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"
63 #ifndef SGV4_FLAG_HIPRI
64 #define SGV4_FLAG_HIPRI 0x800
69 FIO_SG_WRITE_VERIFY = 2,
77 unsigned int writefua;
78 unsigned int write_mode;
81 static struct fio_option options[] = {
84 .lname = "High Priority",
85 .type = FIO_OPT_STR_SET,
86 .off1 = offsetof(struct sg_options, hipri),
87 .help = "Use polled IO completions",
88 .category = FIO_OPT_C_ENGINE,
89 .group = FIO_OPT_G_SG,
93 .lname = "sg engine read fua flag support",
95 .off1 = offsetof(struct sg_options, readfua),
96 .help = "Set FUA flag (force unit access) for all Read operations",
98 .category = FIO_OPT_C_ENGINE,
99 .group = FIO_OPT_G_SG,
103 .lname = "sg engine write fua flag support",
104 .type = FIO_OPT_BOOL,
105 .off1 = offsetof(struct sg_options, writefua),
106 .help = "Set FUA flag (force unit access) for all Write operations",
108 .category = FIO_OPT_C_ENGINE,
109 .group = FIO_OPT_G_SG,
112 .name = "sg_write_mode",
113 .lname = "specify sg write mode",
115 .off1 = offsetof(struct sg_options, write_mode),
116 .help = "Specify SCSI WRITE mode",
120 .oval = FIO_SG_WRITE,
121 .help = "Issue standard SCSI WRITE commands",
124 .oval = FIO_SG_WRITE_VERIFY,
125 .help = "Issue SCSI WRITE AND VERIFY commands",
128 .oval = FIO_SG_WRITE_SAME,
129 .help = "Issue SCSI WRITE SAME commands",
132 .category = FIO_OPT_C_ENGINE,
133 .group = FIO_OPT_G_SG,
140 #define MAX_10B_LBA 0xFFFFFFFFULL
141 #define SCSI_TIMEOUT_MS 30000 // 30 second timeout; currently no method to override
142 #define MAX_SB 64 // sense block maximum return size
144 #define FIO_SGIO_DEBUG
148 unsigned char cdb[16]; // enhanced from 10 to support 16 byte commands
149 unsigned char sb[MAX_SB]; // add sense block to commands
154 uint8_t *unmap_param;
155 unsigned int unmap_range_count;
156 struct io_u **trim_io_us;
160 struct sgio_cmd *cmds;
161 struct io_u **events;
167 struct sgio_trim **trim_queues;
169 #ifdef FIO_SGIO_DEBUG
170 unsigned int *trim_queue_map;
174 static inline uint32_t sgio_get_be32(uint8_t *buf)
176 return be32_to_cpu(*((uint32_t *) buf));
179 static inline uint64_t sgio_get_be64(uint8_t *buf)
181 return be64_to_cpu(*((uint64_t *) buf));
184 static inline void sgio_set_be16(uint16_t val, uint8_t *buf)
186 uint16_t t = cpu_to_be16(val);
188 memcpy(buf, &t, sizeof(uint16_t));
191 static inline void sgio_set_be32(uint32_t val, uint8_t *buf)
193 uint32_t t = cpu_to_be32(val);
195 memcpy(buf, &t, sizeof(uint32_t));
198 static inline void sgio_set_be64(uint64_t val, uint8_t *buf)
200 uint64_t t = cpu_to_be64(val);
202 memcpy(buf, &t, sizeof(uint64_t));
205 static inline bool sgio_unbuffered(struct thread_data *td)
207 return (td->o.odirect || td->o.sync_io);
210 static void sgio_hdr_init(struct sgio_data *sd, struct sg_io_hdr *hdr,
211 struct io_u *io_u, int fs)
213 struct sgio_cmd *sc = &sd->cmds[io_u->index];
215 memset(hdr, 0, sizeof(*hdr));
216 memset(sc->cdb, 0, sizeof(sc->cdb));
218 hdr->interface_id = 'S';
220 hdr->cmd_len = sizeof(sc->cdb);
222 hdr->mx_sb_len = sizeof(sc->sb);
223 hdr->pack_id = io_u->index;
225 hdr->timeout = SCSI_TIMEOUT_MS;
228 hdr->dxferp = io_u->xfer_buf;
229 hdr->dxfer_len = io_u->xfer_buflen;
233 static int pollin_events(struct pollfd *pfds, int fds)
237 for (i = 0; i < fds; i++)
238 if (pfds[i].revents & POLLIN)
244 static int sg_fd_read(int fd, void *data, size_t size)
251 ret = read(fd, data, size);
253 if (errno == EAGAIN || errno == EINTR)
