xref: /openbmc/qemu/block/nvme.c (revision 4e116893)
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, errp);
180     if (r) {
181         error_prepend(errp, "Cannot map queue: ");
182     }
183     return r == 0;
184 }
185 
186 static void nvme_free_queue_pair(NVMeQueuePair *q)
187 {
188     trace_nvme_free_queue_pair(q->index, q);
189     if (q->completion_bh) {
190         qemu_bh_delete(q->completion_bh);
191     }
192     qemu_vfree(q->prp_list_pages);
193     qemu_vfree(q->sq.queue);
194     qemu_vfree(q->cq.queue);
195     qemu_mutex_destroy(&q->lock);
196     g_free(q);
197 }
198 
199 static void nvme_free_req_queue_cb(void *opaque)
200 {
201     NVMeQueuePair *q = opaque;
202 
203     qemu_mutex_lock(&q->lock);
204     while (qemu_co_enter_next(&q->free_req_queue, &q->lock)) {
205         /* Retry all pending requests */
206     }
207     qemu_mutex_unlock(&q->lock);
208 }
209 
210 static NVMeQueuePair *nvme_create_queue_pair(BDRVNVMeState *s,
211                                              AioContext *aio_context,
212                                              unsigned idx, size_t size,
213                                              Error **errp)
214 {
215     int i, r;
216     NVMeQueuePair *q;
217     uint64_t prp_list_iova;
218     size_t bytes;
219 
220     q = g_try_new0(NVMeQueuePair, 1);
221     if (!q) {
222         error_setg(errp, "Cannot allocate queue pair");
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         error_setg(errp, "Cannot allocate PRP page list");
232         goto fail;
233     }
234     memset(q->prp_list_pages, 0, bytes);
235     qemu_mutex_init(&q->lock);
236     q->s = s;
237     q->index = idx;
238     qemu_co_queue_init(&q->free_req_queue);
239     q->completion_bh = aio_bh_new(aio_context, nvme_process_completion_bh, q);
240     r = qemu_vfio_dma_map(s->vfio, q->prp_list_pages, bytes,
241                           false, &prp_list_iova, errp);
242     if (r) {
243         error_prepend(errp, "Cannot map buffer for DMA: ");
244         goto fail;
245     }
246     q->free_req_head = -1;
247     for (i = 0; i < NVME_NUM_REQS; i++) {
248         NVMeRequest *req = &q->reqs[i];
249         req->cid = i + 1;
250         req->free_req_next = q->free_req_head;
251         q->free_req_head = i;
252         req->prp_list_page = q->prp_list_pages + i * s->page_size;
253         req->prp_list_iova = prp_list_iova + i * s->page_size;
254     }
255 
256     if (!nvme_init_queue(s, &q->sq, size, NVME_SQ_ENTRY_BYTES, errp)) {
257         goto fail;
258     }
259     q->sq.doorbell = &s->doorbells[idx * s->doorbell_scale].sq_tail;
260 
261     if (!nvme_init_queue(s, &q->cq, size, NVME_CQ_ENTRY_BYTES, errp)) {
262         goto fail;
263     }
264     q->cq.doorbell = &s->doorbells[idx * s->doorbell_scale].cq_head;
265 
266     return q;
267 fail:
268     nvme_free_queue_pair(q);
269     return NULL;
270 }
271 
272 /* With q->lock */
273 static void nvme_kick(NVMeQueuePair *q)
274 {
275     BDRVNVMeState *s = q->s;
276 
277     if (s->plugged || !q->need_kick) {
278         return;
279     }
280     trace_nvme_kick(s, q->index);
281     assert(!(q->sq.tail & 0xFF00));
282     /* Fence the write to submission queue entry before notifying the device. */
283     smp_wmb();
284     *q->sq.doorbell = cpu_to_le32(q->sq.tail);
285     q->inflight += q->need_kick;
286     q->need_kick = 0;
287 }
288 
289 /* Find a free request element if any, otherwise:
290  * a) if in coroutine context, try to wait for one to become available;
291  * b) if not in coroutine, return NULL;
292  */
293 static NVMeRequest *nvme_get_free_req(NVMeQueuePair *q)
294 {
295     NVMeRequest *req;
296 
297     qemu_mutex_lock(&q->lock);
298 
299     while (q->free_req_head == -1) {
300         if (qemu_in_coroutine()) {
301             trace_nvme_free_req_queue_wait(q->s, q->index);
302             qemu_co_queue_wait(&q->free_req_queue, &q->lock);
303         } else {
304             qemu_mutex_unlock(&q->lock);
305             return NULL;
306         }
307     }
308 
309     req = &q->reqs[q->free_req_head];
310     q->free_req_head = req->free_req_next;
311     req->free_req_next = -1;
312 
313     qemu_mutex_unlock(&q->lock);
314     return req;
315 }
316 
317 /* With q->lock */
318 static void nvme_put_free_req_locked(NVMeQueuePair *q, NVMeRequest *req)
319 {
320     req->free_req_next = q->free_req_head;
321     q->free_req_head = req - q->reqs;
322 }
323 
324 /* With q->lock */
325 static void nvme_wake_free_req_locked(NVMeQueuePair *q)
326 {
327     if (!qemu_co_queue_empty(&q->free_req_queue)) {
328         replay_bh_schedule_oneshot_event(q->s->aio_context,
329                 nvme_free_req_queue_cb, q);
330     }
331 }
332 
333 /* Insert a request in the freelist and wake waiters */
334 static void nvme_put_free_req_and_wake(NVMeQueuePair *q, NVMeRequest *req)
335 {
336     qemu_mutex_lock(&q->lock);
337     nvme_put_free_req_locked(q, req);
338     nvme_wake_free_req_locked(q);
339     qemu_mutex_unlock(&q->lock);
340 }
341 
342 static inline int nvme_translate_error(const NvmeCqe *c)
343 {
344     uint16_t status = (le16_to_cpu(c->status) >> 1) & 0xFF;
345     if (status) {
346         trace_nvme_error(le32_to_cpu(c->result),
347                          le16_to_cpu(c->sq_head),
348                          le16_to_cpu(c->sq_id),
349                          le16_to_cpu(c->cid),
350                          le16_to_cpu(status));
351     }
352     switch (status) {
353     case 0:
354         return 0;
355     case 1:
356         return -ENOSYS;
357     case 2:
358         return -EINVAL;
359     default:
360         return -EIO;
361     }
362 }
363 
364 /* With q->lock */
365 static bool nvme_process_completion(NVMeQueuePair *q)
366 {
367     BDRVNVMeState *s = q->s;
368     bool progress = false;
369     NVMeRequest *preq;
370     NVMeRequest req;
371     NvmeCqe *c;
372 
373     trace_nvme_process_completion(s, q->index, q->inflight);
374     if (s->plugged) {
375         trace_nvme_process_completion_queue_plugged(s, q->index);
376         return false;
377     }
378 
379     /*
380      * Support re-entrancy when a request cb() function invokes aio_poll().
381      * Pending completions must be visible to aio_poll() so that a cb()
382      * function can wait for the completion of another request.
