hw/block/nvme.c

   1 /*
   2  * QEMU NVM Express Controller
   3  *
   4  * Copyright (c) 2012, Intel Corporation
   5  *
   6  * Written by Keith Busch <[email protected]>
   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/error.h"
  30 #include "qapi/visitor.h"
  31 #include "sysemu/block-backend.h"
  32
  33 #include "nvme.h"
  34
  35 static void nvme_process_sq(void *opaque);
  36
  37 static int nvme_check_sqid(NvmeCtrl *n, uint16_t sqid)
  38 {
  39     return sqid < n->num_queues && n->sq[sqid] != NULL ? 0 : -1;
  40 }
  41
  42 static int nvme_check_cqid(NvmeCtrl *n, uint16_t cqid)
  43 {
  44     return cqid < n->num_queues && n->cq[cqid] != NULL ? 0 : -1;
  45 }
  46
  47 static void nvme_inc_cq_tail(NvmeCQueue *cq)
  48 {
  49     cq->tail++;
  50     if (cq->tail >= cq->size) {
  51         cq->tail = 0;
  52         cq->phase = !cq->phase;
  53     }
  54 }
  55
  56 static void nvme_inc_sq_head(NvmeSQueue *sq)
  57 {
  58     sq->head = (sq->head + 1) % sq->size;
  59 }
  60
  61 static uint8_t nvme_cq_full(NvmeCQueue *cq)
  62 {
  63     return (cq->tail + 1) % cq->size == cq->head;
  64 }
  65
  66 static uint8_t nvme_sq_empty(NvmeSQueue *sq)
  67 {
  68     return sq->head == sq->tail;
  69 }
  70
  71 static void nvme_isr_notify(NvmeCtrl *n, NvmeCQueue *cq)
  72 {
  73     if (cq->irq_enabled) {
  74         if (msix_enabled(&(n->parent_obj))) {
  75             msix_notify(&(n->parent_obj), cq->vector);
  76         } else {
  77             pci_irq_pulse(&n->parent_obj);
  78         }
  79     }
  80 }
  81
  82 static uint16_t nvme_map_prp(QEMUSGList *qsg, uint64_t prp1, uint64_t prp2,
  83     uint32_t len, NvmeCtrl *n)
  84 {
  85     hwaddr trans_len = n->page_size - (prp1 % n->page_size);
  86     trans_len = MIN(len, trans_len);
  87     int num_prps = (len >> n->page_bits) + 1;
  88
  89     if (!prp1) {
  90         return NVME_INVALID_FIELD | NVME_DNR;
  91     }
  92
  93     pci_dma_sglist_init(qsg, &n->parent_obj, num_prps);
  94     qemu_sglist_add(qsg, prp1, trans_len);
  95     len -= trans_len;
  96     if (len) {
  97         if (!prp2) {
  98             goto unmap;
  99         }
 100         if (len > n->page_size) {
 101             uint64_t prp_list[n->max_prp_ents];
 102             uint32_t nents, prp_trans;
 103             int i = 0;
 104
 105             nents = (len + n->page_size - 1) >> n->page_bits;
 106             prp_trans = MIN(n->max_prp_ents, nents) * sizeof(uint64_t);
 107             pci_dma_read(&n->parent_obj, prp2, (void *)prp_list, prp_trans);
 108             while (len != 0) {
 109                 uint64_t prp_ent = le64_to_cpu(prp_list[i]);
 110
 111                 if (i == n->max_prp_ents - 1 && len > n->page_size) {
 112                     if (!prp_ent || prp_ent & (n->page_size - 1)) {
 113                         goto unmap;
 114                     }
 115
 116                     i = 0;
 117                     nents = (len + n->page_size - 1) >> n->page_bits;
 118                     prp_trans = MIN(n->max_prp_ents, nents) * sizeof(uint64_t);
 119                     pci_dma_read(&n->parent_obj, prp_ent, (void *)prp_list,
 120                         prp_trans);
 121                     prp_ent = le64_to_cpu(prp_list[i]);
 122                 }
 123
 124                 if (!prp_ent || prp_ent & (n->page_size - 1)) {
 125                     goto unmap;
 126                 }
 127
 128                 trans_len = MIN(len, n->page_size);
 129                 qemu_sglist_add(qsg, prp_ent, trans_len);
 130                 len -= trans_len;
 131                 i++;
 132             }
 133         } else {
 134             if (prp2 & (n->page_size - 1)) {
 135                 goto unmap;
 136             }
 137             qemu_sglist_add(qsg, prp2, len);
 138         }
 139     }
 140     return NVME_SUCCESS;
 141
 142  unmap:
 143     qemu_sglist_destroy(qsg);
 144     return NVME_INVALID_FIELD | NVME_DNR;
 145 }
 146
 147 static uint16_t nvme_dma_read_prp(NvmeCtrl *n, uint8_t *ptr, uint32_t len,
