1 // SPDX-License-Identifier: GPL-2.0
3 * NVM Express device driver
4 * Copyright (c) 2011-2014, Intel Corporation.
7 #include <linux/acpi.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/dmi.h>
14 #include <linux/init.h>
15 #include <linux/interrupt.h>
18 #include <linux/module.h>
19 #include <linux/mutex.h>
20 #include <linux/once.h>
21 #include <linux/pci.h>
22 #include <linux/suspend.h>
23 #include <linux/t10-pi.h>
24 #include <linux/types.h>
25 #include <linux/io-64-nonatomic-lo-hi.h>
26 #include <linux/sed-opal.h>
27 #include <linux/pci-p2pdma.h>
32 #define SQ_SIZE(q) ((q)->q_depth << (q)->sqes)
33 #define CQ_SIZE(q) ((q)->q_depth * sizeof(struct nvme_completion))
35 #define SGES_PER_PAGE (PAGE_SIZE / sizeof(struct nvme_sgl_desc))
38 * These can be higher, but we need to ensure that any command doesn't
39 * require an sg allocation that needs more than a page of data.
41 #define NVME_MAX_KB_SZ 4096
42 #define NVME_MAX_SEGS 127
44 static int use_threaded_interrupts;
45 module_param(use_threaded_interrupts, int, 0);
47 static bool use_cmb_sqes = true;
48 module_param(use_cmb_sqes, bool, 0444);
49 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
51 static unsigned int max_host_mem_size_mb = 128;
52 module_param(max_host_mem_size_mb, uint, 0444);
53 MODULE_PARM_DESC(max_host_mem_size_mb,
54 "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
56 static unsigned int sgl_threshold = SZ_32K;
57 module_param(sgl_threshold, uint, 0644);
58 MODULE_PARM_DESC(sgl_threshold,
59 "Use SGLs when average request segment size is larger or equal to "
60 "this size. Use 0 to disable SGLs.");
62 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
63 static const struct kernel_param_ops io_queue_depth_ops = {
64 .set = io_queue_depth_set,
65 .get = param_get_uint,
68 static unsigned int io_queue_depth = 1024;
69 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
70 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
72 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
77 ret = kstrtouint(val, 10, &n);
78 if (ret != 0 || n > num_possible_cpus())
80 return param_set_uint(val, kp);
83 static const struct kernel_param_ops io_queue_count_ops = {
84 .set = io_queue_count_set,
85 .get = param_get_uint,
88 static unsigned int write_queues;
89 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
90 MODULE_PARM_DESC(write_queues,
91 "Number of queues to use for writes. If not set, reads and writes "
92 "will share a queue set.");
94 static unsigned int poll_queues;
95 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
96 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
99 module_param(noacpi, bool, 0444);
100 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
105 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
106 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
109 * Represents an NVM Express device. Each nvme_dev is a PCI function.
112 struct nvme_queue *queues;
113 struct blk_mq_tag_set tagset;
114 struct blk_mq_tag_set admin_tagset;
117 struct dma_pool *prp_page_pool;
118 struct dma_pool *prp_small_pool;
119 unsigned online_queues;
121 unsigned io_queues[HCTX_MAX_TYPES];
122 unsigned int num_vecs;
127 unsigned long bar_mapped_size;
128 struct work_struct remove_work;
129 struct mutex shutdown_lock;
135 struct nvme_ctrl ctrl;
138 mempool_t *iod_mempool;
140 /* shadow doorbell buffer support: */
142 dma_addr_t dbbuf_dbs_dma_addr;
144 dma_addr_t dbbuf_eis_dma_addr;
146 /* host memory buffer support: */
148 u32 nr_host_mem_descs;
149 dma_addr_t host_mem_descs_dma;
150 struct nvme_host_mem_buf_desc *host_mem_descs;
151 void **host_mem_desc_bufs;
152 unsigned int nr_allocated_queues;
153 unsigned int nr_write_queues;
154 unsigned int nr_poll_queues;
157 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
162 ret = kstrtou32(val, 10, &n);
163 if (ret != 0 || n < 2)
166 return param_set_uint(val, kp);
169 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
171 return qid * 2 * stride;
174 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
176 return (qid * 2 + 1) * stride;
179 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
181 return container_of(ctrl, struct nvme_dev, ctrl);
185 * An NVM Express queue. Each device has at least two (one for admin
186 * commands and one for I/O commands).
189 struct nvme_dev *dev;
192 /* only used for poll queues: */
193 spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
194 struct nvme_completion *cqes;
195 dma_addr_t sq_dma_addr;
196 dma_addr_t cq_dma_addr;
207 #define NVMEQ_ENABLED 0
208 #define NVMEQ_SQ_CMB 1
209 #define NVMEQ_DELETE_ERROR 2
210 #define NVMEQ_POLLED 3
215 struct completion delete_done;
219 * The nvme_iod describes the data in an I/O.
221 * The sg pointer contains the list of PRP/SGL chunk allocations in addition
222 * to the actual struct scatterlist.
225 struct nvme_request req;
226 struct nvme_queue *nvmeq;
229 int npages; /* In the PRP list. 0 means small pool in use */
230 int nents; /* Used in scatterlist */
231 dma_addr_t first_dma;
232 unsigned int dma_len; /* length of single DMA segment mapping */
234 struct scatterlist *sg;
237 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
239 return dev->nr_allocated_queues * 8 * dev->db_stride;
242 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
244 unsigned int mem_size = nvme_dbbuf_size(dev);
249 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
250 &dev->dbbuf_dbs_dma_addr,
254 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
255 &dev->dbbuf_eis_dma_addr,
257 if (!dev->dbbuf_eis) {
258 dma_free_coherent(dev->dev, mem_size,
259 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
260 dev->dbbuf_dbs = NULL;
267 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
269 unsigned int mem_size = nvme_dbbuf_size(dev);
271 if (dev->dbbuf_dbs) {
272 dma_free_coherent(dev->dev, mem_size,
273 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
274 dev->dbbuf_dbs = NULL;
276 if (dev->dbbuf_eis) {
277 dma_free_coherent(dev->dev, mem_size,
278 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
279 dev->dbbuf_eis = NULL;
283 static void nvme_dbbuf_init(struct nvme_dev *dev,
284 struct nvme_queue *nvmeq, int qid)
286 if (!dev->dbbuf_dbs || !qid)
289 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
290 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
291 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
292 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
295 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
300 nvmeq->dbbuf_sq_db = NULL;
301 nvmeq->dbbuf_cq_db = NULL;
302 nvmeq->dbbuf_sq_ei = NULL;
303 nvmeq->dbbuf_cq_ei = NULL;
306 static void nvme_dbbuf_set(struct nvme_dev *dev)
308 struct nvme_command c;
314 memset(&c, 0, sizeof(c));
315 c.dbbuf.opcode = nvme_admin_dbbuf;
316 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
317 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
319 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
320 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
321 /* Free memory and continue on */
322 nvme_dbbuf_dma_free(dev);
324 for (i = 1; i <= dev->online_queues; i++)
325 nvme_dbbuf_free(&dev->queues[i]);
329 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
331 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
334 /* Update dbbuf and return true if an MMIO is required */
335 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
336 volatile u32 *dbbuf_ei)
342 * Ensure that the queue is written before updating
343 * the doorbell in memory
347 old_value = *dbbuf_db;
351 * Ensure that the doorbell is updated before reading the event
352 * index from memory. The controller needs to provide similar
353 * ordering to ensure the envent index is updated before reading
358 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
366 * Will slightly overestimate the number of pages needed. This is OK
367 * as it only leads to a small amount of wasted memory for the lifetime of
370 static int nvme_pci_npages_prp(void)
372 unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
373 NVME_CTRL_PAGE_SIZE);
374 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
378 * Calculates the number of pages needed for the SGL segments. For example a 4k
379 * page can accommodate 256 SGL descriptors.
381 static int nvme_pci_npages_sgl(void)
383 return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc),
387 static size_t nvme_pci_iod_alloc_size(void)
389 size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl());
391 return sizeof(__le64 *) * npages +
392 sizeof(struct scatterlist) * NVME_MAX_SEGS;
395 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
396 unsigned int hctx_idx)
398 struct nvme_dev *dev = data;
399 struct nvme_queue *nvmeq = &dev->queues[0];
401 WARN_ON(hctx_idx != 0);
402 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
404 hctx->driver_data = nvmeq;
408 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
409 unsigned int hctx_idx)
411 struct nvme_dev *dev = data;
412 struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
414 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
415 hctx->driver_data = nvmeq;
419 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
420 unsigned int hctx_idx, unsigned int numa_node)
422 struct nvme_dev *dev = set->driver_data;
423 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
424 int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
425 struct nvme_queue *nvmeq = &dev->queues[queue_idx];
430 nvme_req(req)->ctrl = &dev->ctrl;
434 static int queue_irq_offset(struct nvme_dev *dev)
436 /* if we have more than 1 vec, admin queue offsets us by 1 */
437 if (dev->num_vecs > 1)
443 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
445 struct nvme_dev *dev = set->driver_data;
448 offset = queue_irq_offset(dev);
449 for (i = 0, qoff = 0; i < set->nr_maps; i++) {
450 struct blk_mq_queue_map *map = &set->map[i];
452 map->nr_queues = dev->io_queues[i];
453 if (!map->nr_queues) {
454 BUG_ON(i == HCTX_TYPE_DEFAULT);
459 * The poll queue(s) doesn't have an IRQ (and hence IRQ
460 * affinity), so use the regular blk-mq cpu mapping
462 map->queue_offset = qoff;
463 if (i != HCTX_TYPE_POLL && offset)
464 blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
466 blk_mq_map_queues(map);
467 qoff += map->nr_queues;
468 offset += map->nr_queues;
475 * Write sq tail if we are asked to, or if the next command would wrap.
