |
838 | ||
839 | static void zbd_init_swd(struct fio_file *f) | |
840 | { | |
841 | uint64_t swd; | |
842 | ||
843 | swd = zbd_process_swd(f, SET_SWD); | |
844 | dprint(FD_ZBD, "%s(%s): swd = %" PRIu64 "\n", __func__, f->file_name, | |
845 | swd); | |
615555bb BVA |
846 | } |
847 | ||
bfbdd35b BVA |
848 | void zbd_file_reset(struct thread_data *td, struct fio_file *f) |
849 | { | |
91d25131 | 850 | struct fio_zone_info *zb, *ze; |
bfbdd35b BVA |
851 | uint32_t zone_idx_e; |
852 | ||
853 | if (!f->zbd_info) | |
854 | return; | |
855 | ||
856 | zb = &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)]; | |
857 | zone_idx_e = zbd_zone_idx(f, f->file_offset + f->io_size); | |
858 | ze = &f->zbd_info->zone_info[zone_idx_e]; | |
91d25131 | 859 | zbd_init_swd(f); |
bfbdd35b BVA |
860 | /* |
861 | * If data verification is enabled reset the affected zones before | |
862 | * writing any data to avoid that a zone reset has to be issued while | |
863 | * writing data, which causes data loss. | |
864 | */ | |
865 | zbd_reset_zones(td, f, zb, ze, td->o.verify != VERIFY_NONE && | |
866 | (td->o.td_ddir & TD_DDIR_WRITE) && | |
867 | td->runstate != TD_VERIFYING); | |
a7c2b6fc | 868 | zbd_reset_write_cnt(td, f); |
bfbdd35b BVA |
869 | } |
870 | ||
59b07544 BVA |
871 | /* The caller must hold f->zbd_info->mutex. */ |
872 | static bool is_zone_open(const struct thread_data *td, const struct fio_file *f, | |
873 | unsigned int zone_idx) | |
874 | { | |
875 | struct zoned_block_device_info *zbdi = f->zbd_info; | |
876 | int i; | |
877 | ||
878 | assert(td->o.max_open_zones <= ARRAY_SIZE(zbdi->open_zones)); | |
879 | assert(zbdi->num_open_zones <= td->o.max_open_zones); | |
880 | ||
881 | for (i = 0; i < zbdi->num_open_zones; i++) | |
882 | if (zbdi->open_zones[i] == zone_idx) | |
883 | return true; | |
884 | ||
885 | return false; | |
886 | } | |
887 | ||
888 | /* | |
889 | * Open a ZBD zone if it was not yet open. Returns true if either the zone was | |
890 | * already open or if opening a new zone is allowed. Returns false if the zone | |
891 | * was not yet open and opening a new zone would cause the zone limit to be | |
892 | * exceeded. | |
893 | */ | |
894 | static bool zbd_open_zone(struct thread_data *td, const struct io_u *io_u, | |
895 | uint32_t zone_idx) | |
896 | { | |
897 | const uint32_t min_bs = td->o.min_bs[DDIR_WRITE]; | |
898 | const struct fio_file *f = io_u->file; | |
899 | struct fio_zone_info *z = &f->zbd_info->zone_info[zone_idx]; | |
900 | bool res = true; | |
901 | ||
902 | if (z->cond == BLK_ZONE_COND_OFFLINE) | |
903 | return false; | |
904 | ||
905 | /* | |
906 | * Skip full zones with data verification enabled because resetting a | |
907 | * zone causes data loss and hence causes verification to fail. | |
908 | */ | |
909 | if (td->o.verify != VERIFY_NONE && zbd_zone_full(f, z, min_bs)) | |
910 | return false; | |
911 | ||
912 | /* Zero means no limit */ | |
913 | if (!td->o.max_open_zones) | |
914 | return true; | |
915 | ||
916 | pthread_mutex_lock(&f->zbd_info->mutex); | |
917 | if (is_zone_open(td, f, zone_idx)) | |
918 | goto out; | |
919 | res = false; | |
920 | if (f->zbd_info->num_open_zones >= td->o.max_open_zones) | |
921 | goto out; | |
922 | dprint(FD_ZBD, "%s: opening zone %d\n", f->file_name, zone_idx); | |
923 | f->zbd_info->open_zones[f->zbd_info->num_open_zones++] = zone_idx; | |
