2 * NVM Express device driver
3 * Copyright (c) 2011-2014, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 #include <linux/aer.h>
16 #include <linux/bitops.h>
17 #include <linux/blkdev.h>
18 #include <linux/blk-mq.h>
19 #include <linux/blk-mq-pci.h>
20 #include <linux/dmi.h>
21 #include <linux/init.h>
22 #include <linux/interrupt.h>
25 #include <linux/module.h>
26 #include <linux/mutex.h>
27 #include <linux/pci.h>
28 #include <linux/poison.h>
29 #include <linux/t10-pi.h>
30 #include <linux/timer.h>
31 #include <linux/types.h>
32 #include <linux/io-64-nonatomic-lo-hi.h>
33 #include <asm/unaligned.h>
34 #include <linux/sed-opal.h>
38 #define NVME_Q_DEPTH 1024
39 #define NVME_AQ_DEPTH 256
40 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
41 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
44 * We handle AEN commands ourselves and don't even let the
45 * block layer know about them.
47 #define NVME_AQ_BLKMQ_DEPTH (NVME_AQ_DEPTH - NVME_NR_AERS)
49 static int use_threaded_interrupts;
50 module_param(use_threaded_interrupts, int, 0);
52 static bool use_cmb_sqes = true;
53 module_param(use_cmb_sqes, bool, 0644);
54 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
56 static unsigned int max_host_mem_size_mb = 128;
57 module_param(max_host_mem_size_mb, uint, 0444);
58 MODULE_PARM_DESC(max_host_mem_size_mb,
59 "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
64 static int nvme_reset(struct nvme_dev *dev);
65 static void nvme_process_cq(struct nvme_queue *nvmeq);
66 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
69 * Represents an NVM Express device. Each nvme_dev is a PCI function.
72 struct nvme_queue **queues;
73 struct blk_mq_tag_set tagset;
74 struct blk_mq_tag_set admin_tagset;
77 struct dma_pool *prp_page_pool;
78 struct dma_pool *prp_small_pool;
80 unsigned online_queues;
85 unsigned long bar_mapped_size;
86 struct work_struct reset_work;
87 struct work_struct remove_work;
88 struct mutex shutdown_lock;
91 dma_addr_t cmb_dma_addr;
95 struct nvme_ctrl ctrl;
96 struct completion ioq_wait;
98 /* shadow doorbell buffer support: */
100 dma_addr_t dbbuf_dbs_dma_addr;
102 dma_addr_t dbbuf_eis_dma_addr;
104 /* host memory buffer support: */
106 u32 nr_host_mem_descs;
107 struct nvme_host_mem_buf_desc *host_mem_descs;
108 void **host_mem_desc_bufs;
111 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
113 return qid * 2 * stride;
116 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
118 return (qid * 2 + 1) * stride;
121 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
123 return container_of(ctrl, struct nvme_dev, ctrl);
127 * An NVM Express queue. Each device has at least two (one for admin
128 * commands and one for I/O commands).
131 struct device *q_dmadev;
132 struct nvme_dev *dev;
134 struct nvme_command *sq_cmds;
135 struct nvme_command __iomem *sq_cmds_io;
136 volatile struct nvme_completion *cqes;
137 struct blk_mq_tags **tags;
138 dma_addr_t sq_dma_addr;
139 dma_addr_t cq_dma_addr;
155 * The nvme_iod describes the data in an I/O, including the list of PRP
156 * entries. You can't see it in this data structure because C doesn't let
157 * me express that. Use nvme_init_iod to ensure there's enough space
158 * allocated to store the PRP list.
161 struct nvme_request req;
162 struct nvme_queue *nvmeq;
164 int npages; /* In the PRP list. 0 means small pool in use */
165 int nents; /* Used in scatterlist */
166 int length; /* Of data, in bytes */
167 dma_addr_t first_dma;
168 struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
169 struct scatterlist *sg;
170 struct scatterlist inline_sg[0];
174 * Check we didin't inadvertently grow the command struct
176 static inline void _nvme_check_size(void)
178 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
179 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
180 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
181 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
182 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
183 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
184 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
185 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
186 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
187 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
188 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
189 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
190 BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
193 static inline unsigned int nvme_dbbuf_size(u32 stride)
195 return ((num_possible_cpus() + 1) * 8 * stride);
198 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
200 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
205 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
206 &dev->dbbuf_dbs_dma_addr,
210 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
211 &dev->dbbuf_eis_dma_addr,
213 if (!dev->dbbuf_eis) {
214 dma_free_coherent(dev->dev, mem_size,
215 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
216 dev->dbbuf_dbs = NULL;
223 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
225 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
227 if (dev->dbbuf_dbs) {
228 dma_free_coherent(dev->dev, mem_size,
229 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
230 dev->dbbuf_dbs = NULL;
232 if (dev->dbbuf_eis) {
233 dma_free_coherent(dev->dev, mem_size,
234 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
235 dev->dbbuf_eis = NULL;
239 static void nvme_dbbuf_init(struct nvme_dev *dev,
240 struct nvme_queue *nvmeq, int qid)
242 if (!dev->dbbuf_dbs || !qid)
245 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
246 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
247 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
248 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
251 static void nvme_dbbuf_set(struct nvme_dev *dev)
253 struct nvme_command c;
258 memset(&c, 0, sizeof(c));
259 c.dbbuf.opcode = nvme_admin_dbbuf;
260 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
261 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
263 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
264 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
265 /* Free memory and continue on */
266 nvme_dbbuf_dma_free(dev);
270 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
272 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
275 /* Update dbbuf and return true if an MMIO is required */
276 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
277 volatile u32 *dbbuf_ei)
283 * Ensure that the queue is written before updating
284 * the doorbell in memory
288 old_value = *dbbuf_db;
291 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
299 * Max size of iod being embedded in the request payload
301 #define NVME_INT_PAGES 2
302 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size)
305 * Will slightly overestimate the number of pages needed. This is OK
306 * as it only leads to a small amount of wasted memory for the lifetime of
309 static int nvme_npages(unsigned size, struct nvme_dev *dev)
311 unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
312 dev->ctrl.page_size);
313 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
316 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
317 unsigned int size, unsigned int nseg)
319 return sizeof(__le64 *) * nvme_npages(size, dev) +
320 sizeof(struct scatterlist) * nseg;
323 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
325 return sizeof(struct nvme_iod) +
326 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
329 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
330 unsigned int hctx_idx)
332 struct nvme_dev *dev = data;
333 struct nvme_queue *nvmeq = dev->queues[0];
335 WARN_ON(hctx_idx != 0);
336 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
337 WARN_ON(nvmeq->tags);
339 hctx->driver_data = nvmeq;
340 nvmeq->tags = &dev->admin_tagset.tags[0];
344 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
346 struct nvme_queue *nvmeq = hctx->driver_data;
