1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * NVM Express device driver 4 * Copyright (c) 2011-2014, Intel Corporation. 5 */ 6 7 #include <linux/acpi.h> 8 #include <linux/aer.h> 9 #include <linux/async.h> 10 #include <linux/blkdev.h> 11 #include <linux/blk-mq.h> 12 #include <linux/blk-mq-pci.h> 13 #include <linux/blk-integrity.h> 14 #include <linux/dmi.h> 15 #include <linux/init.h> 16 #include <linux/interrupt.h> 17 #include <linux/io.h> 18 #include <linux/memremap.h> 19 #include <linux/mm.h> 20 #include <linux/module.h> 21 #include <linux/mutex.h> 22 #include <linux/once.h> 23 #include <linux/pci.h> 24 #include <linux/suspend.h> 25 #include <linux/t10-pi.h> 26 #include <linux/types.h> 27 #include <linux/io-64-nonatomic-lo-hi.h> 28 #include <linux/io-64-nonatomic-hi-lo.h> 29 #include <linux/sed-opal.h> 30 #include <linux/pci-p2pdma.h> 31 32 #include "trace.h" 33 #include "nvme.h" 34 35 #define SQ_SIZE(q) ((q)->q_depth << (q)->sqes) 36 #define CQ_SIZE(q) ((q)->q_depth * sizeof(struct nvme_completion)) 37 38 #define SGES_PER_PAGE (PAGE_SIZE / sizeof(struct nvme_sgl_desc)) 39 40 /* 41 * These can be higher, but we need to ensure that any command doesn't 42 * require an sg allocation that needs more than a page of data. 43 */ 44 #define NVME_MAX_KB_SZ 4096 45 #define NVME_MAX_SEGS 127 46 47 static int use_threaded_interrupts; 48 module_param(use_threaded_interrupts, int, 0444); 49 50 static bool use_cmb_sqes = true; 51 module_param(use_cmb_sqes, bool, 0444); 52 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes"); 53 54 static unsigned int max_host_mem_size_mb = 128; 55 module_param(max_host_mem_size_mb, uint, 0444); 56 MODULE_PARM_DESC(max_host_mem_size_mb, 57 "Maximum Host Memory Buffer (HMB) size per controller (in MiB)"); 58 59 static unsigned int sgl_threshold = SZ_32K; 60 module_param(sgl_threshold, uint, 0644); 61 MODULE_PARM_DESC(sgl_threshold, 62 "Use SGLs when average request segment size is larger or equal to " 63 "this size. Use 0 to disable SGLs."); 64 65 #define NVME_PCI_MIN_QUEUE_SIZE 2 66 #define NVME_PCI_MAX_QUEUE_SIZE 4095 67 static int io_queue_depth_set(const char *val, const struct kernel_param *kp); 68 static const struct kernel_param_ops io_queue_depth_ops = { 69 .set = io_queue_depth_set, 70 .get = param_get_uint, 71 }; 72 73 static unsigned int io_queue_depth = 1024; 74 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644); 75 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2 and < 4096"); 76 77 static int io_queue_count_set(const char *val, const struct kernel_param *kp) 78 { 79 unsigned int n; 80 int ret; 81 82 ret = kstrtouint(val, 10, &n); 83 if (ret != 0 || n > num_possible_cpus()) 84 return -EINVAL; 85 return param_set_uint(val, kp); 86 } 87 88 static const struct kernel_param_ops io_queue_count_ops = { 89 .set = io_queue_count_set, 90 .get = param_get_uint, 91 }; 92 93 static unsigned int write_queues; 94 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644); 95 MODULE_PARM_DESC(write_queues, 96 "Number of queues to use for writes. If not set, reads and writes " 97 "will share a queue set."); 98 99 static unsigned int poll_queues; 100 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644); 101 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO."); 102 103 static bool noacpi; 104 module_param(noacpi, bool, 0444); 105 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks"); 106 107 struct nvme_dev; 108 struct nvme_queue; 109 110 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown); 111 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode); 112 113 /* 114 * Represents an NVM Express device. Each nvme_dev is a PCI function. 115 */ 116 struct nvme_dev { 117 struct nvme_queue *queues; 118 struct blk_mq_tag_set tagset; 119 struct blk_mq_tag_set admin_tagset; 120 u32 __iomem *dbs; 121 struct device *dev; 122 struct dma_pool *prp_page_pool; 123 struct dma_pool *prp_small_pool; 124 unsigned online_queues; 125 unsigned max_qid; 126 unsigned io_queues[HCTX_MAX_TYPES]; 127 unsigned int num_vecs; 128 u32 q_depth; 129 int io_sqes; 130 u32 db_stride; 131 void __iomem *bar; 132 unsigned long bar_mapped_size; 133 struct work_struct remove_work; 134 struct mutex shutdown_lock; 135 bool subsystem; 136 u64 cmb_size; 137 bool cmb_use_sqes; 138 u32 cmbsz; 139 u32 cmbloc; 140 struct nvme_ctrl ctrl; 141 u32 last_ps; 142 bool hmb; 143 144 mempool_t *iod_mempool; 145 146 /* shadow doorbell buffer support: */ 147 u32 *dbbuf_dbs; 148 dma_addr_t dbbuf_dbs_dma_addr; 149 u32 *dbbuf_eis; 150 dma_addr_t dbbuf_eis_dma_addr; 151 152 /* host memory buffer support: */ 153 u64 host_mem_size; 154 u32 nr_host_mem_descs; 155 dma_addr_t host_mem_descs_dma; 156 struct nvme_host_mem_buf_desc *host_mem_descs; 157 void **host_mem_desc_bufs; 158 unsigned int nr_allocated_queues; 159 unsigned int nr_write_queues; 160 unsigned int nr_poll_queues; 161 162 bool attrs_added; 163 }; 164 165 static int io_queue_depth_set(const char *val, const struct kernel_param *kp) 166 { 167 return param_set_uint_minmax(val, kp, NVME_PCI_MIN_QUEUE_SIZE, 168 NVME_PCI_MAX_QUEUE_SIZE); 169 } 170 171 static inline unsigned int sq_idx(unsigned int qid, u32 stride) 172 { 173 return qid * 2 * stride; 174 } 175 176 static inline unsigned int cq_idx(unsigned int qid, u32 stride) 177 { 178 return (qid * 2 + 1) * stride; 179 } 180 181 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl) 182 { 183 return container_of(ctrl, struct nvme_dev, ctrl); 184 } 185 186 /* 187 * An NVM Express queue. Each device has at least two (one for admin 188 * commands and one for I/O commands). 189 */ 190 struct nvme_queue { 191 struct nvme_dev *dev; 192 spinlock_t sq_lock; 193 void *sq_cmds; 194 /* only used for poll queues: */ 195 spinlock_t cq_poll_lock ____cacheline_aligned_in_smp; 196 struct nvme_completion *cqes; 197 dma_addr_t sq_dma_addr; 198 dma_addr_t cq_dma_addr; 199 u32 __iomem *q_db; 200 u32 q_depth; 201 u16 cq_vector; 202 u16 sq_tail; 203 u16 last_sq_tail; 204 u16 cq_head; 205 u16 qid; 206 u8 cq_phase; 207 u8 sqes; 208 unsigned long flags; 209 #define NVMEQ_ENABLED 0 210 #define NVMEQ_SQ_CMB 1 211 #define NVMEQ_DELETE_ERROR 2 212 #define NVMEQ_POLLED 3 213 u32 *dbbuf_sq_db; 214 u32 *dbbuf_cq_db; 215 u32 *dbbuf_sq_ei; 216 u32 *dbbuf_cq_ei; 217 struct completion delete_done; 218 }; 219 220 /* 221 * The nvme_iod describes the data in an I/O. 222 * 223 * The sg pointer contains the list of PRP/SGL chunk allocations in addition 224 * to the actual struct scatterlist. 225 */ 226 struct nvme_iod { 227 struct nvme_request req; 228 struct nvme_command cmd; 229 struct nvme_queue *nvmeq; 230 bool use_sgl; 231 int aborted; 232 int npages; /* In the PRP list. 0 means small pool in use */ 233 int nents; /* Used in scatterlist */ 234 dma_addr_t first_dma; 235 unsigned int dma_len; /* length of single DMA segment mapping */ 236 dma_addr_t meta_dma; 237 struct scatterlist *sg; 238 }; 239 240 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev) 241 { 242 return dev->nr_allocated_queues * 8 * dev->db_stride; 243 } 244 245 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev) 246 { 247 unsigned int mem_size = nvme_dbbuf_size(dev); 248 249 if (dev->dbbuf_dbs) { 250 /* 251 * Clear the dbbuf memory so the driver doesn't observe stale 252 * values from the previous instantiation. 253 */ 254 memset(dev->dbbuf_dbs, 0, mem_size); 255 memset(dev->dbbuf_eis, 0, mem_size); 256 return 0; 257 } 258 259 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size, 260 &dev->dbbuf_dbs_dma_addr, 261 GFP_KERNEL); 262 if (!dev->dbbuf_dbs) 263 return -ENOMEM; 264 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size, 265 &dev->dbbuf_eis_dma_addr, 266 GFP_KERNEL); 267 if (!dev->dbbuf_eis) { 268 dma_free_coherent(dev->dev, mem_size, 269 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr); 270 dev->dbbuf_dbs = NULL; 271 return -ENOMEM; 272 } 273 274 return 0; 275 } 276 277 static void nvme_dbbuf_dma_free(struct nvme_dev *dev) 278 { 279 unsigned int mem_size = nvme_dbbuf_size(dev); 280 281 if (dev->dbbuf_dbs) { 282 dma_free_coherent(dev->dev, mem_size, 283 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr); 284 dev->dbbuf_dbs = NULL; 285 } 286 if (dev->dbbuf_eis) { 287 dma_free_coherent(dev->dev, mem_size, 288 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr); 289 dev->dbbuf_eis = NULL; 290 } 291 } 292 293 static void nvme_dbbuf_init(struct nvme_dev *dev, 294 struct nvme_queue *nvmeq, int qid) 295 { 296 if (!dev->dbbuf_dbs || !qid) 297 return; 298 299 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)]; 300 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)]; 301 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)]; 302 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)]; 303 } 304 305 static void nvme_dbbuf_free(struct nvme_queue *nvmeq) 306 { 307 if (!nvmeq->qid) 308 return; 309 310 nvmeq->dbbuf_sq_db = NULL; 311 nvmeq->dbbuf_cq_db = NULL; 312 nvmeq->dbbuf_sq_ei = NULL; 313 nvmeq->dbbuf_cq_ei = NULL; 314 } 315 316 static void nvme_dbbuf_set(struct nvme_dev *dev) 317 { 318 struct nvme_command c = { }; 319 unsigned int i; 320 321 if (!dev->dbbuf_dbs) 322 return; 323 324 c.dbbuf.opcode = nvme_admin_dbbuf; 325 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr); 326 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr); 327 328 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) { 329 dev_warn(dev->ctrl.device, "unable to set dbbuf\n"); 330 /* Free memory and continue on */ 331 nvme_dbbuf_dma_free(dev); 332 333 for (i = 1; i <= dev->online_queues; i++) 334 nvme_dbbuf_free(&dev->queues[i]); 335 } 336 } 337 338 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old) 339 { 340 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old); 341 } 342 343 /* Update dbbuf and return true if an MMIO is required */ 344 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db, 345 volatile u32 *dbbuf_ei) 346 { 347 if (dbbuf_db) { 348 u16 old_value; 349 350 /* 351 * Ensure that the queue is written before updating 352 * the doorbell in memory 353 */ 354 wmb(); 355 356 old_value = *dbbuf_db; 357 *dbbuf_db = value; 358 359 /* 360 * Ensure that the doorbell is updated before reading the event 361 * index from memory. The controller needs to provide similar 362 * ordering to ensure the envent index is updated before reading 363 * the doorbell. 364 */ 365 mb(); 366 367 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value)) 368 return false; 369 } 370 371 return true; 372 } 373 374 /* 375 * Will slightly overestimate the number of pages needed. This is OK 376 * as it only leads to a small amount of wasted memory for the lifetime of 377 * the I/O. 378 */ 379 static int nvme_pci_npages_prp(void) 380 { 381 unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE, 382 NVME_CTRL_PAGE_SIZE); 383 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8); 384 } 385 386 /* 387 * Calculates the number of pages needed for the SGL segments. For example a 4k 388 * page can accommodate 256 SGL descriptors. 389 */ 390 static int nvme_pci_npages_sgl(void) 391 { 392 return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc), 393 PAGE_SIZE); 394 } 395 396 static size_t nvme_pci_iod_alloc_size(void) 397 { 398 size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl()); 399 400 return sizeof(__le64 *) * npages + 401 sizeof(struct scatterlist) * NVME_MAX_SEGS; 402 } 403 404 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 405 unsigned int hctx_idx) 406 { 407 struct nvme_dev *dev = data; 408 struct nvme_queue *nvmeq = &dev->queues[0]; 409 410 WARN_ON(hctx_idx != 0); 411 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags); 412 413 hctx->driver_data = nvmeq; 414 return 0; 415 } 416 417 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, 418 unsigned int hctx_idx) 419 { 420 struct nvme_dev *dev = data; 421 struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1]; 422 423 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags); 424 hctx->driver_data = nvmeq; 425 return 0; 426 } 427 428 static int nvme_pci_init_request(struct blk_mq_tag_set *set, 429 struct request *req, unsigned int hctx_idx, 430 unsigned int numa_node) 431 { 432 struct nvme_dev *dev = set->driver_data; 433 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 434 int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0; 435 struct nvme_queue *nvmeq = &dev->queues[queue_idx]; 436 437 BUG_ON(!nvmeq); 438 iod->nvmeq = nvmeq; 439 440 nvme_req(req)->ctrl = &dev->ctrl; 441 nvme_req(req)->cmd = &iod->cmd; 442 return 0; 443 } 444 445 static int queue_irq_offset(struct nvme_dev *dev) 446 { 447 /* if we have more than 1 vec, admin queue offsets us by 1 */ 448 if (dev->num_vecs > 1) 449 return 1; 450 451 return 0; 452 } 453 454 static int nvme_pci_map_queues(struct blk_mq_tag_set *set) 455 { 456 struct nvme_dev *dev = set->driver_data; 457 int i, qoff, offset; 458 459 offset = queue_irq_offset(dev); 460 for (i = 0, qoff = 0; i < set->nr_maps; i++) { 461 struct blk_mq_queue_map *map = &set->map[i]; 462 463 map->nr_queues = dev->io_queues[i]; 464 if (!map->nr_queues) { 465 BUG_ON(i == HCTX_TYPE_DEFAULT); 466 continue; 467 } 468 469 /* 470 * The poll queue(s) doesn't have an IRQ (and hence IRQ 471 * affinity), so use the regular blk-mq cpu