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;
63 struct msix_entry *entry;
64 struct nvme_bar __iomem *bar;
65 struct list_head namespaces;
72 * An NVM Express namespace is equivalent to a SCSI LUN
75 struct list_head list;
78 struct request_queue *queue;
86 * An NVM Express queue. Each device has at least two (one for admin
87 * commands and one for I/O commands).
90 struct device *q_dmadev;
92 struct nvme_command *sq_cmds;
93 volatile struct nvme_completion *cqes;
94 dma_addr_t sq_dma_addr;
95 dma_addr_t cq_dma_addr;
96 wait_queue_head_t sq_full;
97 struct bio_list sq_cong;
105 unsigned long cmdid_data[];
108 static void nvme_resubmit_bio(struct nvme_queue *nvmeq, struct bio *bio);
111 * Check we didin't inadvertently grow the command struct
113 static inline void _nvme_check_size(void)
115 BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
116 BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
117 BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
118 BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
119 BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
120 BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
121 BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
122 BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
123 BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
126 struct nvme_cmd_info {
128 unsigned long timeout;
131 static struct nvme_cmd_info *nvme_cmd_info(struct nvme_queue *nvmeq)
133 return (void *)&nvmeq->cmdid_data[BITS_TO_LONGS(nvmeq->q_depth)];
137 * alloc_cmdid - Allocate a Command ID
138 * @param nvmeq The queue that will be used for this command
139 * @param ctx A pointer that will be passed to the handler
140 * @param handler The ID of the handler to call
142 * Allocate a Command ID for a queue. The data passed in will
143 * be passed to the completion handler. This is implemented by using
144 * the bottom two bits of the ctx pointer to store the handler ID.
145 * Passing in a pointer that's not 4-byte aligned will cause a BUG.
146 * We can change this if it becomes a problem.
148 static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx, int handler,
151 int depth = nvmeq->q_depth;
152 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
155 BUG_ON((unsigned long)ctx & 3);
158 cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
161 } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
163 info[cmdid].ctx = (unsigned long)ctx | handler;
164 info[cmdid].timeout = jiffies + timeout;
168 static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
169 int handler, unsigned timeout)
172 wait_event_killable(nvmeq->sq_full,
173 (cmdid = alloc_cmdid(nvmeq, ctx, handler, timeout)) >= 0);
174 return (cmdid < 0) ? -EINTR : cmdid;
177 /* If you need more than four handlers, you'll need to change how
178 * alloc_cmdid and nvme_process_cq work. Consider using a special
179 * CMD_CTX value instead, if that works for your situation.
182 sync_completion_id = 0,
186 #define CMD_CTX_BASE (POISON_POINTER_DELTA + sync_completion_id)
187 #define CMD_CTX_CANCELLED (0x30C + CMD_CTX_BASE)
188 #define CMD_CTX_COMPLETED (0x310 + CMD_CTX_BASE)
189 #define CMD_CTX_INVALID (0x314 + CMD_CTX_BASE)
191 static unsigned long free_cmdid(struct nvme_queue *nvmeq, int cmdid)
194 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
196 if (cmdid >= nvmeq->q_depth)
197 return CMD_CTX_INVALID;
198 data = info[cmdid].ctx;
199 info[cmdid].ctx = CMD_CTX_COMPLETED;
200 clear_bit(cmdid, nvmeq->cmdid_data);
201 wake_up(&nvmeq->sq_full);
205 static void cancel_cmdid_data(struct nvme_queue *nvmeq, int cmdid)
207 struct nvme_cmd_info *info = nvme_cmd_info(nvmeq);
208 info[cmdid].ctx = CMD_CTX_CANCELLED;
211 static struct nvme_queue *get_nvmeq(struct nvme_ns *ns)
213 int qid, cpu = get_cpu();
214 if (cpu < ns->dev->queue_count)
217 qid = (cpu % rounddown_pow_of_two(ns->dev->queue_count)) + 1;
218 return ns->dev->queues[qid];
221 static void put_nvmeq(struct nvme_queue *nvmeq)
227 * nvme_submit_cmd: Copy a command into a queue and ring the doorbell
228 * @nvmeq: The queue to use
229 * @cmd: The command to send
231 * Safe to use from interrupt context
