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 remove_work;
96 struct work_struct async_work;
97 struct timer_list watchdog_timer;
98 struct mutex shutdown_lock;
101 dma_addr_t cmb_dma_addr;
104 struct nvme_ctrl ctrl;
105 struct completion ioq_wait;
108 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
110 return container_of(ctrl, struct nvme_dev, ctrl);
114 * An NVM Express queue. Each device has at least two (one for admin
115 * commands and one for I/O commands).
118 struct device *q_dmadev;
119 struct nvme_dev *dev;
120 char irqname[24]; /* nvme4294967295-65535\0 */
122 struct nvme_command *sq_cmds;
123 struct nvme_command __iomem *sq_cmds_io;
124 volatile struct nvme_completion *cqes;
125 struct blk_mq_tags **tags;
126 dma_addr_t sq_dma_addr;
127 dma_addr_t cq_dma_addr;
139 * The nvme_iod describes the data in an I/O, including the list of PRP
140 * entries. You can't see it in this data structure because C doesn't let
141 * me express that. Use nvme_init_iod to ensure there's enough space
142 * allocated to store the PRP list.
145 struct nvme_queue *nvmeq;
147 int npages; /* In the PRP list. 0 means small pool in use */
148 int nents; /* Used in scatterlist */
149 int length; /* Of data, in bytes */
150 dma_addr_t first_dma;
151 struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
152 struct scatterlist *sg;
153 struct scatterlist inline_sg[0];
157 * Check we didin't inadvertently grow the command struct
159 static inline void _nvme_check_size(void)
161 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
162 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
163 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
164 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
165 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
166 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
167 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
168 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
169 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
170 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
171 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
172 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
176 * Max size of iod being embedded in the request payload
178 #define NVME_INT_PAGES 2
179 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size)
182 * Will slightly overestimate the number of pages needed. This is OK
183 * as it only leads to a small amount of wasted memory for the lifetime of
186 static int nvme_npages(unsigned size, struct nvme_dev *dev)
188 unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
189 dev->ctrl.page_size);
190 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
193 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
194 unsigned int size, unsigned int nseg)
196 return sizeof(__le64 *) * nvme_npages(size, dev) +
197 sizeof(struct scatterlist) * nseg;
200 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
202 return sizeof(struct nvme_iod) +
203 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
206 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
207 unsigned int hctx_idx)
209 struct nvme_dev *dev = data;
210 struct nvme_queue *nvmeq = dev->queues[0];
212 WARN_ON(hctx_idx != 0);
213 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
214 WARN_ON(nvmeq->tags);
216 hctx->driver_data = nvmeq;
217 nvmeq->tags = &dev->admin_tagset.tags[0];
221 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
223 struct nvme_queue *nvmeq = hctx->driver_data;
228 static int nvme_admin_init_request(void *data, struct request *req,
229 unsigned int hctx_idx, unsigned int rq_idx,
230 unsigned int numa_node)
232 struct nvme_dev *dev = data;
233 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
234 struct nvme_queue *nvmeq = dev->queues[0];
241 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
242 unsigned int hctx_idx)
244 struct nvme_dev *dev = data;
245 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
248 nvmeq->tags = &dev->tagset.tags[hctx_idx];
250 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
251 hctx->driver_data = nvmeq;
255 static int nvme_init_request(void *data, struct request *req,
256 unsigned int hctx_idx, unsigned int rq_idx,
257 unsigned int numa_node)
259 struct nvme_dev *dev = data;
260 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
261 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
268 static void nvme_complete_async_event(struct nvme_dev *dev,
269 struct nvme_completion *cqe)
271 u16 status = le16_to_cpu(cqe->status) >> 1;
272 u32 result = le32_to_cpu(cqe->result);
274 if (status == NVME_SC_SUCCESS || status == NVME_SC_ABORT_REQ) {
275 ++dev->ctrl.event_limit;
276 queue_work(nvme_workq, &dev->async_work);
279 if (status != NVME_SC_SUCCESS)
282 switch (result & 0xff07) {
283 case NVME_AER_NOTICE_NS_CHANGED:
284 dev_info(dev->ctrl.device, "rescanning\n");
285 nvme_queue_scan(&dev->ctrl);
287 dev_warn(dev->ctrl.device, "async event result %08x\n", result);
292 * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
293 * @nvmeq: The queue to use
294 * @cmd: The command to send
296 * Safe to use from interrupt context
298 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
299 struct nvme_command *cmd)
301 u16 tail = nvmeq->sq_tail;
303 if (nvmeq->sq_cmds_io)
304 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
306 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
308 if (++tail == nvmeq->q_depth)
310 writel(tail, nvmeq->q_db);
311 nvmeq->sq_tail = tail;
314 static __le64 **iod_list(struct request *req)
316 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
317 return (__le64 **)(iod->sg + req->nr_phys_segments);
320 static int nvme_init_iod(struct request *rq, unsigned size,
321 struct nvme_dev *dev)
323 struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
324 int nseg = rq->nr_phys_segments;
326 if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
327 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
329 return BLK_MQ_RQ_QUEUE_BUSY;
331 iod->sg = iod->inline_sg;
341 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
343 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
344 const int last_prp = dev->ctrl.page_size / 8 - 1;
346 __le64 **list = iod_list(req);
347 dma_addr_t prp_dma = iod->first_dma;
349 if (req->cmd_flags & REQ_DISCARD)
350 kfree(req->completion_data);
352 if (iod->npages == 0)
353 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
354 for (i = 0; i < iod->npages; i++) {
355 __le64 *prp_list = list[i];
356 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
357 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
358 prp_dma = next_prp_dma;
361 if (iod->sg != iod->inline_sg)
365 #ifdef CONFIG_BLK_DEV_INTEGRITY
366 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
368 if (be32_to_cpu(pi->ref_tag) == v)
369 pi->ref_tag = cpu_to_be32(p);
372 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
374 if (be32_to_cpu(pi->ref_tag) == p)
375 pi->ref_tag = cpu_to_be32(v);
379 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
381 * The virtual start sector is the one that was originally submitted by the
382 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
383 * start sector may be different. Remap protection information to match the
384 * physical LBA on writes, and back to the original seed on reads.
