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/cpu.h>
21 #include <linux/delay.h>
22 #include <linux/dmi.h>
23 #include <linux/errno.h>
25 #include <linux/genhd.h>
26 #include <linux/hdreg.h>
27 #include <linux/idr.h>
28 #include <linux/init.h>
29 #include <linux/interrupt.h>
31 #include <linux/kdev_t.h>
32 #include <linux/kernel.h>
34 #include <linux/module.h>
35 #include <linux/moduleparam.h>
36 #include <linux/mutex.h>
37 #include <linux/pci.h>
38 #include <linux/poison.h>
39 #include <linux/ptrace.h>
40 #include <linux/sched.h>
41 #include <linux/slab.h>
42 #include <linux/t10-pi.h>
43 #include <linux/timer.h>
44 #include <linux/types.h>
45 #include <linux/io-64-nonatomic-lo-hi.h>
46 #include <asm/unaligned.h>
47 #include <linux/sed-opal.h>
51 #define NVME_Q_DEPTH 1024
52 #define NVME_AQ_DEPTH 256
53 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
54 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
57 * We handle AEN commands ourselves and don't even let the
58 * block layer know about them.
60 #define NVME_AQ_BLKMQ_DEPTH (NVME_AQ_DEPTH - NVME_NR_AERS)
62 static int use_threaded_interrupts;
63 module_param(use_threaded_interrupts, int, 0);
65 static bool use_cmb_sqes = true;
66 module_param(use_cmb_sqes, bool, 0644);
67 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
69 static unsigned int max_host_mem_size_mb = 128;
70 module_param(max_host_mem_size_mb, uint, 0444);
71 MODULE_PARM_DESC(max_host_mem_size_mb,
72 "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
77 static int nvme_reset(struct nvme_dev *dev);
78 static void nvme_process_cq(struct nvme_queue *nvmeq);
79 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
82 * Represents an NVM Express device. Each nvme_dev is a PCI function.
85 struct nvme_queue **queues;
86 struct blk_mq_tag_set tagset;
87 struct blk_mq_tag_set admin_tagset;
90 struct dma_pool *prp_page_pool;
91 struct dma_pool *prp_small_pool;
93 unsigned online_queues;
98 unsigned long bar_mapped_size;
99 struct work_struct reset_work;
100 struct work_struct remove_work;
101 struct timer_list watchdog_timer;
102 struct mutex shutdown_lock;
105 dma_addr_t cmb_dma_addr;
109 struct nvme_ctrl ctrl;
110 struct completion ioq_wait;
112 /* shadow doorbell buffer support: */
114 dma_addr_t dbbuf_dbs_dma_addr;
116 dma_addr_t dbbuf_eis_dma_addr;
118 /* host memory buffer support: */
120 u32 nr_host_mem_descs;
121 struct nvme_host_mem_buf_desc *host_mem_descs;
122 void **host_mem_desc_bufs;
125 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
127 return qid * 2 * stride;
130 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
132 return (qid * 2 + 1) * stride;
135 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
137 return container_of(ctrl, struct nvme_dev, ctrl);
141 * An NVM Express queue. Each device has at least two (one for admin
142 * commands and one for I/O commands).
145 struct device *q_dmadev;
146 struct nvme_dev *dev;
148 struct nvme_command *sq_cmds;
149 struct nvme_command __iomem *sq_cmds_io;
150 volatile struct nvme_completion *cqes;
151 struct blk_mq_tags **tags;
152 dma_addr_t sq_dma_addr;
153 dma_addr_t cq_dma_addr;
169 * The nvme_iod describes the data in an I/O, including the list of PRP
170 * entries. You can't see it in this data structure because C doesn't let
171 * me express that. Use nvme_init_iod to ensure there's enough space
172 * allocated to store the PRP list.
175 struct nvme_request req;
176 struct nvme_queue *nvmeq;
178 int npages; /* In the PRP list. 0 means small pool in use */
179 int nents; /* Used in scatterlist */
180 int length; /* Of data, in bytes */
181 dma_addr_t first_dma;
182 struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
183 struct scatterlist *sg;
184 struct scatterlist inline_sg[0];
188 * Check we didin't inadvertently grow the command struct
190 static inline void _nvme_check_size(void)
192 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
193 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
194 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
195 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
196 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
197 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
198 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
199 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
200 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
201 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
202 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
203 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
204 BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
207 static inline unsigned int nvme_dbbuf_size(u32 stride)
209 return ((num_possible_cpus() + 1) * 8 * stride);
212 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
214 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
219 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
220 &dev->dbbuf_dbs_dma_addr,
224 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
225 &dev->dbbuf_eis_dma_addr,
227 if (!dev->dbbuf_eis) {
228 dma_free_coherent(dev->dev, mem_size,
229 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
230 dev->dbbuf_dbs = NULL;
237 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
239 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
241 if (dev->dbbuf_dbs) {
242 dma_free_coherent(dev->dev, mem_size,
243 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
244 dev->dbbuf_dbs = NULL;
246 if (dev->dbbuf_eis) {
247 dma_free_coherent(dev->dev, mem_size,
248 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
249 dev->dbbuf_eis = NULL;
253 static void nvme_dbbuf_init(struct nvme_dev *dev,
254 struct nvme_queue *nvmeq, int qid)
256 if (!dev->dbbuf_dbs || !qid)
259 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
260 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
261 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
262 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
265 static void nvme_dbbuf_set(struct nvme_dev *dev)
267 struct nvme_command c;
272 memset(&c, 0, sizeof(c));
273 c.dbbuf.opcode = nvme_admin_dbbuf;
274 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
275 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
277 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
278 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
279 /* Free memory and continue on */
280 nvme_dbbuf_dma_free(dev);
284 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
286 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
289 /* Update dbbuf and return true if an MMIO is required */
290 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
291 volatile u32 *dbbuf_ei)
297 * Ensure that the queue is written before updating
298 * the doorbell in memory
302 old_value = *dbbuf_db;
305 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
313 * Max size of iod being embedded in the request payload
315 #define NVME_INT_PAGES 2
316 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size)
319 * Will slightly overestimate the number of pages needed. This is OK
320 * as it only leads to a small amount of wasted memory for the lifetime of
323 static int nvme_npages(unsigned size, struct nvme_dev *dev)
325 unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
326 dev->ctrl.page_size);
327 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
330 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
331 unsigned int size, unsigned int nseg)
333 return sizeof(__le64 *) * nvme_npages(size, dev) +
334 sizeof(struct scatterlist) * nseg;
337 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
339 return sizeof(struct nvme_iod) +
340 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
343 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
344 unsigned int hctx_idx)
346 struct nvme_dev *dev = data;
347 struct nvme_queue *nvmeq = dev->queues[0];
349 WARN_ON(hctx_idx != 0);
350 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
351 WARN_ON(nvmeq->tags);
353 hctx->driver_data = nvmeq;
354 nvmeq->tags = &dev->admin_tagset.tags[0];
358 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
