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 {
264 struct cam_device *dev;
282 CAMDD_FF_NONE = 0x00,
283 CAMDD_FF_CAN_SEEK = 0x01
286 struct camdd_dev_file {
289 char filename[MAXPATHLEN + 1];
290 camdd_file_type file_type;
291 camdd_file_flags file_flags;
295 struct camdd_dev_block {
301 union camdd_dev_spec {
302 struct camdd_dev_pass pass;
303 struct camdd_dev_file file;
304 struct camdd_dev_block block;
308 CAMDD_DEV_FLAG_NONE = 0x00,
309 CAMDD_DEV_FLAG_EOF = 0x01,
310 CAMDD_DEV_FLAG_PEER_EOF = 0x02,
311 CAMDD_DEV_FLAG_ACTIVE = 0x04,
312 CAMDD_DEV_FLAG_EOF_SENT = 0x08,
313 CAMDD_DEV_FLAG_EOF_QUEUED = 0x10
317 camdd_dev_type dev_type;
318 union camdd_dev_spec dev_spec;
319 camdd_dev_flags flags;
320 char device_name[MAXPATHLEN+1];
322 uint32_t sector_size;
324 uint64_t sector_io_limit;
330 uint64_t start_offset_bytes;
331 uint64_t next_io_pos_bytes;
332 uint64_t next_peer_pos_bytes;
333 uint64_t next_completion_pos_bytes;
334 uint64_t peer_bytes_queued;
335 uint64_t bytes_transferred;
336 uint32_t target_queue_depth;
337 uint32_t cur_active_io;
339 uint32_t extra_buf_len;
340 struct camdd_dev *peer_dev;
341 pthread_mutex_t mutex;
345 int (*run)(struct camdd_dev *dev);
346 int (*fetch)(struct camdd_dev *dev);
349 * Buffers that are available for I/O. Uses links.
351 STAILQ_HEAD(,camdd_buf) free_queue;
354 * Free indirect buffers. These are used for breaking a large
355 * buffer into multiple pieces.
357 STAILQ_HEAD(,camdd_buf) free_indirect_queue;
360 * Buffers that have been queued to the kernel. Uses links.
362 STAILQ_HEAD(,camdd_buf) active_queue;
365 * Will generally contain one of our buffers that is waiting for enough
366 * I/O from our partner thread to be able to execute. This will
367 * generally happen when our per-I/O-size is larger than the
368 * partner thread's per-I/O-size. Uses links.
370 STAILQ_HEAD(,camdd_buf) pending_queue;
373 * Number of buffers on the pending queue
375 int num_pending_queue;
378 * Buffers that are filled and ready to execute. This is used when
379 * our partner (reader) thread sends us blocks that are larger than
380 * our blocksize, and so we have to split them into multiple pieces.
382 STAILQ_HEAD(,camdd_buf) run_queue;
385 * Number of buffers on the run queue.
389 STAILQ_HEAD(,camdd_buf) reorder_queue;
391 int num_reorder_queue;
394 * Buffers that have been queued to us by our partner thread
395 * (generally the reader thread) to be written out. Uses
398 STAILQ_HEAD(,camdd_buf) work_queue;
401 * Buffers that have been completed by our partner thread. Uses
404 STAILQ_HEAD(,camdd_buf) peer_done_queue;
407 * Number of buffers on the peer done queue.
409 uint32_t num_peer_done_queue;
412 * A list of buffers that we have queued to our peer thread. Uses
415 STAILQ_HEAD(,camdd_buf) peer_work_queue;
418 * Number of buffers on the peer work queue.
420 uint32_t num_peer_work_queue;
423 static sem_t camdd_sem;
424 static sig_atomic_t need_exit = 0;
425 static sig_atomic_t error_exit = 0;
426 static sig_atomic_t need_status = 0;
429 #define min(a, b) (a < b) ? a : b
433 * XXX KDM private copy of timespecsub(). This is normally defined in
434 * sys/time.h, but is only enabled in the kernel. If that definition is
435 * enabled in userland, it breaks the build of libnetbsd.
438 #define timespecsub(vvp, uvp) \
440 (vvp)->tv_sec -= (uvp)->tv_sec; \
441 (vvp)->tv_nsec -= (uvp)->tv_nsec; \
442 if ((vvp)->tv_nsec < 0) { \
444 (vvp)->tv_nsec += 1000000000; \
450 /* Generically useful offsets into the peripheral private area */
451 #define ppriv_ptr0 periph_priv.entries[0].ptr
452 #define ppriv_ptr1 periph_priv.entries[1].ptr
453 #define ppriv_field0 periph_priv.entries[0].field
454 #define ppriv_field1 periph_priv.entries[1].field
456 #define ccb_buf ppriv_ptr0
458 #define CAMDD_FILE_DEFAULT_BLOCK 524288
459 #define CAMDD_FILE_DEFAULT_DEPTH 1
460 #define CAMDD_PASS_MAX_BLOCK 1048576
461 #define CAMDD_PASS_DEFAULT_DEPTH 6
462 #define CAMDD_PASS_RW_TIMEOUT 60 * 1000
464 static int parse_btl(char *tstr, int *bus, int *target, int *lun,
465 camdd_argmask *arglst);
466 void camdd_free_dev(struct camdd_dev *dev);
467 struct camdd_dev *camdd_alloc_dev(camdd_dev_type dev_type,
468 struct kevent *new_ke, int num_ke,
469 int retry_count, int timeout);
470 static struct camdd_buf *camdd_alloc_buf(struct camdd_dev *dev,
471 camdd_buf_type buf_type);
472 void camdd_release_buf(struct camdd_buf *buf);
473 struct camdd_buf *camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type);
474 int camdd_buf_sg_create(struct camdd_buf *buf, int iovec,
475 uint32_t sector_size, uint32_t *num_sectors_used,
476 int *double_buf_needed);
477 uint32_t camdd_buf_get_len(struct camdd_buf *buf);
478 void camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf);
479 int camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
480 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran);
481 int camdd_probe_pass_scsi(struct cam_device *cam_dev, union ccb *ccb,
482 camdd_argmask arglist, int probe_retry_count,
483 int probe_timeout, uint64_t *maxsector, uint32_t *block_len);
484 struct camdd_dev *camdd_probe_file(int fd, struct camdd_io_opts *io_opts,
485 int retry_count, int timeout);
486 struct camdd_dev *camdd_probe_pass(struct cam_device *cam_dev,
487 struct camdd_io_opts *io_opts,
488 camdd_argmask arglist, int probe_retry_count,
489 int probe_timeout, int io_retry_count,
491 void *camdd_file_worker(void *arg);
492 camdd_buf_status camdd_ccb_status(union ccb *ccb, int protocol);
493 int camdd_get_cgd(struct cam_device *device, struct ccb_getdev *cgd);
494 int camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf);
495 int camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf);
496 void camdd_peer_done(struct camdd_buf *buf);
497 void camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
499 int camdd_pass_fetch(struct camdd_dev *dev);
500 int camdd_file_run(struct camdd_dev *dev);
501 int camdd_pass_run(struct camdd_dev *dev);
502 int camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len);
503 int camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf);
504 void camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
505 uint32_t *peer_depth, uint32_t *our_bytes,
506 uint32_t *peer_bytes);
507 void *camdd_worker(void *arg);
508 void camdd_sig_handler(int sig);
509 void camdd_print_status(struct camdd_dev *camdd_dev,
510 struct camdd_dev *other_dev,
511 struct timespec *start_time);
512 int camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts,
513 uint64_t max_io, int retry_count, int timeout);
514 int camdd_parse_io_opts(char *args, int is_write,
515 struct camdd_io_opts *io_opts);
519 * Parse out a bus, or a bus, target and lun in the following
525 * Returns the number of parsed components, or 0.
528 parse_btl(char *tstr, int *bus, int *target, int *lun, camdd_argmask *arglst)
533 while (isspace(*tstr) && (*tstr != '\0'))
536 tmpstr = (char *)strtok(tstr, ":");
537 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
538 *bus = strtol(tmpstr, NULL, 0);
539 *arglst |= CAMDD_ARG_BUS;
541 tmpstr = (char *)strtok(NULL, ":");
542 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
543 *target = strtol(tmpstr, NULL, 0);
544 *arglst |= CAMDD_ARG_TARGET;
546 tmpstr = (char *)strtok(NULL, ":");
547 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
548 *lun = strtol(tmpstr, NULL, 0);
549 *arglst |= CAMDD_ARG_LUN;
559 * XXX KDM clean up and free all of the buffers on the queue!
562 camdd_free_dev(struct camdd_dev *dev)
567 switch (dev->dev_type) {
568 case CAMDD_DEV_FILE: {
569 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
571 if (file_dev->fd != -1)
573 free(file_dev->tmp_buf);
576 case CAMDD_DEV_PASS: {
577 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
579 if (pass_dev->dev != NULL)
580 cam_close_device(pass_dev->dev);
591 camdd_alloc_dev(camdd_dev_type dev_type, struct kevent *new_ke, int num_ke,
592 int retry_count, int timeout)
594 struct camdd_dev *dev = NULL;
599 dev = calloc(1, sizeof(*dev));
601 warn("%s: unable to malloc %zu bytes", __func__, sizeof(*dev));
605 dev->dev_type = dev_type;
606 dev->io_timeout = timeout;
607 dev->retry_count = retry_count;
608 STAILQ_INIT(&dev->free_queue);
609 STAILQ_INIT(&dev->free_indirect_queue);
610 STAILQ_INIT(&dev->active_queue);
611 STAILQ_INIT(&dev->pending_queue);
612 STAILQ_INIT(&dev->run_queue);
613 STAILQ_INIT(&dev->reorder_queue);
614 STAILQ_INIT(&dev->work_queue);
615 STAILQ_INIT(&dev->peer_done_queue);
616 STAILQ_INIT(&dev->peer_work_queue);
617 retval = pthread_mutex_init(&dev->mutex, NULL);
619 warnc(retval, "%s: failed to initialize mutex", __func__);
623 retval = pthread_cond_init(&dev->cond, NULL);
625 warnc(retval, "%s: failed to initialize condition variable",
632 warn("%s: Unable to create kqueue", __func__);
636 ke_size = sizeof(struct kevent) * (num_ke + 4);
637 ke = calloc(1, ke_size);
639 warn("%s: unable to malloc %zu bytes", __func__, ke_size);
643 bcopy(new_ke, ke, num_ke * sizeof(struct kevent));
645 EV_SET(&ke[num_ke++], (uintptr_t)&dev->work_queue, EVFILT_USER,
646 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
647 EV_SET(&ke[num_ke++], (uintptr_t)&dev->peer_done_queue, EVFILT_USER,
648 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
649 EV_SET(&ke[num_ke++], SIGINFO, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
650 EV_SET(&ke[num_ke++], SIGINT, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
652 retval = kevent(dev->kq, ke, num_ke, NULL, 0, NULL);
654 warn("%s: Unable to register kevents", __func__);
667 static struct camdd_buf *
668 camdd_alloc_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
670 struct camdd_buf *buf = NULL;
671 uint8_t *data_ptr = NULL;
674 * We only need to allocate data space for data buffers.
