hw: virtio-pci: drop DO_UPCAST
[qemu/ar7.git] / block / iscsi.c
blobff473206e6892e71bd680144651e1ab384e9a863
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/config-file.h"
32 #include "qemu/error-report.h"
33 #include "qemu/bitops.h"
34 #include "qemu/bitmap.h"
35 #include "block/block_int.h"
36 #include "block/qdict.h"
37 #include "scsi/constants.h"
38 #include "qemu/iov.h"
39 #include "qemu/option.h"
40 #include "qemu/uuid.h"
41 #include "qapi/error.h"
42 #include "qapi/qapi-commands-misc.h"
43 #include "qapi/qmp/qdict.h"
44 #include "qapi/qmp/qstring.h"
45 #include "crypto/secret.h"
46 #include "scsi/utils.h"
47 #include "trace.h"
49 /* Conflict between scsi/utils.h and libiscsi! :( */
50 #define SCSI_XFER_NONE ISCSI_XFER_NONE
51 #include <iscsi/iscsi.h>
52 #define inline __attribute__((gnu_inline)) /* required for libiscsi v1.9.0 */
53 #include <iscsi/scsi-lowlevel.h>
54 #undef inline
55 #undef SCSI_XFER_NONE
56 QEMU_BUILD_BUG_ON((int)SCSI_XFER_NONE != (int)ISCSI_XFER_NONE);
58 #ifdef __linux__
59 #include <scsi/sg.h>
60 #endif
62 typedef struct IscsiLun {
63 struct iscsi_context *iscsi;
64 AioContext *aio_context;
65 int lun;
66 enum scsi_inquiry_peripheral_device_type type;
67 int block_size;
68 uint64_t num_blocks;
69 int events;
70 QEMUTimer *nop_timer;
71 QEMUTimer *event_timer;
72 QemuMutex mutex;
73 struct scsi_inquiry_logical_block_provisioning lbp;
74 struct scsi_inquiry_block_limits bl;
75 struct scsi_inquiry_device_designator *dd;
76 unsigned char *zeroblock;
77 /* The allocmap tracks which clusters (pages) on the iSCSI target are
78 * allocated and which are not. In case a target returns zeros for
79 * unallocated pages (iscsilun->lprz) we can directly return zeros instead
80 * of reading zeros over the wire if a read request falls within an
81 * unallocated block. As there are 3 possible states we need 2 bitmaps to
82 * track. allocmap_valid keeps track if QEMU's information about a page is
83 * valid. allocmap tracks if a page is allocated or not. In case QEMU has no
84 * valid information about a page the corresponding allocmap entry should be
85 * switched to unallocated as well to force a new lookup of the allocation
86 * status as lookups are generally skipped if a page is suspect to be
87 * allocated. If a iSCSI target is opened with cache.direct = on the
88 * allocmap_valid does not exist turning all cached information invalid so
89 * that a fresh lookup is made for any page even if allocmap entry returns
90 * it's unallocated. */
91 unsigned long *allocmap;
92 unsigned long *allocmap_valid;
93 long allocmap_size;
94 int cluster_size;
95 bool use_16_for_rw;
96 bool write_protected;
97 bool lbpme;
98 bool lbprz;
99 bool dpofua;
100 bool has_write_same;
101 bool request_timed_out;
102 } IscsiLun;
104 typedef struct IscsiTask {
105 int status;
106 int complete;
107 int retries;
108 int do_retry;
109 struct scsi_task *task;
110 Coroutine *co;
111 IscsiLun *iscsilun;
112 QEMUTimer retry_timer;
113 int err_code;
114 char *err_str;
115 } IscsiTask;
117 typedef struct IscsiAIOCB {
118 BlockAIOCB common;
119 QEMUBH *bh;
120 IscsiLun *iscsilun;
121 struct scsi_task *task;
122 int status;
123 int64_t sector_num;
124 int nb_sectors;
125 int ret;
126 #ifdef __linux__
127 sg_io_hdr_t *ioh;
128 #endif
129 bool cancelled;
130 } IscsiAIOCB;
132 /* libiscsi uses time_t so its enough to process events every second */
133 #define EVENT_INTERVAL 1000
134 #define NOP_INTERVAL 5000
135 #define MAX_NOP_FAILURES 3
136 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
137 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048, 8192, 32768};
139 /* this threshold is a trade-off knob to choose between
140 * the potential additional overhead of an extra GET_LBA_STATUS request
141 * vs. unnecessarily reading a lot of zero sectors over the wire.
142 * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
143 * sectors we check the allocation status of the area covered by the
144 * request first if the allocationmap indicates that the area might be
145 * unallocated. */
146 #define ISCSI_CHECKALLOC_THRES 64
148 static void
149 iscsi_bh_cb(void *p)
151 IscsiAIOCB *acb = p;
153 qemu_bh_delete(acb->bh);
155 acb->common.cb(acb->common.opaque, acb->status);
157 if (acb->task != NULL) {
158 scsi_free_scsi_task(acb->task);
159 acb->task = NULL;
162 qemu_aio_unref(acb);
165 static void
166 iscsi_schedule_bh(IscsiAIOCB *acb)
168 if (acb->bh) {
169 return;
171 acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
172 qemu_bh_schedule(acb->bh);
175 static void iscsi_co_generic_bh_cb(void *opaque)
177 struct IscsiTask *iTask = opaque;
179 iTask->complete = 1;
180 aio_co_wake(iTask->co);
183 static void iscsi_retry_timer_expired(void *opaque)
185 struct IscsiTask *iTask = opaque;
186 iTask->complete = 1;
187 if (iTask->co) {
188 aio_co_wake(iTask->co);
192 static inline unsigned exp_random(double mean)
194 return -mean * log((double)rand() / RAND_MAX);
197 /* SCSI_SENSE_ASCQ_INVALID_FIELD_IN_PARAMETER_LIST was introduced in
198 * libiscsi 1.10.0, together with other constants we need. Use it as
199 * a hint that we have to define them ourselves if needed, to keep the
200 * minimum required libiscsi version at 1.9.0. We use an ASCQ macro for
201 * the test because SCSI_STATUS_* is an enum.
203 * To guard against future changes where SCSI_SENSE_ASCQ_* also becomes
204 * an enum, check against the LIBISCSI_API_VERSION macro, which was
205 * introduced in 1.11.0. If it is present, there is no need to define
206 * anything.
208 #if !defined(SCSI_SENSE_ASCQ_INVALID_FIELD_IN_PARAMETER_LIST) && \
209 !defined(LIBISCSI_API_VERSION)
210 #define SCSI_STATUS_TASK_SET_FULL 0x28
211 #define SCSI_STATUS_TIMEOUT 0x0f000002
212 #define SCSI_SENSE_ASCQ_INVALID_FIELD_IN_PARAMETER_LIST 0x2600
213 #define SCSI_SENSE_ASCQ_PARAMETER_LIST_LENGTH_ERROR 0x1a00
214 #endif
216 #ifndef LIBISCSI_API_VERSION
217 #define LIBISCSI_API_VERSION 20130701
218 #endif
220 static int iscsi_translate_sense(struct scsi_sense *sense)
222 return - scsi_sense_to_errno(sense->key,
223 (sense->ascq & 0xFF00) >> 8,
224 sense->ascq & 0xFF);
227 /* Called (via iscsi_service) with QemuMutex held. */
228 static void
229 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
230 void *command_data, void *opaque)
232 struct IscsiTask *iTask = opaque;
233 struct scsi_task *task = command_data;
235 iTask->status = status;
236 iTask->do_retry = 0;
237 iTask->task = task;
239 if (status != SCSI_STATUS_GOOD) {
240 if (iTask->retries++ < ISCSI_CMD_RETRIES) {
241 if (status == SCSI_STATUS_CHECK_CONDITION
242 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
243 error_report("iSCSI CheckCondition: %s",
244 iscsi_get_error(iscsi));
245 iTask->do_retry = 1;
246 goto out;
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;
269 return;
272 iTask->err_code = iscsi_translate_sense(&task->sense);
273 iTask->err_str = g_strdup(iscsi_get_error(iscsi));
276 out:
277 if (iTask->co) {
278 aio_bh_schedule_oneshot(iTask->iscsilun->aio_context,
279 iscsi_co_generic_bh_cb, iTask);
280 } else {
281 iTask->complete = 1;
285 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
287 *iTask = (struct IscsiTask) {
288 .co = qemu_coroutine_self(),
289 .iscsilun = iscsilun,
293 /* Called (via iscsi_service) with QemuMutex held. */
294 static void
295 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
296 void *private_data)
298 IscsiAIOCB *acb = private_data;
300 /* If the command callback hasn't been called yet, drop the task */
301 if (!acb->bh) {
302 /* Call iscsi_aio_ioctl_cb() with SCSI_STATUS_CANCELLED */
303 iscsi_scsi_cancel_task(iscsi, acb->task);
306 qemu_aio_unref(acb); /* acquired in iscsi_aio_cancel() */
309 static void
310 iscsi_aio_cancel(BlockAIOCB *blockacb)
312 IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
313 IscsiLun *iscsilun = acb->iscsilun;
315 qemu_mutex_lock(&iscsilun->mutex);
317 /* If it was cancelled or completed already, our work is done here */
318 if (acb->cancelled || acb->status != -EINPROGRESS) {
319 qemu_mutex_unlock(&iscsilun->mutex);
320 return;
323 acb->cancelled = true;
325 qemu_aio_ref(acb); /* released in iscsi_abort_task_cb() */
327 /* send a task mgmt call to the target to cancel the task on the target */
328 if (iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
329 iscsi_abort_task_cb, acb) < 0) {
330 qemu_aio_unref(acb); /* since iscsi_abort_task_cb() won't be called */
333 qemu_mutex_unlock(&iscsilun->mutex);
336 static const AIOCBInfo iscsi_aiocb_info = {
337 .aiocb_size = sizeof(IscsiAIOCB),
338 .cancel_async = iscsi_aio_cancel,
342 static void iscsi_process_read(void *arg);
343 static void iscsi_process_write(void *arg);
345 /* Called with QemuMutex held. */
346 static void
347 iscsi_set_events(IscsiLun *iscsilun)
349 struct iscsi_context *iscsi = iscsilun->iscsi;
350 int ev = iscsi_which_events(iscsi);
352 if (ev != iscsilun->events) {
353 aio_set_fd_handler(iscsilun->aio_context, iscsi_get_fd(iscsi),
354 false,
355 (ev & POLLIN) ? iscsi_process_read : NULL,
356 (ev & POLLOUT) ? iscsi_process_write : NULL,
357 NULL,
358 iscsilun);
359 iscsilun->events = ev;
363 static void iscsi_timed_check_events(void *opaque)
365 IscsiLun *iscsilun = opaque;
367 qemu_mutex_lock(&iscsilun->mutex);
369 /* check for timed out requests */
370 iscsi_service(iscsilun->iscsi, 0);
372 if (iscsilun->request_timed_out) {
373 iscsilun->request_timed_out = false;
374 iscsi_reconnect(iscsilun->iscsi);
377 /* newer versions of libiscsi may return zero events. Ensure we are able
378 * to return to service once this situation changes. */
379 iscsi_set_events(iscsilun);
381 qemu_mutex_unlock(&iscsilun->mutex);
383 timer_mod(iscsilun->event_timer,