 148     uint64_t prp1, uint64_t prp2)
 149 {
 150     QEMUSGList qsg;
 151
 152     if (nvme_map_prp(&qsg, prp1, prp2, len, n)) {
 153         return NVME_INVALID_FIELD | NVME_DNR;
 154     }
 155     if (dma_buf_read(ptr, len, &qsg)) {
 156         qemu_sglist_destroy(&qsg);
 157         return NVME_INVALID_FIELD | NVME_DNR;
 158     }
 159     qemu_sglist_destroy(&qsg);
 160     return NVME_SUCCESS;
 161 }
 162
 163 static void nvme_post_cqes(void *opaque)
 164 {
 165     NvmeCQueue *cq = opaque;
 166     NvmeCtrl *n = cq->ctrl;
 167     NvmeRequest *req, *next;
 168
 169     QTAILQ_FOREACH_SAFE(req, &cq->req_list, entry, next) {
 170         NvmeSQueue *sq;
 171         hwaddr addr;
 172
 173         if (nvme_cq_full(cq)) {
 174             break;
 175         }
 176
 177         QTAILQ_REMOVE(&cq->req_list, req, entry);
 178         sq = req->sq;
 179         req->cqe.status = cpu_to_le16((req->status << 1) | cq->phase);
 180         req->cqe.sq_id = cpu_to_le16(sq->sqid);
 181         req->cqe.sq_head = cpu_to_le16(sq->head);
 182         addr = cq->dma_addr + cq->tail * n->cqe_size;
 183         nvme_inc_cq_tail(cq);
 184         pci_dma_write(&n->parent_obj, addr, (void *)&req->cqe,
 185             sizeof(req->cqe));
 186         QTAILQ_INSERT_TAIL(&sq->req_list, req, entry);
 187     }
 188     nvme_isr_notify(n, cq);
 189 }
 190
 191 static void nvme_enqueue_req_completion(NvmeCQueue *cq, NvmeRequest *req)
 192 {
 193     assert(cq->cqid == req->sq->cqid);
 194     QTAILQ_REMOVE(&req->sq->out_req_list, req, entry);
 195     QTAILQ_INSERT_TAIL(&cq->req_list, req, entry);
 196     timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 197 }
 198
 199 static void nvme_rw_cb(void *opaque, int ret)
 200 {
 201     NvmeRequest *req = opaque;
 202     NvmeSQueue *sq = req->sq;
 203     NvmeCtrl *n = sq->ctrl;
 204     NvmeCQueue *cq = n->cq[sq->cqid];
 205
 206     if (!ret) {
 207         block_acct_done(blk_get_stats(n->conf.blk), &req->acct);
 208         req->status = NVME_SUCCESS;
 209     } else {
 210         block_acct_failed(blk_get_stats(n->conf.blk), &req->acct);
 211         req->status = NVME_INTERNAL_DEV_ERROR;
 212     }
 213     if (req->has_sg) {
 214         qemu_sglist_destroy(&req->qsg);
 215     }
 216     nvme_enqueue_req_completion(cq, req);
 217 }
 218
 219 static uint16_t nvme_flush(NvmeCtrl *n, NvmeNamespace *ns, NvmeCmd *cmd,
 220     NvmeRequest *req)
 221 {
 222     req->has_sg = false;
 223     block_acct_start(blk_get_stats(n->conf.blk), &req->acct, 0,
 224          BLOCK_ACCT_FLUSH);
 225     req->aiocb = blk_aio_flush(n->conf.blk, nvme_rw_cb, req);
 226
 227     return NVME_NO_COMPLETE;
 228 }
 229
 230 static uint16_t nvme_rw(NvmeCtrl *n, NvmeNamespace *ns, NvmeCmd *cmd,
 231     NvmeRequest *req)
 232 {
 233     NvmeRwCmd *rw = (NvmeRwCmd *)cmd;
 234     uint32_t nlb  = le32_to_cpu(rw->nlb) + 1;
 235     uint64_t slba = le64_to_cpu(rw->slba);
 236     uint64_t prp1 = le64_to_cpu(rw->prp1);
 237     uint64_t prp2 = le64_to_cpu(rw->prp2);
 238
 239     uint8_t lba_index  = NVME_ID_NS_FLBAS_INDEX(ns->id_ns.flbas);
 240     uint8_t data_shift = ns->id_ns.lbaf[lba_index].ds;
 241     uint64_t data_size = (uint64_t)nlb << data_shift;
 242     uint64_t data_offset = slba << data_shift;
 243     int is_write = rw->opcode == NVME_CMD_WRITE ? 1 : 0;
 244     enum BlockAcctType acct = is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ;
 245
 246     if ((slba + nlb) > ns->id_ns.nsze) {
 247         block_acct_invalid(blk_get_stats(n->conf.blk), acct);
 248         return NVME_LBA_RANGE | NVME_DNR;
 249     }
 250
 251     if (nvme_map_prp(&req->qsg, prp1, prp2, data_size, n)) {
 252         block_acct_invalid(blk_get_stats(n->conf.blk), acct);
 253         return NVME_INVALID_FIELD | NVME_DNR;
 254     }
 255
 256     assert((nlb << data_shift) == req->qsg.size);
 257
 258     req->has_sg = true;
 259     dma_acct_start(n->conf.blk, &req->acct, &req->qsg, acct);
 260     req->aiocb = is_write ?
