4 * IO engine that uses the Linux SG v3 interface to talk to SCSI devices
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
11 * What value does queue() return for the different cases?
12 * queue() return value
14 * /dev/sdX RWT FIO_Q_COMPLETED
15 * /dev/sgY RWT FIO_Q_COMPLETED
16 * with direct=1 or sync=1
19 * /dev/sgY RWT FIO_Q_QUEUED
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()
25 * Where are the IO counting functions called for the different cases?
28 * /dev/sdX (commit==NULL)
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()
34 * /dev/sgY with direct=1 or sync=1 (commit does nothing)
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()
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()
59 #include "../optgroup.h"
65 FIO_SG_WRITE_VERIFY = 2,
72 unsigned int writefua;
73 unsigned int write_mode;
76 static struct fio_option options[] = {
79 .lname = "sg engine read fua flag support",
81 .off1 = offsetof(struct sg_options, readfua),
82 .help = "Set FUA flag (force unit access) for all Read operations",
84 .category = FIO_OPT_C_ENGINE,
85 .group = FIO_OPT_G_SG,
89 .lname = "sg engine write fua flag support",
91 .off1 = offsetof(struct sg_options, writefua),
92 .help = "Set FUA flag (force unit access) for all Write operations",
94 .category = FIO_OPT_C_ENGINE,
95 .group = FIO_OPT_G_SG,
98 .name = "sg_write_mode",
99 .lname = "specify sg write mode",
101 .off1 = offsetof(struct sg_options, write_mode),
102 .help = "Specify SCSI WRITE mode",
106 .oval = FIO_SG_WRITE,
107 .help = "Issue standard SCSI WRITE commands",
110 .oval = FIO_SG_WRITE_VERIFY,
111 .help = "Issue SCSI WRITE AND VERIFY commands",
114 .oval = FIO_SG_WRITE_SAME,
115 .help = "Issue SCSI WRITE SAME commands",
118 .category = FIO_OPT_C_ENGINE,
119 .group = FIO_OPT_G_SG,
126 #define MAX_10B_LBA 0xFFFFFFFFULL
127 #define SCSI_TIMEOUT_MS 30000 // 30 second timeout; currently no method to override
128 #define MAX_SB 64 // sense block maximum return size
131 unsigned char cdb[16]; // enhanced from 10 to support 16 byte commands
132 unsigned char sb[MAX_SB]; // add sense block to commands
138 unsigned int unmap_range_count;
139 struct io_u **trim_io_us;
143 struct sgio_cmd *cmds;
144 struct io_u **events;
150 struct sgio_trim **trim_queues;
152 unsigned int *trim_queue_map;
155 static inline bool sgio_unbuffered(struct thread_data *td)
157 return (td->o.odirect || td->o.sync_io);
160 static void sgio_hdr_init(struct sgio_data *sd, struct sg_io_hdr *hdr,
161 struct io_u *io_u, int fs)
163 struct sgio_cmd *sc = &sd->cmds[io_u->index];
165 memset(hdr, 0, sizeof(*hdr));
166 memset(sc->cdb, 0, sizeof(sc->cdb));
168 hdr->interface_id = 'S';
170 hdr->cmd_len = sizeof(sc->cdb);
172 hdr->mx_sb_len = sizeof(sc->sb);
173 hdr->pack_id = io_u->index;
175 hdr->timeout = SCSI_TIMEOUT_MS;
178 hdr->dxferp = io_u->xfer_buf;
179 hdr->dxfer_len = io_u->xfer_buflen;
183 static int pollin_events(struct pollfd *pfds, int fds)
187 for (i = 0; i < fds; i++)
188 if (pfds[i].revents & POLLIN)
194 static int sg_fd_read(int fd, void *data, size_t size)
201 ret = read(fd, data, size);
203 if (errno == EAGAIN || errno == EINTR)
223 static int fio_sgio_getevents(struct thread_data *td, unsigned int min,
225 const struct timespec fio_unused *t)
227 struct sgio_data *sd = td->io_ops_data;
228 int left = max, eventNum, ret, r = 0, trims = 0;
229 void *buf = sd->sgbuf;
