Commit | Line | Data |
---|---|---|
a3ff873e AK |
1 | /* |
2 | * fio xNVMe IO Engine | |
3 | * | |
4 | * IO engine using the xNVMe C API. | |
5 | * | |
6 | * See: http://xnvme.io/ | |
7 | * | |
8 | * SPDX-License-Identifier: Apache-2.0 | |
9 | */ | |
10 | #include <stdlib.h> | |
11 | #include <assert.h> | |
12 | #include <libxnvme.h> | |
13 | #include <libxnvme_libconf.h> | |
14 | #include <libxnvme_nvm.h> | |
15 | #include <libxnvme_znd.h> | |
16 | #include <libxnvme_spec_fs.h> | |
17 | #include "fio.h" | |
18 | #include "zbd_types.h" | |
19 | #include "optgroup.h" | |
20 | ||
21 | static pthread_mutex_t g_serialize = PTHREAD_MUTEX_INITIALIZER; | |
22 | ||
23 | struct xnvme_fioe_fwrap { | |
24 | /* fio file representation */ | |
25 | struct fio_file *fio_file; | |
26 | ||
27 | /* xNVMe device handle */ | |
28 | struct xnvme_dev *dev; | |
29 | /* xNVMe device geometry */ | |
30 | const struct xnvme_geo *geo; | |
31 | ||
32 | struct xnvme_queue *queue; | |
33 | ||
34 | uint32_t ssw; | |
35 | uint32_t lba_nbytes; | |
36 | ||
37 | uint8_t _pad[24]; | |
38 | }; | |
39 | XNVME_STATIC_ASSERT(sizeof(struct xnvme_fioe_fwrap) == 64, "Incorrect size") | |
40 | ||
41 | struct xnvme_fioe_data { | |
42 | /* I/O completion queue */ | |
43 | struct io_u **iocq; | |
44 | ||
45 | /* # of iocq entries; incremented via getevents()/cb_pool() */ | |
46 | uint64_t completed; | |
47 | ||
48 | /* | |
49 | * # of errors; incremented when observed on completion via | |
50 | * getevents()/cb_pool() | |
51 | */ | |
52 | uint64_t ecount; | |
53 | ||
54 | /* Controller which device/file to select */ | |
55 | int32_t prev; | |
56 | int32_t cur; | |
57 | ||
58 | /* Number of devices/files for which open() has been called */ | |
59 | int64_t nopen; | |
60 | /* Number of devices/files allocated in files[] */ | |
61 | uint64_t nallocated; | |
62 | ||
63 | struct iovec *iovec; | |
64 | ||
65 | uint8_t _pad[8]; | |
66 | ||
67 | struct xnvme_fioe_fwrap files[]; | |
68 | }; | |
69 | XNVME_STATIC_ASSERT(sizeof(struct xnvme_fioe_data) == 64, "Incorrect size") | |
70 | ||
71 | struct xnvme_fioe_options { | |
72 | void *padding; | |
73 | unsigned int hipri; | |
74 | unsigned int sqpoll_thread; | |
75 | unsigned int xnvme_dev_nsid; | |
76 | unsigned int xnvme_iovec; | |
77 | char *xnvme_be; | |
78 | char *xnvme_async; | |
79 | char *xnvme_sync; | |
80 | char *xnvme_admin; | |
81 | }; | |
82 | ||
83 | static struct fio_option options[] = { | |
84 | { | |
85 | .name = "hipri", | |
86 | .lname = "High Priority", | |
87 | .type = FIO_OPT_STR_SET, | |
88 | .off1 = offsetof(struct xnvme_fioe_options, hipri), | |
89 | .help = "Use polled IO completions", | |
90 | .category = FIO_OPT_C_ENGINE, | |
91 | .group = FIO_OPT_G_XNVME, | |
92 | }, | |
93 | { | |
94 | .name = "sqthread_poll", | |
95 | .lname = "Kernel SQ thread polling", | |
96 | .type = FIO_OPT_STR_SET, | |
97 | .off1 = offsetof(struct xnvme_fioe_options, sqpoll_thread), | |
98 | .help = "Offload submission/completion to kernel thread", | |
99 | .category = FIO_OPT_C_ENGINE, | |
100 | .group = FIO_OPT_G_XNVME, | |
101 | }, | |
102 | { | |
103 | .name = "xnvme_be", | |
104 | .lname = "xNVMe Backend", | |
105 | .type = FIO_OPT_STR_STORE, | |
106 | .off1 = offsetof(struct xnvme_fioe_options, xnvme_be), | |
107 | .help = "Select xNVMe backend [spdk,linux,fbsd]", | |
108 | .category = FIO_OPT_C_ENGINE, | |
109 | .group = FIO_OPT_G_XNVME, | |
110 | }, | |
111 | { | |
112 | .name = "xnvme_async", | |
113 | .lname = "xNVMe Asynchronous command-interface", | |
114 | .type = FIO_OPT_STR_STORE, | |
115 | .off1 = offsetof(struct xnvme_fioe_options, xnvme_async), | |
203a4c7c AK |
116 | .help = "Select xNVMe async. interface: " |
117 | "[emu,thrpool,io_uring,io_uring_cmd,libaio,posix,vfio,nil]", | |
a3ff873e AK |
118 | .category = FIO_OPT_C_ENGINE, |
119 | .group = FIO_OPT_G_XNVME, | |
120 | }, | |
121 | { | |
122 | .name = "xnvme_sync", | |
123 | .lname = "xNVMe Synchronous. command-interface", | |
124 | .type = FIO_OPT_STR_STORE, | |
125 | .off1 = offsetof(struct xnvme_fioe_options, xnvme_sync), | |
203a4c7c | 126 | .help = "Select xNVMe sync. interface: [nvme,psync,block]", |
a3ff873e AK |
127 | .category = FIO_OPT_C_ENGINE, |