mapping 472 */ 473 map->queue_offset = qoff; 474 if (i != HCTX_TYPE_POLL && offset) 475 blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset); 476 else 477 blk_mq_map_queues(map); 478 qoff += map->nr_queues; 479 offset += map->nr_queues; 480 } 481 482 return 0; 483 } 484 485 /* 486 * Write sq tail if we are asked to, or if the next command would wrap. 487 */ 488 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq) 489 { 490 if (!write_sq) { 491 u16 next_tail = nvmeq->sq_tail + 1; 492 493 if (next_tail == nvmeq->q_depth) 494 next_tail = 0; 495 if (next_tail != nvmeq->last_sq_tail) 496 return; 497 } 498 499 if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail, 500 nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei)) 501 writel(nvmeq->sq_tail, nvmeq->q_db); 502 nvmeq->last_sq_tail = nvmeq->sq_tail; 503 } 504 505 static inline void nvme_sq_copy_cmd(struct nvme_queue *nvmeq, 506 struct nvme_command *cmd) 507 { 508 memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes), 509 absolute_pointer(cmd), sizeof(*cmd)); 510 if (++nvmeq->sq_tail == nvmeq->q_depth) 511 nvmeq->sq_tail = 0; 512 } 513 514 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx) 515 { 516 struct nvme_queue *nvmeq = hctx->driver_data; 517 518 spin_lock(&nvmeq->sq_lock); 519 if (nvmeq->sq_tail != nvmeq->last_sq_tail) 520 nvme_write_sq_db(nvmeq, true); 521 spin_unlock(&nvmeq->sq_lock); 522 } 523 524 static void **nvme_pci_iod_list(struct request *req) 525 { 526 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 527 return (void **)(iod->sg + blk_rq_nr_phys_segments(req)); 528 } 529 530 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req) 531 { 532 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 533 int nseg = blk_rq_nr_phys_segments(req); 534 unsigned int avg_seg_size; 535 536 avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg); 537 538 if (!nvme_ctrl_sgl_supported(&dev->ctrl)) 539 return false; 540 if (!iod->nvmeq->qid) 541 return false; 542 if (!sgl_threshold || avg_seg_size < sgl_threshold) 543 return false; 544 return true; 545 } 546 547 static void nvme_free_prps(struct nvme_dev *dev, struct request *req) 548 { 549 const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1; 550 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 551 dma_addr_t dma_addr = iod->first_dma; 552 int i; 553 554 for (i = 0; i < iod->npages; i++) { 555 __le64 *prp_list = nvme_pci_iod_list(req)[i]; 556 dma_addr_t next_dma_addr = le64_to_cpu(prp_list[last_prp]); 557 558 dma_pool_free(dev->prp_page_pool, prp_list, dma_addr); 559 dma_addr = next_dma_addr; 560 } 561 } 562 563 static void nvme_free_sgls(struct nvme_dev *dev, struct request *req) 564 { 565 const int last_sg = SGES_PER_PAGE - 1; 566 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 567 dma_addr_t dma_addr = iod->first_dma; 568 int i; 569 570 for (i = 0; i < iod->npages; i++) { 571 struct nvme_sgl_desc *sg_list = nvme_pci_iod_list(req)[i]; 572 dma_addr_t next_dma_addr = le64_to_cpu((sg_list[last_sg]).addr); 573 574 dma_pool_free(dev->prp_page_pool, sg_list, dma_addr); 575 dma_addr = next_dma_addr; 576 } 577 } 578 579 static void nvme_unmap_sg(struct nvme_dev *dev, struct request *req) 580 { 581 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 582 583 if (is_pci_p2pdma_page(sg_page(iod->sg))) 584 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents, 585 rq_dma_dir(req)); 586 else 587 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req)); 588 } 589 590 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req) 591 { 592 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 593 594 if (iod->dma_len) { 595 dma_unmap_page(dev->dev, iod->first_dma, iod->dma_len, 596 rq_dma_dir(req)); 597 return; 598 } 599 600 WARN_ON_ONCE(!iod->nents); 601 602 nvme_unmap_sg(dev, req); 603 if (iod->npages == 0) 604 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0], 605 iod->first_dma); 606 else if (iod->use_sgl) 607 nvme_free_sgls(dev, req); 608 else 609 nvme_free_prps(dev, req); 610 mempool_free(iod->sg, dev->iod_mempool); 611 } 612 613 static void nvme_print_sgl(struct scatterlist *sgl, int nents) 614 { 615 int i; 616 struct scatterlist *sg; 617 618 for_each_sg(sgl, sg, nents, i) { 619 dma_addr_t phys = sg_phys(sg); 620 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d " 621 "dma_address:%pad dma_length:%d\n", 622 i, &phys, sg->offset, sg->length, &sg_dma_address(sg), 623 sg_dma_len(sg)); 624 } 625 } 626 627 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev, 628 struct request *req, struct nvme_rw_command *cmnd) 629 { 630 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 631 struct dma_pool *pool; 632 int length = blk_rq_payload_bytes(req); 633 struct scatterlist *sg = iod->sg; 634 int dma_len = sg_dma_len(sg); 635 u64 dma_addr = sg_dma_address(sg); 636 int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1); 637 __le64 *prp_list; 638 void **list = nvme_pci_iod_list(req); 639 dma_addr_t prp_dma; 640 int nprps, i; 641 642 length -= (NVME_CTRL_PAGE_SIZE - offset); 643 if (length <= 0) { 644 iod->first_dma = 0; 645 goto done; 646 } 647 648 dma_len -= (NVME_CTRL_PAGE_SIZE - offset); 649 if (dma_len) { 650 dma_addr += (NVME_CTRL_PAGE_SIZE - offset); 651 } else { 652 sg = sg_next(sg); 653 dma_addr = sg_dma_address(sg); 654 dma_len = sg_dma_len(sg); 655 } 656 657 if (length <= NVME_CTRL_PAGE_SIZE) { 658 iod->first_dma = dma_addr; 659 goto done; 660 } 661 662 nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE); 663 if (nprps <= (256 / 8)) { 664 pool = dev->prp_small_pool; 665 iod->npages = 0; 666 } else { 667 pool = dev->prp_page_pool; 668 iod->npages = 1; 669 } 670 671 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma); 672 if (!prp_list) { 673 iod->first_dma = dma_addr; 674 iod->npages = -1; 675 return BLK_STS_RESOURCE; 676 } 677 list[0] = prp_list; 678 iod->first_dma = prp_dma; 679 i = 0; 680 for (;;) { 681 if (i == NVME_CTRL_PAGE_SIZE >> 3) { 682 __le64 *old_prp_list = prp_list; 683 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma); 684 if (!prp_list) 685 goto free_prps; 686 list[iod->npages++] = prp_list; 687 prp_list[0] = old_prp_list[i - 1]; 688 old_prp_list[i - 1] = cpu_to_le64(prp_dma); 689 i = 1; 690 } 691 prp_list[i++] = cpu_to_le64(dma_addr); 692 dma_len -= NVME_CTRL_PAGE_SIZE; 693 dma_addr += NVME_CTRL_PAGE_SIZE; 694 length -= NVME_CTRL_PAGE_SIZE; 695 if (length <= 0) 696 break; 697 if (dma_len > 0) 698 continue; 699 if (unlikely(dma_len < 0)) 700 goto bad_sgl; 701 sg = sg_next(sg); 702 dma_addr = sg_dma_address(sg); 703 dma_len = sg_dma_len(sg); 704 } 705 done: 706 cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg)); 707 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma); 708 return BLK_STS_OK; 709 free_prps: 710 nvme_free_prps(dev, req); 711 return BLK_STS_RESOURCE; 712 bad_sgl: 713 WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents), 714 "Invalid SGL for payload:%d nents:%d\n", 715 blk_rq_payload_bytes(req), iod->nents); 716 return BLK_STS_IOERR; 717 } 718 719 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge, 720 struct scatterlist *sg) 721 { 722 sge->addr = cpu_to_le64(sg_dma_address(sg)); 723 sge->length = cpu_to_le32(sg_dma_len(sg)); 724 sge->type = NVME_SGL_FMT_DATA_DESC << 4; 725 } 726 727 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge, 728 dma_addr_t dma_addr, int entries) 729 { 730 sge->addr = cpu_to_le64(dma_addr); 731 if (entries < SGES_PER_PAGE) { 732 sge->length = cpu_to_le32(entries * sizeof(*sge)); 733 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4; 734 } else { 735 sge->length = cpu_to_le32(PAGE_SIZE); 736 sge->type = NVME_SGL_FMT_SEG_DESC << 4; 737 } 738 } 739 740 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev, 741 struct request *req, struct nvme_rw_command *cmd, int entries) 742 { 743 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 744 struct dma_pool *pool; 745 struct nvme_sgl_desc *sg_list; 746 struct scatterlist *sg = iod->sg; 747 dma_addr_t sgl_dma; 748 int i = 0; 749 750 /* setting the transfer type as SGL */ 751 cmd->flags = NVME_CMD_SGL_METABUF; 752 753 if (entries == 1) { 754 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg); 755 return BLK_STS_OK; 756 } 757 758 if (entries <= (256 / sizeof(struct nvme_sgl_desc))) { 759 pool = dev->prp_small_pool; 760 iod->npages = 0; 761 } else { 762 pool = dev->prp_page_pool; 763 iod->npages = 1; 764 } 765 766 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma); 767 if (!sg_list) { 768 iod->npages = -1; 769 return BLK_STS_RESOURCE; 770 } 771 772 nvme_pci_iod_list(req)[0] = sg_list; 773 iod->first_dma = sgl_dma; 774 775 nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries); 776 777 do { 778 if (i == SGES_PER_PAGE) { 779 struct nvme_sgl_desc *old_sg_desc = sg_list; 780 struct nvme_sgl_desc *link = &old_sg_desc[i - 1]; 781 782 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma); 783 if (!sg_list) 784 goto free_sgls; 785 786 i = 0; 787 nvme_pci_iod_list(req)[iod->npages++] = sg_list; 788 sg_list[i++] = *link; 789 nvme_pci_sgl_set_seg(link, sgl_dma, entries); 790 } 791 792 nvme_pci_sgl_set_data(&sg_list[i++], sg); 793 sg = sg_next(sg); 794 } while (--entries > 0); 795 796 return BLK_STS_OK; 797 free_sgls: 798 nvme_free_sgls(dev, req); 799 return BLK_STS_RESOURCE; 800 } 801 802 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev, 803 struct request *req, struct nvme_rw_command *cmnd, 804 struct bio_vec *bv) 805 { 806 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 807 unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1); 808 unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset; 809 810 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0); 811 if (dma_mapping_error(dev->dev, iod->first_dma)) 812 return BLK_STS_RESOURCE; 813 iod->dma_len = bv->bv_len; 814 815 cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma); 816 if (bv->bv_len > first_prp_len) 817 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len); 818 return BLK_STS_OK; 819 } 820 821 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev, 822 struct request *req, struct nvme_rw_command *cmnd, 823 struct bio_vec *bv) 824 { 825 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 826 827 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0); 828 if (dma_mapping_error(dev->dev, iod->first_dma)) 829 return BLK_STS_RESOURCE; 830 iod->dma_len = bv->bv_len; 831 832 cmnd->flags = NVME_CMD_SGL_METABUF; 833 cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma); 834 cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len); 835 cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4; 836 return BLK_STS_OK; 837 } 838 839 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req, 840 struct nvme_command *cmnd) 841 { 842 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 843 blk_status_t ret = BLK_STS_RESOURCE; 844 int nr_mapped; 845 846 if (blk_rq_nr_phys_segments(req) == 1) { 847 struct bio_vec bv = req_bvec(req); 848 849 if (!is_pci_p2pdma_page(bv.bv_page)) { 850 if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2) 851 return nvme_setup_prp_simple(dev, req, 852 &cmnd->rw, &bv); 853 854 if (iod->nvmeq->qid && sgl_threshold && 855 nvme_ctrl_sgl_supported(&dev->ctrl)) 856 return nvme_setup_sgl_simple(dev, req, 857 &cmnd->rw, &bv); 858 } 859 } 860 861 iod->dma_len = 0; 862 iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC); 863 if (!iod->sg) 864 return BLK_STS_RESOURCE; 865 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req)); 866 iod->nents = blk_rq_map_sg(req->q, req, iod->sg); 867 if (!iod->nents) 868 goto out_free_sg; 869 870 if (is_pci_p2pdma_page(sg_page(iod->sg))) 871 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg, 872 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN); 873 else 874 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, 875 rq_dma_dir(req), DMA_ATTR_NO_WARN); 876 if (!nr_mapped) 877 goto out_free_sg; 878 879 iod->use_sgl = nvme_pci_use_sgls(dev, req); 880 if (iod->use_sgl) 881 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped); 882 else 883 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw); 884 if (ret != BLK_STS_OK) 885 goto out_unmap_sg; 886 return BLK_STS_OK; 887 888 out_unmap_sg: 889 nvme_unmap_sg(dev, req); 890 out_free_sg: 891 mempool_free(iod->sg, dev->iod_mempool); 892 return ret; 893 } 894 895 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req, 896 struct nvme_command *cmnd) 897 { 898 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 899 900 iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req), 901 rq_dma_dir(req), 0); 902 if (dma_mapping_error(dev->dev, iod->meta_dma)) 903 return BLK_STS_IOERR; 904 cmnd->rw.metadata = cpu_to_le64(iod->meta_dma); 905 return BLK_STS_OK; 906 } 907 908 static blk_status_t nvme_prep_rq(struct nvme_dev *dev, struct request *req) 909 { 910 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 911 blk_status_t ret; 912 913 iod->aborted = 0; 914 iod->npages = -1; 915 iod->nents = 0; 916 917 ret = nvme_setup_cmd(req->q->queuedata, req); 918 if (ret) 919 return ret; 920 921 if (blk_rq_nr_phys_segments(req)) { 922 ret = nvme_map_data(dev, req, &iod->cmd); 923 if (ret) 924 goto out_free_cmd; 925 } 926 927 if (blk_integrity_rq(req)) { 928 ret = nvme_map_metadata(dev, req, &iod->cmd); 929 if (ret) 930 goto out_unmap_data; 931 } 932 933 blk_mq_start_request(req); 934 return BLK_STS_OK; 935 out_unmap_data: 936 nvme_unmap_data(dev, req); 937 out_free_cmd: 938 nvme_cleanup_cmd(req); 939 return ret; 940 } 941 942 /* 943 * NOTE: ns is NULL when called on the admin queue. 