xref: /openbmc/qemu/block/nvme.c (revision 86269823)
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/memalign.h"
25 #include "qemu/vfio-helpers.h"
26 #include "block/block_int.h"
27 #include "sysemu/replay.h"
28 #include "trace.h"
29 
30 #include "block/nvme.h"
31 
32 #define NVME_SQ_ENTRY_BYTES 64
33 #define NVME_CQ_ENTRY_BYTES 16
34 #define NVME_QUEUE_SIZE 128
35 #define NVME_DOORBELL_SIZE 4096
36 
37 /*
38  * We have to leave one slot empty as that is the full queue case where
39  * head == tail + 1.
40  */
41 #define NVME_NUM_REQS (NVME_QUEUE_SIZE - 1)
42 
43 typedef struct BDRVNVMeState BDRVNVMeState;
44 
45 /* Same index is used for queues and IRQs */
46 #define INDEX_ADMIN     0
47 #define INDEX_IO(n)     (1 + n)
48 
49 /* This driver shares a single MSIX IRQ for the admin and I/O queues */
50 enum {
51     MSIX_SHARED_IRQ_IDX = 0,
52     MSIX_IRQ_COUNT = 1
53 };
54 
55 typedef struct {
56     int32_t  head, tail;
57     uint8_t  *queue;
58     uint64_t iova;
59     /* Hardware MMIO register */
60     volatile uint32_t *doorbell;
61 } NVMeQueue;
62 
63 typedef struct {
64     BlockCompletionFunc *cb;
65     void *opaque;
66     int cid;
67     void *prp_list_page;
68     uint64_t prp_list_iova;
69     int free_req_next; /* q->reqs[] index of next free req */
70 } NVMeRequest;
71 
72 typedef struct {
73     QemuMutex   lock;
74 
75     /* Read from I/O code path, initialized under BQL */
76     BDRVNVMeState   *s;
77     int             index;
78 
79     /* Fields protected by BQL */
80     uint8_t     *prp_list_pages;
81 
82     /* Fields protected by @lock */
83     CoQueue     free_req_queue;
84     NVMeQueue   sq, cq;
85     int         cq_phase;
86     int         free_req_head;
87     NVMeRequest reqs[NVME_NUM_REQS];
88     int         need_kick;
89     int         inflight;
90 
91     /* Thread-safe, no lock necessary */
92     QEMUBH      *completion_bh;
93 } NVMeQueuePair;
94 
95 struct BDRVNVMeState {
96     AioContext *aio_context;
97     QEMUVFIOState *vfio;
98     void *bar0_wo_map;
99     /* Memory mapped registers */
100     volatile struct {
101         uint32_t sq_tail;
102         uint32_t cq_head;
103     } *doorbells;
104     /* The submission/completion queue pairs.
105      * [0]: admin queue.
106      * [1..]: io queues.
107      */
108     NVMeQueuePair **queues;
109     unsigned queue_count;
110     size_t page_size;
111     /* How many uint32_t elements does each doorbell entry take. */
112     size_t doorbell_scale;
113     bool write_cache_supported;
114     EventNotifier irq_notifier[MSIX_IRQ_COUNT];
115 
116     uint64_t nsze; /* Namespace size reported by identify command */
117     int nsid;      /* The namespace id to read/write data. */
118     int blkshift;
119 
120     uint64_t max_transfer;
121     bool plugged;
122 
123     bool supports_write_zeroes;
124     bool supports_discard;
125 
126     CoMutex dma_map_lock;
127     CoQueue dma_flush_queue;
128 
129     /* Total size of mapped qiov, accessed under dma_map_lock */
130     int dma_map_count;
131 
132     /* PCI address (required for nvme_refresh_filename()) */
133     char *device;
134 
135     struct {
136         uint64_t completion_errors;
137         uint64_t aligned_accesses;
138         uint64_t unaligned_accesses;
139     } stats;
140 };
141 
142 #define NVME_BLOCK_OPT_DEVICE "device"
143 #define NVME_BLOCK_OPT_NAMESPACE "namespace"
144 
145 static void nvme_process_completion_bh(void *opaque);
146 
147 static QemuOptsList runtime_opts = {
148     .name = "nvme",
149     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
150     .desc = {
151         {
152             .name = NVME_BLOCK_OPT_DEVICE,
153             .type = QEMU_OPT_STRING,
154             .help = "NVMe PCI device address",
155         },
156         {
157             .name = NVME_BLOCK_OPT_NAMESPACE,
158             .type = QEMU_OPT_NUMBER,
159             .help = "NVMe namespace",
160         },
161         { /* end of list */ }
162     },
163 };
164 
165 /* Returns true on success, false on failure. */
166 static bool nvme_init_queue(BDRVNVMeState *s, NVMeQueue *q,
167                             unsigned nentries, size_t entry_bytes, Error **errp)
168 {
169     size_t bytes;
170     int r;
171 
172     bytes = ROUND_UP(nentries * entry_bytes, qemu_real_host_page_size());
173     q->head = q->tail = 0;
174     q->queue = qemu_try_memalign(qemu_real_host_page_size(), bytes);
175     if (!q->queue) {
176         error_setg(errp, "Cannot allocate queue");
177         return false;
178     }
179     memset(q->queue, 0, bytes);
180     r = qemu_vfio_dma_map(s->vfio, q->queue, bytes, false, &q->iova, errp);
181     if (r) {
182         error_prepend(errp, "Cannot map queue: ");
183     }
184     return r == 0;
185 }
186 
187 static void nvme_free_queue(NVMeQueue *q)
188 {
189     qemu_vfree(q->queue);
190 }
191 
192 static void nvme_free_queue_pair(NVMeQueuePair *q)
193 {
194     trace_nvme_free_queue_pair(q->index, q, &q->cq, &q->sq);
195     if (q->completion_bh) {
196         qemu_bh_delete(q->completion_bh);
197     }
198     nvme_free_queue(&q->sq);
199     nvme_free_queue(&q->cq);
200     qemu_vfree(q->prp_list_pages);
201     qemu_mutex_destroy(&q->lock);
202     g_free(q);
203 }
204 
205 static void nvme_free_req_queue_cb(void *opaque)
206 {
207     NVMeQueuePair *q = opaque;
208 
209     qemu_mutex_lock(&q->lock);
210     while (q->free_req_head != -1 &&
211            qemu_co_enter_next(&q->free_req_queue, &q->lock)) {
212         /* Retry waiting requests */
213     }
214     qemu_mutex_unlock(&q->lock);
215 }
216 
217 static NVMeQueuePair *nvme_create_queue_pair(BDRVNVMeState *s,
218                                              AioContext *aio_context,
219                                              unsigned idx, size_t size,
220                                              Error **errp)
221 {
222     int i, r;
223     NVMeQueuePair *q;
224     uint64_t prp_list_iova;
225     size_t bytes;
226 
227     q = g_try_new0(NVMeQueuePair, 1);
228     if (!q) {
229         error_setg(errp, "Cannot allocate queue pair");
230         return NULL;
231     }
232     trace_nvme_create_queue_pair(idx, q, size, aio_context,
233                                  event_notifier_get_fd(s->irq_notifier));
234     bytes = QEMU_ALIGN_UP(s->page_size * NVME_NUM_REQS,
235                           qemu_real_host_page_size());
236     q->prp_list_pages = qemu_try_memalign(qemu_real_host_page_size(), bytes);
237     if (!q->prp_list_pages) {
238         error_setg(errp, "Cannot allocate PRP page list");
239         goto fail;
240     }
241     memset(q->prp_list_pages, 0, bytes);
242     qemu_mutex_init(&q->lock);
243     q->s = s;
244     q->index = idx;
245     qemu_co_queue_init(&q->free_req_queue);
246     q->completion_bh = aio_bh_new(aio_context, nvme_process_completion_bh, q);
247     r = qemu_vfio_dma_map(s->vfio, q->prp_list_pages, bytes,
248                           false, &prp_list_iova, errp);
249     if (r) {
250         error_prepend(errp, "Cannot map buffer for DMA: ");
251         goto fail;
252     }
253     q->free_req_head = -1;
254     for (i = 0; i < NVME_NUM_REQS; i++) {
255         NVMeRequest *req = &q->reqs[i];
256         req->cid = i + 1;
257         req->free_req_next = q->free_req_head;
258         q->free_req_head = i;
259         req->prp_list_page = q->prp_list_pages + i * s->page_size;
260         req->prp_list_iova = prp_list_iova + i * s->page_size;
261     }
262 
263     if (!nvme_init_queue(s, &q->sq, size, NVME_SQ_ENTRY_BYTES, errp)) {
264         goto fail;
265     }
266     q->sq.doorbell = &s->doorbells[idx * s->doorbell_scale].sq_tail;
267 
268     if (!nvme_init_queue(s, &q->cq, size, NVME_CQ_ENTRY_BYTES, errp)) {
269         goto fail;
270     }
271     q->cq.doorbell = &s->doorbells[idx * s->doorbell_scale].cq_head;
272 
273     return q;
274 fail:
275     nvme_free_queue_pair(q);
276     return NULL;
277 }
278 
279 /* With q->lock */
280 static void nvme_kick(NVMeQueuePair *q)
281 {
282     BDRVNVMeState *s = q->s;
283 
284     if (s->plugged || !q->need_kick) {
285         return;
286     }
287     trace_nvme_kick(s, q->index);
288     assert(!(q->sq.tail & 0xFF00));
289     /* Fence the write to submission queue entry before notifying the device. */
290     smp_wmb();
291     *q->sq.doorbell = cpu_to_le32(q->sq.tail);
292     q->inflight += q->need_kick;
293     q->need_kick = 0;
294 }
295 
296 static NVMeRequest *nvme_get_free_req_nofail_locked(NVMeQueuePair *q)
297 {
298     NVMeRequest *req;
299 
300     req = &q->reqs[q->free_req_head];
301     q->free_req_head = req->free_req_next;
302     req->free_req_next = -1;
303     return req;
304 }
305 
306 /* Return a free request element if any, otherwise return NULL.  */
307 static NVMeRequest *nvme_get_free_req_nowait(NVMeQueuePair *q)
308 {
309     QEMU_LOCK_GUARD(&q->lock);
310     if (q->free_req_head == -1) {
311         return NULL;
312     }
313     return nvme_get_free_req_nofail_locked(q);
314 }
315 
316 /*
317  * Wait for a free request to become available if necessary, then
318  * return it.