128 | .group = FIO_OPT_G_XNVME, | |
129 | }, | |
130 | { | |
131 | .name = "xnvme_admin", | |
132 | .lname = "xNVMe Admin command-interface", | |
133 | .type = FIO_OPT_STR_STORE, | |
134 | .off1 = offsetof(struct xnvme_fioe_options, xnvme_admin), | |
203a4c7c | 135 | .help = "Select xNVMe admin. cmd-interface: [nvme,block]", |
a3ff873e AK |
136 | .category = FIO_OPT_C_ENGINE, |
137 | .group = FIO_OPT_G_XNVME, | |
138 | }, | |
139 | { | |
140 | .name = "xnvme_dev_nsid", | |
141 | .lname = "xNVMe Namespace-Identifier, for user-space NVMe driver", | |
142 | .type = FIO_OPT_INT, | |
143 | .off1 = offsetof(struct xnvme_fioe_options, xnvme_dev_nsid), | |
144 | .help = "xNVMe Namespace-Identifier, for user-space NVMe driver", | |
145 | .category = FIO_OPT_C_ENGINE, | |
146 | .group = FIO_OPT_G_XNVME, | |
147 | }, | |
148 | { | |
149 | .name = "xnvme_iovec", | |
150 | .lname = "Vectored IOs", | |
151 | .type = FIO_OPT_STR_SET, | |
152 | .off1 = offsetof(struct xnvme_fioe_options, xnvme_iovec), | |
153 | .help = "Send vectored IOs", | |
154 | .category = FIO_OPT_C_ENGINE, | |
155 | .group = FIO_OPT_G_XNVME, | |
156 | }, | |
157 | ||
158 | { | |
159 | .name = NULL, | |
160 | }, | |
161 | }; | |
162 | ||
163 | static void cb_pool(struct xnvme_cmd_ctx *ctx, void *cb_arg) | |
164 | { | |
165 | struct io_u *io_u = cb_arg; | |
166 | struct xnvme_fioe_data *xd = io_u->mmap_data; | |
167 | ||
168 | if (xnvme_cmd_ctx_cpl_status(ctx)) { | |
169 | xnvme_cmd_ctx_pr(ctx, XNVME_PR_DEF); | |
170 | xd->ecount += 1; | |
171 | io_u->error = EIO; | |
172 | } | |
173 | ||
174 | xd->iocq[xd->completed++] = io_u; | |
175 | xnvme_queue_put_cmd_ctx(ctx->async.queue, ctx); | |
176 | } | |
177 | ||
178 | static struct xnvme_opts xnvme_opts_from_fioe(struct thread_data *td) | |
179 | { | |
180 | struct xnvme_fioe_options *o = td->eo; | |
181 | struct xnvme_opts opts = xnvme_opts_default(); | |
182 | ||
183 | opts.nsid = o->xnvme_dev_nsid; | |
184 | opts.be = o->xnvme_be; | |
185 | opts.async = o->xnvme_async; | |
186 | opts.sync = o->xnvme_sync; | |
187 | opts.admin = o->xnvme_admin; | |
188 | ||
189 | opts.poll_io = o->hipri; | |
190 | opts.poll_sq = o->sqpoll_thread; | |
191 | ||
192 | opts.direct = td->o.odirect; | |
193 | ||
194 | return opts; | |
195 | } | |
196 | ||
197 | static void _dev_close(struct thread_data *td, struct xnvme_fioe_fwrap *fwrap) | |
198 | { | |
199 | if (fwrap->dev) | |
200 | xnvme_queue_term(fwrap->queue); | |
201 | ||
202 | xnvme_dev_close(fwrap->dev); | |
203 | ||
204 | memset(fwrap, 0, sizeof(*fwrap)); | |
205 | } | |
206 | ||
207 | static void xnvme_fioe_cleanup(struct thread_data *td) | |
208 | { | |
fdac9c68 | 209 | struct xnvme_fioe_data *xd = NULL; |
a3ff873e AK |
210 | int err; |
211 | ||
fdac9c68 AK |
212 | if (!td->io_ops_data) |
213 | return; | |
214 | ||
215 | xd = td->io_ops_data; | |
216 | ||
a3ff873e AK |
217 | err = pthread_mutex_lock(&g_serialize); |
218 | if (err) | |
219 | log_err("ioeng->cleanup(): pthread_mutex_lock(), err(%d)\n", err); | |
220 | /* NOTE: not returning here */ | |
221 | ||
222 | for (uint64_t i = 0; i < xd->nallocated; ++i) | |
223 | _dev_close(td, &xd->files[i]); | |
224 | ||
225 | if (!err) { | |
226 | err = pthread_mutex_unlock(&g_serialize); | |
227 | if (err) | |
228 | log_err("ioeng->cleanup(): pthread_mutex_unlock(), err(%d)\n", err); | |
229 | } | |
230 | ||
231 | free(xd->iocq); | |
232 | free(xd->iovec); | |
233 | free(xd); | |
234 | td->io_ops_data = NULL; | |
235 | } | |
236 | ||
237 | /** | |
238 | * Helper function setting up device handles as addressed by the naming | |
239 | * convention of the given `fio_file` filename. | |
240 | * | |
241 | * Checks thread-options for explicit control of asynchronous implementation via | |
242 | * the ``--xnvme_async={thrpool,emu,posix,io_uring,libaio,nil}``. | |
243 | */ | |
244 | static int _dev_open(struct thread_data *td, struct fio_file *f) | |
245 | { | |
246 | struct xnvme_opts opts = xnvme_opts_from_fioe(td); | |
247 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
248 | struct xnvme_fioe_fwrap *fwrap; | |
249 | int flags = 0; | |
250 | int err; | |
251 | ||
252 | if (f->fileno > (int)xd->nallocated) { | |
253 | log_err("ioeng->_dev_open(%s): invalid assumption\n", f->file_name); | |
