2 * Copyright (c) 1997-2007 Kenneth D. Merry
3 * Copyright (c) 2013, 2014, 2015 Spectra Logic Corporation
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions, and the following disclaimer,
11 * without modification.
12 * 2. Redistributions in binary form must reproduce at minimum a disclaimer
13 * substantially similar to the "NO WARRANTY" disclaimer below
14 * ("Disclaimer") and any redistribution must be conditioned upon
15 * including a substantially similar Disclaimer requirement for further
16 * binary redistribution.
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
20 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
21 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTIBILITY AND FITNESS FOR
22 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
23 * HOLDERS OR CONTRIBUTORS BE LIABLE FOR SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
27 * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
28 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGES.
31 * Authors: Ken Merry (Spectra Logic Corporation)
35 * This is eventually intended to be:
36 * - A basic data transfer/copy utility
37 * - A simple benchmark utility
38 * - An example of how to use the asynchronous pass(4) driver interface.
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
43 #include <sys/ioctl.h>
44 #include <sys/stdint.h>
45 #include <sys/types.h>
46 #include <sys/endian.h>
47 #include <sys/param.h>
50 #include <sys/event.h>
54 #include <machine/bus.h>
56 #include <sys/bus_dma.h>
63 #include <semaphore.h>
77 #include <cam/cam_debug.h>
78 #include <cam/cam_ccb.h>
79 #include <cam/scsi/scsi_all.h>
80 #include <cam/scsi/scsi_da.h>
81 #include <cam/scsi/scsi_pass.h>
82 #include <cam/scsi/scsi_message.h>
83 #include <cam/scsi/smp_all.h>
89 CAMDD_CMD_NONE = 0x00000000,
90 CAMDD_CMD_HELP = 0x00000001,
91 CAMDD_CMD_WRITE = 0x00000002,
92 CAMDD_CMD_READ = 0x00000003
96 CAMDD_ARG_NONE = 0x00000000,
97 CAMDD_ARG_VERBOSE = 0x00000001,
98 CAMDD_ARG_DEVICE = 0x00000002,
99 CAMDD_ARG_BUS = 0x00000004,
100 CAMDD_ARG_TARGET = 0x00000008,
101 CAMDD_ARG_LUN = 0x00000010,
102 CAMDD_ARG_UNIT = 0x00000020,
103 CAMDD_ARG_TIMEOUT = 0x00000040,
104 CAMDD_ARG_ERR_RECOVER = 0x00000080,
105 CAMDD_ARG_RETRIES = 0x00000100
109 CAMDD_DEV_NONE = 0x00,
110 CAMDD_DEV_PASS = 0x01,
111 CAMDD_DEV_FILE = 0x02
114 struct camdd_io_opts {
115 camdd_dev_type dev_type;
118 uint64_t queue_depth;
131 struct camdd_buf_indirect {
133 * Pointer to the source buffer.
135 struct camdd_buf *src_buf;
138 * Offset into the source buffer, in bytes.
142 * Pointer to the starting point in the source buffer.
147 * Length of this chunk in bytes.
152 struct camdd_buf_data {
154 * Buffer allocated when we allocate this camdd_buf. This should
155 * be the size of the blocksize for this device.
160 * The amount of backing store allocated in buf. Generally this
161 * will be the blocksize of the device.
166 * The amount of data that was put into the buffer (on reads) or
167 * the amount of data we have put onto the src_list so far (on
173 * The amount of data that was not transferred.
178 * Starting byte offset on the reader.
180 uint64_t src_start_offset;
183 * CCB used for pass(4) device targets.
188 * Number of scatter/gather segments.
193 * Set if we had to tack on an extra buffer to round the transfer
194 * up to a sector size.
199 * Scatter/gather list used generally when we're the writer for a
202 bus_dma_segment_t *segs;
205 * Scatter/gather list used generally when we're the writer for a
206 * file or block device;
211 union camdd_buf_types {
212 struct camdd_buf_indirect indirect;
213 struct camdd_buf_data data;
219 CAMDD_STATUS_SHORT_IO,
225 camdd_buf_type buf_type;
226 union camdd_buf_types buf_type_spec;
228 camdd_buf_status status;
234 * A reference count of how many indirect buffers point to this
240 * A link back to our parent device.
242 struct camdd_dev *dev;
243 STAILQ_ENTRY(camdd_buf) links;
244 STAILQ_ENTRY(camdd_buf) work_links;
247 * A count of the buffers on the src_list.
252 * List of buffers from our partner thread that are the components
253 * of this buffer for the I/O. Uses src_links.
255 STAILQ_HEAD(,camdd_buf) src_list;
256 STAILQ_ENTRY(camdd_buf) src_links;
259 #define NUM_DEV_TYPES 2
261 struct camdd_dev_pass {
263 struct cam_device *dev;
281 CAMDD_FF_NONE = 0x00,
282 CAMDD_FF_CAN_SEEK = 0x01
285 struct camdd_dev_file {
288 char filename[MAXPATHLEN + 1];
289 camdd_file_type file_type;
290 camdd_file_flags file_flags;
294 struct camdd_dev_block {
300 union camdd_dev_spec {
301 struct camdd_dev_pass pass;
302 struct camdd_dev_file file;
303 struct camdd_dev_block block;
307 CAMDD_DEV_FLAG_NONE = 0x00,
308 CAMDD_DEV_FLAG_EOF = 0x01,
309 CAMDD_DEV_FLAG_PEER_EOF = 0x02,
310 CAMDD_DEV_FLAG_ACTIVE = 0x04,
311 CAMDD_DEV_FLAG_EOF_SENT = 0x08,
312 CAMDD_DEV_FLAG_EOF_QUEUED = 0x10
316 camdd_dev_type dev_type;
317 union camdd_dev_spec dev_spec;
318 camdd_dev_flags flags;
319 char device_name[MAXPATHLEN+1];
321 uint32_t sector_size;
323 uint64_t sector_io_limit;
329 uint64_t start_offset_bytes;
330 uint64_t next_io_pos_bytes;
331 uint64_t next_peer_pos_bytes;
332 uint64_t next_completion_pos_bytes;
333 uint64_t peer_bytes_queued;
334 uint64_t bytes_transferred;
335 uint32_t target_queue_depth;
336 uint32_t cur_active_io;
338 uint32_t extra_buf_len;
339 struct camdd_dev *peer_dev;
340 pthread_mutex_t mutex;
344 int (*run)(struct camdd_dev *dev);
345 int (*fetch)(struct camdd_dev *dev);
348 * Buffers that are available for I/O. Uses links.
350 STAILQ_HEAD(,camdd_buf) free_queue;
353 * Free indirect buffers. These are used for breaking a large
354 * buffer into multiple pieces.
356 STAILQ_HEAD(,camdd_buf) free_indirect_queue;
359 * Buffers that have been queued to the kernel. Uses links.
361 STAILQ_HEAD(,camdd_buf) active_queue;
364 * Will generally contain one of our buffers that is waiting for enough
365 * I/O from our partner thread to be able to execute. This will
366 * generally happen when our per-I/O-size is larger than the
367 * partner thread's per-I/O-size. Uses links.
369 STAILQ_HEAD(,camdd_buf) pending_queue;
372 * Number of buffers on the pending queue
374 int num_pending_queue;
377 * Buffers that are filled and ready to execute. This is used when
378 * our partner (reader) thread sends us blocks that are larger than
379 * our blocksize, and so we have to split them into multiple pieces.
381 STAILQ_HEAD(,camdd_buf) run_queue;
384 * Number of buffers on the run queue.
388 STAILQ_HEAD(,camdd_buf) reorder_queue;
390 int num_reorder_queue;
393 * Buffers that have been queued to us by our partner thread
394 * (generally the reader thread) to be written out. Uses
397 STAILQ_HEAD(,camdd_buf) work_queue;
400 * Buffers that have been completed by our partner thread. Uses
403 STAILQ_HEAD(,camdd_buf) peer_done_queue;
406 * Number of buffers on the peer done queue.
408 uint32_t num_peer_done_queue;
411 * A list of buffers that we have queued to our peer thread. Uses
414 STAILQ_HEAD(,camdd_buf) peer_work_queue;
417 * Number of buffers on the peer work queue.
419 uint32_t num_peer_work_queue;
422 static sem_t camdd_sem;
423 static int need_exit = 0;
424 static int error_exit = 0;
425 static int need_status = 0;
428 #define min(a, b) (a < b) ? a : b
432 * XXX KDM private copy of timespecsub(). This is normally defined in
433 * sys/time.h, but is only enabled in the kernel. If that definition is
434 * enabled in userland, it breaks the build of libnetbsd.
437 #define timespecsub(vvp, uvp) \
439 (vvp)->tv_sec -= (uvp)->tv_sec; \
440 (vvp)->tv_nsec -= (uvp)->tv_nsec; \
441 if ((vvp)->tv_nsec < 0) { \
443 (vvp)->tv_nsec += 1000000000; \
449 /* Generically useful offsets into the peripheral private area */
450 #define ppriv_ptr0 periph_priv.entries[0].ptr
451 #define ppriv_ptr1 periph_priv.entries[1].ptr
452 #define ppriv_field0 periph_priv.entries[0].field
453 #define ppriv_field1 periph_priv.entries[1].field
455 #define ccb_buf ppriv_ptr0
457 #define CAMDD_FILE_DEFAULT_BLOCK 524288
458 #define CAMDD_FILE_DEFAULT_DEPTH 1
459 #define CAMDD_PASS_MAX_BLOCK 1048576
460 #define CAMDD_PASS_DEFAULT_DEPTH 6
461 #define CAMDD_PASS_RW_TIMEOUT 60 * 1000
463 static int parse_btl(char *tstr, int *bus, int *target, int *lun,
464 camdd_argmask *arglst);
465 void camdd_free_dev(struct camdd_dev *dev);
466 struct camdd_dev *camdd_alloc_dev(camdd_dev_type dev_type,
467 struct kevent *new_ke, int num_ke,
468 int retry_count, int timeout);
469 static struct camdd_buf *camdd_alloc_buf(struct camdd_dev *dev,
470 camdd_buf_type buf_type);
471 void camdd_release_buf(struct camdd_buf *buf);
472 struct camdd_buf *camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type);
473 int camdd_buf_sg_create(struct camdd_buf *buf, int iovec,
474 uint32_t sector_size, uint32_t *num_sectors_used,
475 int *double_buf_needed);
476 uint32_t camdd_buf_get_len(struct camdd_buf *buf);
477 void camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf);
478 int camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
479 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran);
480 struct camdd_dev *camdd_probe_file(int fd, struct camdd_io_opts *io_opts,
481 int retry_count, int timeout);
482 struct camdd_dev *camdd_probe_pass(struct cam_device *cam_dev,
483 struct camdd_io_opts *io_opts,
484 camdd_argmask arglist, int probe_retry_count,
485 int probe_timeout, int io_retry_count,
487 void *camdd_file_worker(void *arg);
488 camdd_buf_status camdd_ccb_status(union ccb *ccb);
489 int camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf);
490 int camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf);
491 void camdd_peer_done(struct camdd_buf *buf);
492 void camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
494 int camdd_pass_fetch(struct camdd_dev *dev);
495 int camdd_file_run(struct camdd_dev *dev);
496 int camdd_pass_run(struct camdd_dev *dev);
497 int camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len);
498 int camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf);
499 void camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
500 uint32_t *peer_depth, uint32_t *our_bytes,
501 uint32_t *peer_bytes);
502 void *camdd_worker(void *arg);
503 void camdd_sig_handler(int sig);
504 void camdd_print_status(struct camdd_dev *camdd_dev,
505 struct camdd_dev *other_dev,
506 struct timespec *start_time);
507 int camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts,
508 uint64_t max_io, int retry_count, int timeout);
509 int camdd_parse_io_opts(char *args, int is_write,
510 struct camdd_io_opts *io_opts);
514 * Parse out a bus, or a bus, target and lun in the following
520 * Returns the number of parsed components, or 0.