319  */
320 static coroutine_fn NVMeRequest *nvme_get_free_req(NVMeQueuePair *q)
321 {
322     QEMU_LOCK_GUARD(&q->lock);
323 
324     while (q->free_req_head == -1) {
325         trace_nvme_free_req_queue_wait(q->s, q->index);
326         qemu_co_queue_wait(&q->free_req_queue, &q->lock);
327     }
328 
329     return nvme_get_free_req_nofail_locked(q);
330 }
331 
332 /* With q->lock */
333 static void nvme_put_free_req_locked(NVMeQueuePair *q, NVMeRequest *req)
334 {
335     req->free_req_next = q->free_req_head;
336     q->free_req_head = req - q->reqs;
337 }
338 
339 /* With q->lock */
340 static void nvme_wake_free_req_locked(NVMeQueuePair *q)
341 {
342     if (!qemu_co_queue_empty(&q->free_req_queue)) {
343         replay_bh_schedule_oneshot_event(q->s->aio_context,
344                 nvme_free_req_queue_cb, q);
345     }
346 }
347 
348 /* Insert a request in the freelist and wake waiters */
349 static void nvme_put_free_req_and_wake(NVMeQueuePair *q, NVMeRequest *req)
350 {
351     qemu_mutex_lock(&q->lock);
352     nvme_put_free_req_locked(q, req);
353     nvme_wake_free_req_locked(q);
354     qemu_mutex_unlock(&q->lock);
355 }
356 
357 static inline int nvme_translate_error(const NvmeCqe *c)
358 {
359     uint16_t status = (le16_to_cpu(c->status) >> 1) & 0xFF;
360     if (status) {
361         trace_nvme_error(le32_to_cpu(c->result),
362                          le16_to_cpu(c->sq_head),
363                          le16_to_cpu(c->sq_id),
364                          le16_to_cpu(c->cid),
365                          le16_to_cpu(status));
366     }
367     switch (status) {
368     case 0:
369         return 0;
370     case 1:
371         return -ENOSYS;
372     case 2:
373         return -EINVAL;
374     default:
375         return -EIO;
376     }
377 }
378 
379 /* With q->lock */
380 static bool nvme_process_completion(NVMeQueuePair *q)
381 {
382     BDRVNVMeState *s = q->s;
383     bool progress = false;
384     NVMeRequest *preq;
385     NVMeRequest req;
386     NvmeCqe *c;
387 
388     trace_nvme_process_completion(s, q->index, q->inflight);
389     if (s->plugged) {
390         trace_nvme_process_completion_queue_plugged(s, q->index);
391         return false;
392     }
393 
394     /*
395      * Support re-entrancy when a request cb() function invokes aio_poll().
396      * Pending completions must be visible to aio_poll() so that a cb()
397      * function can wait for the completion of another request.
398      *
399      * The aio_poll() loop will execute our BH and we'll resume completion
400      * processing there.
401      */
402     qemu_bh_schedule(q->completion_bh);
403 
404     assert(q->inflight >= 0);
405     while (q->inflight) {
406         int ret;
407         int16_t cid;
408 
409         c = (NvmeCqe *)&q->cq.queue[q->cq.head * NVME_CQ_ENTRY_BYTES];
410         if ((le16_to_cpu(c->status) & 0x1) == q->cq_phase) {
411             break;
412         }
413         ret = nvme_translate_error(c);
414         if (ret) {
415             s->stats.completion_errors++;
416         }
417         q->cq.head = (q->cq.head + 1) % NVME_QUEUE_SIZE;
418         if (!q->cq.head) {
419             q->cq_phase = !q->cq_phase;
420         }
421         cid = le16_to_cpu(c->cid);
422         if (cid == 0 || cid > NVME_QUEUE_SIZE) {
423             warn_report("NVMe: Unexpected CID in completion queue: %"PRIu32", "
424                         "queue size: %u", cid, NVME_QUEUE_SIZE);
425             continue;
426         }
427         trace_nvme_complete_command(s, q->index, cid);
428         preq = &q->reqs[cid - 1];
429         req = *preq;
430         assert(req.cid == cid);
431         assert(req.cb);
432         nvme_put_free_req_locked(q, preq);
433         preq->cb = preq->opaque = NULL;
434         q->inflight--;
435         qemu_mutex_unlock(&q->lock);
436         req.cb(req.opaque, ret);
437         qemu_mutex_lock(&q->lock);
438         progress = true;
439     }
440     if (progress) {
441         /* Notify the device so it can post more completions. */
442         smp_mb_release();
443         *q->cq.doorbell = cpu_to_le32(q->cq.head);
444         nvme_wake_free_req_locked(q);
445     }
446 
447     qemu_bh_cancel(q->completion_bh);
448 
449     return progress;
450 }
451 
452 static void nvme_process_completion_bh(void *opaque)
453 {
454     NVMeQueuePair *q = opaque;
455 
456     /*
457      * We're being invoked because a nvme_process_completion() cb() function
458      * called aio_poll(). The callback may be waiting for further completions
459      * so notify the device that it has space to fill in more completions now.
