drivers/nvme/nvme.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Copyright (C) 2017 NXP Semiconductors
   4  * Copyright (C) 2017 Bin Meng <[email protected]>
   5  */
   6
   7 #include <common.h>
   8 #include <blk.h>
   9 #include <cpu_func.h>
  10 #include <dm.h>
  11 #include <errno.h>
  12 #include <log.h>
  13 #include <malloc.h>
  14 #include <memalign.h>
  15 #include <pci.h>
  16 #include <time.h>
  17 #include <dm/device-internal.h>
  18 #include <linux/compat.h>
  19 #include "nvme.h"
  20
  21 #define NVME_Q_DEPTH            2
  22 #define NVME_AQ_DEPTH           2
  23 #define NVME_SQ_SIZE(depth)     (depth * sizeof(struct nvme_command))
  24 #define NVME_CQ_SIZE(depth)     (depth * sizeof(struct nvme_completion))
  25 #define ADMIN_TIMEOUT           60
  26 #define IO_TIMEOUT              30
  27 #define MAX_PRP_POOL            512
  28
  29 enum nvme_queue_id {
  30         NVME_ADMIN_Q,
  31         NVME_IO_Q,
  32         NVME_Q_NUM,
  33 };
  34
  35 /*
  36  * An NVM Express queue. Each device has at least two (one for admin
  37  * commands and one for I/O commands).
  38  */
  39 struct nvme_queue {
  40         struct nvme_dev *dev;
  41         struct nvme_command *sq_cmds;
  42         struct nvme_completion *cqes;
  43         wait_queue_head_t sq_full;
  44         u32 __iomem *q_db;
  45         u16 q_depth;
  46         s16 cq_vector;
  47         u16 sq_head;
  48         u16 sq_tail;
  49         u16 cq_head;
  50         u16 qid;
  51         u8 cq_phase;
  52         u8 cqe_seen;
  53         unsigned long cmdid_data[];
  54 };
  55
  56 static int nvme_wait_ready(struct nvme_dev *dev, bool enabled)
  57 {
  58         u32 bit = enabled ? NVME_CSTS_RDY : 0;
  59         int timeout;
  60         ulong start;
  61
  62         /* Timeout field in the CAP register is in 500 millisecond units */
  63         timeout = NVME_CAP_TIMEOUT(dev->cap) * 500;
  64
  65         start = get_timer(0);
  66         while (get_timer(start) < timeout) {
  67                 if ((readl(&dev->bar->csts) & NVME_CSTS_RDY) == bit)
  68                         return 0;
  69         }
  70
  71         return -ETIME;
  72 }
  73
  74 static int nvme_setup_prps(struct nvme_dev *dev, u64 *prp2,
  75                            int total_len, u64 dma_addr)
  76 {
  77         u32 page_size = dev->page_size;
  78         int offset = dma_addr & (page_size - 1);
  79         u64 *prp_pool;
  80         int length = total_len;
  81         int i, nprps;
  82         u32 prps_per_page = (page_size >> 3) - 1;
  83         u32 num_pages;
  84
  85         length -= (page_size - offset);
  86
  87         if (length <= 0) {
  88                 *prp2 = 0;
  89                 return 0;
  90         }
  91
  92         if (length)
  93                 dma_addr += (page_size - offset);
  94
  95         if (length <= page_size) {
  96                 *prp2 = dma_addr;
  97                 return 0;
  98         }
  99
 100         nprps = DIV_ROUND_UP(length, page_size);
 101         num_pages = DIV_ROUND_UP(nprps, prps_per_page);
 102
 103         if (nprps > dev->prp_entry_num) {
 104                 free(dev->prp_pool);
 105                 /*
 106                  * Always increase in increments of pages.  It doesn't waste
 107                  * much memory and reduces the number of allocations.