351 static int nvme_admin_init_request(struct blk_mq_tag_set *set,
352 struct request *req, unsigned int hctx_idx,
353 unsigned int numa_node)
355 struct nvme_dev *dev = set->driver_data;
356 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
357 struct nvme_queue *nvmeq = dev->queues[0];
364 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
365 unsigned int hctx_idx)
367 struct nvme_dev *dev = data;
368 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
371 nvmeq->tags = &dev->tagset.tags[hctx_idx];
373 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
374 hctx->driver_data = nvmeq;
378 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
379 unsigned int hctx_idx, unsigned int numa_node)
381 struct nvme_dev *dev = set->driver_data;
382 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
383 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
390 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
392 struct nvme_dev *dev = set->driver_data;
394 return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
398 * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
399 * @nvmeq: The queue to use
400 * @cmd: The command to send
402 * Safe to use from interrupt context
404 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
405 struct nvme_command *cmd)
407 u16 tail = nvmeq->sq_tail;
409 if (nvmeq->sq_cmds_io)
410 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
412 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
414 if (++tail == nvmeq->q_depth)
416 if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
418 writel(tail, nvmeq->q_db);
419 nvmeq->sq_tail = tail;
422 static __le64 **iod_list(struct request *req)
424 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
425 return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
428 static blk_status_t nvme_init_iod(struct request *rq, struct nvme_dev *dev)
430 struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
431 int nseg = blk_rq_nr_phys_segments(rq);
432 unsigned int size = blk_rq_payload_bytes(rq);
434 if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
435 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
437 return BLK_STS_RESOURCE;
439 iod->sg = iod->inline_sg;
450 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
452 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
453 const int last_prp = dev->ctrl.page_size / 8 - 1;
455 __le64 **list = iod_list(req);
456 dma_addr_t prp_dma = iod->first_dma;
458 if (iod->npages == 0)
459 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
460 for (i = 0; i < iod->npages; i++) {
461 __le64 *prp_list = list[i];
462 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
463 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
464 prp_dma = next_prp_dma;
467 if (iod->sg != iod->inline_sg)
471 #ifdef CONFIG_BLK_DEV_INTEGRITY
472 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
474 if (be32_to_cpu(pi->ref_tag) == v)
475 pi->ref_tag = cpu_to_be32(p);
478 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
480 if (be32_to_cpu(pi->ref_tag) == p)
481 pi->ref_tag = cpu_to_be32(v);
485 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
487 * The virtual start sector is the one that was originally submitted by the
488 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
489 * start sector may be different. Remap protection information to match the
490 * physical LBA on writes, and back to the original seed on reads.
492 * Type 0 and 3 do not have a ref tag, so no remapping required.
494 static void nvme_dif_remap(struct request *req,
495 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
497 struct nvme_ns *ns = req->rq_disk->private_data;
498 struct bio_integrity_payload *bip;
499 struct t10_pi_tuple *pi;
501 u32 i, nlb, ts, phys, virt;
503 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
506 bip = bio_integrity(req->bio);
510 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
513 virt = bip_get_seed(bip);
514 phys = nvme_block_nr(ns, blk_rq_pos(req));
515 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
516 ts = ns->disk->queue->integrity.tuple_size;
518 for (i = 0; i < nlb; i++, virt++, phys++) {
519 pi = (struct t10_pi_tuple *)p;
520 dif_swap(phys, virt, pi);
525 #else /* CONFIG_BLK_DEV_INTEGRITY */
526 static void nvme_dif_remap(struct request *req,
527 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
530 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
533 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
538 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req)
540 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
541 struct dma_pool *pool;
542 int length = blk_rq_payload_bytes(req);
543 struct scatterlist *sg = iod->sg;
544 int dma_len = sg_dma_len(sg);
545 u64 dma_addr = sg_dma_address(sg);
546 u32 page_size = dev->ctrl.page_size;
547 int offset = dma_addr & (page_size - 1);
549 __le64 **list = iod_list(req);
553 length -= (page_size - offset);
557 dma_len -= (page_size - offset);
559 dma_addr += (page_size - offset);
562 dma_addr = sg_dma_address(sg);
563 dma_len = sg_dma_len(sg);
566 if (length <= page_size) {
567 iod->first_dma = dma_addr;
571 nprps = DIV_ROUND_UP(length, page_size);
572 if (nprps <= (256 / 8)) {
573 pool = dev->prp_small_pool;
576 pool = dev->prp_page_pool;
580 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
582 iod->first_dma = dma_addr;
587 iod->first_dma = prp_dma;
590 if (i == page_size >> 3) {
591 __le64 *old_prp_list = prp_list;
592 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
595 list[iod->npages++] = prp_list;
596 prp_list[0] = old_prp_list[i - 1];
597 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
600 prp_list[i++] = cpu_to_le64(dma_addr);
601 dma_len -= page_size;
602 dma_addr += page_size;
610 dma_addr = sg_dma_address(sg);
611 dma_len = sg_dma_len(sg);
617 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
618 struct nvme_command *cmnd)
620 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
621 struct request_queue *q = req->q;
622 enum dma_data_direction dma_dir = rq_data_dir(req) ?
623 DMA_TO_DEVICE : DMA_FROM_DEVICE;
624 blk_status_t ret = BLK_STS_IOERR;
626 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
627 iod->nents = blk_rq_map_sg(q, req, iod->sg);
631 ret = BLK_STS_RESOURCE;
632 if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
636 if (!nvme_setup_prps(dev, req))
640 if (blk_integrity_rq(req)) {
641 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
644 sg_init_table(&iod->meta_sg, 1);
645 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
648 if (rq_data_dir(req))
649 nvme_dif_remap(req, nvme_dif_prep);
651 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
655 cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
656 cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
657 if (blk_integrity_rq(req))
658 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
662 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
667 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
669 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
670 enum dma_data_direction dma_dir = rq_data_dir(req) ?
671 DMA_TO_DEVICE : DMA_FROM_DEVICE;
674 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
675 if (blk_integrity_rq(req)) {
676 if (!rq_data_dir(req))
677 nvme_dif_remap(req, nvme_dif_complete);
678 dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
682 nvme_cleanup_cmd(req);
683 nvme_free_iod(dev, req);
687 * NOTE: ns is NULL when called on the admin queue.
689 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
690 const struct blk_mq_queue_data *bd)
692 struct nvme_ns *ns = hctx->queue->queuedata;
693 struct nvme_queue *nvmeq = hctx->driver_data;
694 struct nvme_dev *dev = nvmeq->dev;
695 struct request *req = bd->rq;
696 struct nvme_command cmnd;
697 blk_status_t ret = BLK_STS_OK;
700 * If formated with metadata, require the block layer provide a buffer
701 * unless this namespace is formated such that the metadata can be
702 * stripped/generated by the controller with PRACT=1.