273 static int fio_sgio_getevents(struct thread_data *td, unsigned int min,
275 const struct timespec fio_unused *t)
277 struct sgio_data *sd = td->io_ops_data;
278 int left = max, eventNum, ret, r = 0, trims = 0;
279 void *buf = sd->sgbuf;
280 unsigned int i, j, events;
285 * Fill in the file descriptors
287 for_each_file(td, f, i) {
289 * don't block for min events == 0
292 sd->fd_flags[i] = fio_set_fd_nonblocking(f->fd, "sg");
294 sd->fd_flags[i] = -1;
296 sd->pfds[i].fd = f->fd;
297 sd->pfds[i].events = POLLIN;
301 ** There are two counters here:
302 ** - number of SCSI commands completed
303 ** - number of io_us completed
305 ** These are the same with reads and writes, but
306 ** could differ with trim/unmap commands because
307 ** a single unmap can include multiple io_us
313 dprint(FD_IO, "sgio_getevents: sd %p: min=%d, max=%d, left=%d\n", sd, min, max, left);
319 ret = poll(sd->pfds, td->o.nr_files, -1);
323 td_verror(td, errno, "poll");
328 if (pollin_events(sd->pfds, td->o.nr_files))
338 for_each_file(td, f, i) {
339 for (eventNum = 0; eventNum < left; eventNum++) {
340 ret = sg_fd_read(f->fd, p, sizeof(struct sg_io_hdr));
341 dprint(FD_IO, "sgio_getevents: sg_fd_read ret: %d\n", ret);
344 td_verror(td, r, "sg_read");
347 io_u = ((struct sg_io_hdr *)p)->usr_ptr;
348 if (io_u->ddir == DDIR_TRIM) {
349 events += sd->trim_queues[io_u->index]->unmap_range_count;
350 eventNum += sd->trim_queues[io_u->index]->unmap_range_count - 1;
354 p += sizeof(struct sg_io_hdr);
355 dprint(FD_IO, "sgio_getevents: events: %d, eventNum: %d, left: %d\n", events, eventNum, left);
359 if (r < 0 && !events)
369 for (i = 0; i < events; i++) {
370 struct sg_io_hdr *hdr = (struct sg_io_hdr *) buf + i;
371 sd->events[i + trims] = hdr->usr_ptr;
372 io_u = (struct io_u *)(hdr->usr_ptr);
374 if (hdr->info & SG_INFO_CHECK) {
375 /* record if an io error occurred, ignore resid */
376 memcpy(&io_u->hdr, hdr, sizeof(struct sg_io_hdr));
377 sd->events[i + trims]->error = EIO;
380 if (io_u->ddir == DDIR_TRIM) {
381 struct sgio_trim *st = sd->trim_queues[io_u->index];
382 #ifdef FIO_SGIO_DEBUG
383 assert(st->trim_io_us[0] == io_u);
384 assert(sd->trim_queue_map[io_u->index] == io_u->index);
385 dprint(FD_IO, "sgio_getevents: reaping %d io_us from trim queue %d\n", st->unmap_range_count, io_u->index);
386 dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", io_u->index, i+trims);
388 for (j = 1; j < st->unmap_range_count; j++) {
390 sd->events[i + trims] = st->trim_io_us[j];
391 #ifdef FIO_SGIO_DEBUG
392 dprint(FD_IO, "sgio_getevents: reaped io_u %d and stored in events[%d]\n", st->trim_io_us[j]->index, i+trims);
393 assert(sd->trim_queue_map[st->trim_io_us[j]->index] == io_u->index);
395 if (hdr->info & SG_INFO_CHECK) {
396 /* record if an io error occurred, ignore resid */
397 memcpy(&st->trim_io_us[j]->hdr, hdr, sizeof(struct sg_io_hdr));
398 sd->events[i + trims]->error = EIO;
401 events -= st->unmap_range_count - 1;
402 st->unmap_range_count = 0;
408 for_each_file(td, f, i) {
409 if (sd->fd_flags[i] == -1)
412 if (fcntl(f->fd, F_SETFL, sd->fd_flags[i]) < 0)
413 log_err("fio: sg failed to restore fcntl flags: %s\n", strerror(errno));
420 static enum fio_q_status fio_sgio_ioctl_doio(struct thread_data *td,
424 struct sgio_data *sd = td->io_ops_data;
425 struct sg_io_hdr *hdr = &io_u->hdr;
428 sd->events[0] = io_u;
430 ret = ioctl(f->fd, SG_IO, hdr);
