2 * NVM Express device driver
3 * Copyright (c) 2011, 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
14 * You should have received a copy of the GNU General Public License along with
15 * this program; if not, write to the Free Software Foundation, Inc.,
16 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
19 #include <linux/nvme.h>
20 #include <linux/bio.h>
21 #include <linux/blkdev.h>
22 #include <linux/errno.h>
24 #include <linux/genhd.h>
25 #include <linux/init.h>
26 #include <linux/interrupt.h>
28 #include <linux/kdev_t.h>
29 #include <linux/kernel.h>
31 #include <linux/module.h>
32 #include <linux/moduleparam.h>
33 #include <linux/pci.h>
34 #include <linux/poison.h>
35 #include <linux/sched.h>
36 #include <linux/slab.h>
37 #include <linux/types.h>
38 #include <linux/version.h>
40 #define NVME_Q_DEPTH 1024
41 #define SQ_SIZE(depth) (depth * sizeof(struct nvme_command))
42 #define CQ_SIZE(depth) (depth * sizeof(struct nvme_completion))
43 #define NVME_MINORS 64
44 #define IO_TIMEOUT (5 * HZ)
45 #define ADMIN_TIMEOUT (60 * HZ)
47 static int nvme_major;
48 module_param(nvme_major, int, 0);
50 static int use_threaded_interrupts;
51 module_param(use_threaded_interrupts, int, 0);
54 * Represents an NVM Express device. Each nvme_dev is a PCI function.
57 struct nvme_queue **queues;
59 struct pci_dev *pci_dev;
60 struct dma_pool *prp_page_pool;
61 struct dma_pool *prp_small_pool;
65 struct msix_entry *entry;
66 struct nvme_bar __iomem *bar;
67 struct list_head namespaces;
74 * An NVM Express namespace is equivalent to a SCSI LUN
77 struct list_head list;
80 struct request_queue *queue;
88 * An NVM Express queue. Each device has at least two (one for admin
89 * commands and one for I/O commands).
92 struct device *q_dmadev;
95 struct nvme_command *sq_cmds;
96 volatile struct nvme_completion *cqes;
97 dma_addr_t sq_dma_addr;
98 dma_addr_t cq_dma_addr;
99 wait_queue_head_t sq_full;
100 struct bio_list sq_cong;
108 unsigned long cmdid_data[];
111 static void nvme_resubmit_bio(struct nvme_queue *nvmeq, struct bio *bio);
114 * Check we didin't inadvertently grow the command struct
116 static inline void _nvme_check_size(void)
118 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
119 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
120 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
121 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
122 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
123 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
124 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
125 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
126 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
129 struct nvme_cmd_info {
131 unsigned long timeout;
134 static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
136 return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
140 * alloc_cmdid - Allocate a Command ID
141 * @param nvmeq The queue that will be used for this command
142 * @param ctx A pointer that will be passed to the handler
143 * @param handler The ID of the handler to call
145 * Allocate a Command ID for a queue. The data passed in will
146 * be passed to the completion handler. This is implemented by using
147 * the bottom two bits of the ctx pointer to store the handler ID.
148 * Passing in a pointer that's not 4-byte aligned will cause a BUG.
149 * We can change this if it becomes a problem.
151 static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx, int handler,
154 int depth = nvmeq->q_depth;
155 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
158 BUG_ON((unsigned long)ctx & 3);
161 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
164 } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
166 info[cmdid].ctx = (unsigned long)ctx | handler;
167 info[cmdid].timeout = jiffies + timeout;
171 static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
172 int handler, unsigned timeout)
175 wait_event_killable(nvmeq->sq_full,
176 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
177 return (cmdid < 0) ? -EINTR : cmdid;
180 /* If you need more than four handlers, you'll need to change how
181 * alloc_cmdid and nvme_process_cq work. Consider using a special
182 * CMD_CTX value instead, if that works for your situation.