678 data_ptr = malloc(dev->blocksize);
679 if (data_ptr == NULL) {
680 warn("unable to allocate %u bytes", dev->blocksize);
688 buf = calloc(1, sizeof(*buf));
690 warn("unable to allocate %zu bytes", sizeof(*buf));
694 buf->buf_type = buf_type;
697 case CAMDD_BUF_DATA: {
698 struct camdd_buf_data *data;
700 data = &buf->buf_type_spec.data;
702 data->alloc_len = dev->blocksize;
703 data->buf = data_ptr;
706 case CAMDD_BUF_INDIRECT:
711 STAILQ_INIT(&buf->src_list);
722 camdd_release_buf(struct camdd_buf *buf)
724 struct camdd_dev *dev;
728 switch (buf->buf_type) {
729 case CAMDD_BUF_DATA: {
730 struct camdd_buf_data *data;
732 data = &buf->buf_type_spec.data;
734 if (data->segs != NULL) {
735 if (data->extra_buf != 0) {
739 data->segs[data->sg_count - 1].ds_addr;
746 } else if (data->iovec != NULL) {
747 if (data->extra_buf != 0) {
748 free(data->iovec[data->sg_count - 1].iov_base);
755 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
758 case CAMDD_BUF_INDIRECT:
759 STAILQ_INSERT_TAIL(&dev->free_indirect_queue, buf, links);
762 err(1, "%s: Invalid buffer type %d for released buffer",
763 __func__, buf->buf_type);
769 camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
771 struct camdd_buf *buf = NULL;
775 buf = STAILQ_FIRST(&dev->free_queue);
777 struct camdd_buf_data *data;
781 STAILQ_REMOVE_HEAD(&dev->free_queue, links);
782 data = &buf->buf_type_spec.data;
783 data_ptr = data->buf;
784 alloc_len = data->alloc_len;
785 bzero(buf, sizeof(*buf));
786 data->buf = data_ptr;
787 data->alloc_len = alloc_len;
790 case CAMDD_BUF_INDIRECT:
791 buf = STAILQ_FIRST(&dev->free_indirect_queue);
793 STAILQ_REMOVE_HEAD(&dev->free_indirect_queue, links);
795 bzero(buf, sizeof(*buf));
799 warnx("Unknown buffer type %d requested", buf_type);
805 return (camdd_alloc_buf(dev, buf_type));
807 STAILQ_INIT(&buf->src_list);
809 buf->buf_type = buf_type;
816 camdd_buf_sg_create(struct camdd_buf *buf, int iovec, uint32_t sector_size,
817 uint32_t *num_sectors_used, int *double_buf_needed)
819 struct camdd_buf *tmp_buf;
820 struct camdd_buf_data *data;
821 uint8_t *extra_buf = NULL;
822 size_t extra_buf_len = 0;
823 int extra_buf_attached = 0;
826 data = &buf->buf_type_spec.data;
828 data->sg_count = buf->src_count;
830 * Compose a scatter/gather list from all of the buffers in the list.
831 * If the length of the buffer isn't a multiple of the sector size,
832 * we'll have to add an extra buffer. This should only happen
833 * at the end of a transfer.
835 if ((data->fill_len % sector_size) != 0) {
836 extra_buf_len = sector_size - (data->fill_len % sector_size);
837 extra_buf = calloc(extra_buf_len, 1);
838 if (extra_buf == NULL) {
839 warn("%s: unable to allocate %zu bytes for extra "
840 "buffer space", __func__, extra_buf_len);
848 data->segs = calloc(data->sg_count, sizeof(bus_dma_segment_t));
849 if (data->segs == NULL) {
850 warn("%s: unable to allocate %zu bytes for S/G list",
851 __func__, sizeof(bus_dma_segment_t) *
858 data->iovec = calloc(data->sg_count, sizeof(struct iovec));
859 if (data->iovec == NULL) {
860 warn("%s: unable to allocate %zu bytes for S/G list",
861 __func__, sizeof(struct iovec) * data->sg_count);
867 for (i = 0, tmp_buf = STAILQ_FIRST(&buf->src_list);
868 i < buf->src_count && tmp_buf != NULL; i++,
869 tmp_buf = STAILQ_NEXT(tmp_buf, src_links)) {
871 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
872 struct camdd_buf_data *tmp_data;
874 tmp_data = &tmp_buf->buf_type_spec.data;
876 data->segs[i].ds_addr =
877 (bus_addr_t) tmp_data->buf;
878 data->segs[i].ds_len = tmp_data->fill_len -
881 data->iovec[i].iov_base = tmp_data->buf;
882 data->iovec[i].iov_len = tmp_data->fill_len -
885 if (((tmp_data->fill_len - tmp_data->resid) %
887 *double_buf_needed = 1;
889 struct camdd_buf_indirect *tmp_ind;
891 tmp_ind = &tmp_buf->buf_type_spec.indirect;
893 data->segs[i].ds_addr =
894 (bus_addr_t)tmp_ind->start_ptr;
895 data->segs[i].ds_len = tmp_ind->len;
897 data->iovec[i].iov_base = tmp_ind->start_ptr;
898 data->iovec[i].iov_len = tmp_ind->len;
900 if ((tmp_ind->len % sector_size) != 0)
901 *double_buf_needed = 1;
905 if (extra_buf != NULL) {
907 data->segs[i].ds_addr = (bus_addr_t)extra_buf;
908 data->segs[i].ds_len = extra_buf_len;
910 data->iovec[i].iov_base = extra_buf;
911 data->iovec[i].iov_len = extra_buf_len;
913 extra_buf_attached = 1;
916 if ((tmp_buf != NULL) || (i != data->sg_count)) {
917 warnx("buffer source count does not match "
918 "number of buffers in list!");
925 *num_sectors_used = (data->fill_len + extra_buf_len) /
927 } else if (extra_buf_attached == 0) {
929 * If extra_buf isn't attached yet, we need to free it
940 camdd_buf_get_len(struct camdd_buf *buf)
944 if (buf->buf_type != CAMDD_BUF_DATA) {
945 struct camdd_buf_indirect *indirect;
947 indirect = &buf->buf_type_spec.indirect;
950 struct camdd_buf_data *data;
952 data = &buf->buf_type_spec.data;
953 len = data->fill_len;
960 camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf)
962 struct camdd_buf_data *data;
964 assert(buf->buf_type == CAMDD_BUF_DATA);
966 data = &buf->buf_type_spec.data;
968 STAILQ_INSERT_TAIL(&buf->src_list, child_buf, src_links);
971 data->fill_len += camdd_buf_get_len(child_buf);
979 } camdd_status_item_index;
981 static struct camdd_status_items {
983 struct mt_status_entry *entry;
984 } req_status_items[] = {
987 { "blk_gran", NULL },
988 { "max_effective_iosize", NULL }
992 camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
993 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran)
995 struct mt_status_data status_data;
996 char *xml_str = NULL;
1000 retval = mt_get_xml_str(fd, MTIOCEXTGET, &xml_str);
1002 err(1, "Couldn't get XML string from %s", filename);
1004 retval = mt_get_status(xml_str, &status_data);
1005 if (retval != XML_STATUS_OK) {
1006 warn("couldn't get status for %s", filename);
1012 if (status_data.error != 0) {
1013 warnx("%s", status_data.error_str);
1018 for (i = 0; i < nitems(req_status_items); i++) {
1021 name = __DECONST(char *, req_status_items[i].name);
1022 req_status_items[i].entry = mt_status_entry_find(&status_data,
1024 if (req_status_items[i].entry == NULL) {
1025 errx(1, "Cannot find status entry %s",
1026 req_status_items[i].name);
1030 *max_iosize = req_status_items[CAMDD_TS_EFF_IOSIZE].entry->value_unsigned;
1031 *max_blk= req_status_items[CAMDD_TS_MAX_BLK].entry->value_unsigned;
1032 *min_blk= req_status_items[CAMDD_TS_MIN_BLK].entry->value_unsigned;
1033 *blk_gran = req_status_items[CAMDD_TS_BLK_GRAN].entry->value_unsigned;
1037 mt_status_free(&status_data);
1043 camdd_probe_file(int fd, struct camdd_io_opts *io_opts, int retry_count,
1046 struct camdd_dev *dev = NULL;
1047 struct camdd_dev_file *file_dev;
1048 uint64_t blocksize = io_opts->blocksize;
1050 dev = camdd_alloc_dev(CAMDD_DEV_FILE, NULL, 0, retry_count, timeout);
1054 file_dev = &dev->dev_spec.file;
1056 strlcpy(file_dev->filename, io_opts->dev_name,
1057 sizeof(file_dev->filename));
1058 strlcpy(dev->device_name, io_opts->dev_name, sizeof(dev->device_name));
1060 dev->blocksize = CAMDD_FILE_DEFAULT_BLOCK;
1062 dev->blocksize = blocksize;
1064 if ((io_opts->queue_depth != 0)
1065 && (io_opts->queue_depth != 1)) {
1066 warnx("Queue depth %ju for %s ignored, only 1 outstanding "
1067 "command supported", (uintmax_t)io_opts->queue_depth,
1070 dev->target_queue_depth = CAMDD_FILE_DEFAULT_DEPTH;
1071 dev->run = camdd_file_run;
1075 * We can effectively access files on byte boundaries. We'll reset
1076 * this for devices like disks that can be accessed on sector
1079 dev->sector_size = 1;
1081 if ((fd != STDIN_FILENO)
1082 && (fd != STDOUT_FILENO)) {
1085 retval = fstat(fd, &file_dev->sb);
1087 warn("Cannot stat %s", dev->device_name);
1090 if (S_ISREG(file_dev->sb.st_mode)) {
1091 file_dev->file_type = CAMDD_FILE_REG;
1092 } else if (S_ISCHR(file_dev->sb.st_mode)) {
1095 if (ioctl(fd, FIODTYPE, &type) == -1)
1096 err(1, "FIODTYPE ioctl failed on %s",
1100 file_dev->file_type = CAMDD_FILE_TAPE;
1101 else if (type & D_DISK)
1102 file_dev->file_type = CAMDD_FILE_DISK;
1103 else if (type & D_MEM)
1104 file_dev->file_type = CAMDD_FILE_MEM;
1105 else if (type & D_TTY)
1106 file_dev->file_type = CAMDD_FILE_TTY;
1108 } else if (S_ISDIR(file_dev->sb.st_mode)) {
1109 errx(1, "cannot operate on directory %s",
1111 } else if (S_ISFIFO(file_dev->sb.st_mode)) {
1112 file_dev->file_type = CAMDD_FILE_PIPE;
1114 errx(1, "Cannot determine file type for %s",
1117 switch (file_dev->file_type) {
1118 case CAMDD_FILE_REG:
1119 if (file_dev->sb.st_size != 0)
1120 dev->max_sector = file_dev->sb.st_size - 1;
1122 dev->max_sector = 0;
1123 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1125 case CAMDD_FILE_TAPE: {
1126 uint64_t max_iosize, max_blk, min_blk, blk_gran;
1128 * Check block limits and maximum effective iosize.