924 | z->open = 1; | |
925 | res = true; | |
926 | ||
927 | out: | |
928 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
929 | return res; | |
930 | } | |
931 | ||
932 | /* The caller must hold f->zbd_info->mutex */ | |
933 | static void zbd_close_zone(struct thread_data *td, const struct fio_file *f, | |
934 | unsigned int open_zone_idx) | |
935 | { | |
936 | uint32_t zone_idx; | |
937 | ||
938 | assert(open_zone_idx < f->zbd_info->num_open_zones); | |
939 | zone_idx = f->zbd_info->open_zones[open_zone_idx]; | |
940 | memmove(f->zbd_info->open_zones + open_zone_idx, | |
941 | f->zbd_info->open_zones + open_zone_idx + 1, | |
942 | (FIO_MAX_OPEN_ZBD_ZONES - (open_zone_idx + 1)) * | |
943 | sizeof(f->zbd_info->open_zones[0])); | |
944 | f->zbd_info->num_open_zones--; | |
945 | f->zbd_info->zone_info[zone_idx].open = 0; | |
946 | } | |
947 | ||
948 | /* | |
949 | * Modify the offset of an I/O unit that does not refer to an open zone such | |
950 | * that it refers to an open zone. Close an open zone and open a new zone if | |
951 | * necessary. This algorithm can only work correctly if all write pointers are | |
952 | * a multiple of the fio block size. The caller must neither hold z->mutex | |
953 | * nor f->zbd_info->mutex. Returns with z->mutex held upon success. | |
954 | */ | |
379e5f09 BVA |
955 | static struct fio_zone_info *zbd_convert_to_open_zone(struct thread_data *td, |
956 | struct io_u *io_u) | |
59b07544 BVA |
957 | { |
958 | const uint32_t min_bs = td->o.min_bs[io_u->ddir]; | |
959 | const struct fio_file *f = io_u->file; | |
960 | struct fio_zone_info *z; | |
961 | unsigned int open_zone_idx = -1; | |
962 | uint32_t zone_idx, new_zone_idx; | |
963 | int i; | |
964 | ||
965 | assert(is_valid_offset(f, io_u->offset)); | |
966 | ||
967 | if (td->o.max_open_zones) { | |
968 | /* | |
969 | * This statement accesses f->zbd_info->open_zones[] on purpose | |
970 | * without locking. | |
971 | */ | |
972 | zone_idx = f->zbd_info->open_zones[(io_u->offset - | |
973 | f->file_offset) * | |
974 | f->zbd_info->num_open_zones / f->io_size]; | |
975 | } else { | |
976 | zone_idx = zbd_zone_idx(f, io_u->offset); | |
977 | } | |
978 | dprint(FD_ZBD, "%s(%s): starting from zone %d (offset %lld, buflen %lld)\n", | |
979 | __func__, f->file_name, zone_idx, io_u->offset, io_u->buflen); | |
980 | ||
981 | /* | |
982 | * Since z->mutex is the outer lock and f->zbd_info->mutex the inner | |
983 | * lock it can happen that the state of the zone with index zone_idx | |
984 | * has changed after 'z' has been assigned and before f->zbd_info->mutex | |
985 | * has been obtained. Hence the loop. | |
986 | */ | |
987 | for (;;) { | |
988 | z = &f->zbd_info->zone_info[zone_idx]; | |
989 | ||
990 | pthread_mutex_lock(&z->mutex); | |
991 | pthread_mutex_lock(&f->zbd_info->mutex); | |
992 | if (td->o.max_open_zones == 0) | |
993 | goto examine_zone; | |
994 | if (f->zbd_info->num_open_zones == 0) { | |
995 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
996 | pthread_mutex_unlock(&z->mutex); | |
997 | dprint(FD_ZBD, "%s(%s): no zones are open\n", | |
998 | __func__, f->file_name); | |
999 | return NULL; | |
1000 | } | |
1001 | open_zone_idx = (io_u->offset - f->file_offset) * | |
1002 | f->zbd_info->num_open_zones / f->io_size; | |
1003 | assert(open_zone_idx < f->zbd_info->num_open_zones); | |
