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_queue *nvmeq;
230 int npages; /* In the PRP list. 0 means small pool in use */
231 int nents; /* Used in scatterlist */
232 dma_addr_t first_dma;
233 unsigned int dma_len; /* length of single DMA segment mapping */
235 struct scatterlist *sg;
238 static inline unsigned int nvme_dbbuf_size(struct nvme_dev *dev)
240 return dev->nr_allocated_queues * 8 * dev->db_stride;
243 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
245 unsigned int mem_size = nvme_dbbuf_size(dev);
250 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
251 &dev->dbbuf_dbs_dma_addr,
255 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
256 &dev->dbbuf_eis_dma_addr,
258 if (!dev->dbbuf_eis) {
259 dma_free_coherent(dev->dev, mem_size,
260 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
261 dev->dbbuf_dbs = NULL;
268 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
270 unsigned int mem_size = nvme_dbbuf_size(dev);
272 if (dev->dbbuf_dbs) {
273 dma_free_coherent(dev->dev, mem_size,
274 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
275 dev->dbbuf_dbs = NULL;
277 if (dev->dbbuf_eis) {
278 dma_free_coherent(dev->dev, mem_size,
279 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
280 dev->dbbuf_eis = NULL;
284 static void nvme_dbbuf_init(struct nvme_dev *dev,
285 struct nvme_queue *nvmeq, int qid)
287 if (!dev->dbbuf_dbs || !qid)
290 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
291 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
292 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
293 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
296 static void nvme_dbbuf_free(struct nvme_queue *nvmeq)
301 nvmeq->dbbuf_sq_db = NULL;
302 nvmeq->dbbuf_cq_db = NULL;
303 nvmeq->dbbuf_sq_ei = NULL;
304 nvmeq->dbbuf_cq_ei = NULL;
307 static void nvme_dbbuf_set(struct nvme_dev *dev)
309 struct nvme_command c;
315 memset(&c, 0, sizeof(c));
316 c.dbbuf.opcode = nvme_admin_dbbuf;
317 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
318 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
320 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
321 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
322 /* Free memory and continue on */
323 nvme_dbbuf_dma_free(dev);
325 for (i = 1; i <= dev->online_queues; i++)
326 nvme_dbbuf_free(&dev->queues[i]);
330 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
332 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
335 /* Update dbbuf and return true if an MMIO is required */
336 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
337 volatile u32 *dbbuf_ei)
343 * Ensure that the queue is written before updating
344 * the doorbell in memory
348 old_value = *dbbuf_db;
352 * Ensure that the doorbell is updated before reading the event
353 * index from memory. The controller needs to provide similar
354 * ordering to ensure the envent index is updated before reading
359 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
367 * Will slightly overestimate the number of pages needed. This is OK
368 * as it only leads to a small amount of wasted memory for the lifetime of
371 static int nvme_pci_npages_prp(void)
373 unsigned nprps = DIV_ROUND_UP(NVME_MAX_KB_SZ + NVME_CTRL_PAGE_SIZE,
374 NVME_CTRL_PAGE_SIZE);
375 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
379 * Calculates the number of pages needed for the SGL segments. For example a 4k
380 * page can accommodate 256 SGL descriptors.
382 static int nvme_pci_npages_sgl(void)
384 return DIV_ROUND_UP(NVME_MAX_SEGS * sizeof(struct nvme_sgl_desc),
388 static size_t nvme_pci_iod_alloc_size(void)
390 size_t npages = max(nvme_pci_npages_prp(), nvme_pci_npages_sgl());
392 return sizeof(__le64 *) * npages +
393 sizeof(struct scatterlist) * NVME_MAX_SEGS;
396 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
397 unsigned int hctx_idx)
399 struct nvme_dev *dev = data;
400 struct nvme_queue *nvmeq = &dev->queues[0];
402 WARN_ON(hctx_idx != 0);
403 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
405 hctx->driver_data = nvmeq;
409 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
410 unsigned int hctx_idx)
412 struct nvme_dev *dev = data;
413 struct nvme_queue *nvmeq = &dev->queues[hctx_idx + 1];
415 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
416 hctx->driver_data = nvmeq;
420 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
421 unsigned int hctx_idx, unsigned int numa_node)
423 struct nvme_dev *dev = set->driver_data;
424 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
425 int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
426 struct nvme_queue *nvmeq = &dev->queues[queue_idx];
431 nvme_req(req)->ctrl = &dev->ctrl;
435 static int queue_irq_offset(struct nvme_dev *dev)
437 /* if we have more than 1 vec, admin queue offsets us by 1 */
438 if (dev->num_vecs > 1)
444 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
446 struct nvme_dev *dev = set->driver_data;
449 offset = queue_irq_offset(dev);
450 for (i = 0, qoff = 0; i < set->nr_maps; i++) {
451 struct blk_mq_queue_map *map = &set->map[i];
453 map->nr_queues = dev->io_queues[i];
454 if (!map->nr_queues) {
455 BUG_ON(i == HCTX_TYPE_DEFAULT);
460 * The poll queue(s) doesn't have an IRQ (and hence IRQ
461 * affinity), so use the regular blk-mq cpu mapping
463 map->queue_offset = qoff;
464 if (i != HCTX_TYPE_POLL && offset)
465 blk_mq_pci_map_queues(map, to_pci_dev(dev->dev), offset);
467 blk_mq_map_queues(map);
468 qoff += map->nr_queues;
469 offset += map->nr_queues;
476 * Write sq tail if we are asked to, or if the next command would wrap.
478 static inline void nvme_write_sq_db(struct nvme_queue *nvmeq, bool write_sq)
481 u16 next_tail = nvmeq->sq_tail + 1;
483 if (next_tail == nvmeq->q_depth)
485 if (next_tail != nvmeq->last_sq_tail)
489 if (nvme_dbbuf_update_and_check_event(nvmeq->sq_tail,
490 nvmeq->dbbuf_sq_db, nvmeq->dbbuf_sq_ei))
491 writel(nvmeq->sq_tail, nvmeq->q_db);
492 nvmeq->last_sq_tail = nvmeq->sq_tail;
496 * nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
497 * @nvmeq: The queue to use
498 * @cmd: The command to send
499 * @write_sq: whether to write to the SQ doorbell
501 static void nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd,
504 spin_lock(&nvmeq->sq_lock);
505 memcpy(nvmeq->sq_cmds + (nvmeq->sq_tail << nvmeq->sqes),
507 if (++nvmeq->sq_tail == nvmeq->q_depth)
509 nvme_write_sq_db(nvmeq, write_sq);
510 spin_unlock(&nvmeq->sq_lock);
513 static void nvme_commit_rqs(struct blk_mq_hw_ctx *hctx)
515 struct nvme_queue *nvmeq = hctx->driver_data;
517 spin_lock(&nvmeq->sq_lock);
518 if (nvmeq->sq_tail != nvmeq->last_sq_tail)
519 nvme_write_sq_db(nvmeq, true);
520 spin_unlock(&nvmeq->sq_lock);
523 static void **nvme_pci_iod_list(struct request *req)
525 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
526 return (void **)(iod->sg + blk_rq_nr_phys_segments(req));
529 static inline bool nvme_pci_use_sgls(struct nvme_dev *dev, struct request *req)
531 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
532 int nseg = blk_rq_nr_phys_segments(req);
533 unsigned int avg_seg_size;
535 avg_seg_size = DIV_ROUND_UP(blk_rq_payload_bytes(req), nseg);
537 if (!(dev->ctrl.sgls & ((1 << 0) | (1 << 1))))
539 if (!iod->nvmeq->qid)
541 if (!sgl_threshold || avg_seg_size < sgl_threshold)
546 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
548 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
549 const int last_prp = NVME_CTRL_PAGE_SIZE / sizeof(__le64) - 1;
550 dma_addr_t dma_addr = iod->first_dma, next_dma_addr;
554 dma_unmap_page(dev->dev, dma_addr, iod->dma_len,
559 WARN_ON_ONCE(!iod->nents);
561 if (is_pci_p2pdma_page(sg_page(iod->sg)))
562 pci_p2pdma_unmap_sg(dev->dev, iod->sg, iod->nents,
565 dma_unmap_sg(dev->dev, iod->sg, iod->nents, rq_dma_dir(req));
568 if (iod->npages == 0)
569 dma_pool_free(dev->prp_small_pool, nvme_pci_iod_list(req)[0],
572 for (i = 0; i < iod->npages; i++) {
573 void *addr = nvme_pci_iod_list(req)[i];
576 struct nvme_sgl_desc *sg_list = addr;
579 le64_to_cpu((sg_list[SGES_PER_PAGE - 1]).addr);
581 __le64 *prp_list = addr;
583 next_dma_addr = le64_to_cpu(prp_list[last_prp]);
586 dma_pool_free(dev->prp_page_pool, addr, dma_addr);
587 dma_addr = next_dma_addr;
590 mempool_free(iod->sg, dev->iod_mempool);
593 static void nvme_print_sgl(struct scatterlist *sgl, int nents)
596 struct scatterlist *sg;
598 for_each_sg(sgl, sg, nents, i) {
599 dma_addr_t phys = sg_phys(sg);
600 pr_warn("sg[%d] phys_addr:%pad offset:%d length:%d "
601 "dma_address:%pad dma_length:%d\n",
602 i, &phys, sg->offset, sg->length, &sg_dma_address(sg),
607 static blk_status_t nvme_pci_setup_prps(struct nvme_dev *dev,
608 struct request *req, struct nvme_rw_command *cmnd)
610 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
611 struct dma_pool *pool;
612 int length = blk_rq_payload_bytes(req);
613 struct scatterlist *sg = iod->sg;
614 int dma_len = sg_dma_len(sg);
615 u64 dma_addr = sg_dma_address(sg);
616 int offset = dma_addr & (NVME_CTRL_PAGE_SIZE - 1);
618 void **list = nvme_pci_iod_list(req);
622 length -= (NVME_CTRL_PAGE_SIZE - offset);
628 dma_len -= (NVME_CTRL_PAGE_SIZE - offset);
630 dma_addr += (NVME_CTRL_PAGE_SIZE - offset);
633 dma_addr = sg_dma_address(sg);
634 dma_len = sg_dma_len(sg);
637 if (length <= NVME_CTRL_PAGE_SIZE) {
638 iod->first_dma = dma_addr;
642 nprps = DIV_ROUND_UP(length, NVME_CTRL_PAGE_SIZE);