1129 * Make sure the blocksize is within the block
1130 * limits (and a multiple of the minimum blocksize)
1131 * and that the blocksize is <= maximum effective
1134 retval = camdd_probe_tape(fd, dev->device_name,
1135 &max_iosize, &max_blk, &min_blk, &blk_gran);
1137 errx(1, "Unable to probe tape %s",
1141 * The blocksize needs to be <= the maximum
1142 * effective I/O size of the tape device. Note
1143 * that this also takes into account the maximum
1144 * blocksize reported by READ BLOCK LIMITS.
1146 if (dev->blocksize > max_iosize) {
1147 warnx("Blocksize %u too big for %s, limiting "
1148 "to %ju", dev->blocksize, dev->device_name,
1150 dev->blocksize = max_iosize;
1154 * The blocksize needs to be at least min_blk;
1156 if (dev->blocksize < min_blk) {
1157 warnx("Blocksize %u too small for %s, "
1158 "increasing to %ju", dev->blocksize,
1159 dev->device_name, min_blk);
1160 dev->blocksize = min_blk;
1164 * And the blocksize needs to be a multiple of
1165 * the block granularity.
1168 && (dev->blocksize % (1 << blk_gran))) {
1169 warnx("Blocksize %u for %s not a multiple of "
1170 "%d, adjusting to %d", dev->blocksize,
1171 dev->device_name, (1 << blk_gran),
1172 dev->blocksize & ~((1 << blk_gran) - 1));
1173 dev->blocksize &= ~((1 << blk_gran) - 1);
1176 if (dev->blocksize == 0) {
1177 errx(1, "Unable to derive valid blocksize for "
1178 "%s", dev->device_name);
1182 * For tape drives, set the sector size to the
1183 * blocksize so that we make sure not to write
1184 * less than the blocksize out to the drive.
1186 dev->sector_size = dev->blocksize;
1189 case CAMDD_FILE_DISK: {
1191 unsigned int sector_size;
1193 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1195 if (ioctl(fd, DIOCGSECTORSIZE, §or_size) == -1) {
1196 err(1, "DIOCGSECTORSIZE ioctl failed on %s",
1200 if (sector_size == 0) {
1201 errx(1, "DIOCGSECTORSIZE ioctl returned "
1202 "invalid sector size %u for %s",
1203 sector_size, dev->device_name);
1206 if (ioctl(fd, DIOCGMEDIASIZE, &media_size) == -1) {
1207 err(1, "DIOCGMEDIASIZE ioctl failed on %s",
1211 if (media_size == 0) {
1212 errx(1, "DIOCGMEDIASIZE ioctl returned "
1213 "invalid media size %ju for %s",
1214 (uintmax_t)media_size, dev->device_name);
1217 if (dev->blocksize % sector_size) {
1218 errx(1, "%s blocksize %u not a multiple of "
1219 "sector size %u", dev->device_name,
1220 dev->blocksize, sector_size);
1223 dev->sector_size = sector_size;
1224 dev->max_sector = (media_size / sector_size) - 1;
1227 case CAMDD_FILE_MEM:
1228 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1235 if ((io_opts->offset != 0)
1236 && ((file_dev->file_flags & CAMDD_FF_CAN_SEEK) == 0)) {
1237 warnx("Offset %ju specified for %s, but we cannot seek on %s",
1238 io_opts->offset, io_opts->dev_name, io_opts->dev_name);
1242 else if ((io_opts->offset != 0)
1243 && ((io_opts->offset % dev->sector_size) != 0)) {
1244 warnx("Offset %ju for %s is not a multiple of the "
1245 "sector size %u", io_opts->offset,
1246 io_opts->dev_name, dev->sector_size);
1249 dev->start_offset_bytes = io_opts->offset;
1257 camdd_free_dev(dev);
1262 * Get a get device CCB for the specified device.
1265 camdd_get_cgd(struct cam_device *device, struct ccb_getdev *cgd)
1270 ccb = cam_getccb(device);
1273 warnx("%s: couldn't allocate CCB", __func__);
1277 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cgd);
1279 ccb->ccb_h.func_code = XPT_GDEV_TYPE;
1281 if (cam_send_ccb(device, ccb) < 0) {
1282 warn("%s: error sending Get Device Information CCB", __func__);
1283 cam_error_print(device, ccb, CAM_ESF_ALL,
1284 CAM_EPF_ALL, stderr);
1289 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1290 cam_error_print(device, ccb, CAM_ESF_ALL,
1291 CAM_EPF_ALL, stderr);
1296 bcopy(&ccb->cgd, cgd, sizeof(struct ccb_getdev));
1305 camdd_probe_pass_scsi(struct cam_device *cam_dev, union ccb *ccb,
1306 camdd_argmask arglist, int probe_retry_count,
1307 int probe_timeout, uint64_t *maxsector, uint32_t *block_len)
1309 struct scsi_read_capacity_data rcap;
1310 struct scsi_read_capacity_data_long rcaplong;
1314 warnx("%s: error passed ccb is NULL", __func__);
1318 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
1320 scsi_read_capacity(&ccb->csio,
1321 /*retries*/ probe_retry_count,
1323 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1326 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1328 /* Disable freezing the device queue */
1329 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1331 if (arglist & CAMDD_ARG_ERR_RECOVER)
1332 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1334 if (cam_send_ccb(cam_dev, ccb) < 0) {
1335 warn("error sending READ CAPACITY command");
1337 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1338 CAM_EPF_ALL, stderr);
1343 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1344 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1348 *maxsector = scsi_4btoul(rcap.addr);
1349 *block_len = scsi_4btoul(rcap.length);
1352 * A last block of 2^32-1 means that the true capacity is over 2TB,
1353 * and we need to issue the long READ CAPACITY to get the real
1354 * capacity. Otherwise, we're all set.
1356 if (*maxsector != 0xffffffff) {
1361 scsi_read_capacity_16(&ccb->csio,
1362 /*retries*/ probe_retry_count,
1364 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1368 (uint8_t *)&rcaplong,
1370 /*sense_len*/ SSD_FULL_SIZE,
1371 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1373 /* Disable freezing the device queue */
1374 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1376 if (arglist & CAMDD_ARG_ERR_RECOVER)
1377 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1379 if (cam_send_ccb(cam_dev, ccb) < 0) {
1380 warn("error sending READ CAPACITY (16) command");
1381 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1382 CAM_EPF_ALL, stderr);
1386 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1387 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1391 *maxsector = scsi_8btou64(rcaplong.addr);
1392 *block_len = scsi_4btoul(rcaplong.length);
1401 * Need to implement this. Do a basic probe:
1402 * - Check the inquiry data, make sure we're talking to a device that we
1403 * can reasonably expect to talk to -- direct, RBC, CD, WORM.
1404 * - Send a test unit ready, make sure the device is available.
1405 * - Get the capacity and block size.
1408 camdd_probe_pass(struct cam_device *cam_dev, struct camdd_io_opts *io_opts,
1409 camdd_argmask arglist, int probe_retry_count,
1410 int probe_timeout, int io_retry_count, int io_timeout)
1413 uint64_t maxsector = 0;
1414 uint32_t cpi_maxio, max_iosize, pass_numblocks;
1415 uint32_t block_len = 0;
1416 struct camdd_dev *dev = NULL;
1417 struct camdd_dev_pass *pass_dev;
1419 struct ccb_getdev cgd;
1423 if ((retval = camdd_get_cgd(cam_dev, &cgd)) != 0) {
1424 warnx("%s: error retrieving CGD", __func__);
1428 ccb = cam_getccb(cam_dev);
1431 warnx("%s: error allocating ccb", __func__);
1435 switch (cgd.protocol) {
1437 scsi_dev_type = SID_TYPE(&cam_dev->inq_data);
1440 * For devices that support READ CAPACITY, we'll attempt to get the
1441 * capacity. Otherwise, we really don't support tape or other
1442 * devices via SCSI passthrough, so just return an error in that case.
1444 switch (scsi_dev_type) {
1453 errx(1, "Unsupported SCSI device type %d", scsi_dev_type);
1454 break; /*NOTREACHED*/
1457 if ((retval = camdd_probe_pass_scsi(cam_dev, ccb, probe_retry_count,
1458 arglist, probe_timeout, &maxsector,
1464 errx(1, "Unsupported PROTO type %d", cgd.protocol);
1465 break; /*NOTREACHED*/
1468 if (block_len == 0) {
1469 warnx("Sector size for %s%u is 0, cannot continue",
1470 cam_dev->device_name, cam_dev->dev_unit_num);
1474 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cpi);
1476 ccb->ccb_h.func_code = XPT_PATH_INQ;
1477 ccb->ccb_h.flags = CAM_DIR_NONE;
1478 ccb->ccb_h.retry_count = 1;
1480 if (cam_send_ccb(cam_dev, ccb) < 0) {
1481 warn("error sending XPT_PATH_INQ CCB");
1483 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1484 CAM_EPF_ALL, stderr);
1488 EV_SET(&ke, cam_dev->fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1490 dev = camdd_alloc_dev(CAMDD_DEV_PASS, &ke, 1, io_retry_count,
1495 pass_dev = &dev->dev_spec.pass;
1496 pass_dev->scsi_dev_type = scsi_dev_type;
1497 pass_dev->protocol = cgd.protocol;
1498 pass_dev->dev = cam_dev;
1499 pass_dev->max_sector = maxsector;
1500 pass_dev->block_len = block_len;
1501 pass_dev->cpi_maxio = ccb->cpi.maxio;
1502 snprintf(dev->device_name, sizeof(dev->device_name), "%s%u",
1503 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
1504 dev->sector_size = block_len;
1505 dev->max_sector = maxsector;
1509 * Determine the optimal blocksize to use for this device.
