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 void nvme_process_cq(struct nvme_queue *nvmeq);
65 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
68 * Represents an NVM Express device. Each nvme_dev is a PCI function.
71 struct nvme_queue **queues;
72 struct blk_mq_tag_set tagset;
73 struct blk_mq_tag_set admin_tagset;
76 struct dma_pool *prp_page_pool;
77 struct dma_pool *prp_small_pool;
79 unsigned online_queues;
84 unsigned long bar_mapped_size;
85 struct work_struct remove_work;
86 struct mutex shutdown_lock;
89 dma_addr_t cmb_dma_addr;
93 struct nvme_ctrl ctrl;
94 struct completion ioq_wait;
96 /* shadow doorbell buffer support: */
98 dma_addr_t dbbuf_dbs_dma_addr;
100 dma_addr_t dbbuf_eis_dma_addr;
102 /* host memory buffer support: */
104 u32 nr_host_mem_descs;
105 struct nvme_host_mem_buf_desc *host_mem_descs;
106 void **host_mem_desc_bufs;
109 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
111 return qid * 2 * stride;
114 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
116 return (qid * 2 + 1) * stride;
119 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
121 return container_of(ctrl, struct nvme_dev, ctrl);
125 * An NVM Express queue. Each device has at least two (one for admin
126 * commands and one for I/O commands).
129 struct device *q_dmadev;
130 struct nvme_dev *dev;
132 struct nvme_command *sq_cmds;
133 struct nvme_command __iomem *sq_cmds_io;
134 volatile struct nvme_completion *cqes;
135 struct blk_mq_tags **tags;
136 dma_addr_t sq_dma_addr;
137 dma_addr_t cq_dma_addr;
153 * The nvme_iod describes the data in an I/O, including the list of PRP
154 * entries. You can't see it in this data structure because C doesn't let
155 * me express that. Use nvme_init_iod to ensure there's enough space
156 * allocated to store the PRP list.
159 struct nvme_request req;
160 struct nvme_queue *nvmeq;
162 int npages; /* In the PRP list. 0 means small pool in use */
163 int nents; /* Used in scatterlist */
164 int length; /* Of data, in bytes */
165 dma_addr_t first_dma;
166 struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
167 struct scatterlist *sg;
168 struct scatterlist inline_sg[0];
172 * Check we didin't inadvertently grow the command struct
174 static inline void _nvme_check_size(void)
176 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
177 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
178 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
179 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
180 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
181 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
182 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
183 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
184 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
185 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
186 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
187 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
188 BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
191 static inline unsigned int nvme_dbbuf_size(u32 stride)
193 return ((num_possible_cpus() + 1) * 8 * stride);
196 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
198 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
203 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
204 &dev->dbbuf_dbs_dma_addr,
208 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
209 &dev->dbbuf_eis_dma_addr,
211 if (!dev->dbbuf_eis) {
212 dma_free_coherent(dev->dev, mem_size,
213 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
214 dev->dbbuf_dbs = NULL;
221 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
223 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
225 if (dev->dbbuf_dbs) {
226 dma_free_coherent(dev->dev, mem_size,
227 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
228 dev->dbbuf_dbs = NULL;
230 if (dev->dbbuf_eis) {
231 dma_free_coherent(dev->dev, mem_size,
232 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
233 dev->dbbuf_eis = NULL;
237 static void nvme_dbbuf_init(struct nvme_dev *dev,
238 struct nvme_queue *nvmeq, int qid)
240 if (!dev->dbbuf_dbs || !qid)
243 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
244 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
245 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
246 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
249 static void nvme_dbbuf_set(struct nvme_dev *dev)
251 struct nvme_command c;
256 memset(&c, 0, sizeof(c));
257 c.dbbuf.opcode = nvme_admin_dbbuf;
258 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
259 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
261 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
262 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
263 /* Free memory and continue on */
264 nvme_dbbuf_dma_free(dev);
268 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
270 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
273 /* Update dbbuf and return true if an MMIO is required */
274 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
275 volatile u32 *dbbuf_ei)
281 * Ensure that the queue is written before updating
282 * the doorbell in memory
286 old_value = *dbbuf_db;
289 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
297 * Max size of iod being embedded in the request payload
299 #define NVME_INT_PAGES 2
300 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size)
303 * Will slightly overestimate the number of pages needed. This is OK
304 * as it only leads to a small amount of wasted memory for the lifetime of
307 static int nvme_npages(unsigned size, struct nvme_dev *dev)
309 unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
310 dev->ctrl.page_size);
311 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
314 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
315 unsigned int size, unsigned int nseg)
317 return sizeof(__le64 *) * nvme_npages(size, dev) +
318 sizeof(struct scatterlist) * nseg;
321 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
323 return sizeof(struct nvme_iod) +
324 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
327 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
328 unsigned int hctx_idx)
330 struct nvme_dev *dev = data;
331 struct nvme_queue *nvmeq = dev->queues[0];
333 WARN_ON(hctx_idx != 0);
334 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
335 WARN_ON(nvmeq->tags);
337 hctx->driver_data = nvmeq;
338 nvmeq->tags = &dev->admin_tagset.tags[0];
342 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
344 struct nvme_queue *nvmeq = hctx->driver_data;
349 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
350 unsigned int hctx_idx)
352 struct nvme_dev *dev = data;
353 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
356 nvmeq->tags = &dev->tagset.tags[hctx_idx];
358 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
359 hctx->driver_data = nvmeq;
363 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
364 unsigned int hctx_idx, unsigned int numa_node)
366 struct nvme_dev *dev = set->driver_data;
367 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
368 int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
369 struct nvme_queue *nvmeq = dev->queues[queue_idx];
376 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
378 struct nvme_dev *dev = set->driver_data;
380 return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
384 * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
385 * @nvmeq: The queue to use
386 * @cmd: The command to send
388 * Safe to use from interrupt context
390 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
391 struct nvme_command *cmd)
393 u16 tail = nvmeq->sq_tail;
395 if (nvmeq->sq_cmds_io)
396 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
398 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
400 if (++tail == nvmeq->q_depth)
402 if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
404 writel(tail, nvmeq->q_db);
405 nvmeq->sq_tail = tail;
408 static __le64 **iod_list(struct request *req)
410 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
411 return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
414 static blk_status_t nvme_init_iod(struct request *rq, struct nvme_dev *dev)
416 struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
417 int nseg = blk_rq_nr_phys_segments(rq);
418 unsigned int size = blk_rq_payload_bytes(rq);
420 if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
421 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
423 return BLK_STS_RESOURCE;
425 iod->sg = iod->inline_sg;
436 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
438 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
439 const int last_prp = dev->ctrl.page_size / 8 - 1;
441 __le64 **list = iod_list(req);
442 dma_addr_t prp_dma = iod->first_dma;
444 if (iod->npages == 0)
445 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
446 for (i = 0; i < iod->npages; i++) {
447 __le64 *prp_list = list[i];
448 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
449 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
450 prp_dma = next_prp_dma;
453 if (iod->sg != iod->inline_sg)
457 #ifdef CONFIG_BLK_DEV_INTEGRITY
458 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
460 if (be32_to_cpu(pi->ref_tag) == v)
461 pi->ref_tag = cpu_to_be32(p);
464 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
466 if (be32_to_cpu(pi->ref_tag) == p)
467 pi->ref_tag = cpu_to_be32(v);
471 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
473 * The virtual start sector is the one that was originally submitted by the
474 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
475 * start sector may be different. Remap protection information to match the
476 * physical LBA on writes, and back to the original seed on reads.
