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/cpu.h>
20 #include <linux/delay.h>
21 #include <linux/errno.h>
23 #include <linux/genhd.h>
24 #include <linux/hdreg.h>
25 #include <linux/idr.h>
26 #include <linux/init.h>
27 #include <linux/interrupt.h>
29 #include <linux/kdev_t.h>
30 #include <linux/kernel.h>
32 #include <linux/module.h>
33 #include <linux/moduleparam.h>
34 #include <linux/mutex.h>
35 #include <linux/pci.h>
36 #include <linux/poison.h>
37 #include <linux/ptrace.h>
38 #include <linux/sched.h>
39 #include <linux/slab.h>
40 #include <linux/t10-pi.h>
41 #include <linux/timer.h>
42 #include <linux/types.h>
43 #include <linux/io-64-nonatomic-lo-hi.h>
44 #include <asm/unaligned.h>
48 #define NVME_Q_DEPTH 1024
49 #define NVME_AQ_DEPTH 256
50 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
51 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
54 * We handle AEN commands ourselves and don't even let the
55 * block layer know about them.
57 #define NVME_NR_AEN_COMMANDS 1
58 #define NVME_AQ_BLKMQ_DEPTH (NVME_AQ_DEPTH - NVME_NR_AEN_COMMANDS)
60 static int use_threaded_interrupts;
61 module_param(use_threaded_interrupts, int, 0);
63 static bool use_cmb_sqes = true;
64 module_param(use_cmb_sqes, bool, 0644);
65 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
67 static struct workqueue_struct *nvme_workq;
72 static int nvme_reset(struct nvme_dev *dev);
73 static void nvme_process_cq(struct nvme_queue *nvmeq);
74 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
77 * Represents an NVM Express device. Each nvme_dev is a PCI function.
80 struct nvme_queue **queues;
81 struct blk_mq_tag_set tagset;
82 struct blk_mq_tag_set admin_tagset;
85 struct dma_pool *prp_page_pool;
86 struct dma_pool *prp_small_pool;
88 unsigned online_queues;
92 struct msix_entry *entry;
94 struct work_struct reset_work;
95 struct work_struct scan_work;
96 struct work_struct remove_work;
97 struct work_struct async_work;
98 struct timer_list watchdog_timer;
99 struct mutex shutdown_lock;
102 dma_addr_t cmb_dma_addr;
107 #define NVME_CTRL_RESETTING 0
108 #define NVME_CTRL_REMOVING 1
110 struct nvme_ctrl ctrl;
111 struct completion ioq_wait;
114 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
116 return container_of(ctrl, struct nvme_dev, ctrl);
120 * An NVM Express queue. Each device has at least two (one for admin
121 * commands and one for I/O commands).
124 struct device *q_dmadev;
125 struct nvme_dev *dev;
126 char irqname[24]; /* nvme4294967295-65535\0 */
128 struct nvme_command *sq_cmds;
129 struct nvme_command __iomem *sq_cmds_io;
130 volatile struct nvme_completion *cqes;
131 struct blk_mq_tags **tags;
132 dma_addr_t sq_dma_addr;
133 dma_addr_t cq_dma_addr;
145 * The nvme_iod describes the data in an I/O, including the list of PRP
146 * entries. You can't see it in this data structure because C doesn't let
147 * me express that. Use nvme_init_iod to ensure there's enough space
148 * allocated to store the PRP list.
151 struct nvme_queue *nvmeq;
153 int npages; /* In the PRP list. 0 means small pool in use */
154 int nents; /* Used in scatterlist */
155 int length; /* Of data, in bytes */
156 dma_addr_t first_dma;
157 struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
158 struct scatterlist *sg;
159 struct scatterlist inline_sg[0];
163 * Check we didin't inadvertently grow the command struct
165 static inline void _nvme_check_size(void)
167 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
168 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
169 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
170 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
171 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
172 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
173 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
174 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
175 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
176 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
177 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
178 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
182 * Max size of iod being embedded in the request payload
184 #define NVME_INT_PAGES 2
185 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size)
188 * Will slightly overestimate the number of pages needed. This is OK
189 * as it only leads to a small amount of wasted memory for the lifetime of
192 static int nvme_npages(unsigned size, struct nvme_dev *dev)
194 unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
195 dev->ctrl.page_size);
196 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
199 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
200 unsigned int size, unsigned int nseg)
202 return sizeof(__le64 *) * nvme_npages(size, dev) +
203 sizeof(struct scatterlist) * nseg;
206 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
208 return sizeof(struct nvme_iod) +
209 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
212 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
213 unsigned int hctx_idx)
215 struct nvme_dev *dev = data;
216 struct nvme_queue *nvmeq = dev->queues[0];
218 WARN_ON(hctx_idx != 0);
219 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
220 WARN_ON(nvmeq->tags);
222 hctx->driver_data = nvmeq;
223 nvmeq->tags = &dev->admin_tagset.tags[0];
227 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
229 struct nvme_queue *nvmeq = hctx->driver_data;
234 static int nvme_admin_init_request(void *data, struct request *req,
235 unsigned int hctx_idx, unsigned int rq_idx,
236 unsigned int numa_node)
238 struct nvme_dev *dev = data;
239 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
240 struct nvme_queue *nvmeq = dev->queues[0];
247 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
248 unsigned int hctx_idx)
250 struct nvme_dev *dev = data;
251 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
254 nvmeq->tags = &dev->tagset.tags[hctx_idx];
256 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
257 hctx->driver_data = nvmeq;
261 static int nvme_init_request(void *data, struct request *req,
262 unsigned int hctx_idx, unsigned int rq_idx,
263 unsigned int numa_node)
265 struct nvme_dev *dev = data;
266 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
267 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
274 static void nvme_queue_scan(struct nvme_dev *dev)
277 * Do not queue new scan work when a controller is reset during
280 if (test_bit(NVME_CTRL_REMOVING, &dev->flags))
282 queue_work(nvme_workq, &dev->scan_work);
285 static void nvme_complete_async_event(struct nvme_dev *dev,
286 struct nvme_completion *cqe)
288 u16 status = le16_to_cpu(cqe->status) >> 1;
289 u32 result = le32_to_cpu(cqe->result);
291 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ) {
