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 sig_atomic_t need_exit = 0;
424 static sig_atomic_t error_exit = 0;
425 static sig_atomic_t 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 = calloc(1, sizeof(*dev));
596 warn("%s: unable to malloc %zu bytes", __func__, sizeof(*dev));
600 dev->dev_type = dev_type;
601 dev->io_timeout = timeout;
602 dev->retry_count = retry_count;
603 STAILQ_INIT(&dev->free_queue);
604 STAILQ_INIT(&dev->free_indirect_queue);
605 STAILQ_INIT(&dev->active_queue);
606 STAILQ_INIT(&dev->pending_queue);
607 STAILQ_INIT(&dev->run_queue);
608 STAILQ_INIT(&dev->reorder_queue);
609 STAILQ_INIT(&dev->work_queue);
610 STAILQ_INIT(&dev->peer_done_queue);
611 STAILQ_INIT(&dev->peer_work_queue);
612 retval = pthread_mutex_init(&dev->mutex, NULL);
614 warnc(retval, "%s: failed to initialize mutex", __func__);
618 retval = pthread_cond_init(&dev->cond, NULL);
620 warnc(retval, "%s: failed to initialize condition variable",
627 warn("%s: Unable to create kqueue", __func__);
631 ke_size = sizeof(struct kevent) * (num_ke + 4);
632 ke = calloc(1, ke_size);
634 warn("%s: unable to malloc %zu bytes", __func__, ke_size);
638 bcopy(new_ke, ke, num_ke * sizeof(struct kevent));
640 EV_SET(&ke[num_ke++], (uintptr_t)&dev->work_queue, EVFILT_USER,
641 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
642 EV_SET(&ke[num_ke++], (uintptr_t)&dev->peer_done_queue, EVFILT_USER,
643 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
644 EV_SET(&ke[num_ke++], SIGINFO, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
645 EV_SET(&ke[num_ke++], SIGINT, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
647 retval = kevent(dev->kq, ke, num_ke, NULL, 0, NULL);
649 warn("%s: Unable to register kevents", __func__);
662 static struct camdd_buf *
663 camdd_alloc_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
665 struct camdd_buf *buf = NULL;
666 uint8_t *data_ptr = NULL;
669 * We only need to allocate data space for data buffers.
673 data_ptr = malloc(dev->blocksize);
674 if (data_ptr == NULL) {
675 warn("unable to allocate %u bytes", dev->blocksize);
683 buf = calloc(1, sizeof(*buf));
685 warn("unable to allocate %zu bytes", sizeof(*buf));
689 buf->buf_type = buf_type;
692 case CAMDD_BUF_DATA: {
693 struct camdd_buf_data *data;
695 data = &buf->buf_type_spec.data;
697 data->alloc_len = dev->blocksize;
698 data->buf = data_ptr;
701 case CAMDD_BUF_INDIRECT:
706 STAILQ_INIT(&buf->src_list);
717 camdd_release_buf(struct camdd_buf *buf)
719 struct camdd_dev *dev;
723 switch (buf->buf_type) {
724 case CAMDD_BUF_DATA: {
725 struct camdd_buf_data *data;
727 data = &buf->buf_type_spec.data;
729 if (data->segs != NULL) {
730 if (data->extra_buf != 0) {
734 data->segs[data->sg_count - 1].ds_addr;
741 } else if (data->iovec != NULL) {
742 if (data->extra_buf != 0) {
743 free(data->iovec[data->sg_count - 1].iov_base);
750 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
753 case CAMDD_BUF_INDIRECT:
754 STAILQ_INSERT_TAIL(&dev->free_indirect_queue, buf, links);
757 err(1, "%s: Invalid buffer type %d for released buffer",
758 __func__, buf->buf_type);
764 camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
766 struct camdd_buf *buf = NULL;
770 buf = STAILQ_FIRST(&dev->free_queue);
772 struct camdd_buf_data *data;
776 STAILQ_REMOVE_HEAD(&dev->free_queue, links);
777 data = &buf->buf_type_spec.data;
778 data_ptr = data->buf;
779 alloc_len = data->alloc_len;
780 bzero(buf, sizeof(*buf));
781 data->buf = data_ptr;
782 data->alloc_len = alloc_len;
785 case CAMDD_BUF_INDIRECT:
786 buf = STAILQ_FIRST(&dev->free_indirect_queue);
788 STAILQ_REMOVE_HEAD(&dev->free_indirect_queue, links);
790 bzero(buf, sizeof(*buf));
794 warnx("Unknown buffer type %d requested", buf_type);
800 return (camdd_alloc_buf(dev, buf_type));
802 STAILQ_INIT(&buf->src_list);
804 buf->buf_type = buf_type;
811 camdd_buf_sg_create(struct camdd_buf *buf, int iovec, uint32_t sector_size,
812 uint32_t *num_sectors_used, int *double_buf_needed)
814 struct camdd_buf *tmp_buf;
815 struct camdd_buf_data *data;
816 uint8_t *extra_buf = NULL;
817 size_t extra_buf_len = 0;
820 data = &buf->buf_type_spec.data;
822 data->sg_count = buf->src_count;
824 * Compose a scatter/gather list from all of the buffers in the list.
825 * If the length of the buffer isn't a multiple of the sector size,
826 * we'll have to add an extra buffer. This should only happen
827 * at the end of a transfer.
829 if ((data->fill_len % sector_size) != 0) {
830 extra_buf_len = sector_size - (data->fill_len % sector_size);
831 extra_buf = calloc(extra_buf_len, 1);
832 if (extra_buf == NULL) {
833 warn("%s: unable to allocate %zu bytes for extra "
834 "buffer space", __func__, extra_buf_len);
842 data->segs = calloc(data->sg_count, sizeof(bus_dma_segment_t));
843 if (data->segs == NULL) {
844 warn("%s: unable to allocate %zu bytes for S/G list",
845 __func__, sizeof(bus_dma_segment_t) *
852 data->iovec = calloc(data->sg_count, sizeof(struct iovec));
853 if (data->iovec == NULL) {
854 warn("%s: unable to allocate %zu bytes for S/G list",
855 __func__, sizeof(struct iovec) * data->sg_count);
861 for (i = 0, tmp_buf = STAILQ_FIRST(&buf->src_list);
862 i < buf->src_count && tmp_buf != NULL; i++,
863 tmp_buf = STAILQ_NEXT(tmp_buf, src_links)) {
865 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
866 struct camdd_buf_data *tmp_data;
868 tmp_data = &tmp_buf->buf_type_spec.data;
870 data->segs[i].ds_addr =
871 (bus_addr_t) tmp_data->buf;
872 data->segs[i].ds_len = tmp_data->fill_len -
875 data->iovec[i].iov_base = tmp_data->buf;
876 data->iovec[i].iov_len = tmp_data->fill_len -
879 if (((tmp_data->fill_len - tmp_data->resid) %
881 *double_buf_needed = 1;
883 struct camdd_buf_indirect *tmp_ind;
885 tmp_ind = &tmp_buf->buf_type_spec.indirect;
887 data->segs[i].ds_addr =
888 (bus_addr_t)tmp_ind->start_ptr;
889 data->segs[i].ds_len = tmp_ind->len;
891 data->iovec[i].iov_base = tmp_ind->start_ptr;
892 data->iovec[i].iov_len = tmp_ind->len;
894 if ((tmp_ind->len % sector_size) != 0)
895 *double_buf_needed = 1;
899 if (extra_buf != NULL) {
901 data->segs[i].ds_addr = (bus_addr_t)extra_buf;
902 data->segs[i].ds_len = extra_buf_len;
904 data->iovec[i].iov_base = extra_buf;
905 data->iovec[i].iov_len = extra_buf_len;
909 if ((tmp_buf != NULL) || (i != data->sg_count)) {
910 warnx("buffer source count does not match "
911 "number of buffers in list!");
918 *num_sectors_used = (data->fill_len + extra_buf_len) /
925 camdd_buf_get_len(struct camdd_buf *buf)
929 if (buf->buf_type != CAMDD_BUF_DATA) {
930 struct camdd_buf_indirect *indirect;
932 indirect = &buf->buf_type_spec.indirect;
935 struct camdd_buf_data *data;
937 data = &buf->buf_type_spec.data;
938 len = data->fill_len;
945 camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf)
947 struct camdd_buf_data *data;
949 assert(buf->buf_type == CAMDD_BUF_DATA);
951 data = &buf->buf_type_spec.data;
953 STAILQ_INSERT_TAIL(&buf->src_list, child_buf, src_links);
956 data->fill_len += camdd_buf_get_len(child_buf);
964 } camdd_status_item_index;
966 static struct camdd_status_items {
968 struct mt_status_entry *entry;
969 } req_status_items[] = {
972 { "blk_gran", NULL },
973 { "max_effective_iosize", NULL }
977 camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
978 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran)
980 struct mt_status_data status_data;
981 char *xml_str = NULL;
985 retval = mt_get_xml_str(fd, MTIOCEXTGET, &xml_str);
987 err(1, "Couldn't get XML string from %s", filename);
989 retval = mt_get_status(xml_str, &status_data);
990 if (retval != XML_STATUS_OK) {
991 warn("couldn't get status for %s", filename);
997 if (status_data.error != 0) {
998 warnx("%s", status_data.error_str);
1003 for (i = 0; i < nitems(req_status_items); i++) {
1006 name = __DECONST(char *, req_status_items[i].name);
1007 req_status_items[i].entry = mt_status_entry_find(&status_data,
1009 if (req_status_items[i].entry == NULL) {
