qemu/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 <hw/block/block.h>
  24 #include <hw/hw.h>
  25 #include <hw/pci/msix.h>
  26 #include <hw/pci/pci.h>
  27 #include "sysemu/sysemu.h"
  28 #include "qapi/visitor.h"
  29 #include "sysemu/block-backend.h"
  30
  31 #include "nvme.h"
  32
  33 static void nvme_process_sq(void *opaque);
  34
  35 static int nvme_check_sqid(NvmeCtrl *n, uint16_t sqid)
  36 {
  37     return sqid < n->num_queues && n->sq[sqid] != NULL ? 0 : -1;
  38 }
  39
  40 static int nvme_check_cqid(NvmeCtrl *n, uint16_t cqid)
  41 {
  42     return cqid < n->num_queues && n->cq[cqid] != NULL ? 0 : -1;
  43 }
  44
  45 static void nvme_inc_cq_tail(NvmeCQueue *cq)
  46 {
  47     cq->tail++;
  48     if (cq->tail >= cq->size) {
  49         cq->tail = 0;
  50         cq->phase = !cq->phase;
  51     }
  52 }
  53
  54 static void nvme_inc_sq_head(NvmeSQueue *sq)
  55 {
  56     sq->head = (sq->head + 1) % sq->size;
  57 }
  58
  59 static uint8_t nvme_cq_full(NvmeCQueue *cq)
  60 {
  61     return (cq->tail + 1) % cq->size == cq->head;
  62 }
  63
  64 static uint8_t nvme_sq_empty(NvmeSQueue *sq)
  65 {
  66     return sq->head == sq->tail;
  67 }
  68
  69 static void nvme_isr_notify(NvmeCtrl *n, NvmeCQueue *cq)
  70 {
  71     if (cq->irq_enabled) {
  72         if (msix_enabled(&(n->parent_obj))) {
  73             msix_notify(&(n->parent_obj), cq->vector);
  74         } else {
  75             pci_irq_pulse(&n->parent_obj);
  76         }
  77     }
  78 }
  79
  80 static uint16_t nvme_map_prp(QEMUSGList *qsg, uint64_t prp1, uint64_t prp2,
  81     uint32_t len, NvmeCtrl *n)
  82 {
  83     hwaddr trans_len = n->page_size - (prp1 % n->page_size);
  84     trans_len = MIN(len, trans_len);
  85     int num_prps = (len >> n->page_bits) + 1;
  86
  87     if (!prp1) {
  88         return NVME_INVALID_FIELD | NVME_DNR;
  89     }
  90
  91     pci_dma_sglist_init(qsg, &n->parent_obj, num_prps);
  92     qemu_sglist_add(qsg, prp1, trans_len);
  93     len -= trans_len;
  94     if (len) {
  95         if (!prp2) {
  96             goto unmap;
  97         }
  98         if (len > n->page_size) {
  99             uint64_t prp_list[n->max_prp_ents];
 100             uint32_t nents, prp_trans;
 101             int i = 0;
 102
 103             nents = (len + n->page_size - 1) >> n->page_bits;
 104             prp_trans = MIN(n->max_prp_ents, nents) * sizeof(uint64_t);
 105             pci_dma_read(&n->parent_obj, prp2, (void *)prp_list, prp_trans);
 106             while (len != 0) {
 107                 uint64_t prp_ent = le64_to_cpu(prp_list[i]);
 108
 109                 if (i == n->max_prp_ents - 1 && len > n->page_size) {
 110                     if (!prp_ent || prp_ent & (n->page_size - 1)) {
 111                         goto unmap;
 112                     }
 113
 114                     i = 0;
 115                     nents = (len + n->page_size - 1) >> n->page_bits;
 116                     prp_trans = MIN(n->max_prp_ents, nents) * sizeof(uint64_t);
 117                     pci_dma_read(&n->parent_obj, prp_ent, (void *)prp_list,
 118                         prp_trans);
 119                     prp_ent = le64_to_cpu(prp_list[i]);
 120                 }
 121
 122                 if (!prp_ent || prp_ent & (n->page_size - 1)) {
 123                     goto unmap;
 124                 }
 125
 126                 trans_len = MIN(len, n->page_size);
 127                 qemu_sglist_add(qsg, prp_ent, trans_len);
 128                 len -= trans_len;
 129                 i++;
 130             }
 131         } else {
 132             if (prp2 & (n->page_size - 1)) {
 133                 goto unmap;
 134             }
 135             qemu_sglist_add(qsg, prp2, len);
 136         }
 137     }
 138     return NVME_SUCCESS;
 139
 140  unmap:
 141     qemu_sglist_destroy(qsg);
 142     return NVME_INVALID_FIELD | NVME_DNR;