944 */ 945 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx, 946 const struct blk_mq_queue_data *bd) 947 { 948 struct nvme_queue *nvmeq = hctx->driver_data; 949 struct nvme_dev *dev = nvmeq->dev; 950 struct request *req = bd->rq; 951 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 952 blk_status_t ret; 953 954 /* 955 * We should not need to do this, but we're still using this to 956 * ensure we can drain requests on a dying queue. 957 */ 958 if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags))) 959 return BLK_STS_IOERR; 960 961 if (unlikely(!nvme_check_ready(&dev->ctrl, req, true))) 962 return nvme_fail_nonready_command(&dev->ctrl, req); 963 964 ret = nvme_prep_rq(dev, req); 965 if (unlikely(ret)) 966 return ret; 967 spin_lock(&nvmeq->sq_lock); 968 nvme_sq_copy_cmd(nvmeq, &iod->cmd); 969 nvme_write_sq_db(nvmeq, bd->last); 970 spin_unlock(&nvmeq->sq_lock); 971 return BLK_STS_OK; 972 } 973 974 static void nvme_submit_cmds(struct nvme_queue *nvmeq, struct request **rqlist) 975 { 976 spin_lock(&nvmeq->sq_lock); 977 while (!rq_list_empty(*rqlist)) { 978 struct request *req = rq_list_pop(rqlist); 979 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 980 981 nvme_sq_copy_cmd(nvmeq, &iod->cmd); 982 } 983 nvme_write_sq_db(nvmeq, true); 984 spin_unlock(&nvmeq->sq_lock); 985 } 986 987 static bool nvme_prep_rq_batch(struct nvme_queue *nvmeq, struct request *req) 988 { 989 /* 990 * We should not need to do this, but we're still using this to 991 * ensure we can drain requests on a dying queue. 992 */ 993 if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags))) 994 return false; 995 if (unlikely(!nvme_check_ready(&nvmeq->dev->ctrl, req, true))) 996 return false; 997 998 req->mq_hctx->tags->rqs[req->tag] = req; 999 return nvme_prep_rq(nvmeq->dev, req) == BLK_STS_OK; 1000 } 1001 1002 static void nvme_queue_rqs(struct request **rqlist) 1003 { 1004 struct request *req, *next, *prev = NULL; 1005 struct request *requeue_list = NULL; 1006 1007 rq_list_for_each_safe(rqlist, req, next) { 1008 struct nvme_queue *nvmeq = req->mq_hctx->driver_data; 1009 1010 if (!nvme_prep_rq_batch(nvmeq, req)) { 1011 /* detach 'req' and add to remainder list */ 1012 rq_list_move(rqlist, &requeue_list, req, prev); 1013 1014 req = prev; 1015 if (!req) 1016 continue; 1017 } 1018 1019 if (!next || req->mq_hctx != next->mq_hctx) { 1020 /* detach rest of list, and submit */ 1021 req->rq_next = NULL; 1022 nvme_submit_cmds(nvmeq, rqlist); 1023 *rqlist = next; 1024 prev = NULL; 1025 } else 1026 prev = req; 1027 } 1028 1029 *rqlist = requeue_list; 1030 } 1031 1032 static __always_inline void nvme_pci_unmap_rq(struct request *req) 1033 { 1034 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 1035 struct nvme_dev *dev = iod->nvmeq->dev; 1036 1037 if (blk_integrity_rq(req)) 1038 dma_unmap_page(dev->dev, iod->meta_dma, 1039 rq_integrity_vec(req)->bv_len, rq_data_dir(req)); 1040 if (blk_rq_nr_phys_segments(req)) 1041 nvme_unmap_data(dev, req); 1042 } 1043 1044 static void nvme_pci_complete_rq(struct request *req) 1045 { 1046 nvme_pci_unmap_rq(req); 1047 nvme_complete_rq(req); 1048 } 1049 1050 static void nvme_pci_complete_batch(struct io_comp_batch *iob) 1051 { 1052 nvme_complete_batch(iob, nvme_pci_unmap_rq); 1053 } 1054 1055 /* We read the CQE phase first to check if the rest of the entry is valid */ 1056 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq) 1057 { 1058 struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head]; 1059 1060 return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase; 1061 } 1062 1063 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq) 1064 { 1065 u16 head = nvmeq->cq_head; 1066 1067 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db, 1068 nvmeq->dbbuf_cq_ei)) 1069 writel(head, nvmeq->q_db + nvmeq->dev->db_stride); 1070 } 1071 1072 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq) 1073 { 1074 if (!nvmeq->qid) 1075 return nvmeq->dev->admin_tagset.tags[0]; 1076 return nvmeq->dev->tagset.tags[nvmeq->qid - 1]; 1077 } 1078 1079 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, 1080 struct io_comp_batch *iob, u16 idx) 1081 { 1082 struct nvme_completion *cqe = &nvmeq->cqes[idx]; 1083 __u16 command_id = READ_ONCE(cqe->command_id); 1084 struct request *req; 1085 1086 /* 1087 * AEN requests are special as they don't time out and can 1088 * survive any kind of queue freeze and often don't respond to 1089 * aborts. We don't even bother to allocate a struct request 1090 * for them but rather special case them here. 1091 */ 1092 if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) { 1093 nvme_complete_async_event(&nvmeq->dev->ctrl, 1094 cqe->status, &cqe->result); 1095 return; 1096 } 1097 1098 req = nvme_find_rq(nvme_queue_tagset(nvmeq), command_id); 1099 if (unlikely(!req)) { 1100 dev_warn(nvmeq->dev->ctrl.device, 1101 "invalid id %d completed on queue %d\n", 1102 command_id, le16_to_cpu(cqe->sq_id)); 1103 return; 1104 } 1105 1106 trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail); 1107 if (!nvme_try_complete_req(req, cqe->status, cqe->result) && 1108 !blk_mq_add_to_batch(req, iob, nvme_req(req)->status, 1109 nvme_pci_complete_batch)) 1110 nvme_pci_complete_rq(req); 1111 } 1112 1113 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq) 1114 { 1115 u32 tmp = nvmeq->cq_head + 1; 1116 1117 if (tmp == nvmeq->q_depth) { 1118 nvmeq->cq_head = 0; 1119 nvmeq->cq_phase ^= 1; 1120 } else { 1121 nvmeq->cq_head = tmp; 1122 } 1123 } 1124 1125 static inline int nvme_poll_cq(struct nvme_queue *nvmeq, 1126 struct io_comp_batch *iob) 1127 { 1128 int found = 0; 1129 1130 while (nvme_cqe_pending(nvmeq)) { 1131 found++; 1132 /* 1133 * load-load control dependency between phase and the rest of 1134 * the cqe requires a full read memory barrier 1135 */ 1136 dma_rmb(); 1137 nvme_handle_cqe(nvmeq, iob, nvmeq->cq_head); 1138 nvme_update_cq_head(nvmeq); 1139 } 1140 1141 if (found) 1142 nvme_ring_cq_doorbell(nvmeq); 1143 return found; 1144 } 1145 1146 static irqreturn_t nvme_irq(int irq, void *data) 1147 { 1148 struct nvme_queue *nvmeq = data; 1149 DEFINE_IO_COMP_BATCH(iob); 1150 1151 if (nvme_poll_cq(nvmeq, &iob)) { 1152 if (!rq_list_empty(iob.req_list)) 1153 nvme_pci_complete_batch(&iob); 1154 return IRQ_HANDLED; 1155 } 1156 return IRQ_NONE; 1157 } 1158 1159 static irqreturn_t nvme_irq_check(int irq, void *data) 1160 { 1161 struct nvme_queue *nvmeq = data; 1162 1163 if (nvme_cqe_pending(nvmeq)) 1164 return IRQ_WAKE_THREAD; 1165 return IRQ_NONE; 1166 } 1167 1168 /* 1169 * Poll for completions for any interrupt driven queue 1170 * Can be called from any context. 1171 */ 1172 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq) 1173 { 1174 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev); 1175 1176 WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags)); 1177 1178 disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)); 1179 nvme_poll_cq(nvmeq, NULL); 1180 enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector)); 1181 } 1182 1183 static int nvme_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) 1184 { 1185 struct nvme_queue *nvmeq = hctx->driver_data; 1186 bool found; 1187 1188 if (!nvme_cqe_pending(nvmeq)) 1189 return 0; 1190 1191 spin_lock(&nvmeq->cq_poll_lock); 1192 found = nvme_poll_cq(nvmeq, iob); 1193 spin_unlock(&nvmeq->cq_poll_lock); 1194 1195 return found; 1196 } 1197 1198 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl) 1199 { 1200 struct nvme_dev *dev = to_nvme_dev(ctrl); 1201 struct nvme_queue *nvmeq = &dev->queues[0]; 1202 struct nvme_command c = { }; 1203 1204 c.common.opcode = nvme_admin_async_event; 1205 c.common.command_id = NVME_AQ_BLK_MQ_DEPTH; 1206 1207 spin_lock(&nvmeq->sq_lock); 1208 nvme_sq_copy_cmd(nvmeq, &c); 1209 nvme_write_sq_db(nvmeq, true); 1210 spin_unlock(&nvmeq->sq_lock); 1211 } 1212 1213 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id) 1214 { 1215 struct nvme_command c = { }; 1216 1217 c.delete_queue.opcode = opcode; 1218 c.delete_queue.qid = cpu_to_le16(id); 1219 1220 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 1221 } 1222 1223 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid, 1224 struct nvme_queue *nvmeq, s16 vector) 1225 { 1226 struct nvme_command c = { }; 1227 int flags = NVME_QUEUE_PHYS_CONTIG; 1228 1229 if (!test_bit(NVMEQ_POLLED, &nvmeq->flags)) 1230 flags |= NVME_CQ_IRQ_ENABLED; 1231 1232 /* 1233 * Note: we (ab)use the fact that the prp fields survive if no data 1234 * is attached to the request. 1235 */ 1236 c.create_cq.opcode = nvme_admin_create_cq; 1237 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr); 1238 c.create_cq.cqid = cpu_to_le16(qid); 1239 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1); 1240 c.create_cq.cq_flags = cpu_to_le16(flags); 1241 c.create_cq.irq_vector = cpu_to_le16(vector); 1242 1243 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 1244 } 1245 1246 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid, 1247 struct nvme_queue *nvmeq) 1248 { 1249 struct nvme_ctrl *ctrl = &dev->ctrl; 1250 struct nvme_command c = { }; 1251 int flags = NVME_QUEUE_PHYS_CONTIG; 1252 1253 /* 1254 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't 1255 * set. Since URGENT priority is zeroes, it makes all queues 1256 * URGENT. 1257 */ 1258 if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ) 1259 flags |= NVME_SQ_PRIO_MEDIUM; 1260 1261 /* 1262 * Note: we (ab)use the fact that the prp fields survive if no data 1263 * is attached to the request. 1264 */ 1265 c.create_sq.opcode = nvme_admin_create_sq; 1266 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr); 1267 c.create_sq.sqid = cpu_to_le16(qid); 1268 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1); 1269 c.create_sq.sq_flags = cpu_to_le16(flags); 1270 c.create_sq.cqid = cpu_to_le16(qid); 1271 1272 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 1273 } 1274 1275 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid) 1276 { 1277 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid); 1278 } 1279 1280 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid) 1281 { 1282 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid); 1283 } 1284 1285 static void abort_endio(struct request *req, blk_status_t error) 1286 { 1287 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 1288 struct nvme_queue *nvmeq = iod->nvmeq; 1289 1290 dev_warn(nvmeq->dev->ctrl.device, 1291 "Abort status: 0x%x", nvme_req(req)->status); 1292 atomic_inc(&nvmeq->dev->ctrl.abort_limit); 1293 blk_mq_free_request(req); 1294 } 1295 1296 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts) 1297 { 1298 /* If true, indicates loss of adapter communication, possibly by a 1299 * NVMe Subsystem reset. 1300 */ 1301 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO); 1302 1303 /* If there is a reset/reinit ongoing, we shouldn't reset again. */ 1304 switch (dev->ctrl.state) { 1305 case NVME_CTRL_RESETTING: 1306 case NVME_CTRL_CONNECTING: 1307 return false; 1308 default: 1309 break; 1310 } 1311 1312 /* We shouldn't reset unless the controller is on fatal error state 1313 * _or_ if we lost the communication with it. 1314 */ 1315 if (!(csts & NVME_CSTS_CFS) && !nssro) 1316 return false; 1317 1318 return true; 1319 } 1320 1321 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts) 1322 { 1323 /* Read a config register to help see what died. */ 1324 u16 pci_status; 1325 int result; 1326 1327 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS, 1328 &pci_status); 1329 if (result == PCIBIOS_SUCCESSFUL) 1330 dev_warn(dev->ctrl.device, 1331 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n", 1332 csts, pci_status); 1333 else 1334 dev_warn(dev->ctrl.device, 1335 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n", 1336 csts, result); 1337 } 1338 1339 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved) 1340 { 1341 struct nvme_iod *iod = blk_mq_rq_to_pdu(req); 1342 struct nvme_queue *nvmeq = iod->nvmeq; 1343 struct nvme_dev *dev = nvmeq->dev; 1344 struct request *abort_req; 1345 struct nvme_command cmd = { }; 1346 u32 csts = readl(dev->bar + NVME_REG_CSTS); 1347 1348 /* If PCI error recovery process is happening, we cannot reset or 1349 * the recovery mechanism will surely fail. 1350 */ 1351 mb(); 1352 if (pci_channel_offline(to_pci_dev(dev->dev))) 1353 return BLK_EH_RESET_TIMER; 1354 1355 /* 1356 * Reset immediately if the controller is failed 1357 */ 1358 if (nvme_should_reset(dev, csts)) { 1359 nvme_warn_reset(dev, csts); 1360 nvme_dev_disable(dev, false); 1361 nvme_reset_ctrl(&dev->ctrl); 1362 return BLK_EH_DONE; 1363 } 1364 1365 /* 1366 * Did we miss an interrupt? 