block/nvme.c

   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/cutils.h"
  21 #include "qemu/option.h"
  22 #include "qemu/vfio-helpers.h"
  23 #include "block/block_int.h"
  24 #include "trace.h"
  25
  26 #include "block/nvme.h"
  27
  28 #define NVME_SQ_ENTRY_BYTES 64
  29 #define NVME_CQ_ENTRY_BYTES 16
  30 #define NVME_QUEUE_SIZE 128
  31 #define NVME_BAR_SIZE 8192
  32
  33 typedef struct {
  34     int32_t  head, tail;
  35     uint8_t  *queue;
  36     uint64_t iova;
  37     /* Hardware MMIO register */
  38     volatile uint32_t *doorbell;
  39 } NVMeQueue;
  40
  41 typedef struct {
  42     BlockCompletionFunc *cb;
  43     void *opaque;
  44     int cid;
  45     void *prp_list_page;
  46     uint64_t prp_list_iova;
  47     bool busy;
  48 } NVMeRequest;
  49
  50 typedef struct {
  51     CoQueue     free_req_queue;
  52     QemuMutex   lock;
  53
  54     /* Fields protected by BQL */
  55     int         index;
  56     uint8_t     *prp_list_pages;
  57
  58     /* Fields protected by @lock */
  59     NVMeQueue   sq, cq;
  60     int         cq_phase;
  61     NVMeRequest reqs[NVME_QUEUE_SIZE];
  62     bool        busy;
  63     int         need_kick;
  64     int         inflight;
  65 } NVMeQueuePair;
  66
  67 /* Memory mapped registers */
  68 typedef volatile struct {
  69     uint64_t cap;
  70     uint32_t vs;
  71     uint32_t intms;
  72     uint32_t intmc;
  73     uint32_t cc;
  74     uint32_t reserved0;
  75     uint32_t csts;
  76     uint32_t nssr;
  77     uint32_t aqa;
  78     uint64_t asq;
  79     uint64_t acq;
  80     uint32_t cmbloc;
  81     uint32_t cmbsz;
  82     uint8_t  reserved1[0xec0];
  83     uint8_t  cmd_set_specfic[0x100];
  84     uint32_t doorbells[];
  85 } QEMU_PACKED NVMeRegs;
  86
  87 QEMU_BUILD_BUG_ON(offsetof(NVMeRegs, doorbells) != 0x1000);
  88
  89 typedef struct {
  90     AioContext *aio_context;
  91     QEMUVFIOState *vfio;
  92     NVMeRegs *regs;
  93     /* The submission/completion queue pairs.
  94      * [0]: admin queue.
  95      * [1..]: io queues.
  96      */
  97     NVMeQueuePair **queues;
  98     int nr_queues;
  99     size_t page_size;
 100     /* How many uint32_t elements does each doorbell entry take. */
 101     size_t doorbell_scale;
 102     bool write_cache_supported;
 103     EventNotifier irq_notifier;
 104     uint64_t nsze; /* Namespace size reported by identify command */
 105     int nsid;      /* The namespace id to read/write data. */
 106     uint64_t max_transfer;
 107     int plugged;
 108
 109     CoMutex dma_map_lock;
 110     CoQueue dma_flush_queue;
 111
 112     /* Total size of mapped qiov, accessed under dma_map_lock */
 113     int dma_map_count;
 114 } BDRVNVMeState;
 115
 116 #define NVME_BLOCK_OPT_DEVICE "device"
 117 #define NVME_BLOCK_OPT_NAMESPACE "namespace"
 118
 119 static QemuOptsList runtime_opts = {
 120     .name = "nvme",
 121     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
 122     .desc = {
 123         {
 124             .name = NVME_BLOCK_OPT_DEVICE,
 125             .type = QEMU_OPT_STRING,
 126             .help = "NVMe PCI device address",
 127         },
 128         {
 129             .name = NVME_BLOCK_OPT_NAMESPACE,
 130             .type = QEMU_OPT_NUMBER,
 131             .help = "NVMe namespace",
 132         },
 133         { /* end of list */ }
 134     },
 135 };
 136
 137 static void nvme_init_queue(BlockDriverState *bs, NVMeQueue *q,
 138                             int nentries, int entry_bytes, Error **errp)
 139 {
 140     BDRVNVMeState *s = bs->opaque;
 141     size_t bytes;
 142     int r;
 143
 144     bytes = ROUND_UP(nentries * entry_bytes, s->page_size);
 145     q->head = q->tail = 0;
 146     q->queue = qemu_try_blockalign0(bs, bytes);
 147
 148     if (!q->queue) {
 149         error_setg(errp, "Cannot allocate queue");
 150         return;
 151     }
 152     r = qemu_vfio_dma_map(s->vfio, q->queue, bytes, false, &q->iova);
 153     if (r) {
 154         error_setg(errp, "Cannot map queue");
 155     }
 156 }
 157
 158 static void nvme_free_queue_pair(BlockDriverState *bs, NVMeQueuePair *q)
 159 {
 160     qemu_vfree(q->prp_list_pages);
 161     qemu_vfree(q->sq.queue);
 162     qemu_vfree(q->cq.queue);
 163     qemu_mutex_destroy(&q->lock);
 164     g_free(q);
 165 }
 166
 167 static void nvme_free_req_queue_cb(void *opaque)
 168 {
 169     NVMeQueuePair *q = opaque;
 170
 171     qemu_mutex_lock(&q->lock);
 172     while (qemu_co_enter_next(&q->free_req_queue, &q->lock)) {
 173         /* Retry all pending requests */
 174     }
 175     qemu_mutex_unlock(&q->lock);
 176 }
 177
 178 static NVMeQueuePair *nvme_create_queue_pair(BlockDriverState *bs,
 179                                              int idx, int size,
 180                                              Error **errp)
 181 {
 182     int i, r;
 183     BDRVNVMeState *s = bs->opaque;
 184     Error *local_err = NULL;
 185     NVMeQueuePair *q = g_new0(NVMeQueuePair, 1);
 186     uint64_t prp_list_iova;
 187
 188     qemu_mutex_init(&q->lock);
 189     q->index = idx;