383      *
384      * The aio_poll() loop will execute our BH and we'll resume completion
385      * processing there.
386      */
387     qemu_bh_schedule(q->completion_bh);
388 
389     assert(q->inflight >= 0);
390     while (q->inflight) {
391         int ret;
392         int16_t cid;
393 
394         c = (NvmeCqe *)&q->cq.queue[q->cq.head * NVME_CQ_ENTRY_BYTES];
395         if ((le16_to_cpu(c->status) & 0x1) == q->cq_phase) {
396             break;
397         }
398         ret = nvme_translate_error(c);
399         if (ret) {
400             s->stats.completion_errors++;
401         }
402         q->cq.head = (q->cq.head + 1) % NVME_QUEUE_SIZE;
403         if (!q->cq.head) {
404             q->cq_phase = !q->cq_phase;
405         }
406         cid = le16_to_cpu(c->cid);
407         if (cid == 0 || cid > NVME_QUEUE_SIZE) {
408             warn_report("NVMe: Unexpected CID in completion queue: %"PRIu32", "
409                         "queue size: %u", cid, NVME_QUEUE_SIZE);
410             continue;
411         }
412         trace_nvme_complete_command(s, q->index, cid);
413         preq = &q->reqs[cid - 1];
414         req = *preq;
415         assert(req.cid == cid);
416         assert(req.cb);
417         nvme_put_free_req_locked(q, preq);
418         preq->cb = preq->opaque = NULL;
419         q->inflight--;
420         qemu_mutex_unlock(&q->lock);
421         req.cb(req.opaque, ret);
422         qemu_mutex_lock(&q->lock);
423         progress = true;
424     }
425     if (progress) {
426         /* Notify the device so it can post more completions. */
427         smp_mb_release();
428         *q->cq.doorbell = cpu_to_le32(q->cq.head);
429         nvme_wake_free_req_locked(q);
430     }
431 
432     qemu_bh_cancel(q->completion_bh);
433 
434     return progress;
435 }
436 
437 static void nvme_process_completion_bh(void *opaque)
438 {
439     NVMeQueuePair *q = opaque;
440 
441     /*
442      * We're being invoked because a nvme_process_completion() cb() function
443      * called aio_poll(). The callback may be waiting for further completions
444      * so notify the device that it has space to fill in more completions now.
445      */
446     smp_mb_release();
447     *q->cq.doorbell = cpu_to_le32(q->cq.head);
448     nvme_wake_free_req_locked(q);
449 
450     nvme_process_completion(q);
451 }
452 
453 static void nvme_trace_command(const NvmeCmd *cmd)
454 {
455     int i;
456 
457     if (!trace_event_get_state_backends(TRACE_NVME_SUBMIT_COMMAND_RAW)) {
458         return;
459     }
460     for (i = 0; i < 8; ++i) {
461         uint8_t *cmdp = (uint8_t *)cmd + i * 8;
462         trace_nvme_submit_command_raw(cmdp[0], cmdp[1], cmdp[2], cmdp[3],
463                                       cmdp[4], cmdp[5], cmdp[6], cmdp[7]);
464     }
465 }
466 
467 static void nvme_submit_command(NVMeQueuePair *q, NVMeRequest *req,
468                                 NvmeCmd *cmd, BlockCompletionFunc cb,
469                                 void *opaque)
470 {
471     assert(!req->cb);
472     req->cb = cb;
473     req->opaque = opaque;
474     cmd->cid = cpu_to_le16(req->cid);
475 
476     trace_nvme_submit_command(q->s, q->index, req->cid);
477     nvme_trace_command(cmd);
478     qemu_mutex_lock(&q->lock);
479     memcpy((uint8_t *)q->sq.queue +
480            q->sq.tail * NVME_SQ_ENTRY_BYTES, cmd, sizeof(*cmd));
481     q->sq.tail = (q->sq.tail + 1) % NVME_QUEUE_SIZE;
482     q->need_kick++;
483     nvme_kick(q);
484     nvme_process_completion(q);
485     qemu_mutex_unlock(&q->lock);
486 }
487 
488 static void nvme_admin_cmd_sync_cb(void *opaque, int ret)
489 {
490     int *pret = opaque;
491     *pret = ret;
492     aio_wait_kick();
493 }
494 
495 static int nvme_admin_cmd_sync(BlockDriverState *bs, NvmeCmd *cmd)
496 {
497     BDRVNVMeState *s = bs->opaque;
498     NVMeQueuePair *q = s->queues[INDEX_ADMIN];
499     AioContext *aio_context = bdrv_get_aio_context(bs);
500     NVMeRequest *req;
501     int ret = -EINPROGRESS;
502     req = nvme_get_free_req(q);
503     if (!req) {
504         return -EBUSY;
505     }
506     nvme_submit_command(q, req, cmd, nvme_admin_cmd_sync_cb, &ret);
507 
508     AIO_WAIT_WHILE(aio_context, ret == -EINPROGRESS);
509     return ret;
510 }
511 
512 /* Returns true on success, false on failure. */
513 static bool nvme_identify(BlockDriverState *bs, int namespace, Error **errp)
514 {
515     BDRVNVMeState *s = bs->opaque;
516     bool ret = false;
517     union {
518         NvmeIdCtrl ctrl;
519         NvmeIdNs ns;
520     } *id;
521     NvmeLBAF *lbaf;
522     uint16_t oncs;
523     int r;
524     uint64_t iova;
525     NvmeCmd cmd = {
526         .opcode = NVME_ADM_CMD_IDENTIFY,
527         .cdw10 = cpu_to_le32(0x1),
528     };
529     size_t id_size = QEMU_ALIGN_UP(sizeof(*id), qemu_real_host_page_size);
530 
531     id = qemu_try_memalign(qemu_real_host_page_size, id_size);
532     if (!id) {
533         error_setg(errp, "Cannot allocate buffer for identify response");
534         goto out;
535     }
536     r = qemu_vfio_dma_map(s->vfio, id, id_size, true, &iova, errp);
537     if (r) {
538         error_prepend(errp, "Cannot map buffer for DMA: ");
539         goto out;
540     }
541 
542     memset(id, 0, id_size);
543     cmd.dptr.prp1 = cpu_to_le64(iova);
544     if (nvme_admin_cmd_sync(bs, &cmd)) {
545         error_setg(errp, "Failed to identify controller");
546         goto out;
547     }
548 
549     if (le32_to_cpu(id->ctrl.nn) < namespace) {
550         error_setg(errp, "Invalid namespace");
551         goto out;
552     }
553     s->write_cache_supported = le32_to_cpu(id->ctrl.vwc) & 0x1;
554     s->max_transfer = (id->ctrl.mdts ? 1 << id->ctrl.mdts : 0) * s->page_size;
555     /* For now the page list buffer per command is one page, to hold at most
556      * s->page_size / sizeof(uint64_t) entries. */
557     s->max_transfer = MIN_NON_ZERO(s->max_transfer,
558                           s->page_size / sizeof(uint64_t) * s->page_size);
559 
560     oncs = le16_to_cpu(id->ctrl.oncs);
561     s->supports_write_zeroes = !!(oncs & NVME_ONCS_WRITE_ZEROES);