185 sync_completion_id = 0,
189 #define CMD_CTX_BASE (POISON_POINTER_DELTA + sync_completion_id)
190 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
191 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
192 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
194 static unsigned long free_cmdid(struct nvme_queue *nvmeq, int cmdid)
197 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
199 if (cmdid >= nvmeq->q_depth)
200 return CMD_CTX_INVALID;
201 data = info[cmdid].ctx;
202 info[cmdid].ctx = CMD_CTX_COMPLETED;
203 clear_bit(cmdid, nvmeq->cmdid_data);
204 wake_up(&nvmeq->sq_full);
208 static void cancel_cmdid_data(struct nvme_queue *nvmeq, int cmdid)
210 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
211 info[cmdid].ctx = CMD_CTX_CANCELLED;
214 static struct nvme_queue *get_nvmeq(struct nvme_ns *ns)
216 int qid, cpu = get_cpu();
217 if (cpu < ns->dev->queue_count)
220 qid = (cpu % rounddown_pow_of_two(ns->dev->queue_count)) + 1;
221 return ns->dev->queues[qid];
224 static void put_nvmeq(struct nvme_queue *nvmeq)
230 * nvme_submit_cmd: Copy a command into a queue and ring the doorbell
231 * @nvmeq: The queue to use
232 * @cmd: The command to send
234 * Safe to use from interrupt context
236 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
240 /* XXX: Need to check tail isn't going to overrun head */
241 spin_lock_irqsave(&nvmeq->q_lock, flags);
242 tail = nvmeq->sq_tail;
243 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
244 writel(tail, nvmeq->q_db);
245 if (++tail == nvmeq->q_depth)
247 nvmeq->sq_tail = tail;
248 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
255 dma_addr_t first_dma;
259 static void nvme_free_prps(struct nvme_dev *dev, struct nvme_prps *prps)
261 const int last_prp = PAGE_SIZE / 8 - 1;
268 prp_dma = prps->first_dma;
270 if (prps->npages == 0)
271 dma_pool_free(dev->prp_small_pool, prps->list[0], prp_dma);
272 for (i = 0; i < prps->npages; i++) {
273 __le64 *prp_list = prps->list[i];
274 dma_addr_t next_prp_dma = le64_to_cpu(prp_list[last_prp]);
275 dma_pool_free(dev->prp_page_pool, prp_list, prp_dma);
276 prp_dma = next_prp_dma;
284 struct nvme_prps *prps;
285 struct scatterlist sg[0];
288 /* XXX: use a mempool */
289 static struct nvme_bio *alloc_nbio(unsigned nseg, gfp_t gfp)
291 return kzalloc(sizeof(struct nvme_bio) +
292 sizeof(struct scatterlist) * nseg, gfp);
295 static void free_nbio(struct nvme_queue *nvmeq, struct nvme_bio *nbio)
297 nvme_free_prps(nvmeq->dev, nbio->prps);
301 static void bio_completion(struct nvme_queue *nvmeq, void *ctx,
302 struct nvme_completion *cqe)
304 struct nvme_bio *nbio = ctx;
305 struct bio *bio = nbio->bio;
306 u16 status = le16_to_cpup(&cqe->status) >> 1;
308 dma_unmap_sg(nvmeq->q_dmadev, nbio->sg, nbio->nents,
309 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
310 free_nbio(nvmeq, nbio);
311 bio_endio(bio, status ? -EIO : 0);
312 bio = bio_list_pop(&nvmeq->sq_cong);
314 nvme_resubmit_bio(nvmeq, bio);
317 /* length is in bytes */
318 static struct nvme_prps *nvme_setup_prps(struct nvme_dev *dev,
319 struct nvme_common_command *cmd,
320 struct scatterlist *sg, int length)
322 struct dma_pool *pool;
323 int dma_len = sg_dma_len(sg);
324 u64 dma_addr = sg_dma_address(sg);
325 int offset = offset_in_page(dma_addr);
328 int nprps, npages, i, prp_page;
329 struct nvme_prps *prps = NULL;
331 cmd->prp1 = cpu_to_le64(dma_addr);
332 length -= (PAGE_SIZE - offset);
336 dma_len -= (PAGE_SIZE - offset);
338 dma_addr += (PAGE_SIZE - offset);
341 dma_addr = sg_dma_address(sg);
342 dma_len = sg_dma_len(sg);
345 if (length <= PAGE_SIZE) {
346 cmd->prp2 = cpu_to_le64(dma_addr);
350 nprps = DIV_ROUND_UP(length, PAGE_SIZE);
351 npages = DIV_ROUND_UP(8 * nprps, PAGE_SIZE);
352 prps = kmalloc(sizeof(*prps) + sizeof(__le64 *) * npages, GFP_ATOMIC);
354 if (nprps <= (256 / 8)) {
355 pool = dev->prp_small_pool;
358 pool = dev->prp_page_pool;
359 prps->npages = npages;
362 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
363 prps->list[prp_page++] = prp_list;
364 prps->first_dma = prp_dma;
365 cmd->prp2 = cpu_to_le64(prp_dma);
368 if (i == PAGE_SIZE / 8 - 1) {
369 __le64 *old_prp_list = prp_list;
370 prp_list = dma_pool_alloc(pool, GFP_ATOMIC, &prp_dma);