230 unsigned int i, j, events;
235 * Fill in the file descriptors
237 for_each_file(td, f, i) {
239 * don't block for min events == 0
242 sd->fd_flags[i] = fio_set_fd_nonblocking(f->fd, "sg");
244 sd->fd_flags[i] = -1;
246 sd->pfds[i].fd = f->fd;
247 sd->pfds[i].events = POLLIN;
251 ** There are two counters here:
252 ** - number of SCSI commands completed
253 ** - number of io_us completed
255 ** These are the same with reads and writes, but
256 ** could differ with trim/unmap commands because
257 ** a single unmap can include multiple io_us
263 dprint(FD_IO, "sgio_getevents: sd %p: min=%d, max=%d, left=%d\n", sd, min, max, left);
269 ret = poll(sd->pfds, td->o.nr_files, -1);
273 td_verror(td, errno, "poll");
278 if (pollin_events(sd->pfds, td->o.nr_files))
288 for_each_file(td, f, i) {
289 for (eventNum = 0; eventNum < left; eventNum++) {
290 ret = sg_fd_read(f->fd, p, sizeof(struct sg_io_hdr));
291 dprint(FD_IO, "sgio_getevents: sg_fd_read ret: %d\n", ret);
294 td_verror(td, r, "sg_read");
297 io_u = ((struct sg_io_hdr *)p)->usr_ptr;
298 if (io_u->ddir == DDIR_TRIM) {
299 events += sd->trim_queues[io_u->index]->unmap_range_count;
300 eventNum += sd->trim_queues[io_u->index]->unmap_range_count - 1;
304 p += sizeof(struct sg_io_hdr);
305 dprint(FD_IO, "sgio_getevents: events: %d, eventNum: %d, left: %d\n", events, eventNum, left);
309 if (r < 0 && !events)
319 for (i = 0; i < events; i++) {
320 struct sg_io_hdr *hdr = (struct sg_io_hdr *) buf + i;
321 sd->events[i + trims] = hdr->usr_ptr;
322 io_u = (struct io_u *)(hdr->usr_ptr);
324 if (hdr->info & SG_INFO_CHECK) {
325 /* record if an io error occurred, ignore resid */
326 memcpy(&io_u->hdr, hdr, sizeof(struct sg_io_hdr));
327 sd->events[i + trims]->error = EIO;
330 if (io_u->ddir == DDIR_TRIM) {
331 struct sgio_trim *st = sd->trim_queues[io_u->index];
332 assert(st->trim_io_us[0] == io_u);
333 dprint(FD_IO, "sgio_getevents: reaping %d io_us from trim queue %d\n", st->unmap_range_count, io_u->index);
334 dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", io_u->index, i+trims);
335 for (j = 1; j < st->unmap_range_count; j++) {
337 sd->events[i + trims] = st->trim_io_us[j];
338 dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", st->trim_io_us[j]->index, i+trims);
339 if (hdr->info & SG_INFO_CHECK) {
340 /* record if an io error occurred, ignore resid */
341 memcpy(&st->trim_io_us[j]->hdr, hdr, sizeof(struct sg_io_hdr));
342 sd->events[i + trims]->error = EIO;
345 events -= st->unmap_range_count - 1;
346 st->unmap_range_count = 0;
352 for_each_file(td, f, i) {
353 if (sd->fd_flags[i] == -1)
356 if (fcntl(f->fd, F_SETFL, sd->fd_flags[i]) < 0)
357 log_err("fio: sg failed to restore fcntl flags: %s\n", strerror(errno));
364 static enum fio_q_status fio_sgio_ioctl_doio(struct thread_data *td,
368 struct sgio_data *sd = td->io_ops_data;
369 struct sg_io_hdr *hdr = &io_u->hdr;
372 sd->events[0] = io_u;
374 ret = ioctl(f->fd, SG_IO, hdr);
378 /* record if an io error occurred */
379 if (hdr->info & SG_INFO_CHECK)
382 return FIO_Q_COMPLETED;
385 static enum fio_q_status fio_sgio_rw_doio(struct fio_file *f,
386 struct io_u *io_u, int do_sync)
388 struct sg_io_hdr *hdr = &io_u->hdr;
391 ret = write(f->fd, hdr, sizeof(*hdr));
396 ret = read(f->fd, hdr, sizeof(*hdr));
400 /* record if an io error occurred */
401 if (hdr->info & SG_INFO_CHECK)
404 return FIO_Q_COMPLETED;
