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/io-64-nonatomic-hi-lo.h>
27 #include <linux/sed-opal.h>
28 #include <linux/pci-p2pdma.h>
33 #define SQ_SIZE(q) ((q)->q_depth << (q)->sqes)
34 #define CQ_SIZE(q) ((q)->q_depth * sizeof(struct nvme_completion))
36 #define SGES_PER_PAGE (PAGE_SIZE / sizeof(struct nvme_sgl_desc))
39 * These can be higher, but we need to ensure that any command doesn't
40 * require an sg allocation that needs more than a page of data.
42 #define NVME_MAX_KB_SZ 4096
43 #define NVME_MAX_SEGS 127
45 static int use_threaded_interrupts;
46 module_param(use_threaded_interrupts, int, 0);
48 static bool use_cmb_sqes = true;
49 module_param(use_cmb_sqes, bool, 0444);
50 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
52 static unsigned int max_host_mem_size_mb = 128;
53 module_param(max_host_mem_size_mb, uint, 0444);
54 MODULE_PARM_DESC(max_host_mem_size_mb,
55 "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
57 static unsigned int sgl_threshold = SZ_32K;
58 module_param(sgl_threshold, uint, 0644);
59 MODULE_PARM_DESC(sgl_threshold,
60 "Use SGLs when average request segment size is larger or equal to "
61 "this size. Use 0 to disable SGLs.");
63 static int io_queue_depth_set(const char *val, const struct kernel_param *kp);
64 static const struct kernel_param_ops io_queue_depth_ops = {
65 .set = io_queue_depth_set,
66 .get = param_get_uint,
69 static unsigned int io_queue_depth = 1024;
70 module_param_cb(io_queue_depth, &io_queue_depth_ops, &io_queue_depth, 0644);
71 MODULE_PARM_DESC(io_queue_depth, "set io queue depth, should >= 2");
73 static int io_queue_count_set(const char *val, const struct kernel_param *kp)
78 ret = kstrtouint(val, 10, &n);
79 if (ret != 0 || n > num_possible_cpus())
81 return param_set_uint(val, kp);
84 static const struct kernel_param_ops io_queue_count_ops = {
85 .set = io_queue_count_set,
86 .get = param_get_uint,
89 static unsigned int write_queues;
90 module_param_cb(write_queues, &io_queue_count_ops, &write_queues, 0644);
91 MODULE_PARM_DESC(write_queues,
92 "Number of queues to use for writes. If not set, reads and writes "
93 "will share a queue set.");
95 static unsigned int poll_queues;
96 module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644);
97 MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO.");
100 module_param(noacpi, bool, 0444);
101 MODULE_PARM_DESC(noacpi, "disable acpi bios quirks");
106 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
107 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode);
110 * Represents an NVM Express device. Each nvme_dev is a PCI function.
113 struct nvme_queue *queues;
114 struct blk_mq_tag_set tagset;
115 struct blk_mq_tag_set admin_tagset;
118 struct dma_pool *prp_page_pool;
119 struct dma_pool *prp_small_pool;
120 unsigned online_queues;
122 unsigned io_queues[HCTX_MAX_TYPES];
123 unsigned int num_vecs;
128 unsigned long bar_mapped_size;
129 struct work_struct remove_work;
130 struct mutex shutdown_lock;
136 struct nvme_ctrl ctrl;
139 mempool_t *iod_mempool;
141 /* shadow doorbell buffer support: */
143 dma_addr_t dbbuf_dbs_dma_addr;
145 dma_addr_t dbbuf_eis_dma_addr;
147 /* host memory buffer support: */
149 u32 nr_host_mem_descs;
150 dma_addr_t host_mem_descs_dma;
151 struct nvme_host_mem_buf_desc *host_mem_descs;
152 void **host_mem_desc_bufs;
153 unsigned int nr_allocated_queues;
154 unsigned int nr_write_queues;
155 unsigned int nr_poll_queues;
158 static int io_queue_depth_set(const char *val, const struct kernel_param *kp)
163 ret = kstrtou32(val, 10, &n);
164 if (ret != 0 || n < 2)
167 return param_set_uint(val, kp);
170 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
172 return qid * 2 * stride;
175 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
177 return (qid * 2 + 1) * stride;
180 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
182 return container_of(ctrl, struct nvme_dev, ctrl);
186 * An NVM Express queue. Each device has at least two (one for admin
187 * commands and one for I/O commands).
190 struct nvme_dev *dev;
193 /* only used for poll queues: */
194 spinlock_t cq_poll_lock ____cacheline_aligned_in_smp;
195 struct nvme_completion *cqes;
196 dma_addr_t sq_dma_addr;
197 dma_addr_t cq_dma_addr;
208 #define NVMEQ_ENABLED 0
209 #define NVMEQ_SQ_CMB 1
210 #define NVMEQ_DELETE_ERROR 2
211 #define NVMEQ_POLLED 3
216 struct completion delete_done;
220 * The nvme_iod describes the data in an I/O.
222 * The sg pointer contains the list of PRP/SGL chunk allocations in addition
223 * to the actual struct scatterlist.
226 struct nvme_request req;
227 struct nvme_command cmd;
228 struct nvme_queue *nvmeq;
231 int npages; /* In the PRP list. 0 means small pool in use */
232 int nents; /* Used in scatterlist */
233 dma_addr_t first_dma;
234 unsigned int dma_len; /* length of single DMA segment mapping */
236 struct scatterlist *sg;
239 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
241 return dev->nr_allocated_queues * 8 * dev->db_stride;
244 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
246 unsigned int mem_size = nvme_dbbuf_size(dev);
251 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
252 &dev->dbbuf_dbs_dma_addr,
256 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
257 &dev->dbbuf_eis_dma_addr,
259 if (!dev->dbbuf_eis) {
260 dma_free_coherent(dev->dev, mem_size,
261 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
262 dev->dbbuf_dbs = NULL;
269 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
271 unsigned int mem_size = nvme_dbbuf_size(dev);
273 if (dev->dbbuf_dbs) {
274 dma_free_coherent(dev->dev, mem_size,
275 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
276 dev->dbbuf_dbs = NULL;
278 if (dev->dbbuf_eis) {
279 dma_free_coherent(dev->dev, mem_size,
280 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
281 dev->dbbuf_eis = NULL;
285 static void nvme_dbbuf_init(struct nvme_dev *dev,
286 struct nvme_queue *nvmeq, int qid)
288 if (!dev->dbbuf_dbs || !qid)
291 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
292 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
293 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
294 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
297 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
302 nvmeq->dbbuf_sq_db = NULL;
303 nvmeq->dbbuf_cq_db = NULL;
304 nvmeq->dbbuf_sq_ei = NULL;
305 nvmeq->dbbuf_cq_ei = NULL;
308 static void nvme_dbbuf_set(struct nvme_dev *dev)
310 struct nvme_command c;
316 memset(&c, 0, sizeof(c));
317 c.dbbuf.opcode = nvme_admin_dbbuf;
318 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
319 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
321 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
322 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
323 /* Free memory and continue on */
324 nvme_dbbuf_dma_free(dev);
326 for (i = 1; i <= dev->online_queues; i++)
327 nvme_dbbuf_free(&dev->queues[i]);
331 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
333 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
336 /* Update dbbuf and return true if an MMIO is required */
337 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
338 volatile u32 *dbbuf_ei)
344 * Ensure that the queue is written before updating
345 * the doorbell in memory
349 old_value = *dbbuf_db;
353 * Ensure that the doorbell is updated before reading the event
354 * index from memory. The controller needs to provide similar
355 * ordering to ensure the envent index is updated before reading
360 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
368 * Will slightly overestimate the number of pages needed. This is OK
369 * as it only leads to a small amount of wasted memory for the lifetime of
372 static int nvme_pci_npages_prp(void)
374 unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
375 NVME_CTRL_PAGE_SIZE);
376 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
380 * Calculates the number of pages needed for the SGL segments. For example a 4k
381 * page can accommodate 256 SGL descriptors.
383 static int nvme_pci_npages_sgl(void)
385 return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc),
389 static size_t nvme_pci_iod_alloc_size(void)
391 size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl());
393 return sizeof(__le64 *) * npages +
394 sizeof(struct scatterlist) * NVME_MAX_SEGS;
397 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
398 unsigned int hctx_idx)
400 struct nvme_dev *dev = data;
401 struct nvme_queue *nvmeq = &dev->queues[0];
403 WARN_ON(hctx_idx != 0);
404 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
406 hctx->driver_data = nvmeq;
410 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
411 unsigned int hctx_idx)
413 struct nvme_dev *dev = data;
414 struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
416 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
417 hctx->driver_data = nvmeq;
421 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
422 unsigned int hctx_idx, unsigned int numa_node)
424 struct nvme_dev *dev = set->driver_data;
425 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
426 int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
427 struct nvme_queue *nvmeq = &dev->queues[queue_idx];
432 nvme_req(req)->ctrl = &dev->ctrl;
436 static int queue_irq_offset(struct nvme_dev *dev)
438 /* if we have more than 1 vec, admin queue offsets us by 1 */
439 if (dev->num_vecs > 1)
445 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
447 struct nvme_dev *dev = set->driver_data;
450 offset = queue_irq_offset(dev);
451 for (i = 0, qoff = 0; i < set->nr_maps; i++) {
452 struct blk_mq_queue_map *map = &set->map[i];
454 map->nr_queues = dev->io_queues[i];
455 if (!map->nr_queues) {
456 BUG_ON(i == HCTX_TYPE_DEFAULT);
461 * The poll queue(s) doesn't have an IRQ (and hence IRQ
462 * affinity), so use the regular blk-mq cpu mapping
464 map->queue_offset = qoff;
465 if (i != HCTX_TYPE_POLL && offset)
466 blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
468 blk_mq_map_queues(map);
469 qoff += map->nr_queues;
470 offset += map->nr_queues;
477 * Write sq tail if we are asked to, or if the next command would wrap.
