2 * NVM Express device driver
3 * Copyright (c) 2011-2014, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
15 #include <linux/aer.h>
16 #include <linux/bitops.h>
17 #include <linux/blkdev.h>
18 #include <linux/blk-mq.h>
19 #include <linux/blk-mq-pci.h>
20 #include <linux/dmi.h>
21 #include <linux/init.h>
22 #include <linux/interrupt.h>
25 #include <linux/module.h>
26 #include <linux/mutex.h>
27 #include <linux/pci.h>
28 #include <linux/poison.h>
29 #include <linux/t10-pi.h>
30 #include <linux/timer.h>
31 #include <linux/types.h>
32 #include <linux/io-64-nonatomic-lo-hi.h>
33 #include <asm/unaligned.h>
34 #include <linux/sed-opal.h>
38 #define NVME_Q_DEPTH 1024
39 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
40 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
43 * We handle AEN commands ourselves and don't even let the
44 * block layer know about them.
46 #define NVME_AQ_BLKMQ_DEPTH (NVME_AQ_DEPTH - NVME_NR_AERS)
48 static int use_threaded_interrupts;
49 module_param(use_threaded_interrupts, int, 0);
51 static bool use_cmb_sqes = true;
52 module_param(use_cmb_sqes, bool, 0644);
53 MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
55 static unsigned int max_host_mem_size_mb = 128;
56 module_param(max_host_mem_size_mb, uint, 0444);
57 MODULE_PARM_DESC(max_host_mem_size_mb,
58 "Maximum Host Memory Buffer (HMB) size per controller (in MiB)");
63 static void nvme_process_cq(struct nvme_queue *nvmeq);
64 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown);
67 * Represents an NVM Express device. Each nvme_dev is a PCI function.
70 struct nvme_queue **queues;
71 struct blk_mq_tag_set tagset;
72 struct blk_mq_tag_set admin_tagset;
75 struct dma_pool *prp_page_pool;
76 struct dma_pool *prp_small_pool;
78 unsigned online_queues;
83 unsigned long bar_mapped_size;
84 struct work_struct remove_work;
85 struct mutex shutdown_lock;
88 dma_addr_t cmb_dma_addr;
92 struct nvme_ctrl ctrl;
93 struct completion ioq_wait;
95 /* shadow doorbell buffer support: */
97 dma_addr_t dbbuf_dbs_dma_addr;
99 dma_addr_t dbbuf_eis_dma_addr;
101 /* host memory buffer support: */
103 u32 nr_host_mem_descs;
104 struct nvme_host_mem_buf_desc *host_mem_descs;
105 void **host_mem_desc_bufs;
108 static inline unsigned int sq_idx(unsigned int qid, u32 stride)
110 return qid * 2 * stride;
113 static inline unsigned int cq_idx(unsigned int qid, u32 stride)
115 return (qid * 2 + 1) * stride;
118 static inline struct nvme_dev *to_nvme_dev(struct nvme_ctrl *ctrl)
120 return container_of(ctrl, struct nvme_dev, ctrl);
124 * An NVM Express queue. Each device has at least two (one for admin
125 * commands and one for I/O commands).
128 struct device *q_dmadev;
129 struct nvme_dev *dev;
131 struct nvme_command *sq_cmds;
132 struct nvme_command __iomem *sq_cmds_io;
133 volatile struct nvme_completion *cqes;
134 struct blk_mq_tags **tags;
135 dma_addr_t sq_dma_addr;
136 dma_addr_t cq_dma_addr;
152 * The nvme_iod describes the data in an I/O, including the list of PRP
153 * entries. You can't see it in this data structure because C doesn't let
154 * me express that. Use nvme_init_iod to ensure there's enough space
155 * allocated to store the PRP list.
158 struct nvme_request req;
159 struct nvme_queue *nvmeq;
161 int npages; /* In the PRP list. 0 means small pool in use */
162 int nents; /* Used in scatterlist */
163 int length; /* Of data, in bytes */
164 dma_addr_t first_dma;
165 struct scatterlist meta_sg; /* metadata requires single contiguous buffer */
166 struct scatterlist *sg;
167 struct scatterlist inline_sg[0];
171 * Check we didin't inadvertently grow the command struct
173 static inline void _nvme_check_size(void)
175 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
176 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
177 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
178 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
179 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
180 BUILD_BUG_ON(sizeof(struct nvme_format_cmd) != 64);
181 BUILD_BUG_ON(sizeof(struct nvme_abort_cmd) != 64);
182 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
183 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != NVME_IDENTIFY_DATA_SIZE);
184 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != NVME_IDENTIFY_DATA_SIZE);
185 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
186 BUILD_BUG_ON(sizeof(struct nvme_smart_log) != 512);
187 BUILD_BUG_ON(sizeof(struct nvme_dbbuf) != 64);
190 static inline unsigned int nvme_dbbuf_size(u32 stride)
192 return ((num_possible_cpus() + 1) * 8 * stride);
195 static int nvme_dbbuf_dma_alloc(struct nvme_dev *dev)
197 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
202 dev->dbbuf_dbs = dma_alloc_coherent(dev->dev, mem_size,
203 &dev->dbbuf_dbs_dma_addr,
207 dev->dbbuf_eis = dma_alloc_coherent(dev->dev, mem_size,
208 &dev->dbbuf_eis_dma_addr,
210 if (!dev->dbbuf_eis) {
211 dma_free_coherent(dev->dev, mem_size,
212 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
213 dev->dbbuf_dbs = NULL;
220 static void nvme_dbbuf_dma_free(struct nvme_dev *dev)
222 unsigned int mem_size = nvme_dbbuf_size(dev->db_stride);
224 if (dev->dbbuf_dbs) {
225 dma_free_coherent(dev->dev, mem_size,
226 dev->dbbuf_dbs, dev->dbbuf_dbs_dma_addr);
227 dev->dbbuf_dbs = NULL;
229 if (dev->dbbuf_eis) {
230 dma_free_coherent(dev->dev, mem_size,
231 dev->dbbuf_eis, dev->dbbuf_eis_dma_addr);
232 dev->dbbuf_eis = NULL;
236 static void nvme_dbbuf_init(struct nvme_dev *dev,
237 struct nvme_queue *nvmeq, int qid)
239 if (!dev->dbbuf_dbs || !qid)
242 nvmeq->dbbuf_sq_db = &dev->dbbuf_dbs[sq_idx(qid, dev->db_stride)];
243 nvmeq->dbbuf_cq_db = &dev->dbbuf_dbs[cq_idx(qid, dev->db_stride)];
244 nvmeq->dbbuf_sq_ei = &dev->dbbuf_eis[sq_idx(qid, dev->db_stride)];
245 nvmeq->dbbuf_cq_ei = &dev->dbbuf_eis[cq_idx(qid, dev->db_stride)];
248 static void nvme_dbbuf_set(struct nvme_dev *dev)
250 struct nvme_command c;
255 memset(&c, 0, sizeof(c));
256 c.dbbuf.opcode = nvme_admin_dbbuf;
257 c.dbbuf.prp1 = cpu_to_le64(dev->dbbuf_dbs_dma_addr);
258 c.dbbuf.prp2 = cpu_to_le64(dev->dbbuf_eis_dma_addr);
260 if (nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0)) {
261 dev_warn(dev->ctrl.device, "unable to set dbbuf\n");
262 /* Free memory and continue on */
263 nvme_dbbuf_dma_free(dev);
267 static inline int nvme_dbbuf_need_event(u16 event_idx, u16 new_idx, u16 old)
269 return (u16)(new_idx - event_idx - 1) < (u16)(new_idx - old);
272 /* Update dbbuf and return true if an MMIO is required */
273 static bool nvme_dbbuf_update_and_check_event(u16 value, u32 *dbbuf_db,
274 volatile u32 *dbbuf_ei)
280 * Ensure that the queue is written before updating
281 * the doorbell in memory
285 old_value = *dbbuf_db;
288 if (!nvme_dbbuf_need_event(*dbbuf_ei, value, old_value))
296 * Max size of iod being embedded in the request payload
298 #define NVME_INT_PAGES 2
299 #define NVME_INT_BYTES(dev) (NVME_INT_PAGES * (dev)->ctrl.page_size)
302 * Will slightly overestimate the number of pages needed. This is OK
303 * as it only leads to a small amount of wasted memory for the lifetime of
306 static int nvme_npages(unsigned size, struct nvme_dev *dev)
308 unsigned nprps = DIV_ROUND_UP(size + dev->ctrl.page_size,
309 dev->ctrl.page_size);
310 return DIV_ROUND_UP(8 * nprps, PAGE_SIZE - 8);
313 static unsigned int nvme_iod_alloc_size(struct nvme_dev *dev,
314 unsigned int size, unsigned int nseg)
316 return sizeof(__le64 *) * nvme_npages(size, dev) +
317 sizeof(struct scatterlist) * nseg;
320 static unsigned int nvme_cmd_size(struct nvme_dev *dev)
322 return sizeof(struct nvme_iod) +
323 nvme_iod_alloc_size(dev, NVME_INT_BYTES(dev), NVME_INT_PAGES);
326 static int nvme_admin_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
327 unsigned int hctx_idx)
329 struct nvme_dev *dev = data;