1004 | new_zone_idx = f->zbd_info->open_zones[open_zone_idx]; | |
1005 | if (new_zone_idx == zone_idx) | |
1006 | break; | |
1007 | zone_idx = new_zone_idx; | |
1008 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
1009 | pthread_mutex_unlock(&z->mutex); | |
1010 | } | |
1011 | ||
1012 | /* Both z->mutex and f->zbd_info->mutex are held. */ | |
1013 | ||
1014 | examine_zone: | |
ee3696bd | 1015 | if (z->wp + min_bs <= (z+1)->start) { |
59b07544 BVA |
1016 | pthread_mutex_unlock(&f->zbd_info->mutex); |
1017 | goto out; | |
1018 | } | |
1019 | dprint(FD_ZBD, "%s(%s): closing zone %d\n", __func__, f->file_name, | |
1020 | zone_idx); | |
1021 | if (td->o.max_open_zones) | |
1022 | zbd_close_zone(td, f, open_zone_idx); | |
1023 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
1024 | ||
1025 | /* Only z->mutex is held. */ | |
1026 | ||
1027 | /* Zone 'z' is full, so try to open a new zone. */ | |
1028 | for (i = f->io_size / f->zbd_info->zone_size; i > 0; i--) { | |
1029 | zone_idx++; | |
1030 | pthread_mutex_unlock(&z->mutex); | |
1031 | z++; | |
ee3696bd | 1032 | if (!is_valid_offset(f, z->start)) { |
59b07544 BVA |
1033 | /* Wrap-around. */ |
1034 | zone_idx = zbd_zone_idx(f, f->file_offset); | |
1035 | z = &f->zbd_info->zone_info[zone_idx]; | |
1036 | } | |
ee3696bd | 1037 | assert(is_valid_offset(f, z->start)); |
59b07544 BVA |
1038 | pthread_mutex_lock(&z->mutex); |
1039 | if (z->open) | |
1040 | continue; | |
1041 | if (zbd_open_zone(td, io_u, zone_idx)) | |
1042 | goto out; | |
1043 | } | |
1044 | ||
1045 | /* Only z->mutex is held. */ | |
1046 | ||
1047 | /* Check whether the write fits in any of the already opened zones. */ | |
1048 | pthread_mutex_lock(&f->zbd_info->mutex); | |
1049 | for (i = 0; i < f->zbd_info->num_open_zones; i++) { | |
1050 | zone_idx = f->zbd_info->open_zones[i]; | |
1051 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
1052 | pthread_mutex_unlock(&z->mutex); | |
1053 | ||
1054 | z = &f->zbd_info->zone_info[zone_idx]; | |
1055 | ||
1056 | pthread_mutex_lock(&z->mutex); | |
ee3696bd | 1057 | if (z->wp + min_bs <= (z+1)->start) |
59b07544 BVA |
1058 | goto out; |
1059 | pthread_mutex_lock(&f->zbd_info->mutex); | |
1060 | } | |
1061 | pthread_mutex_unlock(&f->zbd_info->mutex); | |
1062 | pthread_mutex_unlock(&z->mutex); | |
1063 | dprint(FD_ZBD, "%s(%s): did not open another zone\n", __func__, | |
1064 | f->file_name); | |
1065 | return NULL; | |
1066 | ||
1067 | out: | |
1068 | dprint(FD_ZBD, "%s(%s): returning zone %d\n", __func__, f->file_name, | |
1069 | zone_idx); | |
ee3696bd | 1070 | io_u->offset = z->start; |
59b07544 BVA |
1071 | return z; |
1072 | } | |
1073 | ||
bfbdd35b | 1074 | /* The caller must hold z->mutex. */ |
59b07544 BVA |
1075 | static struct fio_zone_info *zbd_replay_write_order(struct thread_data *td, |
1076 | struct io_u *io_u, | |
1077 | struct fio_zone_info *z) | |
bfbdd35b BVA |
1078 | { |
1079 | const struct fio_file *f = io_u->file; | |
1080 | const uint32_t min_bs = td->o.min_bs[DDIR_WRITE]; | |
1081 | ||
59b07544 BVA |
1082 | if (!zbd_open_zone(td, io_u, z - f->zbd_info->zone_info)) { |
1083 | pthread_mutex_unlock(&z->mutex); | |
1084 | z = zbd_convert_to_open_zone(td, io_u); | |
1085 | assert(z); | |
1086 | } | |
1087 | ||
bfbdd35b | 1088 | if (z->verify_block * min_bs >= f->zbd_info->zone_size) |
