t/nvmept_trim: increase transfer size for some tests
[fio.git] / engines / sg.c
... / ...
CommitLineData
1/*
2 * sg engine
3 *
4 * IO engine that uses the Linux SG v3 interface to talk to SCSI devices
5 *
6 * This ioengine can operate in two modes:
7 * sync with block devices (/dev/sdX) or
8 * with character devices (/dev/sgY) with direct=1 or sync=1
9 * async with character devices with direct=0 and sync=0
10 *
11 * What value does queue() return for the different cases?
12 * queue() return value
13 * In sync mode:
14 * /dev/sdX RWT FIO_Q_COMPLETED
15 * /dev/sgY RWT FIO_Q_COMPLETED
16 * with direct=1 or sync=1
17 *
18 * In async mode:
19 * /dev/sgY RWT FIO_Q_QUEUED
20 * direct=0 and sync=0
21 *
22 * Because FIO_SYNCIO is set for this ioengine td_io_queue() will fill in
23 * issue_time *before* each IO is sent to queue()
24 *
25 * Where are the IO counting functions called for the different cases?
26 *
27 * In sync mode:
28 * /dev/sdX (commit==NULL)
29 * RWT
30 * io_u_mark_depth() called in td_io_queue()
31 * io_u_mark_submit/complete() called in td_io_queue()
32 * issue_time set in td_io_queue()
33 *
34 * /dev/sgY with direct=1 or sync=1 (commit does nothing)
35 * RWT
36 * io_u_mark_depth() called in td_io_queue()
37 * io_u_mark_submit/complete() called in queue()
38 * issue_time set in td_io_queue()
39 *
40 * In async mode:
41 * /dev/sgY with direct=0 and sync=0
42 * RW: read and write operations are submitted in queue()
43 * io_u_mark_depth() called in td_io_commit()
44 * io_u_mark_submit() called in queue()
45 * issue_time set in td_io_queue()
46 * T: trim operations are queued in queue() and submitted in commit()
47 * io_u_mark_depth() called in td_io_commit()
48 * io_u_mark_submit() called in commit()
49 * issue_time set in commit()
50 *
51 */
52#include <stdio.h>
53#include <stdlib.h>
54#include <unistd.h>
55#include <errno.h>
56#include <poll.h>
57
58#include "../fio.h"
59#include "../optgroup.h"
60
61#ifdef FIO_HAVE_SGIO
62
63#ifndef SGV4_FLAG_HIPRI
64#define SGV4_FLAG_HIPRI 0x800
65#endif
66
67enum {
68 FIO_SG_WRITE = 1,
69 FIO_SG_WRITE_VERIFY,
70 FIO_SG_WRITE_SAME,
71 FIO_SG_WRITE_SAME_NDOB,
72 FIO_SG_WRITE_STREAM,
73 FIO_SG_VERIFY_BYTCHK_00,
74 FIO_SG_VERIFY_BYTCHK_01,
75 FIO_SG_VERIFY_BYTCHK_11,
76};
77
78struct sg_options {
79 void *pad;
80 unsigned int hipri;
81 unsigned int readfua;
82 unsigned int writefua;
83 unsigned int write_mode;
84 uint16_t stream_id;
85};
86
87static struct fio_option options[] = {
88 {
89 .name = "hipri",
90 .lname = "High Priority",
91 .type = FIO_OPT_STR_SET,
92 .off1 = offsetof(struct sg_options, hipri),
93 .help = "Use polled IO completions",
94 .category = FIO_OPT_C_ENGINE,
95 .group = FIO_OPT_G_SG,
96 },
97 {
98 .name = "readfua",
99 .lname = "sg engine read fua flag support",
100 .type = FIO_OPT_BOOL,
101 .off1 = offsetof(struct sg_options, readfua),
102 .help = "Set FUA flag (force unit access) for all Read operations",
103 .def = "0",
104 .category = FIO_OPT_C_ENGINE,
105 .group = FIO_OPT_G_SG,
106 },
107 {
108 .name = "writefua",
109 .lname = "sg engine write fua flag support",
110 .type = FIO_OPT_BOOL,
111 .off1 = offsetof(struct sg_options, writefua),
112 .help = "Set FUA flag (force unit access) for all Write operations",
113 .def = "0",
114 .category = FIO_OPT_C_ENGINE,
115 .group = FIO_OPT_G_SG,
116 },
117 {
118 .name = "sg_write_mode",
119 .lname = "specify sg write mode",
120 .type = FIO_OPT_STR,
121 .off1 = offsetof(struct sg_options, write_mode),
122 .help = "Specify SCSI WRITE mode",
123 .def = "write",
124 .posval = {
125 { .ival = "write",
126 .oval = FIO_SG_WRITE,
127 .help = "Issue standard SCSI WRITE commands",
128 },
129 { .ival = "write_and_verify",
130 .oval = FIO_SG_WRITE_VERIFY,
131 .help = "Issue SCSI WRITE AND VERIFY commands",
132 },
133 { .ival = "verify",
134 .oval = FIO_SG_WRITE_VERIFY,
135 .help = "Issue SCSI WRITE AND VERIFY commands. This "
136 "option is deprecated. Use write_and_verify instead.",
137 },
138 { .ival = "write_same",
139 .oval = FIO_SG_WRITE_SAME,
140 .help = "Issue SCSI WRITE SAME commands",
141 },
142 { .ival = "same",
143 .oval = FIO_SG_WRITE_SAME,
144 .help = "Issue SCSI WRITE SAME commands. This "
145 "option is deprecated. Use write_same instead.",
146 },
147 { .ival = "write_same_ndob",
148 .oval = FIO_SG_WRITE_SAME_NDOB,
149 .help = "Issue SCSI WRITE SAME(16) commands with NDOB flag set",
150 },
151 { .ival = "verify_bytchk_00",
152 .oval = FIO_SG_VERIFY_BYTCHK_00,
153 .help = "Issue SCSI VERIFY commands with BYTCHK set to 00",
154 },
155 { .ival = "verify_bytchk_01",
156 .oval = FIO_SG_VERIFY_BYTCHK_01,
157 .help = "Issue SCSI VERIFY commands with BYTCHK set to 01",
158 },
159 { .ival = "verify_bytchk_11",
160 .oval = FIO_SG_VERIFY_BYTCHK_11,
161 .help = "Issue SCSI VERIFY commands with BYTCHK set to 11",
162 },
163 { .ival = "write_stream",
164 .oval = FIO_SG_WRITE_STREAM,
165 .help = "Issue SCSI WRITE STREAM(16) commands",
166 },
167 },
168 .category = FIO_OPT_C_ENGINE,
169 .group = FIO_OPT_G_SG,
170 },
171 {
172 .name = "stream_id",