704 if (ns && ns->ms && !blk_integrity_rq(req)) {
705 if (!(ns->pi_type && ns->ms == 8) &&
706 !blk_rq_is_passthrough(req))
707 return BLK_STS_NOTSUPP;
710 ret = nvme_setup_cmd(ns, req, &cmnd);
714 ret = nvme_init_iod(req, dev);
718 if (blk_rq_nr_phys_segments(req)) {
719 ret = nvme_map_data(dev, req, &cmnd);
721 goto out_cleanup_iod;
724 blk_mq_start_request(req);
726 spin_lock_irq(&nvmeq->q_lock);
727 if (unlikely(nvmeq->cq_vector < 0)) {
729 spin_unlock_irq(&nvmeq->q_lock);
730 goto out_cleanup_iod;
732 __nvme_submit_cmd(nvmeq, &cmnd);
733 nvme_process_cq(nvmeq);
734 spin_unlock_irq(&nvmeq->q_lock);
737 nvme_free_iod(dev, req);
739 nvme_cleanup_cmd(req);
743 static void nvme_pci_complete_rq(struct request *req)
745 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
747 nvme_unmap_data(iod->nvmeq->dev, req);
748 nvme_complete_rq(req);
751 /* We read the CQE phase first to check if the rest of the entry is valid */
752 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
755 return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
758 static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag)
762 head = nvmeq->cq_head;
763 phase = nvmeq->cq_phase;
765 while (nvme_cqe_valid(nvmeq, head, phase)) {
766 struct nvme_completion cqe = nvmeq->cqes[head];
769 if (++head == nvmeq->q_depth) {
774 if (tag && *tag == cqe.command_id)
777 if (unlikely(cqe.command_id >= nvmeq->q_depth)) {
778 dev_warn(nvmeq->dev->ctrl.device,
779 "invalid id %d completed on queue %d\n",
780 cqe.command_id, le16_to_cpu(cqe.sq_id));
785 * AEN requests are special as they don't time out and can
786 * survive any kind of queue freeze and often don't respond to
787 * aborts. We don't even bother to allocate a struct request
788 * for them but rather special case them here.
790 if (unlikely(nvmeq->qid == 0 &&
791 cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) {
792 nvme_complete_async_event(&nvmeq->dev->ctrl,
793 cqe.status, &cqe.result);
797 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id);
798 nvme_end_request(req, cqe.status, cqe.result);
801 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
804 if (likely(nvmeq->cq_vector >= 0))
805 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
807 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
808 nvmeq->cq_head = head;
809 nvmeq->cq_phase = phase;
814 static void nvme_process_cq(struct nvme_queue *nvmeq)
816 __nvme_process_cq(nvmeq, NULL);
819 static irqreturn_t nvme_irq(int irq, void *data)
822 struct nvme_queue *nvmeq = data;
823 spin_lock(&nvmeq->q_lock);
824 nvme_process_cq(nvmeq);
825 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
827 spin_unlock(&nvmeq->q_lock);
831 static irqreturn_t nvme_irq_check(int irq, void *data)
833 struct nvme_queue *nvmeq = data;
834 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
835 return IRQ_WAKE_THREAD;
839 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
841 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
842 spin_lock_irq(&nvmeq->q_lock);
843 __nvme_process_cq(nvmeq, &tag);
844 spin_unlock_irq(&nvmeq->q_lock);
853 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
855 struct nvme_queue *nvmeq = hctx->driver_data;
857 return __nvme_poll(nvmeq, tag);
860 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
862 struct nvme_dev *dev = to_nvme_dev(ctrl);
863 struct nvme_queue *nvmeq = dev->queues[0];
864 struct nvme_command c;
866 memset(&c, 0, sizeof(c));
867 c.common.opcode = nvme_admin_async_event;
868 c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
870 spin_lock_irq(&nvmeq->q_lock);
871 __nvme_submit_cmd(nvmeq, &c);
872 spin_unlock_irq(&nvmeq->q_lock);
875 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
877 struct nvme_command c;
879 memset(&c, 0, sizeof(c));
880 c.delete_queue.opcode = opcode;
881 c.delete_queue.qid = cpu_to_le16(id);
883 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
886 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
887 struct nvme_queue *nvmeq)
889 struct nvme_command c;
890 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
893 * Note: we (ab)use the fact the the prp fields survive if no data
894 * is attached to the request.
896 memset(&c, 0, sizeof(c));
897 c.create_cq.opcode = nvme_admin_create_cq;
898 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
899 c.create_cq.cqid = cpu_to_le16(qid);
900 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
901 c.create_cq.cq_flags = cpu_to_le16(flags);
902 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
904 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
907 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
908 struct nvme_queue *nvmeq)
910 struct nvme_command c;
911 int flags = NVME_QUEUE_PHYS_CONTIG;
914 * Note: we (ab)use the fact the the prp fields survive if no data
915 * is attached to the request.
917 memset(&c, 0, sizeof(c));
918 c.create_sq.opcode = nvme_admin_create_sq;
919 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
920 c.create_sq.sqid = cpu_to_le16(qid);
921 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
922 c.create_sq.sq_flags = cpu_to_le16(flags);
923 c.create_sq.cqid = cpu_to_le16(qid);
925 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
928 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
930 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
933 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
935 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
938 static void abort_endio(struct request *req, blk_status_t error)
940 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
941 struct nvme_queue *nvmeq = iod->nvmeq;
943 dev_warn(nvmeq->dev->ctrl.device,
944 "Abort status: 0x%x", nvme_req(req)->status);
945 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
946 blk_mq_free_request(req);
949 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
952 /* If true, indicates loss of adapter communication, possibly by a
953 * NVMe Subsystem reset.
955 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
957 /* If there is a reset ongoing, we shouldn't reset again. */
958 if (dev->ctrl.state == NVME_CTRL_RESETTING)
961 /* We shouldn't reset unless the controller is on fatal error state
962 * _or_ if we lost the communication with it.
964 if (!(csts & NVME_CSTS_CFS) && !nssro)
967 /* If PCI error recovery process is happening, we cannot reset or
968 * the recovery mechanism will surely fail.
970 if (pci_channel_offline(to_pci_dev(dev->dev)))
976 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
978 /* Read a config register to help see what died. */
982 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
984 if (result == PCIBIOS_SUCCESSFUL)
985 dev_warn(dev->ctrl.device,
986 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
989 dev_warn(dev->ctrl.device,
990 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
994 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
996 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
997 struct nvme_queue *nvmeq = iod->nvmeq;
998 struct nvme_dev *dev = nvmeq->dev;
999 struct request *abort_req;
1000 struct nvme_command cmd;
1001 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1004 * Reset immediately if the controller is failed
1006 if (nvme_should_reset(dev, csts)) {
1007 nvme_warn_reset(dev, csts);
1008 nvme_dev_disable(dev, false);
1010 return BLK_EH_HANDLED;
1014 * Did we miss an interrupt?