a0c84dd4 JA |
1089 | log_err("%s: %d * %d >= %llu\n", f->file_name, z->verify_block, |
1090 | min_bs, (unsigned long long) f->zbd_info->zone_size); | |
ee3696bd | 1091 | io_u->offset = z->start + z->verify_block++ * min_bs; |
59b07544 | 1092 | return z; |
bfbdd35b BVA |
1093 | } |
1094 | ||
1095 | /* | |
1096 | * Find another zone for which @io_u fits below the write pointer. Start | |
1097 | * searching in zones @zb + 1 .. @zl and continue searching in zones | |
1098 | * @zf .. @zb - 1. | |
1099 | * | |
1100 | * Either returns NULL or returns a zone pointer and holds the mutex for that | |
1101 | * zone. | |
1102 | */ | |
1103 | static struct fio_zone_info * | |
1104 | zbd_find_zone(struct thread_data *td, struct io_u *io_u, | |
1105 | struct fio_zone_info *zb, struct fio_zone_info *zl) | |
1106 | { | |
1107 | const uint32_t min_bs = td->o.min_bs[io_u->ddir]; | |
1108 | const struct fio_file *f = io_u->file; | |
1109 | struct fio_zone_info *z1, *z2; | |
1110 | const struct fio_zone_info *const zf = | |
1111 | &f->zbd_info->zone_info[zbd_zone_idx(f, f->file_offset)]; | |
1112 | ||
1113 | /* | |
1114 | * Skip to the next non-empty zone in case of sequential I/O and to | |
1115 | * the nearest non-empty zone in case of random I/O. | |
1116 | */ | |
1117 | for (z1 = zb + 1, z2 = zb - 1; z1 < zl || z2 >= zf; z1++, z2--) { | |
1118 | if (z1 < zl && z1->cond != BLK_ZONE_COND_OFFLINE) { | |
1119 | pthread_mutex_lock(&z1->mutex); | |
ee3696bd | 1120 | if (z1->start + min_bs <= z1->wp) |
bfbdd35b BVA |
1121 | return z1; |
1122 | pthread_mutex_unlock(&z1->mutex); | |
1123 | } else if (!td_random(td)) { | |
1124 | break; | |
1125 | } | |
1126 | if (td_random(td) && z2 >= zf && | |
1127 | z2->cond != BLK_ZONE_COND_OFFLINE) { | |
1128 | pthread_mutex_lock(&z2->mutex); | |
ee3696bd | 1129 | if (z2->start + min_bs <= z2->wp) |
bfbdd35b BVA |
1130 | return z2; |
1131 | pthread_mutex_unlock(&z2->mutex); | |
1132 | } | |
1133 | } | |
1134 | dprint(FD_ZBD, "%s: adjusting random read offset failed\n", | |
1135 | f->file_name); | |
1136 | return NULL; | |
1137 | } | |
1138 | ||
bfbdd35b | 1139 | /** |
d9ed3e63 | 1140 | * zbd_queue_io - update the write pointer of a sequential zone |
bfbdd35b | 1141 | * @io_u: I/O unit |
d9ed3e63 DLM |
1142 | * @success: Whether or not the I/O unit has been queued successfully |
1143 | * @q: queueing status (busy, completed or queued). | |
bfbdd35b | 1144 | * |
d9ed3e63 DLM |
1145 | * For write and trim operations, update the write pointer of the I/O unit |
1146 | * target zone. | |
bfbdd35b | 1147 | */ |
d9ed3e63 | 1148 | static void zbd_queue_io(struct io_u *io_u, int q, bool success) |
bfbdd35b | 1149 | { |
d9ed3e63 DLM |
1150 | const struct fio_file *f = io_u->file; |
1151 | struct zoned_block_device_info *zbd_info = f->zbd_info; | |
bfbdd35b BVA |
1152 | struct fio_zone_info *z; |
1153 | uint32_t zone_idx; | |
d9ed3e63 | 1154 | uint64_t zone_end; |
bfbdd35b | 1155 | |
bfbdd35b BVA |
1156 | if (!zbd_info) |
1157 | return; | |
1158 | ||
d9ed3e63 | 1159 | zone_idx = zbd_zone_idx(f, io_u->offset); |
bfbdd35b | 1160 | assert(zone_idx < zbd_info->nr_zones); |
d9ed3e63 DLM |
1161 | z = &zbd_info->zone_info[zone_idx]; |
1162 | ||
bfbdd35b BVA |
1163 | if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ) |
1164 | return; | |
d9ed3e63 | 1165 | |
bfbdd35b BVA |
1166 | if (!success) |