460      */
461     smp_mb_release();
462     *q->cq.doorbell = cpu_to_le32(q->cq.head);
463     nvme_wake_free_req_locked(q);
464 
465     nvme_process_completion(q);
466 }
467 
468 static void nvme_trace_command(const NvmeCmd *cmd)
469 {
470     int i;
471 
472     if (!trace_event_get_state_backends(TRACE_NVME_SUBMIT_COMMAND_RAW)) {
473         return;
474     }
475     for (i = 0; i < 8; ++i) {
476         uint8_t *cmdp = (uint8_t *)cmd + i * 8;
477         trace_nvme_submit_command_raw(cmdp[0], cmdp[1], cmdp[2], cmdp[3],
478                                       cmdp[4], cmdp[5], cmdp[6], cmdp[7]);
479     }
480 }
481 
482 static void nvme_submit_command(NVMeQueuePair *q, NVMeRequest *req,
483                                 NvmeCmd *cmd, BlockCompletionFunc cb,
484                                 void *opaque)
485 {
486     assert(!req->cb);
487     req->cb = cb;
488     req->opaque = opaque;
489     cmd->cid = cpu_to_le16(req->cid);
490 
491     trace_nvme_submit_command(q->s, q->index, req->cid);
492     nvme_trace_command(cmd);
493     qemu_mutex_lock(&q->lock);
494     memcpy((uint8_t *)q->sq.queue +
495            q->sq.tail * NVME_SQ_ENTRY_BYTES, cmd, sizeof(*cmd));
496     q->sq.tail = (q->sq.tail + 1) % NVME_QUEUE_SIZE;
497     q->need_kick++;
498     nvme_kick(q);
499     nvme_process_completion(q);
500     qemu_mutex_unlock(&q->lock);
501 }
502 
503 static void nvme_admin_cmd_sync_cb(void *opaque, int ret)
504 {
505     int *pret = opaque;
506     *pret = ret;
507     aio_wait_kick();
508 }
509 
510 static int nvme_admin_cmd_sync(BlockDriverState *bs, NvmeCmd *cmd)
511 {
512     BDRVNVMeState *s = bs->opaque;
513     NVMeQueuePair *q = s->queues[INDEX_ADMIN];
514     AioContext *aio_context = bdrv_get_aio_context(bs);
515     NVMeRequest *req;
516     int ret = -EINPROGRESS;
517     req = nvme_get_free_req_nowait(q);
518     if (!req) {
519         return -EBUSY;
520     }
521     nvme_submit_command(q, req, cmd, nvme_admin_cmd_sync_cb, &ret);
522 
523     AIO_WAIT_WHILE(aio_context, ret == -EINPROGRESS);
524     return ret;
525 }
526 
527 /* Returns true on success, false on failure. */
528 static bool nvme_identify(BlockDriverState *bs, int namespace, Error **errp)
529 {
530     BDRVNVMeState *s = bs->opaque;
531     bool ret = false;
532     QEMU_AUTO_VFREE union {
533         NvmeIdCtrl ctrl;
534         NvmeIdNs ns;
535     } *id = NULL;
536     NvmeLBAF *lbaf;
537     uint16_t oncs;
538     int r;
539     uint64_t iova;
540     NvmeCmd cmd = {
541         .opcode = NVME_ADM_CMD_IDENTIFY,
542         .cdw10 = cpu_to_le32(0x1),
543     };
544     size_t id_size = QEMU_ALIGN_UP(sizeof(*id), qemu_real_host_page_size());
545 
546     id = qemu_try_memalign(qemu_real_host_page_size(), id_size);
547     if (!id) {
548         error_setg(errp, "Cannot allocate buffer for identify response");
549         goto out;
550     }
551     r = qemu_vfio_dma_map(s->vfio, id, id_size, true, &iova, errp);
552     if (r) {
553         error_prepend(errp, "Cannot map buffer for DMA: ");
554         goto out;
555     }
556 
557     memset(id, 0, id_size);
558     cmd.dptr.prp1 = cpu_to_le64(iova);
559     if (nvme_admin_cmd_sync(bs, &cmd)) {
560         error_setg(errp, "Failed to identify controller");
561         goto out;
562     }
563 
564     if (le32_to_cpu(id->ctrl.nn) < namespace) {
565         error_setg(errp, "Invalid namespace");
566         goto out;
567     }
568     s->write_cache_supported = le32_to_cpu(id->ctrl.vwc) & 0x1;
569     s->max_transfer = (id->ctrl.mdts ? 1 << id->ctrl.mdts : 0) * s->page_size;
570     /* For now the page list buffer per command is one page, to hold at most
571      * s->page_size / sizeof(uint64_t) entries. */
572     s->max_transfer = MIN_NON_ZERO(s->max_transfer,
573                           s->page_size / sizeof(uint64_t) * s->page_size);
574 
575     oncs = le16_to_cpu(id->ctrl.oncs);
576     s->supports_write_zeroes = !!(oncs & NVME_ONCS_WRITE_ZEROES);
577     s->supports_discard = !!(oncs & NVME_ONCS_DSM);
578 
579     memset(id, 0, id_size);
580     cmd.cdw10 = 0;
581     cmd.nsid = cpu_to_le32(namespace);
582     if (nvme_admin_cmd_sync(bs, &cmd)) {
583         error_setg(errp, "Failed to identify namespace");
584         goto out;
585     }
586 
587     s->nsze = le64_to_cpu(id->ns.nsze);
588     lbaf = &id->ns.lbaf[NVME_ID_NS_FLBAS_INDEX(id->ns.flbas)];
589 
590     if (NVME_ID_NS_DLFEAT_WRITE_ZEROES(id->ns.dlfeat) &&
591             NVME_ID_NS_DLFEAT_READ_BEHAVIOR(id->ns.dlfeat) ==
592                     NVME_ID_NS_DLFEAT_READ_BEHAVIOR_ZEROES) {
593         bs->supported_write_flags |= BDRV_REQ_MAY_UNMAP;
594     }
595 
596     if (lbaf->ms) {
597         error_setg(errp, "Namespaces with metadata are not yet supported");
598         goto out;
599     }
600 
601     if (lbaf->ds < BDRV_SECTOR_BITS || lbaf->ds > 12 ||
602         (1 << lbaf->ds) > s->page_size)
603     {
604         error_setg(errp, "Namespace has unsupported block size (2^%d)",
605                    lbaf->ds);
606         goto out;
607     }
608 
609     ret = true;
610     s->blkshift = lbaf->ds;
611 out:
612     qemu_vfio_dma_unmap(s->vfio, id);
613 
614     return ret;
615 }
616 
617 static void nvme_poll_queue(NVMeQueuePair *q)
618 {
619     const size_t cqe_offset = q->cq.head * NVME_CQ_ENTRY_BYTES;
620     NvmeCqe *cqe = (NvmeCqe *)&q->cq.queue[cqe_offset];
621 
622     trace_nvme_poll_queue(q->s, q->index);
623     /*
624      * Do an early check for completions. q->lock isn't needed because
625      * nvme_process_completion() only runs in the event loop thread and
626      * cannot race with itself.
627      */
628     if ((le16_to_cpu(cqe->status) & 0x1) == q->cq_phase) {
629         return;
630     }
631 
632     qemu_mutex_lock(&q->lock);
633     while (nvme_process_completion(q)) {
634         /* Keep polling */
635     }
636     qemu_mutex_unlock(&q->lock);
637 }
638 
639 static void nvme_poll_queues(BDRVNVMeState *s)
640 {
641     int i;
642 
643     for (i = 0; i < s->queue_count; i++) {
644         nvme_poll_queue(s->queues[i]);
645     }
646 }
647 
648 static void nvme_handle_event(EventNotifier *n)
649 {
650     BDRVNVMeState *s = container_of(n, BDRVNVMeState,
651                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
652 
653     trace_nvme_handle_event(s);
654     event_notifier_test_and_clear(n);
655     nvme_poll_queues(s);
656 }
657 
658 static bool nvme_add_io_queue(BlockDriverState *bs, Error **errp)
659 {
660     BDRVNVMeState *s = bs->opaque;
661     unsigned n = s->queue_count;
662     NVMeQueuePair *q;
663     NvmeCmd cmd;
664     unsigned queue_size = NVME_QUEUE_SIZE;
665 
666     assert(n <= UINT16_MAX);
667     q = nvme_create_queue_pair(s, bdrv_get_aio_context(bs),
668                                n, queue_size, errp);
669     if (!q) {
670         return false;
671     }
672     cmd = (NvmeCmd) {
673         .opcode = NVME_ADM_CMD_CREATE_CQ,
674         .dptr.prp1 = cpu_to_le64(q->cq.iova),
675         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | n),
676         .cdw11 = cpu_to_le32(NVME_CQ_IEN | NVME_CQ_PC),
677     };
678     if (nvme_admin_cmd_sync(bs, &cmd)) {
679         error_setg(errp, "Failed to create CQ io queue [%u]", n);
680         goto out_error;
681     }
682     cmd = (NvmeCmd) {
683         .opcode = NVME_ADM_CMD_CREATE_SQ,
684         .dptr.prp1 = cpu_to_le64(q->sq.iova),
685         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | n),
686         .cdw11 = cpu_to_le32(NVME_SQ_PC | (n << 16)),
687     };
688     if (nvme_admin_cmd_sync(bs, &cmd)) {
689         error_setg(errp, "Failed to create SQ io queue [%u]", n);
690         goto out_error;
691     }
692     s->queues = g_renew(NVMeQueuePair *, s->queues, n + 1);
693     s->queues[n] = q;
694     s->queue_count++;
695     return true;
696 out_error:
697     nvme_free_queue_pair(q);
698     return false;
699 }
700 
701 static bool nvme_poll_cb(void *opaque)
702 {
703     EventNotifier *e = opaque;
704     BDRVNVMeState *s = container_of(e, BDRVNVMeState,
705                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
706     int i;
707 
708     for (i = 0; i < s->queue_count; i++) {
709         NVMeQueuePair *q = s->queues[i];
710         const size_t cqe_offset = q->cq.head * NVME_CQ_ENTRY_BYTES;
711         NvmeCqe *cqe = (NvmeCqe *)&q->cq.queue[cqe_offset];
712 
713         /*
714          * q->lock isn't needed because nvme_process_completion() only runs in
715          * the event loop thread and cannot race with itself.