 108                  */
 109                 dev->prp_pool = memalign(page_size, num_pages * page_size);
 110                 if (!dev->prp_pool) {
 111                         printf("Error: malloc prp_pool fail\n");
 112                         return -ENOMEM;
 113                 }
 114                 dev->prp_entry_num = prps_per_page * num_pages;
 115         }
 116
 117         prp_pool = dev->prp_pool;
 118         i = 0;
 119         while (nprps) {
 120                 if (i == ((page_size >> 3) - 1)) {
 121                         *(prp_pool + i) = cpu_to_le64((ulong)prp_pool +
 122                                         page_size);
 123                         i = 0;
 124                         prp_pool += page_size;
 125                 }
 126                 *(prp_pool + i++) = cpu_to_le64(dma_addr);
 127                 dma_addr += page_size;
 128                 nprps--;
 129         }
 130         *prp2 = (ulong)dev->prp_pool;
 131
 132         flush_dcache_range((ulong)dev->prp_pool, (ulong)dev->prp_pool +
 133                            dev->prp_entry_num * sizeof(u64));
 134
 135         return 0;
 136 }
 137
 138 static __le16 nvme_get_cmd_id(void)
 139 {
 140         static unsigned short cmdid;
 141
 142         return cpu_to_le16((cmdid < USHRT_MAX) ? cmdid++ : 0);
 143 }
 144
 145 static u16 nvme_read_completion_status(struct nvme_queue *nvmeq, u16 index)
 146 {
 147         u64 start = (ulong)&nvmeq->cqes[index];
 148         u64 stop = start + sizeof(struct nvme_completion);
 149
 150         invalidate_dcache_range(start, stop);
 151
 152         return le16_to_cpu(readw(&(nvmeq->cqes[index].status)));
 153 }
 154
 155 /**
 156  * nvme_submit_cmd() - copy a command into a queue and ring the doorbell
 157  *
 158  * @nvmeq:      The queue to use
 159  * @cmd:        The command to send
 160  */
 161 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
 162 {
 163         u16 tail = nvmeq->sq_tail;
 164
 165         memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
 166         flush_dcache_range((ulong)&nvmeq->sq_cmds[tail],
 167                            (ulong)&nvmeq->sq_cmds[tail] + sizeof(*cmd));
 168
 169         if (++tail == nvmeq->q_depth)
 170                 tail = 0;
 171         writel(tail, nvmeq->q_db);
 172         nvmeq->sq_tail = tail;
 173 }
 174
 175 static int nvme_submit_sync_cmd(struct nvme_queue *nvmeq,
 176                                 struct nvme_command *cmd,
 177                                 u32 *result, unsigned timeout)
 178 {
 179         u16 head = nvmeq->cq_head;
 180         u16 phase = nvmeq->cq_phase;
 181         u16 status;
 182         ulong start_time;
 183         ulong timeout_us = timeout * 100000;
 184
 185         cmd->common.command_id = nvme_get_cmd_id();
 186         nvme_submit_cmd(nvmeq, cmd);
 187
 188         start_time = timer_get_us();
 189
 190         for (;;) {
 191                 status = nvme_read_completion_status(nvmeq, head);
 192                 if ((status & 0x01) == phase)
 193                         break;
 194                 if (timeout_us > 0 && (timer_get_us() - start_time)
 195                     >= timeout_us)
 196                         return -ETIMEDOUT;
 197         }
 198
 199         status >>= 1;
 200         if (status) {
 201                 printf("ERROR: status = %x, phase = %d, head = %d\n",
 202                        status, phase, head);
 203                 status = 0;
 204                 if (++head == nvmeq->q_depth) {
 205                         head = 0;
 206                         phase = !phase;
 207                 }
 208                 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 209                 nvmeq->cq_head = head;
 210                 nvmeq->cq_phase = phase;
 211
 212                 return -EIO;
 213         }
 214
 215         if (result)