386 * Type 0 and 3 do not have a ref tag, so no remapping required.
388 static void nvme_dif_remap(struct request *req,
389 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
391 struct nvme_ns *ns = req->rq_disk->private_data;
392 struct bio_integrity_payload *bip;
393 struct t10_pi_tuple *pi;
395 u32 i, nlb, ts, phys, virt;
397 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
400 bip = bio_integrity(req->bio);
404 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
407 virt = bip_get_seed(bip);
408 phys = nvme_block_nr(ns, blk_rq_pos(req));
409 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
410 ts = ns->disk->queue->integrity.tuple_size;
412 for (i = 0; i < nlb; i++, virt++, phys++) {
413 pi = (struct t10_pi_tuple *)p;
414 dif_swap(phys, virt, pi);
419 #else /* CONFIG_BLK_DEV_INTEGRITY */
420 static void nvme_dif_remap(struct request *req,
421 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
424 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
427 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
432 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req,
435 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
436 struct dma_pool *pool;
437 int length = total_len;
438 struct scatterlist *sg = iod->sg;
439 int dma_len = sg_dma_len(sg);
440 u64 dma_addr = sg_dma_address(sg);
441 u32 page_size = dev->ctrl.page_size;
442 int offset = dma_addr & (page_size - 1);
444 __le64 **list = iod_list(req);
448 length -= (page_size - offset);
452 dma_len -= (page_size - offset);
454 dma_addr += (page_size - offset);
457 dma_addr = sg_dma_address(sg);
458 dma_len = sg_dma_len(sg);
461 if (length <= page_size) {
462 iod->first_dma = dma_addr;
466 nprps = DIV_ROUND_UP(length, page_size);
467 if (nprps <= (256 / 8)) {
468 pool = dev->prp_small_pool;
471 pool = dev->prp_page_pool;
475 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
477 iod->first_dma = dma_addr;
482 iod->first_dma = prp_dma;
485 if (i == page_size >> 3) {
486 __le64 *old_prp_list = prp_list;
487 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
490 list[iod->npages++] = prp_list;
491 prp_list[0] = old_prp_list[i - 1];
492 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
495 prp_list[i++] = cpu_to_le64(dma_addr);
496 dma_len -= page_size;
497 dma_addr += page_size;
505 dma_addr = sg_dma_address(sg);
506 dma_len = sg_dma_len(sg);
512 static int nvme_map_data(struct nvme_dev *dev, struct request *req,
513 unsigned size, struct nvme_command *cmnd)
515 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
516 struct request_queue *q = req->q;
517 enum dma_data_direction dma_dir = rq_data_dir(req) ?
518 DMA_TO_DEVICE : DMA_FROM_DEVICE;
519 int ret = BLK_MQ_RQ_QUEUE_ERROR;
521 sg_init_table(iod->sg, req->nr_phys_segments);
522 iod->nents = blk_rq_map_sg(q, req, iod->sg);
526 ret = BLK_MQ_RQ_QUEUE_BUSY;
527 if (!dma_map_sg(dev->dev, iod->sg, iod->nents, dma_dir))
530 if (!nvme_setup_prps(dev, req, size))
533 ret = BLK_MQ_RQ_QUEUE_ERROR;
534 if (blk_integrity_rq(req)) {
535 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
538 sg_init_table(&iod->meta_sg, 1);
539 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
542 if (rq_data_dir(req))
543 nvme_dif_remap(req, nvme_dif_prep);
545 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
549 cmnd->rw.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
550 cmnd->rw.prp2 = cpu_to_le64(iod->first_dma);
551 if (blk_integrity_rq(req))
552 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
553 return BLK_MQ_RQ_QUEUE_OK;
556 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
561 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
563 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
564 enum dma_data_direction dma_dir = rq_data_dir(req) ?
565 DMA_TO_DEVICE : DMA_FROM_DEVICE;
568 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
569 if (blk_integrity_rq(req)) {
570 if (!rq_data_dir(req))
571 nvme_dif_remap(req, nvme_dif_complete);
572 dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
576 nvme_free_iod(dev, req);
580 * NOTE: ns is NULL when called on the admin queue.
582 static int nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
583 const struct blk_mq_queue_data *bd)
585 struct nvme_ns *ns = hctx->queue->queuedata;
586 struct nvme_queue *nvmeq = hctx->driver_data;
587 struct nvme_dev *dev = nvmeq->dev;
588 struct request *req = bd->rq;
589 struct nvme_command cmnd;
591 int ret = BLK_MQ_RQ_QUEUE_OK;
594 * If formated with metadata, require the block layer provide a buffer
595 * unless this namespace is formated such that the metadata can be
596 * stripped/generated by the controller with PRACT=1.
