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 /* Generically useful offsets into the peripheral private area */
434 #define ppriv_ptr0 periph_priv.entries[0].ptr
435 #define ppriv_ptr1 periph_priv.entries[1].ptr
436 #define ppriv_field0 periph_priv.entries[0].field
437 #define ppriv_field1 periph_priv.entries[1].field
439 #define ccb_buf ppriv_ptr0
441 #define CAMDD_FILE_DEFAULT_BLOCK 524288
442 #define CAMDD_FILE_DEFAULT_DEPTH 1
443 #define CAMDD_PASS_MAX_BLOCK 1048576
444 #define CAMDD_PASS_DEFAULT_DEPTH 6
445 #define CAMDD_PASS_RW_TIMEOUT 60 * 1000
447 static int parse_btl(char *tstr, int *bus, int *target, int *lun,
448 camdd_argmask *arglst);
449 void camdd_free_dev(struct camdd_dev *dev);
450 struct camdd_dev *camdd_alloc_dev(camdd_dev_type dev_type,
451 struct kevent *new_ke, int num_ke,
452 int retry_count, int timeout);
453 static struct camdd_buf *camdd_alloc_buf(struct camdd_dev *dev,
454 camdd_buf_type buf_type);
455 void camdd_release_buf(struct camdd_buf *buf);
456 struct camdd_buf *camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type);
457 int camdd_buf_sg_create(struct camdd_buf *buf, int iovec,
458 uint32_t sector_size, uint32_t *num_sectors_used,
459 int *double_buf_needed);
460 uint32_t camdd_buf_get_len(struct camdd_buf *buf);
461 void camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf);
462 int camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
463 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran);
464 int camdd_probe_pass_scsi(struct cam_device *cam_dev, union ccb *ccb,
465 camdd_argmask arglist, int probe_retry_count,
466 int probe_timeout, uint64_t *maxsector, uint32_t *block_len);
467 struct camdd_dev *camdd_probe_file(int fd, struct camdd_io_opts *io_opts,
468 int retry_count, int timeout);
469 struct camdd_dev *camdd_probe_pass(struct cam_device *cam_dev,
470 struct camdd_io_opts *io_opts,
471 camdd_argmask arglist, int probe_retry_count,
472 int probe_timeout, int io_retry_count,
474 void *camdd_file_worker(void *arg);
475 camdd_buf_status camdd_ccb_status(union ccb *ccb, int protocol);
476 int camdd_get_cgd(struct cam_device *device, struct ccb_getdev *cgd);
477 int camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf);
478 int camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf);
479 void camdd_peer_done(struct camdd_buf *buf);
480 void camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
482 int camdd_pass_fetch(struct camdd_dev *dev);
483 int camdd_file_run(struct camdd_dev *dev);
484 int camdd_pass_run(struct camdd_dev *dev);
485 int camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len);
486 int camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf);
487 void camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
488 uint32_t *peer_depth, uint32_t *our_bytes,
489 uint32_t *peer_bytes);
490 void *camdd_worker(void *arg);
491 void camdd_sig_handler(int sig);
492 void camdd_print_status(struct camdd_dev *camdd_dev,
493 struct camdd_dev *other_dev,
494 struct timespec *start_time);
495 int camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts,
496 uint64_t max_io, int retry_count, int timeout);
497 int camdd_parse_io_opts(char *args, int is_write,
498 struct camdd_io_opts *io_opts);
502 * Parse out a bus, or a bus, target and lun in the following
508 * Returns the number of parsed components, or 0.
511 parse_btl(char *tstr, int *bus, int *target, int *lun, camdd_argmask *arglst)
516 while (isspace(*tstr) && (*tstr != '\0'))
519 tmpstr = (char *)strtok(tstr, ":");
520 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
521 *bus = strtol(tmpstr, NULL, 0);
522 *arglst |= CAMDD_ARG_BUS;
524 tmpstr = (char *)strtok(NULL, ":");
525 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
526 *target = strtol(tmpstr, NULL, 0);
527 *arglst |= CAMDD_ARG_TARGET;
529 tmpstr = (char *)strtok(NULL, ":");
530 if ((tmpstr != NULL) && (*tmpstr != '\0')) {
531 *lun = strtol(tmpstr, NULL, 0);
532 *arglst |= CAMDD_ARG_LUN;
542 * XXX KDM clean up and free all of the buffers on the queue!
545 camdd_free_dev(struct camdd_dev *dev)
550 switch (dev->dev_type) {
551 case CAMDD_DEV_FILE: {
552 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
554 if (file_dev->fd != -1)
556 free(file_dev->tmp_buf);
559 case CAMDD_DEV_PASS: {
560 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
562 if (pass_dev->dev != NULL)
563 cam_close_device(pass_dev->dev);
574 camdd_alloc_dev(camdd_dev_type dev_type, struct kevent *new_ke, int num_ke,
575 int retry_count, int timeout)
577 struct camdd_dev *dev = NULL;
582 dev = calloc(1, sizeof(*dev));
584 warn("%s: unable to malloc %zu bytes", __func__, sizeof(*dev));
588 dev->dev_type = dev_type;
589 dev->io_timeout = timeout;
590 dev->retry_count = retry_count;
591 STAILQ_INIT(&dev->free_queue);
592 STAILQ_INIT(&dev->free_indirect_queue);
593 STAILQ_INIT(&dev->active_queue);
594 STAILQ_INIT(&dev->pending_queue);
595 STAILQ_INIT(&dev->run_queue);
596 STAILQ_INIT(&dev->reorder_queue);
597 STAILQ_INIT(&dev->work_queue);
598 STAILQ_INIT(&dev->peer_done_queue);
599 STAILQ_INIT(&dev->peer_work_queue);
600 retval = pthread_mutex_init(&dev->mutex, NULL);
602 warnc(retval, "%s: failed to initialize mutex", __func__);
606 retval = pthread_cond_init(&dev->cond, NULL);
608 warnc(retval, "%s: failed to initialize condition variable",
615 warn("%s: Unable to create kqueue", __func__);
619 ke_size = sizeof(struct kevent) * (num_ke + 4);
620 ke = calloc(1, ke_size);
622 warn("%s: unable to malloc %zu bytes", __func__, ke_size);
626 bcopy(new_ke, ke, num_ke * sizeof(struct kevent));
628 EV_SET(&ke[num_ke++], (uintptr_t)&dev->work_queue, EVFILT_USER,
629 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
630 EV_SET(&ke[num_ke++], (uintptr_t)&dev->peer_done_queue, EVFILT_USER,
631 EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
632 EV_SET(&ke[num_ke++], SIGINFO, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
633 EV_SET(&ke[num_ke++], SIGINT, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
635 retval = kevent(dev->kq, ke, num_ke, NULL, 0, NULL);
637 warn("%s: Unable to register kevents", __func__);
650 static struct camdd_buf *
651 camdd_alloc_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
653 struct camdd_buf *buf = NULL;
654 uint8_t *data_ptr = NULL;
657 * We only need to allocate data space for data buffers.
661 data_ptr = malloc(dev->blocksize);
662 if (data_ptr == NULL) {
663 warn("unable to allocate %u bytes", dev->blocksize);
671 buf = calloc(1, sizeof(*buf));
673 warn("unable to allocate %zu bytes", sizeof(*buf));
677 buf->buf_type = buf_type;
680 case CAMDD_BUF_DATA: {
681 struct camdd_buf_data *data;
683 data = &buf->buf_type_spec.data;
685 data->alloc_len = dev->blocksize;
686 data->buf = data_ptr;
689 case CAMDD_BUF_INDIRECT:
694 STAILQ_INIT(&buf->src_list);
705 camdd_release_buf(struct camdd_buf *buf)
707 struct camdd_dev *dev;
711 switch (buf->buf_type) {
712 case CAMDD_BUF_DATA: {
713 struct camdd_buf_data *data;
715 data = &buf->buf_type_spec.data;
717 if (data->segs != NULL) {
718 if (data->extra_buf != 0) {
722 data->segs[data->sg_count - 1].ds_addr;
729 } else if (data->iovec != NULL) {
730 if (data->extra_buf != 0) {
731 free(data->iovec[data->sg_count - 1].iov_base);
738 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
741 case CAMDD_BUF_INDIRECT:
742 STAILQ_INSERT_TAIL(&dev->free_indirect_queue, buf, links);
745 err(1, "%s: Invalid buffer type %d for released buffer",
746 __func__, buf->buf_type);
752 camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
754 struct camdd_buf *buf = NULL;
758 buf = STAILQ_FIRST(&dev->free_queue);
760 struct camdd_buf_data *data;
764 STAILQ_REMOVE_HEAD(&dev->free_queue, links);
765 data = &buf->buf_type_spec.data;
766 data_ptr = data->buf;
767 alloc_len = data->alloc_len;
768 bzero(buf, sizeof(*buf));
769 data->buf = data_ptr;
770 data->alloc_len = alloc_len;
773 case CAMDD_BUF_INDIRECT:
774 buf = STAILQ_FIRST(&dev->free_indirect_queue);
776 STAILQ_REMOVE_HEAD(&dev->free_indirect_queue, links);
778 bzero(buf, sizeof(*buf));
782 warnx("Unknown buffer type %d requested", buf_type);
788 return (camdd_alloc_buf(dev, buf_type));
790 STAILQ_INIT(&buf->src_list);
792 buf->buf_type = buf_type;
799 camdd_buf_sg_create(struct camdd_buf *buf, int iovec, uint32_t sector_size,
800 uint32_t *num_sectors_used, int *double_buf_needed)
802 struct camdd_buf *tmp_buf;
803 struct camdd_buf_data *data;
804 uint8_t *extra_buf = NULL;
805 size_t extra_buf_len = 0;
806 int extra_buf_attached = 0;
809 data = &buf->buf_type_spec.data;
811 data->sg_count = buf->src_count;
813 * Compose a scatter/gather list from all of the buffers in the list.
814 * If the length of the buffer isn't a multiple of the sector size,
815 * we'll have to add an extra buffer. This should only happen
816 * at the end of a transfer.