562     s->supports_discard = !!(oncs & NVME_ONCS_DSM);
563 
564     memset(id, 0, id_size);
565     cmd.cdw10 = 0;
566     cmd.nsid = cpu_to_le32(namespace);
567     if (nvme_admin_cmd_sync(bs, &cmd)) {
568         error_setg(errp, "Failed to identify namespace");
569         goto out;
570     }
571 
572     s->nsze = le64_to_cpu(id->ns.nsze);
573     lbaf = &id->ns.lbaf[NVME_ID_NS_FLBAS_INDEX(id->ns.flbas)];
574 
575     if (NVME_ID_NS_DLFEAT_WRITE_ZEROES(id->ns.dlfeat) &&
576             NVME_ID_NS_DLFEAT_READ_BEHAVIOR(id->ns.dlfeat) ==
577                     NVME_ID_NS_DLFEAT_READ_BEHAVIOR_ZEROES) {
578         bs->supported_write_flags |= BDRV_REQ_MAY_UNMAP;
579     }
580 
581     if (lbaf->ms) {
582         error_setg(errp, "Namespaces with metadata are not yet supported");
583         goto out;
584     }
585 
586     if (lbaf->ds < BDRV_SECTOR_BITS || lbaf->ds > 12 ||
587         (1 << lbaf->ds) > s->page_size)
588     {
589         error_setg(errp, "Namespace has unsupported block size (2^%d)",
590                    lbaf->ds);
591         goto out;
592     }
593 
594     ret = true;
595     s->blkshift = lbaf->ds;
596 out:
597     qemu_vfio_dma_unmap(s->vfio, id);
598     qemu_vfree(id);
599 
600     return ret;
601 }
602 
603 static bool nvme_poll_queue(NVMeQueuePair *q)
604 {
605     bool progress = false;
606 
607     const size_t cqe_offset = q->cq.head * NVME_CQ_ENTRY_BYTES;
608     NvmeCqe *cqe = (NvmeCqe *)&q->cq.queue[cqe_offset];
609 
610     trace_nvme_poll_queue(q->s, q->index);
611     /*
612      * Do an early check for completions. q->lock isn't needed because
613      * nvme_process_completion() only runs in the event loop thread and
614      * cannot race with itself.
615      */
616     if ((le16_to_cpu(cqe->status) & 0x1) == q->cq_phase) {
617         return false;
618     }
619 
620     qemu_mutex_lock(&q->lock);
621     while (nvme_process_completion(q)) {
622         /* Keep polling */
623         progress = true;
624     }
625     qemu_mutex_unlock(&q->lock);
626 
627     return progress;
628 }
629 
630 static bool nvme_poll_queues(BDRVNVMeState *s)
631 {
632     bool progress = false;
633     int i;
634 
635     for (i = 0; i < s->queue_count; i++) {
636         if (nvme_poll_queue(s->queues[i])) {
637             progress = true;
638         }
639     }
640     return progress;
641 }
642 
643 static void nvme_handle_event(EventNotifier *n)
644 {
645     BDRVNVMeState *s = container_of(n, BDRVNVMeState,
646                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
647 
648     trace_nvme_handle_event(s);
649     event_notifier_test_and_clear(n);
650     nvme_poll_queues(s);
651 }
652 
653 static bool nvme_add_io_queue(BlockDriverState *bs, Error **errp)
654 {
655     BDRVNVMeState *s = bs->opaque;
656     unsigned n = s->queue_count;
657     NVMeQueuePair *q;
658     NvmeCmd cmd;
659     unsigned queue_size = NVME_QUEUE_SIZE;
660 
661     assert(n <= UINT16_MAX);
662     q = nvme_create_queue_pair(s, bdrv_get_aio_context(bs),
663                                n, queue_size, errp);
664     if (!q) {
665         return false;
666     }
667     cmd = (NvmeCmd) {
668         .opcode = NVME_ADM_CMD_CREATE_CQ,
669         .dptr.prp1 = cpu_to_le64(q->cq.iova),
670         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | n),
671         .cdw11 = cpu_to_le32(NVME_CQ_IEN | NVME_CQ_PC),
672     };
673     if (nvme_admin_cmd_sync(bs, &cmd)) {
674         error_setg(errp, "Failed to create CQ io queue [%u]", n);
675         goto out_error;
676     }
677     cmd = (NvmeCmd) {
678         .opcode = NVME_ADM_CMD_CREATE_SQ,
679         .dptr.prp1 = cpu_to_le64(q->sq.iova),
680         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | n),
681         .cdw11 = cpu_to_le32(NVME_SQ_PC | (n << 16)),
682     };
683     if (nvme_admin_cmd_sync(bs, &cmd)) {
684         error_setg(errp, "Failed to create SQ io queue [%u]", n);
685         goto out_error;
686     }
687     s->queues = g_renew(NVMeQueuePair *, s->queues, n + 1);
688     s->queues[n] = q;
689     s->queue_count++;
690     return true;
691 out_error:
692     nvme_free_queue_pair(q);
693     return false;
694 }
695 
696 static bool nvme_poll_cb(void *opaque)
697 {
698     EventNotifier *e = opaque;
699     BDRVNVMeState *s = container_of(e, BDRVNVMeState,
700                                     irq_notifier[MSIX_SHARED_IRQ_IDX]);
701 
702     return nvme_poll_queues(s);
703 }
704 
705 static int nvme_init(BlockDriverState *bs, const char *device, int namespace,
706                      Error **errp)
707 {
708     BDRVNVMeState *s = bs->opaque;
709     NVMeQueuePair *q;
710     AioContext *aio_context = bdrv_get_aio_context(bs);
711     int ret;
712     uint64_t cap;
713     uint32_t ver;
714     uint64_t timeout_ms;
715     uint64_t deadline, now;
716     volatile NvmeBar *regs = NULL;
717 
718     qemu_co_mutex_init(&s->dma_map_lock);
719     qemu_co_queue_init(&s->dma_flush_queue);
720     s->device = g_strdup(device);
721     s->nsid = namespace;
722     s->aio_context = bdrv_get_aio_context(bs);
723     ret = event_notifier_init(&s->irq_notifier[MSIX_SHARED_IRQ_IDX], 0);
724     if (ret) {
725         error_setg(errp, "Failed to init event notifier");
726         return ret;
727     }
728 
729     s->vfio = qemu_vfio_open_pci(device, errp);
730     if (!s->vfio) {
731         ret = -EINVAL;
732         goto out;
733     }
734 
735     regs = qemu_vfio_pci_map_bar(s->vfio, 0, 0, sizeof(NvmeBar),
736                                  PROT_READ | PROT_WRITE, errp);
737     if (!regs) {
738         ret = -EINVAL;
739         goto out;
740     }
741     /* Perform initialize sequence as described in NVMe spec "7.6.1
742      * Initialization". */
743 
744     cap = le64_to_cpu(regs->cap);
745     trace_nvme_controller_capability_raw(cap);
746     trace_nvme_controller_capability("Maximum Queue Entries Supported",
747                                      1 + NVME_CAP_MQES(cap));
748     trace_nvme_controller_capability("Contiguous Queues Required",