479 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
482 u16 next_tail = nvmeq->sq_tail + 1;
484 if (next_tail == nvmeq->q_depth)
486 if (next_tail != nvmeq->last_sq_tail)
490 if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
491 nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
492 writel(nvmeq->sq_tail, nvmeq->q_db);
493 nvmeq->last_sq_tail = nvmeq->sq_tail;
497 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
498 * @nvmeq: The queue to use
499 * @cmd: The command to send
500 * @write_sq: whether to write to the SQ doorbell
502 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
505 spin_lock(&nvmeq->sq_lock);
506 memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
508 if (++nvmeq->sq_tail == nvmeq->q_depth)
510 nvme_write_sq_db(nvmeq, write_sq);
511 spin_unlock(&nvmeq->sq_lock);
514 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
516 struct nvme_queue *nvmeq = hctx->driver_data;
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);
524 static void **nvme_pci_iod_list(struct request *req)
526 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
527 return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
530 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
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;
536 avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
538 if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
540 if (!iod->nvmeq->qid)
542 if (!sgl_threshold || avg_seg_size < sgl_threshold)
547 static void nvme_free_prps(struct nvme_dev *dev, struct request *req)
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;
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]);
558 dma_pool_free(dev->prp_page_pool, prp_list, dma_addr);
559 dma_addr = next_dma_addr;
564 static void nvme_free_sgls(struct nvme_dev *dev, struct request *req)
566 const int last_sg = SGES_PER_PAGE - 1;
567 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
568 dma_addr_t dma_addr = iod->first_dma;
571 for (i = 0; i < iod->npages; i++) {
572 struct nvme_sgl_desc *sg_list = nvme_pci_iod_list(req)[i];
573 dma_addr_t next_dma_addr = le64_to_cpu((sg_list[last_sg]).addr);
575 dma_pool_free(dev->prp_page_pool, sg_list, dma_addr);
576 dma_addr = next_dma_addr;
581 static void nvme_unmap_sg(struct nvme_dev *dev, struct request *req)
583 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
585 if (is_pci_p2pdma_page(sg_page(iod->sg)))
586 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
589 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
592 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
594 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
597 dma_unmap_page(dev->dev, iod->first_dma, iod->dma_len,
602 WARN_ON_ONCE(!iod->nents);
604 nvme_unmap_sg(dev, req);
605 if (iod->npages == 0)
606 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
608 else if (iod->use_sgl)
609 nvme_free_sgls(dev, req);
611 nvme_free_prps(dev, req);
612 mempool_free(iod->sg, dev->iod_mempool);
615 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
618 struct scatterlist *sg;
620 for_each_sg(sgl, sg, nents, i) {
621 dma_addr_t phys = sg_phys(sg);
622 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
623 "dma_address:%pad dma_length:%d\n",
624 i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
629 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
630 struct request *req, struct nvme_rw_command *cmnd)
632 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
633 struct dma_pool *pool;
634 int length = blk_rq_payload_bytes(req);
635 struct scatterlist *sg = iod->sg;
636 int dma_len = sg_dma_len(sg);
637 u64 dma_addr = sg_dma_address(sg);
638 int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
640 void **list = nvme_pci_iod_list(req);
644 length -= (NVME_CTRL_PAGE_SIZE - offset);
650 dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
652 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
655 dma_addr = sg_dma_address(sg);
656 dma_len = sg_dma_len(sg);
659 if (length <= NVME_CTRL_PAGE_SIZE) {
660 iod->first_dma = dma_addr;
664 nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
665 if (nprps <= (256 / 8)) {
666 pool = dev->prp_small_pool;
669 pool = dev->prp_page_pool;
673 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
675 iod->first_dma = dma_addr;
677 return BLK_STS_RESOURCE;
680 iod->first_dma = prp_dma;
683 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
684 __le64 *old_prp_list = prp_list;
685 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
688 list[iod->npages++] = prp_list;
689 prp_list[0] = old_prp_list[i - 1];
690 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
693 prp_list[i++] = cpu_to_le64(dma_addr);
694 dma_len -= NVME_CTRL_PAGE_SIZE;
695 dma_addr += NVME_CTRL_PAGE_SIZE;
696 length -= NVME_CTRL_PAGE_SIZE;
701 if (unlikely(dma_len < 0))
704 dma_addr = sg_dma_address(sg);
705 dma_len = sg_dma_len(sg);
708 cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
709 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
712 nvme_free_prps(dev, req);
713 return BLK_STS_RESOURCE;
715 WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
716 "Invalid SGL for payload:%d nents:%d\n",
717 blk_rq_payload_bytes(req), iod->nents);
718 return BLK_STS_IOERR;
721 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
722 struct scatterlist *sg)
724 sge->addr = cpu_to_le64(sg_dma_address(sg));
725 sge->length = cpu_to_le32(sg_dma_len(sg));
726 sge->type = NVME_SGL_FMT_DATA_DESC << 4;
729 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
730 dma_addr_t dma_addr, int entries)
732 sge->addr = cpu_to_le64(dma_addr);
733 if (entries < SGES_PER_PAGE) {
734 sge->length = cpu_to_le32(entries * sizeof(*sge));
735 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
737 sge->length = cpu_to_le32(PAGE_SIZE);
738 sge->type = NVME_SGL_FMT_SEG_DESC << 4;
742 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
743 struct request *req, struct nvme_rw_command *cmd, int entries)
745 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
746 struct dma_pool *pool;
747 struct nvme_sgl_desc *sg_list;
748 struct scatterlist *sg = iod->sg;
752 /* setting the transfer type as SGL */
753 cmd->flags = NVME_CMD_SGL_METABUF;
756 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
760 if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
761 pool = dev->prp_small_pool;
764 pool = dev->prp_page_pool;
768 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
771 return BLK_STS_RESOURCE;
774 nvme_pci_iod_list(req)[0] = sg_list;
775 iod->first_dma = sgl_dma;
777 nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
780 if (i == SGES_PER_PAGE) {
781 struct nvme_sgl_desc *old_sg_desc = sg_list;
782 struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
784 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
789 nvme_pci_iod_list(req)[iod->npages++] = sg_list;
790 sg_list[i++] = *link;
791 nvme_pci_sgl_set_seg(link, sgl_dma, entries);
794 nvme_pci_sgl_set_data(&sg_list[i++], sg);
796 } while (--entries > 0);
800 nvme_free_sgls(dev, req);
801 return BLK_STS_RESOURCE;
804 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
805 struct request *req, struct nvme_rw_command *cmnd,
808 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
809 unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
810 unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
812 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
813 if (dma_mapping_error(dev->dev, iod->first_dma))
814 return BLK_STS_RESOURCE;
815 iod->dma_len = bv->bv_len;
817 cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
818 if (bv->bv_len > first_prp_len)
819 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
823 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
824 struct request *req, struct nvme_rw_command *cmnd,
827 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
829 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
830 if (dma_mapping_error(dev->dev, iod->first_dma))
831 return BLK_STS_RESOURCE;
832 iod->dma_len = bv->bv_len;
834 cmnd->flags = NVME_CMD_SGL_METABUF;
835 cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
836 cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
837 cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
841 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
842 struct nvme_command *cmnd)
844 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
845 blk_status_t ret = BLK_STS_RESOURCE;
848 if (blk_rq_nr_phys_segments(req) == 1) {
849 struct bio_vec bv = req_bvec(req);
851 if (!is_pci_p2pdma_page(bv.bv_page)) {
852 if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
853 return nvme_setup_prp_simple(dev, req,
856 if (iod->nvmeq->qid && sgl_threshold &&
857 dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
858 return nvme_setup_sgl_simple(dev, req,
864 iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
866 return BLK_STS_RESOURCE;
867 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
868 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
872 if (is_pci_p2pdma_page(sg_page(iod->sg)))
873 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
874 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
876 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
877 rq_dma_dir(req), DMA_ATTR_NO_WARN);
881 iod->use_sgl = nvme_pci_use_sgls(dev, req);
883 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
885 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
886 if (ret != BLK_STS_OK)
891 nvme_unmap_sg(dev, req);
893 mempool_free(iod->sg, dev->iod_mempool);
897 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
898 struct nvme_command *cmnd)
900 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
902 iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
904 if (dma_mapping_error(dev->dev, iod->meta_dma))
905 return BLK_STS_IOERR;
906 cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
911 * NOTE: ns is NULL when called on the admin queue.