360 struct nvme_queue *nvmeq = hctx->driver_data;
365 static int nvme_admin_init_request(struct blk_mq_tag_set *set,
366 struct request *req, unsigned int hctx_idx,
367 unsigned int numa_node)
369 struct nvme_dev *dev = set->driver_data;
370 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
371 struct nvme_queue *nvmeq = dev->queues[0];
378 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
379 unsigned int hctx_idx)
381 struct nvme_dev *dev = data;
382 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
385 nvmeq->tags = &dev->tagset.tags[hctx_idx];
387 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
388 hctx->driver_data = nvmeq;
392 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
393 unsigned int hctx_idx, unsigned int numa_node)
395 struct nvme_dev *dev = set->driver_data;
396 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
397 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
404 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
406 struct nvme_dev *dev = set->driver_data;
408 return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
412 * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
413 * @nvmeq: The queue to use
414 * @cmd: The command to send
416 * Safe to use from interrupt context
418 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
419 struct nvme_command *cmd)
421 u16 tail = nvmeq->sq_tail;
423 if (nvmeq->sq_cmds_io)
424 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
426 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
428 if (++tail == nvmeq->q_depth)
430 if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
432 writel(tail, nvmeq->q_db);
433 nvmeq->sq_tail = tail;
436 static __le64 **iod_list(struct request *req)
438 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
439 return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
442 static blk_status_t nvme_init_iod(struct request *rq, struct nvme_dev *dev)
444 struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
445 int nseg = blk_rq_nr_phys_segments(rq);
446 unsigned int size = blk_rq_payload_bytes(rq);
448 if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
449 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
451 return BLK_STS_RESOURCE;
453 iod->sg = iod->inline_sg;
464 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
466 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
467 const int last_prp = dev->ctrl.page_size / 8 - 1;
469 __le64 **list = iod_list(req);
470 dma_addr_t prp_dma = iod->first_dma;
472 if (iod->npages == 0)
473 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
474 for (i = 0; i < iod->npages; i++) {
475 __le64 *prp_list = list[i];
476 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
477 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
478 prp_dma = next_prp_dma;
481 if (iod->sg != iod->inline_sg)
485 #ifdef CONFIG_BLK_DEV_INTEGRITY
486 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
488 if (be32_to_cpu(pi->ref_tag) == v)
489 pi->ref_tag = cpu_to_be32(p);
492 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
494 if (be32_to_cpu(pi->ref_tag) == p)
495 pi->ref_tag = cpu_to_be32(v);
499 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
501 * The virtual start sector is the one that was originally submitted by the
502 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
503 * start sector may be different. Remap protection information to match the
504 * physical LBA on writes, and back to the original seed on reads.
506 * Type 0 and 3 do not have a ref tag, so no remapping required.
508 static void nvme_dif_remap(struct request *req,
509 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
511 struct nvme_ns *ns = req->rq_disk->private_data;
512 struct bio_integrity_payload *bip;
513 struct t10_pi_tuple *pi;
515 u32 i, nlb, ts, phys, virt;
517 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
520 bip = bio_integrity(req->bio);
524 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
527 virt = bip_get_seed(bip);
528 phys = nvme_block_nr(ns, blk_rq_pos(req));
529 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
530 ts = ns->disk->queue->integrity.tuple_size;
532 for (i = 0; i < nlb; i++, virt++, phys++) {
533 pi = (struct t10_pi_tuple *)p;
534 dif_swap(phys, virt, pi);
539 #else /* CONFIG_BLK_DEV_INTEGRITY */
540 static void nvme_dif_remap(struct request *req,
541 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
544 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
547 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
552 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req)
554 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
555 struct dma_pool *pool;
556 int length = blk_rq_payload_bytes(req);
557 struct scatterlist *sg = iod->sg;
558 int dma_len = sg_dma_len(sg);
559 u64 dma_addr = sg_dma_address(sg);
560 u32 page_size = dev->ctrl.page_size;
561 int offset = dma_addr & (page_size - 1);
563 __le64 **list = iod_list(req);
567 length -= (page_size - offset);
571 dma_len -= (page_size - offset);
573 dma_addr += (page_size - offset);
576 dma_addr = sg_dma_address(sg);
577 dma_len = sg_dma_len(sg);
580 if (length <= page_size) {
581 iod->first_dma = dma_addr;
585 nprps = DIV_ROUND_UP(length, page_size);
586 if (nprps <= (256 / 8)) {
587 pool = dev->prp_small_pool;
590 pool = dev->prp_page_pool;
594 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
596 iod->first_dma = dma_addr;
601 iod->first_dma = prp_dma;
604 if (i == page_size >> 3) {
605 __le64 *old_prp_list = prp_list;
606 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
609 list[iod->npages++] = prp_list;
610 prp_list[0] = old_prp_list[i - 1];
611 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
614 prp_list[i++] = cpu_to_le64(dma_addr);
615 dma_len -= page_size;
616 dma_addr += page_size;
624 dma_addr = sg_dma_address(sg);
625 dma_len = sg_dma_len(sg);
631 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
632 struct nvme_command *cmnd)
634 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
635 struct request_queue *q = req->q;
636 enum dma_data_direction dma_dir = rq_data_dir(req) ?
637 DMA_TO_DEVICE : DMA_FROM_DEVICE;
638 blk_status_t ret = BLK_STS_IOERR;
640 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
641 iod->nents = blk_rq_map_sg(q, req, iod->sg);
645 ret = BLK_STS_RESOURCE;
646 if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
650 if (!nvme_setup_prps(dev, req))
654 if (blk_integrity_rq(req)) {
655 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
658 sg_init_table(&iod->meta_sg, 1);
659 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
662 if (rq_data_dir(req))
663 nvme_dif_remap(req, nvme_dif_prep);
665 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
669 cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
670 cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
671 if (blk_integrity_rq(req))
672 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
676 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
681 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
683 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
684 enum dma_data_direction dma_dir = rq_data_dir(req) ?
685 DMA_TO_DEVICE : DMA_FROM_DEVICE;
688 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
689 if (blk_integrity_rq(req)) {
690 if (!rq_data_dir(req))
691 nvme_dif_remap(req, nvme_dif_complete);
692 dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
696 nvme_cleanup_cmd(req);
697 nvme_free_iod(dev, req);
701 * NOTE: ns is NULL when called on the admin queue.
703 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
704 const struct blk_mq_queue_data *bd)
706 struct nvme_ns *ns = hctx->queue->queuedata;
707 struct nvme_queue *nvmeq = hctx->driver_data;
708 struct nvme_dev *dev = nvmeq->dev;
709 struct request *req = bd->rq;
710 struct nvme_command cmnd;
711 blk_status_t ret = BLK_STS_OK;
714 * If formated with metadata, require the block layer provide a buffer
715 * unless this namespace is formated such that the metadata can be
716 * stripped/generated by the controller with PRACT=1.