330 struct nvme_queue *nvmeq = dev->queues[0];
332 WARN_ON(hctx_idx != 0);
333 WARN_ON(dev->admin_tagset.tags[0] != hctx->tags);
334 WARN_ON(nvmeq->tags);
336 hctx->driver_data = nvmeq;
337 nvmeq->tags = &dev->admin_tagset.tags[0];
341 static void nvme_admin_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
343 struct nvme_queue *nvmeq = hctx->driver_data;
348 static int nvme_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
349 unsigned int hctx_idx)
351 struct nvme_dev *dev = data;
352 struct nvme_queue *nvmeq = dev->queues[hctx_idx + 1];
355 nvmeq->tags = &dev->tagset.tags[hctx_idx];
357 WARN_ON(dev->tagset.tags[hctx_idx] != hctx->tags);
358 hctx->driver_data = nvmeq;
362 static int nvme_init_request(struct blk_mq_tag_set *set, struct request *req,
363 unsigned int hctx_idx, unsigned int numa_node)
365 struct nvme_dev *dev = set->driver_data;
366 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
367 int queue_idx = (set == &dev->tagset) ? hctx_idx + 1 : 0;
368 struct nvme_queue *nvmeq = dev->queues[queue_idx];
375 static int nvme_pci_map_queues(struct blk_mq_tag_set *set)
377 struct nvme_dev *dev = set->driver_data;
379 return blk_mq_pci_map_queues(set, to_pci_dev(dev->dev));
383 * __nvme_submit_cmd() - Copy a command into a queue and ring the doorbell
384 * @nvmeq: The queue to use
385 * @cmd: The command to send
387 * Safe to use from interrupt context
389 static void __nvme_submit_cmd(struct nvme_queue *nvmeq,
390 struct nvme_command *cmd)
392 u16 tail = nvmeq->sq_tail;
394 if (nvmeq->sq_cmds_io)
395 memcpy_toio(&nvmeq->sq_cmds_io[tail], cmd, sizeof(*cmd));
397 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
399 if (++tail == nvmeq->q_depth)
401 if (nvme_dbbuf_update_and_check_event(tail, nvmeq->dbbuf_sq_db,
403 writel(tail, nvmeq->q_db);
404 nvmeq->sq_tail = tail;
407 static __le64 **iod_list(struct request *req)
409 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
410 return (__le64 **)(iod->sg + blk_rq_nr_phys_segments(req));
413 static blk_status_t nvme_init_iod(struct request *rq, struct nvme_dev *dev)
415 struct nvme_iod *iod = blk_mq_rq_to_pdu(rq);
416 int nseg = blk_rq_nr_phys_segments(rq);
417 unsigned int size = blk_rq_payload_bytes(rq);
419 if (nseg > NVME_INT_PAGES || size > NVME_INT_BYTES(dev)) {
420 iod->sg = kmalloc(nvme_iod_alloc_size(dev, size, nseg), GFP_ATOMIC);
422 return BLK_STS_RESOURCE;
424 iod->sg = iod->inline_sg;
435 static void nvme_free_iod(struct nvme_dev *dev, struct request *req)
437 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
438 const int last_prp = dev->ctrl.page_size / 8 - 1;
440 __le64 **list = iod_list(req);
441 dma_addr_t prp_dma = iod->first_dma;
443 if (iod->npages == 0)
444 dma_pool_free(dev->prp_small_pool, list[0], prp_dma);
445 for (i = 0; i < iod->npages; i++) {
446 __le64 *prp_list = list[i];
447 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
448 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
449 prp_dma = next_prp_dma;
452 if (iod->sg != iod->inline_sg)
456 #ifdef CONFIG_BLK_DEV_INTEGRITY
457 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
459 if (be32_to_cpu(pi->ref_tag) == v)
460 pi->ref_tag = cpu_to_be32(p);
463 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
465 if (be32_to_cpu(pi->ref_tag) == p)
466 pi->ref_tag = cpu_to_be32(v);
470 * nvme_dif_remap - remaps ref tags to bip seed and physical lba
472 * The virtual start sector is the one that was originally submitted by the
473 * block layer. Due to partitioning, MD/DM cloning, etc. the actual physical
474 * start sector may be different. Remap protection information to match the
475 * physical LBA on writes, and back to the original seed on reads.
477 * Type 0 and 3 do not have a ref tag, so no remapping required.
479 static void nvme_dif_remap(struct request *req,
480 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
482 struct nvme_ns *ns = req->rq_disk->private_data;
483 struct bio_integrity_payload *bip;
484 struct t10_pi_tuple *pi;
486 u32 i, nlb, ts, phys, virt;
488 if (!ns->pi_type || ns->pi_type == NVME_NS_DPS_PI_TYPE3)
491 bip = bio_integrity(req->bio);
495 pmap = kmap_atomic(bip->bip_vec->bv_page) + bip->bip_vec->bv_offset;
498 virt = bip_get_seed(bip);
499 phys = nvme_block_nr(ns, blk_rq_pos(req));
500 nlb = (blk_rq_bytes(req) >> ns->lba_shift);
501 ts = ns->disk->queue->integrity.tuple_size;
503 for (i = 0; i < nlb; i++, virt++, phys++) {
504 pi = (struct t10_pi_tuple *)p;
505 dif_swap(phys, virt, pi);
510 #else /* CONFIG_BLK_DEV_INTEGRITY */
511 static void nvme_dif_remap(struct request *req,
512 void (*dif_swap)(u32 p, u32 v, struct t10_pi_tuple *pi))
515 static void nvme_dif_prep(u32 p, u32 v, struct t10_pi_tuple *pi)
518 static void nvme_dif_complete(u32 p, u32 v, struct t10_pi_tuple *pi)
523 static bool nvme_setup_prps(struct nvme_dev *dev, struct request *req)
525 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
526 struct dma_pool *pool;
527 int length = blk_rq_payload_bytes(req);
528 struct scatterlist *sg = iod->sg;
529 int dma_len = sg_dma_len(sg);
530 u64 dma_addr = sg_dma_address(sg);
531 u32 page_size = dev->ctrl.page_size;
532 int offset = dma_addr & (page_size - 1);
534 __le64 **list = iod_list(req);
538 length -= (page_size - offset);
542 dma_len -= (page_size - offset);
544 dma_addr += (page_size - offset);
547 dma_addr = sg_dma_address(sg);
548 dma_len = sg_dma_len(sg);
551 if (length <= page_size) {
552 iod->first_dma = dma_addr;
556 nprps = DIV_ROUND_UP(length, page_size);
557 if (nprps <= (256 / 8)) {
558 pool = dev->prp_small_pool;
561 pool = dev->prp_page_pool;
565 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
567 iod->first_dma = dma_addr;
572 iod->first_dma = prp_dma;
575 if (i == page_size >> 3) {
576 __le64 *old_prp_list = prp_list;
577 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
580 list[iod->npages++] = prp_list;
581 prp_list[0] = old_prp_list[i - 1];
582 old_prp_list[i - 1] = cpu_to_le64(prp_dma);
585 prp_list[i++] = cpu_to_le64(dma_addr);
586 dma_len -= page_size;
587 dma_addr += page_size;
595 dma_addr = sg_dma_address(sg);
596 dma_len = sg_dma_len(sg);
602 static blk_status_t nvme_map_data(struct nvme_dev *dev, struct request *req,
603 struct nvme_command *cmnd)
605 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
606 struct request_queue *q = req->q;
607 enum dma_data_direction dma_dir = rq_data_dir(req) ?
608 DMA_TO_DEVICE : DMA_FROM_DEVICE;
609 blk_status_t ret = BLK_STS_IOERR;
611 sg_init_table(iod->sg, blk_rq_nr_phys_segments(req));
612 iod->nents = blk_rq_map_sg(q, req, iod->sg);
616 ret = BLK_STS_RESOURCE;
617 if (!dma_map_sg_attrs(dev->dev, iod->sg, iod->nents, dma_dir,
621 if (!nvme_setup_prps(dev, req))
625 if (blk_integrity_rq(req)) {
626 if (blk_rq_count_integrity_sg(q, req->bio) != 1)
629 sg_init_table(&iod->meta_sg, 1);
630 if (blk_rq_map_integrity_sg(q, req->bio, &iod->meta_sg) != 1)
633 if (rq_data_dir(req))
634 nvme_dif_remap(req, nvme_dif_prep);
636 if (!dma_map_sg(dev->dev, &iod->meta_sg, 1, dma_dir))
640 cmnd->rw.dptr.prp1 = cpu_to_le64(sg_dma_address(iod->sg));
641 cmnd->rw.dptr.prp2 = cpu_to_le64(iod->first_dma);
642 if (blk_integrity_rq(req))
643 cmnd->rw.metadata = cpu_to_le64(sg_dma_address(&iod->meta_sg));
647 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
652 static void nvme_unmap_data(struct nvme_dev *dev, struct request *req)
654 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
655 enum dma_data_direction dma_dir = rq_data_dir(req) ?
656 DMA_TO_DEVICE : DMA_FROM_DEVICE;
659 dma_unmap_sg(dev->dev, iod->sg, iod->nents, dma_dir);
660 if (blk_integrity_rq(req)) {
661 if (!rq_data_dir(req))
662 nvme_dif_remap(req, nvme_dif_complete);
663 dma_unmap_sg(dev->dev, &iod->meta_sg, 1, dma_dir);
667 nvme_cleanup_cmd(req);
668 nvme_free_iod(dev, req);
672 * NOTE: ns is NULL when called on the admin queue.
