hw/block/nvme.c

   1 /*
   2  * QEMU NVM Express Controller
   3  *
   4  * Copyright (c) 2012, Intel Corporation
   5  *
   6  * Written by Keith Busch <keith.busch@intel.com>
   7  *
   8  * This code is licensed under the GNU GPL v2 or later.
   9  */
  10
  11 /**
  12  * Reference Specs: http://www.nvmexpress.org, 1.1, 1.0e
  13  *
  14  *  http://www.nvmexpress.org/resources/
  15  */
  16
  17 /**
  18  * Usage: add options:
  19  *      -drive file=<file>,if=none,id=<drive_id>
  20  *      -device nvme,drive=<drive_id>,serial=<serial>,id=<id[optional]>
  21  */
  22
  23 #include "qemu/osdep.h"
  24 #include <hw/block/block.h>
  25 #include <hw/hw.h>
  26 #include <hw/pci/msix.h>
  27 #include <hw/pci/pci.h>
  28 #include "sysemu/sysemu.h"
  29 #include "qapi/visitor.h"
  30 #include "sysemu/block-backend.h"
  31
  32 #include "nvme.h"
  33
  34 static void nvme_process_sq(void *opaque);
  35
  36 static int nvme_check_sqid(NvmeCtrl *n, uint16_t sqid)
  37 {
  38     return sqid < n->num_queues && n->sq[sqid] != NULL ? 0 : -1;
  39 }
  40
  41 static int nvme_check_cqid(NvmeCtrl *n, uint16_t cqid)
  42 {
  43     return cqid < n->num_queues && n->cq[cqid] != NULL ? 0 : -1;
  44 }
  45
  46 static void nvme_inc_cq_tail(NvmeCQueue *cq)
  47 {
  48     cq->tail++;
  49     if (cq->tail >= cq->size) {
  50         cq->tail = 0;
  51         cq->phase = !cq->phase;
  52     }
  53 }
  54
  55 static void nvme_inc_sq_head(NvmeSQueue *sq)
  56 {
  57     sq->head = (sq->head + 1) % sq->size;
  58 }
  59
  60 static uint8_t nvme_cq_full(NvmeCQueue *cq)
  61 {
  62     return (cq->tail + 1) % cq->size == cq->head;
  63 }
  64
  65 static uint8_t nvme_sq_empty(NvmeSQueue *sq)
  66 {
  67     return sq->head == sq->tail;
  68 }
  69
  70 static void nvme_isr_notify(NvmeCtrl *n, NvmeCQueue *cq)
  71 {
  72     if (cq->irq_enabled) {
  73         if (msix_enabled(&(n->parent_obj))) {
  74             msix_notify(&(n->parent_obj), cq->vector);
  75         } else {
  76             pci_irq_pulse(&n->parent_obj);
  77         }
  78     }
  79 }
  80
  81 static uint16_t nvme_map_prp(QEMUSGList *qsg, uint64_t prp1, uint64_t prp2,
  82     uint32_t len, NvmeCtrl *n)
  83 {
  84     hwaddr trans_len = n->page_size - (prp1 % n->page_size);
  85     trans_len = MIN(len, trans_len);
  86     int num_prps = (len >> n->page_bits) + 1;
  87
  88     if (!prp1) {
  89         return NVME_INVALID_FIELD | NVME_DNR;
  90     }
  91
  92     pci_dma_sglist_init(qsg, &n->parent_obj, num_prps);
  93     qemu_sglist_add(qsg, prp1, trans_len);
  94     len -= trans_len;
  95     if (len) {
  96         if (!prp2) {
  97             goto unmap;
  98         }
  99         if (len > n->page_size) {
 100             uint64_t prp_list[n->max_prp_ents];
 101             uint32_t nents, prp_trans;
 102             int i = 0;
 103
 104             nents = (len + n->page_size - 1) >> n->page_bits;
 105             prp_trans = MIN(n->max_prp_ents, nents) * sizeof(uint64_t);
 106             pci_dma_read(&n->parent_obj, prp2, (void *)prp_list, prp_trans);
 107             while (len != 0) {
 108                 uint64_t prp_ent = le64_to_cpu(prp_list[i]);
 109
 110                 if (i == n->max_prp_ents - 1 && len > n->page_size) {
 111                     if (!prp_ent || prp_ent & (n->page_size - 1)) {
 112                         goto unmap;
 113                     }
 114
 115                     i = 0;
 116                     nents = (len + n->page_size - 1) >> n->page_bits;
 117                     prp_trans = MIN(n->max_prp_ents, nents) * sizeof(uint64_t);
 118                     pci_dma_read(&n->parent_obj, prp_ent, (void *)prp_list,
 119                         prp_trans);
 120                     prp_ent = le64_to_cpu(prp_list[i]);
 121                 }
 122
 123                 if (!prp_ent || prp_ent & (n->page_size - 1)) {
 124                     goto unmap;
 125                 }
 126
 127                 trans_len = MIN(len, n->page_size);
 128                 qemu_sglist_add(qsg, prp_ent, trans_len);
 129                 len -= trans_len;
 130                 i++;
 131             }
 132         } else {
 133             if (prp2 & (n->page_size - 1)) {
 134                 goto unmap;
 135             }
