xref: /freebsd/usr.sbin/camdd/camdd.c (revision 5dae51da3da0cc94d17bd67b308fad304ebec7e0)
1 /*-
2  * Copyright (c) 1997-2007 Kenneth D. Merry
3  * Copyright (c) 2013, 2014, 2015 Spectra Logic Corporation
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
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.
17  *
18  * NO WARRANTY
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.
30  *
31  * Authors: Ken Merry           (Spectra Logic Corporation)
32  */
33 
34 /*
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.
39  */
40 #include <sys/cdefs.h>
41 __FBSDID("$FreeBSD$");
42 
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>
48 #include <sys/sbuf.h>
49 #include <sys/stat.h>
50 #include <sys/event.h>
51 #include <sys/time.h>
52 #include <sys/uio.h>
53 #include <vm/vm.h>
54 #include <machine/bus.h>
55 #include <sys/bus.h>
56 #include <sys/bus_dma.h>
57 #include <sys/mtio.h>
58 #include <sys/conf.h>
59 #include <sys/disk.h>
60 
61 #include <stdio.h>
62 #include <stdlib.h>
63 #include <semaphore.h>
64 #include <string.h>
65 #include <unistd.h>
66 #include <inttypes.h>
67 #include <limits.h>
68 #include <fcntl.h>
69 #include <ctype.h>
70 #include <err.h>
71 #include <libutil.h>
72 #include <pthread.h>
73 #include <assert.h>
74 #include <bsdxml.h>
75 
76 #include <cam/cam.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>
84 #include <camlib.h>
85 #include <mtlib.h>
86 #include <zlib.h>
87 
88 typedef enum {
89 	CAMDD_CMD_NONE		= 0x00000000,
90 	CAMDD_CMD_HELP		= 0x00000001,
91 	CAMDD_CMD_WRITE		= 0x00000002,
92 	CAMDD_CMD_READ		= 0x00000003
93 } camdd_cmdmask;
94 
95 typedef enum {
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
106 } camdd_argmask;
107 
108 typedef enum {
109 	CAMDD_DEV_NONE		= 0x00,
110 	CAMDD_DEV_PASS		= 0x01,
111 	CAMDD_DEV_FILE		= 0x02
112 } camdd_dev_type;
113 
114 struct camdd_io_opts {
115 	camdd_dev_type	dev_type;
116 	char		*dev_name;
117 	uint64_t	blocksize;
118 	uint64_t	queue_depth;
119 	uint64_t	offset;
120 	int		min_cmd_size;
121 	int		write_dev;
122 	uint64_t	debug;
123 };
124 
125 typedef enum {
126 	CAMDD_BUF_NONE,
127 	CAMDD_BUF_DATA,
128 	CAMDD_BUF_INDIRECT
129 } camdd_buf_type;
130 
131 struct camdd_buf_indirect {
132 	/*
133 	 * Pointer to the source buffer.
134 	 */
135 	struct camdd_buf *src_buf;
136 
137 	/*
138 	 * Offset into the source buffer, in bytes.
139 	 */
140 	uint64_t	  offset;
141 	/*
142 	 * Pointer to the starting point in the source buffer.
143 	 */
144 	uint8_t		 *start_ptr;
145 
146 	/*
147 	 * Length of this chunk in bytes.
148 	 */
149 	size_t		  len;
150 };
151 
152 struct camdd_buf_data {
153 	/*
154 	 * Buffer allocated when we allocate this camdd_buf.  This should
155 	 * be the size of the blocksize for this device.
156 	 */
157 	uint8_t			*buf;
158 
159 	/*
160 	 * The amount of backing store allocated in buf.  Generally this
161 	 * will be the blocksize of the device.
162 	 */
163 	uint32_t		 alloc_len;
164 
165 	/*
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
168 	 * writes).
169 	 */
170 	uint32_t		 fill_len;
171 
172 	/*
173 	 * The amount of data that was not transferred.
174 	 */
175 	uint32_t		 resid;
176 
177 	/*
178 	 * Starting byte offset on the reader.
179 	 */
180 	uint64_t		 src_start_offset;
181 
182 	/*
183 	 * CCB used for pass(4) device targets.
184 	 */
185 	union ccb		 ccb;
186 
187 	/*
188 	 * Number of scatter/gather segments.
189 	 */
190 	int			 sg_count;
191 
192 	/*
193 	 * Set if we had to tack on an extra buffer to round the transfer
194 	 * up to a sector size.
195 	 */
196 	int			 extra_buf;
197 
198 	/*
199 	 * Scatter/gather list used generally when we're the writer for a
200 	 * pass(4) device.
201 	 */
202 	bus_dma_segment_t	*segs;
203 
204 	/*
205 	 * Scatter/gather list used generally when we're the writer for a
206 	 * file or block device;
207 	 */
208 	struct iovec		*iovec;
209 };
210 
211 union camdd_buf_types {
212 	struct camdd_buf_indirect	indirect;
213 	struct camdd_buf_data		data;
214 };
215 
216 typedef enum {
217 	CAMDD_STATUS_NONE,
218 	CAMDD_STATUS_OK,
219 	CAMDD_STATUS_SHORT_IO,
220 	CAMDD_STATUS_EOF,
221 	CAMDD_STATUS_ERROR
222 } camdd_buf_status;
223 
224 struct camdd_buf {
225 	camdd_buf_type		 buf_type;
226 	union camdd_buf_types	 buf_type_spec;
227 
228 	camdd_buf_status	 status;
229 
230 	uint64_t		 lba;
231 	size_t			 len;
232 
233 	/*
234 	 * A reference count of how many indirect buffers point to this
235 	 * buffer.
236 	 */
237 	int			 refcount;
238 
239 	/*
240 	 * A link back to our parent device.
241 	 */
242 	struct camdd_dev	*dev;
243 	STAILQ_ENTRY(camdd_buf)  links;
244 	STAILQ_ENTRY(camdd_buf)  work_links;
245 
246 	/*
247 	 * A count of the buffers on the src_list.
248 	 */
249 	int			 src_count;
250 
251 	/*
252 	 * List of buffers from our partner thread that are the components
253 	 * of this buffer for the I/O.  Uses src_links.
254 	 */
255 	STAILQ_HEAD(,camdd_buf)	 src_list;
256 	STAILQ_ENTRY(camdd_buf)  src_links;
257 };
258 
259 #define	NUM_DEV_TYPES	2
260 
261 struct camdd_dev_pass {
262 	int			 scsi_dev_type;
263 	struct cam_device	*dev;
264 	uint64_t		 max_sector;
265 	uint32_t		 block_len;
266 	uint32_t		 cpi_maxio;
267 };
268 
269 typedef enum {
270 	CAMDD_FILE_NONE,
271 	CAMDD_FILE_REG,
272 	CAMDD_FILE_STD,
273 	CAMDD_FILE_PIPE,
274 	CAMDD_FILE_DISK,
275 	CAMDD_FILE_TAPE,
276 	CAMDD_FILE_TTY,
277 	CAMDD_FILE_MEM
278 } camdd_file_type;
279 
280 typedef enum {
281 	CAMDD_FF_NONE 		= 0x00,
282 	CAMDD_FF_CAN_SEEK	= 0x01
283 } camdd_file_flags;
284 
285 struct camdd_dev_file {
286 	int			 fd;
287 	struct stat		 sb;
288 	char			 filename[MAXPATHLEN + 1];
289 	camdd_file_type		 file_type;
290 	camdd_file_flags	 file_flags;
291 	uint8_t			*tmp_buf;
292 };
293 
294 struct camdd_dev_block {
295 	int			 fd;
296 	uint64_t		 size_bytes;
297 	uint32_t		 block_len;
298 };
299 
300 union camdd_dev_spec {
301 	struct camdd_dev_pass	pass;
302 	struct camdd_dev_file	file;
303 	struct camdd_dev_block	block;
304 };
305 
306 typedef enum {
307 	CAMDD_DEV_FLAG_NONE		= 0x00,
308 	CAMDD_DEV_FLAG_EOF		= 0x01,
309 	CAMDD_DEV_FLAG_PEER_EOF		= 0x02,
310 	CAMDD_DEV_FLAG_ACTIVE		= 0x04,
311 	CAMDD_DEV_FLAG_EOF_SENT		= 0x08,
312 	CAMDD_DEV_FLAG_EOF_QUEUED	= 0x10
313 } camdd_dev_flags;
314 
315 struct camdd_dev {
316 	camdd_dev_type		 dev_type;
317 	union camdd_dev_spec	 dev_spec;
318 	camdd_dev_flags		 flags;
319 	char			 device_name[MAXPATHLEN+1];
320 	uint32_t		 blocksize;
321 	uint32_t		 sector_size;
322 	uint64_t		 max_sector;
323 	uint64_t		 sector_io_limit;
324 	int			 min_cmd_size;
325 	int			 write_dev;
326 	int			 retry_count;
327 	int			 io_timeout;
328 	int			 debug;
329 	uint64_t		 start_offset_bytes;
330 	uint64_t		 next_io_pos_bytes;
331 	uint64_t		 next_peer_pos_bytes;
332 	uint64_t		 next_completion_pos_bytes;
333 	uint64_t		 peer_bytes_queued;
334 	uint64_t		 bytes_transferred;
335 	uint32_t		 target_queue_depth;
336 	uint32_t		 cur_active_io;
337 	uint8_t			*extra_buf;
338 	uint32_t		 extra_buf_len;
339 	struct camdd_dev	*peer_dev;
340 	pthread_mutex_t		 mutex;
341 	pthread_cond_t		 cond;
342 	int			 kq;
343 
344 	int			 (*run)(struct camdd_dev *dev);
345 	int			 (*fetch)(struct camdd_dev *dev);
346 
347 	/*
348 	 * Buffers that are available for I/O.  Uses links.
349 	 */
350 	STAILQ_HEAD(,camdd_buf)	 free_queue;
351 
352 	/*
353 	 * Free indirect buffers.  These are used for breaking a large
354 	 * buffer into multiple pieces.
355 	 */
356 	STAILQ_HEAD(,camdd_buf)	 free_indirect_queue;
357 
358 	/*
359 	 * Buffers that have been queued to the kernel.  Uses links.
360 	 */
361 	STAILQ_HEAD(,camdd_buf)	 active_queue;
362 
363 	/*
364 	 * Will generally contain one of our buffers that is waiting for enough
365 	 * I/O from our partner thread to be able to execute.  This will
366 	 * generally happen when our per-I/O-size is larger than the
367 	 * partner thread's per-I/O-size.  Uses links.
368 	 */
369 	STAILQ_HEAD(,camdd_buf)	 pending_queue;
370 
371 	/*
372 	 * Number of buffers on the pending queue
373 	 */
374 	int			 num_pending_queue;
375 
376 	/*
377 	 * Buffers that are filled and ready to execute.  This is used when
378 	 * our partner (reader) thread sends us blocks that are larger than
379 	 * our blocksize, and so we have to split them into multiple pieces.
380 	 */
381 	STAILQ_HEAD(,camdd_buf)	 run_queue;
382 
383 	/*
384 	 * Number of buffers on the run queue.
385 	 */
386 	int			 num_run_queue;
387 
388 	STAILQ_HEAD(,camdd_buf)	 reorder_queue;
389 
390 	int			 num_reorder_queue;
391 
392 	/*
393 	 * Buffers that have been queued to us by our partner thread
394 	 * (generally the reader thread) to be written out.  Uses
395 	 * work_links.
396 	 */
397 	STAILQ_HEAD(,camdd_buf)	 work_queue;
398 
399 	/*
400 	 * Buffers that have been completed by our partner thread.  Uses
401 	 * work_links.
402 	 */
403 	STAILQ_HEAD(,camdd_buf)	 peer_done_queue;
404 
405 	/*
406 	 * Number of buffers on the peer done queue.
407 	 */
408 	uint32_t		 num_peer_done_queue;
409 
410 	/*
411 	 * A list of buffers that we have queued to our peer thread.  Uses
412 	 * links.
413 	 */
414 	STAILQ_HEAD(,camdd_buf)	 peer_work_queue;
415 
416 	/*
417 	 * Number of buffers on the peer work queue.
418 	 */
419 	uint32_t		 num_peer_work_queue;
420 };
421 
422 static sem_t camdd_sem;
423 static int need_exit = 0;
424 static int error_exit = 0;
425 static int need_status = 0;
426 
427 #ifndef min
428 #define	min(a, b) (a < b) ? a : b
429 #endif
430 
431 /*
432  * XXX KDM private copy of timespecsub().  This is normally defined in
433  * sys/time.h, but is only enabled in the kernel.  If that definition is
434  * enabled in userland, it breaks the build of libnetbsd.
435  */
436 #ifndef timespecsub
437 #define	timespecsub(vvp, uvp)						\
438 	do {								\
439 		(vvp)->tv_sec -= (uvp)->tv_sec;				\
440 		(vvp)->tv_nsec -= (uvp)->tv_nsec;			\
441 		if ((vvp)->tv_nsec < 0) {				\
442 			(vvp)->tv_sec--;				\
443 			(vvp)->tv_nsec += 1000000000;			\
444 		}							\
445 	} while (0)
446 #endif
447 
448 
449 /* Generically useful offsets into the peripheral private area */
450 #define ppriv_ptr0 periph_priv.entries[0].ptr
451 #define ppriv_ptr1 periph_priv.entries[1].ptr
452 #define ppriv_field0 periph_priv.entries[0].field
453 #define ppriv_field1 periph_priv.entries[1].field
454 
455 #define	ccb_buf	ppriv_ptr0
456 
457 #define	CAMDD_FILE_DEFAULT_BLOCK	524288
458 #define	CAMDD_FILE_DEFAULT_DEPTH	1
459 #define	CAMDD_PASS_MAX_BLOCK		1048576
460 #define	CAMDD_PASS_DEFAULT_DEPTH	6
461 #define	CAMDD_PASS_RW_TIMEOUT		60 * 1000
462 
463 static int parse_btl(char *tstr, int *bus, int *target, int *lun,
464 		     camdd_argmask *arglst);
465 void camdd_free_dev(struct camdd_dev *dev);
466 struct camdd_dev *camdd_alloc_dev(camdd_dev_type dev_type,
467 				  struct kevent *new_ke, int num_ke,
468 				  int retry_count, int timeout);
469 static struct camdd_buf *camdd_alloc_buf(struct camdd_dev *dev,
470 					 camdd_buf_type buf_type);
471 void camdd_release_buf(struct camdd_buf *buf);
472 struct camdd_buf *camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type);
473 int camdd_buf_sg_create(struct camdd_buf *buf, int iovec,
474 			uint32_t sector_size, uint32_t *num_sectors_used,
475 			int *double_buf_needed);
476 uint32_t camdd_buf_get_len(struct camdd_buf *buf);
477 void camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf);
478 int camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
479 		     uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran);
480 struct camdd_dev *camdd_probe_file(int fd, struct camdd_io_opts *io_opts,
481 				   int retry_count, int timeout);
482 struct camdd_dev *camdd_probe_pass(struct cam_device *cam_dev,
483 				   struct camdd_io_opts *io_opts,
484 				   camdd_argmask arglist, int probe_retry_count,
485 				   int probe_timeout, int io_retry_count,
486 				   int io_timeout);
487 void *camdd_file_worker(void *arg);
488 camdd_buf_status camdd_ccb_status(union ccb *ccb);
489 int camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf);
490 int camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf);
491 void camdd_peer_done(struct camdd_buf *buf);
492 void camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
493 			int *error_count);
494 int camdd_pass_fetch(struct camdd_dev *dev);
495 int camdd_file_run(struct camdd_dev *dev);
496 int camdd_pass_run(struct camdd_dev *dev);
497 int camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len);
498 int camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf);
499 void camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
500 		     uint32_t *peer_depth, uint32_t *our_bytes,
501 		     uint32_t *peer_bytes);
502 void *camdd_worker(void *arg);
503 void camdd_sig_handler(int sig);
504 void camdd_print_status(struct camdd_dev *camdd_dev,
505 			struct camdd_dev *other_dev,
506 			struct timespec *start_time);
507 int camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts,
508 	     uint64_t max_io, int retry_count, int timeout);
509 int camdd_parse_io_opts(char *args, int is_write,
510 			struct camdd_io_opts *io_opts);
511 void usage(void);
512 
513 /*
514  * Parse out a bus, or a bus, target and lun in the following
515  * format:
516  * bus
517  * bus:target
518  * bus:target:lun
519  *
520  * Returns the number of parsed components, or 0.