 216                 *result = le32_to_cpu(readl(&(nvmeq->cqes[head].result)));
 217
 218         if (++head == nvmeq->q_depth) {
 219                 head = 0;
 220                 phase = !phase;
 221         }
 222         writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
 223         nvmeq->cq_head = head;
 224         nvmeq->cq_phase = phase;
 225
 226         return status;
 227 }
 228
 229 static int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
 230                                  u32 *result)
 231 {
 232         return nvme_submit_sync_cmd(dev->queues[NVME_ADMIN_Q], cmd,
 233                                     result, ADMIN_TIMEOUT);
 234 }
 235
 236 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev,
 237                                            int qid, int depth)
 238 {
 239         struct nvme_queue *nvmeq = malloc(sizeof(*nvmeq));
 240         if (!nvmeq)
 241                 return NULL;
 242         memset(nvmeq, 0, sizeof(*nvmeq));
 243
 244         nvmeq->cqes = (void *)memalign(4096, NVME_CQ_SIZE(depth));
 245         if (!nvmeq->cqes)
 246                 goto free_nvmeq;
 247         memset((void *)nvmeq->cqes, 0, NVME_CQ_SIZE(depth));
 248
 249         nvmeq->sq_cmds = (void *)memalign(4096, NVME_SQ_SIZE(depth));
 250         if (!nvmeq->sq_cmds)
 251                 goto free_queue;
 252         memset((void *)nvmeq->sq_cmds, 0, NVME_SQ_SIZE(depth));
 253
 254         nvmeq->dev = dev;
 255
 256         nvmeq->cq_head = 0;
 257         nvmeq->cq_phase = 1;
 258         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
 259         nvmeq->q_depth = depth;
 260         nvmeq->qid = qid;
 261         dev->queue_count++;
 262         dev->queues[qid] = nvmeq;
 263
 264         return nvmeq;
 265
 266  free_queue:
 267         free((void *)nvmeq->cqes);
 268  free_nvmeq:
 269         free(nvmeq);
 270
 271         return NULL;
 272 }
 273
 274 static int nvme_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
 275 {
 276         struct nvme_command c;
 277
 278         memset(&c, 0, sizeof(c));
 279         c.delete_queue.opcode = opcode;
 280         c.delete_queue.qid = cpu_to_le16(id);
 281
 282         return nvme_submit_admin_cmd(dev, &c, NULL);
 283 }
 284
 285 static int nvme_delete_sq(struct nvme_dev *dev, u16 sqid)
 286 {
 287         return nvme_delete_queue(dev, nvme_admin_delete_sq, sqid);
 288 }
 289
 290 static int nvme_delete_cq(struct nvme_dev *dev, u16 cqid)
 291 {
 292         return nvme_delete_queue(dev, nvme_admin_delete_cq, cqid);
 293 }
 294
 295 static int nvme_enable_ctrl(struct nvme_dev *dev)
 296 {
 297         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
 298         dev->ctrl_config |= NVME_CC_ENABLE;
 299         writel(cpu_to_le32(dev->ctrl_config), &dev->bar->cc);
 300
 301         return nvme_wait_ready(dev, true);
 302 }
 303
 304 static int nvme_disable_ctrl(struct nvme_dev *dev)
 305 {
 306         dev->ctrl_config &= ~NVME_CC_SHN_MASK;
 307         dev->ctrl_config &= ~NVME_CC_ENABLE;
 308         writel(cpu_to_le32(dev->ctrl_config), &dev->bar->cc);
 309
 310         return nvme_wait_ready(dev, false);
 311 }
 312
 313 static void nvme_free_queue(struct nvme_queue *nvmeq)
 314 {
 315         free((void *)nvmeq->cqes);
 316         free(nvmeq->sq_cmds);
 317         free(nvmeq);
 318 }
 319
 320 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
 321 {
 322         int i;
 323
 324         for (i = dev->queue_count - 1; i >= lowest; i--) {
 325                 struct nvme_queue *nvmeq = dev->queues[i];
 326                 dev->queue_count--;
 327                 dev->queues[i] = NULL;
 328                 nvme_free_queue(nvmeq);
 329         }
 330 }
 331
 332 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
 333 {
 334         struct nvme_dev *dev = nvmeq->dev;
 335
 336         nvmeq->sq_tail = 0;
 337         nvmeq->cq_head = 0;