478 * Type 0 and 3 do not have a ref tag, so no remapping required.
480 static void nvme_dif_remap(struct request *req,
481 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
483 struct nvme_ns *ns = req->rq_disk->private_data;
484 struct bio_integrity_payload *bip;
485 struct t10_pi_tuple *pi;
487 u32 i, nlb, ts, phys, virt;
489 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
492 bip = bio_integrity(req->bio);
496 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
499 virt = bip_get_seed(bip);
500 phys = nvme_block_nr(ns, blk_rq_pos(req));
501 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
502 ts = ns->disk->queue->integrity.tuple_size;
504 for (i = 0; i < nlb; i++, virt++, phys++) {
505 pi = (struct t10_pi_tuple *)p;
506 dif_swap(phys, virt, pi);
511 #else /* CONFIG_BLK_DEV_INTEGRITY */
512 static void nvme_dif_remap(struct request *req,
513 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
516 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
519 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
524 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req)
526 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
527 struct dma_pool *pool;
528 int length = blk_rq_payload_bytes(req);
529 struct scatterlist *sg = iod->sg;
530 int dma_len = sg_dma_len(sg);
531 u64 dma_addr = sg_dma_address(sg);
532 u32 page_size = dev->ctrl.page_size;
533 int offset = dma_addr & (page_size - 1);
535 __le64 **list = iod_list(req);
539 length -= (page_size - offset);
543 dma_len -= (page_size - offset);
545 dma_addr += (page_size - offset);
548 dma_addr = sg_dma_address(sg);
549 dma_len = sg_dma_len(sg);
552 if (length <= page_size) {
553 iod->first_dma = dma_addr;
557 nprps = DIV_ROUND_UP(length, page_size);
558 if (nprps <= (256 / 8)) {
559 pool = dev->prp_small_pool;
562 pool = dev->prp_page_pool;
566 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
568 iod->first_dma = dma_addr;
573 iod->first_dma = prp_dma;
576 if (i == page_size >> 3) {
577 __le64 *old_prp_list = prp_list;
578 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
581 list[iod->npages++] = prp_list;
582 prp_list[0] = old_prp_list[i - 1];
583 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
586 prp_list[i++] = cpu_to_le64(dma_addr);
587 dma_len -= page_size;
588 dma_addr += page_size;
596 dma_addr = sg_dma_address(sg);
597 dma_len = sg_dma_len(sg);
603 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
604 struct nvme_command *cmnd)
606 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
607 struct request_queue *q = req->q;
608 enum dma_data_direction dma_dir = rq_data_dir(req) ?
609 DMA_TO_DEVICE : DMA_FROM_DEVICE;
610 blk_status_t ret = BLK_STS_IOERR;
612 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
613 iod->nents = blk_rq_map_sg(q, req, iod->sg);
617 ret = BLK_STS_RESOURCE;
618 if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
622 if (!nvme_setup_prps(dev, req))
626 if (blk_integrity_rq(req)) {
627 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
630 sg_init_table(&iod->meta_sg, 1);
631 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
634 if (rq_data_dir(req))
635 nvme_dif_remap(req, nvme_dif_prep);
637 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
641 cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
642 cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
643 if (blk_integrity_rq(req))
644 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
648 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
653 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
655 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
656 enum dma_data_direction dma_dir = rq_data_dir(req) ?
657 DMA_TO_DEVICE : DMA_FROM_DEVICE;
660 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
661 if (blk_integrity_rq(req)) {
662 if (!rq_data_dir(req))
663 nvme_dif_remap(req, nvme_dif_complete);
664 dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
668 nvme_cleanup_cmd(req);
669 nvme_free_iod(dev, req);
673 * NOTE: ns is NULL when called on the admin queue.
675 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
676 const struct blk_mq_queue_data *bd)
678 struct nvme_ns *ns = hctx->queue->queuedata;
679 struct nvme_queue *nvmeq = hctx->driver_data;
680 struct nvme_dev *dev = nvmeq->dev;
681 struct request *req = bd->rq;
682 struct nvme_command cmnd;
685 ret = nvme_setup_cmd(ns, req, &cmnd);
689 ret = nvme_init_iod(req, dev);
693 if (blk_rq_nr_phys_segments(req)) {
694 ret = nvme_map_data(dev, req, &cmnd);
696 goto out_cleanup_iod;
699 blk_mq_start_request(req);
701 spin_lock_irq(&nvmeq->q_lock);
702 if (unlikely(nvmeq->cq_vector < 0)) {
704 spin_unlock_irq(&nvmeq->q_lock);
705 goto out_cleanup_iod;
707 __nvme_submit_cmd(nvmeq, &cmnd);
708 nvme_process_cq(nvmeq);
709 spin_unlock_irq(&nvmeq->q_lock);
712 nvme_free_iod(dev, req);
714 nvme_cleanup_cmd(req);
718 static void nvme_pci_complete_rq(struct request *req)
720 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
722 nvme_unmap_data(iod->nvmeq->dev, req);
723 nvme_complete_rq(req);
726 /* We read the CQE phase first to check if the rest of the entry is valid */
727 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
730 return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
733 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
735 u16 head = nvmeq->cq_head;
737 if (likely(nvmeq->cq_vector >= 0)) {
738 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
740 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
744 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
745 struct nvme_completion *cqe)
749 if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
750 dev_warn(nvmeq->dev->ctrl.device,
751 "invalid id %d completed on queue %d\n",
752 cqe->command_id, le16_to_cpu(cqe->sq_id));
757 * AEN requests are special as they don't time out and can
758 * survive any kind of queue freeze and often don't respond to
759 * aborts. We don't even bother to allocate a struct request
760 * for them but rather special case them here.