 136             qemu_sglist_add(qsg, prp2, len);
 137         }
 138     }
 139     return NVME_SUCCESS;
 140
 141  unmap:
 142     qemu_sglist_destroy(qsg);
 143     return NVME_INVALID_FIELD | NVME_DNR;
 144 }
 145
 146 static uint16_t nvme_dma_read_prp(NvmeCtrl *n, uint8_t *ptr, uint32_t len,
 147     uint64_t prp1, uint64_t prp2)
 148 {
 149     QEMUSGList qsg;
 150
 151     if (nvme_map_prp(&qsg, prp1, prp2, len, n)) {
 152         return NVME_INVALID_FIELD | NVME_DNR;
 153     }
 154     if (dma_buf_read(ptr, len, &qsg)) {
 155         qemu_sglist_destroy(&qsg);
 156         return NVME_INVALID_FIELD | NVME_DNR;
 157     }
 158     qemu_sglist_destroy(&qsg);
 159     return NVME_SUCCESS;
 160 }
 161
 162 static void nvme_post_cqes(void *opaque)
 163 {
 164     NvmeCQueue *cq = opaque;
 165     NvmeCtrl *n = cq->ctrl;
 166     NvmeRequest *req, *next;
 167
 168     QTAILQ_FOREACH_SAFE(req, &cq->req_list, entry, next) {
 169         NvmeSQueue *sq;
 170         hwaddr addr;
 171
 172         if (nvme_cq_full(cq)) {
 173             break;
 174         }
 175
 176         QTAILQ_REMOVE(&cq->req_list, req, entry);
 177         sq = req->sq;
 178         req->cqe.status = cpu_to_le16((req->status << 1) | cq->phase);
 179         req->cqe.sq_id = cpu_to_le16(sq->sqid);
 180         req->cqe.sq_head = cpu_to_le16(sq->head);
 181         addr = cq->dma_addr + cq->tail * n->cqe_size;
 182         nvme_inc_cq_tail(cq);
 183         pci_dma_write(&n->parent_obj, addr, (void *)&req->cqe,
 184             sizeof(req->cqe));
 185         QTAILQ_INSERT_TAIL(&sq->req_list, req, entry);
 186     }
 187     nvme_isr_notify(n, cq);
 188 }
 189
 190 static void nvme_enqueue_req_completion(NvmeCQueue *cq, NvmeRequest *req)
 191 {
 192     assert(cq->cqid == req->sq->cqid);
 193     QTAILQ_REMOVE(&req->sq->out_req_list, req, entry);
 194     QTAILQ_INSERT_TAIL(&cq->req_list, req, entry);
 195     timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 196 }
 197
 198 static void nvme_rw_cb(void *opaque, int ret)
 199 {
 200     NvmeRequest *req = opaque;
 201     NvmeSQueue *sq = req->sq;
 202     NvmeCtrl *n = sq->ctrl;
 203     NvmeCQueue *cq = n->cq[sq->cqid];
 204
 205     if (!ret) {
 206         block_acct_done(blk_get_stats(n->conf.blk), &req->acct);
 207         req->status = NVME_SUCCESS;
 208     } else {
 209         block_acct_failed(blk_get_stats(n->conf.blk), &req->acct);
 210         req->status = NVME_INTERNAL_DEV_ERROR;
 211     }
 212     if (req->has_sg) {
 213         qemu_sglist_destroy(&req->qsg);
 214     }
 215     nvme_enqueue_req_completion(cq, req);
 216 }
 217
 218 static uint16_t nvme_flush(NvmeCtrl *n, NvmeNamespace *ns, NvmeCmd *cmd,
 219     NvmeRequest *req)
 220 {
 221     req->has_sg = false;
 222     block_acct_start(blk_get_stats(n->conf.blk), &req->acct, 0,
 223          BLOCK_ACCT_FLUSH);
 224     req->aiocb = blk_aio_flush(n->conf.blk, nvme_rw_cb, req);
 225
 226     return NVME_NO_COMPLETE;
 227 }
 228
 229 static uint16_t nvme_rw(NvmeCtrl *n, NvmeNamespace *ns, NvmeCmd *cmd,
 230     NvmeRequest *req)
 231 {
 232     NvmeRwCmd *rw = (NvmeRwCmd *)cmd;
 233     uint32_t nlb  = le32_to_cpu(rw->nlb) + 1;
 234     uint64_t slba = le64_to_cpu(rw->slba);
 235     uint64_t prp1 = le64_to_cpu(rw->prp1);
 236     uint64_t prp2 = le64_to_cpu(rw->prp2);
 237
 238     uint8_t lba_index  = NVME_ID_NS_FLBAS_INDEX(ns->id_ns.flbas);
 239     uint8_t data_shift = ns->id_ns.lbaf[lba_index].ds;
 240     uint64_t data_size = (uint64_t)nlb << data_shift;
 241     uint64_t aio_slba  = slba << (data_shift - BDRV_SECTOR_BITS);
 242     int is_write = rw->opcode == NVME_CMD_WRITE ? 1 : 0;
 243     enum BlockAcctType acct = is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ;
 244
 245     if ((slba + nlb) > ns->id_ns.nsze) {
 246         block_acct_invalid(blk_get_stats(n->conf.blk), acct);
 247         return NVME_LBA_RANGE | NVME_DNR;
 248     }
 249
 250     if (nvme_map_prp(&req->qsg, prp1, prp2, data_size, n)) {
 251         block_acct_invalid(blk_get_stats(n->conf.blk), acct);
 252         return NVME_INVALID_FIELD | NVME_DNR;
 253     }
 254
 255     assert((nlb << data_shift) == req->qsg.size);
 256
 257     req->has_sg = true;
 258     dma_acct_start(n->conf.blk, &req->acct, &req->qsg, acct);
 259     req->aiocb = is_write ?