523 parse_btl(char *tstr, int *bus, int *target, int *lun, camdd_argmask *arglst)
528 while (isspace(*tstr) && (*tstr != '\0'))
531 tmpstr = (char *)strtok(tstr, ":");
532 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
533 *bus = strtol(tmpstr, NULL, 0);
534 *arglst |= CAMDD_ARG_BUS;
536 tmpstr = (char *)strtok(NULL, ":");
537 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
538 *target = strtol(tmpstr, NULL, 0);
539 *arglst |= CAMDD_ARG_TARGET;
541 tmpstr = (char *)strtok(NULL, ":");
542 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
543 *lun = strtol(tmpstr, NULL, 0);
544 *arglst |= CAMDD_ARG_LUN;
554 * XXX KDM clean up and free all of the buffers on the queue!
557 camdd_free_dev(struct camdd_dev *dev)
562 switch (dev->dev_type) {
563 case CAMDD_DEV_FILE: {
564 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
566 if (file_dev->fd != -1)
568 free(file_dev->tmp_buf);
571 case CAMDD_DEV_PASS: {
572 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
574 if (pass_dev->dev != NULL)
575 cam_close_device(pass_dev->dev);
586 camdd_alloc_dev(camdd_dev_type dev_type, struct kevent *new_ke, int num_ke,
587 int retry_count, int timeout)
589 struct camdd_dev *dev = NULL;
594 dev = malloc(sizeof(*dev));
596 warn("%s: unable to malloc %zu bytes", __func__, sizeof(*dev));
600 bzero(dev, sizeof(*dev));
602 dev->dev_type = dev_type;
603 dev->io_timeout = timeout;
604 dev->retry_count = retry_count;
605 STAILQ_INIT(&dev->free_queue);
606 STAILQ_INIT(&dev->free_indirect_queue);
607 STAILQ_INIT(&dev->active_queue);
608 STAILQ_INIT(&dev->pending_queue);
609 STAILQ_INIT(&dev->run_queue);
610 STAILQ_INIT(&dev->reorder_queue);
611 STAILQ_INIT(&dev->work_queue);
612 STAILQ_INIT(&dev->peer_done_queue);
613 STAILQ_INIT(&dev->peer_work_queue);
614 retval = pthread_mutex_init(&dev->mutex, NULL);
616 warnc(retval, "%s: failed to initialize mutex", __func__);
620 retval = pthread_cond_init(&dev->cond, NULL);
622 warnc(retval, "%s: failed to initialize condition variable",
629 warn("%s: Unable to create kqueue", __func__);
633 ke_size = sizeof(struct kevent) * (num_ke + 4);
634 ke = malloc(ke_size);
636 warn("%s: unable to malloc %zu bytes", __func__, ke_size);
641 bcopy(new_ke, ke, num_ke * sizeof(struct kevent));
643 EV_SET(&ke[num_ke++], (uintptr_t)&dev->work_queue, EVFILT_USER,
644 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
645 EV_SET(&ke[num_ke++], (uintptr_t)&dev->peer_done_queue, EVFILT_USER,
646 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
647 EV_SET(&ke[num_ke++], SIGINFO, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
648 EV_SET(&ke[num_ke++], SIGINT, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
650 retval = kevent(dev->kq, ke, num_ke, NULL, 0, NULL);
652 warn("%s: Unable to register kevents", __func__);
665 static struct camdd_buf *
666 camdd_alloc_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
668 struct camdd_buf *buf = NULL;
669 uint8_t *data_ptr = NULL;
672 * We only need to allocate data space for data buffers.
676 data_ptr = malloc(dev->blocksize);
677 if (data_ptr == NULL) {
678 warn("unable to allocate %u bytes", dev->blocksize);
686 buf = malloc(sizeof(*buf));
688 warn("unable to allocate %zu bytes", sizeof(*buf));
692 bzero(buf, sizeof(*buf));
693 buf->buf_type = buf_type;
696 case CAMDD_BUF_DATA: {
697 struct camdd_buf_data *data;
699 data = &buf->buf_type_spec.data;
701 data->alloc_len = dev->blocksize;
702 data->buf = data_ptr;
705 case CAMDD_BUF_INDIRECT:
710 STAILQ_INIT(&buf->src_list);
715 if (data_ptr != NULL)
725 camdd_release_buf(struct camdd_buf *buf)
727 struct camdd_dev *dev;
731 switch (buf->buf_type) {
732 case CAMDD_BUF_DATA: {
733 struct camdd_buf_data *data;
735 data = &buf->buf_type_spec.data;
737 if (data->segs != NULL) {
738 if (data->extra_buf != 0) {
742 data->segs[data->sg_count - 1].ds_addr;
749 } else if (data->iovec != NULL) {
750 if (data->extra_buf != 0) {
751 free(data->iovec[data->sg_count - 1].iov_base);
758 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
761 case CAMDD_BUF_INDIRECT:
762 STAILQ_INSERT_TAIL(&dev->free_indirect_queue, buf, links);
765 err(1, "%s: Invalid buffer type %d for released buffer",
766 __func__, buf->buf_type);
772 camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
774 struct camdd_buf *buf = NULL;
778 buf = STAILQ_FIRST(&dev->free_queue);
780 struct camdd_buf_data *data;
784 STAILQ_REMOVE_HEAD(&dev->free_queue, links);
785 data = &buf->buf_type_spec.data;
786 data_ptr = data->buf;
787 alloc_len = data->alloc_len;
788 bzero(buf, sizeof(*buf));
789 data->buf = data_ptr;
790 data->alloc_len = alloc_len;
793 case CAMDD_BUF_INDIRECT:
794 buf = STAILQ_FIRST(&dev->free_indirect_queue);
796 STAILQ_REMOVE_HEAD(&dev->free_indirect_queue, links);
798 bzero(buf, sizeof(*buf));
802 warnx("Unknown buffer type %d requested", buf_type);
808 return (camdd_alloc_buf(dev, buf_type));
810 STAILQ_INIT(&buf->src_list);
812 buf->buf_type = buf_type;
819 camdd_buf_sg_create(struct camdd_buf *buf, int iovec, uint32_t sector_size,
820 uint32_t *num_sectors_used, int *double_buf_needed)
822 struct camdd_buf *tmp_buf;
823 struct camdd_buf_data *data;
824 uint8_t *extra_buf = NULL;
825 size_t extra_buf_len = 0;
828 data = &buf->buf_type_spec.data;
830 data->sg_count = buf->src_count;
832 * Compose a scatter/gather list from all of the buffers in the list.
833 * If the length of the buffer isn't a multiple of the sector size,
834 * we'll have to add an extra buffer. This should only happen
835 * at the end of a transfer.