716          */
717         if ((le16_to_cpu(cqe->status) & 0x1) != q->cq_phase) {
718             return true;
719         }
720     }
721     return false;
722 }
723 
724 static void nvme_poll_ready(EventNotifier *e)
725 {
726     BDRVNVMeState *s = container_of(e, BDRVNVMeState,
727                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
728 
729     nvme_poll_queues(s);
730 }
731 
732 static int nvme_init(BlockDriverState *bs, const char *device, int namespace,
733                      Error **errp)
734 {
735     BDRVNVMeState *s = bs->opaque;
736     NVMeQueuePair *q;
737     AioContext *aio_context = bdrv_get_aio_context(bs);
738     int ret;
739     uint64_t cap;
740     uint32_t ver;
741     uint64_t timeout_ms;
742     uint64_t deadline, now;
743     volatile NvmeBar *regs = NULL;
744 
745     qemu_co_mutex_init(&s->dma_map_lock);
746     qemu_co_queue_init(&s->dma_flush_queue);
747     s->device = g_strdup(device);
748     s->nsid = namespace;
749     s->aio_context = bdrv_get_aio_context(bs);
750     ret = event_notifier_init(&s->irq_notifier[MSIX_SHARED_IRQ_IDX], 0);
751     if (ret) {
752         error_setg(errp, "Failed to init event notifier");
753         return ret;
754     }
755 
756     s->vfio = qemu_vfio_open_pci(device, errp);
757     if (!s->vfio) {
758         ret = -EINVAL;
759         goto out;
760     }
761 
762     regs = qemu_vfio_pci_map_bar(s->vfio, 0, 0, sizeof(NvmeBar),
763                                  PROT_READ | PROT_WRITE, errp);
764     if (!regs) {
765         ret = -EINVAL;
766         goto out;
767     }
768     /* Perform initialize sequence as described in NVMe spec "7.6.1
769      * Initialization". */
770 
771     cap = le64_to_cpu(regs->cap);
772     trace_nvme_controller_capability_raw(cap);
773     trace_nvme_controller_capability("Maximum Queue Entries Supported",
774                                      1 + NVME_CAP_MQES(cap));
775     trace_nvme_controller_capability("Contiguous Queues Required",
776                                      NVME_CAP_CQR(cap));
777     trace_nvme_controller_capability("Doorbell Stride",
778                                      1 << (2 + NVME_CAP_DSTRD(cap)));
779     trace_nvme_controller_capability("Subsystem Reset Supported",
780                                      NVME_CAP_NSSRS(cap));
781     trace_nvme_controller_capability("Memory Page Size Minimum",
782                                      1 << (12 + NVME_CAP_MPSMIN(cap)));
783     trace_nvme_controller_capability("Memory Page Size Maximum",
784                                      1 << (12 + NVME_CAP_MPSMAX(cap)));
785     if (!NVME_CAP_CSS(cap)) {
786         error_setg(errp, "Device doesn't support NVMe command set");
787         ret = -EINVAL;
788         goto out;
789     }
790 
791     s->page_size = 1u << (12 + NVME_CAP_MPSMIN(cap));
792     s->doorbell_scale = (4 << NVME_CAP_DSTRD(cap)) / sizeof(uint32_t);
793     bs->bl.opt_mem_alignment = s->page_size;
794     bs->bl.request_alignment = s->page_size;
795     timeout_ms = MIN(500 * NVME_CAP_TO(cap), 30000);
796 
797     ver = le32_to_cpu(regs->vs);
798     trace_nvme_controller_spec_version(extract32(ver, 16, 16),
799                                        extract32(ver, 8, 8),
800                                        extract32(ver, 0, 8));
801 
802     /* Reset device to get a clean state. */
803     regs->cc = cpu_to_le32(le32_to_cpu(regs->cc) & 0xFE);
804     /* Wait for CSTS.RDY = 0. */
805     deadline = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) + timeout_ms * SCALE_MS;
806     while (NVME_CSTS_RDY(le32_to_cpu(regs->csts))) {
807         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
808             error_setg(errp, "Timeout while waiting for device to reset (%"
809                              PRId64 " ms)",
810                        timeout_ms);
811             ret = -ETIMEDOUT;
812             goto out;
813         }
814     }
815 
816     s->bar0_wo_map = qemu_vfio_pci_map_bar(s->vfio, 0, 0,
817                                            sizeof(NvmeBar) + NVME_DOORBELL_SIZE,
818                                            PROT_WRITE, errp);
819     s->doorbells = (void *)((uintptr_t)s->bar0_wo_map + sizeof(NvmeBar));
820     if (!s->doorbells) {
821         ret = -EINVAL;
822         goto out;
823     }
824 
825     /* Set up admin queue. */
826     s->queues = g_new(NVMeQueuePair *, 1);
827     q = nvme_create_queue_pair(s, aio_context, 0, NVME_QUEUE_SIZE, errp);
828     if (!q) {
829         ret = -EINVAL;
830         goto out;
831     }
832     s->queues[INDEX_ADMIN] = q;
833     s->queue_count = 1;
834     QEMU_BUILD_BUG_ON((NVME_QUEUE_SIZE - 1) & 0xF000);
835     regs->aqa = cpu_to_le32(((NVME_QUEUE_SIZE - 1) << AQA_ACQS_SHIFT) |
836                             ((NVME_QUEUE_SIZE - 1) << AQA_ASQS_SHIFT));
837     regs->asq = cpu_to_le64(q->sq.iova);
838     regs->acq = cpu_to_le64(q->cq.iova);
839 
840     /* After setting up all control registers we can enable device now. */
841     regs->cc = cpu_to_le32((ctz32(NVME_CQ_ENTRY_BYTES) << CC_IOCQES_SHIFT) |
842                            (ctz32(NVME_SQ_ENTRY_BYTES) << CC_IOSQES_SHIFT) |
843                            CC_EN_MASK);
844     /* Wait for CSTS.RDY = 1. */
845     now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
846     deadline = now + timeout_ms * SCALE_MS;
847     while (!NVME_CSTS_RDY(le32_to_cpu(regs->csts))) {
848         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
849             error_setg(errp, "Timeout while waiting for device to start (%"
850                              PRId64 " ms)",
851                        timeout_ms);
852             ret = -ETIMEDOUT;
853             goto out;
854         }
855     }
856 
857     ret = qemu_vfio_pci_init_irq(s->vfio, s->irq_notifier,
858                                  VFIO_PCI_MSIX_IRQ_INDEX, errp);
859     if (ret) {
860         goto out;
861     }
862     aio_set_event_notifier(bdrv_get_aio_context(bs),
863                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
864                            false, nvme_handle_event, nvme_poll_cb,
865                            nvme_poll_ready);
866 
867     if (!nvme_identify(bs, namespace, errp)) {
868         ret = -EIO;
869         goto out;
870     }
871 
872     /* Set up command queues. */
873     if (!nvme_add_io_queue(bs, errp)) {
874         ret = -EIO;
875     }
876 out:
877     if (regs) {
878         qemu_vfio_pci_unmap_bar(s->vfio, 0, (void *)regs, 0, sizeof(NvmeBar));
879     }
880 
881     /* Cleaning up is done in nvme_file_open() upon error. */
882     return ret;
883 }
884 
885 /* Parse a filename in the format of nvme://XXXX:XX:XX.X/X. Example:
886  *
887  *     nvme://0000:44:00.0/1
