block: trickle down the fallback image creation function use to the block drivers
[qemu/ar7.git] / block / iscsi.c
blob14680a436aa5f69296b8307c98cf7409870b2570
1 /*
2 * QEMU Block driver for iSCSI images
4 * Copyright (c) 2010-2011 Ronnie Sahlberg <ronniesahlberg@gmail.com>
5 * Copyright (c) 2012-2017 Peter Lieven <pl@kamp.de>
7 * Permission is hereby granted, free of charge, to any person obtaining a copy
8 * of this software and associated documentation files (the "Software"), to deal
9 * in the Software without restriction, including without limitation the rights
10 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
11 * copies of the Software, and to permit persons to whom the Software is
12 * furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
23 * THE SOFTWARE.
26 #include "qemu/osdep.h"
28 #include <poll.h>
29 #include <math.h>
30 #include <arpa/inet.h>
31 #include "qemu-common.h"
32 #include "qemu/config-file.h"
33 #include "qemu/error-report.h"
34 #include "qemu/bitops.h"
35 #include "qemu/bitmap.h"
36 #include "block/block_int.h"
37 #include "block/qdict.h"
38 #include "scsi/constants.h"
39 #include "qemu/iov.h"
40 #include "qemu/module.h"
41 #include "qemu/option.h"
42 #include "qemu/uuid.h"
43 #include "sysemu/replay.h"
44 #include "qapi/error.h"
45 #include "qapi/qapi-commands-misc.h"
46 #include "qapi/qmp/qdict.h"
47 #include "qapi/qmp/qstring.h"
48 #include "crypto/secret.h"
49 #include "scsi/utils.h"
50 #include "trace.h"
52 /* Conflict between scsi/utils.h and libiscsi! :( */
53 #define SCSI_XFER_NONE ISCSI_XFER_NONE
54 #include <iscsi/iscsi.h>
55 #define inline __attribute__((gnu_inline)) /* required for libiscsi v1.9.0 */
56 #include <iscsi/scsi-lowlevel.h>
57 #undef inline
58 #undef SCSI_XFER_NONE
59 QEMU_BUILD_BUG_ON((int)SCSI_XFER_NONE != (int)ISCSI_XFER_NONE);
61 #ifdef __linux__
62 #include <scsi/sg.h>
63 #endif
65 typedef struct IscsiLun {
66 struct iscsi_context *iscsi;
67 AioContext *aio_context;
68 int lun;
69 enum scsi_inquiry_peripheral_device_type type;
70 int block_size;
71 uint64_t num_blocks;
72 int events;
73 QEMUTimer *nop_timer;
74 QEMUTimer *event_timer;
75 QemuMutex mutex;
76 struct scsi_inquiry_logical_block_provisioning lbp;
77 struct scsi_inquiry_block_limits bl;
78 struct scsi_inquiry_device_designator *dd;
79 unsigned char *zeroblock;
80 /* The allocmap tracks which clusters (pages) on the iSCSI target are
81 * allocated and which are not. In case a target returns zeros for
82 * unallocated pages (iscsilun->lprz) we can directly return zeros instead
83 * of reading zeros over the wire if a read request falls within an
84 * unallocated block. As there are 3 possible states we need 2 bitmaps to
85 * track. allocmap_valid keeps track if QEMU's information about a page is
86 * valid. allocmap tracks if a page is allocated or not. In case QEMU has no
87 * valid information about a page the corresponding allocmap entry should be
88 * switched to unallocated as well to force a new lookup of the allocation
89 * status as lookups are generally skipped if a page is suspect to be
90 * allocated. If a iSCSI target is opened with cache.direct = on the
91 * allocmap_valid does not exist turning all cached information invalid so
92 * that a fresh lookup is made for any page even if allocmap entry returns
93 * it's unallocated. */
94 unsigned long *allocmap;
95 unsigned long *allocmap_valid;
96 long allocmap_size;
97 int cluster_size;
98 bool use_16_for_rw;
99 bool write_protected;
100 bool lbpme;
101 bool lbprz;
102 bool dpofua;
103 bool has_write_same;
104 bool request_timed_out;
105 } IscsiLun;
107 typedef struct IscsiTask {
108 int status;
109 int complete;
110 int retries;
111 int do_retry;
112 struct scsi_task *task;
113 Coroutine *co;
114 IscsiLun *iscsilun;
115 QEMUTimer retry_timer;
116 int err_code;
117 char *err_str;
118 } IscsiTask;
120 typedef struct IscsiAIOCB {
121 BlockAIOCB common;
122 QEMUBH *bh;
123 IscsiLun *iscsilun;
124 struct scsi_task *task;
125 int status;
126 int64_t sector_num;
127 int nb_sectors;
128 int ret;
129 #ifdef __linux__
130 sg_io_hdr_t *ioh;
131 #endif
132 bool cancelled;
133 } IscsiAIOCB;
135 /* libiscsi uses time_t so its enough to process events every second */
136 #define EVENT_INTERVAL 1000
137 #define NOP_INTERVAL 5000
138 #define MAX_NOP_FAILURES 3
139 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
140 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048, 8192, 32768};
142 /* this threshold is a trade-off knob to choose between
143 * the potential additional overhead of an extra GET_LBA_STATUS request
144 * vs. unnecessarily reading a lot of zero sectors over the wire.
145 * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
146 * sectors we check the allocation status of the area covered by the
147 * request first if the allocationmap indicates that the area might be
148 * unallocated. */
149 #define ISCSI_CHECKALLOC_THRES 64
151 #ifdef __linux__
153 static void
154 iscsi_bh_cb(void *p)
156 IscsiAIOCB *acb = p;
158 qemu_bh_delete(acb->bh);
160 acb->common.cb(acb->common.opaque, acb->status);
162 if (acb->task != NULL) {
163 scsi_free_scsi_task(acb->task);
164 acb->task = NULL;
167 qemu_aio_unref(acb);
170 static void
171 iscsi_schedule_bh(IscsiAIOCB *acb)
173 if (acb->bh) {
174 return;
176 acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
177 qemu_bh_schedule(acb->bh);
180 #endif
182 static void iscsi_co_generic_bh_cb(void *opaque)
184 struct IscsiTask *iTask = opaque;
186 iTask->complete = 1;
187 aio_co_wake(iTask->co);
190 static void iscsi_retry_timer_expired(void *opaque)
192 struct IscsiTask *iTask = opaque;
193 iTask->complete = 1;
194 if (iTask->co) {
195 aio_co_wake(iTask->co);
199 static inline unsigned exp_random(double mean)
201 return -mean * log((double)rand() / RAND_MAX);
204 /* SCSI_SENSE_ASCQ_INVALID_FIELD_IN_PARAMETER_LIST was introduced in
205 * libiscsi 1.10.0, together with other constants we need. Use it as
206 * a hint that we have to define them ourselves if needed, to keep the
207 * minimum required libiscsi version at 1.9.0. We use an ASCQ macro for
208 * the test because SCSI_STATUS_* is an enum.
210 * To guard against future changes where SCSI_SENSE_ASCQ_* also becomes
211 * an enum, check against the LIBISCSI_API_VERSION macro, which was
212 * introduced in 1.11.0. If it is present, there is no need to define
213 * anything.
215 #if !defined(SCSI_SENSE_ASCQ_INVALID_FIELD_IN_PARAMETER_LIST) && \
216 !defined(LIBISCSI_API_VERSION)
217 #define SCSI_STATUS_TASK_SET_FULL 0x28
218 #define SCSI_STATUS_TIMEOUT 0x0f000002
219 #define SCSI_SENSE_ASCQ_INVALID_FIELD_IN_PARAMETER_LIST 0x2600
220 #define SCSI_SENSE_ASCQ_PARAMETER_LIST_LENGTH_ERROR 0x1a00
221 #endif
223 #ifndef LIBISCSI_API_VERSION
224 #define LIBISCSI_API_VERSION 20130701
225 #endif
227 static int iscsi_translate_sense(struct scsi_sense *sense)
229 return scsi_sense_to_errno(sense->key,
230 (sense->ascq & 0xFF00) >> 8,
231 sense->ascq & 0xFF);
234 /* Called (via iscsi_service) with QemuMutex held. */
235 static void
236 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
237 void *command_data, void *opaque)
239 struct IscsiTask *iTask = opaque;
240 struct scsi_task *task = command_data;
242 iTask->status = status;
243 iTask->do_retry = 0;
244 iTask->task = task;
246 if (status != SCSI_STATUS_GOOD) {
247 if (iTask->retries++ < ISCSI_CMD_RETRIES) {
248 if (status == SCSI_STATUS_BUSY ||
249 status == SCSI_STATUS_TIMEOUT ||
250 status == SCSI_STATUS_TASK_SET_FULL) {
251 unsigned retry_time =
252 exp_random(iscsi_retry_times[iTask->retries - 1]);
253 if (status == SCSI_STATUS_TIMEOUT) {
254 /* make sure the request is rescheduled AFTER the
255 * reconnect is initiated */
256 retry_time = EVENT_INTERVAL * 2;