477 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
480 u16 next_tail = nvmeq->sq_tail + 1;
482 if (next_tail == nvmeq->q_depth)
484 if (next_tail != nvmeq->last_sq_tail)
488 if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
489 nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
490 writel(nvmeq->sq_tail, nvmeq->q_db);
491 nvmeq->last_sq_tail = nvmeq->sq_tail;
495 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
496 * @nvmeq: The queue to use
497 * @cmd: The command to send
498 * @write_sq: whether to write to the SQ doorbell
500 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
503 spin_lock(&nvmeq->sq_lock);
504 memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
506 if (++nvmeq->sq_tail == nvmeq->q_depth)
508 nvme_write_sq_db(nvmeq, write_sq);
509 spin_unlock(&nvmeq->sq_lock);
512 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
514 struct nvme_queue *nvmeq = hctx->driver_data;
516 spin_lock(&nvmeq->sq_lock);
517 if (nvmeq->sq_tail != nvmeq->last_sq_tail)
518 nvme_write_sq_db(nvmeq, true);
519 spin_unlock(&nvmeq->sq_lock);
522 static void **nvme_pci_iod_list(struct request *req)
524 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
525 return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
528 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
530 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
531 int nseg = blk_rq_nr_phys_segments(req);
532 unsigned int avg_seg_size;
534 avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
536 if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
538 if (!iod->nvmeq->qid)
540 if (!sgl_threshold || avg_seg_size < sgl_threshold)
545 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
547 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
548 const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
549 dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
553 dma_unmap_page(dev->dev, dma_addr, iod->dma_len,
558 WARN_ON_ONCE(!iod->nents);
560 if (is_pci_p2pdma_page(sg_page(iod->sg)))
561 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
564 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
567 if (iod->npages == 0)
568 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
571 for (i = 0; i < iod->npages; i++) {
572 void *addr = nvme_pci_iod_list(req)[i];
575 struct nvme_sgl_desc *sg_list = addr;
578 le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
580 __le64 *prp_list = addr;
582 next_dma_addr = le64_to_cpu(prp_list[last_prp]);
585 dma_pool_free(dev->prp_page_pool, addr, dma_addr);
586 dma_addr = next_dma_addr;
589 mempool_free(iod->sg, dev->iod_mempool);
592 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
595 struct scatterlist *sg;
597 for_each_sg(sgl, sg, nents, i) {
598 dma_addr_t phys = sg_phys(sg);
599 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
600 "dma_address:%pad dma_length:%d\n",
601 i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
606 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
607 struct request *req, struct nvme_rw_command *cmnd)
609 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
610 struct dma_pool *pool;
611 int length = blk_rq_payload_bytes(req);
612 struct scatterlist *sg = iod->sg;
613 int dma_len = sg_dma_len(sg);
614 u64 dma_addr = sg_dma_address(sg);
615 int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
617 void **list = nvme_pci_iod_list(req);
621 length -= (NVME_CTRL_PAGE_SIZE - offset);
627 dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
629 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
632 dma_addr = sg_dma_address(sg);
633 dma_len = sg_dma_len(sg);
636 if (length <= NVME_CTRL_PAGE_SIZE) {
637 iod->first_dma = dma_addr;
641 nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
642 if (nprps <= (256 / 8)) {
643 pool = dev->prp_small_pool;
646 pool = dev->prp_page_pool;
650 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
652 iod->first_dma = dma_addr;
654 return BLK_STS_RESOURCE;
657 iod->first_dma = prp_dma;
660 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
661 __le64 *old_prp_list = prp_list;
662 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
664 return BLK_STS_RESOURCE;
665 list[iod->npages++] = prp_list;
666 prp_list[0] = old_prp_list[i - 1];
667 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
670 prp_list[i++] = cpu_to_le64(dma_addr);
671 dma_len -= NVME_CTRL_PAGE_SIZE;
672 dma_addr += NVME_CTRL_PAGE_SIZE;
673 length -= NVME_CTRL_PAGE_SIZE;
678 if (unlikely(dma_len < 0))
681 dma_addr = sg_dma_address(sg);
682 dma_len = sg_dma_len(sg);
686 cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
687 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
692 WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
693 "Invalid SGL for payload:%d nents:%d\n",
694 blk_rq_payload_bytes(req), iod->nents);
695 return BLK_STS_IOERR;
698 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
699 struct scatterlist *sg)
701 sge->addr = cpu_to_le64(sg_dma_address(sg));
702 sge->length = cpu_to_le32(sg_dma_len(sg));
703 sge->type = NVME_SGL_FMT_DATA_DESC << 4;
706 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
707 dma_addr_t dma_addr, int entries)
709 sge->addr = cpu_to_le64(dma_addr);
710 if (entries < SGES_PER_PAGE) {
711 sge->length = cpu_to_le32(entries * sizeof(*sge));
712 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
714 sge->length = cpu_to_le32(PAGE_SIZE);
715 sge->type = NVME_SGL_FMT_SEG_DESC << 4;
719 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
720 struct request *req, struct nvme_rw_command *cmd, int entries)
722 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
723 struct dma_pool *pool;
724 struct nvme_sgl_desc *sg_list;
725 struct scatterlist *sg = iod->sg;
729 /* setting the transfer type as SGL */
730 cmd->flags = NVME_CMD_SGL_METABUF;
733 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
737 if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
738 pool = dev->prp_small_pool;
741 pool = dev->prp_page_pool;
745 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
748 return BLK_STS_RESOURCE;
751 nvme_pci_iod_list(req)[0] = sg_list;
752 iod->first_dma = sgl_dma;
754 nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
757 if (i == SGES_PER_PAGE) {
758 struct nvme_sgl_desc *old_sg_desc = sg_list;
759 struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
761 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
763 return BLK_STS_RESOURCE;
766 nvme_pci_iod_list(req)[iod->npages++] = sg_list;
767 sg_list[i++] = *link;
768 nvme_pci_sgl_set_seg(link, sgl_dma, entries);
771 nvme_pci_sgl_set_data(&sg_list[i++], sg);
773 } while (--entries > 0);
778 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
779 struct request *req, struct nvme_rw_command *cmnd,
782 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
783 unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
784 unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
786 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
787 if (dma_mapping_error(dev->dev, iod->first_dma))
788 return BLK_STS_RESOURCE;
789 iod->dma_len = bv->bv_len;
791 cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
792 if (bv->bv_len > first_prp_len)
793 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
797 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
798 struct request *req, struct nvme_rw_command *cmnd,
801 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
803 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
804 if (dma_mapping_error(dev->dev, iod->first_dma))
805 return BLK_STS_RESOURCE;
806 iod->dma_len = bv->bv_len;
808 cmnd->flags = NVME_CMD_SGL_METABUF;
809 cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
810 cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
811 cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
815 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
816 struct nvme_command *cmnd)
818 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
819 blk_status_t ret = BLK_STS_RESOURCE;
822 if (blk_rq_nr_phys_segments(req) == 1) {
823 struct bio_vec bv = req_bvec(req);
825 if (!is_pci_p2pdma_page(bv.bv_page)) {
826 if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
827 return nvme_setup_prp_simple(dev, req,
830 if (iod->nvmeq->qid &&
831 dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
832 return nvme_setup_sgl_simple(dev, req,
838 iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
840 return BLK_STS_RESOURCE;
841 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
842 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
846 if (is_pci_p2pdma_page(sg_page(iod->sg)))
847 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
848 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
850 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
851 rq_dma_dir(req), DMA_ATTR_NO_WARN);
855 iod->use_sgl = nvme_pci_use_sgls(dev, req);
857 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
859 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
861 if (ret != BLK_STS_OK)
862 nvme_unmap_data(dev, req);
866 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
867 struct nvme_command *cmnd)
869 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
871 iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
873 if (dma_mapping_error(dev->dev, iod->meta_dma))
874 return BLK_STS_IOERR;
875 cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
880 * NOTE: ns is NULL when called on the admin queue.
882 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
883 const struct blk_mq_queue_data *bd)
885 struct nvme_ns *ns = hctx->queue->queuedata;
886 struct nvme_queue *nvmeq = hctx->driver_data;
887 struct nvme_dev *dev = nvmeq->dev;
888 struct request *req = bd->rq;
889 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
890 struct nvme_command cmnd;
898 * We should not need to do this, but we're still using this to
899 * ensure we can drain requests on a dying queue.