410 static enum fio_q_status fio_sgio_doio(struct thread_data *td,
411 struct io_u *io_u, int do_sync)
413 struct fio_file *f = io_u->file;
414 enum fio_q_status ret;
416 if (f->filetype == FIO_TYPE_BLOCK) {
417 ret = fio_sgio_ioctl_doio(td, f, io_u);
418 td_verror(td, io_u->error, __func__);
420 ret = fio_sgio_rw_doio(f, io_u, do_sync);
422 td_verror(td, io_u->error, __func__);
428 static void fio_sgio_rw_lba(struct sg_io_hdr *hdr, unsigned long long lba,
429 unsigned long long nr_blocks)
431 if (lba < MAX_10B_LBA) {
432 hdr->cmdp[2] = (unsigned char) ((lba >> 24) & 0xff);
433 hdr->cmdp[3] = (unsigned char) ((lba >> 16) & 0xff);
434 hdr->cmdp[4] = (unsigned char) ((lba >> 8) & 0xff);
435 hdr->cmdp[5] = (unsigned char) (lba & 0xff);
436 hdr->cmdp[7] = (unsigned char) ((nr_blocks >> 8) & 0xff);
437 hdr->cmdp[8] = (unsigned char) (nr_blocks & 0xff);
439 hdr->cmdp[2] = (unsigned char) ((lba >> 56) & 0xff);
440 hdr->cmdp[3] = (unsigned char) ((lba >> 48) & 0xff);
441 hdr->cmdp[4] = (unsigned char) ((lba >> 40) & 0xff);
442 hdr->cmdp[5] = (unsigned char) ((lba >> 32) & 0xff);
443 hdr->cmdp[6] = (unsigned char) ((lba >> 24) & 0xff);
444 hdr->cmdp[7] = (unsigned char) ((lba >> 16) & 0xff);
445 hdr->cmdp[8] = (unsigned char) ((lba >> 8) & 0xff);
446 hdr->cmdp[9] = (unsigned char) (lba & 0xff);
447 hdr->cmdp[10] = (unsigned char) ((nr_blocks >> 32) & 0xff);
448 hdr->cmdp[11] = (unsigned char) ((nr_blocks >> 16) & 0xff);
449 hdr->cmdp[12] = (unsigned char) ((nr_blocks >> 8) & 0xff);
450 hdr->cmdp[13] = (unsigned char) (nr_blocks & 0xff);
456 static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
458 struct sg_io_hdr *hdr = &io_u->hdr;
459 struct sg_options *o = td->eo;
460 struct sgio_data *sd = td->io_ops_data;
461 unsigned long long nr_blocks, lba;
464 if (io_u->xfer_buflen & (sd->bs - 1)) {
465 log_err("read/write not sector aligned\n");
469 nr_blocks = io_u->xfer_buflen / sd->bs;
470 lba = io_u->offset / sd->bs;
472 if (io_u->ddir == DDIR_READ) {
473 sgio_hdr_init(sd, hdr, io_u, 1);
475 hdr->dxfer_direction = SG_DXFER_FROM_DEV;
476 if (lba < MAX_10B_LBA)
477 hdr->cmdp[0] = 0x28; // read(10)
479 hdr->cmdp[0] = 0x88; // read(16)
482 hdr->cmdp[1] |= 0x08;
484 fio_sgio_rw_lba(hdr, lba, nr_blocks);
486 } else if (io_u->ddir == DDIR_WRITE) {
487 sgio_hdr_init(sd, hdr, io_u, 1);
489 hdr->dxfer_direction = SG_DXFER_TO_DEV;
490 switch(o->write_mode) {
492 if (lba < MAX_10B_LBA)
493 hdr->cmdp[0] = 0x2a; // write(10)
495 hdr->cmdp[0] = 0x8a; // write(16)
497 hdr->cmdp[1] |= 0x08;
499 case FIO_SG_WRITE_VERIFY:
500 if (lba < MAX_10B_LBA)
501 hdr->cmdp[0] = 0x2e; // write and verify(10)
503 hdr->cmdp[0] = 0x8e; // write and verify(16)
505 // BYTCHK is disabled by virtue of the memset in sgio_hdr_init
506 case FIO_SG_WRITE_SAME:
507 hdr->dxfer_len = sd->bs;
508 if (lba < MAX_10B_LBA)
509 hdr->cmdp[0] = 0x41; // write same(10)
511 hdr->cmdp[0] = 0x93; // write same(16)
515 fio_sgio_rw_lba(hdr, lba, nr_blocks);
517 } else if (io_u->ddir == DDIR_TRIM) {
518 struct sgio_trim *st;
520 if (sd->current_queue == -1) {
521 sgio_hdr_init(sd, hdr, io_u, 0);
524 hdr->dxfer_direction = SG_DXFER_TO_DEV;
525 hdr->cmdp[0] = 0x42; // unmap
526 sd->current_queue = io_u->index;
527 st = sd->trim_queues[sd->current_queue];
528 hdr->dxferp = st->unmap_param;
529 assert(sd->trim_queues[io_u->index]->unmap_range_count == 0);