 190     qemu_co_queue_init(&q->free_req_queue);
 191     q->prp_list_pages = qemu_blockalign0(bs, s->page_size * NVME_QUEUE_SIZE);
 192     r = qemu_vfio_dma_map(s->vfio, q->prp_list_pages,
 193                           s->page_size * NVME_QUEUE_SIZE,
 194                           false, &prp_list_iova);
 195     if (r) {
 196         goto fail;
 197     }
 198     for (i = 0; i < NVME_QUEUE_SIZE; i++) {
 199         NVMeRequest *req = &q->reqs[i];
 200         req->cid = i + 1;
 201         req->prp_list_page = q->prp_list_pages + i * s->page_size;
 202         req->prp_list_iova = prp_list_iova + i * s->page_size;
 203     }
 204     nvme_init_queue(bs, &q->sq, size, NVME_SQ_ENTRY_BYTES, &local_err);
 205     if (local_err) {
 206         error_propagate(errp, local_err);
 207         goto fail;
 208     }
 209     q->sq.doorbell = &s->regs->doorbells[idx * 2 * s->doorbell_scale];
 210
 211     nvme_init_queue(bs, &q->cq, size, NVME_CQ_ENTRY_BYTES, &local_err);
 212     if (local_err) {
 213         error_propagate(errp, local_err);
 214         goto fail;
 215     }
 216     q->cq.doorbell = &s->regs->doorbells[idx * 2 * s->doorbell_scale + 1];
 217
 218     return q;
 219 fail:
 220     nvme_free_queue_pair(bs, q);
 221     return NULL;
 222 }
 223
 224 /* With q->lock */
 225 static void nvme_kick(BDRVNVMeState *s, NVMeQueuePair *q)
 226 {
 227     if (s->plugged || !q->need_kick) {
 228         return;
 229     }
 230     trace_nvme_kick(s, q->index);
 231     assert(!(q->sq.tail & 0xFF00));
 232     /* Fence the write to submission queue entry before notifying the device. */
 233     smp_wmb();
 234     *q->sq.doorbell = cpu_to_le32(q->sq.tail);
 235     q->inflight += q->need_kick;
 236     q->need_kick = 0;
 237 }
 238
 239 /* Find a free request element if any, otherwise:
 240  * a) if in coroutine context, try to wait for one to become available;
 241  * b) if not in coroutine, return NULL;
 242  */
 243 static NVMeRequest *nvme_get_free_req(NVMeQueuePair *q)
 244 {
 245     int i;
 246     NVMeRequest *req = NULL;
 247
 248     qemu_mutex_lock(&q->lock);
 249     while (q->inflight + q->need_kick > NVME_QUEUE_SIZE - 2) {
 250         /* We have to leave one slot empty as that is the full queue case (head
 251          * == tail + 1). */
 252         if (qemu_in_coroutine()) {
 253             trace_nvme_free_req_queue_wait(q);
 254             qemu_co_queue_wait(&q->free_req_queue, &q->lock);
 255         } else {
 256             qemu_mutex_unlock(&q->lock);
 257             return NULL;
 258         }
 259     }
 260     for (i = 0; i < NVME_QUEUE_SIZE; i++) {
 261         if (!q->reqs[i].busy) {
 262             q->reqs[i].busy = true;
 263             req = &q->reqs[i];
 264             break;
 265         }
 266     }
 267     /* We have checked inflight and need_kick while holding q->lock, so one
 268      * free req must be available. */
 269     assert(req);
 270     qemu_mutex_unlock(&q->lock);
 271     return req;
 272 }
 273
 274 static inline int nvme_translate_error(const NvmeCqe *c)
 275 {
 276     uint16_t status = (le16_to_cpu(c->status) >> 1) & 0xFF;
 277     if (status) {
 278         trace_nvme_error(le32_to_cpu(c->result),
 279                          le16_to_cpu(c->sq_head),
 280                          le16_to_cpu(c->sq_id),
 281                          le16_to_cpu(c->cid),
 282                          le16_to_cpu(status));
 283     }
 284     switch (status) {
 285     case 0:
 286         return 0;
 287     case 1:
 288         return -ENOSYS;
 289     case 2:
 290         return -EINVAL;
 291     default:
 292         return -EIO;
 293     }
 294 }
 295
 296 /* With q->lock */
 297 static bool nvme_process_completion(BDRVNVMeState *s, NVMeQueuePair *q)
 298 {
 299     bool progress = false;
 300     NVMeRequest *preq;
 301     NVMeRequest req;
 302     NvmeCqe *c;
 303
 304     trace_nvme_process_completion(s, q->index, q->inflight);
 305     if (q->busy || s->plugged) {
 306         trace_nvme_process_completion_queue_busy(s, q->index);
 307         return false;
 308     }
 309     q->busy = true;
 310     assert(q->inflight >= 0);
 311     while (q->inflight) {
 312         int16_t cid;
 313         c = (NvmeCqe *)&q->cq.queue[q->cq.head * NVME_CQ_ENTRY_BYTES];
 314         if (!c->cid || (le16_to_cpu(c->status) & 0x1) == q->cq_phase) {
 315             break;
 316         }
 317         q->cq.head = (q->cq.head + 1) % NVME_QUEUE_SIZE;
 318         if (!q->cq.head) {
 319             q->cq_phase = !q->cq_phase;
 320         }
 321         cid = le16_to_cpu(c->cid);
 322         if (cid == 0 || cid > NVME_QUEUE_SIZE) {
 323             fprintf(stderr, "Unexpected CID in completion queue: %" PRIu32 "\n",
 324                     cid);
 325             continue;
 326         }
 327         assert(cid <= NVME_QUEUE_SIZE);
 328         trace_nvme_complete_command(s, q->index, cid);
 329         preq = &q->reqs[cid - 1];
 330         req = *preq;
 331         assert(req.cid == cid);
 332         assert(req.cb);
 333         preq->busy = false;
