xref: /openbmc/qemu/block/nvme.c (revision c63ca4ff)
1 /*
2  * NVMe block driver based on vfio
3  *
4  * Copyright 2016 - 2018 Red Hat, Inc.
5  *
6  * Authors:
7  *   Fam Zheng <famz@redhat.com>
8  *   Paolo Bonzini <pbonzini@redhat.com>
9  *
10  * This work is licensed under the terms of the GNU GPL, version 2 or later.
11  * See the COPYING file in the top-level directory.
12  */
13 
14 #include "qemu/osdep.h"
15 #include <linux/vfio.h>
16 #include "qapi/error.h"
17 #include "qapi/qmp/qdict.h"
18 #include "qapi/qmp/qstring.h"
19 #include "qemu/error-report.h"
20 #include "qemu/main-loop.h"
21 #include "qemu/module.h"
22 #include "qemu/cutils.h"
23 #include "qemu/option.h"
24 #include "qemu/vfio-helpers.h"
25 #include "block/block_int.h"
26 #include "sysemu/replay.h"
27 #include "trace.h"
28 
29 #include "block/nvme.h"
30 
31 #define NVME_SQ_ENTRY_BYTES 64
32 #define NVME_CQ_ENTRY_BYTES 16
33 #define NVME_QUEUE_SIZE 128
34 #define NVME_DOORBELL_SIZE 4096
35 
36 /*
37  * We have to leave one slot empty as that is the full queue case where
38  * head == tail + 1.
39  */
40 #define NVME_NUM_REQS (NVME_QUEUE_SIZE - 1)
41 
42 typedef struct BDRVNVMeState BDRVNVMeState;
43 
44 /* Same index is used for queues and IRQs */
45 #define INDEX_ADMIN     0
46 #define INDEX_IO(n)     (1 + n)
47 
48 /* This driver shares a single MSIX IRQ for the admin and I/O queues */
49 enum {
50     MSIX_SHARED_IRQ_IDX = 0,
51     MSIX_IRQ_COUNT = 1
52 };
53 
54 typedef struct {
55     int32_t  head, tail;
56     uint8_t  *queue;
57     uint64_t iova;
58     /* Hardware MMIO register */
59     volatile uint32_t *doorbell;
60 } NVMeQueue;
61 
62 typedef struct {
63     BlockCompletionFunc *cb;
64     void *opaque;
65     int cid;
66     void *prp_list_page;
67     uint64_t prp_list_iova;
68     int free_req_next; /* q->reqs[] index of next free req */
69 } NVMeRequest;
70 
71 typedef struct {
72     QemuMutex   lock;
73 
74     /* Read from I/O code path, initialized under BQL */
75     BDRVNVMeState   *s;
76     int             index;
77 
78     /* Fields protected by BQL */
79     uint8_t     *prp_list_pages;
80 
81     /* Fields protected by @lock */
82     CoQueue     free_req_queue;
83     NVMeQueue   sq, cq;
84     int         cq_phase;
85     int         free_req_head;
86     NVMeRequest reqs[NVME_NUM_REQS];
87     int         need_kick;
88     int         inflight;
89 
90     /* Thread-safe, no lock necessary */
91     QEMUBH      *completion_bh;
92 } NVMeQueuePair;
93 
94 struct BDRVNVMeState {
95     AioContext *aio_context;
96     QEMUVFIOState *vfio;
97     void *bar0_wo_map;
98     /* Memory mapped registers */
99     volatile struct {
100         uint32_t sq_tail;
101         uint32_t cq_head;
102     } *doorbells;
103     /* The submission/completion queue pairs.
104      * [0]: admin queue.
105      * [1..]: io queues.
106      */
107     NVMeQueuePair **queues;
108     unsigned queue_count;
109     size_t page_size;
110     /* How many uint32_t elements does each doorbell entry take. */
111     size_t doorbell_scale;
112     bool write_cache_supported;
113     EventNotifier irq_notifier[MSIX_IRQ_COUNT];
114 
115     uint64_t nsze; /* Namespace size reported by identify command */
116     int nsid;      /* The namespace id to read/write data. */
117     int blkshift;
118 
119     uint64_t max_transfer;
120     bool plugged;
121 
122     bool supports_write_zeroes;
123     bool supports_discard;
124 
125     CoMutex dma_map_lock;
126     CoQueue dma_flush_queue;
127 
128     /* Total size of mapped qiov, accessed under dma_map_lock */
129     int dma_map_count;
130 
131     /* PCI address (required for nvme_refresh_filename()) */
132     char *device;
133 
134     struct {
135         uint64_t completion_errors;
136         uint64_t aligned_accesses;
137         uint64_t unaligned_accesses;
138     } stats;
139 };
140 
141 #define NVME_BLOCK_OPT_DEVICE "device"
142 #define NVME_BLOCK_OPT_NAMESPACE "namespace"
143 
144 static void nvme_process_completion_bh(void *opaque);
145 
146 static QemuOptsList runtime_opts = {
147     .name = "nvme",
148     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
149     .desc = {
150         {
151             .name = NVME_BLOCK_OPT_DEVICE,
152             .type = QEMU_OPT_STRING,
153             .help = "NVMe PCI device address",
154         },
155         {
156             .name = NVME_BLOCK_OPT_NAMESPACE,
157             .type = QEMU_OPT_NUMBER,
158             .help = "NVMe namespace",
159         },
160         { /* end of list */ }
161     },
162 };
163 
164 /* Returns true on success, false on failure. */
165 static bool nvme_init_queue(BDRVNVMeState *s, NVMeQueue *q,
166                             unsigned nentries, size_t entry_bytes, Error **errp)
167 {
168     size_t bytes;
169     int r;
170 
171     bytes = ROUND_UP(nentries * entry_bytes, qemu_real_host_page_size);
172     q->head = q->tail = 0;
173     q->queue = qemu_try_memalign(qemu_real_host_page_size, bytes);
174     if (!q->queue) {
175         error_setg(errp, "Cannot allocate queue");
176         return false;
177     }
178     memset(q->queue, 0, bytes);
179     r = qemu_vfio_dma_map(s->vfio, q->queue, bytes, false, &q->iova);
180     if (r) {
181         error_setg(errp, "Cannot map queue");
182         return false;
183     }
184     return true;
185 }
186 
187 static void nvme_free_queue_pair(NVMeQueuePair *q)
188 {
189     trace_nvme_free_queue_pair(q->index, q);
190     if (q->completion_bh) {
191         qemu_bh_delete(q->completion_bh);
192     }
193     qemu_vfree(q->prp_list_pages);
194     qemu_vfree(q->sq.queue);
195     qemu_vfree(q->cq.queue);
196     qemu_mutex_destroy(&q->lock);
197     g_free(q);
198 }
199 
200 static void nvme_free_req_queue_cb(void *opaque)
201 {
202     NVMeQueuePair *q = opaque;
203 
204     qemu_mutex_lock(&q->lock);
205     while (qemu_co_enter_next(&q->free_req_queue, &q->lock)) {
206         /* Retry all pending requests */
207     }
208     qemu_mutex_unlock(&q->lock);
209 }
210 
211 static NVMeQueuePair *nvme_create_queue_pair(BDRVNVMeState *s,
212                                              AioContext *aio_context,
213                                              unsigned idx, size_t size,
214                                              Error **errp)
215 {
216     int i, r;
217     NVMeQueuePair *q;
218     uint64_t prp_list_iova;
219     size_t bytes;
220 
221     q = g_try_new0(NVMeQueuePair, 1);
222     if (!q) {
223         return NULL;
224     }
225     trace_nvme_create_queue_pair(idx, q, size, aio_context,
226                                  event_notifier_get_fd(s->irq_notifier));
227     bytes = QEMU_ALIGN_UP(s->page_size * NVME_NUM_REQS,
228                           qemu_real_host_page_size);
229     q->prp_list_pages = qemu_try_memalign(qemu_real_host_page_size, bytes);
230     if (!q->prp_list_pages) {
231         goto fail;
