xref: /freebsd/usr.sbin/camdd/camdd.c (revision 99429157e8615dc3b7f11afbe3ed92de7476a5db)
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 sig_atomic_t need_exit = 0;
424 static sig_atomic_t error_exit = 0;
425 static sig_atomic_t 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 	free(data_ptr);
716 
717 	return (NULL);
718 }
719 
720 void
721 camdd_release_buf(struct camdd_buf *buf)
722 {
723 	struct camdd_dev *dev;
724 
725 	dev = buf->dev;
726 
727 	switch (buf->buf_type) {
728 	case CAMDD_BUF_DATA: {
729 		struct camdd_buf_data *data;
730 
731 		data = &buf->buf_type_spec.data;
732 
733 		if (data->segs != NULL) {
734 			if (data->extra_buf != 0) {
735 				void *extra_buf;
736 
737 				extra_buf = (void *)
738 				    data->segs[data->sg_count - 1].ds_addr;
739 				free(extra_buf);
740 				data->extra_buf = 0;
741 			}
742 			free(data->segs);
743 			data->segs = NULL;
744 			data->sg_count = 0;
745 		} else if (data->iovec != NULL) {
746 			if (data->extra_buf != 0) {
747 				free(data->iovec[data->sg_count - 1].iov_base);
748 				data->extra_buf = 0;
749 			}
750 			free(data->iovec);
751 			data->iovec = NULL;
752 			data->sg_count = 0;
753 		}
754 		STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
755 		break;
756 	}
757 	case CAMDD_BUF_INDIRECT:
758 		STAILQ_INSERT_TAIL(&dev->free_indirect_queue, buf, links);
759 		break;
760 	default:
761 		err(1, "%s: Invalid buffer type %d for released buffer",
762 		    __func__, buf->buf_type);
763 		break;
764 	}
765 }
766 
767 struct camdd_buf *
768 camdd_get_buf(struct camdd_dev *dev, camdd_buf_type buf_type)
769 {
770 	struct camdd_buf *buf = NULL;
771 
772 	switch (buf_type) {
773 	case CAMDD_BUF_DATA:
774 		buf = STAILQ_FIRST(&dev->free_queue);
775 		if (buf != NULL) {
776 			struct camdd_buf_data *data;
777 			uint8_t *data_ptr;
778 			uint32_t alloc_len;
779 
780 			STAILQ_REMOVE_HEAD(&dev->free_queue, links);
781 			data = &buf->buf_type_spec.data;
782 			data_ptr = data->buf;
783 			alloc_len = data->alloc_len;
784 			bzero(buf, sizeof(*buf));
785 			data->buf = data_ptr;
786 			data->alloc_len = alloc_len;
787 		}
788 		break;
789 	case CAMDD_BUF_INDIRECT:
790 		buf = STAILQ_FIRST(&dev->free_indirect_queue);
791 		if (buf != NULL) {
792 			STAILQ_REMOVE_HEAD(&dev->free_indirect_queue, links);
793 
794 			bzero(buf, sizeof(*buf));
795 		}
796 		break;
797 	default:
798 		warnx("Unknown buffer type %d requested", buf_type);
799 		break;
800 	}
801 
802 
803 	if (buf == NULL)
804 		return (camdd_alloc_buf(dev, buf_type));
805 	else {
806 		STAILQ_INIT(&buf->src_list);
807 		buf->dev = dev;
808 		buf->buf_type = buf_type;
809 
810 		return (buf);
811 	}
812 }
813 
814 int
815 camdd_buf_sg_create(struct camdd_buf *buf, int iovec, uint32_t sector_size,
816 		    uint32_t *num_sectors_used, int *double_buf_needed)
817 {
818 	struct camdd_buf *tmp_buf;
819 	struct camdd_buf_data *data;
820 	uint8_t *extra_buf = NULL;
821 	size_t extra_buf_len = 0;
822 	int i, retval = 0;
823 
824 	data = &buf->buf_type_spec.data;
825 
826 	data->sg_count = buf->src_count;
827 	/*
828 	 * Compose a scatter/gather list from all of the buffers in the list.
829 	 * If the length of the buffer isn't a multiple of the sector size,
830 	 * we'll have to add an extra buffer.  This should only happen
831 	 * at the end of a transfer.
832 	 */
833 	if ((data->fill_len % sector_size) != 0) {
834 		extra_buf_len = sector_size - (data->fill_len % sector_size);
835 		extra_buf = calloc(extra_buf_len, 1);
836 		if (extra_buf == NULL) {
837 			warn("%s: unable to allocate %zu bytes for extra "
838 			    "buffer space", __func__, extra_buf_len);
839 			retval = 1;
840 			goto bailout;
841 		}
842 		data->extra_buf = 1;
843 		data->sg_count++;
844 	}
845 	if (iovec == 0) {
846 		data->segs = calloc(data->sg_count, sizeof(bus_dma_segment_t));
847 		if (data->segs == NULL) {
848 			warn("%s: unable to allocate %zu bytes for S/G list",
849 			    __func__, sizeof(bus_dma_segment_t) *
850 			    data->sg_count);
851 			retval = 1;
852 			goto bailout;
853 		}
854 
855 	} else {
856 		data->iovec = calloc(data->sg_count, sizeof(struct iovec));
857 		if (data->iovec == NULL) {
858 			warn("%s: unable to allocate %zu bytes for S/G list",
859 			    __func__, sizeof(struct iovec) * data->sg_count);
860 			retval = 1;
861 			goto bailout;
862 		}
863 	}
864 
865 	for (i = 0, tmp_buf = STAILQ_FIRST(&buf->src_list);
866 	     i < buf->src_count && tmp_buf != NULL; i++,
867 	     tmp_buf = STAILQ_NEXT(tmp_buf, src_links)) {
868 
869 		if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
870 			struct camdd_buf_data *tmp_data;
871 
872 			tmp_data = &tmp_buf->buf_type_spec.data;
873 			if (iovec == 0) {
874 				data->segs[i].ds_addr =
875 				    (bus_addr_t) tmp_data->buf;
876 				data->segs[i].ds_len = tmp_data->fill_len -
877 				    tmp_data->resid;
878 			} else {
879 				data->iovec[i].iov_base = tmp_data->buf;
880 				data->iovec[i].iov_len = tmp_data->fill_len -
881 				    tmp_data->resid;
882 			}
883 			if (((tmp_data->fill_len - tmp_data->resid) %
884 			     sector_size) != 0)
885 				*double_buf_needed = 1;
886 		} else {
887 			struct camdd_buf_indirect *tmp_ind;
888 
889 			tmp_ind = &tmp_buf->buf_type_spec.indirect;
890 			if (iovec == 0) {
891 				data->segs[i].ds_addr =
892 				    (bus_addr_t)tmp_ind->start_ptr;
893 				data->segs[i].ds_len = tmp_ind->len;
894 			} else {
895 				data->iovec[i].iov_base = tmp_ind->start_ptr;
896 				data->iovec[i].iov_len = tmp_ind->len;
897 			}
898 			if ((tmp_ind->len % sector_size) != 0)
899 				*double_buf_needed = 1;
900 		}
901 	}
902 
903 	if (extra_buf != NULL) {
904 		if (iovec == 0) {
905 			data->segs[i].ds_addr = (bus_addr_t)extra_buf;
906 			data->segs[i].ds_len = extra_buf_len;
907 		} else {
908 			data->iovec[i].iov_base = extra_buf;
909 			data->iovec[i].iov_len = extra_buf_len;
910 		}
911 		i++;
912 	}
913 	if ((tmp_buf != NULL) || (i != data->sg_count)) {
914 		warnx("buffer source count does not match "
915 		      "number of buffers in list!");
916 		retval = 1;
917 		goto bailout;
918 	}
919 
920 bailout:
921 	if (retval == 0) {
922 		*num_sectors_used = (data->fill_len + extra_buf_len) /
923 		    sector_size;
924 	}
925 	return (retval);
926 }
927 
928 uint32_t
929 camdd_buf_get_len(struct camdd_buf *buf)
930 {
931 	uint32_t len = 0;
932 
933 	if (buf->buf_type != CAMDD_BUF_DATA) {
934 		struct camdd_buf_indirect *indirect;
935 
936 		indirect = &buf->buf_type_spec.indirect;
937 		len = indirect->len;
938 	} else {
939 		struct camdd_buf_data *data;
940 
941 		data = &buf->buf_type_spec.data;
942 		len = data->fill_len;
943 	}
944 
945 	return (len);
946 }
947 
948 void
949 camdd_buf_add_child(struct camdd_buf *buf, struct camdd_buf *child_buf)
950 {
951 	struct camdd_buf_data *data;
952 
953 	assert(buf->buf_type == CAMDD_BUF_DATA);
954 
955 	data = &buf->buf_type_spec.data;
956 
957 	STAILQ_INSERT_TAIL(&buf->src_list, child_buf, src_links);
958 	buf->src_count++;
959 
960 	data->fill_len += camdd_buf_get_len(child_buf);
961 }
962 
963 typedef enum {
964 	CAMDD_TS_MAX_BLK,
965 	CAMDD_TS_MIN_BLK,
966 	CAMDD_TS_BLK_GRAN,
967 	CAMDD_TS_EFF_IOSIZE
968 } camdd_status_item_index;
969 
970 static struct camdd_status_items {
971 	const char *name;
972 	struct mt_status_entry *entry;
973 } req_status_items[] = {
974 	{ "max_blk", NULL },
975 	{ "min_blk", NULL },
976 	{ "blk_gran", NULL },
977 	{ "max_effective_iosize", NULL }
978 };
979 
980 int
981 camdd_probe_tape(int fd, char *filename, uint64_t *max_iosize,
982 		 uint64_t *max_blk, uint64_t *min_blk, uint64_t *blk_gran)
983 {
984 	struct mt_status_data status_data;
985 	char *xml_str = NULL;
986 	unsigned int i;
987 	int retval = 0;
988 
989 	retval = mt_get_xml_str(fd, MTIOCEXTGET, &xml_str);
990 	if (retval != 0)
991 		err(1, "Couldn't get XML string from %s", filename);
992 
993 	retval = mt_get_status(xml_str, &status_data);
994 	if (retval != XML_STATUS_OK) {
995 		warn("couldn't get status for %s", filename);
996 		retval = 1;
997 		goto bailout;
998 	} else
999 		retval = 0;
1000 
1001 	if (status_data.error != 0) {
1002 		warnx("%s", status_data.error_str);
1003 		retval = 1;
1004 		goto bailout;
1005 	}
1006 
1007 	for (i = 0; i < nitems(req_status_items); i++) {
1008                 char *name;
1009 
1010 		name = __DECONST(char *, req_status_items[i].name);
1011 		req_status_items[i].entry = mt_status_entry_find(&status_data,
1012 		    name);
1013 		if (req_status_items[i].entry == NULL) {
1014 			errx(1, "Cannot find status entry %s",
1015 			    req_status_items[i].name);
1016 		}
1017 	}
1018 
1019 	*max_iosize = req_status_items[CAMDD_TS_EFF_IOSIZE].entry->value_unsigned;
1020 	*max_blk= req_status_items[CAMDD_TS_MAX_BLK].entry->value_unsigned;
1021 	*min_blk= req_status_items[CAMDD_TS_MIN_BLK].entry->value_unsigned;
1022 	*blk_gran = req_status_items[CAMDD_TS_BLK_GRAN].entry->value_unsigned;
1023 bailout:
1024 
1025 	free(xml_str);
1026 	mt_status_free(&status_data);
1027 
1028 	return (retval);
1029 }
1030 
1031 struct camdd_dev *
1032 camdd_probe_file(int fd, struct camdd_io_opts *io_opts, int retry_count,
1033     int timeout)
1034 {
1035 	struct camdd_dev *dev = NULL;
1036 	struct camdd_dev_file *file_dev;
1037 	uint64_t blocksize = io_opts->blocksize;
1038 
1039 	dev = camdd_alloc_dev(CAMDD_DEV_FILE, NULL, 0, retry_count, timeout);
1040 	if (dev == NULL)
1041 		goto bailout;
1042 
1043 	file_dev = &dev->dev_spec.file;
1044 	file_dev->fd = fd;
1045 	strlcpy(file_dev->filename, io_opts->dev_name,
1046 	    sizeof(file_dev->filename));
1047 	strlcpy(dev->device_name, io_opts->dev_name, sizeof(dev->device_name));
1048 	if (blocksize == 0)
1049 		dev->blocksize = CAMDD_FILE_DEFAULT_BLOCK;
1050 	else
1051 		dev->blocksize = blocksize;
1052 
1053 	if ((io_opts->queue_depth != 0)
1054 	 && (io_opts->queue_depth != 1)) {
1055 		warnx("Queue depth %ju for %s ignored, only 1 outstanding "
1056 		    "command supported", (uintmax_t)io_opts->queue_depth,
1057 		    io_opts->dev_name);
1058 	}
1059 	dev->target_queue_depth = CAMDD_FILE_DEFAULT_DEPTH;
1060 	dev->run = camdd_file_run;
1061 	dev->fetch = NULL;
1062 
1063 	/*
1064 	 * We can effectively access files on byte boundaries.  We'll reset
1065 	 * this for devices like disks that can be accessed on sector
1066 	 * boundaries.