 260         dma_blk_write(n->conf.blk, &req->qsg, aio_slba, nvme_rw_cb, req) :
 261         dma_blk_read(n->conf.blk, &req->qsg, aio_slba, nvme_rw_cb, req);
 262
 263     return NVME_NO_COMPLETE;
 264 }
 265
 266 static uint16_t nvme_io_cmd(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 267 {
 268     NvmeNamespace *ns;
 269     uint32_t nsid = le32_to_cpu(cmd->nsid);
 270
 271     if (nsid == 0 || nsid > n->num_namespaces) {
 272         return NVME_INVALID_NSID | NVME_DNR;
 273     }
 274
 275     ns = &n->namespaces[nsid - 1];
 276     switch (cmd->opcode) {
 277     case NVME_CMD_FLUSH:
 278         return nvme_flush(n, ns, cmd, req);
 279     case NVME_CMD_WRITE:
 280     case NVME_CMD_READ:
 281         return nvme_rw(n, ns, cmd, req);
 282     default:
 283         return NVME_INVALID_OPCODE | NVME_DNR;
 284     }
 285 }
 286
 287 static void nvme_free_sq(NvmeSQueue *sq, NvmeCtrl *n)
 288 {
 289     n->sq[sq->sqid] = NULL;
 290     timer_del(sq->timer);
 291     timer_free(sq->timer);
 292     g_free(sq->io_req);
 293     if (sq->sqid) {
 294         g_free(sq);
 295     }
 296 }
 297
 298 static uint16_t nvme_del_sq(NvmeCtrl *n, NvmeCmd *cmd)
 299 {
 300     NvmeDeleteQ *c = (NvmeDeleteQ *)cmd;
 301     NvmeRequest *req, *next;
 302     NvmeSQueue *sq;
 303     NvmeCQueue *cq;
 304     uint16_t qid = le16_to_cpu(c->qid);
 305
 306     if (!qid || nvme_check_sqid(n, qid)) {
 307         return NVME_INVALID_QID | NVME_DNR;
 308     }
 309
 310     sq = n->sq[qid];
 311     while (!QTAILQ_EMPTY(&sq->out_req_list)) {
 312         req = QTAILQ_FIRST(&sq->out_req_list);
 313         assert(req->aiocb);
 314         blk_aio_cancel(req->aiocb);
 315     }
 316     if (!nvme_check_cqid(n, sq->cqid)) {
 317         cq = n->cq[sq->cqid];
 318         QTAILQ_REMOVE(&cq->sq_list, sq, entry);
 319
 320         nvme_post_cqes(cq);
 321         QTAILQ_FOREACH_SAFE(req, &cq->req_list, entry, next) {
 322             if (req->sq == sq) {
 323                 QTAILQ_REMOVE(&cq->req_list, req, entry);
 324                 QTAILQ_INSERT_TAIL(&sq->req_list, req, entry);
 325             }
 326         }
 327     }
 328
 329     nvme_free_sq(sq, n);
 330     return NVME_SUCCESS;
 331 }
 332
 333 static void nvme_init_sq(NvmeSQueue *sq, NvmeCtrl *n, uint64_t dma_addr,
 334     uint16_t sqid, uint16_t cqid, uint16_t size)
 335 {
 336     int i;
 337     NvmeCQueue *cq;
 338
 339     sq->ctrl = n;
 340     sq->dma_addr = dma_addr;
 341     sq->sqid = sqid;
 342     sq->size = size;
 343     sq->cqid = cqid;
 344     sq->head = sq->tail = 0;
 345     sq->io_req = g_new(NvmeRequest, sq->size);
 346
 347     QTAILQ_INIT(&sq->req_list);
 348     QTAILQ_INIT(&sq->out_req_list);
 349     for (i = 0; i < sq->size; i++) {
 350         sq->io_req[i].sq = sq;
 351         QTAILQ_INSERT_TAIL(&(sq->req_list), &sq->io_req[i], entry);
 352     }
 353     sq->timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, nvme_process_sq, sq);
 354
 355     assert(n->cq[cqid]);
 356     cq = n->cq[cqid];
 357     QTAILQ_INSERT_TAIL(&(cq->sq_list), sq, entry);
 358     n->sq[sqid] = sq;
 359 }
 360
 361 static uint16_t nvme_create_sq(NvmeCtrl *n, NvmeCmd *cmd)
 362 {
 363     NvmeSQueue *sq;
 364     NvmeCreateSq *c = (NvmeCreateSq *)cmd;
 365
 366     uint16_t cqid = le16_to_cpu(c->cqid);
 367     uint16_t sqid = le16_to_cpu(c->sqid);
 368     uint16_t qsize = le16_to_cpu(c->qsize);
 369     uint16_t qflags = le16_to_cpu(c->sq_flags);
 370     uint64_t prp1 = le64_to_cpu(c->prp1);
 371
 372     if (!cqid || nvme_check_cqid(n, cqid)) {
 373         return NVME_INVALID_CQID | NVME_DNR;
 374     }
 375     if (!sqid || (sqid && !nvme_check_sqid(n, sqid))) {
 376         return NVME_INVALID_QID | NVME_DNR;