837 if ((data->fill_len % sector_size) != 0) {
838 extra_buf_len = sector_size - (data->fill_len % sector_size);
839 extra_buf = calloc(extra_buf_len, 1);
840 if (extra_buf == NULL) {
841 warn("%s: unable to allocate %zu bytes for extra "
842 "buffer space", __func__, extra_buf_len);
850 data->segs = calloc(data->sg_count, sizeof(bus_dma_segment_t));
851 if (data->segs == NULL) {
852 warn("%s: unable to allocate %zu bytes for S/G list",
853 __func__, sizeof(bus_dma_segment_t) *
860 data->iovec = calloc(data->sg_count, sizeof(struct iovec));
861 if (data->iovec == NULL) {
862 warn("%s: unable to allocate %zu bytes for S/G list",
863 __func__, sizeof(struct iovec) * data->sg_count);
869 for (i = 0, tmp_buf = STAILQ_FIRST(&buf->src_list);
870 i < buf->src_count && tmp_buf != NULL; i++,
871 tmp_buf = STAILQ_NEXT(tmp_buf, src_links)) {
873 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
874 struct camdd_buf_data *tmp_data;
876 tmp_data = &tmp_buf->buf_type_spec.data;
878 data->segs[i].ds_addr =
879 (bus_addr_t) tmp_data->buf;
880 data->segs[i].ds_len = tmp_data->fill_len -
883 data->iovec[i].iov_base = tmp_data->buf;
884 data->iovec[i].iov_len = tmp_data->fill_len -
887 if (((tmp_data->fill_len - tmp_data->resid) %
889 *double_buf_needed = 1;
891 struct camdd_buf_indirect *tmp_ind;
893 tmp_ind = &tmp_buf->buf_type_spec.indirect;
895 data->segs[i].ds_addr =
896 (bus_addr_t)tmp_ind->start_ptr;
897 data->segs[i].ds_len = tmp_ind->len;
899 data->iovec[i].iov_base = tmp_ind->start_ptr;
900 data->iovec[i].iov_len = tmp_ind->len;
902 if ((tmp_ind->len % sector_size) != 0)
903 *double_buf_needed = 1;
907 if (extra_buf != NULL) {
909 data->segs[i].ds_addr = (bus_addr_t)extra_buf;
910 data->segs[i].ds_len = extra_buf_len;
912 data->iovec[i].iov_base = extra_buf;
913 data->iovec[i].iov_len = extra_buf_len;
917 if ((tmp_buf != NULL) || (i != data->sg_count)) {
918 warnx("buffer source count does not match "
919 "number of buffers in list!");
926 *num_sectors_used = (data->fill_len + extra_buf_len) /
933 camdd_buf_get_len(struct camdd_buf *buf)
937 if (buf->buf_type != CAMDD_BUF_DATA) {
938 struct camdd_buf_indirect *indirect;
940 indirect = &buf->buf_type_spec.indirect;
943 struct camdd_buf_data *data;
945 data = &buf->buf_type_spec.data;
946 len = data->fill_len;
953 camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf)
955 struct camdd_buf_data *data;
957 assert(buf->buf_type == CAMDD_BUF_DATA);
959 data = &buf->buf_type_spec.data;
961 STAILQ_INSERT_TAIL(&buf->src_list, child_buf, src_links);
964 data->fill_len += camdd_buf_get_len(child_buf);
972 } camdd_status_item_index;
974 static struct camdd_status_items {
976 struct mt_status_entry *entry;
977 } req_status_items[] = {
980 { "blk_gran", NULL },
981 { "max_effective_iosize", NULL }
985 camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
986 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran)
988 struct mt_status_data status_data;
989 char *xml_str = NULL;
993 retval = mt_get_xml_str(fd, MTIOCEXTGET, &xml_str);
995 err(1, "Couldn't get XML string from %s", filename);
997 retval = mt_get_status(xml_str, &status_data);
998 if (retval != XML_STATUS_OK) {
999 warn("couldn't get status for %s", filename);
1005 if (status_data.error != 0) {
1006 warnx("%s", status_data.error_str);
1011 for (i = 0; i < sizeof(req_status_items) /
1012 sizeof(req_status_items[0]); i++) {
1015 name = __DECONST(char *, req_status_items[i].name);
1016 req_status_items[i].entry = mt_status_entry_find(&status_data,
1018 if (req_status_items[i].entry == NULL) {
1019 errx(1, "Cannot find status entry %s",
1020 req_status_items[i].name);
1024 *max_iosize = req_status_items[CAMDD_TS_EFF_IOSIZE].entry->value_unsigned;
1025 *max_blk= req_status_items[CAMDD_TS_MAX_BLK].entry->value_unsigned;
1026 *min_blk= req_status_items[CAMDD_TS_MIN_BLK].entry->value_unsigned;
1027 *blk_gran = req_status_items[CAMDD_TS_BLK_GRAN].entry->value_unsigned;
1031 mt_status_free(&status_data);
1037 camdd_probe_file(int fd, struct camdd_io_opts *io_opts, int retry_count,
1040 struct camdd_dev *dev = NULL;
1041 struct camdd_dev_file *file_dev;
1042 uint64_t blocksize = io_opts->blocksize;
1044 dev = camdd_alloc_dev(CAMDD_DEV_FILE, NULL, 0, retry_count, timeout);
1048 file_dev = &dev->dev_spec.file;
1050 strlcpy(file_dev->filename, io_opts->dev_name,
1051 sizeof(file_dev->filename));
1052 strlcpy(dev->device_name, io_opts->dev_name, sizeof(dev->device_name));
1054 dev->blocksize = CAMDD_FILE_DEFAULT_BLOCK;
1056 dev->blocksize = blocksize;
1058 if ((io_opts->queue_depth != 0)
1059 && (io_opts->queue_depth != 1)) {
1060 warnx("Queue depth %ju for %s ignored, only 1 outstanding "
1061 "command supported", (uintmax_t)io_opts->queue_depth,
1064 dev->target_queue_depth = CAMDD_FILE_DEFAULT_DEPTH;
1065 dev->run = camdd_file_run;
1069 * We can effectively access files on byte boundaries. We'll reset
1070 * this for devices like disks that can be accessed on sector
1073 dev->sector_size = 1;
1075 if ((fd != STDIN_FILENO)
1076 && (fd != STDOUT_FILENO)) {
1079 retval = fstat(fd, &file_dev->sb);
1081 warn("Cannot stat %s", dev->device_name);
1084 if (S_ISREG(file_dev->sb.st_mode)) {
1085 file_dev->file_type = CAMDD_FILE_REG;
1086 } else if (S_ISCHR(file_dev->sb.st_mode)) {
1089 if (ioctl(fd, FIODTYPE, &type) == -1)
1090 err(1, "FIODTYPE ioctl failed on %s",
1094 file_dev->file_type = CAMDD_FILE_TAPE;
1095 else if (type & D_DISK)
1096 file_dev->file_type = CAMDD_FILE_DISK;
1097 else if (type & D_MEM)
1098 file_dev->file_type = CAMDD_FILE_MEM;
1099 else if (type & D_TTY)
1100 file_dev->file_type = CAMDD_FILE_TTY;
1102 } else if (S_ISDIR(file_dev->sb.st_mode)) {
1103 errx(1, "cannot operate on directory %s",
1105 } else if (S_ISFIFO(file_dev->sb.st_mode)) {
1106 file_dev->file_type = CAMDD_FILE_PIPE;
1108 errx(1, "Cannot determine file type for %s",
1111 switch (file_dev->file_type) {
1112 case CAMDD_FILE_REG:
1113 if (file_dev->sb.st_size != 0)
1114 dev->max_sector = file_dev->sb.st_size - 1;
1116 dev->max_sector = 0;
1117 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1119 case CAMDD_FILE_TAPE: {
1120 uint64_t max_iosize, max_blk, min_blk, blk_gran;
1122 * Check block limits and maximum effective iosize.
1123 * Make sure the blocksize is within the block
1124 * limits (and a multiple of the minimum blocksize)
1125 * and that the blocksize is <= maximum effective
1128 retval = camdd_probe_tape(fd, dev->device_name,
1129 &max_iosize, &max_blk, &min_blk, &blk_gran);
1131 errx(1, "Unable to probe tape %s",
1135 * The blocksize needs to be <= the maximum
1136 * effective I/O size of the tape device. Note
1137 * that this also takes into account the maximum
1138 * blocksize reported by READ BLOCK LIMITS.
1140 if (dev->blocksize > max_iosize) {
1141 warnx("Blocksize %u too big for %s, limiting "
1142 "to %ju", dev->blocksize, dev->device_name,
1144 dev->blocksize = max_iosize;
1148 * The blocksize needs to be at least min_blk;
1150 if (dev->blocksize < min_blk) {
1151 warnx("Blocksize %u too small for %s, "
1152 "increasing to %ju", dev->blocksize,
1153 dev->device_name, min_blk);
1154 dev->blocksize = min_blk;
1158 * And the blocksize needs to be a multiple of
1159 * the block granularity.
1162 && (dev->blocksize % (1 << blk_gran))) {
1163 warnx("Blocksize %u for %s not a multiple of "
1164 "%d, adjusting to %d", dev->blocksize,
1165 dev->device_name, (1 << blk_gran),
1166 dev->blocksize & ~((1 << blk_gran) - 1));
1167 dev->blocksize &= ~((1 << blk_gran) - 1);
1170 if (dev->blocksize == 0) {
1171 errx(1, "Unable to derive valid blocksize for "
1172 "%s", dev->device_name);
1176 * For tape drives, set the sector size to the
1177 * blocksize so that we make sure not to write
1178 * less than the blocksize out to the drive.
1180 dev->sector_size = dev->blocksize;
1183 case CAMDD_FILE_DISK: {
1185 unsigned int sector_size;
1187 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1189 if (ioctl(fd, DIOCGSECTORSIZE, §or_size) == -1) {
1190 err(1, "DIOCGSECTORSIZE ioctl failed on %s",
1194 if (sector_size == 0) {
1195 errx(1, "DIOCGSECTORSIZE ioctl returned "
1196 "invalid sector size %u for %s",
1197 sector_size, dev->device_name);
1200 if (ioctl(fd, DIOCGMEDIASIZE, &media_size) == -1) {
1201 err(1, "DIOCGMEDIASIZE ioctl failed on %s",
1205 if (media_size == 0) {
1206 errx(1, "DIOCGMEDIASIZE ioctl returned "
1207 "invalid media size %ju for %s",
1208 (uintmax_t)media_size, dev->device_name);
1211 if (dev->blocksize % sector_size) {
1212 errx(1, "%s blocksize %u not a multiple of "
1213 "sector size %u", dev->device_name,
1214 dev->blocksize, sector_size);
1217 dev->sector_size = sector_size;
1218 dev->max_sector = (media_size / sector_size) - 1;
1221 case CAMDD_FILE_MEM:
1222 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1229 if ((io_opts->offset != 0)
1230 && ((file_dev->file_flags & CAMDD_FF_CAN_SEEK) == 0)) {
1231 warnx("Offset %ju specified for %s, but we cannot seek on %s",
1232 io_opts->offset, io_opts->dev_name, io_opts->dev_name);
1236 else if ((io_opts->offset != 0)
1237 && ((io_opts->offset % dev->sector_size) != 0)) {
1238 warnx("Offset %ju for %s is not a multiple of the "
1239 "sector size %u", io_opts->offset,
1240 io_opts->dev_name, dev->sector_size);
1243 dev->start_offset_bytes = io_opts->offset;
1251 camdd_free_dev(dev);
1256 * Need to implement this. Do a basic probe:
1257 * - Check the inquiry data, make sure we're talking to a device that we
1258 * can reasonably expect to talk to -- direct, RBC, CD, WORM.
1259 * - Send a test unit ready, make sure the device is available.
1260 * - Get the capacity and block size.
1263 camdd_probe_pass(struct cam_device *cam_dev, struct camdd_io_opts *io_opts,
1264 camdd_argmask arglist, int probe_retry_count,
1265 int probe_timeout, int io_retry_count, int io_timeout)
1269 uint32_t cpi_maxio, max_iosize, pass_numblocks;
1271 struct scsi_read_capacity_data rcap;
1272 struct scsi_read_capacity_data_long rcaplong;
1273 struct camdd_dev *dev;
1274 struct camdd_dev_pass *pass_dev;
1280 scsi_dev_type = SID_TYPE(&cam_dev->inq_data);
1285 * For devices that support READ CAPACITY, we'll attempt to get the
1286 * capacity. Otherwise, we really don't support tape or other
1287 * devices via SCSI passthrough, so just return an error in that case.
1289 switch (scsi_dev_type) {
1298 errx(1, "Unsupported SCSI device type %d", scsi_dev_type);
1299 break; /*NOTREACHED*/
1302 ccb = cam_getccb(cam_dev);
1305 warnx("%s: error allocating ccb", __func__);
1309 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
1311 scsi_read_capacity(&ccb->csio,
1312 /*retries*/ probe_retry_count,
1314 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1317 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1319 /* Disable freezing the device queue */
1320 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1322 if (arglist & CAMDD_ARG_ERR_RECOVER)
1323 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1325 if (cam_send_ccb(cam_dev, ccb) < 0) {
1326 warn("error sending READ CAPACITY command");
1328 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1329 CAM_EPF_ALL, stderr);
1334 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1335 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1339 maxsector = scsi_4btoul(rcap.addr);
1340 block_len = scsi_4btoul(rcap.length);
1343 * A last block of 2^32-1 means that the true capacity is over 2TB,
1344 * and we need to issue the long READ CAPACITY to get the real
1345 * capacity. Otherwise, we're all set.