674 static blk_status_t nvme_queue_rq(struct blk_mq_hw_ctx *hctx,
675 const struct blk_mq_queue_data *bd)
677 struct nvme_ns *ns = hctx->queue->queuedata;
678 struct nvme_queue *nvmeq = hctx->driver_data;
679 struct nvme_dev *dev = nvmeq->dev;
680 struct request *req = bd->rq;
681 struct nvme_command cmnd;
684 ret = nvme_setup_cmd(ns, req, &cmnd);
688 ret = nvme_init_iod(req, dev);
692 if (blk_rq_nr_phys_segments(req)) {
693 ret = nvme_map_data(dev, req, &cmnd);
695 goto out_cleanup_iod;
698 blk_mq_start_request(req);
700 spin_lock_irq(&nvmeq->q_lock);
701 if (unlikely(nvmeq->cq_vector < 0)) {
703 spin_unlock_irq(&nvmeq->q_lock);
704 goto out_cleanup_iod;
706 __nvme_submit_cmd(nvmeq, &cmnd);
707 nvme_process_cq(nvmeq);
708 spin_unlock_irq(&nvmeq->q_lock);
711 nvme_free_iod(dev, req);
713 nvme_cleanup_cmd(req);
717 static void nvme_pci_complete_rq(struct request *req)
719 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
721 nvme_unmap_data(iod->nvmeq->dev, req);
722 nvme_complete_rq(req);
725 /* We read the CQE phase first to check if the rest of the entry is valid */
726 static inline bool nvme_cqe_valid(struct nvme_queue *nvmeq, u16 head,
729 return (le16_to_cpu(nvmeq->cqes[head].status) & 1) == phase;
732 static inline void nvme_ring_cq_doorbell(struct nvme_queue *nvmeq)
734 u16 head = nvmeq->cq_head;
736 if (likely(nvmeq->cq_vector >= 0)) {
737 if (nvme_dbbuf_update_and_check_event(head, nvmeq->dbbuf_cq_db,
739 writel(head, nvmeq->q_db + nvmeq->dev->db_stride);
743 static inline void nvme_handle_cqe(struct nvme_queue *nvmeq,
744 struct nvme_completion *cqe)
748 if (unlikely(cqe->command_id >= nvmeq->q_depth)) {
749 dev_warn(nvmeq->dev->ctrl.device,
750 "invalid id %d completed on queue %d\n",
751 cqe->command_id, le16_to_cpu(cqe->sq_id));
756 * AEN requests are special as they don't time out and can
757 * survive any kind of queue freeze and often don't respond to
758 * aborts. We don't even bother to allocate a struct request
759 * for them but rather special case them here.
761 if (unlikely(nvmeq->qid == 0 &&
762 cqe->command_id >= NVME_AQ_BLKMQ_DEPTH)) {
763 nvme_complete_async_event(&nvmeq->dev->ctrl,
764 cqe->status, &cqe->result);
768 req = blk_mq_tag_to_rq(*nvmeq->tags, cqe->command_id);
769 nvme_end_request(req, cqe->status, cqe->result);
772 static inline bool nvme_read_cqe(struct nvme_queue *nvmeq,
773 struct nvme_completion *cqe)
775 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase)) {
776 *cqe = nvmeq->cqes[nvmeq->cq_head];
778 if (++nvmeq->cq_head == nvmeq->q_depth) {
780 nvmeq->cq_phase = !nvmeq->cq_phase;
787 static void nvme_process_cq(struct nvme_queue *nvmeq)
789 struct nvme_completion cqe;
792 while (nvme_read_cqe(nvmeq, &cqe)) {
793 nvme_handle_cqe(nvmeq, &cqe);
798 nvme_ring_cq_doorbell(nvmeq);
803 static irqreturn_t nvme_irq(int irq, void *data)
806 struct nvme_queue *nvmeq = data;
807 spin_lock(&nvmeq->q_lock);
808 nvme_process_cq(nvmeq);
809 result = nvmeq->cqe_seen ? IRQ_HANDLED : IRQ_NONE;
811 spin_unlock(&nvmeq->q_lock);
815 static irqreturn_t nvme_irq_check(int irq, void *data)
817 struct nvme_queue *nvmeq = data;
818 if (nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
819 return IRQ_WAKE_THREAD;
823 static int __nvme_poll(struct nvme_queue *nvmeq, unsigned int tag)
825 struct nvme_completion cqe;
826 int found = 0, consumed = 0;
828 if (!nvme_cqe_valid(nvmeq, nvmeq->cq_head, nvmeq->cq_phase))
831 spin_lock_irq(&nvmeq->q_lock);
832 while (nvme_read_cqe(nvmeq, &cqe)) {
833 nvme_handle_cqe(nvmeq, &cqe);
836 if (tag == cqe.command_id) {
843 nvme_ring_cq_doorbell(nvmeq);
844 spin_unlock_irq(&nvmeq->q_lock);
849 static int nvme_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
851 struct nvme_queue *nvmeq = hctx->driver_data;
853 return __nvme_poll(nvmeq, tag);
856 static void nvme_pci_submit_async_event(struct nvme_ctrl *ctrl, int aer_idx)
858 struct nvme_dev *dev = to_nvme_dev(ctrl);
859 struct nvme_queue *nvmeq = dev->queues[0];
860 struct nvme_command c;
862 memset(&c, 0, sizeof(c));
863 c.common.opcode = nvme_admin_async_event;
864 c.common.command_id = NVME_AQ_BLKMQ_DEPTH + aer_idx;
866 spin_lock_irq(&nvmeq->q_lock);
867 __nvme_submit_cmd(nvmeq, &c);
868 spin_unlock_irq(&nvmeq->q_lock);
871 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
873 struct nvme_command c;
875 memset(&c, 0, sizeof(c));
876 c.delete_queue.opcode = opcode;
877 c.delete_queue.qid = cpu_to_le16(id);
879 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
882 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
883 struct nvme_queue *nvmeq)
885 struct nvme_command c;
886 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
889 * Note: we (ab)use the fact the the prp fields survive if no data
890 * is attached to the request.
892 memset(&c, 0, sizeof(c));
893 c.create_cq.opcode = nvme_admin_create_cq;
894 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
895 c.create_cq.cqid = cpu_to_le16(qid);
896 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
897 c.create_cq.cq_flags = cpu_to_le16(flags);
898 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
900 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
903 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
904 struct nvme_queue *nvmeq)
906 struct nvme_command c;
907 int flags = NVME_QUEUE_PHYS_CONTIG;
910 * Note: we (ab)use the fact the the prp fields survive if no data
911 * is attached to the request.
913 memset(&c, 0, sizeof(c));
914 c.create_sq.opcode = nvme_admin_create_sq;
915 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
916 c.create_sq.sqid = cpu_to_le16(qid);
917 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
918 c.create_sq.sq_flags = cpu_to_le16(flags);
919 c.create_sq.cqid = cpu_to_le16(qid);
921 return nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
924 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
926 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
929 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
931 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
934 static void abort_endio(struct request *req, blk_status_t error)
936 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
937 struct nvme_queue *nvmeq = iod->nvmeq;
939 dev_warn(nvmeq->dev->ctrl.device,
940 "Abort status: 0x%x", nvme_req(req)->status);
941 atomic_inc(&nvmeq->dev->ctrl.abort_limit);
942 blk_mq_free_request(req);
945 static bool nvme_should_reset(struct nvme_dev *dev, u32 csts)
948 /* If true, indicates loss of adapter communication, possibly by a
949 * NVMe Subsystem reset.
951 bool nssro = dev->subsystem && (csts & NVME_CSTS_NSSRO);
953 /* If there is a reset ongoing, we shouldn't reset again. */
954 if (dev->ctrl.state == NVME_CTRL_RESETTING)
957 /* We shouldn't reset unless the controller is on fatal error state
958 * _or_ if we lost the communication with it.
960 if (!(csts & NVME_CSTS_CFS) && !nssro)
963 /* If PCI error recovery process is happening, we cannot reset or
964 * the recovery mechanism will surely fail.
966 if (pci_channel_offline(to_pci_dev(dev->dev)))
972 static void nvme_warn_reset(struct nvme_dev *dev, u32 csts)
974 /* Read a config register to help see what died. */
978 result = pci_read_config_word(to_pci_dev(dev->dev), PCI_STATUS,
980 if (result == PCIBIOS_SUCCESSFUL)
981 dev_warn(dev->ctrl.device,
982 "controller is down; will reset: CSTS=0x%x, PCI_STATUS=0x%hx\n",
985 dev_warn(dev->ctrl.device,
986 "controller is down; will reset: CSTS=0x%x, PCI_STATUS read failed (%d)\n",
990 static enum blk_eh_timer_return nvme_timeout(struct request *req, bool reserved)
992 struct nvme_iod *iod = blk_mq_rq_to_pdu(req);
993 struct nvme_queue *nvmeq = iod->nvmeq;
994 struct nvme_dev *dev = nvmeq->dev;
995 struct request *abort_req;
996 struct nvme_command cmd;
997 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1000 * Reset immediately if the controller is failed
1002 if (nvme_should_reset(dev, csts)) {
1003 nvme_warn_reset(dev, csts);
1004 nvme_dev_disable(dev, false);
1005 nvme_reset_ctrl(&dev->ctrl);
1006 return BLK_EH_HANDLED;
1010 * Did we miss an interrupt?
1012 if (__nvme_poll(nvmeq, req->tag)) {
1013 dev_warn(dev->ctrl.device,
1014 "I/O %d QID %d timeout, completion polled\n",
1015 req->tag, nvmeq->qid);
1016 return BLK_EH_HANDLED;
1020 * Shutdown immediately if controller times out while starting. The
1021 * reset work will see the pci device disabled when it gets the forced
1022 * cancellation error. All outstanding requests are completed on
1023 * shutdown, so we return BLK_EH_HANDLED.