1016 if (__nvme_poll(nvmeq, req->tag)) {
1017 dev_warn(dev->ctrl.device,
1018 "I/O %d QID %d timeout, completion polled\n",
1019 req->tag, nvmeq->qid);
1020 return BLK_EH_HANDLED;
1024 * Shutdown immediately if controller times out while starting. The
1025 * reset work will see the pci device disabled when it gets the forced
1026 * cancellation error. All outstanding requests are completed on
1027 * shutdown, so we return BLK_EH_HANDLED.
1029 if (dev->ctrl.state == NVME_CTRL_RESETTING) {
1030 dev_warn(dev->ctrl.device,
1031 "I/O %d QID %d timeout, disable controller\n",
1032 req->tag, nvmeq->qid);
1033 nvme_dev_disable(dev, false);
1034 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1035 return BLK_EH_HANDLED;
1039 * Shutdown the controller immediately and schedule a reset if the
1040 * command was already aborted once before and still hasn't been
1041 * returned to the driver, or if this is the admin queue.
1043 if (!nvmeq->qid || iod->aborted) {
1044 dev_warn(dev->ctrl.device,
1045 "I/O %d QID %d timeout, reset controller\n",
1046 req->tag, nvmeq->qid);
1047 nvme_dev_disable(dev, false);
1051 * Mark the request as handled, since the inline shutdown
1052 * forces all outstanding requests to complete.
1054 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1055 return BLK_EH_HANDLED;
1058 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1059 atomic_inc(&dev->ctrl.abort_limit);
1060 return BLK_EH_RESET_TIMER;
1064 memset(&cmd, 0, sizeof(cmd));
1065 cmd.abort.opcode = nvme_admin_abort_cmd;
1066 cmd.abort.cid = req->tag;
1067 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1069 dev_warn(nvmeq->dev->ctrl.device,
1070 "I/O %d QID %d timeout, aborting\n",
1071 req->tag, nvmeq->qid);
1073 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1074 BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1075 if (IS_ERR(abort_req)) {
1076 atomic_inc(&dev->ctrl.abort_limit);
1077 return BLK_EH_RESET_TIMER;
1080 abort_req->timeout = ADMIN_TIMEOUT;
1081 abort_req->end_io_data = NULL;
1082 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1085 * The aborted req will be completed on receiving the abort req.
1086 * We enable the timer again. If hit twice, it'll cause a device reset,
1087 * as the device then is in a faulty state.
1089 return BLK_EH_RESET_TIMER;
1092 static void nvme_free_queue(struct nvme_queue *nvmeq)
1094 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1095 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1097 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1098 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1102 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1106 for (i = dev->queue_count - 1; i >= lowest; i--) {
1107 struct nvme_queue *nvmeq = dev->queues[i];
1109 dev->queues[i] = NULL;
1110 nvme_free_queue(nvmeq);
1115 * nvme_suspend_queue - put queue into suspended state
1116 * @nvmeq - queue to suspend
1118 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1122 spin_lock_irq(&nvmeq->q_lock);
1123 if (nvmeq->cq_vector == -1) {
1124 spin_unlock_irq(&nvmeq->q_lock);
1127 vector = nvmeq->cq_vector;
1128 nvmeq->dev->online_queues--;
1129 nvmeq->cq_vector = -1;
1130 spin_unlock_irq(&nvmeq->q_lock);
1132 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1133 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1135 pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1140 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1142 struct nvme_queue *nvmeq = dev->queues[0];
1146 if (nvme_suspend_queue(nvmeq))
1150 nvme_shutdown_ctrl(&dev->ctrl);
1152 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1153 dev->bar + NVME_REG_CAP));
1155 spin_lock_irq(&nvmeq->q_lock);
1156 nvme_process_cq(nvmeq);
1157 spin_unlock_irq(&nvmeq->q_lock);
1160 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1163 int q_depth = dev->q_depth;
1164 unsigned q_size_aligned = roundup(q_depth * entry_size,
1165 dev->ctrl.page_size);
1167 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1168 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1169 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1170 q_depth = div_u64(mem_per_q, entry_size);
1173 * Ensure the reduced q_depth is above some threshold where it
1174 * would be better to map queues in system memory with the
1184 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1187 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1188 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1189 dev->ctrl.page_size);
1190 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1191 nvmeq->sq_cmds_io = dev->cmb + offset;
1193 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1194 &nvmeq->sq_dma_addr, GFP_KERNEL);
1195 if (!nvmeq->sq_cmds)
1202 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1203 int depth, int node)
1205 struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
1210 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1211 &nvmeq->cq_dma_addr, GFP_KERNEL);
1215 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1218 nvmeq->q_dmadev = dev->dev;
1220 spin_lock_init(&nvmeq->q_lock);
1222 nvmeq->cq_phase = 1;
1223 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1224 nvmeq->q_depth = depth;
1226 nvmeq->cq_vector = -1;
1227 dev->queues[qid] = nvmeq;
1233 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1234 nvmeq->cq_dma_addr);
1240 static int queue_request_irq(struct nvme_queue *nvmeq)
1242 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1243 int nr = nvmeq->dev->ctrl.instance;
1245 if (use_threaded_interrupts) {
1246 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1247 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1249 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1250 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1254 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1256 struct nvme_dev *dev = nvmeq->dev;
1258 spin_lock_irq(&nvmeq->q_lock);
1261 nvmeq->cq_phase = 1;
1262 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1263 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1264 nvme_dbbuf_init(dev, nvmeq, qid);
1265 dev->online_queues++;
1266 spin_unlock_irq(&nvmeq->q_lock);
1269 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1271 struct nvme_dev *dev = nvmeq->dev;
1274 nvmeq->cq_vector = qid - 1;
1275 result = adapter_alloc_cq(dev, qid, nvmeq);
1279 result = adapter_alloc_sq(dev, qid, nvmeq);
1283 result = queue_request_irq(nvmeq);
1287 nvme_init_queue(nvmeq, qid);
1291 adapter_delete_sq(dev, qid);
1293 adapter_delete_cq(dev, qid);
1297 static const struct blk_mq_ops nvme_mq_admin_ops = {
1298 .queue_rq = nvme_queue_rq,
1299 .complete = nvme_pci_complete_rq,
1300 .init_hctx = nvme_admin_init_hctx,
1301 .exit_hctx = nvme_admin_exit_hctx,
1302 .init_request = nvme_admin_init_request,
1303 .timeout = nvme_timeout,
1306 static const struct blk_mq_ops nvme_mq_ops = {
1307 .queue_rq = nvme_queue_rq,
1308 .complete = nvme_pci_complete_rq,
1309 .init_hctx = nvme_init_hctx,
1310 .init_request = nvme_init_request,
1311 .map_queues = nvme_pci_map_queues,
1312 .timeout = nvme_timeout,
1316 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1318 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1320 * If the controller was reset during removal, it's possible
1321 * user requests may be waiting on a stopped queue. Start the
1322 * queue to flush these to completion.