749                                      NVME_CAP_CQR(cap));
750     trace_nvme_controller_capability("Doorbell Stride",
751                                      1 << (2 + NVME_CAP_DSTRD(cap)));
752     trace_nvme_controller_capability("Subsystem Reset Supported",
753                                      NVME_CAP_NSSRS(cap));
754     trace_nvme_controller_capability("Memory Page Size Minimum",
755                                      1 << (12 + NVME_CAP_MPSMIN(cap)));
756     trace_nvme_controller_capability("Memory Page Size Maximum",
757                                      1 << (12 + NVME_CAP_MPSMAX(cap)));
758     if (!NVME_CAP_CSS(cap)) {
759         error_setg(errp, "Device doesn't support NVMe command set");
760         ret = -EINVAL;
761         goto out;
762     }
763 
764     s->page_size = 1u << (12 + NVME_CAP_MPSMIN(cap));
765     s->doorbell_scale = (4 << NVME_CAP_DSTRD(cap)) / sizeof(uint32_t);
766     bs->bl.opt_mem_alignment = s->page_size;
767     bs->bl.request_alignment = s->page_size;
768     timeout_ms = MIN(500 * NVME_CAP_TO(cap), 30000);
769 
770     ver = le32_to_cpu(regs->vs);
771     trace_nvme_controller_spec_version(extract32(ver, 16, 16),
772                                        extract32(ver, 8, 8),
773                                        extract32(ver, 0, 8));
774 
775     /* Reset device to get a clean state. */
776     regs->cc = cpu_to_le32(le32_to_cpu(regs->cc) & 0xFE);
777     /* Wait for CSTS.RDY = 0. */
778     deadline = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) + timeout_ms * SCALE_MS;
779     while (NVME_CSTS_RDY(le32_to_cpu(regs->csts))) {
780         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
781             error_setg(errp, "Timeout while waiting for device to reset (%"
782                              PRId64 " ms)",
783                        timeout_ms);
784             ret = -ETIMEDOUT;
785             goto out;
786         }
787     }
788 
789     s->bar0_wo_map = qemu_vfio_pci_map_bar(s->vfio, 0, 0,
790                                            sizeof(NvmeBar) + NVME_DOORBELL_SIZE,
791                                            PROT_WRITE, errp);
792     s->doorbells = (void *)((uintptr_t)s->bar0_wo_map + sizeof(NvmeBar));
793     if (!s->doorbells) {
794         ret = -EINVAL;
795         goto out;
796     }
797 
798     /* Set up admin queue. */
799     s->queues = g_new(NVMeQueuePair *, 1);
800     q = nvme_create_queue_pair(s, aio_context, 0, NVME_QUEUE_SIZE, errp);
801     if (!q) {
802         ret = -EINVAL;
803         goto out;
804     }
805     s->queues[INDEX_ADMIN] = q;
806     s->queue_count = 1;
807     QEMU_BUILD_BUG_ON((NVME_QUEUE_SIZE - 1) & 0xF000);
808     regs->aqa = cpu_to_le32(((NVME_QUEUE_SIZE - 1) << AQA_ACQS_SHIFT) |
809                             ((NVME_QUEUE_SIZE - 1) << AQA_ASQS_SHIFT));
810     regs->asq = cpu_to_le64(q->sq.iova);
811     regs->acq = cpu_to_le64(q->cq.iova);
812 
813     /* After setting up all control registers we can enable device now. */
814     regs->cc = cpu_to_le32((ctz32(NVME_CQ_ENTRY_BYTES) << CC_IOCQES_SHIFT) |
815                            (ctz32(NVME_SQ_ENTRY_BYTES) << CC_IOSQES_SHIFT) |
816                            CC_EN_MASK);
817     /* Wait for CSTS.RDY = 1. */
818     now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
819     deadline = now + timeout_ms * SCALE_MS;
820     while (!NVME_CSTS_RDY(le32_to_cpu(regs->csts))) {
821         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
822             error_setg(errp, "Timeout while waiting for device to start (%"
823                              PRId64 " ms)",
824                        timeout_ms);
825             ret = -ETIMEDOUT;
826             goto out;
827         }
828     }
829 
830     ret = qemu_vfio_pci_init_irq(s->vfio, s->irq_notifier,
831                                  VFIO_PCI_MSIX_IRQ_INDEX, errp);
832     if (ret) {
833         goto out;
834     }
835     aio_set_event_notifier(bdrv_get_aio_context(bs),
836                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
837                            false, nvme_handle_event, nvme_poll_cb);
838 
839     if (!nvme_identify(bs, namespace, errp)) {
840         ret = -EIO;
841         goto out;
842     }
843 
844     /* Set up command queues. */
845     if (!nvme_add_io_queue(bs, errp)) {
846         ret = -EIO;
847     }
848 out:
849     if (regs) {
850         qemu_vfio_pci_unmap_bar(s->vfio, 0, (void *)regs, 0, sizeof(NvmeBar));
851     }
852 
853     /* Cleaning up is done in nvme_file_open() upon error. */
854     return ret;
855 }
856 
857 /* Parse a filename in the format of nvme://XXXX:XX:XX.X/X. Example:
858  *
859  *     nvme://0000:44:00.0/1
860  *
861  * where the "nvme://" is a fixed form of the protocol prefix, the middle part
862  * is the PCI address, and the last part is the namespace number starting from
863  * 1 according to the NVMe spec. */
864 static void nvme_parse_filename(const char *filename, QDict *options,
865                                 Error **errp)
866 {
867     int pref = strlen("nvme://");
868 
869     if (strlen(filename) > pref && !strncmp(filename, "nvme://", pref)) {
870         const char *tmp = filename + pref;
871         char *device;
872         const char *namespace;
873         unsigned long ns;
874         const char *slash = strchr(tmp, '/');
875         if (!slash) {
876             qdict_put_str(options, NVME_BLOCK_OPT_DEVICE, tmp);
877             return;
878         }
879         device = g_strndup(tmp, slash - tmp);
880         qdict_put_str(options, NVME_BLOCK_OPT_DEVICE, device);
881         g_free(device);
882         namespace = slash + 1;
883         if (*namespace && qemu_strtoul(namespace, NULL, 10, &ns)) {
884             error_setg(errp, "Invalid namespace '%s', positive number expected",
885                        namespace);
886             return;
887         }
888         qdict_put_str(options, NVME_BLOCK_OPT_NAMESPACE,
889                       *namespace ? namespace : "1");
890     }
891 }
892 
893 static int nvme_enable_disable_write_cache(BlockDriverState *bs, bool enable,