257 iTask->iscsilun->request_timed_out = true;
259 error_report("iSCSI Busy/TaskSetFull/TimeOut"
260 " (retry #%u in %u ms): %s",
261 iTask->retries, retry_time,
262 iscsi_get_error(iscsi));
263 aio_timer_init(iTask->iscsilun->aio_context,
264 &iTask->retry_timer, QEMU_CLOCK_REALTIME,
265 SCALE_MS, iscsi_retry_timer_expired, iTask);
266 timer_mod(&iTask->retry_timer,
267 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + retry_time);
268 iTask->do_retry = 1;
270 } else if (status == SCSI_STATUS_CHECK_CONDITION) {
271 int error = iscsi_translate_sense(&task->sense);
272 if (error == EAGAIN) {
273 error_report("iSCSI CheckCondition: %s",
274 iscsi_get_error(iscsi));
275 iTask->do_retry = 1;
276 } else {
277 iTask->err_code = -error;
278 iTask->err_str = g_strdup(iscsi_get_error(iscsi));
283 if (iTask->co) {
284 replay_bh_schedule_oneshot_event(iTask->iscsilun->aio_context,
285 iscsi_co_generic_bh_cb, iTask);
286 } else {
287 iTask->complete = 1;
291 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
293 *iTask = (struct IscsiTask) {
294 .co = qemu_coroutine_self(),
295 .iscsilun = iscsilun,
299 #ifdef __linux__
301 /* Called (via iscsi_service) with QemuMutex held. */
302 static void
303 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
304 void *private_data)
306 IscsiAIOCB *acb = private_data;
308 /* If the command callback hasn't been called yet, drop the task */
309 if (!acb->bh) {
310 /* Call iscsi_aio_ioctl_cb() with SCSI_STATUS_CANCELLED */
311 iscsi_scsi_cancel_task(iscsi, acb->task);
314 qemu_aio_unref(acb); /* acquired in iscsi_aio_cancel() */
317 static void
318 iscsi_aio_cancel(BlockAIOCB *blockacb)
320 IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
321 IscsiLun *iscsilun = acb->iscsilun;
323 qemu_mutex_lock(&iscsilun->mutex);
325 /* If it was cancelled or completed already, our work is done here */
326 if (acb->cancelled || acb->status != -EINPROGRESS) {
327 qemu_mutex_unlock(&iscsilun->mutex);
328 return;
331 acb->cancelled = true;
333 qemu_aio_ref(acb); /* released in iscsi_abort_task_cb() */
335 /* send a task mgmt call to the target to cancel the task on the target */
336 if (iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
337 iscsi_abort_task_cb, acb) < 0) {
338 qemu_aio_unref(acb); /* since iscsi_abort_task_cb() won't be called */
341 qemu_mutex_unlock(&iscsilun->mutex);
344 static const AIOCBInfo iscsi_aiocb_info = {
345 .aiocb_size = sizeof(IscsiAIOCB),
346 .cancel_async = iscsi_aio_cancel,
349 #endif
351 static void iscsi_process_read(void *arg);
352 static void iscsi_process_write(void *arg);
354 /* Called with QemuMutex held. */
355 static void
356 iscsi_set_events(IscsiLun *iscsilun)
358 struct iscsi_context *iscsi = iscsilun->iscsi;
359 int ev = iscsi_which_events(iscsi);
361 if (ev != iscsilun->events) {
362 aio_set_fd_handler(iscsilun->aio_context, iscsi_get_fd(iscsi),
363 false,
364 (ev & POLLIN) ? iscsi_process_read : NULL,
365 (ev & POLLOUT) ? iscsi_process_write : NULL,
366 NULL,
367 iscsilun);
368 iscsilun->events = ev;
372 static void iscsi_timed_check_events(void *opaque)
374 IscsiLun *iscsilun = opaque;
376 qemu_mutex_lock(&iscsilun->mutex);
378 /* check for timed out requests */
379 iscsi_service(iscsilun->iscsi, 0);
381 if (iscsilun->request_timed_out) {
382 iscsilun->request_timed_out = false;
383 iscsi_reconnect(iscsilun->iscsi);
386 /* newer versions of libiscsi may return zero events. Ensure we are able
387 * to return to service once this situation changes. */
388 iscsi_set_events(iscsilun);
390 qemu_mutex_unlock(&iscsilun->mutex);
392 timer_mod(iscsilun->event_timer,
393 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + EVENT_INTERVAL);
396 static void
397 iscsi_process_read(void *arg)
399 IscsiLun *iscsilun = arg;
400 struct iscsi_context *iscsi = iscsilun->iscsi;
402 qemu_mutex_lock(&iscsilun->mutex);
403 iscsi_service(iscsi, POLLIN);
404 iscsi_set_events(iscsilun);
405 qemu_mutex_unlock(&iscsilun->mutex);
408 static void
409 iscsi_process_write(void *arg)
411 IscsiLun *iscsilun = arg;
412 struct iscsi_context *iscsi = iscsilun->iscsi;
414 qemu_mutex_lock(&iscsilun->mutex);
415 iscsi_service(iscsi, POLLOUT);
416 iscsi_set_events(iscsilun);
417 qemu_mutex_unlock(&iscsilun->mutex);
420 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
422 return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
425 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
427 return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
430 static bool is_byte_request_lun_aligned(int64_t offset, int count,
431 IscsiLun *iscsilun)
433 if (offset % iscsilun->block_size || count % iscsilun->block_size) {
434 error_report("iSCSI misaligned request: "
435 "iscsilun->block_size %u, offset %" PRIi64
436 ", count %d",
437 iscsilun->block_size, offset, count);
438 return false;
440 return true;
443 static bool is_sector_request_lun_aligned(int64_t sector_num, int nb_sectors,
444 IscsiLun *iscsilun)
446 assert(nb_sectors <= BDRV_REQUEST_MAX_SECTORS);
447 return is_byte_request_lun_aligned(sector_num << BDRV_SECTOR_BITS,
448 nb_sectors << BDRV_SECTOR_BITS,
449 iscsilun);
452 static void iscsi_allocmap_free(IscsiLun *iscsilun)
454 g_free(iscsilun->allocmap);
455 g_free(iscsilun->allocmap_valid);
456 iscsilun->allocmap = NULL;
457 iscsilun->allocmap_valid = NULL;
461 static int iscsi_allocmap_init(IscsiLun *iscsilun, int open_flags)
463 iscsi_allocmap_free(iscsilun);
465 assert(iscsilun->cluster_size);
466 iscsilun->allocmap_size =
467 DIV_ROUND_UP(iscsilun->num_blocks * iscsilun->block_size,
468 iscsilun->cluster_size);
470 iscsilun->allocmap = bitmap_try_new(iscsilun->allocmap_size);
471 if (!iscsilun->allocmap) {
472 return -ENOMEM;
475 if (open_flags & BDRV_O_NOCACHE) {
476 /* when cache.direct = on all allocmap entries are
477 * treated as invalid to force a relookup of the block
478 * status on every read request */
479 return 0;
482 iscsilun->allocmap_valid = bitmap_try_new(iscsilun->allocmap_size);
483 if (!iscsilun->allocmap_valid) {
484 /* if we are under memory pressure free the allocmap as well */
485 iscsi_allocmap_free(iscsilun);
486 return -ENOMEM;
489 return 0;
492 static void
493 iscsi_allocmap_update(IscsiLun *iscsilun, int64_t offset,
494 int64_t bytes, bool allocated, bool valid)
496 int64_t cl_num_expanded, nb_cls_expanded, cl_num_shrunk, nb_cls_shrunk;
498 if (iscsilun->allocmap == NULL) {
499 return;
501 /* expand to entirely contain all affected clusters */
502 assert(iscsilun->cluster_size);
503 cl_num_expanded = offset / iscsilun->cluster_size;
504 nb_cls_expanded = DIV_ROUND_UP(offset + bytes,
505 iscsilun->cluster_size) - cl_num_expanded;
506 /* shrink to touch only completely contained clusters */
507 cl_num_shrunk = DIV_ROUND_UP(offset, iscsilun->cluster_size);
508 nb_cls_shrunk = (offset + bytes) / iscsilun->cluster_size - cl_num_shrunk;
509 if (allocated) {
510 bitmap_set(iscsilun->allocmap, cl_num_expanded, nb_cls_expanded);
511 } else {
512 if (nb_cls_shrunk > 0) {
513 bitmap_clear(iscsilun->allocmap, cl_num_shrunk, nb_cls_shrunk);
517 if (iscsilun->allocmap_valid == NULL) {
518 return;
520 if (valid) {
521 if (nb_cls_shrunk > 0) {
522 bitmap_set(iscsilun->allocmap_valid, cl_num_shrunk, nb_cls_shrunk);
524 } else {
525 bitmap_clear(iscsilun->allocmap_valid, cl_num_expanded,
526 nb_cls_expanded);
530 static void
531 iscsi_allocmap_set_allocated(IscsiLun *iscsilun, int64_t offset,
532 int64_t bytes)
534 iscsi_allocmap_update(iscsilun, offset, bytes, true, true);
537 static void
538 iscsi_allocmap_set_unallocated(IscsiLun *iscsilun, int64_t offset,
539 int64_t bytes)
541 /* Note: if cache.direct=on the fifth argument to iscsi_allocmap_update
542 * is ignored, so this will in effect be an iscsi_allocmap_set_invalid.