 338         nvmeq->cq_phase = 1;
 339         nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
 340         memset((void *)nvmeq->cqes, 0, NVME_CQ_SIZE(nvmeq->q_depth));
 341         flush_dcache_range((ulong)nvmeq->cqes,
 342                            (ulong)nvmeq->cqes + NVME_CQ_SIZE(nvmeq->q_depth));
 343         dev->online_queues++;
 344 }
 345
 346 static int nvme_configure_admin_queue(struct nvme_dev *dev)
 347 {
 348         int result;
 349         u32 aqa;
 350         u64 cap = dev->cap;
 351         struct nvme_queue *nvmeq;
 352         /* most architectures use 4KB as the page size */
 353         unsigned page_shift = 12;
 354         unsigned dev_page_min = NVME_CAP_MPSMIN(cap) + 12;
 355         unsigned dev_page_max = NVME_CAP_MPSMAX(cap) + 12;
 356
 357         if (page_shift < dev_page_min) {
 358                 debug("Device minimum page size (%u) too large for host (%u)\n",
 359                       1 << dev_page_min, 1 << page_shift);
 360                 return -ENODEV;
 361         }
 362
 363         if (page_shift > dev_page_max) {
 364                 debug("Device maximum page size (%u) smaller than host (%u)\n",
 365                       1 << dev_page_max, 1 << page_shift);
 366                 page_shift = dev_page_max;
 367         }
 368
 369         result = nvme_disable_ctrl(dev);
 370         if (result < 0)
 371                 return result;
 372
 373         nvmeq = dev->queues[NVME_ADMIN_Q];
 374         if (!nvmeq) {
 375                 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
 376                 if (!nvmeq)
 377                         return -ENOMEM;
 378         }
 379
 380         aqa = nvmeq->q_depth - 1;
 381         aqa |= aqa << 16;
 382         aqa |= aqa << 16;
 383
 384         dev->page_size = 1 << page_shift;
 385
 386         dev->ctrl_config = NVME_CC_CSS_NVM;
 387         dev->ctrl_config |= (page_shift - 12) << NVME_CC_MPS_SHIFT;
 388         dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
 389         dev->ctrl_config |= NVME_CC_IOSQES | NVME_CC_IOCQES;
 390
 391         writel(aqa, &dev->bar->aqa);
 392         nvme_writeq((ulong)nvmeq->sq_cmds, &dev->bar->asq);
 393         nvme_writeq((ulong)nvmeq->cqes, &dev->bar->acq);
 394
 395         result = nvme_enable_ctrl(dev);
 396         if (result)
 397                 goto free_nvmeq;
 398
 399         nvmeq->cq_vector = 0;
 400
 401         nvme_init_queue(dev->queues[NVME_ADMIN_Q], 0);
 402
 403         return result;
 404
 405  free_nvmeq:
 406         nvme_free_queues(dev, 0);
 407
 408         return result;
 409 }
 410
 411 static int nvme_alloc_cq(struct nvme_dev *dev, u16 qid,
 412                             struct nvme_queue *nvmeq)
 413 {
 414         struct nvme_command c;
 415         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
 416
 417         memset(&c, 0, sizeof(c));
 418         c.create_cq.opcode = nvme_admin_create_cq;
 419         c.create_cq.prp1 = cpu_to_le64((ulong)nvmeq->cqes);
 420         c.create_cq.cqid = cpu_to_le16(qid);
 421         c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 422         c.create_cq.cq_flags = cpu_to_le16(flags);
 423         c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
 424
 425         return nvme_submit_admin_cmd(dev, &c, NULL);
 426 }
 427
 428 static int nvme_alloc_sq(struct nvme_dev *dev, u16 qid,
 429                             struct nvme_queue *nvmeq)
 430 {
 431         struct nvme_command c;
 432         int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
 433
 434         memset(&c, 0, sizeof(c));
 435         c.create_sq.opcode = nvme_admin_create_sq;
 436         c.create_sq.prp1 = cpu_to_le64((ulong)nvmeq->sq_cmds);
 437         c.create_sq.sqid = cpu_to_le16(qid);
 438         c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
 439         c.create_sq.sq_flags = cpu_to_le16(flags);