598 if (ns && ns->ms && !blk_integrity_rq(req)) {
599 if (!(ns->pi_type && ns->ms == 8) &&
600 req->cmd_type != REQ_TYPE_DRV_PRIV) {
601 blk_mq_end_request(req, -EFAULT);
602 return BLK_MQ_RQ_QUEUE_OK;
606 map_len = nvme_map_len(req);
607 ret = nvme_init_iod(req, map_len, dev);
611 ret = nvme_setup_cmd(ns, req, &cmnd);
615 if (req->nr_phys_segments)
616 ret = nvme_map_data(dev, req, map_len, &cmnd);
621 cmnd.common.command_id = req->tag;
622 blk_mq_start_request(req);
624 spin_lock_irq(&nvmeq->q_lock);
625 if (unlikely(nvmeq->cq_vector < 0)) {
626 if (ns && !test_bit(NVME_NS_DEAD, &ns->flags))
627 ret = BLK_MQ_RQ_QUEUE_BUSY;
629 ret = BLK_MQ_RQ_QUEUE_ERROR;
630 spin_unlock_irq(&nvmeq->q_lock);
633 __nvme_submit_cmd(nvmeq, &cmnd);
634 nvme_process_cq(nvmeq);
635 spin_unlock_irq(&nvmeq->q_lock);
636 return BLK_MQ_RQ_QUEUE_OK;
638 nvme_free_iod(dev, req);
642 static void nvme_complete_rq(struct request *req)
644 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
645 struct nvme_dev *dev = iod->nvmeq->dev;
648 nvme_unmap_data(dev, req);
650 if (unlikely(req->errors)) {
651 if (nvme_req_needs_retry(req, req->errors)) {
652 nvme_requeue_req(req);
656 if (req->cmd_type == REQ_TYPE_DRV_PRIV)
659 error = nvme_error_status(req->errors);
662 if (unlikely(iod->aborted)) {
663 dev_warn(dev->ctrl.device,
664 "completing aborted command with status: %04x\n",
668 blk_mq_end_request(req, error);
671 /* We read the CQE phase first to check if the rest of the entry is valid */
672 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
675 return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
678 static void __nvme_process_cq(struct nvme_queue *nvmeq, unsigned int *tag)
682 head = nvmeq->cq_head;
683 phase = nvmeq->cq_phase;
685 while (nvme_cqe_valid(nvmeq, head, phase)) {
686 struct nvme_completion cqe = nvmeq->cqes[head];
689 if (++head == nvmeq->q_depth) {
694 if (tag && *tag == cqe.command_id)
697 if (unlikely(cqe.command_id >= nvmeq->q_depth)) {
698 dev_warn(nvmeq->dev->ctrl.device,
699 "invalid id %d completed on queue %d\n",
700 cqe.command_id, le16_to_cpu(cqe.sq_id));
705 * AEN requests are special as they don't time out and can
706 * survive any kind of queue freeze and often don't respond to
707 * aborts. We don't even bother to allocate a struct request
708 * for them but rather special case them here.
710 if (unlikely(nvmeq->qid == 0 &&
711 cqe.command_id >= NVME_AQ_BLKMQ_DEPTH)) {
712 nvme_complete_async_event(nvmeq->dev, &cqe);
716 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe.command_id);
717 if (req->cmd_type == REQ_TYPE_DRV_PRIV && req->special)
718 memcpy(req->special, &cqe, sizeof(cqe));
719 blk_mq_complete_request(req, le16_to_cpu(cqe.status) >> 1);
723 /* If the controller ignores the cq head doorbell and continuously
724 * writes to the queue, it is theoretically possible to wrap around
725 * the queue twice and mistakenly return IRQ_NONE. Linux only
726 * requires that 0.1% of your interrupts are handled, so this isn't
729 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
732 if (likely(nvmeq->cq_vector >= 0))
733 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
734 nvmeq->cq_head = head;
735 nvmeq->cq_phase = phase;
740 static void nvme_process_cq(struct nvme_queue *nvmeq)
742 __nvme_process_cq(nvmeq, NULL);
745 static irqreturn_t nvme_irq(int irq, void *data)
748 struct nvme_queue *nvmeq = data;
749 spin_lock(&nvmeq->q_lock);
750 nvme_process_cq(nvmeq);
751 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
753 spin_unlock(&nvmeq->q_lock);
757 static irqreturn_t nvme_irq_check(int irq, void *data)
759 struct nvme_queue *nvmeq = data;
760 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
761 return IRQ_WAKE_THREAD;
765 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
767 struct nvme_queue *nvmeq = hctx->driver_data;
769 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
770 spin_lock_irq(&nvmeq->q_lock);
771 __nvme_process_cq(nvmeq, &tag);
772 spin_unlock_irq(&nvmeq->q_lock);
781 static void nvme_async_event_work(struct work_struct *work)
783 struct nvme_dev *dev = container_of(work, struct nvme_dev, async_work);
784 struct nvme_queue *nvmeq = dev->queues[0];
785 struct nvme_command c;
787 memset(&c, 0, sizeof(c));
788 c.common.opcode = nvme_admin_async_event;
790 spin_lock_irq(&nvmeq->q_lock);
791 while (dev->ctrl.event_limit > 0) {
792 c.common.command_id = NVME_AQ_BLKMQ_DEPTH +
793 --dev->ctrl.event_limit;
794 __nvme_submit_cmd(nvmeq, &c);
796 spin_unlock_irq(&nvmeq->q_lock);
799 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
801 struct nvme_command c;
803 memset(&c, 0, sizeof(c));
804 c.delete_queue.opcode = opcode;
805 c.delete_queue.qid = cpu_to_le16(id);
807 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
810 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
811 struct nvme_queue *nvmeq)
813 struct nvme_command c;
814 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
817 * Note: we (ab)use the fact the the prp fields survive if no data
818 * is attached to the request.
820 memset(&c, 0, sizeof(c));
821 c.create_cq.opcode = nvme_admin_create_cq;
822 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
823 c.create_cq.cqid = cpu_to_le16(qid);
824 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
825 c.create_cq.cq_flags = cpu_to_le16(flags);
826 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
828 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
831 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
832 struct nvme_queue *nvmeq)
834 struct nvme_command c;
835 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
838 * Note: we (ab)use the fact the the prp fields survive if no data
839 * is attached to the request.
841 memset(&c, 0, sizeof(c));
842 c.create_sq.opcode = nvme_admin_create_sq;
843 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
844 c.create_sq.sqid = cpu_to_le16(qid);
845 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
846 c.create_sq.sq_flags = cpu_to_le16(flags);
847 c.create_sq.cqid = cpu_to_le16(qid);
849 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
852 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
854 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
857 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
859 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
862 static void abort_endio(struct request *req, int error)
864 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
865 struct nvme_queue *nvmeq = iod->nvmeq;
866 u16 status = req->errors;
868 dev_warn(nvmeq->dev->ctrl.device, "Abort status: 0x%x", status);
869 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
870 blk_mq_free_request(req);
873 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
875 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
876 struct nvme_queue *nvmeq = iod->nvmeq;
877 struct nvme_dev *dev = nvmeq->dev;
878 struct request *abort_req;
879 struct nvme_command cmd;
882 * Shutdown immediately if controller times out while starting. The
883 * reset work will see the pci device disabled when it gets the forced
884 * cancellation error. All outstanding requests are completed on
885 * shutdown, so we return BLK_EH_HANDLED.