544 iscsi_allocmap_update(iscsilun, offset, bytes, false, true);
547 static void iscsi_allocmap_set_invalid(IscsiLun *iscsilun, int64_t offset,
548 int64_t bytes)
550 iscsi_allocmap_update(iscsilun, offset, bytes, false, false);
553 static void iscsi_allocmap_invalidate(IscsiLun *iscsilun)
555 if (iscsilun->allocmap) {
556 bitmap_zero(iscsilun->allocmap, iscsilun->allocmap_size);
558 if (iscsilun->allocmap_valid) {
559 bitmap_zero(iscsilun->allocmap_valid, iscsilun->allocmap_size);
563 static inline bool
564 iscsi_allocmap_is_allocated(IscsiLun *iscsilun, int64_t offset,
565 int64_t bytes)
567 unsigned long size;
568 if (iscsilun->allocmap == NULL) {
569 return true;
571 assert(iscsilun->cluster_size);
572 size = DIV_ROUND_UP(offset + bytes, iscsilun->cluster_size);
573 return !(find_next_bit(iscsilun->allocmap, size,
574 offset / iscsilun->cluster_size) == size);
577 static inline bool iscsi_allocmap_is_valid(IscsiLun *iscsilun,
578 int64_t offset, int64_t bytes)
580 unsigned long size;
581 if (iscsilun->allocmap_valid == NULL) {
582 return false;
584 assert(iscsilun->cluster_size);
585 size = DIV_ROUND_UP(offset + bytes, iscsilun->cluster_size);
586 return (find_next_zero_bit(iscsilun->allocmap_valid, size,
587 offset / iscsilun->cluster_size) == size);
590 static void coroutine_fn iscsi_co_wait_for_task(IscsiTask *iTask,
591 IscsiLun *iscsilun)
593 while (!iTask->complete) {
594 iscsi_set_events(iscsilun);
595 qemu_mutex_unlock(&iscsilun->mutex);
596 qemu_coroutine_yield();
597 qemu_mutex_lock(&iscsilun->mutex);
601 static int coroutine_fn
602 iscsi_co_writev(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
603 QEMUIOVector *iov, int flags)
605 IscsiLun *iscsilun = bs->opaque;
606 struct IscsiTask iTask;
607 uint64_t lba;
608 uint32_t num_sectors;
609 bool fua = flags & BDRV_REQ_FUA;
610 int r = 0;
612 if (fua) {
613 assert(iscsilun->dpofua);
615 if (!is_sector_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
616 return -EINVAL;
619 if (bs->bl.max_transfer) {
620 assert(nb_sectors << BDRV_SECTOR_BITS <= bs->bl.max_transfer);
623 lba = sector_qemu2lun(sector_num, iscsilun);
624 num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
625 iscsi_co_init_iscsitask(iscsilun, &iTask);
626 qemu_mutex_lock(&iscsilun->mutex);
627 retry:
628 if (iscsilun->use_16_for_rw) {
629 #if LIBISCSI_API_VERSION >= (20160603)
630 iTask.task = iscsi_write16_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
631 NULL, num_sectors * iscsilun->block_size,
632 iscsilun->block_size, 0, 0, fua, 0, 0,
633 iscsi_co_generic_cb, &iTask,
634 (struct scsi_iovec *)iov->iov, iov->niov);
635 } else {
636 iTask.task = iscsi_write10_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
637 NULL, num_sectors * iscsilun->block_size,
638 iscsilun->block_size, 0, 0, fua, 0, 0,
639 iscsi_co_generic_cb, &iTask,
640 (struct scsi_iovec *)iov->iov, iov->niov);
642 #else
643 iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
644 NULL, num_sectors * iscsilun->block_size,
645 iscsilun->block_size, 0, 0, fua, 0, 0,
646 iscsi_co_generic_cb, &iTask);
647 } else {
648 iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
649 NULL, num_sectors * iscsilun->block_size,
650 iscsilun->block_size, 0, 0, fua, 0, 0,
651 iscsi_co_generic_cb, &iTask);
653 #endif
654 if (iTask.task == NULL) {
655 qemu_mutex_unlock(&iscsilun->mutex);
656 return -ENOMEM;
658 #if LIBISCSI_API_VERSION < (20160603)
659 scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
660 iov->niov);
661 #endif
662 iscsi_co_wait_for_task(&iTask, iscsilun);
664 if (iTask.task != NULL) {
665 scsi_free_scsi_task(iTask.task);
666 iTask.task = NULL;
669 if (iTask.do_retry) {
670 iTask.complete = 0;
671 goto retry;
674 if (iTask.status != SCSI_STATUS_GOOD) {
675 iscsi_allocmap_set_invalid(iscsilun, sector_num * BDRV_SECTOR_SIZE,
676 nb_sectors * BDRV_SECTOR_SIZE);
677 error_report("iSCSI WRITE10/16 failed at lba %" PRIu64 ": %s", lba,
678 iTask.err_str);
679 r = iTask.err_code;
680 goto out_unlock;
683 iscsi_allocmap_set_allocated(iscsilun, sector_num * BDRV_SECTOR_SIZE,
684 nb_sectors * BDRV_SECTOR_SIZE);
686 out_unlock:
687 qemu_mutex_unlock(&iscsilun->mutex);
688 g_free(iTask.err_str);
689 return r;
694 static int coroutine_fn iscsi_co_block_status(BlockDriverState *bs,
695 bool want_zero, int64_t offset,
696 int64_t bytes, int64_t *pnum,
697 int64_t *map,
698 BlockDriverState **file)
700 IscsiLun *iscsilun = bs->opaque;
701 struct scsi_get_lba_status *lbas = NULL;
702 struct scsi_lba_status_descriptor *lbasd = NULL;
703 struct IscsiTask iTask;
704 uint64_t lba, max_bytes;
705 int ret;
707 iscsi_co_init_iscsitask(iscsilun, &iTask);
709 assert(QEMU_IS_ALIGNED(offset | bytes, iscsilun->block_size));
711 /* default to all sectors allocated */
712 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
713 if (map) {
714 *map = offset;
716 *pnum = bytes;
718 /* LUN does not support logical block provisioning */
719 if (!iscsilun->lbpme) {
720 goto out;
723 lba = offset / iscsilun->block_size;
724 max_bytes = (iscsilun->num_blocks - lba) * iscsilun->block_size;
726 qemu_mutex_lock(&iscsilun->mutex);
727 retry:
728 if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
729 lba, 8 + 16, iscsi_co_generic_cb,
730 &iTask) == NULL) {
731 ret = -ENOMEM;
732 goto out_unlock;
734 iscsi_co_wait_for_task(&iTask, iscsilun);
736 if (iTask.do_retry) {
737 if (iTask.task != NULL) {
738 scsi_free_scsi_task(iTask.task);
739 iTask.task = NULL;
741 iTask.complete = 0;
742 goto retry;
745 if (iTask.status != SCSI_STATUS_GOOD) {
746 /* in case the get_lba_status_callout fails (i.e.
747 * because the device is busy or the cmd is not
748 * supported) we pretend all blocks are allocated
749 * for backwards compatibility */
750 error_report("iSCSI GET_LBA_STATUS failed at lba %" PRIu64 ": %s",
751 lba, iTask.err_str);
752 goto out_unlock;
755 lbas = scsi_datain_unmarshall(iTask.task);
756 if (lbas == NULL || lbas->num_descriptors == 0) {
757 ret = -EIO;
758 goto out_unlock;
761 lbasd = &lbas->descriptors[0];
763 if (lba != lbasd->lba) {
764 ret = -EIO;
765 goto out_unlock;
768 *pnum = MIN((int64_t) lbasd->num_blocks * iscsilun->block_size, max_bytes);
770 if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
771 lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
772 ret &= ~BDRV_BLOCK_DATA;
773 if (iscsilun->lbprz) {
774 ret |= BDRV_BLOCK_ZERO;
778 if (ret & BDRV_BLOCK_ZERO) {
779 iscsi_allocmap_set_unallocated(iscsilun, offset, *pnum);
780 } else {
781 iscsi_allocmap_set_allocated(iscsilun, offset, *pnum);
784 if (*pnum > bytes) {
785 *pnum = bytes;
787 out_unlock:
788 qemu_mutex_unlock(&iscsilun->mutex);
789 g_free(iTask.err_str);
790 out:
791 if (iTask.task != NULL) {
792 scsi_free_scsi_task(iTask.task);
794 if (ret > 0 && ret & BDRV_BLOCK_OFFSET_VALID && file) {
795 *file = bs;
797 return ret;
800 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
801 int64_t sector_num, int nb_sectors,
802 QEMUIOVector *iov)
804 IscsiLun *iscsilun = bs->opaque;
805 struct IscsiTask iTask;
806 uint64_t lba;
807 uint32_t num_sectors;
808 int r = 0;
810 if (!is_sector_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
811 return -EINVAL;
814 if (bs->bl.max_transfer) {
815 assert(nb_sectors << BDRV_SECTOR_BITS <= bs->bl.max_transfer);
818 /* if cache.direct is off and we have a valid entry in our allocation map
819 * we can skip checking the block status and directly return zeroes if
820 * the request falls within an unallocated area */
821 if (iscsi_allocmap_is_valid(iscsilun, sector_num * BDRV_SECTOR_SIZE,
822 nb_sectors * BDRV_SECTOR_SIZE) &&
823 !iscsi_allocmap_is_allocated(iscsilun, sector_num * BDRV_SECTOR_SIZE,
824 nb_sectors * BDRV_SECTOR_SIZE)) {
825 qemu_iovec_memset(iov, 0, 0x00, iov->size);
826 return 0;
829 if (nb_sectors >= ISCSI_CHECKALLOC_THRES &&
830 !iscsi_allocmap_is_valid(iscsilun, sector_num * BDRV_SECTOR_SIZE,
831 nb_sectors * BDRV_SECTOR_SIZE) &&
832 !iscsi_allocmap_is_allocated(iscsilun, sector_num * BDRV_SECTOR_SIZE,
833 nb_sectors * BDRV_SECTOR_SIZE)) {
834 int64_t pnum;
835 /* check the block status from the beginning of the cluster
836 * containing the start sector */
837 int64_t head;
838 int ret;
840 assert(iscsilun->cluster_size);
841 head = (sector_num * BDRV_SECTOR_SIZE) % iscsilun->cluster_size;
842 ret = iscsi_co_block_status(bs, true,
843 sector_num * BDRV_SECTOR_SIZE - head,
844 BDRV_REQUEST_MAX_BYTES, &pnum, NULL, NULL);
845 if (ret < 0) {
846 return ret;
848 /* if the whole request falls into an unallocated area we can avoid
849 * reading and directly return zeroes instead */
850 if (ret & BDRV_BLOCK_ZERO &&
851 pnum >= nb_sectors * BDRV_SECTOR_SIZE + head) {
852 qemu_iovec_memset(iov, 0, 0x00, iov->size);
853 return 0;
857 lba = sector_qemu2lun(sector_num, iscsilun);
858 num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
860 iscsi_co_init_iscsitask(iscsilun, &iTask);
861 qemu_mutex_lock(&iscsilun->mutex);
862 retry:
863 if (iscsilun->use_16_for_rw) {
864 #if LIBISCSI_API_VERSION >= (20160603)
865 iTask.task = iscsi_read16_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
866 num_sectors * iscsilun->block_size,
867 iscsilun->block_size, 0, 0, 0, 0, 0,
868 iscsi_co_generic_cb, &iTask,
869 (struct scsi_iovec *)iov->iov, iov->niov);
870 } else {
871 iTask.task = iscsi_read10_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
872 num_sectors * iscsilun->block_size,
873 iscsilun->block_size,
874 0, 0, 0, 0, 0,
875 iscsi_co_generic_cb, &iTask,