901 if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
902 return BLK_STS_IOERR;
904 ret = nvme_setup_cmd(ns, req, &cmnd);
908 if (blk_rq_nr_phys_segments(req)) {
909 ret = nvme_map_data(dev, req, &cmnd);
914 if (blk_integrity_rq(req)) {
915 ret = nvme_map_metadata(dev, req, &cmnd);
920 blk_mq_start_request(req);
921 nvme_submit_cmd(nvmeq, &cmnd, bd->last);
924 nvme_unmap_data(dev, req);
926 nvme_cleanup_cmd(req);
930 static void nvme_pci_complete_rq(struct request *req)
932 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
933 struct nvme_dev *dev = iod->nvmeq->dev;
935 if (blk_integrity_rq(req))
936 dma_unmap_page(dev->dev, iod->meta_dma,
937 rq_integrity_vec(req)->bv_len, rq_data_dir(req));
938 if (blk_rq_nr_phys_segments(req))
939 nvme_unmap_data(dev, req);
940 nvme_complete_rq(req);
943 /* We read the CQE phase first to check if the rest of the entry is valid */
944 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
946 struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
948 return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
951 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
953 u16 head = nvmeq->cq_head;
955 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
957 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
960 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
963 return nvmeq->dev->admin_tagset.tags[0];
964 return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
967 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
969 struct nvme_completion *cqe = &nvmeq->cqes[idx];
973 * AEN requests are special as they don't time out and can
974 * survive any kind of queue freeze and often don't respond to
975 * aborts. We don't even bother to allocate a struct request
976 * for them but rather special case them here.
978 if (unlikely(nvme_is_aen_req(nvmeq->qid, cqe->command_id))) {
979 nvme_complete_async_event(&nvmeq->dev->ctrl,
980 cqe->status, &cqe->result);
984 req = blk_mq_tag_to_rq(nvme_queue_tagset(nvmeq), cqe->command_id);
985 if (unlikely(!req)) {
986 dev_warn(nvmeq->dev->ctrl.device,
987 "invalid id %d completed on queue %d\n",
988 cqe->command_id, le16_to_cpu(cqe->sq_id));
992 trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
993 if (!nvme_try_complete_req(req, cqe->status, cqe->result))
994 nvme_pci_complete_rq(req);
997 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
999 u16 tmp = nvmeq->cq_head + 1;
1001 if (tmp == nvmeq->q_depth) {
1003 nvmeq->cq_phase ^= 1;
1005 nvmeq->cq_head = tmp;
1009 static inline int nvme_process_cq(struct nvme_queue *nvmeq)
1013 while (nvme_cqe_pending(nvmeq)) {
1016 * load-load control dependency between phase and the rest of
1017 * the cqe requires a full read memory barrier
1020 nvme_handle_cqe(nvmeq, nvmeq->cq_head);
1021 nvme_update_cq_head(nvmeq);
1025 nvme_ring_cq_doorbell(nvmeq);
1029 static irqreturn_t nvme_irq(int irq, void *data)
1031 struct nvme_queue *nvmeq = data;
1032 irqreturn_t ret = IRQ_NONE;
1035 * The rmb/wmb pair ensures we see all updates from a previous run of
1036 * the irq handler, even if that was on another CPU.
1039 if (nvme_process_cq(nvmeq))
1046 static irqreturn_t nvme_irq_check(int irq, void *data)
1048 struct nvme_queue *nvmeq = data;
1050 if (nvme_cqe_pending(nvmeq))
1051 return IRQ_WAKE_THREAD;
1056 * Poll for completions for any interrupt driven queue
1057 * Can be called from any context.
1059 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1061 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1063 WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1065 disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1066 nvme_process_cq(nvmeq);
1067 enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1070 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1072 struct nvme_queue *nvmeq = hctx->driver_data;
1075 if (!nvme_cqe_pending(nvmeq))
1078 spin_lock(&nvmeq->cq_poll_lock);
1079 found = nvme_process_cq(nvmeq);
1080 spin_unlock(&nvmeq->cq_poll_lock);
1085 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1087 struct nvme_dev *dev = to_nvme_dev(ctrl);
1088 struct nvme_queue *nvmeq = &dev->queues[0];
1089 struct nvme_command c;
1091 memset(&c, 0, sizeof(c));
1092 c.common.opcode = nvme_admin_async_event;
1093 c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1094 nvme_submit_cmd(nvmeq, &c, true);
1097 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1099 struct nvme_command c;
1101 memset(&c, 0, sizeof(c));
1102 c.delete_queue.opcode = opcode;
1103 c.delete_queue.qid = cpu_to_le16(id);
1105 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1108 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1109 struct nvme_queue *nvmeq, s16 vector)
1111 struct nvme_command c;
1112 int flags = NVME_QUEUE_PHYS_CONTIG;
1114 if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1115 flags |= NVME_CQ_IRQ_ENABLED;
1118 * Note: we (ab)use the fact that the prp fields survive if no data
1119 * is attached to the request.
1121 memset(&c, 0, sizeof(c));
1122 c.create_cq.opcode = nvme_admin_create_cq;
1123 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1124 c.create_cq.cqid = cpu_to_le16(qid);
1125 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1126 c.create_cq.cq_flags = cpu_to_le16(flags);
1127 c.create_cq.irq_vector = cpu_to_le16(vector);
1129 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1132 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1133 struct nvme_queue *nvmeq)
1135 struct nvme_ctrl *ctrl = &dev->ctrl;
1136 struct nvme_command c;
1137 int flags = NVME_QUEUE_PHYS_CONTIG;
1140 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1141 * set. Since URGENT priority is zeroes, it makes all queues
1144 if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1145 flags |= NVME_SQ_PRIO_MEDIUM;
1148 * Note: we (ab)use the fact that the prp fields survive if no data
1149 * is attached to the request.
1151 memset(&c, 0, sizeof(c));
1152 c.create_sq.opcode = nvme_admin_create_sq;
1153 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1154 c.create_sq.sqid = cpu_to_le16(qid);
1155 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1156 c.create_sq.sq_flags = cpu_to_le16(flags);
1157 c.create_sq.cqid = cpu_to_le16(qid);
1159 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1162 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1164 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1167 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1169 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1172 static void abort_endio(struct request *req, blk_status_t error)
1174 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1175 struct nvme_queue *nvmeq = iod->nvmeq;
1177 dev_warn(nvmeq->dev->ctrl.device,
1178 "Abort status: 0x%x", nvme_req(req)->status);
1179 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1180 blk_mq_free_request(req);
1183 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1185 /* If true, indicates loss of adapter communication, possibly by a
1186 * NVMe Subsystem reset.
1188 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1190 /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1191 switch (dev->ctrl.state) {
1192 case NVME_CTRL_RESETTING:
1193 case NVME_CTRL_CONNECTING:
1199 /* We shouldn't reset unless the controller is on fatal error state
1200 * _or_ if we lost the communication with it.
1202 if (!(csts & NVME_CSTS_CFS) && !nssro)
1208 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1210 /* Read a config register to help see what died. */
1214 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1216 if (result == PCIBIOS_SUCCESSFUL)
1217 dev_warn(dev->ctrl.device,
1218 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1221 dev_warn(dev->ctrl.device,
1222 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1226 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1228 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1229 struct nvme_queue *nvmeq = iod->nvmeq;
1230 struct nvme_dev *dev = nvmeq->dev;
1231 struct request *abort_req;
1232 struct nvme_command cmd;
1233 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1235 /* If PCI error recovery process is happening, we cannot reset or
1236 * the recovery mechanism will surely fail.
1239 if (pci_channel_offline(to_pci_dev(dev->dev)))
1240 return BLK_EH_RESET_TIMER;
1243 * Reset immediately if the controller is failed
1245 if (nvme_should_reset(dev, csts)) {
1246 nvme_warn_reset(dev, csts);
1247 nvme_dev_disable(dev, false);
1248 nvme_reset_ctrl(&dev->ctrl);
1253 * Did we miss an interrupt?
1255 if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1256 nvme_poll(req->mq_hctx);
1258 nvme_poll_irqdisable(nvmeq);
1260 if (blk_mq_request_completed(req)) {
1261 dev_warn(dev->ctrl.device,
1262 "I/O %d QID %d timeout, completion polled\n",
1263 req->tag, nvmeq->qid);
1268 * Shutdown immediately if controller times out while starting. The
1269 * reset work will see the pci device disabled when it gets the forced
1270 * cancellation error. All outstanding requests are completed on
1271 * shutdown, so we return BLK_EH_DONE.
1273 switch (dev->ctrl.state) {
1274 case NVME_CTRL_CONNECTING:
1275 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1277 case NVME_CTRL_DELETING:
1278 dev_warn_ratelimited(dev->ctrl.device,
1279 "I/O %d QID %d timeout, disable controller\n",
1280 req->tag, nvmeq->qid);
1281 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1282 nvme_dev_disable(dev, true);
1284 case NVME_CTRL_RESETTING:
1285 return BLK_EH_RESET_TIMER;
1291 * Shutdown the controller immediately and schedule a reset if the
1292 * command was already aborted once before and still hasn't been
1293 * returned to the driver, or if this is the admin queue.