384 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + EVENT_INTERVAL);
387 static void
388 iscsi_process_read(void *arg)
390 IscsiLun *iscsilun = arg;
391 struct iscsi_context *iscsi = iscsilun->iscsi;
393 qemu_mutex_lock(&iscsilun->mutex);
394 iscsi_service(iscsi, POLLIN);
395 iscsi_set_events(iscsilun);
396 qemu_mutex_unlock(&iscsilun->mutex);
399 static void
400 iscsi_process_write(void *arg)
402 IscsiLun *iscsilun = arg;
403 struct iscsi_context *iscsi = iscsilun->iscsi;
405 qemu_mutex_lock(&iscsilun->mutex);
406 iscsi_service(iscsi, POLLOUT);
407 iscsi_set_events(iscsilun);
408 qemu_mutex_unlock(&iscsilun->mutex);
411 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
413 return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
416 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
418 return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
421 static bool is_byte_request_lun_aligned(int64_t offset, int count,
422 IscsiLun *iscsilun)
424 if (offset % iscsilun->block_size || count % iscsilun->block_size) {
425 error_report("iSCSI misaligned request: "
426 "iscsilun->block_size %u, offset %" PRIi64
427 ", count %d",
428 iscsilun->block_size, offset, count);
429 return false;
431 return true;
434 static bool is_sector_request_lun_aligned(int64_t sector_num, int nb_sectors,
435 IscsiLun *iscsilun)
437 assert(nb_sectors <= BDRV_REQUEST_MAX_SECTORS);
438 return is_byte_request_lun_aligned(sector_num << BDRV_SECTOR_BITS,
439 nb_sectors << BDRV_SECTOR_BITS,
440 iscsilun);
443 static void iscsi_allocmap_free(IscsiLun *iscsilun)
445 g_free(iscsilun->allocmap);
446 g_free(iscsilun->allocmap_valid);
447 iscsilun->allocmap = NULL;
448 iscsilun->allocmap_valid = NULL;
452 static int iscsi_allocmap_init(IscsiLun *iscsilun, int open_flags)
454 iscsi_allocmap_free(iscsilun);
456 assert(iscsilun->cluster_size);
457 iscsilun->allocmap_size =
458 DIV_ROUND_UP(iscsilun->num_blocks * iscsilun->block_size,
459 iscsilun->cluster_size);
461 iscsilun->allocmap = bitmap_try_new(iscsilun->allocmap_size);
462 if (!iscsilun->allocmap) {
463 return -ENOMEM;
466 if (open_flags & BDRV_O_NOCACHE) {
467 /* when cache.direct = on all allocmap entries are
468 * treated as invalid to force a relookup of the block
469 * status on every read request */
470 return 0;
473 iscsilun->allocmap_valid = bitmap_try_new(iscsilun->allocmap_size);
474 if (!iscsilun->allocmap_valid) {
475 /* if we are under memory pressure free the allocmap as well */
476 iscsi_allocmap_free(iscsilun);
477 return -ENOMEM;
480 return 0;
483 static void
484 iscsi_allocmap_update(IscsiLun *iscsilun, int64_t offset,
485 int64_t bytes, bool allocated, bool valid)
487 int64_t cl_num_expanded, nb_cls_expanded, cl_num_shrunk, nb_cls_shrunk;
489 if (iscsilun->allocmap == NULL) {
490 return;
492 /* expand to entirely contain all affected clusters */
493 assert(iscsilun->cluster_size);
494 cl_num_expanded = offset / iscsilun->cluster_size;
495 nb_cls_expanded = DIV_ROUND_UP(offset + bytes,
496 iscsilun->cluster_size) - cl_num_expanded;
497 /* shrink to touch only completely contained clusters */
498 cl_num_shrunk = DIV_ROUND_UP(offset, iscsilun->cluster_size);
499 nb_cls_shrunk = (offset + bytes) / iscsilun->cluster_size - cl_num_shrunk;
500 if (allocated) {
501 bitmap_set(iscsilun->allocmap, cl_num_expanded, nb_cls_expanded);
502 } else {
503 if (nb_cls_shrunk > 0) {
504 bitmap_clear(iscsilun->allocmap, cl_num_shrunk, nb_cls_shrunk);
508 if (iscsilun->allocmap_valid == NULL) {
509 return;
511 if (valid) {
512 if (nb_cls_shrunk > 0) {
513 bitmap_set(iscsilun->allocmap_valid, cl_num_shrunk, nb_cls_shrunk);
515 } else {
516 bitmap_clear(iscsilun->allocmap_valid, cl_num_expanded,
517 nb_cls_expanded);
521 static void
522 iscsi_allocmap_set_allocated(IscsiLun *iscsilun, int64_t offset,
523 int64_t bytes)
525 iscsi_allocmap_update(iscsilun, offset, bytes, true, true);
528 static void
529 iscsi_allocmap_set_unallocated(IscsiLun *iscsilun, int64_t offset,
530 int64_t bytes)
532 /* Note: if cache.direct=on the fifth argument to iscsi_allocmap_update
533 * is ignored, so this will in effect be an iscsi_allocmap_set_invalid.
535 iscsi_allocmap_update(iscsilun, offset, bytes, false, true);
538 static void iscsi_allocmap_set_invalid(IscsiLun *iscsilun, int64_t offset,
539 int64_t bytes)
541 iscsi_allocmap_update(iscsilun, offset, bytes, false, false);
544 static void iscsi_allocmap_invalidate(IscsiLun *iscsilun)
546 if (iscsilun->allocmap) {
547 bitmap_zero(iscsilun->allocmap, iscsilun->allocmap_size);
549 if (iscsilun->allocmap_valid) {
550 bitmap_zero(iscsilun->allocmap_valid, iscsilun->allocmap_size);
554 static inline bool
555 iscsi_allocmap_is_allocated(IscsiLun *iscsilun, int64_t offset,
556 int64_t bytes)
558 unsigned long size;
559 if (iscsilun->allocmap == NULL) {
560 return true;
562 assert(iscsilun->cluster_size);
563 size = DIV_ROUND_UP(offset + bytes, iscsilun->cluster_size);
564 return !(find_next_bit(iscsilun->allocmap, size,
565 offset / iscsilun->cluster_size) == size);
568 static inline bool iscsi_allocmap_is_valid(IscsiLun *iscsilun,
569 int64_t offset, int64_t bytes)
571 unsigned long size;
572 if (iscsilun->allocmap_valid == NULL) {
573 return false;
575 assert(iscsilun->cluster_size);
576 size = DIV_ROUND_UP(offset + bytes, iscsilun->cluster_size);
577 return (find_next_zero_bit(iscsilun->allocmap_valid, size,
578 offset / iscsilun->cluster_size) == size);
581 static void coroutine_fn iscsi_co_wait_for_task(IscsiTask *iTask,
582 IscsiLun *iscsilun)
584 while (!iTask->complete) {
585 iscsi_set_events(iscsilun);
586 qemu_mutex_unlock(&iscsilun->mutex);
587 qemu_coroutine_yield();
588 qemu_mutex_lock(&iscsilun->mutex);
592 static int coroutine_fn
593 iscsi_co_writev(BlockDriverState *bs, int64_t sector_num, int nb_sectors,
594 QEMUIOVector *iov, int flags)
596 IscsiLun *iscsilun = bs->opaque;
597 struct IscsiTask iTask;
598 uint64_t lba;
599 uint32_t num_sectors;
600 bool fua = flags & BDRV_REQ_FUA;
601 int r = 0;
603 if (fua) {
604 assert(iscsilun->dpofua);
606 if (!is_sector_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
607 return -EINVAL;
610 if (bs->bl.max_transfer) {
611 assert(nb_sectors << BDRV_SECTOR_BITS <= bs->bl.max_transfer);
614 lba = sector_qemu2lun(sector_num, iscsilun);
615 num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
616 iscsi_co_init_iscsitask(iscsilun, &iTask);
617 qemu_mutex_lock(&iscsilun->mutex);
618 retry:
619 if (iscsilun->use_16_for_rw) {
620 #if LIBISCSI_API_VERSION >= (20160603)
621 iTask.task = iscsi_write16_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
622 NULL, num_sectors * iscsilun->block_size,
623 iscsilun->block_size, 0, 0, fua, 0, 0,
624 iscsi_co_generic_cb, &iTask,
625 (struct scsi_iovec *)iov->iov, iov->niov);
626 } else {
627 iTask.task = iscsi_write10_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
628 NULL, num_sectors * iscsilun->block_size,
629 iscsilun->block_size, 0, 0, fua, 0, 0,
630 iscsi_co_generic_cb, &iTask,
631 (struct scsi_iovec *)iov->iov, iov->niov);
633 #else
634 iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
635 NULL, num_sectors * iscsilun->block_size,
636 iscsilun->block_size, 0, 0, fua, 0, 0,
637 iscsi_co_generic_cb, &iTask);
638 } else {
639 iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
640 NULL, num_sectors * iscsilun->block_size,
641 iscsilun->block_size, 0, 0, fua, 0, 0,
642 iscsi_co_generic_cb, &iTask);
644 #endif
645 if (iTask.task == NULL) {
646 qemu_mutex_unlock(&iscsilun->mutex);
647 return -ENOMEM;
649 #if LIBISCSI_API_VERSION < (20160603)
650 scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
651 iov->niov);
652 #endif
653 iscsi_co_wait_for_task(&iTask, iscsilun);
655 if (iTask.task != NULL) {
656 scsi_free_scsi_task(iTask.task);
657 iTask.task = NULL;
660 if (iTask.do_retry) {
661 iTask.complete = 0;
662 goto retry;
665 if (iTask.status != SCSI_STATUS_GOOD) {
666 iscsi_allocmap_set_invalid(iscsilun, sector_num * BDRV_SECTOR_SIZE,
667 nb_sectors * BDRV_SECTOR_SIZE);
668 error_report("iSCSI WRITE10/16 failed at lba %" PRIu64 ": %s", lba,
669 iTask.err_str);
670 r = iTask.err_code;
671 goto out_unlock;
674 iscsi_allocmap_set_allocated(iscsilun, sector_num * BDRV_SECTOR_SIZE,
675 nb_sectors * BDRV_SECTOR_SIZE);
677 out_unlock:
678 qemu_mutex_unlock(&iscsilun->mutex);
679 g_free(iTask.err_str);
680 return r;
685 static int coroutine_fn iscsi_co_block_status(BlockDriverState *bs,
686 bool want_zero, int64_t offset,
687 int64_t bytes, int64_t *pnum,
688 int64_t *map,
689 BlockDriverState **file)
691 IscsiLun *iscsilun = bs->opaque;
692 struct scsi_get_lba_status *lbas = NULL;
693 struct scsi_lba_status_descriptor *lbasd = NULL;
694 struct IscsiTask iTask;
695 uint64_t lba;
696 int ret;
698 iscsi_co_init_iscsitask(iscsilun, &iTask);
700 assert(QEMU_IS_ALIGNED(offset | bytes, iscsilun->block_size));
702 /* default to all sectors allocated */
703 ret = BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
704 if (map) {
705 *map = offset;
707 *pnum = bytes;
709 /* LUN does not support logical block provisioning */
710 if (!iscsilun->lbpme) {
711 goto out;
714 lba = offset / iscsilun->block_size;
716 qemu_mutex_lock(&iscsilun->mutex);
717 retry:
718 if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
719 lba, 8 + 16, iscsi_co_generic_cb,
720 &iTask) == NULL) {
721 ret = -ENOMEM;
722 goto out_unlock;
724 iscsi_co_wait_for_task(&iTask, iscsilun);
726 if (iTask.do_retry) {
727 if (iTask.task != NULL) {
728 scsi_free_scsi_task(iTask.task);
729 iTask.task = NULL;
731 iTask.complete = 0;
732 goto retry;
735 if (iTask.status != SCSI_STATUS_GOOD) {
736 /* in case the get_lba_status_callout fails (i.e.