530 dprint(FD_IO, "sg: creating new queue based on io_u %d\n", io_u->index);
533 st = sd->trim_queues[sd->current_queue];
535 dprint(FD_IO, "sg: adding io_u %d to trim queue %d\n", io_u->index, sd->current_queue);
536 st->trim_io_us[st->unmap_range_count] = io_u;
537 sd->trim_queue_map[io_u->index] = sd->current_queue;
539 offset = 8 + 16 * st->unmap_range_count;
540 st->unmap_param[offset] = (unsigned char) ((lba >> 56) & 0xff);
541 st->unmap_param[offset+1] = (unsigned char) ((lba >> 48) & 0xff);
542 st->unmap_param[offset+2] = (unsigned char) ((lba >> 40) & 0xff);
543 st->unmap_param[offset+3] = (unsigned char) ((lba >> 32) & 0xff);
544 st->unmap_param[offset+4] = (unsigned char) ((lba >> 24) & 0xff);
545 st->unmap_param[offset+5] = (unsigned char) ((lba >> 16) & 0xff);
546 st->unmap_param[offset+6] = (unsigned char) ((lba >> 8) & 0xff);
547 st->unmap_param[offset+7] = (unsigned char) (lba & 0xff);
548 st->unmap_param[offset+8] = (unsigned char) ((nr_blocks >> 32) & 0xff);
549 st->unmap_param[offset+9] = (unsigned char) ((nr_blocks >> 16) & 0xff);
550 st->unmap_param[offset+10] = (unsigned char) ((nr_blocks >> 8) & 0xff);
551 st->unmap_param[offset+11] = (unsigned char) (nr_blocks & 0xff);
553 st->unmap_range_count++;
555 } else if (ddir_sync(io_u->ddir)) {
556 sgio_hdr_init(sd, hdr, io_u, 0);
557 hdr->dxfer_direction = SG_DXFER_NONE;
558 if (lba < MAX_10B_LBA)
559 hdr->cmdp[0] = 0x35; // synccache(10)
561 hdr->cmdp[0] = 0x91; // synccache(16)
568 static void fio_sgio_unmap_setup(struct sg_io_hdr *hdr, struct sgio_trim *st)
570 hdr->dxfer_len = st->unmap_range_count * 16 + 8;
571 hdr->cmdp[7] = (unsigned char) (((st->unmap_range_count * 16 + 8) >> 8) & 0xff);
572 hdr->cmdp[8] = (unsigned char) ((st->unmap_range_count * 16 + 8) & 0xff);
574 st->unmap_param[0] = (unsigned char) (((16 * st->unmap_range_count + 6) >> 8) & 0xff);
575 st->unmap_param[1] = (unsigned char) ((16 * st->unmap_range_count + 6) & 0xff);
576 st->unmap_param[2] = (unsigned char) (((16 * st->unmap_range_count) >> 8) & 0xff);
577 st->unmap_param[3] = (unsigned char) ((16 * st->unmap_range_count) & 0xff);
582 static enum fio_q_status fio_sgio_queue(struct thread_data *td,
585 struct sg_io_hdr *hdr = &io_u->hdr;
586 struct sgio_data *sd = td->io_ops_data;
587 int ret, do_sync = 0;
589 fio_ro_check(td, io_u);
591 if (sgio_unbuffered(td) || ddir_sync(io_u->ddir))
594 if (io_u->ddir == DDIR_TRIM) {
595 if (do_sync || io_u->file->filetype == FIO_TYPE_BLOCK) {
596 struct sgio_trim *st = sd->trim_queues[sd->current_queue];
598 /* finish cdb setup for unmap because we are
599 ** doing unmap commands synchronously */
600 assert(st->unmap_range_count == 1);
601 assert(io_u == st->trim_io_us[0]);
604 fio_sgio_unmap_setup(hdr, st);
606 st->unmap_range_count = 0;
607 sd->current_queue = -1;
609 /* queue up trim ranges and submit in commit() */
613 ret = fio_sgio_doio(td, io_u, do_sync);
617 else if (hdr->status) {
618 io_u->resid = hdr->resid;
620 } else if (td->io_ops->commit != NULL) {
621 if (do_sync && !ddir_sync(io_u->ddir)) {
622 io_u_mark_submit(td, 1);
623 io_u_mark_complete(td, 1);
624 } else if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
625 io_u_mark_submit(td, 1);
626 io_u_queued(td, io_u);
631 td_verror(td, io_u->error, "xfer");
632 return FIO_Q_COMPLETED;
638 static int fio_sgio_commit(struct thread_data *td)
640 struct sgio_data *sd = td->io_ops_data;