718 if (ns && ns->ms && !blk_integrity_rq(req)) {
719 if (!(ns->pi_type && ns->ms == 8) &&
720 !blk_rq_is_passthrough(req))
721 return BLK_STS_NOTSUPP;
724 ret = nvme_setup_cmd(ns, req, &cmnd);
728 ret = nvme_init_iod(req, dev);
732 if (blk_rq_nr_phys_segments(req)) {
733 ret = nvme_map_data(dev, req, &cmnd);
735 goto out_cleanup_iod;
738 blk_mq_start_request(req);
740 spin_lock_irq(&nvmeq->q_lock);
741 if (unlikely(nvmeq->cq_vector < 0)) {
743 spin_unlock_irq(&nvmeq->q_lock);
744 goto out_cleanup_iod;
746 __nvme_submit_cmd(nvmeq, &cmnd);
747 nvme_process_cq(nvmeq);
748 spin_unlock_irq(&nvmeq->q_lock);
751 nvme_free_iod(dev, req);
753 nvme_cleanup_cmd(req);
757 static void nvme_pci_complete_rq(struct request *req)
759 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
761 nvme_unmap_data(iod->nvmeq->dev, req);
762 nvme_complete_rq(req);
765 /* We read the CQE phase first to check if the rest of the entry is valid */
766 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
769 return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
772 static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag)
776 head = nvmeq->cq_head;
777 phase = nvmeq->cq_phase;
779 while (nvme_cqe_valid(nvmeq, head, phase)) {
780 struct nvme_completion cqe = nvmeq->cqes[head];
783 if (++head == nvmeq->q_depth) {
788 if (tag && *tag == cqe.command_id)
791 if (unlikely(cqe.command_id >= nvmeq->q_depth)) {
792 dev_warn(nvmeq->dev->ctrl.device,
793 "invalid id %d completed on queue %d\n",
794 cqe.command_id, le16_to_cpu(cqe.sq_id));
799 * AEN requests are special as they don't time out and can
800 * survive any kind of queue freeze and often don't respond to
801 * aborts. We don't even bother to allocate a struct request
802 * for them but rather special case them here.
804 if (unlikely(nvmeq->qid == 0 &&
805 cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) {
806 nvme_complete_async_event(&nvmeq->dev->ctrl,
807 cqe.status, &cqe.result);
811 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id);
812 nvme_end_request(req, cqe.status, cqe.result);
815 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
818 if (likely(nvmeq->cq_vector >= 0))
819 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
821 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
822 nvmeq->cq_head = head;
823 nvmeq->cq_phase = phase;
828 static void nvme_process_cq(struct nvme_queue *nvmeq)
830 __nvme_process_cq(nvmeq, NULL);
833 static irqreturn_t nvme_irq(int irq, void *data)
836 struct nvme_queue *nvmeq = data;
837 spin_lock(&nvmeq->q_lock);
838 nvme_process_cq(nvmeq);
839 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
841 spin_unlock(&nvmeq->q_lock);
845 static irqreturn_t nvme_irq_check(int irq, void *data)
847 struct nvme_queue *nvmeq = data;
848 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
849 return IRQ_WAKE_THREAD;
853 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
855 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
856 spin_lock_irq(&nvmeq->q_lock);
857 __nvme_process_cq(nvmeq, &tag);
858 spin_unlock_irq(&nvmeq->q_lock);
867 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
869 struct nvme_queue *nvmeq = hctx->driver_data;
871 return __nvme_poll(nvmeq, tag);
874 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
876 struct nvme_dev *dev = to_nvme_dev(ctrl);
877 struct nvme_queue *nvmeq = dev->queues[0];
878 struct nvme_command c;
880 memset(&c, 0, sizeof(c));
881 c.common.opcode = nvme_admin_async_event;
882 c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
884 spin_lock_irq(&nvmeq->q_lock);
885 __nvme_submit_cmd(nvmeq, &c);
886 spin_unlock_irq(&nvmeq->q_lock);
889 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
891 struct nvme_command c;
893 memset(&c, 0, sizeof(c));
894 c.delete_queue.opcode = opcode;
895 c.delete_queue.qid = cpu_to_le16(id);
897 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
900 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
901 struct nvme_queue *nvmeq)
903 struct nvme_command c;
904 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
907 * Note: we (ab)use the fact the the prp fields survive if no data
908 * is attached to the request.
910 memset(&c, 0, sizeof(c));
911 c.create_cq.opcode = nvme_admin_create_cq;
912 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
913 c.create_cq.cqid = cpu_to_le16(qid);
914 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
915 c.create_cq.cq_flags = cpu_to_le16(flags);
916 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
918 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
921 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
922 struct nvme_queue *nvmeq)
924 struct nvme_command c;
925 int flags = NVME_QUEUE_PHYS_CONTIG;
928 * Note: we (ab)use the fact the the prp fields survive if no data
929 * is attached to the request.
931 memset(&c, 0, sizeof(c));
932 c.create_sq.opcode = nvme_admin_create_sq;
933 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
934 c.create_sq.sqid = cpu_to_le16(qid);
935 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
936 c.create_sq.sq_flags = cpu_to_le16(flags);
937 c.create_sq.cqid = cpu_to_le16(qid);
939 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
942 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
944 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
947 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
949 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
952 static void abort_endio(struct request *req, blk_status_t error)
954 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
955 struct nvme_queue *nvmeq = iod->nvmeq;
957 dev_warn(nvmeq->dev->ctrl.device,
958 "Abort status: 0x%x", nvme_req(req)->status);
959 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
960 blk_mq_free_request(req);
963 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
965 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
966 struct nvme_queue *nvmeq = iod->nvmeq;
967 struct nvme_dev *dev = nvmeq->dev;
968 struct request *abort_req;
969 struct nvme_command cmd;
972 * Did we miss an interrupt?
974 if (__nvme_poll(nvmeq, req->tag)) {
975 dev_warn(dev->ctrl.device,
976 "I/O %d QID %d timeout, completion polled\n",
977 req->tag, nvmeq->qid);
978 return BLK_EH_HANDLED;
982 * Shutdown immediately if controller times out while starting. The
983 * reset work will see the pci device disabled when it gets the forced
984 * cancellation error. All outstanding requests are completed on
985 * shutdown, so we return BLK_EH_HANDLED.
987 if (dev->ctrl.state == NVME_CTRL_RESETTING) {
988 dev_warn(dev->ctrl.device,
989 "I/O %d QID %d timeout, disable controller\n",
990 req->tag, nvmeq->qid);
991 nvme_dev_disable(dev, false);
992 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
993 return BLK_EH_HANDLED;
997 * Shutdown the controller immediately and schedule a reset if the
998 * command was already aborted once before and still hasn't been
999 * returned to the driver, or if this is the admin queue.
1001 if (!nvmeq->qid || iod->aborted) {
1002 dev_warn(dev->ctrl.device,
1003 "I/O %d QID %d timeout, reset controller\n",
1004 req->tag, nvmeq->qid);
1005 nvme_dev_disable(dev, false);
1009 * Mark the request as handled, since the inline shutdown
1010 * forces all outstanding requests to complete.