 143 }
 144
 145 static uint16_t nvme_dma_read_prp(NvmeCtrl *n, uint8_t *ptr, uint32_t len,
 146     uint64_t prp1, uint64_t prp2)
 147 {
 148     QEMUSGList qsg;
 149
 150     if (nvme_map_prp(&qsg, prp1, prp2, len, n)) {
 151         return NVME_INVALID_FIELD | NVME_DNR;
 152     }
 153     if (dma_buf_read(ptr, len, &qsg)) {
 154         qemu_sglist_destroy(&qsg);
 155         return NVME_INVALID_FIELD | NVME_DNR;
 156     }
 157     qemu_sglist_destroy(&qsg);
 158     return NVME_SUCCESS;
 159 }
 160
 161 static void nvme_post_cqes(void *opaque)
 162 {
 163     NvmeCQueue *cq = opaque;
 164     NvmeCtrl *n = cq->ctrl;
 165     NvmeRequest *req, *next;
 166
 167     QTAILQ_FOREACH_SAFE(req, &cq->req_list, entry, next) {
 168         NvmeSQueue *sq;
 169         hwaddr addr;
 170
 171         if (nvme_cq_full(cq)) {
 172             break;
 173         }
 174
 175         QTAILQ_REMOVE(&cq->req_list, req, entry);
 176         sq = req->sq;
 177         req->cqe.status = cpu_to_le16((req->status << 1) | cq->phase);
 178         req->cqe.sq_id = cpu_to_le16(sq->sqid);
 179         req->cqe.sq_head = cpu_to_le16(sq->head);
 180         addr = cq->dma_addr + cq->tail * n->cqe_size;
 181         nvme_inc_cq_tail(cq);
 182         pci_dma_write(&n->parent_obj, addr, (void *)&req->cqe,
 183             sizeof(req->cqe));
 184         QTAILQ_INSERT_TAIL(&sq->req_list, req, entry);
 185     }
 186     nvme_isr_notify(n, cq);
 187 }
 188
 189 static void nvme_enqueue_req_completion(NvmeCQueue *cq, NvmeRequest *req)
 190 {
 191     assert(cq->cqid == req->sq->cqid);
 192     QTAILQ_REMOVE(&req->sq->out_req_list, req, entry);
 193     QTAILQ_INSERT_TAIL(&cq->req_list, req, entry);
 194     timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 195 }
 196
 197 static void nvme_rw_cb(void *opaque, int ret)
 198 {
 199     NvmeRequest *req = opaque;
 200     NvmeSQueue *sq = req->sq;
 201     NvmeCtrl *n = sq->ctrl;
 202     NvmeCQueue *cq = n->cq[sq->cqid];
 203
 204     block_acct_done(blk_get_stats(n->conf.blk), &req->acct);
 205     if (!ret) {
 206         req->status = NVME_SUCCESS;
 207     } else {
 208         req->status = NVME_INTERNAL_DEV_ERROR;
 209     }
 210     if (req->has_sg) {
 211         qemu_sglist_destroy(&req->qsg);
 212     }
 213     nvme_enqueue_req_completion(cq, req);
 214 }
 215
 216 static uint16_t nvme_flush(NvmeCtrl *n, NvmeNamespace *ns, NvmeCmd *cmd,
 217     NvmeRequest *req)
 218 {
 219     req->has_sg = false;
 220     block_acct_start(blk_get_stats(n->conf.blk), &req->acct, 0,
 221          BLOCK_ACCT_FLUSH);
 222     req->aiocb = blk_aio_flush(n->conf.blk, nvme_rw_cb, req);
 223
 224     return NVME_NO_COMPLETE;
 225 }
 226
 227 static uint16_t nvme_rw(NvmeCtrl *n, NvmeNamespace *ns, NvmeCmd *cmd,
 228     NvmeRequest *req)
 229 {
 230     NvmeRwCmd *rw = (NvmeRwCmd *)cmd;
 231     uint32_t nlb  = le32_to_cpu(rw->nlb) + 1;
 232     uint64_t slba = le64_to_cpu(rw->slba);
 233     uint64_t prp1 = le64_to_cpu(rw->prp1);
 234     uint64_t prp2 = le64_to_cpu(rw->prp2);
 235
 236     uint8_t lba_index  = NVME_ID_NS_FLBAS_INDEX(ns->id_ns.flbas);
 237     uint8_t data_shift = ns->id_ns.lbaf[lba_index].ds;
 238     uint64_t data_size = (uint64_t)nlb << data_shift;
 239     uint64_t aio_slba  = slba << (data_shift - BDRV_SECTOR_BITS);
 240     int is_write = rw->opcode == NVME_CMD_WRITE ? 1 : 0;
 241
 242     if ((slba + nlb) > ns->id_ns.nsze) {
 243         return NVME_LBA_RANGE | NVME_DNR;
 244     }
 245     if (nvme_map_prp(&req->qsg, prp1, prp2, data_size, n)) {
 246         return NVME_INVALID_FIELD | NVME_DNR;
 247     }
 248     assert((nlb << data_shift) == req->qsg.size);
 249
 250     req->has_sg = true;
 251     dma_acct_start(n->conf.blk, &req->acct, &req->qsg,
 252                    is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ);
 253     req->aiocb = is_write ?