1010 errx(1, "Cannot find status entry %s",
1011 req_status_items[i].name);
1015 *max_iosize = req_status_items[CAMDD_TS_EFF_IOSIZE].entry->value_unsigned;
1016 *max_blk= req_status_items[CAMDD_TS_MAX_BLK].entry->value_unsigned;
1017 *min_blk= req_status_items[CAMDD_TS_MIN_BLK].entry->value_unsigned;
1018 *blk_gran = req_status_items[CAMDD_TS_BLK_GRAN].entry->value_unsigned;
1022 mt_status_free(&status_data);
1028 camdd_probe_file(int fd, struct camdd_io_opts *io_opts, int retry_count,
1031 struct camdd_dev *dev = NULL;
1032 struct camdd_dev_file *file_dev;
1033 uint64_t blocksize = io_opts->blocksize;
1035 dev = camdd_alloc_dev(CAMDD_DEV_FILE, NULL, 0, retry_count, timeout);
1039 file_dev = &dev->dev_spec.file;
1041 strlcpy(file_dev->filename, io_opts->dev_name,
1042 sizeof(file_dev->filename));
1043 strlcpy(dev->device_name, io_opts->dev_name, sizeof(dev->device_name));
1045 dev->blocksize = CAMDD_FILE_DEFAULT_BLOCK;
1047 dev->blocksize = blocksize;
1049 if ((io_opts->queue_depth != 0)
1050 && (io_opts->queue_depth != 1)) {
1051 warnx("Queue depth %ju for %s ignored, only 1 outstanding "
1052 "command supported", (uintmax_t)io_opts->queue_depth,
1055 dev->target_queue_depth = CAMDD_FILE_DEFAULT_DEPTH;
1056 dev->run = camdd_file_run;
1060 * We can effectively access files on byte boundaries. We'll reset
1061 * this for devices like disks that can be accessed on sector
1064 dev->sector_size = 1;
1066 if ((fd != STDIN_FILENO)
1067 && (fd != STDOUT_FILENO)) {
1070 retval = fstat(fd, &file_dev->sb);
1072 warn("Cannot stat %s", dev->device_name);
1075 if (S_ISREG(file_dev->sb.st_mode)) {
1076 file_dev->file_type = CAMDD_FILE_REG;
1077 } else if (S_ISCHR(file_dev->sb.st_mode)) {
1080 if (ioctl(fd, FIODTYPE, &type) == -1)
1081 err(1, "FIODTYPE ioctl failed on %s",
1085 file_dev->file_type = CAMDD_FILE_TAPE;
1086 else if (type & D_DISK)
1087 file_dev->file_type = CAMDD_FILE_DISK;
1088 else if (type & D_MEM)
1089 file_dev->file_type = CAMDD_FILE_MEM;
1090 else if (type & D_TTY)
1091 file_dev->file_type = CAMDD_FILE_TTY;
1093 } else if (S_ISDIR(file_dev->sb.st_mode)) {
1094 errx(1, "cannot operate on directory %s",
1096 } else if (S_ISFIFO(file_dev->sb.st_mode)) {
1097 file_dev->file_type = CAMDD_FILE_PIPE;
1099 errx(1, "Cannot determine file type for %s",
1102 switch (file_dev->file_type) {
1103 case CAMDD_FILE_REG:
1104 if (file_dev->sb.st_size != 0)
1105 dev->max_sector = file_dev->sb.st_size - 1;
1107 dev->max_sector = 0;
1108 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1110 case CAMDD_FILE_TAPE: {
1111 uint64_t max_iosize, max_blk, min_blk, blk_gran;
1113 * Check block limits and maximum effective iosize.
1114 * Make sure the blocksize is within the block
1115 * limits (and a multiple of the minimum blocksize)
1116 * and that the blocksize is <= maximum effective
1119 retval = camdd_probe_tape(fd, dev->device_name,
1120 &max_iosize, &max_blk, &min_blk, &blk_gran);
1122 errx(1, "Unable to probe tape %s",
1126 * The blocksize needs to be <= the maximum
1127 * effective I/O size of the tape device. Note
1128 * that this also takes into account the maximum
1129 * blocksize reported by READ BLOCK LIMITS.
1131 if (dev->blocksize > max_iosize) {
1132 warnx("Blocksize %u too big for %s, limiting "
1133 "to %ju", dev->blocksize, dev->device_name,
1135 dev->blocksize = max_iosize;
1139 * The blocksize needs to be at least min_blk;
1141 if (dev->blocksize < min_blk) {
1142 warnx("Blocksize %u too small for %s, "
1143 "increasing to %ju", dev->blocksize,
1144 dev->device_name, min_blk);
1145 dev->blocksize = min_blk;
1149 * And the blocksize needs to be a multiple of
1150 * the block granularity.
1153 && (dev->blocksize % (1 << blk_gran))) {
1154 warnx("Blocksize %u for %s not a multiple of "
1155 "%d, adjusting to %d", dev->blocksize,
1156 dev->device_name, (1 << blk_gran),
1157 dev->blocksize & ~((1 << blk_gran) - 1));
1158 dev->blocksize &= ~((1 << blk_gran) - 1);
1161 if (dev->blocksize == 0) {
1162 errx(1, "Unable to derive valid blocksize for "
1163 "%s", dev->device_name);
1167 * For tape drives, set the sector size to the
1168 * blocksize so that we make sure not to write
1169 * less than the blocksize out to the drive.
1171 dev->sector_size = dev->blocksize;
1174 case CAMDD_FILE_DISK: {
1176 unsigned int sector_size;
1178 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1180 if (ioctl(fd, DIOCGSECTORSIZE, §or_size) == -1) {
1181 err(1, "DIOCGSECTORSIZE ioctl failed on %s",
1185 if (sector_size == 0) {
1186 errx(1, "DIOCGSECTORSIZE ioctl returned "
1187 "invalid sector size %u for %s",
1188 sector_size, dev->device_name);
1191 if (ioctl(fd, DIOCGMEDIASIZE, &media_size) == -1) {
1192 err(1, "DIOCGMEDIASIZE ioctl failed on %s",
1196 if (media_size == 0) {
1197 errx(1, "DIOCGMEDIASIZE ioctl returned "
1198 "invalid media size %ju for %s",
1199 (uintmax_t)media_size, dev->device_name);
1202 if (dev->blocksize % sector_size) {
1203 errx(1, "%s blocksize %u not a multiple of "
1204 "sector size %u", dev->device_name,
1205 dev->blocksize, sector_size);
1208 dev->sector_size = sector_size;
1209 dev->max_sector = (media_size / sector_size) - 1;
1212 case CAMDD_FILE_MEM:
1213 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1220 if ((io_opts->offset != 0)
1221 && ((file_dev->file_flags & CAMDD_FF_CAN_SEEK) == 0)) {
1222 warnx("Offset %ju specified for %s, but we cannot seek on %s",
1223 io_opts->offset, io_opts->dev_name, io_opts->dev_name);
1227 else if ((io_opts->offset != 0)
1228 && ((io_opts->offset % dev->sector_size) != 0)) {
1229 warnx("Offset %ju for %s is not a multiple of the "
1230 "sector size %u", io_opts->offset,
1231 io_opts->dev_name, dev->sector_size);
1234 dev->start_offset_bytes = io_opts->offset;
1242 camdd_free_dev(dev);
1247 * Need to implement this. Do a basic probe:
1248 * - Check the inquiry data, make sure we're talking to a device that we
1249 * can reasonably expect to talk to -- direct, RBC, CD, WORM.
1250 * - Send a test unit ready, make sure the device is available.
1251 * - Get the capacity and block size.
1254 camdd_probe_pass(struct cam_device *cam_dev, struct camdd_io_opts *io_opts,
1255 camdd_argmask arglist, int probe_retry_count,
1256 int probe_timeout, int io_retry_count, int io_timeout)
1260 uint32_t cpi_maxio, max_iosize, pass_numblocks;
1262 struct scsi_read_capacity_data rcap;
1263 struct scsi_read_capacity_data_long rcaplong;
1264 struct camdd_dev *dev;
1265 struct camdd_dev_pass *pass_dev;
1271 scsi_dev_type = SID_TYPE(&cam_dev->inq_data);
1276 * For devices that support READ CAPACITY, we'll attempt to get the
1277 * capacity. Otherwise, we really don't support tape or other
1278 * devices via SCSI passthrough, so just return an error in that case.
1280 switch (scsi_dev_type) {
1289 errx(1, "Unsupported SCSI device type %d", scsi_dev_type);
1290 break; /*NOTREACHED*/
1293 ccb = cam_getccb(cam_dev);
1296 warnx("%s: error allocating ccb", __func__);
1300 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
1302 scsi_read_capacity(&ccb->csio,
1303 /*retries*/ probe_retry_count,
1305 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1308 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1310 /* Disable freezing the device queue */
1311 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1313 if (arglist & CAMDD_ARG_ERR_RECOVER)
1314 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1316 if (cam_send_ccb(cam_dev, ccb) < 0) {
1317 warn("error sending READ CAPACITY command");
1319 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1320 CAM_EPF_ALL, stderr);
1325 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1326 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1330 maxsector = scsi_4btoul(rcap.addr);
1331 block_len = scsi_4btoul(rcap.length);
1334 * A last block of 2^32-1 means that the true capacity is over 2TB,
1335 * and we need to issue the long READ CAPACITY to get the real
1336 * capacity. Otherwise, we're all set.