 377     }
 378     if (!qsize || qsize > NVME_CAP_MQES(n->bar.cap)) {
 379         return NVME_MAX_QSIZE_EXCEEDED | NVME_DNR;
 380     }
 381     if (!prp1 || prp1 & (n->page_size - 1)) {
 382         return NVME_INVALID_FIELD | NVME_DNR;
 383     }
 384     if (!(NVME_SQ_FLAGS_PC(qflags))) {
 385         return NVME_INVALID_FIELD | NVME_DNR;
 386     }
 387     sq = g_malloc0(sizeof(*sq));
 388     nvme_init_sq(sq, n, prp1, sqid, cqid, qsize + 1);
 389     return NVME_SUCCESS;
 390 }
 391
 392 static void nvme_free_cq(NvmeCQueue *cq, NvmeCtrl *n)
 393 {
 394     n->cq[cq->cqid] = NULL;
 395     timer_del(cq->timer);
 396     timer_free(cq->timer);
 397     msix_vector_unuse(&n->parent_obj, cq->vector);
 398     if (cq->cqid) {
 399         g_free(cq);
 400     }
 401 }
 402
 403 static uint16_t nvme_del_cq(NvmeCtrl *n, NvmeCmd *cmd)
 404 {
 405     NvmeDeleteQ *c = (NvmeDeleteQ *)cmd;
 406     NvmeCQueue *cq;
 407     uint16_t qid = le16_to_cpu(c->qid);
 408
 409     if (!qid || nvme_check_cqid(n, qid)) {
 410         return NVME_INVALID_CQID | NVME_DNR;
 411     }
 412
 413     cq = n->cq[qid];
 414     if (!QTAILQ_EMPTY(&cq->sq_list)) {
 415         return NVME_INVALID_QUEUE_DEL;
 416     }
 417     nvme_free_cq(cq, n);
 418     return NVME_SUCCESS;
 419 }
 420
 421 static void nvme_init_cq(NvmeCQueue *cq, NvmeCtrl *n, uint64_t dma_addr,
 422     uint16_t cqid, uint16_t vector, uint16_t size, uint16_t irq_enabled)
 423 {
 424     cq->ctrl = n;
 425     cq->cqid = cqid;
 426     cq->size = size;
 427     cq->dma_addr = dma_addr;
 428     cq->phase = 1;
 429     cq->irq_enabled = irq_enabled;
 430     cq->vector = vector;
 431     cq->head = cq->tail = 0;
 432     QTAILQ_INIT(&cq->req_list);
 433     QTAILQ_INIT(&cq->sq_list);
 434     msix_vector_use(&n->parent_obj, cq->vector);
 435     n->cq[cqid] = cq;
 436     cq->timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, nvme_post_cqes, cq);
 437 }
 438
 439 static uint16_t nvme_create_cq(NvmeCtrl *n, NvmeCmd *cmd)
 440 {
 441     NvmeCQueue *cq;
 442     NvmeCreateCq *c = (NvmeCreateCq *)cmd;
 443     uint16_t cqid = le16_to_cpu(c->cqid);
 444     uint16_t vector = le16_to_cpu(c->irq_vector);
 445     uint16_t qsize = le16_to_cpu(c->qsize);
 446     uint16_t qflags = le16_to_cpu(c->cq_flags);
 447     uint64_t prp1 = le64_to_cpu(c->prp1);
 448
 449     if (!cqid || (cqid && !nvme_check_cqid(n, cqid))) {
 450         return NVME_INVALID_CQID | NVME_DNR;
 451     }
 452     if (!qsize || qsize > NVME_CAP_MQES(n->bar.cap)) {
 453         return NVME_MAX_QSIZE_EXCEEDED | NVME_DNR;
 454     }
 455     if (!prp1) {
 456         return NVME_INVALID_FIELD | NVME_DNR;
 457     }
 458     if (vector > n->num_queues) {
 459         return NVME_INVALID_IRQ_VECTOR | NVME_DNR;
 460     }
 461     if (!(NVME_CQ_FLAGS_PC(qflags))) {
 462         return NVME_INVALID_FIELD | NVME_DNR;
 463     }
 464
 465     cq = g_malloc0(sizeof(*cq));
 466     nvme_init_cq(cq, n, prp1, cqid, vector, qsize + 1,
 467         NVME_CQ_FLAGS_IEN(qflags));
 468     return NVME_SUCCESS;
 469 }
 470
 471 static uint16_t nvme_identify(NvmeCtrl *n, NvmeCmd *cmd)
 472 {
 473     NvmeNamespace *ns;
 474     NvmeIdentify *c = (NvmeIdentify *)cmd;
 475     uint32_t cns  = le32_to_cpu(c->cns);
 476     uint32_t nsid = le32_to_cpu(c->nsid);
 477     uint64_t prp1 = le64_to_cpu(c->prp1);
 478     uint64_t prp2 = le64_to_cpu(c->prp2);
 479
 480     if (cns) {
 481         return nvme_dma_read_prp(n, (uint8_t *)&n->id_ctrl, sizeof(n->id_ctrl),
 482             prp1, prp2);
 483     }
 484     if (nsid == 0 || nsid > n->num_namespaces) {
 485         return NVME_INVALID_NSID | NVME_DNR;
 486     }
 487
 488     ns = &n->namespaces[nsid - 1];