818 if ((data->fill_len % sector_size) != 0) {
819 extra_buf_len = sector_size - (data->fill_len % sector_size);
820 extra_buf = calloc(extra_buf_len, 1);
821 if (extra_buf == NULL) {
822 warn("%s: unable to allocate %zu bytes for extra "
823 "buffer space", __func__, extra_buf_len);
831 data->segs = calloc(data->sg_count, sizeof(bus_dma_segment_t));
832 if (data->segs == NULL) {
833 warn("%s: unable to allocate %zu bytes for S/G list",
834 __func__, sizeof(bus_dma_segment_t) *
841 data->iovec = calloc(data->sg_count, sizeof(struct iovec));
842 if (data->iovec == NULL) {
843 warn("%s: unable to allocate %zu bytes for S/G list",
844 __func__, sizeof(struct iovec) * data->sg_count);
850 for (i = 0, tmp_buf = STAILQ_FIRST(&buf->src_list);
851 i < buf->src_count && tmp_buf != NULL; i++,
852 tmp_buf = STAILQ_NEXT(tmp_buf, src_links)) {
854 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
855 struct camdd_buf_data *tmp_data;
857 tmp_data = &tmp_buf->buf_type_spec.data;
859 data->segs[i].ds_addr =
860 (bus_addr_t) tmp_data->buf;
861 data->segs[i].ds_len = tmp_data->fill_len -
864 data->iovec[i].iov_base = tmp_data->buf;
865 data->iovec[i].iov_len = tmp_data->fill_len -
868 if (((tmp_data->fill_len - tmp_data->resid) %
870 *double_buf_needed = 1;
872 struct camdd_buf_indirect *tmp_ind;
874 tmp_ind = &tmp_buf->buf_type_spec.indirect;
876 data->segs[i].ds_addr =
877 (bus_addr_t)tmp_ind->start_ptr;
878 data->segs[i].ds_len = tmp_ind->len;
880 data->iovec[i].iov_base = tmp_ind->start_ptr;
881 data->iovec[i].iov_len = tmp_ind->len;
883 if ((tmp_ind->len % sector_size) != 0)
884 *double_buf_needed = 1;
888 if (extra_buf != NULL) {
890 data->segs[i].ds_addr = (bus_addr_t)extra_buf;
891 data->segs[i].ds_len = extra_buf_len;
893 data->iovec[i].iov_base = extra_buf;
894 data->iovec[i].iov_len = extra_buf_len;
896 extra_buf_attached = 1;
899 if ((tmp_buf != NULL) || (i != data->sg_count)) {
900 warnx("buffer source count does not match "
901 "number of buffers in list!");
908 *num_sectors_used = (data->fill_len + extra_buf_len) /
910 } else if (extra_buf_attached == 0) {
912 * If extra_buf isn't attached yet, we need to free it
923 camdd_buf_get_len(struct camdd_buf *buf)
927 if (buf->buf_type != CAMDD_BUF_DATA) {
928 struct camdd_buf_indirect *indirect;
930 indirect = &buf->buf_type_spec.indirect;
933 struct camdd_buf_data *data;
935 data = &buf->buf_type_spec.data;
936 len = data->fill_len;
943 camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf)
945 struct camdd_buf_data *data;
947 assert(buf->buf_type == CAMDD_BUF_DATA);
949 data = &buf->buf_type_spec.data;
951 STAILQ_INSERT_TAIL(&buf->src_list, child_buf, src_links);
954 data->fill_len += camdd_buf_get_len(child_buf);
962 } camdd_status_item_index;
964 static struct camdd_status_items {
966 struct mt_status_entry *entry;
967 } req_status_items[] = {
970 { "blk_gran", NULL },
971 { "max_effective_iosize", NULL }
975 camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
976 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran)
978 struct mt_status_data status_data;
979 char *xml_str = NULL;
983 retval = mt_get_xml_str(fd, MTIOCEXTGET, &xml_str);
985 err(1, "Couldn't get XML string from %s", filename);
987 retval = mt_get_status(xml_str, &status_data);
988 if (retval != XML_STATUS_OK) {
989 warn("couldn't get status for %s", filename);
995 if (status_data.error != 0) {
996 warnx("%s", status_data.error_str);
1001 for (i = 0; i < nitems(req_status_items); i++) {
1004 name = __DECONST(char *, req_status_items[i].name);
1005 req_status_items[i].entry = mt_status_entry_find(&status_data,
1007 if (req_status_items[i].entry == NULL) {
1008 errx(1, "Cannot find status entry %s",
1009 req_status_items[i].name);
1013 *max_iosize = req_status_items[CAMDD_TS_EFF_IOSIZE].entry->value_unsigned;
1014 *max_blk= req_status_items[CAMDD_TS_MAX_BLK].entry->value_unsigned;
1015 *min_blk= req_status_items[CAMDD_TS_MIN_BLK].entry->value_unsigned;
1016 *blk_gran = req_status_items[CAMDD_TS_BLK_GRAN].entry->value_unsigned;
1020 mt_status_free(&status_data);
1026 camdd_probe_file(int fd, struct camdd_io_opts *io_opts, int retry_count,
1029 struct camdd_dev *dev = NULL;
1030 struct camdd_dev_file *file_dev;
1031 uint64_t blocksize = io_opts->blocksize;
1033 dev = camdd_alloc_dev(CAMDD_DEV_FILE, NULL, 0, retry_count, timeout);
1037 file_dev = &dev->dev_spec.file;
1039 strlcpy(file_dev->filename, io_opts->dev_name,
1040 sizeof(file_dev->filename));
1041 strlcpy(dev->device_name, io_opts->dev_name, sizeof(dev->device_name));
1043 dev->blocksize = CAMDD_FILE_DEFAULT_BLOCK;
1045 dev->blocksize = blocksize;
1047 if ((io_opts->queue_depth != 0)
1048 && (io_opts->queue_depth != 1)) {
1049 warnx("Queue depth %ju for %s ignored, only 1 outstanding "
1050 "command supported", (uintmax_t)io_opts->queue_depth,
1053 dev->target_queue_depth = CAMDD_FILE_DEFAULT_DEPTH;
1054 dev->run = camdd_file_run;
1058 * We can effectively access files on byte boundaries. We'll reset
1059 * this for devices like disks that can be accessed on sector
1062 dev->sector_size = 1;
1064 if ((fd != STDIN_FILENO)
1065 && (fd != STDOUT_FILENO)) {
1068 retval = fstat(fd, &file_dev->sb);
1070 warn("Cannot stat %s", dev->device_name);
1073 if (S_ISREG(file_dev->sb.st_mode)) {
1074 file_dev->file_type = CAMDD_FILE_REG;
1075 } else if (S_ISCHR(file_dev->sb.st_mode)) {
1078 if (ioctl(fd, FIODTYPE, &type) == -1)
1079 err(1, "FIODTYPE ioctl failed on %s",
1083 file_dev->file_type = CAMDD_FILE_TAPE;
1084 else if (type & D_DISK)
1085 file_dev->file_type = CAMDD_FILE_DISK;
1086 else if (type & D_MEM)
1087 file_dev->file_type = CAMDD_FILE_MEM;
1088 else if (type & D_TTY)
1089 file_dev->file_type = CAMDD_FILE_TTY;
1091 } else if (S_ISDIR(file_dev->sb.st_mode)) {
1092 errx(1, "cannot operate on directory %s",
1094 } else if (S_ISFIFO(file_dev->sb.st_mode)) {
1095 file_dev->file_type = CAMDD_FILE_PIPE;
1097 errx(1, "Cannot determine file type for %s",
1100 switch (file_dev->file_type) {
1101 case CAMDD_FILE_REG:
1102 if (file_dev->sb.st_size != 0)
1103 dev->max_sector = file_dev->sb.st_size - 1;
1105 dev->max_sector = 0;
1106 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1108 case CAMDD_FILE_TAPE: {
1109 uint64_t max_iosize, max_blk, min_blk, blk_gran;
1111 * Check block limits and maximum effective iosize.
1112 * Make sure the blocksize is within the block
1113 * limits (and a multiple of the minimum blocksize)
1114 * and that the blocksize is <= maximum effective
1117 retval = camdd_probe_tape(fd, dev->device_name,
1118 &max_iosize, &max_blk, &min_blk, &blk_gran);
1120 errx(1, "Unable to probe tape %s",
1124 * The blocksize needs to be <= the maximum
1125 * effective I/O size of the tape device. Note
1126 * that this also takes into account the maximum
1127 * blocksize reported by READ BLOCK LIMITS.
1129 if (dev->blocksize > max_iosize) {
1130 warnx("Blocksize %u too big for %s, limiting "
1131 "to %ju", dev->blocksize, dev->device_name,
1133 dev->blocksize = max_iosize;
1137 * The blocksize needs to be at least min_blk;
1139 if (dev->blocksize < min_blk) {
1140 warnx("Blocksize %u too small for %s, "
1141 "increasing to %ju", dev->blocksize,
1142 dev->device_name, min_blk);
1143 dev->blocksize = min_blk;
1147 * And the blocksize needs to be a multiple of
1148 * the block granularity.
1151 && (dev->blocksize % (1 << blk_gran))) {
1152 warnx("Blocksize %u for %s not a multiple of "
1153 "%d, adjusting to %d", dev->blocksize,
1154 dev->device_name, (1 << blk_gran),
1155 dev->blocksize & ~((1 << blk_gran) - 1));
1156 dev->blocksize &= ~((1 << blk_gran) - 1);
1159 if (dev->blocksize == 0) {
1160 errx(1, "Unable to derive valid blocksize for "
1161 "%s", dev->device_name);
1165 * For tape drives, set the sector size to the
1166 * blocksize so that we make sure not to write
1167 * less than the blocksize out to the drive.