1338 if (maxsector != 0xffffffff)
1341 scsi_read_capacity_16(&ccb->csio,
1342 /*retries*/ probe_retry_count,
1344 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1348 (uint8_t *)&rcaplong,
1350 /*sense_len*/ SSD_FULL_SIZE,
1351 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1353 /* Disable freezing the device queue */
1354 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1356 if (arglist & CAMDD_ARG_ERR_RECOVER)
1357 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1359 if (cam_send_ccb(cam_dev, ccb) < 0) {
1360 warn("error sending READ CAPACITY (16) command");
1361 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1362 CAM_EPF_ALL, stderr);
1366 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1367 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1371 maxsector = scsi_8btou64(rcaplong.addr);
1372 block_len = scsi_4btoul(rcaplong.length);
1375 if (block_len == 0) {
1376 warnx("Sector size for %s%u is 0, cannot continue",
1377 cam_dev->device_name, cam_dev->dev_unit_num);
1381 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cpi);
1383 ccb->ccb_h.func_code = XPT_PATH_INQ;
1384 ccb->ccb_h.flags = CAM_DIR_NONE;
1385 ccb->ccb_h.retry_count = 1;
1387 if (cam_send_ccb(cam_dev, ccb) < 0) {
1388 warn("error sending XPT_PATH_INQ CCB");
1390 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1391 CAM_EPF_ALL, stderr);
1395 EV_SET(&ke, cam_dev->fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1397 dev = camdd_alloc_dev(CAMDD_DEV_PASS, &ke, 1, io_retry_count,
1402 pass_dev = &dev->dev_spec.pass;
1403 pass_dev->scsi_dev_type = scsi_dev_type;
1404 pass_dev->dev = cam_dev;
1405 pass_dev->max_sector = maxsector;
1406 pass_dev->block_len = block_len;
1407 pass_dev->cpi_maxio = ccb->cpi.maxio;
1408 snprintf(dev->device_name, sizeof(dev->device_name), "%s%u",
1409 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
1410 dev->sector_size = block_len;
1411 dev->max_sector = maxsector;
1415 * Determine the optimal blocksize to use for this device.
1419 * If the controller has not specified a maximum I/O size,
1420 * just go with 128K as a somewhat conservative value.
1422 if (pass_dev->cpi_maxio == 0)
1425 cpi_maxio = pass_dev->cpi_maxio;
1428 * If the controller has a large maximum I/O size, limit it
1429 * to something smaller so that the kernel doesn't have trouble
1430 * allocating buffers to copy data in and out for us.
1431 * XXX KDM this is until we have unmapped I/O support in the kernel.
1433 max_iosize = min(cpi_maxio, CAMDD_PASS_MAX_BLOCK);
1436 * If we weren't able to get a block size for some reason,
1437 * default to 512 bytes.
1439 block_len = pass_dev->block_len;
1444 * Figure out how many blocksize chunks will fit in the
1447 pass_numblocks = max_iosize / block_len;
1450 * And finally, multiple the number of blocks by the LBA
1451 * length to get our maximum block size;
1453 dev->blocksize = pass_numblocks * block_len;
1455 if (io_opts->blocksize != 0) {
1456 if ((io_opts->blocksize % dev->sector_size) != 0) {
1457 warnx("Blocksize %ju for %s is not a multiple of "
1458 "sector size %u", (uintmax_t)io_opts->blocksize,
1459 dev->device_name, dev->sector_size);
1462 dev->blocksize = io_opts->blocksize;
1464 dev->target_queue_depth = CAMDD_PASS_DEFAULT_DEPTH;
1465 if (io_opts->queue_depth != 0)
1466 dev->target_queue_depth = io_opts->queue_depth;
1468 if (io_opts->offset != 0) {
1469 if (io_opts->offset > (dev->max_sector * dev->sector_size)) {
1470 warnx("Offset %ju is past the end of device %s",
1471 io_opts->offset, dev->device_name);
1475 else if ((io_opts->offset % dev->sector_size) != 0) {
1476 warnx("Offset %ju for %s is not a multiple of the "
1477 "sector size %u", io_opts->offset,
1478 dev->device_name, dev->sector_size);
1481 dev->start_offset_bytes = io_opts->offset;
1485 dev->min_cmd_size = io_opts->min_cmd_size;
1487 dev->run = camdd_pass_run;
1488 dev->fetch = camdd_pass_fetch;
1498 camdd_free_dev(dev);
1504 camdd_worker(void *arg)
1506 struct camdd_dev *dev = arg;
1507 struct camdd_buf *buf;
1508 struct timespec ts, *kq_ts;
1513 pthread_mutex_lock(&dev->mutex);
1515 dev->flags |= CAMDD_DEV_FLAG_ACTIVE;
1522 * XXX KDM check the reorder queue depth?
1524 if (dev->write_dev == 0) {
1525 uint32_t our_depth, peer_depth, peer_bytes, our_bytes;
1526 uint32_t target_depth = dev->target_queue_depth;
1527 uint32_t peer_target_depth =
1528 dev->peer_dev->target_queue_depth;
1529 uint32_t peer_blocksize = dev->peer_dev->blocksize;
1531 camdd_get_depth(dev, &our_depth, &peer_depth,
1532 &our_bytes, &peer_bytes);
1535 while (((our_depth < target_depth)
1536 && (peer_depth < peer_target_depth))
1537 || ((peer_bytes + our_bytes) <
1538 (peer_blocksize * 2))) {
1540 while (((our_depth + peer_depth) <
1541 (target_depth + peer_target_depth))
1542 || ((peer_bytes + our_bytes) <
1543 (peer_blocksize * 3))) {
1545 retval = camdd_queue(dev, NULL);
1548 else if (retval != 0) {
1553 camdd_get_depth(dev, &our_depth, &peer_depth,
1554 &our_bytes, &peer_bytes);
1558 * See if we have any I/O that is ready to execute.
1560 buf = STAILQ_FIRST(&dev->run_queue);
1562 while (dev->target_queue_depth > dev->cur_active_io) {
1563 retval = dev->run(dev);
1565 dev->flags |= CAMDD_DEV_FLAG_EOF;
1568 } else if (retval != 0) {
1575 * We've reached EOF, or our partner has reached EOF.
1577 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
1578 || (dev->flags & CAMDD_DEV_FLAG_PEER_EOF)) {
1579 if (dev->write_dev != 0) {
1580 if ((STAILQ_EMPTY(&dev->work_queue))
1581 && (dev->num_run_queue == 0)
1582 && (dev->cur_active_io == 0)) {
1587 * If we're the reader, and the writer
1588 * got EOF, he is already done. If we got
1589 * the EOF, then we need to wait until
1590 * everything is flushed out for the writer.
1592 if (dev->flags & CAMDD_DEV_FLAG_PEER_EOF) {
1594 } else if ((dev->num_peer_work_queue == 0)
1595 && (dev->num_peer_done_queue == 0)
1596 && (dev->cur_active_io == 0)
1597 && (dev->num_run_queue == 0)) {
1602 * XXX KDM need to do something about the pending
1603 * queue and cleanup resources.
1607 if ((dev->write_dev == 0)
1608 && (dev->cur_active_io == 0)
1609 && (dev->peer_bytes_queued < dev->peer_dev->blocksize))
1615 * Run kevent to see if there are events to process.
1617 pthread_mutex_unlock(&dev->mutex);
1618 retval = kevent(dev->kq, NULL, 0, &ke, 1, kq_ts);
1619 pthread_mutex_lock(&dev->mutex);
1621 warn("%s: error returned from kevent",__func__);
1623 } else if (retval != 0) {
1624 switch (ke.filter) {
1626 if (dev->fetch != NULL) {
1627 retval = dev->fetch(dev);
1636 * We register for this so we don't get
1637 * an error as a result of a SIGINFO or a
1638 * SIGINT. It will actually get handled
1639 * by the signal handler. If we get a
1640 * SIGINT, bail out without printing an
1641 * error message. Any other signals
1642 * will result in the error message above.
1644 if (ke.ident == SIGINT)
1650 * Check to see if the other thread has
1651 * queued any I/O for us to do. (In this
1652 * case we're the writer.)
1654 for (buf = STAILQ_FIRST(&dev->work_queue);
1656 buf = STAILQ_FIRST(&dev->work_queue)) {
1657 STAILQ_REMOVE_HEAD(&dev->work_queue,
1659 retval = camdd_queue(dev, buf);
1661 * We keep going unless we get an
1662 * actual error. If we get EOF, we
1663 * still want to remove the buffers
1664 * from the queue and send the back
1665 * to the reader thread.
1675 * Next check to see if the other thread has
1676 * queued any completed buffers back to us.
1677 * (In this case we're the reader.)