233 static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
237 /* XXX: Need to check tail isn't going to overrun head */
238 spin_lock_irqsave(&nvmeq->q_lock, flags);
239 tail = nvmeq->sq_tail;
240 memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
241 writel(tail, nvmeq->q_db);
242 if (++tail == nvmeq->q_depth)
244 nvmeq->sq_tail = tail;
245 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
250 struct nvme_req_info {
253 struct scatterlist sg[0];
256 /* XXX: use a mempool */
257 static struct nvme_req_info *alloc_info(unsigned nseg, gfp_t gfp)
259 return kmalloc(sizeof(struct nvme_req_info) +
260 sizeof(struct scatterlist) * nseg, gfp);
263 static void free_info(struct nvme_req_info *info)
268 static void bio_completion(struct nvme_queue *nvmeq, void *ctx,
269 struct nvme_completion *cqe)
271 struct nvme_req_info *info = ctx;
272 struct bio *bio = info->bio;
273 u16 status = le16_to_cpup(&cqe->status) >> 1;
275 dma_unmap_sg(nvmeq->q_dmadev, info->sg, info->nents,
276 bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
278 bio_endio(bio, status ? -EIO : 0);
279 bio = bio_list_pop(&nvmeq->sq_cong);
281 nvme_resubmit_bio(nvmeq, bio);
284 /* length is in bytes */
285 static void nvme_setup_prps(struct nvme_common_command *cmd,
286 struct scatterlist *sg, int length)
288 int dma_len = sg_dma_len(sg);
289 u64 dma_addr = sg_dma_address(sg);
290 int offset = offset_in_page(dma_addr);
292 cmd->prp1 = cpu_to_le64(dma_addr);
293 length -= (PAGE_SIZE - offset);
297 dma_len -= (PAGE_SIZE - offset);
299 dma_addr += (PAGE_SIZE - offset);
302 dma_addr = sg_dma_address(sg);
303 dma_len = sg_dma_len(sg);
306 if (length <= PAGE_SIZE) {
307 cmd->prp2 = cpu_to_le64(dma_addr);
311 /* XXX: support PRP lists */
314 static int nvme_map_bio(struct device *dev, struct nvme_req_info *info,
315 struct bio *bio, enum dma_data_direction dma_dir, int psegs)
317 struct bio_vec *bvec;
318 struct scatterlist *sg = info->sg;
321 sg_init_table(sg, psegs);
322 bio_for_each_segment(bvec, bio, i) {
323 sg_set_page(sg, bvec->bv_page, bvec->bv_len, bvec->bv_offset);
324 /* XXX: handle non-mergable here */
329 return dma_map_sg(dev, info->sg, info->nents, dma_dir);
332 static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
335 struct nvme_command *cmnd;
336 struct nvme_req_info *info;
337 enum dma_data_direction dma_dir;
342 int psegs = bio_phys_segments(ns->queue, bio);
344 info = alloc_info(psegs, GFP_NOIO);
349 cmdid = alloc_cmdid(nvmeq, info, bio_completion_id, IO_TIMEOUT);
350 if (unlikely(cmdid < 0))
354 if (bio->bi_rw & REQ_FUA)
355 control |= NVME_RW_FUA;
356 if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
357 control |= NVME_RW_LR;
360 if (bio->bi_rw & REQ_RAHEAD)
361 dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
363 spin_lock_irqsave(&nvmeq->q_lock, flags);
364 cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail];
366 memset(cmnd, 0, sizeof(*cmnd));
367 if (bio_data_dir(bio)) {
368 cmnd->rw.opcode = nvme_cmd_write;
369 dma_dir = DMA_TO_DEVICE;
371 cmnd->rw.opcode = nvme_cmd_read;
372 dma_dir = DMA_FROM_DEVICE;
375 nvme_map_bio(nvmeq->q_dmadev, info, bio, dma_dir, psegs);
378 cmnd->rw.command_id = cmdid;
379 cmnd->rw.nsid = cpu_to_le32(ns->ns_id);
380 nvme_setup_prps(&cmnd->common, info->sg, bio->bi_size);
381 cmnd->rw.slba = cpu_to_le64(bio->bi_sector >> (ns->lba_shift - 9));
382 cmnd->rw.length = cpu_to_le16((bio->bi_size >> ns->lba_shift) - 1);
383 cmnd->rw.control = cpu_to_le16(control);
384 cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
386 writel(nvmeq->sq_tail, nvmeq->q_db);
387 if (++nvmeq->sq_tail == nvmeq->q_depth)
390 spin_unlock_irqrestore(&nvmeq->q_lock, flags);
400 static void nvme_resubmit_bio(struct nvme_queue *nvmeq, struct bio *bio)
402 struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
403 if (nvme_submit_bio_queue(nvmeq, ns, bio))
404 bio_list_add_head(&nvmeq->sq_cong, bio);
405 else if (bio_list_empty(&nvmeq->sq_cong))
406 blk_clear_queue_congested(ns->queue, rw_is_sync(bio->bi_rw));
407 /* XXX: Need to duplicate the logic from __freed_request here */
411 * NB: return value of non-zero would mean that we were a stacking driver.