521  */
522 static int
523 parse_btl(char *tstr, int *bus, int *target, int *lun, camdd_argmask *arglst)
524 {
525 	char *tmpstr;
526 	int convs = 0;
527 
528 	while (isspace(*tstr) && (*tstr != '\0'))
529 		tstr++;
530 
531 	tmpstr = (char *)strtok(tstr, ":");
532 	if ((tmpstr != NULL) && (*tmpstr != '\0')) {
533 		*bus = strtol(tmpstr, NULL, 0);
534 		*arglst |= CAMDD_ARG_BUS;
535 		convs++;
536 		tmpstr = (char *)strtok(NULL, ":");
537 		if ((tmpstr != NULL) && (*tmpstr != '\0')) {
538 			*target = strtol(tmpstr, NULL, 0);
539 			*arglst |= CAMDD_ARG_TARGET;
540 			convs++;
541 			tmpstr = (char *)strtok(NULL, ":");
542 			if ((tmpstr != NULL) && (*tmpstr != '\0')) {
543 				*lun = strtol(tmpstr, NULL, 0);
544 				*arglst |= CAMDD_ARG_LUN;
545 				convs++;
546 			}
547 		}
548 	}
549 
550 	return convs;
551 }
552 
553 /*
554  * XXX KDM clean up and free all of the buffers on the queue!
555  */
556 void
557 camdd_free_dev(struct camdd_dev *dev)
558 {
559 	if (dev == NULL)
560 		return;
561 
562 	switch (dev->dev_type) {
563 	case CAMDD_DEV_FILE: {
564 		struct camdd_dev_file *file_dev = &dev->dev_spec.file;
565 
566 		if (file_dev->fd != -1)
567 			close(file_dev->fd);
568 		free(file_dev->tmp_buf);
569 		break;
570 	}
571 	case CAMDD_DEV_PASS: {
572 		struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
573 
574 		if (pass_dev->dev != NULL)
575 			cam_close_device(pass_dev->dev);
576 		break;
577 	}
578 	default:
579 		break;
580 	}
581 
582 	free(dev);
583 }
584 
585 struct camdd_dev *
586 camdd_alloc_dev(camdd_dev_type dev_type, struct kevent *new_ke, int num_ke,
587 		int retry_count, int timeout)
588 {
589 	struct camdd_dev *dev = NULL;
590 	struct kevent *ke;
591 	size_t ke_size;
592 	int retval = 0;
593 
594 	dev = malloc(sizeof(*dev));
595 	if (dev == NULL) {
596 		warn("%s: unable to malloc %zu bytes", __func__, sizeof(*dev));
597 		goto bailout;
598 	}
599 
600 	bzero(dev, sizeof(*dev));
601 
602 	dev->dev_type = dev_type;
603 	dev->io_timeout = timeout;
604 	dev->retry_count = retry_count;
605 	STAILQ_INIT(&dev->free_queue);
606 	STAILQ_INIT(&dev->free_indirect_queue);
607 	STAILQ_INIT(&dev->active_queue);
608 	STAILQ_INIT(&dev->pending_queue);
609 	STAILQ_INIT(&dev->run_queue);
610 	STAILQ_INIT(&dev->reorder_queue);
611 	STAILQ_INIT(&dev->work_queue);
612 	STAILQ_INIT(&dev->peer_done_queue);
613 	STAILQ_INIT(&dev->peer_work_queue);
614 	retval = pthread_mutex_init(&dev->mutex, NULL);
615 	if (retval != 0) {
616 		warnc(retval, "%s: failed to initialize mutex", __func__);
617 		goto bailout;
618 	}
619 
620 	retval = pthread_cond_init(&dev->cond, NULL);
621 	if (retval != 0) {
622 		warnc(retval, "%s: failed to initialize condition variable",
623 		      __func__);
624 		goto bailout;
625 	}
626 
627 	dev->kq = kqueue();
628 	if (dev->kq == -1) {
629 		warn("%s: Unable to create kqueue", __func__);
630 		goto bailout;
631 	}
632 
633 	ke_size = sizeof(struct kevent) * (num_ke + 4);
634 	ke = malloc(ke_size);
635 	if (ke == NULL) {
636 		warn("%s: unable to malloc %zu bytes", __func__, ke_size);
637 		goto bailout;
638 	}
639 	bzero(ke, ke_size);
640 	if (num_ke > 0)
641 		bcopy(new_ke, ke, num_ke * sizeof(struct kevent));
642 
643 	EV_SET(&ke[num_ke++], (uintptr_t)&dev->work_queue, EVFILT_USER,
644 	       EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
645 	EV_SET(&ke[num_ke++], (uintptr_t)&dev->peer_done_queue, EVFILT_USER,
646 	       EV_ADD|EV_ENABLE|EV_CLEAR, 0,0, 0);
647 	EV_SET(&ke[num_ke++], SIGINFO, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
648 	EV_SET(&ke[num_ke++], SIGINT, EVFILT_SIGNAL, EV_ADD|EV_ENABLE, 0,0,0);
649 
650 	retval = kevent(dev->kq, ke, num_ke, NULL, 0, NULL);
651 	if (retval == -1) {
652 		warn("%s: Unable to register kevents", __func__);
653 		goto bailout;
654 	}
655 
656 
657 	return (dev);
658 
659 bailout:
660 	free(dev);
661 
662 	return (NULL);
663 }
664 
665 static struct camdd_buf *
666 camdd_alloc_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
667 {
668 	struct camdd_buf *buf = NULL;
669 	uint8_t *data_ptr = NULL;
670 
671 	/*
672 	 * We only need to allocate data space for data buffers.
673 	 */
674 	switch (buf_type) {
675 	case CAMDD_BUF_DATA:
676 		data_ptr = malloc(dev->blocksize);
677 		if (data_ptr == NULL) {
678 			warn("unable to allocate %u bytes", dev->blocksize);
679 			goto bailout_error;
680 		}
681 		break;
682 	default:
683 		break;
684 	}
685 
686 	buf = malloc(sizeof(*buf));
687 	if (buf == NULL) {
688 		warn("unable to allocate %zu bytes", sizeof(*buf));
689 		goto bailout_error;
690 	}
691 
692 	bzero(buf, sizeof(*buf));
693 	buf->buf_type = buf_type;
694 	buf->dev = dev;
695 	switch (buf_type) {
696 	case CAMDD_BUF_DATA: {
697 		struct camdd_buf_data *data;
698 
699 		data = &buf->buf_type_spec.data;
700 
701 		data->alloc_len = dev->blocksize;
702 		data->buf = data_ptr;
703 		break;
704 	}
705 	case CAMDD_BUF_INDIRECT:
706 		break;
707 	default:
708 		break;
709 	}
710 	STAILQ_INIT(&buf->src_list);
711 
712 	return (buf);
713 
714 bailout_error:
715 	if (data_ptr != NULL)
716 		free(data_ptr);
717 
718 	if (buf != NULL)
719 		free(buf);
720 
721 	return (NULL);
722 }
723 
724 void
725 camdd_release_buf(struct camdd_buf *buf)
726 {
727 	struct camdd_dev *dev;
728 
729 	dev = buf->dev;
730 
731 	switch (buf->buf_type) {
732 	case CAMDD_BUF_DATA: {
733 		struct camdd_buf_data *data;
734 
735 		data = &buf->buf_type_spec.data;
736 
737 		if (data->segs != NULL) {
738 			if (data->extra_buf != 0) {
739 				void *extra_buf;
740 
741 				extra_buf = (void *)
742 				    data->segs[data->sg_count - 1].ds_addr;
743 				free(extra_buf);
744 				data->extra_buf = 0;
745 			}
746 			free(data->segs);
747 			data->segs = NULL;
748 			data->sg_count = 0;
749 		} else if (data->iovec != NULL) {
750 			if (data->extra_buf != 0) {
751 				free(data->iovec[data->sg_count - 1].iov_base);
752 				data->extra_buf = 0;
753 			}
754 			free(data->iovec);
755 			data->iovec = NULL;
756 			data->sg_count = 0;
757 		}
758 		STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
759 		break;
760 	}
761 	case CAMDD_BUF_INDIRECT:
762 		STAILQ_INSERT_TAIL(&dev->free_indirect_queue, buf, links);
763 		break;
764 	default:
765 		err(1, "%s: Invalid buffer type %d for released buffer",
766 		    __func__, buf->buf_type);
767 		break;
768 	}
769 }
770 
771 struct camdd_buf *
772 camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
773 {
774 	struct camdd_buf *buf = NULL;
775 
776 	switch (buf_type) {
777 	case CAMDD_BUF_DATA:
778 		buf = STAILQ_FIRST(&dev->free_queue);
779 		if (buf != NULL) {
780 			struct camdd_buf_data *data;
781 			uint8_t *data_ptr;
782 			uint32_t alloc_len;
783 
784 			STAILQ_REMOVE_HEAD(&dev->free_queue, links);
785 			data = &buf->buf_type_spec.data;
786 			data_ptr = data->buf;
787 			alloc_len = data->alloc_len;
788 			bzero(buf, sizeof(*buf));
789 			data->buf = data_ptr;
790 			data->alloc_len = alloc_len;
791 		}
792 		break;
793 	case CAMDD_BUF_INDIRECT:
794 		buf = STAILQ_FIRST(&dev->free_indirect_queue);
795 		if (buf != NULL) {
796 			STAILQ_REMOVE_HEAD(&dev->free_indirect_queue, links);
797 
798 			bzero(buf, sizeof(*buf));
799 		}
800 		break;
801 	default:
802 		warnx("Unknown buffer type %d requested", buf_type);
803 		break;
804 	}
805 
806 
807 	if (buf == NULL)
808 		return (camdd_alloc_buf(dev, buf_type));
809 	else {
810 		STAILQ_INIT(&buf->src_list);
811 		buf->dev = dev;
812 		buf->buf_type = buf_type;
813 
814 		return (buf);
815 	}
816 }
817 
818 int
819 camdd_buf_sg_create(struct camdd_buf *buf, int iovec, uint32_t sector_size,
820 		    uint32_t *num_sectors_used, int *double_buf_needed)
821 {
822 	struct camdd_buf *tmp_buf;
823 	struct camdd_buf_data *data;
824 	uint8_t *extra_buf = NULL;
825 	size_t extra_buf_len = 0;
826 	int i, retval = 0;
827 
828 	data = &buf->buf_type_spec.data;
829 
830 	data->sg_count = buf->src_count;
831 	/*
832 	 * Compose a scatter/gather list from all of the buffers in the list.
833 	 * If the length of the buffer isn't a multiple of the sector size,
834 	 * we'll have to add an extra buffer.  This should only happen
835 	 * at the end of a transfer.