887 if (dev->ctrl.state == NVME_CTRL_RESETTING) {
888 dev_warn(dev->ctrl.device,
889 "I/O %d QID %d timeout, disable controller\n",
890 req->tag, nvmeq->qid);
891 nvme_dev_disable(dev, false);
892 req->errors = NVME_SC_CANCELLED;
893 return BLK_EH_HANDLED;
897 * Shutdown the controller immediately and schedule a reset if the
898 * command was already aborted once before and still hasn't been
899 * returned to the driver, or if this is the admin queue.
901 if (!nvmeq->qid || iod->aborted) {
902 dev_warn(dev->ctrl.device,
903 "I/O %d QID %d timeout, reset controller\n",
904 req->tag, nvmeq->qid);
905 nvme_dev_disable(dev, false);
906 queue_work(nvme_workq, &dev->reset_work);
909 * Mark the request as handled, since the inline shutdown
910 * forces all outstanding requests to complete.
912 req->errors = NVME_SC_CANCELLED;
913 return BLK_EH_HANDLED;
918 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
919 atomic_inc(&dev->ctrl.abort_limit);
920 return BLK_EH_RESET_TIMER;
923 memset(&cmd, 0, sizeof(cmd));
924 cmd.abort.opcode = nvme_admin_abort_cmd;
925 cmd.abort.cid = req->tag;
926 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
928 dev_warn(nvmeq->dev->ctrl.device,
929 "I/O %d QID %d timeout, aborting\n",
930 req->tag, nvmeq->qid);
932 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
934 if (IS_ERR(abort_req)) {
935 atomic_inc(&dev->ctrl.abort_limit);
936 return BLK_EH_RESET_TIMER;
939 abort_req->timeout = ADMIN_TIMEOUT;
940 abort_req->end_io_data = NULL;
941 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
944 * The aborted req will be completed on receiving the abort req.
945 * We enable the timer again. If hit twice, it'll cause a device reset,
946 * as the device then is in a faulty state.
948 return BLK_EH_RESET_TIMER;
951 static void nvme_cancel_io(struct request *req, void *data, bool reserved)
955 if (!blk_mq_request_started(req))
958 dev_dbg_ratelimited(((struct nvme_dev *) data)->ctrl.device,
959 "Cancelling I/O %d", req->tag);
961 status = NVME_SC_ABORT_REQ;
962 if (blk_queue_dying(req->q))
963 status |= NVME_SC_DNR;
964 blk_mq_complete_request(req, status);
967 static void nvme_free_queue(struct nvme_queue *nvmeq)
969 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
970 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
972 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
973 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
977 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
981 for (i = dev->queue_count - 1; i >= lowest; i--) {
982 struct nvme_queue *nvmeq = dev->queues[i];
984 dev->queues[i] = NULL;
985 nvme_free_queue(nvmeq);
990 * nvme_suspend_queue - put queue into suspended state
991 * @nvmeq - queue to suspend
993 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
997 spin_lock_irq(&nvmeq->q_lock);
998 if (nvmeq->cq_vector == -1) {
999 spin_unlock_irq(&nvmeq->q_lock);
1002 vector = nvmeq->dev->entry[nvmeq->cq_vector].vector;
1003 nvmeq->dev->online_queues--;
1004 nvmeq->cq_vector = -1;
1005 spin_unlock_irq(&nvmeq->q_lock);
1007 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1008 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1010 irq_set_affinity_hint(vector, NULL);
1011 free_irq(vector, nvmeq);
1016 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1018 struct nvme_queue *nvmeq = dev->queues[0];
1022 if (nvme_suspend_queue(nvmeq))
1026 nvme_shutdown_ctrl(&dev->ctrl);
1028 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1029 dev->bar + NVME_REG_CAP));
1031 spin_lock_irq(&nvmeq->q_lock);
1032 nvme_process_cq(nvmeq);
1033 spin_unlock_irq(&nvmeq->q_lock);
1036 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1039 int q_depth = dev->q_depth;
1040 unsigned q_size_aligned = roundup(q_depth * entry_size,
1041 dev->ctrl.page_size);
1043 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1044 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1045 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1046 q_depth = div_u64(mem_per_q, entry_size);
1049 * Ensure the reduced q_depth is above some threshold where it
1050 * would be better to map queues in system memory with the
1060 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1063 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1064 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1065 dev->ctrl.page_size);
1066 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1067 nvmeq->sq_cmds_io = dev->cmb + offset;
1069 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1070 &nvmeq->sq_dma_addr, GFP_KERNEL);
1071 if (!nvmeq->sq_cmds)
1078 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1081 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq), GFP_KERNEL);
1085 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1086 &nvmeq->cq_dma_addr, GFP_KERNEL);
1090 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1093 nvmeq->q_dmadev = dev->dev;
1095 snprintf(nvmeq->irqname, sizeof(nvmeq->irqname), "nvme%dq%d",
1096 dev->ctrl.instance, qid);
1097 spin_lock_init(&nvmeq->q_lock);
1099 nvmeq->cq_phase = 1;
1100 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1101 nvmeq->q_depth = depth;
1103 nvmeq->cq_vector = -1;
1104 dev->queues[qid] = nvmeq;
1110 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1111 nvmeq->cq_dma_addr);
1117 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1120 if (use_threaded_interrupts)
1121 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
1122 nvme_irq_check, nvme_irq, IRQF_SHARED,
1124 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
1125 IRQF_SHARED, name, nvmeq);
1128 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1130 struct nvme_dev *dev = nvmeq->dev;
1132 spin_lock_irq(&nvmeq->q_lock);
1135 nvmeq->cq_phase = 1;
1136 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1137 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1138 dev->online_queues++;
1139 spin_unlock_irq(&nvmeq->q_lock);
1142 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1144 struct nvme_dev *dev = nvmeq->dev;
1147 nvmeq->cq_vector = qid - 1;
1148 result = adapter_alloc_cq(dev, qid, nvmeq);
1152 result = adapter_alloc_sq(dev, qid, nvmeq);
1156 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1160 nvme_init_queue(nvmeq, qid);
1164 adapter_delete_sq(dev, qid);
1166 adapter_delete_cq(dev, qid);
1170 static struct blk_mq_ops nvme_mq_admin_ops = {
1171 .queue_rq = nvme_queue_rq,
1172 .complete = nvme_complete_rq,
1173 .map_queue = blk_mq_map_queue,
1174 .init_hctx = nvme_admin_init_hctx,
1175 .exit_hctx = nvme_admin_exit_hctx,
1176 .init_request = nvme_admin_init_request,
1177 .timeout = nvme_timeout,
1180 static struct blk_mq_ops nvme_mq_ops = {
1181 .queue_rq = nvme_queue_rq,
1182 .complete = nvme_complete_rq,
1183 .map_queue = blk_mq_map_queue,
1184 .init_hctx = nvme_init_hctx,
1185 .init_request = nvme_init_request,
1186 .timeout = nvme_timeout,
1190 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1192 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1194 * If the controller was reset during removal, it's possible
1195 * user requests may be waiting on a stopped queue. Start the
1196 * queue to flush these to completion.