 440         c.create_sq.cqid = cpu_to_le16(qid);
 441
 442         return nvme_submit_admin_cmd(dev, &c, NULL);
 443 }
 444
 445 int nvme_identify(struct nvme_dev *dev, unsigned nsid,
 446                   unsigned cns, dma_addr_t dma_addr)
 447 {
 448         struct nvme_command c;
 449         u32 page_size = dev->page_size;
 450         int offset = dma_addr & (page_size - 1);
 451         int length = sizeof(struct nvme_id_ctrl);
 452         int ret;
 453
 454         memset(&c, 0, sizeof(c));
 455         c.identify.opcode = nvme_admin_identify;
 456         c.identify.nsid = cpu_to_le32(nsid);
 457         c.identify.prp1 = cpu_to_le64(dma_addr);
 458
 459         length -= (page_size - offset);
 460         if (length <= 0) {
 461                 c.identify.prp2 = 0;
 462         } else {
 463                 dma_addr += (page_size - offset);
 464                 c.identify.prp2 = cpu_to_le64(dma_addr);
 465         }
 466
 467         c.identify.cns = cpu_to_le32(cns);
 468
 469         invalidate_dcache_range(dma_addr,
 470                                 dma_addr + sizeof(struct nvme_id_ctrl));
 471
 472         ret = nvme_submit_admin_cmd(dev, &c, NULL);
 473         if (!ret)
 474                 invalidate_dcache_range(dma_addr,
 475                                         dma_addr + sizeof(struct nvme_id_ctrl));
 476
 477         return ret;
 478 }
 479
 480 int nvme_get_features(struct nvme_dev *dev, unsigned fid, unsigned nsid,
 481                       dma_addr_t dma_addr, u32 *result)
 482 {
 483         struct nvme_command c;
 484
 485         memset(&c, 0, sizeof(c));
 486         c.features.opcode = nvme_admin_get_features;
 487         c.features.nsid = cpu_to_le32(nsid);
 488         c.features.prp1 = cpu_to_le64(dma_addr);
 489         c.features.fid = cpu_to_le32(fid);
 490
 491         /*
 492          * TODO: add cache invalidate operation when the size of
 493          * the DMA buffer is known
 494          */
 495
 496         return nvme_submit_admin_cmd(dev, &c, result);
 497 }
 498
 499 int nvme_set_features(struct nvme_dev *dev, unsigned fid, unsigned dword11,
 500                       dma_addr_t dma_addr, u32 *result)
 501 {
 502         struct nvme_command c;
 503
 504         memset(&c, 0, sizeof(c));
 505         c.features.opcode = nvme_admin_set_features;
 506         c.features.prp1 = cpu_to_le64(dma_addr);
 507         c.features.fid = cpu_to_le32(fid);
 508         c.features.dword11 = cpu_to_le32(dword11);
 509
 510         /*
 511          * TODO: add cache flush operation when the size of
 512          * the DMA buffer is known
 513          */
 514
 515         return nvme_submit_admin_cmd(dev, &c, result);
 516 }
 517
 518 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
 519 {
 520         struct nvme_dev *dev = nvmeq->dev;
 521         int result;
 522
 523         nvmeq->cq_vector = qid - 1;
 524         result = nvme_alloc_cq(dev, qid, nvmeq);
 525         if (result < 0)
 526                 goto release_cq;
 527
 528         result = nvme_alloc_sq(dev, qid, nvmeq);
 529         if (result < 0)
 530                 goto release_sq;
 531
 532         nvme_init_queue(nvmeq, qid);
 533
 534         return result;
 535
 536  release_sq:
 537         nvme_delete_sq(dev, qid);
 538  release_cq:
 539         nvme_delete_cq(dev, qid);
 540
 541         return result;
 542 }
 543
 544 static int nvme_set_queue_count(struct nvme_dev *dev, int count)
 545 {
 546         int status;
 547         u32 result;
 548         u32 q_count = (count - 1) | ((count - 1) << 16);
 549
 550         status = nvme_set_features(dev, NVME_FEAT_NUM_QUEUES,
 551                         q_count, 0, &result);
 552
 553         if (status < 0)
 554                 return status;
 555         if (status > 1)
 556                 return 0;
 557
 558         return min(result & 0xffff, result >> 16) + 1;