1067 	 */
1068 	dev->sector_size = 1;
1069 
1070 	if ((fd != STDIN_FILENO)
1071 	 && (fd != STDOUT_FILENO)) {
1072 		int retval;
1073 
1074 		retval = fstat(fd, &file_dev->sb);
1075 		if (retval != 0) {
1076 			warn("Cannot stat %s", dev->device_name);
1077 			goto bailout_error;
1078 		}
1079 		if (S_ISREG(file_dev->sb.st_mode)) {
1080 			file_dev->file_type = CAMDD_FILE_REG;
1081 		} else if (S_ISCHR(file_dev->sb.st_mode)) {
1082 			int type;
1083 
1084 			if (ioctl(fd, FIODTYPE, &type) == -1)
1085 				err(1, "FIODTYPE ioctl failed on %s",
1086 				    dev->device_name);
1087 			else {
1088 				if (type & D_TAPE)
1089 					file_dev->file_type = CAMDD_FILE_TAPE;
1090 				else if (type & D_DISK)
1091 					file_dev->file_type = CAMDD_FILE_DISK;
1092 				else if (type & D_MEM)
1093 					file_dev->file_type = CAMDD_FILE_MEM;
1094 				else if (type & D_TTY)
1095 					file_dev->file_type = CAMDD_FILE_TTY;
1096 			}
1097 		} else if (S_ISDIR(file_dev->sb.st_mode)) {
1098 			errx(1, "cannot operate on directory %s",
1099 			    dev->device_name);
1100 		} else if (S_ISFIFO(file_dev->sb.st_mode)) {
1101 			file_dev->file_type = CAMDD_FILE_PIPE;
1102 		} else
1103 			errx(1, "Cannot determine file type for %s",
1104 			    dev->device_name);
1105 
1106 		switch (file_dev->file_type) {
1107 		case CAMDD_FILE_REG:
1108 			if (file_dev->sb.st_size != 0)
1109 				dev->max_sector = file_dev->sb.st_size - 1;
1110 			else
1111 				dev->max_sector = 0;
1112 			file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1113 			break;
1114 		case CAMDD_FILE_TAPE: {
1115 			uint64_t max_iosize, max_blk, min_blk, blk_gran;
1116 			/*
1117 			 * Check block limits and maximum effective iosize.
1118 			 * Make sure the blocksize is within the block
1119 			 * limits (and a multiple of the minimum blocksize)
1120 			 * and that the blocksize is <= maximum effective
1121 			 * iosize.
1122 			 */
1123 			retval = camdd_probe_tape(fd, dev->device_name,
1124 			    &max_iosize, &max_blk, &min_blk, &blk_gran);
1125 			if (retval != 0)
1126 				errx(1, "Unable to probe tape %s",
1127 				    dev->device_name);
1128 
1129 			/*
1130 			 * The blocksize needs to be <= the maximum
1131 			 * effective I/O size of the tape device.  Note
1132 			 * that this also takes into account the maximum
1133 			 * blocksize reported by READ BLOCK LIMITS.
1134 			 */
1135 			if (dev->blocksize > max_iosize) {
1136 				warnx("Blocksize %u too big for %s, limiting "
1137 				    "to %ju", dev->blocksize, dev->device_name,
1138 				    max_iosize);
1139 				dev->blocksize = max_iosize;
1140 			}
1141 
1142 			/*
1143 			 * The blocksize needs to be at least min_blk;
1144 			 */
1145 			if (dev->blocksize < min_blk) {
1146 				warnx("Blocksize %u too small for %s, "
1147 				    "increasing to %ju", dev->blocksize,
1148 				    dev->device_name, min_blk);
1149 				dev->blocksize = min_blk;
1150 			}
1151 
1152 			/*
1153 			 * And the blocksize needs to be a multiple of
1154 			 * the block granularity.
1155 			 */
1156 			if ((blk_gran != 0)
1157 			 && (dev->blocksize % (1 << blk_gran))) {
1158 				warnx("Blocksize %u for %s not a multiple of "
1159 				    "%d, adjusting to %d", dev->blocksize,
1160 				    dev->device_name, (1 << blk_gran),
1161 				    dev->blocksize & ~((1 << blk_gran) - 1));
1162 				dev->blocksize &= ~((1 << blk_gran) - 1);
1163 			}
1164 
1165 			if (dev->blocksize == 0) {
1166 				errx(1, "Unable to derive valid blocksize for "
1167 				    "%s", dev->device_name);
1168 			}
1169 
1170 			/*
1171 			 * For tape drives, set the sector size to the
1172 			 * blocksize so that we make sure not to write
1173 			 * less than the blocksize out to the drive.
1174 			 */
1175 			dev->sector_size = dev->blocksize;
1176 			break;
1177 		}
1178 		case CAMDD_FILE_DISK: {
1179 			off_t media_size;
1180 			unsigned int sector_size;
1181 
1182 			file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1183 
1184 			if (ioctl(fd, DIOCGSECTORSIZE, &sector_size) == -1) {
1185 				err(1, "DIOCGSECTORSIZE ioctl failed on %s",
1186 				    dev->device_name);
1187 			}
1188 
1189 			if (sector_size == 0) {
1190 				errx(1, "DIOCGSECTORSIZE ioctl returned "
1191 				    "invalid sector size %u for %s",
1192 				    sector_size, dev->device_name);
1193 			}
1194 
1195 			if (ioctl(fd, DIOCGMEDIASIZE, &media_size) == -1) {
1196 				err(1, "DIOCGMEDIASIZE ioctl failed on %s",
1197 				    dev->device_name);
1198 			}
1199 
1200 			if (media_size == 0) {
1201 				errx(1, "DIOCGMEDIASIZE ioctl returned "
1202 				    "invalid media size %ju for %s",
1203 				    (uintmax_t)media_size, dev->device_name);
1204 			}
1205 
1206 			if (dev->blocksize % sector_size) {
1207 				errx(1, "%s blocksize %u not a multiple of "
1208 				    "sector size %u", dev->device_name,
1209 				    dev->blocksize, sector_size);
1210 			}
1211 
1212 			dev->sector_size = sector_size;
1213 			dev->max_sector = (media_size / sector_size) - 1;
1214 			break;
1215 		}
1216 		case CAMDD_FILE_MEM:
1217 			file_dev->file_flags |= CAMDD_FF_CAN_SEEK;
1218 			break;
1219 		default:
1220 			break;
1221 		}
1222 	}
1223 
1224 	if ((io_opts->offset != 0)
1225 	 && ((file_dev->file_flags & CAMDD_FF_CAN_SEEK) == 0)) {
1226 		warnx("Offset %ju specified for %s, but we cannot seek on %s",
1227 		    io_opts->offset, io_opts->dev_name, io_opts->dev_name);
1228 		goto bailout_error;
1229 	}
1230 #if 0
1231 	else if ((io_opts->offset != 0)
1232 		&& ((io_opts->offset % dev->sector_size) != 0)) {
1233 		warnx("Offset %ju for %s is not a multiple of the "
1234 		      "sector size %u", io_opts->offset,
1235 		      io_opts->dev_name, dev->sector_size);
1236 		goto bailout_error;
1237 	} else {
1238 		dev->start_offset_bytes = io_opts->offset;
1239 	}
1240 #endif
1241 
1242 bailout:
1243 	return (dev);
1244 
1245 bailout_error:
1246 	camdd_free_dev(dev);
1247 	return (NULL);
1248 }
1249 
1250 /*
1251  * Need to implement this.  Do a basic probe:
1252  * - Check the inquiry data, make sure we're talking to a device that we
1253  *   can reasonably expect to talk to -- direct, RBC, CD, WORM.
1254  * - Send a test unit ready, make sure the device is available.
1255  * - Get the capacity and block size.
1256  */
1257 struct camdd_dev *
1258 camdd_probe_pass(struct cam_device *cam_dev, struct camdd_io_opts *io_opts,
1259 		 camdd_argmask arglist, int probe_retry_count,
1260 		 int probe_timeout, int io_retry_count, int io_timeout)
1261 {
1262 	union ccb *ccb;
1263 	uint64_t maxsector;
1264 	uint32_t cpi_maxio, max_iosize, pass_numblocks;
1265 	uint32_t block_len;
1266 	struct scsi_read_capacity_data rcap;
1267 	struct scsi_read_capacity_data_long rcaplong;
1268 	struct camdd_dev *dev;
1269 	struct camdd_dev_pass *pass_dev;
1270 	struct kevent ke;
1271 	int scsi_dev_type;
1272 
1273 	dev = NULL;
1274 
1275 	scsi_dev_type = SID_TYPE(&cam_dev->inq_data);
1276 	maxsector = 0;
1277 	block_len = 0;
1278 
1279 	/*
1280 	 * For devices that support READ CAPACITY, we'll attempt to get the
1281 	 * capacity.  Otherwise, we really don't support tape or other
1282 	 * devices via SCSI passthrough, so just return an error in that case.
1283 	 */
1284 	switch (scsi_dev_type) {
1285 	case T_DIRECT:
1286 	case T_WORM:
1287 	case T_CDROM:
1288 	case T_OPTICAL:
1289 	case T_RBC:
1290 	case T_ZBC_HM:
1291 		break;
1292 	default:
1293 		errx(1, "Unsupported SCSI device type %d", scsi_dev_type);
1294 		break; /*NOTREACHED*/
1295 	}
1296 
1297 	ccb = cam_getccb(cam_dev);
1298 
1299 	if (ccb == NULL) {
1300 		warnx("%s: error allocating ccb", __func__);
1301 		goto bailout;
1302 	}
1303 
1304 	CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
1305 
1306 	scsi_read_capacity(&ccb->csio,
1307 			   /*retries*/ probe_retry_count,
1308 			   /*cbfcnp*/ NULL,
1309 			   /*tag_action*/ MSG_SIMPLE_Q_TAG,
1310 			   &rcap,
1311 			   SSD_FULL_SIZE,
1312 			   /*timeout*/ probe_timeout ? probe_timeout : 5000);
1313 
1314 	/* Disable freezing the device queue */
1315 	ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1316 
1317 	if (arglist & CAMDD_ARG_ERR_RECOVER)
1318 		ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1319 
1320 	if (cam_send_ccb(cam_dev, ccb) < 0) {
1321 		warn("error sending READ CAPACITY command");
1322 
1323 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1324 				CAM_EPF_ALL, stderr);
1325 
1326 		goto bailout;
1327 	}
1328 
1329 	if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1330 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1331 		goto bailout;
1332 	}
1333 
1334 	maxsector = scsi_4btoul(rcap.addr);
1335 	block_len = scsi_4btoul(rcap.length);
1336 
1337 	/*
1338 	 * A last block of 2^32-1 means that the true capacity is over 2TB,
1339 	 * and we need to issue the long READ CAPACITY to get the real
1340 	 * capacity.  Otherwise, we're all set.
1341 	 */
1342 	if (maxsector != 0xffffffff)
1343 		goto rcap_done;
1344 
1345 	scsi_read_capacity_16(&ccb->csio,
1346 			      /*retries*/ probe_retry_count,
1347 			      /*cbfcnp*/ NULL,
1348 			      /*tag_action*/ MSG_SIMPLE_Q_TAG,
1349 			      /*lba*/ 0,
1350 			      /*reladdr*/ 0,
1351 			      /*pmi*/ 0,
1352 			      (uint8_t *)&rcaplong,
1353 			      sizeof(rcaplong),
1354 			      /*sense_len*/ SSD_FULL_SIZE,
1355 			      /*timeout*/ probe_timeout ? probe_timeout : 5000);
1356 
1357 	/* Disable freezing the device queue */
1358 	ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
1359 
1360 	if (arglist & CAMDD_ARG_ERR_RECOVER)
1361 		ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
1362 
1363 	if (cam_send_ccb(cam_dev, ccb) < 0) {
1364 		warn("error sending READ CAPACITY (16) command");
1365 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1366 				CAM_EPF_ALL, stderr);
1367 		goto bailout;
1368 	}
1369 
1370 	if ((ccb->ccb_h.status & CAM_STATUS_MASK) != CAM_REQ_CMP) {
1371 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL, CAM_EPF_ALL, stderr);
1372 		goto bailout;
1373 	}
1374 
1375 	maxsector = scsi_8btou64(rcaplong.addr);
1376 	block_len = scsi_4btoul(rcaplong.length);
1377 
1378 rcap_done:
1379 	if (block_len == 0) {
1380 		warnx("Sector size for %s%u is 0, cannot continue",
1381 		    cam_dev->device_name, cam_dev->dev_unit_num);
1382 		goto bailout_error;
1383 	}
1384 
1385 	CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->cpi);
1386 
1387 	ccb->ccb_h.func_code = XPT_PATH_INQ;
1388 	ccb->ccb_h.flags = CAM_DIR_NONE;
1389 	ccb->ccb_h.retry_count = 1;
1390 
1391 	if (cam_send_ccb(cam_dev, ccb) < 0) {
1392 		warn("error sending XPT_PATH_INQ CCB");
1393 
1394 		cam_error_print(cam_dev, ccb, CAM_ESF_ALL,
1395 				CAM_EPF_ALL, stderr);
1396 		goto bailout;
1397 	}
1398 
1399 	EV_SET(&ke, cam_dev->fd, EVFILT_READ, EV_ADD|EV_ENABLE, 0, 0, 0);
1400 
1401 	dev = camdd_alloc_dev(CAMDD_DEV_PASS, &ke, 1, io_retry_count,
1402 			      io_timeout);
1403 	if (dev == NULL)
1404 		goto bailout;
1405 
1406 	pass_dev = &dev->dev_spec.pass;
1407 	pass_dev->scsi_dev_type = scsi_dev_type;
1408 	pass_dev->dev = cam_dev;
1409 	pass_dev->max_sector = maxsector;
1410 	pass_dev->block_len = block_len;
1411 	pass_dev->cpi_maxio = ccb->cpi.maxio;
1412 	snprintf(dev->device_name, sizeof(dev->device_name), "%s%u",
1413 		 pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
1414 	dev->sector_size = block_len;
1415 	dev->max_sector = maxsector;
1416 
1417 
1418 	/*
1419 	 * Determine the optimal blocksize to use for this device.