232     }
233     memset(q->prp_list_pages, 0, bytes);
234     qemu_mutex_init(&q->lock);
235     q->s = s;
236     q->index = idx;
237     qemu_co_queue_init(&q->free_req_queue);
238     q->completion_bh = aio_bh_new(aio_context, nvme_process_completion_bh, q);
239     r = qemu_vfio_dma_map(s->vfio, q->prp_list_pages, bytes,
240                           false, &prp_list_iova);
241     if (r) {
242         goto fail;
243     }
244     q->free_req_head = -1;
245     for (i = 0; i < NVME_NUM_REQS; i++) {
246         NVMeRequest *req = &q->reqs[i];
247         req->cid = i + 1;
248         req->free_req_next = q->free_req_head;
249         q->free_req_head = i;
250         req->prp_list_page = q->prp_list_pages + i * s->page_size;
251         req->prp_list_iova = prp_list_iova + i * s->page_size;
252     }
253 
254     if (!nvme_init_queue(s, &q->sq, size, NVME_SQ_ENTRY_BYTES, errp)) {
255         goto fail;
256     }
257     q->sq.doorbell = &s->doorbells[idx * s->doorbell_scale].sq_tail;
258 
259     if (!nvme_init_queue(s, &q->cq, size, NVME_CQ_ENTRY_BYTES, errp)) {
260         goto fail;
261     }
262     q->cq.doorbell = &s->doorbells[idx * s->doorbell_scale].cq_head;
263 
264     return q;
265 fail:
266     nvme_free_queue_pair(q);
267     return NULL;
268 }
269 
270 /* With q->lock */
271 static void nvme_kick(NVMeQueuePair *q)
272 {
273     BDRVNVMeState *s = q->s;
274 
275     if (s->plugged || !q->need_kick) {
276         return;
277     }
278     trace_nvme_kick(s, q->index);
279     assert(!(q->sq.tail & 0xFF00));
280     /* Fence the write to submission queue entry before notifying the device. */
281     smp_wmb();
282     *q->sq.doorbell = cpu_to_le32(q->sq.tail);
283     q->inflight += q->need_kick;
284     q->need_kick = 0;
285 }
286 
287 /* Find a free request element if any, otherwise:
288  * a) if in coroutine context, try to wait for one to become available;
289  * b) if not in coroutine, return NULL;
290  */
291 static NVMeRequest *nvme_get_free_req(NVMeQueuePair *q)
292 {
293     NVMeRequest *req;
294 
295     qemu_mutex_lock(&q->lock);
296 
297     while (q->free_req_head == -1) {
298         if (qemu_in_coroutine()) {
299             trace_nvme_free_req_queue_wait(q->s, q->index);
300             qemu_co_queue_wait(&q->free_req_queue, &q->lock);
301         } else {
302             qemu_mutex_unlock(&q->lock);
303             return NULL;
304         }
305     }
306 
307     req = &q->reqs[q->free_req_head];
308     q->free_req_head = req->free_req_next;
309     req->free_req_next = -1;
310 
311     qemu_mutex_unlock(&q->lock);
312     return req;
313 }
314 
315 /* With q->lock */
316 static void nvme_put_free_req_locked(NVMeQueuePair *q, NVMeRequest *req)
317 {
318     req->free_req_next = q->free_req_head;
319     q->free_req_head = req - q->reqs;
320 }
321 
322 /* With q->lock */
323 static void nvme_wake_free_req_locked(NVMeQueuePair *q)
324 {
325     if (!qemu_co_queue_empty(&q->free_req_queue)) {
326         replay_bh_schedule_oneshot_event(q->s->aio_context,
327                 nvme_free_req_queue_cb, q);
328     }
329 }
330 
331 /* Insert a request in the freelist and wake waiters */
332 static void nvme_put_free_req_and_wake(NVMeQueuePair *q, NVMeRequest *req)
333 {
334     qemu_mutex_lock(&q->lock);
335     nvme_put_free_req_locked(q, req);
336     nvme_wake_free_req_locked(q);
337     qemu_mutex_unlock(&q->lock);
338 }
339 
340 static inline int nvme_translate_error(const NvmeCqe *c)
341 {
342     uint16_t status = (le16_to_cpu(c->status) >> 1) & 0xFF;
343     if (status) {
344         trace_nvme_error(le32_to_cpu(c->result),
345                          le16_to_cpu(c->sq_head),
346                          le16_to_cpu(c->sq_id),
347                          le16_to_cpu(c->cid),
348                          le16_to_cpu(status));
349     }
350     switch (status) {
351     case 0:
352         return 0;
353     case 1:
354         return -ENOSYS;
355     case 2:
356         return -EINVAL;
357     default:
358         return -EIO;
359     }
360 }
361 
362 /* With q->lock */
363 static bool nvme_process_completion(NVMeQueuePair *q)
364 {
365     BDRVNVMeState *s = q->s;
366     bool progress = false;
367     NVMeRequest *preq;
368     NVMeRequest req;
369     NvmeCqe *c;
370 
371     trace_nvme_process_completion(s, q->index, q->inflight);
372     if (s->plugged) {
373         trace_nvme_process_completion_queue_plugged(s, q->index);
374         return false;
375     }
376 
377     /*
378      * Support re-entrancy when a request cb() function invokes aio_poll().
379      * Pending completions must be visible to aio_poll() so that a cb()
380      * function can wait for the completion of another request.
381      *
382      * The aio_poll() loop will execute our BH and we'll resume completion
383      * processing there.
384      */
385     qemu_bh_schedule(q->completion_bh);
386 
387     assert(q->inflight >= 0);
388     while (q->inflight) {
389         int ret;
390         int16_t cid;
391 
392         c = (NvmeCqe *)&q->cq.queue[q->cq.head * NVME_CQ_ENTRY_BYTES];
393         if ((le16_to_cpu(c->status) & 0x1) == q->cq_phase) {
394             break;
395         }
396         ret = nvme_translate_error(c);
397         if (ret) {
398             s->stats.completion_errors++;
399         }
400         q->cq.head = (q->cq.head + 1) % NVME_QUEUE_SIZE;
401         if (!q->cq.head) {
402             q->cq_phase = !q->cq_phase;
403         }
404         cid = le16_to_cpu(c->cid);
405         if (cid == 0 || cid > NVME_QUEUE_SIZE) {
406             warn_report("NVMe: Unexpected CID in completion queue: %"PRIu32", "
407                         "queue size: %u", cid, NVME_QUEUE_SIZE);
408             continue;
409         }
410         trace_nvme_complete_command(s, q->index, cid);
411         preq = &q->reqs[cid - 1];
412         req = *preq;
413         assert(req.cid == cid);
414         assert(req.cb);
415         nvme_put_free_req_locked(q, preq);
416         preq->cb = preq->opaque = NULL;
417         q->inflight--;
418         qemu_mutex_unlock(&q->lock);
419         req.cb(req.opaque, ret);
420         qemu_mutex_lock(&q->lock);
421         progress = true;
422     }
423     if (progress) {
424         /* Notify the device so it can post more completions. */
425         smp_mb_release();
426         *q->cq.doorbell = cpu_to_le32(q->cq.head);
427         nvme_wake_free_req_locked(q);
428     }
429 
430     qemu_bh_cancel(q->completion_bh);
431 
432     return progress;
433 }
434 
435 static void nvme_process_completion_bh(void *opaque)
436 {
437     NVMeQueuePair *q = opaque;
438 
439     /*
440      * We're being invoked because a nvme_process_completion() cb() function
441      * called aio_poll(). The callback may be waiting for further completions
442      * so notify the device that it has space to fill in more completions now.