 261         dma_blk_write(n->conf.blk, &req->qsg, data_offset, BDRV_SECTOR_SIZE,
 262                       nvme_rw_cb, req) :
 263         dma_blk_read(n->conf.blk, &req->qsg, data_offset, BDRV_SECTOR_SIZE,
 264                      nvme_rw_cb, req);
 265
 266     return NVME_NO_COMPLETE;
 267 }
 268
 269 static uint16_t nvme_io_cmd(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 270 {
 271     NvmeNamespace *ns;
 272     uint32_t nsid = le32_to_cpu(cmd->nsid);
 273
 274     if (nsid == 0 || nsid > n->num_namespaces) {
 275         return NVME_INVALID_NSID | NVME_DNR;
 276     }
 277
 278     ns = &n->namespaces[nsid - 1];
 279     switch (cmd->opcode) {
 280     case NVME_CMD_FLUSH:
 281         return nvme_flush(n, ns, cmd, req);
 282     case NVME_CMD_WRITE:
 283     case NVME_CMD_READ:
 284         return nvme_rw(n, ns, cmd, req);
 285     default:
 286         return NVME_INVALID_OPCODE | NVME_DNR;
 287     }
 288 }
 289
 290 static void nvme_free_sq(NvmeSQueue *sq, NvmeCtrl *n)
 291 {
 292     n->sq[sq->sqid] = NULL;
 293     timer_del(sq->timer);
 294     timer_free(sq->timer);
 295     g_free(sq->io_req);
 296     if (sq->sqid) {
 297         g_free(sq);
 298     }
 299 }
 300
 301 static uint16_t nvme_del_sq(NvmeCtrl *n, NvmeCmd *cmd)
 302 {
 303     NvmeDeleteQ *c = (NvmeDeleteQ *)cmd;
 304     NvmeRequest *req, *next;
 305     NvmeSQueue *sq;
 306     NvmeCQueue *cq;
 307     uint16_t qid = le16_to_cpu(c->qid);
 308
 309     if (!qid || nvme_check_sqid(n, qid)) {
 310         return NVME_INVALID_QID | NVME_DNR;
 311     }
 312
 313     sq = n->sq[qid];
 314     while (!QTAILQ_EMPTY(&sq->out_req_list)) {
 315         req = QTAILQ_FIRST(&sq->out_req_list);
 316         assert(req->aiocb);
 317         blk_aio_cancel(req->aiocb);
 318     }
 319     if (!nvme_check_cqid(n, sq->cqid)) {
 320         cq = n->cq[sq->cqid];
 321         QTAILQ_REMOVE(&cq->sq_list, sq, entry);
 322
 323         nvme_post_cqes(cq);
 324         QTAILQ_FOREACH_SAFE(req, &cq->req_list, entry, next) {
 325             if (req->sq == sq) {
 326                 QTAILQ_REMOVE(&cq->req_list, req, entry);
 327                 QTAILQ_INSERT_TAIL(&sq->req_list, req, entry);
 328             }
 329         }
 330     }
 331
 332     nvme_free_sq(sq, n);
 333     return NVME_SUCCESS;
 334 }
 335
 336 static void nvme_init_sq(NvmeSQueue *sq, NvmeCtrl *n, uint64_t dma_addr,
 337     uint16_t sqid, uint16_t cqid, uint16_t size)
 338 {
 339     int i;
 340     NvmeCQueue *cq;
 341
 342     sq->ctrl = n;
 343     sq->dma_addr = dma_addr;
 344     sq->sqid = sqid;
 345     sq->size = size;
 346     sq->cqid = cqid;
 347     sq->head = sq->tail = 0;
 348     sq->io_req = g_new(NvmeRequest, sq->size);
 349
 350     QTAILQ_INIT(&sq->req_list);
 351     QTAILQ_INIT(&sq->out_req_list);
 352     for (i = 0; i < sq->size; i++) {
 353         sq->io_req[i].sq = sq;
 354         QTAILQ_INSERT_TAIL(&(sq->req_list), &sq->io_req[i], entry);
 355     }
 356     sq->timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, nvme_process_sq, sq);
 357
 358     assert(n->cq[cqid]);
 359     cq = n->cq[cqid];
 360     QTAILQ_INSERT_TAIL(&(cq->sq_list), sq, entry);
 361     n->sq[sqid] = sq;
 362 }
 363
 364 static uint16_t nvme_create_sq(NvmeCtrl *n, NvmeCmd *cmd)
 365 {
 366     NvmeSQueue *sq;
 367     NvmeCreateSq *c = (NvmeCreateSq *)cmd;
 368
 369     uint16_t cqid = le16_to_cpu(c->cqid);
 370     uint16_t sqid = le16_to_cpu(c->sqid);
 371     uint16_t qsize = le16_to_cpu(c->qsize);
 372     uint16_t qflags = le16_to_cpu(c->sq_flags);
 373     uint64_t prp1 = le64_to_cpu(c->prp1);
 374
 375     if (!cqid || nvme_check_cqid(n, cqid)) {
 376         return NVME_INVALID_CQID | NVME_DNR;
 377     }
 378     if (!sqid || !nvme_check_sqid(n, sqid)) {
 379         return NVME_INVALID_QID | NVME_DNR;
 380     }
 381     if (!qsize || qsize > NVME_CAP_MQES(n->bar.cap)) {
 382         return NVME_MAX_QSIZE_EXCEEDED | NVME_DNR;
 383     }
 384     if (!prp1 || prp1 & (n->page_size - 1)) {