1420 	 */
1421 
1422 	/*
1423 	 * If the controller has not specified a maximum I/O size,
1424 	 * just go with 128K as a somewhat conservative value.
1425 	 */
1426 	if (pass_dev->cpi_maxio == 0)
1427 		cpi_maxio = 131072;
1428 	else
1429 		cpi_maxio = pass_dev->cpi_maxio;
1430 
1431 	/*
1432 	 * If the controller has a large maximum I/O size, limit it
1433 	 * to something smaller so that the kernel doesn't have trouble
1434 	 * allocating buffers to copy data in and out for us.
1435 	 * XXX KDM this is until we have unmapped I/O support in the kernel.
1436 	 */
1437 	max_iosize = min(cpi_maxio, CAMDD_PASS_MAX_BLOCK);
1438 
1439 	/*
1440 	 * If we weren't able to get a block size for some reason,
1441 	 * default to 512 bytes.
1442 	 */
1443 	block_len = pass_dev->block_len;
1444 	if (block_len == 0)
1445 		block_len = 512;
1446 
1447 	/*
1448 	 * Figure out how many blocksize chunks will fit in the
1449 	 * maximum I/O size.
1450 	 */
1451 	pass_numblocks = max_iosize / block_len;
1452 
1453 	/*
1454 	 * And finally, multiple the number of blocks by the LBA
1455 	 * length to get our maximum block size;
1456 	 */
1457 	dev->blocksize = pass_numblocks * block_len;
1458 
1459 	if (io_opts->blocksize != 0) {
1460 		if ((io_opts->blocksize % dev->sector_size) != 0) {
1461 			warnx("Blocksize %ju for %s is not a multiple of "
1462 			      "sector size %u", (uintmax_t)io_opts->blocksize,
1463 			      dev->device_name, dev->sector_size);
1464 			goto bailout_error;
1465 		}
1466 		dev->blocksize = io_opts->blocksize;
1467 	}
1468 	dev->target_queue_depth = CAMDD_PASS_DEFAULT_DEPTH;
1469 	if (io_opts->queue_depth != 0)
1470 		dev->target_queue_depth = io_opts->queue_depth;
1471 
1472 	if (io_opts->offset != 0) {
1473 		if (io_opts->offset > (dev->max_sector * dev->sector_size)) {
1474 			warnx("Offset %ju is past the end of device %s",
1475 			    io_opts->offset, dev->device_name);
1476 			goto bailout_error;
1477 		}
1478 #if 0
1479 		else if ((io_opts->offset % dev->sector_size) != 0) {
1480 			warnx("Offset %ju for %s is not a multiple of the "
1481 			      "sector size %u", io_opts->offset,
1482 			      dev->device_name, dev->sector_size);
1483 			goto bailout_error;
1484 		}
1485 		dev->start_offset_bytes = io_opts->offset;
1486 #endif
1487 	}
1488 
1489 	dev->min_cmd_size = io_opts->min_cmd_size;
1490 
1491 	dev->run = camdd_pass_run;
1492 	dev->fetch = camdd_pass_fetch;
1493 
1494 bailout:
1495 	cam_freeccb(ccb);
1496 
1497 	return (dev);
1498 
1499 bailout_error:
1500 	cam_freeccb(ccb);
1501 
1502 	camdd_free_dev(dev);
1503 
1504 	return (NULL);
1505 }
1506 
1507 void *
1508 camdd_worker(void *arg)
1509 {
1510 	struct camdd_dev *dev = arg;
1511 	struct camdd_buf *buf;
1512 	struct timespec ts, *kq_ts;
1513 
1514 	ts.tv_sec = 0;
1515 	ts.tv_nsec = 0;
1516 
1517 	pthread_mutex_lock(&dev->mutex);
1518 
1519 	dev->flags |= CAMDD_DEV_FLAG_ACTIVE;
1520 
1521 	for (;;) {
1522 		struct kevent ke;
1523 		int retval = 0;
1524 
1525 		/*
1526 		 * XXX KDM check the reorder queue depth?
1527 		 */
1528 		if (dev->write_dev == 0) {
1529 			uint32_t our_depth, peer_depth, peer_bytes, our_bytes;
1530 			uint32_t target_depth = dev->target_queue_depth;
1531 			uint32_t peer_target_depth =
1532 			    dev->peer_dev->target_queue_depth;
1533 			uint32_t peer_blocksize = dev->peer_dev->blocksize;
1534 
1535 			camdd_get_depth(dev, &our_depth, &peer_depth,
1536 					&our_bytes, &peer_bytes);
1537 
1538 #if 0
1539 			while (((our_depth < target_depth)
1540 			     && (peer_depth < peer_target_depth))
1541 			    || ((peer_bytes + our_bytes) <
1542 				 (peer_blocksize * 2))) {
1543 #endif
1544 			while (((our_depth + peer_depth) <
1545 			        (target_depth + peer_target_depth))
1546 			    || ((peer_bytes + our_bytes) <
1547 				(peer_blocksize * 3))) {
1548 
1549 				retval = camdd_queue(dev, NULL);
1550 				if (retval == 1)
1551 					break;
1552 				else if (retval != 0) {
1553 					error_exit = 1;
1554 					goto bailout;
1555 				}
1556 
1557 				camdd_get_depth(dev, &our_depth, &peer_depth,
1558 						&our_bytes, &peer_bytes);
1559 			}
1560 		}
1561 		/*
1562 		 * See if we have any I/O that is ready to execute.
1563 		 */
1564 		buf = STAILQ_FIRST(&dev->run_queue);
1565 		if (buf != NULL) {
1566 			while (dev->target_queue_depth > dev->cur_active_io) {
1567 				retval = dev->run(dev);
1568 				if (retval == -1) {
1569 					dev->flags |= CAMDD_DEV_FLAG_EOF;
1570 					error_exit = 1;
1571 					break;
1572 				} else if (retval != 0) {
1573 					break;
1574 				}
1575 			}
1576 		}
1577 
1578 		/*
1579 		 * We've reached EOF, or our partner has reached EOF.
1580 		 */
1581 		if ((dev->flags & CAMDD_DEV_FLAG_EOF)
1582 		 || (dev->flags & CAMDD_DEV_FLAG_PEER_EOF)) {
1583 			if (dev->write_dev != 0) {
1584 			 	if ((STAILQ_EMPTY(&dev->work_queue))
1585 				 && (dev->num_run_queue == 0)
1586 				 && (dev->cur_active_io == 0)) {
1587 					goto bailout;
1588 				}
1589 			} else {
1590 				/*
1591 				 * If we're the reader, and the writer
1592 				 * got EOF, he is already done.  If we got
1593 				 * the EOF, then we need to wait until
1594 				 * everything is flushed out for the writer.
1595 				 */
1596 				if (dev->flags & CAMDD_DEV_FLAG_PEER_EOF) {
1597 					goto bailout;
1598 				} else if ((dev->num_peer_work_queue == 0)
1599 					&& (dev->num_peer_done_queue == 0)
1600 					&& (dev->cur_active_io == 0)
1601 					&& (dev->num_run_queue == 0)) {
1602 					goto bailout;
1603 				}
1604 			}
1605 			/*
1606 			 * XXX KDM need to do something about the pending
1607 			 * queue and cleanup resources.
1608 			 */
1609 		}
1610 
1611 		if ((dev->write_dev == 0)
1612 		 && (dev->cur_active_io == 0)
1613 		 && (dev->peer_bytes_queued < dev->peer_dev->blocksize))
1614 			kq_ts = &ts;
1615 		else
1616 			kq_ts = NULL;
1617 
1618 		/*
1619 		 * Run kevent to see if there are events to process.
1620 		 */
1621 		pthread_mutex_unlock(&dev->mutex);
1622 		retval = kevent(dev->kq, NULL, 0, &ke, 1, kq_ts);
1623 		pthread_mutex_lock(&dev->mutex);
1624 		if (retval == -1) {
1625 			warn("%s: error returned from kevent",__func__);
1626 			goto bailout;
1627 		} else if (retval != 0) {
1628 			switch (ke.filter) {
1629 			case EVFILT_READ:
1630 				if (dev->fetch != NULL) {
1631 					retval = dev->fetch(dev);
1632 					if (retval == -1) {
1633 						error_exit = 1;
1634 						goto bailout;
1635 					}
1636 				}
1637 				break;
1638 			case EVFILT_SIGNAL:
1639 				/*
1640 				 * We register for this so we don't get
1641 				 * an error as a result of a SIGINFO or a
1642 				 * SIGINT.  It will actually get handled
1643 				 * by the signal handler.  If we get a
1644 				 * SIGINT, bail out without printing an
1645 				 * error message.  Any other signals
1646 				 * will result in the error message above.
1647 				 */
1648 				if (ke.ident == SIGINT)
1649 					goto bailout;
1650 				break;
1651 			case EVFILT_USER:
1652 				retval = 0;
1653 				/*
1654 				 * Check to see if the other thread has
1655 				 * queued any I/O for us to do.  (In this
1656 				 * case we're the writer.)
1657 				 */
1658 				for (buf = STAILQ_FIRST(&dev->work_queue);
1659 				     buf != NULL;
1660 				     buf = STAILQ_FIRST(&dev->work_queue)) {
1661 					STAILQ_REMOVE_HEAD(&dev->work_queue,
1662 							   work_links);
1663 					retval = camdd_queue(dev, buf);
1664 					/*
1665 					 * We keep going unless we get an
1666 					 * actual error.  If we get EOF, we
1667 					 * still want to remove the buffers
1668 					 * from the queue and send the back
1669 					 * to the reader thread.
1670 					 */
1671 					if (retval == -1) {
1672 						error_exit = 1;
1673 						goto bailout;
1674 					} else
1675 						retval = 0;
1676 				}
1677 
1678 				/*
1679 				 * Next check to see if the other thread has
1680 				 * queued any completed buffers back to us.
1681 				 * (In this case we're the reader.)
1682 				 */
1683 				for (buf = STAILQ_FIRST(&dev->peer_done_queue);
1684 				     buf != NULL;
1685 				     buf = STAILQ_FIRST(&dev->peer_done_queue)){
1686 					STAILQ_REMOVE_HEAD(
1687 					    &dev->peer_done_queue, work_links);
1688 					dev->num_peer_done_queue--;
1689 					camdd_peer_done(buf);
1690 				}
1691 				break;
1692 			default:
1693 				warnx("%s: unknown kevent filter %d",
1694 				      __func__, ke.filter);
1695 				break;
1696 			}
1697 		}
1698 	}
1699 
1700 bailout:
1701 
1702 	dev->flags &= ~CAMDD_DEV_FLAG_ACTIVE;
1703 
1704 	/* XXX KDM cleanup resources here? */
1705 
1706 	pthread_mutex_unlock(&dev->mutex);
1707 
1708 	need_exit = 1;
1709 	sem_post(&camdd_sem);
1710 
1711 	return (NULL);
1712 }
1713 
1714 /*
1715  * Simplistic translation of CCB status to our local status.