412 * make_request must always succeed.
414 static int nvme_make_request(struct request_queue *q, struct bio *bio)
416 struct nvme_ns *ns = q->queuedata;
417 struct nvme_queue *nvmeq = get_nvmeq(ns);
419 if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
420 blk_set_queue_congested(q, rw_is_sync(bio->bi_rw));
421 spin_lock_irq(&nvmeq->q_lock);
422 bio_list_add(&nvmeq->sq_cong, bio);
423 spin_unlock_irq(&nvmeq->q_lock);
430 struct sync_cmd_info {
431 struct task_struct *task;
436 static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
437 struct nvme_completion *cqe)
439 struct sync_cmd_info *cmdinfo = ctx;
440 if ((unsigned long)cmdinfo == CMD_CTX_CANCELLED)
442 if (unlikely((unsigned long)cmdinfo == CMD_CTX_COMPLETED)) {
443 dev_warn(nvmeq->q_dmadev,
444 "completed id %d twice on queue %d\n",
445 cqe->command_id, le16_to_cpup(&cqe->sq_id));
448 if (unlikely((unsigned long)cmdinfo == CMD_CTX_INVALID)) {
449 dev_warn(nvmeq->q_dmadev,
450 "invalid id %d completed on queue %d\n",
451 cqe->command_id, le16_to_cpup(&cqe->sq_id));
454 cmdinfo->result = le32_to_cpup(&cqe->result);
455 cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
456 wake_up_process(cmdinfo->task);
459 typedef void (*completion_fn)(struct nvme_queue *, void *,
460 struct nvme_completion *);
462 static irqreturn_t nvme_process_cq(struct nvme_queue *nvmeq)
466 static const completion_fn completions[4] = {
467 [sync_completion_id] = sync_completion,
468 [bio_completion_id] = bio_completion,
471 head = nvmeq->cq_head;
472 phase = nvmeq->cq_phase;
477 unsigned char handler;
478 struct nvme_completion cqe = nvmeq->cqes[head];
479 if ((le16_to_cpu(cqe.status) & 1) != phase)
481 nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
482 if (++head == nvmeq->q_depth) {
487 data = free_cmdid(nvmeq, cqe.command_id);
489 ptr = (void *)(data & ~3UL);
490 completions[handler](nvmeq, ptr, &cqe);
493 /* If the controller ignores the cq head doorbell and continuously
494 * writes to the queue, it is theoretically possible to wrap around
495 * the queue twice and mistakenly return IRQ_NONE. Linux only
496 * requires that 0.1% of your interrupts are handled, so this isn't
499 if (head == nvmeq->cq_head && phase == nvmeq->cq_phase)
502 writel(head, nvmeq->q_db + 1);
503 nvmeq->cq_head = head;
504 nvmeq->cq_phase = phase;
509 static irqreturn_t nvme_irq(int irq, void *data)
512 struct nvme_queue *nvmeq = data;
513 spin_lock(&nvmeq->q_lock);
514 result = nvme_process_cq(nvmeq);
515 spin_unlock(&nvmeq->q_lock);
519 static irqreturn_t nvme_irq_check(int irq, void *data)
521 struct nvme_queue *nvmeq = data;
522 struct nvme_completion cqe = nvmeq->cqes[nvmeq->cq_head];
523 if ((le16_to_cpu(cqe.status) & 1) != nvmeq->cq_phase)
525 return IRQ_WAKE_THREAD;
528 static void nvme_abort_command(struct nvme_queue *nvmeq, int cmdid)
530 spin_lock_irq(&nvmeq->q_lock);
531 cancel_cmdid_data(nvmeq, cmdid);
532 spin_unlock_irq(&nvmeq->q_lock);
536 * Returns 0 on success. If the result is negative, it's a Linux error code;
537 * if the result is positive, it's an NVM Express status code
539 static int nvme_submit_sync_cmd(struct nvme_queue *nvmeq,
540 struct nvme_command *cmd, u32 *result, unsigned timeout)
543 struct sync_cmd_info cmdinfo;
545 cmdinfo.task = current;
546 cmdinfo.status = -EINTR;
548 cmdid = alloc_cmdid_killable(nvmeq, &cmdinfo, sync_completion_id,
552 cmd->common.command_id = cmdid;
554 set_current_state(TASK_KILLABLE);
555 nvme_submit_cmd(nvmeq, cmd);
558 if (cmdinfo.status == -EINTR) {
559 nvme_abort_command(nvmeq, cmdid);
564 *result = cmdinfo.result;
566 return cmdinfo.status;
569 static int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
572 return nvme_submit_sync_cmd(dev->queues[0], cmd, result, ADMIN_TIMEOUT);
575 static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
578 struct nvme_command c;
580 memset(&c, 0, sizeof(c));
581 c.delete_queue.opcode = opcode;
582 c.delete_queue.qid = cpu_to_le16(id);
584 status = nvme_submit_admin_cmd(dev, &c, NULL);