371 prps->list[prp_page++] = prp_list;
372 old_prp_list[i] = cpu_to_le64(prp_dma);
375 prp_list[i++] = cpu_to_le64(dma_addr);
376 dma_len -= PAGE_SIZE;
377 dma_addr += PAGE_SIZE;
385 dma_addr = sg_dma_address(sg);
386 dma_len = sg_dma_len(sg);
392 static int nvme_map_bio(struct device *dev, struct nvme_bio *nbio,
393 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
395 struct bio_vec *bvec, *bvprv = NULL;
396 struct scatterlist *sg = NULL;
399 sg_init_table(nbio->sg, psegs);
400 bio_for_each_segment(bvec, bio, i) {
401 if (bvprv && BIOVEC_PHYS_MERGEABLE(bvprv, bvec)) {
402 sg->length += bvec->bv_len;
404 /* Check bvprv && offset == 0 */
405 sg = sg ? sg + 1 : nbio->sg;
406 sg_set_page(sg, bvec->bv_page, bvec->bv_len,
414 return dma_map_sg(dev, nbio->sg, nbio->nents, dma_dir);
417 static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
420 struct nvme_command *cmnd;
421 struct nvme_bio *nbio;
422 enum dma_data_direction dma_dir;
427 int psegs = bio_phys_segments(ns->queue, bio);
429 nbio = alloc_nbio(psegs, GFP_NOIO);
434 cmdid = alloc_cmdid(nvmeq, nbio, bio_completion_id, IO_TIMEOUT);
435 if (unlikely(cmdid < 0))
439 if (bio->bi_rw & REQ_FUA)
440 control |= NVME_RW_FUA;
441 if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
442 control |= NVME_RW_LR;
445 if (bio->bi_rw & REQ_RAHEAD)
446 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
448 spin_lock_irqsave(&nvmeq->q_lock, flags);
449 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
451 memset(cmnd, 0, sizeof(*cmnd));
452 if (bio_data_dir(bio)) {
453 cmnd->rw.opcode = nvme_cmd_write;
454 dma_dir = DMA_TO_DEVICE;
456 cmnd->rw.opcode = nvme_cmd_read;
457 dma_dir = DMA_FROM_DEVICE;
460 if (nvme_map_bio(nvmeq->q_dmadev, nbio, bio, dma_dir, psegs) == 0)
464 cmnd->rw.command_id = cmdid;
465 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
466 nbio->prps = nvme_setup_prps(nvmeq->dev, &cmnd->common, nbio->sg,
468 cmnd->rw.slba = cpu_to_le64(bio->bi_sector >> (ns->lba_shift - 9));
469 cmnd->rw.length = cpu_to_le16((bio->bi_size >> ns->lba_shift) - 1);
470 cmnd->rw.control = cpu_to_le16(control);
471 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
473 writel(nvmeq->sq_tail, nvmeq->q_db);
474 if (++nvmeq->sq_tail == nvmeq->q_depth)
477 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
482 free_nbio(nvmeq, nbio);
483 bio_endio(bio, -ENOMEM);
487 free_nbio(nvmeq, nbio);
492 static void nvme_resubmit_bio(struct nvme_queue *nvmeq, struct bio *bio)
494 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
495 if (nvme_submit_bio_queue(nvmeq, ns, bio))
496 bio_list_add_head(&nvmeq->sq_cong, bio);
497 else if (bio_list_empty(&nvmeq->sq_cong))
498 blk_clear_queue_congested(ns->queue, rw_is_sync(bio->bi_rw));
499 /* XXX: Need to duplicate the logic from __freed_request here */
503 * NB: return value of non-zero would mean that we were a stacking driver.
504 * make_request must always succeed.
506 static int nvme_make_request(struct request_queue *q, struct bio *bio)
508 struct nvme_ns *ns = q->queuedata;
509 struct nvme_queue *nvmeq = get_nvmeq(ns);
511 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
512 blk_set_queue_congested(q, rw_is_sync(bio->bi_rw));
513 spin_lock_irq(&nvmeq->q_lock);
514 bio_list_add(&nvmeq->sq_cong, bio);
515 spin_unlock_irq(&nvmeq->q_lock);
522 struct sync_cmd_info {
523 struct task_struct *task;
528 static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
529 struct nvme_completion *cqe)
531 struct sync_cmd_info *cmdinfo = ctx;
532 if ((unsigned long)cmdinfo == CMD_CTX_CANCELLED)
534 if (unlikely((unsigned long)cmdinfo == CMD_CTX_COMPLETED)) {
535 dev_warn(nvmeq->q_dmadev,
536 "completed id %d twice on queue %d\n",
537 cqe->command_id, le16_to_cpup(&cqe->sq_id));
540 if (unlikely((unsigned long)cmdinfo == CMD_CTX_INVALID)) {
541 dev_warn(nvmeq->q_dmadev,
542 "invalid id %d completed on queue %d\n",
543 cqe->command_id, le16_to_cpup(&cqe->sq_id));
546 cmdinfo->result = le32_to_cpup(&cqe->result);
547 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
548 wake_up_process(cmdinfo->task);
551 typedef void (*completion_fn)(struct nvme_queue *, void *,
552 struct nvme_completion *);
554 static irqreturn_t nvme_process_cq(struct nvme_queue *nvmeq)