762 if (unlikely(nvmeq->qid == 0 &&
763 cqe->command_id >= NVME_AQ_BLKMQ_DEPTH)) {
764 nvme_complete_async_event(&nvmeq->dev->ctrl,
765 cqe->status, &cqe->result);
769 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
770 nvme_end_request(req, cqe->status, cqe->result);
773 static inline bool nvme_read_cqe(struct nvme_queue *nvmeq,
774 struct nvme_completion *cqe)
776 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
777 *cqe = nvmeq->cqes[nvmeq->cq_head];
779 if (++nvmeq->cq_head == nvmeq->q_depth) {
781 nvmeq->cq_phase = !nvmeq->cq_phase;
788 static void __nvme_process_cq(struct nvme_queue *nvmeq, int *tag)
790 struct nvme_completion cqe;
793 while (nvme_read_cqe(nvmeq, &cqe)) {
794 nvme_handle_cqe(nvmeq, &cqe);
797 if (tag && *tag == cqe.command_id) {
804 nvme_ring_cq_doorbell(nvmeq);
809 static void nvme_process_cq(struct nvme_queue *nvmeq)
811 __nvme_process_cq(nvmeq, NULL);
814 static irqreturn_t nvme_irq(int irq, void *data)
817 struct nvme_queue *nvmeq = data;
818 spin_lock(&nvmeq->q_lock);
819 nvme_process_cq(nvmeq);
820 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
822 spin_unlock(&nvmeq->q_lock);
826 static irqreturn_t nvme_irq_check(int irq, void *data)
828 struct nvme_queue *nvmeq = data;
829 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
830 return IRQ_WAKE_THREAD;
834 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
836 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
837 spin_lock_irq(&nvmeq->q_lock);
838 __nvme_process_cq(nvmeq, &tag);
839 spin_unlock_irq(&nvmeq->q_lock);
848 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
850 struct nvme_queue *nvmeq = hctx->driver_data;
852 return __nvme_poll(nvmeq, tag);
855 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
857 struct nvme_dev *dev = to_nvme_dev(ctrl);
858 struct nvme_queue *nvmeq = dev->queues[0];
859 struct nvme_command c;
861 memset(&c, 0, sizeof(c));
862 c.common.opcode = nvme_admin_async_event;
863 c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
865 spin_lock_irq(&nvmeq->q_lock);
866 __nvme_submit_cmd(nvmeq, &c);
867 spin_unlock_irq(&nvmeq->q_lock);
870 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
872 struct nvme_command c;
874 memset(&c, 0, sizeof(c));
875 c.delete_queue.opcode = opcode;
876 c.delete_queue.qid = cpu_to_le16(id);
878 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
881 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
882 struct nvme_queue *nvmeq)
884 struct nvme_command c;
885 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
888 * Note: we (ab)use the fact the the prp fields survive if no data
889 * is attached to the request.
891 memset(&c, 0, sizeof(c));
892 c.create_cq.opcode = nvme_admin_create_cq;
893 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
894 c.create_cq.cqid = cpu_to_le16(qid);
895 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
896 c.create_cq.cq_flags = cpu_to_le16(flags);
897 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
899 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
902 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
903 struct nvme_queue *nvmeq)
905 struct nvme_command c;
906 int flags = NVME_QUEUE_PHYS_CONTIG;
909 * Note: we (ab)use the fact the the prp fields survive if no data
910 * is attached to the request.
912 memset(&c, 0, sizeof(c));
913 c.create_sq.opcode = nvme_admin_create_sq;
914 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
915 c.create_sq.sqid = cpu_to_le16(qid);
916 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
917 c.create_sq.sq_flags = cpu_to_le16(flags);
918 c.create_sq.cqid = cpu_to_le16(qid);
920 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
923 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
925 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
928 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
930 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
933 static void abort_endio(struct request *req, blk_status_t error)
935 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
936 struct nvme_queue *nvmeq = iod->nvmeq;
938 dev_warn(nvmeq->dev->ctrl.device,
939 "Abort status: 0x%x", nvme_req(req)->status);
940 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
941 blk_mq_free_request(req);
944 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
947 /* If true, indicates loss of adapter communication, possibly by a
948 * NVMe Subsystem reset.
950 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
952 /* If there is a reset ongoing, we shouldn't reset again. */
953 if (dev->ctrl.state == NVME_CTRL_RESETTING)
956 /* We shouldn't reset unless the controller is on fatal error state
957 * _or_ if we lost the communication with it.
959 if (!(csts & NVME_CSTS_CFS) && !nssro)
962 /* If PCI error recovery process is happening, we cannot reset or
963 * the recovery mechanism will surely fail.
965 if (pci_channel_offline(to_pci_dev(dev->dev)))
971 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
973 /* Read a config register to help see what died. */
977 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
979 if (result == PCIBIOS_SUCCESSFUL)
980 dev_warn(dev->ctrl.device,
981 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
984 dev_warn(dev->ctrl.device,
985 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
989 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
991 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
992 struct nvme_queue *nvmeq = iod->nvmeq;
993 struct nvme_dev *dev = nvmeq->dev;
994 struct request *abort_req;
995 struct nvme_command cmd;
996 u32 csts = readl(dev->bar + NVME_REG_CSTS);
999 * Reset immediately if the controller is failed
1001 if (nvme_should_reset(dev, csts)) {
1002 nvme_warn_reset(dev, csts);
1003 nvme_dev_disable(dev, false);
1004 nvme_reset_ctrl(&dev->ctrl);
1005 return BLK_EH_HANDLED;
1009 * Did we miss an interrupt?
1011 if (__nvme_poll(nvmeq, req->tag)) {
1012 dev_warn(dev->ctrl.device,
1013 "I/O %d QID %d timeout, completion polled\n",
1014 req->tag, nvmeq->qid);
1015 return BLK_EH_HANDLED;
1019 * Shutdown immediately if controller times out while starting. The
1020 * reset work will see the pci device disabled when it gets the forced
1021 * cancellation error. All outstanding requests are completed on
1022 * shutdown, so we return BLK_EH_HANDLED.
1024 if (dev->ctrl.state == NVME_CTRL_RESETTING) {
1025 dev_warn(dev->ctrl.device,
1026 "I/O %d QID %d timeout, disable controller\n",
1027 req->tag, nvmeq->qid);
1028 nvme_dev_disable(dev, false);
1029 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1030 return BLK_EH_HANDLED;
1034 * Shutdown the controller immediately and schedule a reset if the
1035 * command was already aborted once before and still hasn't been
1036 * returned to the driver, or if this is the admin queue.