836 	 */
837 	if ((data->fill_len % sector_size) != 0) {
838 		extra_buf_len = sector_size - (data->fill_len % sector_size);
839 		extra_buf = calloc(extra_buf_len, 1);
840 		if (extra_buf == NULL) {
841 			warn("%s: unable to allocate %zu bytes for extra "
842 			    "buffer space", __func__, extra_buf_len);
843 			retval = 1;
844 			goto bailout;
845 		}
846 		data->extra_buf = 1;
847 		data->sg_count++;
848 	}
849 	if (iovec == 0) {
850 		data->segs = calloc(data->sg_count, sizeof(bus_dma_segment_t));
851 		if (data->segs == NULL) {
852 			warn("%s: unable to allocate %zu bytes for S/G list",
853 			    __func__, sizeof(bus_dma_segment_t) *
854 			    data->sg_count);
855 			retval = 1;
856 			goto bailout;
857 		}
858 
859 	} else {
860 		data->iovec = calloc(data->sg_count, sizeof(struct iovec));
861 		if (data->iovec == NULL) {
862 			warn("%s: unable to allocate %zu bytes for S/G list",
863 			    __func__, sizeof(struct iovec) * data->sg_count);
864 			retval = 1;
865 			goto bailout;
866 		}
867 	}
868 
869 	for (i = 0, tmp_buf = STAILQ_FIRST(&buf->src_list);
870 	     i < buf->src_count && tmp_buf != NULL; i++,
871 	     tmp_buf = STAILQ_NEXT(tmp_buf, src_links)) {
872 
873 		if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
874 			struct camdd_buf_data *tmp_data;
875 
876 			tmp_data = &tmp_buf->buf_type_spec.data;
877 			if (iovec == 0) {
878 				data->segs[i].ds_addr =
879 				    (bus_addr_t) tmp_data->buf;
880 				data->segs[i].ds_len = tmp_data->fill_len -
881 				    tmp_data->resid;
882 			} else {
883 				data->iovec[i].iov_base = tmp_data->buf;
884 				data->iovec[i].iov_len = tmp_data->fill_len -
885 				    tmp_data->resid;
886 			}
887 			if (((tmp_data->fill_len - tmp_data->resid) %
888 			     sector_size) != 0)
889 				*double_buf_needed = 1;
890 		} else {
891 			struct camdd_buf_indirect *tmp_ind;
892 
893 			tmp_ind = &tmp_buf->buf_type_spec.indirect;
894 			if (iovec == 0) {
895 				data->segs[i].ds_addr =
896 				    (bus_addr_t)tmp_ind->start_ptr;
897 				data->segs[i].ds_len = tmp_ind->len;
898 			} else {
899 				data->iovec[i].iov_base = tmp_ind->start_ptr;
900 				data->iovec[i].iov_len = tmp_ind->len;
901 			}
902 			if ((tmp_ind->len % sector_size) != 0)
903 				*double_buf_needed = 1;
904 		}
905 	}
906 
907 	if (extra_buf != NULL) {
908 		if (iovec == 0) {
909 			data->segs[i].ds_addr = (bus_addr_t)extra_buf;
910 			data->segs[i].ds_len = extra_buf_len;
911 		} else {
912 			data->iovec[i].iov_base = extra_buf;
913 			data->iovec[i].iov_len = extra_buf_len;
914 		}
915 		i++;
916 	}
917 	if ((tmp_buf != NULL) || (i != data->sg_count)) {
918 		warnx("buffer source count does not match "
919 		      "number of buffers in list!");
920 		retval = 1;
921 		goto bailout;
922 	}
923 
924 bailout:
925 	if (retval == 0) {
926 		*num_sectors_used = (data->fill_len + extra_buf_len) /
927 		    sector_size;
928 	}
929 	return (retval);
930 }
931 
932 uint32_t
933 camdd_buf_get_len(struct camdd_buf *buf)
934 {
935 	uint32_t len = 0;
936 
937 	if (buf->buf_type != CAMDD_BUF_DATA) {
938 		struct camdd_buf_indirect *indirect;
939 
940 		indirect = &buf->buf_type_spec.indirect;
941 		len = indirect->len;
942 	} else {
943 		struct camdd_buf_data *data;
944 
945 		data = &buf->buf_type_spec.data;
946 		len = data->fill_len;
947 	}
948 
949 	return (len);
950 }
951 
952 void
953 camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf)
954 {
955 	struct camdd_buf_data *data;
956 
957 	assert(buf->buf_type == CAMDD_BUF_DATA);
958 
959 	data = &buf->buf_type_spec.data;
960 
961 	STAILQ_INSERT_TAIL(&buf->src_list, child_buf, src_links);
962 	buf->src_count++;
963 
964 	data->fill_len += camdd_buf_get_len(child_buf);
965 }
966 
967 typedef enum {
968 	CAMDD_TS_MAX_BLK,
969 	CAMDD_TS_MIN_BLK,
970 	CAMDD_TS_BLK_GRAN,
971 	CAMDD_TS_EFF_IOSIZE
972 } camdd_status_item_index;
973 
974 static struct camdd_status_items {
975 	const char *name;
976 	struct mt_status_entry *entry;
977 } req_status_items[] = {
978 	{ "max_blk", NULL },
979 	{ "min_blk", NULL },
980 	{ "blk_gran", NULL },
981 	{ "max_effective_iosize", NULL }
982 };
983 
984 int
985 camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
986 		 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran)
987 {
988 	struct mt_status_data status_data;
989 	char *xml_str = NULL;
990 	unsigned int i;
991 	int retval = 0;
992 
993 	retval = mt_get_xml_str(fd, MTIOCEXTGET, &xml_str);
994 	if (retval != 0)
995 		err(1, "Couldn't get XML string from %s", filename);
996 
997 	retval = mt_get_status(xml_str, &status_data);
998 	if (retval != XML_STATUS_OK) {
999 		warn("couldn't get status for %s", filename);
1000 		retval = 1;
1001 		goto bailout;
1002 	} else
1003 		retval = 0;
1004 
1005 	if (status_data.error != 0) {
1006 		warnx("%s", status_data.error_str);
1007 		retval = 1;
1008 		goto bailout;
1009 	}
1010 
1011 	for (i = 0; i < sizeof(req_status_items) /
1012 	     sizeof(req_status_items[0]); i++) {
1013                 char *name;
1014 
1015 		name = __DECONST(char *, req_status_items[i].name);
1016 		req_status_items[i].entry = mt_status_entry_find(&status_data,
1017 		    name);
1018 		if (req_status_items[i].entry == NULL) {
1019 			errx(1, "Cannot find status entry %s",
1020 			    req_status_items[i].name);
1021 		}
1022 	}
1023 
1024 	*max_iosize = req_status_items[CAMDD_TS_EFF_IOSIZE].entry->value_unsigned;
1025 	*max_blk= req_status_items[CAMDD_TS_MAX_BLK].entry->value_unsigned;
1026 	*min_blk= req_status_items[CAMDD_TS_MIN_BLK].entry->value_unsigned;
1027 	*blk_gran = req_status_items[CAMDD_TS_BLK_GRAN].entry->value_unsigned;
1028 bailout:
1029 
1030 	free(xml_str);
1031 	mt_status_free(&status_data);
1032 
1033 	return (retval);
1034 }
1035 
1036 struct camdd_dev *
1037 camdd_probe_file(int fd, struct camdd_io_opts *io_opts, int retry_count,
1038     int timeout)
1039 {
1040 	struct camdd_dev *dev = NULL;
1041 	struct camdd_dev_file *file_dev;
1042 	uint64_t blocksize = io_opts->blocksize;
1043 
1044 	dev = camdd_alloc_dev(CAMDD_DEV_FILE, NULL, 0, retry_count, timeout);
1045 	if (dev == NULL)
1046 		goto bailout;
1047 
1048 	file_dev = &dev->dev_spec.file;
1049 	file_dev->fd = fd;
1050 	strlcpy(file_dev->filename, io_opts->dev_name,
1051 	    sizeof(file_dev->filename));
1052 	strlcpy(dev->device_name, io_opts->dev_name, sizeof(dev->device_name));
1053 	if (blocksize == 0)
1054 		dev->blocksize = CAMDD_FILE_DEFAULT_BLOCK;
1055 	else
1056 		dev->blocksize = blocksize;
1057 
1058 	if ((io_opts->queue_depth != 0)
1059 	 && (io_opts->queue_depth != 1)) {
1060 		warnx("Queue depth %ju for %s ignored, only 1 outstanding "
1061 		    "command supported", (uintmax_t)io_opts->queue_depth,
1062 		    io_opts->dev_name);
1063 	}
1064 	dev->target_queue_depth = CAMDD_FILE_DEFAULT_DEPTH;
1065 	dev->run = camdd_file_run;
1066 	dev->fetch = NULL;
1067 
1068 	/*
1069 	 * We can effectively access files on byte boundaries.  We'll reset
1070 	 * this for devices like disks that can be accessed on sector
1071 	 * boundaries.
1072 	 */
1073 	dev->sector_size = 1;
1074 
1075 	if ((fd != STDIN_FILENO)
1076 	 && (fd != STDOUT_FILENO)) {
1077 		int retval;
1078 
1079 		retval = fstat(fd, &file_dev->sb);
1080 		if (retval != 0) {
1081 			warn("Cannot stat %s", dev->device_name);
1082 			goto bailout_error;
1083 		}
1084 		if (S_ISREG(file_dev->sb.st_mode)) {
1085 			file_dev->file_type = CAMDD_FILE_REG;
1086 		} else if (S_ISCHR(file_dev->sb.st_mode)) {
1087 			int type;
1088 
1089 			if (ioctl(fd, FIODTYPE, &type) == -1)
1090 				err(1, "FIODTYPE ioctl failed on %s",
1091 				    dev->device_name);
1092 			else {
1093 				if (type & D_TAPE)
1094 					file_dev->file_type = CAMDD_FILE_TAPE;
1095 				else if (type & D_DISK)
1096 					file_dev->file_type = CAMDD_FILE_DISK;
1097 				else if (type & D_MEM)
1098 					file_dev->file_type = CAMDD_FILE_MEM;
1099 				else if (type & D_TTY)
1100 					file_dev->file_type = CAMDD_FILE_TTY;
1101 			}
1102 		} else if (S_ISDIR(file_dev->sb.st_mode)) {
1103 			errx(1, "cannot operate on directory %s",
1104 			    dev->device_name);
1105 		} else if (S_ISFIFO(file_dev->sb.st_mode)) {
1106 			file_dev->file_type = CAMDD_FILE_PIPE;
1107 		} else
1108 			errx(1, "Cannot determine file type for %s",
1109 			    dev->device_name);
1110 
1111 		switch (file_dev->file_type) {
1112 		case CAMDD_FILE_REG:
1113 			if (file_dev->sb.st_size != 0)
1114 				dev->max_sector = file_dev->sb.st_size - 1;
1115 			else
1116 				dev->max_sector = 0;
1117 			file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1118 			break;
1119 		case CAMDD_FILE_TAPE: {
1120 			uint64_t max_iosize, max_blk, min_blk, blk_gran;
1121 			/*
1122 			 * Check block limits and maximum effective iosize.
1123 			 * Make sure the blocksize is within the block
1124 			 * limits (and a multiple of the minimum blocksize)
1125 			 * and that the blocksize is <= maximum effective
1126 			 * iosize.
1127 			 */
1128 			retval = camdd_probe_tape(fd, dev->device_name,
1129 			    &max_iosize, &max_blk, &min_blk, &blk_gran);
1130 			if (retval != 0)
1131 				errx(1, "Unable to probe tape %s",
1132 				    dev->device_name);
1133 
1134 			/*
1135 			 * The blocksize needs to be <= the maximum
1136 			 * effective I/O size of the tape device.  Note
1137 			 * that this also takes into account the maximum
1138 			 * blocksize reported by READ BLOCK LIMITS.
1139 			 */
1140 			if (dev->blocksize > max_iosize) {
1141 				warnx("Blocksize %u too big for %s, limiting "
1142 				    "to %ju", dev->blocksize, dev->device_name,
1143 				    max_iosize);
1144 				dev->blocksize = max_iosize;
1145 			}
1146 
1147 			/*
1148 			 * The blocksize needs to be at least min_blk;
1149 			 */
1150 			if (dev->blocksize < min_blk) {
1151 				warnx("Blocksize %u too small for %s, "
1152 				    "increasing to %ju", dev->blocksize,
1153 				    dev->device_name, min_blk);
1154 				dev->blocksize = min_blk;
1155 			}
1156 
1157 			/*
1158 			 * And the blocksize needs to be a multiple of
1159 			 * the block granularity.
1160 			 */
1161 			if ((blk_gran != 0)
1162 			 && (dev->blocksize % (1 << blk_gran))) {
1163 				warnx("Blocksize %u for %s not a multiple of "
1164 				    "%d, adjusting to %d", dev->blocksize,
1165 				    dev->device_name, (1 << blk_gran),
1166 				    dev->blocksize & ~((1 << blk_gran) - 1));
1167 				dev->blocksize &= ~((1 << blk_gran) - 1);
1168 			}
1169 
1170 			if (dev->blocksize == 0) {
1171 				errx(1, "Unable to derive valid blocksize for "
1172 				    "%s", dev->device_name);
1173 			}
1174 
1175 			/*
1176 			 * For tape drives, set the sector size to the
1177 			 * blocksize so that we make sure not to write
1178 			 * less than the blocksize out to the drive.
1179 			 */
1180 			dev->sector_size = dev->blocksize;
1181 			break;
1182 		}
1183 		case CAMDD_FILE_DISK: {
1184 			off_t media_size;
1185 			unsigned int sector_size;
1186 
1187 			file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1188 
1189 			if (ioctl(fd, DIOCGSECTORSIZE, &sector_size) == -1) {
1190 				err(1, "DIOCGSECTORSIZE ioctl failed on %s",
1191 				    dev->device_name);
1192 			}
1193 
1194 			if (sector_size == 0) {
1195 				errx(1, "DIOCGSECTORSIZE ioctl returned "
1196 				    "invalid sector size %u for %s",
1197 				    sector_size, dev->device_name);
1198 			}
1199 
1200 			if (ioctl(fd, DIOCGMEDIASIZE, &media_size) == -1) {
1201 				err(1, "DIOCGMEDIASIZE ioctl failed on %s",
1202 				    dev->device_name);
1203 			}
1204 
1205 			if (media_size == 0) {
1206 				errx(1, "DIOCGMEDIASIZE ioctl returned "
1207 				    "invalid media size %ju for %s",
1208 				    (uintmax_t)media_size, dev->device_name);
1209 			}
1210 
1211 			if (dev->blocksize % sector_size) {
1212 				errx(1, "%s blocksize %u not a multiple of "
1213 				    "sector size %u", dev->device_name,
1214 				    dev->blocksize, sector_size);
1215 			}
1216 
1217 			dev->sector_size = sector_size;
1218 			dev->max_sector = (media_size / sector_size) - 1;
1219 			break;
1220 		}
1221 		case CAMDD_FILE_MEM:
1222 			file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1223 			break;
1224 		default:
1225 			break;
1226 		}
1227 	}
1228 
1229 	if ((io_opts->offset != 0)
1230 	 && ((file_dev->file_flags & CAMDD_FF_CAN_SEEK) == 0)) {
1231 		warnx("Offset %ju specified for %s, but we cannot seek on %s",
1232 		    io_opts->offset, io_opts->dev_name, io_opts->dev_name);
1233 		goto bailout_error;
1234 	}
1235 #if 0
1236 	else if ((io_opts->offset != 0)
1237 		&& ((io_opts->offset % dev->sector_size) != 0)) {
1238 		warnx("Offset %ju for %s is not a multiple of the "
1239 		      "sector size %u", io_opts->offset,
1240 		      io_opts->dev_name, dev->sector_size);
1241 		goto bailout_error;
1242 	} else {
1243 		dev->start_offset_bytes = io_opts->offset;
1244 	}
1245 #endif
1246 
1247 bailout:
1248 	return (dev);
1249 
1250 bailout_error:
1251 	camdd_free_dev(dev);
1252 	return (NULL);
1253 }
1254 
1255 /*
1256  * Need to implement this.  Do a basic probe:
1257  * - Check the inquiry data, make sure we're talking to a device that we
1258  *   can reasonably expect to talk to -- direct, RBC, CD, WORM.
1259  * - Send a test unit ready, make sure the device is available.
1260  * - Get the capacity and block size.
1261  */
1262 struct camdd_dev *
1263 camdd_probe_pass(struct cam_device *cam_dev, struct camdd_io_opts *io_opts,
1264 		 camdd_argmask arglist, int probe_retry_count,
1265 		 int probe_timeout, int io_retry_count, int io_timeout)
1266 {
1267 	union ccb *ccb;
1268 	uint64_t maxsector;
1269 	uint32_t cpi_maxio, max_iosize, pass_numblocks;
1270 	uint32_t block_len;
1271 	struct scsi_read_capacity_data rcap;
1272 	struct scsi_read_capacity_data_long rcaplong;
1273 	struct camdd_dev *dev;
1274 	struct camdd_dev_pass *pass_dev;
1275 	struct kevent ke;
1276 	int scsi_dev_type;
1277 
1278 	dev = NULL;
1279 
1280 	scsi_dev_type = SID_TYPE(&cam_dev->inq_data);
1281 	maxsector = 0;
1282 	block_len = 0;
1283 
1284 	/*
1285 	 * For devices that support READ CAPACITY, we'll attempt to get the
1286 	 * capacity.  Otherwise, we really don't support tape or other
1287 	 * devices via SCSI passthrough, so just return an error in that case.
1288 	 */
1289 	switch (scsi_dev_type) {
1290 	case T_DIRECT:
1291 	case T_WORM:
1292 	case T_CDROM:
1293 	case T_OPTICAL:
1294 	case T_RBC:
1295 	case T_ZBC_HM:
1296 		break;
1297 	default:
1298 		errx(1, "Unsupported SCSI device type %d", scsi_dev_type);
1299 		break; /*NOTREACHED*/
1300 	}
1301 
1302 	ccb = cam_getccb(cam_dev);
1303 
1304 	if (ccb == NULL) {
1305 		warnx("%s: error allocating ccb", __func__);
1306 		goto bailout;
1307 	}
1308 
1309 	CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
1310 
1311 	scsi_read_capacity(&ccb->csio,
1312 			   /*retries*/ probe_retry_count,
1313 			   /*cbfcnp*/ NULL,
1314 			   /*tag_action*/ MSG_SIMPLE_Q_TAG,
1315 			   &rcap,
1316 			   SSD_FULL_SIZE,
1317 			   /*timeout*/ probe_timeout ? probe_timeout : 5000);
1318 
1319 	/* Disable freezing the device queue */
1320 	ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1321 
1322 	if (arglist & CAMDD_ARG_ERR_RECOVER)
1323 		ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1324 
1325 	if (cam_send_ccb(cam_dev, ccb) < 0) {
1326 		warn("error sending READ CAPACITY command");
1327 
1328 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1329 				CAM_EPF_ALL, stderr);
1330 
1331 		goto bailout;
1332 	}
1333 
1334 	if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1335 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1336 		goto bailout;
1337 	}
1338 
1339 	maxsector = scsi_4btoul(rcap.addr);
1340 	block_len = scsi_4btoul(rcap.length);
1341 
1342 	/*
1343 	 * A last block of 2^32-1 means that the true capacity is over 2TB,
1344 	 * and we need to issue the long READ CAPACITY to get the real
1345 	 * capacity.  Otherwise, we're all set.