443      */
444     smp_mb_release();
445     *q->cq.doorbell = cpu_to_le32(q->cq.head);
446     nvme_wake_free_req_locked(q);
447 
448     nvme_process_completion(q);
449 }
450 
451 static void nvme_trace_command(const NvmeCmd *cmd)
452 {
453     int i;
454 
455     if (!trace_event_get_state_backends(TRACE_NVME_SUBMIT_COMMAND_RAW)) {
456         return;
457     }
458     for (i = 0; i < 8; ++i) {
459         uint8_t *cmdp = (uint8_t *)cmd + i * 8;
460         trace_nvme_submit_command_raw(cmdp[0], cmdp[1], cmdp[2], cmdp[3],
461                                       cmdp[4], cmdp[5], cmdp[6], cmdp[7]);
462     }
463 }
464 
465 static void nvme_submit_command(NVMeQueuePair *q, NVMeRequest *req,
466                                 NvmeCmd *cmd, BlockCompletionFunc cb,
467                                 void *opaque)
468 {
469     assert(!req->cb);
470     req->cb = cb;
471     req->opaque = opaque;
472     cmd->cid = cpu_to_le16(req->cid);
473 
474     trace_nvme_submit_command(q->s, q->index, req->cid);
475     nvme_trace_command(cmd);
476     qemu_mutex_lock(&q->lock);
477     memcpy((uint8_t *)q->sq.queue +
478            q->sq.tail * NVME_SQ_ENTRY_BYTES, cmd, sizeof(*cmd));
479     q->sq.tail = (q->sq.tail + 1) % NVME_QUEUE_SIZE;
480     q->need_kick++;
481     nvme_kick(q);
482     nvme_process_completion(q);
483     qemu_mutex_unlock(&q->lock);
484 }
485 
486 static void nvme_admin_cmd_sync_cb(void *opaque, int ret)
487 {
488     int *pret = opaque;
489     *pret = ret;
490     aio_wait_kick();
491 }
492 
493 static int nvme_admin_cmd_sync(BlockDriverState *bs, NvmeCmd *cmd)
494 {
495     BDRVNVMeState *s = bs->opaque;
496     NVMeQueuePair *q = s->queues[INDEX_ADMIN];
497     AioContext *aio_context = bdrv_get_aio_context(bs);
498     NVMeRequest *req;
499     int ret = -EINPROGRESS;
500     req = nvme_get_free_req(q);
501     if (!req) {
502         return -EBUSY;
503     }
504     nvme_submit_command(q, req, cmd, nvme_admin_cmd_sync_cb, &ret);
505 
506     AIO_WAIT_WHILE(aio_context, ret == -EINPROGRESS);
507     return ret;
508 }
509 
510 /* Returns true on success, false on failure. */
511 static bool nvme_identify(BlockDriverState *bs, int namespace, Error **errp)
512 {
513     BDRVNVMeState *s = bs->opaque;
514     bool ret = false;
515     union {
516         NvmeIdCtrl ctrl;
517         NvmeIdNs ns;
518     } *id;
519     NvmeLBAF *lbaf;
520     uint16_t oncs;
521     int r;
522     uint64_t iova;
523     NvmeCmd cmd = {
524         .opcode = NVME_ADM_CMD_IDENTIFY,
525         .cdw10 = cpu_to_le32(0x1),
526     };
527     size_t id_size = QEMU_ALIGN_UP(sizeof(*id), qemu_real_host_page_size);
528 
529     id = qemu_try_memalign(qemu_real_host_page_size, id_size);
530     if (!id) {
531         error_setg(errp, "Cannot allocate buffer for identify response");
532         goto out;
533     }
534     r = qemu_vfio_dma_map(s->vfio, id, id_size, true, &iova);
535     if (r) {
536         error_setg(errp, "Cannot map buffer for DMA");
537         goto out;
538     }
539 
540     memset(id, 0, id_size);
541     cmd.dptr.prp1 = cpu_to_le64(iova);
542     if (nvme_admin_cmd_sync(bs, &cmd)) {
543         error_setg(errp, "Failed to identify controller");
544         goto out;
545     }
546 
547     if (le32_to_cpu(id->ctrl.nn) < namespace) {
548         error_setg(errp, "Invalid namespace");
549         goto out;
550     }
551     s->write_cache_supported = le32_to_cpu(id->ctrl.vwc) & 0x1;
552     s->max_transfer = (id->ctrl.mdts ? 1 << id->ctrl.mdts : 0) * s->page_size;
553     /* For now the page list buffer per command is one page, to hold at most
554      * s->page_size / sizeof(uint64_t) entries. */
555     s->max_transfer = MIN_NON_ZERO(s->max_transfer,
556                           s->page_size / sizeof(uint64_t) * s->page_size);
557 
558     oncs = le16_to_cpu(id->ctrl.oncs);
559     s->supports_write_zeroes = !!(oncs & NVME_ONCS_WRITE_ZEROES);
560     s->supports_discard = !!(oncs & NVME_ONCS_DSM);
561 
562     memset(id, 0, id_size);
563     cmd.cdw10 = 0;
564     cmd.nsid = cpu_to_le32(namespace);
565     if (nvme_admin_cmd_sync(bs, &cmd)) {
566         error_setg(errp, "Failed to identify namespace");
567         goto out;
568     }
569 
570     s->nsze = le64_to_cpu(id->ns.nsze);
571     lbaf = &id->ns.lbaf[NVME_ID_NS_FLBAS_INDEX(id->ns.flbas)];
572 
573     if (NVME_ID_NS_DLFEAT_WRITE_ZEROES(id->ns.dlfeat) &&
574             NVME_ID_NS_DLFEAT_READ_BEHAVIOR(id->ns.dlfeat) ==
575                     NVME_ID_NS_DLFEAT_READ_BEHAVIOR_ZEROES) {
576         bs->supported_write_flags |= BDRV_REQ_MAY_UNMAP;
577     }
578 
579     if (lbaf->ms) {
580         error_setg(errp, "Namespaces with metadata are not yet supported");
581         goto out;
582     }
583 
584     if (lbaf->ds < BDRV_SECTOR_BITS || lbaf->ds > 12 ||
585         (1 << lbaf->ds) > s->page_size)
586     {
587         error_setg(errp, "Namespace has unsupported block size (2^%d)",
588                    lbaf->ds);
589         goto out;
590     }
591 
592     ret = true;
593     s->blkshift = lbaf->ds;
594 out:
595     qemu_vfio_dma_unmap(s->vfio, id);
596     qemu_vfree(id);
597 
598     return ret;
599 }
600 
601 static bool nvme_poll_queue(NVMeQueuePair *q)
602 {
603     bool progress = false;
604 
605     const size_t cqe_offset = q->cq.head * NVME_CQ_ENTRY_BYTES;
606     NvmeCqe *cqe = (NvmeCqe *)&q->cq.queue[cqe_offset];
607 
608     trace_nvme_poll_queue(q->s, q->index);
609     /*
610      * Do an early check for completions. q->lock isn't needed because
611      * nvme_process_completion() only runs in the event loop thread and
612      * cannot race with itself.