876 (struct scsi_iovec *)iov->iov, iov->niov);
878 #else
879 iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
880 num_sectors * iscsilun->block_size,
881 iscsilun->block_size, 0, 0, 0, 0, 0,
882 iscsi_co_generic_cb, &iTask);
883 } else {
884 iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
885 num_sectors * iscsilun->block_size,
886 iscsilun->block_size,
887 0, 0, 0, 0, 0,
888 iscsi_co_generic_cb, &iTask);
890 #endif
891 if (iTask.task == NULL) {
892 qemu_mutex_unlock(&iscsilun->mutex);
893 return -ENOMEM;
895 #if LIBISCSI_API_VERSION < (20160603)
896 scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
897 #endif
899 iscsi_co_wait_for_task(&iTask, iscsilun);
900 if (iTask.task != NULL) {
901 scsi_free_scsi_task(iTask.task);
902 iTask.task = NULL;
905 if (iTask.do_retry) {
906 iTask.complete = 0;
907 goto retry;
910 if (iTask.status != SCSI_STATUS_GOOD) {
911 error_report("iSCSI READ10/16 failed at lba %" PRIu64 ": %s",
912 lba, iTask.err_str);
913 r = iTask.err_code;
916 qemu_mutex_unlock(&iscsilun->mutex);
917 g_free(iTask.err_str);
918 return r;
921 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
923 IscsiLun *iscsilun = bs->opaque;
924 struct IscsiTask iTask;
925 int r = 0;
927 iscsi_co_init_iscsitask(iscsilun, &iTask);
928 qemu_mutex_lock(&iscsilun->mutex);
929 retry:
930 if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
931 0, iscsi_co_generic_cb, &iTask) == NULL) {
932 qemu_mutex_unlock(&iscsilun->mutex);
933 return -ENOMEM;
936 iscsi_co_wait_for_task(&iTask, iscsilun);
938 if (iTask.task != NULL) {
939 scsi_free_scsi_task(iTask.task);
940 iTask.task = NULL;
943 if (iTask.do_retry) {
944 iTask.complete = 0;
945 goto retry;
948 if (iTask.status != SCSI_STATUS_GOOD) {
949 error_report("iSCSI SYNCHRONIZECACHE10 failed: %s", iTask.err_str);
950 r = iTask.err_code;
953 qemu_mutex_unlock(&iscsilun->mutex);
954 g_free(iTask.err_str);
955 return r;
958 #ifdef __linux__
959 /* Called (via iscsi_service) with QemuMutex held. */
960 static void
961 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
962 void *command_data, void *opaque)
964 IscsiAIOCB *acb = opaque;
966 if (status == SCSI_STATUS_CANCELLED) {
967 if (!acb->bh) {
968 acb->status = -ECANCELED;
969 iscsi_schedule_bh(acb);
971 return;
974 acb->status = 0;
975 if (status < 0) {
976 error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
977 iscsi_get_error(iscsi));
978 acb->status = -iscsi_translate_sense(&acb->task->sense);
981 acb->ioh->driver_status = 0;
982 acb->ioh->host_status = 0;
983 acb->ioh->resid = 0;
984 acb->ioh->status = status;
986 #define SG_ERR_DRIVER_SENSE 0x08
988 if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
989 int ss;
991 acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
993 acb->ioh->sb_len_wr = acb->task->datain.size - 2;
994 ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
995 acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
996 memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
999 iscsi_schedule_bh(acb);
1002 static void iscsi_ioctl_bh_completion(void *opaque)
1004 IscsiAIOCB *acb = opaque;
1006 qemu_bh_delete(acb->bh);
1007 acb->common.cb(acb->common.opaque, acb->ret);
1008 qemu_aio_unref(acb);
1011 static void iscsi_ioctl_handle_emulated(IscsiAIOCB *acb, int req, void *buf)
1013 BlockDriverState *bs = acb->common.bs;
1014 IscsiLun *iscsilun = bs->opaque;
1015 int ret = 0;
1017 switch (req) {
1018 case SG_GET_VERSION_NUM:
1019 *(int *)buf = 30000;
1020 break;
1021 case SG_GET_SCSI_ID:
1022 ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
1023 break;
1024 default:
1025 ret = -EINVAL;
1027 assert(!acb->bh);
1028 acb->bh = aio_bh_new(bdrv_get_aio_context(bs),
1029 iscsi_ioctl_bh_completion, acb);
1030 acb->ret = ret;
1031 qemu_bh_schedule(acb->bh);
1034 static BlockAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
1035 unsigned long int req, void *buf,
1036 BlockCompletionFunc *cb, void *opaque)
1038 IscsiLun *iscsilun = bs->opaque;
1039 struct iscsi_context *iscsi = iscsilun->iscsi;
1040 struct iscsi_data data;
1041 IscsiAIOCB *acb;
1043 acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
1045 acb->iscsilun = iscsilun;
1046 acb->bh = NULL;
1047 acb->status = -EINPROGRESS;
1048 acb->ioh = buf;
1049 acb->cancelled = false;
1051 if (req != SG_IO) {
1052 iscsi_ioctl_handle_emulated(acb, req, buf);
1053 return &acb->common;
1056 if (acb->ioh->cmd_len > SCSI_CDB_MAX_SIZE) {
1057 error_report("iSCSI: ioctl error CDB exceeds max size (%d > %d)",
1058 acb->ioh->cmd_len, SCSI_CDB_MAX_SIZE);
1059 qemu_aio_unref(acb);
1060 return NULL;
1063 acb->task = malloc(sizeof(struct scsi_task));
1064 if (acb->task == NULL) {
1065 error_report("iSCSI: Failed to allocate task for scsi command. %s",
1066 iscsi_get_error(iscsi));
1067 qemu_aio_unref(acb);
1068 return NULL;
1070 memset(acb->task, 0, sizeof(struct scsi_task));
1072 switch (acb->ioh->dxfer_direction) {
1073 case SG_DXFER_TO_DEV:
1074 acb->task->xfer_dir = SCSI_XFER_WRITE;
1075 break;
1076 case SG_DXFER_FROM_DEV:
1077 acb->task->xfer_dir = SCSI_XFER_READ;
1078 break;
1079 default:
1080 acb->task->xfer_dir = SCSI_XFER_NONE;
1081 break;
1084 acb->task->cdb_size = acb->ioh->cmd_len;
1085 memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
1086 acb->task->expxferlen = acb->ioh->dxfer_len;
1088 data.size = 0;
1089 qemu_mutex_lock(&iscsilun->mutex);
1090 if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
1091 if (acb->ioh->iovec_count == 0) {
1092 data.data = acb->ioh->dxferp;
1093 data.size = acb->ioh->dxfer_len;
1094 } else {
1095 scsi_task_set_iov_out(acb->task,
1096 (struct scsi_iovec *) acb->ioh->dxferp,
1097 acb->ioh->iovec_count);
1101 if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
1102 iscsi_aio_ioctl_cb,
1103 (data.size > 0) ? &data : NULL,
1104 acb) != 0) {
1105 qemu_mutex_unlock(&iscsilun->mutex);
1106 scsi_free_scsi_task(acb->task);
1107 qemu_aio_unref(acb);
1108 return NULL;
1111 /* tell libiscsi to read straight into the buffer we got from ioctl */
1112 if (acb->task->xfer_dir == SCSI_XFER_READ) {
1113 if (acb->ioh->iovec_count == 0) {
1114 scsi_task_add_data_in_buffer(acb->task,
1115 acb->ioh->dxfer_len,
1116 acb->ioh->dxferp);
1117 } else {
1118 scsi_task_set_iov_in(acb->task,
1119 (struct scsi_iovec *) acb->ioh->dxferp,
1120 acb->ioh->iovec_count);
1124 iscsi_set_events(iscsilun);
1125 qemu_mutex_unlock(&iscsilun->mutex);
1127 return &acb->common;
1130 #endif
1132 static int64_t
1133 iscsi_getlength(BlockDriverState *bs)
1135 IscsiLun *iscsilun = bs->opaque;
1136 int64_t len;
1138 len = iscsilun->num_blocks;
1139 len *= iscsilun->block_size;
1141 return len;
1144 static int
1145 coroutine_fn iscsi_co_pdiscard(BlockDriverState *bs, int64_t offset, int bytes)
1147 IscsiLun *iscsilun = bs->opaque;
1148 struct IscsiTask iTask;
1149 struct unmap_list list;
1150 int r = 0;
1152 if (!is_byte_request_lun_aligned(offset, bytes, iscsilun)) {
1153 return -ENOTSUP;
1156 if (!iscsilun->lbp.lbpu) {
1157 /* UNMAP is not supported by the target */
1158 return 0;
1161 list.lba = offset / iscsilun->block_size;
1162 list.num = bytes / iscsilun->block_size;
1164 iscsi_co_init_iscsitask(iscsilun, &iTask);
1165 qemu_mutex_lock(&iscsilun->mutex);
1166 retry:
1167 if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
1168 iscsi_co_generic_cb, &iTask) == NULL) {
1169 r = -ENOMEM;
1170 goto out_unlock;
1173 iscsi_co_wait_for_task(&iTask, iscsilun);
1175 if (iTask.task != NULL) {
1176 scsi_free_scsi_task(iTask.task);
1177 iTask.task = NULL;
1180 if (iTask.do_retry) {
1181 iTask.complete = 0;
1182 goto retry;
1185 iscsi_allocmap_set_invalid(iscsilun, offset, bytes);
1187 if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
1188 /* the target might fail with a check condition if it
1189 is not happy with the alignment of the UNMAP request
1190 we silently fail in this case */
1191 goto out_unlock;
1194 if (iTask.status != SCSI_STATUS_GOOD) {
1195 error_report("iSCSI UNMAP failed at lba %" PRIu64 ": %s",
1196 list.lba, iTask.err_str);
1197 r = iTask.err_code;
1198 goto out_unlock;
1201 out_unlock:
1202 qemu_mutex_unlock(&iscsilun->mutex);
1203 g_free(iTask.err_str);
1204 return r;
1207 static int
1208 coroutine_fn iscsi_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1209 int bytes, BdrvRequestFlags flags)
1211 IscsiLun *iscsilun = bs->opaque;
1212 struct IscsiTask iTask;
1213 uint64_t lba;
1214 uint32_t nb_blocks;
1215 bool use_16_for_ws = iscsilun->use_16_for_rw;
1216 int r = 0;
1218 if (!is_byte_request_lun_aligned(offset, bytes, iscsilun)) {
1219 return -ENOTSUP;
1222 if (flags & BDRV_REQ_MAY_UNMAP) {
1223 if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
1224 /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
1225 use_16_for_ws = true;
1227 if (use_16_for_ws && !iscsilun->lbp.lbpws) {
1228 /* WRITESAME16 with UNMAP is not supported by the target,
1229 * fall back and try WRITESAME10/16 without UNMAP */
1230 flags &= ~BDRV_REQ_MAY_UNMAP;
1231 use_16_for_ws = iscsilun->use_16_for_rw;
1235 if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
1236 /* WRITESAME without UNMAP is not supported by the target */
1237 return -ENOTSUP;
1240 lba = offset / iscsilun->block_size;
1241 nb_blocks = bytes / iscsilun->block_size;
1243 if (iscsilun->zeroblock == NULL) {
1244 iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
1245 if (iscsilun->zeroblock == NULL) {
1246 return -ENOMEM;
1250 qemu_mutex_lock(&iscsilun->mutex);
1251 iscsi_co_init_iscsitask(iscsilun, &iTask);