173 .lname = "stream id for WRITE STREAM(16) commands",
174 .type = FIO_OPT_INT,
175 .off1 = offsetof(struct sg_options, stream_id),
176 .help = "Stream ID for WRITE STREAM(16) commands",
177 .def = "0",
178 .category = FIO_OPT_C_ENGINE,
179 .group = FIO_OPT_G_SG,
180 },
181 {
182 .name = NULL,
183 },
184};
185
186#define MAX_10B_LBA 0xFFFFFFFFULL
187#define SCSI_TIMEOUT_MS 30000 // 30 second timeout; currently no method to override
188#define MAX_SB 64 // sense block maximum return size
189/*
190#define FIO_SGIO_DEBUG
191*/
192
193struct sgio_cmd {
194 unsigned char cdb[16]; // enhanced from 10 to support 16 byte commands
195 unsigned char sb[MAX_SB]; // add sense block to commands
196 int nr;
197};
198
199struct sgio_trim {
200 uint8_t *unmap_param;
201 unsigned int unmap_range_count;
202 struct io_u **trim_io_us;
203};
204
205struct sgio_data {
206 struct sgio_cmd *cmds;
207 struct io_u **events;
208 struct pollfd *pfds;
209 int *fd_flags;
210 void *sgbuf;
211 unsigned int bs;
212 int type_checked;
213 struct sgio_trim **trim_queues;
214 int current_queue;
215#ifdef FIO_SGIO_DEBUG
216 unsigned int *trim_queue_map;
217#endif
218};
219
220static inline uint16_t sgio_get_be16(uint8_t *buf)
221{
222 return be16_to_cpu(*((uint16_t *) buf));
223}
224
225static inline uint32_t sgio_get_be32(uint8_t *buf)
226{
227 return be32_to_cpu(*((uint32_t *) buf));
228}
229
230static inline uint64_t sgio_get_be64(uint8_t *buf)
231{
232 return be64_to_cpu(*((uint64_t *) buf));
233}
234
235static inline void sgio_set_be16(uint16_t val, uint8_t *buf)
236{
237 uint16_t t = cpu_to_be16(val);
238
239 memcpy(buf, &t, sizeof(uint16_t));
240}
241
242static inline void sgio_set_be32(uint32_t val, uint8_t *buf)
243{
244 uint32_t t = cpu_to_be32(val);
245
246 memcpy(buf, &t, sizeof(uint32_t));
247}
248
249static inline void sgio_set_be64(uint64_t val, uint8_t *buf)
250{
251 uint64_t t = cpu_to_be64(val);
252
253 memcpy(buf, &t, sizeof(uint64_t));
254}
255
256static inline bool sgio_unbuffered(struct thread_data *td)
257{
258 return (td->o.odirect || td->o.sync_io);
259}
260
261static void sgio_hdr_init(struct sgio_data *sd, struct sg_io_hdr *hdr,
262 struct io_u *io_u, int fs)
263{
264 struct sgio_cmd *sc = &sd->cmds[io_u->index];
265
266 memset(hdr, 0, sizeof(*hdr));
267 memset(sc->cdb, 0, sizeof(sc->cdb));
268
269 hdr->interface_id = 'S';
270 hdr->cmdp = sc->cdb;
271 hdr->cmd_len = sizeof(sc->cdb);
272 hdr->sbp = sc->sb;
273 hdr->mx_sb_len = sizeof(sc->sb);
274 hdr->pack_id = io_u->index;
275 hdr->usr_ptr = io_u;
276 hdr->timeout = SCSI_TIMEOUT_MS;
277
278 if (fs) {
279 hdr->dxferp = io_u->xfer_buf;
280 hdr->dxfer_len = io_u->xfer_buflen;
281 }
282}
283
284static int pollin_events(struct pollfd *pfds, int fds)
285{
286 int i;
287
288 for (i = 0; i < fds; i++)
289 if (pfds[i].revents & POLLIN)
290 return 1;
291
292 return 0;
293}
294
295static int sg_fd_read(int fd, void *data, size_t size)
296{
297 int err = 0;
298
299 while (size) {
300 ssize_t ret;
301
302 ret = read(fd, data, size);
303 if (ret < 0) {
304 if (errno == EAGAIN || errno == EINTR)
305 continue;
306 err = errno;
307 break;
308 } else if (!ret)
309 break;
310 else {
311 data += ret;
312 size -= ret;
313 }
314 }
315
316 if (err)
317 return err;
318 if (size)
319 return EAGAIN;
320
321 return 0;
322}
323
324static int fio_sgio_getevents(struct thread_data *td, unsigned int min,
325 unsigned int max,
326 const struct timespec fio_unused *t)
327{
328 struct sgio_data *sd = td->io_ops_data;
329 int left = max, eventNum, ret, r = 0, trims = 0;
330 void *buf = sd->sgbuf;
331 unsigned int i, j, events;
332 struct fio_file *f;
333 struct io_u *io_u;
334
335 /*
336 * Fill in the file descriptors
337 */
338 for_each_file(td, f, i) {
339 /*
340 * don't block for min events == 0
341 */
342 if (!min)
343 sd->fd_flags[i] = fio_set_fd_nonblocking(f->fd, "sg");
344 else
345 sd->fd_flags[i] = -1;
346
347 sd->pfds[i].fd = f->fd;
348 sd->pfds[i].events = POLLIN;
349 }
350
351 /*
352 ** There are two counters here:
353 ** - number of SCSI commands completed
354 ** - number of io_us completed
355 **
356 ** These are the same with reads and writes, but
357 ** could differ with trim/unmap commands because
358 ** a single unmap can include multiple io_us
359 */
360
361 while (left > 0) {
362 char *p;
363
364 dprint(FD_IO, "sgio_getevents: sd %p: min=%d, max=%d, left=%d\n", sd, min, max, left);
365
366 do {
367 if (!min)
368 break;
369
370 ret = poll(sd->pfds, td->o.nr_files, -1);
371 if (ret < 0) {
372 if (!r)
373 r = -errno;
374 td_verror(td, errno, "poll");
375 break;
376 } else if (!ret)
377 continue;
378
379 if (pollin_events(sd->pfds, td->o.nr_files))
380 break;
381 } while (1);
382
383 if (r < 0)
384 break;
385
386re_read:
387 p = buf;
388 events = 0;
389 for_each_file(td, f, i) {
390 for (eventNum = 0; eventNum < left; eventNum++) {
391 ret = sg_fd_read(f->fd, p, sizeof(struct sg_io_hdr));
392 dprint(FD_IO, "sgio_getevents: sg_fd_read ret: %d\n", ret);
393 if (ret) {
394 r = -ret;
395 td_verror(td, r, "sg_read");