558 static const completion_fn completions[4] = {
559 [sync_completion_id] = sync_completion,
560 [bio_completion_id] = bio_completion,
563 head = nvmeq->cq_head;
564 phase = nvmeq->cq_phase;
569 unsigned char handler;
570 struct nvme_completion cqe = nvmeq->cqes[head];
571 if ((le16_to_cpu(cqe.status) & 1) != phase)
573 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
574 if (++head == nvmeq->q_depth) {
579 data = free_cmdid(nvmeq, cqe.command_id);
581 ptr = (void *)(data & ~3UL);
582 completions[handler](nvmeq, ptr, &cqe);
585 /* If the controller ignores the cq head doorbell and continuously
586 * writes to the queue, it is theoretically possible to wrap around
587 * the queue twice and mistakenly return IRQ_NONE. Linux only
588 * requires that 0.1% of your interrupts are handled, so this isn't
591 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
594 writel(head, nvmeq->q_db + 1);
595 nvmeq->cq_head = head;
596 nvmeq->cq_phase = phase;
601 static irqreturn_t nvme_irq(int irq, void *data)
604 struct nvme_queue *nvmeq = data;
605 spin_lock(&nvmeq->q_lock);
606 result = nvme_process_cq(nvmeq);
607 spin_unlock(&nvmeq->q_lock);
611 static irqreturn_t nvme_irq_check(int irq, void *data)
613 struct nvme_queue *nvmeq = data;
614 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
615 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
617 return IRQ_WAKE_THREAD;
620 static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
622 spin_lock_irq(&nvmeq->q_lock);
623 cancel_cmdid_data(nvmeq, cmdid);
624 spin_unlock_irq(&nvmeq->q_lock);
628 * Returns 0 on success. If the result is negative, it's a Linux error code;
629 * if the result is positive, it's an NVM Express status code
631 static int nvme_submit_sync_cmd(struct nvme_queue *nvmeq,
632 struct nvme_command *cmd, u32 *result, unsigned timeout)
635 struct sync_cmd_info cmdinfo;
637 cmdinfo.task = current;
638 cmdinfo.status = -EINTR;
640 cmdid = alloc_cmdid_killable(nvmeq, &cmdinfo, sync_completion_id,
644 cmd->common.command_id = cmdid;
646 set_current_state(TASK_KILLABLE);
647 nvme_submit_cmd(nvmeq, cmd);
650 if (cmdinfo.status == -EINTR) {
651 nvme_abort_command(nvmeq, cmdid);
656 *result = cmdinfo.result;
658 return cmdinfo.status;
661 static int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
664 return nvme_submit_sync_cmd(dev->queues[0], cmd, result, ADMIN_TIMEOUT);
667 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
670 struct nvme_command c;
672 memset(&c, 0, sizeof(c));
673 c.delete_queue.opcode = opcode;
674 c.delete_queue.qid = cpu_to_le16(id);
676 status = nvme_submit_admin_cmd(dev, &c, NULL);
682 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
683 struct nvme_queue *nvmeq)
686 struct nvme_command c;
687 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
689 memset(&c, 0, sizeof(c));
690 c.create_cq.opcode = nvme_admin_create_cq;
691 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
692 c.create_cq.cqid = cpu_to_le16(qid);
693 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
694 c.create_cq.cq_flags = cpu_to_le16(flags);
695 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
697 status = nvme_submit_admin_cmd(dev, &c, NULL);
703 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
704 struct nvme_queue *nvmeq)
707 struct nvme_command c;
708 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
710 memset(&c, 0, sizeof(c));
711 c.create_sq.opcode = nvme_admin_create_sq;
712 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
713 c.create_sq.sqid = cpu_to_le16(qid);
714 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
715 c.create_sq.sq_flags = cpu_to_le16(flags);
716 c.create_sq.cqid = cpu_to_le16(qid);
718 status = nvme_submit_admin_cmd(dev, &c, NULL);
724 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
726 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
729 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
731 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
734 static void nvme_free_queue(struct nvme_dev *dev, int qid)
736 struct nvme_queue *nvmeq = dev->queues[qid];
738 free_irq(dev->entry[nvmeq->cq_vector].vector, nvmeq);
740 /* Don't tell the adapter to delete the admin queue */
742 adapter_delete_sq(dev, qid);