643 if (nprps <= (256 / 8)) {
644 pool = dev->prp_small_pool;
647 pool = dev->prp_page_pool;
651 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
653 iod->first_dma = dma_addr;
655 return BLK_STS_RESOURCE;
658 iod->first_dma = prp_dma;
661 if (i == NVME_CTRL_PAGE_SIZE >> 3) {
662 __le64 *old_prp_list = prp_list;
663 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
665 return BLK_STS_RESOURCE;
666 list[iod->npages++] = prp_list;
667 prp_list[0] = old_prp_list[i - 1];
668 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
671 prp_list[i++] = cpu_to_le64(dma_addr);
672 dma_len -= NVME_CTRL_PAGE_SIZE;
673 dma_addr += NVME_CTRL_PAGE_SIZE;
674 length -= NVME_CTRL_PAGE_SIZE;
679 if (unlikely(dma_len < 0))
682 dma_addr = sg_dma_address(sg);
683 dma_len = sg_dma_len(sg);
687 cmnd->dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
688 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma);
693 WARN(DO_ONCE(nvme_print_sgl, iod->sg, iod->nents),
694 "Invalid SGL for payload:%d nents:%d\n",
695 blk_rq_payload_bytes(req), iod->nents);
696 return BLK_STS_IOERR;
699 static void nvme_pci_sgl_set_data(struct nvme_sgl_desc *sge,
700 struct scatterlist *sg)
702 sge->addr = cpu_to_le64(sg_dma_address(sg));
703 sge->length = cpu_to_le32(sg_dma_len(sg));
704 sge->type = NVME_SGL_FMT_DATA_DESC << 4;
707 static void nvme_pci_sgl_set_seg(struct nvme_sgl_desc *sge,
708 dma_addr_t dma_addr, int entries)
710 sge->addr = cpu_to_le64(dma_addr);
711 if (entries < SGES_PER_PAGE) {
712 sge->length = cpu_to_le32(entries * sizeof(*sge));
713 sge->type = NVME_SGL_FMT_LAST_SEG_DESC << 4;
715 sge->length = cpu_to_le32(PAGE_SIZE);
716 sge->type = NVME_SGL_FMT_SEG_DESC << 4;
720 static blk_status_t nvme_pci_setup_sgls(struct nvme_dev *dev,
721 struct request *req, struct nvme_rw_command *cmd, int entries)
723 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
724 struct dma_pool *pool;
725 struct nvme_sgl_desc *sg_list;
726 struct scatterlist *sg = iod->sg;
730 /* setting the transfer type as SGL */
731 cmd->flags = NVME_CMD_SGL_METABUF;
734 nvme_pci_sgl_set_data(&cmd->dptr.sgl, sg);
738 if (entries <= (256 / sizeof(struct nvme_sgl_desc))) {
739 pool = dev->prp_small_pool;
742 pool = dev->prp_page_pool;
746 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
749 return BLK_STS_RESOURCE;
752 nvme_pci_iod_list(req)[0] = sg_list;
753 iod->first_dma = sgl_dma;
755 nvme_pci_sgl_set_seg(&cmd->dptr.sgl, sgl_dma, entries);
758 if (i == SGES_PER_PAGE) {
759 struct nvme_sgl_desc *old_sg_desc = sg_list;
760 struct nvme_sgl_desc *link = &old_sg_desc[i - 1];
762 sg_list = dma_pool_alloc(pool, GFP_ATOMIC, &sgl_dma);
764 return BLK_STS_RESOURCE;
767 nvme_pci_iod_list(req)[iod->npages++] = sg_list;
768 sg_list[i++] = *link;
769 nvme_pci_sgl_set_seg(link, sgl_dma, entries);
772 nvme_pci_sgl_set_data(&sg_list[i++], sg);
774 } while (--entries > 0);
779 static blk_status_t nvme_setup_prp_simple(struct nvme_dev *dev,
780 struct request *req, struct nvme_rw_command *cmnd,
783 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
784 unsigned int offset = bv->bv_offset & (NVME_CTRL_PAGE_SIZE - 1);
785 unsigned int first_prp_len = NVME_CTRL_PAGE_SIZE - offset;
787 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
788 if (dma_mapping_error(dev->dev, iod->first_dma))
789 return BLK_STS_RESOURCE;
790 iod->dma_len = bv->bv_len;
792 cmnd->dptr.prp1 = cpu_to_le64(iod->first_dma);
793 if (bv->bv_len > first_prp_len)
794 cmnd->dptr.prp2 = cpu_to_le64(iod->first_dma + first_prp_len);
798 static blk_status_t nvme_setup_sgl_simple(struct nvme_dev *dev,
799 struct request *req, struct nvme_rw_command *cmnd,
802 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
804 iod->first_dma = dma_map_bvec(dev->dev, bv, rq_dma_dir(req), 0);
805 if (dma_mapping_error(dev->dev, iod->first_dma))
806 return BLK_STS_RESOURCE;
807 iod->dma_len = bv->bv_len;
809 cmnd->flags = NVME_CMD_SGL_METABUF;
810 cmnd->dptr.sgl.addr = cpu_to_le64(iod->first_dma);
811 cmnd->dptr.sgl.length = cpu_to_le32(iod->dma_len);
812 cmnd->dptr.sgl.type = NVME_SGL_FMT_DATA_DESC << 4;
816 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
817 struct nvme_command *cmnd)
819 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
820 blk_status_t ret = BLK_STS_RESOURCE;
823 if (blk_rq_nr_phys_segments(req) == 1) {
824 struct bio_vec bv = req_bvec(req);
826 if (!is_pci_p2pdma_page(bv.bv_page)) {
827 if (bv.bv_offset + bv.bv_len <= NVME_CTRL_PAGE_SIZE * 2)
828 return nvme_setup_prp_simple(dev, req,
831 if (iod->nvmeq->qid &&
832 dev->ctrl.sgls & ((1 << 0) | (1 << 1)))
833 return nvme_setup_sgl_simple(dev, req,
839 iod->sg = mempool_alloc(dev->iod_mempool, GFP_ATOMIC);
841 return BLK_STS_RESOURCE;
842 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
843 iod->nents = blk_rq_map_sg(req->q, req, iod->sg);
847 if (is_pci_p2pdma_page(sg_page(iod->sg)))
848 nr_mapped = pci_p2pdma_map_sg_attrs(dev->dev, iod->sg,
849 iod->nents, rq_dma_dir(req), DMA_ATTR_NO_WARN);
851 nr_mapped = dma_map_sg_attrs(dev->dev, iod->sg, iod->nents,
852 rq_dma_dir(req), DMA_ATTR_NO_WARN);
856 iod->use_sgl = nvme_pci_use_sgls(dev, req);
858 ret = nvme_pci_setup_sgls(dev, req, &cmnd->rw, nr_mapped);
860 ret = nvme_pci_setup_prps(dev, req, &cmnd->rw);
862 if (ret != BLK_STS_OK)
863 nvme_unmap_data(dev, req);
867 static blk_status_t nvme_map_metadata(struct nvme_dev *dev, struct request *req,
868 struct nvme_command *cmnd)
870 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
872 iod->meta_dma = dma_map_bvec(dev->dev, rq_integrity_vec(req),
874 if (dma_mapping_error(dev->dev, iod->meta_dma))
875 return BLK_STS_IOERR;
876 cmnd->rw.metadata = cpu_to_le64(iod->meta_dma);
881 * NOTE: ns is NULL when called on the admin queue.
883 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
884 const struct blk_mq_queue_data *bd)
886 struct nvme_ns *ns = hctx->queue->queuedata;
887 struct nvme_queue *nvmeq = hctx->driver_data;
888 struct nvme_dev *dev = nvmeq->dev;
889 struct request *req = bd->rq;
890 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
891 struct nvme_command cmnd;
899 * We should not need to do this, but we're still using this to
900 * ensure we can drain requests on a dying queue.
902 if (unlikely(!test_bit(NVMEQ_ENABLED, &nvmeq->flags)))
903 return BLK_STS_IOERR;
905 ret = nvme_setup_cmd(ns, req, &cmnd);
909 if (blk_rq_nr_phys_segments(req)) {
910 ret = nvme_map_data(dev, req, &cmnd);
915 if (blk_integrity_rq(req)) {
916 ret = nvme_map_metadata(dev, req, &cmnd);
921 blk_mq_start_request(req);
922 nvme_submit_cmd(nvmeq, &cmnd, bd->last);
925 nvme_unmap_data(dev, req);
927 nvme_cleanup_cmd(req);
931 static void nvme_pci_complete_rq(struct request *req)
933 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
934 struct nvme_dev *dev = iod->nvmeq->dev;
936 if (blk_integrity_rq(req))
937 dma_unmap_page(dev->dev, iod->meta_dma,
938 rq_integrity_vec(req)->bv_len, rq_data_dir(req));
939 if (blk_rq_nr_phys_segments(req))
940 nvme_unmap_data(dev, req);
941 nvme_complete_rq(req);
944 /* We read the CQE phase first to check if the rest of the entry is valid */
945 static inline bool nvme_cqe_pending(struct nvme_queue *nvmeq)
947 struct nvme_completion *hcqe = &nvmeq->cqes[nvmeq->cq_head];
949 return (le16_to_cpu(READ_ONCE(hcqe->status)) & 1) == nvmeq->cq_phase;
952 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
954 u16 head = nvmeq->cq_head;
956 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
958 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
961 static inline struct blk_mq_tags *nvme_queue_tagset(struct nvme_queue *nvmeq)
964 return nvmeq->dev->admin_tagset.tags[0];
965 return nvmeq->dev->tagset.tags[nvmeq->qid - 1];
968 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq, u16 idx)
970 struct nvme_completion *cqe = &nvmeq->cqes[idx];
971 __u16 command_id = READ_ONCE(cqe->command_id);
975 * AEN requests are special as they don't time out and can
976 * survive any kind of queue freeze and often don't respond to
977 * aborts. We don't even bother to allocate a struct request
978 * for them but rather special case them here.
980 if (unlikely(nvme_is_aen_req(nvmeq->qid, command_id))) {
981 nvme_complete_async_event(&nvmeq->dev->ctrl,
982 cqe->status, &cqe->result);
986 req = blk_mq_tag_to_rq(nvme_queue_tagset(nvmeq), command_id);
987 if (unlikely(!req)) {
988 dev_warn(nvmeq->dev->ctrl.device,
989 "invalid id %d completed on queue %d\n",
990 command_id, le16_to_cpu(cqe->sq_id));
994 trace_nvme_sq(req, cqe->sq_head, nvmeq->sq_tail);
995 if (!nvme_try_complete_req(req, cqe->status, cqe->result))
996 nvme_pci_complete_rq(req);
999 static inline void nvme_update_cq_head(struct nvme_queue *nvmeq)
1001 u16 tmp = nvmeq->cq_head + 1;
1003 if (tmp == nvmeq->q_depth) {
1005 nvmeq->cq_phase ^= 1;
1007 nvmeq->cq_head = tmp;
1011 static inline int nvme_process_cq(struct nvme_queue *nvmeq)
1015 while (nvme_cqe_pending(nvmeq)) {
1018 * load-load control dependency between phase and the rest of
1019 * the cqe requires a full read memory barrier
1022 nvme_handle_cqe(nvmeq, nvmeq->cq_head);
1023 nvme_update_cq_head(nvmeq);
1027 nvme_ring_cq_doorbell(nvmeq);
1031 static irqreturn_t nvme_irq(int irq, void *data)
1033 struct nvme_queue *nvmeq = data;
1034 irqreturn_t ret = IRQ_NONE;
1037 * The rmb/wmb pair ensures we see all updates from a previous run of
1038 * the irq handler, even if that was on another CPU.