913 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
914 const struct blk_mq_queue_data *bd)
916 struct nvme_ns *ns = hctx->queue->queuedata;
917 struct nvme_queue *nvmeq = hctx->driver_data;
918 struct nvme_dev *dev = nvmeq->dev;
919 struct request *req = bd->rq;
920 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
921 struct nvme_command *cmnd = &iod->cmd;
929 * We should not need to do this, but we're still using this to
930 * ensure we can drain requests on a dying queue.
932 if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
933 return BLK_STS_IOERR;
935 ret = nvme_setup_cmd(ns, req, cmnd);
939 if (blk_rq_nr_phys_segments(req)) {
940 ret = nvme_map_data(dev, req, cmnd);
945 if (blk_integrity_rq(req)) {
946 ret = nvme_map_metadata(dev, req, cmnd);
951 blk_mq_start_request(req);
952 nvme_submit_cmd(nvmeq, cmnd, bd->last);
955 nvme_unmap_data(dev, req);
957 nvme_cleanup_cmd(req);
961 static void nvme_pci_complete_rq(struct request *req)
963 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
964 struct nvme_dev *dev = iod->nvmeq->dev;
966 if (blk_integrity_rq(req))
967 dma_unmap_page(dev->dev, iod->meta_dma,
968 rq_integrity_vec(req)->bv_len, rq_data_dir(req));
969 if (blk_rq_nr_phys_segments(req))
970 nvme_unmap_data(dev, req);
971 nvme_complete_rq(req);
974 /* We read the CQE phase first to check if the rest of the entry is valid */
975 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
977 struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
979 return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
982 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
984 u16 head = nvmeq->cq_head;
986 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
988 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
991 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
994 return nvmeq->dev->admin_tagset.tags[0];
995 return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
998 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
1000 struct nvme_completion *cqe = &nvmeq->cqes[idx];
1001 __u16 command_id = READ_ONCE(cqe->command_id);
1002 struct request *req;
1005 * AEN requests are special as they don't time out and can
1006 * survive any kind of queue freeze and often don't respond to
1007 * aborts. We don't even bother to allocate a struct request
1008 * for them but rather special case them here.
1010 if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
1011 nvme_complete_async_event(&nvmeq->dev->ctrl,
1012 cqe->status, &cqe->result);
1016 req = nvme_find_rq(nvme_queue_tagset(nvmeq), command_id);
1017 if (unlikely(!req)) {
1018 dev_warn(nvmeq->dev->ctrl.device,
1019 "invalid id %d completed on queue %d\n",
1020 command_id, le16_to_cpu(cqe->sq_id));
1024 trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
1025 if (!nvme_try_complete_req(req, cqe->status, cqe->result))
1026 nvme_pci_complete_rq(req);
1029 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1031 u32 tmp = nvmeq->cq_head + 1;
1033 if (tmp == nvmeq->q_depth) {
1035 nvmeq->cq_phase ^= 1;
1037 nvmeq->cq_head = tmp;
1041 static inline int nvme_process_cq(struct nvme_queue *nvmeq)
1045 while (nvme_cqe_pending(nvmeq)) {
1048 * load-load control dependency between phase and the rest of
1049 * the cqe requires a full read memory barrier
1052 nvme_handle_cqe(nvmeq, nvmeq->cq_head);
1053 nvme_update_cq_head(nvmeq);
1057 nvme_ring_cq_doorbell(nvmeq);
1061 static irqreturn_t nvme_irq(int irq, void *data)
1063 struct nvme_queue *nvmeq = data;
1064 irqreturn_t ret = IRQ_NONE;
1067 * The rmb/wmb pair ensures we see all updates from a previous run of
1068 * the irq handler, even if that was on another CPU.
1071 if (nvme_process_cq(nvmeq))
1078 static irqreturn_t nvme_irq_check(int irq, void *data)
1080 struct nvme_queue *nvmeq = data;
1082 if (nvme_cqe_pending(nvmeq))
1083 return IRQ_WAKE_THREAD;
1088 * Poll for completions for any interrupt driven queue
1089 * Can be called from any context.
1091 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1093 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1095 WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1097 disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1098 nvme_process_cq(nvmeq);
1099 enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1102 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1104 struct nvme_queue *nvmeq = hctx->driver_data;
1107 if (!nvme_cqe_pending(nvmeq))
1110 spin_lock(&nvmeq->cq_poll_lock);
1111 found = nvme_process_cq(nvmeq);
1112 spin_unlock(&nvmeq->cq_poll_lock);
1117 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1119 struct nvme_dev *dev = to_nvme_dev(ctrl);
1120 struct nvme_queue *nvmeq = &dev->queues[0];
1121 struct nvme_command c;
1123 memset(&c, 0, sizeof(c));
1124 c.common.opcode = nvme_admin_async_event;
1125 c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1126 nvme_submit_cmd(nvmeq, &c, true);
1129 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1131 struct nvme_command c;
1133 memset(&c, 0, sizeof(c));
1134 c.delete_queue.opcode = opcode;
1135 c.delete_queue.qid = cpu_to_le16(id);
1137 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1140 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1141 struct nvme_queue *nvmeq, s16 vector)
1143 struct nvme_command c;
1144 int flags = NVME_QUEUE_PHYS_CONTIG;
1146 if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1147 flags |= NVME_CQ_IRQ_ENABLED;
1150 * Note: we (ab)use the fact that the prp fields survive if no data
1151 * is attached to the request.
1153 memset(&c, 0, sizeof(c));
1154 c.create_cq.opcode = nvme_admin_create_cq;
1155 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1156 c.create_cq.cqid = cpu_to_le16(qid);
1157 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1158 c.create_cq.cq_flags = cpu_to_le16(flags);
1159 c.create_cq.irq_vector = cpu_to_le16(vector);
1161 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1164 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1165 struct nvme_queue *nvmeq)
1167 struct nvme_ctrl *ctrl = &dev->ctrl;
1168 struct nvme_command c;
1169 int flags = NVME_QUEUE_PHYS_CONTIG;
1172 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1173 * set. Since URGENT priority is zeroes, it makes all queues
1176 if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1177 flags |= NVME_SQ_PRIO_MEDIUM;
1180 * Note: we (ab)use the fact that the prp fields survive if no data
1181 * is attached to the request.
1183 memset(&c, 0, sizeof(c));
1184 c.create_sq.opcode = nvme_admin_create_sq;
1185 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1186 c.create_sq.sqid = cpu_to_le16(qid);
1187 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1188 c.create_sq.sq_flags = cpu_to_le16(flags);
1189 c.create_sq.cqid = cpu_to_le16(qid);
1191 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1194 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1196 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1199 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1201 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1204 static void abort_endio(struct request *req, blk_status_t error)
1206 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1207 struct nvme_queue *nvmeq = iod->nvmeq;
1209 dev_warn(nvmeq->dev->ctrl.device,
1210 "Abort status: 0x%x", nvme_req(req)->status);
1211 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1212 blk_mq_free_request(req);
1215 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1217 /* If true, indicates loss of adapter communication, possibly by a
1218 * NVMe Subsystem reset.
1220 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1222 /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1223 switch (dev->ctrl.state) {
1224 case NVME_CTRL_RESETTING:
1225 case NVME_CTRL_CONNECTING:
1231 /* We shouldn't reset unless the controller is on fatal error state
1232 * _or_ if we lost the communication with it.
1234 if (!(csts & NVME_CSTS_CFS) && !nssro)
1240 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1242 /* Read a config register to help see what died. */
1246 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1248 if (result == PCIBIOS_SUCCESSFUL)
1249 dev_warn(dev->ctrl.device,
1250 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1253 dev_warn(dev->ctrl.device,
1254 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1258 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1260 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1261 struct nvme_queue *nvmeq = iod->nvmeq;
1262 struct nvme_dev *dev = nvmeq->dev;
1263 struct request *abort_req;
1264 struct nvme_command cmd;
1265 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1267 /* If PCI error recovery process is happening, we cannot reset or
1268 * the recovery mechanism will surely fail.
1271 if (pci_channel_offline(to_pci_dev(dev->dev)))
1272 return BLK_EH_RESET_TIMER;
1275 * Reset immediately if the controller is failed
1277 if (nvme_should_reset(dev, csts)) {
1278 nvme_warn_reset(dev, csts);
1279 nvme_dev_disable(dev, false);
1280 nvme_reset_ctrl(&dev->ctrl);
1285 * Did we miss an interrupt?
1287 if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1288 nvme_poll(req->mq_hctx);
1290 nvme_poll_irqdisable(nvmeq);
1292 if (blk_mq_request_completed(req)) {
1293 dev_warn(dev->ctrl.device,
1294 "I/O %d QID %d timeout, completion polled\n",
1295 req->tag, nvmeq->qid);
1300 * Shutdown immediately if controller times out while starting. The
1301 * reset work will see the pci device disabled when it gets the forced
1302 * cancellation error. All outstanding requests are completed on
1303 * shutdown, so we return BLK_EH_DONE.
1305 switch (dev->ctrl.state) {
1306 case NVME_CTRL_CONNECTING:
1307 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1309 case NVME_CTRL_DELETING:
1310 dev_warn_ratelimited(dev->ctrl.device,
1311 "I/O %d QID %d timeout, disable controller\n",
1312 req->tag, nvmeq->qid);
1313 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1314 nvme_dev_disable(dev, true);
1316 case NVME_CTRL_RESETTING:
1317 return BLK_EH_RESET_TIMER;
1323 * Shutdown the controller immediately and schedule a reset if the
1324 * command was already aborted once before and still hasn't been
1325 * returned to the driver, or if this is the admin queue.