1012 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1013 return BLK_EH_HANDLED;
1016 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1017 atomic_inc(&dev->ctrl.abort_limit);
1018 return BLK_EH_RESET_TIMER;
1022 memset(&cmd, 0, sizeof(cmd));
1023 cmd.abort.opcode = nvme_admin_abort_cmd;
1024 cmd.abort.cid = req->tag;
1025 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1027 dev_warn(nvmeq->dev->ctrl.device,
1028 "I/O %d QID %d timeout, aborting\n",
1029 req->tag, nvmeq->qid);
1031 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1032 BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1033 if (IS_ERR(abort_req)) {
1034 atomic_inc(&dev->ctrl.abort_limit);
1035 return BLK_EH_RESET_TIMER;
1038 abort_req->timeout = ADMIN_TIMEOUT;
1039 abort_req->end_io_data = NULL;
1040 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1043 * The aborted req will be completed on receiving the abort req.
1044 * We enable the timer again. If hit twice, it'll cause a device reset,
1045 * as the device then is in a faulty state.
1047 return BLK_EH_RESET_TIMER;
1050 static void nvme_free_queue(struct nvme_queue *nvmeq)
1052 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1053 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1055 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1056 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1060 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1064 for (i = dev->queue_count - 1; i >= lowest; i--) {
1065 struct nvme_queue *nvmeq = dev->queues[i];
1067 dev->queues[i] = NULL;
1068 nvme_free_queue(nvmeq);
1073 * nvme_suspend_queue - put queue into suspended state
1074 * @nvmeq - queue to suspend
1076 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1080 spin_lock_irq(&nvmeq->q_lock);
1081 if (nvmeq->cq_vector == -1) {
1082 spin_unlock_irq(&nvmeq->q_lock);
1085 vector = nvmeq->cq_vector;
1086 nvmeq->dev->online_queues--;
1087 nvmeq->cq_vector = -1;
1088 spin_unlock_irq(&nvmeq->q_lock);
1090 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1091 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1093 pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1098 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1100 struct nvme_queue *nvmeq = dev->queues[0];
1104 if (nvme_suspend_queue(nvmeq))
1108 nvme_shutdown_ctrl(&dev->ctrl);
1110 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1111 dev->bar + NVME_REG_CAP));
1113 spin_lock_irq(&nvmeq->q_lock);
1114 nvme_process_cq(nvmeq);
1115 spin_unlock_irq(&nvmeq->q_lock);
1118 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1121 int q_depth = dev->q_depth;
1122 unsigned q_size_aligned = roundup(q_depth * entry_size,
1123 dev->ctrl.page_size);
1125 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1126 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1127 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1128 q_depth = div_u64(mem_per_q, entry_size);
1131 * Ensure the reduced q_depth is above some threshold where it
1132 * would be better to map queues in system memory with the
1142 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1145 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1146 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1147 dev->ctrl.page_size);
1148 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1149 nvmeq->sq_cmds_io = dev->cmb + offset;
1151 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1152 &nvmeq->sq_dma_addr, GFP_KERNEL);
1153 if (!nvmeq->sq_cmds)
1160 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1161 int depth, int node)
1163 struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
1168 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1169 &nvmeq->cq_dma_addr, GFP_KERNEL);
1173 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1176 nvmeq->q_dmadev = dev->dev;
1178 spin_lock_init(&nvmeq->q_lock);
1180 nvmeq->cq_phase = 1;
1181 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1182 nvmeq->q_depth = depth;
1184 nvmeq->cq_vector = -1;
1185 dev->queues[qid] = nvmeq;
1191 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1192 nvmeq->cq_dma_addr);
1198 static int queue_request_irq(struct nvme_queue *nvmeq)
1200 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1201 int nr = nvmeq->dev->ctrl.instance;
1203 if (use_threaded_interrupts) {
1204 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1205 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1207 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1208 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1212 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1214 struct nvme_dev *dev = nvmeq->dev;
1216 spin_lock_irq(&nvmeq->q_lock);
1219 nvmeq->cq_phase = 1;
1220 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1221 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1222 nvme_dbbuf_init(dev, nvmeq, qid);
1223 dev->online_queues++;
1224 spin_unlock_irq(&nvmeq->q_lock);
1227 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1229 struct nvme_dev *dev = nvmeq->dev;
1232 nvmeq->cq_vector = qid - 1;
1233 result = adapter_alloc_cq(dev, qid, nvmeq);
1237 result = adapter_alloc_sq(dev, qid, nvmeq);
1241 result = queue_request_irq(nvmeq);
1245 nvme_init_queue(nvmeq, qid);
1249 adapter_delete_sq(dev, qid);
1251 adapter_delete_cq(dev, qid);
1255 static const struct blk_mq_ops nvme_mq_admin_ops = {
1256 .queue_rq = nvme_queue_rq,
1257 .complete = nvme_pci_complete_rq,
1258 .init_hctx = nvme_admin_init_hctx,
1259 .exit_hctx = nvme_admin_exit_hctx,
1260 .init_request = nvme_admin_init_request,
1261 .timeout = nvme_timeout,
1264 static const struct blk_mq_ops nvme_mq_ops = {
1265 .queue_rq = nvme_queue_rq,
1266 .complete = nvme_pci_complete_rq,
1267 .init_hctx = nvme_init_hctx,
1268 .init_request = nvme_init_request,
1269 .map_queues = nvme_pci_map_queues,
1270 .timeout = nvme_timeout,
1274 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1276 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1278 * If the controller was reset during removal, it's possible
1279 * user requests may be waiting on a stopped queue. Start the
1280 * queue to flush these to completion.
1282 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1283 blk_cleanup_queue(dev->ctrl.admin_q);
1284 blk_mq_free_tag_set(&dev->admin_tagset);
1288 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1290 if (!dev->ctrl.admin_q) {
1291 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1292 dev->admin_tagset.nr_hw_queues = 1;
1295 * Subtract one to leave an empty queue entry for 'Full Queue'
1296 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1298 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1299 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1300 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1301 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1302 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1303 dev->admin_tagset.driver_data = dev;
1305 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1308 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1309 if (IS_ERR(dev->ctrl.admin_q)) {
1310 blk_mq_free_tag_set(&dev->admin_tagset);
1313 if (!blk_get_queue(dev->ctrl.admin_q)) {
1314 nvme_dev_remove_admin(dev);
1315 dev->ctrl.admin_q = NULL;
1319 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1324 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1326 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1329 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1331 struct pci_dev *pdev = to_pci_dev(dev->dev);
1333 if (size <= dev->bar_mapped_size)
1335 if (size > pci_resource_len(pdev, 0))
1339 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1341 dev->bar_mapped_size = 0;
1344 dev->bar_mapped_size = size;
1345 dev->dbs = dev->bar + NVME_REG_DBS;
1350 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1354 u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1355 struct nvme_queue *nvmeq;
1357 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1361 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1362 NVME_CAP_NSSRC(cap) : 0;
1364 if (dev->subsystem &&
1365 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1366 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1368 result = nvme_disable_ctrl(&dev->ctrl, cap);
1372 nvmeq = dev->queues[0];
1374 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
1375 dev_to_node(dev->dev));
1380 aqa = nvmeq->q_depth - 1;
1383 writel(aqa, dev->bar + NVME_REG_AQA);
1384 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1385 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1387 result = nvme_enable_ctrl(&dev->ctrl, cap);
1391 nvmeq->cq_vector = 0;
1392 result = queue_request_irq(nvmeq);
1394 nvmeq->cq_vector = -1;
1401 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1404 /* If true, indicates loss of adapter communication, possibly by a
1405 * NVMe Subsystem reset.