641 struct sgio_trim *st;
643 struct sg_io_hdr *hdr;
648 if (sd->current_queue == -1)
651 st = sd->trim_queues[sd->current_queue];
652 io_u = st->trim_io_us[0];
655 fio_sgio_unmap_setup(hdr, st);
657 sd->current_queue = -1;
659 ret = fio_sgio_rw_doio(io_u->file, io_u, 0);
662 for (i = 0; i < st->unmap_range_count; i++)
663 st->trim_io_us[i]->error = errno;
664 else if (hdr->status)
665 for (i = 0; i < st->unmap_range_count; i++) {
666 st->trim_io_us[i]->resid = hdr->resid;
667 st->trim_io_us[i]->error = EIO;
670 if (fio_fill_issue_time(td)) {
671 fio_gettime(&now, NULL);
672 for (i = 0; i < st->unmap_range_count; i++) {
673 struct io_u *io_u = st->trim_io_us[i];
675 memcpy(&io_u->issue_time, &now, sizeof(now));
676 io_u_queued(td, io_u);
679 io_u_mark_submit(td, st->unmap_range_count);
683 td_verror(td, io_u->error, "xfer");
687 if (ret == FIO_Q_QUEUED)
693 static struct io_u *fio_sgio_event(struct thread_data *td, int event)
695 struct sgio_data *sd = td->io_ops_data;
697 return sd->events[event];
700 static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
701 unsigned long long *max_lba)
704 * need to do read capacity operation w/o benefit of sd or
705 * io_u structures, which are not initialized until later.
707 struct sg_io_hdr hdr;
708 unsigned char cmd[16];
709 unsigned char sb[64];
710 unsigned char buf[32]; // read capacity return
714 struct fio_file *f = td->files[0];
716 /* open file independent of rest of application */
717 fd = open(f->file_name, O_RDONLY);
721 memset(&hdr, 0, sizeof(hdr));
722 memset(cmd, 0, sizeof(cmd));
723 memset(sb, 0, sizeof(sb));
724 memset(buf, 0, sizeof(buf));
726 /* First let's try a 10 byte read capacity. */
727 hdr.interface_id = 'S';
731 hdr.mx_sb_len = sizeof(sb);
732 hdr.timeout = SCSI_TIMEOUT_MS;
733 hdr.cmdp[0] = 0x25; // Read Capacity(10)
734 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
736 hdr.dxfer_len = sizeof(buf);
738 ret = ioctl(fd, SG_IO, &hdr);
744 *bs = ((unsigned long) buf[4] << 24) | ((unsigned long) buf[5] << 16) |
745 ((unsigned long) buf[6] << 8) | (unsigned long) buf[7];
746 *max_lba = ((unsigned long) buf[0] << 24) | ((unsigned long) buf[1] << 16) |
747 ((unsigned long) buf[2] << 8) | (unsigned long) buf[3];
750 * If max lba masked by MAX_10B_LBA equals MAX_10B_LBA,
751 * then need to retry with 16 byte Read Capacity command.
753 if (*max_lba == MAX_10B_LBA) {
755 hdr.cmdp[0] = 0x9e; // service action
756 hdr.cmdp[1] = 0x10; // Read Capacity(16)
757 hdr.cmdp[10] = (unsigned char) ((sizeof(buf) >> 24) & 0xff);
758 hdr.cmdp[11] = (unsigned char) ((sizeof(buf) >> 16) & 0xff);
759 hdr.cmdp[12] = (unsigned char) ((sizeof(buf) >> 8) & 0xff);
760 hdr.cmdp[13] = (unsigned char) (sizeof(buf) & 0xff);
762 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
764 hdr.dxfer_len = sizeof(buf);
766 ret = ioctl(fd, SG_IO, &hdr);
772 /* record if an io error occurred */
773 if (hdr.info & SG_INFO_CHECK)
774 td_verror(td, EIO, "fio_sgio_read_capacity");
776 *bs = (buf[8] << 24) | (buf[9] << 16) | (buf[10] << 8) | buf[11];
777 *max_lba = ((unsigned long long)buf[0] << 56) |
778 ((unsigned long long)buf[1] << 48) |
779 ((unsigned long long)buf[2] << 40) |
780 ((unsigned long long)buf[3] << 32) |
781 ((unsigned long long)buf[4] << 24) |
782 ((unsigned long long)buf[5] << 16) |
783 ((unsigned long long)buf[6] << 8) |
784 (unsigned long long)buf[7];