1324 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1325 blk_cleanup_queue(dev->ctrl.admin_q);
1326 blk_mq_free_tag_set(&dev->admin_tagset);
1330 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1332 if (!dev->ctrl.admin_q) {
1333 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1334 dev->admin_tagset.nr_hw_queues = 1;
1337 * Subtract one to leave an empty queue entry for 'Full Queue'
1338 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1340 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1341 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1342 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1343 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1344 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1345 dev->admin_tagset.driver_data = dev;
1347 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1350 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1351 if (IS_ERR(dev->ctrl.admin_q)) {
1352 blk_mq_free_tag_set(&dev->admin_tagset);
1355 if (!blk_get_queue(dev->ctrl.admin_q)) {
1356 nvme_dev_remove_admin(dev);
1357 dev->ctrl.admin_q = NULL;
1361 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1366 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1368 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1371 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1373 struct pci_dev *pdev = to_pci_dev(dev->dev);
1375 if (size <= dev->bar_mapped_size)
1377 if (size > pci_resource_len(pdev, 0))
1381 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1383 dev->bar_mapped_size = 0;
1386 dev->bar_mapped_size = size;
1387 dev->dbs = dev->bar + NVME_REG_DBS;
1392 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1396 u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1397 struct nvme_queue *nvmeq;
1399 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1403 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1404 NVME_CAP_NSSRC(cap) : 0;
1406 if (dev->subsystem &&
1407 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1408 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1410 result = nvme_disable_ctrl(&dev->ctrl, cap);
1414 nvmeq = dev->queues[0];
1416 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
1417 dev_to_node(dev->dev));
1422 aqa = nvmeq->q_depth - 1;
1425 writel(aqa, dev->bar + NVME_REG_AQA);
1426 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1427 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1429 result = nvme_enable_ctrl(&dev->ctrl, cap);
1433 nvmeq->cq_vector = 0;
1434 result = queue_request_irq(nvmeq);
1436 nvmeq->cq_vector = -1;
1443 static int nvme_create_io_queues(struct nvme_dev *dev)
1448 for (i = dev->queue_count; i <= dev->max_qid; i++) {
1449 /* vector == qid - 1, match nvme_create_queue */
1450 if (!nvme_alloc_queue(dev, i, dev->q_depth,
1451 pci_irq_get_node(to_pci_dev(dev->dev), i - 1))) {
1457 max = min(dev->max_qid, dev->queue_count - 1);
1458 for (i = dev->online_queues; i <= max; i++) {
1459 ret = nvme_create_queue(dev->queues[i], i);
1465 * Ignore failing Create SQ/CQ commands, we can continue with less
1466 * than the desired aount of queues, and even a controller without
1467 * I/O queues an still be used to issue admin commands. This might
1468 * be useful to upgrade a buggy firmware for example.
1470 return ret >= 0 ? 0 : ret;
1473 static ssize_t nvme_cmb_show(struct device *dev,
1474 struct device_attribute *attr,
1477 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1479 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1480 ndev->cmbloc, ndev->cmbsz);
1482 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1484 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1486 u64 szu, size, offset;
1487 resource_size_t bar_size;
1488 struct pci_dev *pdev = to_pci_dev(dev->dev);
1490 dma_addr_t dma_addr;
1492 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1493 if (!(NVME_CMB_SZ(dev->cmbsz)))
1495 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1500 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1501 size = szu * NVME_CMB_SZ(dev->cmbsz);
1502 offset = szu * NVME_CMB_OFST(dev->cmbloc);
1503 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1505 if (offset > bar_size)
1509 * Controllers may support a CMB size larger than their BAR,
1510 * for example, due to being behind a bridge. Reduce the CMB to
1511 * the reported size of the BAR
1513 if (size > bar_size - offset)
1514 size = bar_size - offset;
1516 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1517 cmb = ioremap_wc(dma_addr, size);
1521 dev->cmb_dma_addr = dma_addr;
1522 dev->cmb_size = size;
1526 static inline void nvme_release_cmb(struct nvme_dev *dev)
1532 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1533 &dev_attr_cmb.attr, NULL);
1539 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1541 size_t len = dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs);
1542 struct nvme_command c;
1546 dma_addr = dma_map_single(dev->dev, dev->host_mem_descs, len,
1548 if (dma_mapping_error(dev->dev, dma_addr))
1551 memset(&c, 0, sizeof(c));
1552 c.features.opcode = nvme_admin_set_features;
1553 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1554 c.features.dword11 = cpu_to_le32(bits);
1555 c.features.dword12 = cpu_to_le32(dev->host_mem_size >>
1556 ilog2(dev->ctrl.page_size));
1557 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1558 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1559 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1561 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1563 dev_warn(dev->ctrl.device,
1564 "failed to set host mem (err %d, flags %#x).\n",
1567 dma_unmap_single(dev->dev, dma_addr, len, DMA_TO_DEVICE);
1571 static void nvme_free_host_mem(struct nvme_dev *dev)
1575 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1576 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1577 size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1579 dma_free_coherent(dev->dev, size, dev->host_mem_desc_bufs[i],
1580 le64_to_cpu(desc->addr));
1583 kfree(dev->host_mem_desc_bufs);
1584 dev->host_mem_desc_bufs = NULL;
1585 kfree(dev->host_mem_descs);
1586 dev->host_mem_descs = NULL;
1589 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1591 struct nvme_host_mem_buf_desc *descs;
1592 u32 chunk_size, max_entries, i = 0;
1596 /* start big and work our way down */
1597 chunk_size = min(preferred, (u64)PAGE_SIZE << MAX_ORDER);
1599 tmp = (preferred + chunk_size - 1);
1600 do_div(tmp, chunk_size);
1602 descs = kcalloc(max_entries, sizeof(*descs), GFP_KERNEL);
1606 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1608 goto out_free_descs;
1610 for (size = 0; size < preferred; size += chunk_size) {
1611 u32 len = min_t(u64, chunk_size, preferred - size);
1612 dma_addr_t dma_addr;
1614 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1615 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1619 descs[i].addr = cpu_to_le64(dma_addr);
1620 descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1624 if (!size || (min && size < min)) {
1625 dev_warn(dev->ctrl.device,
1626 "failed to allocate host memory buffer.\n");
1630 dev_info(dev->ctrl.device,
1631 "allocated %lld MiB host memory buffer.\n",
1632 size >> ilog2(SZ_1M));
1633 dev->nr_host_mem_descs = i;
1634 dev->host_mem_size = size;
1635 dev->host_mem_descs = descs;
1636 dev->host_mem_desc_bufs = bufs;
1641 size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1643 dma_free_coherent(dev->dev, size, bufs[i],
1644 le64_to_cpu(descs[i].addr));
1651 /* try a smaller chunk size if we failed early */
1652 if (chunk_size >= PAGE_SIZE * 2 && (i == 0 || size < min)) {
1656 dev->host_mem_descs = NULL;
1660 static void nvme_setup_host_mem(struct nvme_dev *dev)
1662 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1663 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1664 u64 min = (u64)dev->ctrl.hmmin * 4096;
1665 u32 enable_bits = NVME_HOST_MEM_ENABLE;
1667 preferred = min(preferred, max);
1669 dev_warn(dev->ctrl.device,
1670 "min host memory (%lld MiB) above limit (%d MiB).\n",
1671 min >> ilog2(SZ_1M), max_host_mem_size_mb);
1672 nvme_free_host_mem(dev);
1677 * If we already have a buffer allocated check if we can reuse it.