1407 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1409 /* If there is a reset ongoing, we shouldn't reset again. */
1410 if (dev->ctrl.state == NVME_CTRL_RESETTING)
1413 /* We shouldn't reset unless the controller is on fatal error state
1414 * _or_ if we lost the communication with it.
1416 if (!(csts & NVME_CSTS_CFS) && !nssro)
1419 /* If PCI error recovery process is happening, we cannot reset or
1420 * the recovery mechanism will surely fail.
1422 if (pci_channel_offline(to_pci_dev(dev->dev)))
1428 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
1430 /* Read a config register to help see what died. */
1434 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
1436 if (result == PCIBIOS_SUCCESSFUL)
1437 dev_warn(dev->ctrl.device,
1438 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
1441 dev_warn(dev->ctrl.device,
1442 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
1446 static void nvme_watchdog_timer(unsigned long data)
1448 struct nvme_dev *dev = (struct nvme_dev *)data;
1449 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1451 /* Skip controllers under certain specific conditions. */
1452 if (nvme_should_reset(dev, csts)) {
1453 if (!nvme_reset(dev))
1454 nvme_warn_reset(dev, csts);
1458 mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1461 static int nvme_create_io_queues(struct nvme_dev *dev)
1466 for (i = dev->queue_count; i <= dev->max_qid; i++) {
1467 /* vector == qid - 1, match nvme_create_queue */
1468 if (!nvme_alloc_queue(dev, i, dev->q_depth,
1469 pci_irq_get_node(to_pci_dev(dev->dev), i - 1))) {
1475 max = min(dev->max_qid, dev->queue_count - 1);
1476 for (i = dev->online_queues; i <= max; i++) {
1477 ret = nvme_create_queue(dev->queues[i], i);
1483 * Ignore failing Create SQ/CQ commands, we can continue with less
1484 * than the desired aount of queues, and even a controller without
1485 * I/O queues an still be used to issue admin commands. This might
1486 * be useful to upgrade a buggy firmware for example.
1488 return ret >= 0 ? 0 : ret;
1491 static ssize_t nvme_cmb_show(struct device *dev,
1492 struct device_attribute *attr,
1495 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1497 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1498 ndev->cmbloc, ndev->cmbsz);
1500 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1502 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1504 u64 szu, size, offset;
1505 resource_size_t bar_size;
1506 struct pci_dev *pdev = to_pci_dev(dev->dev);
1508 dma_addr_t dma_addr;
1510 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1511 if (!(NVME_CMB_SZ(dev->cmbsz)))
1513 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1518 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1519 size = szu * NVME_CMB_SZ(dev->cmbsz);
1520 offset = szu * NVME_CMB_OFST(dev->cmbloc);
1521 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1523 if (offset > bar_size)
1527 * Controllers may support a CMB size larger than their BAR,
1528 * for example, due to being behind a bridge. Reduce the CMB to
1529 * the reported size of the BAR
1531 if (size > bar_size - offset)
1532 size = bar_size - offset;
1534 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1535 cmb = ioremap_wc(dma_addr, size);
1539 dev->cmb_dma_addr = dma_addr;
1540 dev->cmb_size = size;
1544 static inline void nvme_release_cmb(struct nvme_dev *dev)
1550 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1551 &dev_attr_cmb.attr, NULL);
1557 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1559 size_t len = dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs);
1560 struct nvme_command c;
1564 dma_addr = dma_map_single(dev->dev, dev->host_mem_descs, len,
1566 if (dma_mapping_error(dev->dev, dma_addr))
1569 memset(&c, 0, sizeof(c));
1570 c.features.opcode = nvme_admin_set_features;
1571 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1572 c.features.dword11 = cpu_to_le32(bits);
1573 c.features.dword12 = cpu_to_le32(dev->host_mem_size >>
1574 ilog2(dev->ctrl.page_size));
1575 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1576 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1577 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1579 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1581 dev_warn(dev->ctrl.device,
1582 "failed to set host mem (err %d, flags %#x).\n",
1585 dma_unmap_single(dev->dev, dma_addr, len, DMA_TO_DEVICE);
1589 static void nvme_free_host_mem(struct nvme_dev *dev)
1593 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1594 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1595 size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1597 dma_free_coherent(dev->dev, size, dev->host_mem_desc_bufs[i],
1598 le64_to_cpu(desc->addr));
1601 kfree(dev->host_mem_desc_bufs);
1602 dev->host_mem_desc_bufs = NULL;
1603 kfree(dev->host_mem_descs);
1604 dev->host_mem_descs = NULL;
1607 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1609 struct nvme_host_mem_buf_desc *descs;
1610 u32 chunk_size, max_entries, i = 0;
1614 /* start big and work our way down */
1615 chunk_size = min(preferred, (u64)PAGE_SIZE << MAX_ORDER);
1617 tmp = (preferred + chunk_size - 1);
1618 do_div(tmp, chunk_size);
1620 descs = kcalloc(max_entries, sizeof(*descs), GFP_KERNEL);
1624 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1626 goto out_free_descs;
1628 for (size = 0; size < preferred; size += chunk_size) {
1629 u32 len = min_t(u64, chunk_size, preferred - size);
1630 dma_addr_t dma_addr;
1632 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1633 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1637 descs[i].addr = cpu_to_le64(dma_addr);
1638 descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1642 if (!size || (min && size < min)) {
1643 dev_warn(dev->ctrl.device,
1644 "failed to allocate host memory buffer.\n");
1648 dev_info(dev->ctrl.device,
1649 "allocated %lld MiB host memory buffer.\n",
1650 size >> ilog2(SZ_1M));
1651 dev->nr_host_mem_descs = i;
1652 dev->host_mem_size = size;
1653 dev->host_mem_descs = descs;
1654 dev->host_mem_desc_bufs = bufs;
1659 size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1661 dma_free_coherent(dev->dev, size, bufs[i],
1662 le64_to_cpu(descs[i].addr));
1669 /* try a smaller chunk size if we failed early */
1670 if (chunk_size >= PAGE_SIZE * 2 && (i == 0 || size < min)) {
1674 dev->host_mem_descs = NULL;
1678 static void nvme_setup_host_mem(struct nvme_dev *dev)
1680 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1681 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1682 u64 min = (u64)dev->ctrl.hmmin * 4096;
1683 u32 enable_bits = NVME_HOST_MEM_ENABLE;
1685 preferred = min(preferred, max);
1687 dev_warn(dev->ctrl.device,
1688 "min host memory (%lld MiB) above limit (%d MiB).\n",
1689 min >> ilog2(SZ_1M), max_host_mem_size_mb);
1690 nvme_free_host_mem(dev);
1695 * If we already have a buffer allocated check if we can reuse it.