1198 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1199 blk_cleanup_queue(dev->ctrl.admin_q);
1200 blk_mq_free_tag_set(&dev->admin_tagset);
1204 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1206 if (!dev->ctrl.admin_q) {
1207 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1208 dev->admin_tagset.nr_hw_queues = 1;
1211 * Subtract one to leave an empty queue entry for 'Full Queue'
1212 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1214 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1215 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1216 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1217 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1218 dev->admin_tagset.driver_data = dev;
1220 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1223 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1224 if (IS_ERR(dev->ctrl.admin_q)) {
1225 blk_mq_free_tag_set(&dev->admin_tagset);
1228 if (!blk_get_queue(dev->ctrl.admin_q)) {
1229 nvme_dev_remove_admin(dev);
1230 dev->ctrl.admin_q = NULL;
1234 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1239 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1243 u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1244 struct nvme_queue *nvmeq;
1246 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1) ?
1247 NVME_CAP_NSSRC(cap) : 0;
1249 if (dev->subsystem &&
1250 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1251 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1253 result = nvme_disable_ctrl(&dev->ctrl, cap);
1257 nvmeq = dev->queues[0];
1259 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH);
1264 aqa = nvmeq->q_depth - 1;
1267 writel(aqa, dev->bar + NVME_REG_AQA);
1268 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1269 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1271 result = nvme_enable_ctrl(&dev->ctrl, cap);
1275 nvmeq->cq_vector = 0;
1276 result = queue_request_irq(dev, nvmeq, nvmeq->irqname);
1278 nvmeq->cq_vector = -1;
1285 nvme_free_queues(dev, 0);
1289 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
1292 /* If true, indicates loss of adapter communication, possibly by a
1293 * NVMe Subsystem reset.
1295 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
1297 /* If there is a reset ongoing, we shouldn't reset again. */
1298 if (work_busy(&dev->reset_work))
1301 /* We shouldn't reset unless the controller is on fatal error state
1302 * _or_ if we lost the communication with it.
1304 if (!(csts & NVME_CSTS_CFS) && !nssro)
1307 /* If PCI error recovery process is happening, we cannot reset or
1308 * the recovery mechanism will surely fail.
1310 if (pci_channel_offline(to_pci_dev(dev->dev)))
1316 static void nvme_watchdog_timer(unsigned long data)
1318 struct nvme_dev *dev = (struct nvme_dev *)data;
1319 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1321 /* Skip controllers under certain specific conditions. */
1322 if (nvme_should_reset(dev, csts)) {
1323 if (queue_work(nvme_workq, &dev->reset_work))
1325 "Failed status: 0x%x, reset controller.\n",
1330 mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1333 static int nvme_create_io_queues(struct nvme_dev *dev)
1338 for (i = dev->queue_count; i <= dev->max_qid; i++) {
1339 if (!nvme_alloc_queue(dev, i, dev->q_depth)) {
1345 max = min(dev->max_qid, dev->queue_count - 1);
1346 for (i = dev->online_queues; i <= max; i++) {
1347 ret = nvme_create_queue(dev->queues[i], i);
1349 nvme_free_queues(dev, i);
1355 * Ignore failing Create SQ/CQ commands, we can continue with less
1356 * than the desired aount of queues, and even a controller without
1357 * I/O queues an still be used to issue admin commands. This might
1358 * be useful to upgrade a buggy firmware for example.
1360 return ret >= 0 ? 0 : ret;
1363 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1365 u64 szu, size, offset;
1367 resource_size_t bar_size;
1368 struct pci_dev *pdev = to_pci_dev(dev->dev);
1370 dma_addr_t dma_addr;
1375 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1376 if (!(NVME_CMB_SZ(dev->cmbsz)))
1379 cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1381 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1382 size = szu * NVME_CMB_SZ(dev->cmbsz);
1383 offset = szu * NVME_CMB_OFST(cmbloc);
1384 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(cmbloc));
1386 if (offset > bar_size)
1390 * Controllers may support a CMB size larger than their BAR,
1391 * for example, due to being behind a bridge. Reduce the CMB to
1392 * the reported size of the BAR
1394 if (size > bar_size - offset)
1395 size = bar_size - offset;
1397 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(cmbloc)) + offset;
1398 cmb = ioremap_wc(dma_addr, size);
1402 dev->cmb_dma_addr = dma_addr;
1403 dev->cmb_size = size;
1407 static inline void nvme_release_cmb(struct nvme_dev *dev)
1415 static size_t db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1417 return 4096 + ((nr_io_queues + 1) * 8 * dev->db_stride);
1420 static int nvme_setup_io_queues(struct nvme_dev *dev)
1422 struct nvme_queue *adminq = dev->queues[0];
1423 struct pci_dev *pdev = to_pci_dev(dev->dev);
1424 int result, i, vecs, nr_io_queues, size;
1426 nr_io_queues = num_possible_cpus();
1427 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1432 * Degraded controllers might return an error when setting the queue
1433 * count. We still want to be able to bring them online and offer
1434 * access to the admin queue, as that might be only way to fix them up.