1367 */ 1368 if (test_bit(NVMEQ_POLLED, &nvmeq->flags)) 1369 nvme_poll(req->mq_hctx, NULL); 1370 else 1371 nvme_poll_irqdisable(nvmeq); 1372 1373 if (blk_mq_request_completed(req)) { 1374 dev_warn(dev->ctrl.device, 1375 "I/O %d QID %d timeout, completion polled\n", 1376 req->tag, nvmeq->qid); 1377 return BLK_EH_DONE; 1378 } 1379 1380 /* 1381 * Shutdown immediately if controller times out while starting. The 1382 * reset work will see the pci device disabled when it gets the forced 1383 * cancellation error. All outstanding requests are completed on 1384 * shutdown, so we return BLK_EH_DONE. 1385 */ 1386 switch (dev->ctrl.state) { 1387 case NVME_CTRL_CONNECTING: 1388 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING); 1389 fallthrough; 1390 case NVME_CTRL_DELETING: 1391 dev_warn_ratelimited(dev->ctrl.device, 1392 "I/O %d QID %d timeout, disable controller\n", 1393 req->tag, nvmeq->qid); 1394 nvme_req(req)->flags |= NVME_REQ_CANCELLED; 1395 nvme_dev_disable(dev, true); 1396 return BLK_EH_DONE; 1397 case NVME_CTRL_RESETTING: 1398 return BLK_EH_RESET_TIMER; 1399 default: 1400 break; 1401 } 1402 1403 /* 1404 * Shutdown the controller immediately and schedule a reset if the 1405 * command was already aborted once before and still hasn't been 1406 * returned to the driver, or if this is the admin queue. 1407 */ 1408 if (!nvmeq->qid || iod->aborted) { 1409 dev_warn(dev->ctrl.device, 1410 "I/O %d QID %d timeout, reset controller\n", 1411 req->tag, nvmeq->qid); 1412 nvme_req(req)->flags |= NVME_REQ_CANCELLED; 1413 nvme_dev_disable(dev, false); 1414 nvme_reset_ctrl(&dev->ctrl); 1415 1416 return BLK_EH_DONE; 1417 } 1418 1419 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) { 1420 atomic_inc(&dev->ctrl.abort_limit); 1421 return BLK_EH_RESET_TIMER; 1422 } 1423 iod->aborted = 1; 1424 1425 cmd.abort.opcode = nvme_admin_abort_cmd; 1426 cmd.abort.cid = nvme_cid(req); 1427 cmd.abort.sqid = cpu_to_le16(nvmeq->qid); 1428 1429 dev_warn(nvmeq->dev->ctrl.device, 1430 "I/O %d QID %d timeout, aborting\n", 1431 req->tag, nvmeq->qid); 1432 1433 abort_req = blk_mq_alloc_request(dev->ctrl.admin_q, nvme_req_op(&cmd), 1434 BLK_MQ_REQ_NOWAIT); 1435 if (IS_ERR(abort_req)) { 1436 atomic_inc(&dev->ctrl.abort_limit); 1437 return BLK_EH_RESET_TIMER; 1438 } 1439 nvme_init_request(abort_req, &cmd); 1440 1441 abort_req->end_io = abort_endio; 1442 abort_req->end_io_data = NULL; 1443 abort_req->rq_flags |= RQF_QUIET; 1444 blk_execute_rq_nowait(abort_req, false); 1445 1446 /* 1447 * The aborted req will be completed on receiving the abort req. 1448 * We enable the timer again. If hit twice, it'll cause a device reset, 1449 * as the device then is in a faulty state. 1450 */ 1451 return BLK_EH_RESET_TIMER; 1452 } 1453 1454 static void nvme_free_queue(struct nvme_queue *nvmeq) 1455 { 1456 dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq), 1457 (void *)nvmeq->cqes, nvmeq->cq_dma_addr); 1458 if (!nvmeq->sq_cmds) 1459 return; 1460 1461 if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) { 1462 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev), 1463 nvmeq->sq_cmds, SQ_SIZE(nvmeq)); 1464 } else { 1465 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq), 1466 nvmeq->sq_cmds, nvmeq->sq_dma_addr); 1467 } 1468 } 1469 1470 static void nvme_free_queues(struct nvme_dev *dev, int lowest) 1471 { 1472 int i; 1473 1474 for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) { 1475 dev->ctrl.queue_count--; 1476 nvme_free_queue(&dev->queues[i]); 1477 } 1478 } 1479 1480 /** 1481 * nvme_suspend_queue - put queue into suspended state 1482 * @nvmeq: queue to suspend 1483 */ 1484 static int nvme_suspend_queue(struct nvme_queue *nvmeq) 1485 { 1486 if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags)) 1487 return 1; 1488 1489 /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */ 1490 mb(); 1491 1492 nvmeq->dev->online_queues--; 1493 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q) 1494 nvme_stop_admin_queue(&nvmeq->dev->ctrl); 1495 if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags)) 1496 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq); 1497 return 0; 1498 } 1499 1500 static void nvme_suspend_io_queues(struct nvme_dev *dev) 1501 { 1502 int i; 1503 1504 for (i = dev->ctrl.queue_count - 1; i > 0; i--) 1505 nvme_suspend_queue(&dev->queues[i]); 1506 } 1507 1508 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown) 1509 { 1510 struct nvme_queue *nvmeq = &dev->queues[0]; 1511 1512 if (shutdown) 1513 nvme_shutdown_ctrl(&dev->ctrl); 1514 else 1515 nvme_disable_ctrl(&dev->ctrl); 1516 1517 nvme_poll_irqdisable(nvmeq); 1518 } 1519 1520 /* 1521 * Called only on a device that has been disabled and after all other threads 1522 * that can check this device's completion queues have synced, except 1523 * nvme_poll(). This is the last chance for the driver to see a natural 1524 * completion before nvme_cancel_request() terminates all incomplete requests. 1525 */ 1526 static void nvme_reap_pending_cqes(struct nvme_dev *dev) 1527 { 1528 int i; 1529 1530 for (i = dev->ctrl.queue_count - 1; i > 0; i--) { 1531 spin_lock(&dev->queues[i].cq_poll_lock); 1532 nvme_poll_cq(&dev->queues[i], NULL); 1533 spin_unlock(&dev->queues[i].cq_poll_lock); 1534 } 1535 } 1536 1537 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues, 1538 int entry_size) 1539 { 1540 int q_depth = dev->q_depth; 1541 unsigned q_size_aligned = roundup(q_depth * entry_size, 1542 NVME_CTRL_PAGE_SIZE); 1543 1544 if (q_size_aligned * nr_io_queues > dev->cmb_size) { 1545 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues); 1546 1547 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE); 1548 q_depth = div_u64(mem_per_q, entry_size); 1549 1550 /* 1551 * Ensure the reduced q_depth is above some threshold where it 1552 * would be better to map queues in system memory with the 1553 * original depth 1554 */ 1555 if (q_depth < 64) 1556 return -ENOMEM; 1557 } 1558 1559 return q_depth; 1560 } 1561 1562 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq, 1563 int qid) 1564 { 1565 struct pci_dev *pdev = to_pci_dev(dev->dev); 1566 1567 if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) { 1568 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq)); 1569 if (nvmeq->sq_cmds) { 1570 nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev, 1571 nvmeq->sq_cmds); 1572 if (nvmeq->sq_dma_addr) { 1573 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags); 1574 return 0; 1575 } 1576 1577 pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq)); 1578 } 1579 } 1580 1581 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq), 1582 &nvmeq->sq_dma_addr, GFP_KERNEL); 1583 if (!nvmeq->sq_cmds) 1584 return -ENOMEM; 1585 return 0; 1586 } 1587 1588 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth) 1589 { 1590 struct nvme_queue *nvmeq = &dev->queues[qid]; 1591 1592 if (dev->ctrl.queue_count > qid) 1593 return 0; 1594 1595 nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES; 1596 nvmeq->q_depth = depth; 1597 nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq), 1598 &nvmeq->cq_dma_addr, GFP_KERNEL); 1599 if (!nvmeq->cqes) 1600 goto free_nvmeq; 1601 1602 if (nvme_alloc_sq_cmds(dev, nvmeq, qid)) 1603 goto free_cqdma; 1604 1605 nvmeq->dev = dev; 1606 spin_lock_init(&nvmeq->sq_lock); 1607 spin_lock_init(&nvmeq->cq_poll_lock); 1608 nvmeq->cq_head = 0; 1609 nvmeq->cq_phase = 1; 1610 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride]; 1611 nvmeq->qid = qid; 1612 dev->ctrl.queue_count++; 1613 1614 return 0; 1615 1616 free_cqdma: 1617 dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes, 1618 nvmeq->cq_dma_addr); 1619 free_nvmeq: 1620 return -ENOMEM; 1621 } 1622 1623 static int queue_request_irq(struct nvme_queue *nvmeq) 1624 { 1625 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev); 1626 int nr = nvmeq->dev->ctrl.instance; 1627 1628 if (use_threaded_interrupts) { 1629 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check, 1630 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid); 1631 } else { 1632 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq, 1633 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid); 1634 } 1635 } 1636 1637 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid) 1638 { 1639 struct nvme_dev *dev = nvmeq->dev; 1640 1641 nvmeq->sq_tail = 0; 1642 nvmeq->last_sq_tail = 0; 1643 nvmeq->cq_head = 0; 1644 nvmeq->cq_phase = 1; 1645 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride]; 1646 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq)); 1647 nvme_dbbuf_init(dev, nvmeq, qid); 1648 dev->online_queues++; 1649 wmb(); /* ensure the first interrupt sees the initialization */ 1650 } 1651 1652 /* 1653 * Try getting shutdown_lock while setting up IO queues. 1654 */ 1655 static int nvme_setup_io_queues_trylock(struct nvme_dev *dev) 1656 { 1657 /* 1658 * Give up if the lock is being held by nvme_dev_disable. 1659 */ 1660 if (!mutex_trylock(&dev->shutdown_lock)) 1661 return -ENODEV; 1662 1663 /* 1664 * Controller is in wrong state, fail early. 1665 */ 1666 if (dev->ctrl.state != NVME_CTRL_CONNECTING) { 1667 mutex_unlock(&dev->shutdown_lock); 1668 return -ENODEV; 1669 } 1670 1671 return 0; 1672 } 1673 1674 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled) 1675 { 1676 struct nvme_dev *dev = nvmeq->dev; 1677 int result; 1678 u16 vector = 0; 1679 1680 clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags); 1681 1682 /* 1683 * A queue's vector matches the queue identifier unless the controller 1684 * has only one vector available. 1685 */ 1686 if (!polled) 1687 vector = dev->num_vecs == 1 ? 0 : qid; 1688 else 1689 set_bit(NVMEQ_POLLED, &nvmeq->flags); 1690 1691 result = adapter_alloc_cq(dev, qid, nvmeq, vector); 1692 if (result) 1693 return result; 1694 1695 result = adapter_alloc_sq(dev, qid, nvmeq); 1696 if (result < 0) 1697 return result; 1698 if (result) 1699 goto release_cq; 1700 1701 nvmeq->cq_vector = vector; 1702 1703 result = nvme_setup_io_queues_trylock(dev); 1704 if (result) 1705 return result; 1706 nvme_init_queue(nvmeq, qid); 1707 if (!polled) { 1708 result = queue_request_irq(nvmeq); 1709 if (result < 0) 1710 goto release_sq; 1711 } 1712 1713 set_bit(NVMEQ_ENABLED, &nvmeq->flags); 1714 mutex_unlock(&dev->shutdown_lock); 1715 return result; 1716 1717 release_sq: 1718 dev->online_queues--; 1719 mutex_unlock(&dev->shutdown_lock); 1720 adapter_delete_sq(dev, qid); 1721 release_cq: 1722 adapter_delete_cq(dev, qid); 1723 return result; 1724 } 1725 1726 static const struct blk_mq_ops nvme_mq_admin_ops = { 1727 .queue_rq = nvme_queue_rq, 1728 .complete = nvme_pci_complete_rq, 1729 .init_hctx = nvme_admin_init_hctx, 1730 .init_request = nvme_pci_init_request, 1731 .timeout = nvme_timeout, 1732 }; 1733 1734 static const struct blk_mq_ops nvme_mq_ops = { 1735 .queue_rq = nvme_queue_rq, 1736 .queue_rqs = nvme_queue_rqs, 1737 .complete = nvme_pci_complete_rq, 1738 .commit_rqs = nvme_commit_rqs, 1739 .init_hctx = nvme_init_hctx, 1740 .init_request = nvme_pci_init_request, 1741 .map_queues = nvme_pci_map_queues, 1742 .timeout = nvme_timeout, 1743 .poll = nvme_poll, 1744 }; 1745 1746 static void nvme_dev_remove_admin(struct nvme_dev *dev) 1747 { 1748 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) { 1749 /* 1750 * If the controller was reset during removal, it's possible 1751 * user requests may be waiting on a stopped queue. Start the 1752 * queue to flush these to completion. 1753 */ 1754 nvme_start_admin_queue(&dev->ctrl); 1755 blk_cleanup_queue(dev->ctrl.admin_q); 1756 blk_mq_free_tag_set(&dev->admin_tagset); 1757 } 1758 } 1759 1760 static int nvme_alloc_admin_tags(struct nvme_dev *dev) 1761 { 1762 if (!dev->ctrl.admin_q) { 1763 dev->admin_tagset.ops = &nvme_mq_admin_ops; 1764 dev->admin_tagset.nr_hw_queues = 1; 1765 1766 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH; 1767 dev->admin_tagset.timeout = NVME_ADMIN_TIMEOUT; 1768 dev->admin_tagset.numa_node = dev->ctrl.numa_node; 1769 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod); 1770 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED; 1771 dev->admin_tagset.driver_data = dev; 1772 1773 if (blk_mq_alloc_tag_set(&dev->admin_tagset)) 1774 return -ENOMEM; 1775 dev->ctrl.admin_tagset = &dev->admin_tagset; 1776 1777 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset); 1778 if (IS_ERR(dev->ctrl.admin_q)) { 1779 blk_mq_free_tag_set(&dev->admin_tagset); 1780 dev->ctrl.admin_q = NULL; 1781 return -ENOMEM; 1782 } 1783 if (!blk_get_queue(dev->ctrl.admin_q)) { 1784 nvme_dev_remove_admin(dev); 1785 dev->ctrl.admin_q = NULL; 1786 return -ENODEV; 1787 } 1788 } else 1789 nvme_start_admin_queue(&dev->ctrl); 1790 1791 return 0; 1792 } 1793 1794 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues) 1795 { 1796 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride); 1797 } 1798 1799 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size) 1800 { 1801 struct pci_dev *pdev = to_pci_dev(dev->dev); 1802 1803 if (size <= dev->bar_mapped_size) 1804 return 0; 1805 if (size > pci_resource_len(pdev, 0)) 1806 return -ENOMEM; 1807 if (dev->bar) 1808 iounmap(dev->bar); 1809 dev->bar = ioremap(pci_resource_start(pdev, 0), size); 1810 if (!dev->bar) { 1811 dev->bar_mapped_size = 0; 1812 return -ENOMEM; 1813 } 1814 dev->bar_mapped_size = size; 1815 dev->dbs = dev->bar + NVME_REG_DBS; 1816 1817 return 0; 1818 } 1819 1820 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev) 1821 { 1822 int result; 1823 u32 aqa; 1824 struct nvme_queue *nvmeq; 1825 1826 result = nvme_remap_bar(dev, db_bar_size(dev, 0)); 1827 if (result < 0) 1828 return result; 1829 1830 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ? 