1346 	 */
1347 	if (maxsector != 0xffffffff)
1348 		goto rcap_done;
1349 
1350 	scsi_read_capacity_16(&ccb->csio,
1351 			      /*retries*/ probe_retry_count,
1352 			      /*cbfcnp*/ NULL,
1353 			      /*tag_action*/ MSG_SIMPLE_Q_TAG,
1354 			      /*lba*/ 0,
1355 			      /*reladdr*/ 0,
1356 			      /*pmi*/ 0,
1357 			      (uint8_t *)&rcaplong,
1358 			      sizeof(rcaplong),
1359 			      /*sense_len*/ SSD_FULL_SIZE,
1360 			      /*timeout*/ probe_timeout ? probe_timeout : 5000);
1361 
1362 	/* Disable freezing the device queue */
1363 	ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1364 
1365 	if (arglist & CAMDD_ARG_ERR_RECOVER)
1366 		ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1367 
1368 	if (cam_send_ccb(cam_dev, ccb) < 0) {
1369 		warn("error sending READ CAPACITY (16) command");
1370 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1371 				CAM_EPF_ALL, stderr);
1372 		goto bailout;
1373 	}
1374 
1375 	if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1376 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1377 		goto bailout;
1378 	}
1379 
1380 	maxsector = scsi_8btou64(rcaplong.addr);
1381 	block_len = scsi_4btoul(rcaplong.length);
1382 
1383 rcap_done:
1384 	if (block_len == 0) {
1385 		warnx("Sector size for %s%u is 0, cannot continue",
1386 		    cam_dev->device_name, cam_dev->dev_unit_num);
1387 		goto bailout_error;
1388 	}
1389 
1390 	CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cpi);
1391 
1392 	ccb->ccb_h.func_code = XPT_PATH_INQ;
1393 	ccb->ccb_h.flags = CAM_DIR_NONE;
1394 	ccb->ccb_h.retry_count = 1;
1395 
1396 	if (cam_send_ccb(cam_dev, ccb) < 0) {
1397 		warn("error sending XPT_PATH_INQ CCB");
1398 
1399 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1400 				CAM_EPF_ALL, stderr);
1401 		goto bailout;
1402 	}
1403 
1404 	EV_SET(&ke, cam_dev->fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1405 
1406 	dev = camdd_alloc_dev(CAMDD_DEV_PASS, &ke, 1, io_retry_count,
1407 			      io_timeout);
1408 	if (dev == NULL)
1409 		goto bailout;
1410 
1411 	pass_dev = &dev->dev_spec.pass;
1412 	pass_dev->scsi_dev_type = scsi_dev_type;
1413 	pass_dev->dev = cam_dev;
1414 	pass_dev->max_sector = maxsector;
1415 	pass_dev->block_len = block_len;
1416 	pass_dev->cpi_maxio = ccb->cpi.maxio;
1417 	snprintf(dev->device_name, sizeof(dev->device_name), "%s%u",
1418 		 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
1419 	dev->sector_size = block_len;
1420 	dev->max_sector = maxsector;
1421 
1422 
1423 	/*
1424 	 * Determine the optimal blocksize to use for this device.
1425 	 */
1426 
1427 	/*
1428 	 * If the controller has not specified a maximum I/O size,
1429 	 * just go with 128K as a somewhat conservative value.
1430 	 */
1431 	if (pass_dev->cpi_maxio == 0)
1432 		cpi_maxio = 131072;
1433 	else
1434 		cpi_maxio = pass_dev->cpi_maxio;
1435 
1436 	/*
1437 	 * If the controller has a large maximum I/O size, limit it
1438 	 * to something smaller so that the kernel doesn't have trouble
1439 	 * allocating buffers to copy data in and out for us.
1440 	 * XXX KDM this is until we have unmapped I/O support in the kernel.
1441 	 */
1442 	max_iosize = min(cpi_maxio, CAMDD_PASS_MAX_BLOCK);
1443 
1444 	/*
1445 	 * If we weren't able to get a block size for some reason,
1446 	 * default to 512 bytes.
1447 	 */
1448 	block_len = pass_dev->block_len;
1449 	if (block_len == 0)
1450 		block_len = 512;
1451 
1452 	/*
1453 	 * Figure out how many blocksize chunks will fit in the
1454 	 * maximum I/O size.
1455 	 */
1456 	pass_numblocks = max_iosize / block_len;
1457 
1458 	/*
1459 	 * And finally, multiple the number of blocks by the LBA
1460 	 * length to get our maximum block size;
1461 	 */
1462 	dev->blocksize = pass_numblocks * block_len;
1463 
1464 	if (io_opts->blocksize != 0) {
1465 		if ((io_opts->blocksize % dev->sector_size) != 0) {
1466 			warnx("Blocksize %ju for %s is not a multiple of "
1467 			      "sector size %u", (uintmax_t)io_opts->blocksize,
1468 			      dev->device_name, dev->sector_size);
1469 			goto bailout_error;
1470 		}
1471 		dev->blocksize = io_opts->blocksize;
1472 	}
1473 	dev->target_queue_depth = CAMDD_PASS_DEFAULT_DEPTH;
1474 	if (io_opts->queue_depth != 0)
1475 		dev->target_queue_depth = io_opts->queue_depth;
1476 
1477 	if (io_opts->offset != 0) {
1478 		if (io_opts->offset > (dev->max_sector * dev->sector_size)) {
1479 			warnx("Offset %ju is past the end of device %s",
1480 			    io_opts->offset, dev->device_name);
1481 			goto bailout_error;
1482 		}
1483 #if 0
1484 		else if ((io_opts->offset % dev->sector_size) != 0) {
1485 			warnx("Offset %ju for %s is not a multiple of the "
1486 			      "sector size %u", io_opts->offset,
1487 			      dev->device_name, dev->sector_size);
1488 			goto bailout_error;
1489 		}
1490 		dev->start_offset_bytes = io_opts->offset;
1491 #endif
1492 	}
1493 
1494 	dev->min_cmd_size = io_opts->min_cmd_size;
1495 
1496 	dev->run = camdd_pass_run;
1497 	dev->fetch = camdd_pass_fetch;
1498 
1499 bailout:
1500 	cam_freeccb(ccb);
1501 
1502 	return (dev);
1503 
1504 bailout_error:
1505 	cam_freeccb(ccb);
1506 
1507 	camdd_free_dev(dev);
1508 
1509 	return (NULL);
1510 }
1511 
1512 void *
1513 camdd_worker(void *arg)
1514 {
1515 	struct camdd_dev *dev = arg;
1516 	struct camdd_buf *buf;
1517 	struct timespec ts, *kq_ts;
1518 
1519 	ts.tv_sec = 0;
1520 	ts.tv_nsec = 0;
1521 
1522 	pthread_mutex_lock(&dev->mutex);
1523 
1524 	dev->flags |= CAMDD_DEV_FLAG_ACTIVE;
1525 
1526 	for (;;) {
1527 		struct kevent ke;
1528 		int retval = 0;
1529 
1530 		/*
1531 		 * XXX KDM check the reorder queue depth?
1532 		 */
1533 		if (dev->write_dev == 0) {
1534 			uint32_t our_depth, peer_depth, peer_bytes, our_bytes;
1535 			uint32_t target_depth = dev->target_queue_depth;
1536 			uint32_t peer_target_depth =
1537 			    dev->peer_dev->target_queue_depth;
1538 			uint32_t peer_blocksize = dev->peer_dev->blocksize;
1539 
1540 			camdd_get_depth(dev, &our_depth, &peer_depth,
1541 					&our_bytes, &peer_bytes);
1542 
1543 #if 0
1544 			while (((our_depth < target_depth)
1545 			     && (peer_depth < peer_target_depth))
1546 			    || ((peer_bytes + our_bytes) <
1547 				 (peer_blocksize * 2))) {
1548 #endif
1549 			while (((our_depth + peer_depth) <
1550 			        (target_depth + peer_target_depth))
1551 			    || ((peer_bytes + our_bytes) <
1552 				(peer_blocksize * 3))) {
1553 
1554 				retval = camdd_queue(dev, NULL);
1555 				if (retval == 1)
1556 					break;
1557 				else if (retval != 0) {
1558 					error_exit = 1;
1559 					goto bailout;
1560 				}
1561 
1562 				camdd_get_depth(dev, &our_depth, &peer_depth,
1563 						&our_bytes, &peer_bytes);
1564 			}
1565 		}
1566 		/*
1567 		 * See if we have any I/O that is ready to execute.
1568 		 */
1569 		buf = STAILQ_FIRST(&dev->run_queue);
1570 		if (buf != NULL) {
1571 			while (dev->target_queue_depth > dev->cur_active_io) {
1572 				retval = dev->run(dev);
1573 				if (retval == -1) {
1574 					dev->flags |= CAMDD_DEV_FLAG_EOF;
1575 					error_exit = 1;
1576 					break;
1577 				} else if (retval != 0) {
1578 					break;
1579 				}
1580 			}
1581 		}
1582 
1583 		/*
1584 		 * We've reached EOF, or our partner has reached EOF.
1585 		 */
1586 		if ((dev->flags & CAMDD_DEV_FLAG_EOF)
1587 		 || (dev->flags & CAMDD_DEV_FLAG_PEER_EOF)) {
1588 			if (dev->write_dev != 0) {
1589 			 	if ((STAILQ_EMPTY(&dev->work_queue))
1590 				 && (dev->num_run_queue == 0)
1591 				 && (dev->cur_active_io == 0)) {
1592 					goto bailout;
1593 				}
1594 			} else {
1595 				/*
1596 				 * If we're the reader, and the writer
1597 				 * got EOF, he is already done.  If we got
1598 				 * the EOF, then we need to wait until
1599 				 * everything is flushed out for the writer.
1600 				 */
1601 				if (dev->flags & CAMDD_DEV_FLAG_PEER_EOF) {
1602 					goto bailout;
1603 				} else if ((dev->num_peer_work_queue == 0)
1604 					&& (dev->num_peer_done_queue == 0)
1605 					&& (dev->cur_active_io == 0)
1606 					&& (dev->num_run_queue == 0)) {
1607 					goto bailout;
1608 				}
1609 			}
1610 			/*
1611 			 * XXX KDM need to do something about the pending
1612 			 * queue and cleanup resources.
1613 			 */
1614 		}
1615 
1616 		if ((dev->write_dev == 0)
1617 		 && (dev->cur_active_io == 0)
1618 		 && (dev->peer_bytes_queued < dev->peer_dev->blocksize))
1619 			kq_ts = &ts;
1620 		else
1621 			kq_ts = NULL;
1622 
1623 		/*
1624 		 * Run kevent to see if there are events to process.
1625 		 */
1626 		pthread_mutex_unlock(&dev->mutex);
1627 		retval = kevent(dev->kq, NULL, 0, &ke, 1, kq_ts);
1628 		pthread_mutex_lock(&dev->mutex);
1629 		if (retval == -1) {
1630 			warn("%s: error returned from kevent",__func__);
1631 			goto bailout;
1632 		} else if (retval != 0) {
1633 			switch (ke.filter) {
1634 			case EVFILT_READ:
1635 				if (dev->fetch != NULL) {
1636 					retval = dev->fetch(dev);
1637 					if (retval == -1) {
1638 						error_exit = 1;
1639 						goto bailout;
1640 					}
1641 				}
1642 				break;
1643 			case EVFILT_SIGNAL:
1644 				/*
1645 				 * We register for this so we don't get
1646 				 * an error as a result of a SIGINFO or a
1647 				 * SIGINT.  It will actually get handled
1648 				 * by the signal handler.  If we get a
1649 				 * SIGINT, bail out without printing an
1650 				 * error message.  Any other signals
1651 				 * will result in the error message above.
1652 				 */
1653 				if (ke.ident == SIGINT)
1654 					goto bailout;
1655 				break;
1656 			case EVFILT_USER:
1657 				retval = 0;
1658 				/*
1659 				 * Check to see if the other thread has
1660 				 * queued any I/O for us to do.  (In this
1661 				 * case we're the writer.)
1662 				 */
1663 				for (buf = STAILQ_FIRST(&dev->work_queue);
1664 				     buf != NULL;
1665 				     buf = STAILQ_FIRST(&dev->work_queue)) {
1666 					STAILQ_REMOVE_HEAD(&dev->work_queue,
1667 							   work_links);
1668 					retval = camdd_queue(dev, buf);
1669 					/*
1670 					 * We keep going unless we get an
1671 					 * actual error.  If we get EOF, we
1672 					 * still want to remove the buffers
1673 					 * from the queue and send the back
1674 					 * to the reader thread.
1675 					 */
1676 					if (retval == -1) {
1677 						error_exit = 1;
1678 						goto bailout;
1679 					} else
1680 						retval = 0;
1681 				}
1682 
1683 				/*
1684 				 * Next check to see if the other thread has
1685 				 * queued any completed buffers back to us.
1686 				 * (In this case we're the reader.)
1687 				 */
1688 				for (buf = STAILQ_FIRST(&dev->peer_done_queue);
1689 				     buf != NULL;
1690 				     buf = STAILQ_FIRST(&dev->peer_done_queue)){
1691 					STAILQ_REMOVE_HEAD(
1692 					    &dev->peer_done_queue, work_links);
1693 					dev->num_peer_done_queue--;
1694 					camdd_peer_done(buf);
1695 				}
1696 				break;
1697 			default:
1698 				warnx("%s: unknown kevent filter %d",
1699 				      __func__, ke.filter);
1700 				break;
1701 			}
1702 		}
1703 	}
1704 
1705 bailout:
1706 
1707 	dev->flags &= ~CAMDD_DEV_FLAG_ACTIVE;
1708 
1709 	/* XXX KDM cleanup resources here? */
1710 
1711 	pthread_mutex_unlock(&dev->mutex);
1712 
1713 	need_exit = 1;
1714 	sem_post(&camdd_sem);
1715 
1716 	return (NULL);
1717 }
1718 
1719 /*
1720  * Simplistic translation of CCB status to our local status.
1721  */
1722 camdd_buf_status
1723 camdd_ccb_status(union ccb *ccb)
1724 {
1725 	camdd_buf_status status = CAMDD_STATUS_NONE;
1726 	cam_status ccb_status;
1727 
1728 	ccb_status = ccb->ccb_h.status & CAM_STATUS_MASK;
1729 
1730 	switch (ccb_status) {
1731 	case CAM_REQ_CMP: {
1732 		if (ccb->csio.resid == 0) {
1733 			status = CAMDD_STATUS_OK;
1734 		} else if (ccb->csio.dxfer_len > ccb->csio.resid) {
1735 			status = CAMDD_STATUS_SHORT_IO;
1736 		} else {
1737 			status = CAMDD_STATUS_EOF;
1738 		}
1739 		break;
1740 	}
1741 	case CAM_SCSI_STATUS_ERROR: {
1742 		switch (ccb->csio.scsi_status) {
1743 		case SCSI_STATUS_OK:
1744 		case SCSI_STATUS_COND_MET:
1745 		case SCSI_STATUS_INTERMED:
1746 		case SCSI_STATUS_INTERMED_COND_MET:
1747 			status = CAMDD_STATUS_OK;
1748 			break;
1749 		case SCSI_STATUS_CMD_TERMINATED:
1750 		case SCSI_STATUS_CHECK_COND:
1751 		case SCSI_STATUS_QUEUE_FULL:
1752 		case SCSI_STATUS_BUSY:
1753 		case SCSI_STATUS_RESERV_CONFLICT:
1754 		default:
1755 			status = CAMDD_STATUS_ERROR;
1756 			break;
1757 		}
1758 		break;
1759 	}
1760 	default:
1761 		status = CAMDD_STATUS_ERROR;
1762 		break;
1763 	}
1764 
1765 	return (status);
1766 }
1767 
1768 /*
1769  * Queue a buffer to our peer's work thread for writing.