254 | return 1; | |
255 | } | |
256 | ||
257 | fwrap = &xd->files[f->fileno]; | |
258 | ||
259 | err = pthread_mutex_lock(&g_serialize); | |
260 | if (err) { | |
261 | log_err("ioeng->_dev_open(%s): pthread_mutex_lock(), err(%d)\n", f->file_name, | |
262 | err); | |
263 | return -err; | |
264 | } | |
265 | ||
266 | fwrap->dev = xnvme_dev_open(f->file_name, &opts); | |
267 | if (!fwrap->dev) { | |
268 | log_err("ioeng->_dev_open(%s): xnvme_dev_open(), err(%d)\n", f->file_name, errno); | |
269 | goto failure; | |
270 | } | |
271 | fwrap->geo = xnvme_dev_get_geo(fwrap->dev); | |
272 | ||
273 | if (xnvme_queue_init(fwrap->dev, td->o.iodepth, flags, &(fwrap->queue))) { | |
274 | log_err("ioeng->_dev_open(%s): xnvme_queue_init(), err(?)\n", f->file_name); | |
275 | goto failure; | |
276 | } | |
277 | xnvme_queue_set_cb(fwrap->queue, cb_pool, NULL); | |
278 | ||
279 | fwrap->ssw = xnvme_dev_get_ssw(fwrap->dev); | |
280 | fwrap->lba_nbytes = fwrap->geo->lba_nbytes; | |
281 | ||
282 | fwrap->fio_file = f; | |
283 | fwrap->fio_file->filetype = FIO_TYPE_BLOCK; | |
284 | fwrap->fio_file->real_file_size = fwrap->geo->tbytes; | |
285 | fio_file_set_size_known(fwrap->fio_file); | |
286 | ||
287 | err = pthread_mutex_unlock(&g_serialize); | |
288 | if (err) | |
289 | log_err("ioeng->_dev_open(%s): pthread_mutex_unlock(), err(%d)\n", f->file_name, | |
290 | err); | |
291 | ||
292 | return 0; | |
293 | ||
294 | failure: | |
295 | xnvme_queue_term(fwrap->queue); | |
296 | xnvme_dev_close(fwrap->dev); | |
297 | ||
298 | err = pthread_mutex_unlock(&g_serialize); | |
299 | if (err) | |
300 | log_err("ioeng->_dev_open(%s): pthread_mutex_unlock(), err(%d)\n", f->file_name, | |
301 | err); | |
302 | ||
303 | return 1; | |
304 | } | |
305 | ||
306 | static int xnvme_fioe_init(struct thread_data *td) | |
307 | { | |
308 | struct xnvme_fioe_data *xd = NULL; | |
309 | struct fio_file *f; | |
310 | unsigned int i; | |
311 | ||
312 | if (!td->o.use_thread) { | |
313 | log_err("ioeng->init(): --thread=1 is required\n"); | |
314 | return 1; | |
315 | } | |
316 | ||
317 | /* Allocate xd and iocq */ | |
318 | xd = calloc(1, sizeof(*xd) + sizeof(*xd->files) * td->o.nr_files); | |
319 | if (!xd) { | |
320 | log_err("ioeng->init(): !calloc(), err(%d)\n", errno); | |
321 | return 1; | |
322 | } | |
323 | ||
324 | xd->iocq = calloc(td->o.iodepth, sizeof(struct io_u *)); | |
325 | if (!xd->iocq) { | |
eb3570b5 AK |
326 | free(xd); |
327 | log_err("ioeng->init(): !calloc(xd->iocq), err(%d)\n", errno); | |
a3ff873e AK |
328 | return 1; |
329 | } | |
330 | ||
331 | xd->iovec = calloc(td->o.iodepth, sizeof(*xd->iovec)); | |
332 | if (!xd->iovec) { | |
eb3570b5 AK |
333 | free(xd->iocq); |
334 | free(xd); | |
a3ff873e AK |
335 | log_err("ioeng->init(): !calloc(xd->iovec), err(%d)\n", errno); |
336 | return 1; | |
337 | } | |
338 | ||
339 | xd->prev = -1; | |
340 | td->io_ops_data = xd; | |
341 | ||
342 | for_each_file(td, f, i) | |
343 | { | |
344 | if (_dev_open(td, f)) { | |
eb3570b5 AK |
345 | /* |
346 | * Note: We are not freeing xd, iocq and iovec. This | |
347 | * will be done as part of cleanup routine. | |
348 | */ | |
a3ff873e AK |
349 | log_err("ioeng->init(): failed; _dev_open(%s)\n", f->file_name); |
350 | return 1; | |
351 | } | |
352 | ||
353 | ++(xd->nallocated); | |
354 | } | |
355 | ||
356 | if (xd->nallocated != td->o.nr_files) { | |
357 | log_err("ioeng->init(): failed; nallocated != td->o.nr_files\n"); | |
358 | return 1; | |
359 | } | |
360 | ||
361 | return 0; | |
362 | } | |
363 | ||
364 | /* NOTE: using the first device for buffer-allocators) */ | |
365 | static int xnvme_fioe_iomem_alloc(struct thread_data *td, size_t total_mem) | |
366 | { | |
367 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
368 | struct xnvme_fioe_fwrap *fwrap = &xd->files[0]; | |
369 | ||
370 | if (!fwrap->dev) { | |
371 | log_err("ioeng->iomem_alloc(): failed; no dev-handle\n"); | |
372 | return 1; | |
373 | } | |
374 | ||
375 | td->orig_buffer = xnvme_buf_alloc(fwrap->dev, total_mem); | |
376 | ||
377 | return td->orig_buffer == NULL; | |
378 | } | |
379 | ||
380 | /* NOTE: using the first device for buffer-allocators) */ | |