737 * because the device is busy or the cmd is not
738 * supported) we pretend all blocks are allocated
739 * for backwards compatibility */
740 error_report("iSCSI GET_LBA_STATUS failed at lba %" PRIu64 ": %s",
741 lba, iTask.err_str);
742 goto out_unlock;
745 lbas = scsi_datain_unmarshall(iTask.task);
746 if (lbas == NULL) {
747 ret = -EIO;
748 goto out_unlock;
751 lbasd = &lbas->descriptors[0];
753 if (lba != lbasd->lba) {
754 ret = -EIO;
755 goto out_unlock;
758 *pnum = (int64_t) lbasd->num_blocks * iscsilun->block_size;
760 if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
761 lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
762 ret &= ~BDRV_BLOCK_DATA;
763 if (iscsilun->lbprz) {
764 ret |= BDRV_BLOCK_ZERO;
768 if (ret & BDRV_BLOCK_ZERO) {
769 iscsi_allocmap_set_unallocated(iscsilun, offset, *pnum);
770 } else {
771 iscsi_allocmap_set_allocated(iscsilun, offset, *pnum);
774 if (*pnum > bytes) {
775 *pnum = bytes;
777 out_unlock:
778 qemu_mutex_unlock(&iscsilun->mutex);
779 g_free(iTask.err_str);
780 out:
781 if (iTask.task != NULL) {
782 scsi_free_scsi_task(iTask.task);
784 if (ret > 0 && ret & BDRV_BLOCK_OFFSET_VALID && file) {
785 *file = bs;
787 return ret;
790 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
791 int64_t sector_num, int nb_sectors,
792 QEMUIOVector *iov)
794 IscsiLun *iscsilun = bs->opaque;
795 struct IscsiTask iTask;
796 uint64_t lba;
797 uint32_t num_sectors;
798 int r = 0;
800 if (!is_sector_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
801 return -EINVAL;
804 if (bs->bl.max_transfer) {
805 assert(nb_sectors << BDRV_SECTOR_BITS <= bs->bl.max_transfer);
808 /* if cache.direct is off and we have a valid entry in our allocation map
809 * we can skip checking the block status and directly return zeroes if
810 * the request falls within an unallocated area */
811 if (iscsi_allocmap_is_valid(iscsilun, sector_num * BDRV_SECTOR_SIZE,
812 nb_sectors * BDRV_SECTOR_SIZE) &&
813 !iscsi_allocmap_is_allocated(iscsilun, sector_num * BDRV_SECTOR_SIZE,
814 nb_sectors * BDRV_SECTOR_SIZE)) {
815 qemu_iovec_memset(iov, 0, 0x00, iov->size);
816 return 0;
819 if (nb_sectors >= ISCSI_CHECKALLOC_THRES &&
820 !iscsi_allocmap_is_valid(iscsilun, sector_num * BDRV_SECTOR_SIZE,
821 nb_sectors * BDRV_SECTOR_SIZE) &&
822 !iscsi_allocmap_is_allocated(iscsilun, sector_num * BDRV_SECTOR_SIZE,
823 nb_sectors * BDRV_SECTOR_SIZE)) {
824 int64_t pnum;
825 /* check the block status from the beginning of the cluster
826 * containing the start sector */
827 int64_t head;
828 int ret;
830 assert(iscsilun->cluster_size);
831 head = (sector_num * BDRV_SECTOR_SIZE) % iscsilun->cluster_size;
832 ret = iscsi_co_block_status(bs, true,
833 sector_num * BDRV_SECTOR_SIZE - head,
834 BDRV_REQUEST_MAX_BYTES, &pnum, NULL, NULL);
835 if (ret < 0) {
836 return ret;
838 /* if the whole request falls into an unallocated area we can avoid
839 * reading and directly return zeroes instead */
840 if (ret & BDRV_BLOCK_ZERO &&
841 pnum >= nb_sectors * BDRV_SECTOR_SIZE + head) {
842 qemu_iovec_memset(iov, 0, 0x00, iov->size);
843 return 0;
847 lba = sector_qemu2lun(sector_num, iscsilun);
848 num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
850 iscsi_co_init_iscsitask(iscsilun, &iTask);
851 qemu_mutex_lock(&iscsilun->mutex);
852 retry:
853 if (iscsilun->use_16_for_rw) {
854 #if LIBISCSI_API_VERSION >= (20160603)
855 iTask.task = iscsi_read16_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
856 num_sectors * iscsilun->block_size,
857 iscsilun->block_size, 0, 0, 0, 0, 0,
858 iscsi_co_generic_cb, &iTask,
859 (struct scsi_iovec *)iov->iov, iov->niov);
860 } else {
861 iTask.task = iscsi_read10_iov_task(iscsilun->iscsi, iscsilun->lun, lba,
862 num_sectors * iscsilun->block_size,
863 iscsilun->block_size,
864 0, 0, 0, 0, 0,
865 iscsi_co_generic_cb, &iTask,
866 (struct scsi_iovec *)iov->iov, iov->niov);
868 #else
869 iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
870 num_sectors * iscsilun->block_size,
871 iscsilun->block_size, 0, 0, 0, 0, 0,
872 iscsi_co_generic_cb, &iTask);
873 } else {
874 iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
875 num_sectors * iscsilun->block_size,
876 iscsilun->block_size,
877 0, 0, 0, 0, 0,
878 iscsi_co_generic_cb, &iTask);
880 #endif
881 if (iTask.task == NULL) {
882 qemu_mutex_unlock(&iscsilun->mutex);
883 return -ENOMEM;
885 #if LIBISCSI_API_VERSION < (20160603)
886 scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
887 #endif
889 iscsi_co_wait_for_task(&iTask, iscsilun);
890 if (iTask.task != NULL) {
891 scsi_free_scsi_task(iTask.task);
892 iTask.task = NULL;
895 if (iTask.do_retry) {
896 iTask.complete = 0;
897 goto retry;
900 if (iTask.status != SCSI_STATUS_GOOD) {
901 error_report("iSCSI READ10/16 failed at lba %" PRIu64 ": %s",
902 lba, iTask.err_str);
903 r = iTask.err_code;
906 qemu_mutex_unlock(&iscsilun->mutex);
907 g_free(iTask.err_str);
908 return r;
911 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
913 IscsiLun *iscsilun = bs->opaque;
914 struct IscsiTask iTask;
915 int r = 0;
917 iscsi_co_init_iscsitask(iscsilun, &iTask);
918 qemu_mutex_lock(&iscsilun->mutex);
919 retry:
920 if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
921 0, iscsi_co_generic_cb, &iTask) == NULL) {
922 qemu_mutex_unlock(&iscsilun->mutex);
923 return -ENOMEM;
926 iscsi_co_wait_for_task(&iTask, iscsilun);
928 if (iTask.task != NULL) {
929 scsi_free_scsi_task(iTask.task);
930 iTask.task = NULL;
933 if (iTask.do_retry) {
934 iTask.complete = 0;
935 goto retry;
938 if (iTask.status != SCSI_STATUS_GOOD) {
939 error_report("iSCSI SYNCHRONIZECACHE10 failed: %s", iTask.err_str);
940 r = iTask.err_code;
943 qemu_mutex_unlock(&iscsilun->mutex);
944 g_free(iTask.err_str);
945 return r;
948 #ifdef __linux__
949 /* Called (via iscsi_service) with QemuMutex held. */
950 static void
951 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
952 void *command_data, void *opaque)
954 IscsiAIOCB *acb = opaque;
956 if (status == SCSI_STATUS_CANCELLED) {
957 if (!acb->bh) {
958 acb->status = -ECANCELED;
959 iscsi_schedule_bh(acb);
961 return;
964 acb->status = 0;
965 if (status < 0) {
966 error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
967 iscsi_get_error(iscsi));
968 acb->status = iscsi_translate_sense(&acb->task->sense);
971 acb->ioh->driver_status = 0;
972 acb->ioh->host_status = 0;
973 acb->ioh->resid = 0;
974 acb->ioh->status = status;
976 #define SG_ERR_DRIVER_SENSE 0x08
978 if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
979 int ss;
981 acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
983 acb->ioh->sb_len_wr = acb->task->datain.size - 2;
984 ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
985 acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
986 memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
989 iscsi_schedule_bh(acb);
992 static void iscsi_ioctl_bh_completion(void *opaque)
994 IscsiAIOCB *acb = opaque;
996 qemu_bh_delete(acb->bh);
997 acb->common.cb(acb->common.opaque, acb->ret);
998 qemu_aio_unref(acb);
1001 static void iscsi_ioctl_handle_emulated(IscsiAIOCB *acb, int req, void *buf)
1003 BlockDriverState *bs = acb->common.bs;
1004 IscsiLun *iscsilun = bs->opaque;
1005 int ret = 0;
1007 switch (req) {
1008 case SG_GET_VERSION_NUM:
1009 *(int *)buf = 30000;
1010 break;
1011 case SG_GET_SCSI_ID:
1012 ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
1013 break;
1014 default:
1015 ret = -EINVAL;
1017 assert(!acb->bh);
1018 acb->bh = aio_bh_new(bdrv_get_aio_context(bs),
1019 iscsi_ioctl_bh_completion, acb);
1020 acb->ret = ret;
1021 qemu_bh_schedule(acb->bh);
1024 static BlockAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
1025 unsigned long int req, void *buf,
1026 BlockCompletionFunc *cb, void *opaque)
1028 IscsiLun *iscsilun = bs->opaque;
1029 struct iscsi_context *iscsi = iscsilun->iscsi;
1030 struct iscsi_data data;
1031 IscsiAIOCB *acb;
1033 acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
1035 acb->iscsilun = iscsilun;
1036 acb->bh = NULL;
1037 acb->status = -EINPROGRESS;
1038 acb->ioh = buf;