743 adapter_delete_cq(dev, qid);
746 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
747 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
748 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
749 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
753 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
754 int depth, int vector)
756 struct device *dmadev = &dev->pci_dev->dev;
757 unsigned extra = (depth / 8) + (depth * sizeof(struct nvme_cmd_info));
758 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
762 nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
763 &nvmeq->cq_dma_addr, GFP_KERNEL);
766 memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
768 nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
769 &nvmeq->sq_dma_addr, GFP_KERNEL);
773 nvmeq->q_dmadev = dmadev;
775 spin_lock_init(&nvmeq->q_lock);
778 init_waitqueue_head(&nvmeq->sq_full);
779 bio_list_init(&nvmeq->sq_cong);
780 nvmeq->q_db = &dev->dbs[qid * 2];
781 nvmeq->q_depth = depth;
782 nvmeq->cq_vector = vector;
787 dma_free_coherent(dmadev, CQ_SIZE(nvmeq->q_depth), (void *)nvmeq->cqes,
794 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
797 if (use_threaded_interrupts)
798 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
799 nvme_irq_check, nvme_irq,
800 IRQF_DISABLED | IRQF_SHARED,
802 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
803 IRQF_DISABLED | IRQF_SHARED, name, nvmeq);
806 static __devinit struct nvme_queue *nvme_create_queue(struct nvme_dev *dev,
807 int qid, int cq_size, int vector)
810 struct nvme_queue *nvmeq = nvme_alloc_queue(dev, qid, cq_size, vector);
815 result = adapter_alloc_cq(dev, qid, nvmeq);
819 result = adapter_alloc_sq(dev, qid, nvmeq);
823 result = queue_request_irq(dev, nvmeq, "nvme");
830 adapter_delete_sq(dev, qid);
832 adapter_delete_cq(dev, qid);
834 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
835 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
836 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
837 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
842 static int __devinit nvme_configure_admin_queue(struct nvme_dev *dev)
846 struct nvme_queue *nvmeq;
848 dev->dbs = ((void __iomem *)dev->bar) + 4096;
850 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
854 aqa = nvmeq->q_depth - 1;
857 dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
858 dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
859 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
861 writel(0, &dev->bar->cc);
862 writel(aqa, &dev->bar->aqa);
863 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
864 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
865 writel(dev->ctrl_config, &dev->bar->cc);
867 while (!(readl(&dev->bar->csts) & NVME_CSTS_RDY)) {
869 if (fatal_signal_pending(current))
873 result = queue_request_irq(dev, nvmeq, "nvme admin");
874 dev->queues[0] = nvmeq;
878 static int nvme_map_user_pages(struct nvme_dev *dev, int write,
879 unsigned long addr, unsigned length,
880 struct scatterlist **sgp)
882 int i, err, count, nents, offset;
883 struct scatterlist *sg;
891 offset = offset_in_page(addr);
892 count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
893 pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
895 err = get_user_pages_fast(addr, count, 1, pages);
902 sg = kcalloc(count, sizeof(*sg), GFP_KERNEL);
903 sg_init_table(sg, count);
904 sg_set_page(&sg[0], pages[0], PAGE_SIZE - offset, offset);
905 length -= (PAGE_SIZE - offset);
906 for (i = 1; i < count; i++) {
907 sg_set_page(&sg[i], pages[i], min_t(int, length, PAGE_SIZE), 0);
912 nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
913 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
922 for (i = 0; i < count; i++)
928 static void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
929 unsigned long addr, int length,
930 struct scatterlist *sg, int nents)
934 count = DIV_ROUND_UP(offset_in_page(addr) + length, PAGE_SIZE);
935 dma_unmap_sg(&dev->pci_dev->dev, sg, nents, DMA_FROM_DEVICE);
937 for (i = 0; i < count; i++)
938 put_page(sg_page(&sg[i]));
941 static int nvme_submit_user_admin_command(struct nvme_dev *dev,
942 unsigned long addr, unsigned length,
943 struct nvme_command *cmd)
946 struct scatterlist *sg;