613      */
614     if ((le16_to_cpu(cqe->status) & 0x1) == q->cq_phase) {
615         return false;
616     }
617 
618     qemu_mutex_lock(&q->lock);
619     while (nvme_process_completion(q)) {
620         /* Keep polling */
621         progress = true;
622     }
623     qemu_mutex_unlock(&q->lock);
624 
625     return progress;
626 }
627 
628 static bool nvme_poll_queues(BDRVNVMeState *s)
629 {
630     bool progress = false;
631     int i;
632 
633     for (i = 0; i < s->queue_count; i++) {
634         if (nvme_poll_queue(s->queues[i])) {
635             progress = true;
636         }
637     }
638     return progress;
639 }
640 
641 static void nvme_handle_event(EventNotifier *n)
642 {
643     BDRVNVMeState *s = container_of(n, BDRVNVMeState,
644                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
645 
646     trace_nvme_handle_event(s);
647     event_notifier_test_and_clear(n);
648     nvme_poll_queues(s);
649 }
650 
651 static bool nvme_add_io_queue(BlockDriverState *bs, Error **errp)
652 {
653     BDRVNVMeState *s = bs->opaque;
654     unsigned n = s->queue_count;
655     NVMeQueuePair *q;
656     NvmeCmd cmd;
657     unsigned queue_size = NVME_QUEUE_SIZE;
658 
659     assert(n <= UINT16_MAX);
660     q = nvme_create_queue_pair(s, bdrv_get_aio_context(bs),
661                                n, queue_size, errp);
662     if (!q) {
663         return false;
664     }
665     cmd = (NvmeCmd) {
666         .opcode = NVME_ADM_CMD_CREATE_CQ,
667         .dptr.prp1 = cpu_to_le64(q->cq.iova),
668         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | n),
669         .cdw11 = cpu_to_le32(NVME_CQ_IEN | NVME_CQ_PC),
670     };
671     if (nvme_admin_cmd_sync(bs, &cmd)) {
672         error_setg(errp, "Failed to create CQ io queue [%u]", n);
673         goto out_error;
674     }
675     cmd = (NvmeCmd) {
676         .opcode = NVME_ADM_CMD_CREATE_SQ,
677         .dptr.prp1 = cpu_to_le64(q->sq.iova),
678         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | n),
679         .cdw11 = cpu_to_le32(NVME_SQ_PC | (n << 16)),
680     };
681     if (nvme_admin_cmd_sync(bs, &cmd)) {
682         error_setg(errp, "Failed to create SQ io queue [%u]", n);
683         goto out_error;
684     }
685     s->queues = g_renew(NVMeQueuePair *, s->queues, n + 1);
686     s->queues[n] = q;
687     s->queue_count++;
688     return true;
689 out_error:
690     nvme_free_queue_pair(q);
691     return false;
692 }
693 
694 static bool nvme_poll_cb(void *opaque)
695 {
696     EventNotifier *e = opaque;
697     BDRVNVMeState *s = container_of(e, BDRVNVMeState,
698                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
699 
700     return nvme_poll_queues(s);
701 }
702 
703 static int nvme_init(BlockDriverState *bs, const char *device, int namespace,
704                      Error **errp)
705 {
706     BDRVNVMeState *s = bs->opaque;
707     NVMeQueuePair *q;
708     AioContext *aio_context = bdrv_get_aio_context(bs);
709     int ret;
710     uint64_t cap;
711     uint64_t timeout_ms;
712     uint64_t deadline, now;
713     volatile NvmeBar *regs = NULL;
714 
715     qemu_co_mutex_init(&s->dma_map_lock);
716     qemu_co_queue_init(&s->dma_flush_queue);
717     s->device = g_strdup(device);
718     s->nsid = namespace;
719     s->aio_context = bdrv_get_aio_context(bs);
720     ret = event_notifier_init(&s->irq_notifier[MSIX_SHARED_IRQ_IDX], 0);
721     if (ret) {
722         error_setg(errp, "Failed to init event notifier");
723         return ret;
724     }
725 
726     s->vfio = qemu_vfio_open_pci(device, errp);
727     if (!s->vfio) {
728         ret = -EINVAL;
729         goto out;
730     }
731 
732     regs = qemu_vfio_pci_map_bar(s->vfio, 0, 0, sizeof(NvmeBar),
733                                  PROT_READ | PROT_WRITE, errp);
734     if (!regs) {
735         ret = -EINVAL;
736         goto out;
737     }
738     /* Perform initialize sequence as described in NVMe spec "7.6.1
739      * Initialization". */
740 
741     cap = le64_to_cpu(regs->cap);
742     trace_nvme_controller_capability_raw(cap);
743     trace_nvme_controller_capability("Maximum Queue Entries Supported",
744                                      1 + NVME_CAP_MQES(cap));
745     trace_nvme_controller_capability("Contiguous Queues Required",
746                                      NVME_CAP_CQR(cap));
747     trace_nvme_controller_capability("Doorbell Stride",
748                                      2 << (2 + NVME_CAP_DSTRD(cap)));
749     trace_nvme_controller_capability("Subsystem Reset Supported",
750                                      NVME_CAP_NSSRS(cap));
751     trace_nvme_controller_capability("Memory Page Size Minimum",
752                                      1 << (12 + NVME_CAP_MPSMIN(cap)));
753     trace_nvme_controller_capability("Memory Page Size Maximum",
754                                      1 << (12 + NVME_CAP_MPSMAX(cap)));
755     if (!NVME_CAP_CSS(cap)) {
756         error_setg(errp, "Device doesn't support NVMe command set");
757         ret = -EINVAL;
758         goto out;
759     }
760 
761     s->page_size = 1u << (12 + NVME_CAP_MPSMIN(cap));
762     s->doorbell_scale = (4 << NVME_CAP_DSTRD(cap)) / sizeof(uint32_t);
763     bs->bl.opt_mem_alignment = s->page_size;
764     bs->bl.request_alignment = s->page_size;
765     timeout_ms = MIN(500 * NVME_CAP_TO(cap), 30000);
766 
767     /* Reset device to get a clean state. */
768     regs->cc = cpu_to_le32(le32_to_cpu(regs->cc) & 0xFE);
769     /* Wait for CSTS.RDY = 0. */
770     deadline = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) + timeout_ms * SCALE_MS;
771     while (NVME_CSTS_RDY(le32_to_cpu(regs->csts))) {
772         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
773             error_setg(errp, "Timeout while waiting for device to reset (%"
774                              PRId64 " ms)",
775                        timeout_ms);
776             ret = -ETIMEDOUT;
777             goto out;
778         }
779     }
780 
781     s->bar0_wo_map = qemu_vfio_pci_map_bar(s->vfio, 0, 0,
782                                            sizeof(NvmeBar) + NVME_DOORBELL_SIZE,
783                                            PROT_WRITE, errp);
784     s->doorbells = (void *)((uintptr_t)s->bar0_wo_map + sizeof(NvmeBar));
785     if (!s->doorbells) {
786         ret = -EINVAL;
787         goto out;
788     }
789 
790     /* Set up admin queue. */
791     s->queues = g_new(NVMeQueuePair *, 1);
792     q = nvme_create_queue_pair(s, aio_context, 0, NVME_QUEUE_SIZE, errp);
793     if (!q) {