1770  *
1771  * Returns 0 for success, -1 for failure, 1 if the other thread exited.
1772  */
1773 int
1774 camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf)
1775 {
1776 	struct kevent ke;
1777 	STAILQ_HEAD(, camdd_buf) local_queue;
1778 	struct camdd_buf *buf1, *buf2;
1779 	struct camdd_buf_data *data = NULL;
1780 	uint64_t peer_bytes_queued = 0;
1781 	int active = 1;
1782 	int retval = 0;
1783 
1784 	STAILQ_INIT(&local_queue);
1785 
1786 	/*
1787 	 * Since we're the reader, we need to queue our I/O to the writer
1788 	 * in sequential order in order to make sure it gets written out
1789 	 * in sequential order.
1790 	 *
1791 	 * Check the next expected I/O starting offset.  If this doesn't
1792 	 * match, put it on the reorder queue.
1793 	 */
1794 	if ((buf->lba * dev->sector_size) != dev->next_completion_pos_bytes) {
1795 
1796 		/*
1797 		 * If there is nothing on the queue, there is no sorting
1798 		 * needed.
1799 		 */
1800 		if (STAILQ_EMPTY(&dev->reorder_queue)) {
1801 			STAILQ_INSERT_TAIL(&dev->reorder_queue, buf, links);
1802 			dev->num_reorder_queue++;
1803 			goto bailout;
1804 		}
1805 
1806 		/*
1807 		 * Sort in ascending order by starting LBA.  There should
1808 		 * be no identical LBAs.
1809 		 */
1810 		for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1811 		     buf1 = buf2) {
1812 			buf2 = STAILQ_NEXT(buf1, links);
1813 			if (buf->lba < buf1->lba) {
1814 				/*
1815 				 * If we're less than the first one, then
1816 				 * we insert at the head of the list
1817 				 * because this has to be the first element
1818 				 * on the list.
1819 				 */
1820 				STAILQ_INSERT_HEAD(&dev->reorder_queue,
1821 						   buf, links);
1822 				dev->num_reorder_queue++;
1823 				break;
1824 			} else if (buf->lba > buf1->lba) {
1825 				if (buf2 == NULL) {
1826 					STAILQ_INSERT_TAIL(&dev->reorder_queue,
1827 					    buf, links);
1828 					dev->num_reorder_queue++;
1829 					break;
1830 				} else if (buf->lba < buf2->lba) {
1831 					STAILQ_INSERT_AFTER(&dev->reorder_queue,
1832 					    buf1, buf, links);
1833 					dev->num_reorder_queue++;
1834 					break;
1835 				}
1836 			} else {
1837 				errx(1, "Found buffers with duplicate LBA %ju!",
1838 				     buf->lba);
1839 			}
1840 		}
1841 		goto bailout;
1842 	} else {
1843 
1844 		/*
1845 		 * We're the next expected I/O completion, so put ourselves
1846 		 * on the local queue to be sent to the writer.  We use
1847 		 * work_links here so that we can queue this to the
1848 		 * peer_work_queue before taking the buffer off of the
1849 		 * local_queue.
1850 		 */
1851 		dev->next_completion_pos_bytes += buf->len;
1852 		STAILQ_INSERT_TAIL(&local_queue, buf, work_links);
1853 
1854 		/*
1855 		 * Go through the reorder queue looking for more sequential
1856 		 * I/O and add it to the local queue.
1857 		 */
1858 		for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1859 		     buf1 = STAILQ_FIRST(&dev->reorder_queue)) {
1860 			/*
1861 			 * As soon as we see an I/O that is out of sequence,
1862 			 * we're done.
1863 			 */
1864 			if ((buf1->lba * dev->sector_size) !=
1865 			     dev->next_completion_pos_bytes)
1866 				break;
1867 
1868 			STAILQ_REMOVE_HEAD(&dev->reorder_queue, links);
1869 			dev->num_reorder_queue--;
1870 			STAILQ_INSERT_TAIL(&local_queue, buf1, work_links);
1871 			dev->next_completion_pos_bytes += buf1->len;
1872 		}
1873 	}
1874 
1875 	/*
1876 	 * Setup the event to let the other thread know that it has work
1877 	 * pending.
1878 	 */
1879 	EV_SET(&ke, (uintptr_t)&dev->peer_dev->work_queue, EVFILT_USER, 0,
1880 	       NOTE_TRIGGER, 0, NULL);
1881 
1882 	/*
1883 	 * Put this on our shadow queue so that we know what we've queued
1884 	 * to the other thread.
1885 	 */
1886 	STAILQ_FOREACH_SAFE(buf1, &local_queue, work_links, buf2) {
1887 		if (buf1->buf_type != CAMDD_BUF_DATA) {
1888 			errx(1, "%s: should have a data buffer, not an "
1889 			    "indirect buffer", __func__);
1890 		}
1891 		data = &buf1->buf_type_spec.data;
1892 
1893 		/*
1894 		 * We only need to send one EOF to the writer, and don't
1895 		 * need to continue sending EOFs after that.
1896 		 */
1897 		if (buf1->status == CAMDD_STATUS_EOF) {
1898 			if (dev->flags & CAMDD_DEV_FLAG_EOF_SENT) {
1899 				STAILQ_REMOVE(&local_queue, buf1, camdd_buf,
1900 				    work_links);
1901 				camdd_release_buf(buf1);
1902 				retval = 1;
1903 				continue;
1904 			}
1905 			dev->flags |= CAMDD_DEV_FLAG_EOF_SENT;
1906 		}
1907 
1908 
1909 		STAILQ_INSERT_TAIL(&dev->peer_work_queue, buf1, links);
1910 		peer_bytes_queued += (data->fill_len - data->resid);
1911 		dev->peer_bytes_queued += (data->fill_len - data->resid);
1912 		dev->num_peer_work_queue++;
1913 	}
1914 
1915 	if (STAILQ_FIRST(&local_queue) == NULL)
1916 		goto bailout;
1917 
1918 	/*
1919 	 * Drop our mutex and pick up the other thread's mutex.  We need to
1920 	 * do this to avoid deadlocks.
1921 	 */
1922 	pthread_mutex_unlock(&dev->mutex);
1923 	pthread_mutex_lock(&dev->peer_dev->mutex);
1924 
1925 	if (dev->peer_dev->flags & CAMDD_DEV_FLAG_ACTIVE) {
1926 		/*
1927 		 * Put the buffers on the other thread's incoming work queue.
1928 		 */
1929 		for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1930 		     buf1 = STAILQ_FIRST(&local_queue)) {
1931 			STAILQ_REMOVE_HEAD(&local_queue, work_links);
1932 			STAILQ_INSERT_TAIL(&dev->peer_dev->work_queue, buf1,
1933 					   work_links);
1934 		}
1935 		/*
1936 		 * Send an event to the other thread's kqueue to let it know
1937 		 * that there is something on the work queue.
1938 		 */
1939 		retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
1940 		if (retval == -1)
1941 			warn("%s: unable to add peer work_queue kevent",
1942 			     __func__);
1943 		else
1944 			retval = 0;
1945 	} else
1946 		active = 0;
1947 
1948 	pthread_mutex_unlock(&dev->peer_dev->mutex);
1949 	pthread_mutex_lock(&dev->mutex);
1950 
1951 	/*
1952 	 * If the other side isn't active, run through the queue and
1953 	 * release all of the buffers.
1954 	 */
1955 	if (active == 0) {
1956 		for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1957 		     buf1 = STAILQ_FIRST(&local_queue)) {
1958 			STAILQ_REMOVE_HEAD(&local_queue, work_links);
1959 			STAILQ_REMOVE(&dev->peer_work_queue, buf1, camdd_buf,
1960 				      links);
1961 			dev->num_peer_work_queue--;
1962 			camdd_release_buf(buf1);
1963 		}
1964 		dev->peer_bytes_queued -= peer_bytes_queued;
1965 		retval = 1;
1966 	}
1967 
1968 bailout:
1969 	return (retval);
1970 }
1971 
1972 /*
1973  * Return a buffer to the reader thread when we have completed writing it.
1974  */
1975 int
1976 camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf)
1977 {
1978 	struct kevent ke;
1979 	int retval = 0;
1980 
1981 	/*
1982 	 * Setup the event to let the other thread know that we have
1983 	 * completed a buffer.
1984 	 */
1985 	EV_SET(&ke, (uintptr_t)&dev->peer_dev->peer_done_queue, EVFILT_USER, 0,
1986 	       NOTE_TRIGGER, 0, NULL);
1987 
1988 	/*
1989 	 * Drop our lock and acquire the other thread's lock before
1990 	 * manipulating
1991 	 */
1992 	pthread_mutex_unlock(&dev->mutex);
1993 	pthread_mutex_lock(&dev->peer_dev->mutex);
1994 
1995 	/*
1996 	 * Put the buffer on the reader thread's peer done queue now that
1997 	 * we have completed it.
1998 	 */
1999 	STAILQ_INSERT_TAIL(&dev->peer_dev->peer_done_queue, peer_buf,
2000 			   work_links);
2001 	dev->peer_dev->num_peer_done_queue++;
2002 
2003 	/*
2004 	 * Send an event to the peer thread to let it know that we've added
2005 	 * something to its peer done queue.
2006 	 */
2007 	retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2008 	if (retval == -1)
2009 		warn("%s: unable to add peer_done_queue kevent", __func__);
2010 	else
2011 		retval = 0;
2012 
2013 	/*
2014 	 * Drop the other thread's lock and reacquire ours.
2015 	 */
2016 	pthread_mutex_unlock(&dev->peer_dev->mutex);
2017 	pthread_mutex_lock(&dev->mutex);
2018 
2019 	return (retval);
2020 }
2021 
2022 /*
2023  * Free a buffer that was written out by the writer thread and returned to
2024  * the reader thread.
2025  */
2026 void
2027 camdd_peer_done(struct camdd_buf *buf)
2028 {
2029 	struct camdd_dev *dev;
2030 	struct camdd_buf_data *data;
2031 
2032 	dev = buf->dev;
2033 	if (buf->buf_type != CAMDD_BUF_DATA) {
2034 		errx(1, "%s: should have a data buffer, not an "
2035 		    "indirect buffer", __func__);
2036 	}
2037 
2038 	data = &buf->buf_type_spec.data;
2039 
2040 	STAILQ_REMOVE(&dev->peer_work_queue, buf, camdd_buf, links);
2041 	dev->num_peer_work_queue--;
2042 	dev->peer_bytes_queued -= (data->fill_len - data->resid);
2043 
2044 	if (buf->status == CAMDD_STATUS_EOF)
2045 		dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2046 
2047 	STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2048 }
2049 
2050 /*
2051  * Assumes caller holds the lock for this device.
2052  */
2053 void
2054 camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
2055 		   int *error_count)
2056 {
2057 	int retval = 0;
2058 
2059 	/*
2060 	 * If we're the reader, we need to send the completed I/O
2061 	 * to the writer.  If we're the writer, we need to just
2062 	 * free up resources, or let the reader know if we've
2063 	 * encountered an error.
2064 	 */
2065 	if (dev->write_dev == 0) {
2066 		retval = camdd_queue_peer_buf(dev, buf);
2067 		if (retval != 0)
2068 			(*error_count)++;
2069 	} else {
2070 		struct camdd_buf *tmp_buf, *next_buf;
2071 
2072 		STAILQ_FOREACH_SAFE(tmp_buf, &buf->src_list, src_links,
2073 				    next_buf) {
2074 			struct camdd_buf *src_buf;
2075 			struct camdd_buf_indirect *indirect;
2076 
2077 			STAILQ_REMOVE(&buf->src_list, tmp_buf,
2078 				      camdd_buf, src_links);
2079 
2080 			tmp_buf->status = buf->status;
2081 
2082 			if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
2083 				camdd_complete_peer_buf(dev, tmp_buf);
2084 				continue;
2085 			}
2086 
2087 			indirect = &tmp_buf->buf_type_spec.indirect;
2088 			src_buf = indirect->src_buf;
2089 			src_buf->refcount--;
2090 			/*
2091 			 * XXX KDM we probably need to account for
2092 			 * exactly how many bytes we were able to
2093 			 * write.  Allocate the residual to the
2094 			 * first N buffers?  Or just track the
2095 			 * number of bytes written?  Right now the reader
2096 			 * doesn't do anything with a residual.
2097 			 */
2098 			src_buf->status = buf->status;
2099 			if (src_buf->refcount <= 0)
2100 				camdd_complete_peer_buf(dev, src_buf);
2101 			STAILQ_INSERT_TAIL(&dev->free_indirect_queue,
2102 					   tmp_buf, links);
2103 		}
2104 
2105 		STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2106 	}
2107 }
2108 
2109 /*
2110  * Fetch all completed commands from the pass(4) device.
2111  *
2112  * Returns the number of commands received, or -1 if any of the commands
2113  * completed with an error.  Returns 0 if no commands are available.
2114  */
2115 int
2116 camdd_pass_fetch(struct camdd_dev *dev)
2117 {
2118 	struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2119 	union ccb ccb;
2120 	int retval = 0, num_fetched = 0, error_count = 0;
2121 
2122 	pthread_mutex_unlock(&dev->mutex);
2123 	/*
2124 	 * XXX KDM we don't distinguish between EFAULT and ENOENT.
2125 	 */
2126 	while ((retval = ioctl(pass_dev->dev->fd, CAMIOGET, &ccb)) != -1) {
2127 		struct camdd_buf *buf;
2128 		struct camdd_buf_data *data;
2129 		cam_status ccb_status;
2130 		union ccb *buf_ccb;
2131 
2132 		buf = ccb.ccb_h.ccb_buf;
2133 		data = &buf->buf_type_spec.data;
2134 		buf_ccb = &data->ccb;
2135 
2136 		num_fetched++;
2137 
2138 		/*
2139 		 * Copy the CCB back out so we get status, sense data, etc.
2140 		 */
2141 		bcopy(&ccb, buf_ccb, sizeof(ccb));
2142 
2143 		pthread_mutex_lock(&dev->mutex);
2144 
2145 		/*
2146 		 * We're now done, so take this off the active queue.
2147 		 */
2148 		STAILQ_REMOVE(&dev->active_queue, buf, camdd_buf, links);
2149 		dev->cur_active_io--;
2150 
2151 		ccb_status = ccb.ccb_h.status & CAM_STATUS_MASK;
2152 		if (ccb_status != CAM_REQ_CMP) {
2153 			cam_error_print(pass_dev->dev, &ccb, CAM_ESF_ALL,
2154 					CAM_EPF_ALL, stderr);
2155 		}
2156 
2157 		data->resid = ccb.csio.resid;
2158 		dev->bytes_transferred += (ccb.csio.dxfer_len - ccb.csio.resid);
2159 
2160 		if (buf->status == CAMDD_STATUS_NONE)
2161 			buf->status = camdd_ccb_status(&ccb);
2162 		if (buf->status == CAMDD_STATUS_ERROR)
2163 			error_count++;
2164 		else if (buf->status == CAMDD_STATUS_EOF) {
2165 			/*
2166 			 * Once we queue this buffer to our partner thread,
2167 			 * he will know that we've hit EOF.
2168 			 */
2169 			dev->flags |= CAMDD_DEV_FLAG_EOF;
2170 		}
2171 
2172 		camdd_complete_buf(dev, buf, &error_count);
2173 
2174 		/*
2175 		 * Unlock in preparation for the ioctl call.