1169 dev->sector_size = dev->blocksize;
1172 case CAMDD_FILE_DISK: {
1174 unsigned int sector_size;
1176 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1178 if (ioctl(fd, DIOCGSECTORSIZE, §or_size) == -1) {
1179 err(1, "DIOCGSECTORSIZE ioctl failed on %s",
1183 if (sector_size == 0) {
1184 errx(1, "DIOCGSECTORSIZE ioctl returned "
1185 "invalid sector size %u for %s",
1186 sector_size, dev->device_name);
1189 if (ioctl(fd, DIOCGMEDIASIZE, &media_size) == -1) {
1190 err(1, "DIOCGMEDIASIZE ioctl failed on %s",
1194 if (media_size == 0) {
1195 errx(1, "DIOCGMEDIASIZE ioctl returned "
1196 "invalid media size %ju for %s",
1197 (uintmax_t)media_size, dev->device_name);
1200 if (dev->blocksize % sector_size) {
1201 errx(1, "%s blocksize %u not a multiple of "
1202 "sector size %u", dev->device_name,
1203 dev->blocksize, sector_size);
1206 dev->sector_size = sector_size;
1207 dev->max_sector = (media_size / sector_size) - 1;
1210 case CAMDD_FILE_MEM:
1211 file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1218 if ((io_opts->offset != 0)
1219 && ((file_dev->file_flags & CAMDD_FF_CAN_SEEK) == 0)) {
1220 warnx("Offset %ju specified for %s, but we cannot seek on %s",
1221 io_opts->offset, io_opts->dev_name, io_opts->dev_name);
1225 else if ((io_opts->offset != 0)
1226 && ((io_opts->offset % dev->sector_size) != 0)) {
1227 warnx("Offset %ju for %s is not a multiple of the "
1228 "sector size %u", io_opts->offset,
1229 io_opts->dev_name, dev->sector_size);
1232 dev->start_offset_bytes = io_opts->offset;
1240 camdd_free_dev(dev);
1245 * Get a get device CCB for the specified device.
1248 camdd_get_cgd(struct cam_device *device, struct ccb_getdev *cgd)
1253 ccb = cam_getccb(device);
1256 warnx("%s: couldn't allocate CCB", __func__);
1260 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cgd);
1262 ccb->ccb_h.func_code = XPT_GDEV_TYPE;
1264 if (cam_send_ccb(device, ccb) < 0) {
1265 warn("%s: error sending Get Device Information CCB", __func__);
1266 cam_error_print(device, ccb, CAM_ESF_ALL,
1267 CAM_EPF_ALL, stderr);
1272 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1273 cam_error_print(device, ccb, CAM_ESF_ALL,
1274 CAM_EPF_ALL, stderr);
1279 bcopy(&ccb->cgd, cgd, sizeof(struct ccb_getdev));
1288 camdd_probe_pass_scsi(struct cam_device *cam_dev, union ccb *ccb,
1289 camdd_argmask arglist, int probe_retry_count,
1290 int probe_timeout, uint64_t *maxsector, uint32_t *block_len)
1292 struct scsi_read_capacity_data rcap;
1293 struct scsi_read_capacity_data_long rcaplong;
1297 warnx("%s: error passed ccb is NULL", __func__);
1301 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
1303 scsi_read_capacity(&ccb->csio,
1304 /*retries*/ probe_retry_count,
1306 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1309 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1311 /* Disable freezing the device queue */
1312 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1314 if (arglist & CAMDD_ARG_ERR_RECOVER)
1315 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1317 if (cam_send_ccb(cam_dev, ccb) < 0) {
1318 warn("error sending READ CAPACITY command");
1320 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1321 CAM_EPF_ALL, stderr);
1326 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1327 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1331 *maxsector = scsi_4btoul(rcap.addr);
1332 *block_len = scsi_4btoul(rcap.length);
1335 * A last block of 2^32-1 means that the true capacity is over 2TB,
1336 * and we need to issue the long READ CAPACITY to get the real
1337 * capacity. Otherwise, we're all set.
1339 if (*maxsector != 0xffffffff) {
1344 scsi_read_capacity_16(&ccb->csio,
1345 /*retries*/ probe_retry_count,
1347 /*tag_action*/ MSG_SIMPLE_Q_TAG,
1351 (uint8_t *)&rcaplong,
1353 /*sense_len*/ SSD_FULL_SIZE,
1354 /*timeout*/ probe_timeout ? probe_timeout : 5000);
1356 /* Disable freezing the device queue */
1357 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1359 if (arglist & CAMDD_ARG_ERR_RECOVER)
1360 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1362 if (cam_send_ccb(cam_dev, ccb) < 0) {
1363 warn("error sending READ CAPACITY (16) command");
1364 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1365 CAM_EPF_ALL, stderr);
1369 if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1370 cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1374 *maxsector = scsi_8btou64(rcaplong.addr);
1375 *block_len = scsi_4btoul(rcaplong.length);
1384 * Need to implement this. Do a basic probe:
1385 * - Check the inquiry data, make sure we're talking to a device that we
1386 * can reasonably expect to talk to -- direct, RBC, CD, WORM.
1387 * - Send a test unit ready, make sure the device is available.
1388 * - Get the capacity and block size.
1391 camdd_probe_pass(struct cam_device *cam_dev, struct camdd_io_opts *io_opts,
1392 camdd_argmask arglist, int probe_retry_count,
1393 int probe_timeout, int io_retry_count, int io_timeout)
1396 uint64_t maxsector = 0;
1397 uint32_t cpi_maxio, max_iosize, pass_numblocks;
1398 uint32_t block_len = 0;
1399 struct camdd_dev *dev = NULL;
1400 struct camdd_dev_pass *pass_dev;
1402 struct ccb_getdev cgd;
1406 if ((retval = camdd_get_cgd(cam_dev, &cgd)) != 0) {
1407 warnx("%s: error retrieving CGD", __func__);
1411 ccb = cam_getccb(cam_dev);
1414 warnx("%s: error allocating ccb", __func__);
1418 switch (cgd.protocol) {
1420 scsi_dev_type = SID_TYPE(&cam_dev->inq_data);
1423 * For devices that support READ CAPACITY, we'll attempt to get the
1424 * capacity. Otherwise, we really don't support tape or other
1425 * devices via SCSI passthrough, so just return an error in that case.
1427 switch (scsi_dev_type) {
1436 errx(1, "Unsupported SCSI device type %d", scsi_dev_type);
1437 break; /*NOTREACHED*/
1440 if ((retval = camdd_probe_pass_scsi(cam_dev, ccb, probe_retry_count,
1441 arglist, probe_timeout, &maxsector,
1447 errx(1, "Unsupported PROTO type %d", cgd.protocol);
1448 break; /*NOTREACHED*/
1451 if (block_len == 0) {
1452 warnx("Sector size for %s%u is 0, cannot continue",
1453 cam_dev->device_name, cam_dev->dev_unit_num);
1457 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cpi);
1459 ccb->ccb_h.func_code = XPT_PATH_INQ;
1460 ccb->ccb_h.flags = CAM_DIR_NONE;
1461 ccb->ccb_h.retry_count = 1;
1463 if (cam_send_ccb(cam_dev, ccb) < 0) {
1464 warn("error sending XPT_PATH_INQ CCB");
1466 cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1467 CAM_EPF_ALL, stderr);
1471 EV_SET(&ke, cam_dev->fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1473 dev = camdd_alloc_dev(CAMDD_DEV_PASS, &ke, 1, io_retry_count,
1478 pass_dev = &dev->dev_spec.pass;
1479 pass_dev->scsi_dev_type = scsi_dev_type;
1480 pass_dev->protocol = cgd.protocol;
1481 pass_dev->dev = cam_dev;
1482 pass_dev->max_sector = maxsector;
1483 pass_dev->block_len = block_len;
1484 pass_dev->cpi_maxio = ccb->cpi.maxio;
1485 snprintf(dev->device_name, sizeof(dev->device_name), "%s%u",
1486 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
1487 dev->sector_size = block_len;
1488 dev->max_sector = maxsector;
1492 * Determine the optimal blocksize to use for this device.
1496 * If the controller has not specified a maximum I/O size,
1497 * just go with 128K as a somewhat conservative value.
1499 if (pass_dev->cpi_maxio == 0)
1502 cpi_maxio = pass_dev->cpi_maxio;
1505 * If the controller has a large maximum I/O size, limit it
1506 * to something smaller so that the kernel doesn't have trouble
1507 * allocating buffers to copy data in and out for us.
1508 * XXX KDM this is until we have unmapped I/O support in the kernel.
1510 max_iosize = min(cpi_maxio, CAMDD_PASS_MAX_BLOCK);
1513 * If we weren't able to get a block size for some reason,
1514 * default to 512 bytes.
1516 block_len = pass_dev->block_len;
1521 * Figure out how many blocksize chunks will fit in the
1524 pass_numblocks = max_iosize / block_len;
1527 * And finally, multiple the number of blocks by the LBA
1528 * length to get our maximum block size;
1530 dev->blocksize = pass_numblocks * block_len;
1532 if (io_opts->blocksize != 0) {
1533 if ((io_opts->blocksize % dev->sector_size) != 0) {
1534 warnx("Blocksize %ju for %s is not a multiple of "
1535 "sector size %u", (uintmax_t)io_opts->blocksize,
1536 dev->device_name, dev->sector_size);
1539 dev->blocksize = io_opts->blocksize;
1541 dev->target_queue_depth = CAMDD_PASS_DEFAULT_DEPTH;
1542 if (io_opts->queue_depth != 0)
1543 dev->target_queue_depth = io_opts->queue_depth;
1545 if (io_opts->offset != 0) {
1546 if (io_opts->offset > (dev->max_sector * dev->sector_size)) {
1547 warnx("Offset %ju is past the end of device %s",
1548 io_opts->offset, dev->device_name);
1552 else if ((io_opts->offset % dev->sector_size) != 0) {
1553 warnx("Offset %ju for %s is not a multiple of the "
1554 "sector size %u", io_opts->offset,
1555 dev->device_name, dev->sector_size);
1558 dev->start_offset_bytes = io_opts->offset;
1562 dev->min_cmd_size = io_opts->min_cmd_size;
1564 dev->run = camdd_pass_run;
1565 dev->fetch = camdd_pass_fetch;
1575 camdd_free_dev(dev);
1581 camdd_worker(void *arg)
1583 struct camdd_dev *dev = arg;
1584 struct camdd_buf *buf;
1585 struct timespec ts, *kq_ts;
1590 pthread_mutex_lock(&dev->mutex);
1592 dev->flags |= CAMDD_DEV_FLAG_ACTIVE;
1599 * XXX KDM check the reorder queue depth?