1347 if (maxsector != 0xffffffff)
1350 scsi_read_capacity_16(&ccb->csio,
1351 /*retries*/ probe_retry_count,
1353 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1357 (uint8_t *)&rcaplong,
1359 /*sense_len*/ SSD_FULL_SIZE,
1360 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1362 /* Disable freezing the device queue */
1363 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1365 if (arglist & CAMDD_ARG_ERR_RECOVER)
1366 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1368 if (cam_send_ccb(cam_dev, ccb) < 0) {
1369 warn("error sending READ CAPACITY (16) command");
1370 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1371 CAM_EPF_ALL, stderr);
1375 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1376 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1380 maxsector = scsi_8btou64(rcaplong.addr);
1381 block_len = scsi_4btoul(rcaplong.length);
1384 if (block_len == 0) {
1385 warnx("Sector size for %s%u is 0, cannot continue",
1386 cam_dev->device_name, cam_dev->dev_unit_num);
1390 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cpi);
1392 ccb->ccb_h.func_code = XPT_PATH_INQ;
1393 ccb->ccb_h.flags = CAM_DIR_NONE;
1394 ccb->ccb_h.retry_count = 1;
1396 if (cam_send_ccb(cam_dev, ccb) < 0) {
1397 warn("error sending XPT_PATH_INQ CCB");
1399 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1400 CAM_EPF_ALL, stderr);
1404 EV_SET(&ke, cam_dev->fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1406 dev = camdd_alloc_dev(CAMDD_DEV_PASS, &ke, 1, io_retry_count,
1411 pass_dev = &dev->dev_spec.pass;
1412 pass_dev->scsi_dev_type = scsi_dev_type;
1413 pass_dev->dev = cam_dev;
1414 pass_dev->max_sector = maxsector;
1415 pass_dev->block_len = block_len;
1416 pass_dev->cpi_maxio = ccb->cpi.maxio;
1417 snprintf(dev->device_name, sizeof(dev->device_name), "%s%u",
1418 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
1419 dev->sector_size = block_len;
1420 dev->max_sector = maxsector;
1424 * Determine the optimal blocksize to use for this device.
1428 * If the controller has not specified a maximum I/O size,
1429 * just go with 128K as a somewhat conservative value.
1431 if (pass_dev->cpi_maxio == 0)
1434 cpi_maxio = pass_dev->cpi_maxio;
1437 * If the controller has a large maximum I/O size, limit it
1438 * to something smaller so that the kernel doesn't have trouble
1439 * allocating buffers to copy data in and out for us.
1440 * XXX KDM this is until we have unmapped I/O support in the kernel.
1442 max_iosize = min(cpi_maxio, CAMDD_PASS_MAX_BLOCK);
1445 * If we weren't able to get a block size for some reason,
1446 * default to 512 bytes.
1448 block_len = pass_dev->block_len;
1453 * Figure out how many blocksize chunks will fit in the
1456 pass_numblocks = max_iosize / block_len;
1459 * And finally, multiple the number of blocks by the LBA
1460 * length to get our maximum block size;
1462 dev->blocksize = pass_numblocks * block_len;
1464 if (io_opts->blocksize != 0) {
1465 if ((io_opts->blocksize % dev->sector_size) != 0) {
1466 warnx("Blocksize %ju for %s is not a multiple of "
1467 "sector size %u", (uintmax_t)io_opts->blocksize,
1468 dev->device_name, dev->sector_size);
1471 dev->blocksize = io_opts->blocksize;
1473 dev->target_queue_depth = CAMDD_PASS_DEFAULT_DEPTH;
1474 if (io_opts->queue_depth != 0)
1475 dev->target_queue_depth = io_opts->queue_depth;
1477 if (io_opts->offset != 0) {
1478 if (io_opts->offset > (dev->max_sector * dev->sector_size)) {
1479 warnx("Offset %ju is past the end of device %s",
1480 io_opts->offset, dev->device_name);
1484 else if ((io_opts->offset % dev->sector_size) != 0) {
1485 warnx("Offset %ju for %s is not a multiple of the "
1486 "sector size %u", io_opts->offset,
1487 dev->device_name, dev->sector_size);
1490 dev->start_offset_bytes = io_opts->offset;
1494 dev->min_cmd_size = io_opts->min_cmd_size;
1496 dev->run = camdd_pass_run;
1497 dev->fetch = camdd_pass_fetch;
1507 camdd_free_dev(dev);
1513 camdd_worker(void *arg)
1515 struct camdd_dev *dev = arg;
1516 struct camdd_buf *buf;
1517 struct timespec ts, *kq_ts;
1522 pthread_mutex_lock(&dev->mutex);
1524 dev->flags |= CAMDD_DEV_FLAG_ACTIVE;
1531 * XXX KDM check the reorder queue depth?
1533 if (dev->write_dev == 0) {
1534 uint32_t our_depth, peer_depth, peer_bytes, our_bytes;
1535 uint32_t target_depth = dev->target_queue_depth;
1536 uint32_t peer_target_depth =
1537 dev->peer_dev->target_queue_depth;
1538 uint32_t peer_blocksize = dev->peer_dev->blocksize;
1540 camdd_get_depth(dev, &our_depth, &peer_depth,
1541 &our_bytes, &peer_bytes);
1544 while (((our_depth < target_depth)
1545 && (peer_depth < peer_target_depth))
1546 || ((peer_bytes + our_bytes) <
1547 (peer_blocksize * 2))) {
1549 while (((our_depth + peer_depth) <
1550 (target_depth + peer_target_depth))
1551 || ((peer_bytes + our_bytes) <
1552 (peer_blocksize * 3))) {
1554 retval = camdd_queue(dev, NULL);
1557 else if (retval != 0) {
1562 camdd_get_depth(dev, &our_depth, &peer_depth,
1563 &our_bytes, &peer_bytes);
1567 * See if we have any I/O that is ready to execute.
1569 buf = STAILQ_FIRST(&dev->run_queue);
1571 while (dev->target_queue_depth > dev->cur_active_io) {
1572 retval = dev->run(dev);
1574 dev->flags |= CAMDD_DEV_FLAG_EOF;
1577 } else if (retval != 0) {
1584 * We've reached EOF, or our partner has reached EOF.
1586 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
1587 || (dev->flags & CAMDD_DEV_FLAG_PEER_EOF)) {
1588 if (dev->write_dev != 0) {
1589 if ((STAILQ_EMPTY(&dev->work_queue))
1590 && (dev->num_run_queue == 0)
1591 && (dev->cur_active_io == 0)) {
1596 * If we're the reader, and the writer
1597 * got EOF, he is already done. If we got
1598 * the EOF, then we need to wait until
1599 * everything is flushed out for the writer.
1601 if (dev->flags & CAMDD_DEV_FLAG_PEER_EOF) {
1603 } else if ((dev->num_peer_work_queue == 0)
1604 && (dev->num_peer_done_queue == 0)
1605 && (dev->cur_active_io == 0)
1606 && (dev->num_run_queue == 0)) {
1611 * XXX KDM need to do something about the pending
1612 * queue and cleanup resources.
1616 if ((dev->write_dev == 0)
1617 && (dev->cur_active_io == 0)
1618 && (dev->peer_bytes_queued < dev->peer_dev->blocksize))
1624 * Run kevent to see if there are events to process.
1626 pthread_mutex_unlock(&dev->mutex);
1627 retval = kevent(dev->kq, NULL, 0, &ke, 1, kq_ts);
1628 pthread_mutex_lock(&dev->mutex);
1630 warn("%s: error returned from kevent",__func__);
1632 } else if (retval != 0) {
1633 switch (ke.filter) {
1635 if (dev->fetch != NULL) {
1636 retval = dev->fetch(dev);
1645 * We register for this so we don't get
1646 * an error as a result of a SIGINFO or a
1647 * SIGINT. It will actually get handled
1648 * by the signal handler. If we get a
1649 * SIGINT, bail out without printing an
1650 * error message. Any other signals
1651 * will result in the error message above.
1653 if (ke.ident == SIGINT)
1659 * Check to see if the other thread has
1660 * queued any I/O for us to do. (In this
1661 * case we're the writer.)
1663 for (buf = STAILQ_FIRST(&dev->work_queue);
1665 buf = STAILQ_FIRST(&dev->work_queue)) {
1666 STAILQ_REMOVE_HEAD(&dev->work_queue,
1668 retval = camdd_queue(dev, buf);
1670 * We keep going unless we get an
1671 * actual error. If we get EOF, we
1672 * still want to remove the buffers
1673 * from the queue and send the back
1674 * to the reader thread.
1684 * Next check to see if the other thread has
1685 * queued any completed buffers back to us.
1686 * (In this case we're the reader.)
1688 for (buf = STAILQ_FIRST(&dev->peer_done_queue);
1690 buf = STAILQ_FIRST(&dev->peer_done_queue)){
1692 &dev->peer_done_queue, work_links);
1693 dev->num_peer_done_queue--;
1694 camdd_peer_done(buf);
1698 warnx("%s: unknown kevent filter %d",
1699 __func__, ke.filter);
1707 dev->flags &= ~CAMDD_DEV_FLAG_ACTIVE;
1709 /* XXX KDM cleanup resources here? */
1711 pthread_mutex_unlock(&dev->mutex);
1714 sem_post(&camdd_sem);
1720 * Simplistic translation of CCB status to our local status.
1723 camdd_ccb_status(union ccb *ccb)
1725 camdd_buf_status status = CAMDD_STATUS_NONE;
1726 cam_status ccb_status;
1728 ccb_status = ccb->ccb_h.status & CAM_STATUS_MASK;
1730 switch (ccb_status) {
1732 if (ccb->csio.resid == 0) {
1733 status = CAMDD_STATUS_OK;
1734 } else if (ccb->csio.dxfer_len > ccb->csio.resid) {
1735 status = CAMDD_STATUS_SHORT_IO;
1737 status = CAMDD_STATUS_EOF;
1741 case CAM_SCSI_STATUS_ERROR: {
1742 switch (ccb->csio.scsi_status) {
1743 case SCSI_STATUS_OK:
1744 case SCSI_STATUS_COND_MET:
1745 case SCSI_STATUS_INTERMED:
1746 case SCSI_STATUS_INTERMED_COND_MET:
1747 status = CAMDD_STATUS_OK;
1749 case SCSI_STATUS_CMD_TERMINATED:
1750 case SCSI_STATUS_CHECK_COND:
1751 case SCSI_STATUS_QUEUE_FULL:
1752 case SCSI_STATUS_BUSY:
1753 case SCSI_STATUS_RESERV_CONFLICT:
1755 status = CAMDD_STATUS_ERROR;
1761 status = CAMDD_STATUS_ERROR;
1769 * Queue a buffer to our peer's work thread for writing.
1771 * Returns 0 for success, -1 for failure, 1 if the other thread exited.
1774 camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf)
1777 STAILQ_HEAD(, camdd_buf) local_queue;
1778 struct camdd_buf *buf1, *buf2;
1779 struct camdd_buf_data *data = NULL;
1780 uint64_t peer_bytes_queued = 0;
1784 STAILQ_INIT(&local_queue);
1787 * Since we're the reader, we need to queue our I/O to the writer
1788 * in sequential order in order to make sure it gets written out
1789 * in sequential order.
1791 * Check the next expected I/O starting offset. If this doesn't
1792 * match, put it on the reorder queue.