396 break;
397 }
398 io_u = ((struct sg_io_hdr *)p)->usr_ptr;
399 if (io_u->ddir == DDIR_TRIM) {
400 events += sd->trim_queues[io_u->index]->unmap_range_count;
401 eventNum += sd->trim_queues[io_u->index]->unmap_range_count - 1;
402 } else
403 events++;
404
405 p += sizeof(struct sg_io_hdr);
406 dprint(FD_IO, "sgio_getevents: events: %d, eventNum: %d, left: %d\n", events, eventNum, left);
407 }
408 }
409
410 if (r < 0 && !events)
411 break;
412 if (!events) {
413 usleep(1000);
414 goto re_read;
415 }
416
417 left -= events;
418 r += events;
419
420 for (i = 0; i < events; i++) {
421 struct sg_io_hdr *hdr = (struct sg_io_hdr *) buf + i;
422 sd->events[i + trims] = hdr->usr_ptr;
423 io_u = (struct io_u *)(hdr->usr_ptr);
424
425 if (hdr->info & SG_INFO_CHECK) {
426 /* record if an io error occurred, ignore resid */
427 memcpy(&io_u->hdr, hdr, sizeof(struct sg_io_hdr));
428 sd->events[i + trims]->error = EIO;
429 }
430
431 if (io_u->ddir == DDIR_TRIM) {
432 struct sgio_trim *st = sd->trim_queues[io_u->index];
433#ifdef FIO_SGIO_DEBUG
434 assert(st->trim_io_us[0] == io_u);
435 assert(sd->trim_queue_map[io_u->index] == io_u->index);
436 dprint(FD_IO, "sgio_getevents: reaping %d io_us from trim queue %d\n", st->unmap_range_count, io_u->index);
437 dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", io_u->index, i+trims);
438#endif
439 for (j = 1; j < st->unmap_range_count; j++) {
440 ++trims;
441 sd->events[i + trims] = st->trim_io_us[j];
442#ifdef FIO_SGIO_DEBUG
443 dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", st->trim_io_us[j]->index, i+trims);
444 assert(sd->trim_queue_map[st->trim_io_us[j]->index] == io_u->index);
445#endif
446 if (hdr->info & SG_INFO_CHECK) {
447 /* record if an io error occurred, ignore resid */
448 memcpy(&st->trim_io_us[j]->hdr, hdr, sizeof(struct sg_io_hdr));
449 sd->events[i + trims]->error = EIO;
450 }
451 }
452 events -= st->unmap_range_count - 1;
453 st->unmap_range_count = 0;
454 }
455 }
456 }
457
458 if (!min) {
459 for_each_file(td, f, i) {
460 if (sd->fd_flags[i] == -1)
461 continue;
462
463 if (fcntl(f->fd, F_SETFL, sd->fd_flags[i]) < 0)
464 log_err("fio: sg failed to restore fcntl flags: %s\n", strerror(errno));
465 }
466 }
467
468 return r;
469}
470
471static enum fio_q_status fio_sgio_ioctl_doio(struct thread_data *td,
472 struct fio_file *f,
473 struct io_u *io_u)
474{
475 struct sgio_data *sd = td->io_ops_data;
476 struct sg_io_hdr *hdr = &io_u->hdr;
477 int ret;
478
479 sd->events[0] = io_u;
480
481 ret = ioctl(f->fd, SG_IO, hdr);
482 if (ret < 0)
483 return ret;
484
485 /* record if an io error occurred */
486 if (hdr->info & SG_INFO_CHECK)
487 io_u->error = EIO;
488
489 return FIO_Q_COMPLETED;
490}
491
492static enum fio_q_status fio_sgio_rw_doio(struct thread_data *td,
493 struct fio_file *f,
494 struct io_u *io_u, int do_sync)
495{
496 struct sg_io_hdr *hdr = &io_u->hdr;
497 int ret;
498
499 ret = write(f->fd, hdr, sizeof(*hdr));
500 if (ret < 0)
501 return ret;
502
503 if (do_sync) {
504 /*
505 * We can't just read back the first command that completes
506 * and assume it's the one we need, it could be any command
507 * that is inflight.
508 */
509 do {
510 struct io_u *__io_u;
511
512 ret = read(f->fd, hdr, sizeof(*hdr));
513 if (ret < 0)
514 return ret;
515
516 __io_u = hdr->usr_ptr;
517
518 /* record if an io error occurred */
519 if (hdr->info & SG_INFO_CHECK)
520 __io_u->error = EIO;
521
522 if (__io_u == io_u)
523 break;
524
525 if (io_u_sync_complete(td, __io_u))
526 break;
527
528 } while (1);
529
530 return FIO_Q_COMPLETED;
531 }
532
533 return FIO_Q_QUEUED;
534}
535
536static enum fio_q_status fio_sgio_doio(struct thread_data *td,
537 struct io_u *io_u, int do_sync)
538{
539 struct fio_file *f = io_u->file;
540 enum fio_q_status ret;
541
542 if (f->filetype == FIO_TYPE_BLOCK) {
543 ret = fio_sgio_ioctl_doio(td, f, io_u);
544 if (io_u->error)
545 td_verror(td, io_u->error, __func__);
546 } else {
547 ret = fio_sgio_rw_doio(td, f, io_u, do_sync);
548 if (io_u->error && do_sync)
549 td_verror(td, io_u->error, __func__);
550 }
551
552 return ret;
553}
554
555static void fio_sgio_rw_lba(struct sg_io_hdr *hdr, unsigned long long lba,
556 unsigned long long nr_blocks, bool override16)
557{
558 if (lba < MAX_10B_LBA && !override16) {
559 sgio_set_be32((uint32_t) lba, &hdr->cmdp[2]);
560 sgio_set_be16((uint16_t) nr_blocks, &hdr->cmdp[7]);
561 } else {
562 sgio_set_be64(lba, &hdr->cmdp[2]);
563 sgio_set_be32((uint32_t) nr_blocks, &hdr->cmdp[10]);
564 }
565
566 return;
567}
568
569static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
570{
571 struct sg_io_hdr *hdr = &io_u->hdr;
572 struct sg_options *o = td->eo;
573 struct sgio_data *sd = td->io_ops_data;
574 unsigned long long nr_blocks, lba;
575 int offset;
576
577 if (io_u->xfer_buflen & (sd->bs - 1)) {
578 log_err("read/write not sector aligned\n");
579 return EINVAL;
580 }
581
582 nr_blocks = io_u->xfer_buflen / sd->bs;
583 lba = io_u->offset / sd->bs;
584
585 if (io_u->ddir == DDIR_READ) {
586 sgio_hdr_init(sd, hdr, io_u, 1);