590 static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
591 struct nvme_queue *nvmeq)
594 struct nvme_command c;
595 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
597 memset(&c, 0, sizeof(c));
598 c.create_cq.opcode = nvme_admin_create_cq;
599 c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
600 c.create_cq.cqid = cpu_to_le16(qid);
601 c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
602 c.create_cq.cq_flags = cpu_to_le16(flags);
603 c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
605 status = nvme_submit_admin_cmd(dev, &c, NULL);
611 static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
612 struct nvme_queue *nvmeq)
615 struct nvme_command c;
616 int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
618 memset(&c, 0, sizeof(c));
619 c.create_sq.opcode = nvme_admin_create_sq;
620 c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
621 c.create_sq.sqid = cpu_to_le16(qid);
622 c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
623 c.create_sq.sq_flags = cpu_to_le16(flags);
624 c.create_sq.cqid = cpu_to_le16(qid);
626 status = nvme_submit_admin_cmd(dev, &c, NULL);
632 static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
634 return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
637 static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
639 return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
642 static void nvme_free_queue(struct nvme_dev *dev, int qid)
644 struct nvme_queue *nvmeq = dev->queues[qid];
646 free_irq(dev->entry[nvmeq->cq_vector].vector, nvmeq);
648 /* Don't tell the adapter to delete the admin queue */
650 adapter_delete_sq(dev, qid);
651 adapter_delete_cq(dev, qid);
654 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
655 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
656 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
657 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
661 static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
662 int depth, int vector)
664 struct device *dmadev = &dev->pci_dev->dev;
665 unsigned extra = (depth / 8) + (depth * sizeof(struct nvme_cmd_info));
666 struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
670 nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
671 &nvmeq->cq_dma_addr, GFP_KERNEL);
674 memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
676 nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
677 &nvmeq->sq_dma_addr, GFP_KERNEL);
681 nvmeq->q_dmadev = dmadev;
682 spin_lock_init(&nvmeq->q_lock);
685 init_waitqueue_head(&nvmeq->sq_full);
686 bio_list_init(&nvmeq->sq_cong);
687 nvmeq->q_db = &dev->dbs[qid * 2];
688 nvmeq->q_depth = depth;
689 nvmeq->cq_vector = vector;
694 dma_free_coherent(dmadev, CQ_SIZE(nvmeq->q_depth), (void *)nvmeq->cqes,
701 static int queue_request_irq(struct nvme_dev *dev, struct nvme_queue *nvmeq,
704 if (use_threaded_interrupts)
705 return request_threaded_irq(dev->entry[nvmeq->cq_vector].vector,
706 nvme_irq_check, nvme_irq,
707 IRQF_DISABLED | IRQF_SHARED,
709 return request_irq(dev->entry[nvmeq->cq_vector].vector, nvme_irq,
710 IRQF_DISABLED | IRQF_SHARED, name, nvmeq);
713 static __devinit struct nvme_queue *nvme_create_queue(struct nvme_dev *dev,
714 int qid, int cq_size, int vector)
717 struct nvme_queue *nvmeq = nvme_alloc_queue(dev, qid, cq_size, vector);
722 result = adapter_alloc_cq(dev, qid, nvmeq);
726 result = adapter_alloc_sq(dev, qid, nvmeq);
730 result = queue_request_irq(dev, nvmeq, "nvme");
737 adapter_delete_sq(dev, qid);
739 adapter_delete_cq(dev, qid);
741 dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
742 (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
743 dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
744 nvmeq->sq_cmds, nvmeq->sq_dma_addr);
749 static int __devinit nvme_configure_admin_queue(struct nvme_dev *dev)
753 struct nvme_queue *nvmeq;
755 dev->dbs = ((void __iomem *)dev->bar) + 4096;
757 nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
761 aqa = nvmeq->q_depth - 1;
764 dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