791 static void fio_sgio_cleanup(struct thread_data *td)
793 struct sgio_data *sd = td->io_ops_data;
802 free(sd->trim_queue_map);
804 for (i = 0; i < td->o.iodepth; i++) {
805 free(sd->trim_queues[i]->unmap_param);
806 free(sd->trim_queues[i]->trim_io_us);
807 free(sd->trim_queues[i]);
810 free(sd->trim_queues);
815 static int fio_sgio_init(struct thread_data *td)
817 struct sgio_data *sd;
818 struct sgio_trim *st;
821 sd = calloc(1, sizeof(*sd));
822 sd->cmds = calloc(td->o.iodepth, sizeof(struct sgio_cmd));
823 sd->sgbuf = calloc(td->o.iodepth, sizeof(struct sg_io_hdr));
824 sd->events = calloc(td->o.iodepth, sizeof(struct io_u *));
825 sd->pfds = calloc(td->o.nr_files, sizeof(struct pollfd));
826 sd->fd_flags = calloc(td->o.nr_files, sizeof(int));
827 sd->type_checked = 0;
829 sd->trim_queues = calloc(td->o.iodepth, sizeof(struct sgio_trim *));
830 sd->current_queue = -1;
831 sd->trim_queue_map = calloc(td->o.iodepth, sizeof(int));
832 for (i = 0; i < td->o.iodepth; i++) {
833 sd->trim_queues[i] = calloc(1, sizeof(struct sgio_trim));
834 st = sd->trim_queues[i];
835 st->unmap_param = calloc(td->o.iodepth + 1, sizeof(char[16]));
836 st->unmap_range_count = 0;
837 st->trim_io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
840 td->io_ops_data = sd;
843 * we want to do it, regardless of whether odirect is set or not
845 td->o.override_sync = 1;
849 static int fio_sgio_type_check(struct thread_data *td, struct fio_file *f)
851 struct sgio_data *sd = td->io_ops_data;
853 unsigned long long max_lba = 0;
855 if (f->filetype == FIO_TYPE_BLOCK) {
856 if (ioctl(f->fd, BLKSSZGET, &bs) < 0) {
857 td_verror(td, errno, "ioctl");
860 } else if (f->filetype == FIO_TYPE_CHAR) {
863 if (ioctl(f->fd, SG_GET_VERSION_NUM, &version) < 0) {
864 td_verror(td, errno, "ioctl");
868 ret = fio_sgio_read_capacity(td, &bs, &max_lba);
870 td_verror(td, td->error, "fio_sgio_read_capacity");
871 log_err("ioengine sg unable to read capacity successfully\n");
875 td_verror(td, EINVAL, "wrong file type");
876 log_err("ioengine sg only works on block or character devices\n");
881 // Determine size of commands needed based on max_lba
882 if (max_lba >= MAX_10B_LBA) {
883 dprint(FD_IO, "sgio_type_check: using 16 byte read/write "
884 "commands for lba above 0x%016llx/0x%016llx\n",
885 MAX_10B_LBA, max_lba);
888 if (f->filetype == FIO_TYPE_BLOCK) {
889 td->io_ops->getevents = NULL;
890 td->io_ops->event = NULL;
891 td->io_ops->commit = NULL;
893 ** Setting these functions to null may cause problems
894 ** with filename=/dev/sda:/dev/sg0 since we are only
895 ** considering a single file
898 sd->type_checked = 1;
903 static int fio_sgio_open(struct thread_data *td, struct fio_file *f)
905 struct sgio_data *sd = td->io_ops_data;
908 ret = generic_open_file(td, f);
912 if (sd && !sd->type_checked && fio_sgio_type_check(td, f)) {
913 ret = generic_close_file(td, f);
921 * Build an error string with details about the driver, host or scsi
922 * error contained in the sg header Caller will use as necessary.
924 static char *fio_sgio_errdetails(struct io_u *io_u)
926 struct sg_io_hdr *hdr = &io_u->hdr;
927 #define MAXERRDETAIL 1024
928 #define MAXMSGCHUNK 128
929 char *msg, msgchunk[MAXMSGCHUNK];
932 msg = calloc(1, MAXERRDETAIL);
936 * can't seem to find sg_err.h, so I'll just echo the define values
937 * so others can search on internet to find clearer clues of meaning.