381 | static void xnvme_fioe_iomem_free(struct thread_data *td) | |
382 | { | |
fdac9c68 AK |
383 | struct xnvme_fioe_data *xd = NULL; |
384 | struct xnvme_fioe_fwrap *fwrap = NULL; | |
385 | ||
386 | if (!td->io_ops_data) | |
387 | return; | |
388 | ||
389 | xd = td->io_ops_data; | |
390 | fwrap = &xd->files[0]; | |
a3ff873e AK |
391 | |
392 | if (!fwrap->dev) { | |
393 | log_err("ioeng->iomem_free(): failed no dev-handle\n"); | |
394 | return; | |
395 | } | |
396 | ||
397 | xnvme_buf_free(fwrap->dev, td->orig_buffer); | |
398 | } | |
399 | ||
400 | static int xnvme_fioe_io_u_init(struct thread_data *td, struct io_u *io_u) | |
401 | { | |
402 | io_u->mmap_data = td->io_ops_data; | |
403 | ||
404 | return 0; | |
405 | } | |
406 | ||
407 | static void xnvme_fioe_io_u_free(struct thread_data *td, struct io_u *io_u) | |
408 | { | |
409 | io_u->mmap_data = NULL; | |
410 | } | |
411 | ||
412 | static struct io_u *xnvme_fioe_event(struct thread_data *td, int event) | |
413 | { | |
414 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
415 | ||
416 | assert(event >= 0); | |
417 | assert((unsigned)event < xd->completed); | |
418 | ||
419 | return xd->iocq[event]; | |
420 | } | |
421 | ||
422 | static int xnvme_fioe_getevents(struct thread_data *td, unsigned int min, unsigned int max, | |
423 | const struct timespec *t) | |
424 | { | |
425 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
426 | struct xnvme_fioe_fwrap *fwrap = NULL; | |
427 | int nfiles = xd->nallocated; | |
428 | int err = 0; | |
429 | ||
430 | if (xd->prev != -1 && ++xd->prev < nfiles) { | |
431 | fwrap = &xd->files[xd->prev]; | |
432 | xd->cur = xd->prev; | |
433 | } | |
434 | ||
435 | xd->completed = 0; | |
436 | for (;;) { | |
437 | if (fwrap == NULL || xd->cur == nfiles) { | |
438 | fwrap = &xd->files[0]; | |
439 | xd->cur = 0; | |
440 | } | |
441 | ||
442 | while (fwrap != NULL && xd->cur < nfiles && err >= 0) { | |
443 | err = xnvme_queue_poke(fwrap->queue, max - xd->completed); | |
444 | if (err < 0) { | |
445 | switch (err) { | |
446 | case -EBUSY: | |
447 | case -EAGAIN: | |
448 | usleep(1); | |
449 | break; | |
450 | ||
451 | default: | |
452 | log_err("ioeng->getevents(): unhandled IO error\n"); | |
453 | assert(false); | |
454 | return 0; | |
455 | } | |
456 | } | |
457 | if (xd->completed >= min) { | |
458 | xd->prev = xd->cur; | |
459 | return xd->completed; | |
460 | } | |
461 | xd->cur++; | |
462 | fwrap = &xd->files[xd->cur]; | |
463 | ||
464 | if (err < 0) { | |
465 | switch (err) { | |
466 | case -EBUSY: | |
467 | case -EAGAIN: | |
468 | usleep(1); | |
469 | break; | |
470 | } | |
471 | } | |
472 | } | |
473 | } | |
474 | ||
475 | xd->cur = 0; | |
476 | ||
477 | return xd->completed; | |
478 | } | |
479 | ||
480 | static enum fio_q_status xnvme_fioe_queue(struct thread_data *td, struct io_u *io_u) | |
481 | { | |
482 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
483 | struct xnvme_fioe_fwrap *fwrap; | |
484 | struct xnvme_cmd_ctx *ctx; | |
485 | uint32_t nsid; | |
486 | uint64_t slba; | |
487 | uint16_t nlb; | |
488 | int err; | |
489 | bool vectored_io = ((struct xnvme_fioe_options *)td->eo)->xnvme_iovec; | |
490 | ||
491 | fio_ro_check(td, io_u); | |
492 | ||
493 | fwrap = &xd->files[io_u->file->fileno]; | |
494 | nsid = xnvme_dev_get_nsid(fwrap->dev); | |
495 | ||
496 | slba = io_u->offset >> fwrap->ssw; | |
497 | nlb = (io_u->xfer_buflen >> fwrap->ssw) - 1; | |
498 | ||
499 | ctx = xnvme_queue_get_cmd_ctx(fwrap->queue); | |
500 | ctx->async.cb_arg = io_u; | |
501 | ||
502 | ctx->cmd.common.nsid = nsid; | |
503 | ctx->cmd.nvm.slba = slba; | |
504 | ctx->cmd.nvm.nlb = nlb; | |
505 | ||
506 | switch (io_u->ddir) { | |
507 | case DDIR_READ: | |
508 | ctx->cmd.common.opcode = XNVME_SPEC_NVM_OPC_READ; | |
509 | break; | |
510 | ||
511 | case DDIR_WRITE: | |
512 | ctx->cmd.common.opcode = XNVME_SPEC_NVM_OPC_WRITE; | |
513 | break; | |
514 | ||
515 | default: | |
516 | log_err("ioeng->queue(): ENOSYS: %u\n", io_u->ddir); | |
eb3570b5 AK |
517 | xnvme_queue_put_cmd_ctx(ctx->async.queue, ctx); |
518 | ||
519 | io_u->error = ENOSYS; | |
a3ff873e | 520 | assert(false); |
eb3570b5 | 521 | return FIO_Q_COMPLETED; |
a3ff873e AK |
522 | } |
523 | ||
524 | if (vectored_io) { | |