765 dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
766 dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
768 writel(0, &dev->bar->cc);
769 writel(aqa, &dev->bar->aqa);
770 writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
771 writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
772 writel(dev->ctrl_config, &dev->bar->cc);
774 while (!(readl(&dev->bar->csts) & NVME_CSTS_RDY)) {
776 if (fatal_signal_pending(current))
780 result = queue_request_irq(dev, nvmeq, "nvme admin");
781 dev->queues[0] = nvmeq;
785 static int nvme_map_user_pages(struct nvme_dev *dev, int write,
786 unsigned long addr, unsigned length,
787 struct scatterlist **sgp)
789 int i, err, count, nents, offset;
790 struct scatterlist *sg;
798 offset = offset_in_page(addr);
799 count = DIV_ROUND_UP(offset + length, PAGE_SIZE);
800 pages = kcalloc(count, sizeof(*pages), GFP_KERNEL);
802 err = get_user_pages_fast(addr, count, 1, pages);
809 sg = kcalloc(count, sizeof(*sg), GFP_KERNEL);
810 sg_init_table(sg, count);
811 sg_set_page(&sg[0], pages[0], PAGE_SIZE - offset, offset);
812 length -= (PAGE_SIZE - offset);
813 for (i = 1; i < count; i++) {
814 sg_set_page(&sg[i], pages[i], min_t(int, length, PAGE_SIZE), 0);
819 nents = dma_map_sg(&dev->pci_dev->dev, sg, count,
820 write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
829 for (i = 0; i < count; i++)
835 static void nvme_unmap_user_pages(struct nvme_dev *dev, int write,
836 unsigned long addr, int length,
837 struct scatterlist *sg, int nents)
841 count = DIV_ROUND_UP(offset_in_page(addr) + length, PAGE_SIZE);
842 dma_unmap_sg(&dev->pci_dev->dev, sg, nents, DMA_FROM_DEVICE);
844 for (i = 0; i < count; i++)
845 put_page(sg_page(&sg[i]));
848 static int nvme_submit_user_admin_command(struct nvme_dev *dev,
849 unsigned long addr, unsigned length,
850 struct nvme_command *cmd)
853 struct scatterlist *sg;
855 nents = nvme_map_user_pages(dev, 0, addr, length, &sg);
858 nvme_setup_prps(&cmd->common, sg, length);
859 err = nvme_submit_admin_cmd(dev, cmd, NULL);
860 nvme_unmap_user_pages(dev, 0, addr, length, sg, nents);
861 return err ? -EIO : 0;
864 static int nvme_identify(struct nvme_ns *ns, unsigned long addr, int cns)
866 struct nvme_command c;
868 memset(&c, 0, sizeof(c));
869 c.identify.opcode = nvme_admin_identify;
870 c.identify.nsid = cns ? 0 : cpu_to_le32(ns->ns_id);
871 c.identify.cns = cpu_to_le32(cns);
873 return nvme_submit_user_admin_command(ns->dev, addr, 4096, &c);
876 static int nvme_get_range_type(struct nvme_ns *ns, unsigned long addr)
878 struct nvme_command c;
880 memset(&c, 0, sizeof(c));
881 c.features.opcode = nvme_admin_get_features;
882 c.features.nsid = cpu_to_le32(ns->ns_id);
883 c.features.fid = cpu_to_le32(NVME_FEAT_LBA_RANGE);
885 return nvme_submit_user_admin_command(ns->dev, addr, 4096, &c);
888 static int nvme_submit_io(struct nvme_ns *ns, struct nvme_user_io __user *uio)
890 struct nvme_dev *dev = ns->dev;
891 struct nvme_queue *nvmeq;
892 struct nvme_user_io io;
893 struct nvme_command c;
897 struct scatterlist *sg;
899 if (copy_from_user(&io, uio, sizeof(io)))
901 length = io.nblocks << io.block_shift;
902 nents = nvme_map_user_pages(dev, io.opcode & 1, io.addr, length, &sg);
906 memset(&c, 0, sizeof(c));
907 c.rw.opcode = io.opcode;
908 c.rw.flags = io.flags;
909 c.rw.nsid = cpu_to_le32(io.nsid);
910 c.rw.slba = cpu_to_le64(io.slba);
911 c.rw.length = cpu_to_le16(io.nblocks - 1);
912 c.rw.control = cpu_to_le16(io.control);
913 c.rw.dsmgmt = cpu_to_le16(io.dsmgmt);
914 c.rw.reftag = cpu_to_le32(io.reftag); /* XXX: endian? */
915 c.rw.apptag = cpu_to_le16(io.apptag);
916 c.rw.appmask = cpu_to_le16(io.appmask);
918 nvme_setup_prps(&c.common, sg, length);
920 nvmeq = get_nvmeq(ns);
921 /* Since nvme_submit_sync_cmd sleeps, we can't keep preemption
922 * disabled. We may be preempted at any point, and be rescheduled
923 * to a different CPU. That will cause cacheline bouncing, but no
924 * additional races since q_lock already protects against other CPUs.