1697 if (dev->host_mem_descs) {
1698 if (dev->host_mem_size >= min)
1699 enable_bits |= NVME_HOST_MEM_RETURN;
1701 nvme_free_host_mem(dev);
1704 if (!dev->host_mem_descs) {
1705 if (nvme_alloc_host_mem(dev, min, preferred))
1709 if (nvme_set_host_mem(dev, enable_bits))
1710 nvme_free_host_mem(dev);
1713 static int nvme_setup_io_queues(struct nvme_dev *dev)
1715 struct nvme_queue *adminq = dev->queues[0];
1716 struct pci_dev *pdev = to_pci_dev(dev->dev);
1717 int result, nr_io_queues;
1720 nr_io_queues = num_online_cpus();
1721 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1725 if (nr_io_queues == 0)
1728 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1729 result = nvme_cmb_qdepth(dev, nr_io_queues,
1730 sizeof(struct nvme_command));
1732 dev->q_depth = result;
1734 nvme_release_cmb(dev);
1738 size = db_bar_size(dev, nr_io_queues);
1739 result = nvme_remap_bar(dev, size);
1742 if (!--nr_io_queues)
1745 adminq->q_db = dev->dbs;
1747 /* Deregister the admin queue's interrupt */
1748 pci_free_irq(pdev, 0, adminq);
1751 * If we enable msix early due to not intx, disable it again before
1752 * setting up the full range we need.
1754 pci_free_irq_vectors(pdev);
1755 nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1756 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1757 if (nr_io_queues <= 0)
1759 dev->max_qid = nr_io_queues;
1762 * Should investigate if there's a performance win from allocating
1763 * more queues than interrupt vectors; it might allow the submission
1764 * path to scale better, even if the receive path is limited by the
1765 * number of interrupts.
1768 result = queue_request_irq(adminq);
1770 adminq->cq_vector = -1;
1773 return nvme_create_io_queues(dev);
1776 static void nvme_del_queue_end(struct request *req, blk_status_t error)
1778 struct nvme_queue *nvmeq = req->end_io_data;
1780 blk_mq_free_request(req);
1781 complete(&nvmeq->dev->ioq_wait);
1784 static void nvme_del_cq_end(struct request *req, blk_status_t error)
1786 struct nvme_queue *nvmeq = req->end_io_data;
1789 unsigned long flags;
1792 * We might be called with the AQ q_lock held
1793 * and the I/O queue q_lock should always
1794 * nest inside the AQ one.
1796 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1797 SINGLE_DEPTH_NESTING);
1798 nvme_process_cq(nvmeq);
1799 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1802 nvme_del_queue_end(req, error);
1805 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1807 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1808 struct request *req;
1809 struct nvme_command cmd;
1811 memset(&cmd, 0, sizeof(cmd));
1812 cmd.delete_queue.opcode = opcode;
1813 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1815 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1817 return PTR_ERR(req);
1819 req->timeout = ADMIN_TIMEOUT;
1820 req->end_io_data = nvmeq;
1822 blk_execute_rq_nowait(q, NULL, req, false,
1823 opcode == nvme_admin_delete_cq ?
1824 nvme_del_cq_end : nvme_del_queue_end);
1828 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1831 unsigned long timeout;
1832 u8 opcode = nvme_admin_delete_sq;
1834 for (pass = 0; pass < 2; pass++) {
1835 int sent = 0, i = queues;
1837 reinit_completion(&dev->ioq_wait);
1839 timeout = ADMIN_TIMEOUT;
1840 for (; i > 0; i--, sent++)
1841 if (nvme_delete_queue(dev->queues[i], opcode))
1845 timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1851 opcode = nvme_admin_delete_cq;
1856 * Return: error value if an error occurred setting up the queues or calling
1857 * Identify Device. 0 if these succeeded, even if adding some of the
1858 * namespaces failed. At the moment, these failures are silent. TBD which
1859 * failures should be reported.
1861 static int nvme_dev_add(struct nvme_dev *dev)
1863 if (!dev->ctrl.tagset) {
1864 dev->tagset.ops = &nvme_mq_ops;
1865 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1866 dev->tagset.timeout = NVME_IO_TIMEOUT;
1867 dev->tagset.numa_node = dev_to_node(dev->dev);
1868 dev->tagset.queue_depth =
1869 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1870 dev->tagset.cmd_size = nvme_cmd_size(dev);
1871 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1872 dev->tagset.driver_data = dev;
1874 if (blk_mq_alloc_tag_set(&dev->tagset))
1876 dev->ctrl.tagset = &dev->tagset;
1878 nvme_dbbuf_set(dev);
1880 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1882 /* Free previously allocated queues that are no longer usable */
1883 nvme_free_queues(dev, dev->online_queues);
1889 static int nvme_pci_enable(struct nvme_dev *dev)
1892 int result = -ENOMEM;
1893 struct pci_dev *pdev = to_pci_dev(dev->dev);
1895 if (pci_enable_device_mem(pdev))
1898 pci_set_master(pdev);
1900 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1901 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1904 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1910 * Some devices and/or platforms don't advertise or work with INTx
1911 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1912 * adjust this later.
1914 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1918 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1920 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1921 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1922 dev->dbs = dev->bar + 4096;
1925 * Temporary fix for the Apple controller found in the MacBook8,1 and
1926 * some MacBook7,1 to avoid controller resets and data loss.
1928 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1930 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
1931 "set queue depth=%u to work around controller resets\n",
1936 * CMBs can currently only exist on >=1.2 PCIe devices. We only
1937 * populate sysfs if a CMB is implemented. Note that we add the
1938 * CMB attribute to the nvme_ctrl kobj which removes the need to remove
1939 * it on exit. Since nvme_dev_attrs_group has no name we can pass
1940 * NULL as final argument to sysfs_add_file_to_group.