434 /* record if an io error occurred */
435 if (hdr->info & SG_INFO_CHECK)
438 return FIO_Q_COMPLETED;
441 static enum fio_q_status fio_sgio_rw_doio(struct thread_data *td,
443 struct io_u *io_u, int do_sync)
445 struct sg_io_hdr *hdr = &io_u->hdr;
448 ret = write(f->fd, hdr, sizeof(*hdr));
454 * We can't just read back the first command that completes
455 * and assume it's the one we need, it could be any command
461 ret = read(f->fd, hdr, sizeof(*hdr));
465 __io_u = hdr->usr_ptr;
467 /* record if an io error occurred */
468 if (hdr->info & SG_INFO_CHECK)
474 if (io_u_sync_complete(td, __io_u)) {
480 return FIO_Q_COMPLETED;
486 static enum fio_q_status fio_sgio_doio(struct thread_data *td,
487 struct io_u *io_u, int do_sync)
489 struct fio_file *f = io_u->file;
490 enum fio_q_status ret;
492 if (f->filetype == FIO_TYPE_BLOCK) {
493 ret = fio_sgio_ioctl_doio(td, f, io_u);
495 td_verror(td, io_u->error, __func__);
497 ret = fio_sgio_rw_doio(td, f, io_u, do_sync);
498 if (io_u->error && do_sync)
499 td_verror(td, io_u->error, __func__);
505 static void fio_sgio_rw_lba(struct sg_io_hdr *hdr, unsigned long long lba,
506 unsigned long long nr_blocks)
508 if (lba < MAX_10B_LBA) {
509 sgio_set_be32((uint32_t) lba, &hdr->cmdp[2]);
510 sgio_set_be16((uint16_t) nr_blocks, &hdr->cmdp[7]);
512 sgio_set_be64(lba, &hdr->cmdp[2]);
513 sgio_set_be32((uint32_t) nr_blocks, &hdr->cmdp[10]);
519 static int fio_sgio_prep(struct thread_data *td, struct io_u *io_u)
521 struct sg_io_hdr *hdr = &io_u->hdr;
522 struct sg_options *o = td->eo;
523 struct sgio_data *sd = td->io_ops_data;
524 unsigned long long nr_blocks, lba;
527 if (io_u->xfer_buflen & (sd->bs - 1)) {
528 log_err("read/write not sector aligned\n");
532 nr_blocks = io_u->xfer_buflen / sd->bs;
533 lba = io_u->offset / sd->bs;
535 if (io_u->ddir == DDIR_READ) {
536 sgio_hdr_init(sd, hdr, io_u, 1);
538 hdr->dxfer_direction = SG_DXFER_FROM_DEV;
539 if (lba < MAX_10B_LBA)
540 hdr->cmdp[0] = 0x28; // read(10)
542 hdr->cmdp[0] = 0x88; // read(16)
545 hdr->flags |= SGV4_FLAG_HIPRI;
547 hdr->cmdp[1] |= 0x08;
549 fio_sgio_rw_lba(hdr, lba, nr_blocks);
551 } else if (io_u->ddir == DDIR_WRITE) {
552 sgio_hdr_init(sd, hdr, io_u, 1);
554 hdr->dxfer_direction = SG_DXFER_TO_DEV;
555 switch(o->write_mode) {
557 if (lba < MAX_10B_LBA)
558 hdr->cmdp[0] = 0x2a; // write(10)
560 hdr->cmdp[0] = 0x8a; // write(16)
562 hdr->flags |= SGV4_FLAG_HIPRI;
564 hdr->cmdp[1] |= 0x08;
566 case FIO_SG_WRITE_VERIFY:
567 if (lba < MAX_10B_LBA)
568 hdr->cmdp[0] = 0x2e; // write and verify(10)
570 hdr->cmdp[0] = 0x8e; // write and verify(16)
572 // BYTCHK is disabled by virtue of the memset in sgio_hdr_init
573 case FIO_SG_WRITE_SAME:
574 hdr->dxfer_len = sd->bs;
575 if (lba < MAX_10B_LBA)
576 hdr->cmdp[0] = 0x41; // write same(10)
578 hdr->cmdp[0] = 0x93; // write same(16)
582 fio_sgio_rw_lba(hdr, lba, nr_blocks);
584 } else if (io_u->ddir == DDIR_TRIM) {
585 struct sgio_trim *st;
587 if (sd->current_queue == -1) {
588 sgio_hdr_init(sd, hdr, io_u, 0);
591 hdr->dxfer_direction = SG_DXFER_TO_DEV;
592 hdr->cmdp[0] = 0x42; // unmap
593 sd->current_queue = io_u->index;
594 st = sd->trim_queues[sd->current_queue];
595 hdr->dxferp = st->unmap_param;
596 #ifdef FIO_SGIO_DEBUG
597 assert(sd->trim_queues[io_u->index]->unmap_range_count == 0);
598 dprint(FD_IO, "sg: creating new queue based on io_u %d\n", io_u->index);
602 st = sd->trim_queues[sd->current_queue];