587
588 hdr->dxfer_direction = SG_DXFER_FROM_DEV;
589 if (lba < MAX_10B_LBA)
590 hdr->cmdp[0] = 0x28; // read(10)
591 else
592 hdr->cmdp[0] = 0x88; // read(16)
593
594 if (o->hipri)
595 hdr->flags |= SGV4_FLAG_HIPRI;
596 if (o->readfua)
597 hdr->cmdp[1] |= 0x08;
598
599 fio_sgio_rw_lba(hdr, lba, nr_blocks, false);
600
601 } else if (io_u->ddir == DDIR_WRITE) {
602 sgio_hdr_init(sd, hdr, io_u, 1);
603
604 hdr->dxfer_direction = SG_DXFER_TO_DEV;
605 switch(o->write_mode) {
606 case FIO_SG_WRITE:
607 if (lba < MAX_10B_LBA)
608 hdr->cmdp[0] = 0x2a; // write(10)
609 else
610 hdr->cmdp[0] = 0x8a; // write(16)
611 if (o->hipri)
612 hdr->flags |= SGV4_FLAG_HIPRI;
613 if (o->writefua)
614 hdr->cmdp[1] |= 0x08;
615 break;
616 case FIO_SG_WRITE_VERIFY:
617 if (lba < MAX_10B_LBA)
618 hdr->cmdp[0] = 0x2e; // write and verify(10)
619 else
620 hdr->cmdp[0] = 0x8e; // write and verify(16)
621 break;
622 // BYTCHK is disabled by virtue of the memset in sgio_hdr_init
623 case FIO_SG_WRITE_SAME:
624 hdr->dxfer_len = sd->bs;
625 if (lba < MAX_10B_LBA)
626 hdr->cmdp[0] = 0x41; // write same(10)
627 else
628 hdr->cmdp[0] = 0x93; // write same(16)
629 break;
630 case FIO_SG_WRITE_SAME_NDOB:
631 hdr->cmdp[0] = 0x93; // write same(16)
632 hdr->cmdp[1] |= 0x1; // no data output buffer
633 hdr->dxfer_len = 0;
634 break;
635 case FIO_SG_WRITE_STREAM:
636 hdr->cmdp[0] = 0x9a; // write stream (16)
637 if (o->writefua)
638 hdr->cmdp[1] |= 0x08;
639 sgio_set_be64(lba, &hdr->cmdp[2]);
640 sgio_set_be16((uint16_t) io_u->file->engine_pos, &hdr->cmdp[10]);
641 sgio_set_be16((uint16_t) nr_blocks, &hdr->cmdp[12]);
642 break;
643 case FIO_SG_VERIFY_BYTCHK_00:
644 if (lba < MAX_10B_LBA)
645 hdr->cmdp[0] = 0x2f; // VERIFY(10)
646 else
647 hdr->cmdp[0] = 0x8f; // VERIFY(16)
648 hdr->dxfer_len = 0;
649 break;
650 case FIO_SG_VERIFY_BYTCHK_01:
651 if (lba < MAX_10B_LBA)
652 hdr->cmdp[0] = 0x2f; // VERIFY(10)
653 else
654 hdr->cmdp[0] = 0x8f; // VERIFY(16)
655 hdr->cmdp[1] |= 0x02; // BYTCHK = 01b
656 break;
657 case FIO_SG_VERIFY_BYTCHK_11:
658 if (lba < MAX_10B_LBA)
659 hdr->cmdp[0] = 0x2f; // VERIFY(10)
660 else
661 hdr->cmdp[0] = 0x8f; // VERIFY(16)
662 hdr->cmdp[1] |= 0x06; // BYTCHK = 11b
663 hdr->dxfer_len = sd->bs;
664 break;
665 };
666
667 if (o->write_mode != FIO_SG_WRITE_STREAM)
668 fio_sgio_rw_lba(hdr, lba, nr_blocks,
669 o->write_mode == FIO_SG_WRITE_SAME_NDOB);
670
671 } else if (io_u->ddir == DDIR_TRIM) {
672 struct sgio_trim *st;
673
674 if (sd->current_queue == -1) {
675 sgio_hdr_init(sd, hdr, io_u, 0);
676
677 hdr->cmd_len = 10;
678 hdr->dxfer_direction = SG_DXFER_TO_DEV;
679 hdr->cmdp[0] = 0x42; // unmap
680 sd->current_queue = io_u->index;
681 st = sd->trim_queues[sd->current_queue];
682 hdr->dxferp = st->unmap_param;
683#ifdef FIO_SGIO_DEBUG
684 assert(sd->trim_queues[io_u->index]->unmap_range_count == 0);
685 dprint(FD_IO, "sg: creating new queue based on io_u %d\n", io_u->index);
686#endif
687 }
688 else
689 st = sd->trim_queues[sd->current_queue];
690
691 dprint(FD_IO, "sg: adding io_u %d to trim queue %d\n", io_u->index, sd->current_queue);
692 st->trim_io_us[st->unmap_range_count] = io_u;
693#ifdef FIO_SGIO_DEBUG
694 sd->trim_queue_map[io_u->index] = sd->current_queue;
695#endif
696
697 offset = 8 + 16 * st->unmap_range_count;
698 sgio_set_be64(lba, &st->unmap_param[offset]);
699 sgio_set_be32((uint32_t) nr_blocks, &st->unmap_param[offset + 8]);
700
701 st->unmap_range_count++;
702
703 } else if (ddir_sync(io_u->ddir)) {
704 sgio_hdr_init(sd, hdr, io_u, 0);
705 hdr->dxfer_direction = SG_DXFER_NONE;
706 if (lba < MAX_10B_LBA)
707 hdr->cmdp[0] = 0x35; // synccache(10)
708 else
709 hdr->cmdp[0] = 0x91; // synccache(16)
710 } else
711 assert(0);
712
713 return 0;
714}
715
716static void fio_sgio_unmap_setup(struct sg_io_hdr *hdr, struct sgio_trim *st)
717{
718 uint16_t cnt = st->unmap_range_count * 16;
719
720 hdr->dxfer_len = cnt + 8;
721 sgio_set_be16(cnt + 8, &hdr->cmdp[7]);
722 sgio_set_be16(cnt + 6, st->unmap_param);
723 sgio_set_be16(cnt, &st->unmap_param[2]);
724
725 return;
726}
727
728static enum fio_q_status fio_sgio_queue(struct thread_data *td,
729 struct io_u *io_u)
730{
731 struct sg_io_hdr *hdr = &io_u->hdr;
732 struct sgio_data *sd = td->io_ops_data;
733 int ret, do_sync = 0;
734
735 fio_ro_check(td, io_u);
736
737 if (sgio_unbuffered(td) || ddir_sync(io_u->ddir))
738 do_sync = 1;
739
740 if (io_u->ddir == DDIR_TRIM) {
741 if (do_sync || io_u->file->filetype == FIO_TYPE_BLOCK) {
742 struct sgio_trim *st = sd->trim_queues[sd->current_queue];
743
744 /* finish cdb setup for unmap because we are
745 ** doing unmap commands synchronously */
746#ifdef FIO_SGIO_DEBUG
747 assert(st->unmap_range_count == 1);
748 assert(io_u == st->trim_io_us[0]);
749#endif
750 hdr = &io_u->hdr;
751
752 fio_sgio_unmap_setup(hdr, st);
753
754 st->unmap_range_count = 0;
755 sd->current_queue = -1;
756 } else
757 /* queue up trim ranges and submit in commit() */
758 return FIO_Q_QUEUED;
759 }
760
761 ret = fio_sgio_doio(td, io_u, do_sync);
762
763 if (ret < 0)
764 io_u->error = errno;