947 struct nvme_prps *prps;
949 nents = nvme_map_user_pages(dev, 0, addr, length, &sg);
952 prps = nvme_setup_prps(dev, &cmd->common, sg, length);
953 err = nvme_submit_admin_cmd(dev, cmd, NULL);
954 nvme_unmap_user_pages(dev, 0, addr, length, sg, nents);
955 nvme_free_prps(dev, prps);
956 return err ? -EIO : 0;
959 static int nvme_identify(struct nvme_ns *ns, unsigned long addr, int cns)
961 struct nvme_command c;
963 memset(&c, 0, sizeof(c));
964 c.identify.opcode = nvme_admin_identify;
965 c.identify.nsid = cns ? 0 : cpu_to_le32(ns->ns_id);
966 c.identify.cns = cpu_to_le32(cns);
968 return nvme_submit_user_admin_command(ns->dev, addr, 4096, &c);
971 static int nvme_get_range_type(struct nvme_ns *ns, unsigned long addr)
973 struct nvme_command c;
975 memset(&c, 0, sizeof(c));
976 c.features.opcode = nvme_admin_get_features;
977 c.features.nsid = cpu_to_le32(ns->ns_id);
978 c.features.fid = cpu_to_le32(NVME_FEAT_LBA_RANGE);
980 return nvme_submit_user_admin_command(ns->dev, addr, 4096, &c);
983 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
985 struct nvme_dev *dev = ns->dev;
986 struct nvme_queue *nvmeq;
987 struct nvme_user_io io;
988 struct nvme_command c;
992 struct scatterlist *sg;
993 struct nvme_prps *prps;
995 if (copy_from_user(&io, uio, sizeof(io)))
997 length = io.nblocks << io.block_shift;
998 nents = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length, &sg);
1002 memset(&c, 0, sizeof(c));
1003 c.rw.opcode = io.opcode;
1004 c.rw.flags = io.flags;
1005 c.rw.nsid = cpu_to_le32(io.nsid);
1006 c.rw.slba = cpu_to_le64(io.slba);
1007 c.rw.length = cpu_to_le16(io.nblocks - 1);
1008 c.rw.control = cpu_to_le16(io.control);
1009 c.rw.dsmgmt = cpu_to_le16(io.dsmgmt);
1010 c.rw.reftag = cpu_to_le32(io.reftag); /* XXX: endian? */
1011 c.rw.apptag = cpu_to_le16(io.apptag);
1012 c.rw.appmask = cpu_to_le16(io.appmask);
1014 prps = nvme_setup_prps(dev, &c.common, sg, length);
1016 nvmeq = get_nvmeq(ns);
1017 /* Since nvme_submit_sync_cmd sleeps, we can't keep preemption
1018 * disabled. We may be preempted at any point, and be rescheduled
1019 * to a different CPU. That will cause cacheline bouncing, but no
1020 * additional races since q_lock already protects against other CPUs.
1023 status = nvme_submit_sync_cmd(nvmeq, &c, &result, IO_TIMEOUT);
1025 nvme_unmap_user_pages(dev, io.opcode & 1, io.addr, length, sg, nents);
1026 nvme_free_prps(dev, prps);
1027 put_user(result, &uio->result);
1031 static int nvme_download_firmware(struct nvme_ns *ns,
1032 struct nvme_dlfw __user *udlfw)
1034 struct nvme_dev *dev = ns->dev;
1035 struct nvme_dlfw dlfw;
1036 struct nvme_command c;
1038 struct scatterlist *sg;
1039 struct nvme_prps *prps;
1041 if (copy_from_user(&dlfw, udlfw, sizeof(dlfw)))
1043 if (dlfw.length >= (1 << 30))
1046 nents = nvme_map_user_pages(dev, 1, dlfw.addr, dlfw.length * 4, &sg);
1050 memset(&c, 0, sizeof(c));
1051 c.dlfw.opcode = nvme_admin_download_fw;
1052 c.dlfw.numd = cpu_to_le32(dlfw.length);
1053 c.dlfw.offset = cpu_to_le32(dlfw.offset);
1054 prps = nvme_setup_prps(dev, &c.common, sg, dlfw.length * 4);
1056 status = nvme_submit_admin_cmd(dev, &c, NULL);
1057 nvme_unmap_user_pages(dev, 0, dlfw.addr, dlfw.length * 4, sg, nents);
1058 nvme_free_prps(dev, prps);
1062 static int nvme_activate_firmware(struct nvme_ns *ns, unsigned long arg)
1064 struct nvme_dev *dev = ns->dev;
1065 struct nvme_command c;
1067 memset(&c, 0, sizeof(c));
1068 c.common.opcode = nvme_admin_activate_fw;
1069 c.common.rsvd10[0] = cpu_to_le32(arg);
1071 return nvme_submit_admin_cmd(dev, &c, NULL);
1074 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
1077 struct nvme_ns *ns = bdev->bd_disk->private_data;
1080 case NVME_IOCTL_IDENTIFY_NS:
1081 return nvme_identify(ns, arg, 0);
1082 case NVME_IOCTL_IDENTIFY_CTRL:
1083 return nvme_identify(ns, arg, 1);
1084 case NVME_IOCTL_GET_RANGE_TYPE:
1085 return nvme_get_range_type(ns, arg);
1086 case NVME_IOCTL_SUBMIT_IO:
1087 return nvme_submit_io(ns, (void __user *)arg);
1088 case NVME_IOCTL_DOWNLOAD_FW:
1089 return nvme_download_firmware(ns, (void __user *)arg);
1090 case NVME_IOCTL_ACTIVATE_FW:
1091 return nvme_activate_firmware(ns, arg);
1097 static const struct block_device_operations nvme_fops = {