939 if (hdr->info & SG_INFO_CHECK) {
940 if (hdr->host_status) {
941 snprintf(msgchunk, MAXMSGCHUNK, "SG Host Status: 0x%02x; ", hdr->host_status);
942 strlcat(msg, msgchunk, MAXERRDETAIL);
943 switch (hdr->host_status) {
945 strlcat(msg, "SG_ERR_DID_NO_CONNECT", MAXERRDETAIL);
948 strlcat(msg, "SG_ERR_DID_BUS_BUSY", MAXERRDETAIL);
951 strlcat(msg, "SG_ERR_DID_TIME_OUT", MAXERRDETAIL);
954 strlcat(msg, "SG_ERR_DID_BAD_TARGET", MAXERRDETAIL);
957 strlcat(msg, "SG_ERR_DID_ABORT", MAXERRDETAIL);
960 strlcat(msg, "SG_ERR_DID_PARITY", MAXERRDETAIL);
963 strlcat(msg, "SG_ERR_DID_ERROR (internal error)", MAXERRDETAIL);
966 strlcat(msg, "SG_ERR_DID_RESET", MAXERRDETAIL);
969 strlcat(msg, "SG_ERR_DID_BAD_INTR (unexpected)", MAXERRDETAIL);
972 strlcat(msg, "SG_ERR_DID_PASSTHROUGH", MAXERRDETAIL);
975 strlcat(msg, "SG_ERR_DID_SOFT_ERROR (driver retry?)", MAXERRDETAIL);
978 strlcat(msg, "SG_ERR_DID_IMM_RETRY", MAXERRDETAIL);
981 strlcat(msg, "SG_ERR_DID_REQUEUE", MAXERRDETAIL);
984 strlcat(msg, "SG_ERR_DID_TRANSPORT_DISRUPTED", MAXERRDETAIL);
987 strlcat(msg, "SG_ERR_DID_TRANSPORT_FAILFAST", MAXERRDETAIL);
990 strlcat(msg, "SG_ERR_DID_TARGET_FAILURE", MAXERRDETAIL);
993 strlcat(msg, "SG_ERR_DID_NEXUS_FAILURE", MAXERRDETAIL);
996 strlcat(msg, "SG_ERR_DID_ALLOC_FAILURE", MAXERRDETAIL);
999 strlcat(msg, "SG_ERR_DID_MEDIUM_ERROR", MAXERRDETAIL);
1002 strlcat(msg, "Unknown", MAXERRDETAIL);
1005 strlcat(msg, ". ", MAXERRDETAIL);
1007 if (hdr->driver_status) {
1008 snprintf(msgchunk, MAXMSGCHUNK, "SG Driver Status: 0x%02x; ", hdr->driver_status);
1009 strlcat(msg, msgchunk, MAXERRDETAIL);
1010 switch (hdr->driver_status & 0x0F) {
1012 strlcat(msg, "SG_ERR_DRIVER_BUSY", MAXERRDETAIL);
1015 strlcat(msg, "SG_ERR_DRIVER_SOFT", MAXERRDETAIL);
1018 strlcat(msg, "SG_ERR_DRIVER_MEDIA", MAXERRDETAIL);
1021 strlcat(msg, "SG_ERR_DRIVER_ERROR", MAXERRDETAIL);
1024 strlcat(msg, "SG_ERR_DRIVER_INVALID", MAXERRDETAIL);
1027 strlcat(msg, "SG_ERR_DRIVER_TIMEOUT", MAXERRDETAIL);
1030 strlcat(msg, "SG_ERR_DRIVER_HARD", MAXERRDETAIL);
1033 strlcat(msg, "SG_ERR_DRIVER_SENSE", MAXERRDETAIL);
1036 strlcat(msg, "Unknown", MAXERRDETAIL);
1039 strlcat(msg, "; ", MAXERRDETAIL);
1040 switch (hdr->driver_status & 0xF0) {
1042 strlcat(msg, "SG_ERR_SUGGEST_RETRY", MAXERRDETAIL);
1045 strlcat(msg, "SG_ERR_SUGGEST_ABORT", MAXERRDETAIL);
1048 strlcat(msg, "SG_ERR_SUGGEST_REMAP", MAXERRDETAIL);
1051 strlcat(msg, "SG_ERR_SUGGEST_DIE", MAXERRDETAIL);
1054 strlcat(msg, "SG_ERR_SUGGEST_SENSE", MAXERRDETAIL);
1057 strlcat(msg, ". ", MAXERRDETAIL);
1060 snprintf(msgchunk, MAXMSGCHUNK, "SG SCSI Status: 0x%02x; ", hdr->status);
1061 strlcat(msg, msgchunk, MAXERRDETAIL);
1062 // SCSI 3 status codes
1063 switch (hdr->status) {
1065 strlcat(msg, "CHECK_CONDITION", MAXERRDETAIL);
1068 strlcat(msg, "CONDITION_MET", MAXERRDETAIL);
1071 strlcat(msg, "BUSY", MAXERRDETAIL);
1074 strlcat(msg, "INTERMEDIATE", MAXERRDETAIL);
1077 strlcat(msg, "INTERMEDIATE_CONDITION_MET", MAXERRDETAIL);
1080 strlcat(msg, "RESERVATION_CONFLICT", MAXERRDETAIL);
1083 strlcat(msg, "COMMAND_TERMINATED", MAXERRDETAIL);
1086 strlcat(msg, "TASK_SET_FULL", MAXERRDETAIL);
1089 strlcat(msg, "ACA_ACTIVE", MAXERRDETAIL);