1513 * If the controller has not specified a maximum I/O size,
1514 * just go with 128K as a somewhat conservative value.
1516 if (pass_dev->cpi_maxio == 0)
1519 cpi_maxio = pass_dev->cpi_maxio;
1522 * If the controller has a large maximum I/O size, limit it
1523 * to something smaller so that the kernel doesn't have trouble
1524 * allocating buffers to copy data in and out for us.
1525 * XXX KDM this is until we have unmapped I/O support in the kernel.
1527 max_iosize = min(cpi_maxio, CAMDD_PASS_MAX_BLOCK);
1530 * If we weren't able to get a block size for some reason,
1531 * default to 512 bytes.
1533 block_len = pass_dev->block_len;
1538 * Figure out how many blocksize chunks will fit in the
1541 pass_numblocks = max_iosize / block_len;
1544 * And finally, multiple the number of blocks by the LBA
1545 * length to get our maximum block size;
1547 dev->blocksize = pass_numblocks * block_len;
1549 if (io_opts->blocksize != 0) {
1550 if ((io_opts->blocksize % dev->sector_size) != 0) {
1551 warnx("Blocksize %ju for %s is not a multiple of "
1552 "sector size %u", (uintmax_t)io_opts->blocksize,
1553 dev->device_name, dev->sector_size);
1556 dev->blocksize = io_opts->blocksize;
1558 dev->target_queue_depth = CAMDD_PASS_DEFAULT_DEPTH;
1559 if (io_opts->queue_depth != 0)
1560 dev->target_queue_depth = io_opts->queue_depth;
1562 if (io_opts->offset != 0) {
1563 if (io_opts->offset > (dev->max_sector * dev->sector_size)) {
1564 warnx("Offset %ju is past the end of device %s",
1565 io_opts->offset, dev->device_name);
1569 else if ((io_opts->offset % dev->sector_size) != 0) {
1570 warnx("Offset %ju for %s is not a multiple of the "
1571 "sector size %u", io_opts->offset,
1572 dev->device_name, dev->sector_size);
1575 dev->start_offset_bytes = io_opts->offset;
1579 dev->min_cmd_size = io_opts->min_cmd_size;
1581 dev->run = camdd_pass_run;
1582 dev->fetch = camdd_pass_fetch;
1592 camdd_free_dev(dev);
1598 camdd_worker(void *arg)
1600 struct camdd_dev *dev = arg;
1601 struct camdd_buf *buf;
1602 struct timespec ts, *kq_ts;
1607 pthread_mutex_lock(&dev->mutex);
1609 dev->flags |= CAMDD_DEV_FLAG_ACTIVE;
1616 * XXX KDM check the reorder queue depth?
1618 if (dev->write_dev == 0) {
1619 uint32_t our_depth, peer_depth, peer_bytes, our_bytes;
1620 uint32_t target_depth = dev->target_queue_depth;
1621 uint32_t peer_target_depth =
1622 dev->peer_dev->target_queue_depth;
1623 uint32_t peer_blocksize = dev->peer_dev->blocksize;
1625 camdd_get_depth(dev, &our_depth, &peer_depth,
1626 &our_bytes, &peer_bytes);
1629 while (((our_depth < target_depth)
1630 && (peer_depth < peer_target_depth))
1631 || ((peer_bytes + our_bytes) <
1632 (peer_blocksize * 2))) {
1634 while (((our_depth + peer_depth) <
1635 (target_depth + peer_target_depth))
1636 || ((peer_bytes + our_bytes) <
1637 (peer_blocksize * 3))) {
1639 retval = camdd_queue(dev, NULL);
1642 else if (retval != 0) {
1647 camdd_get_depth(dev, &our_depth, &peer_depth,
1648 &our_bytes, &peer_bytes);
1652 * See if we have any I/O that is ready to execute.
1654 buf = STAILQ_FIRST(&dev->run_queue);
1656 while (dev->target_queue_depth > dev->cur_active_io) {
1657 retval = dev->run(dev);
1659 dev->flags |= CAMDD_DEV_FLAG_EOF;
1662 } else if (retval != 0) {
1669 * We've reached EOF, or our partner has reached EOF.
1671 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
1672 || (dev->flags & CAMDD_DEV_FLAG_PEER_EOF)) {
1673 if (dev->write_dev != 0) {
1674 if ((STAILQ_EMPTY(&dev->work_queue))
1675 && (dev->num_run_queue == 0)
1676 && (dev->cur_active_io == 0)) {
1681 * If we're the reader, and the writer
1682 * got EOF, he is already done. If we got
1683 * the EOF, then we need to wait until
1684 * everything is flushed out for the writer.
1686 if (dev->flags & CAMDD_DEV_FLAG_PEER_EOF) {
1688 } else if ((dev->num_peer_work_queue == 0)
1689 && (dev->num_peer_done_queue == 0)
1690 && (dev->cur_active_io == 0)
1691 && (dev->num_run_queue == 0)) {
1696 * XXX KDM need to do something about the pending
1697 * queue and cleanup resources.
1701 if ((dev->write_dev == 0)
1702 && (dev->cur_active_io == 0)
1703 && (dev->peer_bytes_queued < dev->peer_dev->blocksize))
1709 * Run kevent to see if there are events to process.
1711 pthread_mutex_unlock(&dev->mutex);
1712 retval = kevent(dev->kq, NULL, 0, &ke, 1, kq_ts);
1713 pthread_mutex_lock(&dev->mutex);
1715 warn("%s: error returned from kevent",__func__);
1717 } else if (retval != 0) {
1718 switch (ke.filter) {
1720 if (dev->fetch != NULL) {
1721 retval = dev->fetch(dev);
1730 * We register for this so we don't get
1731 * an error as a result of a SIGINFO or a
1732 * SIGINT. It will actually get handled
1733 * by the signal handler. If we get a
1734 * SIGINT, bail out without printing an
1735 * error message. Any other signals
1736 * will result in the error message above.
1738 if (ke.ident == SIGINT)
1744 * Check to see if the other thread has
1745 * queued any I/O for us to do. (In this
1746 * case we're the writer.)
1748 for (buf = STAILQ_FIRST(&dev->work_queue);
1750 buf = STAILQ_FIRST(&dev->work_queue)) {
1751 STAILQ_REMOVE_HEAD(&dev->work_queue,
1753 retval = camdd_queue(dev, buf);
1755 * We keep going unless we get an
1756 * actual error. If we get EOF, we
1757 * still want to remove the buffers
1758 * from the queue and send the back
1759 * to the reader thread.
1769 * Next check to see if the other thread has
1770 * queued any completed buffers back to us.
1771 * (In this case we're the reader.)
1773 for (buf = STAILQ_FIRST(&dev->peer_done_queue);
1775 buf = STAILQ_FIRST(&dev->peer_done_queue)){
1777 &dev->peer_done_queue, work_links);
1778 dev->num_peer_done_queue--;
1779 camdd_peer_done(buf);
1783 warnx("%s: unknown kevent filter %d",
1784 __func__, ke.filter);
1792 dev->flags &= ~CAMDD_DEV_FLAG_ACTIVE;
1794 /* XXX KDM cleanup resources here? */
1796 pthread_mutex_unlock(&dev->mutex);
1799 sem_post(&camdd_sem);
1805 * Simplistic translation of CCB status to our local status.
1808 camdd_ccb_status(union ccb *ccb, int protocol)
1810 camdd_buf_status status = CAMDD_STATUS_NONE;
1811 cam_status ccb_status;
1813 ccb_status = ccb->ccb_h.status & CAM_STATUS_MASK;
1817 switch (ccb_status) {
1819 if (ccb->csio.resid == 0) {
1820 status = CAMDD_STATUS_OK;
1821 } else if (ccb->csio.dxfer_len > ccb->csio.resid) {
1822 status = CAMDD_STATUS_SHORT_IO;
1824 status = CAMDD_STATUS_EOF;
1828 case CAM_SCSI_STATUS_ERROR: {
1829 switch (ccb->csio.scsi_status) {
1830 case SCSI_STATUS_OK:
1831 case SCSI_STATUS_COND_MET:
1832 case SCSI_STATUS_INTERMED:
1833 case SCSI_STATUS_INTERMED_COND_MET:
1834 status = CAMDD_STATUS_OK;
1836 case SCSI_STATUS_CMD_TERMINATED:
1837 case SCSI_STATUS_CHECK_COND:
1838 case SCSI_STATUS_QUEUE_FULL:
1839 case SCSI_STATUS_BUSY:
1840 case SCSI_STATUS_RESERV_CONFLICT:
1842 status = CAMDD_STATUS_ERROR;
1848 status = CAMDD_STATUS_ERROR;
1853 status = CAMDD_STATUS_ERROR;
1861 * Queue a buffer to our peer's work thread for writing.
1863 * Returns 0 for success, -1 for failure, 1 if the other thread exited.
1866 camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf)
1869 STAILQ_HEAD(, camdd_buf) local_queue;
1870 struct camdd_buf *buf1, *buf2;
1871 struct camdd_buf_data *data = NULL;
1872 uint64_t peer_bytes_queued = 0;
1876 STAILQ_INIT(&local_queue);
1879 * Since we're the reader, we need to queue our I/O to the writer
1880 * in sequential order in order to make sure it gets written out
1881 * in sequential order.
1883 * Check the next expected I/O starting offset. If this doesn't
1884 * match, put it on the reorder queue.
1886 if ((buf->lba * dev->sector_size) != dev->next_completion_pos_bytes) {
1889 * If there is nothing on the queue, there is no sorting
1892 if (STAILQ_EMPTY(&dev->reorder_queue)) {
1893 STAILQ_INSERT_TAIL(&dev->reorder_queue, buf, links);
1894 dev->num_reorder_queue++;
1899 * Sort in ascending order by starting LBA. There should
1900 * be no identical LBAs.
1902 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1904 buf2 = STAILQ_NEXT(buf1, links);
1905 if (buf->lba < buf1->lba) {
1907 * If we're less than the first one, then
1908 * we insert at the head of the list
1909 * because this has to be the first element
1912 STAILQ_INSERT_HEAD(&dev->reorder_queue,
1914 dev->num_reorder_queue++;
1916 } else if (buf->lba > buf1->lba) {
1918 STAILQ_INSERT_TAIL(&dev->reorder_queue,
1920 dev->num_reorder_queue++;
1922 } else if (buf->lba < buf2->lba) {
1923 STAILQ_INSERT_AFTER(&dev->reorder_queue,
1925 dev->num_reorder_queue++;
1929 errx(1, "Found buffers with duplicate LBA %ju!",
1937 * We're the next expected I/O completion, so put ourselves
1938 * on the local queue to be sent to the writer. We use
1939 * work_links here so that we can queue this to the
1940 * peer_work_queue before taking the buffer off of the
1943 dev->next_completion_pos_bytes += buf->len;
1944 STAILQ_INSERT_TAIL(&local_queue, buf, work_links);
1947 * Go through the reorder queue looking for more sequential
1948 * I/O and add it to the local queue.