894                                            Error **errp)
895 {
896     int ret;
897     BDRVNVMeState *s = bs->opaque;
898     NvmeCmd cmd = {
899         .opcode = NVME_ADM_CMD_SET_FEATURES,
900         .nsid = cpu_to_le32(s->nsid),
901         .cdw10 = cpu_to_le32(0x06),
902         .cdw11 = cpu_to_le32(enable ? 0x01 : 0x00),
903     };
904 
905     ret = nvme_admin_cmd_sync(bs, &cmd);
906     if (ret) {
907         error_setg(errp, "Failed to configure NVMe write cache");
908     }
909     return ret;
910 }
911 
912 static void nvme_close(BlockDriverState *bs)
913 {
914     BDRVNVMeState *s = bs->opaque;
915 
916     for (unsigned i = 0; i < s->queue_count; ++i) {
917         nvme_free_queue_pair(s->queues[i]);
918     }
919     g_free(s->queues);
920     aio_set_event_notifier(bdrv_get_aio_context(bs),
921                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
922                            false, NULL, NULL);
923     event_notifier_cleanup(&s->irq_notifier[MSIX_SHARED_IRQ_IDX]);
924     qemu_vfio_pci_unmap_bar(s->vfio, 0, s->bar0_wo_map,
925                             0, sizeof(NvmeBar) + NVME_DOORBELL_SIZE);
926     qemu_vfio_close(s->vfio);
927 
928     g_free(s->device);
929 }
930 
931 static int nvme_file_open(BlockDriverState *bs, QDict *options, int flags,
932                           Error **errp)
933 {
934     const char *device;
935     QemuOpts *opts;
936     int namespace;
937     int ret;
938     BDRVNVMeState *s = bs->opaque;
939 
940     bs->supported_write_flags = BDRV_REQ_FUA;
941 
942     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
943     qemu_opts_absorb_qdict(opts, options, &error_abort);
944     device = qemu_opt_get(opts, NVME_BLOCK_OPT_DEVICE);
945     if (!device) {
946         error_setg(errp, "'" NVME_BLOCK_OPT_DEVICE "' option is required");
947         qemu_opts_del(opts);
948         return -EINVAL;
949     }
950 
951     namespace = qemu_opt_get_number(opts, NVME_BLOCK_OPT_NAMESPACE, 1);
952     ret = nvme_init(bs, device, namespace, errp);
953     qemu_opts_del(opts);
954     if (ret) {
955         goto fail;
956     }
957     if (flags & BDRV_O_NOCACHE) {
958         if (!s->write_cache_supported) {
959             error_setg(errp,
960                        "NVMe controller doesn't support write cache configuration");
961             ret = -EINVAL;
962         } else {
963             ret = nvme_enable_disable_write_cache(bs, !(flags & BDRV_O_NOCACHE),
964                                                   errp);
965         }
966         if (ret) {
967             goto fail;
968         }
969     }
970     return 0;
971 fail:
972     nvme_close(bs);
973     return ret;
974 }
975 
976 static int64_t nvme_getlength(BlockDriverState *bs)
977 {
978     BDRVNVMeState *s = bs->opaque;
979     return s->nsze << s->blkshift;
980 }
981 
982 static uint32_t nvme_get_blocksize(BlockDriverState *bs)
983 {
984     BDRVNVMeState *s = bs->opaque;
985     assert(s->blkshift >= BDRV_SECTOR_BITS && s->blkshift <= 12);
986     return UINT32_C(1) << s->blkshift;
987 }
988 
989 static int nvme_probe_blocksizes(BlockDriverState *bs, BlockSizes *bsz)
990 {
991     uint32_t blocksize = nvme_get_blocksize(bs);
992     bsz->phys = blocksize;
993     bsz->log = blocksize;
994     return 0;
995 }
996 
997 /* Called with s->dma_map_lock */
998 static coroutine_fn int nvme_cmd_unmap_qiov(BlockDriverState *bs,
999                                             QEMUIOVector *qiov)
1000 {
1001     int r = 0;
1002     BDRVNVMeState *s = bs->opaque;
1003 
1004     s->dma_map_count -= qiov->size;
1005     if (!s->dma_map_count && !qemu_co_queue_empty(&s->dma_flush_queue)) {
1006         r = qemu_vfio_dma_reset_temporary(s->vfio);
1007         if (!r) {
1008             qemu_co_queue_restart_all(&s->dma_flush_queue);
1009         }
1010     }
1011     return r;
1012 }
1013 
1014 /* Called with s->dma_map_lock */
1015 static coroutine_fn int nvme_cmd_map_qiov(BlockDriverState *bs, NvmeCmd *cmd,
1016                                           NVMeRequest *req, QEMUIOVector *qiov)
1017 {
1018     BDRVNVMeState *s = bs->opaque;
1019     uint64_t *pagelist = req->prp_list_page;
1020     int i, j, r;
1021     int entries = 0;
1022     Error *local_err = NULL, **errp = NULL;
1023 
1024     assert(qiov->size);
1025     assert(QEMU_IS_ALIGNED(qiov->size, s->page_size));
1026     assert(qiov->size / s->page_size <= s->page_size / sizeof(uint64_t));
1027     for (i = 0; i < qiov->niov; ++i) {
1028         bool retry = true;
1029         uint64_t iova;
1030         size_t len = QEMU_ALIGN_UP(qiov->iov[i].iov_len,
1031                                    qemu_real_host_page_size);
1032 try_map:
1033         r = qemu_vfio_dma_map(s->vfio,
1034                               qiov->iov[i].iov_base,
1035                               len, true, &iova, errp);
1036         if (r == -ENOSPC) {
1037             /*
1038              * In addition to the -ENOMEM error, the VFIO_IOMMU_MAP_DMA
1039              * ioctl returns -ENOSPC to signal the user exhausted the DMA
1040              * mappings available for a container since Linux kernel commit
1041              * 492855939bdb ("vfio/type1: Limit DMA mappings per container",
1042              * April 2019, see CVE-2019-3882).
1043              *
1044              * This block driver already handles this error path by checking
1045              * for the -ENOMEM error, so we directly replace -ENOSPC by
1046              * -ENOMEM. Beside, -ENOSPC has a specific meaning for blockdev
1047              * coroutines: it triggers BLOCKDEV_ON_ERROR_ENOSPC and
1048              * BLOCK_ERROR_ACTION_STOP which stops the VM, asking the operator
1049              * to add more storage to the blockdev. Not something we can do
1050              * easily with an IOMMU :)
1051              */
1052             r = -ENOMEM;
1053         }
1054         if (r == -ENOMEM && retry) {
1055             /*
1056              * We exhausted the DMA mappings available for our container:
1057              * recycle the volatile IOVA mappings.