 489     return nvme_dma_read_prp(n, (uint8_t *)&ns->id_ns, sizeof(ns->id_ns),
 490         prp1, prp2);
 491 }
 492
 493 static uint16_t nvme_get_feature(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 494 {
 495     uint32_t dw10 = le32_to_cpu(cmd->cdw10);
 496     uint32_t result;
 497
 498     switch (dw10) {
 499     case NVME_VOLATILE_WRITE_CACHE:
 500         result = blk_enable_write_cache(n->conf.blk);
 501         break;
 502     case NVME_NUMBER_OF_QUEUES:
 503         result = cpu_to_le32((n->num_queues - 1) | ((n->num_queues - 1) << 16));
 504         break;
 505     default:
 506         return NVME_INVALID_FIELD | NVME_DNR;
 507     }
 508
 509     req->cqe.result = result;
 510     return NVME_SUCCESS;
 511 }
 512
 513 static uint16_t nvme_set_feature(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 514 {
 515     uint32_t dw10 = le32_to_cpu(cmd->cdw10);
 516     uint32_t dw11 = le32_to_cpu(cmd->cdw11);
 517
 518     switch (dw10) {
 519     case NVME_VOLATILE_WRITE_CACHE:
 520         blk_set_enable_write_cache(n->conf.blk, dw11 & 1);
 521         break;
 522     case NVME_NUMBER_OF_QUEUES:
 523         req->cqe.result =
 524             cpu_to_le32((n->num_queues - 1) | ((n->num_queues - 1) << 16));
 525         break;
 526     default:
 527         return NVME_INVALID_FIELD | NVME_DNR;
 528     }
 529     return NVME_SUCCESS;
 530 }
 531
 532 static uint16_t nvme_admin_cmd(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 533 {
 534     switch (cmd->opcode) {
 535     case NVME_ADM_CMD_DELETE_SQ:
 536         return nvme_del_sq(n, cmd);
 537     case NVME_ADM_CMD_CREATE_SQ:
 538         return nvme_create_sq(n, cmd);
 539     case NVME_ADM_CMD_DELETE_CQ:
 540         return nvme_del_cq(n, cmd);
 541     case NVME_ADM_CMD_CREATE_CQ:
 542         return nvme_create_cq(n, cmd);
 543     case NVME_ADM_CMD_IDENTIFY:
 544         return nvme_identify(n, cmd);
 545     case NVME_ADM_CMD_SET_FEATURES:
 546         return nvme_set_feature(n, cmd, req);
 547     case NVME_ADM_CMD_GET_FEATURES:
 548         return nvme_get_feature(n, cmd, req);
 549     default:
 550         return NVME_INVALID_OPCODE | NVME_DNR;
 551     }
 552 }
 553
 554 static void nvme_process_sq(void *opaque)
 555 {
 556     NvmeSQueue *sq = opaque;
 557     NvmeCtrl *n = sq->ctrl;
 558     NvmeCQueue *cq = n->cq[sq->cqid];
 559
 560     uint16_t status;
 561     hwaddr addr;
 562     NvmeCmd cmd;
 563     NvmeRequest *req;
 564
 565     while (!(nvme_sq_empty(sq) || QTAILQ_EMPTY(&sq->req_list))) {
 566         addr = sq->dma_addr + sq->head * n->sqe_size;
 567         pci_dma_read(&n->parent_obj, addr, (void *)&cmd, sizeof(cmd));
 568         nvme_inc_sq_head(sq);
 569
 570         req = QTAILQ_FIRST(&sq->req_list);
 571         QTAILQ_REMOVE(&sq->req_list, req, entry);
 572         QTAILQ_INSERT_TAIL(&sq->out_req_list, req, entry);
 573         memset(&req->cqe, 0, sizeof(req->cqe));
 574         req->cqe.cid = cmd.cid;
 575
 576         status = sq->sqid ? nvme_io_cmd(n, &cmd, req) :
 577             nvme_admin_cmd(n, &cmd, req);
 578         if (status != NVME_NO_COMPLETE) {
 579             req->status = status;
 580             nvme_enqueue_req_completion(cq, req);
 581         }
 582     }
 583 }
 584
 585 static void nvme_clear_ctrl(NvmeCtrl *n)
 586 {
 587     int i;
 588
 589     for (i = 0; i < n->num_queues; i++) {
 590         if (n->sq[i] != NULL) {
 591             nvme_free_sq(n->sq[i], n);
 592         }
 593     }
 594     for (i = 0; i < n->num_queues; i++) {
 595         if (n->cq[i] != NULL) {
 596             nvme_free_cq(n->cq[i], n);
 597         }
 598     }
 599
 600     blk_flush(n->conf.blk);
 601     n->bar.cc = 0;
 602 }
 603
 604 static int nvme_start_ctrl(NvmeCtrl *n)