1039 acb->cancelled = false;
1041 if (req != SG_IO) {
1042 iscsi_ioctl_handle_emulated(acb, req, buf);
1043 return &acb->common;
1046 if (acb->ioh->cmd_len > SCSI_CDB_MAX_SIZE) {
1047 error_report("iSCSI: ioctl error CDB exceeds max size (%d > %d)",
1048 acb->ioh->cmd_len, SCSI_CDB_MAX_SIZE);
1049 qemu_aio_unref(acb);
1050 return NULL;
1053 acb->task = malloc(sizeof(struct scsi_task));
1054 if (acb->task == NULL) {
1055 error_report("iSCSI: Failed to allocate task for scsi command. %s",
1056 iscsi_get_error(iscsi));
1057 qemu_aio_unref(acb);
1058 return NULL;
1060 memset(acb->task, 0, sizeof(struct scsi_task));
1062 switch (acb->ioh->dxfer_direction) {
1063 case SG_DXFER_TO_DEV:
1064 acb->task->xfer_dir = SCSI_XFER_WRITE;
1065 break;
1066 case SG_DXFER_FROM_DEV:
1067 acb->task->xfer_dir = SCSI_XFER_READ;
1068 break;
1069 default:
1070 acb->task->xfer_dir = SCSI_XFER_NONE;
1071 break;
1074 acb->task->cdb_size = acb->ioh->cmd_len;
1075 memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
1076 acb->task->expxferlen = acb->ioh->dxfer_len;
1078 data.size = 0;
1079 qemu_mutex_lock(&iscsilun->mutex);
1080 if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
1081 if (acb->ioh->iovec_count == 0) {
1082 data.data = acb->ioh->dxferp;
1083 data.size = acb->ioh->dxfer_len;
1084 } else {
1085 scsi_task_set_iov_out(acb->task,
1086 (struct scsi_iovec *) acb->ioh->dxferp,
1087 acb->ioh->iovec_count);
1091 if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
1092 iscsi_aio_ioctl_cb,
1093 (data.size > 0) ? &data : NULL,
1094 acb) != 0) {
1095 qemu_mutex_unlock(&iscsilun->mutex);
1096 scsi_free_scsi_task(acb->task);
1097 qemu_aio_unref(acb);
1098 return NULL;
1101 /* tell libiscsi to read straight into the buffer we got from ioctl */
1102 if (acb->task->xfer_dir == SCSI_XFER_READ) {
1103 if (acb->ioh->iovec_count == 0) {
1104 scsi_task_add_data_in_buffer(acb->task,
1105 acb->ioh->dxfer_len,
1106 acb->ioh->dxferp);
1107 } else {
1108 scsi_task_set_iov_in(acb->task,
1109 (struct scsi_iovec *) acb->ioh->dxferp,
1110 acb->ioh->iovec_count);
1114 iscsi_set_events(iscsilun);
1115 qemu_mutex_unlock(&iscsilun->mutex);
1117 return &acb->common;
1120 #endif
1122 static int64_t
1123 iscsi_getlength(BlockDriverState *bs)
1125 IscsiLun *iscsilun = bs->opaque;
1126 int64_t len;
1128 len = iscsilun->num_blocks;
1129 len *= iscsilun->block_size;
1131 return len;
1134 static int
1135 coroutine_fn iscsi_co_pdiscard(BlockDriverState *bs, int64_t offset, int bytes)
1137 IscsiLun *iscsilun = bs->opaque;
1138 struct IscsiTask iTask;
1139 struct unmap_list list;
1140 int r = 0;
1142 if (!is_byte_request_lun_aligned(offset, bytes, iscsilun)) {
1143 return -ENOTSUP;
1146 if (!iscsilun->lbp.lbpu) {
1147 /* UNMAP is not supported by the target */
1148 return 0;
1151 list.lba = offset / iscsilun->block_size;
1152 list.num = bytes / iscsilun->block_size;
1154 iscsi_co_init_iscsitask(iscsilun, &iTask);
1155 qemu_mutex_lock(&iscsilun->mutex);
1156 retry:
1157 if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
1158 iscsi_co_generic_cb, &iTask) == NULL) {
1159 r = -ENOMEM;
1160 goto out_unlock;
1163 iscsi_co_wait_for_task(&iTask, iscsilun);
1165 if (iTask.task != NULL) {
1166 scsi_free_scsi_task(iTask.task);
1167 iTask.task = NULL;
1170 if (iTask.do_retry) {
1171 iTask.complete = 0;
1172 goto retry;
1175 iscsi_allocmap_set_invalid(iscsilun, offset, bytes);
1177 if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
1178 /* the target might fail with a check condition if it
1179 is not happy with the alignment of the UNMAP request
1180 we silently fail in this case */
1181 goto out_unlock;
1184 if (iTask.status != SCSI_STATUS_GOOD) {
1185 error_report("iSCSI UNMAP failed at lba %" PRIu64 ": %s",
1186 list.lba, iTask.err_str);
1187 r = iTask.err_code;
1188 goto out_unlock;
1191 out_unlock:
1192 qemu_mutex_unlock(&iscsilun->mutex);
1193 g_free(iTask.err_str);
1194 return r;
1197 static int
1198 coroutine_fn iscsi_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1199 int bytes, BdrvRequestFlags flags)
1201 IscsiLun *iscsilun = bs->opaque;
1202 struct IscsiTask iTask;
1203 uint64_t lba;
1204 uint32_t nb_blocks;
1205 bool use_16_for_ws = iscsilun->use_16_for_rw;
1206 int r = 0;
1208 if (!is_byte_request_lun_aligned(offset, bytes, iscsilun)) {
1209 return -ENOTSUP;
1212 if (flags & BDRV_REQ_MAY_UNMAP) {
1213 if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
1214 /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
1215 use_16_for_ws = true;
1217 if (use_16_for_ws && !iscsilun->lbp.lbpws) {
1218 /* WRITESAME16 with UNMAP is not supported by the target,
1219 * fall back and try WRITESAME10/16 without UNMAP */
1220 flags &= ~BDRV_REQ_MAY_UNMAP;
1221 use_16_for_ws = iscsilun->use_16_for_rw;
1225 if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
1226 /* WRITESAME without UNMAP is not supported by the target */
1227 return -ENOTSUP;
1230 lba = offset / iscsilun->block_size;
1231 nb_blocks = bytes / iscsilun->block_size;
1233 if (iscsilun->zeroblock == NULL) {
1234 iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
1235 if (iscsilun->zeroblock == NULL) {
1236 return -ENOMEM;
1240 qemu_mutex_lock(&iscsilun->mutex);
1241 iscsi_co_init_iscsitask(iscsilun, &iTask);
1242 retry:
1243 if (use_16_for_ws) {
1244 iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
1245 iscsilun->zeroblock, iscsilun->block_size,
1246 nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
1247 0, 0, iscsi_co_generic_cb, &iTask);
1248 } else {
1249 iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
1250 iscsilun->zeroblock, iscsilun->block_size,
1251 nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
1252 0, 0, iscsi_co_generic_cb, &iTask);
1254 if (iTask.task == NULL) {
1255 qemu_mutex_unlock(&iscsilun->mutex);
1256 return -ENOMEM;
1259 iscsi_co_wait_for_task(&iTask, iscsilun);
1261 if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
1262 iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
1263 (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
1264 iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
1265 /* WRITE SAME is not supported by the target */
1266 iscsilun->has_write_same = false;
1267 scsi_free_scsi_task(iTask.task);
1268 r = -ENOTSUP;
1269 goto out_unlock;
1272 if (iTask.task != NULL) {
1273 scsi_free_scsi_task(iTask.task);
1274 iTask.task = NULL;
1277 if (iTask.do_retry) {
1278 iTask.complete = 0;
1279 goto retry;
1282 if (iTask.status != SCSI_STATUS_GOOD) {
1283 iscsi_allocmap_set_invalid(iscsilun, offset, bytes);
1284 error_report("iSCSI WRITESAME10/16 failed at lba %" PRIu64 ": %s",
1285 lba, iTask.err_str);
1286 r = iTask.err_code;
1287 goto out_unlock;
1290 if (flags & BDRV_REQ_MAY_UNMAP) {
1291 iscsi_allocmap_set_invalid(iscsilun, offset, bytes);
1292 } else {
1293 iscsi_allocmap_set_allocated(iscsilun, offset, bytes);
1296 out_unlock:
1297 qemu_mutex_unlock(&iscsilun->mutex);
1298 g_free(iTask.err_str);
1299 return r;
1302 static void apply_chap(struct iscsi_context *iscsi, QemuOpts *opts,
1303 Error **errp)
1305 const char *user = NULL;
1306 const char *password = NULL;
1307 const char *secretid;
1308 char *secret = NULL;
1310 user = qemu_opt_get(opts, "user");
1311 if (!user) {
1312 return;
1315 secretid = qemu_opt_get(opts, "password-secret");
1316 password = qemu_opt_get(opts, "password");
1317 if (secretid && password) {
1318 error_setg(errp, "'password' and 'password-secret' properties are "
1319 "mutually exclusive");
1320 return;
1322 if (secretid) {
1323 secret = qcrypto_secret_lookup_as_utf8(secretid, errp);
1324 if (!secret) {
1325 return;
1327 password = secret;
1328 } else if (!password) {
1329 error_setg(errp, "CHAP username specified but no password was given");
1330 return;
1333 if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
1334 error_setg(errp, "Failed to set initiator username and password");
1337 g_free(secret);
1340 static void apply_header_digest(struct iscsi_context *iscsi, QemuOpts *opts,
1341 Error **errp)
1343 const char *digest = NULL;
1345 digest = qemu_opt_get(opts, "header-digest");