765 else if (hdr->status) {
766 io_u->resid = hdr->resid;
767 io_u->error = EIO;
768 } else if (td->io_ops->commit != NULL) {
769 if (do_sync && !ddir_sync(io_u->ddir)) {
770 io_u_mark_submit(td, 1);
771 io_u_mark_complete(td, 1);
772 } else if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
773 io_u_mark_submit(td, 1);
774 io_u_queued(td, io_u);
775 }
776 }
777
778 if (io_u->error) {
779 td_verror(td, io_u->error, "xfer");
780 return FIO_Q_COMPLETED;
781 }
782
783 return ret;
784}
785
786static int fio_sgio_commit(struct thread_data *td)
787{
788 struct sgio_data *sd = td->io_ops_data;
789 struct sgio_trim *st;
790 struct io_u *io_u;
791 struct sg_io_hdr *hdr;
792 struct timespec now;
793 unsigned int i;
794 int ret;
795
796 if (sd->current_queue == -1)
797 return 0;
798
799 st = sd->trim_queues[sd->current_queue];
800 io_u = st->trim_io_us[0];
801 hdr = &io_u->hdr;
802
803 fio_sgio_unmap_setup(hdr, st);
804
805 sd->current_queue = -1;
806
807 ret = fio_sgio_rw_doio(td, io_u->file, io_u, 0);
808
809 if (ret < 0 || hdr->status) {
810 int error;
811
812 if (ret < 0)
813 error = errno;
814 else {
815 error = EIO;
816 ret = -EIO;
817 }
818
819 for (i = 0; i < st->unmap_range_count; i++) {
820 st->trim_io_us[i]->error = error;
821 clear_io_u(td, st->trim_io_us[i]);
822 if (hdr->status)
823 st->trim_io_us[i]->resid = hdr->resid;
824 }
825
826 td_verror(td, error, "xfer");
827 return ret;
828 }
829
830 if (fio_fill_issue_time(td)) {
831 fio_gettime(&now, NULL);
832 for (i = 0; i < st->unmap_range_count; i++) {
833 memcpy(&st->trim_io_us[i]->issue_time, &now, sizeof(now));
834 io_u_queued(td, io_u);
835 }
836 }
837 io_u_mark_submit(td, st->unmap_range_count);
838
839 return 0;
840}
841
842static struct io_u *fio_sgio_event(struct thread_data *td, int event)
843{
844 struct sgio_data *sd = td->io_ops_data;
845
846 return sd->events[event];
847}
848
849static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
850 unsigned long long *max_lba)
851{
852 /*
853 * need to do read capacity operation w/o benefit of sd or
854 * io_u structures, which are not initialized until later.
855 */
856 struct sg_io_hdr hdr;
857 unsigned long long hlba;
858 unsigned int blksz = 0;
859 unsigned char cmd[16];
860 unsigned char sb[64];
861 unsigned char buf[32]; // read capacity return
862 int ret;
863 int fd = -1;
864
865 struct fio_file *f = td->files[0];
866
867 /* open file independent of rest of application */
868 fd = open(f->file_name, O_RDONLY);
869 if (fd < 0)
870 return -errno;
871
872 memset(&hdr, 0, sizeof(hdr));
873 memset(cmd, 0, sizeof(cmd));
874 memset(sb, 0, sizeof(sb));
875 memset(buf, 0, sizeof(buf));
876
877 /* First let's try a 10 byte read capacity. */
878 hdr.interface_id = 'S';
879 hdr.cmdp = cmd;
880 hdr.cmd_len = 10;
881 hdr.sbp = sb;
882 hdr.mx_sb_len = sizeof(sb);
883 hdr.timeout = SCSI_TIMEOUT_MS;
884 hdr.cmdp[0] = 0x25; // Read Capacity(10)
885 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
886 hdr.dxferp = buf;
887 hdr.dxfer_len = sizeof(buf);
888
889 ret = ioctl(fd, SG_IO, &hdr);
890 if (ret < 0) {
891 close(fd);
892 return ret;
893 }
894
895 if (hdr.info & SG_INFO_CHECK) {
896 /* RCAP(10) might be unsupported by device. Force RCAP(16) */
897 hlba = MAX_10B_LBA;
898 } else {
899 blksz = sgio_get_be32(&buf[4]);
900 hlba = sgio_get_be32(buf);
901 }
902
903 /*
904 * If max lba masked by MAX_10B_LBA equals MAX_10B_LBA,
905 * then need to retry with 16 byte Read Capacity command.
906 */
907 if (hlba == MAX_10B_LBA) {
908 hdr.cmd_len = 16;
909 hdr.cmdp[0] = 0x9e; // service action
910 hdr.cmdp[1] = 0x10; // Read Capacity(16)
911 sgio_set_be32(sizeof(buf), &hdr.cmdp[10]);
912
913 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
914 hdr.dxferp = buf;
915 hdr.dxfer_len = sizeof(buf);
916
917 ret = ioctl(fd, SG_IO, &hdr);
918 if (ret < 0) {
919 close(fd);
920 return ret;
921 }
922
923 /* record if an io error occurred */
924 if (hdr.info & SG_INFO_CHECK)
925 td_verror(td, EIO, "fio_sgio_read_capacity");
926
927 blksz = sgio_get_be32(&buf[8]);
928 hlba = sgio_get_be64(buf);
929 }
930
931 if (blksz) {
932 *bs = blksz;
933 *max_lba = hlba;
934 ret = 0;
935 } else {
936 ret = EIO;
937 }
938
939 close(fd);
940 return ret;
941}
942
943static void fio_sgio_cleanup(struct thread_data *td)
944{
945 struct sgio_data *sd = td->io_ops_data;
946 int i;
947
948 if (sd) {
949 free(sd->events);
950 free(sd->cmds);
951 free(sd->fd_flags);
952 free(sd->pfds);
953 free(sd->sgbuf);
954#ifdef FIO_SGIO_DEBUG
955 free(sd->trim_queue_map);
956#endif
957
958 for (i = 0; i < td->o.iodepth; i++) {
959 free(sd->trim_queues[i]->unmap_param);
960 free(sd->trim_queues[i]->trim_io_us);
961 free(sd->trim_queues[i]);
962 }
963
964 free(sd->trim_queues);
965 free(sd);
966 }
967}
968
969static int fio_sgio_init(struct thread_data *td)
970{
971 struct sgio_data *sd;
972 struct sgio_trim *st;
973 struct sg_io_hdr *h3p;
974 int i;
975
976 sd = calloc(1, sizeof(*sd));
977 sd->cmds = calloc(td->o.iodepth, sizeof(struct sgio_cmd));