292 ++dev->ctrl.event_limit;
293 queue_work(nvme_workq, &dev->async_work);
296 if (status != NVME_SC_SUCCESS)
299 switch (result & 0xff07) {
300 case NVME_AER_NOTICE_NS_CHANGED:
301 dev_info(dev->ctrl.device, "rescanning\n");
302 nvme_queue_scan(dev);
304 dev_warn(dev->ctrl.device, "async event result %08x\n", result);
309 * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
310 * @nvmeq: The queue to use
311 * @cmd: The command to send
313 * Safe to use from interrupt context
315 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
316 struct nvme_command *cmd)
318 u16 tail = nvmeq->sq_tail;
320 if (nvmeq->sq_cmds_io)
321 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
323 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
325 if (++tail == nvmeq->q_depth)
327 writel(tail, nvmeq->q_db);
328 nvmeq->sq_tail = tail;
331 static __le64 **iod_list(struct request *req)
333 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
334 return (__le64 **)(iod->sg + req->nr_phys_segments);
337 static int nvme_init_iod(struct request *rq, unsigned size,
338 struct nvme_dev *dev)
340 struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
341 int nseg = rq->nr_phys_segments;
343 if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
344 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
346 return BLK_MQ_RQ_QUEUE_BUSY;
348 iod->sg = iod->inline_sg;
358 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
360 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
361 const int last_prp = dev->ctrl.page_size / 8 - 1;
363 __le64 **list = iod_list(req);
364 dma_addr_t prp_dma = iod->first_dma;
366 if (req->cmd_flags & REQ_DISCARD)
367 kfree(req->completion_data);
369 if (iod->npages == 0)
370 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
371 for (i = 0; i < iod->npages; i++) {
372 __le64 *prp_list = list[i];
373 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
374 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
375 prp_dma = next_prp_dma;
378 if (iod->sg != iod->inline_sg)
382 #ifdef CONFIG_BLK_DEV_INTEGRITY
383 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
385 if (be32_to_cpu(pi->ref_tag) == v)
386 pi->ref_tag = cpu_to_be32(p);
389 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
391 if (be32_to_cpu(pi->ref_tag) == p)
392 pi->ref_tag = cpu_to_be32(v);
396 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
398 * The virtual start sector is the one that was originally submitted by the
399 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
400 * start sector may be different. Remap protection information to match the
401 * physical LBA on writes, and back to the original seed on reads.
403 * Type 0 and 3 do not have a ref tag, so no remapping required.
405 static void nvme_dif_remap(struct request *req,
406 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
408 struct nvme_ns *ns = req->rq_disk->private_data;
409 struct bio_integrity_payload *bip;
410 struct t10_pi_tuple *pi;
412 u32 i, nlb, ts, phys, virt;
414 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
417 bip = bio_integrity(req->bio);
421 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
424 virt = bip_get_seed(bip);
425 phys = nvme_block_nr(ns, blk_rq_pos(req));
426 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
427 ts = ns->disk->queue->integrity.tuple_size;
429 for (i = 0; i < nlb; i++, virt++, phys++) {
430 pi = (struct t10_pi_tuple *)p;
431 dif_swap(phys, virt, pi);
436 #else /* CONFIG_BLK_DEV_INTEGRITY */
437 static void nvme_dif_remap(struct request *req,
438 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
441 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
444 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
449 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req,
452 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
453 struct dma_pool *pool;
454 int length = total_len;
455 struct scatterlist *sg = iod->sg;
456 int dma_len = sg_dma_len(sg);
457 u64 dma_addr = sg_dma_address(sg);
458 u32 page_size = dev->ctrl.page_size;
459 int offset = dma_addr & (page_size - 1);
461 __le64 **list = iod_list(req);
465 length -= (page_size - offset);
469 dma_len -= (page_size - offset);
471 dma_addr += (page_size - offset);
474 dma_addr = sg_dma_address(sg);
475 dma_len = sg_dma_len(sg);
478 if (length <= page_size) {
479 iod->first_dma = dma_addr;
483 nprps = DIV_ROUND_UP(length, page_size);
484 if (nprps <= (256 / 8)) {
485 pool = dev->prp_small_pool;
488 pool = dev->prp_page_pool;
492 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
494 iod->first_dma = dma_addr;
499 iod->first_dma = prp_dma;
502 if (i == page_size >> 3) {
503 __le64 *old_prp_list = prp_list;
504 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
507 list[iod->npages++] = prp_list;
508 prp_list[0] = old_prp_list[i - 1];
509 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
512 prp_list[i++] = cpu_to_le64(dma_addr);
513 dma_len -= page_size;
514 dma_addr += page_size;
522 dma_addr = sg_dma_address(sg);
523 dma_len = sg_dma_len(sg);
529 static int nvme_map_data(struct nvme_dev *dev, struct request *req,
530 unsigned size, struct nvme_command *cmnd)
532 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
533 struct request_queue *q = req->q;
534 enum dma_data_direction dma_dir = rq_data_dir(req) ?
535 DMA_TO_DEVICE : DMA_FROM_DEVICE;
536 int ret = BLK_MQ_RQ_QUEUE_ERROR;
538 sg_init_table(iod->sg, req->nr_phys_segments);
539 iod->nents = blk_rq_map_sg(q, req, iod->sg);
543 ret = BLK_MQ_RQ_QUEUE_BUSY;
544 if (!dma_map_sg(dev->dev, iod->sg, iod->nents, dma_dir))
547 if (!nvme_setup_prps(dev, req, size))
550 ret = BLK_MQ_RQ_QUEUE_ERROR;
551 if (blk_integrity_rq(req)) {
552 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
555 sg_init_table(&iod->meta_sg, 1);
556 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
559 if (rq_data_dir(req))
560 nvme_dif_remap(req, nvme_dif_prep);
562 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
566 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
567 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
568 if (blk_integrity_rq(req))
569 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
570 return BLK_MQ_RQ_QUEUE_OK;
573 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
578 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
580 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
581 enum dma_data_direction dma_dir = rq_data_dir(req) ?
582 DMA_TO_DEVICE : DMA_FROM_DEVICE;
585 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
586 if (blk_integrity_rq(req)) {
587 if (!rq_data_dir(req))
588 nvme_dif_remap(req, nvme_dif_complete);
589 dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
593 nvme_free_iod(dev, req);
597 * NOTE: ns is NULL when called on the admin queue.