1041 if (nvme_process_cq(nvmeq))
1048 static irqreturn_t nvme_irq_check(int irq, void *data)
1050 struct nvme_queue *nvmeq = data;
1052 if (nvme_cqe_pending(nvmeq))
1053 return IRQ_WAKE_THREAD;
1058 * Poll for completions for any interrupt driven queue
1059 * Can be called from any context.
1061 static void nvme_poll_irqdisable(struct nvme_queue *nvmeq)
1063 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1065 WARN_ON_ONCE(test_bit(NVMEQ_POLLED, &nvmeq->flags));
1067 disable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1068 nvme_process_cq(nvmeq);
1069 enable_irq(pci_irq_vector(pdev, nvmeq->cq_vector));
1072 static int nvme_poll(struct blk_mq_hw_ctx *hctx)
1074 struct nvme_queue *nvmeq = hctx->driver_data;
1077 if (!nvme_cqe_pending(nvmeq))
1080 spin_lock(&nvmeq->cq_poll_lock);
1081 found = nvme_process_cq(nvmeq);
1082 spin_unlock(&nvmeq->cq_poll_lock);
1087 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl)
1089 struct nvme_dev *dev = to_nvme_dev(ctrl);
1090 struct nvme_queue *nvmeq = &dev->queues[0];
1091 struct nvme_command c;
1093 memset(&c, 0, sizeof(c));
1094 c.common.opcode = nvme_admin_async_event;
1095 c.common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1096 nvme_submit_cmd(nvmeq, &c, true);
1099 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
1101 struct nvme_command c;
1103 memset(&c, 0, sizeof(c));
1104 c.delete_queue.opcode = opcode;
1105 c.delete_queue.qid = cpu_to_le16(id);
1107 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1110 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
1111 struct nvme_queue *nvmeq, s16 vector)
1113 struct nvme_command c;
1114 int flags = NVME_QUEUE_PHYS_CONTIG;
1116 if (!test_bit(NVMEQ_POLLED, &nvmeq->flags))
1117 flags |= NVME_CQ_IRQ_ENABLED;
1120 * Note: we (ab)use the fact that the prp fields survive if no data
1121 * is attached to the request.
1123 memset(&c, 0, sizeof(c));
1124 c.create_cq.opcode = nvme_admin_create_cq;
1125 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
1126 c.create_cq.cqid = cpu_to_le16(qid);
1127 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1128 c.create_cq.cq_flags = cpu_to_le16(flags);
1129 c.create_cq.irq_vector = cpu_to_le16(vector);
1131 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1134 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
1135 struct nvme_queue *nvmeq)
1137 struct nvme_ctrl *ctrl = &dev->ctrl;
1138 struct nvme_command c;
1139 int flags = NVME_QUEUE_PHYS_CONTIG;
1142 * Some drives have a bug that auto-enables WRRU if MEDIUM isn't
1143 * set. Since URGENT priority is zeroes, it makes all queues
1146 if (ctrl->quirks & NVME_QUIRK_MEDIUM_PRIO_SQ)
1147 flags |= NVME_SQ_PRIO_MEDIUM;
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_sq.opcode = nvme_admin_create_sq;
1155 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
1156 c.create_sq.sqid = cpu_to_le16(qid);
1157 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
1158 c.create_sq.sq_flags = cpu_to_le16(flags);
1159 c.create_sq.cqid = cpu_to_le16(qid);
1161 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1164 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
1166 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
1169 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
1171 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
1174 static void abort_endio(struct request *req, blk_status_t error)
1176 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1177 struct nvme_queue *nvmeq = iod->nvmeq;
1179 dev_warn(nvmeq->dev->ctrl.device,
1180 "Abort status: 0x%x", nvme_req(req)->status);
1181 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
1182 blk_mq_free_request(req);
1185 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1187 /* If true, indicates loss of adapter communication, possibly by a
1188 * NVMe Subsystem reset.
1190 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1192 /* If there is a reset/reinit ongoing, we shouldn't reset again. */
1193 switch (dev->ctrl.state) {
1194 case NVME_CTRL_RESETTING:
1195 case NVME_CTRL_CONNECTING:
1201 /* We shouldn't reset unless the controller is on fatal error state
1202 * _or_ if we lost the communication with it.
1204 if (!(csts & NVME_CSTS_CFS) && !nssro)
1210 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1212 /* Read a config register to help see what died. */
1216 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1218 if (result == PCIBIOS_SUCCESSFUL)
1219 dev_warn(dev->ctrl.device,
1220 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1223 dev_warn(dev->ctrl.device,
1224 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1228 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
1230 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
1231 struct nvme_queue *nvmeq = iod->nvmeq;
1232 struct nvme_dev *dev = nvmeq->dev;
1233 struct request *abort_req;
1234 struct nvme_command cmd;
1235 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1237 /* If PCI error recovery process is happening, we cannot reset or
1238 * the recovery mechanism will surely fail.
1241 if (pci_channel_offline(to_pci_dev(dev->dev)))
1242 return BLK_EH_RESET_TIMER;
1245 * Reset immediately if the controller is failed
1247 if (nvme_should_reset(dev, csts)) {
1248 nvme_warn_reset(dev, csts);
1249 nvme_dev_disable(dev, false);
1250 nvme_reset_ctrl(&dev->ctrl);
1255 * Did we miss an interrupt?
1257 if (test_bit(NVMEQ_POLLED, &nvmeq->flags))
1258 nvme_poll(req->mq_hctx);
1260 nvme_poll_irqdisable(nvmeq);
1262 if (blk_mq_request_completed(req)) {
1263 dev_warn(dev->ctrl.device,
1264 "I/O %d QID %d timeout, completion polled\n",
1265 req->tag, nvmeq->qid);
1270 * Shutdown immediately if controller times out while starting. The
1271 * reset work will see the pci device disabled when it gets the forced
1272 * cancellation error. All outstanding requests are completed on
1273 * shutdown, so we return BLK_EH_DONE.
1275 switch (dev->ctrl.state) {
1276 case NVME_CTRL_CONNECTING:
1277 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
1279 case NVME_CTRL_DELETING:
1280 dev_warn_ratelimited(dev->ctrl.device,
1281 "I/O %d QID %d timeout, disable controller\n",
1282 req->tag, nvmeq->qid);
1283 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1284 nvme_dev_disable(dev, true);
1286 case NVME_CTRL_RESETTING:
1287 return BLK_EH_RESET_TIMER;
1293 * Shutdown the controller immediately and schedule a reset if the
1294 * command was already aborted once before and still hasn't been
1295 * returned to the driver, or if this is the admin queue.
1297 if (!nvmeq->qid || iod->aborted) {
1298 dev_warn(dev->ctrl.device,
1299 "I/O %d QID %d timeout, reset controller\n",
1300 req->tag, nvmeq->qid);
1301 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1302 nvme_dev_disable(dev, false);
1303 nvme_reset_ctrl(&dev->ctrl);
1308 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1309 atomic_inc(&dev->ctrl.abort_limit);
1310 return BLK_EH_RESET_TIMER;
1314 memset(&cmd, 0, sizeof(cmd));
1315 cmd.abort.opcode = nvme_admin_abort_cmd;
1316 cmd.abort.cid = req->tag;
1317 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1319 dev_warn(nvmeq->dev->ctrl.device,
1320 "I/O %d QID %d timeout, aborting\n",
1321 req->tag, nvmeq->qid);
1323 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1325 if (IS_ERR(abort_req)) {
1326 atomic_inc(&dev->ctrl.abort_limit);
1327 return BLK_EH_RESET_TIMER;
1330 abort_req->end_io_data = NULL;
1331 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1334 * The aborted req will be completed on receiving the abort req.
1335 * We enable the timer again. If hit twice, it'll cause a device reset,
1336 * as the device then is in a faulty state.
1338 return BLK_EH_RESET_TIMER;
1341 static void nvme_free_queue(struct nvme_queue *nvmeq)
1343 dma_free_coherent(nvmeq->dev->dev, CQ_SIZE(nvmeq),
1344 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1345 if (!nvmeq->sq_cmds)
1348 if (test_and_clear_bit(NVMEQ_SQ_CMB, &nvmeq->flags)) {
1349 pci_free_p2pmem(to_pci_dev(nvmeq->dev->dev),
1350 nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1352 dma_free_coherent(nvmeq->dev->dev, SQ_SIZE(nvmeq),
1353 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1357 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1361 for (i = dev->ctrl.queue_count - 1; i >= lowest; i--) {
1362 dev->ctrl.queue_count--;
1363 nvme_free_queue(&dev->queues[i]);
1368 * nvme_suspend_queue - put queue into suspended state
1369 * @nvmeq: queue to suspend
1371 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1373 if (!test_and_clear_bit(NVMEQ_ENABLED, &nvmeq->flags))
1376 /* ensure that nvme_queue_rq() sees NVMEQ_ENABLED cleared */
1379 nvmeq->dev->online_queues--;
1380 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1381 blk_mq_quiesce_queue(nvmeq->dev->ctrl.admin_q);
1382 if (!test_and_clear_bit(NVMEQ_POLLED, &nvmeq->flags))
1383 pci_free_irq(to_pci_dev(nvmeq->dev->dev), nvmeq->cq_vector, nvmeq);
1387 static void nvme_suspend_io_queues(struct nvme_dev *dev)
1391 for (i = dev->ctrl.queue_count - 1; i > 0; i--)
1392 nvme_suspend_queue(&dev->queues[i]);
1395 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1397 struct nvme_queue *nvmeq = &dev->queues[0];
1400 nvme_shutdown_ctrl(&dev->ctrl);
1402 nvme_disable_ctrl(&dev->ctrl);
1404 nvme_poll_irqdisable(nvmeq);
1408 * Called only on a device that has been disabled and after all other threads
1409 * that can check this device's completion queues have synced, except
1410 * nvme_poll(). This is the last chance for the driver to see a natural
1411 * completion before nvme_cancel_request() terminates all incomplete requests.