 254         dma_blk_write(n->conf.blk, &req->qsg, aio_slba, nvme_rw_cb, req) :
 255         dma_blk_read(n->conf.blk, &req->qsg, aio_slba, nvme_rw_cb, req);
 256
 257     return NVME_NO_COMPLETE;
 258 }
 259
 260 static uint16_t nvme_io_cmd(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 261 {
 262     NvmeNamespace *ns;
 263     uint32_t nsid = le32_to_cpu(cmd->nsid);
 264
 265     if (nsid == 0 || nsid > n->num_namespaces) {
 266         return NVME_INVALID_NSID | NVME_DNR;
 267     }
 268
 269     ns = &n->namespaces[nsid - 1];
 270     switch (cmd->opcode) {
 271     case NVME_CMD_FLUSH:
 272         return nvme_flush(n, ns, cmd, req);
 273     case NVME_CMD_WRITE:
 274     case NVME_CMD_READ:
 275         return nvme_rw(n, ns, cmd, req);
 276     default:
 277         return NVME_INVALID_OPCODE | NVME_DNR;
 278     }
 279 }
 280
 281 static void nvme_free_sq(NvmeSQueue *sq, NvmeCtrl *n)
 282 {
 283     n->sq[sq->sqid] = NULL;
 284     timer_del(sq->timer);
 285     timer_free(sq->timer);
 286     g_free(sq->io_req);
 287     if (sq->sqid) {
 288         g_free(sq);
 289     }
 290 }
 291
 292 static uint16_t nvme_del_sq(NvmeCtrl *n, NvmeCmd *cmd)
 293 {
 294     NvmeDeleteQ *c = (NvmeDeleteQ *)cmd;
 295     NvmeRequest *req, *next;
 296     NvmeSQueue *sq;
 297     NvmeCQueue *cq;
 298     uint16_t qid = le16_to_cpu(c->qid);
 299
 300     if (!qid || nvme_check_sqid(n, qid)) {
 301         return NVME_INVALID_QID | NVME_DNR;
 302     }
 303
 304     sq = n->sq[qid];
 305     while (!QTAILQ_EMPTY(&sq->out_req_list)) {
 306         req = QTAILQ_FIRST(&sq->out_req_list);
 307         assert(req->aiocb);
 308         blk_aio_cancel(req->aiocb);
 309     }
 310     if (!nvme_check_cqid(n, sq->cqid)) {
 311         cq = n->cq[sq->cqid];
 312         QTAILQ_REMOVE(&cq->sq_list, sq, entry);
 313
 314         nvme_post_cqes(cq);
 315         QTAILQ_FOREACH_SAFE(req, &cq->req_list, entry, next) {
 316             if (req->sq == sq) {
 317                 QTAILQ_REMOVE(&cq->req_list, req, entry);
 318                 QTAILQ_INSERT_TAIL(&sq->req_list, req, entry);
 319             }
 320         }
 321     }
 322
 323     nvme_free_sq(sq, n);
 324     return NVME_SUCCESS;
 325 }
 326
 327 static void nvme_init_sq(NvmeSQueue *sq, NvmeCtrl *n, uint64_t dma_addr,
 328     uint16_t sqid, uint16_t cqid, uint16_t size)
 329 {
 330     int i;
 331     NvmeCQueue *cq;
 332
 333     sq->ctrl = n;
 334     sq->dma_addr = dma_addr;
 335     sq->sqid = sqid;
 336     sq->size = size;
 337     sq->cqid = cqid;
 338     sq->head = sq->tail = 0;
 339     sq->io_req = g_new(NvmeRequest, sq->size);
 340
 341     QTAILQ_INIT(&sq->req_list);
 342     QTAILQ_INIT(&sq->out_req_list);
 343     for (i = 0; i < sq->size; i++) {
 344         sq->io_req[i].sq = sq;
 345         QTAILQ_INSERT_TAIL(&(sq->req_list), &sq->io_req[i], entry);
 346     }
 347     sq->timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, nvme_process_sq, sq);
 348
 349     assert(n->cq[cqid]);
 350     cq = n->cq[cqid];
 351     QTAILQ_INSERT_TAIL(&(cq->sq_list), sq, entry);
 352     n->sq[sqid] = sq;
 353 }
 354
 355 static uint16_t nvme_create_sq(NvmeCtrl *n, NvmeCmd *cmd)
 356 {
 357     NvmeSQueue *sq;
 358     NvmeCreateSq *c = (NvmeCreateSq *)cmd;
 359
 360     uint16_t cqid = le16_to_cpu(c->cqid);
 361     uint16_t sqid = le16_to_cpu(c->sqid);
 362     uint16_t qsize = le16_to_cpu(c->qsize);
 363     uint16_t qflags = le16_to_cpu(c->sq_flags);
 364     uint64_t prp1 = le64_to_cpu(c->prp1);
 365
 366     if (!cqid || nvme_check_cqid(n, cqid)) {
 367         return NVME_INVALID_CQID | NVME_DNR;
 368     }
 369     if (!sqid || (sqid && !nvme_check_sqid(n, sqid))) {
 370         return NVME_INVALID_QID | NVME_DNR;
 371     }
 372     if (!qsize || qsize > NVME_CAP_MQES(n->bar.cap)) {
 373         return NVME_MAX_QSIZE_EXCEEDED | NVME_DNR;
 374     }
 375     if (!prp1 || prp1 & (n->page_size - 1)) {