 559 }
 560
 561 static void nvme_create_io_queues(struct nvme_dev *dev)
 562 {
 563         unsigned int i;
 564
 565         for (i = dev->queue_count; i <= dev->max_qid; i++)
 566                 if (!nvme_alloc_queue(dev, i, dev->q_depth))
 567                         break;
 568
 569         for (i = dev->online_queues; i <= dev->queue_count - 1; i++)
 570                 if (nvme_create_queue(dev->queues[i], i))
 571                         break;
 572 }
 573
 574 static int nvme_setup_io_queues(struct nvme_dev *dev)
 575 {
 576         int nr_io_queues;
 577         int result;
 578
 579         nr_io_queues = 1;
 580         result = nvme_set_queue_count(dev, nr_io_queues);
 581         if (result <= 0)
 582                 return result;
 583
 584         dev->max_qid = nr_io_queues;
 585
 586         /* Free previously allocated queues */
 587         nvme_free_queues(dev, nr_io_queues + 1);
 588         nvme_create_io_queues(dev);
 589
 590         return 0;
 591 }
 592
 593 static int nvme_get_info_from_identify(struct nvme_dev *dev)
 594 {
 595         struct nvme_id_ctrl *ctrl;
 596         int ret;
 597         int shift = NVME_CAP_MPSMIN(dev->cap) + 12;
 598
 599         ctrl = memalign(dev->page_size, sizeof(struct nvme_id_ctrl));
 600         if (!ctrl)
 601                 return -ENOMEM;
 602
 603         ret = nvme_identify(dev, 0, 1, (dma_addr_t)(long)ctrl);
 604         if (ret) {
 605                 free(ctrl);
 606                 return -EIO;
 607         }
 608
 609         dev->nn = le32_to_cpu(ctrl->nn);
 610         dev->vwc = ctrl->vwc;
 611         memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
 612         memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
 613         memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
 614         if (ctrl->mdts)
 615                 dev->max_transfer_shift = (ctrl->mdts + shift);
 616         else {
 617                 /*
 618                  * Maximum Data Transfer Size (MDTS) field indicates the maximum
 619                  * data transfer size between the host and the controller. The
 620                  * host should not submit a command that exceeds this transfer
 621                  * size. The value is in units of the minimum memory page size
 622                  * and is reported as a power of two (2^n).
 623                  *
 624                  * The spec also says: a value of 0h indicates no restrictions
 625                  * on transfer size. But in nvme_blk_read/write() below we have
 626                  * the following algorithm for maximum number of logic blocks
 627                  * per transfer:
 628                  *
 629                  * u16 lbas = 1 << (dev->max_transfer_shift - ns->lba_shift);
 630                  *
 631                  * In order for lbas not to overflow, the maximum number is 15
 632                  * which means dev->max_transfer_shift = 15 + 9 (ns->lba_shift).
 633                  * Let's use 20 which provides 1MB size.
 634                  */
 635                 dev->max_transfer_shift = 20;
 636         }
 637
 638         free(ctrl);
 639         return 0;
 640 }
 641
 642 int nvme_get_namespace_id(struct udevice *udev, u32 *ns_id, u8 *eui64)
 643 {
 644         struct nvme_ns *ns = dev_get_priv(udev);
 645
 646         if (ns_id)
 647                 *ns_id = ns->ns_id;
 648         if (eui64)
 649                 memcpy(eui64, ns->eui64, sizeof(ns->eui64));
 650
 651         return 0;
 652 }
 653
 654 int nvme_scan_namespace(void)
 655 {
 656         struct uclass *uc;
 657         struct udevice *dev;
 658         int ret;
 659
 660         ret = uclass_get(UCLASS_NVME, &uc);
 661         if (ret)
 662                 return ret;
 663
 664         uclass_foreach_dev(dev, uc) {
 665                 ret = device_probe(dev);
 666                 if (ret)