1716  */
1717 camdd_buf_status
1718 camdd_ccb_status(union ccb *ccb)
1719 {
1720 	camdd_buf_status status = CAMDD_STATUS_NONE;
1721 	cam_status ccb_status;
1722 
1723 	ccb_status = ccb->ccb_h.status & CAM_STATUS_MASK;
1724 
1725 	switch (ccb_status) {
1726 	case CAM_REQ_CMP: {
1727 		if (ccb->csio.resid == 0) {
1728 			status = CAMDD_STATUS_OK;
1729 		} else if (ccb->csio.dxfer_len > ccb->csio.resid) {
1730 			status = CAMDD_STATUS_SHORT_IO;
1731 		} else {
1732 			status = CAMDD_STATUS_EOF;
1733 		}
1734 		break;
1735 	}
1736 	case CAM_SCSI_STATUS_ERROR: {
1737 		switch (ccb->csio.scsi_status) {
1738 		case SCSI_STATUS_OK:
1739 		case SCSI_STATUS_COND_MET:
1740 		case SCSI_STATUS_INTERMED:
1741 		case SCSI_STATUS_INTERMED_COND_MET:
1742 			status = CAMDD_STATUS_OK;
1743 			break;
1744 		case SCSI_STATUS_CMD_TERMINATED:
1745 		case SCSI_STATUS_CHECK_COND:
1746 		case SCSI_STATUS_QUEUE_FULL:
1747 		case SCSI_STATUS_BUSY:
1748 		case SCSI_STATUS_RESERV_CONFLICT:
1749 		default:
1750 			status = CAMDD_STATUS_ERROR;
1751 			break;
1752 		}
1753 		break;
1754 	}
1755 	default:
1756 		status = CAMDD_STATUS_ERROR;
1757 		break;
1758 	}
1759 
1760 	return (status);
1761 }
1762 
1763 /*
1764  * Queue a buffer to our peer's work thread for writing.
1765  *
1766  * Returns 0 for success, -1 for failure, 1 if the other thread exited.
1767  */
1768 int
1769 camdd_queue_peer_buf(struct camdd_dev *dev, struct camdd_buf *buf)
1770 {
1771 	struct kevent ke;
1772 	STAILQ_HEAD(, camdd_buf) local_queue;
1773 	struct camdd_buf *buf1, *buf2;
1774 	struct camdd_buf_data *data = NULL;
1775 	uint64_t peer_bytes_queued = 0;
1776 	int active = 1;
1777 	int retval = 0;
1778 
1779 	STAILQ_INIT(&local_queue);
1780 
1781 	/*
1782 	 * Since we're the reader, we need to queue our I/O to the writer
1783 	 * in sequential order in order to make sure it gets written out
1784 	 * in sequential order.
1785 	 *
1786 	 * Check the next expected I/O starting offset.  If this doesn't
1787 	 * match, put it on the reorder queue.
1788 	 */
1789 	if ((buf->lba * dev->sector_size) != dev->next_completion_pos_bytes) {
1790 
1791 		/*
1792 		 * If there is nothing on the queue, there is no sorting
1793 		 * needed.
1794 		 */
1795 		if (STAILQ_EMPTY(&dev->reorder_queue)) {
1796 			STAILQ_INSERT_TAIL(&dev->reorder_queue, buf, links);
1797 			dev->num_reorder_queue++;
1798 			goto bailout;
1799 		}
1800 
1801 		/*
1802 		 * Sort in ascending order by starting LBA.  There should
1803 		 * be no identical LBAs.
1804 		 */
1805 		for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1806 		     buf1 = buf2) {
1807 			buf2 = STAILQ_NEXT(buf1, links);
1808 			if (buf->lba < buf1->lba) {
1809 				/*
1810 				 * If we're less than the first one, then
1811 				 * we insert at the head of the list
1812 				 * because this has to be the first element
1813 				 * on the list.
1814 				 */
1815 				STAILQ_INSERT_HEAD(&dev->reorder_queue,
1816 						   buf, links);
1817 				dev->num_reorder_queue++;
1818 				break;
1819 			} else if (buf->lba > buf1->lba) {
1820 				if (buf2 == NULL) {
1821 					STAILQ_INSERT_TAIL(&dev->reorder_queue,
1822 					    buf, links);
1823 					dev->num_reorder_queue++;
1824 					break;
1825 				} else if (buf->lba < buf2->lba) {
1826 					STAILQ_INSERT_AFTER(&dev->reorder_queue,
1827 					    buf1, buf, links);
1828 					dev->num_reorder_queue++;
1829 					break;
1830 				}
1831 			} else {
1832 				errx(1, "Found buffers with duplicate LBA %ju!",
1833 				     buf->lba);
1834 			}
1835 		}
1836 		goto bailout;
1837 	} else {
1838 
1839 		/*
1840 		 * We're the next expected I/O completion, so put ourselves
1841 		 * on the local queue to be sent to the writer.  We use
1842 		 * work_links here so that we can queue this to the
1843 		 * peer_work_queue before taking the buffer off of the
1844 		 * local_queue.
1845 		 */
1846 		dev->next_completion_pos_bytes += buf->len;
1847 		STAILQ_INSERT_TAIL(&local_queue, buf, work_links);
1848 
1849 		/*
1850 		 * Go through the reorder queue looking for more sequential
1851 		 * I/O and add it to the local queue.
1852 		 */
1853 		for (buf1 = STAILQ_FIRST(&dev->reorder_queue); buf1 != NULL;
1854 		     buf1 = STAILQ_FIRST(&dev->reorder_queue)) {
1855 			/*
1856 			 * As soon as we see an I/O that is out of sequence,
1857 			 * we're done.
1858 			 */
1859 			if ((buf1->lba * dev->sector_size) !=
1860 			     dev->next_completion_pos_bytes)
1861 				break;
1862 
1863 			STAILQ_REMOVE_HEAD(&dev->reorder_queue, links);
1864 			dev->num_reorder_queue--;
1865 			STAILQ_INSERT_TAIL(&local_queue, buf1, work_links);
1866 			dev->next_completion_pos_bytes += buf1->len;
1867 		}
1868 	}
1869 
1870 	/*
1871 	 * Setup the event to let the other thread know that it has work
1872 	 * pending.
1873 	 */
1874 	EV_SET(&ke, (uintptr_t)&dev->peer_dev->work_queue, EVFILT_USER, 0,
1875 	       NOTE_TRIGGER, 0, NULL);
1876 
1877 	/*
1878 	 * Put this on our shadow queue so that we know what we've queued
1879 	 * to the other thread.
1880 	 */
1881 	STAILQ_FOREACH_SAFE(buf1, &local_queue, work_links, buf2) {
1882 		if (buf1->buf_type != CAMDD_BUF_DATA) {
1883 			errx(1, "%s: should have a data buffer, not an "
1884 			    "indirect buffer", __func__);
1885 		}
1886 		data = &buf1->buf_type_spec.data;
1887 
1888 		/*
1889 		 * We only need to send one EOF to the writer, and don't
1890 		 * need to continue sending EOFs after that.
1891 		 */
1892 		if (buf1->status == CAMDD_STATUS_EOF) {
1893 			if (dev->flags & CAMDD_DEV_FLAG_EOF_SENT) {
1894 				STAILQ_REMOVE(&local_queue, buf1, camdd_buf,
1895 				    work_links);
1896 				camdd_release_buf(buf1);
1897 				retval = 1;
1898 				continue;
1899 			}
1900 			dev->flags |= CAMDD_DEV_FLAG_EOF_SENT;
1901 		}
1902 
1903 
1904 		STAILQ_INSERT_TAIL(&dev->peer_work_queue, buf1, links);
1905 		peer_bytes_queued += (data->fill_len - data->resid);
1906 		dev->peer_bytes_queued += (data->fill_len - data->resid);
1907 		dev->num_peer_work_queue++;
1908 	}
1909 
1910 	if (STAILQ_FIRST(&local_queue) == NULL)
1911 		goto bailout;
1912 
1913 	/*
1914 	 * Drop our mutex and pick up the other thread's mutex.  We need to
1915 	 * do this to avoid deadlocks.
1916 	 */
1917 	pthread_mutex_unlock(&dev->mutex);
1918 	pthread_mutex_lock(&dev->peer_dev->mutex);
1919 
1920 	if (dev->peer_dev->flags & CAMDD_DEV_FLAG_ACTIVE) {
1921 		/*
1922 		 * Put the buffers on the other thread's incoming work queue.
1923 		 */
1924 		for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1925 		     buf1 = STAILQ_FIRST(&local_queue)) {
1926 			STAILQ_REMOVE_HEAD(&local_queue, work_links);
1927 			STAILQ_INSERT_TAIL(&dev->peer_dev->work_queue, buf1,
1928 					   work_links);
1929 		}
1930 		/*
1931 		 * Send an event to the other thread's kqueue to let it know
1932 		 * that there is something on the work queue.
1933 		 */
1934 		retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
1935 		if (retval == -1)
1936 			warn("%s: unable to add peer work_queue kevent",
1937 			     __func__);
1938 		else
1939 			retval = 0;
1940 	} else
1941 		active = 0;
1942 
1943 	pthread_mutex_unlock(&dev->peer_dev->mutex);
1944 	pthread_mutex_lock(&dev->mutex);
1945 
1946 	/*
1947 	 * If the other side isn't active, run through the queue and
1948 	 * release all of the buffers.
1949 	 */
1950 	if (active == 0) {
1951 		for (buf1 = STAILQ_FIRST(&local_queue); buf1 != NULL;
1952 		     buf1 = STAILQ_FIRST(&local_queue)) {
1953 			STAILQ_REMOVE_HEAD(&local_queue, work_links);
1954 			STAILQ_REMOVE(&dev->peer_work_queue, buf1, camdd_buf,
1955 				      links);
1956 			dev->num_peer_work_queue--;
1957 			camdd_release_buf(buf1);
1958 		}
1959 		dev->peer_bytes_queued -= peer_bytes_queued;
1960 		retval = 1;
1961 	}
1962 
1963 bailout:
1964 	return (retval);
1965 }
1966 
1967 /*
1968  * Return a buffer to the reader thread when we have completed writing it.
1969  */
1970 int
1971 camdd_complete_peer_buf(struct camdd_dev *dev, struct camdd_buf *peer_buf)
1972 {
1973 	struct kevent ke;
1974 	int retval = 0;
1975 
1976 	/*
1977 	 * Setup the event to let the other thread know that we have
1978 	 * completed a buffer.
1979 	 */
1980 	EV_SET(&ke, (uintptr_t)&dev->peer_dev->peer_done_queue, EVFILT_USER, 0,
1981 	       NOTE_TRIGGER, 0, NULL);
1982 
1983 	/*
1984 	 * Drop our lock and acquire the other thread's lock before
1985 	 * manipulating
1986 	 */
1987 	pthread_mutex_unlock(&dev->mutex);
1988 	pthread_mutex_lock(&dev->peer_dev->mutex);
1989 
1990 	/*
1991 	 * Put the buffer on the reader thread's peer done queue now that
1992 	 * we have completed it.
1993 	 */
1994 	STAILQ_INSERT_TAIL(&dev->peer_dev->peer_done_queue, peer_buf,
1995 			   work_links);
1996 	dev->peer_dev->num_peer_done_queue++;
1997 
1998 	/*
1999 	 * Send an event to the peer thread to let it know that we've added
2000 	 * something to its peer done queue.
2001 	 */
2002 	retval = kevent(dev->peer_dev->kq, &ke, 1, NULL, 0, NULL);
2003 	if (retval == -1)
2004 		warn("%s: unable to add peer_done_queue kevent", __func__);
2005 	else
2006 		retval = 0;
2007 
2008 	/*
2009 	 * Drop the other thread's lock and reacquire ours.
2010 	 */
2011 	pthread_mutex_unlock(&dev->peer_dev->mutex);
2012 	pthread_mutex_lock(&dev->mutex);
2013 
2014 	return (retval);
2015 }
2016 
2017 /*
2018  * Free a buffer that was written out by the writer thread and returned to
2019  * the reader thread.
2020  */
2021 void
2022 camdd_peer_done(struct camdd_buf *buf)
2023 {
2024 	struct camdd_dev *dev;
2025 	struct camdd_buf_data *data;
2026 
2027 	dev = buf->dev;
2028 	if (buf->buf_type != CAMDD_BUF_DATA) {
2029 		errx(1, "%s: should have a data buffer, not an "
2030 		    "indirect buffer", __func__);
2031 	}
2032 
2033 	data = &buf->buf_type_spec.data;
2034 
2035 	STAILQ_REMOVE(&dev->peer_work_queue, buf, camdd_buf, links);
2036 	dev->num_peer_work_queue--;
2037 	dev->peer_bytes_queued -= (data->fill_len - data->resid);
2038 
2039 	if (buf->status == CAMDD_STATUS_EOF)
2040 		dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2041 
2042 	STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2043 }
2044 
2045 /*
2046  * Assumes caller holds the lock for this device.
2047  */
2048 void
2049 camdd_complete_buf(struct camdd_dev *dev, struct camdd_buf *buf,
2050 		   int *error_count)
2051 {
2052 	int retval = 0;
2053 
2054 	/*
2055 	 * If we're the reader, we need to send the completed I/O
2056 	 * to the writer.  If we're the writer, we need to just
2057 	 * free up resources, or let the reader know if we've
2058 	 * encountered an error.