1943 if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1944 dev->cmb = nvme_map_cmb(dev);
1947 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1948 &dev_attr_cmb.attr, NULL))
1949 dev_warn(dev->ctrl.device,
1950 "failed to add sysfs attribute for CMB\n");
1954 pci_enable_pcie_error_reporting(pdev);
1955 pci_save_state(pdev);
1959 pci_disable_device(pdev);
1963 static void nvme_dev_unmap(struct nvme_dev *dev)
1967 pci_release_mem_regions(to_pci_dev(dev->dev));
1970 static void nvme_pci_disable(struct nvme_dev *dev)
1972 struct pci_dev *pdev = to_pci_dev(dev->dev);
1974 nvme_release_cmb(dev);
1975 pci_free_irq_vectors(pdev);
1977 if (pci_is_enabled(pdev)) {
1978 pci_disable_pcie_error_reporting(pdev);
1979 pci_disable_device(pdev);
1983 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1987 struct pci_dev *pdev = to_pci_dev(dev->dev);
1989 del_timer_sync(&dev->watchdog_timer);
1991 mutex_lock(&dev->shutdown_lock);
1992 if (pci_is_enabled(pdev)) {
1993 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1995 if (dev->ctrl.state == NVME_CTRL_LIVE)
1996 nvme_start_freeze(&dev->ctrl);
1997 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
1998 pdev->error_state != pci_channel_io_normal);
2002 * Give the controller a chance to complete all entered requests if
2003 * doing a safe shutdown.
2007 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
2010 * If the controller is still alive tell it to stop using the
2011 * host memory buffer. In theory the shutdown / reset should
2012 * make sure that it doesn't access the host memoery anymore,
2013 * but I'd rather be safe than sorry..
2015 if (dev->host_mem_descs)
2016 nvme_set_host_mem(dev, 0);
2019 nvme_stop_queues(&dev->ctrl);
2021 queues = dev->online_queues - 1;
2022 for (i = dev->queue_count - 1; i > 0; i--)
2023 nvme_suspend_queue(dev->queues[i]);
2026 /* A device might become IO incapable very soon during
2027 * probe, before the admin queue is configured. Thus,
2028 * queue_count can be 0 here.
2030 if (dev->queue_count)
2031 nvme_suspend_queue(dev->queues[0]);
2033 nvme_disable_io_queues(dev, queues);
2034 nvme_disable_admin_queue(dev, shutdown);
2036 nvme_pci_disable(dev);
2038 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2039 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2042 * The driver will not be starting up queues again if shutting down so
2043 * must flush all entered requests to their failed completion to avoid
2044 * deadlocking blk-mq hot-cpu notifier.
2047 nvme_start_queues(&dev->ctrl);
2048 mutex_unlock(&dev->shutdown_lock);
2051 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2053 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2054 PAGE_SIZE, PAGE_SIZE, 0);
2055 if (!dev->prp_page_pool)
2058 /* Optimisation for I/Os between 4k and 128k */
2059 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2061 if (!dev->prp_small_pool) {
2062 dma_pool_destroy(dev->prp_page_pool);
2068 static void nvme_release_prp_pools(struct nvme_dev *dev)
2070 dma_pool_destroy(dev->prp_page_pool);
2071 dma_pool_destroy(dev->prp_small_pool);
2074 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2076 struct nvme_dev *dev = to_nvme_dev(ctrl);
2078 nvme_dbbuf_dma_free(dev);
2079 put_device(dev->dev);
2080 if (dev->tagset.tags)
2081 blk_mq_free_tag_set(&dev->tagset);
2082 if (dev->ctrl.admin_q)
2083 blk_put_queue(dev->ctrl.admin_q);
2085 free_opal_dev(dev->ctrl.opal_dev);
2089 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
2091 dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
2093 kref_get(&dev->ctrl.kref);
2094 nvme_dev_disable(dev, false);
2095 if (!schedule_work(&dev->remove_work))
2096 nvme_put_ctrl(&dev->ctrl);
2099 static void nvme_reset_work(struct work_struct *work)
2101 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
2102 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2103 int result = -ENODEV;
2105 if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
2109 * If we're called to reset a live controller first shut it down before
2112 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2113 nvme_dev_disable(dev, false);
2115 result = nvme_pci_enable(dev);
2119 result = nvme_configure_admin_queue(dev);
2123 nvme_init_queue(dev->queues[0], 0);
2124 result = nvme_alloc_admin_tags(dev);
2128 result = nvme_init_identify(&dev->ctrl);
2132 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2133 if (!dev->ctrl.opal_dev)
2134 dev->ctrl.opal_dev =
2135 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2136 else if (was_suspend)
2137 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2139 free_opal_dev(dev->ctrl.opal_dev);
2140 dev->ctrl.opal_dev = NULL;
2143 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2144 result = nvme_dbbuf_dma_alloc(dev);
2147 "unable to allocate dma for dbbuf\n");
2150 if (dev->ctrl.hmpre)
2151 nvme_setup_host_mem(dev);
2153 result = nvme_setup_io_queues(dev);
2158 * A controller that can not execute IO typically requires user
2159 * intervention to correct. For such degraded controllers, the driver
2160 * should not submit commands the user did not request, so skip
2161 * registering for asynchronous event notification on this condition.
2163 if (dev->online_queues > 1)
2164 nvme_queue_async_events(&dev->ctrl);
2166 mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
2169 * Keep the controller around but remove all namespaces if we don't have
2170 * any working I/O queue.
2172 if (dev->online_queues < 2) {
2173 dev_warn(dev->ctrl.device, "IO queues not created\n");
2174 nvme_kill_queues(&dev->ctrl);
2175 nvme_remove_namespaces(&dev->ctrl);
2177 nvme_start_queues(&dev->ctrl);
2178 nvme_wait_freeze(&dev->ctrl);
2180 nvme_unfreeze(&dev->ctrl);
2183 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2184 dev_warn(dev->ctrl.device, "failed to mark controller live\n");
2188 if (dev->online_queues > 1)
2189 nvme_queue_scan(&dev->ctrl);
2193 nvme_remove_dead_ctrl(dev, result);
2196 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2198 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2199 struct pci_dev *pdev = to_pci_dev(dev->dev);
2201 nvme_kill_queues(&dev->ctrl);
2202 if (pci_get_drvdata(pdev))
2203 device_release_driver(&pdev->dev);
2204 nvme_put_ctrl(&dev->ctrl);
2207 static int nvme_reset(struct nvme_dev *dev)
2209 if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q))
2211 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING))
2213 if (!queue_work(nvme_wq, &dev->reset_work))
2218 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2220 *val = readl(to_nvme_dev(ctrl)->bar + off);
2224 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2226 writel(val, to_nvme_dev(ctrl)->bar + off);
2230 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2232 *val = readq(to_nvme_dev(ctrl)->bar + off);
2236 static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl)
2238 struct nvme_dev *dev = to_nvme_dev(ctrl);
2239 int ret = nvme_reset(dev);
2242 flush_work(&dev->reset_work);
2246 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2248 .module = THIS_MODULE,
2249 .flags = NVME_F_METADATA_SUPPORTED,
2250 .reg_read32 = nvme_pci_reg_read32,
2251 .reg_write32 = nvme_pci_reg_write32,
2252 .reg_read64 = nvme_pci_reg_read64,
2253 .reset_ctrl = nvme_pci_reset_ctrl,
2254 .free_ctrl = nvme_pci_free_ctrl,
2255 .submit_async_event = nvme_pci_submit_async_event,
2258 static int nvme_dev_map(struct nvme_dev *dev)
2260 struct pci_dev *pdev = to_pci_dev(dev->dev);
2262 if (pci_request_mem_regions(pdev, "nvme"))
2265 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2270 pci_release_mem_regions(pdev);
2274 static unsigned long check_dell_samsung_bug(struct pci_dev *pdev)
2276 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2278 * Several Samsung devices seem to drop off the PCIe bus
2279 * randomly when APST is on and uses the deepest sleep state.