1098 .owner = THIS_MODULE,
1099 .ioctl = nvme_ioctl,
1102 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, int index,
1103 struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1106 struct gendisk *disk;
1109 if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1112 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
1115 ns->queue = blk_alloc_queue(GFP_KERNEL);
1118 ns->queue->queue_flags = QUEUE_FLAG_DEFAULT | QUEUE_FLAG_NOMERGES |
1119 QUEUE_FLAG_NONROT | QUEUE_FLAG_DISCARD;
1120 blk_queue_make_request(ns->queue, nvme_make_request);
1122 ns->queue->queuedata = ns;
1124 disk = alloc_disk(NVME_MINORS);
1126 goto out_free_queue;
1129 lbaf = id->flbas & 0xf;
1130 ns->lba_shift = id->lbaf[lbaf].ds;
1132 disk->major = nvme_major;
1133 disk->minors = NVME_MINORS;
1134 disk->first_minor = NVME_MINORS * index;
1135 disk->fops = &nvme_fops;
1136 disk->private_data = ns;
1137 disk->queue = ns->queue;
1138 disk->driverfs_dev = &dev->pci_dev->dev;
1139 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, index);
1140 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1145 blk_cleanup_queue(ns->queue);
1151 static void nvme_ns_free(struct nvme_ns *ns)
1154 blk_cleanup_queue(ns->queue);
1158 static int set_queue_count(struct nvme_dev *dev, int count)
1162 struct nvme_command c;
1163 u32 q_count = (count - 1) | ((count - 1) << 16);
1165 memset(&c, 0, sizeof(c));
1166 c.features.opcode = nvme_admin_get_features;
1167 c.features.fid = cpu_to_le32(NVME_FEAT_NUM_QUEUES);
1168 c.features.dword11 = cpu_to_le32(q_count);
1170 status = nvme_submit_admin_cmd(dev, &c, &result);
1173 return min(result & 0xffff, result >> 16) + 1;
1176 static int __devinit nvme_setup_io_queues(struct nvme_dev *dev)
1178 int result, cpu, i, nr_queues;
1180 nr_queues = num_online_cpus();
1181 result = set_queue_count(dev, nr_queues);
1184 if (result < nr_queues)
1187 /* Deregister the admin queue's interrupt */
1188 free_irq(dev->entry[0].vector, dev->queues[0]);
1190 for (i = 0; i < nr_queues; i++)
1191 dev->entry[i].entry = i;
1193 result = pci_enable_msix(dev->pci_dev, dev->entry, nr_queues);
1196 } else if (result > 0) {
1205 result = queue_request_irq(dev, dev->queues[0], "nvme admin");
1206 /* XXX: handle failure here */
1208 cpu = cpumask_first(cpu_online_mask);
1209 for (i = 0; i < nr_queues; i++) {
1210 irq_set_affinity_hint(dev->entry[i].vector, get_cpu_mask(cpu));
1211 cpu = cpumask_next(cpu, cpu_online_mask);
1214 for (i = 0; i < nr_queues; i++) {
1215 dev->queues[i + 1] = nvme_create_queue(dev, i + 1,
1217 if (!dev->queues[i + 1])
1225 static void nvme_free_queues(struct nvme_dev *dev)
1229 for (i = dev->queue_count - 1; i >= 0; i--)
1230 nvme_free_queue(dev, i);
1233 static int __devinit nvme_dev_add(struct nvme_dev *dev)
1236 struct nvme_ns *ns, *next;
1237 struct nvme_id_ctrl *ctrl;
1239 dma_addr_t dma_addr;
1240 struct nvme_command cid, crt;
1242 res = nvme_setup_io_queues(dev);
1246 /* XXX: Switch to a SG list once prp2 works */
1247 id = dma_alloc_coherent(&dev->pci_dev->dev, 8192, &dma_addr,
1250 memset(&cid, 0, sizeof(cid));
1251 cid.identify.opcode = nvme_admin_identify;
1252 cid.identify.nsid = 0;
1253 cid.identify.prp1 = cpu_to_le64(dma_addr);
1254 cid.identify.cns = cpu_to_le32(1);
1256 res = nvme_submit_admin_cmd(dev, &cid, NULL);
1263 nn = le32_to_cpup(&ctrl->nn);
1264 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
1265 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
1266 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
1268 cid.identify.cns = 0;
1269 memset(&crt, 0, sizeof(crt));
1270 crt.features.opcode = nvme_admin_get_features;
1271 crt.features.prp1 = cpu_to_le64(dma_addr + 4096);
1272 crt.features.fid = cpu_to_le32(NVME_FEAT_LBA_RANGE);
1274 for (i = 0; i < nn; i++) {
1275 cid.identify.nsid = cpu_to_le32(i);
1276 res = nvme_submit_admin_cmd(dev, &cid, NULL);
1280 if (((struct nvme_id_ns *)id)->ncap == 0)
1283 crt.features.nsid = cpu_to_le32(i);
1284 res = nvme_submit_admin_cmd(dev, &crt, NULL);
1288 ns = nvme_alloc_ns(dev, i, id, id + 4096);
1290 list_add_tail(&ns->list, &dev->namespaces);
1292 list_for_each_entry(ns, &dev->namespaces, list)
1295 dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1299 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