1295 if (!nvmeq->qid || iod->aborted) {
1296 dev_warn(dev->ctrl.device,
1297 "I/O %d QID %d timeout, reset controller\n",
1298 req->tag, nvmeq->qid);
1299 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1300 nvme_dev_disable(dev, false);
1301 nvme_reset_ctrl(&dev->ctrl);
1306 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1307 atomic_inc(&dev->ctrl.abort_limit);
1308 return BLK_EH_RESET_TIMER;
1312 memset(&cmd, 0, sizeof(cmd));
1313 cmd.abort.opcode = nvme_admin_abort_cmd;
1314 cmd.abort.cid = req->tag;
1315 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1317 dev_warn(nvmeq->dev->ctrl.device,
1318 "I/O %d QID %d timeout, aborting\n",
1319 req->tag, nvmeq->qid);
1321 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1323 if (IS_ERR(abort_req)) {
1324 atomic_inc(&dev->ctrl.abort_limit);
1325 return BLK_EH_RESET_TIMER;
1328 abort_req->end_io_data = NULL;
1329 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1332 * The aborted req will be completed on receiving the abort req.
1333 * We enable the timer again. If hit twice, it'll cause a device reset,
1334 * as the device then is in a faulty state.
1336 return BLK_EH_RESET_TIMER;
1339 static void nvme_free_queue(struct nvme_queue *nvmeq)
1341 dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1342 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1343 if (!nvmeq->sq_cmds)
1346 if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1347 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1348 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1350 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1351 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1355 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1359 for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1360 dev->ctrl.queue_count--;
1361 nvme_free_queue(&dev->queues[i]);
1366 * nvme_suspend_queue - put queue into suspended state
1367 * @nvmeq: queue to suspend
1369 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1371 if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1374 /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1377 nvmeq->dev->online_queues--;
1378 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1379 blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1380 if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1381 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1385 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1389 for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1390 nvme_suspend_queue(&dev->queues[i]);
1393 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1395 struct nvme_queue *nvmeq = &dev->queues[0];
1398 nvme_shutdown_ctrl(&dev->ctrl);
1400 nvme_disable_ctrl(&dev->ctrl);
1402 nvme_poll_irqdisable(nvmeq);
1406 * Called only on a device that has been disabled and after all other threads
1407 * that can check this device's completion queues have synced, except
1408 * nvme_poll(). This is the last chance for the driver to see a natural
1409 * completion before nvme_cancel_request() terminates all incomplete requests.
1411 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1415 for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1416 spin_lock(&dev->queues[i].cq_poll_lock);
1417 nvme_process_cq(&dev->queues[i]);
1418 spin_unlock(&dev->queues[i].cq_poll_lock);
1422 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1425 int q_depth = dev->q_depth;
1426 unsigned q_size_aligned = roundup(q_depth * entry_size,
1427 NVME_CTRL_PAGE_SIZE);
1429 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1430 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1432 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1433 q_depth = div_u64(mem_per_q, entry_size);
1436 * Ensure the reduced q_depth is above some threshold where it
1437 * would be better to map queues in system memory with the
1447 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1450 struct pci_dev *pdev = to_pci_dev(dev->dev);
1452 if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1453 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1454 if (nvmeq->sq_cmds) {
1455 nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1457 if (nvmeq->sq_dma_addr) {
1458 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1462 pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1466 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1467 &nvmeq->sq_dma_addr, GFP_KERNEL);
1468 if (!nvmeq->sq_cmds)
1473 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1475 struct nvme_queue *nvmeq = &dev->queues[qid];
1477 if (dev->ctrl.queue_count > qid)
1480 nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1481 nvmeq->q_depth = depth;
1482 nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1483 &nvmeq->cq_dma_addr, GFP_KERNEL);
1487 if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1491 spin_lock_init(&nvmeq->sq_lock);
1492 spin_lock_init(&nvmeq->cq_poll_lock);
1494 nvmeq->cq_phase = 1;
1495 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1497 dev->ctrl.queue_count++;
1502 dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1503 nvmeq->cq_dma_addr);
1508 static int queue_request_irq(struct nvme_queue *nvmeq)
1510 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1511 int nr = nvmeq->dev->ctrl.instance;
1513 if (use_threaded_interrupts) {
1514 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1515 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1517 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1518 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1522 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1524 struct nvme_dev *dev = nvmeq->dev;
1527 nvmeq->last_sq_tail = 0;
1529 nvmeq->cq_phase = 1;
1530 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1531 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1532 nvme_dbbuf_init(dev, nvmeq, qid);
1533 dev->online_queues++;
1534 wmb(); /* ensure the first interrupt sees the initialization */
1537 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1539 struct nvme_dev *dev = nvmeq->dev;
1543 clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1546 * A queue's vector matches the queue identifier unless the controller
1547 * has only one vector available.
1550 vector = dev->num_vecs == 1 ? 0 : qid;
1552 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1554 result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1558 result = adapter_alloc_sq(dev, qid, nvmeq);
1564 nvmeq->cq_vector = vector;
1565 nvme_init_queue(nvmeq, qid);
1568 result = queue_request_irq(nvmeq);
1573 set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1577 dev->online_queues--;
1578 adapter_delete_sq(dev, qid);
1580 adapter_delete_cq(dev, qid);
1584 static const struct blk_mq_ops nvme_mq_admin_ops = {
1585 .queue_rq = nvme_queue_rq,
1586 .complete = nvme_pci_complete_rq,
1587 .init_hctx = nvme_admin_init_hctx,
1588 .init_request = nvme_init_request,
1589 .timeout = nvme_timeout,
1592 static const struct blk_mq_ops nvme_mq_ops = {
1593 .queue_rq = nvme_queue_rq,
1594 .complete = nvme_pci_complete_rq,
1595 .commit_rqs = nvme_commit_rqs,
1596 .init_hctx = nvme_init_hctx,
1597 .init_request = nvme_init_request,
1598 .map_queues = nvme_pci_map_queues,
1599 .timeout = nvme_timeout,
1603 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1605 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1607 * If the controller was reset during removal, it's possible
1608 * user requests may be waiting on a stopped queue. Start the
1609 * queue to flush these to completion.
1611 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1612 blk_cleanup_queue(dev->ctrl.admin_q);
1613 blk_mq_free_tag_set(&dev->admin_tagset);
1617 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1619 if (!dev->ctrl.admin_q) {
1620 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1621 dev->admin_tagset.nr_hw_queues = 1;
1623 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1624 dev->admin_tagset.timeout = NVME_ADMIN_TIMEOUT;
1625 dev->admin_tagset.numa_node = dev->ctrl.numa_node;
1626 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1627 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1628 dev->admin_tagset.driver_data = dev;
1630 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1632 dev->ctrl.admin_tagset = &dev->admin_tagset;
1634 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1635 if (IS_ERR(dev->ctrl.admin_q)) {
1636 blk_mq_free_tag_set(&dev->admin_tagset);
1639 if (!blk_get_queue(dev->ctrl.admin_q)) {
1640 nvme_dev_remove_admin(dev);
1641 dev->ctrl.admin_q = NULL;
1645 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1650 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1652 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1655 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1657 struct pci_dev *pdev = to_pci_dev(dev->dev);
1659 if (size <= dev->bar_mapped_size)
1661 if (size > pci_resource_len(pdev, 0))
1665 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1667 dev->bar_mapped_size = 0;
1670 dev->bar_mapped_size = size;
1671 dev->dbs = dev->bar + NVME_REG_DBS;
1676 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1680 struct nvme_queue *nvmeq;
1682 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1686 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1687 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1689 if (dev->subsystem &&
1690 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1691 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1693 result = nvme_disable_ctrl(&dev->ctrl);
1697 result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1701 dev->ctrl.numa_node = dev_to_node(dev->dev);
1703 nvmeq = &dev->queues[0];
1704 aqa = nvmeq->q_depth - 1;
1707 writel(aqa, dev->bar + NVME_REG_AQA);
1708 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1709 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1711 result = nvme_enable_ctrl(&dev->ctrl);
1715 nvmeq->cq_vector = 0;
1716 nvme_init_queue(nvmeq, 0);
1717 result = queue_request_irq(nvmeq);
1719 dev->online_queues--;
1723 set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1727 static int nvme_create_io_queues(struct nvme_dev *dev)
1729 unsigned i, max, rw_queues;
1732 for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1733 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1739 max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1740 if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1741 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1742 dev->io_queues[HCTX_TYPE_READ];
1747 for (i = dev->online_queues; i <= max; i++) {
1748 bool polled = i > rw_queues;
1750 ret = nvme_create_queue(&dev->queues[i], i, polled);
1756 * Ignore failing Create SQ/CQ commands, we can continue with less
1757 * than the desired amount of queues, and even a controller without
1758 * I/O queues can still be used to issue admin commands. This might
1759 * be useful to upgrade a buggy firmware for example.