1327 if (!nvmeq->qid || iod->aborted) {
1328 dev_warn(dev->ctrl.device,
1329 "I/O %d QID %d timeout, reset controller\n",
1330 req->tag, nvmeq->qid);
1331 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1332 nvme_dev_disable(dev, false);
1333 nvme_reset_ctrl(&dev->ctrl);
1338 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1339 atomic_inc(&dev->ctrl.abort_limit);
1340 return BLK_EH_RESET_TIMER;
1344 memset(&cmd, 0, sizeof(cmd));
1345 cmd.abort.opcode = nvme_admin_abort_cmd;
1346 cmd.abort.cid = nvme_cid(req);
1347 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1349 dev_warn(nvmeq->dev->ctrl.device,
1350 "I/O %d QID %d timeout, aborting\n",
1351 req->tag, nvmeq->qid);
1353 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1355 if (IS_ERR(abort_req)) {
1356 atomic_inc(&dev->ctrl.abort_limit);
1357 return BLK_EH_RESET_TIMER;
1360 abort_req->end_io_data = NULL;
1361 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1364 * The aborted req will be completed on receiving the abort req.
1365 * We enable the timer again. If hit twice, it'll cause a device reset,
1366 * as the device then is in a faulty state.
1368 return BLK_EH_RESET_TIMER;
1371 static void nvme_free_queue(struct nvme_queue *nvmeq)
1373 dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1374 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1375 if (!nvmeq->sq_cmds)
1378 if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1379 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1380 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1382 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1383 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1387 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1391 for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1392 dev->ctrl.queue_count--;
1393 nvme_free_queue(&dev->queues[i]);
1398 * nvme_suspend_queue - put queue into suspended state
1399 * @nvmeq: queue to suspend
1401 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1403 if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1406 /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1409 nvmeq->dev->online_queues--;
1410 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1411 blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1412 if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1413 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1417 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1421 for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1422 nvme_suspend_queue(&dev->queues[i]);
1425 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1427 struct nvme_queue *nvmeq = &dev->queues[0];
1430 nvme_shutdown_ctrl(&dev->ctrl);
1432 nvme_disable_ctrl(&dev->ctrl);
1434 nvme_poll_irqdisable(nvmeq);
1438 * Called only on a device that has been disabled and after all other threads
1439 * that can check this device's completion queues have synced, except
1440 * nvme_poll(). This is the last chance for the driver to see a natural
1441 * completion before nvme_cancel_request() terminates all incomplete requests.
1443 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1447 for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1448 spin_lock(&dev->queues[i].cq_poll_lock);
1449 nvme_process_cq(&dev->queues[i]);
1450 spin_unlock(&dev->queues[i].cq_poll_lock);
1454 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1457 int q_depth = dev->q_depth;
1458 unsigned q_size_aligned = roundup(q_depth * entry_size,
1459 NVME_CTRL_PAGE_SIZE);
1461 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1462 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1464 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1465 q_depth = div_u64(mem_per_q, entry_size);
1468 * Ensure the reduced q_depth is above some threshold where it
1469 * would be better to map queues in system memory with the
1479 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1482 struct pci_dev *pdev = to_pci_dev(dev->dev);
1484 if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1485 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1486 if (nvmeq->sq_cmds) {
1487 nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1489 if (nvmeq->sq_dma_addr) {
1490 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1494 pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1498 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1499 &nvmeq->sq_dma_addr, GFP_KERNEL);
1500 if (!nvmeq->sq_cmds)
1505 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1507 struct nvme_queue *nvmeq = &dev->queues[qid];
1509 if (dev->ctrl.queue_count > qid)
1512 nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1513 nvmeq->q_depth = depth;
1514 nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1515 &nvmeq->cq_dma_addr, GFP_KERNEL);
1519 if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1523 spin_lock_init(&nvmeq->sq_lock);
1524 spin_lock_init(&nvmeq->cq_poll_lock);
1526 nvmeq->cq_phase = 1;
1527 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1529 dev->ctrl.queue_count++;
1534 dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1535 nvmeq->cq_dma_addr);
1540 static int queue_request_irq(struct nvme_queue *nvmeq)
1542 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1543 int nr = nvmeq->dev->ctrl.instance;
1545 if (use_threaded_interrupts) {
1546 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1547 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1549 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1550 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1554 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1556 struct nvme_dev *dev = nvmeq->dev;
1559 nvmeq->last_sq_tail = 0;
1561 nvmeq->cq_phase = 1;
1562 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1563 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1564 nvme_dbbuf_init(dev, nvmeq, qid);
1565 dev->online_queues++;
1566 wmb(); /* ensure the first interrupt sees the initialization */
1569 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1571 struct nvme_dev *dev = nvmeq->dev;
1575 clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1578 * A queue's vector matches the queue identifier unless the controller
1579 * has only one vector available.
1582 vector = dev->num_vecs == 1 ? 0 : qid;
1584 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1586 result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1590 result = adapter_alloc_sq(dev, qid, nvmeq);
1596 nvmeq->cq_vector = vector;
1597 nvme_init_queue(nvmeq, qid);
1600 result = queue_request_irq(nvmeq);
1605 set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1609 dev->online_queues--;
1610 adapter_delete_sq(dev, qid);
1612 adapter_delete_cq(dev, qid);
1616 static const struct blk_mq_ops nvme_mq_admin_ops = {
1617 .queue_rq = nvme_queue_rq,
1618 .complete = nvme_pci_complete_rq,
1619 .init_hctx = nvme_admin_init_hctx,
1620 .init_request = nvme_init_request,
1621 .timeout = nvme_timeout,
1624 static const struct blk_mq_ops nvme_mq_ops = {
1625 .queue_rq = nvme_queue_rq,
1626 .complete = nvme_pci_complete_rq,
1627 .commit_rqs = nvme_commit_rqs,
1628 .init_hctx = nvme_init_hctx,
1629 .init_request = nvme_init_request,
1630 .map_queues = nvme_pci_map_queues,
1631 .timeout = nvme_timeout,
1635 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1637 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1639 * If the controller was reset during removal, it's possible
1640 * user requests may be waiting on a stopped queue. Start the
1641 * queue to flush these to completion.
1643 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1644 blk_cleanup_queue(dev->ctrl.admin_q);
1645 blk_mq_free_tag_set(&dev->admin_tagset);
1649 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1651 if (!dev->ctrl.admin_q) {
1652 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1653 dev->admin_tagset.nr_hw_queues = 1;
1655 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1656 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1657 dev->admin_tagset.numa_node = dev->ctrl.numa_node;
1658 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1659 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1660 dev->admin_tagset.driver_data = dev;
1662 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1664 dev->ctrl.admin_tagset = &dev->admin_tagset;
1666 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1667 if (IS_ERR(dev->ctrl.admin_q)) {
1668 blk_mq_free_tag_set(&dev->admin_tagset);
1669 dev->ctrl.admin_q = NULL;
1672 if (!blk_get_queue(dev->ctrl.admin_q)) {
1673 nvme_dev_remove_admin(dev);
1674 dev->ctrl.admin_q = NULL;
1678 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1683 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1685 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1688 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1690 struct pci_dev *pdev = to_pci_dev(dev->dev);
1692 if (size <= dev->bar_mapped_size)
1694 if (size > pci_resource_len(pdev, 0))
1698 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1700 dev->bar_mapped_size = 0;
1703 dev->bar_mapped_size = size;
1704 dev->dbs = dev->bar + NVME_REG_DBS;
1709 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1713 struct nvme_queue *nvmeq;
1715 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1719 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1720 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1722 if (dev->subsystem &&
1723 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1724 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1726 result = nvme_disable_ctrl(&dev->ctrl);
1730 result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1734 dev->ctrl.numa_node = dev_to_node(dev->dev);
1736 nvmeq = &dev->queues[0];
1737 aqa = nvmeq->q_depth - 1;
1740 writel(aqa, dev->bar + NVME_REG_AQA);
1741 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1742 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1744 result = nvme_enable_ctrl(&dev->ctrl);
1748 nvmeq->cq_vector = 0;
1749 nvme_init_queue(nvmeq, 0);
1750 result = queue_request_irq(nvmeq);
1752 dev->online_queues--;
1756 set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1760 static int nvme_create_io_queues(struct nvme_dev *dev)
1762 unsigned i, max, rw_queues;
1765 for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1766 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1772 max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1773 if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1774 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1775 dev->io_queues[HCTX_TYPE_READ];
1780 for (i = dev->online_queues; i <= max; i++) {
1781 bool polled = i > rw_queues;
1783 ret = nvme_create_queue(&dev->queues[i], i, polled);
1789 * Ignore failing Create SQ/CQ commands, we can continue with less
1790 * than the desired amount of queues, and even a controller without
1791 * I/O queues can still be used to issue admin commands. This might
1792 * be useful to upgrade a buggy firmware for example.