1413 static void nvme_reap_pending_cqes(struct nvme_dev *dev)
1417 for (i = dev->ctrl.queue_count - 1; i > 0; i--) {
1418 spin_lock(&dev->queues[i].cq_poll_lock);
1419 nvme_process_cq(&dev->queues[i]);
1420 spin_unlock(&dev->queues[i].cq_poll_lock);
1424 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1427 int q_depth = dev->q_depth;
1428 unsigned q_size_aligned = roundup(q_depth * entry_size,
1429 NVME_CTRL_PAGE_SIZE);
1431 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1432 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1434 mem_per_q = round_down(mem_per_q, NVME_CTRL_PAGE_SIZE);
1435 q_depth = div_u64(mem_per_q, entry_size);
1438 * Ensure the reduced q_depth is above some threshold where it
1439 * would be better to map queues in system memory with the
1449 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1452 struct pci_dev *pdev = to_pci_dev(dev->dev);
1454 if (qid && dev->cmb_use_sqes && (dev->cmbsz & NVME_CMBSZ_SQS)) {
1455 nvmeq->sq_cmds = pci_alloc_p2pmem(pdev, SQ_SIZE(nvmeq));
1456 if (nvmeq->sq_cmds) {
1457 nvmeq->sq_dma_addr = pci_p2pmem_virt_to_bus(pdev,
1459 if (nvmeq->sq_dma_addr) {
1460 set_bit(NVMEQ_SQ_CMB, &nvmeq->flags);
1464 pci_free_p2pmem(pdev, nvmeq->sq_cmds, SQ_SIZE(nvmeq));
1468 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(nvmeq),
1469 &nvmeq->sq_dma_addr, GFP_KERNEL);
1470 if (!nvmeq->sq_cmds)
1475 static int nvme_alloc_queue(struct nvme_dev *dev, int qid, int depth)
1477 struct nvme_queue *nvmeq = &dev->queues[qid];
1479 if (dev->ctrl.queue_count > qid)
1482 nvmeq->sqes = qid ? dev->io_sqes : NVME_ADM_SQES;
1483 nvmeq->q_depth = depth;
1484 nvmeq->cqes = dma_alloc_coherent(dev->dev, CQ_SIZE(nvmeq),
1485 &nvmeq->cq_dma_addr, GFP_KERNEL);
1489 if (nvme_alloc_sq_cmds(dev, nvmeq, qid))
1493 spin_lock_init(&nvmeq->sq_lock);
1494 spin_lock_init(&nvmeq->cq_poll_lock);
1496 nvmeq->cq_phase = 1;
1497 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1499 dev->ctrl.queue_count++;
1504 dma_free_coherent(dev->dev, CQ_SIZE(nvmeq), (void *)nvmeq->cqes,
1505 nvmeq->cq_dma_addr);
1510 static int queue_request_irq(struct nvme_queue *nvmeq)
1512 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1513 int nr = nvmeq->dev->ctrl.instance;
1515 if (use_threaded_interrupts) {
1516 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1517 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1519 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1520 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1524 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1526 struct nvme_dev *dev = nvmeq->dev;
1529 nvmeq->last_sq_tail = 0;
1531 nvmeq->cq_phase = 1;
1532 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1533 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq));
1534 nvme_dbbuf_init(dev, nvmeq, qid);
1535 dev->online_queues++;
1536 wmb(); /* ensure the first interrupt sees the initialization */
1539 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid, bool polled)
1541 struct nvme_dev *dev = nvmeq->dev;
1545 clear_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
1548 * A queue's vector matches the queue identifier unless the controller
1549 * has only one vector available.
1552 vector = dev->num_vecs == 1 ? 0 : qid;
1554 set_bit(NVMEQ_POLLED, &nvmeq->flags);
1556 result = adapter_alloc_cq(dev, qid, nvmeq, vector);
1560 result = adapter_alloc_sq(dev, qid, nvmeq);
1566 nvmeq->cq_vector = vector;
1567 nvme_init_queue(nvmeq, qid);
1570 result = queue_request_irq(nvmeq);
1575 set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1579 dev->online_queues--;
1580 adapter_delete_sq(dev, qid);
1582 adapter_delete_cq(dev, qid);
1586 static const struct blk_mq_ops nvme_mq_admin_ops = {
1587 .queue_rq = nvme_queue_rq,
1588 .complete = nvme_pci_complete_rq,
1589 .init_hctx = nvme_admin_init_hctx,
1590 .init_request = nvme_init_request,
1591 .timeout = nvme_timeout,
1594 static const struct blk_mq_ops nvme_mq_ops = {
1595 .queue_rq = nvme_queue_rq,
1596 .complete = nvme_pci_complete_rq,
1597 .commit_rqs = nvme_commit_rqs,
1598 .init_hctx = nvme_init_hctx,
1599 .init_request = nvme_init_request,
1600 .map_queues = nvme_pci_map_queues,
1601 .timeout = nvme_timeout,
1605 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1607 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1609 * If the controller was reset during removal, it's possible
1610 * user requests may be waiting on a stopped queue. Start the
1611 * queue to flush these to completion.
1613 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1614 blk_cleanup_queue(dev->ctrl.admin_q);
1615 blk_mq_free_tag_set(&dev->admin_tagset);
1619 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1621 if (!dev->ctrl.admin_q) {
1622 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1623 dev->admin_tagset.nr_hw_queues = 1;
1625 dev->admin_tagset.queue_depth = NVME_AQ_MQ_TAG_DEPTH;
1626 dev->admin_tagset.timeout = NVME_ADMIN_TIMEOUT;
1627 dev->admin_tagset.numa_node = dev->ctrl.numa_node;
1628 dev->admin_tagset.cmd_size = sizeof(struct nvme_iod);
1629 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1630 dev->admin_tagset.driver_data = dev;
1632 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1634 dev->ctrl.admin_tagset = &dev->admin_tagset;
1636 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1637 if (IS_ERR(dev->ctrl.admin_q)) {
1638 blk_mq_free_tag_set(&dev->admin_tagset);
1641 if (!blk_get_queue(dev->ctrl.admin_q)) {
1642 nvme_dev_remove_admin(dev);
1643 dev->ctrl.admin_q = NULL;
1647 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
1652 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1654 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1657 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1659 struct pci_dev *pdev = to_pci_dev(dev->dev);
1661 if (size <= dev->bar_mapped_size)
1663 if (size > pci_resource_len(pdev, 0))
1667 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1669 dev->bar_mapped_size = 0;
1672 dev->bar_mapped_size = size;
1673 dev->dbs = dev->bar + NVME_REG_DBS;
1678 static int nvme_pci_configure_admin_queue(struct nvme_dev *dev)
1682 struct nvme_queue *nvmeq;
1684 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1688 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1689 NVME_CAP_NSSRC(dev->ctrl.cap) : 0;
1691 if (dev->subsystem &&
1692 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1693 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1695 result = nvme_disable_ctrl(&dev->ctrl);
1699 result = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1703 dev->ctrl.numa_node = dev_to_node(dev->dev);
1705 nvmeq = &dev->queues[0];
1706 aqa = nvmeq->q_depth - 1;
1709 writel(aqa, dev->bar + NVME_REG_AQA);
1710 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1711 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1713 result = nvme_enable_ctrl(&dev->ctrl);
1717 nvmeq->cq_vector = 0;
1718 nvme_init_queue(nvmeq, 0);
1719 result = queue_request_irq(nvmeq);
1721 dev->online_queues--;
1725 set_bit(NVMEQ_ENABLED, &nvmeq->flags);
1729 static int nvme_create_io_queues(struct nvme_dev *dev)
1731 unsigned i, max, rw_queues;
1734 for (i = dev->ctrl.queue_count; i <= dev->max_qid; i++) {
1735 if (nvme_alloc_queue(dev, i, dev->q_depth)) {
1741 max = min(dev->max_qid, dev->ctrl.queue_count - 1);
1742 if (max != 1 && dev->io_queues[HCTX_TYPE_POLL]) {
1743 rw_queues = dev->io_queues[HCTX_TYPE_DEFAULT] +
1744 dev->io_queues[HCTX_TYPE_READ];
1749 for (i = dev->online_queues; i <= max; i++) {
1750 bool polled = i > rw_queues;
1752 ret = nvme_create_queue(&dev->queues[i], i, polled);
1758 * Ignore failing Create SQ/CQ commands, we can continue with less
1759 * than the desired amount of queues, and even a controller without
1760 * I/O queues can still be used to issue admin commands. This might
1761 * be useful to upgrade a buggy firmware for example.