 385         return NVME_INVALID_FIELD | NVME_DNR;
 386     }
 387     if (!(NVME_SQ_FLAGS_PC(qflags))) {
 388         return NVME_INVALID_FIELD | NVME_DNR;
 389     }
 390     sq = g_malloc0(sizeof(*sq));
 391     nvme_init_sq(sq, n, prp1, sqid, cqid, qsize + 1);
 392     return NVME_SUCCESS;
 393 }
 394
 395 static void nvme_free_cq(NvmeCQueue *cq, NvmeCtrl *n)
 396 {
 397     n->cq[cq->cqid] = NULL;
 398     timer_del(cq->timer);
 399     timer_free(cq->timer);
 400     msix_vector_unuse(&n->parent_obj, cq->vector);
 401     if (cq->cqid) {
 402         g_free(cq);
 403     }
 404 }
 405
 406 static uint16_t nvme_del_cq(NvmeCtrl *n, NvmeCmd *cmd)
 407 {
 408     NvmeDeleteQ *c = (NvmeDeleteQ *)cmd;
 409     NvmeCQueue *cq;
 410     uint16_t qid = le16_to_cpu(c->qid);
 411
 412     if (!qid || nvme_check_cqid(n, qid)) {
 413         return NVME_INVALID_CQID | NVME_DNR;
 414     }
 415
 416     cq = n->cq[qid];
 417     if (!QTAILQ_EMPTY(&cq->sq_list)) {
 418         return NVME_INVALID_QUEUE_DEL;
 419     }
 420     nvme_free_cq(cq, n);
 421     return NVME_SUCCESS;
 422 }
 423
 424 static void nvme_init_cq(NvmeCQueue *cq, NvmeCtrl *n, uint64_t dma_addr,
 425     uint16_t cqid, uint16_t vector, uint16_t size, uint16_t irq_enabled)
 426 {
 427     cq->ctrl = n;
 428     cq->cqid = cqid;
 429     cq->size = size;
 430     cq->dma_addr = dma_addr;
 431     cq->phase = 1;
 432     cq->irq_enabled = irq_enabled;
 433     cq->vector = vector;
 434     cq->head = cq->tail = 0;
 435     QTAILQ_INIT(&cq->req_list);
 436     QTAILQ_INIT(&cq->sq_list);
 437     msix_vector_use(&n->parent_obj, cq->vector);
 438     n->cq[cqid] = cq;
 439     cq->timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, nvme_post_cqes, cq);
 440 }
 441
 442 static uint16_t nvme_create_cq(NvmeCtrl *n, NvmeCmd *cmd)
 443 {
 444     NvmeCQueue *cq;
 445     NvmeCreateCq *c = (NvmeCreateCq *)cmd;
 446     uint16_t cqid = le16_to_cpu(c->cqid);
 447     uint16_t vector = le16_to_cpu(c->irq_vector);
 448     uint16_t qsize = le16_to_cpu(c->qsize);
 449     uint16_t qflags = le16_to_cpu(c->cq_flags);
 450     uint64_t prp1 = le64_to_cpu(c->prp1);
 451
 452     if (!cqid || !nvme_check_cqid(n, cqid)) {
 453         return NVME_INVALID_CQID | NVME_DNR;
 454     }
 455     if (!qsize || qsize > NVME_CAP_MQES(n->bar.cap)) {
 456         return NVME_MAX_QSIZE_EXCEEDED | NVME_DNR;
 457     }
 458     if (!prp1) {
 459         return NVME_INVALID_FIELD | NVME_DNR;
 460     }
 461     if (vector > n->num_queues) {
 462         return NVME_INVALID_IRQ_VECTOR | NVME_DNR;
 463     }
 464     if (!(NVME_CQ_FLAGS_PC(qflags))) {
 465         return NVME_INVALID_FIELD | NVME_DNR;
 466     }
 467
 468     cq = g_malloc0(sizeof(*cq));
 469     nvme_init_cq(cq, n, prp1, cqid, vector, qsize + 1,
 470         NVME_CQ_FLAGS_IEN(qflags));
 471     return NVME_SUCCESS;
 472 }
 473
 474 static uint16_t nvme_identify_ctrl(NvmeCtrl *n, NvmeIdentify *c)
 475 {
 476     uint64_t prp1 = le64_to_cpu(c->prp1);
 477     uint64_t prp2 = le64_to_cpu(c->prp2);
 478
 479     return nvme_dma_read_prp(n, (uint8_t *)&n->id_ctrl, sizeof(n->id_ctrl),
 480         prp1, prp2);
 481 }
 482
 483 static uint16_t nvme_identify_ns(NvmeCtrl *n, NvmeIdentify *c)
 484 {
 485     NvmeNamespace *ns;
 486     uint32_t nsid = le32_to_cpu(c->nsid);
 487     uint64_t prp1 = le64_to_cpu(c->prp1);
 488     uint64_t prp2 = le64_to_cpu(c->prp2);
 489
 490     if (nsid == 0 || nsid > n->num_namespaces) {
 491         return NVME_INVALID_NSID | NVME_DNR;
 492     }
 493
 494     ns = &n->namespaces[nsid - 1];
 495     return nvme_dma_read_prp(n, (uint8_t *)&ns->id_ns, sizeof(ns->id_ns),
 496         prp1, prp2);
 497 }
 498
 499 static uint16_t nvme_identify_nslist(NvmeCtrl *n, NvmeIdentify *c)
 500 {
 501     static const int data_len = 4096;
 502     uint32_t min_nsid = le32_to_cpu(c->nsid);
 503     uint64_t prp1 = le64_to_cpu(c->prp1);
 504     uint64_t prp2 = le64_to_cpu(c->prp2);
 505     uint32_t *list;
 506     uint16_t ret;