 376         return NVME_INVALID_FIELD | NVME_DNR;
 377     }
 378     if (!(NVME_SQ_FLAGS_PC(qflags))) {
 379         return NVME_INVALID_FIELD | NVME_DNR;
 380     }
 381     sq = g_malloc0(sizeof(*sq));
 382     nvme_init_sq(sq, n, prp1, sqid, cqid, qsize + 1);
 383     return NVME_SUCCESS;
 384 }
 385
 386 static void nvme_free_cq(NvmeCQueue *cq, NvmeCtrl *n)
 387 {
 388     n->cq[cq->cqid] = NULL;
 389     timer_del(cq->timer);
 390     timer_free(cq->timer);
 391     msix_vector_unuse(&n->parent_obj, cq->vector);
 392     if (cq->cqid) {
 393         g_free(cq);
 394     }
 395 }
 396
 397 static uint16_t nvme_del_cq(NvmeCtrl *n, NvmeCmd *cmd)
 398 {
 399     NvmeDeleteQ *c = (NvmeDeleteQ *)cmd;
 400     NvmeCQueue *cq;
 401     uint16_t qid = le16_to_cpu(c->qid);
 402
 403     if (!qid || nvme_check_cqid(n, qid)) {
 404         return NVME_INVALID_CQID | NVME_DNR;
 405     }
 406
 407     cq = n->cq[qid];
 408     if (!QTAILQ_EMPTY(&cq->sq_list)) {
 409         return NVME_INVALID_QUEUE_DEL;
 410     }
 411     nvme_free_cq(cq, n);
 412     return NVME_SUCCESS;
 413 }
 414
 415 static void nvme_init_cq(NvmeCQueue *cq, NvmeCtrl *n, uint64_t dma_addr,
 416     uint16_t cqid, uint16_t vector, uint16_t size, uint16_t irq_enabled)
 417 {
 418     cq->ctrl = n;
 419     cq->cqid = cqid;
 420     cq->size = size;
 421     cq->dma_addr = dma_addr;
 422     cq->phase = 1;
 423     cq->irq_enabled = irq_enabled;
 424     cq->vector = vector;
 425     cq->head = cq->tail = 0;
 426     QTAILQ_INIT(&cq->req_list);
 427     QTAILQ_INIT(&cq->sq_list);
 428     msix_vector_use(&n->parent_obj, cq->vector);
 429     n->cq[cqid] = cq;
 430     cq->timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, nvme_post_cqes, cq);
 431 }
 432
 433 static uint16_t nvme_create_cq(NvmeCtrl *n, NvmeCmd *cmd)
 434 {
 435     NvmeCQueue *cq;
 436     NvmeCreateCq *c = (NvmeCreateCq *)cmd;
 437     uint16_t cqid = le16_to_cpu(c->cqid);
 438     uint16_t vector = le16_to_cpu(c->irq_vector);
 439     uint16_t qsize = le16_to_cpu(c->qsize);
 440     uint16_t qflags = le16_to_cpu(c->cq_flags);
 441     uint64_t prp1 = le64_to_cpu(c->prp1);
 442
 443     if (!cqid || (cqid && !nvme_check_cqid(n, cqid))) {
 444         return NVME_INVALID_CQID | NVME_DNR;
 445     }
 446     if (!qsize || qsize > NVME_CAP_MQES(n->bar.cap)) {
 447         return NVME_MAX_QSIZE_EXCEEDED | NVME_DNR;
 448     }
 449     if (!prp1) {
 450         return NVME_INVALID_FIELD | NVME_DNR;
 451     }
 452     if (vector > n->num_queues) {
 453         return NVME_INVALID_IRQ_VECTOR | NVME_DNR;
 454     }
 455     if (!(NVME_CQ_FLAGS_PC(qflags))) {
 456         return NVME_INVALID_FIELD | NVME_DNR;
 457     }
 458
 459     cq = g_malloc0(sizeof(*cq));
 460     nvme_init_cq(cq, n, prp1, cqid, vector, qsize + 1,
 461         NVME_CQ_FLAGS_IEN(qflags));
 462     return NVME_SUCCESS;
 463 }
 464
 465 static uint16_t nvme_identify(NvmeCtrl *n, NvmeCmd *cmd)
 466 {
 467     NvmeNamespace *ns;
 468     NvmeIdentify *c = (NvmeIdentify *)cmd;
 469     uint32_t cns  = le32_to_cpu(c->cns);
 470     uint32_t nsid = le32_to_cpu(c->nsid);
 471     uint64_t prp1 = le64_to_cpu(c->prp1);
 472     uint64_t prp2 = le64_to_cpu(c->prp2);
 473
 474     if (cns) {
 475         return nvme_dma_read_prp(n, (uint8_t *)&n->id_ctrl, sizeof(n->id_ctrl),
 476             prp1, prp2);
 477     }
 478     if (nsid == 0 || nsid > n->num_namespaces) {
 479         return NVME_INVALID_NSID | NVME_DNR;
 480     }
 481
 482     ns = &n->namespaces[nsid - 1];
 483     return nvme_dma_read_prp(n, (uint8_t *)&ns->id_ns, sizeof(ns->id_ns),
 484         prp1, prp2);
 485 }
 486
 487 static uint16_t nvme_get_feature(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 488 {
 489     uint32_t dw10 = le32_to_cpu(cmd->cdw10);
 490     uint32_t result;
 491
 492     switch (dw10) {
 493     case NVME_VOLATILE_WRITE_CACHE:
 494         result = blk_enable_write_cache(n->conf.blk);
 495         break;
 496     case NVME_NUMBER_OF_QUEUES:
 497         result = cpu_to_le32((n->num_queues - 1) | ((n->num_queues - 1) << 16));
 498         break;
 499     default:
 500         return NVME_INVALID_FIELD | NVME_DNR;
 501     }
 502
 503     req->cqe.result = result;
 504     return NVME_SUCCESS;
 505 }
 506
 507 static uint16_t nvme_set_feature(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 508 {
 509     uint32_t dw10 = le32_to_cpu(cmd->cdw10);
 510     uint32_t dw11 = le32_to_cpu(cmd->cdw11);
 511
 512     switch (dw10) {
 513     case NVME_VOLATILE_WRITE_CACHE:
 514         blk_set_enable_write_cache(n->conf.blk, dw11 & 1);
 515         break;
 516     case NVME_NUMBER_OF_QUEUES:
 517         req->cqe.result =
 518             cpu_to_le32((n->num_queues - 1) | ((n->num_queues - 1) << 16));
 519         break;
 520     default:
 521         return NVME_INVALID_FIELD | NVME_DNR;
 522     }
 523     return NVME_SUCCESS;
 524 }
 525
 526 static uint16_t nvme_admin_cmd(NvmeCtrl *n, NvmeCmd *cmd, NvmeRequest *req)
 527 {
 528     switch (cmd->opcode) {
 529     case NVME_ADM_CMD_DELETE_SQ:
 530         return nvme_del_sq(n, cmd);
 531     case NVME_ADM_CMD_CREATE_SQ:
 532         return nvme_create_sq(n, cmd);
 533     case NVME_ADM_CMD_DELETE_CQ:
 534         return nvme_del_cq(n, cmd);
 535     case NVME_ADM_CMD_CREATE_CQ:
 536         return nvme_create_cq(n, cmd);
 537     case NVME_ADM_CMD_IDENTIFY:
 538         return nvme_identify(n, cmd);
 539     case NVME_ADM_CMD_SET_FEATURES:
 540         return nvme_set_feature(n, cmd, req);
 541     case NVME_ADM_CMD_GET_FEATURES:
 542         return nvme_get_feature(n, cmd, req);
 543     default:
 544         return NVME_INVALID_OPCODE | NVME_DNR;
 545     }
 546 }
 547
 548 static void nvme_process_sq(void *opaque)
 549 {
 550     NvmeSQueue *sq = opaque;
 551     NvmeCtrl *n = sq->ctrl;
 552     NvmeCQueue *cq = n->cq[sq->cqid];
 553
 554     uint16_t status;
 555     hwaddr addr;
 556     NvmeCmd cmd;
 557     NvmeRequest *req;
 558
 559     while (!(nvme_sq_empty(sq) || QTAILQ_EMPTY(&sq->req_list))) {
 560         addr = sq->dma_addr + sq->head * n->sqe_size;
 561         pci_dma_read(&n->parent_obj, addr, (void *)&cmd, sizeof(cmd));
 562         nvme_inc_sq_head(sq);
 563
 564         req = QTAILQ_FIRST(&sq->req_list);
 565         QTAILQ_REMOVE(&sq->req_list, req, entry);
 566         QTAILQ_INSERT_TAIL(&sq->out_req_list, req, entry);
 567         memset(&req->cqe, 0, sizeof(req->cqe));
 568         req->cqe.cid = cmd.cid;
 569
 570         status = sq->sqid ? nvme_io_cmd(n, &cmd, req) :
 571             nvme_admin_cmd(n, &cmd, req);
 572         if (status != NVME_NO_COMPLETE) {
 573             req->status = status;
 574             nvme_enqueue_req_completion(cq, req);
 575         }
 576     }
 577 }
 578
 579 static void nvme_clear_ctrl(NvmeCtrl *n)
 580 {
 581     int i;
 582
 583     for (i = 0; i < n->num_queues; i++) {
 584         if (n->sq[i] != NULL) {
 585             nvme_free_sq(n->sq[i], n);
 586         }
 587     }
 588     for (i = 0; i < n->num_queues; i++) {
 589         if (n->cq[i] != NULL) {
 590             nvme_free_cq(n->cq[i], n);
 591         }
 592     }
 593
 594     blk_flush(n->conf.blk);
 595     n->bar.cc = 0;
 596 }
 597
 598 static int nvme_start_ctrl(NvmeCtrl *n)
 599 {
 600     uint32_t page_bits = NVME_CC_MPS(n->bar.cc) + 12;
 601     uint32_t page_size = 1 << page_bits;
 602
 603     if (n->cq[0] || n->sq[0] || !n->bar.asq || !n->bar.acq ||
 604             n->bar.asq & (page_size - 1) || n->bar.acq & (page_size - 1) ||
 605             NVME_CC_MPS(n->bar.cc) < NVME_CAP_MPSMIN(n->bar.cap) ||
 606             NVME_CC_MPS(n->bar.cc) > NVME_CAP_MPSMAX(n->bar.cap) ||
 607             NVME_CC_IOCQES(n->bar.cc) < NVME_CTRL_CQES_MIN(n->id_ctrl.cqes) ||