599 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
600 const struct blk_mq_queue_data *bd)
602 struct nvme_ns *ns = hctx->queue->queuedata;
603 struct nvme_queue *nvmeq = hctx->driver_data;
604 struct nvme_dev *dev = nvmeq->dev;
605 struct request *req = bd->rq;
606 struct nvme_command cmnd;
608 int ret = BLK_MQ_RQ_QUEUE_OK;
611 * If formated with metadata, require the block layer provide a buffer
612 * unless this namespace is formated such that the metadata can be
613 * stripped/generated by the controller with PRACT=1.
615 if (ns && ns->ms && !blk_integrity_rq(req)) {
616 if (!(ns->pi_type && ns->ms == 8) &&
617 req->cmd_type != REQ_TYPE_DRV_PRIV) {
618 blk_mq_end_request(req, -EFAULT);
619 return BLK_MQ_RQ_QUEUE_OK;
623 map_len = nvme_map_len(req);
624 ret = nvme_init_iod(req, map_len, dev);
628 ret = nvme_setup_cmd(ns, req, &cmnd);
632 if (req->nr_phys_segments)
633 ret = nvme_map_data(dev, req, map_len, &cmnd);
638 cmnd.common.command_id = req->tag;
639 blk_mq_start_request(req);
641 spin_lock_irq(&nvmeq->q_lock);
642 if (unlikely(nvmeq->cq_vector < 0)) {
643 if (ns && !test_bit(NVME_NS_DEAD, &ns->flags))
644 ret = BLK_MQ_RQ_QUEUE_BUSY;
646 ret = BLK_MQ_RQ_QUEUE_ERROR;
647 spin_unlock_irq(&nvmeq->q_lock);
650 __nvme_submit_cmd(nvmeq, &cmnd);
651 nvme_process_cq(nvmeq);
652 spin_unlock_irq(&nvmeq->q_lock);
653 return BLK_MQ_RQ_QUEUE_OK;
655 nvme_free_iod(dev, req);
659 static void nvme_complete_rq(struct request *req)
661 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
662 struct nvme_dev *dev = iod->nvmeq->dev;
665 nvme_unmap_data(dev, req);
667 if (unlikely(req->errors)) {
668 if (nvme_req_needs_retry(req, req->errors)) {
669 nvme_requeue_req(req);
673 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
676 error = nvme_error_status(req->errors);
679 if (unlikely(iod->aborted)) {
680 dev_warn(dev->ctrl.device,
681 "completing aborted command with status: %04x\n",
685 blk_mq_end_request(req, error);
688 /* We read the CQE phase first to check if the rest of the entry is valid */
689 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
692 return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
695 static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag)
699 head = nvmeq->cq_head;
700 phase = nvmeq->cq_phase;
702 while (nvme_cqe_valid(nvmeq, head, phase)) {
703 struct nvme_completion cqe = nvmeq->cqes[head];
706 if (++head == nvmeq->q_depth) {
711 if (tag && *tag == cqe.command_id)
714 if (unlikely(cqe.command_id >= nvmeq->q_depth)) {
715 dev_warn(nvmeq->dev->ctrl.device,
716 "invalid id %d completed on queue %d\n",
717 cqe.command_id, le16_to_cpu(cqe.sq_id));
722 * AEN requests are special as they don't time out and can
723 * survive any kind of queue freeze and often don't respond to
724 * aborts. We don't even bother to allocate a struct request
725 * for them but rather special case them here.
727 if (unlikely(nvmeq->qid == 0 &&
728 cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) {
729 nvme_complete_async_event(nvmeq->dev, &cqe);
733 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id);
734 if (req->cmd_type == REQ_TYPE_DRV_PRIV && req->special)
735 memcpy(req->special, &cqe, sizeof(cqe));
736 blk_mq_complete_request(req, le16_to_cpu(cqe.status) >> 1);
740 /* If the controller ignores the cq head doorbell and continuously
741 * writes to the queue, it is theoretically possible to wrap around
742 * the queue twice and mistakenly return IRQ_NONE. Linux only
743 * requires that 0.1% of your interrupts are handled, so this isn't
746 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
749 if (likely(nvmeq->cq_vector >= 0))
750 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
751 nvmeq->cq_head = head;
752 nvmeq->cq_phase = phase;
757 static void nvme_process_cq(struct nvme_queue *nvmeq)
759 __nvme_process_cq(nvmeq, NULL);
762 static irqreturn_t nvme_irq(int irq, void *data)
765 struct nvme_queue *nvmeq = data;
766 spin_lock(&nvmeq->q_lock);
767 nvme_process_cq(nvmeq);
768 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
770 spin_unlock(&nvmeq->q_lock);
774 static irqreturn_t nvme_irq_check(int irq, void *data)
776 struct nvme_queue *nvmeq = data;
777 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
778 return IRQ_WAKE_THREAD;
782 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
784 struct nvme_queue *nvmeq = hctx->driver_data;
786 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
787 spin_lock_irq(&nvmeq->q_lock);
788 __nvme_process_cq(nvmeq, &tag);
789 spin_unlock_irq(&nvmeq->q_lock);
798 static void nvme_async_event_work(struct work_struct *work)
800 struct nvme_dev *dev = container_of(work, struct nvme_dev, async_work);
801 struct nvme_queue *nvmeq = dev->queues[0];
802 struct nvme_command c;
804 memset(&c, 0, sizeof(c));
805 c.common.opcode = nvme_admin_async_event;
807 spin_lock_irq(&nvmeq->q_lock);
808 while (dev->ctrl.event_limit > 0) {
809 c.common.command_id = NVME_AQ_BLKMQ_DEPTH +
810 --dev->ctrl.event_limit;
811 __nvme_submit_cmd(nvmeq, &c);
813 spin_unlock_irq(&nvmeq->q_lock);
816 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
818 struct nvme_command c;
820 memset(&c, 0, sizeof(c));
821 c.delete_queue.opcode = opcode;
822 c.delete_queue.qid = cpu_to_le16(id);
824 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
827 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
828 struct nvme_queue *nvmeq)
830 struct nvme_command c;
831 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
834 * Note: we (ab)use the fact the the prp fields survive if no data
835 * is attached to the request.