604 dprint(FD_IO, "sg: adding io_u %d to trim queue %d\n", io_u->index, sd->current_queue);
605 st->trim_io_us[st->unmap_range_count] = io_u;
606 #ifdef FIO_SGIO_DEBUG
607 sd->trim_queue_map[io_u->index] = sd->current_queue;
610 offset = 8 + 16 * st->unmap_range_count;
611 sgio_set_be64(lba, &st->unmap_param[offset]);
612 sgio_set_be32((uint32_t) nr_blocks, &st->unmap_param[offset + 8]);
614 st->unmap_range_count++;
616 } else if (ddir_sync(io_u->ddir)) {
617 sgio_hdr_init(sd, hdr, io_u, 0);
618 hdr->dxfer_direction = SG_DXFER_NONE;
619 if (lba < MAX_10B_LBA)
620 hdr->cmdp[0] = 0x35; // synccache(10)
622 hdr->cmdp[0] = 0x91; // synccache(16)
629 static void fio_sgio_unmap_setup(struct sg_io_hdr *hdr, struct sgio_trim *st)
631 uint16_t cnt = st->unmap_range_count * 16;
633 hdr->dxfer_len = cnt + 8;
634 sgio_set_be16(cnt + 8, &hdr->cmdp[7]);
635 sgio_set_be16(cnt + 6, st->unmap_param);
636 sgio_set_be16(cnt, &st->unmap_param[2]);
641 static enum fio_q_status fio_sgio_queue(struct thread_data *td,
644 struct sg_io_hdr *hdr = &io_u->hdr;
645 struct sgio_data *sd = td->io_ops_data;
646 int ret, do_sync = 0;
648 fio_ro_check(td, io_u);
650 if (sgio_unbuffered(td) || ddir_sync(io_u->ddir))
653 if (io_u->ddir == DDIR_TRIM) {
654 if (do_sync || io_u->file->filetype == FIO_TYPE_BLOCK) {
655 struct sgio_trim *st = sd->trim_queues[sd->current_queue];
657 /* finish cdb setup for unmap because we are
658 ** doing unmap commands synchronously */
659 #ifdef FIO_SGIO_DEBUG
660 assert(st->unmap_range_count == 1);
661 assert(io_u == st->trim_io_us[0]);
665 fio_sgio_unmap_setup(hdr, st);
667 st->unmap_range_count = 0;
668 sd->current_queue = -1;
670 /* queue up trim ranges and submit in commit() */
674 ret = fio_sgio_doio(td, io_u, do_sync);
678 else if (hdr->status) {
679 io_u->resid = hdr->resid;
681 } else if (td->io_ops->commit != NULL) {
682 if (do_sync && !ddir_sync(io_u->ddir)) {
683 io_u_mark_submit(td, 1);
684 io_u_mark_complete(td, 1);
685 } else if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
686 io_u_mark_submit(td, 1);
687 io_u_queued(td, io_u);
692 td_verror(td, io_u->error, "xfer");
693 return FIO_Q_COMPLETED;
699 static int fio_sgio_commit(struct thread_data *td)
701 struct sgio_data *sd = td->io_ops_data;
702 struct sgio_trim *st;
704 struct sg_io_hdr *hdr;
709 if (sd->current_queue == -1)
712 st = sd->trim_queues[sd->current_queue];
713 io_u = st->trim_io_us[0];
716 fio_sgio_unmap_setup(hdr, st);
718 sd->current_queue = -1;
720 ret = fio_sgio_rw_doio(td, io_u->file, io_u, 0);
722 if (ret < 0 || hdr->status) {
732 for (i = 0; i < st->unmap_range_count; i++) {
733 st->trim_io_us[i]->error = error;
734 clear_io_u(td, st->trim_io_us[i]);
736 st->trim_io_us[i]->resid = hdr->resid;
739 td_verror(td, error, "xfer");
743 if (fio_fill_issue_time(td)) {
744 fio_gettime(&now, NULL);
745 for (i = 0; i < st->unmap_range_count; i++) {
746 memcpy(&st->trim_io_us[i]->issue_time, &now, sizeof(now));
747 io_u_queued(td, io_u);
750 io_u_mark_submit(td, st->unmap_range_count);
755 static struct io_u *fio_sgio_event(struct thread_data *td, int event)
757 struct sgio_data *sd = td->io_ops_data;
759 return sd->events[event];
762 static int fio_sgio_read_capacity(struct thread_data *td, unsigned int *bs,
763 unsigned long long *max_lba)
766 * need to do read capacity operation w/o benefit of sd or
767 * io_u structures, which are not initialized until later.