1831 NVME_CAP_NSSRC(dev->ctrl.cap) : 0; 1832 1833 if (dev->subsystem && 1834 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO)) 1835 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS); 1836 1837 result = nvme_disable_ctrl(&dev->ctrl); 1838 if (result < 0) 1839 return result; 1840 1841 result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH); 1842 if (result) 1843 return result; 1844 1845 dev->ctrl.numa_node = dev_to_node(dev->dev); 1846 1847 nvmeq = &dev->queues[0]; 1848 aqa = nvmeq->q_depth - 1; 1849 aqa |= aqa << 16; 1850 1851 writel(aqa, dev->bar + NVME_REG_AQA); 1852 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ); 1853 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ); 1854 1855 result = nvme_enable_ctrl(&dev->ctrl); 1856 if (result) 1857 return result; 1858 1859 nvmeq->cq_vector = 0; 1860 nvme_init_queue(nvmeq, 0); 1861 result = queue_request_irq(nvmeq); 1862 if (result) { 1863 dev->online_queues--; 1864 return result; 1865 } 1866 1867 set_bit(NVMEQ_ENABLED, &nvmeq->flags); 1868 return result; 1869 } 1870 1871 static int nvme_create_io_queues(struct nvme_dev *dev) 1872 { 1873 unsigned i, max, rw_queues; 1874 int ret = 0; 1875 1876 for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) { 1877 if (nvme_alloc_queue(dev, i, dev->q_depth)) { 1878 ret = -ENOMEM; 1879 break; 1880 } 1881 } 1882 1883 max = min(dev->max_qid, dev->ctrl.queue_count - 1); 1884 if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) { 1885 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] + 1886 dev->io_queues[HCTX_TYPE_READ]; 1887 } else { 1888 rw_queues = max; 1889 } 1890 1891 for (i = dev->online_queues; i <= max; i++) { 1892 bool polled = i > rw_queues; 1893 1894 ret = nvme_create_queue(&dev->queues[i], i, polled); 1895 if (ret) 1896 break; 1897 } 1898 1899 /* 1900 * Ignore failing Create SQ/CQ commands, we can continue with less 1901 * than the desired amount of queues, and even a controller without 1902 * I/O queues can still be used to issue admin commands. This might 1903 * be useful to upgrade a buggy firmware for example. 1904 */ 1905 return ret >= 0 ? 0 : ret; 1906 } 1907 1908 static u64 nvme_cmb_size_unit(struct nvme_dev *dev) 1909 { 1910 u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK; 1911 1912 return 1ULL << (12 + 4 * szu); 1913 } 1914 1915 static u32 nvme_cmb_size(struct nvme_dev *dev) 1916 { 1917 return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK; 1918 } 1919 1920 static void nvme_map_cmb(struct nvme_dev *dev) 1921 { 1922 u64 size, offset; 1923 resource_size_t bar_size; 1924 struct pci_dev *pdev = to_pci_dev(dev->dev); 1925 int bar; 1926 1927 if (dev->cmb_size) 1928 return; 1929 1930 if (NVME_CAP_CMBS(dev->ctrl.cap)) 1931 writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC); 1932 1933 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ); 1934 if (!dev->cmbsz) 1935 return; 1936 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC); 1937 1938 size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev); 1939 offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc); 1940 bar = NVME_CMB_BIR(dev->cmbloc); 1941 bar_size = pci_resource_len(pdev, bar); 1942 1943 if (offset > bar_size) 1944 return; 1945 1946 /* 1947 * Tell the controller about the host side address mapping the CMB, 1948 * and enable CMB decoding for the NVMe 1.4+ scheme: 1949 */ 1950 if (NVME_CAP_CMBS(dev->ctrl.cap)) { 1951 hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE | 1952 (pci_bus_address(pdev, bar) + offset), 1953 dev->bar + NVME_REG_CMBMSC); 1954 } 1955 1956 /* 1957 * Controllers may support a CMB size larger than their BAR, 1958 * for example, due to being behind a bridge. Reduce the CMB to 1959 * the reported size of the BAR 1960 */ 1961 if (size > bar_size - offset) 1962 size = bar_size - offset; 1963 1964 if (pci_p2pdma_add_resource(pdev, bar, size, offset)) { 1965 dev_warn(dev->ctrl.device, 1966 "failed to register the CMB\n"); 1967 return; 1968 } 1969 1970 dev->cmb_size = size; 1971 dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS); 1972 1973 if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) == 1974 (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) 1975 pci_p2pmem_publish(pdev, true); 1976 } 1977 1978 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits) 1979 { 1980 u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT; 1981 u64 dma_addr = dev->host_mem_descs_dma; 1982 struct nvme_command c = { }; 1983 int ret; 1984 1985 c.features.opcode = nvme_admin_set_features; 1986 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF); 1987 c.features.dword11 = cpu_to_le32(bits); 1988 c.features.dword12 = cpu_to_le32(host_mem_size); 1989 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr)); 1990 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr)); 1991 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs); 1992 1993 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0); 1994 if (ret) { 1995 dev_warn(dev->ctrl.device, 1996 "failed to set host mem (err %d, flags %#x).\n", 1997 ret, bits); 1998 } else 1999 dev->hmb = bits & NVME_HOST_MEM_ENABLE; 2000 2001 return ret; 2002 } 2003 2004 static void nvme_free_host_mem(struct nvme_dev *dev) 2005 { 2006 int i; 2007 2008 for (i = 0; i < dev->nr_host_mem_descs; i++) { 2009 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i]; 2010 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE; 2011 2012 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i], 2013 le64_to_cpu(desc->addr), 2014 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN); 2015 } 2016 2017 kfree(dev->host_mem_desc_bufs); 2018 dev->host_mem_desc_bufs = NULL; 2019 dma_free_coherent(dev->dev, 2020 dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs), 2021 dev->host_mem_descs, dev->host_mem_descs_dma); 2022 dev->host_mem_descs = NULL; 2023 dev->nr_host_mem_descs = 0; 2024 } 2025 2026 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred, 2027 u32 chunk_size) 2028 { 2029 struct nvme_host_mem_buf_desc *descs; 2030 u32 max_entries, len; 2031 dma_addr_t descs_dma; 2032 int i = 0; 2033 void **bufs; 2034 u64 size, tmp; 2035 2036 tmp = (preferred + chunk_size - 1); 2037 do_div(tmp, chunk_size); 2038 max_entries = tmp; 2039 2040 if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries) 2041 max_entries = dev->ctrl.hmmaxd; 2042 2043 descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs), 2044 &descs_dma, GFP_KERNEL); 2045 if (!descs) 2046 goto out; 2047 2048 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL); 2049 if (!bufs) 2050 goto out_free_descs; 2051 2052 for (size = 0; size < preferred && i < max_entries; size += len) { 2053 dma_addr_t dma_addr; 2054 2055 len = min_t(u64, chunk_size, preferred - size); 2056 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL, 2057 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN); 2058 if (!bufs[i]) 2059 break; 2060 2061 descs[i].addr = cpu_to_le64(dma_addr); 2062 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE); 2063 i++; 2064 } 2065 2066 if (!size) 2067 goto out_free_bufs; 2068 2069 dev->nr_host_mem_descs = i; 2070 dev->host_mem_size = size; 2071 dev->host_mem_descs = descs; 2072 dev->host_mem_descs_dma = descs_dma; 2073 dev->host_mem_desc_bufs = bufs; 2074 return 0; 2075 2076 out_free_bufs: 2077 while (--i >= 0) { 2078 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE; 2079 2080 dma_free_attrs(dev->dev, size, bufs[i], 2081 le64_to_cpu(descs[i].addr), 2082 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN); 2083 } 2084 2085 kfree(bufs); 2086 out_free_descs: 2087 dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs, 2088 descs_dma); 2089 out: 2090 dev->host_mem_descs = NULL; 2091 return -ENOMEM; 2092 } 2093 2094 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred) 2095 { 2096 u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES); 2097 u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2); 2098 u64 chunk_size; 2099 2100 /* start big and work our way down */ 2101 for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) { 2102 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) { 2103 if (!min || dev->host_mem_size >= min) 2104 return 0; 2105 nvme_free_host_mem(dev); 2106 } 2107 } 2108 2109 return -ENOMEM; 2110 } 2111 2112 static int nvme_setup_host_mem(struct nvme_dev *dev) 2113 { 2114 u64 max = (u64)max_host_mem_size_mb * SZ_1M; 2115 u64 preferred = (u64)dev->ctrl.hmpre * 4096; 2116 u64 min = (u64)dev->ctrl.hmmin * 4096; 2117 u32 enable_bits = NVME_HOST_MEM_ENABLE; 2118 int ret; 2119 2120 preferred = min(preferred, max); 2121 if (min > max) { 2122 dev_warn(dev->ctrl.device, 2123 "min host memory (%lld MiB) above limit (%d MiB).\n", 2124 min >> ilog2(SZ_1M), max_host_mem_size_mb); 2125 nvme_free_host_mem(dev); 2126 return 0; 2127 } 2128 2129 /* 2130 * If we already have a buffer allocated check if we can reuse it. 2131 */ 2132 if (dev->host_mem_descs) { 2133 if (dev->host_mem_size >= min) 2134 enable_bits |= NVME_HOST_MEM_RETURN; 2135 else 2136 nvme_free_host_mem(dev); 2137 } 2138 2139 if (!dev->host_mem_descs) { 2140 if (nvme_alloc_host_mem(dev, min, preferred)) { 2141 dev_warn(dev->ctrl.device, 2142 "failed to allocate host memory buffer.\n"); 2143 return 0; /* controller must work without HMB */ 2144 } 2145 2146 dev_info(dev->ctrl.device, 2147 "allocated %lld MiB host memory buffer.\n", 2148 dev->host_mem_size >> ilog2(SZ_1M)); 2149 } 2150 2151 ret = nvme_set_host_mem(dev, enable_bits); 2152 if (ret) 2153 nvme_free_host_mem(dev); 2154 return ret; 2155 } 2156 2157 static ssize_t cmb_show(struct device *dev, struct device_attribute *attr, 2158 char *buf) 2159 { 2160 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev)); 2161 2162 return sysfs_emit(buf, "cmbloc : x%08x\ncmbsz : x%08x\n", 2163 ndev->cmbloc, ndev->cmbsz); 2164 } 2165 static DEVICE_ATTR_RO(cmb); 2166 2167 static ssize_t cmbloc_show(struct device *dev, struct device_attribute *attr, 2168 char *buf) 2169 { 2170 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev)); 2171 2172 return sysfs_emit(buf, "%u\n", ndev->cmbloc); 2173 } 2174 static DEVICE_ATTR_RO(cmbloc); 2175 2176 static ssize_t cmbsz_show(struct device *dev, struct device_attribute *attr, 2177 char *buf) 2178 { 2179 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev)); 2180 2181 return sysfs_emit(buf, "%u\n", ndev->cmbsz); 2182 } 2183 static DEVICE_ATTR_RO(cmbsz); 2184 2185 static ssize_t hmb_show(struct device *dev, struct device_attribute *attr, 2186 char *buf) 2187 { 2188 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev)); 2189 2190 return sysfs_emit(buf, "%d\n", ndev->hmb); 2191 } 2192 2193 static ssize_t hmb_store(struct device *dev, struct device_attribute *attr, 2194 const char *buf, size_t count) 2195 { 2196 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev)); 2197 bool new; 2198 int ret; 2199 2200 if (strtobool(buf, &new) < 0) 2201 return -EINVAL; 2202 2203 if (new == ndev->hmb) 2204 return count; 2205 2206 if (new) { 2207 ret = nvme_setup_host_mem(ndev); 2208 } else { 2209 ret = nvme_set_host_mem(ndev, 0); 2210 if (!ret) 2211 nvme_free_host_mem(ndev); 2212 } 2213 2214 if (ret < 0) 2215 return ret; 2216 2217 return count; 2218 } 2219 static DEVICE_ATTR_RW(hmb); 2220 2221 static umode_t nvme_pci_attrs_are_visible(struct kobject *kobj, 2222 struct attribute *a, int n) 2223 { 2224 struct nvme_ctrl *ctrl = 2225 dev_get_drvdata(container_of(kobj, struct device, kobj)); 2226 struct nvme_dev *dev = to_nvme_dev(ctrl); 2227 2228 if (a == &dev_attr_cmb.attr || 2229 a == &dev_attr_cmbloc.attr || 2230 a == &dev_attr_cmbsz.attr) { 2231 if (!dev->cmbsz) 2232 return 0; 2233 } 2234 if (a == &dev_attr_hmb.attr && !ctrl->hmpre) 2235 return 0; 2236 2237 return a->mode; 2238 } 2239 2240 static struct attribute *nvme_pci_attrs[] = { 2241 &dev_attr_cmb.attr, 2242 &dev_attr_cmbloc.attr, 2243 &dev_attr_cmbsz.attr, 2244 &dev_attr_hmb.attr, 2245 NULL, 2246 }; 2247 2248 static const struct attribute_group nvme_pci_attr_group = { 2249 .attrs = nvme_pci_attrs, 2250 .is_visible = nvme_pci_attrs_are_visible, 2251 }; 2252 2253 /* 2254 * nirqs is the number of interrupts available for write and read 2255 * queues. The core already reserved an interrupt for the admin queue. 2256 */ 2257 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs) 2258 { 2259 struct nvme_dev *dev = affd->priv; 2260 unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues; 2261 2262 /* 2263 * If there is no interrupt available for queues, ensure that 2264 * the default queue is set to 1. The affinity set size is 2265 * also set to one, but the irq core ignores it for this case. 2266 * 2267 * If only one interrupt is available or 'write_queue' == 0, combine 2268 * write and read queues. 2269 * 2270 * If 'write_queues' > 0, ensure it leaves room for at least one read 2271 * queue. 2272 */ 2273 if (!nrirqs) { 2274 nrirqs = 1; 2275 nr_read_queues = 0; 2276 } else if (nrirqs == 1 || !nr_write_queues) { 2277 nr_read_queues = 0; 2278 } else if (nr_write_queues >= nrirqs) { 2279 nr_read_queues = 1; 2280 } else { 2281 nr_read_queues = nrirqs - nr_write_queues; 2282 } 2283 2284 dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues; 2285 affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues; 2286 dev->io_queues[HCTX_TYPE_READ] = nr_read_queues; 2287 affd->set_size[HCTX_TYPE_READ] = nr_read_queues; 2288 affd->nr_sets = nr_read_queues ? 