1025 if (dev->ctrl.state == NVME_CTRL_RESETTING) {
1026 dev_warn(dev->ctrl.device,
1027 "I/O %d QID %d timeout, disable controller\n",
1028 req->tag, nvmeq->qid);
1029 nvme_dev_disable(dev, false);
1030 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1031 return BLK_EH_HANDLED;
1035 * Shutdown the controller immediately and schedule a reset if the
1036 * command was already aborted once before and still hasn't been
1037 * returned to the driver, or if this is the admin queue.
1039 if (!nvmeq->qid || iod->aborted) {
1040 dev_warn(dev->ctrl.device,
1041 "I/O %d QID %d timeout, reset controller\n",
1042 req->tag, nvmeq->qid);
1043 nvme_dev_disable(dev, false);
1044 nvme_reset_ctrl(&dev->ctrl);
1047 * Mark the request as handled, since the inline shutdown
1048 * forces all outstanding requests to complete.
1050 nvme_req(req)->flags |= NVME_REQ_CANCELLED;
1051 return BLK_EH_HANDLED;
1054 if (atomic_dec_return(&dev->ctrl.abort_limit) < 0) {
1055 atomic_inc(&dev->ctrl.abort_limit);
1056 return BLK_EH_RESET_TIMER;
1060 memset(&cmd, 0, sizeof(cmd));
1061 cmd.abort.opcode = nvme_admin_abort_cmd;
1062 cmd.abort.cid = req->tag;
1063 cmd.abort.sqid = cpu_to_le16(nvmeq->qid);
1065 dev_warn(nvmeq->dev->ctrl.device,
1066 "I/O %d QID %d timeout, aborting\n",
1067 req->tag, nvmeq->qid);
1069 abort_req = nvme_alloc_request(dev->ctrl.admin_q, &cmd,
1070 BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1071 if (IS_ERR(abort_req)) {
1072 atomic_inc(&dev->ctrl.abort_limit);
1073 return BLK_EH_RESET_TIMER;
1076 abort_req->timeout = ADMIN_TIMEOUT;
1077 abort_req->end_io_data = NULL;
1078 blk_execute_rq_nowait(abort_req->q, NULL, abort_req, 0, abort_endio);
1081 * The aborted req will be completed on receiving the abort req.
1082 * We enable the timer again. If hit twice, it'll cause a device reset,
1083 * as the device then is in a faulty state.
1085 return BLK_EH_RESET_TIMER;
1088 static void nvme_free_queue(struct nvme_queue *nvmeq)
1090 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
1091 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
1093 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
1094 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
1098 static void nvme_free_queues(struct nvme_dev *dev, int lowest)
1102 for (i = dev->queue_count - 1; i >= lowest; i--) {
1103 struct nvme_queue *nvmeq = dev->queues[i];
1105 dev->queues[i] = NULL;
1106 nvme_free_queue(nvmeq);
1111 * nvme_suspend_queue - put queue into suspended state
1112 * @nvmeq - queue to suspend
1114 static int nvme_suspend_queue(struct nvme_queue *nvmeq)
1118 spin_lock_irq(&nvmeq->q_lock);
1119 if (nvmeq->cq_vector == -1) {
1120 spin_unlock_irq(&nvmeq->q_lock);
1123 vector = nvmeq->cq_vector;
1124 nvmeq->dev->online_queues--;
1125 nvmeq->cq_vector = -1;
1126 spin_unlock_irq(&nvmeq->q_lock);
1128 if (!nvmeq->qid && nvmeq->dev->ctrl.admin_q)
1129 blk_mq_stop_hw_queues(nvmeq->dev->ctrl.admin_q);
1131 pci_free_irq(to_pci_dev(nvmeq->dev->dev), vector, nvmeq);
1136 static void nvme_disable_admin_queue(struct nvme_dev *dev, bool shutdown)
1138 struct nvme_queue *nvmeq = dev->queues[0];
1142 if (nvme_suspend_queue(nvmeq))
1146 nvme_shutdown_ctrl(&dev->ctrl);
1148 nvme_disable_ctrl(&dev->ctrl, lo_hi_readq(
1149 dev->bar + NVME_REG_CAP));
1151 spin_lock_irq(&nvmeq->q_lock);
1152 nvme_process_cq(nvmeq);
1153 spin_unlock_irq(&nvmeq->q_lock);
1156 static int nvme_cmb_qdepth(struct nvme_dev *dev, int nr_io_queues,
1159 int q_depth = dev->q_depth;
1160 unsigned q_size_aligned = roundup(q_depth * entry_size,
1161 dev->ctrl.page_size);
1163 if (q_size_aligned * nr_io_queues > dev->cmb_size) {
1164 u64 mem_per_q = div_u64(dev->cmb_size, nr_io_queues);
1165 mem_per_q = round_down(mem_per_q, dev->ctrl.page_size);
1166 q_depth = div_u64(mem_per_q, entry_size);
1169 * Ensure the reduced q_depth is above some threshold where it
1170 * would be better to map queues in system memory with the
1180 static int nvme_alloc_sq_cmds(struct nvme_dev *dev, struct nvme_queue *nvmeq,
1183 if (qid && dev->cmb && use_cmb_sqes && NVME_CMB_SQS(dev->cmbsz)) {
1184 unsigned offset = (qid - 1) * roundup(SQ_SIZE(depth),
1185 dev->ctrl.page_size);
1186 nvmeq->sq_dma_addr = dev->cmb_dma_addr + offset;
1187 nvmeq->sq_cmds_io = dev->cmb + offset;
1189 nvmeq->sq_cmds = dma_alloc_coherent(dev->dev, SQ_SIZE(depth),
1190 &nvmeq->sq_dma_addr, GFP_KERNEL);
1191 if (!nvmeq->sq_cmds)
1198 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
1199 int depth, int node)
1201 struct nvme_queue *nvmeq = kzalloc_node(sizeof(*nvmeq), GFP_KERNEL,
1206 nvmeq->cqes = dma_zalloc_coherent(dev->dev, CQ_SIZE(depth),
1207 &nvmeq->cq_dma_addr, GFP_KERNEL);
1211 if (nvme_alloc_sq_cmds(dev, nvmeq, qid, depth))
1214 nvmeq->q_dmadev = dev->dev;
1216 spin_lock_init(&nvmeq->q_lock);
1218 nvmeq->cq_phase = 1;
1219 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1220 nvmeq->q_depth = depth;
1222 nvmeq->cq_vector = -1;
1223 dev->queues[qid] = nvmeq;
1229 dma_free_coherent(dev->dev, CQ_SIZE(depth), (void *)nvmeq->cqes,
1230 nvmeq->cq_dma_addr);
1236 static int queue_request_irq(struct nvme_queue *nvmeq)
1238 struct pci_dev *pdev = to_pci_dev(nvmeq->dev->dev);
1239 int nr = nvmeq->dev->ctrl.instance;
1241 if (use_threaded_interrupts) {
1242 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq_check,
1243 nvme_irq, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1245 return pci_request_irq(pdev, nvmeq->cq_vector, nvme_irq,
1246 NULL, nvmeq, "nvme%dq%d", nr, nvmeq->qid);
1250 static void nvme_init_queue(struct nvme_queue *nvmeq, u16 qid)
1252 struct nvme_dev *dev = nvmeq->dev;
1254 spin_lock_irq(&nvmeq->q_lock);
1257 nvmeq->cq_phase = 1;
1258 nvmeq->q_db = &dev->dbs[qid * 2 * dev->db_stride];
1259 memset((void *)nvmeq->cqes, 0, CQ_SIZE(nvmeq->q_depth));
1260 nvme_dbbuf_init(dev, nvmeq, qid);
1261 dev->online_queues++;
1262 spin_unlock_irq(&nvmeq->q_lock);
1265 static int nvme_create_queue(struct nvme_queue *nvmeq, int qid)
1267 struct nvme_dev *dev = nvmeq->dev;
1270 nvmeq->cq_vector = qid - 1;
1271 result = adapter_alloc_cq(dev, qid, nvmeq);
1275 result = adapter_alloc_sq(dev, qid, nvmeq);
1279 result = queue_request_irq(nvmeq);
1283 nvme_init_queue(nvmeq, qid);
1287 adapter_delete_sq(dev, qid);
1289 adapter_delete_cq(dev, qid);
1293 static const struct blk_mq_ops nvme_mq_admin_ops = {
1294 .queue_rq = nvme_queue_rq,
1295 .complete = nvme_pci_complete_rq,
1296 .init_hctx = nvme_admin_init_hctx,
1297 .exit_hctx = nvme_admin_exit_hctx,
1298 .init_request = nvme_init_request,
1299 .timeout = nvme_timeout,
1302 static const struct blk_mq_ops nvme_mq_ops = {
1303 .queue_rq = nvme_queue_rq,
1304 .complete = nvme_pci_complete_rq,
1305 .init_hctx = nvme_init_hctx,
1306 .init_request = nvme_init_request,
1307 .map_queues = nvme_pci_map_queues,
1308 .timeout = nvme_timeout,
1312 static void nvme_dev_remove_admin(struct nvme_dev *dev)
1314 if (dev->ctrl.admin_q && !blk_queue_dying(dev->ctrl.admin_q)) {
1316 * If the controller was reset during removal, it's possible
1317 * user requests may be waiting on a stopped queue. Start the
1318 * queue to flush these to completion.