1950 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1951 buf1 = STAILQ_FIRST(&dev->reorder_queue)) {
1953 * As soon as we see an I/O that is out of sequence,
1956 if ((buf1->lba * dev->sector_size) !=
1957 dev->next_completion_pos_bytes)
1960 STAILQ_REMOVE_HEAD(&dev->reorder_queue, links);
1961 dev->num_reorder_queue--;
1962 STAILQ_INSERT_TAIL(&local_queue, buf1, work_links);
1963 dev->next_completion_pos_bytes += buf1->len;
1968 * Setup the event to let the other thread know that it has work
1971 EV_SET(&ke, (uintptr_t)&dev->peer_dev->work_queue, EVFILT_USER, 0,
1972 NOTE_TRIGGER, 0, NULL);
1975 * Put this on our shadow queue so that we know what we've queued
1976 * to the other thread.
1978 STAILQ_FOREACH_SAFE(buf1, &local_queue, work_links, buf2) {
1979 if (buf1->buf_type != CAMDD_BUF_DATA) {
1980 errx(1, "%s: should have a data buffer, not an "
1981 "indirect buffer", __func__);
1983 data = &buf1->buf_type_spec.data;
1986 * We only need to send one EOF to the writer, and don't
1987 * need to continue sending EOFs after that.
1989 if (buf1->status == CAMDD_STATUS_EOF) {
1990 if (dev->flags & CAMDD_DEV_FLAG_EOF_SENT) {
1991 STAILQ_REMOVE(&local_queue, buf1, camdd_buf,
1993 camdd_release_buf(buf1);
1997 dev->flags |= CAMDD_DEV_FLAG_EOF_SENT;
2001 STAILQ_INSERT_TAIL(&dev->peer_work_queue, buf1, links);
2002 peer_bytes_queued += (data->fill_len - data->resid);
2003 dev->peer_bytes_queued += (data->fill_len - data->resid);
2004 dev->num_peer_work_queue++;
2007 if (STAILQ_FIRST(&local_queue) == NULL)
2011 * Drop our mutex and pick up the other thread's mutex. We need to
2012 * do this to avoid deadlocks.
2014 pthread_mutex_unlock(&dev->mutex);
2015 pthread_mutex_lock(&dev->peer_dev->mutex);
2017 if (dev->peer_dev->flags & CAMDD_DEV_FLAG_ACTIVE) {
2019 * Put the buffers on the other thread's incoming work queue.
2021 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
2022 buf1 = STAILQ_FIRST(&local_queue)) {
2023 STAILQ_REMOVE_HEAD(&local_queue, work_links);
2024 STAILQ_INSERT_TAIL(&dev->peer_dev->work_queue, buf1,
2028 * Send an event to the other thread's kqueue to let it know
2029 * that there is something on the work queue.
2031 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2033 warn("%s: unable to add peer work_queue kevent",
2040 pthread_mutex_unlock(&dev->peer_dev->mutex);
2041 pthread_mutex_lock(&dev->mutex);
2044 * If the other side isn't active, run through the queue and
2045 * release all of the buffers.
2048 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
2049 buf1 = STAILQ_FIRST(&local_queue)) {
2050 STAILQ_REMOVE_HEAD(&local_queue, work_links);
2051 STAILQ_REMOVE(&dev->peer_work_queue, buf1, camdd_buf,
2053 dev->num_peer_work_queue--;
2054 camdd_release_buf(buf1);
2056 dev->peer_bytes_queued -= peer_bytes_queued;
2065 * Return a buffer to the reader thread when we have completed writing it.
2068 camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf)
2074 * Setup the event to let the other thread know that we have
2075 * completed a buffer.
2077 EV_SET(&ke, (uintptr_t)&dev->peer_dev->peer_done_queue, EVFILT_USER, 0,
2078 NOTE_TRIGGER, 0, NULL);
2081 * Drop our lock and acquire the other thread's lock before
2084 pthread_mutex_unlock(&dev->mutex);
2085 pthread_mutex_lock(&dev->peer_dev->mutex);
2088 * Put the buffer on the reader thread's peer done queue now that
2089 * we have completed it.
2091 STAILQ_INSERT_TAIL(&dev->peer_dev->peer_done_queue, peer_buf,
2093 dev->peer_dev->num_peer_done_queue++;
2096 * Send an event to the peer thread to let it know that we've added
2097 * something to its peer done queue.
2099 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2101 warn("%s: unable to add peer_done_queue kevent", __func__);
2106 * Drop the other thread's lock and reacquire ours.
2108 pthread_mutex_unlock(&dev->peer_dev->mutex);
2109 pthread_mutex_lock(&dev->mutex);
2115 * Free a buffer that was written out by the writer thread and returned to
2116 * the reader thread.
2119 camdd_peer_done(struct camdd_buf *buf)
2121 struct camdd_dev *dev;
2122 struct camdd_buf_data *data;
2125 if (buf->buf_type != CAMDD_BUF_DATA) {
2126 errx(1, "%s: should have a data buffer, not an "
2127 "indirect buffer", __func__);
2130 data = &buf->buf_type_spec.data;
2132 STAILQ_REMOVE(&dev->peer_work_queue, buf, camdd_buf, links);
2133 dev->num_peer_work_queue--;
2134 dev->peer_bytes_queued -= (data->fill_len - data->resid);
2136 if (buf->status == CAMDD_STATUS_EOF)
2137 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2139 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2143 * Assumes caller holds the lock for this device.
2146 camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
2152 * If we're the reader, we need to send the completed I/O
2153 * to the writer. If we're the writer, we need to just
2154 * free up resources, or let the reader know if we've
2155 * encountered an error.
2157 if (dev->write_dev == 0) {
2158 retval = camdd_queue_peer_buf(dev, buf);
2162 struct camdd_buf *tmp_buf, *next_buf;
2164 STAILQ_FOREACH_SAFE(tmp_buf, &buf->src_list, src_links,
2166 struct camdd_buf *src_buf;
2167 struct camdd_buf_indirect *indirect;
2169 STAILQ_REMOVE(&buf->src_list, tmp_buf,
2170 camdd_buf, src_links);
2172 tmp_buf->status = buf->status;
2174 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
2175 camdd_complete_peer_buf(dev, tmp_buf);
2179 indirect = &tmp_buf->buf_type_spec.indirect;
2180 src_buf = indirect->src_buf;
2181 src_buf->refcount--;
2183 * XXX KDM we probably need to account for
2184 * exactly how many bytes we were able to
2185 * write. Allocate the residual to the
2186 * first N buffers? Or just track the
2187 * number of bytes written? Right now the reader
2188 * doesn't do anything with a residual.
2190 src_buf->status = buf->status;
2191 if (src_buf->refcount <= 0)
2192 camdd_complete_peer_buf(dev, src_buf);
2193 STAILQ_INSERT_TAIL(&dev->free_indirect_queue,
2197 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2202 * Fetch all completed commands from the pass(4) device.
2204 * Returns the number of commands received, or -1 if any of the commands
2205 * completed with an error. Returns 0 if no commands are available.
2208 camdd_pass_fetch(struct camdd_dev *dev)
2210 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2212 int retval = 0, num_fetched = 0, error_count = 0;
2214 pthread_mutex_unlock(&dev->mutex);
2216 * XXX KDM we don't distinguish between EFAULT and ENOENT.
2218 while ((retval = ioctl(pass_dev->dev->fd, CAMIOGET, &ccb)) != -1) {
2219 struct camdd_buf *buf;
2220 struct camdd_buf_data *data;
2221 cam_status ccb_status;
2224 buf = ccb.ccb_h.ccb_buf;
2225 data = &buf->buf_type_spec.data;
2226 buf_ccb = &data->ccb;
2231 * Copy the CCB back out so we get status, sense data, etc.
2233 bcopy(&ccb, buf_ccb, sizeof(ccb));
2235 pthread_mutex_lock(&dev->mutex);
2238 * We're now done, so take this off the active queue.
2240 STAILQ_REMOVE(&dev->active_queue, buf, camdd_buf, links);
2241 dev->cur_active_io--;
2243 ccb_status = ccb.ccb_h.status & CAM_STATUS_MASK;
2244 if (ccb_status != CAM_REQ_CMP) {
2245 cam_error_print(pass_dev->dev, &ccb, CAM_ESF_ALL,
2246 CAM_EPF_ALL, stderr);
2249 switch (pass_dev->protocol) {
2251 data->resid = ccb.csio.resid;
2252 dev->bytes_transferred += (ccb.csio.dxfer_len - ccb.csio.resid);
2259 if (buf->status == CAMDD_STATUS_NONE)
2260 buf->status = camdd_ccb_status(&ccb, pass_dev->protocol);
2261 if (buf->status == CAMDD_STATUS_ERROR)
2263 else if (buf->status == CAMDD_STATUS_EOF) {
2265 * Once we queue this buffer to our partner thread,
2266 * he will know that we've hit EOF.
2268 dev->flags |= CAMDD_DEV_FLAG_EOF;
2271 camdd_complete_buf(dev, buf, &error_count);
2274 * Unlock in preparation for the ioctl call.
2276 pthread_mutex_unlock(&dev->mutex);
2279 pthread_mutex_lock(&dev->mutex);
2281 if (error_count > 0)
2284 return (num_fetched);
2288 * Returns -1 for error, 0 for success/continue, and 1 for resource
2289 * shortage/stop processing.
2292 camdd_file_run(struct camdd_dev *dev)
2294 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
2295 struct camdd_buf_data *data;
2296 struct camdd_buf *buf;
2298 int retval = 0, write_dev = dev->write_dev;
2299 int error_count = 0, no_resources = 0, double_buf_needed = 0;
2300 uint32_t num_sectors = 0, db_len = 0;
2302 buf = STAILQ_FIRST(&dev->run_queue);
2306 } else if ((dev->write_dev == 0)
2307 && (dev->flags & (CAMDD_DEV_FLAG_EOF |
2308 CAMDD_DEV_FLAG_EOF_SENT))) {
2309 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2310 dev->num_run_queue--;
2311 buf->status = CAMDD_STATUS_EOF;
2317 * If we're writing, we need to go through the source buffer list
2318 * and create an S/G list.