2 : 1; 2289 } 2290 2291 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues) 2292 { 2293 struct pci_dev *pdev = to_pci_dev(dev->dev); 2294 struct irq_affinity affd = { 2295 .pre_vectors = 1, 2296 .calc_sets = nvme_calc_irq_sets, 2297 .priv = dev, 2298 }; 2299 unsigned int irq_queues, poll_queues; 2300 2301 /* 2302 * Poll queues don't need interrupts, but we need at least one I/O queue 2303 * left over for non-polled I/O. 2304 */ 2305 poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1); 2306 dev->io_queues[HCTX_TYPE_POLL] = poll_queues; 2307 2308 /* 2309 * Initialize for the single interrupt case, will be updated in 2310 * nvme_calc_irq_sets(). 2311 */ 2312 dev->io_queues[HCTX_TYPE_DEFAULT] = 1; 2313 dev->io_queues[HCTX_TYPE_READ] = 0; 2314 2315 /* 2316 * We need interrupts for the admin queue and each non-polled I/O queue, 2317 * but some Apple controllers require all queues to use the first 2318 * vector. 2319 */ 2320 irq_queues = 1; 2321 if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR)) 2322 irq_queues += (nr_io_queues - poll_queues); 2323 return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues, 2324 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd); 2325 } 2326 2327 static void nvme_disable_io_queues(struct nvme_dev *dev) 2328 { 2329 if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq)) 2330 __nvme_disable_io_queues(dev, nvme_admin_delete_cq); 2331 } 2332 2333 static unsigned int nvme_max_io_queues(struct nvme_dev *dev) 2334 { 2335 /* 2336 * If tags are shared with admin queue (Apple bug), then 2337 * make sure we only use one IO queue. 2338 */ 2339 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) 2340 return 1; 2341 return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues; 2342 } 2343 2344 static int nvme_setup_io_queues(struct nvme_dev *dev) 2345 { 2346 struct nvme_queue *adminq = &dev->queues[0]; 2347 struct pci_dev *pdev = to_pci_dev(dev->dev); 2348 unsigned int nr_io_queues; 2349 unsigned long size; 2350 int result; 2351 2352 /* 2353 * Sample the module parameters once at reset time so that we have 2354 * stable values to work with. 2355 */ 2356 dev->nr_write_queues = write_queues; 2357 dev->nr_poll_queues = poll_queues; 2358 2359 nr_io_queues = dev->nr_allocated_queues - 1; 2360 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues); 2361 if (result < 0) 2362 return result; 2363 2364 if (nr_io_queues == 0) 2365 return 0; 2366 2367 /* 2368 * Free IRQ resources as soon as NVMEQ_ENABLED bit transitions 2369 * from set to unset. If there is a window to it is truely freed, 2370 * pci_free_irq_vectors() jumping into this window will crash. 2371 * And take lock to avoid racing with pci_free_irq_vectors() in 2372 * nvme_dev_disable() path. 2373 */ 2374 result = nvme_setup_io_queues_trylock(dev); 2375 if (result) 2376 return result; 2377 if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags)) 2378 pci_free_irq(pdev, 0, adminq); 2379 2380 if (dev->cmb_use_sqes) { 2381 result = nvme_cmb_qdepth(dev, nr_io_queues, 2382 sizeof(struct nvme_command)); 2383 if (result > 0) 2384 dev->q_depth = result; 2385 else 2386 dev->cmb_use_sqes = false; 2387 } 2388 2389 do { 2390 size = db_bar_size(dev, nr_io_queues); 2391 result = nvme_remap_bar(dev, size); 2392 if (!result) 2393 break; 2394 if (!--nr_io_queues) { 2395 result = -ENOMEM; 2396 goto out_unlock; 2397 } 2398 } while (1); 2399 adminq->q_db = dev->dbs; 2400 2401 retry: 2402 /* Deregister the admin queue's interrupt */ 2403 if (test_and_clear_bit(NVMEQ_ENABLED, &adminq->flags)) 2404 pci_free_irq(pdev, 0, adminq); 2405 2406 /* 2407 * If we enable msix early due to not intx, disable it again before 2408 * setting up the full range we need. 2409 */ 2410 pci_free_irq_vectors(pdev); 2411 2412 result = nvme_setup_irqs(dev, nr_io_queues); 2413 if (result <= 0) { 2414 result = -EIO; 2415 goto out_unlock; 2416 } 2417 2418 dev->num_vecs = result; 2419 result = max(result - 1, 1); 2420 dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL]; 2421 2422 /* 2423 * Should investigate if there's a performance win from allocating 2424 * more queues than interrupt vectors; it might allow the submission 2425 * path to scale better, even if the receive path is limited by the 2426 * number of interrupts. 2427 */ 2428 result = queue_request_irq(adminq); 2429 if (result) 2430 goto out_unlock; 2431 set_bit(NVMEQ_ENABLED, &adminq->flags); 2432 mutex_unlock(&dev->shutdown_lock); 2433 2434 result = nvme_create_io_queues(dev); 2435 if (result || dev->online_queues < 2) 2436 return result; 2437 2438 if (dev->online_queues - 1 < dev->max_qid) { 2439 nr_io_queues = dev->online_queues - 1; 2440 nvme_disable_io_queues(dev); 2441 result = nvme_setup_io_queues_trylock(dev); 2442 if (result) 2443 return result; 2444 nvme_suspend_io_queues(dev); 2445 goto retry; 2446 } 2447 dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n", 2448 dev->io_queues[HCTX_TYPE_DEFAULT], 2449 dev->io_queues[HCTX_TYPE_READ], 2450 dev->io_queues[HCTX_TYPE_POLL]); 2451 return 0; 2452 out_unlock: 2453 mutex_unlock(&dev->shutdown_lock); 2454 return result; 2455 } 2456 2457 static void nvme_del_queue_end(struct request *req, blk_status_t error) 2458 { 2459 struct nvme_queue *nvmeq = req->end_io_data; 2460 2461 blk_mq_free_request(req); 2462 complete(&nvmeq->delete_done); 2463 } 2464 2465 static void nvme_del_cq_end(struct request *req, blk_status_t error) 2466 { 2467 struct nvme_queue *nvmeq = req->end_io_data; 2468 2469 if (error) 2470 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags); 2471 2472 nvme_del_queue_end(req, error); 2473 } 2474 2475 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode) 2476 { 2477 struct request_queue *q = nvmeq->dev->ctrl.admin_q; 2478 struct request *req; 2479 struct nvme_command cmd = { }; 2480 2481 cmd.delete_queue.opcode = opcode; 2482 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid); 2483 2484 req = blk_mq_alloc_request(q, nvme_req_op(&cmd), BLK_MQ_REQ_NOWAIT); 2485 if (IS_ERR(req)) 2486 return PTR_ERR(req); 2487 nvme_init_request(req, &cmd); 2488 2489 if (opcode == nvme_admin_delete_cq) 2490 req->end_io = nvme_del_cq_end; 2491 else 2492 req->end_io = nvme_del_queue_end; 2493 req->end_io_data = nvmeq; 2494 2495 init_completion(&nvmeq->delete_done); 2496 req->rq_flags |= RQF_QUIET; 2497 blk_execute_rq_nowait(req, false); 2498 return 0; 2499 } 2500 2501 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode) 2502 { 2503 int nr_queues = dev->online_queues - 1, sent = 0; 2504 unsigned long timeout; 2505 2506 retry: 2507 timeout = NVME_ADMIN_TIMEOUT; 2508 while (nr_queues > 0) { 2509 if (nvme_delete_queue(&dev->queues[nr_queues], opcode)) 2510 break; 2511 nr_queues--; 2512 sent++; 2513 } 2514 while (sent) { 2515 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent]; 2516 2517 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done, 2518 timeout); 2519 if (timeout == 0) 2520 return false; 2521 2522 sent--; 2523 if (nr_queues) 2524 goto retry; 2525 } 2526 return true; 2527 } 2528 2529 static void nvme_dev_add(struct nvme_dev *dev) 2530 { 2531 int ret; 2532 2533 if (!dev->ctrl.tagset) { 2534 dev->tagset.ops = &nvme_mq_ops; 2535 dev->tagset.nr_hw_queues = dev->online_queues - 1; 2536 dev->tagset.nr_maps = 2; /* default + read */ 2537 if (dev->io_queues[HCTX_TYPE_POLL]) 2538 dev->tagset.nr_maps++; 2539 dev->tagset.timeout = NVME_IO_TIMEOUT; 2540 dev->tagset.numa_node = dev->ctrl.numa_node; 2541 dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth, 2542 BLK_MQ_MAX_DEPTH) - 1; 2543 dev->tagset.cmd_size = sizeof(struct nvme_iod); 2544 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE; 2545 dev->tagset.driver_data = dev; 2546 2547 /* 2548 * Some Apple controllers requires tags to be unique 2549 * across admin and IO queue, so reserve the first 32 2550 * tags of the IO queue. 2551 */ 2552 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) 2553 dev->tagset.reserved_tags = NVME_AQ_DEPTH; 2554 2555 ret = blk_mq_alloc_tag_set(&dev->tagset); 2556 if (ret) { 2557 dev_warn(dev->ctrl.device, 2558 "IO queues tagset allocation failed %d\n", ret); 2559 return; 2560 } 2561 dev->ctrl.tagset = &dev->tagset; 2562 } else { 2563 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1); 2564 2565 /* Free previously allocated queues that are no longer usable */ 2566 nvme_free_queues(dev, dev->online_queues); 2567 } 2568 2569 nvme_dbbuf_set(dev); 2570 } 2571 2572 static int nvme_pci_enable(struct nvme_dev *dev) 2573 { 2574 int result = -ENOMEM; 2575 struct pci_dev *pdev = to_pci_dev(dev->dev); 2576 int dma_address_bits = 64; 2577 2578 if (pci_enable_device_mem(pdev)) 2579 return result; 2580 2581 pci_set_master(pdev); 2582 2583 if (dev->ctrl.quirks & NVME_QUIRK_DMA_ADDRESS_BITS_48) 2584 dma_address_bits = 48; 2585 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(dma_address_bits))) 2586 goto disable; 2587 2588 if (readl(dev->bar + NVME_REG_CSTS) == -1) { 2589 result = -ENODEV; 2590 goto disable; 2591 } 2592 2593 /* 2594 * Some devices and/or platforms don't advertise or work with INTx 2595 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll 2596 * adjust this later. 2597 */ 2598 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES); 2599 if (result < 0) 2600 return result; 2601 2602 dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP); 2603 2604 dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1, 2605 io_queue_depth); 2606 dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */ 2607 dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap); 2608 dev->dbs = dev->bar + 4096; 2609 2610 /* 2611 * Some Apple controllers require a non-standard SQE size. 2612 * Interestingly they also seem to ignore the CC:IOSQES register 2613 * so we don't bother updating it here. 2614 */ 2615 if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES) 2616 dev->io_sqes = 7; 2617 else 2618 dev->io_sqes = NVME_NVM_IOSQES; 2619 2620 /* 2621 * Temporary fix for the Apple controller found in the MacBook8,1 and 2622 * some MacBook7,1 to avoid controller resets and data loss. 2623 */ 2624 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) { 2625 dev->q_depth = 2; 2626 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, " 2627 "set queue depth=%u to work around controller resets\n", 2628 dev->q_depth); 2629 } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG && 2630 (pdev->device == 0xa821 || pdev->device == 0xa822) && 2631 NVME_CAP_MQES(dev->ctrl.cap) == 0) { 2632 dev->q_depth = 64; 2633 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, " 2634 "set queue depth=%u\n", dev->q_depth); 2635 } 2636 2637 /* 2638 * Controllers with the shared tags quirk need the IO queue to be 2639 * big enough so that we get 32 tags for the admin queue 2640 */ 2641 if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) && 2642 (dev->q_depth < (NVME_AQ_DEPTH + 2))) { 2643 dev->q_depth = NVME_AQ_DEPTH + 2; 2644 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n", 2645 dev->q_depth); 2646 } 2647 2648 2649 nvme_map_cmb(dev); 2650 2651 pci_enable_pcie_error_reporting(pdev); 2652 pci_save_state(pdev); 2653 return 0; 2654 2655 disable: 2656 pci_disable_device(pdev); 2657 return result; 2658 } 2659 2660 static void nvme_dev_unmap(struct nvme_dev *dev) 2661 { 2662 if (dev->bar) 2663 iounmap(dev->bar); 2664 pci_release_mem_regions(to_pci_dev(dev->dev)); 2665 } 2666 2667 static void nvme_pci_disable(struct nvme_dev *dev) 2668 { 2669 struct pci_dev *pdev = to_pci_dev(dev->dev); 2670 2671 pci_free_irq_vectors(pdev); 2672 2673 if (pci_is_enabled(pdev)) { 2674 pci_disable_pcie_error_reporting(pdev); 2675 pci_disable_device(pdev); 2676 } 2677 } 2678 2679 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) 2680 { 2681 bool dead = true, freeze = false; 2682 struct pci_dev *pdev = to_pci_dev(dev->dev); 2683 2684 mutex_lock(&dev->shutdown_lock); 2685 if (pci_device_is_present(pdev) && pci_is_enabled(pdev)) { 2686 u32 csts = readl(dev->bar + NVME_REG_CSTS); 2687 2688 if (dev->ctrl.state == NVME_CTRL_LIVE || 2689 dev->ctrl.state == NVME_CTRL_RESETTING) { 2690 freeze = true; 2691 nvme_start_freeze(&dev->ctrl); 2692 } 2693 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) || 2694 pdev->error_state != pci_channel_io_normal); 2695 } 2696 2697 /* 2698 * Give the controller a chance to complete all entered requests if 2699 * doing a safe shutdown. 2700 */ 2701 if (!dead && shutdown && freeze) 2702 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT); 2703 2704 nvme_stop_queues(&dev->ctrl); 2705 2706 if (!dead && dev->ctrl.queue_count > 0) { 2707 nvme_disable_io_queues(dev); 2708 nvme_disable_admin_queue(dev, shutdown); 2709 } 2710 nvme_suspend_io_queues(dev); 2711 nvme_suspend_queue(&dev->queues[0]); 2712 nvme_pci_disable(dev); 2713 nvme_reap_pending_cqes(dev); 2714 2715 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl); 2716 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl); 2717 blk_mq_tagset_wait_completed_request(&dev->tagset); 2718 blk_mq_tagset_wait_completed_request(&dev->admin_tagset); 2719 2720 /* 2721 * The driver will not be starting up queues again if shutting down so 2722 * must flush all entered requests to their failed completion to avoid 2723 * deadlocking blk-mq hot-cpu notifier. 