1320 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1321 blk_cleanup_queue(dev->ctrl.admin_q);
1322 blk_mq_free_tag_set(&dev->admin_tagset);
1326 static int nvme_alloc_admin_tags(struct nvme_dev *dev)
1328 if (!dev->ctrl.admin_q) {
1329 dev->admin_tagset.ops = &nvme_mq_admin_ops;
1330 dev->admin_tagset.nr_hw_queues = 1;
1333 * Subtract one to leave an empty queue entry for 'Full Queue'
1334 * condition. See NVM-Express 1.2 specification, section 4.1.2.
1336 dev->admin_tagset.queue_depth = NVME_AQ_BLKMQ_DEPTH - 1;
1337 dev->admin_tagset.timeout = ADMIN_TIMEOUT;
1338 dev->admin_tagset.numa_node = dev_to_node(dev->dev);
1339 dev->admin_tagset.cmd_size = nvme_cmd_size(dev);
1340 dev->admin_tagset.flags = BLK_MQ_F_NO_SCHED;
1341 dev->admin_tagset.driver_data = dev;
1343 if (blk_mq_alloc_tag_set(&dev->admin_tagset))
1346 dev->ctrl.admin_q = blk_mq_init_queue(&dev->admin_tagset);
1347 if (IS_ERR(dev->ctrl.admin_q)) {
1348 blk_mq_free_tag_set(&dev->admin_tagset);
1351 if (!blk_get_queue(dev->ctrl.admin_q)) {
1352 nvme_dev_remove_admin(dev);
1353 dev->ctrl.admin_q = NULL;
1357 blk_mq_start_stopped_hw_queues(dev->ctrl.admin_q, true);
1362 static unsigned long db_bar_size(struct nvme_dev *dev, unsigned nr_io_queues)
1364 return NVME_REG_DBS + ((nr_io_queues + 1) * 8 * dev->db_stride);
1367 static int nvme_remap_bar(struct nvme_dev *dev, unsigned long size)
1369 struct pci_dev *pdev = to_pci_dev(dev->dev);
1371 if (size <= dev->bar_mapped_size)
1373 if (size > pci_resource_len(pdev, 0))
1377 dev->bar = ioremap(pci_resource_start(pdev, 0), size);
1379 dev->bar_mapped_size = 0;
1382 dev->bar_mapped_size = size;
1383 dev->dbs = dev->bar + NVME_REG_DBS;
1388 static int nvme_configure_admin_queue(struct nvme_dev *dev)
1392 u64 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1393 struct nvme_queue *nvmeq;
1395 result = nvme_remap_bar(dev, db_bar_size(dev, 0));
1399 dev->subsystem = readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 1, 0) ?
1400 NVME_CAP_NSSRC(cap) : 0;
1402 if (dev->subsystem &&
1403 (readl(dev->bar + NVME_REG_CSTS) & NVME_CSTS_NSSRO))
1404 writel(NVME_CSTS_NSSRO, dev->bar + NVME_REG_CSTS);
1406 result = nvme_disable_ctrl(&dev->ctrl, cap);
1410 nvmeq = dev->queues[0];
1412 nvmeq = nvme_alloc_queue(dev, 0, NVME_AQ_DEPTH,
1413 dev_to_node(dev->dev));
1418 aqa = nvmeq->q_depth - 1;
1421 writel(aqa, dev->bar + NVME_REG_AQA);
1422 lo_hi_writeq(nvmeq->sq_dma_addr, dev->bar + NVME_REG_ASQ);
1423 lo_hi_writeq(nvmeq->cq_dma_addr, dev->bar + NVME_REG_ACQ);
1425 result = nvme_enable_ctrl(&dev->ctrl, cap);
1429 nvmeq->cq_vector = 0;
1430 result = queue_request_irq(nvmeq);
1432 nvmeq->cq_vector = -1;
1439 static int nvme_create_io_queues(struct nvme_dev *dev)
1444 for (i = dev->queue_count; i <= dev->max_qid; i++) {
1445 /* vector == qid - 1, match nvme_create_queue */
1446 if (!nvme_alloc_queue(dev, i, dev->q_depth,
1447 pci_irq_get_node(to_pci_dev(dev->dev), i - 1))) {
1453 max = min(dev->max_qid, dev->queue_count - 1);
1454 for (i = dev->online_queues; i <= max; i++) {
1455 ret = nvme_create_queue(dev->queues[i], i);
1461 * Ignore failing Create SQ/CQ commands, we can continue with less
1462 * than the desired aount of queues, and even a controller without
1463 * I/O queues an still be used to issue admin commands. This might
1464 * be useful to upgrade a buggy firmware for example.
1466 return ret >= 0 ? 0 : ret;
1469 static ssize_t nvme_cmb_show(struct device *dev,
1470 struct device_attribute *attr,
1473 struct nvme_dev *ndev = to_nvme_dev(dev_get_drvdata(dev));
1475 return scnprintf(buf, PAGE_SIZE, "cmbloc : x%08x\ncmbsz : x%08x\n",
1476 ndev->cmbloc, ndev->cmbsz);
1478 static DEVICE_ATTR(cmb, S_IRUGO, nvme_cmb_show, NULL);
1480 static void __iomem *nvme_map_cmb(struct nvme_dev *dev)
1482 u64 szu, size, offset;
1483 resource_size_t bar_size;
1484 struct pci_dev *pdev = to_pci_dev(dev->dev);
1486 dma_addr_t dma_addr;
1488 dev->cmbsz = readl(dev->bar + NVME_REG_CMBSZ);
1489 if (!(NVME_CMB_SZ(dev->cmbsz)))
1491 dev->cmbloc = readl(dev->bar + NVME_REG_CMBLOC);
1496 szu = (u64)1 << (12 + 4 * NVME_CMB_SZU(dev->cmbsz));
1497 size = szu * NVME_CMB_SZ(dev->cmbsz);
1498 offset = szu * NVME_CMB_OFST(dev->cmbloc);
1499 bar_size = pci_resource_len(pdev, NVME_CMB_BIR(dev->cmbloc));
1501 if (offset > bar_size)
1505 * Controllers may support a CMB size larger than their BAR,
1506 * for example, due to being behind a bridge. Reduce the CMB to
1507 * the reported size of the BAR
1509 if (size > bar_size - offset)
1510 size = bar_size - offset;
1512 dma_addr = pci_resource_start(pdev, NVME_CMB_BIR(dev->cmbloc)) + offset;
1513 cmb = ioremap_wc(dma_addr, size);
1517 dev->cmb_dma_addr = dma_addr;
1518 dev->cmb_size = size;
1522 static inline void nvme_release_cmb(struct nvme_dev *dev)
1528 sysfs_remove_file_from_group(&dev->ctrl.device->kobj,
1529 &dev_attr_cmb.attr, NULL);
1535 static int nvme_set_host_mem(struct nvme_dev *dev, u32 bits)
1537 size_t len = dev->nr_host_mem_descs * sizeof(*dev->host_mem_descs);
1538 struct nvme_command c;
1542 dma_addr = dma_map_single(dev->dev, dev->host_mem_descs, len,
1544 if (dma_mapping_error(dev->dev, dma_addr))
1547 memset(&c, 0, sizeof(c));
1548 c.features.opcode = nvme_admin_set_features;
1549 c.features.fid = cpu_to_le32(NVME_FEAT_HOST_MEM_BUF);
1550 c.features.dword11 = cpu_to_le32(bits);
1551 c.features.dword12 = cpu_to_le32(dev->host_mem_size >>
1552 ilog2(dev->ctrl.page_size));
1553 c.features.dword13 = cpu_to_le32(lower_32_bits(dma_addr));
1554 c.features.dword14 = cpu_to_le32(upper_32_bits(dma_addr));
1555 c.features.dword15 = cpu_to_le32(dev->nr_host_mem_descs);
1557 ret = nvme_submit_sync_cmd(dev->ctrl.admin_q, &c, NULL, 0);
1559 dev_warn(dev->ctrl.device,
1560 "failed to set host mem (err %d, flags %#x).\n",
1563 dma_unmap_single(dev->dev, dma_addr, len, DMA_TO_DEVICE);
1567 static void nvme_free_host_mem(struct nvme_dev *dev)
1571 for (i = 0; i < dev->nr_host_mem_descs; i++) {
1572 struct nvme_host_mem_buf_desc *desc = &dev->host_mem_descs[i];
1573 size_t size = le32_to_cpu(desc->size) * dev->ctrl.page_size;
1575 dma_free_coherent(dev->dev, size, dev->host_mem_desc_bufs[i],
1576 le64_to_cpu(desc->addr));
1579 kfree(dev->host_mem_desc_bufs);
1580 dev->host_mem_desc_bufs = NULL;
1581 kfree(dev->host_mem_descs);
1582 dev->host_mem_descs = NULL;
1585 static int nvme_alloc_host_mem(struct nvme_dev *dev, u64 min, u64 preferred)
1587 struct nvme_host_mem_buf_desc *descs;
1588 u32 chunk_size, max_entries, i = 0;
1592 /* start big and work our way down */
1593 chunk_size = min(preferred, (u64)PAGE_SIZE << MAX_ORDER);
1595 tmp = (preferred + chunk_size - 1);
1596 do_div(tmp, chunk_size);
1598 descs = kcalloc(max_entries, sizeof(*descs), GFP_KERNEL);
1602 bufs = kcalloc(max_entries, sizeof(*bufs), GFP_KERNEL);
1604 goto out_free_descs;
1606 for (size = 0; size < preferred; size += chunk_size) {
1607 u32 len = min_t(u64, chunk_size, preferred - size);
1608 dma_addr_t dma_addr;
1610 bufs[i] = dma_alloc_attrs(dev->dev, len, &dma_addr, GFP_KERNEL,
1611 DMA_ATTR_NO_KERNEL_MAPPING | DMA_ATTR_NO_WARN);
1615 descs[i].addr = cpu_to_le64(dma_addr);
1616 descs[i].size = cpu_to_le32(len / dev->ctrl.page_size);
1620 if (!size || (min && size < min)) {
1621 dev_warn(dev->ctrl.device,
1622 "failed to allocate host memory buffer.\n");
1626 dev_info(dev->ctrl.device,
1627 "allocated %lld MiB host memory buffer.\n",
1628 size >> ilog2(SZ_1M));
1629 dev->nr_host_mem_descs = i;
1630 dev->host_mem_size = size;
1631 dev->host_mem_descs = descs;
1632 dev->host_mem_desc_bufs = bufs;
1637 size_t size = le32_to_cpu(descs[i].size) * dev->ctrl.page_size;
1639 dma_free_coherent(dev->dev, size, bufs[i],
1640 le64_to_cpu(descs[i].addr));
1647 /* try a smaller chunk size if we failed early */
1648 if (chunk_size >= PAGE_SIZE * 2 && (i == 0 || size < min)) {
1652 dev->host_mem_descs = NULL;
1656 static void nvme_setup_host_mem(struct nvme_dev *dev)
1658 u64 max = (u64)max_host_mem_size_mb * SZ_1M;
1659 u64 preferred = (u64)dev->ctrl.hmpre * 4096;
1660 u64 min = (u64)dev->ctrl.hmmin * 4096;
1661 u32 enable_bits = NVME_HOST_MEM_ENABLE;
1663 preferred = min(preferred, max);
1665 dev_warn(dev->ctrl.device,
1666 "min host memory (%lld MiB) above limit (%d MiB).\n",
1667 min >> ilog2(SZ_1M), max_host_mem_size_mb);
1668 nvme_free_host_mem(dev);
1673 * If we already have a buffer allocated check if we can reuse it.