 608             NVME_CC_IOCQES(n->bar.cc) > NVME_CTRL_CQES_MAX(n->id_ctrl.cqes) ||
 609             NVME_CC_IOSQES(n->bar.cc) < NVME_CTRL_SQES_MIN(n->id_ctrl.sqes) ||
 610             NVME_CC_IOSQES(n->bar.cc) > NVME_CTRL_SQES_MAX(n->id_ctrl.sqes) ||
 611             !NVME_AQA_ASQS(n->bar.aqa) || !NVME_AQA_ACQS(n->bar.aqa)) {
 612         return -1;
 613     }
 614
 615     n->page_bits = page_bits;
 616     n->page_size = page_size;
 617     n->max_prp_ents = n->page_size / sizeof(uint64_t);
 618     n->cqe_size = 1 << NVME_CC_IOCQES(n->bar.cc);
 619     n->sqe_size = 1 << NVME_CC_IOSQES(n->bar.cc);
 620     nvme_init_cq(&n->admin_cq, n, n->bar.acq, 0, 0,
 621         NVME_AQA_ACQS(n->bar.aqa) + 1, 1);
 622     nvme_init_sq(&n->admin_sq, n, n->bar.asq, 0, 0,
 623         NVME_AQA_ASQS(n->bar.aqa) + 1);
 624
 625     return 0;
 626 }
 627
 628 static void nvme_write_bar(NvmeCtrl *n, hwaddr offset, uint64_t data,
 629     unsigned size)
 630 {
 631     switch (offset) {
 632     case 0xc:
 633         n->bar.intms |= data & 0xffffffff;
 634         n->bar.intmc = n->bar.intms;
 635         break;
 636     case 0x10:
 637         n->bar.intms &= ~(data & 0xffffffff);
 638         n->bar.intmc = n->bar.intms;
 639         break;
 640     case 0x14:
 641         /* Windows first sends data, then sends enable bit */
 642         if (!NVME_CC_EN(data) && !NVME_CC_EN(n->bar.cc) &&
 643             !NVME_CC_SHN(data) && !NVME_CC_SHN(n->bar.cc))
 644         {
 645             n->bar.cc = data;
 646         }
 647
 648         if (NVME_CC_EN(data) && !NVME_CC_EN(n->bar.cc)) {
 649             n->bar.cc = data;
 650             if (nvme_start_ctrl(n)) {
 651                 n->bar.csts = NVME_CSTS_FAILED;
 652             } else {
 653                 n->bar.csts = NVME_CSTS_READY;
 654             }
 655         } else if (!NVME_CC_EN(data) && NVME_CC_EN(n->bar.cc)) {
 656             nvme_clear_ctrl(n);
 657             n->bar.csts &= ~NVME_CSTS_READY;
 658         }
 659         if (NVME_CC_SHN(data) && !(NVME_CC_SHN(n->bar.cc))) {
 660                 nvme_clear_ctrl(n);
 661                 n->bar.cc = data;
 662                 n->bar.csts |= NVME_CSTS_SHST_COMPLETE;
 663         } else if (!NVME_CC_SHN(data) && NVME_CC_SHN(n->bar.cc)) {
 664                 n->bar.csts &= ~NVME_CSTS_SHST_COMPLETE;
 665                 n->bar.cc = data;
 666         }
 667         break;
 668     case 0x24:
 669         n->bar.aqa = data & 0xffffffff;
 670         break;
 671     case 0x28:
 672         n->bar.asq = data;
 673         break;
 674     case 0x2c:
 675         n->bar.asq |= data << 32;
 676         break;
 677     case 0x30:
 678         n->bar.acq = data;
 679         break;
 680     case 0x34:
 681         n->bar.acq |= data << 32;
 682         break;
 683     default:
 684         break;
 685     }
 686 }
 687
 688 static uint64_t nvme_mmio_read(void *opaque, hwaddr addr, unsigned size)
 689 {
 690     NvmeCtrl *n = (NvmeCtrl *)opaque;
 691     uint8_t *ptr = (uint8_t *)&n->bar;
 692     uint64_t val = 0;
 693
 694     if (addr < sizeof(n->bar)) {
 695         memcpy(&val, ptr + addr, size);
 696     }
 697     return val;
 698 }
 699
 700 static void nvme_process_db(NvmeCtrl *n, hwaddr addr, int val)
 701 {
 702     uint32_t qid;
 703
 704     if (addr & ((1 << 2) - 1)) {
 705         return;
 706     }
 707
 708     if (((addr - 0x1000) >> 2) & 1) {
 709         uint16_t new_head = val & 0xffff;
 710         int start_sqs;
 711         NvmeCQueue *cq;
 712
 713         qid = (addr - (0x1000 + (1 << 2))) >> 3;
 714         if (nvme_check_cqid(n, qid)) {
 715             return;
 716         }
 717
 718         cq = n->cq[qid];
 719         if (new_head >= cq->size) {
 720             return;
 721         }
 722
 723         start_sqs = nvme_cq_full(cq) ? 1 : 0;
 724         cq->head = new_head;
 725         if (start_sqs) {
 726             NvmeSQueue *sq;