794         ret = -EINVAL;
795         goto out;
796     }
797     s->queues[INDEX_ADMIN] = q;
798     s->queue_count = 1;
799     QEMU_BUILD_BUG_ON((NVME_QUEUE_SIZE - 1) & 0xF000);
800     regs->aqa = cpu_to_le32(((NVME_QUEUE_SIZE - 1) << AQA_ACQS_SHIFT) |
801                             ((NVME_QUEUE_SIZE - 1) << AQA_ASQS_SHIFT));
802     regs->asq = cpu_to_le64(q->sq.iova);
803     regs->acq = cpu_to_le64(q->cq.iova);
804 
805     /* After setting up all control registers we can enable device now. */
806     regs->cc = cpu_to_le32((ctz32(NVME_CQ_ENTRY_BYTES) << CC_IOCQES_SHIFT) |
807                            (ctz32(NVME_SQ_ENTRY_BYTES) << CC_IOSQES_SHIFT) |
808                            CC_EN_MASK);
809     /* Wait for CSTS.RDY = 1. */
810     now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
811     deadline = now + timeout_ms * SCALE_MS;
812     while (!NVME_CSTS_RDY(le32_to_cpu(regs->csts))) {
813         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
814             error_setg(errp, "Timeout while waiting for device to start (%"
815                              PRId64 " ms)",
816                        timeout_ms);
817             ret = -ETIMEDOUT;
818             goto out;
819         }
820     }
821 
822     ret = qemu_vfio_pci_init_irq(s->vfio, s->irq_notifier,
823                                  VFIO_PCI_MSIX_IRQ_INDEX, errp);
824     if (ret) {
825         goto out;
826     }
827     aio_set_event_notifier(bdrv_get_aio_context(bs),
828                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
829                            false, nvme_handle_event, nvme_poll_cb);
830 
831     if (!nvme_identify(bs, namespace, errp)) {
832         ret = -EIO;
833         goto out;
834     }
835 
836     /* Set up command queues. */
837     if (!nvme_add_io_queue(bs, errp)) {
838         ret = -EIO;
839     }
840 out:
841     if (regs) {
842         qemu_vfio_pci_unmap_bar(s->vfio, 0, (void *)regs, 0, sizeof(NvmeBar));
843     }
844 
845     /* Cleaning up is done in nvme_file_open() upon error. */
846     return ret;
847 }
848 
849 /* Parse a filename in the format of nvme://XXXX:XX:XX.X/X. Example:
850  *
851  *     nvme://0000:44:00.0/1
852  *
853  * where the "nvme://" is a fixed form of the protocol prefix, the middle part
854  * is the PCI address, and the last part is the namespace number starting from
855  * 1 according to the NVMe spec. */
856 static void nvme_parse_filename(const char *filename, QDict *options,
857                                 Error **errp)
858 {
859     int pref = strlen("nvme://");
860 
861     if (strlen(filename) > pref && !strncmp(filename, "nvme://", pref)) {
862         const char *tmp = filename + pref;
863         char *device;
864         const char *namespace;
865         unsigned long ns;
866         const char *slash = strchr(tmp, '/');
867         if (!slash) {
868             qdict_put_str(options, NVME_BLOCK_OPT_DEVICE, tmp);
869             return;
870         }
871         device = g_strndup(tmp, slash - tmp);
872         qdict_put_str(options, NVME_BLOCK_OPT_DEVICE, device);
873         g_free(device);
874         namespace = slash + 1;
875         if (*namespace && qemu_strtoul(namespace, NULL, 10, &ns)) {
876             error_setg(errp, "Invalid namespace '%s', positive number expected",
877                        namespace);
878             return;
879         }
880         qdict_put_str(options, NVME_BLOCK_OPT_NAMESPACE,
881                       *namespace ? namespace : "1");
882     }
883 }
884 
885 static int nvme_enable_disable_write_cache(BlockDriverState *bs, bool enable,
886                                            Error **errp)
887 {
888     int ret;
889     BDRVNVMeState *s = bs->opaque;
890     NvmeCmd cmd = {
891         .opcode = NVME_ADM_CMD_SET_FEATURES,
892         .nsid = cpu_to_le32(s->nsid),
893         .cdw10 = cpu_to_le32(0x06),
894         .cdw11 = cpu_to_le32(enable ? 0x01 : 0x00),
895     };
896 
897     ret = nvme_admin_cmd_sync(bs, &cmd);
898     if (ret) {
899         error_setg(errp, "Failed to configure NVMe write cache");
900     }
901     return ret;
902 }
903 
904 static void nvme_close(BlockDriverState *bs)
905 {
906     BDRVNVMeState *s = bs->opaque;
907 
908     for (unsigned i = 0; i < s->queue_count; ++i) {
909         nvme_free_queue_pair(s->queues[i]);
910     }
911     g_free(s->queues);
912     aio_set_event_notifier(bdrv_get_aio_context(bs),
913                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
914                            false, NULL, NULL);
915     event_notifier_cleanup(&s->irq_notifier[MSIX_SHARED_IRQ_IDX]);
916     qemu_vfio_pci_unmap_bar(s->vfio, 0, s->bar0_wo_map,
917                             0, sizeof(NvmeBar) + NVME_DOORBELL_SIZE);
918     qemu_vfio_close(s->vfio);
919 
920     g_free(s->device);
921 }
922 
923 static int nvme_file_open(BlockDriverState *bs, QDict *options, int flags,
924                           Error **errp)
925 {
926     const char *device;
927     QemuOpts *opts;
928     int namespace;
929     int ret;
930     BDRVNVMeState *s = bs->opaque;
931 
932     bs->supported_write_flags = BDRV_REQ_FUA;
933 
934     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
935     qemu_opts_absorb_qdict(opts, options, &error_abort);
936     device = qemu_opt_get(opts, NVME_BLOCK_OPT_DEVICE);
937     if (!device) {
938         error_setg(errp, "'" NVME_BLOCK_OPT_DEVICE "' option is required");
939         qemu_opts_del(opts);
940         return -EINVAL;
941     }
942 
943     namespace = qemu_opt_get_number(opts, NVME_BLOCK_OPT_NAMESPACE, 1);
944     ret = nvme_init(bs, device, namespace, errp);
945     qemu_opts_del(opts);
946     if (ret) {
947         goto fail;
948     }
949     if (flags & BDRV_O_NOCACHE) {
950         if (!s->write_cache_supported) {
951             error_setg(errp,
952                        "NVMe controller doesn't support write cache configuration");
953             ret = -EINVAL;
954         } else {
955             ret = nvme_enable_disable_write_cache(bs, !(flags & BDRV_O_NOCACHE),
956                                                   errp);
957         }
958         if (ret) {
959             goto fail;
960         }
961     }
962     return 0;
963 fail:
964     nvme_close(bs);
965     return ret;
966 }
967 
968 static int64_t nvme_getlength(BlockDriverState *bs)
969 {
970     BDRVNVMeState *s = bs->opaque;
971     return s->nsze << s->blkshift;
972 }
973 
974 static uint32_t nvme_get_blocksize(BlockDriverState *bs)
975 {
976     BDRVNVMeState *s = bs->opaque;