1761 return ret >= 0 ? 0 : ret;
1764 static ssize_t nvme_cmb_show(struct device *dev,
1765 struct device_attribute *attr,
1768 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1770 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1771 ndev->cmbloc, ndev->cmbsz);
1773 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1775 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1777 u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1779 return 1ULL << (12 + 4 * szu);
1782 static u32 nvme_cmb_size(struct nvme_dev *dev)
1784 return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1787 static void nvme_map_cmb(struct nvme_dev *dev)
1790 resource_size_t bar_size;
1791 struct pci_dev *pdev = to_pci_dev(dev->dev);
1797 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1800 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1802 size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1803 offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1804 bar = NVME_CMB_BIR(dev->cmbloc);
1805 bar_size = pci_resource_len(pdev, bar);
1807 if (offset > bar_size)
1811 * Controllers may support a CMB size larger than their BAR,
1812 * for example, due to being behind a bridge. Reduce the CMB to
1813 * the reported size of the BAR
1815 if (size > bar_size - offset)
1816 size = bar_size - offset;
1818 if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1819 dev_warn(dev->ctrl.device,
1820 "failed to register the CMB\n");
1824 dev->cmb_size = size;
1825 dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1827 if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1828 (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1829 pci_p2pmem_publish(pdev, true);
1831 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1832 &dev_attr_cmb.attr, NULL))
1833 dev_warn(dev->ctrl.device,
1834 "failed to add sysfs attribute for CMB\n");
1837 static inline void nvme_release_cmb(struct nvme_dev *dev)
1839 if (dev->cmb_size) {
1840 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1841 &dev_attr_cmb.attr, NULL);
1846 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1848 u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1849 u64 dma_addr = dev->host_mem_descs_dma;
1850 struct nvme_command c;
1853 memset(&c, 0, sizeof(c));
1854 c.features.opcode = nvme_admin_set_features;
1855 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1856 c.features.dword11 = cpu_to_le32(bits);
1857 c.features.dword12 = cpu_to_le32(host_mem_size);
1858 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1859 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1860 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1862 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1864 dev_warn(dev->ctrl.device,
1865 "failed to set host mem (err %d, flags %#x).\n",
1871 static void nvme_free_host_mem(struct nvme_dev *dev)
1875 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1876 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1877 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
1879 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1880 le64_to_cpu(desc->addr),
1881 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1884 kfree(dev->host_mem_desc_bufs);
1885 dev->host_mem_desc_bufs = NULL;
1886 dma_free_coherent(dev->dev,
1887 dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1888 dev->host_mem_descs, dev->host_mem_descs_dma);
1889 dev->host_mem_descs = NULL;
1890 dev->nr_host_mem_descs = 0;
1893 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1896 struct nvme_host_mem_buf_desc *descs;
1897 u32 max_entries, len;
1898 dma_addr_t descs_dma;
1903 tmp = (preferred + chunk_size - 1);
1904 do_div(tmp, chunk_size);
1907 if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1908 max_entries = dev->ctrl.hmmaxd;
1910 descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1911 &descs_dma, GFP_KERNEL);
1915 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1917 goto out_free_descs;
1919 for (size = 0; size < preferred && i < max_entries; size += len) {
1920 dma_addr_t dma_addr;
1922 len = min_t(u64, chunk_size, preferred - size);
1923 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1924 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1928 descs[i].addr = cpu_to_le64(dma_addr);
1929 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
1936 dev->nr_host_mem_descs = i;
1937 dev->host_mem_size = size;
1938 dev->host_mem_descs = descs;
1939 dev->host_mem_descs_dma = descs_dma;
1940 dev->host_mem_desc_bufs = bufs;
1945 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
1947 dma_free_attrs(dev->dev, size, bufs[i],
1948 le64_to_cpu(descs[i].addr),
1949 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1954 dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1957 dev->host_mem_descs = NULL;
1961 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1963 u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1964 u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1967 /* start big and work our way down */
1968 for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
1969 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1970 if (!min || dev->host_mem_size >= min)
1972 nvme_free_host_mem(dev);
1979 static int nvme_setup_host_mem(struct nvme_dev *dev)
1981 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1982 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1983 u64 min = (u64)dev->ctrl.hmmin * 4096;
1984 u32 enable_bits = NVME_HOST_MEM_ENABLE;
1987 preferred = min(preferred, max);
1989 dev_warn(dev->ctrl.device,
1990 "min host memory (%lld MiB) above limit (%d MiB).\n",
1991 min >> ilog2(SZ_1M), max_host_mem_size_mb);
1992 nvme_free_host_mem(dev);
1997 * If we already have a buffer allocated check if we can reuse it.
1999 if (dev->host_mem_descs) {
2000 if (dev->host_mem_size >= min)
2001 enable_bits |= NVME_HOST_MEM_RETURN;
2003 nvme_free_host_mem(dev);
2006 if (!dev->host_mem_descs) {
2007 if (nvme_alloc_host_mem(dev, min, preferred)) {
2008 dev_warn(dev->ctrl.device,
2009 "failed to allocate host memory buffer.\n");
2010 return 0; /* controller must work without HMB */
2013 dev_info(dev->ctrl.device,
2014 "allocated %lld MiB host memory buffer.\n",
2015 dev->host_mem_size >> ilog2(SZ_1M));
2018 ret = nvme_set_host_mem(dev, enable_bits);
2020 nvme_free_host_mem(dev);
2025 * nirqs is the number of interrupts available for write and read
2026 * queues. The core already reserved an interrupt for the admin queue.
2028 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2030 struct nvme_dev *dev = affd->priv;
2031 unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2034 * If there is no interrupt available for queues, ensure that
2035 * the default queue is set to 1. The affinity set size is
2036 * also set to one, but the irq core ignores it for this case.
2038 * If only one interrupt is available or 'write_queue' == 0, combine
2039 * write and read queues.
2041 * If 'write_queues' > 0, ensure it leaves room for at least one read
2047 } else if (nrirqs == 1 || !nr_write_queues) {
2049 } else if (nr_write_queues >= nrirqs) {
2052 nr_read_queues = nrirqs - nr_write_queues;
2055 dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2056 affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2057 dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2058 affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2059 affd->nr_sets = nr_read_queues ? 2 : 1;
2062 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2064 struct pci_dev *pdev = to_pci_dev(dev->dev);
2065 struct irq_affinity affd = {
2067 .calc_sets = nvme_calc_irq_sets,
2070 unsigned int irq_queues, poll_queues;
2073 * Poll queues don't need interrupts, but we need at least one I/O queue
2074 * left over for non-polled I/O.
2076 poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2077 dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2080 * Initialize for the single interrupt case, will be updated in
2081 * nvme_calc_irq_sets().
2083 dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2084 dev->io_queues[HCTX_TYPE_READ] = 0;
2087 * We need interrupts for the admin queue and each non-polled I/O queue,
2088 * but some Apple controllers require all queues to use the first
2092 if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2093 irq_queues += (nr_io_queues - poll_queues);
2094 return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2095 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2098 static void nvme_disable_io_queues(struct nvme_dev *dev)
2100 if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2101 __nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2104 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2107 * If tags are shared with admin queue (Apple bug), then
2108 * make sure we only use one IO queue.
2110 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2112 return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2115 static int nvme_setup_io_queues(struct nvme_dev *dev)
2117 struct nvme_queue *adminq = &dev->queues[0];
2118 struct pci_dev *pdev = to_pci_dev(dev->dev);
2119 unsigned int nr_io_queues;
2124 * Sample the module parameters once at reset time so that we have
2125 * stable values to work with.
2127 dev->nr_write_queues = write_queues;
2128 dev->nr_poll_queues = poll_queues;
2130 nr_io_queues = dev->nr_allocated_queues - 1;
2131 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2135 if (nr_io_queues == 0)
2138 clear_bit(NVMEQ_ENABLED, &adminq->flags);
2140 if (dev->cmb_use_sqes) {
2141 result = nvme_cmb_qdepth(dev, nr_io_queues,
2142 sizeof(struct nvme_command));
2144 dev->q_depth = result;
2146 dev->cmb_use_sqes = false;
2150 size = db_bar_size(dev, nr_io_queues);
2151 result = nvme_remap_bar(dev, size);
2154 if (!--nr_io_queues)
2157 adminq->q_db = dev->dbs;
2160 /* Deregister the admin queue's interrupt */
2161 pci_free_irq(pdev, 0, adminq);
2164 * If we enable msix early due to not intx, disable it again before
2165 * setting up the full range we need.
2167 pci_free_irq_vectors(pdev);
2169 result = nvme_setup_irqs(dev, nr_io_queues);
2173 dev->num_vecs = result;
2174 result = max(result - 1, 1);
2175 dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2178 * Should investigate if there's a performance win from allocating
2179 * more queues than interrupt vectors; it might allow the submission
2180 * path to scale better, even if the receive path is limited by the
2181 * number of interrupts.