1679 for (buf = STAILQ_FIRST(&dev->peer_done_queue);
1681 buf = STAILQ_FIRST(&dev->peer_done_queue)){
1683 &dev->peer_done_queue, work_links);
1684 dev->num_peer_done_queue--;
1685 camdd_peer_done(buf);
1689 warnx("%s: unknown kevent filter %d",
1690 __func__, ke.filter);
1698 dev->flags &= ~CAMDD_DEV_FLAG_ACTIVE;
1700 /* XXX KDM cleanup resources here? */
1702 pthread_mutex_unlock(&dev->mutex);
1705 sem_post(&camdd_sem);
1711 * Simplistic translation of CCB status to our local status.
1714 camdd_ccb_status(union ccb *ccb)
1716 camdd_buf_status status = CAMDD_STATUS_NONE;
1717 cam_status ccb_status;
1719 ccb_status = ccb->ccb_h.status & CAM_STATUS_MASK;
1721 switch (ccb_status) {
1723 if (ccb->csio.resid == 0) {
1724 status = CAMDD_STATUS_OK;
1725 } else if (ccb->csio.dxfer_len > ccb->csio.resid) {
1726 status = CAMDD_STATUS_SHORT_IO;
1728 status = CAMDD_STATUS_EOF;
1732 case CAM_SCSI_STATUS_ERROR: {
1733 switch (ccb->csio.scsi_status) {
1734 case SCSI_STATUS_OK:
1735 case SCSI_STATUS_COND_MET:
1736 case SCSI_STATUS_INTERMED:
1737 case SCSI_STATUS_INTERMED_COND_MET:
1738 status = CAMDD_STATUS_OK;
1740 case SCSI_STATUS_CMD_TERMINATED:
1741 case SCSI_STATUS_CHECK_COND:
1742 case SCSI_STATUS_QUEUE_FULL:
1743 case SCSI_STATUS_BUSY:
1744 case SCSI_STATUS_RESERV_CONFLICT:
1746 status = CAMDD_STATUS_ERROR;
1752 status = CAMDD_STATUS_ERROR;
1760 * Queue a buffer to our peer's work thread for writing.
1762 * Returns 0 for success, -1 for failure, 1 if the other thread exited.
1765 camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf)
1768 STAILQ_HEAD(, camdd_buf) local_queue;
1769 struct camdd_buf *buf1, *buf2;
1770 struct camdd_buf_data *data = NULL;
1771 uint64_t peer_bytes_queued = 0;
1775 STAILQ_INIT(&local_queue);
1778 * Since we're the reader, we need to queue our I/O to the writer
1779 * in sequential order in order to make sure it gets written out
1780 * in sequential order.
1782 * Check the next expected I/O starting offset. If this doesn't
1783 * match, put it on the reorder queue.
1785 if ((buf->lba * dev->sector_size) != dev->next_completion_pos_bytes) {
1788 * If there is nothing on the queue, there is no sorting
1791 if (STAILQ_EMPTY(&dev->reorder_queue)) {
1792 STAILQ_INSERT_TAIL(&dev->reorder_queue, buf, links);
1793 dev->num_reorder_queue++;
1798 * Sort in ascending order by starting LBA. There should
1799 * be no identical LBAs.
1801 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1803 buf2 = STAILQ_NEXT(buf1, links);
1804 if (buf->lba < buf1->lba) {
1806 * If we're less than the first one, then
1807 * we insert at the head of the list
1808 * because this has to be the first element
1811 STAILQ_INSERT_HEAD(&dev->reorder_queue,
1813 dev->num_reorder_queue++;
1815 } else if (buf->lba > buf1->lba) {
1817 STAILQ_INSERT_TAIL(&dev->reorder_queue,
1819 dev->num_reorder_queue++;
1821 } else if (buf->lba < buf2->lba) {
1822 STAILQ_INSERT_AFTER(&dev->reorder_queue,
1824 dev->num_reorder_queue++;
1828 errx(1, "Found buffers with duplicate LBA %ju!",
1836 * We're the next expected I/O completion, so put ourselves
1837 * on the local queue to be sent to the writer. We use
1838 * work_links here so that we can queue this to the
1839 * peer_work_queue before taking the buffer off of the
1842 dev->next_completion_pos_bytes += buf->len;
1843 STAILQ_INSERT_TAIL(&local_queue, buf, work_links);
1846 * Go through the reorder queue looking for more sequential
1847 * I/O and add it to the local queue.
1849 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1850 buf1 = STAILQ_FIRST(&dev->reorder_queue)) {
1852 * As soon as we see an I/O that is out of sequence,
1855 if ((buf1->lba * dev->sector_size) !=
1856 dev->next_completion_pos_bytes)
1859 STAILQ_REMOVE_HEAD(&dev->reorder_queue, links);
1860 dev->num_reorder_queue--;
1861 STAILQ_INSERT_TAIL(&local_queue, buf1, work_links);
1862 dev->next_completion_pos_bytes += buf1->len;
1867 * Setup the event to let the other thread know that it has work
1870 EV_SET(&ke, (uintptr_t)&dev->peer_dev->work_queue, EVFILT_USER, 0,
1871 NOTE_TRIGGER, 0, NULL);
1874 * Put this on our shadow queue so that we know what we've queued
1875 * to the other thread.
1877 STAILQ_FOREACH_SAFE(buf1, &local_queue, work_links, buf2) {
1878 if (buf1->buf_type != CAMDD_BUF_DATA) {
1879 errx(1, "%s: should have a data buffer, not an "
1880 "indirect buffer", __func__);
1882 data = &buf1->buf_type_spec.data;
1885 * We only need to send one EOF to the writer, and don't
1886 * need to continue sending EOFs after that.
1888 if (buf1->status == CAMDD_STATUS_EOF) {
1889 if (dev->flags & CAMDD_DEV_FLAG_EOF_SENT) {
1890 STAILQ_REMOVE(&local_queue, buf1, camdd_buf,
1892 camdd_release_buf(buf1);
1896 dev->flags |= CAMDD_DEV_FLAG_EOF_SENT;
1900 STAILQ_INSERT_TAIL(&dev->peer_work_queue, buf1, links);
1901 peer_bytes_queued += (data->fill_len - data->resid);
1902 dev->peer_bytes_queued += (data->fill_len - data->resid);
1903 dev->num_peer_work_queue++;
1906 if (STAILQ_FIRST(&local_queue) == NULL)
1910 * Drop our mutex and pick up the other thread's mutex. We need to
1911 * do this to avoid deadlocks.
1913 pthread_mutex_unlock(&dev->mutex);
1914 pthread_mutex_lock(&dev->peer_dev->mutex);
1916 if (dev->peer_dev->flags & CAMDD_DEV_FLAG_ACTIVE) {
1918 * Put the buffers on the other thread's incoming work queue.
1920 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1921 buf1 = STAILQ_FIRST(&local_queue)) {
1922 STAILQ_REMOVE_HEAD(&local_queue, work_links);
1923 STAILQ_INSERT_TAIL(&dev->peer_dev->work_queue, buf1,
1927 * Send an event to the other thread's kqueue to let it know
1928 * that there is something on the work queue.
1930 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
1932 warn("%s: unable to add peer work_queue kevent",
1939 pthread_mutex_unlock(&dev->peer_dev->mutex);
1940 pthread_mutex_lock(&dev->mutex);
1943 * If the other side isn't active, run through the queue and
1944 * release all of the buffers.
1947 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1948 buf1 = STAILQ_FIRST(&local_queue)) {
1949 STAILQ_REMOVE_HEAD(&local_queue, work_links);
1950 STAILQ_REMOVE(&dev->peer_work_queue, buf1, camdd_buf,
1952 dev->num_peer_work_queue--;
1953 camdd_release_buf(buf1);
1955 dev->peer_bytes_queued -= peer_bytes_queued;
1964 * Return a buffer to the reader thread when we have completed writing it.
1967 camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf)
1973 * Setup the event to let the other thread know that we have
1974 * completed a buffer.
1976 EV_SET(&ke, (uintptr_t)&dev->peer_dev->peer_done_queue, EVFILT_USER, 0,
1977 NOTE_TRIGGER, 0, NULL);
1980 * Drop our lock and acquire the other thread's lock before
1983 pthread_mutex_unlock(&dev->mutex);
1984 pthread_mutex_lock(&dev->peer_dev->mutex);
1987 * Put the buffer on the reader thread's peer done queue now that
1988 * we have completed it.
1990 STAILQ_INSERT_TAIL(&dev->peer_dev->peer_done_queue, peer_buf,
1992 dev->peer_dev->num_peer_done_queue++;
1995 * Send an event to the peer thread to let it know that we've added
1996 * something to its peer done queue.
1998 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2000 warn("%s: unable to add peer_done_queue kevent", __func__);
2005 * Drop the other thread's lock and reacquire ours.
2007 pthread_mutex_unlock(&dev->peer_dev->mutex);
2008 pthread_mutex_lock(&dev->mutex);
2014 * Free a buffer that was written out by the writer thread and returned to
2015 * the reader thread.
2018 camdd_peer_done(struct camdd_buf *buf)
2020 struct camdd_dev *dev;
2021 struct camdd_buf_data *data;
2024 if (buf->buf_type != CAMDD_BUF_DATA) {
2025 errx(1, "%s: should have a data buffer, not an "
2026 "indirect buffer", __func__);
2029 data = &buf->buf_type_spec.data;
2031 STAILQ_REMOVE(&dev->peer_work_queue, buf, camdd_buf, links);
2032 dev->num_peer_work_queue--;
2033 dev->peer_bytes_queued -= (data->fill_len - data->resid);
2035 if (buf->status == CAMDD_STATUS_EOF)
2036 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2038 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2042 * Assumes caller holds the lock for this device.