1038 if (!nvmeq->qid || iod->aborted) {
1039 dev_warn(dev->ctrl.device,
1040 "I/O %d QID %d timeout, reset controller\n",
1041 req->tag, nvmeq->qid);
1042 nvme_dev_disable(dev, false);
1043 nvme_reset_ctrl(&dev->ctrl);
1046 * Mark the request as handled, since the inline shutdown
1047 * forces all outstanding requests to complete.
1049 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1050 return BLK_EH_HANDLED;
1053 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1054 atomic_inc(&dev->ctrl.abort_limit);
1055 return BLK_EH_RESET_TIMER;
1059 memset(&cmd, 0, sizeof(cmd));
1060 cmd.abort.opcode = nvme_admin_abort_cmd;
1061 cmd.abort.cid = req->tag;
1062 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1064 dev_warn(nvmeq->dev->ctrl.device,
1065 "I/O %d QID %d timeout, aborting\n",
1066 req->tag, nvmeq->qid);
1068 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1069 BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1070 if (IS_ERR(abort_req)) {
1071 atomic_inc(&dev->ctrl.abort_limit);
1072 return BLK_EH_RESET_TIMER;
1075 abort_req->timeout = ADMIN_TIMEOUT;
1076 abort_req->end_io_data = NULL;
1077 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1080 * The aborted req will be completed on receiving the abort req.
1081 * We enable the timer again. If hit twice, it'll cause a device reset,
1082 * as the device then is in a faulty state.
1084 return BLK_EH_RESET_TIMER;
1087 static void nvme_free_queue(struct nvme_queue *nvmeq)
1089 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1090 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1092 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1093 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1097 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1101 for (i = dev->queue_count - 1; i >= lowest; i--) {
1102 struct nvme_queue *nvmeq = dev->queues[i];
1104 dev->queues[i] = NULL;
1105 nvme_free_queue(nvmeq);
1110 * nvme_suspend_queue - put queue into suspended state
1111 * @nvmeq - queue to suspend
1113 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1117 spin_lock_irq(&nvmeq->q_lock);
1118 if (nvmeq->cq_vector == -1) {
1119 spin_unlock_irq(&nvmeq->q_lock);
1122 vector = nvmeq->cq_vector;
1123 nvmeq->dev->online_queues--;
1124 nvmeq->cq_vector = -1;
1125 spin_unlock_irq(&nvmeq->q_lock);
1127 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1128 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1130 pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1135 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1137 struct nvme_queue *nvmeq = dev->queues[0];
1141 if (nvme_suspend_queue(nvmeq))
1145 nvme_shutdown_ctrl(&dev->ctrl);
1147 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1148 dev->bar + NVME_REG_CAP));
1150 spin_lock_irq(&nvmeq->q_lock);
1151 nvme_process_cq(nvmeq);
1152 spin_unlock_irq(&nvmeq->q_lock);
1155 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1158 int q_depth = dev->q_depth;
1159 unsigned q_size_aligned = roundup(q_depth * entry_size,
1160 dev->ctrl.page_size);
1162 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1163 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1164 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1165 q_depth = div_u64(mem_per_q, entry_size);
1168 * Ensure the reduced q_depth is above some threshold where it
1169 * would be better to map queues in system memory with the
1179 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1182 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1183 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1184 dev->ctrl.page_size);
1185 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1186 nvmeq->sq_cmds_io = dev->cmb + offset;
1188 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1189 &nvmeq->sq_dma_addr, GFP_KERNEL);
1190 if (!nvmeq->sq_cmds)
1197 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1198 int depth, int node)
1200 struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
1205 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1206 &nvmeq->cq_dma_addr, GFP_KERNEL);
1210 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1213 nvmeq->q_dmadev = dev->dev;
1215 spin_lock_init(&nvmeq->q_lock);
1217 nvmeq->cq_phase = 1;
1218 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1219 nvmeq->q_depth = depth;
1221 nvmeq->cq_vector = -1;
1222 dev->queues[qid] = nvmeq;
1228 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1229 nvmeq->cq_dma_addr);
1235 static int queue_request_irq(struct nvme_queue *nvmeq)
1237 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1238 int nr = nvmeq->dev->ctrl.instance;
1240 if (use_threaded_interrupts) {
1241 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1242 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1244 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1245 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1249 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1251 struct nvme_dev *dev = nvmeq->dev;
1253 spin_lock_irq(&nvmeq->q_lock);
1256 nvmeq->cq_phase = 1;
1257 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1258 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1259 nvme_dbbuf_init(dev, nvmeq, qid);
1260 dev->online_queues++;
1261 spin_unlock_irq(&nvmeq->q_lock);
1264 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1266 struct nvme_dev *dev = nvmeq->dev;
1269 nvmeq->cq_vector = qid - 1;
1270 result = adapter_alloc_cq(dev, qid, nvmeq);
1274 result = adapter_alloc_sq(dev, qid, nvmeq);
1278 result = queue_request_irq(nvmeq);
1282 nvme_init_queue(nvmeq, qid);
1286 adapter_delete_sq(dev, qid);
1288 adapter_delete_cq(dev, qid);
1292 static const struct blk_mq_ops nvme_mq_admin_ops = {
1293 .queue_rq = nvme_queue_rq,
1294 .complete = nvme_pci_complete_rq,
1295 .init_hctx = nvme_admin_init_hctx,
1296 .exit_hctx = nvme_admin_exit_hctx,
1297 .init_request = nvme_init_request,
1298 .timeout = nvme_timeout,
1301 static const struct blk_mq_ops nvme_mq_ops = {
1302 .queue_rq = nvme_queue_rq,
1303 .complete = nvme_pci_complete_rq,
1304 .init_hctx = nvme_init_hctx,
1305 .init_request = nvme_init_request,
1306 .map_queues = nvme_pci_map_queues,
1307 .timeout = nvme_timeout,
1311 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1313 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1315 * If the controller was reset during removal, it's possible
1316 * user requests may be waiting on a stopped queue. Start the
1317 * queue to flush these to completion.