1763 return ret >= 0 ? 0 : ret;
1766 static ssize_t nvme_cmb_show(struct device *dev,
1767 struct device_attribute *attr,
1770 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1772 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1773 ndev->cmbloc, ndev->cmbsz);
1775 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1777 static u64 nvme_cmb_size_unit(struct nvme_dev *dev)
1779 u8 szu = (dev->cmbsz >> NVME_CMBSZ_SZU_SHIFT) & NVME_CMBSZ_SZU_MASK;
1781 return 1ULL << (12 + 4 * szu);
1784 static u32 nvme_cmb_size(struct nvme_dev *dev)
1786 return (dev->cmbsz >> NVME_CMBSZ_SZ_SHIFT) & NVME_CMBSZ_SZ_MASK;
1789 static void nvme_map_cmb(struct nvme_dev *dev)
1792 resource_size_t bar_size;
1793 struct pci_dev *pdev = to_pci_dev(dev->dev);
1799 if (NVME_CAP_CMBS(dev->ctrl.cap))
1800 writel(NVME_CMBMSC_CRE, dev->bar + NVME_REG_CMBMSC);
1802 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1805 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1807 size = nvme_cmb_size_unit(dev) * nvme_cmb_size(dev);
1808 offset = nvme_cmb_size_unit(dev) * NVME_CMB_OFST(dev->cmbloc);
1809 bar = NVME_CMB_BIR(dev->cmbloc);
1810 bar_size = pci_resource_len(pdev, bar);
1812 if (offset > bar_size)
1816 * Tell the controller about the host side address mapping the CMB,
1817 * and enable CMB decoding for the NVMe 1.4+ scheme:
1819 if (NVME_CAP_CMBS(dev->ctrl.cap)) {
1820 hi_lo_writeq(NVME_CMBMSC_CRE | NVME_CMBMSC_CMSE |
1821 (pci_bus_address(pdev, bar) + offset),
1822 dev->bar + NVME_REG_CMBMSC);
1826 * Controllers may support a CMB size larger than their BAR,
1827 * for example, due to being behind a bridge. Reduce the CMB to
1828 * the reported size of the BAR
1830 if (size > bar_size - offset)
1831 size = bar_size - offset;
1833 if (pci_p2pdma_add_resource(pdev, bar, size, offset)) {
1834 dev_warn(dev->ctrl.device,
1835 "failed to register the CMB\n");
1839 dev->cmb_size = size;
1840 dev->cmb_use_sqes = use_cmb_sqes && (dev->cmbsz & NVME_CMBSZ_SQS);
1842 if ((dev->cmbsz & (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS)) ==
1843 (NVME_CMBSZ_WDS | NVME_CMBSZ_RDS))
1844 pci_p2pmem_publish(pdev, true);
1846 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1847 &dev_attr_cmb.attr, NULL))
1848 dev_warn(dev->ctrl.device,
1849 "failed to add sysfs attribute for CMB\n");
1852 static inline void nvme_release_cmb(struct nvme_dev *dev)
1854 if (dev->cmb_size) {
1855 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1856 &dev_attr_cmb.attr, NULL);
1861 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1863 u32 host_mem_size = dev->host_mem_size >> NVME_CTRL_PAGE_SHIFT;
1864 u64 dma_addr = dev->host_mem_descs_dma;
1865 struct nvme_command c;
1868 memset(&c, 0, sizeof(c));
1869 c.features.opcode = nvme_admin_set_features;
1870 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1871 c.features.dword11 = cpu_to_le32(bits);
1872 c.features.dword12 = cpu_to_le32(host_mem_size);
1873 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1874 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1875 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1877 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1879 dev_warn(dev->ctrl.device,
1880 "failed to set host mem (err %d, flags %#x).\n",
1886 static void nvme_free_host_mem(struct nvme_dev *dev)
1890 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1891 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1892 size_t size = le32_to_cpu(desc->size) * NVME_CTRL_PAGE_SIZE;
1894 dma_free_attrs(dev->dev, size, dev->host_mem_desc_bufs[i],
1895 le64_to_cpu(desc->addr),
1896 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1899 kfree(dev->host_mem_desc_bufs);
1900 dev->host_mem_desc_bufs = NULL;
1901 dma_free_coherent(dev->dev,
1902 dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs),
1903 dev->host_mem_descs, dev->host_mem_descs_dma);
1904 dev->host_mem_descs = NULL;
1905 dev->nr_host_mem_descs = 0;
1908 static int __nvme_alloc_host_mem(struct nvme_dev *dev, u64 preferred,
1911 struct nvme_host_mem_buf_desc *descs;
1912 u32 max_entries, len;
1913 dma_addr_t descs_dma;
1918 tmp = (preferred + chunk_size - 1);
1919 do_div(tmp, chunk_size);
1922 if (dev->ctrl.hmmaxd && dev->ctrl.hmmaxd < max_entries)
1923 max_entries = dev->ctrl.hmmaxd;
1925 descs = dma_alloc_coherent(dev->dev, max_entries * sizeof(*descs),
1926 &descs_dma, GFP_KERNEL);
1930 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1932 goto out_free_descs;
1934 for (size = 0; size < preferred && i < max_entries; size += len) {
1935 dma_addr_t dma_addr;
1937 len = min_t(u64, chunk_size, preferred - size);
1938 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1939 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1943 descs[i].addr = cpu_to_le64(dma_addr);
1944 descs[i].size = cpu_to_le32(len / NVME_CTRL_PAGE_SIZE);
1951 dev->nr_host_mem_descs = i;
1952 dev->host_mem_size = size;
1953 dev->host_mem_descs = descs;
1954 dev->host_mem_descs_dma = descs_dma;
1955 dev->host_mem_desc_bufs = bufs;
1960 size_t size = le32_to_cpu(descs[i].size) * NVME_CTRL_PAGE_SIZE;
1962 dma_free_attrs(dev->dev, size, bufs[i],
1963 le64_to_cpu(descs[i].addr),
1964 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1969 dma_free_coherent(dev->dev, max_entries * sizeof(*descs), descs,
1972 dev->host_mem_descs = NULL;
1976 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1978 u64 min_chunk = min_t(u64, preferred, PAGE_SIZE * MAX_ORDER_NR_PAGES);
1979 u64 hmminds = max_t(u32, dev->ctrl.hmminds * 4096, PAGE_SIZE * 2);
1982 /* start big and work our way down */
1983 for (chunk_size = min_chunk; chunk_size >= hmminds; chunk_size /= 2) {
1984 if (!__nvme_alloc_host_mem(dev, preferred, chunk_size)) {
1985 if (!min || dev->host_mem_size >= min)
1987 nvme_free_host_mem(dev);
1994 static int nvme_setup_host_mem(struct nvme_dev *dev)
1996 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1997 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1998 u64 min = (u64)dev->ctrl.hmmin * 4096;
1999 u32 enable_bits = NVME_HOST_MEM_ENABLE;
2002 preferred = min(preferred, max);
2004 dev_warn(dev->ctrl.device,
2005 "min host memory (%lld MiB) above limit (%d MiB).\n",
2006 min >> ilog2(SZ_1M), max_host_mem_size_mb);
2007 nvme_free_host_mem(dev);
2012 * If we already have a buffer allocated check if we can reuse it.
2014 if (dev->host_mem_descs) {
2015 if (dev->host_mem_size >= min)
2016 enable_bits |= NVME_HOST_MEM_RETURN;
2018 nvme_free_host_mem(dev);
2021 if (!dev->host_mem_descs) {
2022 if (nvme_alloc_host_mem(dev, min, preferred)) {
2023 dev_warn(dev->ctrl.device,
2024 "failed to allocate host memory buffer.\n");
2025 return 0; /* controller must work without HMB */
2028 dev_info(dev->ctrl.device,
2029 "allocated %lld MiB host memory buffer.\n",
2030 dev->host_mem_size >> ilog2(SZ_1M));
2033 ret = nvme_set_host_mem(dev, enable_bits);
2035 nvme_free_host_mem(dev);
2040 * nirqs is the number of interrupts available for write and read
2041 * queues. The core already reserved an interrupt for the admin queue.
2043 static void nvme_calc_irq_sets(struct irq_affinity *affd, unsigned int nrirqs)
2045 struct nvme_dev *dev = affd->priv;
2046 unsigned int nr_read_queues, nr_write_queues = dev->nr_write_queues;
2049 * If there is no interrupt available for queues, ensure that
2050 * the default queue is set to 1. The affinity set size is
2051 * also set to one, but the irq core ignores it for this case.
2053 * If only one interrupt is available or 'write_queue' == 0, combine
2054 * write and read queues.
2056 * If 'write_queues' > 0, ensure it leaves room for at least one read
2062 } else if (nrirqs == 1 || !nr_write_queues) {
2064 } else if (nr_write_queues >= nrirqs) {
2067 nr_read_queues = nrirqs - nr_write_queues;
2070 dev->io_queues[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2071 affd->set_size[HCTX_TYPE_DEFAULT] = nrirqs - nr_read_queues;
2072 dev->io_queues[HCTX_TYPE_READ] = nr_read_queues;
2073 affd->set_size[HCTX_TYPE_READ] = nr_read_queues;
2074 affd->nr_sets = nr_read_queues ? 2 : 1;
2077 static int nvme_setup_irqs(struct nvme_dev *dev, unsigned int nr_io_queues)
2079 struct pci_dev *pdev = to_pci_dev(dev->dev);
2080 struct irq_affinity affd = {
2082 .calc_sets = nvme_calc_irq_sets,
2085 unsigned int irq_queues, poll_queues;
2088 * Poll queues don't need interrupts, but we need at least one I/O queue
2089 * left over for non-polled I/O.
2091 poll_queues = min(dev->nr_poll_queues, nr_io_queues - 1);
2092 dev->io_queues[HCTX_TYPE_POLL] = poll_queues;
2095 * Initialize for the single interrupt case, will be updated in
2096 * nvme_calc_irq_sets().
2098 dev->io_queues[HCTX_TYPE_DEFAULT] = 1;
2099 dev->io_queues[HCTX_TYPE_READ] = 0;
2102 * We need interrupts for the admin queue and each non-polled I/O queue,
2103 * but some Apple controllers require all queues to use the first
2107 if (!(dev->ctrl.quirks & NVME_QUIRK_SINGLE_VECTOR))
2108 irq_queues += (nr_io_queues - poll_queues);
2109 return pci_alloc_irq_vectors_affinity(pdev, 1, irq_queues,
2110 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY, &affd);
2113 static void nvme_disable_io_queues(struct nvme_dev *dev)
2115 if (__nvme_disable_io_queues(dev, nvme_admin_delete_sq))
2116 __nvme_disable_io_queues(dev, nvme_admin_delete_cq);
2119 static unsigned int nvme_max_io_queues(struct nvme_dev *dev)
2122 * If tags are shared with admin queue (Apple bug), then
2123 * make sure we only use one IO queue.
2125 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2127 return num_possible_cpus() + dev->nr_write_queues + dev->nr_poll_queues;
2130 static int nvme_setup_io_queues(struct nvme_dev *dev)
2132 struct nvme_queue *adminq = &dev->queues[0];
2133 struct pci_dev *pdev = to_pci_dev(dev->dev);
2134 unsigned int nr_io_queues;
2139 * Sample the module parameters once at reset time so that we have
2140 * stable values to work with.
2142 dev->nr_write_queues = write_queues;
2143 dev->nr_poll_queues = poll_queues;
2145 nr_io_queues = dev->nr_allocated_queues - 1;
2146 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
2150 if (nr_io_queues == 0)
2153 clear_bit(NVMEQ_ENABLED, &adminq->flags);
2155 if (dev->cmb_use_sqes) {
2156 result = nvme_cmb_qdepth(dev, nr_io_queues,
2157 sizeof(struct nvme_command));
2159 dev->q_depth = result;
2161 dev->cmb_use_sqes = false;
2165 size = db_bar_size(dev, nr_io_queues);
2166 result = nvme_remap_bar(dev, size);
2169 if (!--nr_io_queues)
2172 adminq->q_db = dev->dbs;
2175 /* Deregister the admin queue's interrupt */
2176 pci_free_irq(pdev, 0, adminq);
2179 * If we enable msix early due to not intx, disable it again before
2180 * setting up the full range we need.
2182 pci_free_irq_vectors(pdev);
2184 result = nvme_setup_irqs(dev, nr_io_queues);
2188 dev->num_vecs = result;
2189 result = max(result - 1, 1);
2190 dev->max_qid = result + dev->io_queues[HCTX_TYPE_POLL];
2193 * Should investigate if there's a performance win from allocating
2194 * more queues than interrupt vectors; it might allow the submission
2195 * path to scale better, even if the receive path is limited by the
2196 * number of interrupts.