1058              */
1059             retry = false;
1060             trace_nvme_dma_flush_queue_wait(s);
1061             if (s->dma_map_count) {
1062                 trace_nvme_dma_map_flush(s);
1063                 qemu_co_queue_wait(&s->dma_flush_queue, &s->dma_map_lock);
1064             } else {
1065                 r = qemu_vfio_dma_reset_temporary(s->vfio);
1066                 if (r) {
1067                     goto fail;
1068                 }
1069             }
1070             errp = &local_err;
1071 
1072             goto try_map;
1073         }
1074         if (r) {
1075             goto fail;
1076         }
1077 
1078         for (j = 0; j < qiov->iov[i].iov_len / s->page_size; j++) {
1079             pagelist[entries++] = cpu_to_le64(iova + j * s->page_size);
1080         }
1081         trace_nvme_cmd_map_qiov_iov(s, i, qiov->iov[i].iov_base,
1082                                     qiov->iov[i].iov_len / s->page_size);
1083     }
1084 
1085     s->dma_map_count += qiov->size;
1086 
1087     assert(entries <= s->page_size / sizeof(uint64_t));
1088     switch (entries) {
1089     case 0:
1090         abort();
1091     case 1:
1092         cmd->dptr.prp1 = pagelist[0];
1093         cmd->dptr.prp2 = 0;
1094         break;
1095     case 2:
1096         cmd->dptr.prp1 = pagelist[0];
1097         cmd->dptr.prp2 = pagelist[1];
1098         break;
1099     default:
1100         cmd->dptr.prp1 = pagelist[0];
1101         cmd->dptr.prp2 = cpu_to_le64(req->prp_list_iova + sizeof(uint64_t));
1102         break;
1103     }
1104     trace_nvme_cmd_map_qiov(s, cmd, req, qiov, entries);
1105     for (i = 0; i < entries; ++i) {
1106         trace_nvme_cmd_map_qiov_pages(s, i, pagelist[i]);
1107     }
1108     return 0;
1109 fail:
1110     /* No need to unmap [0 - i) iovs even if we've failed, since we don't
1111      * increment s->dma_map_count. This is okay for fixed mapping memory areas
1112      * because they are already mapped before calling this function; for
1113      * temporary mappings, a later nvme_cmd_(un)map_qiov will reclaim by
1114      * calling qemu_vfio_dma_reset_temporary when necessary. */
1115     if (local_err) {
1116         error_reportf_err(local_err, "Cannot map buffer for DMA: ");
1117     }
1118     return r;
1119 }
1120 
1121 typedef struct {
1122     Coroutine *co;
1123     int ret;
1124     AioContext *ctx;
1125 } NVMeCoData;
1126 
1127 static void nvme_rw_cb_bh(void *opaque)
1128 {
1129     NVMeCoData *data = opaque;
1130     qemu_coroutine_enter(data->co);
1131 }
1132 
1133 static void nvme_rw_cb(void *opaque, int ret)
1134 {
1135     NVMeCoData *data = opaque;
1136     data->ret = ret;
1137     if (!data->co) {
1138         /* The rw coroutine hasn't yielded, don't try to enter. */
1139         return;
1140     }
1141     replay_bh_schedule_oneshot_event(data->ctx, nvme_rw_cb_bh, data);
1142 }
1143 
1144 static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
1145                                             uint64_t offset, uint64_t bytes,
1146                                             QEMUIOVector *qiov,
1147                                             bool is_write,
1148                                             int flags)
1149 {
1150     int r;
1151     BDRVNVMeState *s = bs->opaque;
1152     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1153     NVMeRequest *req;
1154 
1155     uint32_t cdw12 = (((bytes >> s->blkshift) - 1) & 0xFFFF) |
1156                        (flags & BDRV_REQ_FUA ? 1 << 30 : 0);
1157     NvmeCmd cmd = {
1158         .opcode = is_write ? NVME_CMD_WRITE : NVME_CMD_READ,
1159         .nsid = cpu_to_le32(s->nsid),
1160         .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
1161         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
1162         .cdw12 = cpu_to_le32(cdw12),
1163     };
1164     NVMeCoData data = {
1165         .ctx = bdrv_get_aio_context(bs),
1166         .ret = -EINPROGRESS,
1167     };
1168 
1169     trace_nvme_prw_aligned(s, is_write, offset, bytes, flags, qiov->niov);
1170     assert(s->queue_count > 1);
1171     req = nvme_get_free_req(ioq);
1172     assert(req);
1173 
1174     qemu_co_mutex_lock(&s->dma_map_lock);
1175     r = nvme_cmd_map_qiov(bs, &cmd, req, qiov);
1176     qemu_co_mutex_unlock(&s->dma_map_lock);
1177     if (r) {
1178         nvme_put_free_req_and_wake(ioq, req);
1179         return r;
1180     }
1181     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1182 
1183     data.co = qemu_coroutine_self();
1184     while (data.ret == -EINPROGRESS) {
1185         qemu_coroutine_yield();
1186     }
1187 
1188     qemu_co_mutex_lock(&s->dma_map_lock);
1189     r = nvme_cmd_unmap_qiov(bs, qiov);
1190     qemu_co_mutex_unlock(&s->dma_map_lock);
1191     if (r) {
1192         return r;
1193     }
1194 
1195     trace_nvme_rw_done(s, is_write, offset, bytes, data.ret);
1196     return data.ret;
1197 }
1198 
1199 static inline bool nvme_qiov_aligned(BlockDriverState *bs,
1200                                      const QEMUIOVector *qiov)
1201 {
1202     int i;
1203     BDRVNVMeState *s = bs->opaque;
1204 
1205     for (i = 0; i < qiov->niov; ++i) {
1206         if (!QEMU_PTR_IS_ALIGNED(qiov->iov[i].iov_base,
1207                                  qemu_real_host_page_size) ||
1208             !QEMU_IS_ALIGNED(qiov->iov[i].iov_len, qemu_real_host_page_size)) {
1209             trace_nvme_qiov_unaligned(qiov, i, qiov->iov[i].iov_base,
1210                                       qiov->iov[i].iov_len, s->page_size);
1211             return false;
1212         }
1213     }
1214     return true;
1215 }
1216 
1217 static int nvme_co_prw(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
1218                        QEMUIOVector *qiov, bool is_write, int flags)
1219 {
1220     BDRVNVMeState *s = bs->opaque;
1221     int r;
1222     uint8_t *buf = NULL;
1223     QEMUIOVector local_qiov;
1224     size_t len = QEMU_ALIGN_UP(bytes, qemu_real_host_page_size);
1225     assert(QEMU_IS_ALIGNED(offset, s->page_size));
1226     assert(QEMU_IS_ALIGNED(bytes, s->page_size));
1227     assert(bytes <= s->max_transfer);
1228     if (nvme_qiov_aligned(bs, qiov)) {
1229         s->stats.aligned_accesses++;
1230         return nvme_co_prw_aligned(bs, offset, bytes, qiov, is_write, flags);
1231     }
1232     s->stats.unaligned_accesses++;