1319 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1320 blk_cleanup_queue(dev->ctrl.admin_q);
1321 blk_mq_free_tag_set(&dev->admin_tagset);
1325 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1327 if (!dev->ctrl.admin_q) {
1328 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1329 dev->admin_tagset.nr_hw_queues = 1;
1332 * Subtract one to leave an empty queue entry for 'Full Queue'
1333 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1335 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1336 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1337 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1338 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1339 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1340 dev->admin_tagset.driver_data = dev;
1342 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1345 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1346 if (IS_ERR(dev->ctrl.admin_q)) {
1347 blk_mq_free_tag_set(&dev->admin_tagset);
1350 if (!blk_get_queue(dev->ctrl.admin_q)) {
1351 nvme_dev_remove_admin(dev);
1352 dev->ctrl.admin_q = NULL;
1356 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1361 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1363 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1366 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1368 struct pci_dev *pdev = to_pci_dev(dev->dev);
1370 if (size <= dev->bar_mapped_size)
1372 if (size > pci_resource_len(pdev, 0))
1376 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1378 dev->bar_mapped_size = 0;
1381 dev->bar_mapped_size = size;
1382 dev->dbs = dev->bar + NVME_REG_DBS;
1387 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1391 u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1392 struct nvme_queue *nvmeq;
1394 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1398 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1399 NVME_CAP_NSSRC(cap) : 0;
1401 if (dev->subsystem &&
1402 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1403 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1405 result = nvme_disable_ctrl(&dev->ctrl, cap);
1409 nvmeq = dev->queues[0];
1411 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
1412 dev_to_node(dev->dev));
1417 aqa = nvmeq->q_depth - 1;
1420 writel(aqa, dev->bar + NVME_REG_AQA);
1421 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1422 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1424 result = nvme_enable_ctrl(&dev->ctrl, cap);
1428 nvmeq->cq_vector = 0;
1429 result = queue_request_irq(nvmeq);
1431 nvmeq->cq_vector = -1;
1438 static int nvme_create_io_queues(struct nvme_dev *dev)
1443 for (i = dev->queue_count; i <= dev->max_qid; i++) {
1444 /* vector == qid - 1, match nvme_create_queue */
1445 if (!nvme_alloc_queue(dev, i, dev->q_depth,
1446 pci_irq_get_node(to_pci_dev(dev->dev), i - 1))) {
1452 max = min(dev->max_qid, dev->queue_count - 1);
1453 for (i = dev->online_queues; i <= max; i++) {
1454 ret = nvme_create_queue(dev->queues[i], i);
1460 * Ignore failing Create SQ/CQ commands, we can continue with less
1461 * than the desired aount of queues, and even a controller without
1462 * I/O queues an still be used to issue admin commands. This might
1463 * be useful to upgrade a buggy firmware for example.
1465 return ret >= 0 ? 0 : ret;
1468 static ssize_t nvme_cmb_show(struct device *dev,
1469 struct device_attribute *attr,
1472 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1474 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1475 ndev->cmbloc, ndev->cmbsz);
1477 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1479 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1481 u64 szu, size, offset;
1482 resource_size_t bar_size;
1483 struct pci_dev *pdev = to_pci_dev(dev->dev);
1485 dma_addr_t dma_addr;
1487 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1488 if (!(NVME_CMB_SZ(dev->cmbsz)))
1490 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1495 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1496 size = szu * NVME_CMB_SZ(dev->cmbsz);
1497 offset = szu * NVME_CMB_OFST(dev->cmbloc);
1498 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1500 if (offset > bar_size)
1504 * Controllers may support a CMB size larger than their BAR,
1505 * for example, due to being behind a bridge. Reduce the CMB to
1506 * the reported size of the BAR
1508 if (size > bar_size - offset)
1509 size = bar_size - offset;
1511 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1512 cmb = ioremap_wc(dma_addr, size);
1516 dev->cmb_dma_addr = dma_addr;
1517 dev->cmb_size = size;
1521 static inline void nvme_release_cmb(struct nvme_dev *dev)
1527 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1528 &dev_attr_cmb.attr, NULL);
1534 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1536 size_t len = dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs);
1537 struct nvme_command c;
1541 dma_addr = dma_map_single(dev->dev, dev->host_mem_descs, len,
1543 if (dma_mapping_error(dev->dev, dma_addr))
1546 memset(&c, 0, sizeof(c));
1547 c.features.opcode = nvme_admin_set_features;
1548 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1549 c.features.dword11 = cpu_to_le32(bits);
1550 c.features.dword12 = cpu_to_le32(dev->host_mem_size >>
1551 ilog2(dev->ctrl.page_size));
1552 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1553 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1554 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1556 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1558 dev_warn(dev->ctrl.device,
1559 "failed to set host mem (err %d, flags %#x).\n",
1562 dma_unmap_single(dev->dev, dma_addr, len, DMA_TO_DEVICE);
1566 static void nvme_free_host_mem(struct nvme_dev *dev)
1570 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1571 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1572 size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1574 dma_free_coherent(dev->dev, size, dev->host_mem_desc_bufs[i],
1575 le64_to_cpu(desc->addr));
1578 kfree(dev->host_mem_desc_bufs);
1579 dev->host_mem_desc_bufs = NULL;
1580 kfree(dev->host_mem_descs);
1581 dev->host_mem_descs = NULL;
1584 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1586 struct nvme_host_mem_buf_desc *descs;
1587 u32 chunk_size, max_entries, i = 0;
1591 /* start big and work our way down */
1592 chunk_size = min(preferred, (u64)PAGE_SIZE << MAX_ORDER);
1594 tmp = (preferred + chunk_size - 1);
1595 do_div(tmp, chunk_size);
1597 descs = kcalloc(max_entries, sizeof(*descs), GFP_KERNEL);
1601 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1603 goto out_free_descs;
1605 for (size = 0; size < preferred; size += chunk_size) {
1606 u32 len = min_t(u64, chunk_size, preferred - size);
1607 dma_addr_t dma_addr;
1609 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1610 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1614 descs[i].addr = cpu_to_le64(dma_addr);
1615 descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1619 if (!size || (min && size < min)) {
1620 dev_warn(dev->ctrl.device,
1621 "failed to allocate host memory buffer.\n");
1625 dev_info(dev->ctrl.device,
1626 "allocated %lld MiB host memory buffer.\n",
1627 size >> ilog2(SZ_1M));
1628 dev->nr_host_mem_descs = i;
1629 dev->host_mem_size = size;
1630 dev->host_mem_descs = descs;
1631 dev->host_mem_desc_bufs = bufs;
1636 size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1638 dma_free_coherent(dev->dev, size, bufs[i],
1639 le64_to_cpu(descs[i].addr));
1646 /* try a smaller chunk size if we failed early */
1647 if (chunk_size >= PAGE_SIZE * 2 && (i == 0 || size < min)) {
1651 dev->host_mem_descs = NULL;
1655 static void nvme_setup_host_mem(struct nvme_dev *dev)
1657 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1658 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1659 u64 min = (u64)dev->ctrl.hmmin * 4096;
1660 u32 enable_bits = NVME_HOST_MEM_ENABLE;
1662 preferred = min(preferred, max);
1664 dev_warn(dev->ctrl.device,
1665 "min host memory (%lld MiB) above limit (%d MiB).\n",
1666 min >> ilog2(SZ_1M), max_host_mem_size_mb);
1667 nvme_free_host_mem(dev);
1672 * If we already have a buffer allocated check if we can reuse it.