 334         preq->cb = preq->opaque = NULL;
 335         qemu_mutex_unlock(&q->lock);
 336         req.cb(req.opaque, nvme_translate_error(c));
 337         qemu_mutex_lock(&q->lock);
 338         c->cid = cpu_to_le16(0);
 339         q->inflight--;
 340         /* Flip Phase Tag bit. */
 341         c->status = cpu_to_le16(le16_to_cpu(c->status) ^ 0x1);
 342         progress = true;
 343     }
 344     if (progress) {
 345         /* Notify the device so it can post more completions. */
 346         smp_mb_release();
 347         *q->cq.doorbell = cpu_to_le32(q->cq.head);
 348         if (!qemu_co_queue_empty(&q->free_req_queue)) {
 349             aio_bh_schedule_oneshot(s->aio_context, nvme_free_req_queue_cb, q);
 350         }
 351     }
 352     q->busy = false;
 353     return progress;
 354 }
 355
 356 static void nvme_trace_command(const NvmeCmd *cmd)
 357 {
 358     int i;
 359
 360     for (i = 0; i < 8; ++i) {
 361         uint8_t *cmdp = (uint8_t *)cmd + i * 8;
 362         trace_nvme_submit_command_raw(cmdp[0], cmdp[1], cmdp[2], cmdp[3],
 363                                       cmdp[4], cmdp[5], cmdp[6], cmdp[7]);
 364     }
 365 }
 366
 367 static void nvme_submit_command(BDRVNVMeState *s, NVMeQueuePair *q,
 368                                 NVMeRequest *req,
 369                                 NvmeCmd *cmd, BlockCompletionFunc cb,
 370                                 void *opaque)
 371 {
 372     assert(!req->cb);
 373     req->cb = cb;
 374     req->opaque = opaque;
 375     cmd->cid = cpu_to_le32(req->cid);
 376
 377     trace_nvme_submit_command(s, q->index, req->cid);
 378     nvme_trace_command(cmd);
 379     qemu_mutex_lock(&q->lock);
 380     memcpy((uint8_t *)q->sq.queue +
 381            q->sq.tail * NVME_SQ_ENTRY_BYTES, cmd, sizeof(*cmd));
 382     q->sq.tail = (q->sq.tail + 1) % NVME_QUEUE_SIZE;
 383     q->need_kick++;
 384     nvme_kick(s, q);
 385     nvme_process_completion(s, q);
 386     qemu_mutex_unlock(&q->lock);
 387 }
 388
 389 static void nvme_cmd_sync_cb(void *opaque, int ret)
 390 {
 391     int *pret = opaque;
 392     *pret = ret;
 393 }
 394
 395 static int nvme_cmd_sync(BlockDriverState *bs, NVMeQueuePair *q,
 396                          NvmeCmd *cmd)
 397 {
 398     NVMeRequest *req;
 399     BDRVNVMeState *s = bs->opaque;
 400     int ret = -EINPROGRESS;
 401     req = nvme_get_free_req(q);
 402     if (!req) {
 403         return -EBUSY;
 404     }
 405     nvme_submit_command(s, q, req, cmd, nvme_cmd_sync_cb, &ret);
 406
 407     BDRV_POLL_WHILE(bs, ret == -EINPROGRESS);
 408     return ret;
 409 }
 410
 411 static void nvme_identify(BlockDriverState *bs, int namespace, Error **errp)
 412 {
 413     BDRVNVMeState *s = bs->opaque;
 414     NvmeIdCtrl *idctrl;
 415     NvmeIdNs *idns;
 416     uint8_t *resp;
 417     int r;
 418     uint64_t iova;
 419     NvmeCmd cmd = {
 420         .opcode = NVME_ADM_CMD_IDENTIFY,
 421         .cdw10 = cpu_to_le32(0x1),
 422     };
 423
 424     resp = qemu_try_blockalign0(bs, sizeof(NvmeIdCtrl));
 425     if (!resp) {
 426         error_setg(errp, "Cannot allocate buffer for identify response");
 427         goto out;
 428     }
 429     idctrl = (NvmeIdCtrl *)resp;
 430     idns = (NvmeIdNs *)resp;
 431     r = qemu_vfio_dma_map(s->vfio, resp, sizeof(NvmeIdCtrl), true, &iova);
 432     if (r) {
 433         error_setg(errp, "Cannot map buffer for DMA");
 434         goto out;
 435     }
 436     cmd.prp1 = cpu_to_le64(iova);
 437
 438     if (nvme_cmd_sync(bs, s->queues[0], &cmd)) {
 439         error_setg(errp, "Failed to identify controller");
 440         goto out;
 441     }
 442
 443     if (le32_to_cpu(idctrl->nn) < namespace) {
 444         error_setg(errp, "Invalid namespace");
 445         goto out;
 446     }
 447     s->write_cache_supported = le32_to_cpu(idctrl->vwc) & 0x1;
 448     s->max_transfer = (idctrl->mdts ? 1 << idctrl->mdts : 0) * s->page_size;
 449     /* For now the page list buffer per command is one page, to hold at most
 450      * s->page_size / sizeof(uint64_t) entries. */
 451     s->max_transfer = MIN_NON_ZERO(s->max_transfer,
 452                           s->page_size / sizeof(uint64_t) * s->page_size);
 453
 454     memset(resp, 0, 4096);
 455
 456     cmd.cdw10 = 0;
 457     cmd.nsid = cpu_to_le32(namespace);
 458     if (nvme_cmd_sync(bs, s->queues[0], &cmd)) {
 459         error_setg(errp, "Failed to identify namespace");
 460         goto out;
 461     }
 462
 463     s->nsze = le64_to_cpu(idns->nsze);
 464
 465 out:
 466     qemu_vfio_dma_unmap(s->vfio, resp);
 467     qemu_vfree(resp);
 468 }
 469
 470 static bool nvme_poll_queues(BDRVNVMeState *s)
 471 {
 472     bool progress = false;
 473     int i;
 474
 475     for (i = 0; i < s->nr_queues; i++) {
 476         NVMeQueuePair *q = s->queues[i];
 477         qemu_mutex_lock(&q->lock);
 478         while (nvme_process_completion(s, q)) {
 479             /* Keep polling */
 480             progress = true;
 481         }
 482         qemu_mutex_unlock(&q->lock);
 483     }
 484     return progress;
 485 }
 486