1675 if (dev->host_mem_descs) {
1676 if (dev->host_mem_size >= min)
1677 enable_bits |= NVME_HOST_MEM_RETURN;
1679 nvme_free_host_mem(dev);
1682 if (!dev->host_mem_descs) {
1683 if (nvme_alloc_host_mem(dev, min, preferred))
1687 if (nvme_set_host_mem(dev, enable_bits))
1688 nvme_free_host_mem(dev);
1691 static int nvme_setup_io_queues(struct nvme_dev *dev)
1693 struct nvme_queue *adminq = dev->queues[0];
1694 struct pci_dev *pdev = to_pci_dev(dev->dev);
1695 int result, nr_io_queues;
1698 nr_io_queues = num_present_cpus();
1699 result = nvme_set_queue_count(&dev->ctrl, &nr_io_queues);
1703 if (nr_io_queues == 0)
1706 if (dev->cmb && NVME_CMB_SQS(dev->cmbsz)) {
1707 result = nvme_cmb_qdepth(dev, nr_io_queues,
1708 sizeof(struct nvme_command));
1710 dev->q_depth = result;
1712 nvme_release_cmb(dev);
1716 size = db_bar_size(dev, nr_io_queues);
1717 result = nvme_remap_bar(dev, size);
1720 if (!--nr_io_queues)
1723 adminq->q_db = dev->dbs;
1725 /* Deregister the admin queue's interrupt */
1726 pci_free_irq(pdev, 0, adminq);
1729 * If we enable msix early due to not intx, disable it again before
1730 * setting up the full range we need.
1732 pci_free_irq_vectors(pdev);
1733 nr_io_queues = pci_alloc_irq_vectors(pdev, 1, nr_io_queues,
1734 PCI_IRQ_ALL_TYPES | PCI_IRQ_AFFINITY);
1735 if (nr_io_queues <= 0)
1737 dev->max_qid = nr_io_queues;
1740 * Should investigate if there's a performance win from allocating
1741 * more queues than interrupt vectors; it might allow the submission
1742 * path to scale better, even if the receive path is limited by the
1743 * number of interrupts.
1746 result = queue_request_irq(adminq);
1748 adminq->cq_vector = -1;
1751 return nvme_create_io_queues(dev);
1754 static void nvme_del_queue_end(struct request *req, blk_status_t error)
1756 struct nvme_queue *nvmeq = req->end_io_data;
1758 blk_mq_free_request(req);
1759 complete(&nvmeq->dev->ioq_wait);
1762 static void nvme_del_cq_end(struct request *req, blk_status_t error)
1764 struct nvme_queue *nvmeq = req->end_io_data;
1767 unsigned long flags;
1770 * We might be called with the AQ q_lock held
1771 * and the I/O queue q_lock should always
1772 * nest inside the AQ one.
1774 spin_lock_irqsave_nested(&nvmeq->q_lock, flags,
1775 SINGLE_DEPTH_NESTING);
1776 nvme_process_cq(nvmeq);
1777 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
1780 nvme_del_queue_end(req, error);
1783 static int nvme_delete_queue(struct nvme_queue *nvmeq, u8 opcode)
1785 struct request_queue *q = nvmeq->dev->ctrl.admin_q;
1786 struct request *req;
1787 struct nvme_command cmd;
1789 memset(&cmd, 0, sizeof(cmd));
1790 cmd.delete_queue.opcode = opcode;
1791 cmd.delete_queue.qid = cpu_to_le16(nvmeq->qid);
1793 req = nvme_alloc_request(q, &cmd, BLK_MQ_REQ_NOWAIT, NVME_QID_ANY);
1795 return PTR_ERR(req);
1797 req->timeout = ADMIN_TIMEOUT;
1798 req->end_io_data = nvmeq;
1800 blk_execute_rq_nowait(q, NULL, req, false,
1801 opcode == nvme_admin_delete_cq ?
1802 nvme_del_cq_end : nvme_del_queue_end);
1806 static void nvme_disable_io_queues(struct nvme_dev *dev, int queues)
1809 unsigned long timeout;
1810 u8 opcode = nvme_admin_delete_sq;
1812 for (pass = 0; pass < 2; pass++) {
1813 int sent = 0, i = queues;
1815 reinit_completion(&dev->ioq_wait);
1817 timeout = ADMIN_TIMEOUT;
1818 for (; i > 0; i--, sent++)
1819 if (nvme_delete_queue(dev->queues[i], opcode))
1823 timeout = wait_for_completion_io_timeout(&dev->ioq_wait, timeout);
1829 opcode = nvme_admin_delete_cq;
1834 * Return: error value if an error occurred setting up the queues or calling
1835 * Identify Device. 0 if these succeeded, even if adding some of the
1836 * namespaces failed. At the moment, these failures are silent. TBD which
1837 * failures should be reported.
1839 static int nvme_dev_add(struct nvme_dev *dev)
1841 if (!dev->ctrl.tagset) {
1842 dev->tagset.ops = &nvme_mq_ops;
1843 dev->tagset.nr_hw_queues = dev->online_queues - 1;
1844 dev->tagset.timeout = NVME_IO_TIMEOUT;
1845 dev->tagset.numa_node = dev_to_node(dev->dev);
1846 dev->tagset.queue_depth =
1847 min_t(int, dev->q_depth, BLK_MQ_MAX_DEPTH) - 1;
1848 dev->tagset.cmd_size = nvme_cmd_size(dev);
1849 dev->tagset.flags = BLK_MQ_F_SHOULD_MERGE;
1850 dev->tagset.driver_data = dev;
1852 if (blk_mq_alloc_tag_set(&dev->tagset))
1854 dev->ctrl.tagset = &dev->tagset;
1856 nvme_dbbuf_set(dev);
1858 blk_mq_update_nr_hw_queues(&dev->tagset, dev->online_queues - 1);
1860 /* Free previously allocated queues that are no longer usable */
1861 nvme_free_queues(dev, dev->online_queues);
1867 static int nvme_pci_enable(struct nvme_dev *dev)
1870 int result = -ENOMEM;
1871 struct pci_dev *pdev = to_pci_dev(dev->dev);
1873 if (pci_enable_device_mem(pdev))
1876 pci_set_master(pdev);
1878 if (dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(64)) &&
1879 dma_set_mask_and_coherent(dev->dev, DMA_BIT_MASK(32)))
1882 if (readl(dev->bar + NVME_REG_CSTS) == -1) {
1888 * Some devices and/or platforms don't advertise or work with INTx
1889 * interrupts. Pre-enable a single MSIX or MSI vec for setup. We'll
1890 * adjust this later.
1892 result = pci_alloc_irq_vectors(pdev, 1, 1, PCI_IRQ_ALL_TYPES);
1896 cap = lo_hi_readq(dev->bar + NVME_REG_CAP);
1898 dev->q_depth = min_t(int, NVME_CAP_MQES(cap) + 1, NVME_Q_DEPTH);
1899 dev->db_stride = 1 << NVME_CAP_STRIDE(cap);
1900 dev->dbs = dev->bar + 4096;
1903 * Temporary fix for the Apple controller found in the MacBook8,1 and
1904 * some MacBook7,1 to avoid controller resets and data loss.