1252 retry:
1253 if (use_16_for_ws) {
1254 iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
1255 iscsilun->zeroblock, iscsilun->block_size,
1256 nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
1257 0, 0, iscsi_co_generic_cb, &iTask);
1258 } else {
1259 iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
1260 iscsilun->zeroblock, iscsilun->block_size,
1261 nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
1262 0, 0, iscsi_co_generic_cb, &iTask);
1264 if (iTask.task == NULL) {
1265 qemu_mutex_unlock(&iscsilun->mutex);
1266 return -ENOMEM;
1269 iscsi_co_wait_for_task(&iTask, iscsilun);
1271 if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
1272 iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
1273 (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
1274 iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
1275 /* WRITE SAME is not supported by the target */
1276 iscsilun->has_write_same = false;
1277 scsi_free_scsi_task(iTask.task);
1278 r = -ENOTSUP;
1279 goto out_unlock;
1282 if (iTask.task != NULL) {
1283 scsi_free_scsi_task(iTask.task);
1284 iTask.task = NULL;
1287 if (iTask.do_retry) {
1288 iTask.complete = 0;
1289 goto retry;
1292 if (iTask.status != SCSI_STATUS_GOOD) {
1293 iscsi_allocmap_set_invalid(iscsilun, offset, bytes);
1294 error_report("iSCSI WRITESAME10/16 failed at lba %" PRIu64 ": %s",
1295 lba, iTask.err_str);
1296 r = iTask.err_code;
1297 goto out_unlock;
1300 if (flags & BDRV_REQ_MAY_UNMAP) {
1301 iscsi_allocmap_set_invalid(iscsilun, offset, bytes);
1302 } else {
1303 iscsi_allocmap_set_allocated(iscsilun, offset, bytes);
1306 out_unlock:
1307 qemu_mutex_unlock(&iscsilun->mutex);
1308 g_free(iTask.err_str);
1309 return r;
1312 static void apply_chap(struct iscsi_context *iscsi, QemuOpts *opts,
1313 Error **errp)
1315 const char *user = NULL;
1316 const char *password = NULL;
1317 const char *secretid;
1318 char *secret = NULL;
1320 user = qemu_opt_get(opts, "user");
1321 if (!user) {
1322 return;
1325 secretid = qemu_opt_get(opts, "password-secret");
1326 password = qemu_opt_get(opts, "password");
1327 if (secretid && password) {
1328 error_setg(errp, "'password' and 'password-secret' properties are "
1329 "mutually exclusive");
1330 return;
1332 if (secretid) {
1333 secret = qcrypto_secret_lookup_as_utf8(secretid, errp);
1334 if (!secret) {
1335 return;
1337 password = secret;
1338 } else if (!password) {
1339 error_setg(errp, "CHAP username specified but no password was given");
1340 return;
1343 if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
1344 error_setg(errp, "Failed to set initiator username and password");
1347 g_free(secret);
1350 static void apply_header_digest(struct iscsi_context *iscsi, QemuOpts *opts,
1351 Error **errp)
1353 const char *digest = NULL;
1355 digest = qemu_opt_get(opts, "header-digest");
1356 if (!digest) {
1357 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1358 } else if (!strcmp(digest, "crc32c")) {
1359 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1360 } else if (!strcmp(digest, "none")) {
1361 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1362 } else if (!strcmp(digest, "crc32c-none")) {
1363 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1364 } else if (!strcmp(digest, "none-crc32c")) {
1365 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1366 } else {
1367 error_setg(errp, "Invalid header-digest setting : %s", digest);
1371 static char *get_initiator_name(QemuOpts *opts)
1373 const char *name;
1374 char *iscsi_name;
1375 UuidInfo *uuid_info;
1377 name = qemu_opt_get(opts, "initiator-name");
1378 if (name) {
1379 return g_strdup(name);
1382 uuid_info = qmp_query_uuid(NULL);
1383 if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1384 name = qemu_get_vm_name();
1385 } else {
1386 name = uuid_info->UUID;
1388 iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1389 name ? ":" : "", name ? name : "");
1390 qapi_free_UuidInfo(uuid_info);
1391 return iscsi_name;
1394 static void iscsi_nop_timed_event(void *opaque)
1396 IscsiLun *iscsilun = opaque;
1398 qemu_mutex_lock(&iscsilun->mutex);
1399 if (iscsi_get_nops_in_flight(iscsilun->iscsi) >= MAX_NOP_FAILURES) {
1400 error_report("iSCSI: NOP timeout. Reconnecting...");
1401 iscsilun->request_timed_out = true;
1402 } else if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1403 error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1404 goto out;
1407 timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1408 iscsi_set_events(iscsilun);
1410 out:
1411 qemu_mutex_unlock(&iscsilun->mutex);
1414 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1416 struct scsi_task *task = NULL;
1417 struct scsi_readcapacity10 *rc10 = NULL;
1418 struct scsi_readcapacity16 *rc16 = NULL;
1419 int retries = ISCSI_CMD_RETRIES;
1421 do {
1422 if (task != NULL) {
1423 scsi_free_scsi_task(task);
1424 task = NULL;
1427 switch (iscsilun->type) {
1428 case TYPE_DISK:
1429 task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1430 if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1431 rc16 = scsi_datain_unmarshall(task);
1432 if (rc16 == NULL) {
1433 error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1434 } else {
1435 iscsilun->block_size = rc16->block_length;
1436 iscsilun->num_blocks = rc16->returned_lba + 1;
1437 iscsilun->lbpme = !!rc16->lbpme;
1438 iscsilun->lbprz = !!rc16->lbprz;
1439 iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1441 break;
1443 if (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1444 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
1445 break;
1447 /* Fall through and try READ CAPACITY(10) instead. */
1448 case TYPE_ROM:
1449 task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1450 if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1451 rc10 = scsi_datain_unmarshall(task);
1452 if (rc10 == NULL) {
1453 error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1454 } else {
1455 iscsilun->block_size = rc10->block_size;
1456 if (rc10->lba == 0) {
1457 /* blank disk loaded */
1458 iscsilun->num_blocks = 0;
1459 } else {
1460 iscsilun->num_blocks = rc10->lba + 1;
1464 break;
1465 default:
1466 return;
1468 } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1469 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1470 && retries-- > 0);
1472 if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1473 error_setg(errp, "iSCSI: failed to send readcapacity10/16 command");
1474 } else if (!iscsilun->block_size ||
1475 iscsilun->block_size % BDRV_SECTOR_SIZE) {
1476 error_setg(errp, "iSCSI: the target returned an invalid "
1477 "block size of %d.", iscsilun->block_size);
1479 if (task) {
1480 scsi_free_scsi_task(task);
1484 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1485 int evpd, int pc, void **inq, Error **errp)
1487 int full_size;
1488 struct scsi_task *task = NULL;
1489 task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1490 if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1491 goto fail;
1493 full_size = scsi_datain_getfullsize(task);
1494 if (full_size > task->datain.size) {
1495 scsi_free_scsi_task(task);
1497 /* we need more data for the full list */
1498 task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1499 if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1500 goto fail;
1504 *inq = scsi_datain_unmarshall(task);
1505 if (*inq == NULL) {
1506 error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1507 goto fail_with_err;
1510 return task;
1512 fail:
1513 error_setg(errp, "iSCSI: Inquiry command failed : %s",
1514 iscsi_get_error(iscsi));
1515 fail_with_err:
1516 if (task != NULL) {
1517 scsi_free_scsi_task(task);
1519 return NULL;
1522 static void iscsi_detach_aio_context(BlockDriverState *bs)
1524 IscsiLun *iscsilun = bs->opaque;
1526 aio_set_fd_handler(iscsilun->aio_context, iscsi_get_fd(iscsilun->iscsi),
1527 false, NULL, NULL, NULL, NULL);
1528 iscsilun->events = 0;
1530 if (iscsilun->nop_timer) {
1531 timer_del(iscsilun->nop_timer);
1532 timer_free(iscsilun->nop_timer);
1533 iscsilun->nop_timer = NULL;
1535 if (iscsilun->event_timer) {
1536 timer_del(iscsilun->event_timer);
1537 timer_free(iscsilun->event_timer);
1538 iscsilun->event_timer = NULL;
1542 static void iscsi_attach_aio_context(BlockDriverState *bs,
1543 AioContext *new_context)
1545 IscsiLun *iscsilun = bs->opaque;
1547 iscsilun->aio_context = new_context;
1548 iscsi_set_events(iscsilun);
1550 /* Set up a timer for sending out iSCSI NOPs */
1551 iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1552 QEMU_CLOCK_REALTIME, SCALE_MS,
1553 iscsi_nop_timed_event, iscsilun);
1554 timer_mod(iscsilun->nop_timer,
1555 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1557 /* Set up a timer for periodic calls to iscsi_set_events and to
1558 * scan for command timeout */
1559 iscsilun->event_timer = aio_timer_new(iscsilun->aio_context,
1560 QEMU_CLOCK_REALTIME, SCALE_MS,
1561 iscsi_timed_check_events, iscsilun);
1562 timer_mod(iscsilun->event_timer,
1563 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + EVENT_INTERVAL);
1566 static void iscsi_modesense_sync(IscsiLun *iscsilun)
1568 struct scsi_task *task;
1569 struct scsi_mode_sense *ms = NULL;
1570 iscsilun->write_protected = false;
1571 iscsilun->dpofua = false;
1573 task = iscsi_modesense6_sync(iscsilun->iscsi, iscsilun->lun,
1574 1, SCSI_MODESENSE_PC_CURRENT,
1575 0x3F, 0, 255);
1576 if (task == NULL) {
1577 error_report("iSCSI: Failed to send MODE_SENSE(6) command: %s",
1578 iscsi_get_error(iscsilun->iscsi));