2059 	 */
2060 	if (dev->write_dev == 0) {
2061 		retval = camdd_queue_peer_buf(dev, buf);
2062 		if (retval != 0)
2063 			(*error_count)++;
2064 	} else {
2065 		struct camdd_buf *tmp_buf, *next_buf;
2066 
2067 		STAILQ_FOREACH_SAFE(tmp_buf, &buf->src_list, src_links,
2068 				    next_buf) {
2069 			struct camdd_buf *src_buf;
2070 			struct camdd_buf_indirect *indirect;
2071 
2072 			STAILQ_REMOVE(&buf->src_list, tmp_buf,
2073 				      camdd_buf, src_links);
2074 
2075 			tmp_buf->status = buf->status;
2076 
2077 			if (tmp_buf->buf_type == CAMDD_BUF_DATA) {
2078 				camdd_complete_peer_buf(dev, tmp_buf);
2079 				continue;
2080 			}
2081 
2082 			indirect = &tmp_buf->buf_type_spec.indirect;
2083 			src_buf = indirect->src_buf;
2084 			src_buf->refcount--;
2085 			/*
2086 			 * XXX KDM we probably need to account for
2087 			 * exactly how many bytes we were able to
2088 			 * write.  Allocate the residual to the
2089 			 * first N buffers?  Or just track the
2090 			 * number of bytes written?  Right now the reader
2091 			 * doesn't do anything with a residual.
2092 			 */
2093 			src_buf->status = buf->status;
2094 			if (src_buf->refcount <= 0)
2095 				camdd_complete_peer_buf(dev, src_buf);
2096 			STAILQ_INSERT_TAIL(&dev->free_indirect_queue,
2097 					   tmp_buf, links);
2098 		}
2099 
2100 		STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2101 	}
2102 }
2103 
2104 /*
2105  * Fetch all completed commands from the pass(4) device.
2106  *
2107  * Returns the number of commands received, or -1 if any of the commands
2108  * completed with an error.  Returns 0 if no commands are available.
2109  */
2110 int
2111 camdd_pass_fetch(struct camdd_dev *dev)
2112 {
2113 	struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2114 	union ccb ccb;
2115 	int retval = 0, num_fetched = 0, error_count = 0;
2116 
2117 	pthread_mutex_unlock(&dev->mutex);
2118 	/*
2119 	 * XXX KDM we don't distinguish between EFAULT and ENOENT.
2120 	 */
2121 	while ((retval = ioctl(pass_dev->dev->fd, CAMIOGET, &ccb)) != -1) {
2122 		struct camdd_buf *buf;
2123 		struct camdd_buf_data *data;
2124 		cam_status ccb_status;
2125 		union ccb *buf_ccb;
2126 
2127 		buf = ccb.ccb_h.ccb_buf;
2128 		data = &buf->buf_type_spec.data;
2129 		buf_ccb = &data->ccb;
2130 
2131 		num_fetched++;
2132 
2133 		/*
2134 		 * Copy the CCB back out so we get status, sense data, etc.
2135 		 */
2136 		bcopy(&ccb, buf_ccb, sizeof(ccb));
2137 
2138 		pthread_mutex_lock(&dev->mutex);
2139 
2140 		/*
2141 		 * We're now done, so take this off the active queue.
2142 		 */
2143 		STAILQ_REMOVE(&dev->active_queue, buf, camdd_buf, links);
2144 		dev->cur_active_io--;
2145 
2146 		ccb_status = ccb.ccb_h.status & CAM_STATUS_MASK;
2147 		if (ccb_status != CAM_REQ_CMP) {
2148 			cam_error_print(pass_dev->dev, &ccb, CAM_ESF_ALL,
2149 					CAM_EPF_ALL, stderr);
2150 		}
2151 
2152 		data->resid = ccb.csio.resid;
2153 		dev->bytes_transferred += (ccb.csio.dxfer_len - ccb.csio.resid);
2154 
2155 		if (buf->status == CAMDD_STATUS_NONE)
2156 			buf->status = camdd_ccb_status(&ccb);
2157 		if (buf->status == CAMDD_STATUS_ERROR)
2158 			error_count++;
2159 		else if (buf->status == CAMDD_STATUS_EOF) {
2160 			/*
2161 			 * Once we queue this buffer to our partner thread,
2162 			 * he will know that we've hit EOF.
2163 			 */
2164 			dev->flags |= CAMDD_DEV_FLAG_EOF;
2165 		}
2166 
2167 		camdd_complete_buf(dev, buf, &error_count);
2168 
2169 		/*
2170 		 * Unlock in preparation for the ioctl call.
2171 		 */
2172 		pthread_mutex_unlock(&dev->mutex);
2173 	}
2174 
2175 	pthread_mutex_lock(&dev->mutex);
2176 
2177 	if (error_count > 0)
2178 		return (-1);
2179 	else
2180 		return (num_fetched);
2181 }
2182 
2183 /*
2184  * Returns -1 for error, 0 for success/continue, and 1 for resource
2185  * shortage/stop processing.
2186  */
2187 int
2188 camdd_file_run(struct camdd_dev *dev)
2189 {
2190 	struct camdd_dev_file *file_dev = &dev->dev_spec.file;
2191 	struct camdd_buf_data *data;
2192 	struct camdd_buf *buf;
2193 	off_t io_offset;
2194 	int retval = 0, write_dev = dev->write_dev;
2195 	int error_count = 0, no_resources = 0, double_buf_needed = 0;
2196 	uint32_t num_sectors = 0, db_len = 0;
2197 
2198 	buf = STAILQ_FIRST(&dev->run_queue);
2199 	if (buf == NULL) {
2200 		no_resources = 1;
2201 		goto bailout;
2202 	} else if ((dev->write_dev == 0)
2203 		&& (dev->flags & (CAMDD_DEV_FLAG_EOF |
2204 				  CAMDD_DEV_FLAG_EOF_SENT))) {
2205 		STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2206 		dev->num_run_queue--;
2207 		buf->status = CAMDD_STATUS_EOF;
2208 		error_count++;
2209 		goto bailout;
2210 	}
2211 
2212 	/*
2213 	 * If we're writing, we need to go through the source buffer list
2214 	 * and create an S/G list.
2215 	 */
2216 	if (write_dev != 0) {
2217 		retval = camdd_buf_sg_create(buf, /*iovec*/ 1,
2218 		    dev->sector_size, &num_sectors, &double_buf_needed);
2219 		if (retval != 0) {
2220 			no_resources = 1;
2221 			goto bailout;
2222 		}
2223 	}
2224 
2225 	STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2226 	dev->num_run_queue--;
2227 
2228 	data = &buf->buf_type_spec.data;
2229 
2230 	/*
2231 	 * pread(2) and pwrite(2) offsets are byte offsets.
2232 	 */
2233 	io_offset = buf->lba * dev->sector_size;
2234 
2235 	/*
2236 	 * Unlock the mutex while we read or write.
2237 	 */
2238 	pthread_mutex_unlock(&dev->mutex);
2239 
2240 	/*
2241 	 * Note that we don't need to double buffer if we're the reader
2242 	 * because in that case, we have allocated a single buffer of
2243 	 * sufficient size to do the read.  This copy is necessary on
2244 	 * writes because if one of the components of the S/G list is not
2245 	 * a sector size multiple, the kernel will reject the write.  This
2246 	 * is unfortunate but not surprising.  So this will make sure that
2247 	 * we're using a single buffer that is a multiple of the sector size.
2248 	 */
2249 	if ((double_buf_needed != 0)
2250 	 && (data->sg_count > 1)
2251 	 && (write_dev != 0)) {
2252 		uint32_t cur_offset;
2253 		int i;
2254 
2255 		if (file_dev->tmp_buf == NULL)
2256 			file_dev->tmp_buf = calloc(dev->blocksize, 1);
2257 		if (file_dev->tmp_buf == NULL) {
2258 			buf->status = CAMDD_STATUS_ERROR;
2259 			error_count++;
2260 			pthread_mutex_lock(&dev->mutex);
2261 			goto bailout;
2262 		}
2263 		for (i = 0, cur_offset = 0; i < data->sg_count; i++) {
2264 			bcopy(data->iovec[i].iov_base,
2265 			    &file_dev->tmp_buf[cur_offset],
2266 			    data->iovec[i].iov_len);
2267 			cur_offset += data->iovec[i].iov_len;
2268 		}
2269 		db_len = cur_offset;
2270 	}
2271 
2272 	if (file_dev->file_flags & CAMDD_FF_CAN_SEEK) {
2273 		if (write_dev == 0) {
2274 			/*
2275 			 * XXX KDM is there any way we would need a S/G
2276 			 * list here?
2277 			 */
2278 			retval = pread(file_dev->fd, data->buf,
2279 			    buf->len, io_offset);
2280 		} else {
2281 			if (double_buf_needed != 0) {
2282 				retval = pwrite(file_dev->fd, file_dev->tmp_buf,
2283 				    db_len, io_offset);
2284 			} else if (data->sg_count == 0) {
2285 				retval = pwrite(file_dev->fd, data->buf,
2286 				    data->fill_len, io_offset);
2287 			} else {
2288 				retval = pwritev(file_dev->fd, data->iovec,
2289 				    data->sg_count, io_offset);
2290 			}
2291 		}
2292 	} else {
2293 		if (write_dev == 0) {
2294 			/*
2295 			 * XXX KDM is there any way we would need a S/G
2296 			 * list here?
2297 			 */
2298 			retval = read(file_dev->fd, data->buf, buf->len);
2299 		} else {
2300 			if (double_buf_needed != 0) {
2301 				retval = write(file_dev->fd, file_dev->tmp_buf,
2302 				    db_len);
2303 			} else if (data->sg_count == 0) {
2304 				retval = write(file_dev->fd, data->buf,
2305 				    data->fill_len);
2306 			} else {
2307 				retval = writev(file_dev->fd, data->iovec,
2308 				    data->sg_count);
2309 			}
2310 		}
2311 	}
2312 
2313 	/* We're done, re-acquire the lock */
2314 	pthread_mutex_lock(&dev->mutex);
2315 
2316 	if (retval >= (ssize_t)data->fill_len) {
2317 		/*
2318 		 * If the bytes transferred is more than the request size,
2319 		 * that indicates an overrun, which should only happen at
2320 		 * the end of a transfer if we have to round up to a sector
2321 		 * boundary.
2322 		 */
2323 		if (buf->status == CAMDD_STATUS_NONE)
2324 			buf->status = CAMDD_STATUS_OK;
2325 		data->resid = 0;
2326 		dev->bytes_transferred += retval;
2327 	} else if (retval == -1) {
2328 		warn("Error %s %s", (write_dev) ? "writing to" :
2329 		    "reading from", file_dev->filename);
2330 
2331 		buf->status = CAMDD_STATUS_ERROR;
2332 		data->resid = data->fill_len;
2333 		error_count++;
2334 
2335 		if (dev->debug == 0)
2336 			goto bailout;
2337 
2338 		if ((double_buf_needed != 0)
2339 		 && (write_dev != 0)) {
2340 			fprintf(stderr, "%s: fd %d, DB buf %p, len %u lba %ju "
2341 			    "offset %ju\n", __func__, file_dev->fd,
2342 			    file_dev->tmp_buf, db_len, (uintmax_t)buf->lba,
2343 			    (uintmax_t)io_offset);
2344 		} else if (data->sg_count == 0) {
2345 			fprintf(stderr, "%s: fd %d, buf %p, len %u, lba %ju "
2346 			    "offset %ju\n", __func__, file_dev->fd, data->buf,
2347 			    data->fill_len, (uintmax_t)buf->lba,
2348 			    (uintmax_t)io_offset);
2349 		} else {
2350 			int i;
2351 
2352 			fprintf(stderr, "%s: fd %d, len %u, lba %ju "
2353 			    "offset %ju\n", __func__, file_dev->fd,
2354 			    data->fill_len, (uintmax_t)buf->lba,
2355 			    (uintmax_t)io_offset);
2356 
2357 			for (i = 0; i < data->sg_count; i++) {
2358 				fprintf(stderr, "index %d ptr %p len %zu\n",
2359 				    i, data->iovec[i].iov_base,
2360 				    data->iovec[i].iov_len);
2361 			}
2362 		}
2363 	} else if (retval == 0) {
2364 		buf->status = CAMDD_STATUS_EOF;
2365 		if (dev->debug != 0)
2366 			printf("%s: got EOF from %s!\n", __func__,
2367 			    file_dev->filename);
2368 		data->resid = data->fill_len;
2369 		error_count++;
2370 	} else if (retval < (ssize_t)data->fill_len) {
2371 		if (buf->status == CAMDD_STATUS_NONE)
2372 			buf->status = CAMDD_STATUS_SHORT_IO;
2373 		data->resid = data->fill_len - retval;
2374 		dev->bytes_transferred += retval;
2375 	}
2376 
2377 bailout:
2378 	if (buf != NULL) {
2379 		if (buf->status == CAMDD_STATUS_EOF) {
2380 			struct camdd_buf *buf2;
2381 			dev->flags |= CAMDD_DEV_FLAG_EOF;
2382 			STAILQ_FOREACH(buf2, &dev->run_queue, links)
2383 				buf2->status = CAMDD_STATUS_EOF;
2384 		}
2385 
2386 		camdd_complete_buf(dev, buf, &error_count);
2387 	}
2388 
2389 	if (error_count != 0)
2390 		return (-1);
2391 	else if (no_resources != 0)
2392 		return (1);
2393 	else
2394 		return (0);
2395 }
2396 
2397 /*
2398  * Execute one command from the run queue.  Returns 0 for success, 1 for
2399  * stop processing, and -1 for error.