1601 if (dev->write_dev == 0) {
1602 uint32_t our_depth, peer_depth, peer_bytes, our_bytes;
1603 uint32_t target_depth = dev->target_queue_depth;
1604 uint32_t peer_target_depth =
1605 dev->peer_dev->target_queue_depth;
1606 uint32_t peer_blocksize = dev->peer_dev->blocksize;
1608 camdd_get_depth(dev, &our_depth, &peer_depth,
1609 &our_bytes, &peer_bytes);
1612 while (((our_depth < target_depth)
1613 && (peer_depth < peer_target_depth))
1614 || ((peer_bytes + our_bytes) <
1615 (peer_blocksize * 2))) {
1617 while (((our_depth + peer_depth) <
1618 (target_depth + peer_target_depth))
1619 || ((peer_bytes + our_bytes) <
1620 (peer_blocksize * 3))) {
1622 retval = camdd_queue(dev, NULL);
1625 else if (retval != 0) {
1630 camdd_get_depth(dev, &our_depth, &peer_depth,
1631 &our_bytes, &peer_bytes);
1635 * See if we have any I/O that is ready to execute.
1637 buf = STAILQ_FIRST(&dev->run_queue);
1639 while (dev->target_queue_depth > dev->cur_active_io) {
1640 retval = dev->run(dev);
1642 dev->flags |= CAMDD_DEV_FLAG_EOF;
1645 } else if (retval != 0) {
1652 * We've reached EOF, or our partner has reached EOF.
1654 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
1655 || (dev->flags & CAMDD_DEV_FLAG_PEER_EOF)) {
1656 if (dev->write_dev != 0) {
1657 if ((STAILQ_EMPTY(&dev->work_queue))
1658 && (dev->num_run_queue == 0)
1659 && (dev->cur_active_io == 0)) {
1664 * If we're the reader, and the writer
1665 * got EOF, he is already done. If we got
1666 * the EOF, then we need to wait until
1667 * everything is flushed out for the writer.
1669 if (dev->flags & CAMDD_DEV_FLAG_PEER_EOF) {
1671 } else if ((dev->num_peer_work_queue == 0)
1672 && (dev->num_peer_done_queue == 0)
1673 && (dev->cur_active_io == 0)
1674 && (dev->num_run_queue == 0)) {
1679 * XXX KDM need to do something about the pending
1680 * queue and cleanup resources.
1684 if ((dev->write_dev == 0)
1685 && (dev->cur_active_io == 0)
1686 && (dev->peer_bytes_queued < dev->peer_dev->blocksize))
1692 * Run kevent to see if there are events to process.
1694 pthread_mutex_unlock(&dev->mutex);
1695 retval = kevent(dev->kq, NULL, 0, &ke, 1, kq_ts);
1696 pthread_mutex_lock(&dev->mutex);
1698 warn("%s: error returned from kevent",__func__);
1700 } else if (retval != 0) {
1701 switch (ke.filter) {
1703 if (dev->fetch != NULL) {
1704 retval = dev->fetch(dev);
1713 * We register for this so we don't get
1714 * an error as a result of a SIGINFO or a
1715 * SIGINT. It will actually get handled
1716 * by the signal handler. If we get a
1717 * SIGINT, bail out without printing an
1718 * error message. Any other signals
1719 * will result in the error message above.
1721 if (ke.ident == SIGINT)
1727 * Check to see if the other thread has
1728 * queued any I/O for us to do. (In this
1729 * case we're the writer.)
1731 for (buf = STAILQ_FIRST(&dev->work_queue);
1733 buf = STAILQ_FIRST(&dev->work_queue)) {
1734 STAILQ_REMOVE_HEAD(&dev->work_queue,
1736 retval = camdd_queue(dev, buf);
1738 * We keep going unless we get an
1739 * actual error. If we get EOF, we
1740 * still want to remove the buffers
1741 * from the queue and send the back
1742 * to the reader thread.
1752 * Next check to see if the other thread has
1753 * queued any completed buffers back to us.
1754 * (In this case we're the reader.)
1756 for (buf = STAILQ_FIRST(&dev->peer_done_queue);
1758 buf = STAILQ_FIRST(&dev->peer_done_queue)){
1760 &dev->peer_done_queue, work_links);
1761 dev->num_peer_done_queue--;
1762 camdd_peer_done(buf);
1766 warnx("%s: unknown kevent filter %d",
1767 __func__, ke.filter);
1775 dev->flags &= ~CAMDD_DEV_FLAG_ACTIVE;
1777 /* XXX KDM cleanup resources here? */
1779 pthread_mutex_unlock(&dev->mutex);
1782 sem_post(&camdd_sem);
1788 * Simplistic translation of CCB status to our local status.
1791 camdd_ccb_status(union ccb *ccb, int protocol)
1793 camdd_buf_status status = CAMDD_STATUS_NONE;
1794 cam_status ccb_status;
1796 ccb_status = ccb->ccb_h.status & CAM_STATUS_MASK;
1800 switch (ccb_status) {
1802 if (ccb->csio.resid == 0) {
1803 status = CAMDD_STATUS_OK;
1804 } else if (ccb->csio.dxfer_len > ccb->csio.resid) {
1805 status = CAMDD_STATUS_SHORT_IO;
1807 status = CAMDD_STATUS_EOF;
1811 case CAM_SCSI_STATUS_ERROR: {
1812 switch (ccb->csio.scsi_status) {
1813 case SCSI_STATUS_OK:
1814 case SCSI_STATUS_COND_MET:
1815 case SCSI_STATUS_INTERMED:
1816 case SCSI_STATUS_INTERMED_COND_MET:
1817 status = CAMDD_STATUS_OK;
1819 case SCSI_STATUS_CMD_TERMINATED:
1820 case SCSI_STATUS_CHECK_COND:
1821 case SCSI_STATUS_QUEUE_FULL:
1822 case SCSI_STATUS_BUSY:
1823 case SCSI_STATUS_RESERV_CONFLICT:
1825 status = CAMDD_STATUS_ERROR;
1831 status = CAMDD_STATUS_ERROR;
1836 status = CAMDD_STATUS_ERROR;
1844 * Queue a buffer to our peer's work thread for writing.
1846 * Returns 0 for success, -1 for failure, 1 if the other thread exited.
1849 camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf)
1852 STAILQ_HEAD(, camdd_buf) local_queue;
1853 struct camdd_buf *buf1, *buf2;
1854 struct camdd_buf_data *data = NULL;
1855 uint64_t peer_bytes_queued = 0;
1859 STAILQ_INIT(&local_queue);
1862 * Since we're the reader, we need to queue our I/O to the writer
1863 * in sequential order in order to make sure it gets written out
1864 * in sequential order.
1866 * Check the next expected I/O starting offset. If this doesn't
1867 * match, put it on the reorder queue.
1869 if ((buf->lba * dev->sector_size) != dev->next_completion_pos_bytes) {
1872 * If there is nothing on the queue, there is no sorting
1875 if (STAILQ_EMPTY(&dev->reorder_queue)) {
1876 STAILQ_INSERT_TAIL(&dev->reorder_queue, buf, links);
1877 dev->num_reorder_queue++;
1882 * Sort in ascending order by starting LBA. There should
1883 * be no identical LBAs.
1885 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1887 buf2 = STAILQ_NEXT(buf1, links);
1888 if (buf->lba < buf1->lba) {
1890 * If we're less than the first one, then
1891 * we insert at the head of the list
1892 * because this has to be the first element
1895 STAILQ_INSERT_HEAD(&dev->reorder_queue,
1897 dev->num_reorder_queue++;
1899 } else if (buf->lba > buf1->lba) {
1901 STAILQ_INSERT_TAIL(&dev->reorder_queue,
1903 dev->num_reorder_queue++;
1905 } else if (buf->lba < buf2->lba) {
1906 STAILQ_INSERT_AFTER(&dev->reorder_queue,
1908 dev->num_reorder_queue++;
1912 errx(1, "Found buffers with duplicate LBA %ju!",
1920 * We're the next expected I/O completion, so put ourselves
1921 * on the local queue to be sent to the writer. We use
1922 * work_links here so that we can queue this to the
1923 * peer_work_queue before taking the buffer off of the
1926 dev->next_completion_pos_bytes += buf->len;
1927 STAILQ_INSERT_TAIL(&local_queue, buf, work_links);
1930 * Go through the reorder queue looking for more sequential
1931 * I/O and add it to the local queue.
1933 for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1934 buf1 = STAILQ_FIRST(&dev->reorder_queue)) {
1936 * As soon as we see an I/O that is out of sequence,
1939 if ((buf1->lba * dev->sector_size) !=
1940 dev->next_completion_pos_bytes)
1943 STAILQ_REMOVE_HEAD(&dev->reorder_queue, links);
1944 dev->num_reorder_queue--;
1945 STAILQ_INSERT_TAIL(&local_queue, buf1, work_links);
1946 dev->next_completion_pos_bytes += buf1->len;
1951 * Setup the event to let the other thread know that it has work
1954 EV_SET(&ke, (uintptr_t)&dev->peer_dev->work_queue, EVFILT_USER, 0,
1955 NOTE_TRIGGER, 0, NULL);
1958 * Put this on our shadow queue so that we know what we've queued
1959 * to the other thread.
1961 STAILQ_FOREACH_SAFE(buf1, &local_queue, work_links, buf2) {
1962 if (buf1->buf_type != CAMDD_BUF_DATA) {
1963 errx(1, "%s: should have a data buffer, not an "
1964 "indirect buffer", __func__);
1966 data = &buf1->buf_type_spec.data;
1969 * We only need to send one EOF to the writer, and don't
1970 * need to continue sending EOFs after that.