 507     int i, j = 0;
 508
 509     list = g_malloc0(data_len);
 510     for (i = 0; i < n->num_namespaces; i++) {
 511         if (i < min_nsid) {
 512             continue;
 513         }
 514         list[j++] = cpu_to_le32(i + 1);
 515         if (j == data_len / sizeof(uint32_t)) {
 516             break;
 517         }
 518     }
 519     ret = nvme_dma_read_prp(n, (uint8_t *)list, data_len, prp1, prp2);
 520     g_free(list);
 521     return ret;
 522 }
 523
 524
 525 static uint16_t nvme_identify(NvmeCtrl *n, NvmeCmd *cmd)
 526 {
 527     NvmeIdentify *c = (NvmeIdentify *)cmd;
 528
 529     switch (le32_to_cpu(c->cns)) {
 530     case 0x00:
 531         return nvme_identify_ns(n, c);
 532     case 0x01:
 533         return nvme_identify_ctrl(n, c);
 534     case 0x02:
 535         return nvme_identify_nslist(n, c);
 536     default:
 537         return NVME_INVALID_FIELD | NVME_DNR;
 538     }
 539 }
 540
 541 static uint16_t nvme_get_feature(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 542 {
 543     uint32_t dw10 = le32_to_cpu(cmd->cdw10);
 544     uint32_t result;
 545
 546     switch (dw10) {
 547     case NVME_VOLATILE_WRITE_CACHE:
 548         result = blk_enable_write_cache(n->conf.blk);
 549         break;
 550     case NVME_NUMBER_OF_QUEUES:
 551         result = cpu_to_le32((n->num_queues - 1) | ((n->num_queues - 1) << 16));
 552         break;
 553     default:
 554         return NVME_INVALID_FIELD | NVME_DNR;
 555     }
 556
 557     req->cqe.result = result;
 558     return NVME_SUCCESS;
 559 }
 560
 561 static uint16_t nvme_set_feature(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 562 {
 563     uint32_t dw10 = le32_to_cpu(cmd->cdw10);
 564     uint32_t dw11 = le32_to_cpu(cmd->cdw11);
 565
 566     switch (dw10) {
 567     case NVME_VOLATILE_WRITE_CACHE:
 568         blk_set_enable_write_cache(n->conf.blk, dw11 & 1);
 569         break;
 570     case NVME_NUMBER_OF_QUEUES:
 571         req->cqe.result =
 572             cpu_to_le32((n->num_queues - 1) | ((n->num_queues - 1) << 16));
 573         break;
 574     default:
 575         return NVME_INVALID_FIELD | NVME_DNR;
 576     }
 577     return NVME_SUCCESS;
 578 }
 579
 580 static uint16_t nvme_admin_cmd(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 581 {
 582     switch (cmd->opcode) {
 583     case NVME_ADM_CMD_DELETE_SQ:
 584         return nvme_del_sq(n, cmd);
 585     case NVME_ADM_CMD_CREATE_SQ:
 586         return nvme_create_sq(n, cmd);
 587     case NVME_ADM_CMD_DELETE_CQ:
 588         return nvme_del_cq(n, cmd);
 589     case NVME_ADM_CMD_CREATE_CQ:
 590         return nvme_create_cq(n, cmd);
 591     case NVME_ADM_CMD_IDENTIFY:
 592         return nvme_identify(n, cmd);
 593     case NVME_ADM_CMD_SET_FEATURES:
 594         return nvme_set_feature(n, cmd, req);
 595     case NVME_ADM_CMD_GET_FEATURES:
 596         return nvme_get_feature(n, cmd, req);
 597     default:
 598         return NVME_INVALID_OPCODE | NVME_DNR;
 599     }
 600 }
 601
 602 static void nvme_process_sq(void *opaque)
 603 {
 604     NvmeSQueue *sq = opaque;
 605     NvmeCtrl *n = sq->ctrl;
 606     NvmeCQueue *cq = n->cq[sq->cqid];
 607
 608     uint16_t status;
 609     hwaddr addr;
 610     NvmeCmd cmd;
 611     NvmeRequest *req;
 612
 613     while (!(nvme_sq_empty(sq) || QTAILQ_EMPTY(&sq->req_list))) {
 614         addr = sq->dma_addr + sq->head * n->sqe_size;
 615         pci_dma_read(&n->parent_obj, addr, (void *)&cmd, sizeof(cmd));
 616         nvme_inc_sq_head(sq);
 617
 618         req = QTAILQ_FIRST(&sq->req_list);
 619         QTAILQ_REMOVE(&sq->req_list, req, entry);
 620         QTAILQ_INSERT_TAIL(&sq->out_req_list, req, entry);
 621         memset(&req->cqe, 0, sizeof(req->cqe));
 622         req->cqe.cid = cmd.cid;
 623
 624         status = sq->sqid ? nvme_io_cmd(n, &cmd, req) :
 625             nvme_admin_cmd(n, &cmd, req);
 626         if (status != NVME_NO_COMPLETE) {
 627             req->status = status;
 628             nvme_enqueue_req_completion(cq, req);