978 sd->sgbuf = calloc(td->o.iodepth, sizeof(struct sg_io_hdr));
979 sd->events = calloc(td->o.iodepth, sizeof(struct io_u *));
980 sd->pfds = calloc(td->o.nr_files, sizeof(struct pollfd));
981 sd->fd_flags = calloc(td->o.nr_files, sizeof(int));
982 sd->type_checked = 0;
983
984 sd->trim_queues = calloc(td->o.iodepth, sizeof(struct sgio_trim *));
985 sd->current_queue = -1;
986#ifdef FIO_SGIO_DEBUG
987 sd->trim_queue_map = calloc(td->o.iodepth, sizeof(int));
988#endif
989 for (i = 0, h3p = sd->sgbuf; i < td->o.iodepth; i++, ++h3p) {
990 sd->trim_queues[i] = calloc(1, sizeof(struct sgio_trim));
991 st = sd->trim_queues[i];
992 st->unmap_param = calloc(td->o.iodepth + 1, sizeof(char[16]));
993 st->unmap_range_count = 0;
994 st->trim_io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
995 h3p->interface_id = 'S';
996 }
997
998 td->io_ops_data = sd;
999
1000 /*
1001 * we want to do it, regardless of whether odirect is set or not
1002 */
1003 td->o.override_sync = 1;
1004 return 0;
1005}
1006
1007static int fio_sgio_type_check(struct thread_data *td, struct fio_file *f)
1008{
1009 struct sgio_data *sd = td->io_ops_data;
1010 unsigned int bs = 0;
1011 unsigned long long max_lba = 0;
1012
1013 if (f->filetype == FIO_TYPE_BLOCK) {
1014 if (ioctl(f->fd, BLKSSZGET, &bs) < 0) {
1015 td_verror(td, errno, "ioctl");
1016 return 1;
1017 }
1018 } else if (f->filetype == FIO_TYPE_CHAR) {
1019 int version, ret;
1020
1021 if (ioctl(f->fd, SG_GET_VERSION_NUM, &version) < 0) {
1022 td_verror(td, errno, "ioctl");
1023 return 1;
1024 }
1025
1026 ret = fio_sgio_read_capacity(td, &bs, &max_lba);
1027 if (ret) {
1028 td_verror(td, td->error, "fio_sgio_read_capacity");
1029 log_err("ioengine sg unable to read capacity successfully\n");
1030 return 1;
1031 }
1032 } else {
1033 td_verror(td, EINVAL, "wrong file type");
1034 log_err("ioengine sg only works on block or character devices\n");
1035 return 1;
1036 }
1037
1038 sd->bs = bs;
1039 // Determine size of commands needed based on max_lba
1040 if (max_lba >= MAX_10B_LBA) {
1041 dprint(FD_IO, "sgio_type_check: using 16 byte read/write "
1042 "commands for lba above 0x%016llx/0x%016llx\n",
1043 MAX_10B_LBA, max_lba);
1044 }
1045
1046 if (f->filetype == FIO_TYPE_BLOCK) {
1047 td->io_ops->getevents = NULL;
1048 td->io_ops->event = NULL;
1049 td->io_ops->commit = NULL;
1050 /*
1051 ** Setting these functions to null may cause problems
1052 ** with filename=/dev/sda:/dev/sg0 since we are only
1053 ** considering a single file
1054 */
1055 }
1056 sd->type_checked = 1;
1057
1058 return 0;
1059}
1060
1061static int fio_sgio_stream_control(struct fio_file *f, bool open_stream, uint16_t *stream_id)
1062{
1063 struct sg_io_hdr hdr;
1064 unsigned char cmd[16];
1065 unsigned char sb[64];
1066 unsigned char buf[8];
1067 int ret;
1068
1069 memset(&hdr, 0, sizeof(hdr));
1070 memset(cmd, 0, sizeof(cmd));
1071 memset(sb, 0, sizeof(sb));
1072 memset(buf, 0, sizeof(buf));
1073
1074 hdr.interface_id = 'S';
1075 hdr.cmdp = cmd;
1076 hdr.cmd_len = 16;
1077 hdr.sbp = sb;
1078 hdr.mx_sb_len = sizeof(sb);
1079 hdr.timeout = SCSI_TIMEOUT_MS;
1080 hdr.cmdp[0] = 0x9e;
1081 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
1082 hdr.dxferp = buf;
1083 hdr.dxfer_len = sizeof(buf);
1084 sgio_set_be32(sizeof(buf), &hdr.cmdp[10]);
1085
1086 if (open_stream)
1087 hdr.cmdp[1] = 0x34;
1088 else {
1089 hdr.cmdp[1] = 0x54;
1090 sgio_set_be16(*stream_id, &hdr.cmdp[4]);
1091 }
1092
1093 ret = ioctl(f->fd, SG_IO, &hdr);
1094
1095 if (ret < 0)
1096 return ret;
1097
1098 if (hdr.info & SG_INFO_CHECK)
1099 return 1;
1100
1101 if (open_stream) {
1102 *stream_id = sgio_get_be16(&buf[4]);
1103 dprint(FD_FILE, "sgio_stream_control: opened stream %u\n", (unsigned int) *stream_id);
1104 assert(*stream_id != 0);
1105 } else
1106 dprint(FD_FILE, "sgio_stream_control: closed stream %u\n", (unsigned int) *stream_id);
1107
1108 return 0;
1109}
1110
1111static int fio_sgio_open(struct thread_data *td, struct fio_file *f)
1112{
1113 struct sgio_data *sd = td->io_ops_data;
1114 struct sg_options *o = td->eo;
1115 int ret;
1116
1117 ret = generic_open_file(td, f);
1118 if (ret)
1119 return ret;
1120
1121 if (sd && !sd->type_checked && fio_sgio_type_check(td, f)) {
1122 ret = generic_close_file(td, f);
1123 return ret;
1124 }
1125
1126 if (o->write_mode == FIO_SG_WRITE_STREAM) {
1127 if (o->stream_id)
1128 f->engine_pos = o->stream_id;
1129 else {
1130 ret = fio_sgio_stream_control(f, true, (uint16_t *) &f->engine_pos);
1131 if (ret)
1132 return ret;
1133 }
1134 }
1135
1136 return 0;
1137}
1138
1139int fio_sgio_close(struct thread_data *td, struct fio_file *f)
1140{
1141 struct sg_options *o = td->eo;
1142 int ret;
1143
1144 if (!o->stream_id && o->write_mode == FIO_SG_WRITE_STREAM) {
1145 ret = fio_sgio_stream_control(f, false, (uint16_t *) &f->engine_pos);
1146 if (ret)
1147 return ret;
1148 }
1149
1150 return generic_close_file(td, f);
1151}
1152
1153/*
1154 * Build an error string with details about the driver, host or scsi
1155 * error contained in the sg header Caller will use as necessary.