1972 if (buf1->status == CAMDD_STATUS_EOF) {
1973 if (dev->flags & CAMDD_DEV_FLAG_EOF_SENT) {
1974 STAILQ_REMOVE(&local_queue, buf1, camdd_buf,
1976 camdd_release_buf(buf1);
1980 dev->flags |= CAMDD_DEV_FLAG_EOF_SENT;
1984 STAILQ_INSERT_TAIL(&dev->peer_work_queue, buf1, links);
1985 peer_bytes_queued += (data->fill_len - data->resid);
1986 dev->peer_bytes_queued += (data->fill_len - data->resid);
1987 dev->num_peer_work_queue++;
1990 if (STAILQ_FIRST(&local_queue) == NULL)
1994 * Drop our mutex and pick up the other thread's mutex. We need to
1995 * do this to avoid deadlocks.
1997 pthread_mutex_unlock(&dev->mutex);
1998 pthread_mutex_lock(&dev->peer_dev->mutex);
2000 if (dev->peer_dev->flags & CAMDD_DEV_FLAG_ACTIVE) {
2002 * Put the buffers on the other thread's incoming work queue.
2004 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
2005 buf1 = STAILQ_FIRST(&local_queue)) {
2006 STAILQ_REMOVE_HEAD(&local_queue, work_links);
2007 STAILQ_INSERT_TAIL(&dev->peer_dev->work_queue, buf1,
2011 * Send an event to the other thread's kqueue to let it know
2012 * that there is something on the work queue.
2014 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2016 warn("%s: unable to add peer work_queue kevent",
2023 pthread_mutex_unlock(&dev->peer_dev->mutex);
2024 pthread_mutex_lock(&dev->mutex);
2027 * If the other side isn't active, run through the queue and
2028 * release all of the buffers.
2031 for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
2032 buf1 = STAILQ_FIRST(&local_queue)) {
2033 STAILQ_REMOVE_HEAD(&local_queue, work_links);
2034 STAILQ_REMOVE(&dev->peer_work_queue, buf1, camdd_buf,
2036 dev->num_peer_work_queue--;
2037 camdd_release_buf(buf1);
2039 dev->peer_bytes_queued -= peer_bytes_queued;
2048 * Return a buffer to the reader thread when we have completed writing it.
2051 camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf)
2057 * Setup the event to let the other thread know that we have
2058 * completed a buffer.
2060 EV_SET(&ke, (uintptr_t)&dev->peer_dev->peer_done_queue, EVFILT_USER, 0,
2061 NOTE_TRIGGER, 0, NULL);
2064 * Drop our lock and acquire the other thread's lock before
2067 pthread_mutex_unlock(&dev->mutex);
2068 pthread_mutex_lock(&dev->peer_dev->mutex);
2071 * Put the buffer on the reader thread's peer done queue now that
2072 * we have completed it.
2074 STAILQ_INSERT_TAIL(&dev->peer_dev->peer_done_queue, peer_buf,
2076 dev->peer_dev->num_peer_done_queue++;
2079 * Send an event to the peer thread to let it know that we've added
2080 * something to its peer done queue.
2082 retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2084 warn("%s: unable to add peer_done_queue kevent", __func__);
2089 * Drop the other thread's lock and reacquire ours.
2091 pthread_mutex_unlock(&dev->peer_dev->mutex);
2092 pthread_mutex_lock(&dev->mutex);
2098 * Free a buffer that was written out by the writer thread and returned to
2099 * the reader thread.
2102 camdd_peer_done(struct camdd_buf *buf)
2104 struct camdd_dev *dev;
2105 struct camdd_buf_data *data;
2108 if (buf->buf_type != CAMDD_BUF_DATA) {
2109 errx(1, "%s: should have a data buffer, not an "
2110 "indirect buffer", __func__);
2113 data = &buf->buf_type_spec.data;
2115 STAILQ_REMOVE(&dev->peer_work_queue, buf, camdd_buf, links);
2116 dev->num_peer_work_queue--;
2117 dev->peer_bytes_queued -= (data->fill_len - data->resid);
2119 if (buf->status == CAMDD_STATUS_EOF)
2120 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2122 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2126 * Assumes caller holds the lock for this device.
2129 camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
2135 * If we're the reader, we need to send the completed I/O
2136 * to the writer. If we're the writer, we need to just
2137 * free up resources, or let the reader know if we've
2138 * encountered an error.
2140 if (dev->write_dev == 0) {
2141 retval = camdd_queue_peer_buf(dev, buf);
2145 struct camdd_buf *tmp_buf, *next_buf;
2147 STAILQ_FOREACH_SAFE(tmp_buf, &buf->src_list, src_links,
2149 struct camdd_buf *src_buf;
2150 struct camdd_buf_indirect *indirect;
2152 STAILQ_REMOVE(&buf->src_list, tmp_buf,
2153 camdd_buf, src_links);
2155 tmp_buf->status = buf->status;
2157 if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
2158 camdd_complete_peer_buf(dev, tmp_buf);
2162 indirect = &tmp_buf->buf_type_spec.indirect;
2163 src_buf = indirect->src_buf;
2164 src_buf->refcount--;
2166 * XXX KDM we probably need to account for
2167 * exactly how many bytes we were able to
2168 * write. Allocate the residual to the
2169 * first N buffers? Or just track the
2170 * number of bytes written? Right now the reader
2171 * doesn't do anything with a residual.
2173 src_buf->status = buf->status;
2174 if (src_buf->refcount <= 0)
2175 camdd_complete_peer_buf(dev, src_buf);
2176 STAILQ_INSERT_TAIL(&dev->free_indirect_queue,
2180 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2185 * Fetch all completed commands from the pass(4) device.
2187 * Returns the number of commands received, or -1 if any of the commands
2188 * completed with an error. Returns 0 if no commands are available.
2191 camdd_pass_fetch(struct camdd_dev *dev)
2193 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2195 int retval = 0, num_fetched = 0, error_count = 0;
2197 pthread_mutex_unlock(&dev->mutex);
2199 * XXX KDM we don't distinguish between EFAULT and ENOENT.
2201 while ((retval = ioctl(pass_dev->dev->fd, CAMIOGET, &ccb)) != -1) {
2202 struct camdd_buf *buf;
2203 struct camdd_buf_data *data;
2204 cam_status ccb_status;
2207 buf = ccb.ccb_h.ccb_buf;
2208 data = &buf->buf_type_spec.data;
2209 buf_ccb = &data->ccb;
2214 * Copy the CCB back out so we get status, sense data, etc.
2216 bcopy(&ccb, buf_ccb, sizeof(ccb));
2218 pthread_mutex_lock(&dev->mutex);
2221 * We're now done, so take this off the active queue.
2223 STAILQ_REMOVE(&dev->active_queue, buf, camdd_buf, links);
2224 dev->cur_active_io--;
2226 ccb_status = ccb.ccb_h.status & CAM_STATUS_MASK;
2227 if (ccb_status != CAM_REQ_CMP) {
2228 cam_error_print(pass_dev->dev, &ccb, CAM_ESF_ALL,
2229 CAM_EPF_ALL, stderr);
2232 switch (pass_dev->protocol) {
2234 data->resid = ccb.csio.resid;
2235 dev->bytes_transferred += (ccb.csio.dxfer_len - ccb.csio.resid);
2242 if (buf->status == CAMDD_STATUS_NONE)
2243 buf->status = camdd_ccb_status(&ccb, pass_dev->protocol);
2244 if (buf->status == CAMDD_STATUS_ERROR)
2246 else if (buf->status == CAMDD_STATUS_EOF) {
2248 * Once we queue this buffer to our partner thread,
2249 * he will know that we've hit EOF.
2251 dev->flags |= CAMDD_DEV_FLAG_EOF;
2254 camdd_complete_buf(dev, buf, &error_count);
2257 * Unlock in preparation for the ioctl call.
2259 pthread_mutex_unlock(&dev->mutex);
2262 pthread_mutex_lock(&dev->mutex);
2264 if (error_count > 0)
2267 return (num_fetched);
2271 * Returns -1 for error, 0 for success/continue, and 1 for resource
2272 * shortage/stop processing.
2275 camdd_file_run(struct camdd_dev *dev)
2277 struct camdd_dev_file *file_dev = &dev->dev_spec.file;
2278 struct camdd_buf_data *data;
2279 struct camdd_buf *buf;
2281 int retval = 0, write_dev = dev->write_dev;
2282 int error_count = 0, no_resources = 0, double_buf_needed = 0;
2283 uint32_t num_sectors = 0, db_len = 0;
2285 buf = STAILQ_FIRST(&dev->run_queue);
2289 } else if ((dev->write_dev == 0)
2290 && (dev->flags & (CAMDD_DEV_FLAG_EOF |
2291 CAMDD_DEV_FLAG_EOF_SENT))) {
2292 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2293 dev->num_run_queue--;
2294 buf->status = CAMDD_STATUS_EOF;
2300 * If we're writing, we need to go through the source buffer list
2301 * and create an S/G list.
2303 if (write_dev != 0) {
2304 retval = camdd_buf_sg_create(buf, /*iovec*/ 1,
2305 dev->sector_size, &num_sectors, &double_buf_needed);
2312 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2313 dev->num_run_queue--;
2315 data = &buf->buf_type_spec.data;
2318 * pread(2) and pwrite(2) offsets are byte offsets.
2320 io_offset = buf->lba * dev->sector_size;
2323 * Unlock the mutex while we read or write.
2325 pthread_mutex_unlock(&dev->mutex);
2328 * Note that we don't need to double buffer if we're the reader
2329 * because in that case, we have allocated a single buffer of
2330 * sufficient size to do the read. This copy is necessary on
2331 * writes because if one of the components of the S/G list is not
2332 * a sector size multiple, the kernel will reject the write. This
2333 * is unfortunate but not surprising. So this will make sure that
2334 * we're using a single buffer that is a multiple of the sector size.