1579 goto out;
1582 if (task->status != SCSI_STATUS_GOOD) {
1583 error_report("iSCSI: Failed MODE_SENSE(6), LUN assumed writable");
1584 goto out;
1586 ms = scsi_datain_unmarshall(task);
1587 if (!ms) {
1588 error_report("iSCSI: Failed to unmarshall MODE_SENSE(6) data: %s",
1589 iscsi_get_error(iscsilun->iscsi));
1590 goto out;
1592 iscsilun->write_protected = ms->device_specific_parameter & 0x80;
1593 iscsilun->dpofua = ms->device_specific_parameter & 0x10;
1595 out:
1596 if (task) {
1597 scsi_free_scsi_task(task);
1601 static void iscsi_parse_iscsi_option(const char *target, QDict *options)
1603 QemuOptsList *list;
1604 QemuOpts *opts;
1605 const char *user, *password, *password_secret, *initiator_name,
1606 *header_digest, *timeout;
1608 list = qemu_find_opts("iscsi");
1609 if (!list) {
1610 return;
1613 opts = qemu_opts_find(list, target);
1614 if (opts == NULL) {
1615 opts = QTAILQ_FIRST(&list->head);
1616 if (!opts) {
1617 return;
1621 user = qemu_opt_get(opts, "user");
1622 if (user) {
1623 qdict_set_default_str(options, "user", user);
1626 password = qemu_opt_get(opts, "password");
1627 if (password) {
1628 qdict_set_default_str(options, "password", password);
1631 password_secret = qemu_opt_get(opts, "password-secret");
1632 if (password_secret) {
1633 qdict_set_default_str(options, "password-secret", password_secret);
1636 initiator_name = qemu_opt_get(opts, "initiator-name");
1637 if (initiator_name) {
1638 qdict_set_default_str(options, "initiator-name", initiator_name);
1641 header_digest = qemu_opt_get(opts, "header-digest");
1642 if (header_digest) {
1643 /* -iscsi takes upper case values, but QAPI only supports lower case
1644 * enum constant names, so we have to convert here. */
1645 char *qapi_value = g_ascii_strdown(header_digest, -1);
1646 qdict_set_default_str(options, "header-digest", qapi_value);
1647 g_free(qapi_value);
1650 timeout = qemu_opt_get(opts, "timeout");
1651 if (timeout) {
1652 qdict_set_default_str(options, "timeout", timeout);
1657 * We support iscsi url's on the form
1658 * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1660 static void iscsi_parse_filename(const char *filename, QDict *options,
1661 Error **errp)
1663 struct iscsi_url *iscsi_url;
1664 const char *transport_name;
1665 char *lun_str;
1667 iscsi_url = iscsi_parse_full_url(NULL, filename);
1668 if (iscsi_url == NULL) {
1669 error_setg(errp, "Failed to parse URL : %s", filename);
1670 return;
1673 #if LIBISCSI_API_VERSION >= (20160603)
1674 switch (iscsi_url->transport) {
1675 case TCP_TRANSPORT:
1676 transport_name = "tcp";
1677 break;
1678 case ISER_TRANSPORT:
1679 transport_name = "iser";
1680 break;
1681 default:
1682 error_setg(errp, "Unknown transport type (%d)",
1683 iscsi_url->transport);
1684 return;
1686 #else
1687 transport_name = "tcp";
1688 #endif
1690 qdict_set_default_str(options, "transport", transport_name);
1691 qdict_set_default_str(options, "portal", iscsi_url->portal);
1692 qdict_set_default_str(options, "target", iscsi_url->target);
1694 lun_str = g_strdup_printf("%d", iscsi_url->lun);
1695 qdict_set_default_str(options, "lun", lun_str);
1696 g_free(lun_str);
1698 /* User/password from -iscsi take precedence over those from the URL */
1699 iscsi_parse_iscsi_option(iscsi_url->target, options);
1701 if (iscsi_url->user[0] != '\0') {
1702 qdict_set_default_str(options, "user", iscsi_url->user);
1703 qdict_set_default_str(options, "password", iscsi_url->passwd);
1706 iscsi_destroy_url(iscsi_url);
1709 static QemuOptsList runtime_opts = {
1710 .name = "iscsi",
1711 .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1712 .desc = {
1714 .name = "transport",
1715 .type = QEMU_OPT_STRING,
1718 .name = "portal",
1719 .type = QEMU_OPT_STRING,
1722 .name = "target",
1723 .type = QEMU_OPT_STRING,
1726 .name = "user",
1727 .type = QEMU_OPT_STRING,
1730 .name = "password",
1731 .type = QEMU_OPT_STRING,
1734 .name = "password-secret",
1735 .type = QEMU_OPT_STRING,
1738 .name = "lun",
1739 .type = QEMU_OPT_NUMBER,
1742 .name = "initiator-name",
1743 .type = QEMU_OPT_STRING,
1746 .name = "header-digest",
1747 .type = QEMU_OPT_STRING,
1750 .name = "timeout",
1751 .type = QEMU_OPT_NUMBER,
1753 { /* end of list */ }
1757 static void iscsi_save_designator(IscsiLun *lun,
1758 struct scsi_inquiry_device_identification *inq_di)
1760 struct scsi_inquiry_device_designator *desig, *copy = NULL;
1762 for (desig = inq_di->designators; desig; desig = desig->next) {
1763 if (desig->association ||
1764 desig->designator_type > SCSI_DESIGNATOR_TYPE_NAA) {
1765 continue;
1767 /* NAA works better than T10 vendor ID based designator. */
1768 if (!copy || copy->designator_type < desig->designator_type) {
1769 copy = desig;
1772 if (copy) {
1773 lun->dd = g_new(struct scsi_inquiry_device_designator, 1);
1774 *lun->dd = *copy;
1775 lun->dd->next = NULL;
1776 lun->dd->designator = g_malloc(copy->designator_length);
1777 memcpy(lun->dd->designator, copy->designator, copy->designator_length);
1781 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1782 Error **errp)
1784 IscsiLun *iscsilun = bs->opaque;
1785 struct iscsi_context *iscsi = NULL;
1786 struct scsi_task *task = NULL;
1787 struct scsi_inquiry_standard *inq = NULL;
1788 struct scsi_inquiry_supported_pages *inq_vpd;
1789 char *initiator_name = NULL;
1790 QemuOpts *opts;
1791 Error *local_err = NULL;
1792 const char *transport_name, *portal, *target;
1793 #if LIBISCSI_API_VERSION >= (20160603)
1794 enum iscsi_transport_type transport;
1795 #endif
1796 int i, ret = 0, timeout = 0, lun;
1798 opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1799 qemu_opts_absorb_qdict(opts, options, &local_err);
1800 if (local_err) {
1801 error_propagate(errp, local_err);
1802 ret = -EINVAL;
1803 goto out;
1806 transport_name = qemu_opt_get(opts, "transport");
1807 portal = qemu_opt_get(opts, "portal");
1808 target = qemu_opt_get(opts, "target");
1809 lun = qemu_opt_get_number(opts, "lun", 0);
1811 if (!transport_name || !portal || !target) {
1812 error_setg(errp, "Need all of transport, portal and target options");
1813 ret = -EINVAL;
1814 goto out;
1817 if (!strcmp(transport_name, "tcp")) {
1818 #if LIBISCSI_API_VERSION >= (20160603)
1819 transport = TCP_TRANSPORT;
1820 } else if (!strcmp(transport_name, "iser")) {
1821 transport = ISER_TRANSPORT;
1822 #else
1823 /* TCP is what older libiscsi versions always use */
1824 #endif
1825 } else {
1826 error_setg(errp, "Unknown transport: %s", transport_name);
1827 ret = -EINVAL;
1828 goto out;
1831 memset(iscsilun, 0, sizeof(IscsiLun));
1833 initiator_name = get_initiator_name(opts);
1835 iscsi = iscsi_create_context(initiator_name);
1836 if (iscsi == NULL) {
1837 error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1838 ret = -ENOMEM;
1839 goto out;
1841 #if LIBISCSI_API_VERSION >= (20160603)
1842 if (iscsi_init_transport(iscsi, transport)) {
1843 error_setg(errp, ("Error initializing transport."));
1844 ret = -EINVAL;
1845 goto out;
1847 #endif
1848 if (iscsi_set_targetname(iscsi, target)) {
1849 error_setg(errp, "iSCSI: Failed to set target name.");
1850 ret = -EINVAL;
1851 goto out;
1854 /* check if we got CHAP username/password via the options */
1855 apply_chap(iscsi, opts, &local_err);
1856 if (local_err != NULL) {
1857 error_propagate(errp, local_err);
1858 ret = -EINVAL;
1859 goto out;
1862 if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1863 error_setg(errp, "iSCSI: Failed to set session type to normal.");
1864 ret = -EINVAL;
1865 goto out;
1868 /* check if we got HEADER_DIGEST via the options */
1869 apply_header_digest(iscsi, opts, &local_err);
1870 if (local_err != NULL) {
1871 error_propagate(errp, local_err);
1872 ret = -EINVAL;
1873 goto out;
1876 /* timeout handling is broken in libiscsi before 1.15.0 */
1877 timeout = qemu_opt_get_number(opts, "timeout", 0);
1878 #if LIBISCSI_API_VERSION >= 20150621
1879 iscsi_set_timeout(iscsi, timeout);
1880 #else
1881 if (timeout) {
1882 warn_report("iSCSI: ignoring timeout value for libiscsi <1.15.0");
1884 #endif
1886 if (iscsi_full_connect_sync(iscsi, portal, lun) != 0) {
1887 error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1888 iscsi_get_error(iscsi));
1889 ret = -EINVAL;
1890 goto out;
1893 iscsilun->iscsi = iscsi;
1894 iscsilun->aio_context = bdrv_get_aio_context(bs);
1895 iscsilun->lun = lun;
1896 iscsilun->has_write_same = true;
1898 task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1899 (void **) &inq, errp);
1900 if (task == NULL) {
1901 ret = -EINVAL;
1902 goto out;
1904 iscsilun->type = inq->periperal_device_type;
1905 scsi_free_scsi_task(task);
1906 task = NULL;
1908 iscsi_modesense_sync(iscsilun);
1909 if (iscsilun->dpofua) {
1910 bs->supported_write_flags = BDRV_REQ_FUA;
1913 /* Check the write protect flag of the LUN if we want to write */
1914 if (iscsilun->type == TYPE_DISK && (flags & BDRV_O_RDWR) &&
1915 iscsilun->write_protected) {
1916 ret = bdrv_apply_auto_read_only(bs, "LUN is write protected", errp);
1917 if (ret < 0) {
1918 goto out;
1920 flags &= ~BDRV_O_RDWR;
1923 iscsi_readcapacity_sync(iscsilun, &local_err);
1924 if (local_err != NULL) {
1925 error_propagate(errp, local_err);
1926 ret = -EINVAL;
1927 goto out;
1929 bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1931 /* We don't have any emulation for devices other than disks and CD-ROMs, so
1932 * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1933 * will try to read from the device to guess the image format.