769 struct sg_io_hdr hdr;
770 unsigned long long hlba;
771 unsigned int blksz = 0;
772 unsigned char cmd[16];
773 unsigned char sb[64];
774 unsigned char buf[32]; // read capacity return
778 struct fio_file *f = td->files[0];
780 /* open file independent of rest of application */
781 fd = open(f->file_name, O_RDONLY);
785 memset(&hdr, 0, sizeof(hdr));
786 memset(cmd, 0, sizeof(cmd));
787 memset(sb, 0, sizeof(sb));
788 memset(buf, 0, sizeof(buf));
790 /* First let's try a 10 byte read capacity. */
791 hdr.interface_id = 'S';
795 hdr.mx_sb_len = sizeof(sb);
796 hdr.timeout = SCSI_TIMEOUT_MS;
797 hdr.cmdp[0] = 0x25; // Read Capacity(10)
798 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
800 hdr.dxfer_len = sizeof(buf);
802 ret = ioctl(fd, SG_IO, &hdr);
808 if (hdr.info & SG_INFO_CHECK) {
809 /* RCAP(10) might be unsupported by device. Force RCAP(16) */
812 blksz = sgio_get_be32(&buf[4]);
813 hlba = sgio_get_be32(buf);
817 * If max lba masked by MAX_10B_LBA equals MAX_10B_LBA,
818 * then need to retry with 16 byte Read Capacity command.
820 if (hlba == MAX_10B_LBA) {
822 hdr.cmdp[0] = 0x9e; // service action
823 hdr.cmdp[1] = 0x10; // Read Capacity(16)
824 sgio_set_be32(sizeof(buf), &hdr.cmdp[10]);
826 hdr.dxfer_direction = SG_DXFER_FROM_DEV;
828 hdr.dxfer_len = sizeof(buf);
830 ret = ioctl(fd, SG_IO, &hdr);
836 /* record if an io error occurred */
837 if (hdr.info & SG_INFO_CHECK)
838 td_verror(td, EIO, "fio_sgio_read_capacity");
840 blksz = sgio_get_be32(&buf[8]);
841 hlba = sgio_get_be64(buf);
856 static void fio_sgio_cleanup(struct thread_data *td)
858 struct sgio_data *sd = td->io_ops_data;
867 #ifdef FIO_SGIO_DEBUG
868 free(sd->trim_queue_map);
871 for (i = 0; i < td->o.iodepth; i++) {
872 free(sd->trim_queues[i]->unmap_param);
873 free(sd->trim_queues[i]->trim_io_us);
874 free(sd->trim_queues[i]);
877 free(sd->trim_queues);
882 static int fio_sgio_init(struct thread_data *td)
884 struct sgio_data *sd;
885 struct sgio_trim *st;
886 struct sg_io_hdr *h3p;
889 sd = calloc(1, sizeof(*sd));
890 sd->cmds = calloc(td->o.iodepth, sizeof(struct sgio_cmd));
891 sd->sgbuf = calloc(td->o.iodepth, sizeof(struct sg_io_hdr));
892 sd->events = calloc(td->o.iodepth, sizeof(struct io_u *));
893 sd->pfds = calloc(td->o.nr_files, sizeof(struct pollfd));
894 sd->fd_flags = calloc(td->o.nr_files, sizeof(int));
895 sd->type_checked = 0;
897 sd->trim_queues = calloc(td->o.iodepth, sizeof(struct sgio_trim *));
898 sd->current_queue = -1;
899 #ifdef FIO_SGIO_DEBUG
900 sd->trim_queue_map = calloc(td->o.iodepth, sizeof(int));
902 for (i = 0, h3p = sd->sgbuf; i < td->o.iodepth; i++, ++h3p) {
903 sd->trim_queues[i] = calloc(1, sizeof(struct sgio_trim));
904 st = sd->trim_queues[i];
905 st->unmap_param = calloc(td->o.iodepth + 1, sizeof(char[16]));
906 st->unmap_range_count = 0;
907 st->trim_io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
908 h3p->interface_id = 'S';
911 td->io_ops_data = sd;
914 * we want to do it, regardless of whether odirect is set or not