 487 static void nvme_handle_event(EventNotifier *n)
 488 {
 489     BDRVNVMeState *s = container_of(n, BDRVNVMeState, irq_notifier);
 490
 491     trace_nvme_handle_event(s);
 492     aio_context_acquire(s->aio_context);
 493     event_notifier_test_and_clear(n);
 494     nvme_poll_queues(s);
 495     aio_context_release(s->aio_context);
 496 }
 497
 498 static bool nvme_add_io_queue(BlockDriverState *bs, Error **errp)
 499 {
 500     BDRVNVMeState *s = bs->opaque;
 501     int n = s->nr_queues;
 502     NVMeQueuePair *q;
 503     NvmeCmd cmd;
 504     int queue_size = NVME_QUEUE_SIZE;
 505
 506     q = nvme_create_queue_pair(bs, n, queue_size, errp);
 507     if (!q) {
 508         return false;
 509     }
 510     cmd = (NvmeCmd) {
 511         .opcode = NVME_ADM_CMD_CREATE_CQ,
 512         .prp1 = cpu_to_le64(q->cq.iova),
 513         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | (n & 0xFFFF)),
 514         .cdw11 = cpu_to_le32(0x3),
 515     };
 516     if (nvme_cmd_sync(bs, s->queues[0], &cmd)) {
 517         error_setg(errp, "Failed to create io queue [%d]", n);
 518         nvme_free_queue_pair(bs, q);
 519         return false;
 520     }
 521     cmd = (NvmeCmd) {
 522         .opcode = NVME_ADM_CMD_CREATE_SQ,
 523         .prp1 = cpu_to_le64(q->sq.iova),
 524         .cdw10 = cpu_to_le32(((queue_size - 1) << 16) | (n & 0xFFFF)),
 525         .cdw11 = cpu_to_le32(0x1 | (n << 16)),
 526     };
 527     if (nvme_cmd_sync(bs, s->queues[0], &cmd)) {
 528         error_setg(errp, "Failed to create io queue [%d]", n);
 529         nvme_free_queue_pair(bs, q);
 530         return false;
 531     }
 532     s->queues = g_renew(NVMeQueuePair *, s->queues, n + 1);
 533     s->queues[n] = q;
 534     s->nr_queues++;
 535     return true;
 536 }
 537
 538 static bool nvme_poll_cb(void *opaque)
 539 {
 540     EventNotifier *e = opaque;
 541     BDRVNVMeState *s = container_of(e, BDRVNVMeState, irq_notifier);
 542     bool progress = false;
 543
 544     trace_nvme_poll_cb(s);
 545     progress = nvme_poll_queues(s);
 546     return progress;
 547 }
 548
 549 static int nvme_init(BlockDriverState *bs, const char *device, int namespace,
 550                      Error **errp)
 551 {
 552     BDRVNVMeState *s = bs->opaque;
 553     int ret;
 554     uint64_t cap;
 555     uint64_t timeout_ms;
 556     uint64_t deadline, now;
 557     Error *local_err = NULL;
 558
 559     qemu_co_mutex_init(&s->dma_map_lock);
 560     qemu_co_queue_init(&s->dma_flush_queue);
 561     s->nsid = namespace;
 562     s->aio_context = bdrv_get_aio_context(bs);
 563     ret = event_notifier_init(&s->irq_notifier, 0);
 564     if (ret) {
 565         error_setg(errp, "Failed to init event notifier");
 566         return ret;
 567     }
 568
 569     s->vfio = qemu_vfio_open_pci(device, errp);
 570     if (!s->vfio) {
 571         ret = -EINVAL;
 572         goto fail;
 573     }
 574
 575     s->regs = qemu_vfio_pci_map_bar(s->vfio, 0, 0, NVME_BAR_SIZE, errp);
 576     if (!s->regs) {
 577         ret = -EINVAL;
 578         goto fail;
 579     }
 580
 581     /* Perform initialize sequence as described in NVMe spec "7.6.1
 582      * Initialization". */
 583
 584     cap = le64_to_cpu(s->regs->cap);
 585     if (!(cap & (1ULL << 37))) {
 586         error_setg(errp, "Device doesn't support NVMe command set");
 587         ret = -EINVAL;
 588         goto fail;
 589     }
 590
 591     s->page_size = MAX(4096, 1 << (12 + ((cap >> 48) & 0xF)));
 592     s->doorbell_scale = (4 << (((cap >> 32) & 0xF))) / sizeof(uint32_t);
 593     bs->bl.opt_mem_alignment = s->page_size;
 594     timeout_ms = MIN(500 * ((cap >> 24) & 0xFF), 30000);
 595
 596     /* Reset device to get a clean state. */
 597     s->regs->cc = cpu_to_le32(le32_to_cpu(s->regs->cc) & 0xFE);
 598     /* Wait for CSTS.RDY = 0. */
 599     deadline = qemu_clock_get_ns(QEMU_CLOCK_REALTIME) + timeout_ms * 1000000ULL;
 600     while (le32_to_cpu(s->regs->csts) & 0x1) {
 601         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
 602             error_setg(errp, "Timeout while waiting for device to reset (%"
 603                              PRId64 " ms)",
 604                        timeout_ms);
 605             ret = -ETIMEDOUT;
 606             goto fail;
 607         }
 608     }
 609
 610     /* Set up admin queue. */
 611     s->queues = g_new(NVMeQueuePair *, 1);
 612     s->nr_queues = 1;
 613     s->queues[0] = nvme_create_queue_pair(bs, 0, NVME_QUEUE_SIZE, errp);
 614     if (!s->queues[0]) {
 615         ret = -EINVAL;
 616         goto fail;
 617     }
 618     QEMU_BUILD_BUG_ON(NVME_QUEUE_SIZE & 0xF000);
 619     s->regs->aqa = cpu_to_le32((NVME_QUEUE_SIZE << 16) | NVME_QUEUE_SIZE);
 620     s->regs->asq = cpu_to_le64(s->queues[0]->sq.iova);
 621     s->regs->acq = cpu_to_le64(s->queues[0]->cq.iova);
 622
 623     /* After setting up all control registers we can enable device now. */
 624     s->regs->cc = cpu_to_le32((ctz32(NVME_CQ_ENTRY_BYTES) << 20) |
 625                               (ctz32(NVME_SQ_ENTRY_BYTES) << 16) |
 626                               0x1);