2320 if (write_dev != 0) {
2321 retval = camdd_buf_sg_create(buf, /*iovec*/ 1,
2322 dev->sector_size, &num_sectors, &double_buf_needed);
2329 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2330 dev->num_run_queue--;
2332 data = &buf->buf_type_spec.data;
2335 * pread(2) and pwrite(2) offsets are byte offsets.
2337 io_offset = buf->lba * dev->sector_size;
2340 * Unlock the mutex while we read or write.
2342 pthread_mutex_unlock(&dev->mutex);
2345 * Note that we don't need to double buffer if we're the reader
2346 * because in that case, we have allocated a single buffer of
2347 * sufficient size to do the read. This copy is necessary on
2348 * writes because if one of the components of the S/G list is not
2349 * a sector size multiple, the kernel will reject the write. This
2350 * is unfortunate but not surprising. So this will make sure that
2351 * we're using a single buffer that is a multiple of the sector size.
2353 if ((double_buf_needed != 0)
2354 && (data->sg_count > 1)
2355 && (write_dev != 0)) {
2356 uint32_t cur_offset;
2359 if (file_dev->tmp_buf == NULL)
2360 file_dev->tmp_buf = calloc(dev->blocksize, 1);
2361 if (file_dev->tmp_buf == NULL) {
2362 buf->status = CAMDD_STATUS_ERROR;
2364 pthread_mutex_lock(&dev->mutex);
2367 for (i = 0, cur_offset = 0; i < data->sg_count; i++) {
2368 bcopy(data->iovec[i].iov_base,
2369 &file_dev->tmp_buf[cur_offset],
2370 data->iovec[i].iov_len);
2371 cur_offset += data->iovec[i].iov_len;
2373 db_len = cur_offset;
2376 if (file_dev->file_flags & CAMDD_FF_CAN_SEEK) {
2377 if (write_dev == 0) {
2379 * XXX KDM is there any way we would need a S/G
2382 retval = pread(file_dev->fd, data->buf,
2383 buf->len, io_offset);
2385 if (double_buf_needed != 0) {
2386 retval = pwrite(file_dev->fd, file_dev->tmp_buf,
2388 } else if (data->sg_count == 0) {
2389 retval = pwrite(file_dev->fd, data->buf,
2390 data->fill_len, io_offset);
2392 retval = pwritev(file_dev->fd, data->iovec,
2393 data->sg_count, io_offset);
2397 if (write_dev == 0) {
2399 * XXX KDM is there any way we would need a S/G
2402 retval = read(file_dev->fd, data->buf, buf->len);
2404 if (double_buf_needed != 0) {
2405 retval = write(file_dev->fd, file_dev->tmp_buf,
2407 } else if (data->sg_count == 0) {
2408 retval = write(file_dev->fd, data->buf,
2411 retval = writev(file_dev->fd, data->iovec,
2417 /* We're done, re-acquire the lock */
2418 pthread_mutex_lock(&dev->mutex);
2420 if (retval >= (ssize_t)data->fill_len) {
2422 * If the bytes transferred is more than the request size,
2423 * that indicates an overrun, which should only happen at
2424 * the end of a transfer if we have to round up to a sector
2427 if (buf->status == CAMDD_STATUS_NONE)
2428 buf->status = CAMDD_STATUS_OK;
2430 dev->bytes_transferred += retval;
2431 } else if (retval == -1) {
2432 warn("Error %s %s", (write_dev) ? "writing to" :
2433 "reading from", file_dev->filename);
2435 buf->status = CAMDD_STATUS_ERROR;
2436 data->resid = data->fill_len;
2439 if (dev->debug == 0)
2442 if ((double_buf_needed != 0)
2443 && (write_dev != 0)) {
2444 fprintf(stderr, "%s: fd %d, DB buf %p, len %u lba %ju "
2445 "offset %ju\n", __func__, file_dev->fd,
2446 file_dev->tmp_buf, db_len, (uintmax_t)buf->lba,
2447 (uintmax_t)io_offset);
2448 } else if (data->sg_count == 0) {
2449 fprintf(stderr, "%s: fd %d, buf %p, len %u, lba %ju "
2450 "offset %ju\n", __func__, file_dev->fd, data->buf,
2451 data->fill_len, (uintmax_t)buf->lba,
2452 (uintmax_t)io_offset);
2456 fprintf(stderr, "%s: fd %d, len %u, lba %ju "
2457 "offset %ju\n", __func__, file_dev->fd,
2458 data->fill_len, (uintmax_t)buf->lba,
2459 (uintmax_t)io_offset);
2461 for (i = 0; i < data->sg_count; i++) {
2462 fprintf(stderr, "index %d ptr %p len %zu\n",
2463 i, data->iovec[i].iov_base,
2464 data->iovec[i].iov_len);
2467 } else if (retval == 0) {
2468 buf->status = CAMDD_STATUS_EOF;
2469 if (dev->debug != 0)
2470 printf("%s: got EOF from %s!\n", __func__,
2471 file_dev->filename);
2472 data->resid = data->fill_len;
2474 } else if (retval < (ssize_t)data->fill_len) {
2475 if (buf->status == CAMDD_STATUS_NONE)
2476 buf->status = CAMDD_STATUS_SHORT_IO;
2477 data->resid = data->fill_len - retval;
2478 dev->bytes_transferred += retval;
2483 if (buf->status == CAMDD_STATUS_EOF) {
2484 struct camdd_buf *buf2;
2485 dev->flags |= CAMDD_DEV_FLAG_EOF;
2486 STAILQ_FOREACH(buf2, &dev->run_queue, links)
2487 buf2->status = CAMDD_STATUS_EOF;
2490 camdd_complete_buf(dev, buf, &error_count);
2493 if (error_count != 0)
2495 else if (no_resources != 0)
2502 * Execute one command from the run queue. Returns 0 for success, 1 for
2503 * stop processing, and -1 for error.
2506 camdd_pass_run(struct camdd_dev *dev)
2508 struct camdd_buf *buf = NULL;
2509 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2510 struct camdd_buf_data *data;
2511 uint32_t num_blocks, sectors_used = 0;
2513 int retval = 0, is_write = dev->write_dev;
2514 int double_buf_needed = 0;
2516 buf = STAILQ_FIRST(&dev->run_queue);
2523 * If we're writing, we need to go through the source buffer list
2524 * and create an S/G list.
2526 if (is_write != 0) {
2527 retval = camdd_buf_sg_create(buf, /*iovec*/ 0,dev->sector_size,
2528 §ors_used, &double_buf_needed);
2535 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2536 dev->num_run_queue--;
2538 data = &buf->buf_type_spec.data;
2541 * In almost every case the number of blocks should be the device
2542 * block size. The exception may be at the end of an I/O stream
2543 * for a partial block or at the end of a device.
2546 num_blocks = sectors_used;
2548 num_blocks = data->fill_len / pass_dev->block_len;
2552 switch (pass_dev->protocol) {
2554 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
2556 scsi_read_write(&ccb->csio,
2557 /*retries*/ dev->retry_count,
2559 /*tag_action*/ MSG_SIMPLE_Q_TAG,
2560 /*readop*/ (dev->write_dev == 0) ? SCSI_RW_READ :
2563 /*minimum_cmd_size*/ dev->min_cmd_size,
2565 /*block_count*/ num_blocks,
2566 /*data_ptr*/ (data->sg_count != 0) ?
2567 (uint8_t *)data->segs : data->buf,
2568 /*dxfer_len*/ (num_blocks * pass_dev->block_len),
2569 /*sense_len*/ SSD_FULL_SIZE,
2570 /*timeout*/ dev->io_timeout);
2572 if (data->sg_count != 0) {
2573 ccb->csio.sglist_cnt = data->sg_count;
2581 /* Disable freezing the device queue */
2582 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
2584 if (dev->retry_count != 0)
2585 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
2587 if (data->sg_count != 0) {
2588 ccb->ccb_h.flags |= CAM_DATA_SG;
2592 * Store a pointer to the buffer in the CCB. The kernel will
2593 * restore this when we get it back, and we'll use it to identify
2594 * the buffer this CCB came from.
2596 ccb->ccb_h.ccb_buf = buf;
2599 * Unlock our mutex in preparation for issuing the ioctl.
2601 pthread_mutex_unlock(&dev->mutex);
2603 * Queue the CCB to the pass(4) driver.
2605 if (ioctl(pass_dev->dev->fd, CAMIOQUEUE, ccb) == -1) {
2606 pthread_mutex_lock(&dev->mutex);
2608 warn("%s: error sending CAMIOQUEUE ioctl to %s%u", __func__,
2609 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
2610 warn("%s: CCB address is %p", __func__, ccb);
2613 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2615 pthread_mutex_lock(&dev->mutex);
2617 dev->cur_active_io++;
2618 STAILQ_INSERT_TAIL(&dev->active_queue, buf, links);
2626 camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len)
2628 struct camdd_dev_pass *pass_dev;
2629 uint32_t num_blocks;
2632 pass_dev = &dev->dev_spec.pass;
2634 *lba = dev->next_io_pos_bytes / dev->sector_size;
2635 *len = dev->blocksize;
2636 num_blocks = *len / dev->sector_size;
2639 * If max_sector is 0, then we have no set limit. This can happen
2640 * if we're writing to a file in a filesystem, or reading from
2641 * something like /dev/zero.
2643 if ((dev->max_sector != 0)
2644 || (dev->sector_io_limit != 0)) {
2645 uint64_t max_sector;
2647 if ((dev->max_sector != 0)
2648 && (dev->sector_io_limit != 0))
2649 max_sector = min(dev->sector_io_limit, dev->max_sector);
2650 else if (dev->max_sector != 0)
2651 max_sector = dev->max_sector;
2653 max_sector = dev->sector_io_limit;
2657 * Check to see whether we're starting off past the end of
2658 * the device. If so, we need to just send an EOF
2659 * notification to the writer.
2661 if (*lba > max_sector) {
2664 } else if (((*lba + num_blocks) > max_sector + 1)
2665 || ((*lba + num_blocks) < *lba)) {
2667 * If we get here (but pass the first check), we
2668 * can trim the request length down to go to the
2669 * end of the device.
2671 num_blocks = (max_sector + 1) - *lba;
2672 *len = num_blocks * dev->sector_size;
2677 dev->next_io_pos_bytes += *len;
2683 * Returns 0 for success, 1 for EOF detected, and -1 for failure.