2198 result = queue_request_irq(adminq);
2201 set_bit(NVMEQ_ENABLED, &adminq->flags);
2203 result = nvme_create_io_queues(dev);
2204 if (result || dev->online_queues < 2)
2207 if (dev->online_queues - 1 < dev->max_qid) {
2208 nr_io_queues = dev->online_queues - 1;
2209 nvme_disable_io_queues(dev);
2210 nvme_suspend_io_queues(dev);
2213 dev_info(dev->ctrl.device, "%d/%d/%d default/read/poll queues\n",
2214 dev->io_queues[HCTX_TYPE_DEFAULT],
2215 dev->io_queues[HCTX_TYPE_READ],
2216 dev->io_queues[HCTX_TYPE_POLL]);
2220 static void nvme_del_queue_end(struct request *req, blk_status_t error)
2222 struct nvme_queue *nvmeq = req->end_io_data;
2224 blk_mq_free_request(req);
2225 complete(&nvmeq->delete_done);
2228 static void nvme_del_cq_end(struct request *req, blk_status_t error)
2230 struct nvme_queue *nvmeq = req->end_io_data;
2233 set_bit(NVMEQ_DELETE_ERROR, &nvmeq->flags);
2235 nvme_del_queue_end(req, error);
2238 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
2240 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
2241 struct request *req;
2242 struct nvme_command cmd;
2244 memset(&cmd, 0, sizeof(cmd));
2245 cmd.delete_queue.opcode = opcode;
2246 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
2248 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT);
2250 return PTR_ERR(req);
2252 req->end_io_data = nvmeq;
2254 init_completion(&nvmeq->delete_done);
2255 blk_execute_rq_nowait(q, NULL, req, false,
2256 opcode == nvme_admin_delete_cq ?
2257 nvme_del_cq_end : nvme_del_queue_end);
2261 static bool __nvme_disable_io_queues(struct nvme_dev *dev, u8 opcode)
2263 int nr_queues = dev->online_queues - 1, sent = 0;
2264 unsigned long timeout;
2267 timeout = NVME_ADMIN_TIMEOUT;
2268 while (nr_queues > 0) {
2269 if (nvme_delete_queue(&dev->queues[nr_queues], opcode))
2275 struct nvme_queue *nvmeq = &dev->queues[nr_queues + sent];
2277 timeout = wait_for_completion_io_timeout(&nvmeq->delete_done,
2289 static void nvme_dev_add(struct nvme_dev *dev)
2293 if (!dev->ctrl.tagset) {
2294 dev->tagset.ops = &nvme_mq_ops;
2295 dev->tagset.nr_hw_queues = dev->online_queues - 1;
2296 dev->tagset.nr_maps = 2; /* default + read */
2297 if (dev->io_queues[HCTX_TYPE_POLL])
2298 dev->tagset.nr_maps++;
2299 dev->tagset.timeout = NVME_IO_TIMEOUT;
2300 dev->tagset.numa_node = dev->ctrl.numa_node;
2301 dev->tagset.queue_depth = min_t(unsigned int, dev->q_depth,
2302 BLK_MQ_MAX_DEPTH) - 1;
2303 dev->tagset.cmd_size = sizeof(struct nvme_iod);
2304 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
2305 dev->tagset.driver_data = dev;
2308 * Some Apple controllers requires tags to be unique
2309 * across admin and IO queue, so reserve the first 32
2310 * tags of the IO queue.
2312 if (dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS)
2313 dev->tagset.reserved_tags = NVME_AQ_DEPTH;
2315 ret = blk_mq_alloc_tag_set(&dev->tagset);
2317 dev_warn(dev->ctrl.device,
2318 "IO queues tagset allocation failed %d\n", ret);
2321 dev->ctrl.tagset = &dev->tagset;
2323 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
2325 /* Free previously allocated queues that are no longer usable */
2326 nvme_free_queues(dev, dev->online_queues);
2329 nvme_dbbuf_set(dev);
2332 static int nvme_pci_enable(struct nvme_dev *dev)
2334 int result = -ENOMEM;
2335 struct pci_dev *pdev = to_pci_dev(dev->dev);
2337 if (pci_enable_device_mem(pdev))
2340 pci_set_master(pdev);
2342 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)))
2345 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
2351 * Some devices and/or platforms don't advertise or work with INTx
2352 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
2353 * adjust this later.
2355 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
2359 dev->ctrl.cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
2361 dev->q_depth = min_t(u32, NVME_CAP_MQES(dev->ctrl.cap) + 1,
2363 dev->ctrl.sqsize = dev->q_depth - 1; /* 0's based queue depth */
2364 dev->db_stride = 1 << NVME_CAP_STRIDE(dev->ctrl.cap);
2365 dev->dbs = dev->bar + 4096;
2368 * Some Apple controllers require a non-standard SQE size.
2369 * Interestingly they also seem to ignore the CC:IOSQES register
2370 * so we don't bother updating it here.
2372 if (dev->ctrl.quirks & NVME_QUIRK_128_BYTES_SQES)
2375 dev->io_sqes = NVME_NVM_IOSQES;
2378 * Temporary fix for the Apple controller found in the MacBook8,1 and
2379 * some MacBook7,1 to avoid controller resets and data loss.
2381 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
2383 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
2384 "set queue depth=%u to work around controller resets\n",
2386 } else if (pdev->vendor == PCI_VENDOR_ID_SAMSUNG &&
2387 (pdev->device == 0xa821 || pdev->device == 0xa822) &&
2388 NVME_CAP_MQES(dev->ctrl.cap) == 0) {
2390 dev_err(dev->ctrl.device, "detected PM1725 NVMe controller, "
2391 "set queue depth=%u\n", dev->q_depth);
2395 * Controllers with the shared tags quirk need the IO queue to be
2396 * big enough so that we get 32 tags for the admin queue
2398 if ((dev->ctrl.quirks & NVME_QUIRK_SHARED_TAGS) &&
2399 (dev->q_depth < (NVME_AQ_DEPTH + 2))) {
2400 dev->q_depth = NVME_AQ_DEPTH + 2;
2401 dev_warn(dev->ctrl.device, "IO queue depth clamped to %d\n",
2408 pci_enable_pcie_error_reporting(pdev);
2409 pci_save_state(pdev);
2413 pci_disable_device(pdev);
2417 static void nvme_dev_unmap(struct nvme_dev *dev)
2421 pci_release_mem_regions(to_pci_dev(dev->dev));
2424 static void nvme_pci_disable(struct nvme_dev *dev)
2426 struct pci_dev *pdev = to_pci_dev(dev->dev);
2428 pci_free_irq_vectors(pdev);
2430 if (pci_is_enabled(pdev)) {
2431 pci_disable_pcie_error_reporting(pdev);
2432 pci_disable_device(pdev);
2436 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
2438 bool dead = true, freeze = false;
2439 struct pci_dev *pdev = to_pci_dev(dev->dev);
2441 mutex_lock(&dev->shutdown_lock);
2442 if (pci_is_enabled(pdev)) {
2443 u32 csts = readl(dev->bar + NVME_REG_CSTS);
2445 if (dev->ctrl.state == NVME_CTRL_LIVE ||
2446 dev->ctrl.state == NVME_CTRL_RESETTING) {
2448 nvme_start_freeze(&dev->ctrl);
2450 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
2451 pdev->error_state != pci_channel_io_normal);
2455 * Give the controller a chance to complete all entered requests if
2456 * doing a safe shutdown.
2458 if (!dead && shutdown && freeze)
2459 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2461 nvme_stop_queues(&dev->ctrl);
2463 if (!dead && dev->ctrl.queue_count > 0) {
2464 nvme_disable_io_queues(dev);
2465 nvme_disable_admin_queue(dev, shutdown);
2467 nvme_suspend_io_queues(dev);
2468 nvme_suspend_queue(&dev->queues[0]);
2469 nvme_pci_disable(dev);
2470 nvme_reap_pending_cqes(dev);
2472 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2473 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2474 blk_mq_tagset_wait_completed_request(&dev->tagset);
2475 blk_mq_tagset_wait_completed_request(&dev->admin_tagset);
2478 * The driver will not be starting up queues again if shutting down so
2479 * must flush all entered requests to their failed completion to avoid
2480 * deadlocking blk-mq hot-cpu notifier.
2483 nvme_start_queues(&dev->ctrl);
2484 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q))
2485 blk_mq_unquiesce_queue(dev->ctrl.admin_q);
2487 mutex_unlock(&dev->shutdown_lock);
2490 static int nvme_disable_prepare_reset(struct nvme_dev *dev, bool shutdown)
2492 if (!nvme_wait_reset(&dev->ctrl))
2494 nvme_dev_disable(dev, shutdown);
2498 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2500 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2501 NVME_CTRL_PAGE_SIZE,
2502 NVME_CTRL_PAGE_SIZE, 0);
2503 if (!dev->prp_page_pool)
2506 /* Optimisation for I/Os between 4k and 128k */
2507 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2509 if (!dev->prp_small_pool) {
2510 dma_pool_destroy(dev->prp_page_pool);
2516 static void nvme_release_prp_pools(struct nvme_dev *dev)
2518 dma_pool_destroy(dev->prp_page_pool);
2519 dma_pool_destroy(dev->prp_small_pool);
2522 static void nvme_free_tagset(struct nvme_dev *dev)
2524 if (dev->tagset.tags)
2525 blk_mq_free_tag_set(&dev->tagset);
2526 dev->ctrl.tagset = NULL;
2529 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2531 struct nvme_dev *dev = to_nvme_dev(ctrl);
2533 nvme_dbbuf_dma_free(dev);
2534 nvme_free_tagset(dev);
2535 if (dev->ctrl.admin_q)
2536 blk_put_queue(dev->ctrl.admin_q);
2537 free_opal_dev(dev->ctrl.opal_dev);
2538 mempool_destroy(dev->iod_mempool);
2539 put_device(dev->dev);
2544 static void nvme_remove_dead_ctrl(struct nvme_dev *dev)
2547 * Set state to deleting now to avoid blocking nvme_wait_reset(), which
2548 * may be holding this pci_dev's device lock.
2550 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2551 nvme_get_ctrl(&dev->ctrl);
2552 nvme_dev_disable(dev, false);
2553 nvme_kill_queues(&dev->ctrl);
2554 if (!queue_work(nvme_wq, &dev->remove_work))
2555 nvme_put_ctrl(&dev->ctrl);
2558 static void nvme_reset_work(struct work_struct *work)
2560 struct nvme_dev *dev =
2561 container_of(work, struct nvme_dev, ctrl.reset_work);
2562 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2565 if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING)) {
2571 * If we're called to reset a live controller first shut it down before
2574 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2575 nvme_dev_disable(dev, false);
2576 nvme_sync_queues(&dev->ctrl);
2578 mutex_lock(&dev->shutdown_lock);
2579 result = nvme_pci_enable(dev);
2583 result = nvme_pci_configure_admin_queue(dev);
2587 result = nvme_alloc_admin_tags(dev);
2592 * Limit the max command size to prevent iod->sg allocations going
2593 * over a single page.