1679 if (dev->host_mem_descs) {
1680 if (dev->host_mem_size >= min)
1681 enable_bits |= NVME_HOST_MEM_RETURN;
1683 nvme_free_host_mem(dev);
1686 if (!dev->host_mem_descs) {
1687 if (nvme_alloc_host_mem(dev, min, preferred))
1691 if (nvme_set_host_mem(dev, enable_bits))
1692 nvme_free_host_mem(dev);
1695 static int nvme_setup_io_queues(struct nvme_dev *dev)
1697 struct nvme_queue *adminq = dev->queues[0];
1698 struct pci_dev *pdev = to_pci_dev(dev->dev);
1699 int result, nr_io_queues;
1702 nr_io_queues = num_online_cpus();
1703 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1707 if (nr_io_queues == 0)
1710 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1711 result = nvme_cmb_qdepth(dev, nr_io_queues,
1712 sizeof(struct nvme_command));
1714 dev->q_depth = result;
1716 nvme_release_cmb(dev);
1720 size = db_bar_size(dev, nr_io_queues);
1721 result = nvme_remap_bar(dev, size);
1724 if (!--nr_io_queues)
1727 adminq->q_db = dev->dbs;
1729 /* Deregister the admin queue's interrupt */
1730 pci_free_irq(pdev, 0, adminq);
1733 * If we enable msix early due to not intx, disable it again before
1734 * setting up the full range we need.
1736 pci_free_irq_vectors(pdev);
1737 nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1738 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1739 if (nr_io_queues <= 0)
1741 dev->max_qid = nr_io_queues;
1744 * Should investigate if there's a performance win from allocating
1745 * more queues than interrupt vectors; it might allow the submission
1746 * path to scale better, even if the receive path is limited by the
1747 * number of interrupts.
1750 result = queue_request_irq(adminq);
1752 adminq->cq_vector = -1;
1755 return nvme_create_io_queues(dev);
1758 static void nvme_del_queue_end(struct request *req, blk_status_t error)
1760 struct nvme_queue *nvmeq = req->end_io_data;
1762 blk_mq_free_request(req);
1763 complete(&nvmeq->dev->ioq_wait);
1766 static void nvme_del_cq_end(struct request *req, blk_status_t error)
1768 struct nvme_queue *nvmeq = req->end_io_data;
1771 unsigned long flags;
1774 * We might be called with the AQ q_lock held
1775 * and the I/O queue q_lock should always
1776 * nest inside the AQ one.
1778 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1779 SINGLE_DEPTH_NESTING);
1780 nvme_process_cq(nvmeq);
1781 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1784 nvme_del_queue_end(req, error);
1787 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1789 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1790 struct request *req;
1791 struct nvme_command cmd;
1793 memset(&cmd, 0, sizeof(cmd));
1794 cmd.delete_queue.opcode = opcode;
1795 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1797 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1799 return PTR_ERR(req);
1801 req->timeout = ADMIN_TIMEOUT;
1802 req->end_io_data = nvmeq;
1804 blk_execute_rq_nowait(q, NULL, req, false,
1805 opcode == nvme_admin_delete_cq ?
1806 nvme_del_cq_end : nvme_del_queue_end);
1810 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1813 unsigned long timeout;
1814 u8 opcode = nvme_admin_delete_sq;
1816 for (pass = 0; pass < 2; pass++) {
1817 int sent = 0, i = queues;
1819 reinit_completion(&dev->ioq_wait);
1821 timeout = ADMIN_TIMEOUT;
1822 for (; i > 0; i--, sent++)
1823 if (nvme_delete_queue(dev->queues[i], opcode))
1827 timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1833 opcode = nvme_admin_delete_cq;
1838 * Return: error value if an error occurred setting up the queues or calling
1839 * Identify Device. 0 if these succeeded, even if adding some of the
1840 * namespaces failed. At the moment, these failures are silent. TBD which
1841 * failures should be reported.
1843 static int nvme_dev_add(struct nvme_dev *dev)
1845 if (!dev->ctrl.tagset) {
1846 dev->tagset.ops = &nvme_mq_ops;
1847 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1848 dev->tagset.timeout = NVME_IO_TIMEOUT;
1849 dev->tagset.numa_node = dev_to_node(dev->dev);
1850 dev->tagset.queue_depth =
1851 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1852 dev->tagset.cmd_size = nvme_cmd_size(dev);
1853 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1854 dev->tagset.driver_data = dev;
1856 if (blk_mq_alloc_tag_set(&dev->tagset))
1858 dev->ctrl.tagset = &dev->tagset;
1860 nvme_dbbuf_set(dev);
1862 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1864 /* Free previously allocated queues that are no longer usable */
1865 nvme_free_queues(dev, dev->online_queues);
1871 static int nvme_pci_enable(struct nvme_dev *dev)
1874 int result = -ENOMEM;
1875 struct pci_dev *pdev = to_pci_dev(dev->dev);
1877 if (pci_enable_device_mem(pdev))
1880 pci_set_master(pdev);
1882 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1883 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1886 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1892 * Some devices and/or platforms don't advertise or work with INTx
1893 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1894 * adjust this later.
1896 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1900 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1902 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1903 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1904 dev->dbs = dev->bar + 4096;
1907 * Temporary fix for the Apple controller found in the MacBook8,1 and
1908 * some MacBook7,1 to avoid controller resets and data loss.
1910 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1912 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
1913 "set queue depth=%u to work around controller resets\n",
1918 * CMBs can currently only exist on >=1.2 PCIe devices. We only
1919 * populate sysfs if a CMB is implemented. Note that we add the
1920 * CMB attribute to the nvme_ctrl kobj which removes the need to remove
1921 * it on exit. Since nvme_dev_attrs_group has no name we can pass
1922 * NULL as final argument to sysfs_add_file_to_group.