1794 if ((buf->lba * dev->sector_size) != dev->next_completion_pos_bytes) {
1797 * If there is nothing on the queue, there is no sorting
1800 if (STAILQ_EMPTY(&dev->reorder_queue)) {
1801 STAILQ_INSERT_TAIL(&dev->reorder_queue, buf, links);
1802 dev->num_reorder_queue++;
1807 * Sort in ascending order by starting LBA. There should
1808 * be no identical LBAs.
1810 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1812 buf2 = STAILQ_NEXT(buf1, links);
1813 if (buf->lba < buf1->lba) {
1815 * If we're less than the first one, then
1816 * we insert at the head of the list
1817 * because this has to be the first element
1820 STAILQ_INSERT_HEAD(&dev->reorder_queue,
1822 dev->num_reorder_queue++;
1824 } else if (buf->lba > buf1->lba) {
1826 STAILQ_INSERT_TAIL(&dev->reorder_queue,
1828 dev->num_reorder_queue++;
1830 } else if (buf->lba < buf2->lba) {
1831 STAILQ_INSERT_AFTER(&dev->reorder_queue,
1833 dev->num_reorder_queue++;
1837 errx(1, "Found buffers with duplicate LBA %ju!",
1845 * We're the next expected I/O completion, so put ourselves
1846 * on the local queue to be sent to the writer. We use
1847 * work_links here so that we can queue this to the
1848 * peer_work_queue before taking the buffer off of the
1851 dev->next_completion_pos_bytes += buf->len;
1852 STAILQ_INSERT_TAIL(&local_queue, buf, work_links);
1855 * Go through the reorder queue looking for more sequential
1856 * I/O and add it to the local queue.
1858 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1859 buf1 = STAILQ_FIRST(&dev->reorder_queue)) {
1861 * As soon as we see an I/O that is out of sequence,
1864 if ((buf1->lba * dev->sector_size) !=
1865 dev->next_completion_pos_bytes)
1868 STAILQ_REMOVE_HEAD(&dev->reorder_queue, links);
1869 dev->num_reorder_queue--;
1870 STAILQ_INSERT_TAIL(&local_queue, buf1, work_links);
1871 dev->next_completion_pos_bytes += buf1->len;
1876 * Setup the event to let the other thread know that it has work
1879 EV_SET(&ke, (uintptr_t)&dev->peer_dev->work_queue, EVFILT_USER, 0,
1880 NOTE_TRIGGER, 0, NULL);
1883 * Put this on our shadow queue so that we know what we've queued
1884 * to the other thread.
1886 STAILQ_FOREACH_SAFE(buf1, &local_queue, work_links, buf2) {
1887 if (buf1->buf_type != CAMDD_BUF_DATA) {
1888 errx(1, "%s: should have a data buffer, not an "
1889 "indirect buffer", __func__);
1891 data = &buf1->buf_type_spec.data;
1894 * We only need to send one EOF to the writer, and don't
1895 * need to continue sending EOFs after that.
1897 if (buf1->status == CAMDD_STATUS_EOF) {
1898 if (dev->flags & CAMDD_DEV_FLAG_EOF_SENT) {
1899 STAILQ_REMOVE(&local_queue, buf1, camdd_buf,
1901 camdd_release_buf(buf1);
1905 dev->flags |= CAMDD_DEV_FLAG_EOF_SENT;
1909 STAILQ_INSERT_TAIL(&dev->peer_work_queue, buf1, links);
1910 peer_bytes_queued += (data->fill_len - data->resid);
1911 dev->peer_bytes_queued += (data->fill_len - data->resid);
1912 dev->num_peer_work_queue++;
1915 if (STAILQ_FIRST(&local_queue) == NULL)
1919 * Drop our mutex and pick up the other thread's mutex. We need to
1920 * do this to avoid deadlocks.
1922 pthread_mutex_unlock(&dev->mutex);
1923 pthread_mutex_lock(&dev->peer_dev->mutex);
1925 if (dev->peer_dev->flags & CAMDD_DEV_FLAG_ACTIVE) {
1927 * Put the buffers on the other thread's incoming work queue.
1929 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1930 buf1 = STAILQ_FIRST(&local_queue)) {
1931 STAILQ_REMOVE_HEAD(&local_queue, work_links);
1932 STAILQ_INSERT_TAIL(&dev->peer_dev->work_queue, buf1,
1936 * Send an event to the other thread's kqueue to let it know
1937 * that there is something on the work queue.
1939 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
1941 warn("%s: unable to add peer work_queue kevent",
1948 pthread_mutex_unlock(&dev->peer_dev->mutex);
1949 pthread_mutex_lock(&dev->mutex);
1952 * If the other side isn't active, run through the queue and
1953 * release all of the buffers.
1956 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1957 buf1 = STAILQ_FIRST(&local_queue)) {
1958 STAILQ_REMOVE_HEAD(&local_queue, work_links);
1959 STAILQ_REMOVE(&dev->peer_work_queue, buf1, camdd_buf,
1961 dev->num_peer_work_queue--;
1962 camdd_release_buf(buf1);
1964 dev->peer_bytes_queued -= peer_bytes_queued;
1973 * Return a buffer to the reader thread when we have completed writing it.
1976 camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf)
1982 * Setup the event to let the other thread know that we have
1983 * completed a buffer.
1985 EV_SET(&ke, (uintptr_t)&dev->peer_dev->peer_done_queue, EVFILT_USER, 0,
1986 NOTE_TRIGGER, 0, NULL);
1989 * Drop our lock and acquire the other thread's lock before
1992 pthread_mutex_unlock(&dev->mutex);
1993 pthread_mutex_lock(&dev->peer_dev->mutex);
1996 * Put the buffer on the reader thread's peer done queue now that
1997 * we have completed it.
1999 STAILQ_INSERT_TAIL(&dev->peer_dev->peer_done_queue, peer_buf,
2001 dev->peer_dev->num_peer_done_queue++;
2004 * Send an event to the peer thread to let it know that we've added
2005 * something to its peer done queue.
2007 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2009 warn("%s: unable to add peer_done_queue kevent", __func__);
2014 * Drop the other thread's lock and reacquire ours.
2016 pthread_mutex_unlock(&dev->peer_dev->mutex);
2017 pthread_mutex_lock(&dev->mutex);
2023 * Free a buffer that was written out by the writer thread and returned to
2024 * the reader thread.
2027 camdd_peer_done(struct camdd_buf *buf)
2029 struct camdd_dev *dev;
2030 struct camdd_buf_data *data;
2033 if (buf->buf_type != CAMDD_BUF_DATA) {
2034 errx(1, "%s: should have a data buffer, not an "
2035 "indirect buffer", __func__);
2038 data = &buf->buf_type_spec.data;
2040 STAILQ_REMOVE(&dev->peer_work_queue, buf, camdd_buf, links);
2041 dev->num_peer_work_queue--;
2042 dev->peer_bytes_queued -= (data->fill_len - data->resid);
2044 if (buf->status == CAMDD_STATUS_EOF)
2045 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2047 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2051 * Assumes caller holds the lock for this device.
2054 camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
2060 * If we're the reader, we need to send the completed I/O
2061 * to the writer. If we're the writer, we need to just
2062 * free up resources, or let the reader know if we've
2063 * encountered an error.
2065 if (dev->write_dev == 0) {
2066 retval = camdd_queue_peer_buf(dev, buf);
2070 struct camdd_buf *tmp_buf, *next_buf;
2072 STAILQ_FOREACH_SAFE(tmp_buf, &buf->src_list, src_links,
2074 struct camdd_buf *src_buf;
2075 struct camdd_buf_indirect *indirect;
2077 STAILQ_REMOVE(&buf->src_list, tmp_buf,
2078 camdd_buf, src_links);
2080 tmp_buf->status = buf->status;
2082 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
2083 camdd_complete_peer_buf(dev, tmp_buf);
2087 indirect = &tmp_buf->buf_type_spec.indirect;
2088 src_buf = indirect->src_buf;
2089 src_buf->refcount--;
2091 * XXX KDM we probably need to account for
2092 * exactly how many bytes we were able to
2093 * write. Allocate the residual to the
2094 * first N buffers? Or just track the
2095 * number of bytes written? Right now the reader
2096 * doesn't do anything with a residual.
2098 src_buf->status = buf->status;
2099 if (src_buf->refcount <= 0)
2100 camdd_complete_peer_buf(dev, src_buf);
2101 STAILQ_INSERT_TAIL(&dev->free_indirect_queue,
2105 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2110 * Fetch all completed commands from the pass(4) device.
2112 * Returns the number of commands received, or -1 if any of the commands
2113 * completed with an error. Returns 0 if no commands are available.
2116 camdd_pass_fetch(struct camdd_dev *dev)
2118 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2120 int retval = 0, num_fetched = 0, error_count = 0;
2122 pthread_mutex_unlock(&dev->mutex);
2124 * XXX KDM we don't distinguish between EFAULT and ENOENT.
2126 while ((retval = ioctl(pass_dev->dev->fd, CAMIOGET, &ccb)) != -1) {
2127 struct camdd_buf *buf;
2128 struct camdd_buf_data *data;
2129 cam_status ccb_status;
2132 buf = ccb.ccb_h.ccb_buf;
2133 data = &buf->buf_type_spec.data;
2134 buf_ccb = &data->ccb;
2139 * Copy the CCB back out so we get status, sense data, etc.
2141 bcopy(&ccb, buf_ccb, sizeof(ccb));
2143 pthread_mutex_lock(&dev->mutex);
2146 * We're now done, so take this off the active queue.
2148 STAILQ_REMOVE(&dev->active_queue, buf, camdd_buf, links);
2149 dev->cur_active_io--;
2151 ccb_status = ccb.ccb_h.status & CAM_STATUS_MASK;
2152 if (ccb_status != CAM_REQ_CMP) {
2153 cam_error_print(pass_dev->dev, &ccb, CAM_ESF_ALL,
2154 CAM_EPF_ALL, stderr);
2157 data->resid = ccb.csio.resid;
2158 dev->bytes_transferred += (ccb.csio.dxfer_len - ccb.csio.resid);
2160 if (buf->status == CAMDD_STATUS_NONE)
2161 buf->status = camdd_ccb_status(&ccb);
2162 if (buf->status == CAMDD_STATUS_ERROR)
2164 else if (buf->status == CAMDD_STATUS_EOF) {
2166 * Once we queue this buffer to our partner thread,
2167 * he will know that we've hit EOF.
2169 dev->flags |= CAMDD_DEV_FLAG_EOF;
2172 camdd_complete_buf(dev, buf, &error_count);
2175 * Unlock in preparation for the ioctl call.
2177 pthread_mutex_unlock(&dev->mutex);
2180 pthread_mutex_lock(&dev->mutex);
2182 if (error_count > 0)
2185 return (num_fetched);
2189 * Returns -1 for error, 0 for success/continue, and 1 for resource
2190 * shortage/stop processing.