2686 camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf)
2688 struct camdd_buf *buf = NULL;
2689 struct camdd_buf_data *data;
2690 struct camdd_dev_pass *pass_dev;
2692 struct camdd_buf_data *rb_data;
2693 int is_write = dev->write_dev;
2694 int eof_flush_needed = 0;
2698 pass_dev = &dev->dev_spec.pass;
2701 * If we've gotten EOF or our partner has, we should not continue
2702 * queueing I/O. If we're a writer, though, we should continue
2703 * to write any buffers that don't have EOF status.
2705 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
2706 || ((dev->flags & CAMDD_DEV_FLAG_PEER_EOF)
2707 && (is_write == 0))) {
2709 * Tell the worker thread that we have seen EOF.
2714 * If we're the writer, send the buffer back with EOF status.
2717 read_buf->status = CAMDD_STATUS_EOF;
2719 error = camdd_complete_peer_buf(dev, read_buf);
2724 if (is_write == 0) {
2725 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2730 data = &buf->buf_type_spec.data;
2732 retval = camdd_get_next_lba_len(dev, &buf->lba, &buf->len);
2734 buf->status = CAMDD_STATUS_EOF;
2737 && ((dev->flags & (CAMDD_DEV_FLAG_EOF_SENT |
2738 CAMDD_DEV_FLAG_EOF_QUEUED)) != 0)) {
2739 camdd_release_buf(buf);
2742 dev->flags |= CAMDD_DEV_FLAG_EOF_QUEUED;
2745 data->fill_len = buf->len;
2746 data->src_start_offset = buf->lba * dev->sector_size;
2749 * Put this on the run queue.
2751 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2752 dev->num_run_queue++;
2759 * Check for new EOF status from the reader.
2761 if ((read_buf->status == CAMDD_STATUS_EOF)
2762 || (read_buf->status == CAMDD_STATUS_ERROR)) {
2763 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2764 if ((STAILQ_FIRST(&dev->pending_queue) == NULL)
2765 && (read_buf->len == 0)) {
2766 camdd_complete_peer_buf(dev, read_buf);
2770 eof_flush_needed = 1;
2774 * See if we have a buffer we're composing with pieces from our
2777 buf = STAILQ_FIRST(&dev->pending_queue);
2782 retval = camdd_get_next_lba_len(dev, &lba, &len);
2784 read_buf->status = CAMDD_STATUS_EOF;
2787 dev->flags |= CAMDD_DEV_FLAG_EOF;
2788 error = camdd_complete_peer_buf(dev, read_buf);
2794 * If we don't have a pending buffer, we need to grab a new
2795 * one from the free list or allocate another one.
2797 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2806 STAILQ_INSERT_TAIL(&dev->pending_queue, buf, links);
2807 dev->num_pending_queue++;
2810 data = &buf->buf_type_spec.data;
2812 rb_data = &read_buf->buf_type_spec.data;
2814 if ((rb_data->src_start_offset != dev->next_peer_pos_bytes)
2815 && (dev->debug != 0)) {
2816 printf("%s: WARNING: reader offset %#jx != expected offset "
2817 "%#jx\n", __func__, (uintmax_t)rb_data->src_start_offset,
2818 (uintmax_t)dev->next_peer_pos_bytes);
2820 dev->next_peer_pos_bytes = rb_data->src_start_offset +
2821 (rb_data->fill_len - rb_data->resid);
2823 new_len = (rb_data->fill_len - rb_data->resid) + data->fill_len;
2824 if (new_len < buf->len) {
2826 * There are three cases here:
2827 * 1. We need more data to fill up a block, so we put
2828 * this I/O on the queue and wait for more I/O.
2829 * 2. We have a pending buffer in the queue that is
2830 * smaller than our blocksize, but we got an EOF. So we
2831 * need to go ahead and flush the write out.
2832 * 3. We got an error.
2836 * Increment our fill length.
2838 data->fill_len += (rb_data->fill_len - rb_data->resid);
2841 * Add the new read buffer to the list for writing.
2843 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2845 /* Increment the count */
2848 if (eof_flush_needed == 0) {
2850 * We need to exit, because we don't have enough
2856 * Take the buffer off of the pending queue.
2858 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2860 dev->num_pending_queue--;
2863 * If we need an EOF flush, but there is no data
2864 * to flush, go ahead and return this buffer.
2866 if (data->fill_len == 0) {
2867 camdd_complete_buf(dev, buf, /*error_count*/0);
2873 * Put this on the next queue for execution.
2875 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2876 dev->num_run_queue++;
2878 } else if (new_len == buf->len) {
2880 * We have enough data to completey fill one block,
2881 * so we're ready to issue the I/O.
2885 * Take the buffer off of the pending queue.
2887 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf, links);
2888 dev->num_pending_queue--;
2891 * Add the new read buffer to the list for writing.
2893 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2895 /* Increment the count */
2899 * Increment our fill length.
2901 data->fill_len += (rb_data->fill_len - rb_data->resid);
2904 * Put this on the next queue for execution.
2906 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2907 dev->num_run_queue++;
2909 struct camdd_buf *idb;
2910 struct camdd_buf_indirect *indirect;
2911 uint32_t len_to_go, cur_offset;
2914 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2919 indirect = &idb->buf_type_spec.indirect;
2920 indirect->src_buf = read_buf;
2921 read_buf->refcount++;
2922 indirect->offset = 0;
2923 indirect->start_ptr = rb_data->buf;
2925 * We've already established that there is more
2926 * data in read_buf than we have room for in our
2927 * current write request. So this particular chunk
2928 * of the request should just be the remainder
2929 * needed to fill up a block.
2931 indirect->len = buf->len - (data->fill_len - data->resid);
2933 camdd_buf_add_child(buf, idb);
2936 * This buffer is ready to execute, so we can take
2937 * it off the pending queue and put it on the run
2940 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2942 dev->num_pending_queue--;
2943 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2944 dev->num_run_queue++;
2946 cur_offset = indirect->offset + indirect->len;
2949 * The resulting I/O would be too large to fit in
2950 * one block. We need to split this I/O into
2951 * multiple pieces. Allocate as many buffers as needed.
2953 for (len_to_go = rb_data->fill_len - rb_data->resid -
2954 indirect->len; len_to_go > 0;) {
2955 struct camdd_buf *new_buf;
2956 struct camdd_buf_data *new_data;
2960 retval = camdd_get_next_lba_len(dev, &lba, &len);
2964 * The device has already been marked
2965 * as EOF, and there is no space left.
2970 new_buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2971 if (new_buf == NULL) {
2979 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2985 indirect = &idb->buf_type_spec.indirect;
2987 indirect->src_buf = read_buf;
2988 read_buf->refcount++;
2989 indirect->offset = cur_offset;
2990 indirect->start_ptr = rb_data->buf + cur_offset;
2991 indirect->len = min(len_to_go, new_buf->len);
2993 if (((indirect->len % dev->sector_size) != 0)
2994 || ((indirect->offset % dev->sector_size) != 0)) {
2995 warnx("offset %ju len %ju not aligned with "
2996 "sector size %u", indirect->offset,
2997 (uintmax_t)indirect->len, dev->sector_size);
3000 cur_offset += indirect->len;
3001 len_to_go -= indirect->len;
3003 camdd_buf_add_child(new_buf, idb);
3005 new_data = &new_buf->buf_type_spec.data;
3007 if ((new_data->fill_len == new_buf->len)
3008 || (eof_flush_needed != 0)) {
3009 STAILQ_INSERT_TAIL(&dev->run_queue,
3011 dev->num_run_queue++;
3012 } else if (new_data->fill_len < buf->len) {
3013 STAILQ_INSERT_TAIL(&dev->pending_queue,
3015 dev->num_pending_queue++;
3017 warnx("%s: too much data in new "
3018 "buffer!", __func__);
3030 camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
3031 uint32_t *peer_depth, uint32_t *our_bytes, uint32_t *peer_bytes)
3033 *our_depth = dev->cur_active_io + dev->num_run_queue;
3034 if (dev->num_peer_work_queue >
3035 dev->num_peer_done_queue)
3036 *peer_depth = dev->num_peer_work_queue -
3037 dev->num_peer_done_queue;
3040 *our_bytes = *our_depth * dev->blocksize;
3041 *peer_bytes = dev->peer_bytes_queued;
3045 camdd_sig_handler(int sig)
3054 sem_post(&camdd_sem);
3058 camdd_print_status(struct camdd_dev *camdd_dev, struct camdd_dev *other_dev,
3059 struct timespec *start_time)
3061 struct timespec done_time;
3063 long double mb_sec, total_sec;
3066 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &done_time);
3068 warn("Unable to get done time");
3072 timespecsub(&done_time, start_time);
3074 total_ns = done_time.tv_nsec + (done_time.tv_sec * 1000000000);
3075 total_sec = total_ns;
3076 total_sec /= 1000000000;
3078 fprintf(stderr, "%ju bytes %s %s\n%ju bytes %s %s\n"
3079 "%.4Lf seconds elapsed\n",
3080 (uintmax_t)camdd_dev->bytes_transferred,
3081 (camdd_dev->write_dev == 0) ? "read from" : "written to",
3082 camdd_dev->device_name,
3083 (uintmax_t)other_dev->bytes_transferred,
3084 (other_dev->write_dev == 0) ? "read from" : "written to",
3085 other_dev->device_name, total_sec);
3087 mb_sec = min(other_dev->bytes_transferred,camdd_dev->bytes_transferred);
3088 mb_sec /= 1024 * 1024;
3089 mb_sec *= 1000000000;
3091 fprintf(stderr, "%.2Lf MB/sec\n", mb_sec);