 605 {
 606     uint32_t page_bits = NVME_CC_MPS(n->bar.cc) + 12;
 607     uint32_t page_size = 1 << page_bits;
 608
 609     if (n->cq[0] || n->sq[0] || !n->bar.asq || !n->bar.acq ||
 610             n->bar.asq & (page_size - 1) || n->bar.acq & (page_size - 1) ||
 611             NVME_CC_MPS(n->bar.cc) < NVME_CAP_MPSMIN(n->bar.cap) ||
 612             NVME_CC_MPS(n->bar.cc) > NVME_CAP_MPSMAX(n->bar.cap) ||
 613             NVME_CC_IOCQES(n->bar.cc) < NVME_CTRL_CQES_MIN(n->id_ctrl.cqes) ||
 614             NVME_CC_IOCQES(n->bar.cc) > NVME_CTRL_CQES_MAX(n->id_ctrl.cqes) ||
 615             NVME_CC_IOSQES(n->bar.cc) < NVME_CTRL_SQES_MIN(n->id_ctrl.sqes) ||
 616             NVME_CC_IOSQES(n->bar.cc) > NVME_CTRL_SQES_MAX(n->id_ctrl.sqes) ||
 617             !NVME_AQA_ASQS(n->bar.aqa) || !NVME_AQA_ACQS(n->bar.aqa)) {
 618         return -1;
 619     }
 620
 621     n->page_bits = page_bits;
 622     n->page_size = page_size;
 623     n->max_prp_ents = n->page_size / sizeof(uint64_t);
 624     n->cqe_size = 1 << NVME_CC_IOCQES(n->bar.cc);
 625     n->sqe_size = 1 << NVME_CC_IOSQES(n->bar.cc);
 626     nvme_init_cq(&n->admin_cq, n, n->bar.acq, 0, 0,
 627         NVME_AQA_ACQS(n->bar.aqa) + 1, 1);
 628     nvme_init_sq(&n->admin_sq, n, n->bar.asq, 0, 0,
 629         NVME_AQA_ASQS(n->bar.aqa) + 1);
 630
 631     return 0;
 632 }
 633
 634 static void nvme_write_bar(NvmeCtrl *n, hwaddr offset, uint64_t data,
 635     unsigned size)
 636 {
 637     switch (offset) {
 638     case 0xc:
 639         n->bar.intms |= data & 0xffffffff;
 640         n->bar.intmc = n->bar.intms;
 641         break;
 642     case 0x10:
 643         n->bar.intms &= ~(data & 0xffffffff);
 644         n->bar.intmc = n->bar.intms;
 645         break;
 646     case 0x14:
 647         /* Windows first sends data, then sends enable bit */
 648         if (!NVME_CC_EN(data) && !NVME_CC_EN(n->bar.cc) &&
 649             !NVME_CC_SHN(data) && !NVME_CC_SHN(n->bar.cc))
 650         {
 651             n->bar.cc = data;
 652         }
 653
 654         if (NVME_CC_EN(data) && !NVME_CC_EN(n->bar.cc)) {
 655             n->bar.cc = data;
 656             if (nvme_start_ctrl(n)) {
 657                 n->bar.csts = NVME_CSTS_FAILED;
 658             } else {
 659                 n->bar.csts = NVME_CSTS_READY;
 660             }
 661         } else if (!NVME_CC_EN(data) && NVME_CC_EN(n->bar.cc)) {
 662             nvme_clear_ctrl(n);
 663             n->bar.csts &= ~NVME_CSTS_READY;
 664         }
 665         if (NVME_CC_SHN(data) && !(NVME_CC_SHN(n->bar.cc))) {
 666                 nvme_clear_ctrl(n);
 667                 n->bar.cc = data;
 668                 n->bar.csts |= NVME_CSTS_SHST_COMPLETE;
 669         } else if (!NVME_CC_SHN(data) && NVME_CC_SHN(n->bar.cc)) {
 670                 n->bar.csts &= ~NVME_CSTS_SHST_COMPLETE;
 671                 n->bar.cc = data;
 672         }
 673         break;
 674     case 0x24:
 675         n->bar.aqa = data & 0xffffffff;
 676         break;
 677     case 0x28:
 678         n->bar.asq = data;
 679         break;
 680     case 0x2c:
 681         n->bar.asq |= data << 32;
 682         break;
 683     case 0x30:
 684         n->bar.acq = data;
 685         break;
 686     case 0x34:
 687         n->bar.acq |= data << 32;
 688         break;
 689     default:
 690         break;
 691     }
 692 }
 693
 694 static uint64_t nvme_mmio_read(void *opaque, hwaddr addr, unsigned size)
 695 {
 696     NvmeCtrl *n = (NvmeCtrl *)opaque;
 697     uint8_t *ptr = (uint8_t *)&n->bar;
 698     uint64_t val = 0;
 699
 700     if (addr < sizeof(n->bar)) {
 701         memcpy(&val, ptr + addr, size);
 702     }
 703     return val;
 704 }
 705
 706 static void nvme_process_db(NvmeCtrl *n, hwaddr addr, int val)