1674 if (dev->host_mem_descs) {
1675 if (dev->host_mem_size >= min)
1676 enable_bits |= NVME_HOST_MEM_RETURN;
1678 nvme_free_host_mem(dev);
1681 if (!dev->host_mem_descs) {
1682 if (nvme_alloc_host_mem(dev, min, preferred))
1686 if (nvme_set_host_mem(dev, enable_bits))
1687 nvme_free_host_mem(dev);
1690 static int nvme_setup_io_queues(struct nvme_dev *dev)
1692 struct nvme_queue *adminq = dev->queues[0];
1693 struct pci_dev *pdev = to_pci_dev(dev->dev);
1694 int result, nr_io_queues;
1697 nr_io_queues = num_online_cpus();
1698 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1702 if (nr_io_queues == 0)
1705 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1706 result = nvme_cmb_qdepth(dev, nr_io_queues,
1707 sizeof(struct nvme_command));
1709 dev->q_depth = result;
1711 nvme_release_cmb(dev);
1715 size = db_bar_size(dev, nr_io_queues);
1716 result = nvme_remap_bar(dev, size);
1719 if (!--nr_io_queues)
1722 adminq->q_db = dev->dbs;
1724 /* Deregister the admin queue's interrupt */
1725 pci_free_irq(pdev, 0, adminq);
1728 * If we enable msix early due to not intx, disable it again before
1729 * setting up the full range we need.
1731 pci_free_irq_vectors(pdev);
1732 nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1733 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1734 if (nr_io_queues <= 0)
1736 dev->max_qid = nr_io_queues;
1739 * Should investigate if there's a performance win from allocating
1740 * more queues than interrupt vectors; it might allow the submission
1741 * path to scale better, even if the receive path is limited by the
1742 * number of interrupts.
1745 result = queue_request_irq(adminq);
1747 adminq->cq_vector = -1;
1750 return nvme_create_io_queues(dev);
1753 static void nvme_del_queue_end(struct request *req, blk_status_t error)
1755 struct nvme_queue *nvmeq = req->end_io_data;
1757 blk_mq_free_request(req);
1758 complete(&nvmeq->dev->ioq_wait);
1761 static void nvme_del_cq_end(struct request *req, blk_status_t error)
1763 struct nvme_queue *nvmeq = req->end_io_data;
1766 unsigned long flags;
1769 * We might be called with the AQ q_lock held
1770 * and the I/O queue q_lock should always
1771 * nest inside the AQ one.
1773 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1774 SINGLE_DEPTH_NESTING);
1775 nvme_process_cq(nvmeq);
1776 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1779 nvme_del_queue_end(req, error);
1782 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1784 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1785 struct request *req;
1786 struct nvme_command cmd;
1788 memset(&cmd, 0, sizeof(cmd));
1789 cmd.delete_queue.opcode = opcode;
1790 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1792 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1794 return PTR_ERR(req);
1796 req->timeout = ADMIN_TIMEOUT;
1797 req->end_io_data = nvmeq;
1799 blk_execute_rq_nowait(q, NULL, req, false,
1800 opcode == nvme_admin_delete_cq ?
1801 nvme_del_cq_end : nvme_del_queue_end);
1805 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1808 unsigned long timeout;
1809 u8 opcode = nvme_admin_delete_sq;
1811 for (pass = 0; pass < 2; pass++) {
1812 int sent = 0, i = queues;
1814 reinit_completion(&dev->ioq_wait);
1816 timeout = ADMIN_TIMEOUT;
1817 for (; i > 0; i--, sent++)
1818 if (nvme_delete_queue(dev->queues[i], opcode))
1822 timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1828 opcode = nvme_admin_delete_cq;
1833 * Return: error value if an error occurred setting up the queues or calling
1834 * Identify Device. 0 if these succeeded, even if adding some of the
1835 * namespaces failed. At the moment, these failures are silent. TBD which
1836 * failures should be reported.
1838 static int nvme_dev_add(struct nvme_dev *dev)
1840 if (!dev->ctrl.tagset) {
1841 dev->tagset.ops = &nvme_mq_ops;
1842 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1843 dev->tagset.timeout = NVME_IO_TIMEOUT;
1844 dev->tagset.numa_node = dev_to_node(dev->dev);
1845 dev->tagset.queue_depth =
1846 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1847 dev->tagset.cmd_size = nvme_cmd_size(dev);
1848 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1849 dev->tagset.driver_data = dev;
1851 if (blk_mq_alloc_tag_set(&dev->tagset))
1853 dev->ctrl.tagset = &dev->tagset;
1855 nvme_dbbuf_set(dev);
1857 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1859 /* Free previously allocated queues that are no longer usable */
1860 nvme_free_queues(dev, dev->online_queues);
1866 static int nvme_pci_enable(struct nvme_dev *dev)
1869 int result = -ENOMEM;
1870 struct pci_dev *pdev = to_pci_dev(dev->dev);
1872 if (pci_enable_device_mem(pdev))
1875 pci_set_master(pdev);
1877 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1878 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1881 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1887 * Some devices and/or platforms don't advertise or work with INTx
1888 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1889 * adjust this later.
1891 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1895 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1897 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1898 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1899 dev->dbs = dev->bar + 4096;
1902 * Temporary fix for the Apple controller found in the MacBook8,1 and
1903 * some MacBook7,1 to avoid controller resets and data loss.
1905 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1907 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
1908 "set queue depth=%u to work around controller resets\n",
1913 * CMBs can currently only exist on >=1.2 PCIe devices. We only
1914 * populate sysfs if a CMB is implemented. Note that we add the
1915 * CMB attribute to the nvme_ctrl kobj which removes the need to remove
1916 * it on exit. Since nvme_dev_attrs_group has no name we can pass
1917 * NULL as final argument to sysfs_add_file_to_group.
1920 if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1921 dev->cmb = nvme_map_cmb(dev);
1924 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1925 &dev_attr_cmb.attr, NULL))
1926 dev_warn(dev->ctrl.device,
1927 "failed to add sysfs attribute for CMB\n");
1931 pci_enable_pcie_error_reporting(pdev);
1932 pci_save_state(pdev);
1936 pci_disable_device(pdev);
1940 static void nvme_dev_unmap(struct nvme_dev *dev)
1944 pci_release_mem_regions(to_pci_dev(dev->dev));
1947 static void nvme_pci_disable(struct nvme_dev *dev)
1949 struct pci_dev *pdev = to_pci_dev(dev->dev);
1951 nvme_release_cmb(dev);
1952 pci_free_irq_vectors(pdev);
1954 if (pci_is_enabled(pdev)) {
1955 pci_disable_pcie_error_reporting(pdev);
1956 pci_disable_device(pdev);
1960 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1964 struct pci_dev *pdev = to_pci_dev(dev->dev);
1966 mutex_lock(&dev->shutdown_lock);
1967 if (pci_is_enabled(pdev)) {
1968 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1970 if (dev->ctrl.state == NVME_CTRL_LIVE)
1971 nvme_start_freeze(&dev->ctrl);
1972 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
1973 pdev->error_state != pci_channel_io_normal);
1977 * Give the controller a chance to complete all entered requests if
1978 * doing a safe shutdown.