1300 list_del(&ns->list);
1304 dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1308 static int nvme_dev_remove(struct nvme_dev *dev)
1310 struct nvme_ns *ns, *next;
1312 /* TODO: wait all I/O finished or cancel them */
1314 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
1315 list_del(&ns->list);
1316 del_gendisk(ns->disk);
1320 nvme_free_queues(dev);
1325 static int nvme_setup_prp_pools(struct nvme_dev *dev)
1327 struct device *dmadev = &dev->pci_dev->dev;
1328 dev->prp_page_pool = dma_pool_create("prp list page", dmadev,
1329 PAGE_SIZE, PAGE_SIZE, 0);
1330 if (!dev->prp_page_pool)
1333 /* Optimisation for I/Os between 4k and 128k */
1334 dev->prp_small_pool = dma_pool_create("prp list 256", dmadev,
1336 if (!dev->prp_small_pool) {
1337 dma_pool_destroy(dev->prp_page_pool);
1343 static void nvme_release_prp_pools(struct nvme_dev *dev)
1345 dma_pool_destroy(dev->prp_page_pool);
1346 dma_pool_destroy(dev->prp_small_pool);
1349 /* XXX: Use an ida or something to let remove / add work correctly */
1350 static void nvme_set_instance(struct nvme_dev *dev)
1352 static int instance;
1353 dev->instance = instance++;
1356 static void nvme_release_instance(struct nvme_dev *dev)
1360 static int __devinit nvme_probe(struct pci_dev *pdev,
1361 const struct pci_device_id *id)
1363 int bars, result = -ENOMEM;
1364 struct nvme_dev *dev;
1366 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1369 dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
1373 dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
1378 if (pci_enable_device_mem(pdev))
1380 pci_set_master(pdev);
1381 bars = pci_select_bars(pdev, IORESOURCE_MEM);
1382 if (pci_request_selected_regions(pdev, bars, "nvme"))
1385 INIT_LIST_HEAD(&dev->namespaces);
1386 dev->pci_dev = pdev;
1387 pci_set_drvdata(pdev, dev);
1388 dma_set_mask(&pdev->dev, DMA_BIT_MASK(64));
1389 dma_set_coherent_mask(&pdev->dev, DMA_BIT_MASK(64));
1390 nvme_set_instance(dev);
1391 dev->entry[0].vector = pdev->irq;
1393 result = nvme_setup_prp_pools(dev);
1397 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
1403 result = nvme_configure_admin_queue(dev);
1408 result = nvme_dev_add(dev);
1414 nvme_free_queues(dev);
1418 pci_disable_msix(pdev);
1419 nvme_release_instance(dev);
1420 nvme_release_prp_pools(dev);
1422 pci_disable_device(pdev);
1423 pci_release_regions(pdev);
1431 static void __devexit nvme_remove(struct pci_dev *pdev)
1433 struct nvme_dev *dev = pci_get_drvdata(pdev);
1434 nvme_dev_remove(dev);
1435 pci_disable_msix(pdev);
1437 nvme_release_instance(dev);
1438 nvme_release_prp_pools(dev);
1439 pci_disable_device(pdev);
1440 pci_release_regions(pdev);
1446 /* These functions are yet to be implemented */
1447 #define nvme_error_detected NULL
1448 #define nvme_dump_registers NULL
1449 #define nvme_link_reset NULL
1450 #define nvme_slot_reset NULL
1451 #define nvme_error_resume NULL
1452 #define nvme_suspend NULL
1453 #define nvme_resume NULL
1455 static struct pci_error_handlers nvme_err_handler = {
1456 .error_detected = nvme_error_detected,
1457 .mmio_enabled = nvme_dump_registers,
1458 .link_reset = nvme_link_reset,
1459 .slot_reset = nvme_slot_reset,
1460 .resume = nvme_error_resume,
1463 /* Move to pci_ids.h later */
1464 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
1466 static DEFINE_PCI_DEVICE_TABLE(nvme_id_table) = {
1467 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
1470 MODULE_DEVICE_TABLE(pci, nvme_id_table);
1472 static struct pci_driver nvme_driver = {
1474 .id_table = nvme_id_table,
1475 .probe = nvme_probe,
1476 .remove = __devexit_p(nvme_remove),
1477 .suspend = nvme_suspend,
1478 .resume = nvme_resume,
1479 .err_handler = &nvme_err_handler,
1482 static int __init nvme_init(void)
1486 nvme_major = register_blkdev(nvme_major, "nvme");
1487 if (nvme_major <= 0)
1490 result = pci_register_driver(&nvme_driver);
1494 unregister_blkdev(nvme_major, "nvme");
1498 static void __exit nvme_exit(void)
1500 pci_unregister_driver(&nvme_driver);
1501 unregister_blkdev(nvme_major, "nvme");
1504 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
1505 MODULE_LICENSE("GPL");
1506 MODULE_VERSION("0.2");
1507 module_init(nvme_init);
1508 module_exit(nvme_exit);