2045 camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
2051 * If we're the reader, we need to send the completed I/O
2052 * to the writer. If we're the writer, we need to just
2053 * free up resources, or let the reader know if we've
2054 * encountered an error.
2056 if (dev->write_dev == 0) {
2057 retval = camdd_queue_peer_buf(dev, buf);
2061 struct camdd_buf *tmp_buf, *next_buf;
2063 STAILQ_FOREACH_SAFE(tmp_buf, &buf->src_list, src_links,
2065 struct camdd_buf *src_buf;
2066 struct camdd_buf_indirect *indirect;
2068 STAILQ_REMOVE(&buf->src_list, tmp_buf,
2069 camdd_buf, src_links);
2071 tmp_buf->status = buf->status;
2073 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
2074 camdd_complete_peer_buf(dev, tmp_buf);
2078 indirect = &tmp_buf->buf_type_spec.indirect;
2079 src_buf = indirect->src_buf;
2080 src_buf->refcount--;
2082 * XXX KDM we probably need to account for
2083 * exactly how many bytes we were able to
2084 * write. Allocate the residual to the
2085 * first N buffers? Or just track the
2086 * number of bytes written? Right now the reader
2087 * doesn't do anything with a residual.
2089 src_buf->status = buf->status;
2090 if (src_buf->refcount <= 0)
2091 camdd_complete_peer_buf(dev, src_buf);
2092 STAILQ_INSERT_TAIL(&dev->free_indirect_queue,
2096 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2101 * Fetch all completed commands from the pass(4) device.
2103 * Returns the number of commands received, or -1 if any of the commands
2104 * completed with an error. Returns 0 if no commands are available.
2107 camdd_pass_fetch(struct camdd_dev *dev)
2109 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2111 int retval = 0, num_fetched = 0, error_count = 0;
2113 pthread_mutex_unlock(&dev->mutex);
2115 * XXX KDM we don't distinguish between EFAULT and ENOENT.
2117 while ((retval = ioctl(pass_dev->dev->fd, CAMIOGET, &ccb)) != -1) {
2118 struct camdd_buf *buf;
2119 struct camdd_buf_data *data;
2120 cam_status ccb_status;
2123 buf = ccb.ccb_h.ccb_buf;
2124 data = &buf->buf_type_spec.data;
2125 buf_ccb = &data->ccb;
2130 * Copy the CCB back out so we get status, sense data, etc.
2132 bcopy(&ccb, buf_ccb, sizeof(ccb));
2134 pthread_mutex_lock(&dev->mutex);
2137 * We're now done, so take this off the active queue.
2139 STAILQ_REMOVE(&dev->active_queue, buf, camdd_buf, links);
2140 dev->cur_active_io--;
2142 ccb_status = ccb.ccb_h.status & CAM_STATUS_MASK;
2143 if (ccb_status != CAM_REQ_CMP) {
2144 cam_error_print(pass_dev->dev, &ccb, CAM_ESF_ALL,
2145 CAM_EPF_ALL, stderr);
2148 data->resid = ccb.csio.resid;
2149 dev->bytes_transferred += (ccb.csio.dxfer_len - ccb.csio.resid);
2151 if (buf->status == CAMDD_STATUS_NONE)
2152 buf->status = camdd_ccb_status(&ccb);
2153 if (buf->status == CAMDD_STATUS_ERROR)
2155 else if (buf->status == CAMDD_STATUS_EOF) {
2157 * Once we queue this buffer to our partner thread,
2158 * he will know that we've hit EOF.
2160 dev->flags |= CAMDD_DEV_FLAG_EOF;
2163 camdd_complete_buf(dev, buf, &error_count);
2166 * Unlock in preparation for the ioctl call.
2168 pthread_mutex_unlock(&dev->mutex);
2171 pthread_mutex_lock(&dev->mutex);
2173 if (error_count > 0)
2176 return (num_fetched);
2180 * Returns -1 for error, 0 for success/continue, and 1 for resource
2181 * shortage/stop processing.
2184 camdd_file_run(struct camdd_dev *dev)
2186 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
2187 struct camdd_buf_data *data;
2188 struct camdd_buf *buf;
2190 int retval = 0, write_dev = dev->write_dev;
2191 int error_count = 0, no_resources = 0, double_buf_needed = 0;
2192 uint32_t num_sectors = 0, db_len = 0;
2194 buf = STAILQ_FIRST(&dev->run_queue);
2198 } else if ((dev->write_dev == 0)
2199 && (dev->flags & (CAMDD_DEV_FLAG_EOF |
2200 CAMDD_DEV_FLAG_EOF_SENT))) {
2201 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2202 dev->num_run_queue--;
2203 buf->status = CAMDD_STATUS_EOF;
2209 * If we're writing, we need to go through the source buffer list
2210 * and create an S/G list.
2212 if (write_dev != 0) {
2213 retval = camdd_buf_sg_create(buf, /*iovec*/ 1,
2214 dev->sector_size, &num_sectors, &double_buf_needed);
2221 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2222 dev->num_run_queue--;
2224 data = &buf->buf_type_spec.data;
2227 * pread(2) and pwrite(2) offsets are byte offsets.
2229 io_offset = buf->lba * dev->sector_size;
2232 * Unlock the mutex while we read or write.
2234 pthread_mutex_unlock(&dev->mutex);
2237 * Note that we don't need to double buffer if we're the reader
2238 * because in that case, we have allocated a single buffer of
2239 * sufficient size to do the read. This copy is necessary on
2240 * writes because if one of the components of the S/G list is not
2241 * a sector size multiple, the kernel will reject the write. This
2242 * is unfortunate but not surprising. So this will make sure that
2243 * we're using a single buffer that is a multiple of the sector size.
2245 if ((double_buf_needed != 0)
2246 && (data->sg_count > 1)
2247 && (write_dev != 0)) {
2248 uint32_t cur_offset;
2251 if (file_dev->tmp_buf == NULL)
2252 file_dev->tmp_buf = calloc(dev->blocksize, 1);
2253 if (file_dev->tmp_buf == NULL) {
2254 buf->status = CAMDD_STATUS_ERROR;
2256 pthread_mutex_lock(&dev->mutex);
2259 for (i = 0, cur_offset = 0; i < data->sg_count; i++) {
2260 bcopy(data->iovec[i].iov_base,
2261 &file_dev->tmp_buf[cur_offset],
2262 data->iovec[i].iov_len);
2263 cur_offset += data->iovec[i].iov_len;
2265 db_len = cur_offset;
2268 if (file_dev->file_flags & CAMDD_FF_CAN_SEEK) {
2269 if (write_dev == 0) {
2271 * XXX KDM is there any way we would need a S/G
2274 retval = pread(file_dev->fd, data->buf,
2275 buf->len, io_offset);
2277 if (double_buf_needed != 0) {
2278 retval = pwrite(file_dev->fd, file_dev->tmp_buf,
2280 } else if (data->sg_count == 0) {
2281 retval = pwrite(file_dev->fd, data->buf,
2282 data->fill_len, io_offset);
2284 retval = pwritev(file_dev->fd, data->iovec,
2285 data->sg_count, io_offset);
2289 if (write_dev == 0) {
2291 * XXX KDM is there any way we would need a S/G
2294 retval = read(file_dev->fd, data->buf, buf->len);
2296 if (double_buf_needed != 0) {
2297 retval = write(file_dev->fd, file_dev->tmp_buf,
2299 } else if (data->sg_count == 0) {
2300 retval = write(file_dev->fd, data->buf,
2303 retval = writev(file_dev->fd, data->iovec,
2309 /* We're done, re-acquire the lock */
2310 pthread_mutex_lock(&dev->mutex);
2312 if (retval >= (ssize_t)data->fill_len) {
2314 * If the bytes transferred is more than the request size,
2315 * that indicates an overrun, which should only happen at
2316 * the end of a transfer if we have to round up to a sector
2319 if (buf->status == CAMDD_STATUS_NONE)
2320 buf->status = CAMDD_STATUS_OK;
2322 dev->bytes_transferred += retval;
2323 } else if (retval == -1) {
2324 warn("Error %s %s", (write_dev) ? "writing to" :
2325 "reading from", file_dev->filename);
2327 buf->status = CAMDD_STATUS_ERROR;
2328 data->resid = data->fill_len;
2331 if (dev->debug == 0)
2334 if ((double_buf_needed != 0)
2335 && (write_dev != 0)) {
2336 fprintf(stderr, "%s: fd %d, DB buf %p, len %u lba %ju "
2337 "offset %ju\n", __func__, file_dev->fd,
2338 file_dev->tmp_buf, db_len, (uintmax_t)buf->lba,
2339 (uintmax_t)io_offset);
2340 } else if (data->sg_count == 0) {
2341 fprintf(stderr, "%s: fd %d, buf %p, len %u, lba %ju "
2342 "offset %ju\n", __func__, file_dev->fd, data->buf,
2343 data->fill_len, (uintmax_t)buf->lba,
2344 (uintmax_t)io_offset);
2348 fprintf(stderr, "%s: fd %d, len %u, lba %ju "
2349 "offset %ju\n", __func__, file_dev->fd,
2350 data->fill_len, (uintmax_t)buf->lba,
2351 (uintmax_t)io_offset);
2353 for (i = 0; i < data->sg_count; i++) {
2354 fprintf(stderr, "index %d ptr %p len %zu\n",
2355 i, data->iovec[i].iov_base,
2356 data->iovec[i].iov_len);
2359 } else if (retval == 0) {
2360 buf->status = CAMDD_STATUS_EOF;
2361 if (dev->debug != 0)
2362 printf("%s: got EOF from %s!\n", __func__,
2363 file_dev->filename);
2364 data->resid = data->fill_len;
2366 } else if (retval < (ssize_t)data->fill_len) {
2367 if (buf->status == CAMDD_STATUS_NONE)
2368 buf->status = CAMDD_STATUS_SHORT_IO;
2369 data->resid = data->fill_len - retval;
2370 dev->bytes_transferred += retval;
2375 if (buf->status == CAMDD_STATUS_EOF) {
2376 struct camdd_buf *buf2;
2377 dev->flags |= CAMDD_DEV_FLAG_EOF;
2378 STAILQ_FOREACH(buf2, &dev->run_queue, links)
2379 buf2->status = CAMDD_STATUS_EOF;
2382 camdd_complete_buf(dev, buf, &error_count);
2385 if (error_count != 0)
2387 else if (no_resources != 0)
2394 * Execute one command from the run queue. Returns 0 for success, 1 for
2395 * stop processing, and -1 for error.