977     assert(s->blkshift >= BDRV_SECTOR_BITS && s->blkshift <= 12);
978     return UINT32_C(1) << s->blkshift;
979 }
980 
981 static int nvme_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
982 {
983     uint32_t blocksize = nvme_get_blocksize(bs);
984     bsz->phys = blocksize;
985     bsz->log = blocksize;
986     return 0;
987 }
988 
989 /* Called with s->dma_map_lock */
990 static coroutine_fn int nvme_cmd_unmap_qiov(BlockDriverState *bs,
991                                             QEMUIOVector *qiov)
992 {
993     int r = 0;
994     BDRVNVMeState *s = bs->opaque;
995 
996     s->dma_map_count -= qiov->size;
997     if (!s->dma_map_count && !qemu_co_queue_empty(&s->dma_flush_queue)) {
998         r = qemu_vfio_dma_reset_temporary(s->vfio);
999         if (!r) {
1000             qemu_co_queue_restart_all(&s->dma_flush_queue);
1001         }
1002     }
1003     return r;
1004 }
1005 
1006 /* Called with s->dma_map_lock */
1007 static coroutine_fn int nvme_cmd_map_qiov(BlockDriverState *bs, NvmeCmd *cmd,
1008                                           NVMeRequest *req, QEMUIOVector *qiov)
1009 {
1010     BDRVNVMeState *s = bs->opaque;
1011     uint64_t *pagelist = req->prp_list_page;
1012     int i, j, r;
1013     int entries = 0;
1014 
1015     assert(qiov->size);
1016     assert(QEMU_IS_ALIGNED(qiov->size, s->page_size));
1017     assert(qiov->size / s->page_size <= s->page_size / sizeof(uint64_t));
1018     for (i = 0; i < qiov->niov; ++i) {
1019         bool retry = true;
1020         uint64_t iova;
1021         size_t len = QEMU_ALIGN_UP(qiov->iov[i].iov_len,
1022                                    qemu_real_host_page_size);
1023 try_map:
1024         r = qemu_vfio_dma_map(s->vfio,
1025                               qiov->iov[i].iov_base,
1026                               len, true, &iova);
1027         if (r == -ENOMEM && retry) {
1028             retry = false;
1029             trace_nvme_dma_flush_queue_wait(s);
1030             if (s->dma_map_count) {
1031                 trace_nvme_dma_map_flush(s);
1032                 qemu_co_queue_wait(&s->dma_flush_queue, &s->dma_map_lock);
1033             } else {
1034                 r = qemu_vfio_dma_reset_temporary(s->vfio);
1035                 if (r) {
1036                     goto fail;
1037                 }
1038             }
1039             goto try_map;
1040         }
1041         if (r) {
1042             goto fail;
1043         }
1044 
1045         for (j = 0; j < qiov->iov[i].iov_len / s->page_size; j++) {
1046             pagelist[entries++] = cpu_to_le64(iova + j * s->page_size);
1047         }
1048         trace_nvme_cmd_map_qiov_iov(s, i, qiov->iov[i].iov_base,
1049                                     qiov->iov[i].iov_len / s->page_size);
1050     }
1051 
1052     s->dma_map_count += qiov->size;
1053 
1054     assert(entries <= s->page_size / sizeof(uint64_t));
1055     switch (entries) {
1056     case 0:
1057         abort();
1058     case 1:
1059         cmd->dptr.prp1 = pagelist[0];
1060         cmd->dptr.prp2 = 0;
1061         break;
1062     case 2:
1063         cmd->dptr.prp1 = pagelist[0];
1064         cmd->dptr.prp2 = pagelist[1];
1065         break;
1066     default:
1067         cmd->dptr.prp1 = pagelist[0];
1068         cmd->dptr.prp2 = cpu_to_le64(req->prp_list_iova + sizeof(uint64_t));
1069         break;
1070     }
1071     trace_nvme_cmd_map_qiov(s, cmd, req, qiov, entries);
1072     for (i = 0; i < entries; ++i) {
1073         trace_nvme_cmd_map_qiov_pages(s, i, pagelist[i]);
1074     }
1075     return 0;
1076 fail:
1077     /* No need to unmap [0 - i) iovs even if we've failed, since we don't
1078      * increment s->dma_map_count. This is okay for fixed mapping memory areas
1079      * because they are already mapped before calling this function; for
1080      * temporary mappings, a later nvme_cmd_(un)map_qiov will reclaim by
1081      * calling qemu_vfio_dma_reset_temporary when necessary. */
1082     return r;
1083 }
1084 
1085 typedef struct {
1086     Coroutine *co;
1087     int ret;
1088     AioContext *ctx;
1089 } NVMeCoData;
1090 
1091 static void nvme_rw_cb_bh(void *opaque)
1092 {
1093     NVMeCoData *data = opaque;
1094     qemu_coroutine_enter(data->co);
1095 }
1096 
1097 static void nvme_rw_cb(void *opaque, int ret)
1098 {
1099     NVMeCoData *data = opaque;
1100     data->ret = ret;
1101     if (!data->co) {
1102         /* The rw coroutine hasn't yielded, don't try to enter. */
1103         return;
1104     }
1105     replay_bh_schedule_oneshot_event(data->ctx, nvme_rw_cb_bh, data);
1106 }
1107 
1108 static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
1109                                             uint64_t offset, uint64_t bytes,
1110                                             QEMUIOVector *qiov,
1111                                             bool is_write,
1112                                             int flags)
1113 {
1114     int r;
1115     BDRVNVMeState *s = bs->opaque;
1116     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1117     NVMeRequest *req;
1118 
1119     uint32_t cdw12 = (((bytes >> s->blkshift) - 1) & 0xFFFF) |
1120                        (flags & BDRV_REQ_FUA ? 1 << 30 : 0);
1121     NvmeCmd cmd = {
1122         .opcode = is_write ? NVME_CMD_WRITE : NVME_CMD_READ,
1123         .nsid = cpu_to_le32(s->nsid),
1124         .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
1125         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
1126         .cdw12 = cpu_to_le32(cdw12),
1127     };
1128     NVMeCoData data = {
1129         .ctx = bdrv_get_aio_context(bs),
1130         .ret = -EINPROGRESS,
1131     };
1132 
1133     trace_nvme_prw_aligned(s, is_write, offset, bytes, flags, qiov->niov);
1134     assert(s->queue_count > 1);
1135     req = nvme_get_free_req(ioq);
1136     assert(req);
1137 
1138     qemu_co_mutex_lock(&s->dma_map_lock);
1139     r = nvme_cmd_map_qiov(bs, &cmd, req, qiov);
1140     qemu_co_mutex_unlock(&s->dma_map_lock);
1141     if (r) {
1142         nvme_put_free_req_and_wake(ioq, req);
1143         return r;
1144     }
1145     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1146 
1147     data.co = qemu_coroutine_self();
1148     while (data.ret == -EINPROGRESS) {
1149         qemu_coroutine_yield();
1150     }
1151 
1152     qemu_co_mutex_lock(&s->dma_map_lock);
1153     r = nvme_cmd_unmap_qiov(bs, qiov);
1154     qemu_co_mutex_unlock(&s->dma_map_lock);
1155     if (r) {
1156         return r;
1157     }
1158 
1159     trace_nvme_rw_done(s, is_write, offset, bytes, data.ret);
1160     return data.ret;
1161 }