1437 dev_err(dev->ctrl.device,
1438 "Could not set queue count (%d)\n", result);
1442 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1443 result = nvme_cmb_qdepth(dev, nr_io_queues,
1444 sizeof(struct nvme_command));
1446 dev->q_depth = result;
1448 nvme_release_cmb(dev);
1451 size = db_bar_size(dev, nr_io_queues);
1455 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1458 if (!--nr_io_queues)
1460 size = db_bar_size(dev, nr_io_queues);
1462 dev->dbs = dev->bar + 4096;
1463 adminq->q_db = dev->dbs;
1466 /* Deregister the admin queue's interrupt */
1467 free_irq(dev->entry[0].vector, adminq);
1470 * If we enable msix early due to not intx, disable it again before
1471 * setting up the full range we need.
1473 if (pdev->msi_enabled)
1474 pci_disable_msi(pdev);
1475 else if (pdev->msix_enabled)
1476 pci_disable_msix(pdev);
1478 for (i = 0; i < nr_io_queues; i++)
1479 dev->entry[i].entry = i;
1480 vecs = pci_enable_msix_range(pdev, dev->entry, 1, nr_io_queues);
1482 vecs = pci_enable_msi_range(pdev, 1, min(nr_io_queues, 32));
1486 for (i = 0; i < vecs; i++)
1487 dev->entry[i].vector = i + pdev->irq;
1492 * Should investigate if there's a performance win from allocating
1493 * more queues than interrupt vectors; it might allow the submission
1494 * path to scale better, even if the receive path is limited by the
1495 * number of interrupts.
1497 nr_io_queues = vecs;
1498 dev->max_qid = nr_io_queues;
1500 result = queue_request_irq(dev, adminq, adminq->irqname);
1502 adminq->cq_vector = -1;
1505 return nvme_create_io_queues(dev);
1508 nvme_free_queues(dev, 1);
1512 static void nvme_pci_post_scan(struct nvme_ctrl *ctrl)
1514 struct nvme_dev *dev = to_nvme_dev(ctrl);
1515 struct nvme_queue *nvmeq;
1518 for (i = 0; i < dev->online_queues; i++) {
1519 nvmeq = dev->queues[i];
1521 if (!nvmeq->tags || !(*nvmeq->tags))
1524 irq_set_affinity_hint(dev->entry[nvmeq->cq_vector].vector,
1525 blk_mq_tags_cpumask(*nvmeq->tags));
1529 static void nvme_del_queue_end(struct request *req, int error)
1531 struct nvme_queue *nvmeq = req->end_io_data;
1533 blk_mq_free_request(req);
1534 complete(&nvmeq->dev->ioq_wait);
1537 static void nvme_del_cq_end(struct request *req, int error)
1539 struct nvme_queue *nvmeq = req->end_io_data;
1542 unsigned long flags;
1545 * We might be called with the AQ q_lock held
1546 * and the I/O queue q_lock should always
1547 * nest inside the AQ one.
1549 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1550 SINGLE_DEPTH_NESTING);
1551 nvme_process_cq(nvmeq);
1552 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1555 nvme_del_queue_end(req, error);
1558 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1560 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1561 struct request *req;
1562 struct nvme_command cmd;
1564 memset(&cmd, 0, sizeof(cmd));
1565 cmd.delete_queue.opcode = opcode;
1566 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1568 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT);
1570 return PTR_ERR(req);
1572 req->timeout = ADMIN_TIMEOUT;
1573 req->end_io_data = nvmeq;
1575 blk_execute_rq_nowait(q, NULL, req, false,
1576 opcode == nvme_admin_delete_cq ?
1577 nvme_del_cq_end : nvme_del_queue_end);
1581 static void nvme_disable_io_queues(struct nvme_dev *dev)
1584 unsigned long timeout;
1585 u8 opcode = nvme_admin_delete_sq;
1587 for (pass = 0; pass < 2; pass++) {
1588 int sent = 0, i = dev->queue_count - 1;
1590 reinit_completion(&dev->ioq_wait);
1592 timeout = ADMIN_TIMEOUT;
1593 for (; i > 0; i--) {
1594 struct nvme_queue *nvmeq = dev->queues[i];
1597 nvme_suspend_queue(nvmeq);
1598 if (nvme_delete_queue(nvmeq, opcode))
1603 timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1609 opcode = nvme_admin_delete_cq;
1614 * Return: error value if an error occurred setting up the queues or calling
1615 * Identify Device. 0 if these succeeded, even if adding some of the
1616 * namespaces failed. At the moment, these failures are silent. TBD which
1617 * failures should be reported.
1619 static int nvme_dev_add(struct nvme_dev *dev)
1621 if (!dev->ctrl.tagset) {
1622 dev->tagset.ops = &nvme_mq_ops;
1623 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1624 dev->tagset.timeout = NVME_IO_TIMEOUT;
1625 dev->tagset.numa_node = dev_to_node(dev->dev);
1626 dev->tagset.queue_depth =
1627 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1628 dev->tagset.cmd_size = nvme_cmd_size(dev);
1629 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1630 dev->tagset.driver_data = dev;
1632 if (blk_mq_alloc_tag_set(&dev->tagset))
1634 dev->ctrl.tagset = &dev->tagset;
1636 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1638 /* Free previously allocated queues that are no longer usable */
1639 nvme_free_queues(dev, dev->online_queues);
1645 static int nvme_pci_enable(struct nvme_dev *dev)
1648 int result = -ENOMEM;
1649 struct pci_dev *pdev = to_pci_dev(dev->dev);
1651 if (pci_enable_device_mem(pdev))
1654 pci_set_master(pdev);
1656 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1657 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1660 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1666 * Some devices and/or platforms don't advertise or work with INTx
1667 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1668 * adjust this later.
1670 if (pci_enable_msix(pdev, dev->entry, 1)) {
1671 pci_enable_msi(pdev);
1672 dev->entry[0].vector = pdev->irq;
1675 if (!dev->entry[0].vector) {
1680 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1682 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1683 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1684 dev->dbs = dev->bar + 4096;
1687 * Temporary fix for the Apple controller found in the MacBook8,1 and
1688 * some MacBook7,1 to avoid controller resets and data loss.