1935 if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1936 bs->sg = true;
1939 task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1940 SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1941 (void **) &inq_vpd, errp);
1942 if (task == NULL) {
1943 ret = -EINVAL;
1944 goto out;
1946 for (i = 0; i < inq_vpd->num_pages; i++) {
1947 struct scsi_task *inq_task;
1948 struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1949 struct scsi_inquiry_block_limits *inq_bl;
1950 struct scsi_inquiry_device_identification *inq_di;
1951 switch (inq_vpd->pages[i]) {
1952 case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1953 inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1954 SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1955 (void **) &inq_lbp, errp);
1956 if (inq_task == NULL) {
1957 ret = -EINVAL;
1958 goto out;
1960 memcpy(&iscsilun->lbp, inq_lbp,
1961 sizeof(struct scsi_inquiry_logical_block_provisioning));
1962 scsi_free_scsi_task(inq_task);
1963 break;
1964 case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1965 inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1966 SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1967 (void **) &inq_bl, errp);
1968 if (inq_task == NULL) {
1969 ret = -EINVAL;
1970 goto out;
1972 memcpy(&iscsilun->bl, inq_bl,
1973 sizeof(struct scsi_inquiry_block_limits));
1974 scsi_free_scsi_task(inq_task);
1975 break;
1976 case SCSI_INQUIRY_PAGECODE_DEVICE_IDENTIFICATION:
1977 inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1978 SCSI_INQUIRY_PAGECODE_DEVICE_IDENTIFICATION,
1979 (void **) &inq_di, errp);
1980 if (inq_task == NULL) {
1981 ret = -EINVAL;
1982 goto out;
1984 iscsi_save_designator(iscsilun, inq_di);
1985 scsi_free_scsi_task(inq_task);
1986 break;
1987 default:
1988 break;
1991 scsi_free_scsi_task(task);
1992 task = NULL;
1994 qemu_mutex_init(&iscsilun->mutex);
1995 iscsi_attach_aio_context(bs, iscsilun->aio_context);
1997 /* Guess the internal cluster (page) size of the iscsi target by the means
1998 * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1999 * reasonable size */
2000 if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
2001 iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
2002 iscsilun->cluster_size = iscsilun->bl.opt_unmap_gran *
2003 iscsilun->block_size;
2004 if (iscsilun->lbprz) {
2005 ret = iscsi_allocmap_init(iscsilun, bs->open_flags);
2009 if (iscsilun->lbprz && iscsilun->lbp.lbpws) {
2010 bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP;
2013 out:
2014 qemu_opts_del(opts);
2015 g_free(initiator_name);
2016 if (task != NULL) {
2017 scsi_free_scsi_task(task);
2020 if (ret) {
2021 if (iscsi != NULL) {
2022 if (iscsi_is_logged_in(iscsi)) {
2023 iscsi_logout_sync(iscsi);
2025 iscsi_destroy_context(iscsi);
2027 memset(iscsilun, 0, sizeof(IscsiLun));
2030 return ret;
2033 static void iscsi_close(BlockDriverState *bs)
2035 IscsiLun *iscsilun = bs->opaque;
2036 struct iscsi_context *iscsi = iscsilun->iscsi;
2038 iscsi_detach_aio_context(bs);
2039 if (iscsi_is_logged_in(iscsi)) {
2040 iscsi_logout_sync(iscsi);
2042 iscsi_destroy_context(iscsi);
2043 if (iscsilun->dd) {
2044 g_free(iscsilun->dd->designator);
2045 g_free(iscsilun->dd);
2047 g_free(iscsilun->zeroblock);
2048 iscsi_allocmap_free(iscsilun);
2049 qemu_mutex_destroy(&iscsilun->mutex);
2050 memset(iscsilun, 0, sizeof(IscsiLun));
2053 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
2055 /* We don't actually refresh here, but just return data queried in
2056 * iscsi_open(): iscsi targets don't change their limits. */
2058 IscsiLun *iscsilun = bs->opaque;
2059 uint64_t max_xfer_len = iscsilun->use_16_for_rw ? 0xffffffff : 0xffff;
2060 unsigned int block_size = MAX(BDRV_SECTOR_SIZE, iscsilun->block_size);
2062 assert(iscsilun->block_size >= BDRV_SECTOR_SIZE || bs->sg);
2064 bs->bl.request_alignment = block_size;
2066 if (iscsilun->bl.max_xfer_len) {
2067 max_xfer_len = MIN(max_xfer_len, iscsilun->bl.max_xfer_len);
2070 if (max_xfer_len * block_size < INT_MAX) {
2071 bs->bl.max_transfer = max_xfer_len * iscsilun->block_size;
2074 if (iscsilun->lbp.lbpu) {
2075 if (iscsilun->bl.max_unmap < 0xffffffff / block_size) {
2076 bs->bl.max_pdiscard =
2077 iscsilun->bl.max_unmap * iscsilun->block_size;
2079 bs->bl.pdiscard_alignment =
2080 iscsilun->bl.opt_unmap_gran * iscsilun->block_size;
2081 } else {
2082 bs->bl.pdiscard_alignment = iscsilun->block_size;
2085 if (iscsilun->bl.max_ws_len < 0xffffffff / block_size) {
2086 bs->bl.max_pwrite_zeroes =
2087 iscsilun->bl.max_ws_len * iscsilun->block_size;
2089 if (iscsilun->lbp.lbpws) {
2090 bs->bl.pwrite_zeroes_alignment =
2091 iscsilun->bl.opt_unmap_gran * iscsilun->block_size;
2092 } else {
2093 bs->bl.pwrite_zeroes_alignment = iscsilun->block_size;
2095 if (iscsilun->bl.opt_xfer_len &&
2096 iscsilun->bl.opt_xfer_len < INT_MAX / block_size) {
2097 bs->bl.opt_transfer = pow2floor(iscsilun->bl.opt_xfer_len *
2098 iscsilun->block_size);
2102 /* Note that this will not re-establish a connection with an iSCSI target - it
2103 * is effectively a NOP. */
2104 static int iscsi_reopen_prepare(BDRVReopenState *state,
2105 BlockReopenQueue *queue, Error **errp)
2107 IscsiLun *iscsilun = state->bs->opaque;
2109 if (state->flags & BDRV_O_RDWR && iscsilun->write_protected) {
2110 error_setg(errp, "Cannot open a write protected LUN as read-write");
2111 return -EACCES;
2113 return 0;
2116 static void iscsi_reopen_commit(BDRVReopenState *reopen_state)
2118 IscsiLun *iscsilun = reopen_state->bs->opaque;
2120 /* the cache.direct status might have changed */
2121 if (iscsilun->allocmap != NULL) {
2122 iscsi_allocmap_init(iscsilun, reopen_state->flags);
2126 static int coroutine_fn iscsi_co_truncate(BlockDriverState *bs, int64_t offset,
2127 bool exact, PreallocMode prealloc,
2128 Error **errp)
2130 IscsiLun *iscsilun = bs->opaque;
2131 int64_t cur_length;
2132 Error *local_err = NULL;
2134 if (prealloc != PREALLOC_MODE_OFF) {
2135 error_setg(errp, "Unsupported preallocation mode '%s'",
2136 PreallocMode_str(prealloc));
2137 return -ENOTSUP;
2140 if (iscsilun->type != TYPE_DISK) {
2141 error_setg(errp, "Cannot resize non-disk iSCSI devices");
2142 return -ENOTSUP;
2145 iscsi_readcapacity_sync(iscsilun, &local_err);
2146 if (local_err != NULL) {
2147 error_propagate(errp, local_err);
2148 return -EIO;
2151 cur_length = iscsi_getlength(bs);
2152 if (offset != cur_length && exact) {
2153 error_setg(errp, "Cannot resize iSCSI devices");
2154 return -ENOTSUP;
2155 } else if (offset > cur_length) {
2156 error_setg(errp, "Cannot grow iSCSI devices");
2157 return -EINVAL;
2160 if (iscsilun->allocmap != NULL) {
2161 iscsi_allocmap_init(iscsilun, bs->open_flags);
2164 return 0;
2167 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
2169 IscsiLun *iscsilun = bs->opaque;
2170 bdi->unallocated_blocks_are_zero = iscsilun->lbprz;
2171 bdi->cluster_size = iscsilun->cluster_size;
2172 return 0;
2175 static void coroutine_fn iscsi_co_invalidate_cache(BlockDriverState *bs,
2176 Error **errp)
2178 IscsiLun *iscsilun = bs->opaque;
2179 iscsi_allocmap_invalidate(iscsilun);
2182 static int coroutine_fn iscsi_co_copy_range_from(BlockDriverState *bs,
2183 BdrvChild *src,
2184 uint64_t src_offset,
2185 BdrvChild *dst,
2186 uint64_t dst_offset,
2187 uint64_t bytes,
2188 BdrvRequestFlags read_flags,
2189 BdrvRequestFlags write_flags)
2191 return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
2192 read_flags, write_flags);
2195 static struct scsi_task *iscsi_xcopy_task(int param_len)
2197 struct scsi_task *task;
2199 task = g_new0(struct scsi_task, 1);
2201 task->cdb[0] = EXTENDED_COPY;
2202 task->cdb[10] = (param_len >> 24) & 0xFF;
2203 task->cdb[11] = (param_len >> 16) & 0xFF;
2204 task->cdb[12] = (param_len >> 8) & 0xFF;
2205 task->cdb[13] = param_len & 0xFF;
2206 task->cdb_size = 16;
2207 task->xfer_dir = SCSI_XFER_WRITE;
2208 task->expxferlen = param_len;
2210 return task;
2213 static void iscsi_populate_target_desc(unsigned char *desc, IscsiLun *lun)
2215 struct scsi_inquiry_device_designator *dd = lun->dd;
2217 memset(desc, 0, 32);
2218 desc[0] = 0xE4; /* IDENT_DESCR_TGT_DESCR */
2219 desc[4] = dd->code_set;
2220 desc[5] = (dd->designator_type & 0xF)
2221 | ((dd->association & 3) << 4);
2222 desc[7] = dd->designator_length;
2223 memcpy(desc + 8, dd->designator, MIN(dd->designator_length, 20));
2225 desc[28] = 0;
2226 desc[29] = (lun->block_size >> 16) & 0xFF;