1162 
1163 static inline bool nvme_qiov_aligned(BlockDriverState *bs,
1164                                      const QEMUIOVector *qiov)
1165 {
1166     int i;
1167     BDRVNVMeState *s = bs->opaque;
1168 
1169     for (i = 0; i < qiov->niov; ++i) {
1170         if (!QEMU_PTR_IS_ALIGNED(qiov->iov[i].iov_base,
1171                                  qemu_real_host_page_size) ||
1172             !QEMU_IS_ALIGNED(qiov->iov[i].iov_len, qemu_real_host_page_size)) {
1173             trace_nvme_qiov_unaligned(qiov, i, qiov->iov[i].iov_base,
1174                                       qiov->iov[i].iov_len, s->page_size);
1175             return false;
1176         }
1177     }
1178     return true;
1179 }
1180 
1181 static int nvme_co_prw(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
1182                        QEMUIOVector *qiov, bool is_write, int flags)
1183 {
1184     BDRVNVMeState *s = bs->opaque;
1185     int r;
1186     uint8_t *buf = NULL;
1187     QEMUIOVector local_qiov;
1188     size_t len = QEMU_ALIGN_UP(bytes, qemu_real_host_page_size);
1189     assert(QEMU_IS_ALIGNED(offset, s->page_size));
1190     assert(QEMU_IS_ALIGNED(bytes, s->page_size));
1191     assert(bytes <= s->max_transfer);
1192     if (nvme_qiov_aligned(bs, qiov)) {
1193         s->stats.aligned_accesses++;
1194         return nvme_co_prw_aligned(bs, offset, bytes, qiov, is_write, flags);
1195     }
1196     s->stats.unaligned_accesses++;
1197     trace_nvme_prw_buffered(s, offset, bytes, qiov->niov, is_write);
1198     buf = qemu_try_memalign(qemu_real_host_page_size, len);
1199 
1200     if (!buf) {
1201         return -ENOMEM;
1202     }
1203     qemu_iovec_init(&local_qiov, 1);
1204     if (is_write) {
1205         qemu_iovec_to_buf(qiov, 0, buf, bytes);
1206     }
1207     qemu_iovec_add(&local_qiov, buf, bytes);
1208     r = nvme_co_prw_aligned(bs, offset, bytes, &local_qiov, is_write, flags);
1209     qemu_iovec_destroy(&local_qiov);
1210     if (!r && !is_write) {
1211         qemu_iovec_from_buf(qiov, 0, buf, bytes);
1212     }
1213     qemu_vfree(buf);
1214     return r;
1215 }
1216 
1217 static coroutine_fn int nvme_co_preadv(BlockDriverState *bs,
1218                                        uint64_t offset, uint64_t bytes,
1219                                        QEMUIOVector *qiov, int flags)
1220 {
1221     return nvme_co_prw(bs, offset, bytes, qiov, false, flags);
1222 }
1223 
1224 static coroutine_fn int nvme_co_pwritev(BlockDriverState *bs,
1225                                         uint64_t offset, uint64_t bytes,
1226                                         QEMUIOVector *qiov, int flags)
1227 {
1228     return nvme_co_prw(bs, offset, bytes, qiov, true, flags);
1229 }
1230 
1231 static coroutine_fn int nvme_co_flush(BlockDriverState *bs)
1232 {
1233     BDRVNVMeState *s = bs->opaque;
1234     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1235     NVMeRequest *req;
1236     NvmeCmd cmd = {
1237         .opcode = NVME_CMD_FLUSH,
1238         .nsid = cpu_to_le32(s->nsid),
1239     };
1240     NVMeCoData data = {
1241         .ctx = bdrv_get_aio_context(bs),
1242         .ret = -EINPROGRESS,
1243     };
1244 
1245     assert(s->queue_count > 1);
1246     req = nvme_get_free_req(ioq);
1247     assert(req);
1248     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1249 
1250     data.co = qemu_coroutine_self();
1251     if (data.ret == -EINPROGRESS) {
1252         qemu_coroutine_yield();
1253     }
1254 
1255     return data.ret;
1256 }
1257 
1258 
1259 static coroutine_fn int nvme_co_pwrite_zeroes(BlockDriverState *bs,
1260                                               int64_t offset,
1261                                               int bytes,
1262                                               BdrvRequestFlags flags)
1263 {
1264     BDRVNVMeState *s = bs->opaque;
1265     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1266     NVMeRequest *req;
1267 
1268     uint32_t cdw12 = ((bytes >> s->blkshift) - 1) & 0xFFFF;
1269 
1270     if (!s->supports_write_zeroes) {
1271         return -ENOTSUP;
1272     }
1273 
1274     NvmeCmd cmd = {
1275         .opcode = NVME_CMD_WRITE_ZEROES,
1276         .nsid = cpu_to_le32(s->nsid),
1277         .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
1278         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
1279     };
1280 
1281     NVMeCoData data = {
1282         .ctx = bdrv_get_aio_context(bs),
1283         .ret = -EINPROGRESS,
1284     };
1285 
1286     if (flags & BDRV_REQ_MAY_UNMAP) {
1287         cdw12 |= (1 << 25);
1288     }
1289 
1290     if (flags & BDRV_REQ_FUA) {
1291         cdw12 |= (1 << 30);
1292     }
1293 
1294     cmd.cdw12 = cpu_to_le32(cdw12);
1295 
1296     trace_nvme_write_zeroes(s, offset, bytes, flags);
1297     assert(s->queue_count > 1);
1298     req = nvme_get_free_req(ioq);
1299     assert(req);
1300 
1301     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1302 
1303     data.co = qemu_coroutine_self();
1304     while (data.ret == -EINPROGRESS) {
1305         qemu_coroutine_yield();
1306     }
1307 
1308     trace_nvme_rw_done(s, true, offset, bytes, data.ret);
1309     return data.ret;
1310 }
1311 
1312 
1313 static int coroutine_fn nvme_co_pdiscard(BlockDriverState *bs,
1314                                          int64_t offset,
1315                                          int bytes)
1316 {
1317     BDRVNVMeState *s = bs->opaque;
1318     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1319     NVMeRequest *req;
1320     NvmeDsmRange *buf;
1321     QEMUIOVector local_qiov;
1322     int ret;
1323 
1324     NvmeCmd cmd = {
1325         .opcode = NVME_CMD_DSM,
1326         .nsid = cpu_to_le32(s->nsid),
1327         .cdw10 = cpu_to_le32(0), /*number of ranges - 0 based*/
1328         .cdw11 = cpu_to_le32(1 << 2), /*deallocate bit*/
1329     };
1330 
1331     NVMeCoData data = {
1332         .ctx = bdrv_get_aio_context(bs),
1333         .ret = -EINPROGRESS,
1334     };
1335 
1336     if (!s->supports_discard) {
1337         return -ENOTSUP;
1338     }
1339 
1340     assert(s->queue_count > 1);
1341 
1342     buf = qemu_try_memalign(s->page_size, s->page_size);
1343     if (!buf) {
1344         return -ENOMEM;
1345     }
1346     memset(buf, 0, s->page_size);
1347     buf->nlb = cpu_to_le32(bytes >> s->blkshift);
1348     buf->slba = cpu_to_le64(offset >> s->blkshift);
1349     buf->cattr = 0;
1350 
1351     qemu_iovec_init(&local_qiov, 1);
1352     qemu_iovec_add(&local_qiov, buf, 4096);
1353 
1354     req = nvme_get_free_req(ioq);
1355     assert(req);
1356 