1346 if (!digest) {
1347 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1348 } else if (!strcmp(digest, "crc32c")) {
1349 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1350 } else if (!strcmp(digest, "none")) {
1351 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1352 } else if (!strcmp(digest, "crc32c-none")) {
1353 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1354 } else if (!strcmp(digest, "none-crc32c")) {
1355 iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1356 } else {
1357 error_setg(errp, "Invalid header-digest setting : %s", digest);
1361 static char *get_initiator_name(QemuOpts *opts)
1363 const char *name;
1364 char *iscsi_name;
1365 UuidInfo *uuid_info;
1367 name = qemu_opt_get(opts, "initiator-name");
1368 if (name) {
1369 return g_strdup(name);
1372 uuid_info = qmp_query_uuid(NULL);
1373 if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1374 name = qemu_get_vm_name();
1375 } else {
1376 name = uuid_info->UUID;
1378 iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1379 name ? ":" : "", name ? name : "");
1380 qapi_free_UuidInfo(uuid_info);
1381 return iscsi_name;
1384 static void iscsi_nop_timed_event(void *opaque)
1386 IscsiLun *iscsilun = opaque;
1388 qemu_mutex_lock(&iscsilun->mutex);
1389 if (iscsi_get_nops_in_flight(iscsilun->iscsi) >= MAX_NOP_FAILURES) {
1390 error_report("iSCSI: NOP timeout. Reconnecting...");
1391 iscsilun->request_timed_out = true;
1392 } else if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1393 error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1394 goto out;
1397 timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1398 iscsi_set_events(iscsilun);
1400 out:
1401 qemu_mutex_unlock(&iscsilun->mutex);
1404 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1406 struct scsi_task *task = NULL;
1407 struct scsi_readcapacity10 *rc10 = NULL;
1408 struct scsi_readcapacity16 *rc16 = NULL;
1409 int retries = ISCSI_CMD_RETRIES;
1411 do {
1412 if (task != NULL) {
1413 scsi_free_scsi_task(task);
1414 task = NULL;
1417 switch (iscsilun->type) {
1418 case TYPE_DISK:
1419 task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1420 if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1421 rc16 = scsi_datain_unmarshall(task);
1422 if (rc16 == NULL) {
1423 error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1424 } else {
1425 iscsilun->block_size = rc16->block_length;
1426 iscsilun->num_blocks = rc16->returned_lba + 1;
1427 iscsilun->lbpme = !!rc16->lbpme;
1428 iscsilun->lbprz = !!rc16->lbprz;
1429 iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1431 break;
1433 if (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1434 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
1435 break;
1437 /* Fall through and try READ CAPACITY(10) instead. */
1438 case TYPE_ROM:
1439 task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1440 if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1441 rc10 = scsi_datain_unmarshall(task);
1442 if (rc10 == NULL) {
1443 error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1444 } else {
1445 iscsilun->block_size = rc10->block_size;
1446 if (rc10->lba == 0) {
1447 /* blank disk loaded */
1448 iscsilun->num_blocks = 0;
1449 } else {
1450 iscsilun->num_blocks = rc10->lba + 1;
1454 break;
1455 default:
1456 return;
1458 } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1459 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1460 && retries-- > 0);
1462 if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1463 error_setg(errp, "iSCSI: failed to send readcapacity10/16 command");
1464 } else if (!iscsilun->block_size ||
1465 iscsilun->block_size % BDRV_SECTOR_SIZE) {
1466 error_setg(errp, "iSCSI: the target returned an invalid "
1467 "block size of %d.", iscsilun->block_size);
1469 if (task) {
1470 scsi_free_scsi_task(task);
1474 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1475 int evpd, int pc, void **inq, Error **errp)
1477 int full_size;
1478 struct scsi_task *task = NULL;
1479 task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1480 if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1481 goto fail;
1483 full_size = scsi_datain_getfullsize(task);
1484 if (full_size > task->datain.size) {
1485 scsi_free_scsi_task(task);
1487 /* we need more data for the full list */
1488 task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1489 if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1490 goto fail;
1494 *inq = scsi_datain_unmarshall(task);
1495 if (*inq == NULL) {
1496 error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1497 goto fail_with_err;
1500 return task;
1502 fail:
1503 error_setg(errp, "iSCSI: Inquiry command failed : %s",
1504 iscsi_get_error(iscsi));
1505 fail_with_err:
1506 if (task != NULL) {
1507 scsi_free_scsi_task(task);
1509 return NULL;
1512 static void iscsi_detach_aio_context(BlockDriverState *bs)
1514 IscsiLun *iscsilun = bs->opaque;
1516 aio_set_fd_handler(iscsilun->aio_context, iscsi_get_fd(iscsilun->iscsi),
1517 false, NULL, NULL, NULL, NULL);
1518 iscsilun->events = 0;
1520 if (iscsilun->nop_timer) {
1521 timer_del(iscsilun->nop_timer);
1522 timer_free(iscsilun->nop_timer);
1523 iscsilun->nop_timer = NULL;
1525 if (iscsilun->event_timer) {
1526 timer_del(iscsilun->event_timer);
1527 timer_free(iscsilun->event_timer);
1528 iscsilun->event_timer = NULL;
1532 static void iscsi_attach_aio_context(BlockDriverState *bs,
1533 AioContext *new_context)
1535 IscsiLun *iscsilun = bs->opaque;
1537 iscsilun->aio_context = new_context;
1538 iscsi_set_events(iscsilun);
1540 /* Set up a timer for sending out iSCSI NOPs */
1541 iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1542 QEMU_CLOCK_REALTIME, SCALE_MS,
1543 iscsi_nop_timed_event, iscsilun);
1544 timer_mod(iscsilun->nop_timer,
1545 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1547 /* Set up a timer for periodic calls to iscsi_set_events and to
1548 * scan for command timeout */
1549 iscsilun->event_timer = aio_timer_new(iscsilun->aio_context,
1550 QEMU_CLOCK_REALTIME, SCALE_MS,
1551 iscsi_timed_check_events, iscsilun);
1552 timer_mod(iscsilun->event_timer,
1553 qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + EVENT_INTERVAL);
1556 static void iscsi_modesense_sync(IscsiLun *iscsilun)
1558 struct scsi_task *task;
1559 struct scsi_mode_sense *ms = NULL;
1560 iscsilun->write_protected = false;
1561 iscsilun->dpofua = false;
1563 task = iscsi_modesense6_sync(iscsilun->iscsi, iscsilun->lun,
1564 1, SCSI_MODESENSE_PC_CURRENT,
1565 0x3F, 0, 255);
1566 if (task == NULL) {
1567 error_report("iSCSI: Failed to send MODE_SENSE(6) command: %s",
1568 iscsi_get_error(iscsilun->iscsi));
1569 goto out;
1572 if (task->status != SCSI_STATUS_GOOD) {
1573 error_report("iSCSI: Failed MODE_SENSE(6), LUN assumed writable");
1574 goto out;
1576 ms = scsi_datain_unmarshall(task);
1577 if (!ms) {
1578 error_report("iSCSI: Failed to unmarshall MODE_SENSE(6) data: %s",
1579 iscsi_get_error(iscsilun->iscsi));
1580 goto out;
1582 iscsilun->write_protected = ms->device_specific_parameter & 0x80;
1583 iscsilun->dpofua = ms->device_specific_parameter & 0x10;
1585 out:
1586 if (task) {
1587 scsi_free_scsi_task(task);
1591 static void iscsi_parse_iscsi_option(const char *target, QDict *options)
1593 QemuOptsList *list;
1594 QemuOpts *opts;
1595 const char *user, *password, *password_secret, *initiator_name,
1596 *header_digest, *timeout;
1598 list = qemu_find_opts("iscsi");
1599 if (!list) {
1600 return;
1603 opts = qemu_opts_find(list, target);
1604 if (opts == NULL) {
1605 opts = QTAILQ_FIRST(&list->head);
1606 if (!opts) {
1607 return;
1611 user = qemu_opt_get(opts, "user");
1612 if (user) {
1613 qdict_set_default_str(options, "user", user);
1616 password = qemu_opt_get(opts, "password");
1617 if (password) {
1618 qdict_set_default_str(options, "password", password);
1621 password_secret = qemu_opt_get(opts, "password-secret");
1622 if (password_secret) {
1623 qdict_set_default_str(options, "password-secret", password_secret);
1626 initiator_name = qemu_opt_get(opts, "initiator-name");