2280 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2281 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2282 * 950 PRO 256GB", but it seems to be restricted to two Dell
2285 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2286 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2287 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2288 return NVME_QUIRK_NO_DEEPEST_PS;
2294 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2296 int node, result = -ENOMEM;
2297 struct nvme_dev *dev;
2298 unsigned long quirks = id->driver_data;
2300 node = dev_to_node(&pdev->dev);
2301 if (node == NUMA_NO_NODE)
2302 set_dev_node(&pdev->dev, first_memory_node);
2304 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2307 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2312 dev->dev = get_device(&pdev->dev);
2313 pci_set_drvdata(pdev, dev);
2315 result = nvme_dev_map(dev);
2319 INIT_WORK(&dev->reset_work, nvme_reset_work);
2320 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2321 setup_timer(&dev->watchdog_timer, nvme_watchdog_timer,
2322 (unsigned long)dev);
2323 mutex_init(&dev->shutdown_lock);
2324 init_completion(&dev->ioq_wait);
2326 result = nvme_setup_prp_pools(dev);
2330 quirks |= check_dell_samsung_bug(pdev);
2332 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2337 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING);
2338 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2340 queue_work(nvme_wq, &dev->reset_work);
2344 nvme_release_prp_pools(dev);
2346 put_device(dev->dev);
2347 nvme_dev_unmap(dev);
2354 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2356 struct nvme_dev *dev = pci_get_drvdata(pdev);
2359 nvme_dev_disable(dev, false);
2364 static void nvme_shutdown(struct pci_dev *pdev)
2366 struct nvme_dev *dev = pci_get_drvdata(pdev);
2367 nvme_dev_disable(dev, true);
2371 * The driver's remove may be called on a device in a partially initialized
2372 * state. This function must not have any dependencies on the device state in
2375 static void nvme_remove(struct pci_dev *pdev)
2377 struct nvme_dev *dev = pci_get_drvdata(pdev);
2379 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2381 cancel_work_sync(&dev->reset_work);
2382 pci_set_drvdata(pdev, NULL);
2384 if (!pci_device_is_present(pdev)) {
2385 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2386 nvme_dev_disable(dev, false);
2389 flush_work(&dev->reset_work);
2390 nvme_uninit_ctrl(&dev->ctrl);
2391 nvme_dev_disable(dev, true);
2392 nvme_free_host_mem(dev);
2393 nvme_dev_remove_admin(dev);
2394 nvme_free_queues(dev, 0);
2395 nvme_release_prp_pools(dev);
2396 nvme_dev_unmap(dev);
2397 nvme_put_ctrl(&dev->ctrl);
2400 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2405 if (pci_vfs_assigned(pdev)) {
2406 dev_warn(&pdev->dev,
2407 "Cannot disable SR-IOV VFs while assigned\n");
2410 pci_disable_sriov(pdev);
2414 ret = pci_enable_sriov(pdev, numvfs);
2415 return ret ? ret : numvfs;
2418 #ifdef CONFIG_PM_SLEEP
2419 static int nvme_suspend(struct device *dev)
2421 struct pci_dev *pdev = to_pci_dev(dev);
2422 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2424 nvme_dev_disable(ndev, true);
2428 static int nvme_resume(struct device *dev)
2430 struct pci_dev *pdev = to_pci_dev(dev);
2431 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2438 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2440 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2441 pci_channel_state_t state)
2443 struct nvme_dev *dev = pci_get_drvdata(pdev);
2446 * A frozen channel requires a reset. When detected, this method will
2447 * shutdown the controller to quiesce. The controller will be restarted
2448 * after the slot reset through driver's slot_reset callback.
2451 case pci_channel_io_normal:
2452 return PCI_ERS_RESULT_CAN_RECOVER;
2453 case pci_channel_io_frozen:
2454 dev_warn(dev->ctrl.device,
2455 "frozen state error detected, reset controller\n");
2456 nvme_dev_disable(dev, false);
2457 return PCI_ERS_RESULT_NEED_RESET;
2458 case pci_channel_io_perm_failure:
2459 dev_warn(dev->ctrl.device,
2460 "failure state error detected, request disconnect\n");
2461 return PCI_ERS_RESULT_DISCONNECT;
2463 return PCI_ERS_RESULT_NEED_RESET;
2466 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2468 struct nvme_dev *dev = pci_get_drvdata(pdev);
2470 dev_info(dev->ctrl.device, "restart after slot reset\n");
2471 pci_restore_state(pdev);
2473 return PCI_ERS_RESULT_RECOVERED;
2476 static void nvme_error_resume(struct pci_dev *pdev)
2478 pci_cleanup_aer_uncorrect_error_status(pdev);
2481 static const struct pci_error_handlers nvme_err_handler = {
2482 .error_detected = nvme_error_detected,
2483 .slot_reset = nvme_slot_reset,
2484 .resume = nvme_error_resume,
2485 .reset_notify = nvme_reset_notify,
2488 static const struct pci_device_id nvme_id_table[] = {
2489 { PCI_VDEVICE(INTEL, 0x0953),
2490 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2491 NVME_QUIRK_DEALLOCATE_ZEROES, },
2492 { PCI_VDEVICE(INTEL, 0x0a53),
2493 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2494 NVME_QUIRK_DEALLOCATE_ZEROES, },
2495 { PCI_VDEVICE(INTEL, 0x0a54),
2496 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2497 NVME_QUIRK_DEALLOCATE_ZEROES, },
2498 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
2499 .driver_data = NVME_QUIRK_NO_DEEPEST_PS },
2500 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
2501 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2502 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
2503 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2504 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
2505 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2506 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2507 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2508 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2511 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2513 static struct pci_driver nvme_driver = {
2515 .id_table = nvme_id_table,
2516 .probe = nvme_probe,
2517 .remove = nvme_remove,
2518 .shutdown = nvme_shutdown,
2520 .pm = &nvme_dev_pm_ops,
2522 .sriov_configure = nvme_pci_sriov_configure,
2523 .err_handler = &nvme_err_handler,
2526 static int __init nvme_init(void)
2528 return pci_register_driver(&nvme_driver);
2531 static void __exit nvme_exit(void)
2533 pci_unregister_driver(&nvme_driver);
2537 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2538 MODULE_LICENSE("GPL");
2539 MODULE_VERSION("1.0");
2540 module_init(nvme_init);
2541 module_exit(nvme_exit);