1357     qemu_co_mutex_lock(&s->dma_map_lock);
1358     ret = nvme_cmd_map_qiov(bs, &cmd, req, &local_qiov);
1359     qemu_co_mutex_unlock(&s->dma_map_lock);
1360 
1361     if (ret) {
1362         nvme_put_free_req_and_wake(ioq, req);
1363         goto out;
1364     }
1365 
1366     trace_nvme_dsm(s, offset, bytes);
1367 
1368     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1369 
1370     data.co = qemu_coroutine_self();
1371     while (data.ret == -EINPROGRESS) {
1372         qemu_coroutine_yield();
1373     }
1374 
1375     qemu_co_mutex_lock(&s->dma_map_lock);
1376     ret = nvme_cmd_unmap_qiov(bs, &local_qiov);
1377     qemu_co_mutex_unlock(&s->dma_map_lock);
1378 
1379     if (ret) {
1380         goto out;
1381     }
1382 
1383     ret = data.ret;
1384     trace_nvme_dsm_done(s, offset, bytes, ret);
1385 out:
1386     qemu_iovec_destroy(&local_qiov);
1387     qemu_vfree(buf);
1388     return ret;
1389 
1390 }
1391 
1392 
1393 static int nvme_reopen_prepare(BDRVReopenState *reopen_state,
1394                                BlockReopenQueue *queue, Error **errp)
1395 {
1396     return 0;
1397 }
1398 
1399 static void nvme_refresh_filename(BlockDriverState *bs)
1400 {
1401     BDRVNVMeState *s = bs->opaque;
1402 
1403     snprintf(bs->exact_filename, sizeof(bs->exact_filename), "nvme://%s/%i",
1404              s->device, s->nsid);
1405 }
1406 
1407 static void nvme_refresh_limits(BlockDriverState *bs, Error **errp)
1408 {
1409     BDRVNVMeState *s = bs->opaque;
1410 
1411     bs->bl.opt_mem_alignment = s->page_size;
1412     bs->bl.request_alignment = s->page_size;
1413     bs->bl.max_transfer = s->max_transfer;
1414 }
1415 
1416 static void nvme_detach_aio_context(BlockDriverState *bs)
1417 {
1418     BDRVNVMeState *s = bs->opaque;
1419 
1420     for (unsigned i = 0; i < s->queue_count; i++) {
1421         NVMeQueuePair *q = s->queues[i];
1422 
1423         qemu_bh_delete(q->completion_bh);
1424         q->completion_bh = NULL;
1425     }
1426 
1427     aio_set_event_notifier(bdrv_get_aio_context(bs),
1428                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
1429                            false, NULL, NULL);
1430 }
1431 
1432 static void nvme_attach_aio_context(BlockDriverState *bs,
1433                                     AioContext *new_context)
1434 {
1435     BDRVNVMeState *s = bs->opaque;
1436 
1437     s->aio_context = new_context;
1438     aio_set_event_notifier(new_context, &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
1439                            false, nvme_handle_event, nvme_poll_cb);
1440 
1441     for (unsigned i = 0; i < s->queue_count; i++) {
1442         NVMeQueuePair *q = s->queues[i];
1443 
1444         q->completion_bh =
1445             aio_bh_new(new_context, nvme_process_completion_bh, q);
1446     }
1447 }
1448 
1449 static void nvme_aio_plug(BlockDriverState *bs)
1450 {
1451     BDRVNVMeState *s = bs->opaque;
1452     assert(!s->plugged);
1453     s->plugged = true;
1454 }
1455 
1456 static void nvme_aio_unplug(BlockDriverState *bs)
1457 {
1458     BDRVNVMeState *s = bs->opaque;
1459     assert(s->plugged);
1460     s->plugged = false;
1461     for (unsigned i = INDEX_IO(0); i < s->queue_count; i++) {
1462         NVMeQueuePair *q = s->queues[i];
1463         qemu_mutex_lock(&q->lock);
1464         nvme_kick(q);
1465         nvme_process_completion(q);
1466         qemu_mutex_unlock(&q->lock);
1467     }
1468 }
1469 
1470 static void nvme_register_buf(BlockDriverState *bs, void *host, size_t size)
1471 {
1472     int ret;
1473     BDRVNVMeState *s = bs->opaque;
1474 
1475     ret = qemu_vfio_dma_map(s->vfio, host, size, false, NULL);
1476     if (ret) {
1477         /* FIXME: we may run out of IOVA addresses after repeated
1478          * bdrv_register_buf/bdrv_unregister_buf, because nvme_vfio_dma_unmap
1479          * doesn't reclaim addresses for fixed mappings. */
1480         error_report("nvme_register_buf failed: %s", strerror(-ret));
1481     }
1482 }
1483 
1484 static void nvme_unregister_buf(BlockDriverState *bs, void *host)
1485 {
1486     BDRVNVMeState *s = bs->opaque;
1487 
1488     qemu_vfio_dma_unmap(s->vfio, host);
1489 }
1490 
1491 static BlockStatsSpecific *nvme_get_specific_stats(BlockDriverState *bs)
1492 {
1493     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
1494     BDRVNVMeState *s = bs->opaque;
1495 
1496     stats->driver = BLOCKDEV_DRIVER_NVME;
1497     stats->u.nvme = (BlockStatsSpecificNvme) {
1498         .completion_errors = s->stats.completion_errors,
1499         .aligned_accesses = s->stats.aligned_accesses,
1500         .unaligned_accesses = s->stats.unaligned_accesses,
1501     };
1502 
1503     return stats;
1504 }
1505 
1506 static const char *const nvme_strong_runtime_opts[] = {
1507     NVME_BLOCK_OPT_DEVICE,
1508     NVME_BLOCK_OPT_NAMESPACE,
1509 
1510     NULL
1511 };
1512 
1513 static BlockDriver bdrv_nvme = {
1514     .format_name              = "nvme",
1515     .protocol_name            = "nvme",
1516     .instance_size            = sizeof(BDRVNVMeState),
1517 
1518     .bdrv_co_create_opts      = bdrv_co_create_opts_simple,
1519     .create_opts              = &bdrv_create_opts_simple,
1520 
1521     .bdrv_parse_filename      = nvme_parse_filename,
1522     .bdrv_file_open           = nvme_file_open,
1523     .bdrv_close               = nvme_close,
1524     .bdrv_getlength           = nvme_getlength,
1525     .bdrv_probe_blocksizes    = nvme_probe_blocksizes,
1526 
1527     .bdrv_co_preadv           = nvme_co_preadv,
1528     .bdrv_co_pwritev          = nvme_co_pwritev,
1529 
1530     .bdrv_co_pwrite_zeroes    = nvme_co_pwrite_zeroes,
1531     .bdrv_co_pdiscard         = nvme_co_pdiscard,
1532 
1533     .bdrv_co_flush_to_disk    = nvme_co_flush,
1534     .bdrv_reopen_prepare      = nvme_reopen_prepare,
1535 
1536     .bdrv_refresh_filename    = nvme_refresh_filename,
1537     .bdrv_refresh_limits      = nvme_refresh_limits,
1538     .strong_runtime_opts      = nvme_strong_runtime_opts,
1539     .bdrv_get_specific_stats  = nvme_get_specific_stats,
1540 
1541     .bdrv_detach_aio_context  = nvme_detach_aio_context,
1542     .bdrv_attach_aio_context  = nvme_attach_aio_context,
1543 
1544     .bdrv_io_plug             = nvme_aio_plug,
1545     .bdrv_io_unplug           = nvme_aio_unplug,
1546 
1547     .bdrv_register_buf        = nvme_register_buf,
1548     .bdrv_unregister_buf      = nvme_unregister_buf,
1549 };
1550 
1551 static void bdrv_nvme_init(void)
1552 {
1553     bdrv_register(&bdrv_nvme);
1554 }
1555 
1556 block_init(bdrv_nvme_init);
1557