888  *
889  * where the "nvme://" is a fixed form of the protocol prefix, the middle part
890  * is the PCI address, and the last part is the namespace number starting from
891  * 1 according to the NVMe spec. */
892 static void nvme_parse_filename(const char *filename, QDict *options,
893                                 Error **errp)
894 {
895     int pref = strlen("nvme://");
896 
897     if (strlen(filename) > pref && !strncmp(filename, "nvme://", pref)) {
898         const char *tmp = filename + pref;
899         char *device;
900         const char *namespace;
901         unsigned long ns;
902         const char *slash = strchr(tmp, '/');
903         if (!slash) {
904             qdict_put_str(options, NVME_BLOCK_OPT_DEVICE, tmp);
905             return;
906         }
907         device = g_strndup(tmp, slash - tmp);
908         qdict_put_str(options, NVME_BLOCK_OPT_DEVICE, device);
909         g_free(device);
910         namespace = slash + 1;
911         if (*namespace && qemu_strtoul(namespace, NULL, 10, &ns)) {
912             error_setg(errp, "Invalid namespace '%s', positive number expected",
913                        namespace);
914             return;
915         }
916         qdict_put_str(options, NVME_BLOCK_OPT_NAMESPACE,
917                       *namespace ? namespace : "1");
918     }
919 }
920 
921 static int nvme_enable_disable_write_cache(BlockDriverState *bs, bool enable,
922                                            Error **errp)
923 {
924     int ret;
925     BDRVNVMeState *s = bs->opaque;
926     NvmeCmd cmd = {
927         .opcode = NVME_ADM_CMD_SET_FEATURES,
928         .nsid = cpu_to_le32(s->nsid),
929         .cdw10 = cpu_to_le32(0x06),
930         .cdw11 = cpu_to_le32(enable ? 0x01 : 0x00),
931     };
932 
933     ret = nvme_admin_cmd_sync(bs, &cmd);
934     if (ret) {
935         error_setg(errp, "Failed to configure NVMe write cache");
936     }
937     return ret;
938 }
939 
940 static void nvme_close(BlockDriverState *bs)
941 {
942     BDRVNVMeState *s = bs->opaque;
943 
944     for (unsigned i = 0; i < s->queue_count; ++i) {
945         nvme_free_queue_pair(s->queues[i]);
946     }
947     g_free(s->queues);
948     aio_set_event_notifier(bdrv_get_aio_context(bs),
949                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
950                            false, NULL, NULL, NULL);
951     event_notifier_cleanup(&s->irq_notifier[MSIX_SHARED_IRQ_IDX]);
952     qemu_vfio_pci_unmap_bar(s->vfio, 0, s->bar0_wo_map,
953                             0, sizeof(NvmeBar) + NVME_DOORBELL_SIZE);
954     qemu_vfio_close(s->vfio);
955 
956     g_free(s->device);
957 }
958 
959 static int nvme_file_open(BlockDriverState *bs, QDict *options, int flags,
960                           Error **errp)
961 {
962     const char *device;
963     QemuOpts *opts;
964     int namespace;
965     int ret;
966     BDRVNVMeState *s = bs->opaque;
967 
968     bs->supported_write_flags = BDRV_REQ_FUA;
969 
970     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
971     qemu_opts_absorb_qdict(opts, options, &error_abort);
972     device = qemu_opt_get(opts, NVME_BLOCK_OPT_DEVICE);
973     if (!device) {
974         error_setg(errp, "'" NVME_BLOCK_OPT_DEVICE "' option is required");
975         qemu_opts_del(opts);
976         return -EINVAL;
977     }
978 
979     namespace = qemu_opt_get_number(opts, NVME_BLOCK_OPT_NAMESPACE, 1);
980     ret = nvme_init(bs, device, namespace, errp);
981     qemu_opts_del(opts);
982     if (ret) {
983         goto fail;
984     }
985     if (flags & BDRV_O_NOCACHE) {
986         if (!s->write_cache_supported) {
987             error_setg(errp,
988                        "NVMe controller doesn't support write cache configuration");
989             ret = -EINVAL;
990         } else {
991             ret = nvme_enable_disable_write_cache(bs, !(flags & BDRV_O_NOCACHE),
992                                                   errp);
993         }
994         if (ret) {
995             goto fail;
996         }
997     }
998     return 0;
999 fail:
1000     nvme_close(bs);
1001     return ret;
1002 }
1003 
1004 static int64_t nvme_getlength(BlockDriverState *bs)
1005 {
1006     BDRVNVMeState *s = bs->opaque;
1007     return s->nsze << s->blkshift;
1008 }
1009 
1010 static uint32_t nvme_get_blocksize(BlockDriverState *bs)
1011 {
1012     BDRVNVMeState *s = bs->opaque;
1013     assert(s->blkshift >= BDRV_SECTOR_BITS && s->blkshift <= 12);
1014     return UINT32_C(1) << s->blkshift;
1015 }
1016 
1017 static int nvme_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
1018 {
1019     uint32_t blocksize = nvme_get_blocksize(bs);
1020     bsz->phys = blocksize;
1021     bsz->log = blocksize;
1022     return 0;
1023 }
1024 
1025 /* Called with s->dma_map_lock */
1026 static coroutine_fn int nvme_cmd_unmap_qiov(BlockDriverState *bs,
1027                                             QEMUIOVector *qiov)
1028 {
1029     int r = 0;
1030     BDRVNVMeState *s = bs->opaque;
1031 
1032     s->dma_map_count -= qiov->size;
1033     if (!s->dma_map_count && !qemu_co_queue_empty(&s->dma_flush_queue)) {
1034         r = qemu_vfio_dma_reset_temporary(s->vfio);
1035         if (!r) {
1036             qemu_co_queue_restart_all(&s->dma_flush_queue);
1037         }
1038     }
1039     return r;
1040 }
1041 
1042 /* Called with s->dma_map_lock */
1043 static coroutine_fn int nvme_cmd_map_qiov(BlockDriverState *bs, NvmeCmd *cmd,
1044                                           NVMeRequest *req, QEMUIOVector *qiov)
1045 {
1046     BDRVNVMeState *s = bs->opaque;
1047     uint64_t *pagelist = req->prp_list_page;
1048     int i, j, r;
1049     int entries = 0;
1050     Error *local_err = NULL, **errp = NULL;
1051 
1052     assert(qiov->size);
1053     assert(QEMU_IS_ALIGNED(qiov->size, s->page_size));
1054     assert(qiov->size / s->page_size <= s->page_size / sizeof(uint64_t));
1055     for (i = 0; i < qiov->niov; ++i) {
1056         bool retry = true;
1057         uint64_t iova;
1058         size_t len = QEMU_ALIGN_UP(qiov->iov[i].iov_len,
1059                                    qemu_real_host_page_size());
1060 try_map:
1061         r = qemu_vfio_dma_map(s->vfio,
1062                               qiov->iov[i].iov_base,
1063                               len, true, &iova, errp);
1064         if (r == -ENOSPC) {
1065             /*
1066              * In addition to the -ENOMEM error, the VFIO_IOMMU_MAP_DMA
1067              * ioctl returns -ENOSPC to signal the user exhausted the DMA
1068              * mappings available for a container since Linux kernel commit
1069              * 492855939bdb ("vfio/type1: Limit DMA mappings per container",
1070              * April 2019, see CVE-2019-3882).