2336 if ((double_buf_needed != 0)
2337 && (data->sg_count > 1)
2338 && (write_dev != 0)) {
2339 uint32_t cur_offset;
2342 if (file_dev->tmp_buf == NULL)
2343 file_dev->tmp_buf = calloc(dev->blocksize, 1);
2344 if (file_dev->tmp_buf == NULL) {
2345 buf->status = CAMDD_STATUS_ERROR;
2347 pthread_mutex_lock(&dev->mutex);
2350 for (i = 0, cur_offset = 0; i < data->sg_count; i++) {
2351 bcopy(data->iovec[i].iov_base,
2352 &file_dev->tmp_buf[cur_offset],
2353 data->iovec[i].iov_len);
2354 cur_offset += data->iovec[i].iov_len;
2356 db_len = cur_offset;
2359 if (file_dev->file_flags & CAMDD_FF_CAN_SEEK) {
2360 if (write_dev == 0) {
2362 * XXX KDM is there any way we would need a S/G
2365 retval = pread(file_dev->fd, data->buf,
2366 buf->len, io_offset);
2368 if (double_buf_needed != 0) {
2369 retval = pwrite(file_dev->fd, file_dev->tmp_buf,
2371 } else if (data->sg_count == 0) {
2372 retval = pwrite(file_dev->fd, data->buf,
2373 data->fill_len, io_offset);
2375 retval = pwritev(file_dev->fd, data->iovec,
2376 data->sg_count, io_offset);
2380 if (write_dev == 0) {
2382 * XXX KDM is there any way we would need a S/G
2385 retval = read(file_dev->fd, data->buf, buf->len);
2387 if (double_buf_needed != 0) {
2388 retval = write(file_dev->fd, file_dev->tmp_buf,
2390 } else if (data->sg_count == 0) {
2391 retval = write(file_dev->fd, data->buf,
2394 retval = writev(file_dev->fd, data->iovec,
2400 /* We're done, re-acquire the lock */
2401 pthread_mutex_lock(&dev->mutex);
2403 if (retval >= (ssize_t)data->fill_len) {
2405 * If the bytes transferred is more than the request size,
2406 * that indicates an overrun, which should only happen at
2407 * the end of a transfer if we have to round up to a sector
2410 if (buf->status == CAMDD_STATUS_NONE)
2411 buf->status = CAMDD_STATUS_OK;
2413 dev->bytes_transferred += retval;
2414 } else if (retval == -1) {
2415 warn("Error %s %s", (write_dev) ? "writing to" :
2416 "reading from", file_dev->filename);
2418 buf->status = CAMDD_STATUS_ERROR;
2419 data->resid = data->fill_len;
2422 if (dev->debug == 0)
2425 if ((double_buf_needed != 0)
2426 && (write_dev != 0)) {
2427 fprintf(stderr, "%s: fd %d, DB buf %p, len %u lba %ju "
2428 "offset %ju\n", __func__, file_dev->fd,
2429 file_dev->tmp_buf, db_len, (uintmax_t)buf->lba,
2430 (uintmax_t)io_offset);
2431 } else if (data->sg_count == 0) {
2432 fprintf(stderr, "%s: fd %d, buf %p, len %u, lba %ju "
2433 "offset %ju\n", __func__, file_dev->fd, data->buf,
2434 data->fill_len, (uintmax_t)buf->lba,
2435 (uintmax_t)io_offset);
2439 fprintf(stderr, "%s: fd %d, len %u, lba %ju "
2440 "offset %ju\n", __func__, file_dev->fd,
2441 data->fill_len, (uintmax_t)buf->lba,
2442 (uintmax_t)io_offset);
2444 for (i = 0; i < data->sg_count; i++) {
2445 fprintf(stderr, "index %d ptr %p len %zu\n",
2446 i, data->iovec[i].iov_base,
2447 data->iovec[i].iov_len);
2450 } else if (retval == 0) {
2451 buf->status = CAMDD_STATUS_EOF;
2452 if (dev->debug != 0)
2453 printf("%s: got EOF from %s!\n", __func__,
2454 file_dev->filename);
2455 data->resid = data->fill_len;
2457 } else if (retval < (ssize_t)data->fill_len) {
2458 if (buf->status == CAMDD_STATUS_NONE)
2459 buf->status = CAMDD_STATUS_SHORT_IO;
2460 data->resid = data->fill_len - retval;
2461 dev->bytes_transferred += retval;
2466 if (buf->status == CAMDD_STATUS_EOF) {
2467 struct camdd_buf *buf2;
2468 dev->flags |= CAMDD_DEV_FLAG_EOF;
2469 STAILQ_FOREACH(buf2, &dev->run_queue, links)
2470 buf2->status = CAMDD_STATUS_EOF;
2473 camdd_complete_buf(dev, buf, &error_count);
2476 if (error_count != 0)
2478 else if (no_resources != 0)
2485 * Execute one command from the run queue. Returns 0 for success, 1 for
2486 * stop processing, and -1 for error.
2489 camdd_pass_run(struct camdd_dev *dev)
2491 struct camdd_buf *buf = NULL;
2492 struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2493 struct camdd_buf_data *data;
2494 uint32_t num_blocks, sectors_used = 0;
2496 int retval = 0, is_write = dev->write_dev;
2497 int double_buf_needed = 0;
2499 buf = STAILQ_FIRST(&dev->run_queue);
2506 * If we're writing, we need to go through the source buffer list
2507 * and create an S/G list.
2509 if (is_write != 0) {
2510 retval = camdd_buf_sg_create(buf, /*iovec*/ 0,dev->sector_size,
2511 §ors_used, &double_buf_needed);
2518 STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2519 dev->num_run_queue--;
2521 data = &buf->buf_type_spec.data;
2524 * In almost every case the number of blocks should be the device
2525 * block size. The exception may be at the end of an I/O stream
2526 * for a partial block or at the end of a device.
2529 num_blocks = sectors_used;
2531 num_blocks = data->fill_len / pass_dev->block_len;
2535 switch (pass_dev->protocol) {
2537 CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
2539 scsi_read_write(&ccb->csio,
2540 /*retries*/ dev->retry_count,
2542 /*tag_action*/ MSG_SIMPLE_Q_TAG,
2543 /*readop*/ (dev->write_dev == 0) ? SCSI_RW_READ :
2546 /*minimum_cmd_size*/ dev->min_cmd_size,
2548 /*block_count*/ num_blocks,
2549 /*data_ptr*/ (data->sg_count != 0) ?
2550 (uint8_t *)data->segs : data->buf,
2551 /*dxfer_len*/ (num_blocks * pass_dev->block_len),
2552 /*sense_len*/ SSD_FULL_SIZE,
2553 /*timeout*/ dev->io_timeout);
2555 if (data->sg_count != 0) {
2556 ccb->csio.sglist_cnt = data->sg_count;
2564 /* Disable freezing the device queue */
2565 ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
2567 if (dev->retry_count != 0)
2568 ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
2570 if (data->sg_count != 0) {
2571 ccb->ccb_h.flags |= CAM_DATA_SG;
2575 * Store a pointer to the buffer in the CCB. The kernel will
2576 * restore this when we get it back, and we'll use it to identify
2577 * the buffer this CCB came from.
2579 ccb->ccb_h.ccb_buf = buf;
2582 * Unlock our mutex in preparation for issuing the ioctl.
2584 pthread_mutex_unlock(&dev->mutex);
2586 * Queue the CCB to the pass(4) driver.
2588 if (ioctl(pass_dev->dev->fd, CAMIOQUEUE, ccb) == -1) {
2589 pthread_mutex_lock(&dev->mutex);
2591 warn("%s: error sending CAMIOQUEUE ioctl to %s%u", __func__,
2592 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
2593 warn("%s: CCB address is %p", __func__, ccb);
2596 STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2598 pthread_mutex_lock(&dev->mutex);
2600 dev->cur_active_io++;
2601 STAILQ_INSERT_TAIL(&dev->active_queue, buf, links);
2609 camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len)
2611 struct camdd_dev_pass *pass_dev;
2612 uint32_t num_blocks;
2615 pass_dev = &dev->dev_spec.pass;
2617 *lba = dev->next_io_pos_bytes / dev->sector_size;
2618 *len = dev->blocksize;
2619 num_blocks = *len / dev->sector_size;
2622 * If max_sector is 0, then we have no set limit. This can happen
2623 * if we're writing to a file in a filesystem, or reading from
2624 * something like /dev/zero.
2626 if ((dev->max_sector != 0)
2627 || (dev->sector_io_limit != 0)) {
2628 uint64_t max_sector;
2630 if ((dev->max_sector != 0)
2631 && (dev->sector_io_limit != 0))
2632 max_sector = min(dev->sector_io_limit, dev->max_sector);
2633 else if (dev->max_sector != 0)
2634 max_sector = dev->max_sector;
2636 max_sector = dev->sector_io_limit;
2640 * Check to see whether we're starting off past the end of
2641 * the device. If so, we need to just send an EOF
2642 * notification to the writer.
2644 if (*lba > max_sector) {
2647 } else if (((*lba + num_blocks) > max_sector + 1)
2648 || ((*lba + num_blocks) < *lba)) {
2650 * If we get here (but pass the first check), we
2651 * can trim the request length down to go to the
2652 * end of the device.
2654 num_blocks = (max_sector + 1) - *lba;
2655 *len = num_blocks * dev->sector_size;
2660 dev->next_io_pos_bytes += *len;
2666 * Returns 0 for success, 1 for EOF detected, and -1 for failure.
2669 camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf)
2671 struct camdd_buf *buf = NULL;
2672 struct camdd_buf_data *data;
2673 struct camdd_dev_pass *pass_dev;
2675 struct camdd_buf_data *rb_data;
2676 int is_write = dev->write_dev;
2677 int eof_flush_needed = 0;
2681 pass_dev = &dev->dev_spec.pass;
2684 * If we've gotten EOF or our partner has, we should not continue
2685 * queueing I/O. If we're a writer, though, we should continue
2686 * to write any buffers that don't have EOF status.
2688 if ((dev->flags & CAMDD_DEV_FLAG_EOF)
2689 || ((dev->flags & CAMDD_DEV_FLAG_PEER_EOF)
2690 && (is_write == 0))) {
2692 * Tell the worker thread that we have seen EOF.
2697 * If we're the writer, send the buffer back with EOF status.