2183 result = queue_request_irq(adminq);
2186 set_bit(NVMEQ_ENABLED, &adminq->flags);
2188 result = nvme_create_io_queues(dev);
2189 if (result || dev->online_queues < 2)
2192 if (dev->online_queues - 1 < dev->max_qid) {
2193 nr_io_queues = dev->online_queues - 1;
2194 nvme_disable_io_queues(dev);
2195 nvme_suspend_io_queues(dev);
2198 dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2199 dev->io_queues[HCTX_TYPE_DEFAULT],
2200 dev->io_queues[HCTX_TYPE_READ],
2201 dev->io_queues[HCTX_TYPE_POLL]);
2205 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2207 struct nvme_queue *nvmeq = req->end_io_data;
2209 blk_mq_free_request(req);
2210 complete(&nvmeq->delete_done);
2213 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2215 struct nvme_queue *nvmeq = req->end_io_data;
2218 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2220 nvme_del_queue_end(req, error);
2223 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2225 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2226 struct request *req;
2227 struct nvme_command cmd;
2229 memset(&cmd, 0, sizeof(cmd));
2230 cmd.delete_queue.opcode = opcode;
2231 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2233 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT);
2235 return PTR_ERR(req);
2237 req->end_io_data = nvmeq;
2239 init_completion(&nvmeq->delete_done);
2240 blk_execute_rq_nowait(q, NULL, req, false,
2241 opcode == nvme_admin_delete_cq ?
2242 nvme_del_cq_end : nvme_del_queue_end);
2246 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2248 int nr_queues = dev->online_queues - 1, sent = 0;
2249 unsigned long timeout;
2252 timeout = NVME_ADMIN_TIMEOUT;
2253 while (nr_queues > 0) {
2254 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2260 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2262 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2274 static void nvme_dev_add(struct nvme_dev *dev)
2278 if (!dev->ctrl.tagset) {
2279 dev->tagset.ops = &nvme_mq_ops;
2280 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2281 dev->tagset.nr_maps = 2; /* default + read */
2282 if (dev->io_queues[HCTX_TYPE_POLL])
2283 dev->tagset.nr_maps++;
2284 dev->tagset.timeout = NVME_IO_TIMEOUT;
2285 dev->tagset.numa_node = dev->ctrl.numa_node;
2286 dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth,
2287 BLK_MQ_MAX_DEPTH) - 1;
2288 dev->tagset.cmd_size = sizeof(struct nvme_iod);
2289 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2290 dev->tagset.driver_data = dev;
2293 * Some Apple controllers requires tags to be unique
2294 * across admin and IO queue, so reserve the first 32
2295 * tags of the IO queue.
2297 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2298 dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2300 ret = blk_mq_alloc_tag_set(&dev->tagset);
2302 dev_warn(dev->ctrl.device,
2303 "IO queues tagset allocation failed %d\n", ret);
2306 dev->ctrl.tagset = &dev->tagset;
2308 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2310 /* Free previously allocated queues that are no longer usable */
2311 nvme_free_queues(dev, dev->online_queues);
2314 nvme_dbbuf_set(dev);
2317 static int nvme_pci_enable(struct nvme_dev *dev)
2319 int result = -ENOMEM;
2320 struct pci_dev *pdev = to_pci_dev(dev->dev);
2322 if (pci_enable_device_mem(pdev))
2325 pci_set_master(pdev);
2327 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2330 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2336 * Some devices and/or platforms don't advertise or work with INTx
2337 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2338 * adjust this later.
2340 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2344 dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2346 dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2348 dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2349 dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2350 dev->dbs = dev->bar + 4096;
2353 * Some Apple controllers require a non-standard SQE size.
2354 * Interestingly they also seem to ignore the CC:IOSQES register
2355 * so we don't bother updating it here.
2357 if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2360 dev->io_sqes = NVME_NVM_IOSQES;
2363 * Temporary fix for the Apple controller found in the MacBook8,1 and
2364 * some MacBook7,1 to avoid controller resets and data loss.
2366 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2368 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2369 "set queue depth=%u to work around controller resets\n",
2371 } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2372 (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2373 NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2375 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2376 "set queue depth=%u\n", dev->q_depth);
2380 * Controllers with the shared tags quirk need the IO queue to be
2381 * big enough so that we get 32 tags for the admin queue
2383 if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2384 (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2385 dev->q_depth = NVME_AQ_DEPTH + 2;
2386 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2393 pci_enable_pcie_error_reporting(pdev);
2394 pci_save_state(pdev);
2398 pci_disable_device(pdev);
2402 static void nvme_dev_unmap(struct nvme_dev *dev)
2406 pci_release_mem_regions(to_pci_dev(dev->dev));
2409 static void nvme_pci_disable(struct nvme_dev *dev)
2411 struct pci_dev *pdev = to_pci_dev(dev->dev);
2413 pci_free_irq_vectors(pdev);
2415 if (pci_is_enabled(pdev)) {
2416 pci_disable_pcie_error_reporting(pdev);
2417 pci_disable_device(pdev);
2421 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2423 bool dead = true, freeze = false;
2424 struct pci_dev *pdev = to_pci_dev(dev->dev);
2426 mutex_lock(&dev->shutdown_lock);
2427 if (pci_is_enabled(pdev)) {
2428 u32 csts = readl(dev->bar + NVME_REG_CSTS);
2430 if (dev->ctrl.state == NVME_CTRL_LIVE ||
2431 dev->ctrl.state == NVME_CTRL_RESETTING) {
2433 nvme_start_freeze(&dev->ctrl);
2435 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2436 pdev->error_state != pci_channel_io_normal);
2440 * Give the controller a chance to complete all entered requests if
2441 * doing a safe shutdown.
2443 if (!dead && shutdown && freeze)
2444 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2446 nvme_stop_queues(&dev->ctrl);
2448 if (!dead && dev->ctrl.queue_count > 0) {
2449 nvme_disable_io_queues(dev);
2450 nvme_disable_admin_queue(dev, shutdown);
2452 nvme_suspend_io_queues(dev);
2453 nvme_suspend_queue(&dev->queues[0]);
2454 nvme_pci_disable(dev);
2455 nvme_reap_pending_cqes(dev);
2457 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2458 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2459 blk_mq_tagset_wait_completed_request(&dev->tagset);
2460 blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2463 * The driver will not be starting up queues again if shutting down so
2464 * must flush all entered requests to their failed completion to avoid
2465 * deadlocking blk-mq hot-cpu notifier.
2468 nvme_start_queues(&dev->ctrl);
2469 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2470 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2472 mutex_unlock(&dev->shutdown_lock);
2475 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2477 if (!nvme_wait_reset(&dev->ctrl))
2479 nvme_dev_disable(dev, shutdown);
2483 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2485 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2486 NVME_CTRL_PAGE_SIZE,
2487 NVME_CTRL_PAGE_SIZE, 0);
2488 if (!dev->prp_page_pool)
2491 /* Optimisation for I/Os between 4k and 128k */
2492 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2494 if (!dev->prp_small_pool) {
2495 dma_pool_destroy(dev->prp_page_pool);
2501 static void nvme_release_prp_pools(struct nvme_dev *dev)
2503 dma_pool_destroy(dev->prp_page_pool);
2504 dma_pool_destroy(dev->prp_small_pool);
2507 static void nvme_free_tagset(struct nvme_dev *dev)
2509 if (dev->tagset.tags)
2510 blk_mq_free_tag_set(&dev->tagset);
2511 dev->ctrl.tagset = NULL;
2514 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2516 struct nvme_dev *dev = to_nvme_dev(ctrl);
2518 nvme_dbbuf_dma_free(dev);
2519 nvme_free_tagset(dev);
2520 if (dev->ctrl.admin_q)
2521 blk_put_queue(dev->ctrl.admin_q);
2522 free_opal_dev(dev->ctrl.opal_dev);
2523 mempool_destroy(dev->iod_mempool);
2524 put_device(dev->dev);
2529 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2532 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2533 * may be holding this pci_dev's device lock.
2535 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2536 nvme_get_ctrl(&dev->ctrl);
2537 nvme_dev_disable(dev, false);
2538 nvme_kill_queues(&dev->ctrl);
2539 if (!queue_work(nvme_wq, &dev->remove_work))
2540 nvme_put_ctrl(&dev->ctrl);
2543 static void nvme_reset_work(struct work_struct *work)
2545 struct nvme_dev *dev =
2546 container_of(work, struct nvme_dev, ctrl.reset_work);
2547 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2550 if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2556 * If we're called to reset a live controller first shut it down before
2559 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2560 nvme_dev_disable(dev, false);
2561 nvme_sync_queues(&dev->ctrl);
2563 mutex_lock(&dev->shutdown_lock);
2564 result = nvme_pci_enable(dev);
2568 result = nvme_pci_configure_admin_queue(dev);
2572 result = nvme_alloc_admin_tags(dev);
2577 * Limit the max command size to prevent iod->sg allocations going
2578 * over a single page.