1690 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1692 dev_warn(dev->dev, "detected Apple NVMe controller, set "
1693 "queue depth=%u to work around controller resets\n",
1697 if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2))
1698 dev->cmb = nvme_map_cmb(dev);
1700 pci_enable_pcie_error_reporting(pdev);
1701 pci_save_state(pdev);
1705 pci_disable_device(pdev);
1709 static void nvme_dev_unmap(struct nvme_dev *dev)
1713 pci_release_regions(to_pci_dev(dev->dev));
1716 static void nvme_pci_disable(struct nvme_dev *dev)
1718 struct pci_dev *pdev = to_pci_dev(dev->dev);
1720 if (pdev->msi_enabled)
1721 pci_disable_msi(pdev);
1722 else if (pdev->msix_enabled)
1723 pci_disable_msix(pdev);
1725 if (pci_is_enabled(pdev)) {
1726 pci_disable_pcie_error_reporting(pdev);
1727 pci_disable_device(pdev);
1731 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1736 del_timer_sync(&dev->watchdog_timer);
1738 mutex_lock(&dev->shutdown_lock);
1739 if (pci_is_enabled(to_pci_dev(dev->dev))) {
1740 nvme_stop_queues(&dev->ctrl);
1741 csts = readl(dev->bar + NVME_REG_CSTS);
1743 if (csts & NVME_CSTS_CFS || !(csts & NVME_CSTS_RDY)) {
1744 for (i = dev->queue_count - 1; i >= 0; i--) {
1745 struct nvme_queue *nvmeq = dev->queues[i];
1746 nvme_suspend_queue(nvmeq);
1749 nvme_disable_io_queues(dev);
1750 nvme_disable_admin_queue(dev, shutdown);
1752 nvme_pci_disable(dev);
1754 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_io, dev);
1755 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_io, dev);
1756 mutex_unlock(&dev->shutdown_lock);
1759 static int nvme_setup_prp_pools(struct nvme_dev *dev)
1761 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
1762 PAGE_SIZE, PAGE_SIZE, 0);
1763 if (!dev->prp_page_pool)
1766 /* Optimisation for I/Os between 4k and 128k */
1767 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
1769 if (!dev->prp_small_pool) {
1770 dma_pool_destroy(dev->prp_page_pool);
1776 static void nvme_release_prp_pools(struct nvme_dev *dev)
1778 dma_pool_destroy(dev->prp_page_pool);
1779 dma_pool_destroy(dev->prp_small_pool);
1782 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
1784 struct nvme_dev *dev = to_nvme_dev(ctrl);
1786 put_device(dev->dev);
1787 if (dev->tagset.tags)
1788 blk_mq_free_tag_set(&dev->tagset);
1789 if (dev->ctrl.admin_q)
1790 blk_put_queue(dev->ctrl.admin_q);
1796 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
1798 dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
1800 kref_get(&dev->ctrl.kref);
1801 nvme_dev_disable(dev, false);
1802 if (!schedule_work(&dev->remove_work))
1803 nvme_put_ctrl(&dev->ctrl);
1806 static void nvme_reset_work(struct work_struct *work)
1808 struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
1809 int result = -ENODEV;
1811 if (WARN_ON(dev->ctrl.state == NVME_CTRL_RESETTING))
1815 * If we're called to reset a live controller first shut it down before
1818 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
1819 nvme_dev_disable(dev, false);
1821 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING))
1824 result = nvme_pci_enable(dev);
1828 result = nvme_configure_admin_queue(dev);
1832 nvme_init_queue(dev->queues[0], 0);
1833 result = nvme_alloc_admin_tags(dev);
1837 result = nvme_init_identify(&dev->ctrl);
1841 result = nvme_setup_io_queues(dev);
1846 * A controller that can not execute IO typically requires user
1847 * intervention to correct. For such degraded controllers, the driver
1848 * should not submit commands the user did not request, so skip
1849 * registering for asynchronous event notification on this condition.
1851 if (dev->online_queues > 1) {
1852 dev->ctrl.event_limit = NVME_NR_AEN_COMMANDS;
1853 queue_work(nvme_workq, &dev->async_work);
1856 mod_timer(&dev->watchdog_timer, round_jiffies(jiffies + HZ));
1859 * Keep the controller around but remove all namespaces if we don't have
1860 * any working I/O queue.
1862 if (dev->online_queues < 2) {
1863 dev_warn(dev->ctrl.device, "IO queues not created\n");
1864 nvme_kill_queues(&dev->ctrl);
1865 nvme_remove_namespaces(&dev->ctrl);
1867 nvme_start_queues(&dev->ctrl);
1871 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
1872 dev_warn(dev->ctrl.device, "failed to mark controller live\n");
1876 if (dev->online_queues > 1)
1877 nvme_queue_scan(&dev->ctrl);
1881 nvme_remove_dead_ctrl(dev, result);
1884 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
1886 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
1887 struct pci_dev *pdev = to_pci_dev(dev->dev);
1889 nvme_kill_queues(&dev->ctrl);
1890 if (pci_get_drvdata(pdev))
1891 pci_stop_and_remove_bus_device_locked(pdev);
1892 nvme_put_ctrl(&dev->ctrl);
1895 static int nvme_reset(struct nvme_dev *dev)
1897 if (!dev->ctrl.admin_q || blk_queue_dying(dev->ctrl.admin_q))
1900 if (!queue_work(nvme_workq, &dev->reset_work))
1903 flush_work(&dev->reset_work);
1907 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
1909 *val = readl(to_nvme_dev(ctrl)->bar + off);
1913 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
1915 writel(val, to_nvme_dev(ctrl)->bar + off);
1919 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
1921 *val = readq(to_nvme_dev(ctrl)->bar + off);
1925 static int nvme_pci_reset_ctrl(struct nvme_ctrl *ctrl)
1927 return nvme_reset(to_nvme_dev(ctrl));
1930 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
1931 .module = THIS_MODULE,
1932 .reg_read32 = nvme_pci_reg_read32,
1933 .reg_write32 = nvme_pci_reg_write32,
1934 .reg_read64 = nvme_pci_reg_read64,
1935 .reset_ctrl = nvme_pci_reset_ctrl,
1936 .free_ctrl = nvme_pci_free_ctrl,
1937 .post_scan = nvme_pci_post_scan,