525 | xd->iovec[io_u->index].iov_base = io_u->xfer_buf; | |
526 | xd->iovec[io_u->index].iov_len = io_u->xfer_buflen; | |
527 | ||
528 | err = xnvme_cmd_passv(ctx, &xd->iovec[io_u->index], 1, io_u->xfer_buflen, NULL, 0, | |
529 | 0); | |
530 | } else { | |
531 | err = xnvme_cmd_pass(ctx, io_u->xfer_buf, io_u->xfer_buflen, NULL, 0); | |
532 | } | |
533 | switch (err) { | |
534 | case 0: | |
535 | return FIO_Q_QUEUED; | |
536 | ||
537 | case -EBUSY: | |
538 | case -EAGAIN: | |
539 | xnvme_queue_put_cmd_ctx(ctx->async.queue, ctx); | |
540 | return FIO_Q_BUSY; | |
541 | ||
542 | default: | |
543 | log_err("ioeng->queue(): err: '%d'\n", err); | |
544 | ||
545 | xnvme_queue_put_cmd_ctx(ctx->async.queue, ctx); | |
546 | ||
547 | io_u->error = abs(err); | |
548 | assert(false); | |
549 | return FIO_Q_COMPLETED; | |
550 | } | |
551 | } | |
552 | ||
553 | static int xnvme_fioe_close(struct thread_data *td, struct fio_file *f) | |
554 | { | |
555 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
556 | ||
557 | dprint(FD_FILE, "xnvme close %s -- nopen: %ld\n", f->file_name, xd->nopen); | |
558 | ||
559 | --(xd->nopen); | |
560 | ||
561 | return 0; | |
562 | } | |
563 | ||
564 | static int xnvme_fioe_open(struct thread_data *td, struct fio_file *f) | |
565 | { | |
566 | struct xnvme_fioe_data *xd = td->io_ops_data; | |
567 | ||
568 | dprint(FD_FILE, "xnvme open %s -- nopen: %ld\n", f->file_name, xd->nopen); | |
569 | ||
570 | if (f->fileno > (int)xd->nallocated) { | |
571 | log_err("ioeng->open(): f->fileno > xd->nallocated; invalid assumption\n"); | |
572 | return 1; | |
573 | } | |
574 | if (xd->files[f->fileno].fio_file != f) { | |
575 | log_err("ioeng->open(): fio_file != f; invalid assumption\n"); | |
576 | return 1; | |
577 | } | |
578 | ||
579 | ++(xd->nopen); | |
580 | ||
581 | return 0; | |
582 | } | |
583 | ||
584 | static int xnvme_fioe_invalidate(struct thread_data *td, struct fio_file *f) | |
585 | { | |
586 | /* Consider only doing this with be:spdk */ | |
587 | return 0; | |
588 | } | |
589 | ||
590 | static int xnvme_fioe_get_max_open_zones(struct thread_data *td, struct fio_file *f, | |
591 | unsigned int *max_open_zones) | |
592 | { | |
593 | struct xnvme_opts opts = xnvme_opts_from_fioe(td); | |
594 | struct xnvme_dev *dev; | |
595 | const struct xnvme_spec_znd_idfy_ns *zns; | |
596 | int err = 0, err_lock; | |
597 | ||
598 | if (f->filetype != FIO_TYPE_FILE && f->filetype != FIO_TYPE_BLOCK && | |
599 | f->filetype != FIO_TYPE_CHAR) { | |
600 | log_info("ioeng->get_max_open_zoned(): ignoring filetype: %d\n", f->filetype); | |
601 | return 0; | |
602 | } | |
603 | err_lock = pthread_mutex_lock(&g_serialize); | |
604 | if (err_lock) { | |
605 | log_err("ioeng->get_max_open_zones(): pthread_mutex_lock(), err(%d)\n", err_lock); | |
606 | return -err_lock; | |
607 | } | |
608 | ||
609 | dev = xnvme_dev_open(f->file_name, &opts); | |
610 | if (!dev) { | |
611 | log_err("ioeng->get_max_open_zones(): xnvme_dev_open(), err(%d)\n", err_lock); | |
612 | err = -errno; | |
613 | goto exit; | |
614 | } | |
615 | if (xnvme_dev_get_geo(dev)->type != XNVME_GEO_ZONED) { | |
616 | errno = EINVAL; | |
617 | err = -errno; | |
618 | goto exit; | |
619 | } | |
620 | ||
621 | zns = (void *)xnvme_dev_get_ns_css(dev); | |
622 | if (!zns) { | |
623 | log_err("ioeng->get_max_open_zones(): xnvme_dev_get_ns_css(), err(%d)\n", errno); | |
624 | err = -errno; | |
625 | goto exit; | |
626 | } | |
627 | ||
628 | /* | |
629 | * intentional overflow as the value is zero-based and NVMe | |
630 | * defines 0xFFFFFFFF as unlimited thus overflowing to 0 which | |
631 | * is how fio indicates unlimited and otherwise just converting | |
632 | * to one-based. | |
633 | */ | |
634 | *max_open_zones = zns->mor + 1; | |
635 | ||
636 | exit: | |
637 | xnvme_dev_close(dev); | |
638 | err_lock = pthread_mutex_unlock(&g_serialize); | |
639 | if (err_lock) | |
640 | log_err("ioeng->get_max_open_zones(): pthread_mutex_unlock(), err(%d)\n", | |
641 | err_lock); | |
642 | ||
643 | return err; | |
644 | } | |
645 | ||
646 | /** | |
647 | * Currently, this function is called before of I/O engine initialization, so, | |
648 | * we cannot consult the file-wrapping done when 'fioe' initializes. | |