2595 dev->ctrl.max_hw_sectors = min_t(u32,
2596 NVME_MAX_KB_SZ << 1, dma_max_mapping_size(dev->dev) >> 9);
2597 dev->ctrl.max_segments = NVME_MAX_SEGS;
2600 * Don't limit the IOMMU merged segment size.
2602 dma_set_max_seg_size(dev->dev, 0xffffffff);
2604 mutex_unlock(&dev->shutdown_lock);
2607 * Introduce CONNECTING state from nvme-fc/rdma transports to mark the
2608 * initializing procedure here.
2610 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_CONNECTING)) {
2611 dev_warn(dev->ctrl.device,
2612 "failed to mark controller CONNECTING\n");
2618 * We do not support an SGL for metadata (yet), so we are limited to a
2619 * single integrity segment for the separate metadata pointer.
2621 dev->ctrl.max_integrity_segments = 1;
2623 result = nvme_init_identify(&dev->ctrl);
2627 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2628 if (!dev->ctrl.opal_dev)
2629 dev->ctrl.opal_dev =
2630 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2631 else if (was_suspend)
2632 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2634 free_opal_dev(dev->ctrl.opal_dev);
2635 dev->ctrl.opal_dev = NULL;
2638 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2639 result = nvme_dbbuf_dma_alloc(dev);
2642 "unable to allocate dma for dbbuf\n");
2645 if (dev->ctrl.hmpre) {
2646 result = nvme_setup_host_mem(dev);
2651 result = nvme_setup_io_queues(dev);
2656 * Keep the controller around but remove all namespaces if we don't have
2657 * any working I/O queue.
2659 if (dev->online_queues < 2) {
2660 dev_warn(dev->ctrl.device, "IO queues not created\n");
2661 nvme_kill_queues(&dev->ctrl);
2662 nvme_remove_namespaces(&dev->ctrl);
2663 nvme_free_tagset(dev);
2665 nvme_start_queues(&dev->ctrl);
2666 nvme_wait_freeze(&dev->ctrl);
2668 nvme_unfreeze(&dev->ctrl);
2672 * If only admin queue live, keep it to do further investigation or
2675 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2676 dev_warn(dev->ctrl.device,
2677 "failed to mark controller live state\n");
2682 nvme_start_ctrl(&dev->ctrl);
2686 mutex_unlock(&dev->shutdown_lock);
2689 dev_warn(dev->ctrl.device,
2690 "Removing after probe failure status: %d\n", result);
2691 nvme_remove_dead_ctrl(dev);
2694 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2696 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2697 struct pci_dev *pdev = to_pci_dev(dev->dev);
2699 if (pci_get_drvdata(pdev))
2700 device_release_driver(&pdev->dev);
2701 nvme_put_ctrl(&dev->ctrl);
2704 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2706 *val = readl(to_nvme_dev(ctrl)->bar + off);
2710 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2712 writel(val, to_nvme_dev(ctrl)->bar + off);
2716 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2718 *val = lo_hi_readq(to_nvme_dev(ctrl)->bar + off);
2722 static int nvme_pci_get_address(struct nvme_ctrl *ctrl, char *buf, int size)
2724 struct pci_dev *pdev = to_pci_dev(to_nvme_dev(ctrl)->dev);
2726 return snprintf(buf, size, "%s\n", dev_name(&pdev->dev));
2729 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2731 .module = THIS_MODULE,
2732 .flags = NVME_F_METADATA_SUPPORTED |
2734 .reg_read32 = nvme_pci_reg_read32,
2735 .reg_write32 = nvme_pci_reg_write32,
2736 .reg_read64 = nvme_pci_reg_read64,
2737 .free_ctrl = nvme_pci_free_ctrl,
2738 .submit_async_event = nvme_pci_submit_async_event,
2739 .get_address = nvme_pci_get_address,
2742 static int nvme_dev_map(struct nvme_dev *dev)
2744 struct pci_dev *pdev = to_pci_dev(dev->dev);
2746 if (pci_request_mem_regions(pdev, "nvme"))
2749 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2754 pci_release_mem_regions(pdev);
2758 static unsigned long check_vendor_combination_bug(struct pci_dev *pdev)
2760 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2762 * Several Samsung devices seem to drop off the PCIe bus
2763 * randomly when APST is on and uses the deepest sleep state.
2764 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2765 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2766 * 950 PRO 256GB", but it seems to be restricted to two Dell
2769 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2770 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2771 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2772 return NVME_QUIRK_NO_DEEPEST_PS;
2773 } else if (pdev->vendor == 0x144d && pdev->device == 0xa804) {
2775 * Samsung SSD 960 EVO drops off the PCIe bus after system
2776 * suspend on a Ryzen board, ASUS PRIME B350M-A, as well as
2777 * within few minutes after bootup on a Coffee Lake board -
2780 if (dmi_match(DMI_BOARD_VENDOR, "ASUSTeK COMPUTER INC.") &&
2781 (dmi_match(DMI_BOARD_NAME, "PRIME B350M-A") ||
2782 dmi_match(DMI_BOARD_NAME, "PRIME Z370-A")))
2783 return NVME_QUIRK_NO_APST;
2784 } else if ((pdev->vendor == 0x144d && (pdev->device == 0xa801 ||
2785 pdev->device == 0xa808 || pdev->device == 0xa809)) ||
2786 (pdev->vendor == 0x1e0f && pdev->device == 0x0001)) {
2788 * Forcing to use host managed nvme power settings for
2789 * lowest idle power with quick resume latency on
2790 * Samsung and Toshiba SSDs based on suspend behavior
2791 * on Coffee Lake board for LENOVO C640
2793 if ((dmi_match(DMI_BOARD_VENDOR, "LENOVO")) &&
2794 dmi_match(DMI_BOARD_NAME, "LNVNB161216"))
2795 return NVME_QUIRK_SIMPLE_SUSPEND;
2802 static bool nvme_acpi_storage_d3(struct pci_dev *dev)
2804 struct acpi_device *adev;
2805 struct pci_dev *root;
2811 * Look for _DSD property specifying that the storage device on the port
2812 * must use D3 to support deep platform power savings during
2815 root = pcie_find_root_port(dev);
2819 adev = ACPI_COMPANION(&root->dev);
2824 * The property is defined in the PXSX device for South complex ports
2825 * and in the PEGP device for North complex ports.
2827 status = acpi_get_handle(adev->handle, "PXSX", &handle);
2828 if (ACPI_FAILURE(status)) {
2829 status = acpi_get_handle(adev->handle, "PEGP", &handle);
2830 if (ACPI_FAILURE(status))
2834 if (acpi_bus_get_device(handle, &adev))
2837 if (fwnode_property_read_u8(acpi_fwnode_handle(adev), "StorageD3Enable",
2843 static inline bool nvme_acpi_storage_d3(struct pci_dev *dev)
2847 #endif /* CONFIG_ACPI */
2849 static void nvme_async_probe(void *data, async_cookie_t cookie)
2851 struct nvme_dev *dev = data;
2853 flush_work(&dev->ctrl.reset_work);
2854 flush_work(&dev->ctrl.scan_work);
2855 nvme_put_ctrl(&dev->ctrl);
2858 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2860 int node, result = -ENOMEM;
2861 struct nvme_dev *dev;
2862 unsigned long quirks = id->driver_data;
2865 node = dev_to_node(&pdev->dev);
2866 if (node == NUMA_NO_NODE)
2867 set_dev_node(&pdev->dev, first_memory_node);
2869 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2873 dev->nr_write_queues = write_queues;
2874 dev->nr_poll_queues = poll_queues;
2875 dev->nr_allocated_queues = nvme_max_io_queues(dev) + 1;
2876 dev->queues = kcalloc_node(dev->nr_allocated_queues,
2877 sizeof(struct nvme_queue), GFP_KERNEL, node);
2881 dev->dev = get_device(&pdev->dev);
2882 pci_set_drvdata(pdev, dev);
2884 result = nvme_dev_map(dev);
2888 INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2889 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2890 mutex_init(&dev->shutdown_lock);
2892 result = nvme_setup_prp_pools(dev);
2896 quirks |= check_vendor_combination_bug(pdev);
2898 if (!noacpi && nvme_acpi_storage_d3(pdev)) {
2900 * Some systems use a bios work around to ask for D3 on
2901 * platforms that support kernel managed suspend.
2903 dev_info(&pdev->dev,
2904 "platform quirk: setting simple suspend\n");
2905 quirks |= NVME_QUIRK_SIMPLE_SUSPEND;
2909 * Double check that our mempool alloc size will cover the biggest
2910 * command we support.