2227 desc[30] = (lun->block_size >> 8) & 0xFF;
2228 desc[31] = lun->block_size & 0xFF;
2231 static void iscsi_xcopy_desc_hdr(uint8_t *hdr, int dc, int cat, int src_index,
2232 int dst_index)
2234 hdr[0] = 0x02; /* BLK_TO_BLK_SEG_DESCR */
2235 hdr[1] = ((dc << 1) | cat) & 0xFF;
2236 hdr[2] = (XCOPY_BLK2BLK_SEG_DESC_SIZE >> 8) & 0xFF;
2237 /* don't account for the first 4 bytes in descriptor header*/
2238 hdr[3] = (XCOPY_BLK2BLK_SEG_DESC_SIZE - 4 /* SEG_DESC_SRC_INDEX_OFFSET */) & 0xFF;
2239 hdr[4] = (src_index >> 8) & 0xFF;
2240 hdr[5] = src_index & 0xFF;
2241 hdr[6] = (dst_index >> 8) & 0xFF;
2242 hdr[7] = dst_index & 0xFF;
2245 static void iscsi_xcopy_populate_desc(uint8_t *desc, int dc, int cat,
2246 int src_index, int dst_index, int num_blks,
2247 uint64_t src_lba, uint64_t dst_lba)
2249 iscsi_xcopy_desc_hdr(desc, dc, cat, src_index, dst_index);
2251 /* The caller should verify the request size */
2252 assert(num_blks < 65536);
2253 desc[10] = (num_blks >> 8) & 0xFF;
2254 desc[11] = num_blks & 0xFF;
2255 desc[12] = (src_lba >> 56) & 0xFF;
2256 desc[13] = (src_lba >> 48) & 0xFF;
2257 desc[14] = (src_lba >> 40) & 0xFF;
2258 desc[15] = (src_lba >> 32) & 0xFF;
2259 desc[16] = (src_lba >> 24) & 0xFF;
2260 desc[17] = (src_lba >> 16) & 0xFF;
2261 desc[18] = (src_lba >> 8) & 0xFF;
2262 desc[19] = src_lba & 0xFF;
2263 desc[20] = (dst_lba >> 56) & 0xFF;
2264 desc[21] = (dst_lba >> 48) & 0xFF;
2265 desc[22] = (dst_lba >> 40) & 0xFF;
2266 desc[23] = (dst_lba >> 32) & 0xFF;
2267 desc[24] = (dst_lba >> 24) & 0xFF;
2268 desc[25] = (dst_lba >> 16) & 0xFF;
2269 desc[26] = (dst_lba >> 8) & 0xFF;
2270 desc[27] = dst_lba & 0xFF;
2273 static void iscsi_xcopy_populate_header(unsigned char *buf, int list_id, int str,
2274 int list_id_usage, int prio,
2275 int tgt_desc_len,
2276 int seg_desc_len, int inline_data_len)
2278 buf[0] = list_id;
2279 buf[1] = ((str & 1) << 5) | ((list_id_usage & 3) << 3) | (prio & 7);
2280 buf[2] = (tgt_desc_len >> 8) & 0xFF;
2281 buf[3] = tgt_desc_len & 0xFF;
2282 buf[8] = (seg_desc_len >> 24) & 0xFF;
2283 buf[9] = (seg_desc_len >> 16) & 0xFF;
2284 buf[10] = (seg_desc_len >> 8) & 0xFF;
2285 buf[11] = seg_desc_len & 0xFF;
2286 buf[12] = (inline_data_len >> 24) & 0xFF;
2287 buf[13] = (inline_data_len >> 16) & 0xFF;
2288 buf[14] = (inline_data_len >> 8) & 0xFF;
2289 buf[15] = inline_data_len & 0xFF;
2292 static void iscsi_xcopy_data(struct iscsi_data *data,
2293 IscsiLun *src, int64_t src_lba,
2294 IscsiLun *dst, int64_t dst_lba,
2295 uint16_t num_blocks)
2297 uint8_t *buf;
2298 const int src_offset = XCOPY_DESC_OFFSET;
2299 const int dst_offset = XCOPY_DESC_OFFSET + IDENT_DESCR_TGT_DESCR_SIZE;
2300 const int seg_offset = dst_offset + IDENT_DESCR_TGT_DESCR_SIZE;
2302 data->size = XCOPY_DESC_OFFSET +
2303 IDENT_DESCR_TGT_DESCR_SIZE * 2 +
2304 XCOPY_BLK2BLK_SEG_DESC_SIZE;
2305 data->data = g_malloc0(data->size);
2306 buf = data->data;
2308 /* Initialise the parameter list header */
2309 iscsi_xcopy_populate_header(buf, 1, 0, 2 /* LIST_ID_USAGE_DISCARD */,
2310 0, 2 * IDENT_DESCR_TGT_DESCR_SIZE,
2311 XCOPY_BLK2BLK_SEG_DESC_SIZE,
2314 /* Initialise CSCD list with one src + one dst descriptor */
2315 iscsi_populate_target_desc(&buf[src_offset], src);
2316 iscsi_populate_target_desc(&buf[dst_offset], dst);
2318 /* Initialise one segment descriptor */
2319 iscsi_xcopy_populate_desc(&buf[seg_offset], 0, 0, 0, 1, num_blocks,
2320 src_lba, dst_lba);
2323 static int coroutine_fn iscsi_co_copy_range_to(BlockDriverState *bs,
2324 BdrvChild *src,
2325 uint64_t src_offset,
2326 BdrvChild *dst,
2327 uint64_t dst_offset,
2328 uint64_t bytes,
2329 BdrvRequestFlags read_flags,
2330 BdrvRequestFlags write_flags)
2332 IscsiLun *dst_lun = dst->bs->opaque;
2333 IscsiLun *src_lun;
2334 struct IscsiTask iscsi_task;
2335 struct iscsi_data data;
2336 int r = 0;
2337 int block_size;
2339 if (src->bs->drv->bdrv_co_copy_range_to != iscsi_co_copy_range_to) {
2340 return -ENOTSUP;
2342 src_lun = src->bs->opaque;
2344 if (!src_lun->dd || !dst_lun->dd) {
2345 return -ENOTSUP;
2347 if (!is_byte_request_lun_aligned(dst_offset, bytes, dst_lun)) {
2348 return -ENOTSUP;
2350 if (!is_byte_request_lun_aligned(src_offset, bytes, src_lun)) {
2351 return -ENOTSUP;
2353 if (dst_lun->block_size != src_lun->block_size ||
2354 !dst_lun->block_size) {
2355 return -ENOTSUP;
2358 block_size = dst_lun->block_size;
2359 if (bytes / block_size > 65535) {
2360 return -ENOTSUP;
2363 iscsi_xcopy_data(&data,
2364 src_lun, src_offset / block_size,
2365 dst_lun, dst_offset / block_size,
2366 bytes / block_size);
2368 iscsi_co_init_iscsitask(dst_lun, &iscsi_task);
2370 qemu_mutex_lock(&dst_lun->mutex);
2371 iscsi_task.task = iscsi_xcopy_task(data.size);
2372 retry:
2373 if (iscsi_scsi_command_async(dst_lun->iscsi, dst_lun->lun,
2374 iscsi_task.task, iscsi_co_generic_cb,
2375 &data,
2376 &iscsi_task) != 0) {
2377 r = -EIO;
2378 goto out_unlock;
2381 iscsi_co_wait_for_task(&iscsi_task, dst_lun);
2383 if (iscsi_task.do_retry) {
2384 iscsi_task.complete = 0;
2385 goto retry;
2388 if (iscsi_task.status != SCSI_STATUS_GOOD) {
2389 r = iscsi_task.err_code;
2390 goto out_unlock;
2393 out_unlock:
2395 trace_iscsi_xcopy(src_lun, src_offset, dst_lun, dst_offset, bytes, r);
2396 g_free(iscsi_task.task);
2397 qemu_mutex_unlock(&dst_lun->mutex);
2398 g_free(iscsi_task.err_str);
2399 return r;
2403 static const char *const iscsi_strong_runtime_opts[] = {
2404 "transport",
2405 "portal",
2406 "target",
2407 "user",
2408 "password",
2409 "password-secret",
2410 "lun",
2411 "initiator-name",
2412 "header-digest",
2414 NULL
2417 static BlockDriver bdrv_iscsi = {
2418 .format_name = "iscsi",
2419 .protocol_name = "iscsi",
2421 .instance_size = sizeof(IscsiLun),
2422 .bdrv_parse_filename = iscsi_parse_filename,
2423 .bdrv_file_open = iscsi_open,
2424 .bdrv_close = iscsi_close,
2425 .bdrv_co_create_opts = bdrv_co_create_opts_simple,
2426 .create_opts = &bdrv_create_opts_simple,
2427 .bdrv_reopen_prepare = iscsi_reopen_prepare,
2428 .bdrv_reopen_commit = iscsi_reopen_commit,
2429 .bdrv_co_invalidate_cache = iscsi_co_invalidate_cache,
2431 .bdrv_getlength = iscsi_getlength,
2432 .bdrv_get_info = iscsi_get_info,
2433 .bdrv_co_truncate = iscsi_co_truncate,
2434 .bdrv_refresh_limits = iscsi_refresh_limits,
2436 .bdrv_co_block_status = iscsi_co_block_status,
2437 .bdrv_co_pdiscard = iscsi_co_pdiscard,
2438 .bdrv_co_copy_range_from = iscsi_co_copy_range_from,
2439 .bdrv_co_copy_range_to = iscsi_co_copy_range_to,
2440 .bdrv_co_pwrite_zeroes = iscsi_co_pwrite_zeroes,
2441 .bdrv_co_readv = iscsi_co_readv,
2442 .bdrv_co_writev = iscsi_co_writev,
2443 .bdrv_co_flush_to_disk = iscsi_co_flush,
2445 #ifdef __linux__
2446 .bdrv_aio_ioctl = iscsi_aio_ioctl,
2447 #endif
2449 .bdrv_detach_aio_context = iscsi_detach_aio_context,
2450 .bdrv_attach_aio_context = iscsi_attach_aio_context,
2452 .strong_runtime_opts = iscsi_strong_runtime_opts,
2455 #if LIBISCSI_API_VERSION >= (20160603)
2456 static BlockDriver bdrv_iser = {
2457 .format_name = "iser",
2458 .protocol_name = "iser",
2460 .instance_size = sizeof(IscsiLun),
2461 .bdrv_parse_filename = iscsi_parse_filename,
2462 .bdrv_file_open = iscsi_open,
2463 .bdrv_close = iscsi_close,
2464 .bdrv_co_create_opts = bdrv_co_create_opts_simple,
2465 .create_opts = &bdrv_create_opts_simple,
2466 .bdrv_reopen_prepare = iscsi_reopen_prepare,
2467 .bdrv_reopen_commit = iscsi_reopen_commit,
2468 .bdrv_co_invalidate_cache = iscsi_co_invalidate_cache,
2470 .bdrv_getlength = iscsi_getlength,
2471 .bdrv_get_info = iscsi_get_info,
2472 .bdrv_co_truncate = iscsi_co_truncate,
2473 .bdrv_refresh_limits = iscsi_refresh_limits,
2475 .bdrv_co_block_status = iscsi_co_block_status,
2476 .bdrv_co_pdiscard = iscsi_co_pdiscard,
2477 .bdrv_co_copy_range_from = iscsi_co_copy_range_from,
2478 .bdrv_co_copy_range_to = iscsi_co_copy_range_to,
2479 .bdrv_co_pwrite_zeroes = iscsi_co_pwrite_zeroes,
2480 .bdrv_co_readv = iscsi_co_readv,
2481 .bdrv_co_writev = iscsi_co_writev,
2482 .bdrv_co_flush_to_disk = iscsi_co_flush,
2484 #ifdef __linux__
2485 .bdrv_aio_ioctl = iscsi_aio_ioctl,
2486 #endif
2488 .bdrv_detach_aio_context = iscsi_detach_aio_context,
2489 .bdrv_attach_aio_context = iscsi_attach_aio_context,
2491 .strong_runtime_opts = iscsi_strong_runtime_opts,
2493 #endif
2495 static void iscsi_block_init(void)
2497 bdrv_register(&bdrv_iscsi);
2498 #if LIBISCSI_API_VERSION >= (20160603)
2499 bdrv_register(&bdrv_iser);
2500 #endif
2503 block_init(iscsi_block_init);