1071              *
1072              * This block driver already handles this error path by checking
1073              * for the -ENOMEM error, so we directly replace -ENOSPC by
1074              * -ENOMEM. Beside, -ENOSPC has a specific meaning for blockdev
1075              * coroutines: it triggers BLOCKDEV_ON_ERROR_ENOSPC and
1076              * BLOCK_ERROR_ACTION_STOP which stops the VM, asking the operator
1077              * to add more storage to the blockdev. Not something we can do
1078              * easily with an IOMMU :)
1079              */
1080             r = -ENOMEM;
1081         }
1082         if (r == -ENOMEM && retry) {
1083             /*
1084              * We exhausted the DMA mappings available for our container:
1085              * recycle the volatile IOVA mappings.
1086              */
1087             retry = false;
1088             trace_nvme_dma_flush_queue_wait(s);
1089             if (s->dma_map_count) {
1090                 trace_nvme_dma_map_flush(s);
1091                 qemu_co_queue_wait(&s->dma_flush_queue, &s->dma_map_lock);
1092             } else {
1093                 r = qemu_vfio_dma_reset_temporary(s->vfio);
1094                 if (r) {
1095                     goto fail;
1096                 }
1097             }
1098             errp = &local_err;
1099 
1100             goto try_map;
1101         }
1102         if (r) {
1103             goto fail;
1104         }
1105 
1106         for (j = 0; j < qiov->iov[i].iov_len / s->page_size; j++) {
1107             pagelist[entries++] = cpu_to_le64(iova + j * s->page_size);
1108         }
1109         trace_nvme_cmd_map_qiov_iov(s, i, qiov->iov[i].iov_base,
1110                                     qiov->iov[i].iov_len / s->page_size);
1111     }
1112 
1113     s->dma_map_count += qiov->size;
1114 
1115     assert(entries <= s->page_size / sizeof(uint64_t));
1116     switch (entries) {
1117     case 0:
1118         abort();
1119     case 1:
1120         cmd->dptr.prp1 = pagelist[0];
1121         cmd->dptr.prp2 = 0;
1122         break;
1123     case 2:
1124         cmd->dptr.prp1 = pagelist[0];
1125         cmd->dptr.prp2 = pagelist[1];
1126         break;
1127     default:
1128         cmd->dptr.prp1 = pagelist[0];
1129         cmd->dptr.prp2 = cpu_to_le64(req->prp_list_iova + sizeof(uint64_t));
1130         break;
1131     }
1132     trace_nvme_cmd_map_qiov(s, cmd, req, qiov, entries);
1133     for (i = 0; i < entries; ++i) {
1134         trace_nvme_cmd_map_qiov_pages(s, i, pagelist[i]);
1135     }
1136     return 0;
1137 fail:
1138     /* No need to unmap [0 - i) iovs even if we've failed, since we don't
1139      * increment s->dma_map_count. This is okay for fixed mapping memory areas
1140      * because they are already mapped before calling this function; for
1141      * temporary mappings, a later nvme_cmd_(un)map_qiov will reclaim by
1142      * calling qemu_vfio_dma_reset_temporary when necessary. */
1143     if (local_err) {
1144         error_reportf_err(local_err, "Cannot map buffer for DMA: ");
1145     }
1146     return r;
1147 }
1148 
1149 typedef struct {
1150     Coroutine *co;
1151     int ret;
1152     AioContext *ctx;
1153 } NVMeCoData;
1154 
1155 static void nvme_rw_cb_bh(void *opaque)
1156 {
1157     NVMeCoData *data = opaque;
1158     qemu_coroutine_enter(data->co);
1159 }
1160 
1161 static void nvme_rw_cb(void *opaque, int ret)
1162 {
1163     NVMeCoData *data = opaque;
1164     data->ret = ret;
1165     if (!data->co) {
1166         /* The rw coroutine hasn't yielded, don't try to enter. */
1167         return;
1168     }
1169     replay_bh_schedule_oneshot_event(data->ctx, nvme_rw_cb_bh, data);
1170 }
1171 
1172 static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
1173                                             uint64_t offset, uint64_t bytes,
1174                                             QEMUIOVector *qiov,
1175                                             bool is_write,
1176                                             int flags)
1177 {
1178     int r;
1179     BDRVNVMeState *s = bs->opaque;
1180     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1181     NVMeRequest *req;
1182 
1183     uint32_t cdw12 = (((bytes >> s->blkshift) - 1) & 0xFFFF) |
1184                        (flags & BDRV_REQ_FUA ? 1 << 30 : 0);
1185     NvmeCmd cmd = {
1186         .opcode = is_write ? NVME_CMD_WRITE : NVME_CMD_READ,
1187         .nsid = cpu_to_le32(s->nsid),
1188         .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
1189         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
1190         .cdw12 = cpu_to_le32(cdw12),
1191     };
1192     NVMeCoData data = {
1193         .ctx = bdrv_get_aio_context(bs),
1194         .ret = -EINPROGRESS,
1195     };
1196 
1197     trace_nvme_prw_aligned(s, is_write, offset, bytes, flags, qiov->niov);
1198     assert(s->queue_count > 1);
1199     req = nvme_get_free_req(ioq);
1200     assert(req);
1201 
1202     qemu_co_mutex_lock(&s->dma_map_lock);
1203     r = nvme_cmd_map_qiov(bs, &cmd, req, qiov);
1204     qemu_co_mutex_unlock(&s->dma_map_lock);
1205     if (r) {
1206         nvme_put_free_req_and_wake(ioq, req);
1207         return r;
1208     }
1209     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1210 
1211     data.co = qemu_coroutine_self();
1212     while (data.ret == -EINPROGRESS) {
1213         qemu_coroutine_yield();
1214     }
1215 
1216     qemu_co_mutex_lock(&s->dma_map_lock);
1217     r = nvme_cmd_unmap_qiov(bs, qiov);
1218     qemu_co_mutex_unlock(&s->dma_map_lock);
1219     if (r) {
1220         return r;
1221     }
1222 
1223     trace_nvme_rw_done(s, is_write, offset, bytes, data.ret);
1224     return data.ret;
1225 }
1226 
1227 static inline bool nvme_qiov_aligned(BlockDriverState *bs,
1228                                      const QEMUIOVector *qiov)
1229 {
1230     int i;
1231     BDRVNVMeState *s = bs->opaque;
1232 
1233     for (i = 0; i < qiov->niov; ++i) {
1234         if (!QEMU_PTR_IS_ALIGNED(qiov->iov[i].iov_base,
1235                                  qemu_real_host_page_size()) ||
1236             !QEMU_IS_ALIGNED(qiov->iov[i].iov_len, qemu_real_host_page_size())) {
1237             trace_nvme_qiov_unaligned(qiov, i, qiov->iov[i].iov_base,