 629         }
 630     }
 631 }
 632
 633 static void nvme_clear_ctrl(NvmeCtrl *n)
 634 {
 635     int i;
 636
 637     for (i = 0; i < n->num_queues; i++) {
 638         if (n->sq[i] != NULL) {
 639             nvme_free_sq(n->sq[i], n);
 640         }
 641     }
 642     for (i = 0; i < n->num_queues; i++) {
 643         if (n->cq[i] != NULL) {
 644             nvme_free_cq(n->cq[i], n);
 645         }
 646     }
 647
 648     blk_flush(n->conf.blk);
 649     n->bar.cc = 0;
 650 }
 651
 652 static int nvme_start_ctrl(NvmeCtrl *n)
 653 {
 654     uint32_t page_bits = NVME_CC_MPS(n->bar.cc) + 12;
 655     uint32_t page_size = 1 << page_bits;
 656
 657     if (n->cq[0] || n->sq[0] || !n->bar.asq || !n->bar.acq ||
 658             n->bar.asq & (page_size - 1) || n->bar.acq & (page_size - 1) ||
 659             NVME_CC_MPS(n->bar.cc) < NVME_CAP_MPSMIN(n->bar.cap) ||
 660             NVME_CC_MPS(n->bar.cc) > NVME_CAP_MPSMAX(n->bar.cap) ||
 661             NVME_CC_IOCQES(n->bar.cc) < NVME_CTRL_CQES_MIN(n->id_ctrl.cqes) ||
 662             NVME_CC_IOCQES(n->bar.cc) > NVME_CTRL_CQES_MAX(n->id_ctrl.cqes) ||
 663             NVME_CC_IOSQES(n->bar.cc) < NVME_CTRL_SQES_MIN(n->id_ctrl.sqes) ||
 664             NVME_CC_IOSQES(n->bar.cc) > NVME_CTRL_SQES_MAX(n->id_ctrl.sqes) ||
 665             !NVME_AQA_ASQS(n->bar.aqa) || !NVME_AQA_ACQS(n->bar.aqa)) {
 666         return -1;
 667     }
 668
 669     n->page_bits = page_bits;
 670     n->page_size = page_size;
 671     n->max_prp_ents = n->page_size / sizeof(uint64_t);
 672     n->cqe_size = 1 << NVME_CC_IOCQES(n->bar.cc);
 673     n->sqe_size = 1 << NVME_CC_IOSQES(n->bar.cc);
 674     nvme_init_cq(&n->admin_cq, n, n->bar.acq, 0, 0,
 675         NVME_AQA_ACQS(n->bar.aqa) + 1, 1);
 676     nvme_init_sq(&n->admin_sq, n, n->bar.asq, 0, 0,
 677         NVME_AQA_ASQS(n->bar.aqa) + 1);
 678
 679     return 0;
 680 }
 681
 682 static void nvme_write_bar(NvmeCtrl *n, hwaddr offset, uint64_t data,
 683     unsigned size)
 684 {
 685     switch (offset) {
 686     case 0xc:
 687         n->bar.intms |= data & 0xffffffff;
 688         n->bar.intmc = n->bar.intms;
 689         break;
 690     case 0x10:
 691         n->bar.intms &= ~(data & 0xffffffff);
 692         n->bar.intmc = n->bar.intms;
 693         break;
 694     case 0x14:
 695         /* Windows first sends data, then sends enable bit */
 696         if (!NVME_CC_EN(data) && !NVME_CC_EN(n->bar.cc) &&
 697             !NVME_CC_SHN(data) && !NVME_CC_SHN(n->bar.cc))
 698         {
 699             n->bar.cc = data;
 700         }
 701
 702         if (NVME_CC_EN(data) && !NVME_CC_EN(n->bar.cc)) {
 703             n->bar.cc = data;
 704             if (nvme_start_ctrl(n)) {
 705                 n->bar.csts = NVME_CSTS_FAILED;
 706             } else {
 707                 n->bar.csts = NVME_CSTS_READY;
 708             }
 709         } else if (!NVME_CC_EN(data) && NVME_CC_EN(n->bar.cc)) {
 710             nvme_clear_ctrl(n);
 711             n->bar.csts &= ~NVME_CSTS_READY;
 712         }
 713         if (NVME_CC_SHN(data) && !(NVME_CC_SHN(n->bar.cc))) {
 714                 nvme_clear_ctrl(n);
 715                 n->bar.cc = data;
 716                 n->bar.csts |= NVME_CSTS_SHST_COMPLETE;
 717         } else if (!NVME_CC_SHN(data) && NVME_CC_SHN(n->bar.cc)) {
 718                 n->bar.csts &= ~NVME_CSTS_SHST_COMPLETE;
 719                 n->bar.cc = data;
 720         }
 721         break;
 722     case 0x24:
 723         n->bar.aqa = data & 0xffffffff;
 724         break;
 725     case 0x28:
 726         n->bar.asq = data;
 727         break;
 728     case 0x2c:
 729         n->bar.asq |= data << 32;
 730         break;
 731     case 0x30:
 732         n->bar.acq = data;
 733         break;
 734     case 0x34:
 735         n->bar.acq |= data << 32;
 736         break;
 737     default:
 738         break;
 739     }
 740 }
 741
 742 static uint64_t nvme_mmio_read(void *opaque, hwaddr addr, unsigned size)
 743 {
 744     NvmeCtrl *n = (NvmeCtrl *)opaque;