 627     /* Wait for CSTS.RDY = 1. */
 628     now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
 629     deadline = now + timeout_ms * 1000000;
 630     while (!(le32_to_cpu(s->regs->csts) & 0x1)) {
 631         if (qemu_clock_get_ns(QEMU_CLOCK_REALTIME) > deadline) {
 632             error_setg(errp, "Timeout while waiting for device to start (%"
 633                              PRId64 " ms)",
 634                        timeout_ms);
 635             ret = -ETIMEDOUT;
 636             goto fail_queue;
 637         }
 638     }
 639
 640     ret = qemu_vfio_pci_init_irq(s->vfio, &s->irq_notifier,
 641                                  VFIO_PCI_MSIX_IRQ_INDEX, errp);
 642     if (ret) {
 643         goto fail_queue;
 644     }
 645     aio_set_event_notifier(bdrv_get_aio_context(bs), &s->irq_notifier,
 646                            false, nvme_handle_event, nvme_poll_cb);
 647
 648     nvme_identify(bs, namespace, errp);
 649     if (local_err) {
 650         error_propagate(errp, local_err);
 651         ret = -EIO;
 652         goto fail_handler;
 653     }
 654
 655     /* Set up command queues. */
 656     if (!nvme_add_io_queue(bs, errp)) {
 657         ret = -EIO;
 658         goto fail_handler;
 659     }
 660     return 0;
 661
 662 fail_handler:
 663     aio_set_event_notifier(bdrv_get_aio_context(bs), &s->irq_notifier,
 664                            false, NULL, NULL);
 665 fail_queue:
 666     nvme_free_queue_pair(bs, s->queues[0]);
 667 fail:
 668     g_free(s->queues);
 669     qemu_vfio_pci_unmap_bar(s->vfio, 0, (void *)s->regs, 0, NVME_BAR_SIZE);
 670     qemu_vfio_close(s->vfio);
 671     event_notifier_cleanup(&s->irq_notifier);
 672     return ret;
 673 }
 674
 675 /* Parse a filename in the format of nvme://XXXX:XX:XX.X/X. Example:
 676  *
 677  *     nvme://0000:44:00.0/1
 678  *
 679  * where the "nvme://" is a fixed form of the protocol prefix, the middle part
 680  * is the PCI address, and the last part is the namespace number starting from
 681  * 1 according to the NVMe spec. */
 682 static void nvme_parse_filename(const char *filename, QDict *options,
 683                                 Error **errp)
 684 {
 685     int pref = strlen("nvme://");
 686
 687     if (strlen(filename) > pref && !strncmp(filename, "nvme://", pref)) {
 688         const char *tmp = filename + pref;
 689         char *device;
 690         const char *namespace;
 691         unsigned long ns;
 692         const char *slash = strchr(tmp, '/');
 693         if (!slash) {
 694             qdict_put(options, NVME_BLOCK_OPT_DEVICE,
 695                       qstring_from_str(tmp));
 696             return;
 697         }
 698         device = g_strndup(tmp, slash - tmp);
 699         qdict_put(options, NVME_BLOCK_OPT_DEVICE, qstring_from_str(device));
 700         g_free(device);
 701         namespace = slash + 1;
 702         if (*namespace && qemu_strtoul(namespace, NULL, 10, &ns)) {
 703             error_setg(errp, "Invalid namespace '%s', positive number expected",
 704                        namespace);
 705             return;
 706         }
 707         qdict_put(options, NVME_BLOCK_OPT_NAMESPACE,
 708                   qstring_from_str(*namespace ? namespace : "1"));
 709     }
 710 }
 711
 712 static int nvme_enable_disable_write_cache(BlockDriverState *bs, bool enable,
 713                                            Error **errp)
 714 {
 715     int ret;
 716     BDRVNVMeState *s = bs->opaque;
 717     NvmeCmd cmd = {
 718         .opcode = NVME_ADM_CMD_SET_FEATURES,
 719         .nsid = cpu_to_le32(s->nsid),
 720         .cdw10 = cpu_to_le32(0x06),
 721         .cdw11 = cpu_to_le32(enable ? 0x01 : 0x00),
 722     };
 723
 724     ret = nvme_cmd_sync(bs, s->queues[0], &cmd);
 725     if (ret) {
 726         error_setg(errp, "Failed to configure NVMe write cache");
 727     }
 728     return ret;
 729 }
 730
 731 static void nvme_close(BlockDriverState *bs)
 732 {
 733     int i;
 734     BDRVNVMeState *s = bs->opaque;
 735
 736     for (i = 0; i < s->nr_queues; ++i) {
 737         nvme_free_queue_pair(bs, s->queues[i]);
 738     }
 739     aio_set_event_notifier(bdrv_get_aio_context(bs), &s->irq_notifier,
 740                            false, NULL, NULL);
 741     qemu_vfio_pci_unmap_bar(s->vfio, 0, (void *)s->regs, 0, NVME_BAR_SIZE);
 742     qemu_vfio_close(s->vfio);
 743 }
 744
 745 static int nvme_file_open(BlockDriverState *bs, QDict *options, int flags,
 746                           Error **errp)
 747 {
 748     const char *device;
 749     QemuOpts *opts;
 750     int namespace;
 751     int ret;
 752     BDRVNVMeState *s = bs->opaque;
 753
 754     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
 755     qemu_opts_absorb_qdict(opts, options, &error_abort);
 756     device = qemu_opt_get(opts, NVME_BLOCK_OPT_DEVICE);
 757     if (!device) {
 758         error_setg(errp, "'" NVME_BLOCK_OPT_DEVICE "' option is required");
 759         qemu_opts_del(opts);
 760         return -EINVAL;
 761     }
 762
 763     namespace = qemu_opt_get_number(opts, NVME_BLOCK_OPT_NAMESPACE, 1);