 727             QTAILQ_FOREACH(sq, &cq->sq_list, entry) {
 728                 timer_mod(sq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 729             }
 730             timer_mod(cq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 731         }
 732
 733         if (cq->tail != cq->head) {
 734             nvme_isr_notify(n, cq);
 735         }
 736     } else {
 737         uint16_t new_tail = val & 0xffff;
 738         NvmeSQueue *sq;
 739
 740         qid = (addr - 0x1000) >> 3;
 741         if (nvme_check_sqid(n, qid)) {
 742             return;
 743         }
 744
 745         sq = n->sq[qid];
 746         if (new_tail >= sq->size) {
 747             return;
 748         }
 749
 750         sq->tail = new_tail;
 751         timer_mod(sq->timer, qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + 500);
 752     }
 753 }
 754
 755 static void nvme_mmio_write(void *opaque, hwaddr addr, uint64_t data,
 756     unsigned size)
 757 {
 758     NvmeCtrl *n = (NvmeCtrl *)opaque;
 759     if (addr < sizeof(n->bar)) {
 760         nvme_write_bar(n, addr, data, size);
 761     } else if (addr >= 0x1000) {
 762         nvme_process_db(n, addr, data);
 763     }
 764 }
 765
 766 static const MemoryRegionOps nvme_mmio_ops = {
 767     .read = nvme_mmio_read,
 768     .write = nvme_mmio_write,
 769     .endianness = DEVICE_LITTLE_ENDIAN,
 770     .impl = {
 771         .min_access_size = 2,
 772         .max_access_size = 8,
 773     },
 774 };
 775
 776 static int nvme_init(PCIDevice *pci_dev)
 777 {
 778     NvmeCtrl *n = NVME(pci_dev);
 779     NvmeIdCtrl *id = &n->id_ctrl;
 780
 781     int i;
 782     int64_t bs_size;
 783     uint8_t *pci_conf;
 784
 785     if (!n->conf.blk) {
 786         return -1;
 787     }
 788
 789     bs_size = blk_getlength(n->conf.blk);
 790     if (bs_size < 0) {
 791         return -1;
 792     }
 793
 794     blkconf_serial(&n->conf, &n->serial);
 795     if (!n->serial) {
 796         return -1;
 797     }
 798     blkconf_blocksizes(&n->conf);
 799
 800     pci_conf = pci_dev->config;
 801     pci_conf[PCI_INTERRUPT_PIN] = 1;
 802     pci_config_set_prog_interface(pci_dev->config, 0x2);
 803     pci_config_set_class(pci_dev->config, PCI_CLASS_STORAGE_EXPRESS);
 804     pcie_endpoint_cap_init(&n->parent_obj, 0x80);
 805
 806     n->num_namespaces = 1;
 807     n->num_queues = 64;
 808     n->reg_size = 1 << qemu_fls(0x1004 + 2 * (n->num_queues + 1) * 4);
 809     n->ns_size = bs_size / (uint64_t)n->num_namespaces;
 810
 811     n->namespaces = g_new0(NvmeNamespace, n->num_namespaces);
 812     n->sq = g_new0(NvmeSQueue *, n->num_queues);
 813     n->cq = g_new0(NvmeCQueue *, n->num_queues);
 814
 815     memory_region_init_io(&n->iomem, OBJECT(n), &nvme_mmio_ops, n,
 816                           "nvme", n->reg_size);
 817     pci_register_bar(&n->parent_obj, 0,
 818         PCI_BASE_ADDRESS_SPACE_MEMORY | PCI_BASE_ADDRESS_MEM_TYPE_64,
 819         &n->iomem);
 820     msix_init_exclusive_bar(&n->parent_obj, n->num_queues, 4);
 821
 822     id->vid = cpu_to_le16(pci_get_word(pci_conf + PCI_VENDOR_ID));
 823     id->ssvid = cpu_to_le16(pci_get_word(pci_conf + PCI_SUBSYSTEM_VENDOR_ID));
 824     strpadcpy((char *)id->mn, sizeof(id->mn), "QEMU NVMe Ctrl", ' ');
 825     strpadcpy((char *)id->fr, sizeof(id->fr), "1.0", ' ');
 826     strpadcpy((char *)id->sn, sizeof(id->sn), n->serial, ' ');
 827     id->rab = 6;
 828     id->ieee[0] = 0x00;
 829     id->ieee[1] = 0x02;
 830     id->ieee[2] = 0xb3;
 831     id->oacs = cpu_to_le16(0);
 832     id->frmw = 7 << 1;
 833     id->lpa = 1 << 0;
 834     id->sqes = (0x6 << 4) | 0x6;
 835     id->cqes = (0x4 << 4) | 0x4;
 836     id->nn = cpu_to_le32(n->num_namespaces);
 837     id->psd[0].mp = cpu_to_le16(0x9c4);
 838     id->psd[0].enlat = cpu_to_le32(0x10);
 839     id->psd[0].exlat = cpu_to_le32(0x4);
 840     if (blk_enable_write_cache(n->conf.blk)) {
 841         id->vwc = 1;