927 status = nvme_submit_sync_cmd(nvmeq, &c, &result, IO_TIMEOUT);
929 nvme_unmap_user_pages(dev, io.opcode & 1, io.addr, length, sg, nents);
930 put_user(result, &uio->result);
934 static int nvme_download_firmware(struct nvme_ns *ns,
935 struct nvme_dlfw __user *udlfw)
937 struct nvme_dev *dev = ns->dev;
938 struct nvme_dlfw dlfw;
939 struct nvme_command c;
941 struct scatterlist *sg;
943 if (copy_from_user(&dlfw, udlfw, sizeof(dlfw)))
945 if (dlfw.length >= (1 << 30))
948 nents = nvme_map_user_pages(dev, 1, dlfw.addr, dlfw.length * 4, &sg);
952 memset(&c, 0, sizeof(c));
953 c.dlfw.opcode = nvme_admin_download_fw;
954 c.dlfw.numd = cpu_to_le32(dlfw.length);
955 c.dlfw.offset = cpu_to_le32(dlfw.offset);
956 nvme_setup_prps(&c.common, sg, dlfw.length * 4);
958 status = nvme_submit_admin_cmd(dev, &c, NULL);
959 nvme_unmap_user_pages(dev, 0, dlfw.addr, dlfw.length * 4, sg, nents);
963 static int nvme_activate_firmware(struct nvme_ns *ns, unsigned long arg)
965 struct nvme_dev *dev = ns->dev;
966 struct nvme_command c;
968 memset(&c, 0, sizeof(c));
969 c.common.opcode = nvme_admin_activate_fw;
970 c.common.rsvd10[0] = cpu_to_le32(arg);
972 return nvme_submit_admin_cmd(dev, &c, NULL);
975 static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
978 struct nvme_ns *ns = bdev->bd_disk->private_data;
981 case NVME_IOCTL_IDENTIFY_NS:
982 return nvme_identify(ns, arg, 0);
983 case NVME_IOCTL_IDENTIFY_CTRL:
984 return nvme_identify(ns, arg, 1);
985 case NVME_IOCTL_GET_RANGE_TYPE:
986 return nvme_get_range_type(ns, arg);
987 case NVME_IOCTL_SUBMIT_IO:
988 return nvme_submit_io(ns, (void __user *)arg);
989 case NVME_IOCTL_DOWNLOAD_FW:
990 return nvme_download_firmware(ns, (void __user *)arg);
991 case NVME_IOCTL_ACTIVATE_FW:
992 return nvme_activate_firmware(ns, arg);
998 static const struct block_device_operations nvme_fops = {
999 .owner = THIS_MODULE,
1000 .ioctl = nvme_ioctl,
1003 static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, int index,
1004 struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
1007 struct gendisk *disk;
1010 if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
1013 ns = kzalloc(sizeof(*ns), GFP_KERNEL);
1016 ns->queue = blk_alloc_queue(GFP_KERNEL);
1019 ns->queue->queue_flags = QUEUE_FLAG_DEFAULT | QUEUE_FLAG_NOMERGES |
1020 QUEUE_FLAG_NONROT | QUEUE_FLAG_DISCARD;
1021 blk_queue_make_request(ns->queue, nvme_make_request);
1023 ns->queue->queuedata = ns;
1025 disk = alloc_disk(NVME_MINORS);
1027 goto out_free_queue;
1030 lbaf = id->flbas & 0xf;
1031 ns->lba_shift = id->lbaf[lbaf].ds;
1033 disk->major = nvme_major;
1034 disk->minors = NVME_MINORS;
1035 disk->first_minor = NVME_MINORS * index;
1036 disk->fops = &nvme_fops;
1037 disk->private_data = ns;
1038 disk->queue = ns->queue;
1039 disk->driverfs_dev = &dev->pci_dev->dev;
1040 sprintf(disk->disk_name, "nvme%dn%d", dev->instance, index);
1041 set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
1046 blk_cleanup_queue(ns->queue);
1052 static void nvme_ns_free(struct nvme_ns *ns)
1055 blk_cleanup_queue(ns->queue);
1059 static int set_queue_count(struct nvme_dev *dev, int count)
1063 struct nvme_command c;
1064 u32 q_count = (count - 1) | ((count - 1) << 16);