837 memset(&c, 0, sizeof(c));
838 c.create_cq.opcode = nvme_admin_create_cq;
839 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
840 c.create_cq.cqid = cpu_to_le16(qid);
841 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
842 c.create_cq.cq_flags = cpu_to_le16(flags);
843 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
845 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
848 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
849 struct nvme_queue *nvmeq)
851 struct nvme_command c;
852 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
855 * Note: we (ab)use the fact the the prp fields survive if no data
856 * is attached to the request.
858 memset(&c, 0, sizeof(c));
859 c.create_sq.opcode = nvme_admin_create_sq;
860 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
861 c.create_sq.sqid = cpu_to_le16(qid);
862 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
863 c.create_sq.sq_flags = cpu_to_le16(flags);
864 c.create_sq.cqid = cpu_to_le16(qid);
866 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
869 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
871 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
874 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
876 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
879 static void abort_endio(struct request *req, int error)
881 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
882 struct nvme_queue *nvmeq = iod->nvmeq;
883 u16 status = req->errors;
885 dev_warn(nvmeq->dev->ctrl.device, "Abort status: 0x%x", status);
886 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
887 blk_mq_free_request(req);
890 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
892 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
893 struct nvme_queue *nvmeq = iod->nvmeq;
894 struct nvme_dev *dev = nvmeq->dev;
895 struct request *abort_req;
896 struct nvme_command cmd;
899 * Shutdown immediately if controller times out while starting. The
900 * reset work will see the pci device disabled when it gets the forced
901 * cancellation error. All outstanding requests are completed on
902 * shutdown, so we return BLK_EH_HANDLED.
904 if (test_bit(NVME_CTRL_RESETTING, &dev->flags)) {
905 dev_warn(dev->ctrl.device,
906 "I/O %d QID %d timeout, disable controller\n",
907 req->tag, nvmeq->qid);
908 nvme_dev_disable(dev, false);
909 req->errors = NVME_SC_CANCELLED;
910 return BLK_EH_HANDLED;
914 * Shutdown the controller immediately and schedule a reset if the
915 * command was already aborted once before and still hasn't been
916 * returned to the driver, or if this is the admin queue.
918 if (!nvmeq->qid || iod->aborted) {
919 dev_warn(dev->ctrl.device,
920 "I/O %d QID %d timeout, reset controller\n",
921 req->tag, nvmeq->qid);
922 nvme_dev_disable(dev, false);
923 queue_work(nvme_workq, &dev->reset_work);
926 * Mark the request as handled, since the inline shutdown
927 * forces all outstanding requests to complete.
929 req->errors = NVME_SC_CANCELLED;
930 return BLK_EH_HANDLED;
935 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
936 atomic_inc(&dev->ctrl.abort_limit);
937 return BLK_EH_RESET_TIMER;
940 memset(&cmd, 0, sizeof(cmd));
941 cmd.abort.opcode = nvme_admin_abort_cmd;
942 cmd.abort.cid = req->tag;
943 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
945 dev_warn(nvmeq->dev->ctrl.device,
946 "I/O %d QID %d timeout, aborting\n",
947 req->tag, nvmeq->qid);
949 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
951 if (IS_ERR(abort_req)) {
952 atomic_inc(&dev->ctrl.abort_limit);
953 return BLK_EH_RESET_TIMER;
956 abort_req->timeout = ADMIN_TIMEOUT;
957 abort_req->end_io_data = NULL;
958 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
961 * The aborted req will be completed on receiving the abort req.
962 * We enable the timer again. If hit twice, it'll cause a device reset,
963 * as the device then is in a faulty state.
965 return BLK_EH_RESET_TIMER;
968 static void nvme_cancel_io(struct request *req, void *data, bool reserved)
972 if (!blk_mq_request_started(req))
975 dev_dbg_ratelimited(((struct nvme_dev *) data)->ctrl.device,
976 "Cancelling I/O %d", req->tag);
978 status = NVME_SC_ABORT_REQ;
979 if (blk_queue_dying(req->q))
980 status |= NVME_SC_DNR;
981 blk_mq_complete_request(req, status);
984 static void nvme_free_queue(struct nvme_queue *nvmeq)
986 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
987 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
989 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
990 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
994 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
998 for (i = dev->queue_count - 1; i >= lowest; i--) {
999 struct nvme_queue *nvmeq = dev->queues[i];
1001 dev->queues[i] = NULL;
1002 nvme_free_queue(nvmeq);
1007 * nvme_suspend_queue - put queue into suspended state
1008 * @nvmeq - queue to suspend
1010 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1014 spin_lock_irq(&nvmeq->q_lock);
1015 if (nvmeq->cq_vector == -1) {
1016 spin_unlock_irq(&nvmeq->q_lock);
1019 vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1020 nvmeq->dev->online_queues--;
1021 nvmeq->cq_vector = -1;
1022 spin_unlock_irq(&nvmeq->q_lock);
1024 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1025 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1027 irq_set_affinity_hint(vector, NULL);
1028 free_irq(vector, nvmeq);
1033 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1035 struct nvme_queue *nvmeq = dev->queues[0];
1039 if (nvme_suspend_queue(nvmeq))
1043 nvme_shutdown_ctrl(&dev->ctrl);
1045 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1046 dev->bar + NVME_REG_CAP));
1048 spin_lock_irq(&nvmeq->q_lock);
1049 nvme_process_cq(nvmeq);
1050 spin_unlock_irq(&nvmeq->q_lock);
1053 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1056 int q_depth = dev->q_depth;
1057 unsigned q_size_aligned = roundup(q_depth * entry_size,
1058 dev->ctrl.page_size);
1060 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1061 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1062 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1063 q_depth = div_u64(mem_per_q, entry_size);
1066 * Ensure the reduced q_depth is above some threshold where it
1067 * would be better to map queues in system memory with the
1077 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1080 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1081 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1082 dev->ctrl.page_size);
1083 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1084 nvmeq->sq_cmds_io = dev->cmb + offset;
1086 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1087 &nvmeq->sq_dma_addr, GFP_KERNEL);
1088 if (!nvmeq->sq_cmds)
1095 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1098 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1102 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1103 &nvmeq->cq_dma_addr, GFP_KERNEL);
1107 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1110 nvmeq->q_dmadev = dev->dev;
1112 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1113 dev->ctrl.instance, qid);
1114 spin_lock_init(&nvmeq->q_lock);
1116 nvmeq->cq_phase = 1;
1117 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1118 nvmeq->q_depth = depth;
1120 nvmeq->cq_vector = -1;
1121 dev->queues[qid] = nvmeq;
1127 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1128 nvmeq->cq_dma_addr);