1156 */
1157static char *fio_sgio_errdetails(struct io_u *io_u)
1158{
1159 struct sg_io_hdr *hdr = &io_u->hdr;
1160#define MAXERRDETAIL 1024
1161#define MAXMSGCHUNK 128
1162 char *msg, msgchunk[MAXMSGCHUNK];
1163 int i;
1164
1165 msg = calloc(1, MAXERRDETAIL);
1166 strcpy(msg, "");
1167
1168 /*
1169 * can't seem to find sg_err.h, so I'll just echo the define values
1170 * so others can search on internet to find clearer clues of meaning.
1171 */
1172 if (hdr->info & SG_INFO_CHECK) {
1173 if (hdr->host_status) {
1174 snprintf(msgchunk, MAXMSGCHUNK, "SG Host Status: 0x%02x; ", hdr->host_status);
1175 strlcat(msg, msgchunk, MAXERRDETAIL);
1176 switch (hdr->host_status) {
1177 case 0x01:
1178 strlcat(msg, "SG_ERR_DID_NO_CONNECT", MAXERRDETAIL);
1179 break;
1180 case 0x02:
1181 strlcat(msg, "SG_ERR_DID_BUS_BUSY", MAXERRDETAIL);
1182 break;
1183 case 0x03:
1184 strlcat(msg, "SG_ERR_DID_TIME_OUT", MAXERRDETAIL);
1185 break;
1186 case 0x04:
1187 strlcat(msg, "SG_ERR_DID_BAD_TARGET", MAXERRDETAIL);
1188 break;
1189 case 0x05:
1190 strlcat(msg, "SG_ERR_DID_ABORT", MAXERRDETAIL);
1191 break;
1192 case 0x06:
1193 strlcat(msg, "SG_ERR_DID_PARITY", MAXERRDETAIL);
1194 break;
1195 case 0x07:
1196 strlcat(msg, "SG_ERR_DID_ERROR (internal error)", MAXERRDETAIL);
1197 break;
1198 case 0x08:
1199 strlcat(msg, "SG_ERR_DID_RESET", MAXERRDETAIL);
1200 break;
1201 case 0x09:
1202 strlcat(msg, "SG_ERR_DID_BAD_INTR (unexpected)", MAXERRDETAIL);
1203 break;
1204 case 0x0a:
1205 strlcat(msg, "SG_ERR_DID_PASSTHROUGH", MAXERRDETAIL);
1206 break;
1207 case 0x0b:
1208 strlcat(msg, "SG_ERR_DID_SOFT_ERROR (driver retry?)", MAXERRDETAIL);
1209 break;
1210 case 0x0c:
1211 strlcat(msg, "SG_ERR_DID_IMM_RETRY", MAXERRDETAIL);
1212 break;
1213 case 0x0d:
1214 strlcat(msg, "SG_ERR_DID_REQUEUE", MAXERRDETAIL);
1215 break;
1216 case 0x0e:
1217 strlcat(msg, "SG_ERR_DID_TRANSPORT_DISRUPTED", MAXERRDETAIL);
1218 break;
1219 case 0x0f:
1220 strlcat(msg, "SG_ERR_DID_TRANSPORT_FAILFAST", MAXERRDETAIL);
1221 break;
1222 case 0x10:
1223 strlcat(msg, "SG_ERR_DID_TARGET_FAILURE", MAXERRDETAIL);
1224 break;
1225 case 0x11:
1226 strlcat(msg, "SG_ERR_DID_NEXUS_FAILURE", MAXERRDETAIL);
1227 break;
1228 case 0x12:
1229 strlcat(msg, "SG_ERR_DID_ALLOC_FAILURE", MAXERRDETAIL);
1230 break;
1231 case 0x13:
1232 strlcat(msg, "SG_ERR_DID_MEDIUM_ERROR", MAXERRDETAIL);
1233 break;
1234 default:
1235 strlcat(msg, "Unknown", MAXERRDETAIL);
1236 break;
1237 }
1238 strlcat(msg, ". ", MAXERRDETAIL);
1239 }
1240 if (hdr->driver_status) {
1241 snprintf(msgchunk, MAXMSGCHUNK, "SG Driver Status: 0x%02x; ", hdr->driver_status);
1242 strlcat(msg, msgchunk, MAXERRDETAIL);
1243 switch (hdr->driver_status & 0x0F) {
1244 case 0x01:
1245 strlcat(msg, "SG_ERR_DRIVER_BUSY", MAXERRDETAIL);
1246 break;
1247 case 0x02:
1248 strlcat(msg, "SG_ERR_DRIVER_SOFT", MAXERRDETAIL);
1249 break;
1250 case 0x03:
1251 strlcat(msg, "SG_ERR_DRIVER_MEDIA", MAXERRDETAIL);
1252 break;
1253 case 0x04:
1254 strlcat(msg, "SG_ERR_DRIVER_ERROR", MAXERRDETAIL);
1255 break;
1256 case 0x05:
1257 strlcat(msg, "SG_ERR_DRIVER_INVALID", MAXERRDETAIL);
1258 break;
1259 case 0x06:
1260 strlcat(msg, "SG_ERR_DRIVER_TIMEOUT", MAXERRDETAIL);
1261 break;
1262 case 0x07:
1263 strlcat(msg, "SG_ERR_DRIVER_HARD", MAXERRDETAIL);
1264 break;
1265 case 0x08:
1266 strlcat(msg, "SG_ERR_DRIVER_SENSE", MAXERRDETAIL);
1267 break;
1268 default:
1269 strlcat(msg, "Unknown", MAXERRDETAIL);
1270 break;
1271 }
1272 strlcat(msg, "; ", MAXERRDETAIL);
1273 switch (hdr->driver_status & 0xF0) {
1274 case 0x10:
1275 strlcat(msg, "SG_ERR_SUGGEST_RETRY", MAXERRDETAIL);
1276 break;
1277 case 0x20:
1278 strlcat(msg, "SG_ERR_SUGGEST_ABORT", MAXERRDETAIL);
1279 break;
1280 case 0x30:
1281 strlcat(msg, "SG_ERR_SUGGEST_REMAP", MAXERRDETAIL);
1282 break;
1283 case 0x40:
1284 strlcat(msg, "SG_ERR_SUGGEST_DIE", MAXERRDETAIL);
1285 break;
1286 case 0x80:
1287 strlcat(msg, "SG_ERR_SUGGEST_SENSE", MAXERRDETAIL);
1288 break;
1289 }
1290 strlcat(msg, ". ", MAXERRDETAIL);
1291 }
1292 if (hdr->status) {
1293 snprintf(msgchunk, MAXMSGCHUNK, "SG SCSI Status: 0x%02x; ", hdr->status);
1294 strlcat(msg, msgchunk, MAXERRDETAIL);
1295 // SCSI 3 status codes
1296 switch (hdr->status) {
1297 case 0x02:
1298 strlcat(msg, "CHECK_CONDITION", MAXERRDETAIL);
1299 break;
1300 case 0x04:
1301 strlcat(msg, "CONDITION_MET", MAXERRDETAIL);
1302 break;
1303 case 0x08:
1304 strlcat(msg, "BUSY", MAXERRDETAIL);
1305 break;
1306 case 0x10:
1307 strlcat(msg, "INTERMEDIATE", MAXERRDETAIL);
1308 break;
1309 case 0x14:
1310 strlcat(msg, "INTERMEDIATE_CONDITION_MET", MAXERRDETAIL);
1311 break;
1312 case 0x18:
1313 strlcat(msg, "RESERVATION_CONFLICT", MAXERRDETAIL);