1982 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
1985 * If the controller is still alive tell it to stop using the
1986 * host memory buffer. In theory the shutdown / reset should
1987 * make sure that it doesn't access the host memoery anymore,
1988 * but I'd rather be safe than sorry..
1990 if (dev->host_mem_descs)
1991 nvme_set_host_mem(dev, 0);
1994 nvme_stop_queues(&dev->ctrl);
1996 queues = dev->online_queues - 1;
1997 for (i = dev->queue_count - 1; i > 0; i--)
1998 nvme_suspend_queue(dev->queues[i]);
2001 /* A device might become IO incapable very soon during
2002 * probe, before the admin queue is configured. Thus,
2003 * queue_count can be 0 here.
2005 if (dev->queue_count)
2006 nvme_suspend_queue(dev->queues[0]);
2008 nvme_disable_io_queues(dev, queues);
2009 nvme_disable_admin_queue(dev, shutdown);
2011 nvme_pci_disable(dev);
2013 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2014 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2017 * The driver will not be starting up queues again if shutting down so
2018 * must flush all entered requests to their failed completion to avoid
2019 * deadlocking blk-mq hot-cpu notifier.
2022 nvme_start_queues(&dev->ctrl);
2023 mutex_unlock(&dev->shutdown_lock);
2026 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2028 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2029 PAGE_SIZE, PAGE_SIZE, 0);
2030 if (!dev->prp_page_pool)
2033 /* Optimisation for I/Os between 4k and 128k */
2034 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2036 if (!dev->prp_small_pool) {
2037 dma_pool_destroy(dev->prp_page_pool);
2043 static void nvme_release_prp_pools(struct nvme_dev *dev)
2045 dma_pool_destroy(dev->prp_page_pool);
2046 dma_pool_destroy(dev->prp_small_pool);
2049 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2051 struct nvme_dev *dev = to_nvme_dev(ctrl);
2053 nvme_dbbuf_dma_free(dev);
2054 put_device(dev->dev);
2055 if (dev->tagset.tags)
2056 blk_mq_free_tag_set(&dev->tagset);
2057 if (dev->ctrl.admin_q)
2058 blk_put_queue(dev->ctrl.admin_q);
2060 free_opal_dev(dev->ctrl.opal_dev);
2064 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
2066 dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
2068 kref_get(&dev->ctrl.kref);
2069 nvme_dev_disable(dev, false);
2070 if (!schedule_work(&dev->remove_work))
2071 nvme_put_ctrl(&dev->ctrl);
2074 static void nvme_reset_work(struct work_struct *work)
2076 struct nvme_dev *dev =
2077 container_of(work, struct nvme_dev, ctrl.reset_work);
2078 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2079 int result = -ENODEV;
2081 if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
2085 * If we're called to reset a live controller first shut it down before
2088 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2089 nvme_dev_disable(dev, false);
2091 result = nvme_pci_enable(dev);
2095 result = nvme_configure_admin_queue(dev);
2099 nvme_init_queue(dev->queues[0], 0);
2100 result = nvme_alloc_admin_tags(dev);
2104 result = nvme_init_identify(&dev->ctrl);
2108 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2109 if (!dev->ctrl.opal_dev)
2110 dev->ctrl.opal_dev =
2111 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2112 else if (was_suspend)
2113 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2115 free_opal_dev(dev->ctrl.opal_dev);
2116 dev->ctrl.opal_dev = NULL;
2119 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2120 result = nvme_dbbuf_dma_alloc(dev);
2123 "unable to allocate dma for dbbuf\n");
2126 if (dev->ctrl.hmpre)
2127 nvme_setup_host_mem(dev);
2129 result = nvme_setup_io_queues(dev);
2134 * A controller that can not execute IO typically requires user
2135 * intervention to correct. For such degraded controllers, the driver
2136 * should not submit commands the user did not request, so skip
2137 * registering for asynchronous event notification on this condition.
2139 if (dev->online_queues > 1)
2140 nvme_queue_async_events(&dev->ctrl);
2143 * Keep the controller around but remove all namespaces if we don't have
2144 * any working I/O queue.
2146 if (dev->online_queues < 2) {
2147 dev_warn(dev->ctrl.device, "IO queues not created\n");
2148 nvme_kill_queues(&dev->ctrl);
2149 nvme_remove_namespaces(&dev->ctrl);
2151 nvme_start_queues(&dev->ctrl);
2152 nvme_wait_freeze(&dev->ctrl);
2154 nvme_unfreeze(&dev->ctrl);
2157 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2158 dev_warn(dev->ctrl.device, "failed to mark controller live\n");
2162 if (dev->online_queues > 1)
2163 nvme_queue_scan(&dev->ctrl);
2167 nvme_remove_dead_ctrl(dev, result);
2170 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2172 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2173 struct pci_dev *pdev = to_pci_dev(dev->dev);
2175 nvme_kill_queues(&dev->ctrl);
2176 if (pci_get_drvdata(pdev))
2177 device_release_driver(&pdev->dev);
2178 nvme_put_ctrl(&dev->ctrl);
2181 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2183 *val = readl(to_nvme_dev(ctrl)->bar + off);
2187 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2189 writel(val, to_nvme_dev(ctrl)->bar + off);
2193 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2195 *val = readq(to_nvme_dev(ctrl)->bar + off);
2199 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2201 .module = THIS_MODULE,
2202 .flags = NVME_F_METADATA_SUPPORTED,
2203 .reg_read32 = nvme_pci_reg_read32,
2204 .reg_write32 = nvme_pci_reg_write32,
2205 .reg_read64 = nvme_pci_reg_read64,
2206 .free_ctrl = nvme_pci_free_ctrl,
2207 .submit_async_event = nvme_pci_submit_async_event,
2210 static int nvme_dev_map(struct nvme_dev *dev)
2212 struct pci_dev *pdev = to_pci_dev(dev->dev);
2214 if (pci_request_mem_regions(pdev, "nvme"))
2217 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2222 pci_release_mem_regions(pdev);
2226 static unsigned long check_dell_samsung_bug(struct pci_dev *pdev)
2228 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2230 * Several Samsung devices seem to drop off the PCIe bus
2231 * randomly when APST is on and uses the deepest sleep state.