1233     trace_nvme_prw_buffered(s, offset, bytes, qiov->niov, is_write);
1234     buf = qemu_try_memalign(qemu_real_host_page_size, len);
1235 
1236     if (!buf) {
1237         return -ENOMEM;
1238     }
1239     qemu_iovec_init(&local_qiov, 1);
1240     if (is_write) {
1241         qemu_iovec_to_buf(qiov, 0, buf, bytes);
1242     }
1243     qemu_iovec_add(&local_qiov, buf, bytes);
1244     r = nvme_co_prw_aligned(bs, offset, bytes, &local_qiov, is_write, flags);
1245     qemu_iovec_destroy(&local_qiov);
1246     if (!r && !is_write) {
1247         qemu_iovec_from_buf(qiov, 0, buf, bytes);
1248     }
1249     qemu_vfree(buf);
1250     return r;
1251 }
1252 
1253 static coroutine_fn int nvme_co_preadv(BlockDriverState *bs,
1254                                        uint64_t offset, uint64_t bytes,
1255                                        QEMUIOVector *qiov, int flags)
1256 {
1257     return nvme_co_prw(bs, offset, bytes, qiov, false, flags);
1258 }
1259 
1260 static coroutine_fn int nvme_co_pwritev(BlockDriverState *bs,
1261                                         uint64_t offset, uint64_t bytes,
1262                                         QEMUIOVector *qiov, int flags)
1263 {
1264     return nvme_co_prw(bs, offset, bytes, qiov, true, flags);
1265 }
1266 
1267 static coroutine_fn int nvme_co_flush(BlockDriverState *bs)
1268 {
1269     BDRVNVMeState *s = bs->opaque;
1270     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1271     NVMeRequest *req;
1272     NvmeCmd cmd = {
1273         .opcode = NVME_CMD_FLUSH,
1274         .nsid = cpu_to_le32(s->nsid),
1275     };
1276     NVMeCoData data = {
1277         .ctx = bdrv_get_aio_context(bs),
1278         .ret = -EINPROGRESS,
1279     };
1280 
1281     assert(s->queue_count > 1);
1282     req = nvme_get_free_req(ioq);
1283     assert(req);
1284     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1285 
1286     data.co = qemu_coroutine_self();
1287     if (data.ret == -EINPROGRESS) {
1288         qemu_coroutine_yield();
1289     }
1290 
1291     return data.ret;
1292 }
1293 
1294 
1295 static coroutine_fn int nvme_co_pwrite_zeroes(BlockDriverState *bs,
1296                                               int64_t offset,
1297                                               int bytes,
1298                                               BdrvRequestFlags flags)
1299 {
1300     BDRVNVMeState *s = bs->opaque;
1301     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1302     NVMeRequest *req;
1303 
1304     uint32_t cdw12 = ((bytes >> s->blkshift) - 1) & 0xFFFF;
1305 
1306     if (!s->supports_write_zeroes) {
1307         return -ENOTSUP;
1308     }
1309 
1310     NvmeCmd cmd = {
1311         .opcode = NVME_CMD_WRITE_ZEROES,
1312         .nsid = cpu_to_le32(s->nsid),
1313         .cdw10 = cpu_to_le32((offset >> s->blkshift) & 0xFFFFFFFF),
1314         .cdw11 = cpu_to_le32(((offset >> s->blkshift) >> 32) & 0xFFFFFFFF),
1315     };
1316 
1317     NVMeCoData data = {
1318         .ctx = bdrv_get_aio_context(bs),
1319         .ret = -EINPROGRESS,
1320     };
1321 
1322     if (flags & BDRV_REQ_MAY_UNMAP) {
1323         cdw12 |= (1 << 25);
1324     }
1325 
1326     if (flags & BDRV_REQ_FUA) {
1327         cdw12 |= (1 << 30);
1328     }
1329 
1330     cmd.cdw12 = cpu_to_le32(cdw12);
1331 
1332     trace_nvme_write_zeroes(s, offset, bytes, flags);
1333     assert(s->queue_count > 1);
1334     req = nvme_get_free_req(ioq);
1335     assert(req);
1336 
1337     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1338 
1339     data.co = qemu_coroutine_self();
1340     while (data.ret == -EINPROGRESS) {
1341         qemu_coroutine_yield();
1342     }
1343 
1344     trace_nvme_rw_done(s, true, offset, bytes, data.ret);
1345     return data.ret;
1346 }
1347 
1348 
1349 static int coroutine_fn nvme_co_pdiscard(BlockDriverState *bs,
1350                                          int64_t offset,
1351                                          int bytes)
1352 {
1353     BDRVNVMeState *s = bs->opaque;
1354     NVMeQueuePair *ioq = s->queues[INDEX_IO(0)];
1355     NVMeRequest *req;
1356     NvmeDsmRange *buf;
1357     QEMUIOVector local_qiov;
1358     int ret;
1359 
1360     NvmeCmd cmd = {
1361         .opcode = NVME_CMD_DSM,
1362         .nsid = cpu_to_le32(s->nsid),
1363         .cdw10 = cpu_to_le32(0), /*number of ranges - 0 based*/
1364         .cdw11 = cpu_to_le32(1 << 2), /*deallocate bit*/
1365     };
1366 
1367     NVMeCoData data = {
1368         .ctx = bdrv_get_aio_context(bs),
1369         .ret = -EINPROGRESS,
1370     };
1371 
1372     if (!s->supports_discard) {
1373         return -ENOTSUP;
1374     }
1375 
1376     assert(s->queue_count > 1);
1377 
1378     buf = qemu_try_memalign(s->page_size, s->page_size);
1379     if (!buf) {
1380         return -ENOMEM;
1381     }
1382     memset(buf, 0, s->page_size);
1383     buf->nlb = cpu_to_le32(bytes >> s->blkshift);
1384     buf->slba = cpu_to_le64(offset >> s->blkshift);
1385     buf->cattr = 0;
1386 
1387     qemu_iovec_init(&local_qiov, 1);
1388     qemu_iovec_add(&local_qiov, buf, 4096);
1389 
1390     req = nvme_get_free_req(ioq);
1391     assert(req);
1392 
1393     qemu_co_mutex_lock(&s->dma_map_lock);
1394     ret = nvme_cmd_map_qiov(bs, &cmd, req, &local_qiov);
1395     qemu_co_mutex_unlock(&s->dma_map_lock);
1396 
1397     if (ret) {
1398         nvme_put_free_req_and_wake(ioq, req);
1399         goto out;
1400     }
1401 
1402     trace_nvme_dsm(s, offset, bytes);
1403 
1404     nvme_submit_command(ioq, req, &cmd, nvme_rw_cb, &data);
1405 
1406     data.co = qemu_coroutine_self();
1407     while (data.ret == -EINPROGRESS) {
1408         qemu_coroutine_yield();
1409     }
1410 
1411     qemu_co_mutex_lock(&s->dma_map_lock);
1412     ret = nvme_cmd_unmap_qiov(bs, &local_qiov);
1413     qemu_co_mutex_unlock(&s->dma_map_lock);
1414 
1415     if (ret) {
1416         goto out;
1417     }
1418 
1419     ret = data.ret;
1420     trace_nvme_dsm_done(s, offset, bytes, ret);
1421 out:
1422     qemu_iovec_destroy(&local_qiov);
1423     qemu_vfree(buf);
1424     return ret;
1425 
1426 }
1427 
1428 static int coroutine_fn nvme_co_truncate(BlockDriverState *bs, int64_t offset,
1429                                          bool exact, PreallocMode prealloc,
1430                                          BdrvRequestFlags flags, Error **errp)
1431 {