2193 camdd_file_run(struct camdd_dev *dev)
2195 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
2196 struct camdd_buf_data *data;
2197 struct camdd_buf *buf;
2199 int retval = 0, write_dev = dev->write_dev;
2200 int error_count = 0, no_resources = 0, double_buf_needed = 0;
2201 uint32_t num_sectors = 0, db_len = 0;
2203 buf = STAILQ_FIRST(&dev->run_queue);
2207 } else if ((dev->write_dev == 0)
2208 && (dev->flags & (CAMDD_DEV_FLAG_EOF |
2209 CAMDD_DEV_FLAG_EOF_SENT))) {
2210 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2211 dev->num_run_queue--;
2212 buf->status = CAMDD_STATUS_EOF;
2218 * If we're writing, we need to go through the source buffer list
2219 * and create an S/G list.
2221 if (write_dev != 0) {
2222 retval = camdd_buf_sg_create(buf, /*iovec*/ 1,
2223 dev->sector_size, &num_sectors, &double_buf_needed);
2230 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2231 dev->num_run_queue--;
2233 data = &buf->buf_type_spec.data;
2236 * pread(2) and pwrite(2) offsets are byte offsets.
2238 io_offset = buf->lba * dev->sector_size;
2241 * Unlock the mutex while we read or write.
2243 pthread_mutex_unlock(&dev->mutex);
2246 * Note that we don't need to double buffer if we're the reader
2247 * because in that case, we have allocated a single buffer of
2248 * sufficient size to do the read. This copy is necessary on
2249 * writes because if one of the components of the S/G list is not
2250 * a sector size multiple, the kernel will reject the write. This
2251 * is unfortunate but not surprising. So this will make sure that
2252 * we're using a single buffer that is a multiple of the sector size.
2254 if ((double_buf_needed != 0)
2255 && (data->sg_count > 1)
2256 && (write_dev != 0)) {
2257 uint32_t cur_offset;
2260 if (file_dev->tmp_buf == NULL)
2261 file_dev->tmp_buf = calloc(dev->blocksize, 1);
2262 if (file_dev->tmp_buf == NULL) {
2263 buf->status = CAMDD_STATUS_ERROR;
2267 for (i = 0, cur_offset = 0; i < data->sg_count; i++) {
2268 bcopy(data->iovec[i].iov_base,
2269 &file_dev->tmp_buf[cur_offset],
2270 data->iovec[i].iov_len);
2271 cur_offset += data->iovec[i].iov_len;
2273 db_len = cur_offset;
2276 if (file_dev->file_flags & CAMDD_FF_CAN_SEEK) {
2277 if (write_dev == 0) {
2279 * XXX KDM is there any way we would need a S/G
2282 retval = pread(file_dev->fd, data->buf,
2283 buf->len, io_offset);
2285 if (double_buf_needed != 0) {
2286 retval = pwrite(file_dev->fd, file_dev->tmp_buf,
2288 } else if (data->sg_count == 0) {
2289 retval = pwrite(file_dev->fd, data->buf,
2290 data->fill_len, io_offset);
2292 retval = pwritev(file_dev->fd, data->iovec,
2293 data->sg_count, io_offset);
2297 if (write_dev == 0) {
2299 * XXX KDM is there any way we would need a S/G
2302 retval = read(file_dev->fd, data->buf, buf->len);
2304 if (double_buf_needed != 0) {
2305 retval = write(file_dev->fd, file_dev->tmp_buf,
2307 } else if (data->sg_count == 0) {
2308 retval = write(file_dev->fd, data->buf,
2311 retval = writev(file_dev->fd, data->iovec,
2317 /* We're done, re-acquire the lock */
2318 pthread_mutex_lock(&dev->mutex);
2320 if (retval >= (ssize_t)data->fill_len) {
2322 * If the bytes transferred is more than the request size,
2323 * that indicates an overrun, which should only happen at
2324 * the end of a transfer if we have to round up to a sector
2327 if (buf->status == CAMDD_STATUS_NONE)
2328 buf->status = CAMDD_STATUS_OK;
2330 dev->bytes_transferred += retval;
2331 } else if (retval == -1) {
2332 warn("Error %s %s", (write_dev) ? "writing to" :
2333 "reading from", file_dev->filename);
2335 buf->status = CAMDD_STATUS_ERROR;
2336 data->resid = data->fill_len;
2339 if (dev->debug == 0)
2342 if ((double_buf_needed != 0)
2343 && (write_dev != 0)) {
2344 fprintf(stderr, "%s: fd %d, DB buf %p, len %u lba %ju "
2345 "offset %ju\n", __func__, file_dev->fd,
2346 file_dev->tmp_buf, db_len, (uintmax_t)buf->lba,
2347 (uintmax_t)io_offset);
2348 } else if (data->sg_count == 0) {
2349 fprintf(stderr, "%s: fd %d, buf %p, len %u, lba %ju "
2350 "offset %ju\n", __func__, file_dev->fd, data->buf,
2351 data->fill_len, (uintmax_t)buf->lba,
2352 (uintmax_t)io_offset);
2356 fprintf(stderr, "%s: fd %d, len %u, lba %ju "
2357 "offset %ju\n", __func__, file_dev->fd,
2358 data->fill_len, (uintmax_t)buf->lba,
2359 (uintmax_t)io_offset);
2361 for (i = 0; i < data->sg_count; i++) {
2362 fprintf(stderr, "index %d ptr %p len %zu\n",
2363 i, data->iovec[i].iov_base,
2364 data->iovec[i].iov_len);
2367 } else if (retval == 0) {
2368 buf->status = CAMDD_STATUS_EOF;
2369 if (dev->debug != 0)
2370 printf("%s: got EOF from %s!\n", __func__,
2371 file_dev->filename);
2372 data->resid = data->fill_len;
2374 } else if (retval < (ssize_t)data->fill_len) {
2375 if (buf->status == CAMDD_STATUS_NONE)
2376 buf->status = CAMDD_STATUS_SHORT_IO;
2377 data->resid = data->fill_len - retval;
2378 dev->bytes_transferred += retval;
2383 if (buf->status == CAMDD_STATUS_EOF) {
2384 struct camdd_buf *buf2;
2385 dev->flags |= CAMDD_DEV_FLAG_EOF;
2386 STAILQ_FOREACH(buf2, &dev->run_queue, links)
2387 buf2->status = CAMDD_STATUS_EOF;
2390 camdd_complete_buf(dev, buf, &error_count);
2393 if (error_count != 0)
2395 else if (no_resources != 0)
2402 * Execute one command from the run queue. Returns 0 for success, 1 for
2403 * stop processing, and -1 for error.
2406 camdd_pass_run(struct camdd_dev *dev)
2408 struct camdd_buf *buf = NULL;
2409 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2410 struct camdd_buf_data *data;
2411 uint32_t num_blocks, sectors_used = 0;
2413 int retval = 0, is_write = dev->write_dev;
2414 int double_buf_needed = 0;
2416 buf = STAILQ_FIRST(&dev->run_queue);
2423 * If we're writing, we need to go through the source buffer list
2424 * and create an S/G list.
2426 if (is_write != 0) {
2427 retval = camdd_buf_sg_create(buf, /*iovec*/ 0,dev->sector_size,
2428 §ors_used, &double_buf_needed);
2435 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2436 dev->num_run_queue--;
2438 data = &buf->buf_type_spec.data;
2441 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
2444 * In almost every case the number of blocks should be the device
2445 * block size. The exception may be at the end of an I/O stream
2446 * for a partial block or at the end of a device.
2449 num_blocks = sectors_used;
2451 num_blocks = data->fill_len / pass_dev->block_len;
2453 scsi_read_write(&ccb->csio,
2454 /*retries*/ dev->retry_count,
2456 /*tag_action*/ MSG_SIMPLE_Q_TAG,
2457 /*readop*/ (dev->write_dev == 0) ? SCSI_RW_READ :
2460 /*minimum_cmd_size*/ dev->min_cmd_size,
2462 /*block_count*/ num_blocks,
2463 /*data_ptr*/ (data->sg_count != 0) ?
2464 (uint8_t *)data->segs : data->buf,
2465 /*dxfer_len*/ (num_blocks * pass_dev->block_len),
2466 /*sense_len*/ SSD_FULL_SIZE,
2467 /*timeout*/ dev->io_timeout);
2469 /* Disable freezing the device queue */
2470 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
2472 if (dev->retry_count != 0)
2473 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
2475 if (data->sg_count != 0) {
2476 ccb->csio.sglist_cnt = data->sg_count;
2477 ccb->ccb_h.flags |= CAM_DATA_SG;
2481 * Store a pointer to the buffer in the CCB. The kernel will
2482 * restore this when we get it back, and we'll use it to identify
2483 * the buffer this CCB came from.
2485 ccb->ccb_h.ccb_buf = buf;
2488 * Unlock our mutex in preparation for issuing the ioctl.
2490 pthread_mutex_unlock(&dev->mutex);
2492 * Queue the CCB to the pass(4) driver.
2494 if (ioctl(pass_dev->dev->fd, CAMIOQUEUE, ccb) == -1) {
2495 pthread_mutex_lock(&dev->mutex);
2497 warn("%s: error sending CAMIOQUEUE ioctl to %s%u", __func__,
2498 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
2499 warn("%s: CCB address is %p", __func__, ccb);
2502 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2504 pthread_mutex_lock(&dev->mutex);
2506 dev->cur_active_io++;
2507 STAILQ_INSERT_TAIL(&dev->active_queue, buf, links);
2515 camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len)
2517 struct camdd_dev_pass *pass_dev;
2518 uint32_t num_blocks;
2521 pass_dev = &dev->dev_spec.pass;
2523 *lba = dev->next_io_pos_bytes / dev->sector_size;
2524 *len = dev->blocksize;
2525 num_blocks = *len / dev->sector_size;
2528 * If max_sector is 0, then we have no set limit. This can happen
2529 * if we're writing to a file in a filesystem, or reading from
2530 * something like /dev/zero.
2532 if ((dev->max_sector != 0)
2533 || (dev->sector_io_limit != 0)) {
2534 uint64_t max_sector;
2536 if ((dev->max_sector != 0)
2537 && (dev->sector_io_limit != 0))
2538 max_sector = min(dev->sector_io_limit, dev->max_sector);
2539 else if (dev->max_sector != 0)
2540 max_sector = dev->max_sector;
2542 max_sector = dev->sector_io_limit;
2546 * Check to see whether we're starting off past the end of
2547 * the device. If so, we need to just send an EOF
2548 * notification to the writer.
2550 if (*lba > max_sector) {
2553 } else if (((*lba + num_blocks) > max_sector + 1)
2554 || ((*lba + num_blocks) < *lba)) {
2556 * If we get here (but pass the first check), we
2557 * can trim the request length down to go to the
2558 * end of the device.
2560 num_blocks = (max_sector + 1) - *lba;
2561 *len = num_blocks * dev->sector_size;
2566 dev->next_io_pos_bytes += *len;
2572 * Returns 0 for success, 1 for EOF detected, and -1 for failure.