1794 return ret >= 0 ? 0 : ret;
1797 static ssize_t nvme_cmb_show(struct device *dev,
1798 struct device_attribute *attr,
1801 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1803 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1804 ndev->cmbloc, ndev->cmbsz);
1806 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1808 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1810 u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1812 return 1ULL << (12 + 4 * szu);
1815 static u32 nvme_cmb_size(struct nvme_dev *dev)
1817 return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1820 static void nvme_map_cmb(struct nvme_dev *dev)
1823 resource_size_t bar_size;
1824 struct pci_dev *pdev = to_pci_dev(dev->dev);
1830 if (NVME_CAP_CMBS(dev->ctrl.cap))
1831 writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
1833 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1836 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1838 size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1839 offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1840 bar = NVME_CMB_BIR(dev->cmbloc);
1841 bar_size = pci_resource_len(pdev, bar);
1843 if (offset > bar_size)
1847 * Tell the controller about the host side address mapping the CMB,
1848 * and enable CMB decoding for the NVMe 1.4+ scheme:
1850 if (NVME_CAP_CMBS(dev->ctrl.cap)) {
1851 hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
1852 (pci_bus_address(pdev, bar) + offset),
1853 dev->bar + NVME_REG_CMBMSC);
1857 * Controllers may support a CMB size larger than their BAR,
1858 * for example, due to being behind a bridge. Reduce the CMB to
1859 * the reported size of the BAR
1861 if (size > bar_size - offset)
1862 size = bar_size - offset;
1864 if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1865 dev_warn(dev->ctrl.device,
1866 "failed to register the CMB\n");
1870 dev->cmb_size = size;
1871 dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1873 if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1874 (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1875 pci_p2pmem_publish(pdev, true);
1877 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1878 &dev_attr_cmb.attr, NULL))
1879 dev_warn(dev->ctrl.device,
1880 "failed to add sysfs attribute for CMB\n");
1883 static inline void nvme_release_cmb(struct nvme_dev *dev)
1885 if (dev->cmb_size) {
1886 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1887 &dev_attr_cmb.attr, NULL);
1892 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1894 u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1895 u64 dma_addr = dev->host_mem_descs_dma;
1896 struct nvme_command c;
1899 memset(&c, 0, sizeof(c));
1900 c.features.opcode = nvme_admin_set_features;
1901 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1902 c.features.dword11 = cpu_to_le32(bits);
1903 c.features.dword12 = cpu_to_le32(host_mem_size);
1904 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1905 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1906 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1908 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1910 dev_warn(dev->ctrl.device,
1911 "failed to set host mem (err %d, flags %#x).\n",
1917 static void nvme_free_host_mem(struct nvme_dev *dev)
1921 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1922 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1923 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
1925 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1926 le64_to_cpu(desc->addr),
1927 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1930 kfree(dev->host_mem_desc_bufs);
1931 dev->host_mem_desc_bufs = NULL;
1932 dma_free_coherent(dev->dev,
1933 dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1934 dev->host_mem_descs, dev->host_mem_descs_dma);
1935 dev->host_mem_descs = NULL;
1936 dev->nr_host_mem_descs = 0;
1939 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1942 struct nvme_host_mem_buf_desc *descs;
1943 u32 max_entries, len;
1944 dma_addr_t descs_dma;
1949 tmp = (preferred + chunk_size - 1);
1950 do_div(tmp, chunk_size);
1953 if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1954 max_entries = dev->ctrl.hmmaxd;
1956 descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1957 &descs_dma, GFP_KERNEL);
1961 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1963 goto out_free_descs;
1965 for (size = 0; size < preferred && i < max_entries; size += len) {
1966 dma_addr_t dma_addr;
1968 len = min_t(u64, chunk_size, preferred - size);
1969 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1970 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1974 descs[i].addr = cpu_to_le64(dma_addr);
1975 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
1982 dev->nr_host_mem_descs = i;
1983 dev->host_mem_size = size;
1984 dev->host_mem_descs = descs;
1985 dev->host_mem_descs_dma = descs_dma;
1986 dev->host_mem_desc_bufs = bufs;
1991 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
1993 dma_free_attrs(dev->dev, size, bufs[i],
1994 le64_to_cpu(descs[i].addr),
1995 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
2000 dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
2003 dev->host_mem_descs = NULL;
2007 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
2009 u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
2010 u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
2013 /* start big and work our way down */
2014 for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
2015 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
2016 if (!min || dev->host_mem_size >= min)
2018 nvme_free_host_mem(dev);
2025 static int nvme_setup_host_mem(struct nvme_dev *dev)
2027 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
2028 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
2029 u64 min = (u64)dev->ctrl.hmmin * 4096;
2030 u32 enable_bits = NVME_HOST_MEM_ENABLE;
2033 preferred = min(preferred, max);
2035 dev_warn(dev->ctrl.device,
2036 "min host memory (%lld MiB) above limit (%d MiB).\n",
2037 min >> ilog2(SZ_1M), max_host_mem_size_mb);
2038 nvme_free_host_mem(dev);
2043 * If we already have a buffer allocated check if we can reuse it.
2045 if (dev->host_mem_descs) {
2046 if (dev->host_mem_size >= min)
2047 enable_bits |= NVME_HOST_MEM_RETURN;
2049 nvme_free_host_mem(dev);
2052 if (!dev->host_mem_descs) {
2053 if (nvme_alloc_host_mem(dev, min, preferred)) {
2054 dev_warn(dev->ctrl.device,
2055 "failed to allocate host memory buffer.\n");
2056 return 0; /* controller must work without HMB */
2059 dev_info(dev->ctrl.device,
2060 "allocated %lld MiB host memory buffer.\n",
2061 dev->host_mem_size >> ilog2(SZ_1M));
2064 ret = nvme_set_host_mem(dev, enable_bits);
2066 nvme_free_host_mem(dev);
2071 * nirqs is the number of interrupts available for write and read
2072 * queues. The core already reserved an interrupt for the admin queue.
2074 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2076 struct nvme_dev *dev = affd->priv;
2077 unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2080 * If there is no interrupt available for queues, ensure that
2081 * the default queue is set to 1. The affinity set size is
2082 * also set to one, but the irq core ignores it for this case.
2084 * If only one interrupt is available or 'write_queue' == 0, combine
2085 * write and read queues.
2087 * If 'write_queues' > 0, ensure it leaves room for at least one read
2093 } else if (nrirqs == 1 || !nr_write_queues) {
2095 } else if (nr_write_queues >= nrirqs) {
2098 nr_read_queues = nrirqs - nr_write_queues;
2101 dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2102 affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2103 dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2104 affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2105 affd->nr_sets = nr_read_queues ? 2 : 1;
2108 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2110 struct pci_dev *pdev = to_pci_dev(dev->dev);
2111 struct irq_affinity affd = {
2113 .calc_sets = nvme_calc_irq_sets,
2116 unsigned int irq_queues, poll_queues;
2119 * Poll queues don't need interrupts, but we need at least one I/O queue
2120 * left over for non-polled I/O.
2122 poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2123 dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2126 * Initialize for the single interrupt case, will be updated in
2127 * nvme_calc_irq_sets().
2129 dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2130 dev->io_queues[HCTX_TYPE_READ] = 0;
2133 * We need interrupts for the admin queue and each non-polled I/O queue,
2134 * but some Apple controllers require all queues to use the first
2138 if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2139 irq_queues += (nr_io_queues - poll_queues);
2140 return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2141 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2144 static void nvme_disable_io_queues(struct nvme_dev *dev)
2146 if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2147 __nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2150 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2152 return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2155 static int nvme_setup_io_queues(struct nvme_dev *dev)
2157 struct nvme_queue *adminq = &dev->queues[0];
2158 struct pci_dev *pdev = to_pci_dev(dev->dev);
2159 unsigned int nr_io_queues;
2164 * Sample the module parameters once at reset time so that we have
2165 * stable values to work with.
2167 dev->nr_write_queues = write_queues;
2168 dev->nr_poll_queues = poll_queues;
2171 * If tags are shared with admin queue (Apple bug), then
2172 * make sure we only use one IO queue.
2174 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2177 nr_io_queues = min(nvme_max_io_queues(dev),
2178 dev->nr_allocated_queues - 1);
2180 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2184 if (nr_io_queues == 0)
2187 clear_bit(NVMEQ_ENABLED, &adminq->flags);
2189 if (dev->cmb_use_sqes) {
2190 result = nvme_cmb_qdepth(dev, nr_io_queues,
2191 sizeof(struct nvme_command));
2193 dev->q_depth = result;
2195 dev->cmb_use_sqes = false;
2199 size = db_bar_size(dev, nr_io_queues);
2200 result = nvme_remap_bar(dev, size);
2203 if (!--nr_io_queues)
2206 adminq->q_db = dev->dbs;
2209 /* Deregister the admin queue's interrupt */
2210 pci_free_irq(pdev, 0, adminq);
2213 * If we enable msix early due to not intx, disable it again before
2214 * setting up the full range we need.
2216 pci_free_irq_vectors(pdev);
2218 result = nvme_setup_irqs(dev, nr_io_queues);
2222 dev->num_vecs = result;
2223 result = max(result - 1, 1);
2224 dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2227 * Should investigate if there's a performance win from allocating
2228 * more queues than interrupt vectors; it might allow the submission
2229 * path to scale better, even if the receive path is limited by the
2230 * number of interrupts.