 842     }
 843
 844     n->bar.cap = 0;
 845     NVME_CAP_SET_MQES(n->bar.cap, 0x7ff);
 846     NVME_CAP_SET_CQR(n->bar.cap, 1);
 847     NVME_CAP_SET_AMS(n->bar.cap, 1);
 848     NVME_CAP_SET_TO(n->bar.cap, 0xf);
 849     NVME_CAP_SET_CSS(n->bar.cap, 1);
 850     NVME_CAP_SET_MPSMAX(n->bar.cap, 4);
 851
 852     n->bar.vs = 0x00010100;
 853     n->bar.intmc = n->bar.intms = 0;
 854
 855     for (i = 0; i < n->num_namespaces; i++) {
 856         NvmeNamespace *ns = &n->namespaces[i];
 857         NvmeIdNs *id_ns = &ns->id_ns;
 858         id_ns->nsfeat = 0;
 859         id_ns->nlbaf = 0;
 860         id_ns->flbas = 0;
 861         id_ns->mc = 0;
 862         id_ns->dpc = 0;
 863         id_ns->dps = 0;
 864         id_ns->lbaf[0].ds = BDRV_SECTOR_BITS;
 865         id_ns->ncap  = id_ns->nuse = id_ns->nsze =
 866             cpu_to_le64(n->ns_size >>
 867                 id_ns->lbaf[NVME_ID_NS_FLBAS_INDEX(ns->id_ns.flbas)].ds);
 868     }
 869     return 0;
 870 }
 871
 872 static void nvme_exit(PCIDevice *pci_dev)
 873 {
 874     NvmeCtrl *n = NVME(pci_dev);
 875
 876     nvme_clear_ctrl(n);
 877     g_free(n->namespaces);
 878     g_free(n->cq);
 879     g_free(n->sq);
 880     msix_uninit_exclusive_bar(pci_dev);
 881 }
 882
 883 static Property nvme_props[] = {
 884     DEFINE_BLOCK_PROPERTIES(NvmeCtrl, conf),
 885     DEFINE_PROP_STRING("serial", NvmeCtrl, serial),
 886     DEFINE_PROP_END_OF_LIST(),
 887 };
 888
 889 static const VMStateDescription nvme_vmstate = {
 890     .name = "nvme",
 891     .unmigratable = 1,
 892 };
 893
 894 static void nvme_class_init(ObjectClass *oc, void *data)
 895 {
 896     DeviceClass *dc = DEVICE_CLASS(oc);
 897     PCIDeviceClass *pc = PCI_DEVICE_CLASS(oc);
 898
 899     pc->init = nvme_init;
 900     pc->exit = nvme_exit;
 901     pc->class_id = PCI_CLASS_STORAGE_EXPRESS;
 902     pc->vendor_id = PCI_VENDOR_ID_INTEL;
 903     pc->device_id = 0x5845;
 904     pc->revision = 1;
 905     pc->is_express = 1;
 906
 907     set_bit(DEVICE_CATEGORY_STORAGE, dc->categories);
 908     dc->desc = "Non-Volatile Memory Express";
 909     dc->props = nvme_props;
 910     dc->vmsd = &nvme_vmstate;
 911 }
 912
 913 static void nvme_get_bootindex(Object *obj, Visitor *v, void *opaque,
 914                                   const char *name, Error **errp)
 915 {
 916     NvmeCtrl *s = NVME(obj);
 917
 918     visit_type_int32(v, &s->conf.bootindex, name, errp);
 919 }
 920
 921 static void nvme_set_bootindex(Object *obj, Visitor *v, void *opaque,
 922                                   const char *name, Error **errp)
 923 {
 924     NvmeCtrl *s = NVME(obj);
 925     int32_t boot_index;
 926     Error *local_err = NULL;
 927
 928     visit_type_int32(v, &boot_index, name, &local_err);
 929     if (local_err) {
 930         goto out;
 931     }
 932     /* check whether bootindex is present in fw_boot_order list  */
 933     check_boot_index(boot_index, &local_err);
 934     if (local_err) {
 935         goto out;
 936     }
 937     /* change bootindex to a new one */
 938     s->conf.bootindex = boot_index;
 939
 940 out:
 941     if (local_err) {
 942         error_propagate(errp, local_err);
 943     }
 944 }
 945
 946 static void nvme_instance_init(Object *obj)
 947 {
 948     object_property_add(obj, "bootindex", "int32",
 949                         nvme_get_bootindex,
 950                         nvme_set_bootindex, NULL, NULL, NULL);
 951     object_property_set_int(obj, -1, "bootindex", NULL);
 952 }
 953
 954 static const TypeInfo nvme_info = {
 955     .name          = "nvme",
 956     .parent        = TYPE_PCI_DEVICE,
 957     .instance_size = sizeof(NvmeCtrl),
 958     .class_init    = nvme_class_init,
 959     .instance_init = nvme_instance_init,
 960 };
 961
 962 static void nvme_register_types(void)
 963 {
 964     type_register_static(&nvme_info);
 965 }
 966
 967 type_init(nvme_register_types)