2176 		 */
2177 		pthread_mutex_unlock(&dev->mutex);
2178 	}
2179 
2180 	pthread_mutex_lock(&dev->mutex);
2181 
2182 	if (error_count > 0)
2183 		return (-1);
2184 	else
2185 		return (num_fetched);
2186 }
2187 
2188 /*
2189  * Returns -1 for error, 0 for success/continue, and 1 for resource
2190  * shortage/stop processing.
2191  */
2192 int
2193 camdd_file_run(struct camdd_dev *dev)
2194 {
2195 	struct camdd_dev_file *file_dev = &dev->dev_spec.file;
2196 	struct camdd_buf_data *data;
2197 	struct camdd_buf *buf;
2198 	off_t io_offset;
2199 	int retval = 0, write_dev = dev->write_dev;
2200 	int error_count = 0, no_resources = 0, double_buf_needed = 0;
2201 	uint32_t num_sectors = 0, db_len = 0;
2202 
2203 	buf = STAILQ_FIRST(&dev->run_queue);
2204 	if (buf == NULL) {
2205 		no_resources = 1;
2206 		goto bailout;
2207 	} else if ((dev->write_dev == 0)
2208 		&& (dev->flags & (CAMDD_DEV_FLAG_EOF |
2209 				  CAMDD_DEV_FLAG_EOF_SENT))) {
2210 		STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2211 		dev->num_run_queue--;
2212 		buf->status = CAMDD_STATUS_EOF;
2213 		error_count++;
2214 		goto bailout;
2215 	}
2216 
2217 	/*
2218 	 * If we're writing, we need to go through the source buffer list
2219 	 * and create an S/G list.
2220 	 */
2221 	if (write_dev != 0) {
2222 		retval = camdd_buf_sg_create(buf, /*iovec*/ 1,
2223 		    dev->sector_size, &num_sectors, &double_buf_needed);
2224 		if (retval != 0) {
2225 			no_resources = 1;
2226 			goto bailout;
2227 		}
2228 	}
2229 
2230 	STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2231 	dev->num_run_queue--;
2232 
2233 	data = &buf->buf_type_spec.data;
2234 
2235 	/*
2236 	 * pread(2) and pwrite(2) offsets are byte offsets.
2237 	 */
2238 	io_offset = buf->lba * dev->sector_size;
2239 
2240 	/*
2241 	 * Unlock the mutex while we read or write.
2242 	 */
2243 	pthread_mutex_unlock(&dev->mutex);
2244 
2245 	/*
2246 	 * Note that we don't need to double buffer if we're the reader
2247 	 * because in that case, we have allocated a single buffer of
2248 	 * sufficient size to do the read.  This copy is necessary on
2249 	 * writes because if one of the components of the S/G list is not
2250 	 * a sector size multiple, the kernel will reject the write.  This
2251 	 * is unfortunate but not surprising.  So this will make sure that
2252 	 * we're using a single buffer that is a multiple of the sector size.
2253 	 */
2254 	if ((double_buf_needed != 0)
2255 	 && (data->sg_count > 1)
2256 	 && (write_dev != 0)) {
2257 		uint32_t cur_offset;
2258 		int i;
2259 
2260 		if (file_dev->tmp_buf == NULL)
2261 			file_dev->tmp_buf = calloc(dev->blocksize, 1);
2262 		if (file_dev->tmp_buf == NULL) {
2263 			buf->status = CAMDD_STATUS_ERROR;
2264 			error_count++;
2265 			goto bailout;
2266 		}
2267 		for (i = 0, cur_offset = 0; i < data->sg_count; i++) {
2268 			bcopy(data->iovec[i].iov_base,
2269 			    &file_dev->tmp_buf[cur_offset],
2270 			    data->iovec[i].iov_len);
2271 			cur_offset += data->iovec[i].iov_len;
2272 		}
2273 		db_len = cur_offset;
2274 	}
2275 
2276 	if (file_dev->file_flags & CAMDD_FF_CAN_SEEK) {
2277 		if (write_dev == 0) {
2278 			/*
2279 			 * XXX KDM is there any way we would need a S/G
2280 			 * list here?
2281 			 */
2282 			retval = pread(file_dev->fd, data->buf,
2283 			    buf->len, io_offset);
2284 		} else {
2285 			if (double_buf_needed != 0) {
2286 				retval = pwrite(file_dev->fd, file_dev->tmp_buf,
2287 				    db_len, io_offset);
2288 			} else if (data->sg_count == 0) {
2289 				retval = pwrite(file_dev->fd, data->buf,
2290 				    data->fill_len, io_offset);
2291 			} else {
2292 				retval = pwritev(file_dev->fd, data->iovec,
2293 				    data->sg_count, io_offset);
2294 			}
2295 		}
2296 	} else {
2297 		if (write_dev == 0) {
2298 			/*
2299 			 * XXX KDM is there any way we would need a S/G
2300 			 * list here?
2301 			 */
2302 			retval = read(file_dev->fd, data->buf, buf->len);
2303 		} else {
2304 			if (double_buf_needed != 0) {
2305 				retval = write(file_dev->fd, file_dev->tmp_buf,
2306 				    db_len);
2307 			} else if (data->sg_count == 0) {
2308 				retval = write(file_dev->fd, data->buf,
2309 				    data->fill_len);
2310 			} else {
2311 				retval = writev(file_dev->fd, data->iovec,
2312 				    data->sg_count);
2313 			}
2314 		}
2315 	}
2316 
2317 	/* We're done, re-acquire the lock */
2318 	pthread_mutex_lock(&dev->mutex);
2319 
2320 	if (retval >= (ssize_t)data->fill_len) {
2321 		/*
2322 		 * If the bytes transferred is more than the request size,
2323 		 * that indicates an overrun, which should only happen at
2324 		 * the end of a transfer if we have to round up to a sector
2325 		 * boundary.
2326 		 */
2327 		if (buf->status == CAMDD_STATUS_NONE)
2328 			buf->status = CAMDD_STATUS_OK;
2329 		data->resid = 0;
2330 		dev->bytes_transferred += retval;
2331 	} else if (retval == -1) {
2332 		warn("Error %s %s", (write_dev) ? "writing to" :
2333 		    "reading from", file_dev->filename);
2334 
2335 		buf->status = CAMDD_STATUS_ERROR;
2336 		data->resid = data->fill_len;
2337 		error_count++;
2338 
2339 		if (dev->debug == 0)
2340 			goto bailout;
2341 
2342 		if ((double_buf_needed != 0)
2343 		 && (write_dev != 0)) {
2344 			fprintf(stderr, "%s: fd %d, DB buf %p, len %u lba %ju "
2345 			    "offset %ju\n", __func__, file_dev->fd,
2346 			    file_dev->tmp_buf, db_len, (uintmax_t)buf->lba,
2347 			    (uintmax_t)io_offset);
2348 		} else if (data->sg_count == 0) {
2349 			fprintf(stderr, "%s: fd %d, buf %p, len %u, lba %ju "
2350 			    "offset %ju\n", __func__, file_dev->fd, data->buf,
2351 			    data->fill_len, (uintmax_t)buf->lba,
2352 			    (uintmax_t)io_offset);
2353 		} else {
2354 			int i;
2355 
2356 			fprintf(stderr, "%s: fd %d, len %u, lba %ju "
2357 			    "offset %ju\n", __func__, file_dev->fd,
2358 			    data->fill_len, (uintmax_t)buf->lba,
2359 			    (uintmax_t)io_offset);
2360 
2361 			for (i = 0; i < data->sg_count; i++) {
2362 				fprintf(stderr, "index %d ptr %p len %zu\n",
2363 				    i, data->iovec[i].iov_base,
2364 				    data->iovec[i].iov_len);
2365 			}
2366 		}
2367 	} else if (retval == 0) {
2368 		buf->status = CAMDD_STATUS_EOF;
2369 		if (dev->debug != 0)
2370 			printf("%s: got EOF from %s!\n", __func__,
2371 			    file_dev->filename);
2372 		data->resid = data->fill_len;
2373 		error_count++;
2374 	} else if (retval < (ssize_t)data->fill_len) {
2375 		if (buf->status == CAMDD_STATUS_NONE)
2376 			buf->status = CAMDD_STATUS_SHORT_IO;
2377 		data->resid = data->fill_len - retval;
2378 		dev->bytes_transferred += retval;
2379 	}
2380 
2381 bailout:
2382 	if (buf != NULL) {
2383 		if (buf->status == CAMDD_STATUS_EOF) {
2384 			struct camdd_buf *buf2;
2385 			dev->flags |= CAMDD_DEV_FLAG_EOF;
2386 			STAILQ_FOREACH(buf2, &dev->run_queue, links)
2387 				buf2->status = CAMDD_STATUS_EOF;
2388 		}
2389 
2390 		camdd_complete_buf(dev, buf, &error_count);
2391 	}
2392 
2393 	if (error_count != 0)
2394 		return (-1);
2395 	else if (no_resources != 0)
2396 		return (1);
2397 	else
2398 		return (0);
2399 }
2400 
2401 /*
2402  * Execute one command from the run queue.  Returns 0 for success, 1 for
2403  * stop processing, and -1 for error.
2404  */
2405 int
2406 camdd_pass_run(struct camdd_dev *dev)
2407 {
2408 	struct camdd_buf *buf = NULL;
2409 	struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2410 	struct camdd_buf_data *data;
2411 	uint32_t num_blocks, sectors_used = 0;
2412 	union ccb *ccb;
2413 	int retval = 0, is_write = dev->write_dev;
2414 	int double_buf_needed = 0;
2415 
2416 	buf = STAILQ_FIRST(&dev->run_queue);
2417 	if (buf == NULL) {
2418 		retval = 1;
2419 		goto bailout;
2420 	}
2421 
2422 	/*
2423 	 * If we're writing, we need to go through the source buffer list
2424 	 * and create an S/G list.
2425 	 */
2426 	if (is_write != 0) {
2427 		retval = camdd_buf_sg_create(buf, /*iovec*/ 0,dev->sector_size,
2428 		    &sectors_used, &double_buf_needed);
2429 		if (retval != 0) {
2430 			retval = -1;
2431 			goto bailout;
2432 		}
2433 	}
2434 
2435 	STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2436 	dev->num_run_queue--;
2437 
2438 	data = &buf->buf_type_spec.data;
2439 
2440 	ccb = &data->ccb;
2441 	CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
2442 
2443 	/*
2444 	 * In almost every case the number of blocks should be the device
2445 	 * block size.  The exception may be at the end of an I/O stream
2446 	 * for a partial block or at the end of a device.
2447 	 */
2448 	if (is_write != 0)
2449 		num_blocks = sectors_used;
2450 	else
2451 		num_blocks = data->fill_len / pass_dev->block_len;
2452 
2453 	scsi_read_write(&ccb->csio,
2454 			/*retries*/ dev->retry_count,
2455 			/*cbfcnp*/ NULL,
2456 			/*tag_action*/ MSG_SIMPLE_Q_TAG,
2457 			/*readop*/ (dev->write_dev == 0) ? SCSI_RW_READ :
2458 				   SCSI_RW_WRITE,
2459 			/*byte2*/ 0,
2460 			/*minimum_cmd_size*/ dev->min_cmd_size,
2461 			/*lba*/ buf->lba,
2462 			/*block_count*/ num_blocks,
2463 			/*data_ptr*/ (data->sg_count != 0) ?
2464 				     (uint8_t *)data->segs : data->buf,
2465 			/*dxfer_len*/ (num_blocks * pass_dev->block_len),
2466 			/*sense_len*/ SSD_FULL_SIZE,
2467 			/*timeout*/ dev->io_timeout);
2468 
2469 	/* Disable freezing the device queue */
2470 	ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
2471 
2472 	if (dev->retry_count != 0)
2473 		ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
2474 
2475 	if (data->sg_count != 0) {
2476 		ccb->csio.sglist_cnt = data->sg_count;
2477 		ccb->ccb_h.flags |= CAM_DATA_SG;
2478 	}
2479 
2480 	/*
2481 	 * Store a pointer to the buffer in the CCB.  The kernel will
2482 	 * restore this when we get it back, and we'll use it to identify
2483 	 * the buffer this CCB came from.
2484 	 */
2485 	ccb->ccb_h.ccb_buf = buf;
2486 
2487 	/*
2488 	 * Unlock our mutex in preparation for issuing the ioctl.
2489 	 */
2490 	pthread_mutex_unlock(&dev->mutex);
2491 	/*
2492 	 * Queue the CCB to the pass(4) driver.
2493 	 */
2494 	if (ioctl(pass_dev->dev->fd, CAMIOQUEUE, ccb) == -1) {
2495 		pthread_mutex_lock(&dev->mutex);
2496 
2497 		warn("%s: error sending CAMIOQUEUE ioctl to %s%u", __func__,
2498 		     pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
2499 		warn("%s: CCB address is %p", __func__, ccb);
2500 		retval = -1;
2501 
2502 		STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2503 	} else {
2504 		pthread_mutex_lock(&dev->mutex);
2505 
2506 		dev->cur_active_io++;
2507 		STAILQ_INSERT_TAIL(&dev->active_queue, buf, links);
2508 	}
2509 
2510 bailout:
2511 	return (retval);
2512 }
2513 
2514 int
2515 camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len)
2516 {
2517 	struct camdd_dev_pass *pass_dev;
2518 	uint32_t num_blocks;
2519 	int retval = 0;
2520 
2521 	pass_dev = &dev->dev_spec.pass;
2522 
2523 	*lba = dev->next_io_pos_bytes / dev->sector_size;
2524 	*len = dev->blocksize;
2525 	num_blocks = *len / dev->sector_size;
2526 
2527 	/*
2528 	 * If max_sector is 0, then we have no set limit.  This can happen
2529 	 * if we're writing to a file in a filesystem, or reading from
2530 	 * something like /dev/zero.
2531 	 */
2532 	if ((dev->max_sector != 0)
2533 	 || (dev->sector_io_limit != 0)) {
2534 		uint64_t max_sector;
2535 
2536 		if ((dev->max_sector != 0)
2537 		 && (dev->sector_io_limit != 0))
2538 			max_sector = min(dev->sector_io_limit, dev->max_sector);
2539 		else if (dev->max_sector != 0)
2540 			max_sector = dev->max_sector;
2541 		else
2542 			max_sector = dev->sector_io_limit;
2543 
2544 
2545 		/*
2546 		 * Check to see whether we're starting off past the end of
2547 		 * the device.  If so, we need to just send an EOF
2548 		 * notification to the writer.
2549 		 */
2550 		if (*lba > max_sector) {
2551 			*len = 0;
2552 			retval = 1;
2553 		} else if (((*lba + num_blocks) > max_sector + 1)
2554 			|| ((*lba + num_blocks) < *lba)) {
2555 			/*
2556 			 * If we get here (but pass the first check), we
2557 			 * can trim the request length down to go to the
2558 			 * end of the device.
2559 			 */
2560 			num_blocks = (max_sector + 1) - *lba;
2561 			*len = num_blocks * dev->sector_size;
2562 			retval = 1;
2563 		}
2564 	}
2565 
2566 	dev->next_io_pos_bytes += *len;
2567 
2568 	return (retval);
2569 }
2570 
2571 /*
2572  * Returns 0 for success, 1 for EOF detected, and -1 for failure.
2573  */
2574 int
2575 camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf)
2576 {
2577 	struct camdd_buf *buf = NULL;
2578 	struct camdd_buf_data *data;
2579 	struct camdd_dev_pass *pass_dev;
2580 	size_t new_len;
2581 	struct camdd_buf_data *rb_data;
2582 	int is_write = dev->write_dev;
2583 	int eof_flush_needed = 0;
2584 	int retval = 0;
2585 	int error;
2586 
2587 	pass_dev = &dev->dev_spec.pass;
2588 
2589 	/*
2590 	 * If we've gotten EOF or our partner has, we should not continue
2591 	 * queueing I/O.  If we're a writer, though, we should continue
2592 	 * to write any buffers that don't have EOF status.