2232 result = queue_request_irq(adminq);
2235 set_bit(NVMEQ_ENABLED, &adminq->flags);
2237 result = nvme_create_io_queues(dev);
2238 if (result || dev->online_queues < 2)
2241 if (dev->online_queues - 1 < dev->max_qid) {
2242 nr_io_queues = dev->online_queues - 1;
2243 nvme_disable_io_queues(dev);
2244 nvme_suspend_io_queues(dev);
2247 dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2248 dev->io_queues[HCTX_TYPE_DEFAULT],
2249 dev->io_queues[HCTX_TYPE_READ],
2250 dev->io_queues[HCTX_TYPE_POLL]);
2254 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2256 struct nvme_queue *nvmeq = req->end_io_data;
2258 blk_mq_free_request(req);
2259 complete(&nvmeq->delete_done);
2262 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2264 struct nvme_queue *nvmeq = req->end_io_data;
2267 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2269 nvme_del_queue_end(req, error);
2272 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2274 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2275 struct request *req;
2276 struct nvme_command cmd;
2278 memset(&cmd, 0, sizeof(cmd));
2279 cmd.delete_queue.opcode = opcode;
2280 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2282 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT);
2284 return PTR_ERR(req);
2286 req->end_io_data = nvmeq;
2288 init_completion(&nvmeq->delete_done);
2289 blk_execute_rq_nowait(q, NULL, req, false,
2290 opcode == nvme_admin_delete_cq ?
2291 nvme_del_cq_end : nvme_del_queue_end);
2295 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2297 int nr_queues = dev->online_queues - 1, sent = 0;
2298 unsigned long timeout;
2301 timeout = ADMIN_TIMEOUT;
2302 while (nr_queues > 0) {
2303 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2309 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2311 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2323 static void nvme_dev_add(struct nvme_dev *dev)
2327 if (!dev->ctrl.tagset) {
2328 dev->tagset.ops = &nvme_mq_ops;
2329 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2330 dev->tagset.nr_maps = 2; /* default + read */
2331 if (dev->io_queues[HCTX_TYPE_POLL])
2332 dev->tagset.nr_maps++;
2333 dev->tagset.timeout = NVME_IO_TIMEOUT;
2334 dev->tagset.numa_node = dev->ctrl.numa_node;
2335 dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth,
2336 BLK_MQ_MAX_DEPTH) - 1;
2337 dev->tagset.cmd_size = sizeof(struct nvme_iod);
2338 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2339 dev->tagset.driver_data = dev;
2342 * Some Apple controllers requires tags to be unique
2343 * across admin and IO queue, so reserve the first 32
2344 * tags of the IO queue.
2346 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2347 dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2349 ret = blk_mq_alloc_tag_set(&dev->tagset);
2351 dev_warn(dev->ctrl.device,
2352 "IO queues tagset allocation failed %d\n", ret);
2355 dev->ctrl.tagset = &dev->tagset;
2357 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2359 /* Free previously allocated queues that are no longer usable */
2360 nvme_free_queues(dev, dev->online_queues);
2363 nvme_dbbuf_set(dev);
2366 static int nvme_pci_enable(struct nvme_dev *dev)
2368 int result = -ENOMEM;
2369 struct pci_dev *pdev = to_pci_dev(dev->dev);
2371 if (pci_enable_device_mem(pdev))
2374 pci_set_master(pdev);
2376 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2379 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2385 * Some devices and/or platforms don't advertise or work with INTx
2386 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2387 * adjust this later.
2389 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2393 dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2395 dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2397 dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2398 dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2399 dev->dbs = dev->bar + 4096;
2402 * Some Apple controllers require a non-standard SQE size.
2403 * Interestingly they also seem to ignore the CC:IOSQES register
2404 * so we don't bother updating it here.
2406 if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2409 dev->io_sqes = NVME_NVM_IOSQES;
2412 * Temporary fix for the Apple controller found in the MacBook8,1 and
2413 * some MacBook7,1 to avoid controller resets and data loss.
2415 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2417 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2418 "set queue depth=%u to work around controller resets\n",
2420 } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2421 (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2422 NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2424 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2425 "set queue depth=%u\n", dev->q_depth);
2429 * Controllers with the shared tags quirk need the IO queue to be
2430 * big enough so that we get 32 tags for the admin queue
2432 if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2433 (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2434 dev->q_depth = NVME_AQ_DEPTH + 2;
2435 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2442 pci_enable_pcie_error_reporting(pdev);
2443 pci_save_state(pdev);
2447 pci_disable_device(pdev);
2451 static void nvme_dev_unmap(struct nvme_dev *dev)
2455 pci_release_mem_regions(to_pci_dev(dev->dev));
2458 static void nvme_pci_disable(struct nvme_dev *dev)
2460 struct pci_dev *pdev = to_pci_dev(dev->dev);
2462 pci_free_irq_vectors(pdev);
2464 if (pci_is_enabled(pdev)) {
2465 pci_disable_pcie_error_reporting(pdev);
2466 pci_disable_device(pdev);
2470 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2472 bool dead = true, freeze = false;
2473 struct pci_dev *pdev = to_pci_dev(dev->dev);
2475 mutex_lock(&dev->shutdown_lock);
2476 if (pci_is_enabled(pdev)) {
2477 u32 csts = readl(dev->bar + NVME_REG_CSTS);
2479 if (dev->ctrl.state == NVME_CTRL_LIVE ||
2480 dev->ctrl.state == NVME_CTRL_RESETTING) {
2482 nvme_start_freeze(&dev->ctrl);
2484 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2485 pdev->error_state != pci_channel_io_normal);
2489 * Give the controller a chance to complete all entered requests if
2490 * doing a safe shutdown.
2492 if (!dead && shutdown && freeze)
2493 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2495 nvme_stop_queues(&dev->ctrl);
2497 if (!dead && dev->ctrl.queue_count > 0) {
2498 nvme_disable_io_queues(dev);
2499 nvme_disable_admin_queue(dev, shutdown);
2501 nvme_suspend_io_queues(dev);
2502 nvme_suspend_queue(&dev->queues[0]);
2503 nvme_pci_disable(dev);
2504 nvme_reap_pending_cqes(dev);
2506 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2507 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2508 blk_mq_tagset_wait_completed_request(&dev->tagset);
2509 blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2512 * The driver will not be starting up queues again if shutting down so
2513 * must flush all entered requests to their failed completion to avoid
2514 * deadlocking blk-mq hot-cpu notifier.
2517 nvme_start_queues(&dev->ctrl);
2518 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2519 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2521 mutex_unlock(&dev->shutdown_lock);
2524 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2526 if (!nvme_wait_reset(&dev->ctrl))
2528 nvme_dev_disable(dev, shutdown);
2532 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2534 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2535 NVME_CTRL_PAGE_SIZE,
2536 NVME_CTRL_PAGE_SIZE, 0);
2537 if (!dev->prp_page_pool)
2540 /* Optimisation for I/Os between 4k and 128k */
2541 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2543 if (!dev->prp_small_pool) {
2544 dma_pool_destroy(dev->prp_page_pool);
2550 static void nvme_release_prp_pools(struct nvme_dev *dev)
2552 dma_pool_destroy(dev->prp_page_pool);
2553 dma_pool_destroy(dev->prp_small_pool);
2556 static void nvme_free_tagset(struct nvme_dev *dev)
2558 if (dev->tagset.tags)
2559 blk_mq_free_tag_set(&dev->tagset);
2560 dev->ctrl.tagset = NULL;
2563 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2565 struct nvme_dev *dev = to_nvme_dev(ctrl);
2567 nvme_dbbuf_dma_free(dev);
2568 nvme_free_tagset(dev);
2569 if (dev->ctrl.admin_q)
2570 blk_put_queue(dev->ctrl.admin_q);
2571 free_opal_dev(dev->ctrl.opal_dev);
2572 mempool_destroy(dev->iod_mempool);
2573 put_device(dev->dev);
2578 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2581 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2582 * may be holding this pci_dev's device lock.
2584 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2585 nvme_get_ctrl(&dev->ctrl);
2586 nvme_dev_disable(dev, false);
2587 nvme_kill_queues(&dev->ctrl);
2588 if (!queue_work(nvme_wq, &dev->remove_work))
2589 nvme_put_ctrl(&dev->ctrl);
2592 static void nvme_reset_work(struct work_struct *work)
2594 struct nvme_dev *dev =
2595 container_of(work, struct nvme_dev, ctrl.reset_work);
2596 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2599 if (dev->ctrl.state != NVME_CTRL_RESETTING) {
2600 dev_warn(dev->ctrl.device, "ctrl state %d is not RESETTING\n",
2607 * If we're called to reset a live controller first shut it down before
2610 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2611 nvme_dev_disable(dev, false);
2612 nvme_sync_queues(&dev->ctrl);
2614 mutex_lock(&dev->shutdown_lock);
2615 result = nvme_pci_enable(dev);
2619 result = nvme_pci_configure_admin_queue(dev);
2623 result = nvme_alloc_admin_tags(dev);
2627 dma_set_min_align_mask(dev->dev, NVME_CTRL_PAGE_SIZE - 1);
2630 * Limit the max command size to prevent iod->sg allocations going
2631 * over a single page.
2633 dev->ctrl.max_hw_sectors = min_t(u32,
2634 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2635 dev->ctrl.max_segments = NVME_MAX_SEGS;
2638 * Don't limit the IOMMU merged segment size.