1925 if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1926 dev->cmb = nvme_map_cmb(dev);
1929 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1930 &dev_attr_cmb.attr, NULL))
1931 dev_warn(dev->ctrl.device,
1932 "failed to add sysfs attribute for CMB\n");
1936 pci_enable_pcie_error_reporting(pdev);
1937 pci_save_state(pdev);
1941 pci_disable_device(pdev);
1945 static void nvme_dev_unmap(struct nvme_dev *dev)
1949 pci_release_mem_regions(to_pci_dev(dev->dev));
1952 static void nvme_pci_disable(struct nvme_dev *dev)
1954 struct pci_dev *pdev = to_pci_dev(dev->dev);
1956 nvme_release_cmb(dev);
1957 pci_free_irq_vectors(pdev);
1959 if (pci_is_enabled(pdev)) {
1960 pci_disable_pcie_error_reporting(pdev);
1961 pci_disable_device(pdev);
1965 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1969 struct pci_dev *pdev = to_pci_dev(dev->dev);
1971 mutex_lock(&dev->shutdown_lock);
1972 if (pci_is_enabled(pdev)) {
1973 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1975 if (dev->ctrl.state == NVME_CTRL_LIVE)
1976 nvme_start_freeze(&dev->ctrl);
1977 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
1978 pdev->error_state != pci_channel_io_normal);
1982 * Give the controller a chance to complete all entered requests if
1983 * doing a safe shutdown.
1987 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
1990 * If the controller is still alive tell it to stop using the
1991 * host memory buffer. In theory the shutdown / reset should
1992 * make sure that it doesn't access the host memoery anymore,
1993 * but I'd rather be safe than sorry..
1995 if (dev->host_mem_descs)
1996 nvme_set_host_mem(dev, 0);
1999 nvme_stop_queues(&dev->ctrl);
2001 queues = dev->online_queues - 1;
2002 for (i = dev->queue_count - 1; i > 0; i--)
2003 nvme_suspend_queue(dev->queues[i]);
2006 /* A device might become IO incapable very soon during
2007 * probe, before the admin queue is configured. Thus,
2008 * queue_count can be 0 here.
2010 if (dev->queue_count)
2011 nvme_suspend_queue(dev->queues[0]);
2013 nvme_disable_io_queues(dev, queues);
2014 nvme_disable_admin_queue(dev, shutdown);
2016 nvme_pci_disable(dev);
2018 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2019 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2022 * The driver will not be starting up queues again if shutting down so
2023 * must flush all entered requests to their failed completion to avoid
2024 * deadlocking blk-mq hot-cpu notifier.
2027 nvme_start_queues(&dev->ctrl);
2028 mutex_unlock(&dev->shutdown_lock);
2031 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2033 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2034 PAGE_SIZE, PAGE_SIZE, 0);
2035 if (!dev->prp_page_pool)
2038 /* Optimisation for I/Os between 4k and 128k */
2039 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2041 if (!dev->prp_small_pool) {
2042 dma_pool_destroy(dev->prp_page_pool);
2048 static void nvme_release_prp_pools(struct nvme_dev *dev)
2050 dma_pool_destroy(dev->prp_page_pool);
2051 dma_pool_destroy(dev->prp_small_pool);
2054 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2056 struct nvme_dev *dev = to_nvme_dev(ctrl);
2058 nvme_dbbuf_dma_free(dev);
2059 put_device(dev->dev);
2060 if (dev->tagset.tags)
2061 blk_mq_free_tag_set(&dev->tagset);
2062 if (dev->ctrl.admin_q)
2063 blk_put_queue(dev->ctrl.admin_q);
2065 free_opal_dev(dev->ctrl.opal_dev);
2069 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
2071 dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
2073 kref_get(&dev->ctrl.kref);
2074 nvme_dev_disable(dev, false);
2075 if (!schedule_work(&dev->remove_work))
2076 nvme_put_ctrl(&dev->ctrl);
2079 static void nvme_reset_work(struct work_struct *work)
2081 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2082 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2083 int result = -ENODEV;
2085 if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
2089 * If we're called to reset a live controller first shut it down before
2092 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2093 nvme_dev_disable(dev, false);
2095 result = nvme_pci_enable(dev);
2099 result = nvme_configure_admin_queue(dev);
2103 nvme_init_queue(dev->queues[0], 0);
2104 result = nvme_alloc_admin_tags(dev);
2108 result = nvme_init_identify(&dev->ctrl);
2112 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2113 if (!dev->ctrl.opal_dev)
2114 dev->ctrl.opal_dev =
2115 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2116 else if (was_suspend)
2117 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2119 free_opal_dev(dev->ctrl.opal_dev);
2120 dev->ctrl.opal_dev = NULL;
2123 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2124 result = nvme_dbbuf_dma_alloc(dev);
2127 "unable to allocate dma for dbbuf\n");
2130 if (dev->ctrl.hmpre)
2131 nvme_setup_host_mem(dev);
2133 result = nvme_setup_io_queues(dev);
2138 * A controller that can not execute IO typically requires user
2139 * intervention to correct. For such degraded controllers, the driver
2140 * should not submit commands the user did not request, so skip
2141 * registering for asynchronous event notification on this condition.
2143 if (dev->online_queues > 1)
2144 nvme_queue_async_events(&dev->ctrl);
2147 * Keep the controller around but remove all namespaces if we don't have
2148 * any working I/O queue.
2150 if (dev->online_queues < 2) {
2151 dev_warn(dev->ctrl.device, "IO queues not created\n");
2152 nvme_kill_queues(&dev->ctrl);
2153 nvme_remove_namespaces(&dev->ctrl);
2155 nvme_start_queues(&dev->ctrl);
2156 nvme_wait_freeze(&dev->ctrl);
2158 nvme_unfreeze(&dev->ctrl);
2161 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2162 dev_warn(dev->ctrl.device, "failed to mark controller live\n");
2166 if (dev->online_queues > 1)
2167 nvme_queue_scan(&dev->ctrl);
2171 nvme_remove_dead_ctrl(dev, result);
2174 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2176 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2177 struct pci_dev *pdev = to_pci_dev(dev->dev);
2179 nvme_kill_queues(&dev->ctrl);
2180 if (pci_get_drvdata(pdev))
2181 device_release_driver(&pdev->dev);
2182 nvme_put_ctrl(&dev->ctrl);
2185 static int nvme_reset(struct nvme_dev *dev)
2187 if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q))
2189 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING))
2191 if (!queue_work(nvme_wq, &dev->reset_work))
2196 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2198 *val = readl(to_nvme_dev(ctrl)->bar + off);
2202 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2204 writel(val, to_nvme_dev(ctrl)->bar + off);
2208 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2210 *val = readq(to_nvme_dev(ctrl)->bar + off);
2214 static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl)
2216 struct nvme_dev *dev = to_nvme_dev(ctrl);
2217 int ret = nvme_reset(dev);
2220 flush_work(&dev->reset_work);
2224 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2226 .module = THIS_MODULE,
2227 .flags = NVME_F_METADATA_SUPPORTED,
2228 .reg_read32 = nvme_pci_reg_read32,
2229 .reg_write32 = nvme_pci_reg_write32,
2230 .reg_read64 = nvme_pci_reg_read64,
2231 .reset_ctrl = nvme_pci_reset_ctrl,
2232 .free_ctrl = nvme_pci_free_ctrl,
2233 .submit_async_event = nvme_pci_submit_async_event,
2236 static int nvme_dev_map(struct nvme_dev *dev)
2238 struct pci_dev *pdev = to_pci_dev(dev->dev);
2240 if (pci_request_mem_regions(pdev, "nvme"))
2243 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2248 pci_release_mem_regions(pdev);
2252 static unsigned long check_dell_samsung_bug(struct pci_dev *pdev)
2254 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2256 * Several Samsung devices seem to drop off the PCIe bus
2257 * randomly when APST is on and uses the deepest sleep state.