1134 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1137 if (use_threaded_interrupts)
1138 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1139 nvme_irq_check, nvme_irq, IRQF_SHARED,
1141 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1142 IRQF_SHARED, name, nvmeq);
1145 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1147 struct nvme_dev *dev = nvmeq->dev;
1149 spin_lock_irq(&nvmeq->q_lock);
1152 nvmeq->cq_phase = 1;
1153 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1154 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1155 dev->online_queues++;
1156 spin_unlock_irq(&nvmeq->q_lock);
1159 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1161 struct nvme_dev *dev = nvmeq->dev;
1164 nvmeq->cq_vector = qid - 1;
1165 result = adapter_alloc_cq(dev, qid, nvmeq);
1169 result = adapter_alloc_sq(dev, qid, nvmeq);
1173 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1177 nvme_init_queue(nvmeq, qid);
1181 adapter_delete_sq(dev, qid);
1183 adapter_delete_cq(dev, qid);
1187 static struct blk_mq_ops nvme_mq_admin_ops = {
1188 .queue_rq = nvme_queue_rq,
1189 .complete = nvme_complete_rq,
1190 .map_queue = blk_mq_map_queue,
1191 .init_hctx = nvme_admin_init_hctx,
1192 .exit_hctx = nvme_admin_exit_hctx,
1193 .init_request = nvme_admin_init_request,
1194 .timeout = nvme_timeout,
1197 static struct blk_mq_ops nvme_mq_ops = {
1198 .queue_rq = nvme_queue_rq,
1199 .complete = nvme_complete_rq,
1200 .map_queue = blk_mq_map_queue,
1201 .init_hctx = nvme_init_hctx,
1202 .init_request = nvme_init_request,
1203 .timeout = nvme_timeout,
1207 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1209 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1211 * If the controller was reset during removal, it's possible
1212 * user requests may be waiting on a stopped queue. Start the
1213 * queue to flush these to completion.
1215 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1216 blk_cleanup_queue(dev->ctrl.admin_q);
1217 blk_mq_free_tag_set(&dev->admin_tagset);
1221 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1223 if (!dev->ctrl.admin_q) {
1224 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1225 dev->admin_tagset.nr_hw_queues = 1;
1228 * Subtract one to leave an empty queue entry for 'Full Queue'
1229 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1231 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1232 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1233 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1234 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1235 dev->admin_tagset.driver_data = dev;
1237 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1240 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1241 if (IS_ERR(dev->ctrl.admin_q)) {
1242 blk_mq_free_tag_set(&dev->admin_tagset);
1245 if (!blk_get_queue(dev->ctrl.admin_q)) {
1246 nvme_dev_remove_admin(dev);
1247 dev->ctrl.admin_q = NULL;
1251 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1256 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1260 u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1261 struct nvme_queue *nvmeq;
1263 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1) ?
1264 NVME_CAP_NSSRC(cap) : 0;
1266 if (dev->subsystem &&
1267 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1268 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1270 result = nvme_disable_ctrl(&dev->ctrl, cap);
1274 nvmeq = dev->queues[0];
1276 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1281 aqa = nvmeq->q_depth - 1;
1284 writel(aqa, dev->bar + NVME_REG_AQA);
1285 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1286 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1288 result = nvme_enable_ctrl(&dev->ctrl, cap);
1292 nvmeq->cq_vector = 0;
1293 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1295 nvmeq->cq_vector = -1;
1302 nvme_free_queues(dev, 0);
1306 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1309 /* If true, indicates loss of adapter communication, possibly by a
1310 * NVMe Subsystem reset.
1312 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1314 /* If there is a reset ongoing, we shouldn't reset again. */
1315 if (work_busy(&dev->reset_work))
1318 /* We shouldn't reset unless the controller is on fatal error state
1319 * _or_ if we lost the communication with it.
1321 if (!(csts & NVME_CSTS_CFS) && !nssro)
1324 /* If PCI error recovery process is happening, we cannot reset or
1325 * the recovery mechanism will surely fail.
1327 if (pci_channel_offline(to_pci_dev(dev->dev)))
1333 static void nvme_watchdog_timer(unsigned long data)
1335 struct nvme_dev *dev = (struct nvme_dev *)data;
1336 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1338 /* Skip controllers under certain specific conditions. */
1339 if (nvme_should_reset(dev, csts)) {
1340 if (queue_work(nvme_workq, &dev->reset_work))
1342 "Failed status: 0x%x, reset controller.\n",
1347 mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1350 static int nvme_create_io_queues(struct nvme_dev *dev)
1355 for (i = dev->queue_count; i <= dev->max_qid; i++) {
1356 if (!nvme_alloc_queue(dev, i, dev->q_depth)) {
1362 max = min(dev->max_qid, dev->queue_count - 1);
1363 for (i = dev->online_queues; i <= max; i++) {
1364 ret = nvme_create_queue(dev->queues[i], i);
1366 nvme_free_queues(dev, i);
1372 * Ignore failing Create SQ/CQ commands, we can continue with less
1373 * than the desired aount of queues, and even a controller without
1374 * I/O queues an still be used to issue admin commands. This might
1375 * be useful to upgrade a buggy firmware for example.
1377 return ret >= 0 ? 0 : ret;
1380 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1382 u64 szu, size, offset;
1384 resource_size_t bar_size;
1385 struct pci_dev *pdev = to_pci_dev(dev->dev);
1387 dma_addr_t dma_addr;
1392 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1393 if (!(NVME_CMB_SZ(dev->cmbsz)))
1396 cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1398 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1399 size = szu * NVME_CMB_SZ(dev->cmbsz);
1400 offset = szu * NVME_CMB_OFST(cmbloc);
1401 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(cmbloc));
1403 if (offset > bar_size)
1407 * Controllers may support a CMB size larger than their BAR,
1408 * for example, due to being behind a bridge. Reduce the CMB to
1409 * the reported size of the BAR
1411 if (size > bar_size - offset)
1412 size = bar_size - offset;
1414 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(cmbloc)) + offset;
1415 cmb = ioremap_wc(dma_addr, size);
1419 dev->cmb_dma_addr = dma_addr;
1420 dev->cmb_size = size;
1424 static inline void nvme_release_cmb(struct nvme_dev *dev)
1432 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1434 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1437 static int nvme_setup_io_queues(struct nvme_dev *dev)
1439 struct nvme_queue *adminq = dev->queues[0];
1440 struct pci_dev *pdev = to_pci_dev(dev->dev);
1441 int result, i, vecs, nr_io_queues, size;
1443 nr_io_queues = num_possible_cpus();
1444 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1449 * Degraded controllers might return an error when setting the queue
1450 * count. We still want to be able to bring them online and offer
1451 * access to the admin queue, as that might be only way to fix them up.