2575 camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf)
2577 struct camdd_buf *buf = NULL;
2578 struct camdd_buf_data *data;
2579 struct camdd_dev_pass *pass_dev;
2581 struct camdd_buf_data *rb_data;
2582 int is_write = dev->write_dev;
2583 int eof_flush_needed = 0;
2587 pass_dev = &dev->dev_spec.pass;
2590 * If we've gotten EOF or our partner has, we should not continue
2591 * queueing I/O. If we're a writer, though, we should continue
2592 * to write any buffers that don't have EOF status.
2594 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
2595 || ((dev->flags & CAMDD_DEV_FLAG_PEER_EOF)
2596 && (is_write == 0))) {
2598 * Tell the worker thread that we have seen EOF.
2603 * If we're the writer, send the buffer back with EOF status.
2606 read_buf->status = CAMDD_STATUS_EOF;
2608 error = camdd_complete_peer_buf(dev, read_buf);
2613 if (is_write == 0) {
2614 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2619 data = &buf->buf_type_spec.data;
2621 retval = camdd_get_next_lba_len(dev, &buf->lba, &buf->len);
2623 buf->status = CAMDD_STATUS_EOF;
2626 && ((dev->flags & (CAMDD_DEV_FLAG_EOF_SENT |
2627 CAMDD_DEV_FLAG_EOF_QUEUED)) != 0)) {
2628 camdd_release_buf(buf);
2631 dev->flags |= CAMDD_DEV_FLAG_EOF_QUEUED;
2634 data->fill_len = buf->len;
2635 data->src_start_offset = buf->lba * dev->sector_size;
2638 * Put this on the run queue.
2640 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2641 dev->num_run_queue++;
2648 * Check for new EOF status from the reader.
2650 if ((read_buf->status == CAMDD_STATUS_EOF)
2651 || (read_buf->status == CAMDD_STATUS_ERROR)) {
2652 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2653 if ((STAILQ_FIRST(&dev->pending_queue) == NULL)
2654 && (read_buf->len == 0)) {
2655 camdd_complete_peer_buf(dev, read_buf);
2659 eof_flush_needed = 1;
2663 * See if we have a buffer we're composing with pieces from our
2666 buf = STAILQ_FIRST(&dev->pending_queue);
2671 retval = camdd_get_next_lba_len(dev, &lba, &len);
2673 read_buf->status = CAMDD_STATUS_EOF;
2676 dev->flags |= CAMDD_DEV_FLAG_EOF;
2677 error = camdd_complete_peer_buf(dev, read_buf);
2683 * If we don't have a pending buffer, we need to grab a new
2684 * one from the free list or allocate another one.
2686 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2695 STAILQ_INSERT_TAIL(&dev->pending_queue, buf, links);
2696 dev->num_pending_queue++;
2699 data = &buf->buf_type_spec.data;
2701 rb_data = &read_buf->buf_type_spec.data;
2703 if ((rb_data->src_start_offset != dev->next_peer_pos_bytes)
2704 && (dev->debug != 0)) {
2705 printf("%s: WARNING: reader offset %#jx != expected offset "
2706 "%#jx\n", __func__, (uintmax_t)rb_data->src_start_offset,
2707 (uintmax_t)dev->next_peer_pos_bytes);
2709 dev->next_peer_pos_bytes = rb_data->src_start_offset +
2710 (rb_data->fill_len - rb_data->resid);
2712 new_len = (rb_data->fill_len - rb_data->resid) + data->fill_len;
2713 if (new_len < buf->len) {
2715 * There are three cases here:
2716 * 1. We need more data to fill up a block, so we put
2717 * this I/O on the queue and wait for more I/O.
2718 * 2. We have a pending buffer in the queue that is
2719 * smaller than our blocksize, but we got an EOF. So we
2720 * need to go ahead and flush the write out.
2721 * 3. We got an error.
2725 * Increment our fill length.
2727 data->fill_len += (rb_data->fill_len - rb_data->resid);
2730 * Add the new read buffer to the list for writing.
2732 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2734 /* Increment the count */
2737 if (eof_flush_needed == 0) {
2739 * We need to exit, because we don't have enough
2745 * Take the buffer off of the pending queue.
2747 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2749 dev->num_pending_queue--;
2752 * If we need an EOF flush, but there is no data
2753 * to flush, go ahead and return this buffer.
2755 if (data->fill_len == 0) {
2756 camdd_complete_buf(dev, buf, /*error_count*/0);
2762 * Put this on the next queue for execution.
2764 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2765 dev->num_run_queue++;
2767 } else if (new_len == buf->len) {
2769 * We have enough data to completey fill one block,
2770 * so we're ready to issue the I/O.
2774 * Take the buffer off of the pending queue.
2776 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf, links);
2777 dev->num_pending_queue--;
2780 * Add the new read buffer to the list for writing.
2782 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2784 /* Increment the count */
2788 * Increment our fill length.
2790 data->fill_len += (rb_data->fill_len - rb_data->resid);
2793 * Put this on the next queue for execution.
2795 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2796 dev->num_run_queue++;
2798 struct camdd_buf *idb;
2799 struct camdd_buf_indirect *indirect;
2800 uint32_t len_to_go, cur_offset;
2803 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2808 indirect = &idb->buf_type_spec.indirect;
2809 indirect->src_buf = read_buf;
2810 read_buf->refcount++;
2811 indirect->offset = 0;
2812 indirect->start_ptr = rb_data->buf;
2814 * We've already established that there is more
2815 * data in read_buf than we have room for in our
2816 * current write request. So this particular chunk
2817 * of the request should just be the remainder
2818 * needed to fill up a block.
2820 indirect->len = buf->len - (data->fill_len - data->resid);
2822 camdd_buf_add_child(buf, idb);
2825 * This buffer is ready to execute, so we can take
2826 * it off the pending queue and put it on the run
2829 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2831 dev->num_pending_queue--;
2832 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2833 dev->num_run_queue++;
2835 cur_offset = indirect->offset + indirect->len;
2838 * The resulting I/O would be too large to fit in
2839 * one block. We need to split this I/O into
2840 * multiple pieces. Allocate as many buffers as needed.
2842 for (len_to_go = rb_data->fill_len - rb_data->resid -
2843 indirect->len; len_to_go > 0;) {
2844 struct camdd_buf *new_buf;
2845 struct camdd_buf_data *new_data;
2849 retval = camdd_get_next_lba_len(dev, &lba, &len);
2853 * The device has already been marked
2854 * as EOF, and there is no space left.
2859 new_buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2860 if (new_buf == NULL) {
2868 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2874 indirect = &idb->buf_type_spec.indirect;
2876 indirect->src_buf = read_buf;
2877 read_buf->refcount++;
2878 indirect->offset = cur_offset;
2879 indirect->start_ptr = rb_data->buf + cur_offset;
2880 indirect->len = min(len_to_go, new_buf->len);
2882 if (((indirect->len % dev->sector_size) != 0)
2883 || ((indirect->offset % dev->sector_size) != 0)) {
2884 warnx("offset %ju len %ju not aligned with "
2885 "sector size %u", indirect->offset,
2886 (uintmax_t)indirect->len, dev->sector_size);
2889 cur_offset += indirect->len;
2890 len_to_go -= indirect->len;
2892 camdd_buf_add_child(new_buf, idb);
2894 new_data = &new_buf->buf_type_spec.data;
2896 if ((new_data->fill_len == new_buf->len)
2897 || (eof_flush_needed != 0)) {
2898 STAILQ_INSERT_TAIL(&dev->run_queue,
2900 dev->num_run_queue++;
2901 } else if (new_data->fill_len < buf->len) {
2902 STAILQ_INSERT_TAIL(&dev->pending_queue,
2904 dev->num_pending_queue++;
2906 warnx("%s: too much data in new "
2907 "buffer!", __func__);
2919 camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
2920 uint32_t *peer_depth, uint32_t *our_bytes, uint32_t *peer_bytes)
2922 *our_depth = dev->cur_active_io + dev->num_run_queue;
2923 if (dev->num_peer_work_queue >
2924 dev->num_peer_done_queue)
2925 *peer_depth = dev->num_peer_work_queue -
2926 dev->num_peer_done_queue;
2929 *our_bytes = *our_depth * dev->blocksize;
2930 *peer_bytes = dev->peer_bytes_queued;
2934 camdd_sig_handler(int sig)
2943 sem_post(&camdd_sem);
2947 camdd_print_status(struct camdd_dev *camdd_dev, struct camdd_dev *other_dev,
2948 struct timespec *start_time)
2950 struct timespec done_time;
2952 long double mb_sec, total_sec;
2955 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &done_time);
2957 warn("Unable to get done time");
2961 timespecsub(&done_time, start_time);
2963 total_ns = done_time.tv_nsec + (done_time.tv_sec * 1000000000);
2964 total_sec = total_ns;
2965 total_sec /= 1000000000;
2967 fprintf(stderr, "%ju bytes %s %s\n%ju bytes %s %s\n"
2968 "%.4Lf seconds elapsed\n",
2969 (uintmax_t)camdd_dev->bytes_transferred,
2970 (camdd_dev->write_dev == 0) ? "read from" : "written to",
2971 camdd_dev->device_name,
2972 (uintmax_t)other_dev->bytes_transferred,
2973 (other_dev->write_dev == 0) ? "read from" : "written to",
2974 other_dev->device_name, total_sec);
2976 mb_sec = min(other_dev->bytes_transferred,camdd_dev->bytes_transferred);
2977 mb_sec /= 1024 * 1024;
2978 mb_sec *= 1000000000;
2980 fprintf(stderr, "%.2Lf MB/sec\n", mb_sec);
2984 camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts, uint64_t max_io,
2985 int retry_count, int timeout)
2987 char *device = NULL;