1238                                       qiov->iov[i].iov_len, s->page_size);
1239             return false;
1240         }
1241     }
1242     return true;
1243 }
1244 
1245 static coroutine_fn int nvme_co_prw(BlockDriverState *bs,
1246                                     uint64_t offset, uint64_t bytes,
1247                                     QEMUIOVector *qiov, bool is_write,
1248                                     int flags)
1249 {
1250     BDRVNVMeState *s = bs->opaque;
1251     int r;
1252     QEMU_AUTO_VFREE uint8_t *buf = NULL;
1253     QEMUIOVector local_qiov;
1254     size_t len = QEMU_ALIGN_UP(bytes, qemu_real_host_page_size());
1255     assert(QEMU_IS_ALIGNED(offset, s->page_size));
1256     assert(QEMU_IS_ALIGNED(bytes, s->page_size));
1257     assert(bytes <= s->max_transfer);
1258     if (nvme_qiov_aligned(bs, qiov)) {
1259         s->stats.aligned_accesses++;
1260         return nvme_co_prw_aligned(bs, offset, bytes, qiov, is_write, flags);
1261     }
1262     s->stats.unaligned_accesses++;
1263     trace_nvme_prw_buffered(s, offset, bytes, qiov->niov, is_write);
1264     buf = qemu_try_memalign(qemu_real_host_page_size(), len);
1265 
1266     if (!buf) {
1267         return -ENOMEM;
1268     }
1269     qemu_iovec_init(&local_qiov, 1);
1270     if (is_write) {
1271         qemu_iovec_to_buf(qiov, 0, buf, bytes);
1272     }
1273     qemu_iovec_add(&local_qiov, buf, bytes);
1274     r = nvme_co_prw_aligned(bs, offset, bytes, &local_qiov, is_write, flags);
1275     qemu_iovec_destroy(&local_qiov);
1276     if (!r && !is_write) {
1277         qemu_iovec_from_buf(qiov, 0, buf, bytes);
1278     }
1279     return r;
1280 }
1281 
1282 static coroutine_fn int nvme_co_preadv(BlockDriverState *bs,
1283                                        int64_t offset, int64_t bytes,
1284                                        QEMUIOVector *qiov,
1285                                        BdrvRequestFlags flags)
1286 {
1287     return nvme_co_prw(bs, offset, bytes, qiov, false, flags);
1288 }
1289 
1290 static coroutine_fn int nvme_co_pwritev(BlockDriverState *bs,
1291                                         int64_t offset, int64_t bytes,
1292                                         QEMUIOVector *qiov,
1293                                         BdrvRequestFlags flags)
1294 {
1295     return nvme_co_prw(bs, offset, bytes, qiov, true, flags);
1296 }
1297 
1298 static coroutine_fn int nvme_co_flush(BlockDriverState *bs)
1299 {
1300     BDRVNVMeState *s = bs->opaque;
1301     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1302     NVMeRequest *req;
1303     NvmeCmd cmd = {
1304         .opcode = NVME_CMD_FLUSH,
1305         .nsid = cpu_to_le32(s->nsid),
1306     };
1307     NVMeCoData data = {
1308         .ctx = bdrv_get_aio_context(bs),
1309         .ret = -EINPROGRESS,
1310     };
1311 
1312     assert(s->queue_count > 1);
1313     req = nvme_get_free_req(ioq);
1314     assert(req);
1315     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1316 
1317     data.co = qemu_coroutine_self();
1318     if (data.ret == -EINPROGRESS) {
1319         qemu_coroutine_yield();
1320     }
1321 
1322     return data.ret;
1323 }
1324 
1325 
1326 static coroutine_fn int nvme_co_pwrite_zeroes(BlockDriverState *bs,
1327                                               int64_t offset,
1328                                               int64_t bytes,
1329                                               BdrvRequestFlags flags)
1330 {
1331     BDRVNVMeState *s = bs->opaque;
1332     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1333     NVMeRequest *req;
1334     uint32_t cdw12;
1335 
1336     if (!s->supports_write_zeroes) {
1337         return -ENOTSUP;
1338     }
1339 
1340     if (bytes == 0) {
1341         return 0;
1342     }
1343 
1344     cdw12 = ((bytes >> s->blkshift) - 1) & 0xFFFF;
1345     /*
1346      * We should not lose information. pwrite_zeroes_alignment and
1347      * max_pwrite_zeroes guarantees it.
1348      */
1349     assert(((cdw12 + 1) << s->blkshift) == bytes);
1350 
1351     NvmeCmd cmd = {
1352         .opcode = NVME_CMD_WRITE_ZEROES,
1353         .nsid = cpu_to_le32(s->nsid),
1354         .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
1355         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
1356     };
1357 
1358     NVMeCoData data = {
1359         .ctx = bdrv_get_aio_context(bs),
1360         .ret = -EINPROGRESS,
1361     };
1362 
1363     if (flags & BDRV_REQ_MAY_UNMAP) {
1364         cdw12 |= (1 << 25);
1365     }
1366 
1367     if (flags & BDRV_REQ_FUA) {
1368         cdw12 |= (1 << 30);
1369     }
1370 
1371     cmd.cdw12 = cpu_to_le32(cdw12);
1372 
1373     trace_nvme_write_zeroes(s, offset, bytes, flags);
1374     assert(s->queue_count > 1);
1375     req = nvme_get_free_req(ioq);
1376     assert(req);
1377 
1378     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1379 
1380     data.co = qemu_coroutine_self();
1381     while (data.ret == -EINPROGRESS) {
1382         qemu_coroutine_yield();
1383     }
1384 
1385     trace_nvme_rw_done(s, true, offset, bytes, data.ret);
1386     return data.ret;
1387 }
1388 
1389 
1390 static int coroutine_fn nvme_co_pdiscard(BlockDriverState *bs,
1391                                          int64_t offset,
1392                                          int64_t bytes)
1393 {
1394     BDRVNVMeState *s = bs->opaque;
1395     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1396     NVMeRequest *req;
1397     QEMU_AUTO_VFREE NvmeDsmRange *buf = NULL;
1398     QEMUIOVector local_qiov;
1399     int ret;
1400 
1401     NvmeCmd cmd = {
1402         .opcode = NVME_CMD_DSM,
1403         .nsid = cpu_to_le32(s->nsid),
1404         .cdw10 = cpu_to_le32(0), /*number of ranges - 0 based*/
1405         .cdw11 = cpu_to_le32(1 << 2), /*deallocate bit*/
1406     };
1407 
1408     NVMeCoData data = {
1409         .ctx = bdrv_get_aio_context(bs),
1410         .ret = -EINPROGRESS,
1411     };
1412 
1413     if (!s->supports_discard) {
1414         return -ENOTSUP;
1415     }
1416 
1417     assert(s->queue_count > 1);
1418 
1419     /*
1420      * Filling the @buf requires @offset and @bytes to satisfy restrictions
1421      * defined in nvme_refresh_limits().