2400  */
2401 int
2402 camdd_pass_run(struct camdd_dev *dev)
2403 {
2404 	struct camdd_buf *buf = NULL;
2405 	struct camdd_dev_pass *pass_dev = &dev->dev_spec.pass;
2406 	struct camdd_buf_data *data;
2407 	uint32_t num_blocks, sectors_used = 0;
2408 	union ccb *ccb;
2409 	int retval = 0, is_write = dev->write_dev;
2410 	int double_buf_needed = 0;
2411 
2412 	buf = STAILQ_FIRST(&dev->run_queue);
2413 	if (buf == NULL) {
2414 		retval = 1;
2415 		goto bailout;
2416 	}
2417 
2418 	/*
2419 	 * If we're writing, we need to go through the source buffer list
2420 	 * and create an S/G list.
2421 	 */
2422 	if (is_write != 0) {
2423 		retval = camdd_buf_sg_create(buf, /*iovec*/ 0,dev->sector_size,
2424 		    &sectors_used, &double_buf_needed);
2425 		if (retval != 0) {
2426 			retval = -1;
2427 			goto bailout;
2428 		}
2429 	}
2430 
2431 	STAILQ_REMOVE(&dev->run_queue, buf, camdd_buf, links);
2432 	dev->num_run_queue--;
2433 
2434 	data = &buf->buf_type_spec.data;
2435 
2436 	ccb = &data->ccb;
2437 	CCB_CLEAR_ALL_EXCEPT_HDR(&ccb->csio);
2438 
2439 	/*
2440 	 * In almost every case the number of blocks should be the device
2441 	 * block size.  The exception may be at the end of an I/O stream
2442 	 * for a partial block or at the end of a device.
2443 	 */
2444 	if (is_write != 0)
2445 		num_blocks = sectors_used;
2446 	else
2447 		num_blocks = data->fill_len / pass_dev->block_len;
2448 
2449 	scsi_read_write(&ccb->csio,
2450 			/*retries*/ dev->retry_count,
2451 			/*cbfcnp*/ NULL,
2452 			/*tag_action*/ MSG_SIMPLE_Q_TAG,
2453 			/*readop*/ (dev->write_dev == 0) ? SCSI_RW_READ :
2454 				   SCSI_RW_WRITE,
2455 			/*byte2*/ 0,
2456 			/*minimum_cmd_size*/ dev->min_cmd_size,
2457 			/*lba*/ buf->lba,
2458 			/*block_count*/ num_blocks,
2459 			/*data_ptr*/ (data->sg_count != 0) ?
2460 				     (uint8_t *)data->segs : data->buf,
2461 			/*dxfer_len*/ (num_blocks * pass_dev->block_len),
2462 			/*sense_len*/ SSD_FULL_SIZE,
2463 			/*timeout*/ dev->io_timeout);
2464 
2465 	/* Disable freezing the device queue */
2466 	ccb->ccb_h.flags |= CAM_DEV_QFRZDIS;
2467 
2468 	if (dev->retry_count != 0)
2469 		ccb->ccb_h.flags |= CAM_PASS_ERR_RECOVER;
2470 
2471 	if (data->sg_count != 0) {
2472 		ccb->csio.sglist_cnt = data->sg_count;
2473 		ccb->ccb_h.flags |= CAM_DATA_SG;
2474 	}
2475 
2476 	/*
2477 	 * Store a pointer to the buffer in the CCB.  The kernel will
2478 	 * restore this when we get it back, and we'll use it to identify
2479 	 * the buffer this CCB came from.
2480 	 */
2481 	ccb->ccb_h.ccb_buf = buf;
2482 
2483 	/*
2484 	 * Unlock our mutex in preparation for issuing the ioctl.
2485 	 */
2486 	pthread_mutex_unlock(&dev->mutex);
2487 	/*
2488 	 * Queue the CCB to the pass(4) driver.
2489 	 */
2490 	if (ioctl(pass_dev->dev->fd, CAMIOQUEUE, ccb) == -1) {
2491 		pthread_mutex_lock(&dev->mutex);
2492 
2493 		warn("%s: error sending CAMIOQUEUE ioctl to %s%u", __func__,
2494 		     pass_dev->dev->device_name, pass_dev->dev->dev_unit_num);
2495 		warn("%s: CCB address is %p", __func__, ccb);
2496 		retval = -1;
2497 
2498 		STAILQ_INSERT_TAIL(&dev->free_queue, buf, links);
2499 	} else {
2500 		pthread_mutex_lock(&dev->mutex);
2501 
2502 		dev->cur_active_io++;
2503 		STAILQ_INSERT_TAIL(&dev->active_queue, buf, links);
2504 	}
2505 
2506 bailout:
2507 	return (retval);
2508 }
2509 
2510 int
2511 camdd_get_next_lba_len(struct camdd_dev *dev, uint64_t *lba, ssize_t *len)
2512 {
2513 	struct camdd_dev_pass *pass_dev;
2514 	uint32_t num_blocks;
2515 	int retval = 0;
2516 
2517 	pass_dev = &dev->dev_spec.pass;
2518 
2519 	*lba = dev->next_io_pos_bytes / dev->sector_size;
2520 	*len = dev->blocksize;
2521 	num_blocks = *len / dev->sector_size;
2522 
2523 	/*
2524 	 * If max_sector is 0, then we have no set limit.  This can happen
2525 	 * if we're writing to a file in a filesystem, or reading from
2526 	 * something like /dev/zero.
2527 	 */
2528 	if ((dev->max_sector != 0)
2529 	 || (dev->sector_io_limit != 0)) {
2530 		uint64_t max_sector;
2531 
2532 		if ((dev->max_sector != 0)
2533 		 && (dev->sector_io_limit != 0))
2534 			max_sector = min(dev->sector_io_limit, dev->max_sector);
2535 		else if (dev->max_sector != 0)
2536 			max_sector = dev->max_sector;
2537 		else
2538 			max_sector = dev->sector_io_limit;
2539 
2540 
2541 		/*
2542 		 * Check to see whether we're starting off past the end of
2543 		 * the device.  If so, we need to just send an EOF
2544 		 * notification to the writer.
2545 		 */
2546 		if (*lba > max_sector) {
2547 			*len = 0;
2548 			retval = 1;
2549 		} else if (((*lba + num_blocks) > max_sector + 1)
2550 			|| ((*lba + num_blocks) < *lba)) {
2551 			/*
2552 			 * If we get here (but pass the first check), we
2553 			 * can trim the request length down to go to the
2554 			 * end of the device.
2555 			 */
2556 			num_blocks = (max_sector + 1) - *lba;
2557 			*len = num_blocks * dev->sector_size;
2558 			retval = 1;
2559 		}
2560 	}
2561 
2562 	dev->next_io_pos_bytes += *len;
2563 
2564 	return (retval);
2565 }
2566 
2567 /*
2568  * Returns 0 for success, 1 for EOF detected, and -1 for failure.
2569  */
2570 int
2571 camdd_queue(struct camdd_dev *dev, struct camdd_buf *read_buf)
2572 {
2573 	struct camdd_buf *buf = NULL;
2574 	struct camdd_buf_data *data;
2575 	struct camdd_dev_pass *pass_dev;
2576 	size_t new_len;
2577 	struct camdd_buf_data *rb_data;
2578 	int is_write = dev->write_dev;
2579 	int eof_flush_needed = 0;
2580 	int retval = 0;
2581 	int error;
2582 
2583 	pass_dev = &dev->dev_spec.pass;
2584 
2585 	/*
2586 	 * If we've gotten EOF or our partner has, we should not continue
2587 	 * queueing I/O.  If we're a writer, though, we should continue
2588 	 * to write any buffers that don't have EOF status.
2589 	 */
2590 	if ((dev->flags & CAMDD_DEV_FLAG_EOF)
2591 	 || ((dev->flags & CAMDD_DEV_FLAG_PEER_EOF)
2592 	  && (is_write == 0))) {
2593 		/*
2594 		 * Tell the worker thread that we have seen EOF.
2595 		 */
2596 		retval = 1;
2597 
2598 		/*
2599 		 * If we're the writer, send the buffer back with EOF status.
2600 		 */
2601 		if (is_write) {
2602 			read_buf->status = CAMDD_STATUS_EOF;
2603 
2604 			error = camdd_complete_peer_buf(dev, read_buf);
2605 		}
2606 		goto bailout;
2607 	}
2608 
2609 	if (is_write == 0) {
2610 		buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2611 		if (buf == NULL) {
2612 			retval = -1;
2613 			goto bailout;
2614 		}
2615 		data = &buf->buf_type_spec.data;
2616 
2617 		retval = camdd_get_next_lba_len(dev, &buf->lba, &buf->len);
2618 		if (retval != 0) {
2619 			buf->status = CAMDD_STATUS_EOF;
2620 
2621 		 	if ((buf->len == 0)
2622 			 && ((dev->flags & (CAMDD_DEV_FLAG_EOF_SENT |
2623 			     CAMDD_DEV_FLAG_EOF_QUEUED)) != 0)) {
2624 				camdd_release_buf(buf);
2625 				goto bailout;
2626 			}
2627 			dev->flags |= CAMDD_DEV_FLAG_EOF_QUEUED;
2628 		}
2629 
2630 		data->fill_len = buf->len;
2631 		data->src_start_offset = buf->lba * dev->sector_size;
2632 
2633 		/*
2634 		 * Put this on the run queue.
2635 		 */
2636 		STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2637 		dev->num_run_queue++;
2638 
2639 		/* We're done. */
2640 		goto bailout;
2641 	}
2642 
2643 	/*
2644 	 * Check for new EOF status from the reader.
2645 	 */
2646 	if ((read_buf->status == CAMDD_STATUS_EOF)
2647 	 || (read_buf->status == CAMDD_STATUS_ERROR)) {
2648 		dev->flags |= CAMDD_DEV_FLAG_PEER_EOF;
2649 		if ((STAILQ_FIRST(&dev->pending_queue) == NULL)
2650 		 && (read_buf->len == 0)) {
2651 			camdd_complete_peer_buf(dev, read_buf);
2652 			retval = 1;
2653 			goto bailout;
2654 		} else
2655 			eof_flush_needed = 1;
2656 	}
2657 
2658 	/*
2659 	 * See if we have a buffer we're composing with pieces from our
2660 	 * partner thread.
2661 	 */
2662 	buf = STAILQ_FIRST(&dev->pending_queue);
2663 	if (buf == NULL) {
2664 		uint64_t lba;
2665 		ssize_t len;
2666 
2667 		retval = camdd_get_next_lba_len(dev, &lba, &len);
2668 		if (retval != 0) {
2669 			read_buf->status = CAMDD_STATUS_EOF;
2670 
2671 			if (len == 0) {
2672 				dev->flags |= CAMDD_DEV_FLAG_EOF;
2673 				error = camdd_complete_peer_buf(dev, read_buf);
2674 				goto bailout;
2675 			}
2676 		}
2677 
2678 		/*
2679 		 * If we don't have a pending buffer, we need to grab a new
2680 		 * one from the free list or allocate another one.
2681 		 */
2682 		buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2683 		if (buf == NULL) {
2684 			retval = 1;
2685 			goto bailout;
2686 		}
2687 
2688 		buf->lba = lba;
2689 		buf->len = len;
2690 
2691 		STAILQ_INSERT_TAIL(&dev->pending_queue, buf, links);
2692 		dev->num_pending_queue++;
2693 	}
2694 
2695 	data = &buf->buf_type_spec.data;
2696 
2697 	rb_data = &read_buf->buf_type_spec.data;
2698 
2699 	if ((rb_data->src_start_offset != dev->next_peer_pos_bytes)
2700 	 && (dev->debug != 0)) {
2701 		printf("%s: WARNING: reader offset %#jx != expected offset "
2702 		    "%#jx\n", __func__, (uintmax_t)rb_data->src_start_offset,
2703 		    (uintmax_t)dev->next_peer_pos_bytes);
2704 	}
2705 	dev->next_peer_pos_bytes = rb_data->src_start_offset +
2706 	    (rb_data->fill_len - rb_data->resid);
2707 
2708 	new_len = (rb_data->fill_len - rb_data->resid) + data->fill_len;
2709 	if (new_len < buf->len) {
2710 		/*
2711 		 * There are three cases here:
2712 		 * 1. We need more data to fill up a block, so we put
2713 		 *    this I/O on the queue and wait for more I/O.
2714 		 * 2. We have a pending buffer in the queue that is
2715 		 *    smaller than our blocksize, but we got an EOF.  So we
2716 		 *    need to go ahead and flush the write out.
2717 		 * 3. We got an error.
2718 		 */
2719 
2720 		/*
2721 		 * Increment our fill length.
2722 		 */
2723 		data->fill_len += (rb_data->fill_len - rb_data->resid);
2724 
2725 		/*
2726 		 * Add the new read buffer to the list for writing.