2640 dma_set_max_seg_size(dev->dev, 0xffffffff);
2642 mutex_unlock(&dev->shutdown_lock);
2645 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2646 * initializing procedure here.
2648 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2649 dev_warn(dev->ctrl.device,
2650 "failed to mark controller CONNECTING\n");
2656 * We do not support an SGL for metadata (yet), so we are limited to a
2657 * single integrity segment for the separate metadata pointer.
2659 dev->ctrl.max_integrity_segments = 1;
2661 result = nvme_init_identify(&dev->ctrl);
2665 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2666 if (!dev->ctrl.opal_dev)
2667 dev->ctrl.opal_dev =
2668 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2669 else if (was_suspend)
2670 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2672 free_opal_dev(dev->ctrl.opal_dev);
2673 dev->ctrl.opal_dev = NULL;
2676 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2677 result = nvme_dbbuf_dma_alloc(dev);
2680 "unable to allocate dma for dbbuf\n");
2683 if (dev->ctrl.hmpre) {
2684 result = nvme_setup_host_mem(dev);
2689 result = nvme_setup_io_queues(dev);
2694 * Keep the controller around but remove all namespaces if we don't have
2695 * any working I/O queue.
2697 if (dev->online_queues < 2) {
2698 dev_warn(dev->ctrl.device, "IO queues not created\n");
2699 nvme_kill_queues(&dev->ctrl);
2700 nvme_remove_namespaces(&dev->ctrl);
2701 nvme_free_tagset(dev);
2703 nvme_start_queues(&dev->ctrl);
2704 nvme_wait_freeze(&dev->ctrl);
2706 nvme_unfreeze(&dev->ctrl);
2710 * If only admin queue live, keep it to do further investigation or
2713 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2714 dev_warn(dev->ctrl.device,
2715 "failed to mark controller live state\n");
2720 nvme_start_ctrl(&dev->ctrl);
2724 mutex_unlock(&dev->shutdown_lock);
2727 dev_warn(dev->ctrl.device,
2728 "Removing after probe failure status: %d\n", result);
2729 nvme_remove_dead_ctrl(dev);
2732 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2734 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2735 struct pci_dev *pdev = to_pci_dev(dev->dev);
2737 if (pci_get_drvdata(pdev))
2738 device_release_driver(&pdev->dev);
2739 nvme_put_ctrl(&dev->ctrl);
2742 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2744 *val = readl(to_nvme_dev(ctrl)->bar + off);
2748 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2750 writel(val, to_nvme_dev(ctrl)->bar + off);
2754 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2756 *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2760 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2762 struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2764 return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2767 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2769 .module = THIS_MODULE,
2770 .flags = NVME_F_METADATA_SUPPORTED |
2772 .reg_read32 = nvme_pci_reg_read32,
2773 .reg_write32 = nvme_pci_reg_write32,
2774 .reg_read64 = nvme_pci_reg_read64,
2775 .free_ctrl = nvme_pci_free_ctrl,
2776 .submit_async_event = nvme_pci_submit_async_event,
2777 .get_address = nvme_pci_get_address,
2780 static int nvme_dev_map(struct nvme_dev *dev)
2782 struct pci_dev *pdev = to_pci_dev(dev->dev);
2784 if (pci_request_mem_regions(pdev, "nvme"))
2787 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2792 pci_release_mem_regions(pdev);
2796 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2798 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2800 * Several Samsung devices seem to drop off the PCIe bus
2801 * randomly when APST is on and uses the deepest sleep state.
2802 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2803 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2804 * 950 PRO 256GB", but it seems to be restricted to two Dell
2807 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2808 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2809 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2810 return NVME_QUIRK_NO_DEEPEST_PS;
2811 } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2813 * Samsung SSD 960 EVO drops off the PCIe bus after system
2814 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2815 * within few minutes after bootup on a Coffee Lake board -
2818 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2819 (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2820 dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2821 return NVME_QUIRK_NO_APST;
2822 } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2823 pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2824 (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2826 * Forcing to use host managed nvme power settings for
2827 * lowest idle power with quick resume latency on
2828 * Samsung and Toshiba SSDs based on suspend behavior
2829 * on Coffee Lake board for LENOVO C640
2831 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2832 dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2833 return NVME_QUIRK_SIMPLE_SUSPEND;
2840 static bool nvme_acpi_storage_d3(struct pci_dev *dev)
2842 struct acpi_device *adev = ACPI_COMPANION(&dev->dev);
2846 * Look for _DSD property specifying that the storage device on the port
2847 * must use D3 to support deep platform power savings during
2853 if (fwnode_property_read_u8(acpi_fwnode_handle(adev), "StorageD3Enable",
2859 static inline bool nvme_acpi_storage_d3(struct pci_dev *dev)
2863 #endif /* CONFIG_ACPI */
2865 static void nvme_async_probe(void *data, async_cookie_t cookie)
2867 struct nvme_dev *dev = data;
2869 flush_work(&dev->ctrl.reset_work);
2870 flush_work(&dev->ctrl.scan_work);
2871 nvme_put_ctrl(&dev->ctrl);
2874 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2876 int node, result = -ENOMEM;
2877 struct nvme_dev *dev;
2878 unsigned long quirks = id->driver_data;
2881 node = dev_to_node(&pdev->dev);
2882 if (node == NUMA_NO_NODE)
2883 set_dev_node(&pdev->dev, first_memory_node);
2885 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2889 dev->nr_write_queues = write_queues;
2890 dev->nr_poll_queues = poll_queues;
2891 dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
2892 dev->queues = kcalloc_node(dev->nr_allocated_queues,
2893 sizeof(struct nvme_queue), GFP_KERNEL, node);
2897 dev->dev = get_device(&pdev->dev);
2898 pci_set_drvdata(pdev, dev);
2900 result = nvme_dev_map(dev);
2904 INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2905 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2906 mutex_init(&dev->shutdown_lock);
2908 result = nvme_setup_prp_pools(dev);
2912 quirks |= check_vendor_combination_bug(pdev);
2914 if (!noacpi && nvme_acpi_storage_d3(pdev)) {
2916 * Some systems use a bios work around to ask for D3 on
2917 * platforms that support kernel managed suspend.
2919 dev_info(&pdev->dev,
2920 "platform quirk: setting simple suspend\n");
2921 quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
2925 * Double check that our mempool alloc size will cover the biggest
2926 * command we support.
2928 alloc_size = nvme_pci_iod_alloc_size();
2929 WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2931 dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2933 (void *) alloc_size,
2935 if (!dev->iod_mempool) {
2940 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2943 goto release_mempool;
2945 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2947 nvme_reset_ctrl(&dev->ctrl);
2948 async_schedule(nvme_async_probe, dev);
2953 mempool_destroy(dev->iod_mempool);
2955 nvme_release_prp_pools(dev);
2957 nvme_dev_unmap(dev);
2959 put_device(dev->dev);
2966 static void nvme_reset_prepare(struct pci_dev *pdev)
2968 struct nvme_dev *dev = pci_get_drvdata(pdev);
2971 * We don't need to check the return value from waiting for the reset
2972 * state as pci_dev device lock is held, making it impossible to race
2975 nvme_disable_prepare_reset(dev, false);
2976 nvme_sync_queues(&dev->ctrl);
2979 static void nvme_reset_done(struct pci_dev *pdev)
2981 struct nvme_dev *dev = pci_get_drvdata(pdev);
2983 if (!nvme_try_sched_reset(&dev->ctrl))
2984 flush_work(&dev->ctrl.reset_work);
2987 static void nvme_shutdown(struct pci_dev *pdev)
2989 struct nvme_dev *dev = pci_get_drvdata(pdev);
2991 nvme_disable_prepare_reset(dev, true);
2995 * The driver's remove may be called on a device in a partially initialized
2996 * state. This function must not have any dependencies on the device state in
2999 static void nvme_remove(struct pci_dev *pdev)
3001 struct nvme_dev *dev = pci_get_drvdata(pdev);
3003 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
3004 pci_set_drvdata(pdev, NULL);
3006 if (!pci_device_is_present(pdev)) {
3007 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
3008 nvme_dev_disable(dev, true);
3011 flush_work(&dev->ctrl.reset_work);
3012 nvme_stop_ctrl(&dev->ctrl);
3013 nvme_remove_namespaces(&dev->ctrl);
3014 nvme_dev_disable(dev, true);
3015 nvme_release_cmb(dev);
3016 nvme_free_host_mem(dev);
3017 nvme_dev_remove_admin(dev);
3018 nvme_free_queues(dev, 0);
3019 nvme_release_prp_pools(dev);
3020 nvme_dev_unmap(dev);
3021 nvme_uninit_ctrl(&dev->ctrl);
3024 #ifdef CONFIG_PM_SLEEP
3025 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
3027 return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3030 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3032 return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3035 static int nvme_resume(struct device *dev)
3037 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3038 struct nvme_ctrl *ctrl = &ndev->ctrl;
3040 if (ndev->last_ps == U32_MAX ||
3041 nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3042 return nvme_try_sched_reset(&ndev->ctrl);
3046 static int nvme_suspend(struct device *dev)
3048 struct pci_dev *pdev = to_pci_dev(dev);
3049 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3050 struct nvme_ctrl *ctrl = &ndev->ctrl;
3053 ndev->last_ps = U32_MAX;
3056 * The platform does not remove power for a kernel managed suspend so
3057 * use host managed nvme power settings for lowest idle power if
3058 * possible. This should have quicker resume latency than a full device
3059 * shutdown. But if the firmware is involved after the suspend or the
3060 * device does not support any non-default power states, shut down the
3063 * If ASPM is not enabled for the device, shut down the device and allow
3064 * the PCI bus layer to put it into D3 in order to take the PCIe link
3065 * down, so as to allow the platform to achieve its minimum low-power
3066 * state (which may not be possible if the link is up).