2700 read_buf->status = CAMDD_STATUS_EOF;
2702 error = camdd_complete_peer_buf(dev, read_buf);
2707 if (is_write == 0) {
2708 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2713 data = &buf->buf_type_spec.data;
2715 retval = camdd_get_next_lba_len(dev, &buf->lba, &buf->len);
2717 buf->status = CAMDD_STATUS_EOF;
2720 && ((dev->flags & (CAMDD_DEV_FLAG_EOF_SENT |
2721 CAMDD_DEV_FLAG_EOF_QUEUED)) != 0)) {
2722 camdd_release_buf(buf);
2725 dev->flags |= CAMDD_DEV_FLAG_EOF_QUEUED;
2728 data->fill_len = buf->len;
2729 data->src_start_offset = buf->lba * dev->sector_size;
2732 * Put this on the run queue.
2734 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2735 dev->num_run_queue++;
2742 * Check for new EOF status from the reader.
2744 if ((read_buf->status == CAMDD_STATUS_EOF)
2745 || (read_buf->status == CAMDD_STATUS_ERROR)) {
2746 dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2747 if ((STAILQ_FIRST(&dev->pending_queue) == NULL)
2748 && (read_buf->len == 0)) {
2749 camdd_complete_peer_buf(dev, read_buf);
2753 eof_flush_needed = 1;
2757 * See if we have a buffer we're composing with pieces from our
2760 buf = STAILQ_FIRST(&dev->pending_queue);
2765 retval = camdd_get_next_lba_len(dev, &lba, &len);
2767 read_buf->status = CAMDD_STATUS_EOF;
2770 dev->flags |= CAMDD_DEV_FLAG_EOF;
2771 error = camdd_complete_peer_buf(dev, read_buf);
2777 * If we don't have a pending buffer, we need to grab a new
2778 * one from the free list or allocate another one.
2780 buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2789 STAILQ_INSERT_TAIL(&dev->pending_queue, buf, links);
2790 dev->num_pending_queue++;
2793 data = &buf->buf_type_spec.data;
2795 rb_data = &read_buf->buf_type_spec.data;
2797 if ((rb_data->src_start_offset != dev->next_peer_pos_bytes)
2798 && (dev->debug != 0)) {
2799 printf("%s: WARNING: reader offset %#jx != expected offset "
2800 "%#jx\n", __func__, (uintmax_t)rb_data->src_start_offset,
2801 (uintmax_t)dev->next_peer_pos_bytes);
2803 dev->next_peer_pos_bytes = rb_data->src_start_offset +
2804 (rb_data->fill_len - rb_data->resid);
2806 new_len = (rb_data->fill_len - rb_data->resid) + data->fill_len;
2807 if (new_len < buf->len) {
2809 * There are three cases here:
2810 * 1. We need more data to fill up a block, so we put
2811 * this I/O on the queue and wait for more I/O.
2812 * 2. We have a pending buffer in the queue that is
2813 * smaller than our blocksize, but we got an EOF. So we
2814 * need to go ahead and flush the write out.
2815 * 3. We got an error.
2819 * Increment our fill length.
2821 data->fill_len += (rb_data->fill_len - rb_data->resid);
2824 * Add the new read buffer to the list for writing.
2826 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2828 /* Increment the count */
2831 if (eof_flush_needed == 0) {
2833 * We need to exit, because we don't have enough
2839 * Take the buffer off of the pending queue.
2841 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2843 dev->num_pending_queue--;
2846 * If we need an EOF flush, but there is no data
2847 * to flush, go ahead and return this buffer.
2849 if (data->fill_len == 0) {
2850 camdd_complete_buf(dev, buf, /*error_count*/0);
2856 * Put this on the next queue for execution.
2858 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2859 dev->num_run_queue++;
2861 } else if (new_len == buf->len) {
2863 * We have enough data to completey fill one block,
2864 * so we're ready to issue the I/O.
2868 * Take the buffer off of the pending queue.
2870 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf, links);
2871 dev->num_pending_queue--;
2874 * Add the new read buffer to the list for writing.
2876 STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2878 /* Increment the count */
2882 * Increment our fill length.
2884 data->fill_len += (rb_data->fill_len - rb_data->resid);
2887 * Put this on the next queue for execution.
2889 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2890 dev->num_run_queue++;
2892 struct camdd_buf *idb;
2893 struct camdd_buf_indirect *indirect;
2894 uint32_t len_to_go, cur_offset;
2897 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2902 indirect = &idb->buf_type_spec.indirect;
2903 indirect->src_buf = read_buf;
2904 read_buf->refcount++;
2905 indirect->offset = 0;
2906 indirect->start_ptr = rb_data->buf;
2908 * We've already established that there is more
2909 * data in read_buf than we have room for in our
2910 * current write request. So this particular chunk
2911 * of the request should just be the remainder
2912 * needed to fill up a block.
2914 indirect->len = buf->len - (data->fill_len - data->resid);
2916 camdd_buf_add_child(buf, idb);
2919 * This buffer is ready to execute, so we can take
2920 * it off the pending queue and put it on the run
2923 STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2925 dev->num_pending_queue--;
2926 STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2927 dev->num_run_queue++;
2929 cur_offset = indirect->offset + indirect->len;
2932 * The resulting I/O would be too large to fit in
2933 * one block. We need to split this I/O into
2934 * multiple pieces. Allocate as many buffers as needed.
2936 for (len_to_go = rb_data->fill_len - rb_data->resid -
2937 indirect->len; len_to_go > 0;) {
2938 struct camdd_buf *new_buf;
2939 struct camdd_buf_data *new_data;
2943 retval = camdd_get_next_lba_len(dev, &lba, &len);
2947 * The device has already been marked
2948 * as EOF, and there is no space left.
2953 new_buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2954 if (new_buf == NULL) {
2962 idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2968 indirect = &idb->buf_type_spec.indirect;
2970 indirect->src_buf = read_buf;
2971 read_buf->refcount++;
2972 indirect->offset = cur_offset;
2973 indirect->start_ptr = rb_data->buf + cur_offset;
2974 indirect->len = min(len_to_go, new_buf->len);
2976 if (((indirect->len % dev->sector_size) != 0)
2977 || ((indirect->offset % dev->sector_size) != 0)) {
2978 warnx("offset %ju len %ju not aligned with "
2979 "sector size %u", indirect->offset,
2980 (uintmax_t)indirect->len, dev->sector_size);
2983 cur_offset += indirect->len;
2984 len_to_go -= indirect->len;
2986 camdd_buf_add_child(new_buf, idb);
2988 new_data = &new_buf->buf_type_spec.data;
2990 if ((new_data->fill_len == new_buf->len)
2991 || (eof_flush_needed != 0)) {
2992 STAILQ_INSERT_TAIL(&dev->run_queue,
2994 dev->num_run_queue++;
2995 } else if (new_data->fill_len < buf->len) {
2996 STAILQ_INSERT_TAIL(&dev->pending_queue,
2998 dev->num_pending_queue++;
3000 warnx("%s: too much data in new "
3001 "buffer!", __func__);
3013 camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
3014 uint32_t *peer_depth, uint32_t *our_bytes, uint32_t *peer_bytes)
3016 *our_depth = dev->cur_active_io + dev->num_run_queue;
3017 if (dev->num_peer_work_queue >
3018 dev->num_peer_done_queue)
3019 *peer_depth = dev->num_peer_work_queue -
3020 dev->num_peer_done_queue;
3023 *our_bytes = *our_depth * dev->blocksize;
3024 *peer_bytes = dev->peer_bytes_queued;
3028 camdd_sig_handler(int sig)
3037 sem_post(&camdd_sem);
3041 camdd_print_status(struct camdd_dev *camdd_dev, struct camdd_dev *other_dev,
3042 struct timespec *start_time)
3044 struct timespec done_time;
3046 long double mb_sec, total_sec;
3049 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &done_time);
3051 warn("Unable to get done time");
3055 timespecsub(&done_time, start_time, &done_time);
3057 total_ns = done_time.tv_nsec + (done_time.tv_sec * 1000000000);
3058 total_sec = total_ns;
3059 total_sec /= 1000000000;
3061 fprintf(stderr, "%ju bytes %s %s\n%ju bytes %s %s\n"
3062 "%.4Lf seconds elapsed\n",
3063 (uintmax_t)camdd_dev->bytes_transferred,
3064 (camdd_dev->write_dev == 0) ? "read from" : "written to",
3065 camdd_dev->device_name,
3066 (uintmax_t)other_dev->bytes_transferred,
3067 (other_dev->write_dev == 0) ? "read from" : "written to",
3068 other_dev->device_name, total_sec);
3070 mb_sec = min(other_dev->bytes_transferred,camdd_dev->bytes_transferred);
3071 mb_sec /= 1024 * 1024;
3072 mb_sec *= 1000000000;
3074 fprintf(stderr, "%.2Lf MB/sec\n", mb_sec);
3078 camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts, uint64_t max_io,
3079 int retry_count, int timeout)
3081 struct cam_device *new_cam_dev = NULL;
3082 struct camdd_dev *devs[2];
3083 struct timespec start_time;
3084 pthread_t threads[2];
3089 if (num_io_opts != 2) {
3090 warnx("Must have one input and one output path");
3095 bzero(devs, sizeof(devs));