2580 dev->ctrl.max_hw_sectors = min_t(u32,
2581 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2582 dev->ctrl.max_segments = NVME_MAX_SEGS;
2585 * Don't limit the IOMMU merged segment size.
2587 dma_set_max_seg_size(dev->dev, 0xffffffff);
2589 mutex_unlock(&dev->shutdown_lock);
2592 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2593 * initializing procedure here.
2595 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2596 dev_warn(dev->ctrl.device,
2597 "failed to mark controller CONNECTING\n");
2603 * We do not support an SGL for metadata (yet), so we are limited to a
2604 * single integrity segment for the separate metadata pointer.
2606 dev->ctrl.max_integrity_segments = 1;
2608 result = nvme_init_identify(&dev->ctrl);
2612 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2613 if (!dev->ctrl.opal_dev)
2614 dev->ctrl.opal_dev =
2615 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2616 else if (was_suspend)
2617 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2619 free_opal_dev(dev->ctrl.opal_dev);
2620 dev->ctrl.opal_dev = NULL;
2623 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2624 result = nvme_dbbuf_dma_alloc(dev);
2627 "unable to allocate dma for dbbuf\n");
2630 if (dev->ctrl.hmpre) {
2631 result = nvme_setup_host_mem(dev);
2636 result = nvme_setup_io_queues(dev);
2641 * Keep the controller around but remove all namespaces if we don't have
2642 * any working I/O queue.
2644 if (dev->online_queues < 2) {
2645 dev_warn(dev->ctrl.device, "IO queues not created\n");
2646 nvme_kill_queues(&dev->ctrl);
2647 nvme_remove_namespaces(&dev->ctrl);
2648 nvme_free_tagset(dev);
2650 nvme_start_queues(&dev->ctrl);
2651 nvme_wait_freeze(&dev->ctrl);
2653 nvme_unfreeze(&dev->ctrl);
2657 * If only admin queue live, keep it to do further investigation or
2660 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2661 dev_warn(dev->ctrl.device,
2662 "failed to mark controller live state\n");
2667 nvme_start_ctrl(&dev->ctrl);
2671 mutex_unlock(&dev->shutdown_lock);
2674 dev_warn(dev->ctrl.device,
2675 "Removing after probe failure status: %d\n", result);
2676 nvme_remove_dead_ctrl(dev);
2679 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2681 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2682 struct pci_dev *pdev = to_pci_dev(dev->dev);
2684 if (pci_get_drvdata(pdev))
2685 device_release_driver(&pdev->dev);
2686 nvme_put_ctrl(&dev->ctrl);
2689 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2691 *val = readl(to_nvme_dev(ctrl)->bar + off);
2695 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2697 writel(val, to_nvme_dev(ctrl)->bar + off);
2701 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2703 *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2707 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2709 struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2711 return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2714 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2716 .module = THIS_MODULE,
2717 .flags = NVME_F_METADATA_SUPPORTED |
2719 .reg_read32 = nvme_pci_reg_read32,
2720 .reg_write32 = nvme_pci_reg_write32,
2721 .reg_read64 = nvme_pci_reg_read64,
2722 .free_ctrl = nvme_pci_free_ctrl,
2723 .submit_async_event = nvme_pci_submit_async_event,
2724 .get_address = nvme_pci_get_address,
2727 static int nvme_dev_map(struct nvme_dev *dev)
2729 struct pci_dev *pdev = to_pci_dev(dev->dev);
2731 if (pci_request_mem_regions(pdev, "nvme"))
2734 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2739 pci_release_mem_regions(pdev);
2743 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2745 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2747 * Several Samsung devices seem to drop off the PCIe bus
2748 * randomly when APST is on and uses the deepest sleep state.
2749 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2750 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2751 * 950 PRO 256GB", but it seems to be restricted to two Dell
2754 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2755 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2756 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2757 return NVME_QUIRK_NO_DEEPEST_PS;
2758 } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2760 * Samsung SSD 960 EVO drops off the PCIe bus after system
2761 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2762 * within few minutes after bootup on a Coffee Lake board -
2765 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2766 (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2767 dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2768 return NVME_QUIRK_NO_APST;
2769 } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2770 pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2771 (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2773 * Forcing to use host managed nvme power settings for
2774 * lowest idle power with quick resume latency on
2775 * Samsung and Toshiba SSDs based on suspend behavior
2776 * on Coffee Lake board for LENOVO C640
2778 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2779 dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2780 return NVME_QUIRK_SIMPLE_SUSPEND;
2787 static bool nvme_acpi_storage_d3(struct pci_dev *dev)
2789 struct acpi_device *adev;
2790 struct pci_dev *root;
2796 * Look for _DSD property specifying that the storage device on the port
2797 * must use D3 to support deep platform power savings during
2800 root = pcie_find_root_port(dev);
2804 adev = ACPI_COMPANION(&root->dev);
2809 * The property is defined in the PXSX device for South complex ports
2810 * and in the PEGP device for North complex ports.
2812 status = acpi_get_handle(adev->handle, "PXSX", &handle);
2813 if (ACPI_FAILURE(status)) {
2814 status = acpi_get_handle(adev->handle, "PEGP", &handle);
2815 if (ACPI_FAILURE(status))
2819 if (acpi_bus_get_device(handle, &adev))
2822 if (fwnode_property_read_u8(acpi_fwnode_handle(adev), "StorageD3Enable",
2828 static inline bool nvme_acpi_storage_d3(struct pci_dev *dev)
2832 #endif /* CONFIG_ACPI */
2834 static void nvme_async_probe(void *data, async_cookie_t cookie)
2836 struct nvme_dev *dev = data;
2838 flush_work(&dev->ctrl.reset_work);
2839 flush_work(&dev->ctrl.scan_work);
2840 nvme_put_ctrl(&dev->ctrl);
2843 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2845 int node, result = -ENOMEM;
2846 struct nvme_dev *dev;
2847 unsigned long quirks = id->driver_data;
2850 node = dev_to_node(&pdev->dev);
2851 if (node == NUMA_NO_NODE)
2852 set_dev_node(&pdev->dev, first_memory_node);
2854 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2858 dev->nr_write_queues = write_queues;
2859 dev->nr_poll_queues = poll_queues;
2860 dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
2861 dev->queues = kcalloc_node(dev->nr_allocated_queues,
2862 sizeof(struct nvme_queue), GFP_KERNEL, node);
2866 dev->dev = get_device(&pdev->dev);
2867 pci_set_drvdata(pdev, dev);
2869 result = nvme_dev_map(dev);
2873 INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2874 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2875 mutex_init(&dev->shutdown_lock);
2877 result = nvme_setup_prp_pools(dev);
2881 quirks |= check_vendor_combination_bug(pdev);
2883 if (!noacpi && nvme_acpi_storage_d3(pdev)) {
2885 * Some systems use a bios work around to ask for D3 on
2886 * platforms that support kernel managed suspend.
2888 dev_info(&pdev->dev,
2889 "platform quirk: setting simple suspend\n");
2890 quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
2894 * Double check that our mempool alloc size will cover the biggest
2895 * command we support.
2897 alloc_size = nvme_pci_iod_alloc_size();
2898 WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2900 dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2902 (void *) alloc_size,
2904 if (!dev->iod_mempool) {
2909 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2912 goto release_mempool;
2914 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2916 nvme_reset_ctrl(&dev->ctrl);
2917 async_schedule(nvme_async_probe, dev);
2922 mempool_destroy(dev->iod_mempool);
2924 nvme_release_prp_pools(dev);
2926 nvme_dev_unmap(dev);
2928 put_device(dev->dev);
2935 static void nvme_reset_prepare(struct pci_dev *pdev)
2937 struct nvme_dev *dev = pci_get_drvdata(pdev);
2940 * We don't need to check the return value from waiting for the reset
2941 * state as pci_dev device lock is held, making it impossible to race
2944 nvme_disable_prepare_reset(dev, false);
2945 nvme_sync_queues(&dev->ctrl);
2948 static void nvme_reset_done(struct pci_dev *pdev)
2950 struct nvme_dev *dev = pci_get_drvdata(pdev);
2952 if (!nvme_try_sched_reset(&dev->ctrl))
2953 flush_work(&dev->ctrl.reset_work);
2956 static void nvme_shutdown(struct pci_dev *pdev)
2958 struct nvme_dev *dev = pci_get_drvdata(pdev);
2960 nvme_disable_prepare_reset(dev, true);
2964 * The driver's remove may be called on a device in a partially initialized
2965 * state. This function must not have any dependencies on the device state in
2968 static void nvme_remove(struct pci_dev *pdev)
2970 struct nvme_dev *dev = pci_get_drvdata(pdev);
2972 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2973 pci_set_drvdata(pdev, NULL);
2975 if (!pci_device_is_present(pdev)) {
2976 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2977 nvme_dev_disable(dev, true);
2978 nvme_dev_remove_admin(dev);
2981 flush_work(&dev->ctrl.reset_work);
2982 nvme_stop_ctrl(&dev->ctrl);
2983 nvme_remove_namespaces(&dev->ctrl);
2984 nvme_dev_disable(dev, true);
2985 nvme_release_cmb(dev);
2986 nvme_free_host_mem(dev);
2987 nvme_dev_remove_admin(dev);
2988 nvme_free_queues(dev, 0);
2989 nvme_release_prp_pools(dev);
2990 nvme_dev_unmap(dev);
2991 nvme_uninit_ctrl(&dev->ctrl);
2994 #ifdef CONFIG_PM_SLEEP
2995 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
2997 return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3000 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3002 return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3005 static int nvme_resume(struct device *dev)
3007 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3008 struct nvme_ctrl *ctrl = &ndev->ctrl;
3010 if (ndev->last_ps == U32_MAX ||
3011 nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3012 return nvme_try_sched_reset(&ndev->ctrl);
3016 static int nvme_suspend(struct device *dev)
3018 struct pci_dev *pdev = to_pci_dev(dev);
3019 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3020 struct nvme_ctrl *ctrl = &ndev->ctrl;
3023 ndev->last_ps = U32_MAX;
3026 * The platform does not remove power for a kernel managed suspend so
3027 * use host managed nvme power settings for lowest idle power if
3028 * possible. This should have quicker resume latency than a full device
3029 * shutdown. But if the firmware is involved after the suspend or the
3030 * device does not support any non-default power states, shut down the
3033 * If ASPM is not enabled for the device, shut down the device and allow
3034 * the PCI bus layer to put it into D3 in order to take the PCIe link
3035 * down, so as to allow the platform to achieve its minimum low-power
3036 * state (which may not be possible if the link is up).