 764     ret = nvme_init(bs, device, namespace, errp);
 765     qemu_opts_del(opts);
 766     if (ret) {
 767         goto fail;
 768     }
 769     if (flags & BDRV_O_NOCACHE) {
 770         if (!s->write_cache_supported) {
 771             error_setg(errp,
 772                        "NVMe controller doesn't support write cache configuration");
 773             ret = -EINVAL;
 774         } else {
 775             ret = nvme_enable_disable_write_cache(bs, !(flags & BDRV_O_NOCACHE),
 776                                                   errp);
 777         }
 778         if (ret) {
 779             goto fail;
 780         }
 781     }
 782     bs->supported_write_flags = BDRV_REQ_FUA;
 783     return 0;
 784 fail:
 785     nvme_close(bs);
 786     return ret;
 787 }
 788
 789 static int64_t nvme_getlength(BlockDriverState *bs)
 790 {
 791     BDRVNVMeState *s = bs->opaque;
 792
 793     return s->nsze << BDRV_SECTOR_BITS;
 794 }
 795
 796 /* Called with s->dma_map_lock */
 797 static coroutine_fn int nvme_cmd_unmap_qiov(BlockDriverState *bs,
 798                                             QEMUIOVector *qiov)
 799 {
 800     int r = 0;
 801     BDRVNVMeState *s = bs->opaque;
 802
 803     s->dma_map_count -= qiov->size;
 804     if (!s->dma_map_count && !qemu_co_queue_empty(&s->dma_flush_queue)) {
 805         r = qemu_vfio_dma_reset_temporary(s->vfio);
 806         if (!r) {
 807             qemu_co_queue_restart_all(&s->dma_flush_queue);
 808         }
 809     }
 810     return r;
 811 }
 812
 813 /* Called with s->dma_map_lock */
 814 static coroutine_fn int nvme_cmd_map_qiov(BlockDriverState *bs, NvmeCmd *cmd,
 815                                           NVMeRequest *req, QEMUIOVector *qiov)
 816 {
 817     BDRVNVMeState *s = bs->opaque;
 818     uint64_t *pagelist = req->prp_list_page;
 819     int i, j, r;
 820     int entries = 0;
 821
 822     assert(qiov->size);
 823     assert(QEMU_IS_ALIGNED(qiov->size, s->page_size));
 824     assert(qiov->size / s->page_size <= s->page_size / sizeof(uint64_t));
 825     for (i = 0; i < qiov->niov; ++i) {
 826         bool retry = true;
 827         uint64_t iova;
 828 try_map:
 829         r = qemu_vfio_dma_map(s->vfio,
 830                               qiov->iov[i].iov_base,
 831                               qiov->iov[i].iov_len,
 832                               true, &iova);
 833         if (r == -ENOMEM && retry) {
 834             retry = false;
 835             trace_nvme_dma_flush_queue_wait(s);
 836             if (s->dma_map_count) {
 837                 trace_nvme_dma_map_flush(s);
 838                 qemu_co_queue_wait(&s->dma_flush_queue, &s->dma_map_lock);
 839             } else {
 840                 r = qemu_vfio_dma_reset_temporary(s->vfio);
 841                 if (r) {
 842                     goto fail;
 843                 }
 844             }
 845             goto try_map;
 846         }
 847         if (r) {
 848             goto fail;
 849         }
 850
 851         for (j = 0; j < qiov->iov[i].iov_len / s->page_size; j++) {
 852             pagelist[entries++] = iova + j * s->page_size;
 853         }
 854         trace_nvme_cmd_map_qiov_iov(s, i, qiov->iov[i].iov_base,
 855                                     qiov->iov[i].iov_len / s->page_size);
 856     }
 857
 858     s->dma_map_count += qiov->size;
 859
 860     assert(entries <= s->page_size / sizeof(uint64_t));
 861     switch (entries) {
 862     case 0:
 863         abort();
 864     case 1:
 865         cmd->prp1 = cpu_to_le64(pagelist[0]);
 866         cmd->prp2 = 0;
 867         break;
 868     case 2:
 869         cmd->prp1 = cpu_to_le64(pagelist[0]);
 870         cmd->prp2 = cpu_to_le64(pagelist[1]);;
 871         break;
 872     default:
 873         cmd->prp1 = cpu_to_le64(pagelist[0]);
 874         cmd->prp2 = cpu_to_le64(req->prp_list_iova);
 875         for (i = 0; i < entries - 1; ++i) {
 876             pagelist[i] = cpu_to_le64(pagelist[i + 1]);
 877         }
 878         pagelist[entries - 1] = 0;
 879         break;
 880     }
 881     trace_nvme_cmd_map_qiov(s, cmd, req, qiov, entries);
 882     for (i = 0; i < entries; ++i) {
 883         trace_nvme_cmd_map_qiov_pages(s, i, pagelist[i]);
 884     }
 885     return 0;
 886 fail:
 887     /* No need to unmap [0 - i) iovs even if we've failed, since we don't
 888      * increment s->dma_map_count. This is okay for fixed mapping memory areas
 889      * because they are already mapped before calling this function; for
 890      * temporary mappings, a later nvme_cmd_(un)map_qiov will reclaim by
 891      * calling qemu_vfio_dma_reset_temporary when necessary. */
 892     return r;
 893 }
 894
 895 typedef struct {
 896     Coroutine *co;
 897     int ret;
 898     AioContext *ctx;
 899 } NVMeCoData;
 900
 901 static void nvme_rw_cb_bh(void *opaque)
 902 {
 903     NVMeCoData *data = opaque;
 904     qemu_coroutine_enter(data->co);
 905 }
 906
 907 static void nvme_rw_cb(void *opaque, int ret)
 908 {
 909     NVMeCoData *data = opaque;
 910     data->ret = ret;
 911     if (!data->co) {