1167 | goto unlock; | |
d9ed3e63 DLM |
1168 | |
1169 | dprint(FD_ZBD, | |
1170 | "%s: queued I/O (%lld, %llu) for zone %u\n", | |
1171 | f->file_name, io_u->offset, io_u->buflen, zone_idx); | |
1172 | ||
bfbdd35b BVA |
1173 | switch (io_u->ddir) { |
1174 | case DDIR_WRITE: | |
d9ed3e63 DLM |
1175 | zone_end = min((uint64_t)(io_u->offset + io_u->buflen), |
1176 | (z + 1)->start); | |
a7c2b6fc BVA |
1177 | pthread_mutex_lock(&zbd_info->mutex); |
1178 | /* | |
1179 | * z->wp > zone_end means that one or more I/O errors | |
1180 | * have occurred. | |
1181 | */ | |
1182 | if (z->wp <= zone_end) | |
1183 | zbd_info->sectors_with_data += zone_end - z->wp; | |
1184 | pthread_mutex_unlock(&zbd_info->mutex); | |
bfbdd35b BVA |
1185 | z->wp = zone_end; |
1186 | break; | |
1187 | case DDIR_TRIM: | |
1188 | assert(z->wp == z->start); | |
1189 | break; | |
1190 | default: | |
1191 | break; | |
1192 | } | |
d9ed3e63 | 1193 | |
bfbdd35b | 1194 | unlock: |
d9ed3e63 DLM |
1195 | if (!success || q != FIO_Q_QUEUED) { |
1196 | /* BUSY or COMPLETED: unlock the zone */ | |
1197 | pthread_mutex_unlock(&z->mutex); | |
1198 | io_u->zbd_put_io = NULL; | |
1199 | } | |
1200 | } | |
1201 | ||
1202 | /** | |
1203 | * zbd_put_io - Unlock an I/O unit target zone lock | |
1204 | * @io_u: I/O unit | |
1205 | */ | |
1206 | static void zbd_put_io(const struct io_u *io_u) | |
1207 | { | |
1208 | const struct fio_file *f = io_u->file; | |
1209 | struct zoned_block_device_info *zbd_info = f->zbd_info; | |
1210 | struct fio_zone_info *z; | |
1211 | uint32_t zone_idx; | |
1212 | ||
1213 | if (!zbd_info) | |
1214 | return; | |
615555bb | 1215 | |
d9ed3e63 DLM |
1216 | zone_idx = zbd_zone_idx(f, io_u->offset); |
1217 | assert(zone_idx < zbd_info->nr_zones); | |
1218 | z = &zbd_info->zone_info[zone_idx]; | |
1219 | ||
1220 | if (z->type != BLK_ZONE_TYPE_SEQWRITE_REQ) | |
1221 | return; | |
1222 | ||
1223 | dprint(FD_ZBD, | |
1224 | "%s: terminate I/O (%lld, %llu) for zone %u\n", | |
1225 | f->file_name, io_u->offset, io_u->buflen, zone_idx); | |
1226 | ||
1227 | assert(pthread_mutex_unlock(&z->mutex) == 0); | |
1228 | zbd_check_swd(f); | |
bfbdd35b BVA |
1229 | } |
1230 | ||
1231 | bool zbd_unaligned_write(int error_code) | |
1232 | { | |
1233 | switch (error_code) { | |
1234 | case EIO: | |
1235 | case EREMOTEIO: | |
1236 | return true; | |
1237 | } | |
1238 | return false; | |
1239 | } | |
1240 | ||
4d37720a DLM |
1241 | /** |
1242 | * setup_zbd_zone_mode - handle zoneskip as necessary for ZBD drives | |
1243 | * @td: FIO thread data. | |
1244 | * @io_u: FIO I/O unit. | |
1245 | * | |
1246 | * For sequential workloads, change the file offset to skip zoneskip bytes when | |
1247 | * no more IO can be performed in the current zone. | |
1248 | * - For read workloads, zoneskip is applied when the io has reached the end of | |
1249 | * the zone or the zone write position (when td->o.read_beyond_wp is false). | |
1250 | * - For write workloads, zoneskip is applied when the zone is full. | |
1251 | * This applies only to read and write operations. | |
1252 | */ | |
1253 | void setup_zbd_zone_mode(struct thread_data *td, struct io_u *io_u) | |
1254 | { | |
1255 | struct fio_file *f = io_u->file; | |
1256 | enum fio_ddir ddir = io_u->ddir; | |
1257 | struct fio_zone_info *z; | |
1258 | uint32_t zone_idx; | |
1259 | ||