2912 alloc_size = nvme_pci_iod_alloc_size();
2913 WARN_ON_ONCE(alloc_size > PAGE_SIZE);
2915 dev->iod_mempool = mempool_create_node(1, mempool_kmalloc,
2917 (void *) alloc_size,
2919 if (!dev->iod_mempool) {
2924 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2927 goto release_mempool;
2929 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2931 nvme_reset_ctrl(&dev->ctrl);
2932 async_schedule(nvme_async_probe, dev);
2937 mempool_destroy(dev->iod_mempool);
2939 nvme_release_prp_pools(dev);
2941 nvme_dev_unmap(dev);
2943 put_device(dev->dev);
2950 static void nvme_reset_prepare(struct pci_dev *pdev)
2952 struct nvme_dev *dev = pci_get_drvdata(pdev);
2955 * We don't need to check the return value from waiting for the reset
2956 * state as pci_dev device lock is held, making it impossible to race
2959 nvme_disable_prepare_reset(dev, false);
2960 nvme_sync_queues(&dev->ctrl);
2963 static void nvme_reset_done(struct pci_dev *pdev)
2965 struct nvme_dev *dev = pci_get_drvdata(pdev);
2967 if (!nvme_try_sched_reset(&dev->ctrl))
2968 flush_work(&dev->ctrl.reset_work);
2971 static void nvme_shutdown(struct pci_dev *pdev)
2973 struct nvme_dev *dev = pci_get_drvdata(pdev);
2975 nvme_disable_prepare_reset(dev, true);
2979 * The driver's remove may be called on a device in a partially initialized
2980 * state. This function must not have any dependencies on the device state in
2983 static void nvme_remove(struct pci_dev *pdev)
2985 struct nvme_dev *dev = pci_get_drvdata(pdev);
2987 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2988 pci_set_drvdata(pdev, NULL);
2990 if (!pci_device_is_present(pdev)) {
2991 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2992 nvme_dev_disable(dev, true);
2993 nvme_dev_remove_admin(dev);
2996 flush_work(&dev->ctrl.reset_work);
2997 nvme_stop_ctrl(&dev->ctrl);
2998 nvme_remove_namespaces(&dev->ctrl);
2999 nvme_dev_disable(dev, true);
3000 nvme_release_cmb(dev);
3001 nvme_free_host_mem(dev);
3002 nvme_dev_remove_admin(dev);
3003 nvme_free_queues(dev, 0);
3004 nvme_release_prp_pools(dev);
3005 nvme_dev_unmap(dev);
3006 nvme_uninit_ctrl(&dev->ctrl);
3009 #ifdef CONFIG_PM_SLEEP
3010 static int nvme_get_power_state(struct nvme_ctrl *ctrl, u32 *ps)
3012 return nvme_get_features(ctrl, NVME_FEAT_POWER_MGMT, 0, NULL, 0, ps);
3015 static int nvme_set_power_state(struct nvme_ctrl *ctrl, u32 ps)
3017 return nvme_set_features(ctrl, NVME_FEAT_POWER_MGMT, ps, NULL, 0, NULL);
3020 static int nvme_resume(struct device *dev)
3022 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3023 struct nvme_ctrl *ctrl = &ndev->ctrl;
3025 if (ndev->last_ps == U32_MAX ||
3026 nvme_set_power_state(ctrl, ndev->last_ps) != 0)
3027 return nvme_try_sched_reset(&ndev->ctrl);
3031 static int nvme_suspend(struct device *dev)
3033 struct pci_dev *pdev = to_pci_dev(dev);
3034 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3035 struct nvme_ctrl *ctrl = &ndev->ctrl;
3038 ndev->last_ps = U32_MAX;
3041 * The platform does not remove power for a kernel managed suspend so
3042 * use host managed nvme power settings for lowest idle power if
3043 * possible. This should have quicker resume latency than a full device
3044 * shutdown. But if the firmware is involved after the suspend or the
3045 * device does not support any non-default power states, shut down the
3048 * If ASPM is not enabled for the device, shut down the device and allow
3049 * the PCI bus layer to put it into D3 in order to take the PCIe link
3050 * down, so as to allow the platform to achieve its minimum low-power
3051 * state (which may not be possible if the link is up).
3053 * If a host memory buffer is enabled, shut down the device as the NVMe
3054 * specification allows the device to access the host memory buffer in
3055 * host DRAM from all power states, but hosts will fail access to DRAM
3058 if (pm_suspend_via_firmware() || !ctrl->npss ||
3059 !pcie_aspm_enabled(pdev) ||
3060 ndev->nr_host_mem_descs ||
3061 (ndev->ctrl.quirks & NVME_QUIRK_SIMPLE_SUSPEND))
3062 return nvme_disable_prepare_reset(ndev, true);
3064 nvme_start_freeze(ctrl);
3065 nvme_wait_freeze(ctrl);
3066 nvme_sync_queues(ctrl);
3068 if (ctrl->state != NVME_CTRL_LIVE)
3071 ret = nvme_get_power_state(ctrl, &ndev->last_ps);
3076 * A saved state prevents pci pm from generically controlling the
3077 * device's power. If we're using protocol specific settings, we don't
3078 * want pci interfering.
3080 pci_save_state(pdev);
3082 ret = nvme_set_power_state(ctrl, ctrl->npss);
3087 /* discard the saved state */
3088 pci_load_saved_state(pdev, NULL);
3091 * Clearing npss forces a controller reset on resume. The
3092 * correct value will be rediscovered then.
3094 ret = nvme_disable_prepare_reset(ndev, true);
3098 nvme_unfreeze(ctrl);
3102 static int nvme_simple_suspend(struct device *dev)
3104 struct nvme_dev *ndev = pci_get_drvdata(to_pci_dev(dev));
3106 return nvme_disable_prepare_reset(ndev, true);
3109 static int nvme_simple_resume(struct device *dev)
3111 struct pci_dev *pdev = to_pci_dev(dev);
3112 struct nvme_dev *ndev = pci_get_drvdata(pdev);
3114 return nvme_try_sched_reset(&ndev->ctrl);
3117 static const struct dev_pm_ops nvme_dev_pm_ops = {
3118 .suspend = nvme_suspend,
3119 .resume = nvme_resume,
3120 .freeze = nvme_simple_suspend,
3121 .thaw = nvme_simple_resume,
3122 .poweroff = nvme_simple_suspend,
3123 .restore = nvme_simple_resume,
3125 #endif /* CONFIG_PM_SLEEP */
3127 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
3128 pci_channel_state_t state)
3130 struct nvme_dev *dev = pci_get_drvdata(pdev);
3133 * A frozen channel requires a reset. When detected, this method will
3134 * shutdown the controller to quiesce. The controller will be restarted
3135 * after the slot reset through driver's slot_reset callback.
3138 case pci_channel_io_normal:
3139 return PCI_ERS_RESULT_CAN_RECOVER;
3140 case pci_channel_io_frozen:
3141 dev_warn(dev->ctrl.device,
3142 "frozen state error detected, reset controller\n");
3143 nvme_dev_disable(dev, false);
3144 return PCI_ERS_RESULT_NEED_RESET;
3145 case pci_channel_io_perm_failure:
3146 dev_warn(dev->ctrl.device,
3147 "failure state error detected, request disconnect\n");
3148 return PCI_ERS_RESULT_DISCONNECT;
3150 return PCI_ERS_RESULT_NEED_RESET;
3153 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
3155 struct nvme_dev *dev = pci_get_drvdata(pdev);
3157 dev_info(dev->ctrl.device, "restart after slot reset\n");
3158 pci_restore_state(pdev);
3159 nvme_reset_ctrl(&dev->ctrl);
3160 return PCI_ERS_RESULT_RECOVERED;
3163 static void nvme_error_resume(struct pci_dev *pdev)
3165 struct nvme_dev *dev = pci_get_drvdata(pdev);
3167 flush_work(&dev->ctrl.reset_work);
3170 static const struct pci_error_handlers nvme_err_handler = {
3171 .error_detected = nvme_error_detected,
3172 .slot_reset = nvme_slot_reset,
3173 .resume = nvme_error_resume,
3174 .reset_prepare = nvme_reset_prepare,
3175 .reset_done = nvme_reset_done,
3178 static const struct pci_device_id nvme_id_table[] = {
3179 { PCI_VDEVICE(INTEL, 0x0953), /* Intel 750/P3500/P3600/P3700 */
3180 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3181 NVME_QUIRK_DEALLOCATE_ZEROES, },
3182 { PCI_VDEVICE(INTEL, 0x0a53), /* Intel P3520 */
3183 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3184 NVME_QUIRK_DEALLOCATE_ZEROES, },
3185 { PCI_VDEVICE(INTEL, 0x0a54), /* Intel P4500/P4600 */
3186 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3187 NVME_QUIRK_DEALLOCATE_ZEROES, },
3188 { PCI_VDEVICE(INTEL, 0x0a55), /* Dell Express Flash P4600 */
3189 .driver_data = NVME_QUIRK_STRIPE_SIZE |
3190 NVME_QUIRK_DEALLOCATE_ZEROES, },
3191 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
3192 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3193 NVME_QUIRK_MEDIUM_PRIO_SQ |
3194 NVME_QUIRK_NO_TEMP_THRESH_CHANGE |
3195 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3196 { PCI_VDEVICE(INTEL, 0xf1a6), /* Intel 760p/Pro 7600p */
3197 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3198 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
3199 .driver_data = NVME_QUIRK_IDENTIFY_CNS |
3200 NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3201 { PCI_DEVICE(0x126f, 0x2263), /* Silicon Motion unidentified */
3202 .driver_data = NVME_QUIRK_NO_NS_DESC_LIST, },
3203 { PCI_DEVICE(0x1bb1, 0x0100), /* Seagate Nytro Flash Storage */
3204 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3205 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
3206 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3207 { PCI_DEVICE(0x1c58, 0x0023), /* WDC SN200 adapter */
3208 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3209 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
3210 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3211 { PCI_DEVICE(0x144d, 0xa821), /* Samsung PM1725 */
3212 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
3213 { PCI_DEVICE(0x144d, 0xa822), /* Samsung PM1725a */
3214 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY |
3215 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3216 { PCI_DEVICE(0x1d1d, 0x1f1f), /* LighNVM qemu device */
3217 .driver_data = NVME_QUIRK_LIGHTNVM, },
3218 { PCI_DEVICE(0x1d1d, 0x2807), /* CNEX WL */
3219 .driver_data = NVME_QUIRK_LIGHTNVM, },
3220 { PCI_DEVICE(0x1d1d, 0x2601), /* CNEX Granby */
3221 .driver_data = NVME_QUIRK_LIGHTNVM, },
3222 { PCI_DEVICE(0x10ec, 0x5762), /* ADATA SX6000LNP */
3223 .driver_data = NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3224 { PCI_DEVICE(0x1cc1, 0x8201), /* ADATA SX8200PNP 512GB */
3225 .driver_data = NVME_QUIRK_NO_DEEPEST_PS |
3226 NVME_QUIRK_IGNORE_DEV_SUBNQN, },
3227 { PCI_DEVICE(0x1c5c, 0x1504), /* SK Hynix PC400 */
3228 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3229 { PCI_DEVICE(0x15b7, 0x2001), /* Sandisk Skyhawk */
3230 .driver_data = NVME_QUIRK_DISABLE_WRITE_ZEROES, },
3231 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001),
3232 .driver_data = NVME_QUIRK_SINGLE_VECTOR },
3233 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
3234 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2005),
3235 .driver_data = NVME_QUIRK_SINGLE_VECTOR |
3236 NVME_QUIRK_128_BYTES_SQES |
3237 NVME_QUIRK_SHARED_TAGS },
3239 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
3242 MODULE_DEVICE_TABLE(pci, nvme_id_table);
3244 static struct pci_driver nvme_driver = {
3246 .id_table = nvme_id_table,
3247 .probe = nvme_probe,
3248 .remove = nvme_remove,
3249 .shutdown = nvme_shutdown,
3250 #ifdef CONFIG_PM_SLEEP
3252 .pm = &nvme_dev_pm_ops,
3255 .sriov_configure = pci_sriov_configure_simple,
3256 .err_handler = &nvme_err_handler,
3259 static int __init nvme_init(void)
3261 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
3262 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
3263 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
3264 BUILD_BUG_ON(IRQ_AFFINITY_MAX_SETS < 2);
3266 return pci_register_driver(&nvme_driver);
3269 static void __exit nvme_exit(void)
3271 pci_unregister_driver(&nvme_driver);
3272 flush_workqueue(nvme_wq);
3275 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
3276 MODULE_LICENSE("GPL");
3277 MODULE_VERSION("1.0");
3278 module_init(nvme_init);
3279 module_exit(nvme_exit);