916 td->o.override_sync = 1;
920 static int fio_sgio_type_check(struct thread_data *td, struct fio_file *f)
922 struct sgio_data *sd = td->io_ops_data;
924 unsigned long long max_lba = 0;
926 if (f->filetype == FIO_TYPE_BLOCK) {
927 if (ioctl(f->fd, BLKSSZGET, &bs) < 0) {
928 td_verror(td, errno, "ioctl");
931 } else if (f->filetype == FIO_TYPE_CHAR) {
934 if (ioctl(f->fd, SG_GET_VERSION_NUM, &version) < 0) {
935 td_verror(td, errno, "ioctl");
939 ret = fio_sgio_read_capacity(td, &bs, &max_lba);
941 td_verror(td, td->error, "fio_sgio_read_capacity");
942 log_err("ioengine sg unable to read capacity successfully\n");
946 td_verror(td, EINVAL, "wrong file type");
947 log_err("ioengine sg only works on block or character devices\n");
952 // Determine size of commands needed based on max_lba
953 if (max_lba >= MAX_10B_LBA) {
954 dprint(FD_IO, "sgio_type_check: using 16 byte read/write "
955 "commands for lba above 0x%016llx/0x%016llx\n",
956 MAX_10B_LBA, max_lba);
959 if (f->filetype == FIO_TYPE_BLOCK) {
960 td->io_ops->getevents = NULL;
961 td->io_ops->event = NULL;
962 td->io_ops->commit = NULL;
964 ** Setting these functions to null may cause problems
965 ** with filename=/dev/sda:/dev/sg0 since we are only
966 ** considering a single file
969 sd->type_checked = 1;
974 static int fio_sgio_open(struct thread_data *td, struct fio_file *f)
976 struct sgio_data *sd = td->io_ops_data;
979 ret = generic_open_file(td, f);
983 if (sd && !sd->type_checked && fio_sgio_type_check(td, f)) {
984 ret = generic_close_file(td, f);
992 * Build an error string with details about the driver, host or scsi
993 * error contained in the sg header Caller will use as necessary.
995 static char *fio_sgio_errdetails(struct io_u *io_u)
997 struct sg_io_hdr *hdr = &io_u->hdr;
998 #define MAXERRDETAIL 1024
999 #define MAXMSGCHUNK 128
1000 char *msg, msgchunk[MAXMSGCHUNK];
1003 msg = calloc(1, MAXERRDETAIL);
1007 * can't seem to find sg_err.h, so I'll just echo the define values
1008 * so others can search on internet to find clearer clues of meaning.
1010 if (hdr->info & SG_INFO_CHECK) {
1011 if (hdr->host_status) {
1012 snprintf(msgchunk, MAXMSGCHUNK, "SG Host Status: 0x%02x; ", hdr->host_status);
1013 strlcat(msg, msgchunk, MAXERRDETAIL);
1014 switch (hdr->host_status) {
1016 strlcat(msg, "SG_ERR_DID_NO_CONNECT", MAXERRDETAIL);
1019 strlcat(msg, "SG_ERR_DID_BUS_BUSY", MAXERRDETAIL);
1022 strlcat(msg, "SG_ERR_DID_TIME_OUT", MAXERRDETAIL);
1025 strlcat(msg, "SG_ERR_DID_BAD_TARGET", MAXERRDETAIL);
1028 strlcat(msg, "SG_ERR_DID_ABORT", MAXERRDETAIL);
1031 strlcat(msg, "SG_ERR_DID_PARITY", MAXERRDETAIL);
1034 strlcat(msg, "SG_ERR_DID_ERROR (internal error)", MAXERRDETAIL);
1037 strlcat(msg, "SG_ERR_DID_RESET", MAXERRDETAIL);
1040 strlcat(msg, "SG_ERR_DID_BAD_INTR (unexpected)", MAXERRDETAIL);
1043 strlcat(msg, "SG_ERR_DID_PASSTHROUGH", MAXERRDETAIL);
1046 strlcat(msg, "SG_ERR_DID_SOFT_ERROR (driver retry?)", MAXERRDETAIL);
1049 strlcat(msg, "SG_ERR_DID_IMM_RETRY", MAXERRDETAIL);
1052 strlcat(msg, "SG_ERR_DID_REQUEUE", MAXERRDETAIL);