1454 dev_err(dev->ctrl.device,
1455 "Could not set queue count (%d)\n", result);
1459 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1460 result = nvme_cmb_qdepth(dev, nr_io_queues,
1461 sizeof(struct nvme_command));
1463 dev->q_depth = result;
1465 nvme_release_cmb(dev);
1468 size = db_bar_size(dev, nr_io_queues);
1472 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1475 if (!--nr_io_queues)
1477 size = db_bar_size(dev, nr_io_queues);
1479 dev->dbs = dev->bar + 4096;
1480 adminq->q_db = dev->dbs;
1483 /* Deregister the admin queue's interrupt */
1484 free_irq(dev->entry[0].vector, adminq);
1487 * If we enable msix early due to not intx, disable it again before
1488 * setting up the full range we need.
1490 if (pdev->msi_enabled)
1491 pci_disable_msi(pdev);
1492 else if (pdev->msix_enabled)
1493 pci_disable_msix(pdev);
1495 for (i = 0; i < nr_io_queues; i++)
1496 dev->entry[i].entry = i;
1497 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
1499 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
1503 for (i = 0; i < vecs; i++)
1504 dev->entry[i].vector = i + pdev->irq;
1509 * Should investigate if there's a performance win from allocating
1510 * more queues than interrupt vectors; it might allow the submission
1511 * path to scale better, even if the receive path is limited by the
1512 * number of interrupts.
1514 nr_io_queues = vecs;
1515 dev->max_qid = nr_io_queues;
1517 result = queue_request_irq(dev, adminq, adminq->irqname);
1519 adminq->cq_vector = -1;
1522 return nvme_create_io_queues(dev);
1525 nvme_free_queues(dev, 1);
1529 static void nvme_set_irq_hints(struct nvme_dev *dev)
1531 struct nvme_queue *nvmeq;
1534 for (i = 0; i < dev->online_queues; i++) {
1535 nvmeq = dev->queues[i];
1537 if (!nvmeq->tags || !(*nvmeq->tags))
1540 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
1541 blk_mq_tags_cpumask(*nvmeq->tags));
1545 static void nvme_dev_scan(struct work_struct *work)
1547 struct nvme_dev *dev = container_of(work, struct nvme_dev, scan_work);
1549 if (!dev->tagset.tags)
1551 nvme_scan_namespaces(&dev->ctrl);
1552 nvme_set_irq_hints(dev);
1555 static void nvme_del_queue_end(struct request *req, int error)
1557 struct nvme_queue *nvmeq = req->end_io_data;
1559 blk_mq_free_request(req);
1560 complete(&nvmeq->dev->ioq_wait);
1563 static void nvme_del_cq_end(struct request *req, int error)
1565 struct nvme_queue *nvmeq = req->end_io_data;
1568 unsigned long flags;
1571 * We might be called with the AQ q_lock held
1572 * and the I/O queue q_lock should always
1573 * nest inside the AQ one.
1575 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1576 SINGLE_DEPTH_NESTING);
1577 nvme_process_cq(nvmeq);
1578 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1581 nvme_del_queue_end(req, error);
1584 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1586 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1587 struct request *req;
1588 struct nvme_command cmd;
1590 memset(&cmd, 0, sizeof(cmd));
1591 cmd.delete_queue.opcode = opcode;
1592 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1594 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT);
1596 return PTR_ERR(req);
1598 req->timeout = ADMIN_TIMEOUT;
1599 req->end_io_data = nvmeq;
1601 blk_execute_rq_nowait(q, NULL, req, false,
1602 opcode == nvme_admin_delete_cq ?
1603 nvme_del_cq_end : nvme_del_queue_end);
1607 static void nvme_disable_io_queues(struct nvme_dev *dev)
1610 unsigned long timeout;
1611 u8 opcode = nvme_admin_delete_sq;
1613 for (pass = 0; pass < 2; pass++) {
1614 int sent = 0, i = dev->queue_count - 1;
1616 reinit_completion(&dev->ioq_wait);
1618 timeout = ADMIN_TIMEOUT;
1619 for (; i > 0; i--) {
1620 struct nvme_queue *nvmeq = dev->queues[i];
1623 nvme_suspend_queue(nvmeq);
1624 if (nvme_delete_queue(nvmeq, opcode))
1629 timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1635 opcode = nvme_admin_delete_cq;
1640 * Return: error value if an error occurred setting up the queues or calling
1641 * Identify Device. 0 if these succeeded, even if adding some of the
1642 * namespaces failed. At the moment, these failures are silent. TBD which
1643 * failures should be reported.
1645 static int nvme_dev_add(struct nvme_dev *dev)
1647 if (!dev->ctrl.tagset) {
1648 dev->tagset.ops = &nvme_mq_ops;
1649 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1650 dev->tagset.timeout = NVME_IO_TIMEOUT;
1651 dev->tagset.numa_node = dev_to_node(dev->dev);
1652 dev->tagset.queue_depth =
1653 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1654 dev->tagset.cmd_size = nvme_cmd_size(dev);
1655 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1656 dev->tagset.driver_data = dev;
1658 if (blk_mq_alloc_tag_set(&dev->tagset))
1660 dev->ctrl.tagset = &dev->tagset;
1662 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1664 /* Free previously allocated queues that are no longer usable */
1665 nvme_free_queues(dev, dev->online_queues);
1668 nvme_queue_scan(dev);
1672 static int nvme_pci_enable(struct nvme_dev *dev)
1675 int result = -ENOMEM;
1676 struct pci_dev *pdev = to_pci_dev(dev->dev);
1678 if (pci_enable_device_mem(pdev))
1681 pci_set_master(pdev);
1683 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1684 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1687 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1693 * Some devices and/or platforms don't advertise or work with INTx
1694 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1695 * adjust this later.