 707 {
 708     uint32_t qid;
 709
 710     if (addr & ((1 << 2) - 1)) {
 711         return;
 712     }
 713
 714     if (((addr - 0x1000) >> 2) & 1) {
 715         uint16_t new_head = val & 0xffff;
 716         int start_sqs;
 717         NvmeCQueue *cq;
 718
 719         qid = (addr - (0x1000 + (1 << 2))) >> 3;
 720         if (nvme_check_cqid(n, qid)) {
 721             return;
 722         }
 723
 724         cq = n->cq[qid];
 725         if (new_head >= cq->size) {
 726             return;
 727         }
 728
 729         start_sqs = nvme_cq_full(cq) ? 1 : 0;
 730         cq->head = new_head;
 731         if (start_sqs) {
 732             NvmeSQueue *sq;
 733             QTAILQ_FOREACH(sq, &cq->sq_list, entry) {
 734                 timer_mod(sq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 735             }
 736             timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 737         }
 738
 739         if (cq->tail != cq->head) {
 740             nvme_isr_notify(n, cq);
 741         }
 742     } else {
 743         uint16_t new_tail = val & 0xffff;
 744         NvmeSQueue *sq;
 745
 746         qid = (addr - 0x1000) >> 3;
 747         if (nvme_check_sqid(n, qid)) {
 748             return;
 749         }
 750
 751         sq = n->sq[qid];
 752         if (new_tail >= sq->size) {
 753             return;
 754         }
 755
 756         sq->tail = new_tail;
 757         timer_mod(sq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 758     }
 759 }
 760
 761 static void nvme_mmio_write(void *opaque, hwaddr addr, uint64_t data,
 762     unsigned size)
 763 {
 764     NvmeCtrl *n = (NvmeCtrl *)opaque;
 765     if (addr < sizeof(n->bar)) {
 766         nvme_write_bar(n, addr, data, size);
 767     } else if (addr >= 0x1000) {
 768         nvme_process_db(n, addr, data);
 769     }
 770 }
 771
 772 static const MemoryRegionOps nvme_mmio_ops = {
 773     .read = nvme_mmio_read,
 774     .write = nvme_mmio_write,
 775     .endianness = DEVICE_LITTLE_ENDIAN,
 776     .impl = {
 777         .min_access_size = 2,
 778         .max_access_size = 8,
 779     },
 780 };
 781
 782 static int nvme_init(PCIDevice *pci_dev)
 783 {
 784     NvmeCtrl *n = NVME(pci_dev);
 785     NvmeIdCtrl *id = &n->id_ctrl;
 786
 787     int i;
 788     int64_t bs_size;
 789     uint8_t *pci_conf;
 790
 791     if (!n->conf.blk) {
 792         return -1;
 793     }
 794
 795     bs_size = blk_getlength(n->conf.blk);
 796     if (bs_size < 0) {
 797         return -1;
 798     }
 799
 800     blkconf_serial(&n->conf, &n->serial);
 801     if (!n->serial) {
 802         return -1;
 803     }
 804     blkconf_blocksizes(&n->conf);
 805
 806     pci_conf = pci_dev->config;
 807     pci_conf[PCI_INTERRUPT_PIN] = 1;
 808     pci_config_set_prog_interface(pci_dev->config, 0x2);
 809     pci_config_set_class(pci_dev->config, PCI_CLASS_STORAGE_EXPRESS);
 810     pcie_endpoint_cap_init(&n->parent_obj, 0x80);
 811
 812     n->num_namespaces = 1;
 813     n->num_queues = 64;
 814     n->reg_size = pow2ceil(0x1004 + 2 * (n->num_queues + 1) * 4);
 815     n->ns_size = bs_size / (uint64_t)n->num_namespaces;
 816
 817     n->namespaces = g_new0(NvmeNamespace, n->num_namespaces);
 818     n->sq = g_new0(NvmeSQueue *, n->num_queues);
 819     n->cq = g_new0(NvmeCQueue *, n->num_queues);
 820
 821     memory_region_init_io(&n->iomem, OBJECT(n), &nvme_mmio_ops, n,
 822                           "nvme", n->reg_size);
 823     pci_register_bar(&n->parent_obj, 0,
 824         PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_64,
 825         &n->iomem);
 826     msix_init_exclusive_bar(&n->parent_obj, n->num_queues, 4);
 827
 828     id->vid = cpu_to_le16(pci_get_word(pci_conf + PCI_VENDOR_ID));