1422      */
1423     assert(QEMU_IS_ALIGNED(bytes, 1UL << s->blkshift));
1424     assert(QEMU_IS_ALIGNED(offset, 1UL << s->blkshift));
1425     assert((bytes >> s->blkshift) <= UINT32_MAX);
1426 
1427     buf = qemu_try_memalign(s->page_size, s->page_size);
1428     if (!buf) {
1429         return -ENOMEM;
1430     }
1431     memset(buf, 0, s->page_size);
1432     buf->nlb = cpu_to_le32(bytes >> s->blkshift);
1433     buf->slba = cpu_to_le64(offset >> s->blkshift);
1434     buf->cattr = 0;
1435 
1436     qemu_iovec_init(&local_qiov, 1);
1437     qemu_iovec_add(&local_qiov, buf, 4096);
1438 
1439     req = nvme_get_free_req(ioq);
1440     assert(req);
1441 
1442     qemu_co_mutex_lock(&s->dma_map_lock);
1443     ret = nvme_cmd_map_qiov(bs, &cmd, req, &local_qiov);
1444     qemu_co_mutex_unlock(&s->dma_map_lock);
1445 
1446     if (ret) {
1447         nvme_put_free_req_and_wake(ioq, req);
1448         goto out;
1449     }
1450 
1451     trace_nvme_dsm(s, offset, bytes);
1452 
1453     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1454 
1455     data.co = qemu_coroutine_self();
1456     while (data.ret == -EINPROGRESS) {
1457         qemu_coroutine_yield();
1458     }
1459 
1460     qemu_co_mutex_lock(&s->dma_map_lock);
1461     ret = nvme_cmd_unmap_qiov(bs, &local_qiov);
1462     qemu_co_mutex_unlock(&s->dma_map_lock);
1463 
1464     if (ret) {
1465         goto out;
1466     }
1467 
1468     ret = data.ret;
1469     trace_nvme_dsm_done(s, offset, bytes, ret);
1470 out:
1471     qemu_iovec_destroy(&local_qiov);
1472     return ret;
1473 
1474 }
1475 
1476 static int coroutine_fn nvme_co_truncate(BlockDriverState *bs, int64_t offset,
1477                                          bool exact, PreallocMode prealloc,
1478                                          BdrvRequestFlags flags, Error **errp)
1479 {
1480     int64_t cur_length;
1481 
1482     if (prealloc != PREALLOC_MODE_OFF) {
1483         error_setg(errp, "Unsupported preallocation mode '%s'",
1484                    PreallocMode_str(prealloc));
1485         return -ENOTSUP;
1486     }
1487 
1488     cur_length = nvme_getlength(bs);
1489     if (offset != cur_length && exact) {
1490         error_setg(errp, "Cannot resize NVMe devices");
1491         return -ENOTSUP;
1492     } else if (offset > cur_length) {
1493         error_setg(errp, "Cannot grow NVMe devices");
1494         return -EINVAL;
1495     }
1496 
1497     return 0;
1498 }
1499 
1500 static int nvme_reopen_prepare(BDRVReopenState *reopen_state,
1501                                BlockReopenQueue *queue, Error **errp)
1502 {
1503     return 0;
1504 }
1505 
1506 static void nvme_refresh_filename(BlockDriverState *bs)
1507 {
1508     BDRVNVMeState *s = bs->opaque;
1509 
1510     snprintf(bs->exact_filename, sizeof(bs->exact_filename), "nvme://%s/%i",
1511              s->device, s->nsid);
1512 }
1513 
1514 static void nvme_refresh_limits(BlockDriverState *bs, Error **errp)
1515 {
1516     BDRVNVMeState *s = bs->opaque;
1517 
1518     bs->bl.opt_mem_alignment = s->page_size;
1519     bs->bl.request_alignment = s->page_size;
1520     bs->bl.max_transfer = s->max_transfer;
1521 
1522     /*
1523      * Look at nvme_co_pwrite_zeroes: after shift and decrement we should get
1524      * at most 0xFFFF
1525      */
1526     bs->bl.max_pwrite_zeroes = 1ULL << (s->blkshift + 16);
1527     bs->bl.pwrite_zeroes_alignment = MAX(bs->bl.request_alignment,
1528                                          1UL << s->blkshift);
1529 
1530     bs->bl.max_pdiscard = (uint64_t)UINT32_MAX << s->blkshift;
1531     bs->bl.pdiscard_alignment = MAX(bs->bl.request_alignment,
1532                                     1UL << s->blkshift);
1533 }
1534 
1535 static void nvme_detach_aio_context(BlockDriverState *bs)
1536 {
1537     BDRVNVMeState *s = bs->opaque;
1538 
1539     for (unsigned i = 0; i < s->queue_count; i++) {
1540         NVMeQueuePair *q = s->queues[i];
1541 
1542         qemu_bh_delete(q->completion_bh);
1543         q->completion_bh = NULL;
1544     }
1545 
1546     aio_set_event_notifier(bdrv_get_aio_context(bs),
1547                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
1548                            false, NULL, NULL, NULL);
1549 }
1550 
1551 static void nvme_attach_aio_context(BlockDriverState *bs,
1552                                     AioContext *new_context)
1553 {
1554     BDRVNVMeState *s = bs->opaque;
1555 
1556     s->aio_context = new_context;
1557     aio_set_event_notifier(new_context, &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
1558                            false, nvme_handle_event, nvme_poll_cb,
1559                            nvme_poll_ready);
1560 
1561     for (unsigned i = 0; i < s->queue_count; i++) {
1562         NVMeQueuePair *q = s->queues[i];
1563 
1564         q->completion_bh =
1565             aio_bh_new(new_context, nvme_process_completion_bh, q);
1566     }
1567 }
1568 
1569 static void nvme_aio_plug(BlockDriverState *bs)
1570 {
1571     BDRVNVMeState *s = bs->opaque;
1572     assert(!s->plugged);
1573     s->plugged = true;
1574 }
1575 
1576 static void nvme_aio_unplug(BlockDriverState *bs)
1577 {
1578     BDRVNVMeState *s = bs->opaque;
1579     assert(s->plugged);
1580     s->plugged = false;
1581     for (unsigned i = INDEX_IO(0); i < s->queue_count; i++) {
1582         NVMeQueuePair *q = s->queues[i];
1583         qemu_mutex_lock(&q->lock);
1584         nvme_kick(q);
1585         nvme_process_completion(q);
1586         qemu_mutex_unlock(&q->lock);
1587     }
1588 }
1589 
1590 static void nvme_register_buf(BlockDriverState *bs, void *host, size_t size)
1591 {
1592     int ret;
1593     Error *local_err = NULL;
1594     BDRVNVMeState *s = bs->opaque;
1595 
1596     ret = qemu_vfio_dma_map(s->vfio, host, size, false, NULL, &local_err);
1597     if (ret) {
1598         /* FIXME: we may run out of IOVA addresses after repeated
1599          * bdrv_register_buf/bdrv_unregister_buf, because nvme_vfio_dma_unmap
1600          * doesn't reclaim addresses for fixed mappings. */
1601         error_reportf_err(local_err, "nvme_register_buf failed: ");
1602     }
1603 }
1604 
1605 static void nvme_unregister_buf(BlockDriverState *bs, void *host)
1606 {
1607     BDRVNVMeState *s = bs->opaque;
1608 
1609     qemu_vfio_dma_unmap(s->vfio, host);
1610 }
1611 
1612 static BlockStatsSpecific *nvme_get_specific_stats(BlockDriverState *bs)
1613 {
1614     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
1615     BDRVNVMeState *s = bs->opaque;
1616 
1617     stats->driver = BLOCKDEV_DRIVER_NVME;
1618     stats->u.nvme = (BlockStatsSpecificNvme) {
1619         .completion_errors = s->stats.completion_errors,
1620         .aligned_accesses = s->stats.aligned_accesses,
1621         .unaligned_accesses = s->stats.unaligned_accesses,
1622     };
1623 
1624     return stats;
1625 }
1626 
1627 static const char *const nvme_strong_runtime_opts[] = {
1628     NVME_BLOCK_OPT_DEVICE,
1629     NVME_BLOCK_OPT_NAMESPACE,
1630 
1631     NULL
1632 };
1633 
1634 static BlockDriver bdrv_nvme = {
1635     .format_name              = "nvme",
1636     .protocol_name            = "nvme",
1637     .instance_size            = sizeof(BDRVNVMeState),
1638 
1639     .bdrv_co_create_opts      = bdrv_co_create_opts_simple,
1640     .create_opts              = &bdrv_create_opts_simple,
1641 
1642     .bdrv_parse_filename      = nvme_parse_filename,
1643     .bdrv_file_open           = nvme_file_open,
1644     .bdrv_close               = nvme_close,
1645     .bdrv_getlength           = nvme_getlength,
1646     .bdrv_probe_blocksizes    = nvme_probe_blocksizes,
1647     .bdrv_co_truncate         = nvme_co_truncate,
1648 
1649     .bdrv_co_preadv           = nvme_co_preadv,
1650     .bdrv_co_pwritev          = nvme_co_pwritev,
1651 
1652     .bdrv_co_pwrite_zeroes    = nvme_co_pwrite_zeroes,
1653     .bdrv_co_pdiscard         = nvme_co_pdiscard,
1654 
1655     .bdrv_co_flush_to_disk    = nvme_co_flush,
1656     .bdrv_reopen_prepare      = nvme_reopen_prepare,
1657 
1658     .bdrv_refresh_filename    = nvme_refresh_filename,
1659     .bdrv_refresh_limits      = nvme_refresh_limits,
1660     .strong_runtime_opts      = nvme_strong_runtime_opts,
1661     .bdrv_get_specific_stats  = nvme_get_specific_stats,
1662 
1663     .bdrv_detach_aio_context  = nvme_detach_aio_context,
1664     .bdrv_attach_aio_context  = nvme_attach_aio_context,
1665 
1666     .bdrv_io_plug             = nvme_aio_plug,
1667     .bdrv_io_unplug           = nvme_aio_unplug,
1668 
1669     .bdrv_register_buf        = nvme_register_buf,
1670     .bdrv_unregister_buf      = nvme_unregister_buf,
1671 };
1672 
1673 static void bdrv_nvme_init(void)
1674 {
1675     bdrv_register(&bdrv_nvme);
1676 }
1677 
1678 block_init(bdrv_nvme_init);
1679