 667                         return ret;
 668         }
 669
 670         return 0;
 671 }
 672
 673 static int nvme_blk_probe(struct udevice *udev)
 674 {
 675         struct nvme_dev *ndev = dev_get_priv(udev->parent);
 676         struct blk_desc *desc = dev_get_uclass_platdata(udev);
 677         struct nvme_ns *ns = dev_get_priv(udev);
 678         u8 flbas;
 679         struct pci_child_platdata *pplat;
 680         struct nvme_id_ns *id;
 681
 682         id = memalign(ndev->page_size, sizeof(struct nvme_id_ns));
 683         if (!id)
 684                 return -ENOMEM;
 685
 686         memset(ns, 0, sizeof(*ns));
 687         ns->dev = ndev;
 688         /* extract the namespace id from the block device name */
 689         ns->ns_id = trailing_strtol(udev->name) + 1;
 690         if (nvme_identify(ndev, ns->ns_id, 0, (dma_addr_t)(long)id)) {
 691                 free(id);
 692                 return -EIO;
 693         }
 694
 695         memcpy(&ns->eui64, &id->eui64, sizeof(id->eui64));
 696         flbas = id->flbas & NVME_NS_FLBAS_LBA_MASK;
 697         ns->flbas = flbas;
 698         ns->lba_shift = id->lbaf[flbas].ds;
 699         ns->mode_select_num_blocks = le64_to_cpu(id->nsze);
 700         ns->mode_select_block_len = 1 << ns->lba_shift;
 701         list_add(&ns->list, &ndev->namespaces);
 702
 703         desc->lba = ns->mode_select_num_blocks;
 704         desc->log2blksz = ns->lba_shift;
 705         desc->blksz = 1 << ns->lba_shift;
 706         desc->bdev = udev;
 707         pplat = dev_get_parent_platdata(udev->parent);
 708         sprintf(desc->vendor, "0x%.4x", pplat->vendor);
 709         memcpy(desc->product, ndev->serial, sizeof(ndev->serial));
 710         memcpy(desc->revision, ndev->firmware_rev, sizeof(ndev->firmware_rev));
 711
 712         free(id);
 713         return 0;
 714 }
 715
 716 static ulong nvme_blk_rw(struct udevice *udev, lbaint_t blknr,
 717                          lbaint_t blkcnt, void *buffer, bool read)
 718 {
 719         struct nvme_ns *ns = dev_get_priv(udev);
 720         struct nvme_dev *dev = ns->dev;
 721         struct nvme_command c;
 722         struct blk_desc *desc = dev_get_uclass_platdata(udev);
 723         int status;
 724         u64 prp2;
 725         u64 total_len = blkcnt << desc->log2blksz;
 726         u64 temp_len = total_len;
 727
 728         u64 slba = blknr;
 729         u16 lbas = 1 << (dev->max_transfer_shift - ns->lba_shift);
 730         u64 total_lbas = blkcnt;
 731
 732         flush_dcache_range((unsigned long)buffer,
 733                            (unsigned long)buffer + total_len);
 734
 735         c.rw.opcode = read ? nvme_cmd_read : nvme_cmd_write;
 736         c.rw.flags = 0;
 737         c.rw.nsid = cpu_to_le32(ns->ns_id);
 738         c.rw.control = 0;
 739         c.rw.dsmgmt = 0;
 740         c.rw.reftag = 0;
 741         c.rw.apptag = 0;
 742         c.rw.appmask = 0;
 743         c.rw.metadata = 0;
 744
 745         while (total_lbas) {
 746                 if (total_lbas < lbas) {
 747                         lbas = (u16)total_lbas;
 748                         total_lbas = 0;
 749                 } else {
 750                         total_lbas -= lbas;
 751                 }
 752
 753                 if (nvme_setup_prps(dev, &prp2,
 754                                     lbas << ns->lba_shift, (ulong)buffer))
 755                         return -EIO;
 756                 c.rw.slba = cpu_to_le64(slba);
 757                 slba += lbas;
 758                 c.rw.length = cpu_to_le16(lbas - 1);
 759                 c.rw.prp1 = cpu_to_le64((ulong)buffer);
 760                 c.rw.prp2 = cpu_to_le64(prp2);
 761                 status = nvme_submit_sync_cmd(dev->queues[NVME_IO_Q],
 762                                 &c, NULL, IO_TIMEOUT);
 763                 if (status)
 764                         break;