 829     id->ssvid = cpu_to_le16(pci_get_word(pci_conf + PCI_SUBSYSTEM_VENDOR_ID));
 830     strpadcpy((char *)id->mn, sizeof(id->mn), "QEMU NVMe Ctrl", ' ');
 831     strpadcpy((char *)id->fr, sizeof(id->fr), "1.0", ' ');
 832     strpadcpy((char *)id->sn, sizeof(id->sn), n->serial, ' ');
 833     id->rab = 6;
 834     id->ieee[0] = 0x00;
 835     id->ieee[1] = 0x02;
 836     id->ieee[2] = 0xb3;
 837     id->oacs = cpu_to_le16(0);
 838     id->frmw = 7 << 1;
 839     id->lpa = 1 << 0;
 840     id->sqes = (0x6 << 4) | 0x6;
 841     id->cqes = (0x4 << 4) | 0x4;
 842     id->nn = cpu_to_le32(n->num_namespaces);
 843     id->psd[0].mp = cpu_to_le16(0x9c4);
 844     id->psd[0].enlat = cpu_to_le32(0x10);
 845     id->psd[0].exlat = cpu_to_le32(0x4);
 846     if (blk_enable_write_cache(n->conf.blk)) {
 847         id->vwc = 1;
 848     }
 849
 850     n->bar.cap = 0;
 851     NVME_CAP_SET_MQES(n->bar.cap, 0x7ff);
 852     NVME_CAP_SET_CQR(n->bar.cap, 1);
 853     NVME_CAP_SET_AMS(n->bar.cap, 1);
 854     NVME_CAP_SET_TO(n->bar.cap, 0xf);
 855     NVME_CAP_SET_CSS(n->bar.cap, 1);
 856     NVME_CAP_SET_MPSMAX(n->bar.cap, 4);
 857
 858     n->bar.vs = 0x00010100;
 859     n->bar.intmc = n->bar.intms = 0;
 860
 861     for (i = 0; i < n->num_namespaces; i++) {
 862         NvmeNamespace *ns = &n->namespaces[i];
 863         NvmeIdNs *id_ns = &ns->id_ns;
 864         id_ns->nsfeat = 0;
 865         id_ns->nlbaf = 0;
 866         id_ns->flbas = 0;
 867         id_ns->mc = 0;
 868         id_ns->dpc = 0;
 869         id_ns->dps = 0;
 870         id_ns->lbaf[0].ds = BDRV_SECTOR_BITS;
 871         id_ns->ncap  = id_ns->nuse = id_ns->nsze =
 872             cpu_to_le64(n->ns_size >>
 873                 id_ns->lbaf[NVME_ID_NS_FLBAS_INDEX(ns->id_ns.flbas)].ds);
 874     }
 875     return 0;
 876 }
 877
 878 static void nvme_exit(PCIDevice *pci_dev)
 879 {
 880     NvmeCtrl *n = NVME(pci_dev);
 881
 882     nvme_clear_ctrl(n);
 883     g_free(n->namespaces);
 884     g_free(n->cq);
 885     g_free(n->sq);
 886     msix_uninit_exclusive_bar(pci_dev);
 887 }
 888
 889 static Property nvme_props[] = {
 890     DEFINE_BLOCK_PROPERTIES(NvmeCtrl, conf),
 891     DEFINE_PROP_STRING("serial", NvmeCtrl, serial),
 892     DEFINE_PROP_END_OF_LIST(),
 893 };
 894
 895 static const VMStateDescription nvme_vmstate = {
 896     .name = "nvme",
 897     .unmigratable = 1,
 898 };
 899
 900 static void nvme_class_init(ObjectClass *oc, void *data)
 901 {
 902     DeviceClass *dc = DEVICE_CLASS(oc);
 903     PCIDeviceClass *pc = PCI_DEVICE_CLASS(oc);
 904
 905     pc->init = nvme_init;
 906     pc->exit = nvme_exit;
 907     pc->class_id = PCI_CLASS_STORAGE_EXPRESS;
 908     pc->vendor_id = PCI_VENDOR_ID_INTEL;
 909     pc->device_id = 0x5845;
 910     pc->revision = 1;
 911     pc->is_express = 1;
 912
 913     set_bit(DEVICE_CATEGORY_STORAGE, dc->categories);
 914     dc->desc = "Non-Volatile Memory Express";
 915     dc->props = nvme_props;
 916     dc->vmsd = &nvme_vmstate;
 917 }
 918
 919 static void nvme_instance_init(Object *obj)
 920 {
 921     NvmeCtrl *s = NVME(obj);
 922
 923     device_add_bootindex_property(obj, &s->conf.bootindex,
 924                                   "bootindex", "/namespace@1,0",
 925                                   DEVICE(obj), &error_abort);
 926 }
 927
 928 static const TypeInfo nvme_info = {
 929     .name          = "nvme",
 930     .parent        = TYPE_PCI_DEVICE,
 931     .instance_size = sizeof(NvmeCtrl),
 932     .class_init    = nvme_class_init,
 933     .instance_init = nvme_instance_init,
 934 };
 935
 936 static void nvme_register_types(void)
 937 {
 938     type_register_static(&nvme_info);
 939 }
 940
 941 type_init(nvme_register_types)