3068 * If a host memory buffer is enabled, shut down the device as the NVMe
3069 * specification allows the device to access the host memory buffer in
3070 * host DRAM from all power states, but hosts will fail access to DRAM
3073 if (pm_suspend_via_firmware() || !ctrl->npss ||
3074 !pcie_aspm_enabled(pdev) ||
3075 ndev->nr_host_mem_descs ||
3076 (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3077 return nvme_disable_prepare_reset(ndev, true);
3079 nvme_start_freeze(ctrl);
3080 nvme_wait_freeze(ctrl);
3081 nvme_sync_queues(ctrl);
3083 if (ctrl->state != NVME_CTRL_LIVE)
3086 ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3091 * A saved state prevents pci pm from generically controlling the
3092 * device's power. If we're using protocol specific settings, we don't
3093 * want pci interfering.
3095 pci_save_state(pdev);
3097 ret = nvme_set_power_state(ctrl, ctrl->npss);
3102 /* discard the saved state */
3103 pci_load_saved_state(pdev, NULL);
3106 * Clearing npss forces a controller reset on resume. The
3107 * correct value will be rediscovered then.
3109 ret = nvme_disable_prepare_reset(ndev, true);
3113 nvme_unfreeze(ctrl);
3117 static int nvme_simple_suspend(struct device *dev)
3119 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3121 return nvme_disable_prepare_reset(ndev, true);
3124 static int nvme_simple_resume(struct device *dev)
3126 struct pci_dev *pdev = to_pci_dev(dev);
3127 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3129 return nvme_try_sched_reset(&ndev->ctrl);
3132 static const struct dev_pm_ops nvme_dev_pm_ops = {
3133 .suspend = nvme_suspend,
3134 .resume = nvme_resume,
3135 .freeze = nvme_simple_suspend,
3136 .thaw = nvme_simple_resume,
3137 .poweroff = nvme_simple_suspend,
3138 .restore = nvme_simple_resume,
3140 #endif /* CONFIG_PM_SLEEP */
3142 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3143 pci_channel_state_t state)
3145 struct nvme_dev *dev = pci_get_drvdata(pdev);
3148 * A frozen channel requires a reset. When detected, this method will
3149 * shutdown the controller to quiesce. The controller will be restarted
3150 * after the slot reset through driver's slot_reset callback.
3153 case pci_channel_io_normal:
3154 return PCI_ERS_RESULT_CAN_RECOVER;
3155 case pci_channel_io_frozen:
3156 dev_warn(dev->ctrl.device,
3157 "frozen state error detected, reset controller\n");
3158 nvme_dev_disable(dev, false);
3159 return PCI_ERS_RESULT_NEED_RESET;
3160 case pci_channel_io_perm_failure:
3161 dev_warn(dev->ctrl.device,
3162 "failure state error detected, request disconnect\n");
3163 return PCI_ERS_RESULT_DISCONNECT;
3165 return PCI_ERS_RESULT_NEED_RESET;
3168 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3170 struct nvme_dev *dev = pci_get_drvdata(pdev);
3172 dev_info(dev->ctrl.device, "restart after slot reset\n");
3173 pci_restore_state(pdev);
3174 nvme_reset_ctrl(&dev->ctrl);
3175 return PCI_ERS_RESULT_RECOVERED;
3178 static void nvme_error_resume(struct pci_dev *pdev)
3180 struct nvme_dev *dev = pci_get_drvdata(pdev);
3182 flush_work(&dev->ctrl.reset_work);
3185 static const struct pci_error_handlers nvme_err_handler = {
3186 .error_detected = nvme_error_detected,
3187 .slot_reset = nvme_slot_reset,
3188 .resume = nvme_error_resume,
3189 .reset_prepare = nvme_reset_prepare,
3190 .reset_done = nvme_reset_done,
3193 static const struct pci_device_id nvme_id_table[] = {
3194 { PCI_VDEVICE(INTEL, 0x0953), /* Intel 750/P3500/P3600/P3700 */
3195 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3196 NVME_QUIRK_DEALLOCATE_ZEROES, },
3197 { PCI_VDEVICE(INTEL, 0x0a53), /* Intel P3520 */
3198 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3199 NVME_QUIRK_DEALLOCATE_ZEROES, },
3200 { PCI_VDEVICE(INTEL, 0x0a54), /* Intel P4500/P4600 */
3201 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3202 NVME_QUIRK_DEALLOCATE_ZEROES |
3203 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3204 { PCI_VDEVICE(INTEL, 0x0a55), /* Dell Express Flash P4600 */
3205 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3206 NVME_QUIRK_DEALLOCATE_ZEROES, },
3207 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
3208 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3209 NVME_QUIRK_MEDIUM_PRIO_SQ |
3210 NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3211 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3212 { PCI_VDEVICE(INTEL, 0xf1a6), /* Intel 760p/Pro 7600p */
3213 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3214 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
3215 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3216 NVME_QUIRK_DISABLE_WRITE_ZEROES |
3217 NVME_QUIRK_BOGUS_NID, },
3218 { PCI_VDEVICE(REDHAT, 0x0010), /* Qemu emulated controller */
3219 .driver_data = NVME_QUIRK_BOGUS_NID, },
3220 { PCI_DEVICE(0x126f, 0x2263), /* Silicon Motion unidentified */
3221 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST, },
3222 { PCI_DEVICE(0x1bb1, 0x0100), /* Seagate Nytro Flash Storage */
3223 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3224 NVME_QUIRK_NO_NS_DESC_LIST, },
3225 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
3226 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3227 { PCI_DEVICE(0x1c58, 0x0023), /* WDC SN200 adapter */
3228 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3229 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
3230 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3231 { PCI_DEVICE(0x144d, 0xa821), /* Samsung PM1725 */
3232 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3233 { PCI_DEVICE(0x144d, 0xa822), /* Samsung PM1725a */
3234 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3235 NVME_QUIRK_DISABLE_WRITE_ZEROES|
3236 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3237 { PCI_DEVICE(0x1987, 0x5016), /* Phison E16 */
3238 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN |
3239 NVME_QUIRK_BOGUS_NID, },
3240 { PCI_DEVICE(0x1b4b, 0x1092), /* Lexar 256 GB SSD */
3241 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST |
3242 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3243 { PCI_DEVICE(0x1d1d, 0x1f1f), /* LighNVM qemu device */
3244 .driver_data = NVME_QUIRK_LIGHTNVM, },
3245 { PCI_DEVICE(0x1d1d, 0x2807), /* CNEX WL */
3246 .driver_data = NVME_QUIRK_LIGHTNVM, },
3247 { PCI_DEVICE(0x1d1d, 0x2601), /* CNEX Granby */
3248 .driver_data = NVME_QUIRK_LIGHTNVM, },
3249 { PCI_DEVICE(0x10ec, 0x5762), /* ADATA SX6000LNP */
3250 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN |
3251 NVME_QUIRK_BOGUS_NID, },
3252 { PCI_DEVICE(0x1cc1, 0x8201), /* ADATA SX8200PNP 512GB */
3253 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3254 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3255 { PCI_DEVICE(0x1c5c, 0x1504), /* SK Hynix PC400 */
3256 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3257 { PCI_DEVICE(0x15b7, 0x2001), /* Sandisk Skyhawk */
3258 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3259 { PCI_DEVICE(0x2646, 0x2262), /* KINGSTON SKC2000 NVMe SSD */
3260 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3261 { PCI_DEVICE(0x2646, 0x2263), /* KINGSTON A2000 NVMe SSD */
3262 .driver_data = NVME_QUIRK_NO_DEEPEST_PS, },
3263 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3264 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3265 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3266 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3267 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3268 NVME_QUIRK_128_BYTES_SQES |
3269 NVME_QUIRK_SHARED_TAGS |
3270 NVME_QUIRK_SKIP_CID_GEN },
3271 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3274 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3276 static struct pci_driver nvme_driver = {
3278 .id_table = nvme_id_table,
3279 .probe = nvme_probe,
3280 .remove = nvme_remove,
3281 .shutdown = nvme_shutdown,
3282 #ifdef CONFIG_PM_SLEEP
3284 .pm = &nvme_dev_pm_ops,
3287 .sriov_configure = pci_sriov_configure_simple,
3288 .err_handler = &nvme_err_handler,
3291 static int __init nvme_init(void)
3293 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3294 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3295 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3296 BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3298 return pci_register_driver(&nvme_driver);
3301 static void __exit nvme_exit(void)
3303 pci_unregister_driver(&nvme_driver);
3304 flush_workqueue(nvme_wq);
3307 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3308 MODULE_LICENSE("GPL");
3309 MODULE_VERSION("1.0");
3310 module_init(nvme_init);
3311 module_exit(nvme_exit);