1906 if (pdev->vendor == PCI_VENDOR_ID_APPLE && pdev->device == 0x2001) {
1908 dev_warn(dev->ctrl.device, "detected Apple NVMe controller, "
1909 "set queue depth=%u to work around controller resets\n",
1914 * CMBs can currently only exist on >=1.2 PCIe devices. We only
1915 * populate sysfs if a CMB is implemented. Note that we add the
1916 * CMB attribute to the nvme_ctrl kobj which removes the need to remove
1917 * it on exit. Since nvme_dev_attrs_group has no name we can pass
1918 * NULL as final argument to sysfs_add_file_to_group.
1921 if (readl(dev->bar + NVME_REG_VS) >= NVME_VS(1, 2, 0)) {
1922 dev->cmb = nvme_map_cmb(dev);
1925 if (sysfs_add_file_to_group(&dev->ctrl.device->kobj,
1926 &dev_attr_cmb.attr, NULL))
1927 dev_warn(dev->ctrl.device,
1928 "failed to add sysfs attribute for CMB\n");
1932 pci_enable_pcie_error_reporting(pdev);
1933 pci_save_state(pdev);
1937 pci_disable_device(pdev);
1941 static void nvme_dev_unmap(struct nvme_dev *dev)
1945 pci_release_mem_regions(to_pci_dev(dev->dev));
1948 static void nvme_pci_disable(struct nvme_dev *dev)
1950 struct pci_dev *pdev = to_pci_dev(dev->dev);
1952 nvme_release_cmb(dev);
1953 pci_free_irq_vectors(pdev);
1955 if (pci_is_enabled(pdev)) {
1956 pci_disable_pcie_error_reporting(pdev);
1957 pci_disable_device(pdev);
1961 static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown)
1965 struct pci_dev *pdev = to_pci_dev(dev->dev);
1967 mutex_lock(&dev->shutdown_lock);
1968 if (pci_is_enabled(pdev)) {
1969 u32 csts = readl(dev->bar + NVME_REG_CSTS);
1971 if (dev->ctrl.state == NVME_CTRL_LIVE ||
1972 dev->ctrl.state == NVME_CTRL_RESETTING)
1973 nvme_start_freeze(&dev->ctrl);
1974 dead = !!((csts & NVME_CSTS_CFS) || !(csts & NVME_CSTS_RDY) ||
1975 pdev->error_state != pci_channel_io_normal);
1979 * Give the controller a chance to complete all entered requests if
1980 * doing a safe shutdown.
1984 nvme_wait_freeze_timeout(&dev->ctrl, NVME_IO_TIMEOUT);
1987 * If the controller is still alive tell it to stop using the
1988 * host memory buffer. In theory the shutdown / reset should
1989 * make sure that it doesn't access the host memoery anymore,
1990 * but I'd rather be safe than sorry..
1992 if (dev->host_mem_descs)
1993 nvme_set_host_mem(dev, 0);
1996 nvme_stop_queues(&dev->ctrl);
1998 queues = dev->online_queues - 1;
1999 for (i = dev->queue_count - 1; i > 0; i--)
2000 nvme_suspend_queue(dev->queues[i]);
2003 /* A device might become IO incapable very soon during
2004 * probe, before the admin queue is configured. Thus,
2005 * queue_count can be 0 here.
2007 if (dev->queue_count)
2008 nvme_suspend_queue(dev->queues[0]);
2010 nvme_disable_io_queues(dev, queues);
2011 nvme_disable_admin_queue(dev, shutdown);
2013 nvme_pci_disable(dev);
2015 blk_mq_tagset_busy_iter(&dev->tagset, nvme_cancel_request, &dev->ctrl);
2016 blk_mq_tagset_busy_iter(&dev->admin_tagset, nvme_cancel_request, &dev->ctrl);
2019 * The driver will not be starting up queues again if shutting down so
2020 * must flush all entered requests to their failed completion to avoid
2021 * deadlocking blk-mq hot-cpu notifier.
2024 nvme_start_queues(&dev->ctrl);
2025 mutex_unlock(&dev->shutdown_lock);
2028 static int nvme_setup_prp_pools(struct nvme_dev *dev)
2030 dev->prp_page_pool = dma_pool_create("prp list page", dev->dev,
2031 PAGE_SIZE, PAGE_SIZE, 0);
2032 if (!dev->prp_page_pool)
2035 /* Optimisation for I/Os between 4k and 128k */
2036 dev->prp_small_pool = dma_pool_create("prp list 256", dev->dev,
2038 if (!dev->prp_small_pool) {
2039 dma_pool_destroy(dev->prp_page_pool);
2045 static void nvme_release_prp_pools(struct nvme_dev *dev)
2047 dma_pool_destroy(dev->prp_page_pool);
2048 dma_pool_destroy(dev->prp_small_pool);
2051 static void nvme_pci_free_ctrl(struct nvme_ctrl *ctrl)
2053 struct nvme_dev *dev = to_nvme_dev(ctrl);
2055 nvme_dbbuf_dma_free(dev);
2056 put_device(dev->dev);
2057 if (dev->tagset.tags)
2058 blk_mq_free_tag_set(&dev->tagset);
2059 if (dev->ctrl.admin_q)
2060 blk_put_queue(dev->ctrl.admin_q);
2062 free_opal_dev(dev->ctrl.opal_dev);
2066 static void nvme_remove_dead_ctrl(struct nvme_dev *dev, int status)
2068 dev_warn(dev->ctrl.device, "Removing after probe failure status: %d\n", status);
2070 kref_get(&dev->ctrl.kref);
2071 nvme_dev_disable(dev, false);
2072 if (!schedule_work(&dev->remove_work))
2073 nvme_put_ctrl(&dev->ctrl);
2076 static void nvme_reset_work(struct work_struct *work)
2078 struct nvme_dev *dev =
2079 container_of(work, struct nvme_dev, ctrl.reset_work);
2080 bool was_suspend = !!(dev->ctrl.ctrl_config & NVME_CC_SHN_NORMAL);
2081 int result = -ENODEV;
2083 if (WARN_ON(dev->ctrl.state != NVME_CTRL_RESETTING))
2087 * If we're called to reset a live controller first shut it down before
2090 if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
2091 nvme_dev_disable(dev, false);
2093 result = nvme_pci_enable(dev);
2097 result = nvme_configure_admin_queue(dev);
2101 nvme_init_queue(dev->queues[0], 0);
2102 result = nvme_alloc_admin_tags(dev);
2106 result = nvme_init_identify(&dev->ctrl);
2110 if (dev->ctrl.oacs & NVME_CTRL_OACS_SEC_SUPP) {
2111 if (!dev->ctrl.opal_dev)
2112 dev->ctrl.opal_dev =
2113 init_opal_dev(&dev->ctrl, &nvme_sec_submit);
2114 else if (was_suspend)
2115 opal_unlock_from_suspend(dev->ctrl.opal_dev);
2117 free_opal_dev(dev->ctrl.opal_dev);
2118 dev->ctrl.opal_dev = NULL;
2121 if (dev->ctrl.oacs & NVME_CTRL_OACS_DBBUF_SUPP) {
2122 result = nvme_dbbuf_dma_alloc(dev);
2125 "unable to allocate dma for dbbuf\n");
2128 if (dev->ctrl.hmpre)
2129 nvme_setup_host_mem(dev);
2131 result = nvme_setup_io_queues(dev);
2136 * A controller that can not execute IO typically requires user
2137 * intervention to correct. For such degraded controllers, the driver
2138 * should not submit commands the user did not request, so skip
2139 * registering for asynchronous event notification on this condition.
2141 if (dev->online_queues > 1)
2142 nvme_queue_async_events(&dev->ctrl);
2145 * Keep the controller around but remove all namespaces if we don't have
2146 * any working I/O queue.
2148 if (dev->online_queues < 2) {
2149 dev_warn(dev->ctrl.device, "IO queues not created\n");
2150 nvme_kill_queues(&dev->ctrl);
2151 nvme_remove_namespaces(&dev->ctrl);
2153 nvme_start_queues(&dev->ctrl);
2154 nvme_wait_freeze(&dev->ctrl);
2156 nvme_unfreeze(&dev->ctrl);
2159 if (!nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_LIVE)) {
2160 dev_warn(dev->ctrl.device, "failed to mark controller live\n");
2164 if (dev->online_queues > 1)
2165 nvme_queue_scan(&dev->ctrl);
2169 nvme_remove_dead_ctrl(dev, result);
2172 static void nvme_remove_dead_ctrl_work(struct work_struct *work)
2174 struct nvme_dev *dev = container_of(work, struct nvme_dev, remove_work);
2175 struct pci_dev *pdev = to_pci_dev(dev->dev);
2177 nvme_kill_queues(&dev->ctrl);
2178 if (pci_get_drvdata(pdev))
2179 device_release_driver(&pdev->dev);
2180 nvme_put_ctrl(&dev->ctrl);
2183 static int nvme_pci_reg_read32(struct nvme_ctrl *ctrl, u32 off, u32 *val)
2185 *val = readl(to_nvme_dev(ctrl)->bar + off);
2189 static int nvme_pci_reg_write32(struct nvme_ctrl *ctrl, u32 off, u32 val)
2191 writel(val, to_nvme_dev(ctrl)->bar + off);
2195 static int nvme_pci_reg_read64(struct nvme_ctrl *ctrl, u32 off, u64 *val)
2197 *val = readq(to_nvme_dev(ctrl)->bar + off);
2201 static const struct nvme_ctrl_ops nvme_pci_ctrl_ops = {
2203 .module = THIS_MODULE,
2204 .flags = NVME_F_METADATA_SUPPORTED,
2205 .reg_read32 = nvme_pci_reg_read32,
2206 .reg_write32 = nvme_pci_reg_write32,
2207 .reg_read64 = nvme_pci_reg_read64,
2208 .free_ctrl = nvme_pci_free_ctrl,
2209 .submit_async_event = nvme_pci_submit_async_event,
2212 static int nvme_dev_map(struct nvme_dev *dev)
2214 struct pci_dev *pdev = to_pci_dev(dev->dev);
2216 if (pci_request_mem_regions(pdev, "nvme"))
2219 if (nvme_remap_bar(dev, NVME_REG_DBS + 4096))
2224 pci_release_mem_regions(pdev);
2228 static unsigned long check_dell_samsung_bug(struct pci_dev *pdev)
2230 if (pdev->vendor == 0x144d && pdev->device == 0xa802) {
2232 * Several Samsung devices seem to drop off the PCIe bus
2233 * randomly when APST is on and uses the deepest sleep state.