1260 | assert(td->o.zone_mode == ZONE_MODE_ZBD); | |
1261 | assert(td->o.zone_size); | |
1262 | ||
1263 | /* | |
1264 | * zone_skip is valid only for sequential workloads. | |
1265 | */ | |
1266 | if (td_random(td) || !td->o.zone_skip) | |
1267 | return; | |
1268 | ||
1269 | /* | |
1270 | * It is time to switch to a new zone if: | |
1271 | * - zone_bytes == zone_size bytes have already been accessed | |
1272 | * - The last position reached the end of the current zone. | |
1273 | * - For reads with td->o.read_beyond_wp == false, the last position | |
1274 | * reached the zone write pointer. | |
1275 | */ | |
1276 | zone_idx = zbd_zone_idx(f, f->last_pos[ddir]); | |
1277 | z = &f->zbd_info->zone_info[zone_idx]; | |
1278 | ||
1279 | if (td->zone_bytes >= td->o.zone_size || | |
1280 | f->last_pos[ddir] >= (z+1)->start || | |
1281 | (ddir == DDIR_READ && | |
1282 | (!td->o.read_beyond_wp) && f->last_pos[ddir] >= z->wp)) { | |
1283 | /* | |
1284 | * Skip zones. | |
1285 | */ | |
1286 | td->zone_bytes = 0; | |
1287 | f->file_offset += td->o.zone_size + td->o.zone_skip; | |
1288 | ||
1289 | /* | |
1290 | * Wrap from the beginning, if we exceed the file size | |
1291 | */ | |
1292 | if (f->file_offset >= f->real_file_size) | |
1293 | f->file_offset = get_start_offset(td, f); | |
1294 | ||
1295 | f->last_pos[ddir] = f->file_offset; | |
1296 | td->io_skip_bytes += td->o.zone_skip; | |
1297 | } | |
1298 | } | |
1299 | ||
bfbdd35b BVA |
1300 | /** |
1301 | * zbd_adjust_block - adjust the offset and length as necessary for ZBD drives | |
1302 | * @td: FIO thread data. | |
1303 | * @io_u: FIO I/O unit. | |
1304 | * | |
1305 | * Locking strategy: returns with z->mutex locked if and only if z refers | |
1306 | * to a sequential zone and if io_u_accept is returned. z is the zone that | |
1307 | * corresponds to io_u->offset at the end of this function. | |
1308 | */ | |
1309 | enum io_u_action zbd_adjust_block(struct thread_data *td, struct io_u *io_u) | |
1310 | { | |
1311 | const struct fio_file *f = io_u->file; | |
1312 | uint32_t zone_idx_b; | |
de65f7b7 | 1313 | struct fio_zone_info *zb, *zl, *orig_zb; |
bfbdd35b BVA |
1314 | uint32_t orig_len = io_u->buflen; |
1315 | uint32_t min_bs = td->o.min_bs[io_u->ddir]; | |
1316 | uint64_t new_len; | |
1317 | int64_t range; | |
1318 | ||
1319 | if (!f->zbd_info) | |
1320 | return io_u_accept; | |
1321 | ||
1322 | assert(is_valid_offset(f, io_u->offset)); | |
1323 | assert(io_u->buflen); | |
1324 | zone_idx_b = zbd_zone_idx(f, io_u->offset); | |
1325 | zb = &f->zbd_info->zone_info[zone_idx_b]; | |
de65f7b7 | 1326 | orig_zb = zb; |
bfbdd35b BVA |
1327 | |
1328 | /* Accept the I/O offset for conventional zones. */ | |
1329 | if (zb->type == BLK_ZONE_TYPE_CONVENTIONAL) | |
1330 | return io_u_accept; | |
1331 | ||
1332 | /* | |
1333 | * Accept the I/O offset for reads if reading beyond the write pointer | |
1334 | * is enabled. | |
1335 | */ | |
1336 | if (zb->cond != BLK_ZONE_COND_OFFLINE && | |
1337 | io_u->ddir == DDIR_READ && td->o.read_beyond_wp) | |
1338 | return io_u_accept; | |
1339 | ||
615555bb BVA |
1340 | zbd_check_swd(f); |
1341 | ||
6f0c6085 DLM |
1342 | /* |
1343 | * Lock the io_u target zone. The zone will be unlocked if io_u offset | |
1344 | * is changed or when io_u completes and zbd_put_io() executed. | |
1345 | * To avoid multiple jobs doing asynchronous I/Os from deadlocking each | |