1066 memset(&c, 0, sizeof(c));
1067 c.features.opcode = nvme_admin_get_features;
1068 c.features.fid = cpu_to_le32(NVME_FEAT_NUM_QUEUES);
1069 c.features.dword11 = cpu_to_le32(q_count);
1071 status = nvme_submit_admin_cmd(dev, &c, &result);
1074 return min(result & 0xffff, result >> 16) + 1;
1077 static int __devinit nvme_setup_io_queues(struct nvme_dev *dev)
1079 int result, cpu, i, nr_queues;
1081 nr_queues = num_online_cpus();
1082 result = set_queue_count(dev, nr_queues);
1085 if (result < nr_queues)
1088 /* Deregister the admin queue's interrupt */
1089 free_irq(dev->entry[0].vector, dev->queues[0]);
1091 for (i = 0; i < nr_queues; i++)
1092 dev->entry[i].entry = i;
1094 result = pci_enable_msix(dev->pci_dev, dev->entry, nr_queues);
1097 } else if (result > 0) {
1106 result = queue_request_irq(dev, dev->queues[0], "nvme admin");
1107 /* XXX: handle failure here */
1109 cpu = cpumask_first(cpu_online_mask);
1110 for (i = 0; i < nr_queues; i++) {
1111 irq_set_affinity_hint(dev->entry[i].vector, get_cpu_mask(cpu));
1112 cpu = cpumask_next(cpu, cpu_online_mask);
1115 for (i = 0; i < nr_queues; i++) {
1116 dev->queues[i + 1] = nvme_create_queue(dev, i + 1,
1118 if (!dev->queues[i + 1])
1126 static void nvme_free_queues(struct nvme_dev *dev)
1130 for (i = dev->queue_count - 1; i >= 0; i--)
1131 nvme_free_queue(dev, i);
1134 static int __devinit nvme_dev_add(struct nvme_dev *dev)
1137 struct nvme_ns *ns, *next;
1138 struct nvme_id_ctrl *ctrl;
1140 dma_addr_t dma_addr;
1141 struct nvme_command cid, crt;
1143 res = nvme_setup_io_queues(dev);
1147 /* XXX: Switch to a SG list once prp2 works */
1148 id = dma_alloc_coherent(&dev->pci_dev->dev, 8192, &dma_addr,
1151 memset(&cid, 0, sizeof(cid));
1152 cid.identify.opcode = nvme_admin_identify;
1153 cid.identify.nsid = 0;
1154 cid.identify.prp1 = cpu_to_le64(dma_addr);
1155 cid.identify.cns = cpu_to_le32(1);
1157 res = nvme_submit_admin_cmd(dev, &cid, NULL);
1164 nn = le32_to_cpup(&ctrl->nn);
1165 memcpy(dev->serial, ctrl->sn, sizeof(ctrl->sn));
1166 memcpy(dev->model, ctrl->mn, sizeof(ctrl->mn));
1167 memcpy(dev->firmware_rev, ctrl->fr, sizeof(ctrl->fr));
1169 cid.identify.cns = 0;
1170 memset(&crt, 0, sizeof(crt));
1171 crt.features.opcode = nvme_admin_get_features;
1172 crt.features.prp1 = cpu_to_le64(dma_addr + 4096);
1173 crt.features.fid = cpu_to_le32(NVME_FEAT_LBA_RANGE);
1175 for (i = 0; i < nn; i++) {
1176 cid.identify.nsid = cpu_to_le32(i);
1177 res = nvme_submit_admin_cmd(dev, &cid, NULL);
1181 if (((struct nvme_id_ns *)id)->ncap == 0)
1184 crt.features.nsid = cpu_to_le32(i);
1185 res = nvme_submit_admin_cmd(dev, &crt, NULL);
1189 ns = nvme_alloc_ns(dev, i, id, id + 4096);
1191 list_add_tail(&ns->list, &dev->namespaces);
1193 list_for_each_entry(ns, &dev->namespaces, list)
1196 dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1200 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
1201 list_del(&ns->list);
1205 dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
1209 static int nvme_dev_remove(struct nvme_dev *dev)
1211 struct nvme_ns *ns, *next;
1213 /* TODO: wait all I/O finished or cancel them */
1215 list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
1216 list_del(&ns->list);
