NVMe: New driver
authorMatthew Wilcox <matthew.r.wilcox@intel.com>
Thu, 20 Jan 2011 17:50:14 +0000 (12:50 -0500)
committerMatthew Wilcox <matthew.r.wilcox@intel.com>
Fri, 4 Nov 2011 19:52:51 +0000 (15:52 -0400)
This driver is for devices that follow the NVM Express standard

Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
Documentation/ioctl/ioctl-number.txt
drivers/block/Kconfig
drivers/block/Makefile
drivers/block/nvme.c [new file with mode: 0644]
include/linux/nvme.h [new file with mode: 0644]

index 54078ed96b3751724acaf0bd3fdab95df962480f..4840334ea97b30705df13a2c642e976a83621800 100644 (file)
@@ -149,6 +149,7 @@ Code  Seq#(hex)     Include File            Comments
 'M'    01-03   drivers/scsi/megaraid/megaraid_sas.h
 'M'    00-0F   drivers/video/fsl-diu-fb.h      conflict!
 'N'    00-1F   drivers/usb/scanner.h
+'N'    40-7F   drivers/block/nvme.c
 'O'     00-06   mtd/ubi-user.h         UBI
 'P'    all     linux/soundcard.h       conflict!
 'P'    60-6F   sound/sscape_ioctl.h    conflict!
index 6f07ec1c2f58d2dabd766463165e32728910faf6..35e56e1c948fa91819eaabb73130b3efa4b68eae 100644 (file)
@@ -315,6 +315,17 @@ config BLK_DEV_NBD
 
          If unsure, say N.
 
+config BLK_DEV_NVME
+       tristate "NVM Express block device"
+       depends on PCI
+       ---help---
+         The NVM Express driver is for solid state drives directly
+         connected to the PCI or PCI Express bus.  If you know you
+         don't have one of these, it is safe to answer N.
+
+         To compile this driver as a module, choose M here: the
+         module will be called nvme.
+
 config BLK_DEV_OSD
        tristate "OSD object-as-blkdev support"
        depends on SCSI_OSD_ULD
index 76646e9a1c91528617c22e2c3f429201531acdf4..349539ad3ad95ef33a8d38e5aa30aeea7321235d 100644 (file)
@@ -23,6 +23,7 @@ obj-$(CONFIG_XILINX_SYSACE)   += xsysace.o
 obj-$(CONFIG_CDROM_PKTCDVD)    += pktcdvd.o
 obj-$(CONFIG_MG_DISK)          += mg_disk.o
 obj-$(CONFIG_SUNVDC)           += sunvdc.o
+obj-$(CONFIG_BLK_DEV_NVME)     += nvme.o
 obj-$(CONFIG_BLK_DEV_OSD)      += osdblk.o
 