1055 strlcat(msg, "SG_ERR_DID_TRANSPORT_DISRUPTED", MAXERRDETAIL);
1058 strlcat(msg, "SG_ERR_DID_TRANSPORT_FAILFAST", MAXERRDETAIL);
1061 strlcat(msg, "SG_ERR_DID_TARGET_FAILURE", MAXERRDETAIL);
1064 strlcat(msg, "SG_ERR_DID_NEXUS_FAILURE", MAXERRDETAIL);
1067 strlcat(msg, "SG_ERR_DID_ALLOC_FAILURE", MAXERRDETAIL);
1070 strlcat(msg, "SG_ERR_DID_MEDIUM_ERROR", MAXERRDETAIL);
1073 strlcat(msg, "Unknown", MAXERRDETAIL);
1076 strlcat(msg, ". ", MAXERRDETAIL);
1078 if (hdr->driver_status) {
1079 snprintf(msgchunk, MAXMSGCHUNK, "SG Driver Status: 0x%02x; ", hdr->driver_status);
1080 strlcat(msg, msgchunk, MAXERRDETAIL);
1081 switch (hdr->driver_status & 0x0F) {
1083 strlcat(msg, "SG_ERR_DRIVER_BUSY", MAXERRDETAIL);
1086 strlcat(msg, "SG_ERR_DRIVER_SOFT", MAXERRDETAIL);
1089 strlcat(msg, "SG_ERR_DRIVER_MEDIA", MAXERRDETAIL);
1092 strlcat(msg, "SG_ERR_DRIVER_ERROR", MAXERRDETAIL);
1095 strlcat(msg, "SG_ERR_DRIVER_INVALID", MAXERRDETAIL);
1098 strlcat(msg, "SG_ERR_DRIVER_TIMEOUT", MAXERRDETAIL);
1101 strlcat(msg, "SG_ERR_DRIVER_HARD", MAXERRDETAIL);
1104 strlcat(msg, "SG_ERR_DRIVER_SENSE", MAXERRDETAIL);
1107 strlcat(msg, "Unknown", MAXERRDETAIL);
1110 strlcat(msg, "; ", MAXERRDETAIL);
1111 switch (hdr->driver_status & 0xF0) {
1113 strlcat(msg, "SG_ERR_SUGGEST_RETRY", MAXERRDETAIL);
1116 strlcat(msg, "SG_ERR_SUGGEST_ABORT", MAXERRDETAIL);
1119 strlcat(msg, "SG_ERR_SUGGEST_REMAP", MAXERRDETAIL);
1122 strlcat(msg, "SG_ERR_SUGGEST_DIE", MAXERRDETAIL);
1125 strlcat(msg, "SG_ERR_SUGGEST_SENSE", MAXERRDETAIL);
1128 strlcat(msg, ". ", MAXERRDETAIL);
1131 snprintf(msgchunk, MAXMSGCHUNK, "SG SCSI Status: 0x%02x; ", hdr->status);
1132 strlcat(msg, msgchunk, MAXERRDETAIL);
1133 // SCSI 3 status codes
1134 switch (hdr->status) {
1136 strlcat(msg, "CHECK_CONDITION", MAXERRDETAIL);
1139 strlcat(msg, "CONDITION_MET", MAXERRDETAIL);
1142 strlcat(msg, "BUSY", MAXERRDETAIL);
1145 strlcat(msg, "INTERMEDIATE", MAXERRDETAIL);
1148 strlcat(msg, "INTERMEDIATE_CONDITION_MET", MAXERRDETAIL);
1151 strlcat(msg, "RESERVATION_CONFLICT", MAXERRDETAIL);
1154 strlcat(msg, "COMMAND_TERMINATED", MAXERRDETAIL);
1157 strlcat(msg, "TASK_SET_FULL", MAXERRDETAIL);
1160 strlcat(msg, "ACA_ACTIVE", MAXERRDETAIL);
1163 strlcat(msg, "TASK_ABORTED", MAXERRDETAIL);
1166 strlcat(msg, "Unknown", MAXERRDETAIL);
1169 strlcat(msg, ". ", MAXERRDETAIL);
1171 if (hdr->sb_len_wr) {
1172 snprintf(msgchunk, MAXMSGCHUNK, "Sense Data (%d bytes):", hdr->sb_len_wr);
1173 strlcat(msg, msgchunk, MAXERRDETAIL);
1174 for (i = 0; i < hdr->sb_len_wr; i++) {
1175 snprintf(msgchunk, MAXMSGCHUNK, " %02x", hdr->sbp[i]);
1176 strlcat(msg, msgchunk, MAXERRDETAIL);
1178 strlcat(msg, ". ", MAXERRDETAIL);
1180 if (hdr->resid != 0) {
1181 snprintf(msgchunk, MAXMSGCHUNK, "SG Driver: %d bytes out of %d not transferred. ", hdr->resid, hdr->dxfer_len);
1182 strlcat(msg, msgchunk, MAXERRDETAIL);
1185 strlcat(msg, "cdb:", MAXERRDETAIL);
1186 for (i = 0; i < hdr->cmd_len; i++) {
1187 snprintf(msgchunk, MAXMSGCHUNK, " %02x", hdr->cmdp[i]);