1346 | * other waiting for zone locks when building an io_u batch, first | |
1347 | * only trylock the zone. If the zone is already locked by another job, | |
1348 | * process the currently queued I/Os so that I/O progress is made and | |
1349 | * zones unlocked. | |
1350 | */ | |
1351 | if (pthread_mutex_trylock(&zb->mutex) != 0) { | |
1352 | if (!td_ioengine_flagged(td, FIO_SYNCIO)) | |
1353 | io_u_quiesce(td); | |
1354 | pthread_mutex_lock(&zb->mutex); | |
1355 | } | |
1356 | ||
bfbdd35b BVA |
1357 | switch (io_u->ddir) { |
1358 | case DDIR_READ: | |
1359 | if (td->runstate == TD_VERIFYING) { | |
59b07544 | 1360 | zb = zbd_replay_write_order(td, io_u, zb); |
bfbdd35b BVA |
1361 | goto accept; |
1362 | } | |
1363 | /* | |
de65f7b7 DLM |
1364 | * Check that there is enough written data in the zone to do an |
1365 | * I/O of at least min_bs B. If there isn't, find a new zone for | |
1366 | * the I/O. | |
bfbdd35b BVA |
1367 | */ |
1368 | range = zb->cond != BLK_ZONE_COND_OFFLINE ? | |
ee3696bd | 1369 | zb->wp - zb->start : 0; |
de65f7b7 | 1370 | if (range < min_bs || |
ee3696bd | 1371 | ((!td_random(td)) && (io_u->offset + min_bs > zb->wp))) { |
bfbdd35b BVA |
1372 | pthread_mutex_unlock(&zb->mutex); |
1373 | zl = &f->zbd_info->zone_info[zbd_zone_idx(f, | |
1374 | f->file_offset + f->io_size)]; | |
1375 | zb = zbd_find_zone(td, io_u, zb, zl); | |
1376 | if (!zb) { | |
1377 | dprint(FD_ZBD, | |
1378 | "%s: zbd_find_zone(%lld, %llu) failed\n", | |
1379 | f->file_name, io_u->offset, | |
1380 | io_u->buflen); | |
1381 | goto eof; | |
1382 | } | |
de65f7b7 DLM |
1383 | /* |
1384 | * zbd_find_zone() returned a zone with a range of at | |
1385 | * least min_bs. | |
1386 | */ | |
ee3696bd | 1387 | range = zb->wp - zb->start; |
de65f7b7 DLM |
1388 | assert(range >= min_bs); |
1389 | ||
1390 | if (!td_random(td)) | |
ee3696bd | 1391 | io_u->offset = zb->start; |
bfbdd35b | 1392 | } |
de65f7b7 DLM |
1393 | /* |
1394 | * Make sure the I/O is within the zone valid data range while | |
1395 | * maximizing the I/O size and preserving randomness. | |
1396 | */ | |
1397 | if (range <= io_u->buflen) | |
ee3696bd | 1398 | io_u->offset = zb->start; |
de65f7b7 | 1399 | else if (td_random(td)) |
ee3696bd DLM |
1400 | io_u->offset = zb->start + |
1401 | ((io_u->offset - orig_zb->start) % | |
de65f7b7 DLM |
1402 | (range - io_u->buflen)) / min_bs * min_bs; |
1403 | /* | |
1404 | * Make sure the I/O does not cross over the zone wp position. | |
1405 | */ | |
1406 | new_len = min((unsigned long long)io_u->buflen, | |
ee3696bd | 1407 | (unsigned long long)(zb->wp - io_u->offset)); |
de65f7b7 DLM |
1408 | new_len = new_len / min_bs * min_bs; |
1409 | if (new_len < io_u->buflen) { | |
1410 | io_u->buflen = new_len; | |
1411 | dprint(FD_IO, "Changed length from %u into %llu\n", | |
1412 | orig_len, io_u->buflen); | |
bfbdd35b | 1413 | } |
ee3696bd DLM |
1414 | assert(zb->start <= io_u->offset); |
1415 | assert(io_u->offset + io_u->buflen <= zb->wp); | |
bfbdd35b BVA |
1416 | goto accept; |
1417 | case DDIR_WRITE: | |
ee3696bd | 1418 | if (io_u->buflen > f->zbd_info->zone_size) |
bfbdd35b | 1419 | goto eof; |
59b07544 BVA |
1420 | if (!zbd_open_zone(td, io_u, zone_idx_b)) { |
1421 | pthread_mutex_unlock(&zb->mutex); | |
1422 | zb = zbd_convert_to_open_zone(td, io_u); | |
1423 | if (!zb) | |