2232 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2233 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2234 * 950 PRO 256GB", but it seems to be restricted to two Dell
2237 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2238 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2239 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2240 return NVME_QUIRK_NO_DEEPEST_PS;
2246 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2248 int node, result = -ENOMEM;
2249 struct nvme_dev *dev;
2250 unsigned long quirks = id->driver_data;
2252 node = dev_to_node(&pdev->dev);
2253 if (node == NUMA_NO_NODE)
2254 set_dev_node(&pdev->dev, first_memory_node);
2256 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2259 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2264 dev->dev = get_device(&pdev->dev);
2265 pci_set_drvdata(pdev, dev);
2267 result = nvme_dev_map(dev);
2271 INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2272 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2273 mutex_init(&dev->shutdown_lock);
2274 init_completion(&dev->ioq_wait);
2276 result = nvme_setup_prp_pools(dev);
2280 quirks |= check_dell_samsung_bug(pdev);
2282 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2287 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING);
2288 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2290 queue_work(nvme_wq, &dev->ctrl.reset_work);
2294 nvme_release_prp_pools(dev);
2296 put_device(dev->dev);
2297 nvme_dev_unmap(dev);
2304 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2306 struct nvme_dev *dev = pci_get_drvdata(pdev);
2309 nvme_dev_disable(dev, false);
2311 nvme_reset_ctrl(&dev->ctrl);
2314 static void nvme_shutdown(struct pci_dev *pdev)
2316 struct nvme_dev *dev = pci_get_drvdata(pdev);
2317 nvme_dev_disable(dev, true);
2321 * The driver's remove may be called on a device in a partially initialized
2322 * state. This function must not have any dependencies on the device state in
2325 static void nvme_remove(struct pci_dev *pdev)
2327 struct nvme_dev *dev = pci_get_drvdata(pdev);
2329 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2331 cancel_work_sync(&dev->ctrl.reset_work);
2332 pci_set_drvdata(pdev, NULL);
2334 if (!pci_device_is_present(pdev)) {
2335 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2336 nvme_dev_disable(dev, false);
2339 flush_work(&dev->ctrl.reset_work);
2340 nvme_uninit_ctrl(&dev->ctrl);
2341 nvme_dev_disable(dev, true);
2342 nvme_free_host_mem(dev);
2343 nvme_dev_remove_admin(dev);
2344 nvme_free_queues(dev, 0);
2345 nvme_release_prp_pools(dev);
2346 nvme_dev_unmap(dev);
2347 nvme_put_ctrl(&dev->ctrl);
2350 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2355 if (pci_vfs_assigned(pdev)) {
2356 dev_warn(&pdev->dev,
2357 "Cannot disable SR-IOV VFs while assigned\n");
2360 pci_disable_sriov(pdev);
2364 ret = pci_enable_sriov(pdev, numvfs);
2365 return ret ? ret : numvfs;
2368 #ifdef CONFIG_PM_SLEEP
2369 static int nvme_suspend(struct device *dev)
2371 struct pci_dev *pdev = to_pci_dev(dev);
2372 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2374 nvme_dev_disable(ndev, true);
2378 static int nvme_resume(struct device *dev)
2380 struct pci_dev *pdev = to_pci_dev(dev);
2381 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2383 nvme_reset_ctrl(&ndev->ctrl);
2388 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2390 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2391 pci_channel_state_t state)
2393 struct nvme_dev *dev = pci_get_drvdata(pdev);
2396 * A frozen channel requires a reset. When detected, this method will
2397 * shutdown the controller to quiesce. The controller will be restarted
2398 * after the slot reset through driver's slot_reset callback.
2401 case pci_channel_io_normal:
2402 return PCI_ERS_RESULT_CAN_RECOVER;
2403 case pci_channel_io_frozen:
2404 dev_warn(dev->ctrl.device,
2405 "frozen state error detected, reset controller\n");
2406 nvme_dev_disable(dev, false);
2407 return PCI_ERS_RESULT_NEED_RESET;
2408 case pci_channel_io_perm_failure:
2409 dev_warn(dev->ctrl.device,
2410 "failure state error detected, request disconnect\n");
2411 return PCI_ERS_RESULT_DISCONNECT;
2413 return PCI_ERS_RESULT_NEED_RESET;
2416 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2418 struct nvme_dev *dev = pci_get_drvdata(pdev);
2420 dev_info(dev->ctrl.device, "restart after slot reset\n");
2421 pci_restore_state(pdev);
2422 nvme_reset_ctrl(&dev->ctrl);
2423 return PCI_ERS_RESULT_RECOVERED;
2426 static void nvme_error_resume(struct pci_dev *pdev)
2428 pci_cleanup_aer_uncorrect_error_status(pdev);
2431 static const struct pci_error_handlers nvme_err_handler = {
2432 .error_detected = nvme_error_detected,
2433 .slot_reset = nvme_slot_reset,
2434 .resume = nvme_error_resume,
2435 .reset_notify = nvme_reset_notify,
2438 static const struct pci_device_id nvme_id_table[] = {
2439 { PCI_VDEVICE(INTEL, 0x0953),
2440 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2441 NVME_QUIRK_DEALLOCATE_ZEROES, },
2442 { PCI_VDEVICE(INTEL, 0x0a53),
2443 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2444 NVME_QUIRK_DEALLOCATE_ZEROES, },
2445 { PCI_VDEVICE(INTEL, 0x0a54),
2446 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2447 NVME_QUIRK_DEALLOCATE_ZEROES, },
2448 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
2449 .driver_data = NVME_QUIRK_NO_DEEPEST_PS },
2450 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
2451 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2452 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
2453 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2454 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
2455 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2456 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2457 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2458 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2461 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2463 static struct pci_driver nvme_driver = {
2465 .id_table = nvme_id_table,
2466 .probe = nvme_probe,
2467 .remove = nvme_remove,
2468 .shutdown = nvme_shutdown,
2470 .pm = &nvme_dev_pm_ops,
2472 .sriov_configure = nvme_pci_sriov_configure,
2473 .err_handler = &nvme_err_handler,
2476 static int __init nvme_init(void)
2478 return pci_register_driver(&nvme_driver);
2481 static void __exit nvme_exit(void)
2483 pci_unregister_driver(&nvme_driver);
2487 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2488 MODULE_LICENSE("GPL");
2489 MODULE_VERSION("1.0");
2490 module_init(nvme_init);
2491 module_exit(nvme_exit);