 745     uint8_t *ptr = (uint8_t *)&n->bar;
 746     uint64_t val = 0;
 747
 748     if (addr < sizeof(n->bar)) {
 749         memcpy(&val, ptr + addr, size);
 750     }
 751     return val;
 752 }
 753
 754 static void nvme_process_db(NvmeCtrl *n, hwaddr addr, int val)
 755 {
 756     uint32_t qid;
 757
 758     if (addr & ((1 << 2) - 1)) {
 759         return;
 760     }
 761
 762     if (((addr - 0x1000) >> 2) & 1) {
 763         uint16_t new_head = val & 0xffff;
 764         int start_sqs;
 765         NvmeCQueue *cq;
 766
 767         qid = (addr - (0x1000 + (1 << 2))) >> 3;
 768         if (nvme_check_cqid(n, qid)) {
 769             return;
 770         }
 771
 772         cq = n->cq[qid];
 773         if (new_head >= cq->size) {
 774             return;
 775         }
 776
 777         start_sqs = nvme_cq_full(cq) ? 1 : 0;
 778         cq->head = new_head;
 779         if (start_sqs) {
 780             NvmeSQueue *sq;
 781             QTAILQ_FOREACH(sq, &cq->sq_list, entry) {
 782                 timer_mod(sq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 783             }
 784             timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 785         }
 786
 787         if (cq->tail != cq->head) {
 788             nvme_isr_notify(n, cq);
 789         }
 790     } else {
 791         uint16_t new_tail = val & 0xffff;
 792         NvmeSQueue *sq;
 793
 794         qid = (addr - 0x1000) >> 3;
 795         if (nvme_check_sqid(n, qid)) {
 796             return;
 797         }
 798
 799         sq = n->sq[qid];
 800         if (new_tail >= sq->size) {
 801             return;
 802         }
 803
 804         sq->tail = new_tail;
 805         timer_mod(sq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 806     }
 807 }
 808
 809 static void nvme_mmio_write(void *opaque, hwaddr addr, uint64_t data,
 810     unsigned size)
 811 {
 812     NvmeCtrl *n = (NvmeCtrl *)opaque;
 813     if (addr < sizeof(n->bar)) {
 814         nvme_write_bar(n, addr, data, size);
 815     } else if (addr >= 0x1000) {
 816         nvme_process_db(n, addr, data);
 817     }
 818 }
 819
 820 static const MemoryRegionOps nvme_mmio_ops = {
 821     .read = nvme_mmio_read,
 822     .write = nvme_mmio_write,
 823     .endianness = DEVICE_LITTLE_ENDIAN,
 824     .impl = {
 825         .min_access_size = 2,
 826         .max_access_size = 8,
 827     },
 828 };
 829
 830 static int nvme_init(PCIDevice *pci_dev)
 831 {
 832     NvmeCtrl *n = NVME(pci_dev);
 833     NvmeIdCtrl *id = &n->id_ctrl;
 834
 835     int i;
 836     int64_t bs_size;
 837     uint8_t *pci_conf;
 838     Error *local_err = NULL;
 839
 840     if (!n->conf.blk) {
 841         return -1;
 842     }
 843
 844     bs_size = blk_getlength(n->conf.blk);
 845     if (bs_size < 0) {
 846         return -1;
 847     }
 848
 849     blkconf_serial(&n->conf, &n->serial);
 850     if (!n->serial) {
 851         return -1;
 852     }
 853     blkconf_blocksizes(&n->conf);
 854     blkconf_apply_backend_options(&n->conf, blk_is_read_only(n->conf.blk),
 855                                   false, &local_err);
 856     if (local_err) {
 857         error_report_err(local_err);
 858         return -1;
 859     }
 860
 861     pci_conf = pci_dev->config;
 862     pci_conf[PCI_INTERRUPT_PIN] = 1;
 863     pci_config_set_prog_interface(pci_dev->config, 0x2);
 864     pci_config_set_class(pci_dev->config, PCI_CLASS_STORAGE_EXPRESS);
 865     pcie_endpoint_cap_init(&n->parent_obj, 0x80);
 866
 867     n->num_namespaces = 1;
 868     n->num_queues = 64;
 869     n->reg_size = pow2ceil(0x1004 + 2 * (n->num_queues + 1) * 4);
 870     n->ns_size = bs_size / (uint64_t)n->num_namespaces;
 871
 872     n->namespaces = g_new0(NvmeNamespace, n->num_namespaces);
 873     n->sq = g_new0(NvmeSQueue *, n->num_queues);
 874     n->cq = g_new0(NvmeCQueue *, n->num_queues);
 875
 876     memory_region_init_io(&n->iomem, OBJECT(n), &nvme_mmio_ops, n,
 877                           "nvme", n->reg_size);