1217 del_gendisk(ns->disk);
1221 nvme_free_queues(dev);
1226 /* XXX: Use an ida or something to let remove / add work correctly */
1227 static void nvme_set_instance(struct nvme_dev *dev)
1229 static int instance;
1230 dev->instance = instance++;
1233 static void nvme_release_instance(struct nvme_dev *dev)
1237 static int __devinit nvme_probe(struct pci_dev *pdev,
1238 const struct pci_device_id *id)
1240 int bars, result = -ENOMEM;
1241 struct nvme_dev *dev;
1243 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1246 dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
1250 dev->queues = kcalloc(num_possible_cpus() + 1, sizeof(void *),
1255 if (pci_enable_device_mem(pdev))
1257 pci_set_master(pdev);
1258 bars = pci_select_bars(pdev, IORESOURCE_MEM);
1259 if (pci_request_selected_regions(pdev, bars, "nvme"))
1262 INIT_LIST_HEAD(&dev->namespaces);
1263 dev->pci_dev = pdev;
1264 pci_set_drvdata(pdev, dev);
1265 dma_set_mask(&pdev->dev, DMA_BIT_MASK(64));
1266 dma_set_coherent_mask(&pdev->dev, DMA_BIT_MASK(64));
1267 nvme_set_instance(dev);
1268 dev->entry[0].vector = pdev->irq;
1270 dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
1276 result = nvme_configure_admin_queue(dev);
1281 result = nvme_dev_add(dev);
1287 nvme_free_queues(dev);
1291 pci_disable_msix(pdev);
1292 nvme_release_instance(dev);
1294 pci_disable_device(pdev);
1295 pci_release_regions(pdev);
1303 static void __devexit nvme_remove(struct pci_dev *pdev)
1305 struct nvme_dev *dev = pci_get_drvdata(pdev);
1306 nvme_dev_remove(dev);
1307 pci_disable_msix(pdev);
1309 nvme_release_instance(dev);
1310 pci_disable_device(pdev);
1311 pci_release_regions(pdev);
1317 /* These functions are yet to be implemented */
1318 #define nvme_error_detected NULL
1319 #define nvme_dump_registers NULL
1320 #define nvme_link_reset NULL
1321 #define nvme_slot_reset NULL
1322 #define nvme_error_resume NULL
1323 #define nvme_suspend NULL
1324 #define nvme_resume NULL
1326 static struct pci_error_handlers nvme_err_handler = {
1327 .error_detected = nvme_error_detected,
1328 .mmio_enabled = nvme_dump_registers,
1329 .link_reset = nvme_link_reset,
1330 .slot_reset = nvme_slot_reset,
1331 .resume = nvme_error_resume,
1334 /* Move to pci_ids.h later */
1335 #define PCI_CLASS_STORAGE_EXPRESS 0x010802
1337 static DEFINE_PCI_DEVICE_TABLE(nvme_id_table) = {
1338 { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
1341 MODULE_DEVICE_TABLE(pci, nvme_id_table);
1343 static struct pci_driver nvme_driver = {
1345 .id_table = nvme_id_table,
1346 .probe = nvme_probe,
1347 .remove = __devexit_p(nvme_remove),
1348 .suspend = nvme_suspend,
1349 .resume = nvme_resume,
1350 .err_handler = &nvme_err_handler,
1353 static int __init nvme_init(void)
1357 nvme_major = register_blkdev(nvme_major, "nvme");
1358 if (nvme_major <= 0)
1361 result = pci_register_driver(&nvme_driver);
1365 unregister_blkdev(nvme_major, "nvme");
1369 static void __exit nvme_exit(void)
1371 pci_unregister_driver(&nvme_driver);
1372 unregister_blkdev(nvme_major, "nvme");
1375 MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
1376 MODULE_LICENSE("GPL");
1377 MODULE_VERSION("0.2");
1378 module_init(nvme_init);
1379 module_exit(nvme_exit);