2234 * This has been observed on a Samsung "SM951 NVMe SAMSUNG
2235 * 256GB", a "PM951 NVMe SAMSUNG 512GB", and a "Samsung SSD
2236 * 950 PRO 256GB", but it seems to be restricted to two Dell
2239 if (dmi_match(DMI_SYS_VENDOR, "Dell Inc.") &&
2240 (dmi_match(DMI_PRODUCT_NAME, "XPS 15 9550") ||
2241 dmi_match(DMI_PRODUCT_NAME, "Precision 5510")))
2242 return NVME_QUIRK_NO_DEEPEST_PS;
2248 static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
2250 int node, result = -ENOMEM;
2251 struct nvme_dev *dev;
2252 unsigned long quirks = id->driver_data;
2254 node = dev_to_node(&pdev->dev);
2255 if (node == NUMA_NO_NODE)
2256 set_dev_node(&pdev->dev, first_memory_node);
2258 dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, node);
2261 dev->queues = kzalloc_node((num_possible_cpus() + 1) * sizeof(void *),
2266 dev->dev = get_device(&pdev->dev);
2267 pci_set_drvdata(pdev, dev);
2269 result = nvme_dev_map(dev);
2273 INIT_WORK(&dev->ctrl.reset_work, nvme_reset_work);
2274 INIT_WORK(&dev->remove_work, nvme_remove_dead_ctrl_work);
2275 mutex_init(&dev->shutdown_lock);
2276 init_completion(&dev->ioq_wait);
2278 result = nvme_setup_prp_pools(dev);
2282 quirks |= check_dell_samsung_bug(pdev);
2284 result = nvme_init_ctrl(&dev->ctrl, &pdev->dev, &nvme_pci_ctrl_ops,
2289 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_RESETTING);
2290 dev_info(dev->ctrl.device, "pci function %s\n", dev_name(&pdev->dev));
2292 queue_work(nvme_wq, &dev->ctrl.reset_work);
2296 nvme_release_prp_pools(dev);
2298 put_device(dev->dev);
2299 nvme_dev_unmap(dev);
2306 static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
2308 struct nvme_dev *dev = pci_get_drvdata(pdev);
2311 nvme_dev_disable(dev, false);
2313 nvme_reset_ctrl(&dev->ctrl);
2316 static void nvme_shutdown(struct pci_dev *pdev)
2318 struct nvme_dev *dev = pci_get_drvdata(pdev);
2319 nvme_dev_disable(dev, true);
2323 * The driver's remove may be called on a device in a partially initialized
2324 * state. This function must not have any dependencies on the device state in
2327 static void nvme_remove(struct pci_dev *pdev)
2329 struct nvme_dev *dev = pci_get_drvdata(pdev);
2331 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DELETING);
2333 cancel_work_sync(&dev->ctrl.reset_work);
2334 pci_set_drvdata(pdev, NULL);
2336 if (!pci_device_is_present(pdev)) {
2337 nvme_change_ctrl_state(&dev->ctrl, NVME_CTRL_DEAD);
2338 nvme_dev_disable(dev, false);
2341 flush_work(&dev->ctrl.reset_work);
2342 nvme_uninit_ctrl(&dev->ctrl);
2343 nvme_dev_disable(dev, true);
2344 nvme_free_host_mem(dev);
2345 nvme_dev_remove_admin(dev);
2346 nvme_free_queues(dev, 0);
2347 nvme_release_prp_pools(dev);
2348 nvme_dev_unmap(dev);
2349 nvme_put_ctrl(&dev->ctrl);
2352 static int nvme_pci_sriov_configure(struct pci_dev *pdev, int numvfs)
2357 if (pci_vfs_assigned(pdev)) {
2358 dev_warn(&pdev->dev,
2359 "Cannot disable SR-IOV VFs while assigned\n");
2362 pci_disable_sriov(pdev);
2366 ret = pci_enable_sriov(pdev, numvfs);
2367 return ret ? ret : numvfs;
2370 #ifdef CONFIG_PM_SLEEP
2371 static int nvme_suspend(struct device *dev)
2373 struct pci_dev *pdev = to_pci_dev(dev);
2374 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2376 nvme_dev_disable(ndev, true);
2380 static int nvme_resume(struct device *dev)
2382 struct pci_dev *pdev = to_pci_dev(dev);
2383 struct nvme_dev *ndev = pci_get_drvdata(pdev);
2385 nvme_reset_ctrl(&ndev->ctrl);
2390 static SIMPLE_DEV_PM_OPS(nvme_dev_pm_ops, nvme_suspend, nvme_resume);
2392 static pci_ers_result_t nvme_error_detected(struct pci_dev *pdev,
2393 pci_channel_state_t state)
2395 struct nvme_dev *dev = pci_get_drvdata(pdev);
2398 * A frozen channel requires a reset. When detected, this method will
2399 * shutdown the controller to quiesce. The controller will be restarted
2400 * after the slot reset through driver's slot_reset callback.
2403 case pci_channel_io_normal:
2404 return PCI_ERS_RESULT_CAN_RECOVER;
2405 case pci_channel_io_frozen:
2406 dev_warn(dev->ctrl.device,
2407 "frozen state error detected, reset controller\n");
2408 nvme_dev_disable(dev, false);
2409 return PCI_ERS_RESULT_NEED_RESET;
2410 case pci_channel_io_perm_failure:
2411 dev_warn(dev->ctrl.device,
2412 "failure state error detected, request disconnect\n");
2413 return PCI_ERS_RESULT_DISCONNECT;
2415 return PCI_ERS_RESULT_NEED_RESET;
2418 static pci_ers_result_t nvme_slot_reset(struct pci_dev *pdev)
2420 struct nvme_dev *dev = pci_get_drvdata(pdev);
2422 dev_info(dev->ctrl.device, "restart after slot reset\n");
2423 pci_restore_state(pdev);
2424 nvme_reset_ctrl(&dev->ctrl);
2425 return PCI_ERS_RESULT_RECOVERED;
2428 static void nvme_error_resume(struct pci_dev *pdev)
2430 pci_cleanup_aer_uncorrect_error_status(pdev);
2433 static const struct pci_error_handlers nvme_err_handler = {
2434 .error_detected = nvme_error_detected,
2435 .slot_reset = nvme_slot_reset,
2436 .resume = nvme_error_resume,
2437 .reset_notify = nvme_reset_notify,
2440 static const struct pci_device_id nvme_id_table[] = {
2441 { PCI_VDEVICE(INTEL, 0x0953),
2442 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2443 NVME_QUIRK_DEALLOCATE_ZEROES, },
2444 { PCI_VDEVICE(INTEL, 0x0a53),
2445 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2446 NVME_QUIRK_DEALLOCATE_ZEROES, },
2447 { PCI_VDEVICE(INTEL, 0x0a54),
2448 .driver_data = NVME_QUIRK_STRIPE_SIZE |
2449 NVME_QUIRK_DEALLOCATE_ZEROES, },
2450 { PCI_VDEVICE(INTEL, 0xf1a5), /* Intel 600P/P3100 */
2451 .driver_data = NVME_QUIRK_NO_DEEPEST_PS },
2452 { PCI_VDEVICE(INTEL, 0x5845), /* Qemu emulated controller */
2453 .driver_data = NVME_QUIRK_IDENTIFY_CNS, },
2454 { PCI_DEVICE(0x1c58, 0x0003), /* HGST adapter */
2455 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2456 { PCI_DEVICE(0x1c5f, 0x0540), /* Memblaze Pblaze4 adapter */
2457 .driver_data = NVME_QUIRK_DELAY_BEFORE_CHK_RDY, },
2458 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
2459 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2001) },
2460 { PCI_DEVICE(PCI_VENDOR_ID_APPLE, 0x2003) },
2463 MODULE_DEVICE_TABLE(pci, nvme_id_table);
2465 static struct pci_driver nvme_driver = {
2467 .id_table = nvme_id_table,
2468 .probe = nvme_probe,
2469 .remove = nvme_remove,
2470 .shutdown = nvme_shutdown,
2472 .pm = &nvme_dev_pm_ops,
2474 .sriov_configure = nvme_pci_sriov_configure,
2475 .err_handler = &nvme_err_handler,
2478 static int __init nvme_init(void)
2480 return pci_register_driver(&nvme_driver);
2483 static void __exit nvme_exit(void)
2485 pci_unregister_driver(&nvme_driver);
2489 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
2490 MODULE_LICENSE("GPL");
2491 MODULE_VERSION("1.0");
2492 module_init(nvme_init);
2493 module_exit(nvme_exit);