2724 */ 2725 if (shutdown) { 2726 nvme_start_queues(&dev->ctrl); 2727 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) 2728 nvme_start_admin_queue(&dev->ctrl); 2729 } 2730 mutex_unlock(&dev->shutdown_lock); 2731 } 2732 2733 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown) 2734 { 2735 if (!nvme_wait_reset(&dev->ctrl)) 2736 return -EBUSY; 2737 nvme_dev_disable(dev, shutdown); 2738 return 0; 2739 } 2740 2741 static int nvme_setup_prp_pools(struct nvme_dev *dev) 2742 { 2743 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev, 2744 NVME_CTRL_PAGE_SIZE, 2745 NVME_CTRL_PAGE_SIZE, 0); 2746 if (!dev->prp_page_pool) 2747 return -ENOMEM; 2748 2749 /* Optimisation for I/Os between 4k and 128k */ 2750 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev, 2751 256, 256, 0); 2752 if (!dev->prp_small_pool) { 2753 dma_pool_destroy(dev->prp_page_pool); 2754 return -ENOMEM; 2755 } 2756 return 0; 2757 } 2758 2759 static void nvme_release_prp_pools(struct nvme_dev *dev) 2760 { 2761 dma_pool_destroy(dev->prp_page_pool); 2762 dma_pool_destroy(dev->prp_small_pool); 2763 } 2764 2765 static void nvme_free_tagset(struct nvme_dev *dev) 2766 { 2767 if (dev->tagset.tags) 2768 blk_mq_free_tag_set(&dev->tagset); 2769 dev->ctrl.tagset = NULL; 2770 } 2771 2772 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl) 2773 { 2774 struct nvme_dev *dev = to_nvme_dev(ctrl); 2775 2776 nvme_dbbuf_dma_free(dev); 2777 nvme_free_tagset(dev); 2778 if (dev->ctrl.admin_q) 2779 blk_put_queue(dev->ctrl.admin_q); 2780 free_opal_dev(dev->ctrl.opal_dev); 2781 mempool_destroy(dev->iod_mempool); 2782 put_device(dev->dev); 2783 kfree(dev->queues); 2784 kfree(dev); 2785 } 2786 2787 static void nvme_remove_dead_ctrl(struct nvme_dev *dev) 2788 { 2789 /* 2790 * Set state to deleting now to avoid blocking nvme_wait_reset(), which 2791 * may be holding this pci_dev's device lock. 2792 */ 2793 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING); 2794 nvme_get_ctrl(&dev->ctrl); 2795 nvme_dev_disable(dev, false); 2796 nvme_kill_queues(&dev->ctrl); 2797 if (!queue_work(nvme_wq, &dev->remove_work)) 2798 nvme_put_ctrl(&dev->ctrl); 2799 } 2800 2801 static void nvme_reset_work(struct work_struct *work) 2802 { 2803 struct nvme_dev *dev = 2804 container_of(work, struct nvme_dev, ctrl.reset_work); 2805 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL); 2806 int result; 2807 2808 if (dev->ctrl.state != NVME_CTRL_RESETTING) { 2809 dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n", 2810 dev->ctrl.state); 2811 result = -ENODEV; 2812 goto out; 2813 } 2814 2815 /* 2816 * If we're called to reset a live controller first shut it down before 2817 * moving on. 2818 */ 2819 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE) 2820 nvme_dev_disable(dev, false); 2821 nvme_sync_queues(&dev->ctrl); 2822 2823 mutex_lock(&dev->shutdown_lock); 2824 result = nvme_pci_enable(dev); 2825 if (result) 2826 goto out_unlock; 2827 2828 result = nvme_pci_configure_admin_queue(dev); 2829 if (result) 2830 goto out_unlock; 2831 2832 result = nvme_alloc_admin_tags(dev); 2833 if (result) 2834 goto out_unlock; 2835 2836 /* 2837 * Limit the max command size to prevent iod->sg allocations going 2838 * over a single page. 2839 */ 2840 dev->ctrl.max_hw_sectors = min_t(u32, 2841 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9); 2842 dev->ctrl.max_segments = NVME_MAX_SEGS; 2843 2844 /* 2845 * Don't limit the IOMMU merged segment size. 2846 */ 2847 dma_set_max_seg_size(dev->dev, 0xffffffff); 2848 dma_set_min_align_mask(dev->dev, NVME_CTRL_PAGE_SIZE - 1); 2849 2850 mutex_unlock(&dev->shutdown_lock); 2851 2852 /* 2853 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the 2854 * initializing procedure here. 2855 */ 2856 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) { 2857 dev_warn(dev->ctrl.device, 2858 "failed to mark controller CONNECTING\n"); 2859 result = -EBUSY; 2860 goto out; 2861 } 2862 2863 /* 2864 * We do not support an SGL for metadata (yet), so we are limited to a 2865 * single integrity segment for the separate metadata pointer. 2866 */ 2867 dev->ctrl.max_integrity_segments = 1; 2868 2869 result = nvme_init_ctrl_finish(&dev->ctrl); 2870 if (result) 2871 goto out; 2872 2873 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) { 2874 if (!dev->ctrl.opal_dev) 2875 dev->ctrl.opal_dev = 2876 init_opal_dev(&dev->ctrl, &nvme_sec_submit); 2877 else if (was_suspend) 2878 opal_unlock_from_suspend(dev->ctrl.opal_dev); 2879 } else { 2880 free_opal_dev(dev->ctrl.opal_dev); 2881 dev->ctrl.opal_dev = NULL; 2882 } 2883 2884 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) { 2885 result = nvme_dbbuf_dma_alloc(dev); 2886 if (result) 2887 dev_warn(dev->dev, 2888 "unable to allocate dma for dbbuf\n"); 2889 } 2890 2891 if (dev->ctrl.hmpre) { 2892 result = nvme_setup_host_mem(dev); 2893 if (result < 0) 2894 goto out; 2895 } 2896 2897 result = nvme_setup_io_queues(dev); 2898 if (result) 2899 goto out; 2900 2901 /* 2902 * Keep the controller around but remove all namespaces if we don't have 2903 * any working I/O queue. 2904 */ 2905 if (dev->online_queues < 2) { 2906 dev_warn(dev->ctrl.device, "IO queues not created\n"); 2907 nvme_kill_queues(&dev->ctrl); 2908 nvme_remove_namespaces(&dev->ctrl); 2909 nvme_free_tagset(dev); 2910 } else { 2911 nvme_start_queues(&dev->ctrl); 2912 nvme_wait_freeze(&dev->ctrl); 2913 nvme_dev_add(dev); 2914 nvme_unfreeze(&dev->ctrl); 2915 } 2916 2917 /* 2918 * If only admin queue live, keep it to do further investigation or 2919 * recovery. 2920 */ 2921 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) { 2922 dev_warn(dev->ctrl.device, 2923 "failed to mark controller live state\n"); 2924 result = -ENODEV; 2925 goto out; 2926 } 2927 2928 if (!dev->attrs_added && !sysfs_create_group(&dev->ctrl.device->kobj, 2929 &nvme_pci_attr_group)) 2930 dev->attrs_added = true; 2931 2932 nvme_start_ctrl(&dev->ctrl); 2933 return; 2934 2935 out_unlock: 2936 mutex_unlock(&dev->shutdown_lock); 2937 out: 2938 if (result) 2939 dev_warn(dev->ctrl.device, 2940 "Removing after probe failure status: %d\n", result); 2941 nvme_remove_dead_ctrl(dev); 2942 } 2943 2944 static void nvme_remove_dead_ctrl_work(struct work_struct *work) 2945 { 2946 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work); 2947 struct pci_dev *pdev = to_pci_dev(dev->dev); 2948 2949 if (pci_get_drvdata(pdev)) 2950 device_release_driver(&pdev->dev); 2951 nvme_put_ctrl(&dev->ctrl); 2952 } 2953 2954 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val) 2955 { 2956 *val = readl(to_nvme_dev(ctrl)->bar + off); 2957 return 0; 2958 } 2959 2960 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val) 2961 { 2962 writel(val, to_nvme_dev(ctrl)->bar + off); 2963 return 0; 2964 } 2965 2966 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val) 2967 { 2968 *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off); 2969 return 0; 2970 } 2971 2972 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size) 2973 { 2974 struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev); 2975 2976 return snprintf(buf, size, "%s\n", dev_name(&pdev->dev)); 2977 } 2978 2979 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = { 2980 .name = "pcie", 2981 .module = THIS_MODULE, 2982 .flags = NVME_F_METADATA_SUPPORTED | 2983 NVME_F_PCI_P2PDMA, 2984 .reg_read32 = nvme_pci_reg_read32, 2985 .reg_write32 = nvme_pci_reg_write32, 2986 .reg_read64 = nvme_pci_reg_read64, 2987 .free_ctrl = nvme_pci_free_ctrl, 2988 .submit_async_event = nvme_pci_submit_async_event, 2989 .get_address = nvme_pci_get_address, 2990 }; 2991 2992 static int nvme_dev_map(struct nvme_dev *dev) 2993 { 2994 struct pci_dev *pdev = to_pci_dev(dev->dev); 2995 2996 if (pci_request_mem_regions(pdev, "nvme")) 2997 return -ENODEV; 2998 2999 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096)) 3000 goto release; 3001 3002 return 0; 3003 release: 3004 pci_release_mem_regions(pdev); 3005 return -ENODEV; 3006 } 3007 3008 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev) 3009 { 3010 if (pdev->vendor == 0x144d && pdev->device == 0xa802) { 3011 /* 3012 * Several Samsung devices seem to drop off the PCIe bus 3013 * randomly when APST is on and uses the deepest sleep state. 3014 * This has been observed on a Samsung "SM951 NVMe SAMSUNG 3015 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD 3016 * 950 PRO 256GB", but it seems to be restricted to two Dell 3017 * laptops. 3018 */ 3019 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") && 3020 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") || 3021 dmi_match(DMI_PRODUCT_NAME, "Precision 5510"))) 3022 return NVME_QUIRK_NO_DEEPEST_PS; 3023 } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) { 3024 /* 3025 * Samsung SSD 960 EVO drops off the PCIe bus after system 3026 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as 3027 * within few minutes after bootup on a Coffee Lake board - 3028 * ASUS PRIME Z370-A 3029 */ 3030 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") && 3031 (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") || 3032 dmi_match(DMI_BOARD_NAME, "PRIME Z370-A"))) 3033 return NVME_QUIRK_NO_APST; 3034 } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 || 3035 pdev->device == 0xa808 || pdev->device == 0xa809)) || 3036 (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) { 3037 /* 3038 * Forcing to use host managed nvme power settings for 3039 * lowest idle power with quick resume latency on 3040 * Samsung and Toshiba SSDs based on suspend behavior 3041 * on Coffee Lake board for LENOVO C640 3042 */ 3043 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) && 3044 dmi_match(DMI_BOARD_NAME, "LNVNB161216")) 3045 return NVME_QUIRK_SIMPLE_SUSPEND; 3046 } 3047 3048 return 0; 3049 } 3050 3051 static void nvme_async_probe(void *data, async_cookie_t cookie) 3052 { 3053 struct nvme_dev *dev = data; 3054 3055 flush_work(&dev->ctrl.reset_work); 3056 flush_work(&dev->ctrl.scan_work); 3057 nvme_put_ctrl(&dev->ctrl); 3058 } 3059 3060 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id) 3061 { 3062 int node, result = -ENOMEM; 3063 struct nvme_dev *dev; 3064 unsigned long quirks = id->driver_data; 3065 size_t alloc_size; 3066 3067 node = dev_to_node(&pdev->dev); 3068 if (node == NUMA_NO_NODE) 3069 set_dev_node(&pdev->dev, first_memory_node); 3070 3071 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node); 3072 if (!dev) 3073 return -ENOMEM; 3074 3075 dev->nr_write_queues = write_queues; 3076 dev->nr_poll_queues = poll_queues; 3077 dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1; 3078 dev->queues = kcalloc_node(dev->nr_allocated_queues, 3079 sizeof(struct nvme_queue), GFP_KERNEL, node); 3080 if (!dev->queues) 3081 goto free; 3082 3083 dev->dev = get_device(&pdev->dev); 3084 pci_set_drvdata(pdev, dev); 3085 3086 result = nvme_dev_map(dev); 3087 if (result) 3088 goto put_pci; 3089 3090 INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work); 3091 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work); 3092 mutex_init(&dev->shutdown_lock); 3093 3094 result = nvme_setup_prp_pools(dev); 3095 if (result) 3096 goto unmap; 3097 3098 quirks |= check_vendor_combination_bug(pdev); 3099 3100 if (!noacpi && acpi_storage_d3(&pdev->dev)) { 3101 /* 3102 * Some systems use a bios work around to ask for D3 on 3103 * platforms that support kernel managed suspend. 3104 */ 3105 dev_info(&pdev->dev, 3106 "platform quirk: setting simple suspend\n"); 3107 quirks |= NVME_QUIRK_SIMPLE_SUSPEND; 3108 } 3109 3110 /* 3111 * Double check that our mempool alloc size will cover the biggest 3112 * command we support. 3113 */ 3114 alloc_size = nvme_pci_iod_alloc_size(); 3115 WARN_ON_ONCE(alloc_size > PAGE_SIZE); 3116 3117 dev->iod_mempool = mempool_create_node(1, mempool_kmalloc, 3118 mempool_kfree, 3119 (void *) alloc_size, 3120 GFP_KERNEL, node); 3121 if (!dev->iod_mempool) { 3122 result = -ENOMEM; 3123 goto release_pools; 3124 } 3125 3126 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops, 3127 quirks); 3128 if (result) 3129 goto release_mempool; 3130 3131 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev)); 3132 3133 nvme_reset_ctrl(&dev->ctrl); 3134 async_schedule(nvme_async_probe, dev); 3135 3136 return 0; 3137 3138 release_mempool: 3139 mempool_destroy(dev->iod_mempool); 3140 release_pools: 3141 nvme_release_prp_pools(dev); 3142 unmap: 3143 nvme_dev_unmap(dev); 3144 put_pci: 3145 put_device(dev->dev); 3146 free: 3147 kfree(dev->queues); 3148 kfree(dev); 3149 return result; 3150 } 3151 3152 static void nvme_reset_prepare(struct pci_dev *pdev) 3153 { 3154 struct nvme_dev *dev = pci_get_drvdata(pdev); 3155 3156 /* 3157 * We don't need to check the return value from waiting for the reset 3158 * state as pci_dev device lock is held, making it impossible to race 3159 * with ->remove(). 