1627 if (initiator_name) {
1628 qdict_set_default_str(options, "initiator-name", initiator_name);
1631 header_digest = qemu_opt_get(opts, "header-digest");
1632 if (header_digest) {
1633 /* -iscsi takes upper case values, but QAPI only supports lower case
1634 * enum constant names, so we have to convert here. */
1635 char *qapi_value = g_ascii_strdown(header_digest, -1);
1636 qdict_set_default_str(options, "header-digest", qapi_value);
1637 g_free(qapi_value);
1640 timeout = qemu_opt_get(opts, "timeout");
1641 if (timeout) {
1642 qdict_set_default_str(options, "timeout", timeout);
1647 * We support iscsi url's on the form
1648 * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1650 static void iscsi_parse_filename(const char *filename, QDict *options,
1651 Error **errp)
1653 struct iscsi_url *iscsi_url;
1654 const char *transport_name;
1655 char *lun_str;
1657 iscsi_url = iscsi_parse_full_url(NULL, filename);
1658 if (iscsi_url == NULL) {
1659 error_setg(errp, "Failed to parse URL : %s", filename);
1660 return;
1663 #if LIBISCSI_API_VERSION >= (20160603)
1664 switch (iscsi_url->transport) {
1665 case TCP_TRANSPORT:
1666 transport_name = "tcp";
1667 break;
1668 case ISER_TRANSPORT:
1669 transport_name = "iser";
1670 break;
1671 default:
1672 error_setg(errp, "Unknown transport type (%d)",
1673 iscsi_url->transport);
1674 return;
1676 #else
1677 transport_name = "tcp";
1678 #endif
1680 qdict_set_default_str(options, "transport", transport_name);
1681 qdict_set_default_str(options, "portal", iscsi_url->portal);
1682 qdict_set_default_str(options, "target", iscsi_url->target);
1684 lun_str = g_strdup_printf("%d", iscsi_url->lun);
1685 qdict_set_default_str(options, "lun", lun_str);
1686 g_free(lun_str);
1688 /* User/password from -iscsi take precedence over those from the URL */
1689 iscsi_parse_iscsi_option(iscsi_url->target, options);
1691 if (iscsi_url->user[0] != '\0') {
1692 qdict_set_default_str(options, "user", iscsi_url->user);
1693 qdict_set_default_str(options, "password", iscsi_url->passwd);
1696 iscsi_destroy_url(iscsi_url);
1699 static QemuOptsList runtime_opts = {
1700 .name = "iscsi",
1701 .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1702 .desc = {
1704 .name = "transport",
1705 .type = QEMU_OPT_STRING,
1708 .name = "portal",
1709 .type = QEMU_OPT_STRING,
1712 .name = "target",
1713 .type = QEMU_OPT_STRING,
1716 .name = "user",
1717 .type = QEMU_OPT_STRING,
1720 .name = "password",
1721 .type = QEMU_OPT_STRING,
1724 .name = "password-secret",
1725 .type = QEMU_OPT_STRING,
1728 .name = "lun",
1729 .type = QEMU_OPT_NUMBER,
1732 .name = "initiator-name",
1733 .type = QEMU_OPT_STRING,
1736 .name = "header-digest",
1737 .type = QEMU_OPT_STRING,
1740 .name = "timeout",
1741 .type = QEMU_OPT_NUMBER,
1743 { /* end of list */ }
1747 static void iscsi_save_designator(IscsiLun *lun,
1748 struct scsi_inquiry_device_identification *inq_di)
1750 struct scsi_inquiry_device_designator *desig, *copy = NULL;
1752 for (desig = inq_di->designators; desig; desig = desig->next) {
1753 if (desig->association ||
1754 desig->designator_type > SCSI_DESIGNATOR_TYPE_NAA) {
1755 continue;
1757 /* NAA works better than T10 vendor ID based designator. */
1758 if (!copy || copy->designator_type < desig->designator_type) {
1759 copy = desig;
1762 if (copy) {
1763 lun->dd = g_new(struct scsi_inquiry_device_designator, 1);
1764 *lun->dd = *copy;
1765 lun->dd->next = NULL;
1766 lun->dd->designator = g_malloc(copy->designator_length);
1767 memcpy(lun->dd->designator, copy->designator, copy->designator_length);
1771 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1772 Error **errp)
1774 IscsiLun *iscsilun = bs->opaque;
1775 struct iscsi_context *iscsi = NULL;
1776 struct scsi_task *task = NULL;
1777 struct scsi_inquiry_standard *inq = NULL;
1778 struct scsi_inquiry_supported_pages *inq_vpd;
1779 char *initiator_name = NULL;
1780 QemuOpts *opts;
1781 Error *local_err = NULL;
1782 const char *transport_name, *portal, *target;
1783 #if LIBISCSI_API_VERSION >= (20160603)
1784 enum iscsi_transport_type transport;
1785 #endif
1786 int i, ret = 0, timeout = 0, lun;
1788 opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1789 qemu_opts_absorb_qdict(opts, options, &local_err);
1790 if (local_err) {
1791 error_propagate(errp, local_err);
1792 ret = -EINVAL;
1793 goto out;
1796 transport_name = qemu_opt_get(opts, "transport");
1797 portal = qemu_opt_get(opts, "portal");
1798 target = qemu_opt_get(opts, "target");
1799 lun = qemu_opt_get_number(opts, "lun", 0);
1801 if (!transport_name || !portal || !target) {
1802 error_setg(errp, "Need all of transport, portal and target options");
1803 ret = -EINVAL;
1804 goto out;
1807 if (!strcmp(transport_name, "tcp")) {
1808 #if LIBISCSI_API_VERSION >= (20160603)
1809 transport = TCP_TRANSPORT;
1810 } else if (!strcmp(transport_name, "iser")) {
1811 transport = ISER_TRANSPORT;
1812 #else
1813 /* TCP is what older libiscsi versions always use */
1814 #endif
1815 } else {
1816 error_setg(errp, "Unknown transport: %s", transport_name);
1817 ret = -EINVAL;
1818 goto out;
1821 memset(iscsilun, 0, sizeof(IscsiLun));
1823 initiator_name = get_initiator_name(opts);
1825 iscsi = iscsi_create_context(initiator_name);
1826 if (iscsi == NULL) {
1827 error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1828 ret = -ENOMEM;
1829 goto out;
1831 #if LIBISCSI_API_VERSION >= (20160603)
1832 if (iscsi_init_transport(iscsi, transport)) {
1833 error_setg(errp, ("Error initializing transport."));
1834 ret = -EINVAL;
1835 goto out;
1837 #endif
1838 if (iscsi_set_targetname(iscsi, target)) {
1839 error_setg(errp, "iSCSI: Failed to set target name.");
1840 ret = -EINVAL;
1841 goto out;
1844 /* check if we got CHAP username/password via the options */
1845 apply_chap(iscsi, opts, &local_err);
1846 if (local_err != NULL) {
1847 error_propagate(errp, local_err);
1848 ret = -EINVAL;
1849 goto out;
1852 if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1853 error_setg(errp, "iSCSI: Failed to set session type to normal.");
1854 ret = -EINVAL;
1855 goto out;
1858 /* check if we got HEADER_DIGEST via the options */
1859 apply_header_digest(iscsi, opts, &local_err);
1860 if (local_err != NULL) {
1861 error_propagate(errp, local_err);
1862 ret = -EINVAL;
1863 goto out;
1866 /* timeout handling is broken in libiscsi before 1.15.0 */
1867 timeout = qemu_opt_get_number(opts, "timeout", 0);
1868 #if LIBISCSI_API_VERSION >= 20150621
1869 iscsi_set_timeout(iscsi, timeout);
1870 #else
1871 if (timeout) {
1872 warn_report("iSCSI: ignoring timeout value for libiscsi <1.15.0");
1874 #endif
1876 if (iscsi_full_connect_sync(iscsi, portal, lun) != 0) {
1877 error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1878 iscsi_get_error(iscsi));
1879 ret = -EINVAL;
1880 goto out;
1883 iscsilun->iscsi = iscsi;
1884 iscsilun->aio_context = bdrv_get_aio_context(bs);
1885 iscsilun->lun = lun;
1886 iscsilun->has_write_same = true;
1888 task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1889 (void **) &inq, errp);
1890 if (task == NULL) {
1891 ret = -EINVAL;
1892 goto out;
1894 iscsilun->type = inq->periperal_device_type;
1895 scsi_free_scsi_task(task);
1896 task = NULL;
1898 iscsi_modesense_sync(iscsilun);
1899 if (iscsilun->dpofua) {
1900 bs->supported_write_flags = BDRV_REQ_FUA;
1903 /* Check the write protect flag of the LUN if we want to write */
1904 if (iscsilun->type == TYPE_DISK && (flags & BDRV_O_RDWR) &&
1905 iscsilun->write_protected) {
1906 ret = bdrv_apply_auto_read_only(bs, "LUN is write protected", errp);
1907 if (ret < 0) {
1908 goto out;
1910 flags &= ~BDRV_O_RDWR;
1913 iscsi_readcapacity_sync(iscsilun, &local_err);
1914 if (local_err != NULL) {
1915 error_propagate(errp, local_err);
1916 ret = -EINVAL;
1917 goto out;
1919 bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1921 /* We don't have any emulation for devices other than disks and CD-ROMs, so
1922 * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1923 * will try to read from the device to guess the image format.