 765                 temp_len -= (u32)lbas << ns->lba_shift;
 766                 buffer += lbas << ns->lba_shift;
 767         }
 768
 769         if (read)
 770                 invalidate_dcache_range((unsigned long)buffer,
 771                                         (unsigned long)buffer + total_len);
 772
 773         return (total_len - temp_len) >> desc->log2blksz;
 774 }
 775
 776 static ulong nvme_blk_read(struct udevice *udev, lbaint_t blknr,
 777                            lbaint_t blkcnt, void *buffer)
 778 {
 779         return nvme_blk_rw(udev, blknr, blkcnt, buffer, true);
 780 }
 781
 782 static ulong nvme_blk_write(struct udevice *udev, lbaint_t blknr,
 783                             lbaint_t blkcnt, const void *buffer)
 784 {
 785         return nvme_blk_rw(udev, blknr, blkcnt, (void *)buffer, false);
 786 }
 787
 788 static const struct blk_ops nvme_blk_ops = {
 789         .read   = nvme_blk_read,
 790         .write  = nvme_blk_write,
 791 };
 792
 793 U_BOOT_DRIVER(nvme_blk) = {
 794         .name   = "nvme-blk",
 795         .id     = UCLASS_BLK,
 796         .probe  = nvme_blk_probe,
 797         .ops    = &nvme_blk_ops,
 798         .priv_auto      = sizeof(struct nvme_ns),
 799 };
 800
 801 static int nvme_bind(struct udevice *udev)
 802 {
 803         static int ndev_num;
 804         char name[20];
 805
 806         sprintf(name, "nvme#%d", ndev_num++);
 807
 808         return device_set_name(udev, name);
 809 }
 810
 811 static int nvme_probe(struct udevice *udev)
 812 {
 813         int ret;
 814         struct nvme_dev *ndev = dev_get_priv(udev);
 815
 816         ndev->instance = trailing_strtol(udev->name);
 817
 818         INIT_LIST_HEAD(&ndev->namespaces);
 819         ndev->bar = dm_pci_map_bar(udev, PCI_BASE_ADDRESS_0,
 820                         PCI_REGION_MEM);
 821         if (readl(&ndev->bar->csts) == -1) {
 822                 ret = -ENODEV;
 823                 printf("Error: %s: Out of memory!\n", udev->name);
 824                 goto free_nvme;
 825         }
 826
 827         ndev->queues = malloc(NVME_Q_NUM * sizeof(struct nvme_queue *));
 828         if (!ndev->queues) {
 829                 ret = -ENOMEM;
 830                 printf("Error: %s: Out of memory!\n", udev->name);
 831                 goto free_nvme;
 832         }
 833         memset(ndev->queues, 0, NVME_Q_NUM * sizeof(struct nvme_queue *));
 834
 835         ndev->cap = nvme_readq(&ndev->bar->cap);
 836         ndev->q_depth = min_t(int, NVME_CAP_MQES(ndev->cap) + 1, NVME_Q_DEPTH);
 837         ndev->db_stride = 1 << NVME_CAP_STRIDE(ndev->cap);
 838         ndev->dbs = ((void __iomem *)ndev->bar) + 4096;
 839
 840         ret = nvme_configure_admin_queue(ndev);
 841         if (ret)
 842                 goto free_queue;
 843
 844         /* Allocate after the page size is known */
 845         ndev->prp_pool = memalign(ndev->page_size, MAX_PRP_POOL);
 846         if (!ndev->prp_pool) {
 847                 ret = -ENOMEM;
 848                 printf("Error: %s: Out of memory!\n", udev->name);
 849                 goto free_nvme;
 850         }
 851         ndev->prp_entry_num = MAX_PRP_POOL >> 3;
 852
 853         ret = nvme_setup_io_queues(ndev);
 854         if (ret)
 855                 goto free_queue;
 856
 857         nvme_get_info_from_identify(ndev);
 858
 859         return 0;
 860
 861 free_queue:
 862         free((void *)ndev->queues);
 863 free_nvme:
 864         return ret;
 865 }
 866
 867 U_BOOT_DRIVER(nvme) = {
 868         .name   = "nvme",
 869         .id     = UCLASS_NVME,
 870         .bind   = nvme_bind,
 871         .probe  = nvme_probe,
 872         .priv_auto      = sizeof(struct nvme_dev),
 873 };
 874
 875 struct pci_device_id nvme_supported[] = {
 876         { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, ~0) },
 877         {}
 878 };
 879
 880 U_BOOT_PCI_DEVICE(nvme, nvme_supported);