2398 camdd_pass_run(struct camdd_dev *dev)
2400 struct camdd_buf *buf = NULL;
2401 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2402 struct camdd_buf_data *data;
2403 uint32_t num_blocks, sectors_used = 0;
2405 int retval = 0, is_write = dev->write_dev;
2406 int double_buf_needed = 0;
2408 buf = STAILQ_FIRST(&dev->run_queue);
2415 * If we're writing, we need to go through the source buffer list
2416 * and create an S/G list.
2418 if (is_write != 0) {
2419 retval = camdd_buf_sg_create(buf, /*iovec*/ 0,dev->sector_size,
2420 §ors_used, &double_buf_needed);
2427 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2428 dev->num_run_queue--;
2430 data = &buf->buf_type_spec.data;
2433 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
2436 * In almost every case the number of blocks should be the device
2437 * block size. The exception may be at the end of an I/O stream
2438 * for a partial block or at the end of a device.
2441 num_blocks = sectors_used;
2443 num_blocks = data->fill_len / pass_dev->block_len;
2445 scsi_read_write(&ccb->csio,
2446 /*retries*/ dev->retry_count,
2448 /*tag_action*/ MSG_SIMPLE_Q_TAG,
2449 /*readop*/ (dev->write_dev == 0) ? SCSI_RW_READ :
2452 /*minimum_cmd_size*/ dev->min_cmd_size,
2454 /*block_count*/ num_blocks,
2455 /*data_ptr*/ (data->sg_count != 0) ?
2456 (uint8_t *)data->segs : data->buf,
2457 /*dxfer_len*/ (num_blocks * pass_dev->block_len),
2458 /*sense_len*/ SSD_FULL_SIZE,
2459 /*timeout*/ dev->io_timeout);
2461 /* Disable freezing the device queue */
2462 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
2464 if (dev->retry_count != 0)
2465 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
2467 if (data->sg_count != 0) {
2468 ccb->csio.sglist_cnt = data->sg_count;
2469 ccb->ccb_h.flags |= CAM_DATA_SG;
2473 * Store a pointer to the buffer in the CCB. The kernel will
2474 * restore this when we get it back, and we'll use it to identify
2475 * the buffer this CCB came from.
2477 ccb->ccb_h.ccb_buf = buf;
2480 * Unlock our mutex in preparation for issuing the ioctl.
2482 pthread_mutex_unlock(&dev->mutex);
2484 * Queue the CCB to the pass(4) driver.
2486 if (ioctl(pass_dev->dev->fd, CAMIOQUEUE, ccb) == -1) {
2487 pthread_mutex_lock(&dev->mutex);
2489 warn("%s: error sending CAMIOQUEUE ioctl to %s%u", __func__,
2490 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
2491 warn("%s: CCB address is %p", __func__, ccb);
2494 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2496 pthread_mutex_lock(&dev->mutex);
2498 dev->cur_active_io++;
2499 STAILQ_INSERT_TAIL(&dev->active_queue, buf, links);
2507 camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len)
2509 struct camdd_dev_pass *pass_dev;
2510 uint32_t num_blocks;
2513 pass_dev = &dev->dev_spec.pass;
2515 *lba = dev->next_io_pos_bytes / dev->sector_size;
2516 *len = dev->blocksize;
2517 num_blocks = *len / dev->sector_size;
2520 * If max_sector is 0, then we have no set limit. This can happen
2521 * if we're writing to a file in a filesystem, or reading from
2522 * something like /dev/zero.
2524 if ((dev->max_sector != 0)
2525 || (dev->sector_io_limit != 0)) {
2526 uint64_t max_sector;
2528 if ((dev->max_sector != 0)
2529 && (dev->sector_io_limit != 0))
2530 max_sector = min(dev->sector_io_limit, dev->max_sector);
2531 else if (dev->max_sector != 0)
2532 max_sector = dev->max_sector;
2534 max_sector = dev->sector_io_limit;
2538 * Check to see whether we're starting off past the end of
2539 * the device. If so, we need to just send an EOF
2540 * notification to the writer.
2542 if (*lba > max_sector) {
2545 } else if (((*lba + num_blocks) > max_sector + 1)
2546 || ((*lba + num_blocks) < *lba)) {
2548 * If we get here (but pass the first check), we
2549 * can trim the request length down to go to the
2550 * end of the device.
2552 num_blocks = (max_sector + 1) - *lba;
2553 *len = num_blocks * dev->sector_size;
2558 dev->next_io_pos_bytes += *len;
2564 * Returns 0 for success, 1 for EOF detected, and -1 for failure.
2567 camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf)
2569 struct camdd_buf *buf = NULL;
2570 struct camdd_buf_data *data;
2571 struct camdd_dev_pass *pass_dev;
2573 struct camdd_buf_data *rb_data;
2574 int is_write = dev->write_dev;
2575 int eof_flush_needed = 0;
2579 pass_dev = &dev->dev_spec.pass;
2582 * If we've gotten EOF or our partner has, we should not continue
2583 * queueing I/O. If we're a writer, though, we should continue
2584 * to write any buffers that don't have EOF status.
2586 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
2587 || ((dev->flags & CAMDD_DEV_FLAG_PEER_EOF)
2588 && (is_write == 0))) {
2590 * Tell the worker thread that we have seen EOF.
2595 * If we're the writer, send the buffer back with EOF status.
2598 read_buf->status = CAMDD_STATUS_EOF;
2600 error = camdd_complete_peer_buf(dev, read_buf);
2605 if (is_write == 0) {
2606 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2611 data = &buf->buf_type_spec.data;
2613 retval = camdd_get_next_lba_len(dev, &buf->lba, &buf->len);
2615 buf->status = CAMDD_STATUS_EOF;
2618 && ((dev->flags & (CAMDD_DEV_FLAG_EOF_SENT |
2619 CAMDD_DEV_FLAG_EOF_QUEUED)) != 0)) {
2620 camdd_release_buf(buf);
2623 dev->flags |= CAMDD_DEV_FLAG_EOF_QUEUED;
2626 data->fill_len = buf->len;
2627 data->src_start_offset = buf->lba * dev->sector_size;
2630 * Put this on the run queue.
2632 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2633 dev->num_run_queue++;
2640 * Check for new EOF status from the reader.
2642 if ((read_buf->status == CAMDD_STATUS_EOF)
2643 || (read_buf->status == CAMDD_STATUS_ERROR)) {
2644 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2645 if ((STAILQ_FIRST(&dev->pending_queue) == NULL)
2646 && (read_buf->len == 0)) {
2647 camdd_complete_peer_buf(dev, read_buf);
2651 eof_flush_needed = 1;
2655 * See if we have a buffer we're composing with pieces from our
2658 buf = STAILQ_FIRST(&dev->pending_queue);
2663 retval = camdd_get_next_lba_len(dev, &lba, &len);
2665 read_buf->status = CAMDD_STATUS_EOF;
2668 dev->flags |= CAMDD_DEV_FLAG_EOF;
2669 error = camdd_complete_peer_buf(dev, read_buf);
2675 * If we don't have a pending buffer, we need to grab a new
2676 * one from the free list or allocate another one.
2678 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2687 STAILQ_INSERT_TAIL(&dev->pending_queue, buf, links);
2688 dev->num_pending_queue++;
2691 data = &buf->buf_type_spec.data;
2693 rb_data = &read_buf->buf_type_spec.data;
2695 if ((rb_data->src_start_offset != dev->next_peer_pos_bytes)
2696 && (dev->debug != 0)) {
2697 printf("%s: WARNING: reader offset %#jx != expected offset "
2698 "%#jx\n", __func__, (uintmax_t)rb_data->src_start_offset,
2699 (uintmax_t)dev->next_peer_pos_bytes);
2701 dev->next_peer_pos_bytes = rb_data->src_start_offset +
2702 (rb_data->fill_len - rb_data->resid);
2704 new_len = (rb_data->fill_len - rb_data->resid) + data->fill_len;
2705 if (new_len < buf->len) {
2707 * There are three cases here:
2708 * 1. We need more data to fill up a block, so we put
2709 * this I/O on the queue and wait for more I/O.