1925 if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1926 bs->sg = true;
1929 task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1930 SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1931 (void **) &inq_vpd, errp);
1932 if (task == NULL) {
1933 ret = -EINVAL;
1934 goto out;
1936 for (i = 0; i < inq_vpd->num_pages; i++) {
1937 struct scsi_task *inq_task;
1938 struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1939 struct scsi_inquiry_block_limits *inq_bl;
1940 struct scsi_inquiry_device_identification *inq_di;
1941 switch (inq_vpd->pages[i]) {
1942 case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1943 inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1944 SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1945 (void **) &inq_lbp, errp);
1946 if (inq_task == NULL) {
1947 ret = -EINVAL;
1948 goto out;
1950 memcpy(&iscsilun->lbp, inq_lbp,
1951 sizeof(struct scsi_inquiry_logical_block_provisioning));
1952 scsi_free_scsi_task(inq_task);
1953 break;
1954 case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1955 inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1956 SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1957 (void **) &inq_bl, errp);
1958 if (inq_task == NULL) {
1959 ret = -EINVAL;
1960 goto out;
1962 memcpy(&iscsilun->bl, inq_bl,
1963 sizeof(struct scsi_inquiry_block_limits));
1964 scsi_free_scsi_task(inq_task);
1965 break;
1966 case SCSI_INQUIRY_PAGECODE_DEVICE_IDENTIFICATION:
1967 inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1968 SCSI_INQUIRY_PAGECODE_DEVICE_IDENTIFICATION,
1969 (void **) &inq_di, errp);
1970 if (inq_task == NULL) {
1971 ret = -EINVAL;
1972 goto out;
1974 iscsi_save_designator(iscsilun, inq_di);
1975 scsi_free_scsi_task(inq_task);
1976 break;
1977 default:
1978 break;
1981 scsi_free_scsi_task(task);
1982 task = NULL;
1984 qemu_mutex_init(&iscsilun->mutex);
1985 iscsi_attach_aio_context(bs, iscsilun->aio_context);
1987 /* Guess the internal cluster (page) size of the iscsi target by the means
1988 * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1989 * reasonable size */
1990 if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1991 iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1992 iscsilun->cluster_size = iscsilun->bl.opt_unmap_gran *
1993 iscsilun->block_size;
1994 if (iscsilun->lbprz) {
1995 ret = iscsi_allocmap_init(iscsilun, bs->open_flags);
1999 if (iscsilun->lbprz && iscsilun->lbp.lbpws) {
2000 bs->supported_zero_flags = BDRV_REQ_MAY_UNMAP;
2003 out:
2004 qemu_opts_del(opts);
2005 g_free(initiator_name);
2006 if (task != NULL) {
2007 scsi_free_scsi_task(task);
2010 if (ret) {
2011 if (iscsi != NULL) {
2012 if (iscsi_is_logged_in(iscsi)) {
2013 iscsi_logout_sync(iscsi);
2015 iscsi_destroy_context(iscsi);
2017 memset(iscsilun, 0, sizeof(IscsiLun));
2020 return ret;
2023 static void iscsi_close(BlockDriverState *bs)
2025 IscsiLun *iscsilun = bs->opaque;
2026 struct iscsi_context *iscsi = iscsilun->iscsi;
2028 iscsi_detach_aio_context(bs);
2029 if (iscsi_is_logged_in(iscsi)) {
2030 iscsi_logout_sync(iscsi);
2032 iscsi_destroy_context(iscsi);
2033 if (iscsilun->dd) {
2034 g_free(iscsilun->dd->designator);
2035 g_free(iscsilun->dd);
2037 g_free(iscsilun->zeroblock);
2038 iscsi_allocmap_free(iscsilun);
2039 qemu_mutex_destroy(&iscsilun->mutex);
2040 memset(iscsilun, 0, sizeof(IscsiLun));
2043 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
2045 /* We don't actually refresh here, but just return data queried in
2046 * iscsi_open(): iscsi targets don't change their limits. */
2048 IscsiLun *iscsilun = bs->opaque;
2049 uint64_t max_xfer_len = iscsilun->use_16_for_rw ? 0xffffffff : 0xffff;
2050 unsigned int block_size = MAX(BDRV_SECTOR_SIZE, iscsilun->block_size);
2052 assert(iscsilun->block_size >= BDRV_SECTOR_SIZE || bs->sg);
2054 bs->bl.request_alignment = block_size;
2056 if (iscsilun->bl.max_xfer_len) {
2057 max_xfer_len = MIN(max_xfer_len, iscsilun->bl.max_xfer_len);
2060 if (max_xfer_len * block_size < INT_MAX) {
2061 bs->bl.max_transfer = max_xfer_len * iscsilun->block_size;
2064 if (iscsilun->lbp.lbpu) {
2065 if (iscsilun->bl.max_unmap < 0xffffffff / block_size) {
2066 bs->bl.max_pdiscard =
2067 iscsilun->bl.max_unmap * iscsilun->block_size;
2069 bs->bl.pdiscard_alignment =
2070 iscsilun->bl.opt_unmap_gran * iscsilun->block_size;
2071 } else {
2072 bs->bl.pdiscard_alignment = iscsilun->block_size;
2075 if (iscsilun->bl.max_ws_len < 0xffffffff / block_size) {
2076 bs->bl.max_pwrite_zeroes =
2077 iscsilun->bl.max_ws_len * iscsilun->block_size;
2079 if (iscsilun->lbp.lbpws) {
2080 bs->bl.pwrite_zeroes_alignment =
2081 iscsilun->bl.opt_unmap_gran * iscsilun->block_size;
2082 } else {
2083 bs->bl.pwrite_zeroes_alignment = iscsilun->block_size;
2085 if (iscsilun->bl.opt_xfer_len &&
2086 iscsilun->bl.opt_xfer_len < INT_MAX / block_size) {
2087 bs->bl.opt_transfer = pow2floor(iscsilun->bl.opt_xfer_len *
2088 iscsilun->block_size);
2092 /* Note that this will not re-establish a connection with an iSCSI target - it
2093 * is effectively a NOP. */
2094 static int iscsi_reopen_prepare(BDRVReopenState *state,
2095 BlockReopenQueue *queue, Error **errp)
2097 IscsiLun *iscsilun = state->bs->opaque;
2099 if (state->flags & BDRV_O_RDWR && iscsilun->write_protected) {
2100 error_setg(errp, "Cannot open a write protected LUN as read-write");
2101 return -EACCES;
2103 return 0;
2106 static void iscsi_reopen_commit(BDRVReopenState *reopen_state)
2108 IscsiLun *iscsilun = reopen_state->bs->opaque;
2110 /* the cache.direct status might have changed */
2111 if (iscsilun->allocmap != NULL) {
2112 iscsi_allocmap_init(iscsilun, reopen_state->flags);
2116 static int coroutine_fn iscsi_co_truncate(BlockDriverState *bs, int64_t offset,
2117 PreallocMode prealloc, Error **errp)
2119 IscsiLun *iscsilun = bs->opaque;
2120 Error *local_err = NULL;
2122 if (prealloc != PREALLOC_MODE_OFF) {
2123 error_setg(errp, "Unsupported preallocation mode '%s'",
2124 PreallocMode_str(prealloc));
2125 return -ENOTSUP;
2128 if (iscsilun->type != TYPE_DISK) {
2129 error_setg(errp, "Cannot resize non-disk iSCSI devices");
2130 return -ENOTSUP;
2133 iscsi_readcapacity_sync(iscsilun, &local_err);
2134 if (local_err != NULL) {
2135 error_propagate(errp, local_err);
2136 return -EIO;
2139 if (offset > iscsi_getlength(bs)) {
2140 error_setg(errp, "Cannot grow iSCSI devices");
2141 return -EINVAL;
2144 if (iscsilun->allocmap != NULL) {
2145 iscsi_allocmap_init(iscsilun, bs->open_flags);
2148 return 0;
2151 static int coroutine_fn iscsi_co_create_opts(const char *filename, QemuOpts *opts,
2152 Error **errp)
2154 int ret = 0;
2155 int64_t total_size = 0;
2156 BlockDriverState *bs;
2157 IscsiLun *iscsilun = NULL;
2158 QDict *bs_options;
2159 Error *local_err = NULL;
2161 bs = bdrv_new();
2163 /* Read out options */
2164 total_size = DIV_ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
2165 BDRV_SECTOR_SIZE);
2166 bs->opaque = g_new0(struct IscsiLun, 1);
2167 iscsilun = bs->opaque;
2169 bs_options = qdict_new();
2170 iscsi_parse_filename(filename, bs_options, &local_err);
2171 if (local_err) {
2172 error_propagate(errp, local_err);
2173 ret = -EINVAL;
2174 } else {
2175 ret = iscsi_open(bs, bs_options, 0, NULL);
2177 qobject_unref(bs_options);
2179 if (ret != 0) {
2180 goto out;
2182 iscsi_detach_aio_context(bs);
2183 if (iscsilun->type != TYPE_DISK) {
2184 ret = -ENODEV;
2185 goto out;
2187 if (bs->total_sectors < total_size) {
2188 ret = -ENOSPC;
2189 goto out;
2192 ret = 0;
2193 out:
2194 if (iscsilun->iscsi != NULL) {
2195 iscsi_destroy_context(iscsilun->iscsi);
2197 g_free(bs->opaque);
2198 bs->opaque = NULL;
2199 bdrv_unref(bs);
2200 return ret;
2203 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
2205 IscsiLun *iscsilun = bs->opaque;
2206 bdi->unallocated_blocks_are_zero = iscsilun->lbprz;
2207 bdi->cluster_size = iscsilun->cluster_size;
2208 return 0;
2211 static void coroutine_fn iscsi_co_invalidate_cache(BlockDriverState *bs,
2212 Error **errp)
2214 IscsiLun *iscsilun = bs->opaque;
2215 iscsi_allocmap_invalidate(iscsilun);
2218 static int coroutine_fn iscsi_co_copy_range_from(BlockDriverState *bs,
2219 BdrvChild *src,
2220 uint64_t src_offset,
2221 BdrvChild *dst,
2222 uint64_t dst_offset,
2223 uint64_t bytes,
2224 BdrvRequestFlags read_flags,
2225 BdrvRequestFlags write_flags)
2227 return bdrv_co_copy_range_to(src, src_offset, dst, dst_offset, bytes,
2228 read_flags, write_flags);
2231 static struct scsi_task *iscsi_xcopy_task(int param_len)
2233 struct scsi_task *task;
2235 task = g_new0(struct scsi_task, 1);
2237 task->cdb[0] = EXTENDED_COPY;
2238 task->cdb[10] = (param_len >> 24) & 0xFF;
2239 task->cdb[11] = (param_len >> 16) & 0xFF;
2240 task->cdb[12] = (param_len >> 8) & 0xFF;
2241 task->cdb[13] = param_len & 0xFF;
2242 task->cdb_size = 16;
2243 task->xfer_dir = SCSI_XFER_WRITE;
2244 task->expxferlen = param_len;