3095 camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts, uint64_t max_io,
3096 int retry_count, int timeout)
3098 struct cam_device *new_cam_dev = NULL;
3099 struct camdd_dev *devs[2];
3100 struct timespec start_time;
3101 pthread_t threads[2];
3106 if (num_io_opts != 2) {
3107 warnx("Must have one input and one output path");
3112 bzero(devs, sizeof(devs));
3114 for (i = 0; i < num_io_opts; i++) {
3115 switch (io_opts[i].dev_type) {
3116 case CAMDD_DEV_PASS: {
3117 if (isdigit(io_opts[i].dev_name[0])) {
3118 camdd_argmask new_arglist = CAMDD_ARG_NONE;
3119 int bus = 0, target = 0, lun = 0;
3122 /* device specified as bus:target[:lun] */
3123 rv = parse_btl(io_opts[i].dev_name, &bus,
3124 &target, &lun, &new_arglist);
3126 warnx("numeric device specification "
3127 "must be either bus:target, or "
3132 /* default to 0 if lun was not specified */
3133 if ((new_arglist & CAMDD_ARG_LUN) == 0) {
3135 new_arglist |= CAMDD_ARG_LUN;
3137 new_cam_dev = cam_open_btl(bus, target, lun,
3142 if (cam_get_device(io_opts[i].dev_name, name,
3143 sizeof name, &unit) == -1) {
3144 warnx("%s", cam_errbuf);
3148 new_cam_dev = cam_open_spec_device(name, unit,
3152 if (new_cam_dev == NULL) {
3153 warnx("%s", cam_errbuf);
3158 devs[i] = camdd_probe_pass(new_cam_dev,
3159 /*io_opts*/ &io_opts[i],
3160 CAMDD_ARG_ERR_RECOVER,
3161 /*probe_retry_count*/ 3,
3162 /*probe_timeout*/ 5000,
3163 /*io_retry_count*/ retry_count,
3164 /*io_timeout*/ timeout);
3165 if (devs[i] == NULL) {
3166 warn("Unable to probe device %s%u",
3167 new_cam_dev->device_name,
3168 new_cam_dev->dev_unit_num);
3174 case CAMDD_DEV_FILE: {
3177 if (io_opts[i].dev_name[0] == '-') {
3178 if (io_opts[i].write_dev != 0)
3183 if (io_opts[i].write_dev != 0) {
3184 fd = open(io_opts[i].dev_name,
3185 O_RDWR | O_CREAT, S_IWUSR |S_IRUSR);
3187 fd = open(io_opts[i].dev_name,
3192 warn("error opening file %s",
3193 io_opts[i].dev_name);
3198 devs[i] = camdd_probe_file(fd, &io_opts[i],
3199 retry_count, timeout);
3200 if (devs[i] == NULL) {
3208 warnx("Unknown device type %d (%s)",
3209 io_opts[i].dev_type, io_opts[i].dev_name);
3212 break; /*NOTREACHED */
3215 devs[i]->write_dev = io_opts[i].write_dev;
3217 devs[i]->start_offset_bytes = io_opts[i].offset;
3220 devs[i]->sector_io_limit =
3221 (devs[i]->start_offset_bytes /
3222 devs[i]->sector_size) +
3223 (max_io / devs[i]->sector_size) - 1;
3226 devs[i]->next_io_pos_bytes = devs[i]->start_offset_bytes;
3227 devs[i]->next_completion_pos_bytes =devs[i]->start_offset_bytes;
3230 devs[0]->peer_dev = devs[1];
3231 devs[1]->peer_dev = devs[0];
3232 devs[0]->next_peer_pos_bytes = devs[0]->peer_dev->next_io_pos_bytes;
3233 devs[1]->next_peer_pos_bytes = devs[1]->peer_dev->next_io_pos_bytes;
3235 sem_init(&camdd_sem, /*pshared*/ 0, 0);
3237 signal(SIGINFO, camdd_sig_handler);
3238 signal(SIGINT, camdd_sig_handler);
3240 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &start_time);
3242 warn("Unable to get start time");
3246 for (i = 0; i < num_io_opts; i++) {
3247 error = pthread_create(&threads[i], NULL, camdd_worker,
3250 warnc(error, "pthread_create() failed");
3256 if ((sem_wait(&camdd_sem) == -1)
3257 || (need_exit != 0)) {
3260 for (i = 0; i < num_io_opts; i++) {
3261 EV_SET(&ke, (uintptr_t)&devs[i]->work_queue,
3262 EVFILT_USER, 0, NOTE_TRIGGER, 0, NULL);
3264 devs[i]->flags |= CAMDD_DEV_FLAG_EOF;
3266 error = kevent(devs[i]->kq, &ke, 1, NULL, 0,
3269 warn("%s: unable to wake up thread",
3274 } else if (need_status != 0) {
3275 camdd_print_status(devs[0], devs[1], &start_time);
3279 for (i = 0; i < num_io_opts; i++) {
3280 pthread_join(threads[i], NULL);
3283 camdd_print_status(devs[0], devs[1], &start_time);
3287 for (i = 0; i < num_io_opts; i++)
3288 camdd_free_dev(devs[i]);
3290 return (error + error_exit);
3297 "usage: camdd <-i|-o pass=pass0,bs=1M,offset=1M,depth=4>\n"
3298 " <-i|-o file=/tmp/file,bs=512K,offset=1M>\n"
3299 " <-i|-o file=/dev/da0,bs=512K,offset=1M>\n"
3300 " <-i|-o file=/dev/nsa0,bs=512K>\n"
3301 " [-C retry_count][-E][-m max_io_amt][-t timeout_secs][-v][-h]\n"
3302 "Option description\n"
3303 "-i <arg=val> Specify input device/file and parameters\n"
3304 "-o <arg=val> Specify output device/file and parameters\n"
3305 "Input and Output parameters\n"
3306 "pass=name Specify a pass(4) device like pass0 or /dev/pass0\n"
3307 "file=name Specify a file or device, /tmp/foo, /dev/da0, /dev/null\n"
3308 " or - for stdin/stdout\n"
3309 "bs=blocksize Specify blocksize in bytes, or using K, M, G, etc. suffix\n"
3310 "offset=len Specify starting offset in bytes or using K, M, G suffix\n"
3311 " NOTE: offset cannot be specified on tapes, pipes, stdin/out\n"
3312 "depth=N Specify a numeric queue depth. This only applies to pass(4)\n"
3313 "mcs=N Specify a minimum cmd size for pass(4) read/write commands\n"
3314 "Optional arguments\n"
3315 "-C retry_cnt Specify a retry count for pass(4) devices\n"
3316 "-E Enable CAM error recovery for pass(4) devices\n"
3317 "-m max_io Specify the maximum amount to be transferred in bytes or\n"
3318 " using K, G, M, etc. suffixes\n"
3319 "-t timeout Specify the I/O timeout to use with pass(4) devices\n"
3320 "-v Enable verbose error recovery\n"
3321 "-h Print this message\n");
3326 camdd_parse_io_opts(char *args, int is_write, struct camdd_io_opts *io_opts)
3328 char *tmpstr, *tmpstr2;
3329 char *orig_tmpstr = NULL;
3332 io_opts->write_dev = is_write;
3334 tmpstr = strdup(args);
3335 if (tmpstr == NULL) {
3336 warn("strdup failed");
3340 orig_tmpstr = tmpstr;
3341 while ((tmpstr2 = strsep(&tmpstr, ",")) != NULL) {
3345 * If the user creates an empty parameter by putting in two
3346 * commas, skip over it and look for the next field.
3348 if (*tmpstr2 == '\0')
3351 name = strsep(&tmpstr2, "=");
3352 if (*name == '\0') {
3353 warnx("Got empty I/O parameter name");
3357 value = strsep(&tmpstr2, "=");
3359 || (*value == '\0')) {
3360 warnx("Empty I/O parameter value for %s", name);
3364 if (strncasecmp(name, "file", 4) == 0) {
3365 io_opts->dev_type = CAMDD_DEV_FILE;
3366 io_opts->dev_name = strdup(value);
3367 if (io_opts->dev_name == NULL) {
3368 warn("Error allocating memory");
3372 } else if (strncasecmp(name, "pass", 4) == 0) {
3373 io_opts->dev_type = CAMDD_DEV_PASS;
3374 io_opts->dev_name = strdup(value);
3375 if (io_opts->dev_name == NULL) {
3376 warn("Error allocating memory");
3380 } else if ((strncasecmp(name, "bs", 2) == 0)
3381 || (strncasecmp(name, "blocksize", 9) == 0)) {
3382 retval = expand_number(value, &io_opts->blocksize);
3384 warn("expand_number(3) failed on %s=%s", name,
3389 } else if (strncasecmp(name, "depth", 5) == 0) {
3392 io_opts->queue_depth = strtoull(value, &endptr, 0);
3393 if (*endptr != '\0') {
3394 warnx("invalid queue depth %s", value);
3398 } else if (strncasecmp(name, "mcs", 3) == 0) {
3401 io_opts->min_cmd_size = strtol(value, &endptr, 0);
3402 if ((*endptr != '\0')
3403 || ((io_opts->min_cmd_size > 16)
3404 || (io_opts->min_cmd_size < 0))) {
3405 warnx("invalid minimum cmd size %s", value);
3409 } else if (strncasecmp(name, "offset", 6) == 0) {
3410 retval = expand_number(value, &io_opts->offset);
3412 warn("expand_number(3) failed on %s=%s", name,
3417 } else if (strncasecmp(name, "debug", 5) == 0) {
3420 io_opts->debug = strtoull(value, &endptr, 0);
3421 if (*endptr != '\0') {
3422 warnx("invalid debug level %s", value);
3427 warnx("Unrecognized parameter %s=%s", name, value);
3437 main(int argc, char **argv)
3440 camdd_argmask arglist = CAMDD_ARG_NONE;
3441 int timeout = 0, retry_count = 1;
3443 uint64_t max_io = 0;
3444 struct camdd_io_opts *opt_list = NULL;
3451 opt_list = calloc(2, sizeof(struct camdd_io_opts));
3452 if (opt_list == NULL) {
3453 warn("Unable to allocate option list");
3458 while ((c = getopt(argc, argv, "C:Ehi:m:o:t:v")) != -1){
3461 retry_count = strtol(optarg, NULL, 0);
3462 if (retry_count < 0)
3463 errx(1, "retry count %d is < 0",
3465 arglist |= CAMDD_ARG_RETRIES;
3468 arglist |= CAMDD_ARG_ERR_RECOVER;
3473 && (opt_list[0].dev_type != CAMDD_DEV_NONE))
3475 && (opt_list[1].dev_type != CAMDD_DEV_NONE))) {
3476 errx(1, "Only one input and output path "
3479 error = camdd_parse_io_opts(optarg, (c == 'o') ? 1 : 0,
3480 (c == 'o') ? &opt_list[1] : &opt_list[0]);
3485 error = expand_number(optarg, &max_io);
3487 warn("invalid maximum I/O amount %s", optarg);
3493 timeout = strtol(optarg, NULL, 0);
3495 errx(1, "invalid timeout %d", timeout);
3496 /* Convert the timeout from seconds to ms */
3498 arglist |= CAMDD_ARG_TIMEOUT;
3501 arglist |= CAMDD_ARG_VERBOSE;
3507 break; /*NOTREACHED*/
3511 if ((opt_list[0].dev_type == CAMDD_DEV_NONE)
3512 || (opt_list[1].dev_type == CAMDD_DEV_NONE))
3513 errx(1, "Must specify both -i and -o");
3516 * Set the timeout if the user hasn't specified one.
3519 timeout = CAMDD_PASS_RW_TIMEOUT;
3521 error = camdd_rw(opt_list, 2, max_io, retry_count, timeout);