2710 * 2. We have a pending buffer in the queue that is
2711 * smaller than our blocksize, but we got an EOF. So we
2712 * need to go ahead and flush the write out.
2713 * 3. We got an error.
2717 * Increment our fill length.
2719 data->fill_len += (rb_data->fill_len - rb_data->resid);
2722 * Add the new read buffer to the list for writing.
2724 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2726 /* Increment the count */
2729 if (eof_flush_needed == 0) {
2731 * We need to exit, because we don't have enough
2737 * Take the buffer off of the pending queue.
2739 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2741 dev->num_pending_queue--;
2744 * If we need an EOF flush, but there is no data
2745 * to flush, go ahead and return this buffer.
2747 if (data->fill_len == 0) {
2748 camdd_complete_buf(dev, buf, /*error_count*/0);
2754 * Put this on the next queue for execution.
2756 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2757 dev->num_run_queue++;
2759 } else if (new_len == buf->len) {
2761 * We have enough data to completey fill one block,
2762 * so we're ready to issue the I/O.
2766 * Take the buffer off of the pending queue.
2768 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf, links);
2769 dev->num_pending_queue--;
2772 * Add the new read buffer to the list for writing.
2774 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2776 /* Increment the count */
2780 * Increment our fill length.
2782 data->fill_len += (rb_data->fill_len - rb_data->resid);
2785 * Put this on the next queue for execution.
2787 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2788 dev->num_run_queue++;
2790 struct camdd_buf *idb;
2791 struct camdd_buf_indirect *indirect;
2792 uint32_t len_to_go, cur_offset;
2795 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2800 indirect = &idb->buf_type_spec.indirect;
2801 indirect->src_buf = read_buf;
2802 read_buf->refcount++;
2803 indirect->offset = 0;
2804 indirect->start_ptr = rb_data->buf;
2806 * We've already established that there is more
2807 * data in read_buf than we have room for in our
2808 * current write request. So this particular chunk
2809 * of the request should just be the remainder
2810 * needed to fill up a block.
2812 indirect->len = buf->len - (data->fill_len - data->resid);
2814 camdd_buf_add_child(buf, idb);
2817 * This buffer is ready to execute, so we can take
2818 * it off the pending queue and put it on the run
2821 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2823 dev->num_pending_queue--;
2824 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2825 dev->num_run_queue++;
2827 cur_offset = indirect->offset + indirect->len;
2830 * The resulting I/O would be too large to fit in
2831 * one block. We need to split this I/O into
2832 * multiple pieces. Allocate as many buffers as needed.
2834 for (len_to_go = rb_data->fill_len - rb_data->resid -
2835 indirect->len; len_to_go > 0;) {
2836 struct camdd_buf *new_buf;
2837 struct camdd_buf_data *new_data;
2841 retval = camdd_get_next_lba_len(dev, &lba, &len);
2845 * The device has already been marked
2846 * as EOF, and there is no space left.
2851 new_buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2852 if (new_buf == NULL) {
2860 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2866 indirect = &idb->buf_type_spec.indirect;
2868 indirect->src_buf = read_buf;
2869 read_buf->refcount++;
2870 indirect->offset = cur_offset;
2871 indirect->start_ptr = rb_data->buf + cur_offset;
2872 indirect->len = min(len_to_go, new_buf->len);
2874 if (((indirect->len % dev->sector_size) != 0)
2875 || ((indirect->offset % dev->sector_size) != 0)) {
2876 warnx("offset %ju len %ju not aligned with "
2877 "sector size %u", indirect->offset,
2878 (uintmax_t)indirect->len, dev->sector_size);
2881 cur_offset += indirect->len;
2882 len_to_go -= indirect->len;
2884 camdd_buf_add_child(new_buf, idb);
2886 new_data = &new_buf->buf_type_spec.data;
2888 if ((new_data->fill_len == new_buf->len)
2889 || (eof_flush_needed != 0)) {
2890 STAILQ_INSERT_TAIL(&dev->run_queue,
2892 dev->num_run_queue++;
2893 } else if (new_data->fill_len < buf->len) {
2894 STAILQ_INSERT_TAIL(&dev->pending_queue,
2896 dev->num_pending_queue++;
2898 warnx("%s: too much data in new "
2899 "buffer!", __func__);
2911 camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
2912 uint32_t *peer_depth, uint32_t *our_bytes, uint32_t *peer_bytes)
2914 *our_depth = dev->cur_active_io + dev->num_run_queue;
2915 if (dev->num_peer_work_queue >
2916 dev->num_peer_done_queue)
2917 *peer_depth = dev->num_peer_work_queue -
2918 dev->num_peer_done_queue;
2921 *our_bytes = *our_depth * dev->blocksize;
2922 *peer_bytes = dev->peer_bytes_queued;
2926 camdd_sig_handler(int sig)
2935 sem_post(&camdd_sem);
2939 camdd_print_status(struct camdd_dev *camdd_dev, struct camdd_dev *other_dev,
2940 struct timespec *start_time)
2942 struct timespec done_time;
2944 long double mb_sec, total_sec;
2947 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &done_time);
2949 warn("Unable to get done time");
2953 timespecsub(&done_time, start_time);
2955 total_ns = done_time.tv_nsec + (done_time.tv_sec * 1000000000);
2956 total_sec = total_ns;
2957 total_sec /= 1000000000;
2959 fprintf(stderr, "%ju bytes %s %s\n%ju bytes %s %s\n"
2960 "%.4Lf seconds elapsed\n",
2961 (uintmax_t)camdd_dev->bytes_transferred,
2962 (camdd_dev->write_dev == 0) ? "read from" : "written to",
2963 camdd_dev->device_name,
2964 (uintmax_t)other_dev->bytes_transferred,
2965 (other_dev->write_dev == 0) ? "read from" : "written to",
2966 other_dev->device_name, total_sec);
2968 mb_sec = min(other_dev->bytes_transferred,camdd_dev->bytes_transferred);
2969 mb_sec /= 1024 * 1024;
2970 mb_sec *= 1000000000;
2972 fprintf(stderr, "%.2Lf MB/sec\n", mb_sec);
2976 camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts, uint64_t max_io,
2977 int retry_count, int timeout)
2979 struct cam_device *new_cam_dev = NULL;
2980 struct camdd_dev *devs[2];
2981 struct timespec start_time;
2982 pthread_t threads[2];
2987 if (num_io_opts != 2) {
2988 warnx("Must have one input and one output path");
2993 bzero(devs, sizeof(devs));
2995 for (i = 0; i < num_io_opts; i++) {
2996 switch (io_opts[i].dev_type) {
2997 case CAMDD_DEV_PASS: {
2998 if (isdigit(io_opts[i].dev_name[0])) {
2999 camdd_argmask new_arglist = CAMDD_ARG_NONE;
3000 int bus = 0, target = 0, lun = 0;
3003 /* device specified as bus:target[:lun] */
3004 rv = parse_btl(io_opts[i].dev_name, &bus,
3005 &target, &lun, &new_arglist);
3007 warnx("numeric device specification "
3008 "must be either bus:target, or "
3013 /* default to 0 if lun was not specified */
3014 if ((new_arglist & CAMDD_ARG_LUN) == 0) {
3016 new_arglist |= CAMDD_ARG_LUN;
3018 new_cam_dev = cam_open_btl(bus, target, lun,
3023 if (cam_get_device(io_opts[i].dev_name, name,
3024 sizeof name, &unit) == -1) {
3025 warnx("%s", cam_errbuf);
3029 new_cam_dev = cam_open_spec_device(name, unit,
3033 if (new_cam_dev == NULL) {
3034 warnx("%s", cam_errbuf);
3039 devs[i] = camdd_probe_pass(new_cam_dev,
3040 /*io_opts*/ &io_opts[i],
3041 CAMDD_ARG_ERR_RECOVER,
3042 /*probe_retry_count*/ 3,
3043 /*probe_timeout*/ 5000,
3044 /*io_retry_count*/ retry_count,
3045 /*io_timeout*/ timeout);
3046 if (devs[i] == NULL) {
3047 warn("Unable to probe device %s%u",
3048 new_cam_dev->device_name,
3049 new_cam_dev->dev_unit_num);
3055 case CAMDD_DEV_FILE: {