1424 | goto eof; | |
1425 | zone_idx_b = zb - f->zbd_info->zone_info; | |
1426 | } | |
a7c2b6fc BVA |
1427 | /* Check whether the zone reset threshold has been exceeded */ |
1428 | if (td->o.zrf.u.f) { | |
ee3696bd | 1429 | if (f->zbd_info->sectors_with_data >= |
a7c2b6fc BVA |
1430 | f->io_size * td->o.zrt.u.f && |
1431 | zbd_dec_and_reset_write_cnt(td, f)) { | |
1432 | zb->reset_zone = 1; | |
1433 | } | |
1434 | } | |
bfbdd35b BVA |
1435 | /* Reset the zone pointer if necessary */ |
1436 | if (zb->reset_zone || zbd_zone_full(f, zb, min_bs)) { | |
1437 | assert(td->o.verify == VERIFY_NONE); | |
1438 | /* | |
1439 | * Since previous write requests may have been submitted | |
1440 | * asynchronously and since we will submit the zone | |
1441 | * reset synchronously, wait until previously submitted | |
1442 | * write requests have completed before issuing a | |
1443 | * zone reset. | |
1444 | */ | |
1445 | io_u_quiesce(td); | |
1446 | zb->reset_zone = 0; | |
1447 | if (zbd_reset_zone(td, f, zb) < 0) | |
1448 | goto eof; | |
1449 | } | |
1450 | /* Make writes occur at the write pointer */ | |
1451 | assert(!zbd_zone_full(f, zb, min_bs)); | |
ee3696bd | 1452 | io_u->offset = zb->wp; |
bfbdd35b BVA |
1453 | if (!is_valid_offset(f, io_u->offset)) { |
1454 | dprint(FD_ZBD, "Dropped request with offset %llu\n", | |
1455 | io_u->offset); | |
1456 | goto eof; | |
1457 | } | |
1458 | /* | |
1459 | * Make sure that the buflen is a multiple of the minimal | |
1460 | * block size. Give up if shrinking would make the request too | |
1461 | * small. | |
1462 | */ | |
1463 | new_len = min((unsigned long long)io_u->buflen, | |
ee3696bd | 1464 | (zb + 1)->start - io_u->offset); |
bfbdd35b BVA |
1465 | new_len = new_len / min_bs * min_bs; |
1466 | if (new_len == io_u->buflen) | |
1467 | goto accept; | |
1468 | if (new_len >= min_bs) { | |
1469 | io_u->buflen = new_len; | |
1470 | dprint(FD_IO, "Changed length from %u into %llu\n", | |
1471 | orig_len, io_u->buflen); | |
1472 | goto accept; | |
1473 | } | |
1474 | log_err("Zone remainder %lld smaller than minimum block size %d\n", | |
ee3696bd | 1475 | ((zb + 1)->start - io_u->offset), |
bfbdd35b BVA |
1476 | min_bs); |
1477 | goto eof; | |
1478 | case DDIR_TRIM: | |
1479 | /* fall-through */ | |
1480 | case DDIR_SYNC: | |
1481 | case DDIR_DATASYNC: | |
1482 | case DDIR_SYNC_FILE_RANGE: | |
1483 | case DDIR_WAIT: | |
1484 | case DDIR_LAST: | |
1485 | case DDIR_INVAL: | |
1486 | goto accept; | |
1487 | } | |
1488 | ||
1489 | assert(false); | |
1490 | ||
1491 | accept: | |
1492 | assert(zb); | |
1493 | assert(zb->cond != BLK_ZONE_COND_OFFLINE); | |
d9ed3e63 DLM |
1494 | assert(!io_u->zbd_queue_io); |
1495 | assert(!io_u->zbd_put_io); | |
1496 | io_u->zbd_queue_io = zbd_queue_io; | |
1497 | io_u->zbd_put_io = zbd_put_io; | |
bfbdd35b BVA |
1498 | return io_u_accept; |
1499 | ||
1500 | eof: | |
1501 | if (zb) | |
1502 | pthread_mutex_unlock(&zb->mutex); | |
1503 | return io_u_eof; | |
1504 | } | |
fd5d733f BVA |
1505 | |
1506 | /* Return a string with ZBD statistics */ | |
1507 | char *zbd_write_status(const struct thread_stat *ts) | |
1508 | { | |
1509 | char *res; | |
1510 | ||
a0c84dd4 | 1511 | if (asprintf(&res, "; %llu zone resets", (unsigned long long) ts->nr_zone_resets) < 0) |
fd5d733f BVA |
1512 | return NULL; |
1513 | return res; | |
1514 | } |