3160 */ 3161 nvme_disable_prepare_reset(dev, false); 3162 nvme_sync_queues(&dev->ctrl); 3163 } 3164 3165 static void nvme_reset_done(struct pci_dev *pdev) 3166 { 3167 struct nvme_dev *dev = pci_get_drvdata(pdev); 3168 3169 if (!nvme_try_sched_reset(&dev->ctrl)) 3170 flush_work(&dev->ctrl.reset_work); 3171 } 3172 3173 static void nvme_shutdown(struct pci_dev *pdev) 3174 { 3175 struct nvme_dev *dev = pci_get_drvdata(pdev); 3176 3177 nvme_disable_prepare_reset(dev, true); 3178 } 3179 3180 static void nvme_remove_attrs(struct nvme_dev *dev) 3181 { 3182 if (dev->attrs_added) 3183 sysfs_remove_group(&dev->ctrl.device->kobj, 3184 &nvme_pci_attr_group); 3185 } 3186 3187 /* 3188 * The driver's remove may be called on a device in a partially initialized 3189 * state. This function must not have any dependencies on the device state in 3190 * order to proceed. 3191 */ 3192 static void nvme_remove(struct pci_dev *pdev) 3193 { 3194 struct nvme_dev *dev = pci_get_drvdata(pdev); 3195 3196 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING); 3197 pci_set_drvdata(pdev, NULL); 3198 3199 if (!pci_device_is_present(pdev)) { 3200 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD); 3201 nvme_dev_disable(dev, true); 3202 } 3203 3204 flush_work(&dev->ctrl.reset_work); 3205 nvme_stop_ctrl(&dev->ctrl); 3206 nvme_remove_namespaces(&dev->ctrl); 3207 nvme_dev_disable(dev, true); 3208 nvme_remove_attrs(dev); 3209 nvme_free_host_mem(dev); 3210 nvme_dev_remove_admin(dev); 3211 nvme_free_queues(dev, 0); 3212 nvme_release_prp_pools(dev); 3213 nvme_dev_unmap(dev); 3214 nvme_uninit_ctrl(&dev->ctrl); 3215 } 3216 3217 #ifdef CONFIG_PM_SLEEP 3218 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps) 3219 { 3220 return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps); 3221 } 3222 3223 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps) 3224 { 3225 return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL); 3226 } 3227 3228 static int nvme_resume(struct device *dev) 3229 { 3230 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev)); 3231 struct nvme_ctrl *ctrl = &ndev->ctrl; 3232 3233 if (ndev->last_ps == U32_MAX || 3234 nvme_set_power_state(ctrl, ndev->last_ps) != 0) 3235 goto reset; 3236 if (ctrl->hmpre && nvme_setup_host_mem(ndev)) 3237 goto reset; 3238 3239 return 0; 3240 reset: 3241 return nvme_try_sched_reset(ctrl); 3242 } 3243 3244 static int nvme_suspend(struct device *dev) 3245 { 3246 struct pci_dev *pdev = to_pci_dev(dev); 3247 struct nvme_dev *ndev = pci_get_drvdata(pdev); 3248 struct nvme_ctrl *ctrl = &ndev->ctrl; 3249 int ret = -EBUSY; 3250 3251 ndev->last_ps = U32_MAX; 3252 3253 /* 3254 * The platform does not remove power for a kernel managed suspend so 3255 * use host managed nvme power settings for lowest idle power if 3256 * possible. This should have quicker resume latency than a full device 3257 * shutdown. But if the firmware is involved after the suspend or the 3258 * device does not support any non-default power states, shut down the 3259 * device fully. 3260 * 3261 * If ASPM is not enabled for the device, shut down the device and allow 3262 * the PCI bus layer to put it into D3 in order to take the PCIe link 3263 * down, so as to allow the platform to achieve its minimum low-power 3264 * state (which may not be possible if the link is up). 3265 */ 3266 if (pm_suspend_via_firmware() || !ctrl->npss || 3267 !pcie_aspm_enabled(pdev) || 3268 (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND)) 3269 return nvme_disable_prepare_reset(ndev, true); 3270 3271 nvme_start_freeze(ctrl); 3272 nvme_wait_freeze(ctrl); 3273 nvme_sync_queues(ctrl); 3274 3275 if (ctrl->state != NVME_CTRL_LIVE) 3276 goto unfreeze; 3277 3278 /* 3279 * Host memory access may not be successful in a system suspend state, 3280 * but the specification allows the controller to access memory in a 3281 * non-operational power state. 3282 */ 3283 if (ndev->hmb) { 3284 ret = nvme_set_host_mem(ndev, 0); 3285 if (ret < 0) 3286 goto unfreeze; 3287 } 3288 3289 ret = nvme_get_power_state(ctrl, &ndev->last_ps); 3290 if (ret < 0) 3291 goto unfreeze; 3292 3293 /* 3294 * A saved state prevents pci pm from generically controlling the 3295 * device's power. If we're using protocol specific settings, we don't 3296 * want pci interfering. 3297 */ 3298 pci_save_state(pdev); 3299 3300 ret = nvme_set_power_state(ctrl, ctrl->npss); 3301 if (ret < 0) 3302 goto unfreeze; 3303 3304 if (ret) { 3305 /* discard the saved state */ 3306 pci_load_saved_state(pdev, NULL); 3307 3308 /* 3309 * Clearing npss forces a controller reset on resume. The 3310 * correct value will be rediscovered then. 3311 */ 3312 ret = nvme_disable_prepare_reset(ndev, true); 3313 ctrl->npss = 0; 3314 } 3315 unfreeze: 3316 nvme_unfreeze(ctrl); 3317 return ret; 3318 } 3319 3320 static int nvme_simple_suspend(struct device *dev) 3321 { 3322 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev)); 3323 3324 return nvme_disable_prepare_reset(ndev, true); 3325 } 3326 3327 static int nvme_simple_resume(struct device *dev) 3328 { 3329 struct pci_dev *pdev = to_pci_dev(dev); 3330 struct nvme_dev *ndev = pci_get_drvdata(pdev); 3331 3332 return nvme_try_sched_reset(&ndev->ctrl); 3333 } 3334 3335 static const struct dev_pm_ops nvme_dev_pm_ops = { 3336 .suspend = nvme_suspend, 3337 .resume = nvme_resume, 3338 .freeze = nvme_simple_suspend, 3339 .thaw = nvme_simple_resume, 3340 .poweroff = nvme_simple_suspend, 3341 .restore = nvme_simple_resume, 3342 }; 3343 #endif /* CONFIG_PM_SLEEP */ 3344 3345 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev, 3346 pci_channel_state_t state) 3347 { 3348 struct nvme_dev *dev = pci_get_drvdata(pdev); 3349 3350 /* 3351 * A frozen channel requires a reset. When detected, this method will 3352 * shutdown the controller to quiesce. The controller will be restarted 3353 * after the slot reset through driver's slot_reset callback. 3354 */ 3355 switch (state) { 3356 case pci_channel_io_normal: 3357 return PCI_ERS_RESULT_CAN_RECOVER; 3358 case pci_channel_io_frozen: 3359 dev_warn(dev->ctrl.device, 3360 "frozen state error detected, reset controller\n"); 3361 nvme_dev_disable(dev, false); 3362 return PCI_ERS_RESULT_NEED_RESET; 3363 case pci_channel_io_perm_failure: 3364 dev_warn(dev->ctrl.device, 3365 "failure state error detected, request disconnect\n"); 3366 return PCI_ERS_RESULT_DISCONNECT; 3367 } 3368 return PCI_ERS_RESULT_NEED_RESET; 3369 } 3370 3371 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev) 3372 { 3373 struct nvme_dev *dev = pci_get_drvdata(pdev); 3374 3375 dev_info(dev->ctrl.device, "restart after slot reset\n"); 3376 pci_restore_state(pdev); 3377 nvme_reset_ctrl(&dev->ctrl); 3378 return PCI_ERS_RESULT_RECOVERED; 3379 } 3380 3381 static void nvme_error_resume(struct pci_dev *pdev) 3382 { 3383 struct nvme_dev *dev = pci_get_drvdata(pdev); 3384 3385 flush_work(&dev->ctrl.reset_work); 3386 } 3387 3388 static const struct pci_error_handlers nvme_err_handler = { 3389 .error_detected = nvme_error_detected, 3390 .slot_reset = nvme_slot_reset, 3391 .resume = nvme_error_resume, 3392 .reset_prepare = nvme_reset_prepare, 3393 .reset_done = nvme_reset_done, 3394 }; 3395 3396 static const struct pci_device_id nvme_id_table[] = { 3397 { PCI_VDEVICE(INTEL, 0x0953), /* Intel 750/P3500/P3600/P3700 */ 3398 .driver_data = NVME_QUIRK_STRIPE_SIZE | 3399 NVME_QUIRK_DEALLOCATE_ZEROES, }, 3400 { PCI_VDEVICE(INTEL, 0x0a53), /* Intel P3520 */ 3401 .driver_data = NVME_QUIRK_STRIPE_SIZE | 3402 NVME_QUIRK_DEALLOCATE_ZEROES, }, 3403 { PCI_VDEVICE(INTEL, 0x0a54), /* Intel P4500/P4600 */ 3404 .driver_data = NVME_QUIRK_STRIPE_SIZE | 3405 NVME_QUIRK_DEALLOCATE_ZEROES | 3406 NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3407 { PCI_VDEVICE(INTEL, 0x0a55), /* Dell Express Flash P4600 */ 3408 .driver_data = NVME_QUIRK_STRIPE_SIZE | 3409 NVME_QUIRK_DEALLOCATE_ZEROES, }, 3410 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */ 3411 .driver_data = NVME_QUIRK_NO_DEEPEST_PS | 3412 NVME_QUIRK_MEDIUM_PRIO_SQ | 3413 NVME_QUIRK_NO_TEMP_THRESH_CHANGE | 3414 NVME_QUIRK_DISABLE_WRITE_ZEROES, }, 3415 { PCI_VDEVICE(INTEL, 0xf1a6), /* Intel 760p/Pro 7600p */ 3416 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3417 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */ 3418 .driver_data = NVME_QUIRK_IDENTIFY_CNS | 3419 NVME_QUIRK_DISABLE_WRITE_ZEROES | 3420 NVME_QUIRK_BOGUS_NID, }, 3421 { PCI_VDEVICE(REDHAT, 0x0010), /* Qemu emulated controller */ 3422 .driver_data = NVME_QUIRK_BOGUS_NID, }, 3423 { PCI_DEVICE(0x126f, 0x2263), /* Silicon Motion unidentified */ 3424 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST, }, 3425 { PCI_DEVICE(0x1bb1, 0x0100), /* Seagate Nytro Flash Storage */ 3426 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY | 3427 NVME_QUIRK_NO_NS_DESC_LIST, }, 3428 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */ 3429 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, }, 3430 { PCI_DEVICE(0x1c58, 0x0023), /* WDC SN200 adapter */ 3431 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, }, 3432 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */ 3433 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, }, 3434 { PCI_DEVICE(0x144d, 0xa821), /* Samsung PM1725 */ 3435 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, }, 3436 { PCI_DEVICE(0x144d, 0xa822), /* Samsung PM1725a */ 3437 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY | 3438 NVME_QUIRK_DISABLE_WRITE_ZEROES| 3439 NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3440 { PCI_DEVICE(0x1987, 0x5016), /* Phison E16 */ 3441 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3442 { PCI_DEVICE(0x1b4b, 0x1092), /* Lexar 256 GB SSD */ 3443 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST | 3444 NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3445 { PCI_DEVICE(0x10ec, 0x5762), /* ADATA SX6000LNP */ 3446 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3447 { PCI_DEVICE(0x1cc1, 0x8201), /* ADATA SX8200PNP 512GB */ 3448 .driver_data = NVME_QUIRK_NO_DEEPEST_PS | 3449 NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3450 { PCI_DEVICE(0x1c5c, 0x1504), /* SK Hynix PC400 */ 3451 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, }, 3452 { PCI_DEVICE(0x15b7, 0x2001), /* Sandisk Skyhawk */ 3453 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, }, 3454 { PCI_DEVICE(0x1d97, 0x2263), /* SPCC */ 3455 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, }, 3456 { PCI_DEVICE(0x2646, 0x2262), /* KINGSTON SKC2000 NVMe SSD */ 3457 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, }, 3458 { PCI_DEVICE(0x2646, 0x2263), /* KINGSTON A2000 NVMe SSD */ 3459 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, }, 3460 { PCI_DEVICE(0x1e4B, 0x1001), /* MAXIO MAP1001 */ 3461 .driver_data = NVME_QUIRK_BOGUS_NID, }, 3462 { PCI_DEVICE(0x1e4B, 0x1002), /* MAXIO MAP1002 */ 3463 .driver_data = NVME_QUIRK_BOGUS_NID, }, 3464 { PCI_DEVICE(0x1e4B, 0x1202), /* MAXIO MAP1202 */ 3465 .driver_data = NVME_QUIRK_BOGUS_NID, }, 3466 { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0061), 3467 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, }, 3468 { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x0065), 3469 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, }, 3470 { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0x8061), 3471 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, }, 3472 { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd00), 3473 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, }, 3474 { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd01), 3475 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, }, 3476 { PCI_DEVICE(PCI_VENDOR_ID_AMAZON, 0xcd02), 3477 .driver_data = NVME_QUIRK_DMA_ADDRESS_BITS_48, }, 3478 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001), 3479 .driver_data = NVME_QUIRK_SINGLE_VECTOR }, 3480 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) }, 3481 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005), 3482 .driver_data = NVME_QUIRK_SINGLE_VECTOR | 3483 NVME_QUIRK_128_BYTES_SQES | 3484 NVME_QUIRK_SHARED_TAGS | 3485 NVME_QUIRK_SKIP_CID_GEN }, 3486 { PCI_DEVICE(0x144d, 0xa808), /* Samsung X5 */ 3487 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY| 3488 NVME_QUIRK_NO_DEEPEST_PS | 3489 NVME_QUIRK_IGNORE_DEV_SUBNQN, }, 3490 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) }, 3491 { 0, } 3492 }; 3493 MODULE_DEVICE_TABLE(pci, nvme_id_table); 3494 3495 static struct pci_driver nvme_driver = { 3496 .name = "nvme", 3497 .id_table = nvme_id_table, 3498 .probe = nvme_probe, 3499 .remove = nvme_remove, 3500 .shutdown = nvme_shutdown, 3501 #ifdef CONFIG_PM_SLEEP 3502 .driver = { 3503 .pm = &nvme_dev_pm_ops, 3504 }, 3505 #endif 3506 .sriov_configure = pci_sriov_configure_simple, 3507 .err_handler = &nvme_err_handler, 3508 }; 3509 3510 static int __init nvme_init(void) 3511 { 3512 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64); 3513 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64); 3514 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64); 3515 BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2); 3516 3517 return pci_register_driver(&nvme_driver); 3518 } 3519 3520 static void __exit nvme_exit(void) 3521 { 3522 pci_unregister_driver(&nvme_driver); 3523 flush_workqueue(nvme_wq); 3524 } 3525 3526 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>"); 3527 MODULE_LICENSE("GPL"); 3528 MODULE_VERSION("1.0"); 3529 module_init(nvme_init); 3530 module_exit(nvme_exit); 3531