2246 return task;
2249 static void iscsi_populate_target_desc(unsigned char *desc, IscsiLun *lun)
2251 struct scsi_inquiry_device_designator *dd = lun->dd;
2253 memset(desc, 0, 32);
2254 desc[0] = 0xE4; /* IDENT_DESCR_TGT_DESCR */
2255 desc[4] = dd->code_set;
2256 desc[5] = (dd->designator_type & 0xF)
2257 | ((dd->association & 3) << 4);
2258 desc[7] = dd->designator_length;
2259 memcpy(desc + 8, dd->designator, MIN(dd->designator_length, 20));
2261 desc[28] = 0;
2262 desc[29] = (lun->block_size >> 16) & 0xFF;
2263 desc[30] = (lun->block_size >> 8) & 0xFF;
2264 desc[31] = lun->block_size & 0xFF;
2267 static void iscsi_xcopy_desc_hdr(uint8_t *hdr, int dc, int cat, int src_index,
2268 int dst_index)
2270 hdr[0] = 0x02; /* BLK_TO_BLK_SEG_DESCR */
2271 hdr[1] = ((dc << 1) | cat) & 0xFF;
2272 hdr[2] = (XCOPY_BLK2BLK_SEG_DESC_SIZE >> 8) & 0xFF;
2273 /* don't account for the first 4 bytes in descriptor header*/
2274 hdr[3] = (XCOPY_BLK2BLK_SEG_DESC_SIZE - 4 /* SEG_DESC_SRC_INDEX_OFFSET */) & 0xFF;
2275 hdr[4] = (src_index >> 8) & 0xFF;
2276 hdr[5] = src_index & 0xFF;
2277 hdr[6] = (dst_index >> 8) & 0xFF;
2278 hdr[7] = dst_index & 0xFF;
2281 static void iscsi_xcopy_populate_desc(uint8_t *desc, int dc, int cat,
2282 int src_index, int dst_index, int num_blks,
2283 uint64_t src_lba, uint64_t dst_lba)
2285 iscsi_xcopy_desc_hdr(desc, dc, cat, src_index, dst_index);
2287 /* The caller should verify the request size */
2288 assert(num_blks < 65536);
2289 desc[10] = (num_blks >> 8) & 0xFF;
2290 desc[11] = num_blks & 0xFF;
2291 desc[12] = (src_lba >> 56) & 0xFF;
2292 desc[13] = (src_lba >> 48) & 0xFF;
2293 desc[14] = (src_lba >> 40) & 0xFF;
2294 desc[15] = (src_lba >> 32) & 0xFF;
2295 desc[16] = (src_lba >> 24) & 0xFF;
2296 desc[17] = (src_lba >> 16) & 0xFF;
2297 desc[18] = (src_lba >> 8) & 0xFF;
2298 desc[19] = src_lba & 0xFF;
2299 desc[20] = (dst_lba >> 56) & 0xFF;
2300 desc[21] = (dst_lba >> 48) & 0xFF;
2301 desc[22] = (dst_lba >> 40) & 0xFF;
2302 desc[23] = (dst_lba >> 32) & 0xFF;
2303 desc[24] = (dst_lba >> 24) & 0xFF;
2304 desc[25] = (dst_lba >> 16) & 0xFF;
2305 desc[26] = (dst_lba >> 8) & 0xFF;
2306 desc[27] = dst_lba & 0xFF;
2309 static void iscsi_xcopy_populate_header(unsigned char *buf, int list_id, int str,
2310 int list_id_usage, int prio,
2311 int tgt_desc_len,
2312 int seg_desc_len, int inline_data_len)
2314 buf[0] = list_id;
2315 buf[1] = ((str & 1) << 5) | ((list_id_usage & 3) << 3) | (prio & 7);
2316 buf[2] = (tgt_desc_len >> 8) & 0xFF;
2317 buf[3] = tgt_desc_len & 0xFF;
2318 buf[8] = (seg_desc_len >> 24) & 0xFF;
2319 buf[9] = (seg_desc_len >> 16) & 0xFF;
2320 buf[10] = (seg_desc_len >> 8) & 0xFF;
2321 buf[11] = seg_desc_len & 0xFF;
2322 buf[12] = (inline_data_len >> 24) & 0xFF;
2323 buf[13] = (inline_data_len >> 16) & 0xFF;
2324 buf[14] = (inline_data_len >> 8) & 0xFF;
2325 buf[15] = inline_data_len & 0xFF;
2328 static void iscsi_xcopy_data(struct iscsi_data *data,
2329 IscsiLun *src, int64_t src_lba,
2330 IscsiLun *dst, int64_t dst_lba,
2331 uint16_t num_blocks)
2333 uint8_t *buf;
2334 const int src_offset = XCOPY_DESC_OFFSET;
2335 const int dst_offset = XCOPY_DESC_OFFSET + IDENT_DESCR_TGT_DESCR_SIZE;
2336 const int seg_offset = dst_offset + IDENT_DESCR_TGT_DESCR_SIZE;
2338 data->size = XCOPY_DESC_OFFSET +
2339 IDENT_DESCR_TGT_DESCR_SIZE * 2 +
2340 XCOPY_BLK2BLK_SEG_DESC_SIZE;
2341 data->data = g_malloc0(data->size);
2342 buf = data->data;
2344 /* Initialise the parameter list header */
2345 iscsi_xcopy_populate_header(buf, 1, 0, 2 /* LIST_ID_USAGE_DISCARD */,
2346 0, 2 * IDENT_DESCR_TGT_DESCR_SIZE,
2347 XCOPY_BLK2BLK_SEG_DESC_SIZE,
2350 /* Initialise CSCD list with one src + one dst descriptor */
2351 iscsi_populate_target_desc(&buf[src_offset], src);
2352 iscsi_populate_target_desc(&buf[dst_offset], dst);
2354 /* Initialise one segment descriptor */
2355 iscsi_xcopy_populate_desc(&buf[seg_offset], 0, 0, 0, 1, num_blocks,
2356 src_lba, dst_lba);
2359 static int coroutine_fn iscsi_co_copy_range_to(BlockDriverState *bs,
2360 BdrvChild *src,
2361 uint64_t src_offset,
2362 BdrvChild *dst,
2363 uint64_t dst_offset,
2364 uint64_t bytes,
2365 BdrvRequestFlags read_flags,
2366 BdrvRequestFlags write_flags)
2368 IscsiLun *dst_lun = dst->bs->opaque;
2369 IscsiLun *src_lun;
2370 struct IscsiTask iscsi_task;
2371 struct iscsi_data data;
2372 int r = 0;
2373 int block_size;
2375 if (src->bs->drv->bdrv_co_copy_range_to != iscsi_co_copy_range_to) {
2376 return -ENOTSUP;
2378 src_lun = src->bs->opaque;
2380 if (!src_lun->dd || !dst_lun->dd) {
2381 return -ENOTSUP;
2383 if (!is_byte_request_lun_aligned(dst_offset, bytes, dst_lun)) {
2384 return -ENOTSUP;
2386 if (!is_byte_request_lun_aligned(src_offset, bytes, src_lun)) {
2387 return -ENOTSUP;
2389 if (dst_lun->block_size != src_lun->block_size ||
2390 !dst_lun->block_size) {
2391 return -ENOTSUP;
2394 block_size = dst_lun->block_size;
2395 if (bytes / block_size > 65535) {
2396 return -ENOTSUP;
2399 iscsi_xcopy_data(&data,
2400 src_lun, src_offset / block_size,
2401 dst_lun, dst_offset / block_size,
2402 bytes / block_size);
2404 iscsi_co_init_iscsitask(dst_lun, &iscsi_task);
2406 qemu_mutex_lock(&dst_lun->mutex);
2407 iscsi_task.task = iscsi_xcopy_task(data.size);
2408 retry:
2409 if (iscsi_scsi_command_async(dst_lun->iscsi, dst_lun->lun,
2410 iscsi_task.task, iscsi_co_generic_cb,
2411 &data,
2412 &iscsi_task) != 0) {
2413 r = -EIO;
2414 goto out_unlock;
2417 iscsi_co_wait_for_task(&iscsi_task, dst_lun);
2419 if (iscsi_task.do_retry) {
2420 iscsi_task.complete = 0;
2421 goto retry;
2424 if (iscsi_task.status != SCSI_STATUS_GOOD) {
2425 r = iscsi_task.err_code;
2426 goto out_unlock;
2429 out_unlock:
2431 trace_iscsi_xcopy(src_lun, src_offset, dst_lun, dst_offset, bytes, r);
2432 g_free(iscsi_task.task);
2433 qemu_mutex_unlock(&dst_lun->mutex);
2434 g_free(iscsi_task.err_str);
2435 return r;
2438 static QemuOptsList iscsi_create_opts = {
2439 .name = "iscsi-create-opts",
2440 .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
2441 .desc = {
2443 .name = BLOCK_OPT_SIZE,
2444 .type = QEMU_OPT_SIZE,
2445 .help = "Virtual disk size"
2447 { /* end of list */ }
2451 static BlockDriver bdrv_iscsi = {
2452 .format_name = "iscsi",
2453 .protocol_name = "iscsi",
2455 .instance_size = sizeof(IscsiLun),
2456 .bdrv_parse_filename = iscsi_parse_filename,
2457 .bdrv_file_open = iscsi_open,
2458 .bdrv_close = iscsi_close,
2459 .bdrv_co_create_opts = iscsi_co_create_opts,
2460 .create_opts = &iscsi_create_opts,
2461 .bdrv_reopen_prepare = iscsi_reopen_prepare,
2462 .bdrv_reopen_commit = iscsi_reopen_commit,
2463 .bdrv_co_invalidate_cache = iscsi_co_invalidate_cache,
2465 .bdrv_getlength = iscsi_getlength,
2466 .bdrv_get_info = iscsi_get_info,
2467 .bdrv_co_truncate = iscsi_co_truncate,
2468 .bdrv_refresh_limits = iscsi_refresh_limits,
2470 .bdrv_co_block_status = iscsi_co_block_status,
2471 .bdrv_co_pdiscard = iscsi_co_pdiscard,
2472 .bdrv_co_copy_range_from = iscsi_co_copy_range_from,
2473 .bdrv_co_copy_range_to = iscsi_co_copy_range_to,
2474 .bdrv_co_pwrite_zeroes = iscsi_co_pwrite_zeroes,
2475 .bdrv_co_readv = iscsi_co_readv,
2476 .bdrv_co_writev = iscsi_co_writev,
2477 .bdrv_co_flush_to_disk = iscsi_co_flush,
2479 #ifdef __linux__
2480 .bdrv_aio_ioctl = iscsi_aio_ioctl,
2481 #endif
2483 .bdrv_detach_aio_context = iscsi_detach_aio_context,
2484 .bdrv_attach_aio_context = iscsi_attach_aio_context,
2487 #if LIBISCSI_API_VERSION >= (20160603)
2488 static BlockDriver bdrv_iser = {
2489 .format_name = "iser",
2490 .protocol_name = "iser",
2492 .instance_size = sizeof(IscsiLun),
2493 .bdrv_parse_filename = iscsi_parse_filename,
2494 .bdrv_file_open = iscsi_open,
2495 .bdrv_close = iscsi_close,
2496 .bdrv_co_create_opts = iscsi_co_create_opts,
2497 .create_opts = &iscsi_create_opts,
2498 .bdrv_reopen_prepare = iscsi_reopen_prepare,
2499 .bdrv_reopen_commit = iscsi_reopen_commit,
2500 .bdrv_co_invalidate_cache = iscsi_co_invalidate_cache,
2502 .bdrv_getlength = iscsi_getlength,
2503 .bdrv_get_info = iscsi_get_info,
2504 .bdrv_co_truncate = iscsi_co_truncate,
2505 .bdrv_refresh_limits = iscsi_refresh_limits,
2507 .bdrv_co_block_status = iscsi_co_block_status,
2508 .bdrv_co_pdiscard = iscsi_co_pdiscard,
2509 .bdrv_co_copy_range_from = iscsi_co_copy_range_from,
2510 .bdrv_co_copy_range_to = iscsi_co_copy_range_to,
2511 .bdrv_co_pwrite_zeroes = iscsi_co_pwrite_zeroes,
2512 .bdrv_co_readv = iscsi_co_readv,
2513 .bdrv_co_writev = iscsi_co_writev,
2514 .bdrv_co_flush_to_disk = iscsi_co_flush,
2516 #ifdef __linux__
2517 .bdrv_aio_ioctl = iscsi_aio_ioctl,
2518 #endif
2520 .bdrv_detach_aio_context = iscsi_detach_aio_context,
2521 .bdrv_attach_aio_context = iscsi_attach_aio_context,
2523 #endif
2525 static void iscsi_block_init(void)
2527 bdrv_register(&bdrv_iscsi);
2528 #if LIBISCSI_API_VERSION >= (20160603)
2529 bdrv_register(&bdrv_iser);
2530 #endif
2533 block_init(iscsi_block_init);