2593 	 */
2594 	if ((dev->flags & CAMDD_DEV_FLAG_EOF)
2595 	 || ((dev->flags & CAMDD_DEV_FLAG_PEER_EOF)
2596 	  && (is_write == 0))) {
2597 		/*
2598 		 * Tell the worker thread that we have seen EOF.
2599 		 */
2600 		retval = 1;
2601 
2602 		/*
2603 		 * If we're the writer, send the buffer back with EOF status.
2604 		 */
2605 		if (is_write) {
2606 			read_buf->status = CAMDD_STATUS_EOF;
2607 
2608 			error = camdd_complete_peer_buf(dev, read_buf);
2609 		}
2610 		goto bailout;
2611 	}
2612 
2613 	if (is_write == 0) {
2614 		buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2615 		if (buf == NULL) {
2616 			retval = -1;
2617 			goto bailout;
2618 		}
2619 		data = &buf->buf_type_spec.data;
2620 
2621 		retval = camdd_get_next_lba_len(dev, &buf->lba, &buf->len);
2622 		if (retval != 0) {
2623 			buf->status = CAMDD_STATUS_EOF;
2624 
2625 		 	if ((buf->len == 0)
2626 			 && ((dev->flags & (CAMDD_DEV_FLAG_EOF_SENT |
2627 			     CAMDD_DEV_FLAG_EOF_QUEUED)) != 0)) {
2628 				camdd_release_buf(buf);
2629 				goto bailout;
2630 			}
2631 			dev->flags |= CAMDD_DEV_FLAG_EOF_QUEUED;
2632 		}
2633 
2634 		data->fill_len = buf->len;
2635 		data->src_start_offset = buf->lba * dev->sector_size;
2636 
2637 		/*
2638 		 * Put this on the run queue.
2639 		 */
2640 		STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2641 		dev->num_run_queue++;
2642 
2643 		/* We're done. */
2644 		goto bailout;
2645 	}
2646 
2647 	/*
2648 	 * Check for new EOF status from the reader.
2649 	 */
2650 	if ((read_buf->status == CAMDD_STATUS_EOF)
2651 	 || (read_buf->status == CAMDD_STATUS_ERROR)) {
2652 		dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2653 		if ((STAILQ_FIRST(&dev->pending_queue) == NULL)
2654 		 && (read_buf->len == 0)) {
2655 			camdd_complete_peer_buf(dev, read_buf);
2656 			retval = 1;
2657 			goto bailout;
2658 		} else
2659 			eof_flush_needed = 1;
2660 	}
2661 
2662 	/*
2663 	 * See if we have a buffer we're composing with pieces from our
2664 	 * partner thread.
2665 	 */
2666 	buf = STAILQ_FIRST(&dev->pending_queue);
2667 	if (buf == NULL) {
2668 		uint64_t lba;
2669 		ssize_t len;
2670 
2671 		retval = camdd_get_next_lba_len(dev, &lba, &len);
2672 		if (retval != 0) {
2673 			read_buf->status = CAMDD_STATUS_EOF;
2674 
2675 			if (len == 0) {
2676 				dev->flags |= CAMDD_DEV_FLAG_EOF;
2677 				error = camdd_complete_peer_buf(dev, read_buf);
2678 				goto bailout;
2679 			}
2680 		}
2681 
2682 		/*
2683 		 * If we don't have a pending buffer, we need to grab a new
2684 		 * one from the free list or allocate another one.
2685 		 */
2686 		buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2687 		if (buf == NULL) {
2688 			retval = 1;
2689 			goto bailout;
2690 		}
2691 
2692 		buf->lba = lba;
2693 		buf->len = len;
2694 
2695 		STAILQ_INSERT_TAIL(&dev->pending_queue, buf, links);
2696 		dev->num_pending_queue++;
2697 	}
2698 
2699 	data = &buf->buf_type_spec.data;
2700 
2701 	rb_data = &read_buf->buf_type_spec.data;
2702 
2703 	if ((rb_data->src_start_offset != dev->next_peer_pos_bytes)
2704 	 && (dev->debug != 0)) {
2705 		printf("%s: WARNING: reader offset %#jx != expected offset "
2706 		    "%#jx\n", __func__, (uintmax_t)rb_data->src_start_offset,
2707 		    (uintmax_t)dev->next_peer_pos_bytes);
2708 	}
2709 	dev->next_peer_pos_bytes = rb_data->src_start_offset +
2710 	    (rb_data->fill_len - rb_data->resid);
2711 
2712 	new_len = (rb_data->fill_len - rb_data->resid) + data->fill_len;
2713 	if (new_len < buf->len) {
2714 		/*
2715 		 * There are three cases here:
2716 		 * 1. We need more data to fill up a block, so we put
2717 		 *    this I/O on the queue and wait for more I/O.
2718 		 * 2. We have a pending buffer in the queue that is
2719 		 *    smaller than our blocksize, but we got an EOF.  So we
2720 		 *    need to go ahead and flush the write out.
2721 		 * 3. We got an error.
2722 		 */
2723 
2724 		/*
2725 		 * Increment our fill length.
2726 		 */
2727 		data->fill_len += (rb_data->fill_len - rb_data->resid);
2728 
2729 		/*
2730 		 * Add the new read buffer to the list for writing.
2731 		 */
2732 		STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2733 
2734 		/* Increment the count */
2735 		buf->src_count++;
2736 
2737 		if (eof_flush_needed == 0) {
2738 			/*
2739 			 * We need to exit, because we don't have enough
2740 			 * data yet.
2741 			 */
2742 			goto bailout;
2743 		} else {
2744 			/*
2745 			 * Take the buffer off of the pending queue.
2746 			 */
2747 			STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2748 				      links);
2749 			dev->num_pending_queue--;
2750 
2751 			/*
2752 			 * If we need an EOF flush, but there is no data
2753 			 * to flush, go ahead and return this buffer.
2754 			 */
2755 			if (data->fill_len == 0) {
2756 				camdd_complete_buf(dev, buf, /*error_count*/0);
2757 				retval = 1;
2758 				goto bailout;
2759 			}
2760 
2761 			/*
2762 			 * Put this on the next queue for execution.
2763 			 */
2764 			STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2765 			dev->num_run_queue++;
2766 		}
2767 	} else if (new_len == buf->len) {
2768 		/*
2769 		 * We have enough data to completey fill one block,
2770 		 * so we're ready to issue the I/O.
2771 		 */
2772 
2773 		/*
2774 		 * Take the buffer off of the pending queue.
2775 		 */
2776 		STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf, links);
2777 		dev->num_pending_queue--;
2778 
2779 		/*
2780 		 * Add the new read buffer to the list for writing.
2781 		 */
2782 		STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2783 
2784 		/* Increment the count */
2785 		buf->src_count++;
2786 
2787 		/*
2788 		 * Increment our fill length.
2789 		 */
2790 		data->fill_len += (rb_data->fill_len - rb_data->resid);
2791 
2792 		/*
2793 		 * Put this on the next queue for execution.
2794 		 */
2795 		STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2796 		dev->num_run_queue++;
2797 	} else {
2798 		struct camdd_buf *idb;
2799 		struct camdd_buf_indirect *indirect;
2800 		uint32_t len_to_go, cur_offset;
2801 
2802 
2803 		idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2804 		if (idb == NULL) {
2805 			retval = 1;
2806 			goto bailout;
2807 		}
2808 		indirect = &idb->buf_type_spec.indirect;
2809 		indirect->src_buf = read_buf;
2810 		read_buf->refcount++;
2811 		indirect->offset = 0;
2812 		indirect->start_ptr = rb_data->buf;
2813 		/*
2814 		 * We've already established that there is more
2815 		 * data in read_buf than we have room for in our
2816 		 * current write request.  So this particular chunk
2817 		 * of the request should just be the remainder
2818 		 * needed to fill up a block.
2819 		 */
2820 		indirect->len = buf->len - (data->fill_len - data->resid);
2821 
2822 		camdd_buf_add_child(buf, idb);
2823 
2824 		/*
2825 		 * This buffer is ready to execute, so we can take
2826 		 * it off the pending queue and put it on the run
2827 		 * queue.
2828 		 */
2829 		STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2830 			      links);
2831 		dev->num_pending_queue--;
2832 		STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2833 		dev->num_run_queue++;
2834 
2835 		cur_offset = indirect->offset + indirect->len;
2836 
2837 		/*
2838 		 * The resulting I/O would be too large to fit in
2839 		 * one block.  We need to split this I/O into
2840 		 * multiple pieces.  Allocate as many buffers as needed.
2841 		 */
2842 		for (len_to_go = rb_data->fill_len - rb_data->resid -
2843 		     indirect->len; len_to_go > 0;) {
2844 			struct camdd_buf *new_buf;
2845 			struct camdd_buf_data *new_data;
2846 			uint64_t lba;
2847 			ssize_t len;
2848 
2849 			retval = camdd_get_next_lba_len(dev, &lba, &len);
2850 			if ((retval != 0)
2851 			 && (len == 0)) {
2852 				/*
2853 				 * The device has already been marked
2854 				 * as EOF, and there is no space left.
2855 				 */
2856 				goto bailout;
2857 			}
2858 
2859 			new_buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2860 			if (new_buf == NULL) {
2861 				retval = 1;
2862 				goto bailout;
2863 			}
2864 
2865 			new_buf->lba = lba;
2866 			new_buf->len = len;
2867 
2868 			idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2869 			if (idb == NULL) {
2870 				retval = 1;
2871 				goto bailout;
2872 			}
2873 
2874 			indirect = &idb->buf_type_spec.indirect;
2875 
2876 			indirect->src_buf = read_buf;
2877 			read_buf->refcount++;
2878 			indirect->offset = cur_offset;
2879 			indirect->start_ptr = rb_data->buf + cur_offset;
2880 			indirect->len = min(len_to_go, new_buf->len);
2881 #if 0
2882 			if (((indirect->len % dev->sector_size) != 0)
2883 			 || ((indirect->offset % dev->sector_size) != 0)) {
2884 				warnx("offset %ju len %ju not aligned with "
2885 				    "sector size %u", indirect->offset,
2886 				    (uintmax_t)indirect->len, dev->sector_size);
2887 			}
2888 #endif
2889 			cur_offset += indirect->len;
2890 			len_to_go -= indirect->len;
2891 
2892 			camdd_buf_add_child(new_buf, idb);
2893 
2894 			new_data = &new_buf->buf_type_spec.data;
2895 
2896 			if ((new_data->fill_len == new_buf->len)
2897 			 || (eof_flush_needed != 0)) {
2898 				STAILQ_INSERT_TAIL(&dev->run_queue,
2899 						   new_buf, links);
2900 				dev->num_run_queue++;
2901 			} else if (new_data->fill_len < buf->len) {
2902 				STAILQ_INSERT_TAIL(&dev->pending_queue,
2903 					   	new_buf, links);
2904 				dev->num_pending_queue++;
2905 			} else {
2906 				warnx("%s: too much data in new "
2907 				      "buffer!", __func__);
2908 				retval = 1;
2909 				goto bailout;
2910 			}
2911 		}
2912 	}
2913 
2914 bailout:
2915 	return (retval);
2916 }
2917 
2918 void
2919 camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
2920 		uint32_t *peer_depth, uint32_t *our_bytes, uint32_t *peer_bytes)
2921 {
2922 	*our_depth = dev->cur_active_io + dev->num_run_queue;
2923 	if (dev->num_peer_work_queue >
2924 	    dev->num_peer_done_queue)
2925 		*peer_depth = dev->num_peer_work_queue -
2926 			      dev->num_peer_done_queue;
2927 	else
2928 		*peer_depth = 0;
2929 	*our_bytes = *our_depth * dev->blocksize;
2930 	*peer_bytes = dev->peer_bytes_queued;
2931 }
2932 
2933 void
2934 camdd_sig_handler(int sig)
2935 {
2936 	if (sig == SIGINFO)
2937 		need_status = 1;
2938 	else {
2939 		need_exit = 1;
2940 		error_exit = 1;
2941 	}
2942 
2943 	sem_post(&camdd_sem);
2944 }
2945 
2946 void
2947 camdd_print_status(struct camdd_dev *camdd_dev, struct camdd_dev *other_dev,
2948 		   struct timespec *start_time)
2949 {
2950 	struct timespec done_time;
2951 	uint64_t total_ns;
2952 	long double mb_sec, total_sec;
2953 	int error = 0;
2954 
2955 	error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &done_time);
2956 	if (error != 0) {
2957 		warn("Unable to get done time");
2958 		return;
2959 	}
2960 
2961 	timespecsub(&done_time, start_time);
2962 
2963 	total_ns = done_time.tv_nsec + (done_time.tv_sec * 1000000000);
2964 	total_sec = total_ns;
2965 	total_sec /= 1000000000;
2966 
2967 	fprintf(stderr, "%ju bytes %s %s\n%ju bytes %s %s\n"
2968 		"%.4Lf seconds elapsed\n",
2969 		(uintmax_t)camdd_dev->bytes_transferred,
2970 		(camdd_dev->write_dev == 0) ?  "read from" : "written to",
2971 		camdd_dev->device_name,
2972 		(uintmax_t)other_dev->bytes_transferred,
2973 		(other_dev->write_dev == 0) ? "read from" : "written to",
2974 		other_dev->device_name, total_sec);
2975 
2976 	mb_sec = min(other_dev->bytes_transferred,camdd_dev->bytes_transferred);
2977 	mb_sec /= 1024 * 1024;
2978 	mb_sec *= 1000000000;
2979 	mb_sec /= total_ns;
2980 	fprintf(stderr, "%.2Lf MB/sec\n", mb_sec);
2981 }
2982 
2983 int
2984 camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts, uint64_t max_io,
2985 	 int retry_count, int timeout)
2986 {
2987 	char *device = NULL;
2988 	struct cam_device *new_cam_dev = NULL;
2989 	struct camdd_dev *devs[2];
2990 	struct timespec start_time;
2991 	pthread_t threads[2];
2992 	int unit = 0;
2993 	int error = 0;
2994 	int i;
2995 
2996 	if (num_io_opts != 2) {
2997 		warnx("Must have one input and one output path");
2998 		error = 1;
2999 		goto bailout;
3000 	}
3001 
3002 	bzero(devs, sizeof(devs));
3003 
3004 	for (i = 0; i < num_io_opts; i++) {
3005 		switch (io_opts[i].dev_type) {
3006 		case CAMDD_DEV_PASS: {
3007 			camdd_argmask new_arglist = CAMDD_ARG_NONE;
3008 			int bus = 0, target = 0, lun = 0;
3009 			char name[30];
3010 			int rv;
3011 
3012 			if (isdigit(io_opts[i].dev_name[0])) {
3013 				/* device specified as bus:target[:lun] */
3014 				rv = parse_btl(io_opts[i].dev_name, &bus,