1697 if (pci_enable_msix(pdev, dev->entry, 1)) {
1698 pci_enable_msi(pdev);
1699 dev->entry[0].vector = pdev->irq;
1702 if (!dev->entry[0].vector) {
1707 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1709 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1710 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1711 dev->dbs = dev->bar + 4096;
1714 * Temporary fix for the Apple controller found in the MacBook8,1 and
1715 * some MacBook7,1 to avoid controller resets and data loss.
1717 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1719 dev_warn(dev->dev, "detected Apple NVMe controller, set "
1720 "queue depth=%u to work around controller resets\n",
1724 if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2))
1725 dev->cmb = nvme_map_cmb(dev);
1727 pci_enable_pcie_error_reporting(pdev);
1728 pci_save_state(pdev);
1732 pci_disable_device(pdev);
1736 static void nvme_dev_unmap(struct nvme_dev *dev)
1740 pci_release_regions(to_pci_dev(dev->dev));
1743 static void nvme_pci_disable(struct nvme_dev *dev)
1745 struct pci_dev *pdev = to_pci_dev(dev->dev);
1747 if (pdev->msi_enabled)
1748 pci_disable_msi(pdev);
1749 else if (pdev->msix_enabled)
1750 pci_disable_msix(pdev);
1752 if (pci_is_enabled(pdev)) {
1753 pci_disable_pcie_error_reporting(pdev);
1754 pci_disable_device(pdev);
1758 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1763 del_timer_sync(&dev->watchdog_timer);
1765 mutex_lock(&dev->shutdown_lock);
1766 if (pci_is_enabled(to_pci_dev(dev->dev))) {
1767 nvme_stop_queues(&dev->ctrl);
1768 csts = readl(dev->bar + NVME_REG_CSTS);
1770 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
1771 for (i = dev->queue_count - 1; i >= 0; i--) {
1772 struct nvme_queue *nvmeq = dev->queues[i];
1773 nvme_suspend_queue(nvmeq);
1776 nvme_disable_io_queues(dev);
1777 nvme_disable_admin_queue(dev, shutdown);
1779 nvme_pci_disable(dev);
1781 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_io, dev);
1782 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_io, dev);
1783 mutex_unlock(&dev->shutdown_lock);
1786 static int nvme_setup_prp_pools(struct nvme_dev *dev)
1788 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
1789 PAGE_SIZE, PAGE_SIZE, 0);
1790 if (!dev->prp_page_pool)
1793 /* Optimisation for I/Os between 4k and 128k */
1794 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
1796 if (!dev->prp_small_pool) {
1797 dma_pool_destroy(dev->prp_page_pool);
1803 static void nvme_release_prp_pools(struct nvme_dev *dev)
1805 dma_pool_destroy(dev->prp_page_pool);
1806 dma_pool_destroy(dev->prp_small_pool);
1809 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
1811 struct nvme_dev *dev = to_nvme_dev(ctrl);
1813 put_device(dev->dev);
1814 if (dev->tagset.tags)
1815 blk_mq_free_tag_set(&dev->tagset);
1816 if (dev->ctrl.admin_q)
1817 blk_put_queue(dev->ctrl.admin_q);
1823 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
1825 dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
1827 kref_get(&dev->ctrl.kref);
1828 nvme_dev_disable(dev, false);
1829 if (!schedule_work(&dev->remove_work))
1830 nvme_put_ctrl(&dev->ctrl);
1833 static void nvme_reset_work(struct work_struct *work)
1835 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
1836 int result = -ENODEV;
1838 if (WARN_ON(test_bit(NVME_CTRL_RESETTING, &dev->flags)))
1842 * If we're called to reset a live controller first shut it down before
1845 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
1846 nvme_dev_disable(dev, false);
1848 set_bit(NVME_CTRL_RESETTING, &dev->flags);
1850 result = nvme_pci_enable(dev);
1854 result = nvme_configure_admin_queue(dev);
1858 nvme_init_queue(dev->queues[0], 0);
1859 result = nvme_alloc_admin_tags(dev);
1863 result = nvme_init_identify(&dev->ctrl);
1867 result = nvme_setup_io_queues(dev);
1872 * A controller that can not execute IO typically requires user
1873 * intervention to correct. For such degraded controllers, the driver
1874 * should not submit commands the user did not request, so skip
1875 * registering for asynchronous event notification on this condition.
1877 if (dev->online_queues > 1) {
1878 dev->ctrl.event_limit = NVME_NR_AEN_COMMANDS;
1879 queue_work(nvme_workq, &dev->async_work);
1882 mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1885 * Keep the controller around but remove all namespaces if we don't have
1886 * any working I/O queue.