3058 if (io_opts[i].dev_name[0] == '-') {
3059 if (io_opts[i].write_dev != 0)
3064 if (io_opts[i].write_dev != 0) {
3065 fd = open(io_opts[i].dev_name,
3066 O_RDWR | O_CREAT, S_IWUSR |S_IRUSR);
3068 fd = open(io_opts[i].dev_name,
3073 warn("error opening file %s",
3074 io_opts[i].dev_name);
3079 devs[i] = camdd_probe_file(fd, &io_opts[i],
3080 retry_count, timeout);
3081 if (devs[i] == NULL) {
3089 warnx("Unknown device type %d (%s)",
3090 io_opts[i].dev_type, io_opts[i].dev_name);
3093 break; /*NOTREACHED */
3096 devs[i]->write_dev = io_opts[i].write_dev;
3098 devs[i]->start_offset_bytes = io_opts[i].offset;
3101 devs[i]->sector_io_limit =
3102 (devs[i]->start_offset_bytes /
3103 devs[i]->sector_size) +
3104 (max_io / devs[i]->sector_size) - 1;
3105 devs[i]->sector_io_limit =
3106 (devs[i]->start_offset_bytes /
3107 devs[i]->sector_size) +
3108 (max_io / devs[i]->sector_size) - 1;
3111 devs[i]->next_io_pos_bytes = devs[i]->start_offset_bytes;
3112 devs[i]->next_completion_pos_bytes =devs[i]->start_offset_bytes;
3115 devs[0]->peer_dev = devs[1];
3116 devs[1]->peer_dev = devs[0];
3117 devs[0]->next_peer_pos_bytes = devs[0]->peer_dev->next_io_pos_bytes;
3118 devs[1]->next_peer_pos_bytes = devs[1]->peer_dev->next_io_pos_bytes;
3120 sem_init(&camdd_sem, /*pshared*/ 0, 0);
3122 signal(SIGINFO, camdd_sig_handler);
3123 signal(SIGINT, camdd_sig_handler);
3125 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &start_time);
3127 warn("Unable to get start time");
3131 for (i = 0; i < num_io_opts; i++) {
3132 error = pthread_create(&threads[i], NULL, camdd_worker,
3135 warnc(error, "pthread_create() failed");
3141 if ((sem_wait(&camdd_sem) == -1)
3142 || (need_exit != 0)) {
3145 for (i = 0; i < num_io_opts; i++) {
3146 EV_SET(&ke, (uintptr_t)&devs[i]->work_queue,
3147 EVFILT_USER, 0, NOTE_TRIGGER, 0, NULL);
3149 devs[i]->flags |= CAMDD_DEV_FLAG_EOF;
3151 error = kevent(devs[i]->kq, &ke, 1, NULL, 0,
3154 warn("%s: unable to wake up thread",
3159 } else if (need_status != 0) {
3160 camdd_print_status(devs[0], devs[1], &start_time);
3164 for (i = 0; i < num_io_opts; i++) {
3165 pthread_join(threads[i], NULL);
3168 camdd_print_status(devs[0], devs[1], &start_time);
3172 for (i = 0; i < num_io_opts; i++)
3173 camdd_free_dev(devs[i]);
3175 return (error + error_exit);
3182 "usage: camdd <-i|-o pass=pass0,bs=1M,offset=1M,depth=4>\n"
3183 " <-i|-o file=/tmp/file,bs=512K,offset=1M>\n"
3184 " <-i|-o file=/dev/da0,bs=512K,offset=1M>\n"
3185 " <-i|-o file=/dev/nsa0,bs=512K>\n"
3186 " [-C retry_count][-E][-m max_io_amt][-t timeout_secs][-v][-h]\n"
3187 "Option description\n"
3188 "-i <arg=val> Specify input device/file and parameters\n"
3189 "-o <arg=val> Specify output device/file and parameters\n"
3190 "Input and Output parameters\n"
3191 "pass=name Specify a pass(4) device like pass0 or /dev/pass0\n"
3192 "file=name Specify a file or device, /tmp/foo, /dev/da0, /dev/null\n"
3193 " or - for stdin/stdout\n"
3194 "bs=blocksize Specify blocksize in bytes, or using K, M, G, etc. suffix\n"
3195 "offset=len Specify starting offset in bytes or using K, M, G suffix\n"
3196 " NOTE: offset cannot be specified on tapes, pipes, stdin/out\n"
3197 "depth=N Specify a numeric queue depth. This only applies to pass(4)\n"
3198 "mcs=N Specify a minimum cmd size for pass(4) read/write commands\n"
3199 "Optional arguments\n"
3200 "-C retry_cnt Specify a retry count for pass(4) devices\n"
3201 "-E Enable CAM error recovery for pass(4) devices\n"
3202 "-m max_io Specify the maximum amount to be transferred in bytes or\n"
3203 " using K, G, M, etc. suffixes\n"
3204 "-t timeout Specify the I/O timeout to use with pass(4) devices\n"
3205 "-v Enable verbose error recovery\n"
3206 "-h Print this message\n");
3211 camdd_parse_io_opts(char *args, int is_write, struct camdd_io_opts *io_opts)
3213 char *tmpstr, *tmpstr2;
3214 char *orig_tmpstr = NULL;
3217 io_opts->write_dev = is_write;
3219 tmpstr = strdup(args);
3220 if (tmpstr == NULL) {
3221 warn("strdup failed");
3225 orig_tmpstr = tmpstr;
3226 while ((tmpstr2 = strsep(&tmpstr, ",")) != NULL) {
3230 * If the user creates an empty parameter by putting in two
3231 * commas, skip over it and look for the next field.
3233 if (*tmpstr2 == '\0')
3236 name = strsep(&tmpstr2, "=");
3237 if (*name == '\0') {
3238 warnx("Got empty I/O parameter name");
3242 value = strsep(&tmpstr2, "=");
3244 || (*value == '\0')) {
3245 warnx("Empty I/O parameter value for %s", name);
3249 if (strncasecmp(name, "file", 4) == 0) {
3250 io_opts->dev_type = CAMDD_DEV_FILE;
3251 io_opts->dev_name = strdup(value);
3252 if (io_opts->dev_name == NULL) {
3253 warn("Error allocating memory");
3257 } else if (strncasecmp(name, "pass", 4) == 0) {
3258 io_opts->dev_type = CAMDD_DEV_PASS;
3259 io_opts->dev_name = strdup(value);
3260 if (io_opts->dev_name == NULL) {
3261 warn("Error allocating memory");
3265 } else if ((strncasecmp(name, "bs", 2) == 0)
3266 || (strncasecmp(name, "blocksize", 9) == 0)) {
3267 retval = expand_number(value, &io_opts->blocksize);
3269 warn("expand_number(3) failed on %s=%s", name,
3274 } else if (strncasecmp(name, "depth", 5) == 0) {
3277 io_opts->queue_depth = strtoull(value, &endptr, 0);
3278 if (*endptr != '\0') {
3279 warnx("invalid queue depth %s", value);
3283 } else if (strncasecmp(name, "mcs", 3) == 0) {
3286 io_opts->min_cmd_size = strtol(value, &endptr, 0);
3287 if ((*endptr != '\0')
3288 || ((io_opts->min_cmd_size > 16)
3289 || (io_opts->min_cmd_size < 0))) {
3290 warnx("invalid minimum cmd size %s", value);
3294 } else if (strncasecmp(name, "offset", 6) == 0) {
3295 retval = expand_number(value, &io_opts->offset);
3297 warn("expand_number(3) failed on %s=%s", name,
3302 } else if (strncasecmp(name, "debug", 5) == 0) {
3305 io_opts->debug = strtoull(value, &endptr, 0);
3306 if (*endptr != '\0') {
3307 warnx("invalid debug level %s", value);
3312 warnx("Unrecognized parameter %s=%s", name, value);
3322 main(int argc, char **argv)
3325 camdd_argmask arglist = CAMDD_ARG_NONE;
3326 int timeout = 0, retry_count = 1;
3328 uint64_t max_io = 0;
3329 struct camdd_io_opts *opt_list = NULL;
3336 opt_list = calloc(2, sizeof(struct camdd_io_opts));
3337 if (opt_list == NULL) {
3338 warn("Unable to allocate option list");
3343 while ((c = getopt(argc, argv, "C:Ehi:m:o:t:v")) != -1){
3346 retry_count = strtol(optarg, NULL, 0);
3347 if (retry_count < 0)
3348 errx(1, "retry count %d is < 0",
3350 arglist |= CAMDD_ARG_RETRIES;
3353 arglist |= CAMDD_ARG_ERR_RECOVER;
3358 && (opt_list[0].dev_type != CAMDD_DEV_NONE))
3360 && (opt_list[1].dev_type != CAMDD_DEV_NONE))) {
3361 errx(1, "Only one input and output path "
3364 error = camdd_parse_io_opts(optarg, (c == 'o') ? 1 : 0,
3365 (c == 'o') ? &opt_list[1] : &opt_list[0]);
3370 error = expand_number(optarg, &max_io);
3372 warn("invalid maximum I/O amount %s", optarg);
3378 timeout = strtol(optarg, NULL, 0);
3380 errx(1, "invalid timeout %d", timeout);
3381 /* Convert the timeout from seconds to ms */
3383 arglist |= CAMDD_ARG_TIMEOUT;
3386 arglist |= CAMDD_ARG_VERBOSE;
3392 break; /*NOTREACHED*/
3396 if ((opt_list[0].dev_type == CAMDD_DEV_NONE)
3397 || (opt_list[1].dev_type == CAMDD_DEV_NONE))
3398 errx(1, "Must specify both -i and -o");
3401 * Set the timeout if the user hasn't specified one.
3404 timeout = CAMDD_PASS_RW_TIMEOUT;
3406 error = camdd_rw(opt_list, 2, max_io, retry_count, timeout);