649 | * Instead we just open based on the given filename. | |
650 | * | |
651 | * TODO: unify the different setup methods, consider keeping the handle around, | |
652 | * and consider how to support the --be option in this usecase | |
653 | */ | |
654 | static int xnvme_fioe_get_zoned_model(struct thread_data *td, struct fio_file *f, | |
655 | enum zbd_zoned_model *model) | |
656 | { | |
657 | struct xnvme_opts opts = xnvme_opts_from_fioe(td); | |
658 | struct xnvme_dev *dev; | |
659 | int err = 0, err_lock; | |
660 | ||
661 | if (f->filetype != FIO_TYPE_FILE && f->filetype != FIO_TYPE_BLOCK && | |
662 | f->filetype != FIO_TYPE_CHAR) { | |
663 | log_info("ioeng->get_zoned_model(): ignoring filetype: %d\n", f->filetype); | |
664 | return -EINVAL; | |
665 | } | |
666 | ||
667 | err = pthread_mutex_lock(&g_serialize); | |
668 | if (err) { | |
669 | log_err("ioeng->get_zoned_model(): pthread_mutex_lock(), err(%d)\n", err); | |
670 | return -err; | |
671 | } | |
672 | ||
673 | dev = xnvme_dev_open(f->file_name, &opts); | |
674 | if (!dev) { | |
675 | log_err("ioeng->get_zoned_model(): xnvme_dev_open(%s) failed, errno: %d\n", | |
676 | f->file_name, errno); | |
677 | err = -errno; | |
678 | goto exit; | |
679 | } | |
680 | ||
681 | switch (xnvme_dev_get_geo(dev)->type) { | |
682 | case XNVME_GEO_UNKNOWN: | |
683 | dprint(FD_ZBD, "%s: got 'unknown', assigning ZBD_NONE\n", f->file_name); | |
684 | *model = ZBD_NONE; | |
685 | break; | |
686 | ||
687 | case XNVME_GEO_CONVENTIONAL: | |
688 | dprint(FD_ZBD, "%s: got 'conventional', assigning ZBD_NONE\n", f->file_name); | |
689 | *model = ZBD_NONE; | |
690 | break; | |
691 | ||
692 | case XNVME_GEO_ZONED: | |
693 | dprint(FD_ZBD, "%s: got 'zoned', assigning ZBD_HOST_MANAGED\n", f->file_name); | |
694 | *model = ZBD_HOST_MANAGED; | |
695 | break; | |
696 | ||
697 | default: | |
698 | dprint(FD_ZBD, "%s: hit-default, assigning ZBD_NONE\n", f->file_name); | |
699 | *model = ZBD_NONE; | |
700 | errno = EINVAL; | |
701 | err = -errno; | |
702 | break; | |
703 | } | |
704 | ||
705 | exit: | |
706 | xnvme_dev_close(dev); | |
707 | ||
708 | err_lock = pthread_mutex_unlock(&g_serialize); | |
709 | if (err_lock) | |
710 | log_err("ioeng->get_zoned_model(): pthread_mutex_unlock(), err(%d)\n", err_lock); | |
711 | ||
712 | return err; | |
713 | } | |
714 | ||
715 | /** | |
716 | * Fills the given ``zbdz`` with at most ``nr_zones`` zone-descriptors. | |
717 | * | |
718 | * The implementation converts the NVMe Zoned Command Set log-pages for Zone | |
719 | * descriptors into the Linux Kernel Zoned Block Report format. | |
720 | * | |
721 | * NOTE: This function is called before I/O engine initialization, that is, | |
722 | * before ``_dev_open`` has been called and file-wrapping is setup. Thus is has | |
723 | * to do the ``_dev_open`` itself, and shut it down again once it is done | |
724 | * retrieving the log-pages and converting them to the report format. | |
725 | * | |
726 | * TODO: unify the different setup methods, consider keeping the handle around, | |
727 | * and consider how to support the --async option in this usecase | |
728 | */ | |
729 | static int xnvme_fioe_report_zones(struct thread_data *td, struct fio_file *f, uint64_t offset, | |
730 | struct zbd_zone *zbdz, unsigned int nr_zones) | |
731 | { | |
732 | struct xnvme_opts opts = xnvme_opts_from_fioe(td); | |
733 | const struct xnvme_spec_znd_idfy_lbafe *lbafe = NULL; | |
734 | struct xnvme_dev *dev = NULL; | |
735 | const struct xnvme_geo *geo = NULL; | |
736 | struct xnvme_znd_report *rprt = NULL; | |
737 | uint32_t ssw; | |
738 | uint64_t slba; | |
739 | unsigned int limit = 0; | |
740 | int err = 0, err_lock; | |
741 | ||
742 | dprint(FD_ZBD, "%s: report_zones() offset: %zu, nr_zones: %u\n", f->file_name, offset, | |
743 | nr_zones); | |
744 | ||
745 | err = pthread_mutex_lock(&g_serialize); | |
746 | if (err) { | |
747 | log_err("ioeng->report_zones(%s): pthread_mutex_lock(), err(%d)\n", f->file_name, | |
748 | err); | |
749 | return -err; | |
750 | } | |
751 | ||
752 | dev = xnvme_dev_open(f->file_name, &opts); | |
753 | if (!dev) { | |
754 | log_err("ioeng->report_zones(%s): xnvme_dev_open(), err(%d)\n", f->file_name, | |