2258 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2259 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2260 * 950 PRO 256GB", but it seems to be restricted to two Dell
2263 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2264 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2265 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2266 return NVME_QUIRK_NO_DEEPEST_PS;
2272 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2274 int node, result = -ENOMEM;
2275 struct nvme_dev *dev;
2276 unsigned long quirks = id->driver_data;
2278 node = dev_to_node(&pdev->dev);
2279 if (node == NUMA_NO_NODE)
2280 set_dev_node(&pdev->dev, first_memory_node);
2282 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2285 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2290 dev->dev = get_device(&pdev->dev);
2291 pci_set_drvdata(pdev, dev);
2293 result = nvme_dev_map(dev);
2297 INIT_WORK(&dev->reset_work, nvme_reset_work);
2298 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2299 mutex_init(&dev->shutdown_lock);
2300 init_completion(&dev->ioq_wait);
2302 result = nvme_setup_prp_pools(dev);
2306 quirks |= check_dell_samsung_bug(pdev);
2308 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2313 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING);
2314 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2316 queue_work(nvme_wq, &dev->reset_work);
2320 nvme_release_prp_pools(dev);
2322 put_device(dev->dev);
2323 nvme_dev_unmap(dev);
2330 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2332 struct nvme_dev *dev = pci_get_drvdata(pdev);
2335 nvme_dev_disable(dev, false);
2340 static void nvme_shutdown(struct pci_dev *pdev)
2342 struct nvme_dev *dev = pci_get_drvdata(pdev);
2343 nvme_dev_disable(dev, true);
2347 * The driver's remove may be called on a device in a partially initialized
2348 * state. This function must not have any dependencies on the device state in
2351 static void nvme_remove(struct pci_dev *pdev)
2353 struct nvme_dev *dev = pci_get_drvdata(pdev);
2355 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2357 cancel_work_sync(&dev->reset_work);
2358 pci_set_drvdata(pdev, NULL);
2360 if (!pci_device_is_present(pdev)) {
2361 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2362 nvme_dev_disable(dev, false);
2365 flush_work(&dev->reset_work);
2366 nvme_uninit_ctrl(&dev->ctrl);
2367 nvme_dev_disable(dev, true);
2368 nvme_free_host_mem(dev);
2369 nvme_dev_remove_admin(dev);
2370 nvme_free_queues(dev, 0);
2371 nvme_release_prp_pools(dev);
2372 nvme_dev_unmap(dev);
2373 nvme_put_ctrl(&dev->ctrl);
2376 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2381 if (pci_vfs_assigned(pdev)) {
2382 dev_warn(&pdev->dev,
2383 "Cannot disable SR-IOV VFs while assigned\n");
2386 pci_disable_sriov(pdev);
2390 ret = pci_enable_sriov(pdev, numvfs);
2391 return ret ? ret : numvfs;
2394 #ifdef CONFIG_PM_SLEEP
2395 static int nvme_suspend(struct device *dev)
2397 struct pci_dev *pdev = to_pci_dev(dev);
2398 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2400 nvme_dev_disable(ndev, true);
2404 static int nvme_resume(struct device *dev)
2406 struct pci_dev *pdev = to_pci_dev(dev);
2407 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2414 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2416 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2417 pci_channel_state_t state)
2419 struct nvme_dev *dev = pci_get_drvdata(pdev);
2422 * A frozen channel requires a reset. When detected, this method will
2423 * shutdown the controller to quiesce. The controller will be restarted
2424 * after the slot reset through driver's slot_reset callback.
2427 case pci_channel_io_normal:
2428 return PCI_ERS_RESULT_CAN_RECOVER;
2429 case pci_channel_io_frozen:
2430 dev_warn(dev->ctrl.device,
2431 "frozen state error detected, reset controller\n");
2432 nvme_dev_disable(dev, false);
2433 return PCI_ERS_RESULT_NEED_RESET;
2434 case pci_channel_io_perm_failure:
2435 dev_warn(dev->ctrl.device,
2436 "failure state error detected, request disconnect\n");
2437 return PCI_ERS_RESULT_DISCONNECT;
2439 return PCI_ERS_RESULT_NEED_RESET;
2442 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2444 struct nvme_dev *dev = pci_get_drvdata(pdev);
2446 dev_info(dev->ctrl.device, "restart after slot reset\n");
2447 pci_restore_state(pdev);
2449 return PCI_ERS_RESULT_RECOVERED;
2452 static void nvme_error_resume(struct pci_dev *pdev)
2454 pci_cleanup_aer_uncorrect_error_status(pdev);
2457 static const struct pci_error_handlers nvme_err_handler = {
2458 .error_detected = nvme_error_detected,
2459 .slot_reset = nvme_slot_reset,
2460 .resume = nvme_error_resume,
2461 .reset_notify = nvme_reset_notify,
2464 static const struct pci_device_id nvme_id_table[] = {
2465 { PCI_VDEVICE(INTEL, 0x0953),
2466 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2467 NVME_QUIRK_DEALLOCATE_ZEROES, },
2468 { PCI_VDEVICE(INTEL, 0x0a53),
2469 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2470 NVME_QUIRK_DEALLOCATE_ZEROES, },
2471 { PCI_VDEVICE(INTEL, 0x0a54),
2472 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2473 NVME_QUIRK_DEALLOCATE_ZEROES, },
2474 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
2475 .driver_data = NVME_QUIRK_NO_DEEPEST_PS },
2476 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
2477 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2478 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
2479 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2480 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
2481 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2482 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2483 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2484 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2487 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2489 static struct pci_driver nvme_driver = {
2491 .id_table = nvme_id_table,
2492 .probe = nvme_probe,
2493 .remove = nvme_remove,
2494 .shutdown = nvme_shutdown,
2496 .pm = &nvme_dev_pm_ops,
2498 .sriov_configure = nvme_pci_sriov_configure,
2499 .err_handler = &nvme_err_handler,
2502 static int __init nvme_init(void)
2504 return pci_register_driver(&nvme_driver);
2507 static void __exit nvme_exit(void)
2509 pci_unregister_driver(&nvme_driver);
2513 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2514 MODULE_LICENSE("GPL");
2515 MODULE_VERSION("1.0");
2516 module_init(nvme_init);
2517 module_exit(nvme_exit);