1940 static int nvme_dev_map(struct nvme_dev *dev)
1943 struct pci_dev *pdev = to_pci_dev(dev->dev);
1945 bars = pci_select_bars(pdev, IORESOURCE_MEM);
1948 if (pci_request_selected_regions(pdev, bars, "nvme"))
1951 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
1957 pci_release_regions(pdev);
1961 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1963 int node, result = -ENOMEM;
1964 struct nvme_dev *dev;
1966 node = dev_to_node(&pdev->dev);
1967 if (node == NUMA_NO_NODE)
1968 set_dev_node(&pdev->dev, 0);
1970 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
1973 dev->entry = kzalloc_node(num_possible_cpus() * sizeof(*dev->entry),
1977 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
1982 dev->dev = get_device(&pdev->dev);
1983 pci_set_drvdata(pdev, dev);
1985 result = nvme_dev_map(dev);
1989 INIT_WORK(&dev->reset_work, nvme_reset_work);
1990 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
1991 INIT_WORK(&dev->async_work, nvme_async_event_work);
1992 setup_timer(&dev->watchdog_timer, nvme_watchdog_timer,
1993 (unsigned long)dev);
1994 mutex_init(&dev->shutdown_lock);
1995 init_completion(&dev->ioq_wait);
1997 result = nvme_setup_prp_pools(dev);
2001 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2006 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2008 queue_work(nvme_workq, &dev->reset_work);
2012 nvme_release_prp_pools(dev);
2014 put_device(dev->dev);
2015 nvme_dev_unmap(dev);
2023 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2025 struct nvme_dev *dev = pci_get_drvdata(pdev);
2028 nvme_dev_disable(dev, false);
2030 queue_work(nvme_workq, &dev->reset_work);
2033 static void nvme_shutdown(struct pci_dev *pdev)
2035 struct nvme_dev *dev = pci_get_drvdata(pdev);
2036 nvme_dev_disable(dev, true);
2040 * The driver's remove may be called on a device in a partially initialized
2041 * state. This function must not have any dependencies on the device state in
2044 static void nvme_remove(struct pci_dev *pdev)
2046 struct nvme_dev *dev = pci_get_drvdata(pdev);
2048 del_timer_sync(&dev->watchdog_timer);
2050 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2052 pci_set_drvdata(pdev, NULL);
2053 flush_work(&dev->async_work);
2054 nvme_uninit_ctrl(&dev->ctrl);
2055 nvme_dev_disable(dev, true);
2056 flush_work(&dev->reset_work);
2057 nvme_dev_remove_admin(dev);
2058 nvme_free_queues(dev, 0);
2059 nvme_release_cmb(dev);
2060 nvme_release_prp_pools(dev);
2061 nvme_dev_unmap(dev);
2062 nvme_put_ctrl(&dev->ctrl);
2065 #ifdef CONFIG_PM_SLEEP
2066 static int nvme_suspend(struct device *dev)
2068 struct pci_dev *pdev = to_pci_dev(dev);
2069 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2071 nvme_dev_disable(ndev, true);
2075 static int nvme_resume(struct device *dev)
2077 struct pci_dev *pdev = to_pci_dev(dev);
2078 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2080 queue_work(nvme_workq, &ndev->reset_work);
2085 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2087 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2088 pci_channel_state_t state)
2090 struct nvme_dev *dev = pci_get_drvdata(pdev);
2093 * A frozen channel requires a reset. When detected, this method will
2094 * shutdown the controller to quiesce. The controller will be restarted
2095 * after the slot reset through driver's slot_reset callback.
2097 dev_warn(dev->ctrl.device, "error detected: state:%d\n", state);
2099 case pci_channel_io_normal:
2100 return PCI_ERS_RESULT_CAN_RECOVER;
2101 case pci_channel_io_frozen:
2102 nvme_dev_disable(dev, false);
2103 return PCI_ERS_RESULT_NEED_RESET;
2104 case pci_channel_io_perm_failure:
2105 return PCI_ERS_RESULT_DISCONNECT;
2107 return PCI_ERS_RESULT_NEED_RESET;
2110 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2112 struct nvme_dev *dev = pci_get_drvdata(pdev);
2114 dev_info(dev->ctrl.device, "restart after slot reset\n");
2115 pci_restore_state(pdev);
2116 queue_work(nvme_workq, &dev->reset_work);
2117 return PCI_ERS_RESULT_RECOVERED;
2120 static void nvme_error_resume(struct pci_dev *pdev)
2122 pci_cleanup_aer_uncorrect_error_status(pdev);
2125 static const struct pci_error_handlers nvme_err_handler = {
2126 .error_detected = nvme_error_detected,
2127 .slot_reset = nvme_slot_reset,
2128 .resume = nvme_error_resume,
2129 .reset_notify = nvme_reset_notify,
2132 /* Move to pci_ids.h later */
2133 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
2135 static const struct pci_device_id nvme_id_table[] = {
2136 { PCI_VDEVICE(INTEL, 0x0953),
2137 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2138 NVME_QUIRK_DISCARD_ZEROES, },
2139 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
2140 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2141 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2142 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2145 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2147 static struct pci_driver nvme_driver = {
2149 .id_table = nvme_id_table,
2150 .probe = nvme_probe,
2151 .remove = nvme_remove,
2152 .shutdown = nvme_shutdown,
2154 .pm = &nvme_dev_pm_ops,
2156 .err_handler = &nvme_err_handler,
2159 static int __init nvme_init(void)
2163 nvme_workq = alloc_workqueue("nvme", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
2167 result = pci_register_driver(&nvme_driver);
2169 destroy_workqueue(nvme_workq);
2173 static void __exit nvme_exit(void)
2175 pci_unregister_driver(&nvme_driver);
2176 destroy_workqueue(nvme_workq);
2180 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2181 MODULE_LICENSE("GPL");
2182 MODULE_VERSION("1.0");
2183 module_init(nvme_init);
2184 module_exit(nvme_exit);