3097 for (i = 0; i < num_io_opts; i++) {
3098 switch (io_opts[i].dev_type) {
3099 case CAMDD_DEV_PASS: {
3100 if (isdigit(io_opts[i].dev_name[0])) {
3101 camdd_argmask new_arglist = CAMDD_ARG_NONE;
3102 int bus = 0, target = 0, lun = 0;
3105 /* device specified as bus:target[:lun] */
3106 rv = parse_btl(io_opts[i].dev_name, &bus,
3107 &target, &lun, &new_arglist);
3109 warnx("numeric device specification "
3110 "must be either bus:target, or "
3115 /* default to 0 if lun was not specified */
3116 if ((new_arglist & CAMDD_ARG_LUN) == 0) {
3118 new_arglist |= CAMDD_ARG_LUN;
3120 new_cam_dev = cam_open_btl(bus, target, lun,
3125 if (cam_get_device(io_opts[i].dev_name, name,
3126 sizeof name, &unit) == -1) {
3127 warnx("%s", cam_errbuf);
3131 new_cam_dev = cam_open_spec_device(name, unit,
3135 if (new_cam_dev == NULL) {
3136 warnx("%s", cam_errbuf);
3141 devs[i] = camdd_probe_pass(new_cam_dev,
3142 /*io_opts*/ &io_opts[i],
3143 CAMDD_ARG_ERR_RECOVER,
3144 /*probe_retry_count*/ 3,
3145 /*probe_timeout*/ 5000,
3146 /*io_retry_count*/ retry_count,
3147 /*io_timeout*/ timeout);
3148 if (devs[i] == NULL) {
3149 warn("Unable to probe device %s%u",
3150 new_cam_dev->device_name,
3151 new_cam_dev->dev_unit_num);
3157 case CAMDD_DEV_FILE: {
3160 if (io_opts[i].dev_name[0] == '-') {
3161 if (io_opts[i].write_dev != 0)
3166 if (io_opts[i].write_dev != 0) {
3167 fd = open(io_opts[i].dev_name,
3168 O_RDWR | O_CREAT, S_IWUSR |S_IRUSR);
3170 fd = open(io_opts[i].dev_name,
3175 warn("error opening file %s",
3176 io_opts[i].dev_name);
3181 devs[i] = camdd_probe_file(fd, &io_opts[i],
3182 retry_count, timeout);
3183 if (devs[i] == NULL) {
3191 warnx("Unknown device type %d (%s)",
3192 io_opts[i].dev_type, io_opts[i].dev_name);
3195 break; /*NOTREACHED */
3198 devs[i]->write_dev = io_opts[i].write_dev;
3200 devs[i]->start_offset_bytes = io_opts[i].offset;
3203 devs[i]->sector_io_limit =
3204 (devs[i]->start_offset_bytes /
3205 devs[i]->sector_size) +
3206 (max_io / devs[i]->sector_size) - 1;
3209 devs[i]->next_io_pos_bytes = devs[i]->start_offset_bytes;
3210 devs[i]->next_completion_pos_bytes =devs[i]->start_offset_bytes;
3213 devs[0]->peer_dev = devs[1];
3214 devs[1]->peer_dev = devs[0];
3215 devs[0]->next_peer_pos_bytes = devs[0]->peer_dev->next_io_pos_bytes;
3216 devs[1]->next_peer_pos_bytes = devs[1]->peer_dev->next_io_pos_bytes;
3218 sem_init(&camdd_sem, /*pshared*/ 0, 0);
3220 signal(SIGINFO, camdd_sig_handler);
3221 signal(SIGINT, camdd_sig_handler);
3223 error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &start_time);
3225 warn("Unable to get start time");
3229 for (i = 0; i < num_io_opts; i++) {
3230 error = pthread_create(&threads[i], NULL, camdd_worker,
3233 warnc(error, "pthread_create() failed");
3239 if ((sem_wait(&camdd_sem) == -1)
3240 || (need_exit != 0)) {
3243 for (i = 0; i < num_io_opts; i++) {
3244 EV_SET(&ke, (uintptr_t)&devs[i]->work_queue,
3245 EVFILT_USER, 0, NOTE_TRIGGER, 0, NULL);
3247 devs[i]->flags |= CAMDD_DEV_FLAG_EOF;
3249 error = kevent(devs[i]->kq, &ke, 1, NULL, 0,
3252 warn("%s: unable to wake up thread",
3257 } else if (need_status != 0) {
3258 camdd_print_status(devs[0], devs[1], &start_time);
3262 for (i = 0; i < num_io_opts; i++) {
3263 pthread_join(threads[i], NULL);
3266 camdd_print_status(devs[0], devs[1], &start_time);
3270 for (i = 0; i < num_io_opts; i++)
3271 camdd_free_dev(devs[i]);
3273 return (error + error_exit);
3280 "usage: camdd <-i|-o pass=pass0,bs=1M,offset=1M,depth=4>\n"
3281 " <-i|-o file=/tmp/file,bs=512K,offset=1M>\n"
3282 " <-i|-o file=/dev/da0,bs=512K,offset=1M>\n"
3283 " <-i|-o file=/dev/nsa0,bs=512K>\n"
3284 " [-C retry_count][-E][-m max_io_amt][-t timeout_secs][-v][-h]\n"
3285 "Option description\n"
3286 "-i <arg=val> Specify input device/file and parameters\n"
3287 "-o <arg=val> Specify output device/file and parameters\n"
3288 "Input and Output parameters\n"
3289 "pass=name Specify a pass(4) device like pass0 or /dev/pass0\n"
3290 "file=name Specify a file or device, /tmp/foo, /dev/da0, /dev/null\n"
3291 " or - for stdin/stdout\n"
3292 "bs=blocksize Specify blocksize in bytes, or using K, M, G, etc. suffix\n"
3293 "offset=len Specify starting offset in bytes or using K, M, G suffix\n"
3294 " NOTE: offset cannot be specified on tapes, pipes, stdin/out\n"
3295 "depth=N Specify a numeric queue depth. This only applies to pass(4)\n"
3296 "mcs=N Specify a minimum cmd size for pass(4) read/write commands\n"
3297 "Optional arguments\n"
3298 "-C retry_cnt Specify a retry count for pass(4) devices\n"
3299 "-E Enable CAM error recovery for pass(4) devices\n"
3300 "-m max_io Specify the maximum amount to be transferred in bytes or\n"
3301 " using K, G, M, etc. suffixes\n"
3302 "-t timeout Specify the I/O timeout to use with pass(4) devices\n"
3303 "-v Enable verbose error recovery\n"
3304 "-h Print this message\n");
3309 camdd_parse_io_opts(char *args, int is_write, struct camdd_io_opts *io_opts)
3311 char *tmpstr, *tmpstr2;
3312 char *orig_tmpstr = NULL;
3315 io_opts->write_dev = is_write;
3317 tmpstr = strdup(args);
3318 if (tmpstr == NULL) {
3319 warn("strdup failed");
3323 orig_tmpstr = tmpstr;
3324 while ((tmpstr2 = strsep(&tmpstr, ",")) != NULL) {
3328 * If the user creates an empty parameter by putting in two
3329 * commas, skip over it and look for the next field.
3331 if (*tmpstr2 == '\0')
3334 name = strsep(&tmpstr2, "=");
3335 if (*name == '\0') {
3336 warnx("Got empty I/O parameter name");
3340 value = strsep(&tmpstr2, "=");
3342 || (*value == '\0')) {
3343 warnx("Empty I/O parameter value for %s", name);
3347 if (strncasecmp(name, "file", 4) == 0) {
3348 io_opts->dev_type = CAMDD_DEV_FILE;
3349 io_opts->dev_name = strdup(value);
3350 if (io_opts->dev_name == NULL) {
3351 warn("Error allocating memory");
3355 } else if (strncasecmp(name, "pass", 4) == 0) {
3356 io_opts->dev_type = CAMDD_DEV_PASS;
3357 io_opts->dev_name = strdup(value);
3358 if (io_opts->dev_name == NULL) {
3359 warn("Error allocating memory");
3363 } else if ((strncasecmp(name, "bs", 2) == 0)
3364 || (strncasecmp(name, "blocksize", 9) == 0)) {
3365 retval = expand_number(value, &io_opts->blocksize);
3367 warn("expand_number(3) failed on %s=%s", name,
3372 } else if (strncasecmp(name, "depth", 5) == 0) {
3375 io_opts->queue_depth = strtoull(value, &endptr, 0);
3376 if (*endptr != '\0') {
3377 warnx("invalid queue depth %s", value);
3381 } else if (strncasecmp(name, "mcs", 3) == 0) {
3384 io_opts->min_cmd_size = strtol(value, &endptr, 0);
3385 if ((*endptr != '\0')
3386 || ((io_opts->min_cmd_size > 16)
3387 || (io_opts->min_cmd_size < 0))) {
3388 warnx("invalid minimum cmd size %s", value);
3392 } else if (strncasecmp(name, "offset", 6) == 0) {
3393 retval = expand_number(value, &io_opts->offset);
3395 warn("expand_number(3) failed on %s=%s", name,
3400 } else if (strncasecmp(name, "debug", 5) == 0) {
3403 io_opts->debug = strtoull(value, &endptr, 0);
3404 if (*endptr != '\0') {
3405 warnx("invalid debug level %s", value);
3410 warnx("Unrecognized parameter %s=%s", name, value);
3420 main(int argc, char **argv)
3423 camdd_argmask arglist = CAMDD_ARG_NONE;
3424 int timeout = 0, retry_count = 1;
3426 uint64_t max_io = 0;
3427 struct camdd_io_opts *opt_list = NULL;
3434 opt_list = calloc(2, sizeof(struct camdd_io_opts));
3435 if (opt_list == NULL) {
3436 warn("Unable to allocate option list");
3441 while ((c = getopt(argc, argv, "C:Ehi:m:o:t:v")) != -1){
3444 retry_count = strtol(optarg, NULL, 0);
3445 if (retry_count < 0)
3446 errx(1, "retry count %d is < 0",
3448 arglist |= CAMDD_ARG_RETRIES;
3451 arglist |= CAMDD_ARG_ERR_RECOVER;
3456 && (opt_list[0].dev_type != CAMDD_DEV_NONE))
3458 && (opt_list[1].dev_type != CAMDD_DEV_NONE))) {
3459 errx(1, "Only one input and output path "
3462 error = camdd_parse_io_opts(optarg, (c == 'o') ? 1 : 0,
3463 (c == 'o') ? &opt_list[1] : &opt_list[0]);
3468 error = expand_number(optarg, &max_io);
3470 warn("invalid maximum I/O amount %s", optarg);
3476 timeout = strtol(optarg, NULL, 0);
3478 errx(1, "invalid timeout %d", timeout);
3479 /* Convert the timeout from seconds to ms */
3481 arglist |= CAMDD_ARG_TIMEOUT;
3484 arglist |= CAMDD_ARG_VERBOSE;
3490 break; /*NOTREACHED*/
3494 if ((opt_list[0].dev_type == CAMDD_DEV_NONE)
3495 || (opt_list[1].dev_type == CAMDD_DEV_NONE))
3496 errx(1, "Must specify both -i and -o");
3499 * Set the timeout if the user hasn't specified one.
3502 timeout = CAMDD_PASS_RW_TIMEOUT;
3504 error = camdd_rw(opt_list, 2, max_io, retry_count, timeout);