1432     int64_t cur_length;
1433 
1434     if (prealloc != PREALLOC_MODE_OFF) {
1435         error_setg(errp, "Unsupported preallocation mode '%s'",
1436                    PreallocMode_str(prealloc));
1437         return -ENOTSUP;
1438     }
1439 
1440     cur_length = nvme_getlength(bs);
1441     if (offset != cur_length && exact) {
1442         error_setg(errp, "Cannot resize NVMe devices");
1443         return -ENOTSUP;
1444     } else if (offset > cur_length) {
1445         error_setg(errp, "Cannot grow NVMe devices");
1446         return -EINVAL;
1447     }
1448 
1449     return 0;
1450 }
1451 
1452 static int nvme_reopen_prepare(BDRVReopenState *reopen_state,
1453                                BlockReopenQueue *queue, Error **errp)
1454 {
1455     return 0;
1456 }
1457 
1458 static void nvme_refresh_filename(BlockDriverState *bs)
1459 {
1460     BDRVNVMeState *s = bs->opaque;
1461 
1462     snprintf(bs->exact_filename, sizeof(bs->exact_filename), "nvme://%s/%i",
1463              s->device, s->nsid);
1464 }
1465 
1466 static void nvme_refresh_limits(BlockDriverState *bs, Error **errp)
1467 {
1468     BDRVNVMeState *s = bs->opaque;
1469 
1470     bs->bl.opt_mem_alignment = s->page_size;
1471     bs->bl.request_alignment = s->page_size;
1472     bs->bl.max_transfer = s->max_transfer;
1473 }
1474 
1475 static void nvme_detach_aio_context(BlockDriverState *bs)
1476 {
1477     BDRVNVMeState *s = bs->opaque;
1478 
1479     for (unsigned i = 0; i < s->queue_count; i++) {
1480         NVMeQueuePair *q = s->queues[i];
1481 
1482         qemu_bh_delete(q->completion_bh);
1483         q->completion_bh = NULL;
1484     }
1485 
1486     aio_set_event_notifier(bdrv_get_aio_context(bs),
1487                            &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
1488                            false, NULL, NULL);
1489 }
1490 
1491 static void nvme_attach_aio_context(BlockDriverState *bs,
1492                                     AioContext *new_context)
1493 {
1494     BDRVNVMeState *s = bs->opaque;
1495 
1496     s->aio_context = new_context;
1497     aio_set_event_notifier(new_context, &s->irq_notifier[MSIX_SHARED_IRQ_IDX],
1498                            false, nvme_handle_event, nvme_poll_cb);
1499 
1500     for (unsigned i = 0; i < s->queue_count; i++) {
1501         NVMeQueuePair *q = s->queues[i];
1502 
1503         q->completion_bh =
1504             aio_bh_new(new_context, nvme_process_completion_bh, q);
1505     }
1506 }
1507 
1508 static void nvme_aio_plug(BlockDriverState *bs)
1509 {
1510     BDRVNVMeState *s = bs->opaque;
1511     assert(!s->plugged);
1512     s->plugged = true;
1513 }
1514 
1515 static void nvme_aio_unplug(BlockDriverState *bs)
1516 {
1517     BDRVNVMeState *s = bs->opaque;
1518     assert(s->plugged);
1519     s->plugged = false;
1520     for (unsigned i = INDEX_IO(0); i < s->queue_count; i++) {
1521         NVMeQueuePair *q = s->queues[i];
1522         qemu_mutex_lock(&q->lock);
1523         nvme_kick(q);
1524         nvme_process_completion(q);
1525         qemu_mutex_unlock(&q->lock);
1526     }
1527 }
1528 
1529 static void nvme_register_buf(BlockDriverState *bs, void *host, size_t size)
1530 {
1531     int ret;
1532     Error *local_err = NULL;
1533     BDRVNVMeState *s = bs->opaque;
1534 
1535     ret = qemu_vfio_dma_map(s->vfio, host, size, false, NULL, &local_err);
1536     if (ret) {
1537         /* FIXME: we may run out of IOVA addresses after repeated
1538          * bdrv_register_buf/bdrv_unregister_buf, because nvme_vfio_dma_unmap
1539          * doesn't reclaim addresses for fixed mappings. */
1540         error_reportf_err(local_err, "nvme_register_buf failed: ");
1541     }
1542 }
1543 
1544 static void nvme_unregister_buf(BlockDriverState *bs, void *host)
1545 {
1546     BDRVNVMeState *s = bs->opaque;
1547 
1548     qemu_vfio_dma_unmap(s->vfio, host);
1549 }
1550 
1551 static BlockStatsSpecific *nvme_get_specific_stats(BlockDriverState *bs)
1552 {
1553     BlockStatsSpecific *stats = g_new(BlockStatsSpecific, 1);
1554     BDRVNVMeState *s = bs->opaque;
1555 
1556     stats->driver = BLOCKDEV_DRIVER_NVME;
1557     stats->u.nvme = (BlockStatsSpecificNvme) {
1558         .completion_errors = s->stats.completion_errors,
1559         .aligned_accesses = s->stats.aligned_accesses,
1560         .unaligned_accesses = s->stats.unaligned_accesses,
1561     };
1562 
1563     return stats;
1564 }
1565 
1566 static const char *const nvme_strong_runtime_opts[] = {
1567     NVME_BLOCK_OPT_DEVICE,
1568     NVME_BLOCK_OPT_NAMESPACE,
1569 
1570     NULL
1571 };
1572 
1573 static BlockDriver bdrv_nvme = {
1574     .format_name              = "nvme",
1575     .protocol_name            = "nvme",
1576     .instance_size            = sizeof(BDRVNVMeState),
1577 
1578     .bdrv_co_create_opts      = bdrv_co_create_opts_simple,
1579     .create_opts              = &bdrv_create_opts_simple,
1580 
1581     .bdrv_parse_filename      = nvme_parse_filename,
1582     .bdrv_file_open           = nvme_file_open,
1583     .bdrv_close               = nvme_close,
1584     .bdrv_getlength           = nvme_getlength,
1585     .bdrv_probe_blocksizes    = nvme_probe_blocksizes,
1586     .bdrv_co_truncate         = nvme_co_truncate,
1587 
1588     .bdrv_co_preadv           = nvme_co_preadv,
1589     .bdrv_co_pwritev          = nvme_co_pwritev,
1590 
1591     .bdrv_co_pwrite_zeroes    = nvme_co_pwrite_zeroes,
1592     .bdrv_co_pdiscard         = nvme_co_pdiscard,
1593 
1594     .bdrv_co_flush_to_disk    = nvme_co_flush,
1595     .bdrv_reopen_prepare      = nvme_reopen_prepare,
1596 
1597     .bdrv_refresh_filename    = nvme_refresh_filename,
1598     .bdrv_refresh_limits      = nvme_refresh_limits,
1599     .strong_runtime_opts      = nvme_strong_runtime_opts,
1600     .bdrv_get_specific_stats  = nvme_get_specific_stats,
1601 
1602     .bdrv_detach_aio_context  = nvme_detach_aio_context,
1603     .bdrv_attach_aio_context  = nvme_attach_aio_context,
1604 
1605     .bdrv_io_plug             = nvme_aio_plug,
1606     .bdrv_io_unplug           = nvme_aio_unplug,
1607 
1608     .bdrv_register_buf        = nvme_register_buf,
1609     .bdrv_unregister_buf      = nvme_unregister_buf,
1610 };
1611 
1612 static void bdrv_nvme_init(void)
1613 {
1614     bdrv_register(&bdrv_nvme);
1615 }
1616 
1617 block_init(bdrv_nvme_init);
1618