1188 strlcat(msg, msgchunk, MAXERRDETAIL);
1190 strlcat(msg, ". ", MAXERRDETAIL);
1191 if (io_u->ddir == DDIR_TRIM) {
1192 unsigned char *param_list = hdr->dxferp;
1193 strlcat(msg, "dxferp:", MAXERRDETAIL);
1194 for (i = 0; i < hdr->dxfer_len; i++) {
1195 snprintf(msgchunk, MAXMSGCHUNK, " %02x", param_list[i]);
1196 strlcat(msg, msgchunk, MAXERRDETAIL);
1198 strlcat(msg, ". ", MAXERRDETAIL);
1203 if (!(hdr->info & SG_INFO_CHECK) && !strlen(msg))
1204 snprintf(msg, MAXERRDETAIL, "%s",
1205 "SG Driver did not report a Host, Driver or Device check");
1211 * get max file size from read capacity.
1213 static int fio_sgio_get_file_size(struct thread_data *td, struct fio_file *f)
1216 * get_file_size is being called even before sgio_init is
1217 * called, so none of the sg_io structures are
1218 * initialized in the thread_data yet. So we need to do the
1219 * ReadCapacity without any of those helpers. One of the effects
1220 * is that ReadCapacity may get called 4 times on each open:
1221 * readcap(10) followed by readcap(16) if needed - just to get
1222 * the file size after the init occurs - it will be called
1223 * again when "type_check" is called during structure
1224 * initialization I'm not sure how to prevent this little
1227 unsigned int bs = 0;
1228 unsigned long long max_lba = 0;
1231 if (fio_file_size_known(f))
1234 if (f->filetype != FIO_TYPE_BLOCK && f->filetype != FIO_TYPE_CHAR) {
1235 td_verror(td, EINVAL, "wrong file type");
1236 log_err("ioengine sg only works on block or character devices\n");
1240 ret = fio_sgio_read_capacity(td, &bs, &max_lba);
1242 td_verror(td, td->error, "fio_sgio_read_capacity");
1243 log_err("ioengine sg unable to successfully execute read capacity to get block size and maximum lba\n");
1247 f->real_file_size = (max_lba + 1) * bs;
1248 fio_file_set_size_known(f);
1253 static struct ioengine_ops ioengine = {
1255 .version = FIO_IOOPS_VERSION,
1256 .init = fio_sgio_init,
1257 .prep = fio_sgio_prep,
1258 .queue = fio_sgio_queue,
1259 .commit = fio_sgio_commit,
1260 .getevents = fio_sgio_getevents,
1261 .errdetails = fio_sgio_errdetails,
1262 .event = fio_sgio_event,
1263 .cleanup = fio_sgio_cleanup,
1264 .open_file = fio_sgio_open,
1265 .close_file = generic_close_file,
1266 .get_file_size = fio_sgio_get_file_size,
1267 .flags = FIO_SYNCIO | FIO_RAWIO,
1269 .option_struct_size = sizeof(struct sg_options)
1272 #else /* FIO_HAVE_SGIO */
1275 * When we have a proper configure system in place, we simply wont build
1276 * and install this io engine. For now install a crippled version that
1277 * just complains and fails to load.
1279 static int fio_sgio_init(struct thread_data fio_unused *td)
1281 log_err("fio: ioengine sg not available\n");
1285 static struct ioengine_ops ioengine = {
1287 .version = FIO_IOOPS_VERSION,
1288 .init = fio_sgio_init,
1293 static void fio_init fio_sgio_register(void)
1295 register_ioengine(&ioengine);
1298 static void fio_exit fio_sgio_unregister(void)
1300 unregister_ioengine(&ioengine);