3038 * If a host memory buffer is enabled, shut down the device as the NVMe
3039 * specification allows the device to access the host memory buffer in
3040 * host DRAM from all power states, but hosts will fail access to DRAM
3043 if (pm_suspend_via_firmware() || !ctrl->npss ||
3044 !pcie_aspm_enabled(pdev) ||
3045 ndev->nr_host_mem_descs ||
3046 (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3047 return nvme_disable_prepare_reset(ndev, true);
3049 nvme_start_freeze(ctrl);
3050 nvme_wait_freeze(ctrl);
3051 nvme_sync_queues(ctrl);
3053 if (ctrl->state != NVME_CTRL_LIVE)
3056 ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3061 * A saved state prevents pci pm from generically controlling the
3062 * device's power. If we're using protocol specific settings, we don't
3063 * want pci interfering.
3065 pci_save_state(pdev);
3067 ret = nvme_set_power_state(ctrl, ctrl->npss);
3072 /* discard the saved state */
3073 pci_load_saved_state(pdev, NULL);
3076 * Clearing npss forces a controller reset on resume. The
3077 * correct value will be rediscovered then.
3079 ret = nvme_disable_prepare_reset(ndev, true);
3083 nvme_unfreeze(ctrl);
3087 static int nvme_simple_suspend(struct device *dev)
3089 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3091 return nvme_disable_prepare_reset(ndev, true);
3094 static int nvme_simple_resume(struct device *dev)
3096 struct pci_dev *pdev = to_pci_dev(dev);
3097 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3099 return nvme_try_sched_reset(&ndev->ctrl);
3102 static const struct dev_pm_ops nvme_dev_pm_ops = {
3103 .suspend = nvme_suspend,
3104 .resume = nvme_resume,
3105 .freeze = nvme_simple_suspend,
3106 .thaw = nvme_simple_resume,
3107 .poweroff = nvme_simple_suspend,
3108 .restore = nvme_simple_resume,
3110 #endif /* CONFIG_PM_SLEEP */
3112 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3113 pci_channel_state_t state)
3115 struct nvme_dev *dev = pci_get_drvdata(pdev);
3118 * A frozen channel requires a reset. When detected, this method will
3119 * shutdown the controller to quiesce. The controller will be restarted
3120 * after the slot reset through driver's slot_reset callback.
3123 case pci_channel_io_normal:
3124 return PCI_ERS_RESULT_CAN_RECOVER;
3125 case pci_channel_io_frozen:
3126 dev_warn(dev->ctrl.device,
3127 "frozen state error detected, reset controller\n");
3128 nvme_dev_disable(dev, false);
3129 return PCI_ERS_RESULT_NEED_RESET;
3130 case pci_channel_io_perm_failure:
3131 dev_warn(dev->ctrl.device,
3132 "failure state error detected, request disconnect\n");
3133 return PCI_ERS_RESULT_DISCONNECT;
3135 return PCI_ERS_RESULT_NEED_RESET;
3138 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3140 struct nvme_dev *dev = pci_get_drvdata(pdev);
3142 dev_info(dev->ctrl.device, "restart after slot reset\n");
3143 pci_restore_state(pdev);
3144 nvme_reset_ctrl(&dev->ctrl);
3145 return PCI_ERS_RESULT_RECOVERED;
3148 static void nvme_error_resume(struct pci_dev *pdev)
3150 struct nvme_dev *dev = pci_get_drvdata(pdev);
3152 flush_work(&dev->ctrl.reset_work);
3155 static const struct pci_error_handlers nvme_err_handler = {
3156 .error_detected = nvme_error_detected,
3157 .slot_reset = nvme_slot_reset,
3158 .resume = nvme_error_resume,
3159 .reset_prepare = nvme_reset_prepare,
3160 .reset_done = nvme_reset_done,
3163 static const struct pci_device_id nvme_id_table[] = {
3164 { PCI_VDEVICE(INTEL, 0x0953), /* Intel 750/P3500/P3600/P3700 */
3165 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3166 NVME_QUIRK_DEALLOCATE_ZEROES, },
3167 { PCI_VDEVICE(INTEL, 0x0a53), /* Intel P3520 */
3168 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3169 NVME_QUIRK_DEALLOCATE_ZEROES, },
3170 { PCI_VDEVICE(INTEL, 0x0a54), /* Intel P4500/P4600 */
3171 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3172 NVME_QUIRK_DEALLOCATE_ZEROES, },
3173 { PCI_VDEVICE(INTEL, 0x0a55), /* Dell Express Flash P4600 */
3174 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3175 NVME_QUIRK_DEALLOCATE_ZEROES, },
3176 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
3177 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3178 NVME_QUIRK_MEDIUM_PRIO_SQ |
3179 NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3180 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3181 { PCI_VDEVICE(INTEL, 0xf1a6), /* Intel 760p/Pro 7600p */
3182 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3183 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
3184 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3185 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3186 { PCI_DEVICE(0x126f, 0x2263), /* Silicon Motion unidentified */
3187 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST, },
3188 { PCI_DEVICE(0x1bb1, 0x0100), /* Seagate Nytro Flash Storage */
3189 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3190 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
3191 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3192 { PCI_DEVICE(0x1c58, 0x0023), /* WDC SN200 adapter */
3193 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3194 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
3195 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3196 { PCI_DEVICE(0x144d, 0xa821), /* Samsung PM1725 */
3197 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3198 { PCI_DEVICE(0x144d, 0xa822), /* Samsung PM1725a */
3199 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3200 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3201 { PCI_DEVICE(0x1d1d, 0x1f1f), /* LighNVM qemu device */
3202 .driver_data = NVME_QUIRK_LIGHTNVM, },
3203 { PCI_DEVICE(0x1d1d, 0x2807), /* CNEX WL */
3204 .driver_data = NVME_QUIRK_LIGHTNVM, },
3205 { PCI_DEVICE(0x1d1d, 0x2601), /* CNEX Granby */
3206 .driver_data = NVME_QUIRK_LIGHTNVM, },
3207 { PCI_DEVICE(0x10ec, 0x5762), /* ADATA SX6000LNP */
3208 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3209 { PCI_DEVICE(0x1cc1, 0x8201), /* ADATA SX8200PNP 512GB */
3210 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3211 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3212 { PCI_DEVICE(0x1c5c, 0x1504), /* SK Hynix PC400 */
3213 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3214 { PCI_DEVICE(0x15b7, 0x2001), /* Sandisk Skyhawk */
3215 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3216 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3217 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3218 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3219 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3220 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3221 NVME_QUIRK_128_BYTES_SQES |
3222 NVME_QUIRK_SHARED_TAGS },
3224 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3227 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3229 static struct pci_driver nvme_driver = {
3231 .id_table = nvme_id_table,
3232 .probe = nvme_probe,
3233 .remove = nvme_remove,
3234 .shutdown = nvme_shutdown,
3235 #ifdef CONFIG_PM_SLEEP
3237 .pm = &nvme_dev_pm_ops,
3240 .sriov_configure = pci_sriov_configure_simple,
3241 .err_handler = &nvme_err_handler,
3244 static int __init nvme_init(void)
3246 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3247 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3248 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3249 BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3251 return pci_register_driver(&nvme_driver);
3254 static void __exit nvme_exit(void)
3256 pci_unregister_driver(&nvme_driver);
3257 flush_workqueue(nvme_wq);
3260 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3261 MODULE_LICENSE("GPL");
3262 MODULE_VERSION("1.0");
3263 module_init(nvme_init);
3264 module_exit(nvme_exit);