3015 				    &target, &lun, &new_arglist);
3016 				if (rv < 2) {
3017 					warnx("numeric device specification "
3018 					     "must be either bus:target, or "
3019 					     "bus:target:lun");
3020 					error = 1;
3021 					goto bailout;
3022 				}
3023 				/* default to 0 if lun was not specified */
3024 				if ((new_arglist & CAMDD_ARG_LUN) == 0) {
3025 					lun = 0;
3026 					new_arglist |= CAMDD_ARG_LUN;
3027 				}
3028 			} else {
3029 				if (cam_get_device(io_opts[i].dev_name, name,
3030 						   sizeof name, &unit) == -1) {
3031 					warnx("%s", cam_errbuf);
3032 					error = 1;
3033 					goto bailout;
3034 				}
3035 				device = strdup(name);
3036 				new_arglist |= CAMDD_ARG_DEVICE |CAMDD_ARG_UNIT;
3037 			}
3038 
3039 			if (new_arglist & (CAMDD_ARG_BUS | CAMDD_ARG_TARGET))
3040 				new_cam_dev = cam_open_btl(bus, target, lun,
3041 				    O_RDWR, NULL);
3042 			else
3043 				new_cam_dev = cam_open_spec_device(device, unit,
3044 				    O_RDWR, NULL);
3045 			if (new_cam_dev == NULL) {
3046 				warnx("%s", cam_errbuf);
3047 				error = 1;
3048 				goto bailout;
3049 			}
3050 
3051 			devs[i] = camdd_probe_pass(new_cam_dev,
3052 			    /*io_opts*/ &io_opts[i],
3053 			    CAMDD_ARG_ERR_RECOVER,
3054 			    /*probe_retry_count*/ 3,
3055 			    /*probe_timeout*/ 5000,
3056 			    /*io_retry_count*/ retry_count,
3057 			    /*io_timeout*/ timeout);
3058 			if (devs[i] == NULL) {
3059 				warn("Unable to probe device %s%u",
3060 				     new_cam_dev->device_name,
3061 				     new_cam_dev->dev_unit_num);
3062 				error = 1;
3063 				goto bailout;
3064 			}
3065 			break;
3066 		}
3067 		case CAMDD_DEV_FILE: {
3068 			int fd = -1;
3069 
3070 			if (io_opts[i].dev_name[0] == '-') {
3071 				if (io_opts[i].write_dev != 0)
3072 					fd = STDOUT_FILENO;
3073 				else
3074 					fd = STDIN_FILENO;
3075 			} else {
3076 				if (io_opts[i].write_dev != 0) {
3077 					fd = open(io_opts[i].dev_name,
3078 					    O_RDWR | O_CREAT, S_IWUSR |S_IRUSR);
3079 				} else {
3080 					fd = open(io_opts[i].dev_name,
3081 					    O_RDONLY);
3082 				}
3083 			}
3084 			if (fd == -1) {
3085 				warn("error opening file %s",
3086 				    io_opts[i].dev_name);
3087 				error = 1;
3088 				goto bailout;
3089 			}
3090 
3091 			devs[i] = camdd_probe_file(fd, &io_opts[i],
3092 			    retry_count, timeout);
3093 			if (devs[i] == NULL) {
3094 				error = 1;
3095 				goto bailout;
3096 			}
3097 
3098 			break;
3099 		}
3100 		default:
3101 			warnx("Unknown device type %d (%s)",
3102 			    io_opts[i].dev_type, io_opts[i].dev_name);
3103 			error = 1;
3104 			goto bailout;
3105 			break; /*NOTREACHED */
3106 		}
3107 
3108 		devs[i]->write_dev = io_opts[i].write_dev;
3109 
3110 		devs[i]->start_offset_bytes = io_opts[i].offset;
3111 
3112 		if (max_io != 0) {
3113 			devs[i]->sector_io_limit =
3114 			    (devs[i]->start_offset_bytes /
3115 			    devs[i]->sector_size) +
3116 			    (max_io / devs[i]->sector_size) - 1;
3117 			devs[i]->sector_io_limit =
3118 			    (devs[i]->start_offset_bytes /
3119 			    devs[i]->sector_size) +
3120 			    (max_io / devs[i]->sector_size) - 1;
3121 		}
3122 
3123 		devs[i]->next_io_pos_bytes = devs[i]->start_offset_bytes;
3124 		devs[i]->next_completion_pos_bytes =devs[i]->start_offset_bytes;
3125 	}
3126 
3127 	devs[0]->peer_dev = devs[1];
3128 	devs[1]->peer_dev = devs[0];
3129 	devs[0]->next_peer_pos_bytes = devs[0]->peer_dev->next_io_pos_bytes;
3130 	devs[1]->next_peer_pos_bytes = devs[1]->peer_dev->next_io_pos_bytes;
3131 
3132 	sem_init(&camdd_sem, /*pshared*/ 0, 0);
3133 
3134 	signal(SIGINFO, camdd_sig_handler);
3135 	signal(SIGINT, camdd_sig_handler);
3136 
3137 	error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &start_time);
3138 	if (error != 0) {
3139 		warn("Unable to get start time");
3140 		goto bailout;
3141 	}
3142 
3143 	for (i = 0; i < num_io_opts; i++) {
3144 		error = pthread_create(&threads[i], NULL, camdd_worker,
3145 				       (void *)devs[i]);
3146 		if (error != 0) {
3147 			warnc(error, "pthread_create() failed");
3148 			goto bailout;
3149 		}
3150 	}
3151 
3152 	for (;;) {
3153 		if ((sem_wait(&camdd_sem) == -1)
3154 		 || (need_exit != 0)) {
3155 			struct kevent ke;
3156 
3157 			for (i = 0; i < num_io_opts; i++) {
3158 				EV_SET(&ke, (uintptr_t)&devs[i]->work_queue,
3159 				    EVFILT_USER, 0, NOTE_TRIGGER, 0, NULL);
3160 
3161 				devs[i]->flags |= CAMDD_DEV_FLAG_EOF;
3162 
3163 				error = kevent(devs[i]->kq, &ke, 1, NULL, 0,
3164 						NULL);
3165 				if (error == -1)
3166 					warn("%s: unable to wake up thread",
3167 					    __func__);
3168 				error = 0;
3169 			}
3170 			break;
3171 		} else if (need_status != 0) {
3172 			camdd_print_status(devs[0], devs[1], &start_time);
3173 			need_status = 0;
3174 		}
3175 	}
3176 	for (i = 0; i < num_io_opts; i++) {
3177 		pthread_join(threads[i], NULL);
3178 	}
3179 
3180 	camdd_print_status(devs[0], devs[1], &start_time);
3181 
3182 bailout:
3183 
3184 	for (i = 0; i < num_io_opts; i++)
3185 		camdd_free_dev(devs[i]);
3186 
3187 	return (error + error_exit);
3188 }
3189 
3190 void
3191 usage(void)
3192 {
3193 	fprintf(stderr,
3194 "usage:  camdd <-i|-o pass=pass0,bs=1M,offset=1M,depth=4>\n"
3195 "              <-i|-o file=/tmp/file,bs=512K,offset=1M>\n"
3196 "              <-i|-o file=/dev/da0,bs=512K,offset=1M>\n"
3197 "              <-i|-o file=/dev/nsa0,bs=512K>\n"
3198 "              [-C retry_count][-E][-m max_io_amt][-t timeout_secs][-v][-h]\n"
3199 "Option description\n"
3200 "-i <arg=val>  Specify input device/file and parameters\n"
3201 "-o <arg=val>  Specify output device/file and parameters\n"
3202 "Input and Output parameters\n"
3203 "pass=name     Specify a pass(4) device like pass0 or /dev/pass0\n"
3204 "file=name     Specify a file or device, /tmp/foo, /dev/da0, /dev/null\n"
3205 "              or - for stdin/stdout\n"
3206 "bs=blocksize  Specify blocksize in bytes, or using K, M, G, etc. suffix\n"
3207 "offset=len    Specify starting offset in bytes or using K, M, G suffix\n"
3208 "              NOTE: offset cannot be specified on tapes, pipes, stdin/out\n"
3209 "depth=N       Specify a numeric queue depth.  This only applies to pass(4)\n"
3210 "mcs=N         Specify a minimum cmd size for pass(4) read/write commands\n"
3211 "Optional arguments\n"
3212 "-C retry_cnt  Specify a retry count for pass(4) devices\n"
3213 "-E            Enable CAM error recovery for pass(4) devices\n"
3214 "-m max_io     Specify the maximum amount to be transferred in bytes or\n"
3215 "              using K, G, M, etc. suffixes\n"
3216 "-t timeout    Specify the I/O timeout to use with pass(4) devices\n"
3217 "-v            Enable verbose error recovery\n"
3218 "-h            Print this message\n");
3219 }
3220 
3221 
3222 int
3223 camdd_parse_io_opts(char *args, int is_write, struct camdd_io_opts *io_opts)
3224 {
3225 	char *tmpstr, *tmpstr2;
3226 	char *orig_tmpstr = NULL;
3227 	int retval = 0;
3228 
3229 	io_opts->write_dev = is_write;
3230 
3231 	tmpstr = strdup(args);
3232 	if (tmpstr == NULL) {
3233 		warn("strdup failed");
3234 		retval = 1;
3235 		goto bailout;
3236 	}
3237 	orig_tmpstr = tmpstr;
3238 	while ((tmpstr2 = strsep(&tmpstr, ",")) != NULL) {
3239 		char *name, *value;
3240 
3241 		/*
3242 		 * If the user creates an empty parameter by putting in two
3243 		 * commas, skip over it and look for the next field.
3244 		 */
3245 		if (*tmpstr2 == '\0')
3246 			continue;
3247 
3248 		name = strsep(&tmpstr2, "=");
3249 		if (*name == '\0') {
3250 			warnx("Got empty I/O parameter name");
3251 			retval = 1;
3252 			goto bailout;
3253 		}
3254 		value = strsep(&tmpstr2, "=");
3255 		if ((value == NULL)
3256 		 || (*value == '\0')) {
3257 			warnx("Empty I/O parameter value for %s", name);
3258 			retval = 1;
3259 			goto bailout;
3260 		}
3261 		if (strncasecmp(name, "file", 4) == 0) {
3262 			io_opts->dev_type = CAMDD_DEV_FILE;
3263 			io_opts->dev_name = strdup(value);
3264 			if (io_opts->dev_name == NULL) {
3265 				warn("Error allocating memory");
3266 				retval = 1;
3267 				goto bailout;
3268 			}
3269 		} else if (strncasecmp(name, "pass", 4) == 0) {
3270 			io_opts->dev_type = CAMDD_DEV_PASS;
3271 			io_opts->dev_name = strdup(value);
3272 			if (io_opts->dev_name == NULL) {
3273 				warn("Error allocating memory");
3274 				retval = 1;
3275 				goto bailout;
3276 			}
3277 		} else if ((strncasecmp(name, "bs", 2) == 0)
3278 			|| (strncasecmp(name, "blocksize", 9) == 0)) {
3279 			retval = expand_number(value, &io_opts->blocksize);
3280 			if (retval == -1) {
3281 				warn("expand_number(3) failed on %s=%s", name,
3282 				    value);
3283 				retval = 1;
3284 				goto bailout;
3285 			}
3286 		} else if (strncasecmp(name, "depth", 5) == 0) {
3287 			char *endptr;
3288 
3289 			io_opts->queue_depth = strtoull(value, &endptr, 0);
3290 			if (*endptr != '\0') {
3291 				warnx("invalid queue depth %s", value);
3292 				retval = 1;
3293 				goto bailout;
3294 			}
3295 		} else if (strncasecmp(name, "mcs", 3) == 0) {
3296 			char *endptr;
3297 
3298 			io_opts->min_cmd_size = strtol(value, &endptr, 0);
3299 			if ((*endptr != '\0')
3300 			 || ((io_opts->min_cmd_size > 16)
3301 			  || (io_opts->min_cmd_size < 0))) {
3302 				warnx("invalid minimum cmd size %s", value);
3303 				retval = 1;
3304 				goto bailout;
3305 			}
3306 		} else if (strncasecmp(name, "offset", 6) == 0) {
3307 			retval = expand_number(value, &io_opts->offset);
3308 			if (retval == -1) {
3309 				warn("expand_number(3) failed on %s=%s", name,
3310 				    value);
3311 				retval = 1;
3312 				goto bailout;
3313 			}
3314 		} else if (strncasecmp(name, "debug", 5) == 0) {
3315 			char *endptr;
3316 
3317 			io_opts->debug = strtoull(value, &endptr, 0);
3318 			if (*endptr != '\0') {
3319 				warnx("invalid debug level %s", value);
3320 				retval = 1;
3321 				goto bailout;
3322 			}
3323 		} else {
3324 			warnx("Unrecognized parameter %s=%s", name, value);
3325 		}
3326 	}
3327 bailout:
3328 	free(orig_tmpstr);
3329 
3330 	return (retval);
3331 }
3332 
3333 int
3334 main(int argc, char **argv)
3335 {
3336 	int c;
3337 	camdd_argmask arglist = CAMDD_ARG_NONE;
3338 	int timeout = 0, retry_count = 1;
3339 	int error = 0;
3340 	uint64_t max_io = 0;
3341 	struct camdd_io_opts *opt_list = NULL;
3342 
3343 	if (argc == 1) {
3344 		usage();
3345 		exit(1);
3346 	}
3347 
3348 	opt_list = calloc(2, sizeof(struct camdd_io_opts));
3349 	if (opt_list == NULL) {
3350 		warn("Unable to allocate option list");
3351 		error = 1;
3352 		goto bailout;
3353 	}
3354 
3355 	while ((c = getopt(argc, argv, "C:Ehi:m:o:t:v")) != -1){
3356 		switch (c) {
3357 		case 'C':
3358 			retry_count = strtol(optarg, NULL, 0);
3359 			if (retry_count < 0)
3360 				errx(1, "retry count %d is < 0",
3361 				     retry_count);
3362 			arglist |= CAMDD_ARG_RETRIES;
3363 			break;
3364 		case 'E':
3365 			arglist |= CAMDD_ARG_ERR_RECOVER;
3366 			break;
3367 		case 'i':
3368 		case 'o':
3369 			if (((c == 'i')
3370 			  && (opt_list[0].dev_type != CAMDD_DEV_NONE))
3371 			 || ((c == 'o')
3372 			  && (opt_list[1].dev_type != CAMDD_DEV_NONE))) {
3373 				errx(1, "Only one input and output path "
3374 				    "allowed");
3375 			}
3376 			error = camdd_parse_io_opts(optarg, (c == 'o') ? 1 : 0,
3377 			    (c == 'o') ? &opt_list[1] : &opt_list[0]);
3378 			if (error != 0)
3379 				goto bailout;
3380 			break;
3381 		case 'm':
3382 			error = expand_number(optarg, &max_io);
3383 			if (error == -1) {
3384 				warn("invalid maximum I/O amount %s", optarg);
3385 				error = 1;
3386 				goto bailout;
3387 			}
3388 			break;
3389 		case 't':
3390 			timeout = strtol(optarg, NULL, 0);
3391 			if (timeout < 0)
3392 				errx(1, "invalid timeout %d", timeout);
3393 			/* Convert the timeout from seconds to ms */
3394 			timeout *= 1000;
3395 			arglist |= CAMDD_ARG_TIMEOUT;
3396 			break;
3397 		case 'v':
3398 			arglist |= CAMDD_ARG_VERBOSE;
3399 			break;
3400 		case 'h':
3401 		default:
3402 			usage();
3403 			exit(1);
3404 			break; /*NOTREACHED*/
3405 		}
3406 	}
3407 
3408 	if ((opt_list[0].dev_type == CAMDD_DEV_NONE)
3409 	 || (opt_list[1].dev_type == CAMDD_DEV_NONE))
3410 		errx(1, "Must specify both -i and -o");
3411 
3412 	/*
3413 	 * Set the timeout if the user hasn't specified one.
3414 	 */
3415 	if (timeout == 0)
3416 		timeout = CAMDD_PASS_RW_TIMEOUT;
3417 
3418 	error = camdd_rw(opt_list, 2, max_io, retry_count, timeout);
3419 
3420 bailout:
3421 	free(opt_list);
3422 
3423 	exit(error);
3424 }
3425