2727 		 */
2728 		STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2729 
2730 		/* Increment the count */
2731 		buf->src_count++;
2732 
2733 		if (eof_flush_needed == 0) {
2734 			/*
2735 			 * We need to exit, because we don't have enough
2736 			 * data yet.
2737 			 */
2738 			goto bailout;
2739 		} else {
2740 			/*
2741 			 * Take the buffer off of the pending queue.
2742 			 */
2743 			STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2744 				      links);
2745 			dev->num_pending_queue--;
2746 
2747 			/*
2748 			 * If we need an EOF flush, but there is no data
2749 			 * to flush, go ahead and return this buffer.
2750 			 */
2751 			if (data->fill_len == 0) {
2752 				camdd_complete_buf(dev, buf, /*error_count*/0);
2753 				retval = 1;
2754 				goto bailout;
2755 			}
2756 
2757 			/*
2758 			 * Put this on the next queue for execution.
2759 			 */
2760 			STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2761 			dev->num_run_queue++;
2762 		}
2763 	} else if (new_len == buf->len) {
2764 		/*
2765 		 * We have enough data to completey fill one block,
2766 		 * so we're ready to issue the I/O.
2767 		 */
2768 
2769 		/*
2770 		 * Take the buffer off of the pending queue.
2771 		 */
2772 		STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf, links);
2773 		dev->num_pending_queue--;
2774 
2775 		/*
2776 		 * Add the new read buffer to the list for writing.
2777 		 */
2778 		STAILQ_INSERT_TAIL(&buf->src_list, read_buf, src_links);
2779 
2780 		/* Increment the count */
2781 		buf->src_count++;
2782 
2783 		/*
2784 		 * Increment our fill length.
2785 		 */
2786 		data->fill_len += (rb_data->fill_len - rb_data->resid);
2787 
2788 		/*
2789 		 * Put this on the next queue for execution.
2790 		 */
2791 		STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2792 		dev->num_run_queue++;
2793 	} else {
2794 		struct camdd_buf *idb;
2795 		struct camdd_buf_indirect *indirect;
2796 		uint32_t len_to_go, cur_offset;
2797 
2798 
2799 		idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2800 		if (idb == NULL) {
2801 			retval = 1;
2802 			goto bailout;
2803 		}
2804 		indirect = &idb->buf_type_spec.indirect;
2805 		indirect->src_buf = read_buf;
2806 		read_buf->refcount++;
2807 		indirect->offset = 0;
2808 		indirect->start_ptr = rb_data->buf;
2809 		/*
2810 		 * We've already established that there is more
2811 		 * data in read_buf than we have room for in our
2812 		 * current write request.  So this particular chunk
2813 		 * of the request should just be the remainder
2814 		 * needed to fill up a block.
2815 		 */
2816 		indirect->len = buf->len - (data->fill_len - data->resid);
2817 
2818 		camdd_buf_add_child(buf, idb);
2819 
2820 		/*
2821 		 * This buffer is ready to execute, so we can take
2822 		 * it off the pending queue and put it on the run
2823 		 * queue.
2824 		 */
2825 		STAILQ_REMOVE(&dev->pending_queue, buf, camdd_buf,
2826 			      links);
2827 		dev->num_pending_queue--;
2828 		STAILQ_INSERT_TAIL(&dev->run_queue, buf, links);
2829 		dev->num_run_queue++;
2830 
2831 		cur_offset = indirect->offset + indirect->len;
2832 
2833 		/*
2834 		 * The resulting I/O would be too large to fit in
2835 		 * one block.  We need to split this I/O into
2836 		 * multiple pieces.  Allocate as many buffers as needed.
2837 		 */
2838 		for (len_to_go = rb_data->fill_len - rb_data->resid -
2839 		     indirect->len; len_to_go > 0;) {
2840 			struct camdd_buf *new_buf;
2841 			struct camdd_buf_data *new_data;
2842 			uint64_t lba;
2843 			ssize_t len;
2844 
2845 			retval = camdd_get_next_lba_len(dev, &lba, &len);
2846 			if ((retval != 0)
2847 			 && (len == 0)) {
2848 				/*
2849 				 * The device has already been marked
2850 				 * as EOF, and there is no space left.
2851 				 */
2852 				goto bailout;
2853 			}
2854 
2855 			new_buf = camdd_get_buf(dev, CAMDD_BUF_DATA);
2856 			if (new_buf == NULL) {
2857 				retval = 1;
2858 				goto bailout;
2859 			}
2860 
2861 			new_buf->lba = lba;
2862 			new_buf->len = len;
2863 
2864 			idb = camdd_get_buf(dev, CAMDD_BUF_INDIRECT);
2865 			if (idb == NULL) {
2866 				retval = 1;
2867 				goto bailout;
2868 			}
2869 
2870 			indirect = &idb->buf_type_spec.indirect;
2871 
2872 			indirect->src_buf = read_buf;
2873 			read_buf->refcount++;
2874 			indirect->offset = cur_offset;
2875 			indirect->start_ptr = rb_data->buf + cur_offset;
2876 			indirect->len = min(len_to_go, new_buf->len);
2877 #if 0
2878 			if (((indirect->len % dev->sector_size) != 0)
2879 			 || ((indirect->offset % dev->sector_size) != 0)) {
2880 				warnx("offset %ju len %ju not aligned with "
2881 				    "sector size %u", indirect->offset,
2882 				    (uintmax_t)indirect->len, dev->sector_size);
2883 			}
2884 #endif
2885 			cur_offset += indirect->len;
2886 			len_to_go -= indirect->len;
2887 
2888 			camdd_buf_add_child(new_buf, idb);
2889 
2890 			new_data = &new_buf->buf_type_spec.data;
2891 
2892 			if ((new_data->fill_len == new_buf->len)
2893 			 || (eof_flush_needed != 0)) {
2894 				STAILQ_INSERT_TAIL(&dev->run_queue,
2895 						   new_buf, links);
2896 				dev->num_run_queue++;
2897 			} else if (new_data->fill_len < buf->len) {
2898 				STAILQ_INSERT_TAIL(&dev->pending_queue,
2899 					   	new_buf, links);
2900 				dev->num_pending_queue++;
2901 			} else {
2902 				warnx("%s: too much data in new "
2903 				      "buffer!", __func__);
2904 				retval = 1;
2905 				goto bailout;
2906 			}
2907 		}
2908 	}
2909 
2910 bailout:
2911 	return (retval);
2912 }
2913 
2914 void
2915 camdd_get_depth(struct camdd_dev *dev, uint32_t *our_depth,
2916 		uint32_t *peer_depth, uint32_t *our_bytes, uint32_t *peer_bytes)
2917 {
2918 	*our_depth = dev->cur_active_io + dev->num_run_queue;
2919 	if (dev->num_peer_work_queue >
2920 	    dev->num_peer_done_queue)
2921 		*peer_depth = dev->num_peer_work_queue -
2922 			      dev->num_peer_done_queue;
2923 	else
2924 		*peer_depth = 0;
2925 	*our_bytes = *our_depth * dev->blocksize;
2926 	*peer_bytes = dev->peer_bytes_queued;
2927 }
2928 
2929 void
2930 camdd_sig_handler(int sig)
2931 {
2932 	if (sig == SIGINFO)
2933 		need_status = 1;
2934 	else {
2935 		need_exit = 1;
2936 		error_exit = 1;
2937 	}
2938 
2939 	sem_post(&camdd_sem);
2940 }
2941 
2942 void
2943 camdd_print_status(struct camdd_dev *camdd_dev, struct camdd_dev *other_dev,
2944 		   struct timespec *start_time)
2945 {
2946 	struct timespec done_time;
2947 	uint64_t total_ns;
2948 	long double mb_sec, total_sec;
2949 	int error = 0;
2950 
2951 	error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &done_time);
2952 	if (error != 0) {
2953 		warn("Unable to get done time");
2954 		return;
2955 	}
2956 
2957 	timespecsub(&done_time, start_time);
2958 
2959 	total_ns = done_time.tv_nsec + (done_time.tv_sec * 1000000000);
2960 	total_sec = total_ns;
2961 	total_sec /= 1000000000;
2962 
2963 	fprintf(stderr, "%ju bytes %s %s\n%ju bytes %s %s\n"
2964 		"%.4Lf seconds elapsed\n",
2965 		(uintmax_t)camdd_dev->bytes_transferred,
2966 		(camdd_dev->write_dev == 0) ?  "read from" : "written to",
2967 		camdd_dev->device_name,
2968 		(uintmax_t)other_dev->bytes_transferred,
2969 		(other_dev->write_dev == 0) ? "read from" : "written to",
2970 		other_dev->device_name, total_sec);
2971 
2972 	mb_sec = min(other_dev->bytes_transferred,camdd_dev->bytes_transferred);
2973 	mb_sec /= 1024 * 1024;
2974 	mb_sec *= 1000000000;
2975 	mb_sec /= total_ns;
2976 	fprintf(stderr, "%.2Lf MB/sec\n", mb_sec);
2977 }
2978 
2979 int
2980 camdd_rw(struct camdd_io_opts *io_opts, int num_io_opts, uint64_t max_io,
2981 	 int retry_count, int timeout)
2982 {
2983 	struct cam_device *new_cam_dev = NULL;
2984 	struct camdd_dev *devs[2];
2985 	struct timespec start_time;
2986 	pthread_t threads[2];
2987 	int unit = 0;
2988 	int error = 0;
2989 	int i;
2990 
2991 	if (num_io_opts != 2) {
2992 		warnx("Must have one input and one output path");
2993 		error = 1;
2994 		goto bailout;
2995 	}
2996 
2997 	bzero(devs, sizeof(devs));
2998 
2999 	for (i = 0; i < num_io_opts; i++) {
3000 		switch (io_opts[i].dev_type) {
3001 		case CAMDD_DEV_PASS: {
3002 			if (isdigit(io_opts[i].dev_name[0])) {
3003 				camdd_argmask new_arglist = CAMDD_ARG_NONE;
3004 				int bus = 0, target = 0, lun = 0;
3005 				int rv;
3006 
3007 				/* device specified as bus:target[:lun] */
3008 				rv = parse_btl(io_opts[i].dev_name, &bus,
3009 				    &target, &lun, &new_arglist);
3010 				if (rv < 2) {
3011 					warnx("numeric device specification "
3012 					     "must be either bus:target, or "
3013 					     "bus:target:lun");
3014 					error = 1;
3015 					goto bailout;
3016 				}
3017 				/* default to 0 if lun was not specified */
3018 				if ((new_arglist & CAMDD_ARG_LUN) == 0) {
3019 					lun = 0;
3020 					new_arglist |= CAMDD_ARG_LUN;
3021 				}
3022 				new_cam_dev = cam_open_btl(bus, target, lun,
3023 				    O_RDWR, NULL);
3024 			} else {
3025 				char name[30];
3026 
3027 				if (cam_get_device(io_opts[i].dev_name, name,
3028 						   sizeof name, &unit) == -1) {
3029 					warnx("%s", cam_errbuf);
3030 					error = 1;
3031 					goto bailout;
3032 				}
3033 				new_cam_dev = cam_open_spec_device(name, unit,
3034 				    O_RDWR, NULL);
3035 			}
3036 
3037 			if (new_cam_dev == NULL) {
3038 				warnx("%s", cam_errbuf);
3039 				error = 1;
3040 				goto bailout;
3041 			}
3042 
3043 			devs[i] = camdd_probe_pass(new_cam_dev,
3044 			    /*io_opts*/ &io_opts[i],
3045 			    CAMDD_ARG_ERR_RECOVER,
3046 			    /*probe_retry_count*/ 3,
3047 			    /*probe_timeout*/ 5000,
3048 			    /*io_retry_count*/ retry_count,
3049 			    /*io_timeout*/ timeout);
3050 			if (devs[i] == NULL) {
3051 				warn("Unable to probe device %s%u",
3052 				     new_cam_dev->device_name,
3053 				     new_cam_dev->dev_unit_num);
3054 				error = 1;
3055 				goto bailout;
3056 			}
3057 			break;
3058 		}
3059 		case CAMDD_DEV_FILE: {
3060 			int fd = -1;
3061 
3062 			if (io_opts[i].dev_name[0] == '-') {
3063 				if (io_opts[i].write_dev != 0)
3064 					fd = STDOUT_FILENO;
3065 				else
3066 					fd = STDIN_FILENO;
3067 			} else {
3068 				if (io_opts[i].write_dev != 0) {
3069 					fd = open(io_opts[i].dev_name,
3070 					    O_RDWR | O_CREAT, S_IWUSR |S_IRUSR);
3071 				} else {
3072 					fd = open(io_opts[i].dev_name,