755 | errno); | |
756 | goto exit; | |
757 | } | |
758 | ||
759 | geo = xnvme_dev_get_geo(dev); | |
760 | ssw = xnvme_dev_get_ssw(dev); | |
761 | lbafe = xnvme_znd_dev_get_lbafe(dev); | |
762 | ||
763 | limit = nr_zones > geo->nzone ? geo->nzone : nr_zones; | |
764 | ||
765 | dprint(FD_ZBD, "%s: limit: %u\n", f->file_name, limit); | |
766 | ||
767 | slba = ((offset >> ssw) / geo->nsect) * geo->nsect; | |
768 | ||
769 | rprt = xnvme_znd_report_from_dev(dev, slba, limit, 0); | |
770 | if (!rprt) { | |
771 | log_err("ioeng->report_zones(%s): xnvme_znd_report_from_dev(), err(%d)\n", | |
772 | f->file_name, errno); | |
773 | err = -errno; | |
774 | goto exit; | |
775 | } | |
776 | if (rprt->nentries != limit) { | |
777 | log_err("ioeng->report_zones(%s): nentries != nr_zones\n", f->file_name); | |
778 | err = 1; | |
779 | goto exit; | |
780 | } | |
781 | if (offset > geo->tbytes) { | |
782 | log_err("ioeng->report_zones(%s): out-of-bounds\n", f->file_name); | |
783 | goto exit; | |
784 | } | |
785 | ||
786 | /* Transform the zone-report */ | |
787 | for (uint32_t idx = 0; idx < rprt->nentries; ++idx) { | |
788 | struct xnvme_spec_znd_descr *descr = XNVME_ZND_REPORT_DESCR(rprt, idx); | |
789 | ||
790 | zbdz[idx].start = descr->zslba << ssw; | |
791 | zbdz[idx].len = lbafe->zsze << ssw; | |
792 | zbdz[idx].capacity = descr->zcap << ssw; | |
793 | zbdz[idx].wp = descr->wp << ssw; | |
794 | ||
795 | switch (descr->zt) { | |
796 | case XNVME_SPEC_ZND_TYPE_SEQWR: | |
797 | zbdz[idx].type = ZBD_ZONE_TYPE_SWR; | |
798 | break; | |
799 | ||
800 | default: | |
801 | log_err("ioeng->report_zones(%s): invalid type for zone at offset(%zu)\n", | |
802 | f->file_name, zbdz[idx].start); | |
803 | err = -EIO; | |
804 | goto exit; | |
805 | } | |
806 | ||
807 | switch (descr->zs) { | |
808 | case XNVME_SPEC_ZND_STATE_EMPTY: | |
809 | zbdz[idx].cond = ZBD_ZONE_COND_EMPTY; | |
810 | break; | |
811 | case XNVME_SPEC_ZND_STATE_IOPEN: | |
812 | zbdz[idx].cond = ZBD_ZONE_COND_IMP_OPEN; | |
813 | break; | |
814 | case XNVME_SPEC_ZND_STATE_EOPEN: | |
815 | zbdz[idx].cond = ZBD_ZONE_COND_EXP_OPEN; | |
816 | break; | |
817 | case XNVME_SPEC_ZND_STATE_CLOSED: | |
818 | zbdz[idx].cond = ZBD_ZONE_COND_CLOSED; | |
819 | break; | |
820 | case XNVME_SPEC_ZND_STATE_FULL: | |
821 | zbdz[idx].cond = ZBD_ZONE_COND_FULL; | |
822 | break; | |
823 | ||
824 | case XNVME_SPEC_ZND_STATE_RONLY: | |
825 | case XNVME_SPEC_ZND_STATE_OFFLINE: | |
826 | default: | |
827 | zbdz[idx].cond = ZBD_ZONE_COND_OFFLINE; | |
828 | break; | |
829 | } | |
830 | } | |
831 | ||
832 | exit: | |
833 | xnvme_buf_virt_free(rprt); | |
834 | ||
835 | xnvme_dev_close(dev); | |
836 | ||
837 | err_lock = pthread_mutex_unlock(&g_serialize); | |
838 | if (err_lock) | |
839 | log_err("ioeng->report_zones(): pthread_mutex_unlock(), err: %d\n", err_lock); | |
840 | ||
841 | dprint(FD_ZBD, "err: %d, nr_zones: %d\n", err, (int)nr_zones); | |
842 | ||
843 | return err ? err : (int)limit; | |
844 | } | |
845 | ||
846 | /** | |
847 | * NOTE: This function may get called before I/O engine initialization, that is, | |
848 | * before ``_dev_open`` has been called and file-wrapping is setup. In such | |
849 | * case it has to do ``_dev_open`` itself, and shut it down again once it is | |
850 | * done resetting write pointer of zones. | |
851 | */ | |
852 | static int xnvme_fioe_reset_wp(struct thread_data *td, struct fio_file *f, uint64_t offset, | |
853 | uint64_t length) | |
854 | { | |
855 | struct xnvme_opts opts = xnvme_opts_from_fioe(td); | |
856 | struct xnvme_fioe_data *xd = NULL; | |
857 | struct xnvme_fioe_fwrap *fwrap = NULL; | |
858 | struct xnvme_dev *dev = NULL; | |
859 | const struct xnvme_geo *geo = NULL; | |
860 | uint64_t first, last; | |
861 | uint32_t ssw; | |
862 | uint32_t nsid; | |
863 | int err = 0, err_lock; | |
864 | ||
865 | if (td->io_ops_data) { | |
866 | xd = td->io_ops_data; | |
867 | fwrap = &xd->files[f->fileno]; | |
868 | ||
869 | assert(fwrap->dev); | |
870 | assert(fwrap->geo); | |
871 | ||
872 | dev = fwrap->dev; | |
873 | geo = fwrap->geo; | |
874 | ssw = fwrap->ssw; | |
875 | } else { | |
876 | err = pthread_mutex_lock(&g_serialize); | |
877 | if (err) { | |