2988 struct cam_device *new_cam_dev = NULL;
2989 struct camdd_dev *devs[2];
2990 struct timespec start_time;
2991 pthread_t threads[2];
2996 if (num_io_opts != 2) {
2997 warnx("Must have one input and one output path");
3002 bzero(devs, sizeof(devs));
3004 for (i = 0; i < num_io_opts; i++) {
3005 switch (io_opts[i].dev_type) {
3006 case CAMDD_DEV_PASS: {
3007 camdd_argmask new_arglist = CAMDD_ARG_NONE;
3008 int bus = 0, target = 0, lun = 0;
3012 if (isdigit(io_opts[i].dev_name[0])) {
3013 /* device specified as bus:target[:lun] */
3014 rv = parse_btl(io_opts[i].dev_name, &bus,
3015 &target, &lun, &new_arglist);
3017 warnx("numeric device specification "
3018 "must be either bus:target, or "
3023 /* default to 0 if lun was not specified */
3024 if ((new_arglist & CAMDD_ARG_LUN) == 0) {
3026 new_arglist |= CAMDD_ARG_LUN;
3029 if (cam_get_device(io_opts[i].dev_name, name,
3030 sizeof name, &unit) == -1) {
3031 warnx("%s", cam_errbuf);
3035 device = strdup(name);
3036 new_arglist |= CAMDD_ARG_DEVICE |CAMDD_ARG_UNIT;
3039 if (new_arglist & (CAMDD_ARG_BUS | CAMDD_ARG_TARGET))
3040 new_cam_dev = cam_open_btl(bus, target, lun,
3043 new_cam_dev = cam_open_spec_device(device, unit,
3045 if (new_cam_dev == NULL) {
3046 warnx("%s", cam_errbuf);
3051 devs[i] = camdd_probe_pass(new_cam_dev,
3052 /*io_opts*/ &io_opts[i],
3053 CAMDD_ARG_ERR_RECOVER,
3054 /*probe_retry_count*/ 3,
3055 /*probe_timeout*/ 5000,
3056 /*io_retry_count*/ retry_count,
3057 /*io_timeout*/ timeout);
3058 if (devs[i] == NULL) {
3059 warn("Unable to probe device %s%u",
3060 new_cam_dev->device_name,
3061 new_cam_dev->dev_unit_num);
3067 case CAMDD_DEV_FILE: {
3070 if (io_opts[i].dev_name[0] == '-') {
3071 if (io_opts[i].write_dev != 0)
3076 if (io_opts[i].write_dev != 0) {
3077 fd = open(io_opts[i].dev_name,
3078 O_RDWR | O_CREAT, S_IWUSR |S_IRUSR);
3080 fd = open(io_opts[i].dev_name,
3085 warn("error opening file %s",
3086 io_opts[i].dev_name);
3091 devs[i] = camdd_probe_file(fd, &io_opts[i],
3092 retry_count, timeout);
3093 if (devs[i] == NULL) {
3101 warnx("Unknown device type %d (%s)",
3102 io_opts[i].dev_type, io_opts[i].dev_name);
3105 break; /*NOTREACHED */
3108 devs[i]->write_dev = io_opts[i].write_dev;
3110 devs[i]->start_offset_bytes = io_opts[i].offset;
3113 devs[i]->sector_io_limit =
3114 (devs[i]->start_offset_bytes /
3115 devs[i]->sector_size) +
3116 (max_io / devs[i]->sector_size) - 1;
3117 devs[i]->sector_io_limit =
3118 (devs[i]->start_offset_bytes /
3119 devs[i]->sector_size) +
3120 (max_io / devs[i]->sector_size) - 1;
3123 devs[i]->next_io_pos_bytes = devs[i]->start_offset_bytes;
3124 devs[i]->next_completion_pos_bytes =devs[i]->start_offset_bytes;
3127 devs[0]->peer_dev = devs[1];
3128 devs[1]->peer_dev = devs[0];
3129 devs[0]->next_peer_pos_bytes = devs[0]->peer_dev->next_io_pos_bytes;
3130 devs[1]->next_peer_pos_bytes = devs[1]->peer_dev->next_io_pos_bytes;
3132 sem_init(&camdd_sem, /*pshared*/ 0, 0);
3134 signal(SIGINFO, camdd_sig_handler);
3135 signal(SIGINT, camdd_sig_handler);
3137 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &start_time);
3139 warn("Unable to get start time");
3143 for (i = 0; i < num_io_opts; i++) {
3144 error = pthread_create(&threads[i], NULL, camdd_worker,
3147 warnc(error, "pthread_create() failed");
3153 if ((sem_wait(&camdd_sem) == -1)
3154 || (need_exit != 0)) {
3157 for (i = 0; i < num_io_opts; i++) {
3158 EV_SET(&ke, (uintptr_t)&devs[i]->work_queue,
3159 EVFILT_USER, 0, NOTE_TRIGGER, 0, NULL);
3161 devs[i]->flags |= CAMDD_DEV_FLAG_EOF;
3163 error = kevent(devs[i]->kq, &ke, 1, NULL, 0,
3166 warn("%s: unable to wake up thread",
3171 } else if (need_status != 0) {
3172 camdd_print_status(devs[0], devs[1], &start_time);
3176 for (i = 0; i < num_io_opts; i++) {
3177 pthread_join(threads[i], NULL);
3180 camdd_print_status(devs[0], devs[1], &start_time);
3184 for (i = 0; i < num_io_opts; i++)
3185 camdd_free_dev(devs[i]);
3187 return (error + error_exit);
3194 "usage: camdd <-i|-o pass=pass0,bs=1M,offset=1M,depth=4>\n"
3195 " <-i|-o file=/tmp/file,bs=512K,offset=1M>\n"
3196 " <-i|-o file=/dev/da0,bs=512K,offset=1M>\n"
3197 " <-i|-o file=/dev/nsa0,bs=512K>\n"
3198 " [-C retry_count][-E][-m max_io_amt][-t timeout_secs][-v][-h]\n"
3199 "Option description\n"
3200 "-i <arg=val> Specify input device/file and parameters\n"
3201 "-o <arg=val> Specify output device/file and parameters\n"
3202 "Input and Output parameters\n"
3203 "pass=name Specify a pass(4) device like pass0 or /dev/pass0\n"
3204 "file=name Specify a file or device, /tmp/foo, /dev/da0, /dev/null\n"
3205 " or - for stdin/stdout\n"
3206 "bs=blocksize Specify blocksize in bytes, or using K, M, G, etc. suffix\n"
3207 "offset=len Specify starting offset in bytes or using K, M, G suffix\n"
3208 " NOTE: offset cannot be specified on tapes, pipes, stdin/out\n"
3209 "depth=N Specify a numeric queue depth. This only applies to pass(4)\n"
3210 "mcs=N Specify a minimum cmd size for pass(4) read/write commands\n"
3211 "Optional arguments\n"
3212 "-C retry_cnt Specify a retry count for pass(4) devices\n"
3213 "-E Enable CAM error recovery for pass(4) devices\n"
3214 "-m max_io Specify the maximum amount to be transferred in bytes or\n"
3215 " using K, G, M, etc. suffixes\n"
3216 "-t timeout Specify the I/O timeout to use with pass(4) devices\n"
3217 "-v Enable verbose error recovery\n"
3218 "-h Print this message\n");
3223 camdd_parse_io_opts(char *args, int is_write, struct camdd_io_opts *io_opts)
3225 char *tmpstr, *tmpstr2;
3226 char *orig_tmpstr = NULL;
3229 io_opts->write_dev = is_write;
3231 tmpstr = strdup(args);
3232 if (tmpstr == NULL) {
3233 warn("strdup failed");
3237 orig_tmpstr = tmpstr;
3238 while ((tmpstr2 = strsep(&tmpstr, ",")) != NULL) {
3242 * If the user creates an empty parameter by putting in two
3243 * commas, skip over it and look for the next field.
3245 if (*tmpstr2 == '\0')
3248 name = strsep(&tmpstr2, "=");
3249 if (*name == '\0') {
3250 warnx("Got empty I/O parameter name");
3254 value = strsep(&tmpstr2, "=");
3256 || (*value == '\0')) {
3257 warnx("Empty I/O parameter value for %s", name);
3261 if (strncasecmp(name, "file", 4) == 0) {
3262 io_opts->dev_type = CAMDD_DEV_FILE;
3263 io_opts->dev_name = strdup(value);
3264 if (io_opts->dev_name == NULL) {
3265 warn("Error allocating memory");
3269 } else if (strncasecmp(name, "pass", 4) == 0) {
3270 io_opts->dev_type = CAMDD_DEV_PASS;
3271 io_opts->dev_name = strdup(value);
3272 if (io_opts->dev_name == NULL) {
3273 warn("Error allocating memory");
3277 } else if ((strncasecmp(name, "bs", 2) == 0)
3278 || (strncasecmp(name, "blocksize", 9) == 0)) {
3279 retval = expand_number(value, &io_opts->blocksize);
3281 warn("expand_number(3) failed on %s=%s", name,
3286 } else if (strncasecmp(name, "depth", 5) == 0) {
3289 io_opts->queue_depth = strtoull(value, &endptr, 0);
3290 if (*endptr != '\0') {
3291 warnx("invalid queue depth %s", value);
3295 } else if (strncasecmp(name, "mcs", 3) == 0) {
3298 io_opts->min_cmd_size = strtol(value, &endptr, 0);
3299 if ((*endptr != '\0')
3300 || ((io_opts->min_cmd_size > 16)
3301 || (io_opts->min_cmd_size < 0))) {
3302 warnx("invalid minimum cmd size %s", value);
3306 } else if (strncasecmp(name, "offset", 6) == 0) {
3307 retval = expand_number(value, &io_opts->offset);
3309 warn("expand_number(3) failed on %s=%s", name,
3314 } else if (strncasecmp(name, "debug", 5) == 0) {
3317 io_opts->debug = strtoull(value, &endptr, 0);
3318 if (*endptr != '\0') {
3319 warnx("invalid debug level %s", value);
3324 warnx("Unrecognized parameter %s=%s", name, value);
3334 main(int argc, char **argv)
3337 camdd_argmask arglist = CAMDD_ARG_NONE;
3338 int timeout = 0, retry_count = 1;
3340 uint64_t max_io = 0;
3341 struct camdd_io_opts *opt_list = NULL;
3348 opt_list = calloc(2, sizeof(struct camdd_io_opts));
3349 if (opt_list == NULL) {
3350 warn("Unable to allocate option list");
3355 while ((c = getopt(argc, argv, "C:Ehi:m:o:t:v")) != -1){
3358 retry_count = strtol(optarg, NULL, 0);
3359 if (retry_count < 0)
3360 errx(1, "retry count %d is < 0",
3362 arglist |= CAMDD_ARG_RETRIES;
3365 arglist |= CAMDD_ARG_ERR_RECOVER;
3370 && (opt_list[0].dev_type != CAMDD_DEV_NONE))
3372 && (opt_list[1].dev_type != CAMDD_DEV_NONE))) {
3373 errx(1, "Only one input and output path "
3376 error = camdd_parse_io_opts(optarg, (c == 'o') ? 1 : 0,
3377 (c == 'o') ? &opt_list[1] : &opt_list[0]);
3382 error = expand_number(optarg, &max_io);
3384 warn("invalid maximum I/O amount %s", optarg);
3390 timeout = strtol(optarg, NULL, 0);
3392 errx(1, "invalid timeout %d", timeout);
3393 /* Convert the timeout from seconds to ms */
3395 arglist |= CAMDD_ARG_TIMEOUT;
3398 arglist |= CAMDD_ARG_VERBOSE;
3404 break; /*NOTREACHED*/
3408 if ((opt_list[0].dev_type == CAMDD_DEV_NONE)
3409 || (opt_list[1].dev_type == CAMDD_DEV_NONE))
3410 errx(1, "Must specify both -i and -o");
3413 * Set the timeout if the user hasn't specified one.
3416 timeout = CAMDD_PASS_RW_TIMEOUT;
3418 error = camdd_rw(opt_list, 2, max_io, retry_count, timeout);