1092 strlcat(msg, "TASK_ABORTED", MAXERRDETAIL);
1095 strlcat(msg, "Unknown", MAXERRDETAIL);
1098 strlcat(msg, ". ", MAXERRDETAIL);
1100 if (hdr->sb_len_wr) {
1101 snprintf(msgchunk, MAXMSGCHUNK, "Sense Data (%d bytes):", hdr->sb_len_wr);
1102 strlcat(msg, msgchunk, MAXERRDETAIL);
1103 for (i = 0; i < hdr->sb_len_wr; i++) {
1104 snprintf(msgchunk, MAXMSGCHUNK, " %02x", hdr->sbp[i]);
1105 strlcat(msg, msgchunk, MAXERRDETAIL);
1107 strlcat(msg, ". ", MAXERRDETAIL);
1109 if (hdr->resid != 0) {
1110 snprintf(msgchunk, MAXMSGCHUNK, "SG Driver: %d bytes out of %d not transferred. ", hdr->resid, hdr->dxfer_len);
1111 strlcat(msg, msgchunk, MAXERRDETAIL);
1115 if (!(hdr->info & SG_INFO_CHECK) && !strlen(msg))
1116 strncpy(msg, "SG Driver did not report a Host, Driver or Device check",
1123 * get max file size from read capacity.
1125 static int fio_sgio_get_file_size(struct thread_data *td, struct fio_file *f)
1128 * get_file_size is being called even before sgio_init is
1129 * called, so none of the sg_io structures are
1130 * initialized in the thread_data yet. So we need to do the
1131 * ReadCapacity without any of those helpers. One of the effects
1132 * is that ReadCapacity may get called 4 times on each open:
1133 * readcap(10) followed by readcap(16) if needed - just to get
1134 * the file size after the init occurs - it will be called
1135 * again when "type_check" is called during structure
1136 * initialization I'm not sure how to prevent this little
1139 unsigned int bs = 0;
1140 unsigned long long max_lba = 0;
1143 if (fio_file_size_known(f))
1146 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) {
1147 td_verror(td, EINVAL, "wrong file type");
1148 log_err("ioengine sg only works on block or character devices\n");
1152 ret = fio_sgio_read_capacity(td, &bs, &max_lba);
1154 td_verror(td, td->error, "fio_sgio_read_capacity");
1155 log_err("ioengine sg unable to successfully execute read capacity to get block size and maximum lba\n");
1159 f->real_file_size = (max_lba + 1) * bs;
1160 fio_file_set_size_known(f);
1165 static struct ioengine_ops ioengine = {
1167 .version = FIO_IOOPS_VERSION,
1168 .init = fio_sgio_init,
1169 .prep = fio_sgio_prep,
1170 .queue = fio_sgio_queue,
1171 .commit = fio_sgio_commit,
1172 .getevents = fio_sgio_getevents,
1173 .errdetails = fio_sgio_errdetails,
1174 .event = fio_sgio_event,
1175 .cleanup = fio_sgio_cleanup,
1176 .open_file = fio_sgio_open,
1177 .close_file = generic_close_file,
1178 .get_file_size = fio_sgio_get_file_size,
1179 .flags = FIO_SYNCIO | FIO_RAWIO,
1181 .option_struct_size = sizeof(struct sg_options)
1184 #else /* FIO_HAVE_SGIO */
1187 * When we have a proper configure system in place, we simply wont build
1188 * and install this io engine. For now install a crippled version that
1189 * just complains and fails to load.
1191 static int fio_sgio_init(struct thread_data fio_unused *td)
1193 log_err("fio: ioengine sg not available\n");
1197 static struct ioengine_ops ioengine = {
1199 .version = FIO_IOOPS_VERSION,
1200 .init = fio_sgio_init,
1205 static void fio_init fio_sgio_register(void)
1207 register_ioengine(&ioengine);
1210 static void fio_exit fio_sgio_unregister(void)
1212 unregister_ioengine(&ioengine);