1888 if (dev->online_queues < 2) {
1889 dev_warn(dev->ctrl.device, "IO queues not created\n");
1890 nvme_kill_queues(&dev->ctrl);
1891 nvme_remove_namespaces(&dev->ctrl);
1893 nvme_start_queues(&dev->ctrl);
1897 clear_bit(NVME_CTRL_RESETTING, &dev->flags);
1901 nvme_remove_dead_ctrl(dev, result);
1904 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
1906 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
1907 struct pci_dev *pdev = to_pci_dev(dev->dev);
1909 nvme_kill_queues(&dev->ctrl);
1910 if (pci_get_drvdata(pdev))
1911 pci_stop_and_remove_bus_device_locked(pdev);
1912 nvme_put_ctrl(&dev->ctrl);
1915 static int nvme_reset(struct nvme_dev *dev)
1917 if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q))
1920 if (!queue_work(nvme_workq, &dev->reset_work))
1923 flush_work(&dev->reset_work);
1927 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
1929 *val = readl(to_nvme_dev(ctrl)->bar + off);
1933 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
1935 writel(val, to_nvme_dev(ctrl)->bar + off);
1939 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
1941 *val = readq(to_nvme_dev(ctrl)->bar + off);
1945 static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl)
1947 return nvme_reset(to_nvme_dev(ctrl));
1950 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
1951 .module = THIS_MODULE,
1952 .reg_read32 = nvme_pci_reg_read32,
1953 .reg_write32 = nvme_pci_reg_write32,
1954 .reg_read64 = nvme_pci_reg_read64,
1955 .reset_ctrl = nvme_pci_reset_ctrl,
1956 .free_ctrl = nvme_pci_free_ctrl,
1959 static int nvme_dev_map(struct nvme_dev *dev)
1962 struct pci_dev *pdev = to_pci_dev(dev->dev);
1964 bars = pci_select_bars(pdev, IORESOURCE_MEM);
1967 if (pci_request_selected_regions(pdev, bars, "nvme"))
1970 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
1976 pci_release_regions(pdev);
1980 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1982 int node, result = -ENOMEM;
1983 struct nvme_dev *dev;
1985 node = dev_to_node(&pdev->dev);
1986 if (node == NUMA_NO_NODE)
1987 set_dev_node(&pdev->dev, 0);
1989 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
1992 dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
1996 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2001 dev->dev = get_device(&pdev->dev);
2002 pci_set_drvdata(pdev, dev);
2004 result = nvme_dev_map(dev);
2008 INIT_WORK(&dev->scan_work, nvme_dev_scan);
2009 INIT_WORK(&dev->reset_work, nvme_reset_work);
2010 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2011 INIT_WORK(&dev->async_work, nvme_async_event_work);
2012 setup_timer(&dev->watchdog_timer, nvme_watchdog_timer,
2013 (unsigned long)dev);
2014 mutex_init(&dev->shutdown_lock);
2015 init_completion(&dev->ioq_wait);
2017 result = nvme_setup_prp_pools(dev);
2021 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2026 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2028 queue_work(nvme_workq, &dev->reset_work);
2032 nvme_release_prp_pools(dev);
2034 put_device(dev->dev);
2035 nvme_dev_unmap(dev);
2043 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2045 struct nvme_dev *dev = pci_get_drvdata(pdev);
2048 nvme_dev_disable(dev, false);
2050 queue_work(nvme_workq, &dev->reset_work);
2053 static void nvme_shutdown(struct pci_dev *pdev)
2055 struct nvme_dev *dev = pci_get_drvdata(pdev);
2056 nvme_dev_disable(dev, true);
2060 * The driver's remove may be called on a device in a partially initialized
2061 * state. This function must not have any dependencies on the device state in
2064 static void nvme_remove(struct pci_dev *pdev)
2066 struct nvme_dev *dev = pci_get_drvdata(pdev);
2068 del_timer_sync(&dev->watchdog_timer);
2070 set_bit(NVME_CTRL_REMOVING, &dev->flags);
2071 pci_set_drvdata(pdev, NULL);
2072 flush_work(&dev->async_work);
2073 flush_work(&dev->scan_work);
2074 nvme_remove_namespaces(&dev->ctrl);
2075 nvme_uninit_ctrl(&dev->ctrl);
2076 nvme_dev_disable(dev, true);
2077 flush_work(&dev->reset_work);
2078 nvme_dev_remove_admin(dev);
2079 nvme_free_queues(dev, 0);
2080 nvme_release_cmb(dev);
2081 nvme_release_prp_pools(dev);
2082 nvme_dev_unmap(dev);
2083 nvme_put_ctrl(&dev->ctrl);
2086 #ifdef CONFIG_PM_SLEEP
2087 static int nvme_suspend(struct device *dev)
2089 struct pci_dev *pdev = to_pci_dev(dev);
2090 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2092 nvme_dev_disable(ndev, true);
2096 static int nvme_resume(struct device *dev)
2098 struct pci_dev *pdev = to_pci_dev(dev);
2099 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2101 queue_work(nvme_workq, &ndev->reset_work);
2106 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2108 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2109 pci_channel_state_t state)
2111 struct nvme_dev *dev = pci_get_drvdata(pdev);
2114 * A frozen channel requires a reset. When detected, this method will
2115 * shutdown the controller to quiesce. The controller will be restarted
2116 * after the slot reset through driver's slot_reset callback.
2118 dev_warn(dev->ctrl.device, "error detected: state:%d\n", state);
2120 case pci_channel_io_normal:
2121 return PCI_ERS_RESULT_CAN_RECOVER;
2122 case pci_channel_io_frozen:
2123 nvme_dev_disable(dev, false);
2124 return PCI_ERS_RESULT_NEED_RESET;
2125 case pci_channel_io_perm_failure:
2126 return PCI_ERS_RESULT_DISCONNECT;
2128 return PCI_ERS_RESULT_NEED_RESET;
2131 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2133 struct nvme_dev *dev = pci_get_drvdata(pdev);
2135 dev_info(dev->ctrl.device, "restart after slot reset\n");
2136 pci_restore_state(pdev);
2137 queue_work(nvme_workq, &dev->reset_work);
2138 return PCI_ERS_RESULT_RECOVERED;
2141 static void nvme_error_resume(struct pci_dev *pdev)
2143 pci_cleanup_aer_uncorrect_error_status(pdev);
2146 static const struct pci_error_handlers nvme_err_handler = {
2147 .error_detected = nvme_error_detected,
2148 .slot_reset = nvme_slot_reset,
2149 .resume = nvme_error_resume,
2150 .reset_notify = nvme_reset_notify,
2153 /* Move to pci_ids.h later */
2154 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
2156 static const struct pci_device_id nvme_id_table[] = {
2157 { PCI_VDEVICE(INTEL, 0x0953),
2158 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2159 NVME_QUIRK_DISCARD_ZEROES, },
2160 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
2161 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2162 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2163 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2166 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2168 static struct pci_driver nvme_driver = {
2170 .id_table = nvme_id_table,
2171 .probe = nvme_probe,
2172 .remove = nvme_remove,
2173 .shutdown = nvme_shutdown,
2175 .pm = &nvme_dev_pm_ops,
2177 .err_handler = &nvme_err_handler,
2180 static int __init nvme_init(void)
2184 nvme_workq = alloc_workqueue("nvme", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
2188 result = pci_register_driver(&nvme_driver);
2190 destroy_workqueue(nvme_workq);
2194 static void __exit nvme_exit(void)
2196 pci_unregister_driver(&nvme_driver);
2197 destroy_workqueue(nvme_workq);
2201 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2202 MODULE_LICENSE("GPL");
2203 MODULE_VERSION("1.0");
2204 module_init(nvme_init);
2205 module_exit(nvme_exit);