1314 break;
1315 case 0x22:
1316 strlcat(msg, "COMMAND_TERMINATED", MAXERRDETAIL);
1317 break;
1318 case 0x28:
1319 strlcat(msg, "TASK_SET_FULL", MAXERRDETAIL);
1320 break;
1321 case 0x30:
1322 strlcat(msg, "ACA_ACTIVE", MAXERRDETAIL);
1323 break;
1324 case 0x40:
1325 strlcat(msg, "TASK_ABORTED", MAXERRDETAIL);
1326 break;
1327 default:
1328 strlcat(msg, "Unknown", MAXERRDETAIL);
1329 break;
1330 }
1331 strlcat(msg, ". ", MAXERRDETAIL);
1332 }
1333 if (hdr->sb_len_wr) {
1334 const uint8_t *const sbp = hdr->sbp;
1335
1336 snprintf(msgchunk, MAXMSGCHUNK, "Sense Data (%d bytes):", hdr->sb_len_wr);
1337 strlcat(msg, msgchunk, MAXERRDETAIL);
1338 for (i = 0; i < hdr->sb_len_wr; i++) {
1339 snprintf(msgchunk, MAXMSGCHUNK, " %02x", sbp[i]);
1340 strlcat(msg, msgchunk, MAXERRDETAIL);
1341 }
1342 strlcat(msg, ". ", MAXERRDETAIL);
1343 }
1344 if (hdr->resid != 0) {
1345 snprintf(msgchunk, MAXMSGCHUNK, "SG Driver: %d bytes out of %d not transferred. ", hdr->resid, hdr->dxfer_len);
1346 strlcat(msg, msgchunk, MAXERRDETAIL);
1347 }
1348 if (hdr->cmdp) {
1349 strlcat(msg, "cdb:", MAXERRDETAIL);
1350 for (i = 0; i < hdr->cmd_len; i++) {
1351 snprintf(msgchunk, MAXMSGCHUNK, " %02x", hdr->cmdp[i]);
1352 strlcat(msg, msgchunk, MAXERRDETAIL);
1353 }
1354 strlcat(msg, ". ", MAXERRDETAIL);
1355 if (io_u->ddir == DDIR_TRIM) {
1356 unsigned char *param_list = hdr->dxferp;
1357 strlcat(msg, "dxferp:", MAXERRDETAIL);
1358 for (i = 0; i < hdr->dxfer_len; i++) {
1359 snprintf(msgchunk, MAXMSGCHUNK, " %02x", param_list[i]);
1360 strlcat(msg, msgchunk, MAXERRDETAIL);
1361 }
1362 strlcat(msg, ". ", MAXERRDETAIL);
1363 }
1364 }
1365 }
1366
1367 if (!(hdr->info & SG_INFO_CHECK) && !strlen(msg))
1368 snprintf(msg, MAXERRDETAIL, "%s",
1369 "SG Driver did not report a Host, Driver or Device check");
1370
1371 return msg;
1372}
1373
1374/*
1375 * get max file size from read capacity.
1376 */
1377static int fio_sgio_get_file_size(struct thread_data *td, struct fio_file *f)
1378{
1379 /*
1380 * get_file_size is being called even before sgio_init is
1381 * called, so none of the sg_io structures are
1382 * initialized in the thread_data yet. So we need to do the
1383 * ReadCapacity without any of those helpers. One of the effects
1384 * is that ReadCapacity may get called 4 times on each open:
1385 * readcap(10) followed by readcap(16) if needed - just to get
1386 * the file size after the init occurs - it will be called
1387 * again when "type_check" is called during structure
1388 * initialization I'm not sure how to prevent this little
1389 * inefficiency.
1390 */
1391 unsigned int bs = 0;
1392 unsigned long long max_lba = 0;
1393 int ret;
1394
1395 if (fio_file_size_known(f))
1396 return 0;
1397
1398 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) {
1399 td_verror(td, EINVAL, "wrong file type");
1400 log_err("ioengine sg only works on block or character devices\n");
1401 return 1;
1402 }
1403
1404 ret = fio_sgio_read_capacity(td, &bs, &max_lba);
1405 if (ret ) {
1406 td_verror(td, td->error, "fio_sgio_read_capacity");
1407 log_err("ioengine sg unable to successfully execute read capacity to get block size and maximum lba\n");
1408 return 1;
1409 }
1410
1411 f->real_file_size = (max_lba + 1) * bs;
1412 fio_file_set_size_known(f);
1413 return 0;
1414}
1415
1416
1417static struct ioengine_ops ioengine = {
1418 .name = "sg",
1419 .version = FIO_IOOPS_VERSION,
1420 .init = fio_sgio_init,
1421 .prep = fio_sgio_prep,
1422 .queue = fio_sgio_queue,
1423 .commit = fio_sgio_commit,
1424 .getevents = fio_sgio_getevents,
1425 .errdetails = fio_sgio_errdetails,
1426 .event = fio_sgio_event,
1427 .cleanup = fio_sgio_cleanup,
1428 .open_file = fio_sgio_open,
1429 .close_file = fio_sgio_close,
1430 .get_file_size = fio_sgio_get_file_size,
1431 .flags = FIO_SYNCIO | FIO_RAWIO | FIO_RO_NEEDS_RW_OPEN,
1432 .options = options,
1433 .option_struct_size = sizeof(struct sg_options)
1434};
1435
1436#else /* FIO_HAVE_SGIO */
1437
1438/*
1439 * When we have a proper configure system in place, we simply wont build
1440 * and install this io engine. For now install a crippled version that
1441 * just complains and fails to load.
1442 */
1443static int fio_sgio_init(struct thread_data fio_unused *td)
1444{
1445 log_err("fio: ioengine sg not available\n");
1446 return 1;
1447}
1448
1449static struct ioengine_ops ioengine = {
1450 .name = "sg",
1451 .version = FIO_IOOPS_VERSION,
1452 .init = fio_sgio_init,
1453};
1454
1455#endif
1456
1457static void fio_init fio_sgio_register(void)
1458{
1459 register_ioengine(&ioengine);
1460}
1461
1462static void fio_exit fio_sgio_unregister(void)
1463{
1464 unregister_ioengine(&ioengine);
1465}