3073 					    O_RDONLY);
3074 				}
3075 			}
3076 			if (fd == -1) {
3077 				warn("error opening file %s",
3078 				    io_opts[i].dev_name);
3079 				error = 1;
3080 				goto bailout;
3081 			}
3082 
3083 			devs[i] = camdd_probe_file(fd, &io_opts[i],
3084 			    retry_count, timeout);
3085 			if (devs[i] == NULL) {
3086 				error = 1;
3087 				goto bailout;
3088 			}
3089 
3090 			break;
3091 		}
3092 		default:
3093 			warnx("Unknown device type %d (%s)",
3094 			    io_opts[i].dev_type, io_opts[i].dev_name);
3095 			error = 1;
3096 			goto bailout;
3097 			break; /*NOTREACHED */
3098 		}
3099 
3100 		devs[i]->write_dev = io_opts[i].write_dev;
3101 
3102 		devs[i]->start_offset_bytes = io_opts[i].offset;
3103 
3104 		if (max_io != 0) {
3105 			devs[i]->sector_io_limit =
3106 			    (devs[i]->start_offset_bytes /
3107 			    devs[i]->sector_size) +
3108 			    (max_io / devs[i]->sector_size) - 1;
3109 		}
3110 
3111 		devs[i]->next_io_pos_bytes = devs[i]->start_offset_bytes;
3112 		devs[i]->next_completion_pos_bytes =devs[i]->start_offset_bytes;
3113 	}
3114 
3115 	devs[0]->peer_dev = devs[1];
3116 	devs[1]->peer_dev = devs[0];
3117 	devs[0]->next_peer_pos_bytes = devs[0]->peer_dev->next_io_pos_bytes;
3118 	devs[1]->next_peer_pos_bytes = devs[1]->peer_dev->next_io_pos_bytes;
3119 
3120 	sem_init(&camdd_sem, /*pshared*/ 0, 0);
3121 
3122 	signal(SIGINFO, camdd_sig_handler);
3123 	signal(SIGINT, camdd_sig_handler);
3124 
3125 	error = clock_gettime(CLOCK_MONOTONIC_PRECISE, &start_time);
3126 	if (error != 0) {
3127 		warn("Unable to get start time");
3128 		goto bailout;
3129 	}
3130 
3131 	for (i = 0; i < num_io_opts; i++) {
3132 		error = pthread_create(&threads[i], NULL, camdd_worker,
3133 				       (void *)devs[i]);
3134 		if (error != 0) {
3135 			warnc(error, "pthread_create() failed");
3136 			goto bailout;
3137 		}
3138 	}
3139 
3140 	for (;;) {
3141 		if ((sem_wait(&camdd_sem) == -1)
3142 		 || (need_exit != 0)) {
3143 			struct kevent ke;
3144 
3145 			for (i = 0; i < num_io_opts; i++) {
3146 				EV_SET(&ke, (uintptr_t)&devs[i]->work_queue,
3147 				    EVFILT_USER, 0, NOTE_TRIGGER, 0, NULL);
3148 
3149 				devs[i]->flags |= CAMDD_DEV_FLAG_EOF;
3150 
3151 				error = kevent(devs[i]->kq, &ke, 1, NULL, 0,
3152 						NULL);
3153 				if (error == -1)
3154 					warn("%s: unable to wake up thread",
3155 					    __func__);
3156 				error = 0;
3157 			}
3158 			break;
3159 		} else if (need_status != 0) {
3160 			camdd_print_status(devs[0], devs[1], &start_time);
3161 			need_status = 0;
3162 		}
3163 	}
3164 	for (i = 0; i < num_io_opts; i++) {
3165 		pthread_join(threads[i], NULL);
3166 	}
3167 
3168 	camdd_print_status(devs[0], devs[1], &start_time);
3169 
3170 bailout:
3171 
3172 	for (i = 0; i < num_io_opts; i++)
3173 		camdd_free_dev(devs[i]);
3174 
3175 	return (error + error_exit);
3176 }
3177 
3178 void
3179 usage(void)
3180 {
3181 	fprintf(stderr,
3182 "usage:  camdd <-i|-o pass=pass0,bs=1M,offset=1M,depth=4>\n"
3183 "              <-i|-o file=/tmp/file,bs=512K,offset=1M>\n"
3184 "              <-i|-o file=/dev/da0,bs=512K,offset=1M>\n"
3185 "              <-i|-o file=/dev/nsa0,bs=512K>\n"
3186 "              [-C retry_count][-E][-m max_io_amt][-t timeout_secs][-v][-h]\n"
3187 "Option description\n"
3188 "-i <arg=val>  Specify input device/file and parameters\n"
3189 "-o <arg=val>  Specify output device/file and parameters\n"
3190 "Input and Output parameters\n"
3191 "pass=name     Specify a pass(4) device like pass0 or /dev/pass0\n"
3192 "file=name     Specify a file or device, /tmp/foo, /dev/da0, /dev/null\n"
3193 "              or - for stdin/stdout\n"
3194 "bs=blocksize  Specify blocksize in bytes, or using K, M, G, etc. suffix\n"
3195 "offset=len    Specify starting offset in bytes or using K, M, G suffix\n"
3196 "              NOTE: offset cannot be specified on tapes, pipes, stdin/out\n"
3197 "depth=N       Specify a numeric queue depth.  This only applies to pass(4)\n"
3198 "mcs=N         Specify a minimum cmd size for pass(4) read/write commands\n"
3199 "Optional arguments\n"
3200 "-C retry_cnt  Specify a retry count for pass(4) devices\n"
3201 "-E            Enable CAM error recovery for pass(4) devices\n"
3202 "-m max_io     Specify the maximum amount to be transferred in bytes or\n"
3203 "              using K, G, M, etc. suffixes\n"
3204 "-t timeout    Specify the I/O timeout to use with pass(4) devices\n"
3205 "-v            Enable verbose error recovery\n"
3206 "-h            Print this message\n");
3207 }
3208 
3209 
3210 int
3211 camdd_parse_io_opts(char *args, int is_write, struct camdd_io_opts *io_opts)
3212 {
3213 	char *tmpstr, *tmpstr2;
3214 	char *orig_tmpstr = NULL;
3215 	int retval = 0;
3216 
3217 	io_opts->write_dev = is_write;
3218 
3219 	tmpstr = strdup(args);
3220 	if (tmpstr == NULL) {
3221 		warn("strdup failed");
3222 		retval = 1;
3223 		goto bailout;
3224 	}
3225 	orig_tmpstr = tmpstr;
3226 	while ((tmpstr2 = strsep(&tmpstr, ",")) != NULL) {
3227 		char *name, *value;
3228 
3229 		/*
3230 		 * If the user creates an empty parameter by putting in two
3231 		 * commas, skip over it and look for the next field.
3232 		 */
3233 		if (*tmpstr2 == '\0')
3234 			continue;
3235 
3236 		name = strsep(&tmpstr2, "=");
3237 		if (*name == '\0') {
3238 			warnx("Got empty I/O parameter name");
3239 			retval = 1;
3240 			goto bailout;
3241 		}
3242 		value = strsep(&tmpstr2, "=");
3243 		if ((value == NULL)
3244 		 || (*value == '\0')) {
3245 			warnx("Empty I/O parameter value for %s", name);
3246 			retval = 1;
3247 			goto bailout;
3248 		}
3249 		if (strncasecmp(name, "file", 4) == 0) {
3250 			io_opts->dev_type = CAMDD_DEV_FILE;
3251 			io_opts->dev_name = strdup(value);
3252 			if (io_opts->dev_name == NULL) {
3253 				warn("Error allocating memory");
3254 				retval = 1;
3255 				goto bailout;
3256 			}
3257 		} else if (strncasecmp(name, "pass", 4) == 0) {
3258 			io_opts->dev_type = CAMDD_DEV_PASS;
3259 			io_opts->dev_name = strdup(value);
3260 			if (io_opts->dev_name == NULL) {
3261 				warn("Error allocating memory");
3262 				retval = 1;
3263 				goto bailout;
3264 			}
3265 		} else if ((strncasecmp(name, "bs", 2) == 0)
3266 			|| (strncasecmp(name, "blocksize", 9) == 0)) {
3267 			retval = expand_number(value, &io_opts->blocksize);
3268 			if (retval == -1) {
3269 				warn("expand_number(3) failed on %s=%s", name,
3270 				    value);
3271 				retval = 1;
3272 				goto bailout;
3273 			}
3274 		} else if (strncasecmp(name, "depth", 5) == 0) {
3275 			char *endptr;
3276 
3277 			io_opts->queue_depth = strtoull(value, &endptr, 0);
3278 			if (*endptr != '\0') {
3279 				warnx("invalid queue depth %s", value);
3280 				retval = 1;
3281 				goto bailout;
3282 			}
3283 		} else if (strncasecmp(name, "mcs", 3) == 0) {
3284 			char *endptr;
3285 
3286 			io_opts->min_cmd_size = strtol(value, &endptr, 0);
3287 			if ((*endptr != '\0')
3288 			 || ((io_opts->min_cmd_size > 16)
3289 			  || (io_opts->min_cmd_size < 0))) {
3290 				warnx("invalid minimum cmd size %s", value);
3291 				retval = 1;
3292 				goto bailout;
3293 			}
3294 		} else if (strncasecmp(name, "offset", 6) == 0) {
3295 			retval = expand_number(value, &io_opts->offset);
3296 			if (retval == -1) {
3297 				warn("expand_number(3) failed on %s=%s", name,
3298 				    value);
3299 				retval = 1;
3300 				goto bailout;
3301 			}
3302 		} else if (strncasecmp(name, "debug", 5) == 0) {
3303 			char *endptr;
3304 
3305 			io_opts->debug = strtoull(value, &endptr, 0);
3306 			if (*endptr != '\0') {
3307 				warnx("invalid debug level %s", value);
3308 				retval = 1;
3309 				goto bailout;
3310 			}
3311 		} else {
3312 			warnx("Unrecognized parameter %s=%s", name, value);
3313 		}
3314 	}
3315 bailout:
3316 	free(orig_tmpstr);
3317 
3318 	return (retval);
3319 }
3320 
3321 int
3322 main(int argc, char **argv)
3323 {
3324 	int c;
3325 	camdd_argmask arglist = CAMDD_ARG_NONE;
3326 	int timeout = 0, retry_count = 1;
3327 	int error = 0;
3328 	uint64_t max_io = 0;
3329 	struct camdd_io_opts *opt_list = NULL;
3330 
3331 	if (argc == 1) {
3332 		usage();
3333 		exit(1);
3334 	}
3335 
3336 	opt_list = calloc(2, sizeof(struct camdd_io_opts));
3337 	if (opt_list == NULL) {
3338 		warn("Unable to allocate option list");
3339 		error = 1;
3340 		goto bailout;
3341 	}
3342 
3343 	while ((c = getopt(argc, argv, "C:Ehi:m:o:t:v")) != -1){
3344 		switch (c) {
3345 		case 'C':
3346 			retry_count = strtol(optarg, NULL, 0);
3347 			if (retry_count < 0)
3348 				errx(1, "retry count %d is < 0",
3349 				     retry_count);
3350 			arglist |= CAMDD_ARG_RETRIES;
3351 			break;
3352 		case 'E':
3353 			arglist |= CAMDD_ARG_ERR_RECOVER;
3354 			break;
3355 		case 'i':
3356 		case 'o':
3357 			if (((c == 'i')
3358 			  && (opt_list[0].dev_type != CAMDD_DEV_NONE))
3359 			 || ((c == 'o')
3360 			  && (opt_list[1].dev_type != CAMDD_DEV_NONE))) {
3361 				errx(1, "Only one input and output path "
3362 				    "allowed");
3363 			}
3364 			error = camdd_parse_io_opts(optarg, (c == 'o') ? 1 : 0,
3365 			    (c == 'o') ? &opt_list[1] : &opt_list[0]);
3366 			if (error != 0)
3367 				goto bailout;
3368 			break;
3369 		case 'm':
3370 			error = expand_number(optarg, &max_io);
3371 			if (error == -1) {
3372 				warn("invalid maximum I/O amount %s", optarg);
3373 				error = 1;
3374 				goto bailout;
3375 			}
3376 			break;
3377 		case 't':
3378 			timeout = strtol(optarg, NULL, 0);
3379 			if (timeout < 0)
3380 				errx(1, "invalid timeout %d", timeout);
3381 			/* Convert the timeout from seconds to ms */
3382 			timeout *= 1000;
3383 			arglist |= CAMDD_ARG_TIMEOUT;
3384 			break;
3385 		case 'v':
3386 			arglist |= CAMDD_ARG_VERBOSE;
3387 			break;
3388 		case 'h':
3389 		default:
3390 			usage();
3391 			exit(1);
3392 			break; /*NOTREACHED*/
3393 		}
3394 	}
3395 
3396 	if ((opt_list[0].dev_type == CAMDD_DEV_NONE)
3397 	 || (opt_list[1].dev_type == CAMDD_DEV_NONE))
3398 		errx(1, "Must specify both -i and -o");
3399 
3400 	/*
3401 	 * Set the timeout if the user hasn't specified one.
3402 	 */
3403 	if (timeout == 0)
3404 		timeout = CAMDD_PASS_RW_TIMEOUT;
3405 
3406 	error = camdd_rw(opt_list, 2, max_io, retry_count, timeout);
3407 
3408 bailout:
3409 	free(opt_list);
3410 
3411 	exit(error);
3412 }
3413