 912         /* The rw coroutine hasn't yielded, don't try to enter. */
 913         return;
 914     }
 915     aio_bh_schedule_oneshot(data->ctx, nvme_rw_cb_bh, data);
 916 }
 917
 918 static coroutine_fn int nvme_co_prw_aligned(BlockDriverState *bs,
 919                                             uint64_t offset, uint64_t bytes,
 920                                             QEMUIOVector *qiov,
 921                                             bool is_write,
 922                                             int flags)
 923 {
 924     int r;
 925     BDRVNVMeState *s = bs->opaque;
 926     NVMeQueuePair *ioq = s->queues[1];
 927     NVMeRequest *req;
 928     uint32_t cdw12 = (((bytes >> BDRV_SECTOR_BITS) - 1) & 0xFFFF) |
 929                        (flags & BDRV_REQ_FUA ? 1 << 30 : 0);
 930     NvmeCmd cmd = {
 931         .opcode = is_write ? NVME_CMD_WRITE : NVME_CMD_READ,
 932         .nsid = cpu_to_le32(s->nsid),
 933         .cdw10 = cpu_to_le32((offset >> BDRV_SECTOR_BITS) & 0xFFFFFFFF),
 934         .cdw11 = cpu_to_le32(((offset >> BDRV_SECTOR_BITS) >> 32) & 0xFFFFFFFF),
 935         .cdw12 = cpu_to_le32(cdw12),
 936     };
 937     NVMeCoData data = {
 938         .ctx = bdrv_get_aio_context(bs),
 939         .ret = -EINPROGRESS,
 940     };
 941
 942     trace_nvme_prw_aligned(s, is_write, offset, bytes, flags, qiov->niov);
 943     assert(s->nr_queues > 1);
 944     req = nvme_get_free_req(ioq);
 945     assert(req);
 946
 947     qemu_co_mutex_lock(&s->dma_map_lock);
 948     r = nvme_cmd_map_qiov(bs, &cmd, req, qiov);
 949     qemu_co_mutex_unlock(&s->dma_map_lock);
 950     if (r) {
 951         req->busy = false;
 952         return r;
 953     }
 954     nvme_submit_command(s, ioq, req, &cmd, nvme_rw_cb, &data);
 955
 956     data.co = qemu_coroutine_self();
 957     while (data.ret == -EINPROGRESS) {
 958         qemu_coroutine_yield();
 959     }
 960
 961     qemu_co_mutex_lock(&s->dma_map_lock);
 962     r = nvme_cmd_unmap_qiov(bs, qiov);
 963     qemu_co_mutex_unlock(&s->dma_map_lock);
 964     if (r) {
 965         return r;
 966     }
 967
 968     trace_nvme_rw_done(s, is_write, offset, bytes, data.ret);
 969     return data.ret;
 970 }
 971
 972 static inline bool nvme_qiov_aligned(BlockDriverState *bs,
 973                                      const QEMUIOVector *qiov)
 974 {
 975     int i;
 976     BDRVNVMeState *s = bs->opaque;
 977
 978     for (i = 0; i < qiov->niov; ++i) {
 979         if (!QEMU_PTR_IS_ALIGNED(qiov->iov[i].iov_base, s->page_size) ||
 980             !QEMU_IS_ALIGNED(qiov->iov[i].iov_len, s->page_size)) {
 981             trace_nvme_qiov_unaligned(qiov, i, qiov->iov[i].iov_base,
 982                                       qiov->iov[i].iov_len, s->page_size);
 983             return false;
 984         }
 985     }
 986     return true;
 987 }
 988
 989 static int nvme_co_prw(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 990                        QEMUIOVector *qiov, bool is_write, int flags)
 991 {
 992     BDRVNVMeState *s = bs->opaque;
 993     int r;
 994     uint8_t *buf = NULL;
 995     QEMUIOVector local_qiov;
 996
 997     assert(QEMU_IS_ALIGNED(offset, s->page_size));
 998     assert(QEMU_IS_ALIGNED(bytes, s->page_size));
 999     assert(bytes <= s->max_transfer);
1000     if (nvme_qiov_aligned(bs, qiov)) {
1001         return nvme_co_prw_aligned(bs, offset, bytes, qiov, is_write, flags);
1002     }
1003     trace_nvme_prw_buffered(s, offset, bytes, qiov->niov, is_write);
1004     buf = qemu_try_blockalign(bs, bytes);
1005
1006     if (!buf) {
1007         return -ENOMEM;
1008     }
1009     qemu_iovec_init(&local_qiov, 1);
1010     if (is_write) {
1011         qemu_iovec_to_buf(qiov, 0, buf, bytes);
1012     }
1013     qemu_iovec_add(&local_qiov, buf, bytes);
1014     r = nvme_co_prw_aligned(bs, offset, bytes, &local_qiov, is_write, flags);
1015     qemu_iovec_destroy(&local_qiov);
1016     if (!r && !is_write) {
1017         qemu_iovec_from_buf(qiov, 0, buf, bytes);
1018     }
1019     qemu_vfree(buf);
1020     return r;
1021 }
1022
1023 static coroutine_fn int nvme_co_preadv(BlockDriverState *bs,
1024                                        uint64_t offset, uint64_t bytes,
1025                                        QEMUIOVector *qiov, int flags)
1026 {
1027     return nvme_co_prw(bs, offset, bytes, qiov, false, flags);
1028 }
1029
1030 static coroutine_fn int nvme_co_pwritev(BlockDriverState *bs,
1031                                         uint64_t offset, uint64_t bytes,
1032                                         QEMUIOVector *qiov, int flags)
1033 {
1034     return nvme_co_prw(bs, offset, bytes, qiov, true, flags);
1035 }
1036
1037 static coroutine_fn int nvme_co_flush(BlockDriverState *bs)
1038 {
1039     BDRVNVMeState *s = bs->opaque;
1040     NVMeQueuePair *ioq = s->queues[1];
1041     NVMeRequest *req;
1042     NvmeCmd cmd = {
1043         .opcode = NVME_CMD_FLUSH,
1044         .nsid = cpu_to_le32(s->nsid),
1045     };
1046     NVMeCoData data = {
1047         .ctx = bdrv_get_aio_context(bs),
1048         .ret = -EINPROGRESS,