878 | log_err("ioeng->reset_wp(): pthread_mutex_lock(), err(%d)\n", err); | |
879 | return -err; | |
880 | } | |
881 | ||
882 | dev = xnvme_dev_open(f->file_name, &opts); | |
883 | if (!dev) { | |
884 | log_err("ioeng->reset_wp(): xnvme_dev_open(%s) failed, errno(%d)\n", | |
885 | f->file_name, errno); | |
886 | goto exit; | |
887 | } | |
888 | geo = xnvme_dev_get_geo(dev); | |
889 | ssw = xnvme_dev_get_ssw(dev); | |
890 | } | |
891 | ||
892 | nsid = xnvme_dev_get_nsid(dev); | |
893 | ||
894 | first = ((offset >> ssw) / geo->nsect) * geo->nsect; | |
895 | last = (((offset + length) >> ssw) / geo->nsect) * geo->nsect; | |
896 | dprint(FD_ZBD, "first: 0x%lx, last: 0x%lx\n", first, last); | |
897 | ||
898 | for (uint64_t zslba = first; zslba < last; zslba += geo->nsect) { | |
899 | struct xnvme_cmd_ctx ctx = xnvme_cmd_ctx_from_dev(dev); | |
900 | ||
901 | if (zslba >= (geo->nsect * geo->nzone)) { | |
902 | log_err("ioeng->reset_wp(): out-of-bounds\n"); | |
903 | err = 0; | |
904 | break; | |
905 | } | |
906 | ||
907 | err = xnvme_znd_mgmt_send(&ctx, nsid, zslba, false, | |
908 | XNVME_SPEC_ZND_CMD_MGMT_SEND_RESET, 0x0, NULL); | |
909 | if (err || xnvme_cmd_ctx_cpl_status(&ctx)) { | |
910 | err = err ? err : -EIO; | |
911 | log_err("ioeng->reset_wp(): err(%d), sc(%d)", err, ctx.cpl.status.sc); | |
912 | goto exit; | |
913 | } | |
914 | } | |
915 | ||
916 | exit: | |
917 | if (!td->io_ops_data) { | |
918 | xnvme_dev_close(dev); | |
919 | ||
920 | err_lock = pthread_mutex_unlock(&g_serialize); | |
921 | if (err_lock) | |
922 | log_err("ioeng->reset_wp(): pthread_mutex_unlock(), err(%d)\n", err_lock); | |
923 | } | |
924 | ||
925 | return err; | |
926 | } | |
927 | ||
928 | static int xnvme_fioe_get_file_size(struct thread_data *td, struct fio_file *f) | |
929 | { | |
930 | struct xnvme_opts opts = xnvme_opts_from_fioe(td); | |
931 | struct xnvme_dev *dev; | |
932 | int ret = 0, err; | |
933 | ||
934 | if (fio_file_size_known(f)) | |
935 | return 0; | |
936 | ||
937 | ret = pthread_mutex_lock(&g_serialize); | |
938 | if (ret) { | |
939 | log_err("ioeng->reset_wp(): pthread_mutex_lock(), err(%d)\n", ret); | |
940 | return -ret; | |
941 | } | |
942 | ||
943 | dev = xnvme_dev_open(f->file_name, &opts); | |
944 | if (!dev) { | |
945 | log_err("%s: failed retrieving device handle, errno: %d\n", f->file_name, errno); | |
946 | ret = -errno; | |
947 | goto exit; | |
948 | } | |
949 | ||
950 | f->real_file_size = xnvme_dev_get_geo(dev)->tbytes; | |
951 | fio_file_set_size_known(f); | |
952 | f->filetype = FIO_TYPE_BLOCK; | |
953 | ||
954 | exit: | |
955 | xnvme_dev_close(dev); | |
956 | err = pthread_mutex_unlock(&g_serialize); | |
957 | if (err) | |
958 | log_err("ioeng->reset_wp(): pthread_mutex_unlock(), err(%d)\n", err); | |
959 | ||
960 | return ret; | |
961 | } | |
962 | ||
963 | FIO_STATIC struct ioengine_ops ioengine = { | |
964 | .name = "xnvme", | |
965 | .version = FIO_IOOPS_VERSION, | |
966 | .options = options, | |
967 | .option_struct_size = sizeof(struct xnvme_fioe_options), | |
968 | .flags = FIO_DISKLESSIO | FIO_NODISKUTIL | FIO_NOEXTEND | FIO_MEMALIGN | FIO_RAWIO, | |
969 | ||
970 | .cleanup = xnvme_fioe_cleanup, | |
971 | .init = xnvme_fioe_init, | |
972 | ||
973 | .iomem_free = xnvme_fioe_iomem_free, | |
974 | .iomem_alloc = xnvme_fioe_iomem_alloc, | |
975 | ||
976 | .io_u_free = xnvme_fioe_io_u_free, | |
977 | .io_u_init = xnvme_fioe_io_u_init, | |
978 | ||
979 | .event = xnvme_fioe_event, | |
980 | .getevents = xnvme_fioe_getevents, | |
981 | .queue = xnvme_fioe_queue, | |
982 | ||
983 | .close_file = xnvme_fioe_close, | |
984 | .open_file = xnvme_fioe_open, | |
985 | .get_file_size = xnvme_fioe_get_file_size, | |
986 | ||
987 | .invalidate = xnvme_fioe_invalidate, | |
988 | .get_max_open_zones = xnvme_fioe_get_max_open_zones, | |
989 | .get_zoned_model = xnvme_fioe_get_zoned_model, | |
990 | .report_zones = xnvme_fioe_report_zones, | |
991 | .reset_wp = xnvme_fioe_reset_wp, | |
992 | }; | |
993 | ||
994 | static void fio_init fio_xnvme_register(void) | |
995 | { | |
996 | register_ioengine(&ioengine); | |
997 | } | |
998 | ||
999 | static void fio_exit fio_xnvme_unregister(void) | |
1000 | { | |
1001 | unregister_ioengine(&ioengine); | |
1002 | } |