 878     pci_register_bar(&n->parent_obj, 0,
 879         PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_64,
 880         &n->iomem);
 881     msix_init_exclusive_bar(&n->parent_obj, n->num_queues, 4, NULL);
 882
 883     id->vid = cpu_to_le16(pci_get_word(pci_conf + PCI_VENDOR_ID));
 884     id->ssvid = cpu_to_le16(pci_get_word(pci_conf + PCI_SUBSYSTEM_VENDOR_ID));
 885     strpadcpy((char *)id->mn, sizeof(id->mn), "QEMU NVMe Ctrl", ' ');
 886     strpadcpy((char *)id->fr, sizeof(id->fr), "1.0", ' ');
 887     strpadcpy((char *)id->sn, sizeof(id->sn), n->serial, ' ');
 888     id->rab = 6;
 889     id->ieee[0] = 0x00;
 890     id->ieee[1] = 0x02;
 891     id->ieee[2] = 0xb3;
 892     id->oacs = cpu_to_le16(0);
 893     id->frmw = 7 << 1;
 894     id->lpa = 1 << 0;
 895     id->sqes = (0x6 << 4) | 0x6;
 896     id->cqes = (0x4 << 4) | 0x4;
 897     id->nn = cpu_to_le32(n->num_namespaces);
 898     id->psd[0].mp = cpu_to_le16(0x9c4);
 899     id->psd[0].enlat = cpu_to_le32(0x10);
 900     id->psd[0].exlat = cpu_to_le32(0x4);
 901     if (blk_enable_write_cache(n->conf.blk)) {
 902         id->vwc = 1;
 903     }
 904
 905     n->bar.cap = 0;
 906     NVME_CAP_SET_MQES(n->bar.cap, 0x7ff);
 907     NVME_CAP_SET_CQR(n->bar.cap, 1);
 908     NVME_CAP_SET_AMS(n->bar.cap, 1);
 909     NVME_CAP_SET_TO(n->bar.cap, 0xf);
 910     NVME_CAP_SET_CSS(n->bar.cap, 1);
 911     NVME_CAP_SET_MPSMAX(n->bar.cap, 4);
 912
 913     n->bar.vs = 0x00010100;
 914     n->bar.intmc = n->bar.intms = 0;
 915
 916     for (i = 0; i < n->num_namespaces; i++) {
 917         NvmeNamespace *ns = &n->namespaces[i];
 918         NvmeIdNs *id_ns = &ns->id_ns;
 919         id_ns->nsfeat = 0;
 920         id_ns->nlbaf = 0;
 921         id_ns->flbas = 0;
 922         id_ns->mc = 0;
 923         id_ns->dpc = 0;
 924         id_ns->dps = 0;
 925         id_ns->lbaf[0].ds = BDRV_SECTOR_BITS;
 926         id_ns->ncap  = id_ns->nuse = id_ns->nsze =
 927             cpu_to_le64(n->ns_size >>
 928                 id_ns->lbaf[NVME_ID_NS_FLBAS_INDEX(ns->id_ns.flbas)].ds);
 929     }
 930     return 0;
 931 }
 932
 933 static void nvme_exit(PCIDevice *pci_dev)
 934 {
 935     NvmeCtrl *n = NVME(pci_dev);
 936
 937     nvme_clear_ctrl(n);
 938     g_free(n->namespaces);
 939     g_free(n->cq);
 940     g_free(n->sq);
 941     msix_uninit_exclusive_bar(pci_dev);
 942 }
 943
 944 static Property nvme_props[] = {
 945     DEFINE_BLOCK_PROPERTIES(NvmeCtrl, conf),
 946     DEFINE_PROP_STRING("serial", NvmeCtrl, serial),
 947     DEFINE_PROP_END_OF_LIST(),
 948 };
 949
 950 static const VMStateDescription nvme_vmstate = {
 951     .name = "nvme",
 952     .unmigratable = 1,
 953 };
 954
 955 static void nvme_class_init(ObjectClass *oc, void *data)
 956 {
 957     DeviceClass *dc = DEVICE_CLASS(oc);
 958     PCIDeviceClass *pc = PCI_DEVICE_CLASS(oc);
 959
 960     pc->init = nvme_init;
 961     pc->exit = nvme_exit;
 962     pc->class_id = PCI_CLASS_STORAGE_EXPRESS;
 963     pc->vendor_id = PCI_VENDOR_ID_INTEL;
 964     pc->device_id = 0x5845;
 965     pc->revision = 2;
 966     pc->is_express = 1;
 967
 968     set_bit(DEVICE_CATEGORY_STORAGE, dc->categories);
 969     dc->desc = "Non-Volatile Memory Express";
 970     dc->props = nvme_props;
 971     dc->vmsd = &nvme_vmstate;
 972 }
 973
 974 static void nvme_instance_init(Object *obj)
 975 {
 976     NvmeCtrl *s = NVME(obj);
 977
 978     device_add_bootindex_property(obj, &s->conf.bootindex,
 979                                   "bootindex", "/namespace@1,0",
 980                                   DEVICE(obj), &error_abort);
 981 }
 982
 983 static const TypeInfo nvme_info = {
 984     .name          = "nvme",
 985     .parent        = TYPE_PCI_DEVICE,
 986     .instance_size = sizeof(NvmeCtrl),
 987     .class_init    = nvme_class_init,
 988     .instance_init = nvme_instance_init,
 989 };
 990
 991 static void nvme_register_types(void)
 992 {
 993     type_register_static(&nvme_info);
 994 }
 995
 996 type_init(nvme_register_types)