1049     };
1050
1051     assert(s->nr_queues > 1);
1052     req = nvme_get_free_req(ioq);
1053     assert(req);
1054     nvme_submit_command(s, ioq, req, &cmd, nvme_rw_cb, &data);
1055
1056     data.co = qemu_coroutine_self();
1057     if (data.ret == -EINPROGRESS) {
1058         qemu_coroutine_yield();
1059     }
1060
1061     return data.ret;
1062 }
1063
1064
1065 static int nvme_reopen_prepare(BDRVReopenState *reopen_state,
1066                                BlockReopenQueue *queue, Error **errp)
1067 {
1068     return 0;
1069 }
1070
1071 static int64_t coroutine_fn nvme_co_get_block_status(BlockDriverState *bs,
1072                                                      int64_t sector_num,
1073                                                      int nb_sectors, int *pnum,
1074                                                      BlockDriverState **file)
1075 {
1076     *pnum = nb_sectors;
1077     *file = bs;
1078
1079     return BDRV_BLOCK_ALLOCATED | BDRV_BLOCK_OFFSET_VALID |
1080            (sector_num << BDRV_SECTOR_BITS);
1081 }
1082
1083 static void nvme_refresh_filename(BlockDriverState *bs, QDict *opts)
1084 {
1085     QINCREF(opts);
1086     qdict_del(opts, "filename");
1087
1088     if (!qdict_size(opts)) {
1089         snprintf(bs->exact_filename, sizeof(bs->exact_filename), "%s://",
1090                  bs->drv->format_name);
1091     }
1092
1093     qdict_put(opts, "driver", qstring_from_str(bs->drv->format_name));
1094     bs->full_open_options = opts;
1095 }
1096
1097 static void nvme_refresh_limits(BlockDriverState *bs, Error **errp)
1098 {
1099     BDRVNVMeState *s = bs->opaque;
1100
1101     bs->bl.opt_mem_alignment = s->page_size;
1102     bs->bl.request_alignment = s->page_size;
1103     bs->bl.max_transfer = s->max_transfer;
1104 }
1105
1106 static void nvme_detach_aio_context(BlockDriverState *bs)
1107 {
1108     BDRVNVMeState *s = bs->opaque;
1109
1110     aio_set_event_notifier(bdrv_get_aio_context(bs), &s->irq_notifier,
1111                            false, NULL, NULL);
1112 }
1113
1114 static void nvme_attach_aio_context(BlockDriverState *bs,
1115                                     AioContext *new_context)
1116 {
1117     BDRVNVMeState *s = bs->opaque;
1118
1119     s->aio_context = new_context;
1120     aio_set_event_notifier(new_context, &s->irq_notifier,
1121                            false, nvme_handle_event, nvme_poll_cb);
1122 }
1123
1124 static void nvme_aio_plug(BlockDriverState *bs)
1125 {
1126     BDRVNVMeState *s = bs->opaque;
1127     s->plugged++;
1128 }
1129
1130 static void nvme_aio_unplug(BlockDriverState *bs)
1131 {
1132     int i;
1133     BDRVNVMeState *s = bs->opaque;
1134     assert(s->plugged);
1135     if (!--s->plugged) {
1136         for (i = 1; i < s->nr_queues; i++) {
1137             NVMeQueuePair *q = s->queues[i];
1138             qemu_mutex_lock(&q->lock);
1139             nvme_kick(s, q);
1140             nvme_process_completion(s, q);
1141             qemu_mutex_unlock(&q->lock);
1142         }
1143     }
1144 }
1145
1146 static void nvme_register_buf(BlockDriverState *bs, void *host, size_t size)
1147 {
1148     int ret;
1149     BDRVNVMeState *s = bs->opaque;
1150
1151     ret = qemu_vfio_dma_map(s->vfio, host, size, false, NULL);
1152     if (ret) {
1153         /* FIXME: we may run out of IOVA addresses after repeated
1154          * bdrv_register_buf/bdrv_unregister_buf, because nvme_vfio_dma_unmap
1155          * doesn't reclaim addresses for fixed mappings. */
1156         error_report("nvme_register_buf failed: %s", strerror(-ret));
1157     }
1158 }
1159
1160 static void nvme_unregister_buf(BlockDriverState *bs, void *host)
1161 {
1162     BDRVNVMeState *s = bs->opaque;
1163
1164     qemu_vfio_dma_unmap(s->vfio, host);
1165 }
1166
1167 static BlockDriver bdrv_nvme = {
1168     .format_name              = "nvme",
1169     .protocol_name            = "nvme",
1170     .instance_size            = sizeof(BDRVNVMeState),
1171
1172     .bdrv_parse_filename      = nvme_parse_filename,
1173     .bdrv_file_open           = nvme_file_open,
1174     .bdrv_close               = nvme_close,
1175     .bdrv_getlength           = nvme_getlength,
1176
1177     .bdrv_co_preadv           = nvme_co_preadv,
1178     .bdrv_co_pwritev          = nvme_co_pwritev,
1179     .bdrv_co_flush_to_disk    = nvme_co_flush,
1180     .bdrv_reopen_prepare      = nvme_reopen_prepare,
1181
1182     .bdrv_co_get_block_status = nvme_co_get_block_status,
1183
1184     .bdrv_refresh_filename    = nvme_refresh_filename,
1185     .bdrv_refresh_limits      = nvme_refresh_limits,
1186
1187     .bdrv_detach_aio_context  = nvme_detach_aio_context,
1188     .bdrv_attach_aio_context  = nvme_attach_aio_context,
1189
1190     .bdrv_io_plug             = nvme_aio_plug,
1191     .bdrv_io_unplug           = nvme_aio_unplug,
1192
1193     .bdrv_register_buf        = nvme_register_buf,
1194     .bdrv_unregister_buf      = nvme_unregister_buf,
1195 };
1196
1197 static void bdrv_nvme_init(void)
1198 {
1199     bdrv_register(&bdrv_nvme);
1200 }
1201
1202 block_init(bdrv_nvme_init);