 obj-$(CONFIG_BLK_DEV_UMEM)     += umem.o
diff --git a/drivers/block/nvme.c b/drivers/block/nvme.c
new file mode 100644 (file)
index 0000000..ef66ecc
--- /dev/null
@@ -0,0 +1,1043 @@
+/*
+ * NVM Express device driver
+ * Copyright (c) 2011, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc.,
+ * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#include <linux/nvme.h>
+#include <linux/bio.h>
+#include <linux/blkdev.h>
+#include <linux/errno.h>
+#include <linux/fs.h>
+#include <linux/genhd.h>
+#include <linux/init.h>
+#include <linux/interrupt.h>
+#include <linux/io.h>
+#include <linux/kdev_t.h>
+#include <linux/kernel.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <linux/moduleparam.h>
+#include <linux/pci.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/types.h>
+#include <linux/version.h>
+
+#define NVME_Q_DEPTH 1024
+#define SQ_SIZE(depth)         (depth * sizeof(struct nvme_command))
+#define CQ_SIZE(depth)         (depth * sizeof(struct nvme_completion))
+#define NVME_MINORS 64
+
+static int nvme_major;
+module_param(nvme_major, int, 0);
+
+/*
+ * Represents an NVM Express device.  Each nvme_dev is a PCI function.
+ */
+struct nvme_dev {
+       struct list_head node;
+       struct nvme_queue **queues;
+       u32 __iomem *dbs;
+       struct pci_dev *pci_dev;
+       int instance;
+       int queue_count;
+       u32 ctrl_config;
+       struct msix_entry *entry;
+       struct nvme_bar __iomem *bar;
+       struct list_head namespaces;
+};
+
+/*
+ * An NVM Express namespace is equivalent to a SCSI LUN
+ */
+struct nvme_ns {
+       struct list_head list;
+
+       struct nvme_dev *dev;
+       struct request_queue *queue;
+       struct gendisk *disk;
+
+       int ns_id;
+       int lba_shift;
+};
+
+/*
+ * An NVM Express queue.  Each device has at least two (one for admin
+ * commands and one for I/O commands).
+ */
+struct nvme_queue {
+       struct device *q_dmadev;
+       spinlock_t q_lock;
+       struct nvme_command *sq_cmds;
+       volatile struct nvme_completion *cqes;
+       dma_addr_t sq_dma_addr;
+       dma_addr_t cq_dma_addr;
+       wait_queue_head_t sq_full;
+       struct bio_list sq_cong;
+       u32 __iomem *q_db;
+       u16 q_depth;
+       u16 cq_vector;
+       u16 sq_head;
+       u16 sq_tail;
+       u16 cq_head;
+       u16 cq_cycle;
+       unsigned long cmdid_data[];
+};
+
+/*
+ * Check we didin't inadvertently grow the command struct
+ */
+static inline void _nvme_check_size(void)
+{
+       BUILD_BUG_ON(sizeof(struct nvme_rw_command) != 64);
+       BUILD_BUG_ON(sizeof(struct nvme_create_cq) != 64);
+       BUILD_BUG_ON(sizeof(struct nvme_create_sq) != 64);
+       BUILD_BUG_ON(sizeof(struct nvme_delete_queue) != 64);
+       BUILD_BUG_ON(sizeof(struct nvme_features) != 64);
+       BUILD_BUG_ON(sizeof(struct nvme_command) != 64);
+       BUILD_BUG_ON(sizeof(struct nvme_id_ctrl) != 4096);
+       BUILD_BUG_ON(sizeof(struct nvme_id_ns) != 4096);
+       BUILD_BUG_ON(sizeof(struct nvme_lba_range_type) != 64);
+}
+
+/**
+ * alloc_cmdid - Allocate a Command ID
+ * @param nvmeq The queue that will be used for this command
+ * @param ctx A pointer that will be passed to the handler
+ * @param handler The ID of the handler to call
+ *
+ * Allocate a Command ID for a queue.  The data passed in will
+ * be passed to the completion handler.  This is implemented by using
+ * the bottom two bits of the ctx pointer to store the handler ID.
+ * Passing in a pointer that's not 4-byte aligned will cause a BUG.
+ * We can change this if it becomes a problem.
+ */
+static int alloc_cmdid(struct nvme_queue *nvmeq, void *ctx, int handler)
+{
+       int depth = nvmeq->q_depth;
+       unsigned long data = (unsigned long)ctx | handler;
+       int cmdid;
+
+       BUG_ON((unsigned long)ctx & 3);
+
+       do {
+               cmdid = find_first_zero_bit(nvmeq->cmdid_data, depth);
+               if (cmdid >= depth)
+                       return -EBUSY;
+       } while (test_and_set_bit(cmdid, nvmeq->cmdid_data));
+
+       nvmeq->cmdid_data[cmdid + BITS_TO_LONGS(depth)] = data;
+       return cmdid;
+}
+
+static int alloc_cmdid_killable(struct nvme_queue *nvmeq, void *ctx,
+                                                               int handler)
+{
+       int cmdid;
+       wait_event_killable(nvmeq->sq_full,
+                       (cmdid = alloc_cmdid(nvmeq, ctx, handler)) >= 0);
+       return (cmdid < 0) ? -EINTR : cmdid;
+}
+
+/* If you need more than four handlers, you'll need to change how
+ * alloc_cmdid and nvme_process_cq work
+ */
+enum {
+       sync_completion_id = 0,
+       bio_completion_id,
+};
+
+static unsigned long free_cmdid(struct nvme_queue *nvmeq, int cmdid)
+{
+       unsigned long data;
+
+       data = nvmeq->cmdid_data[cmdid + BITS_TO_LONGS(nvmeq->q_depth)];
+       clear_bit(cmdid, nvmeq->cmdid_data);
+       wake_up(&nvmeq->sq_full);
+       return data;
+}
+
+static struct nvme_queue *get_nvmeq(struct nvme_ns *ns)
+{
+       return ns->dev->queues[1];
+}
+
+static void put_nvmeq(struct nvme_queue *nvmeq)
+{
+}
+
+/**
+ * nvme_submit_cmd: Copy a command into a queue and ring the doorbell
+ * @nvmeq: The queue to use
+ * @cmd: The command to send
+ *
+ * Safe to use from interrupt context
+ */
+static int nvme_submit_cmd(struct nvme_queue *nvmeq, struct nvme_command *cmd)
+{
+       unsigned long flags;
+       u16 tail;
+       /* XXX: Need to check tail isn't going to overrun head */
+       spin_lock_irqsave(&nvmeq->q_lock, flags);
+       tail = nvmeq->sq_tail;
+       memcpy(&nvmeq->sq_cmds[tail], cmd, sizeof(*cmd));
+       writel(tail, nvmeq->q_db);
+       if (++tail == nvmeq->q_depth)
+               tail = 0;
+       nvmeq->sq_tail = tail;
+       spin_unlock_irqrestore(&nvmeq->q_lock, flags);
+
+       return 0;
+}
+
+struct nvme_req_info {
+       struct bio *bio;
+       int nents;
+       struct scatterlist sg[0];
+};
+
+/* XXX: use a mempool */
+static struct nvme_req_info *alloc_info(unsigned nseg, gfp_t gfp)
+{
+       return kmalloc(sizeof(struct nvme_req_info) +
+                       sizeof(struct scatterlist) * nseg, gfp);
+}
+
+static void free_info(struct nvme_req_info *info)
+{
+       kfree(info);
+}
+
+static void bio_completion(struct nvme_queue *nvmeq, void *ctx,
+                                               struct nvme_completion *cqe)
+{
+       struct nvme_req_info *info = ctx;
+       struct bio *bio = info->bio;
+       u16 status = le16_to_cpup(&cqe->status) >> 1;
+
+       dma_unmap_sg(nvmeq->q_dmadev, info->sg, info->nents,
+                       bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
+       free_info(info);
+       bio_endio(bio, status ? -EIO : 0);
+}
+
+static int nvme_map_bio(struct device *dev, struct nvme_req_info *info,
+               struct bio *bio, enum dma_data_direction dma_dir, int psegs)
+{
+       struct bio_vec *bvec;
+       struct scatterlist *sg = info->sg;
+       int i, nsegs;
+
+       sg_init_table(sg, psegs);
+       bio_for_each_segment(bvec, bio, i) {
+               sg_set_page(sg, bvec->bv_page, bvec->bv_len, bvec->bv_offset);
+               /* XXX: handle non-mergable here */
+               nsegs++;
+       }
+       info->nents = nsegs;
+
+       return dma_map_sg(dev, info->sg, info->nents, dma_dir);
+}
+
+static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
+                                                               struct bio *bio)
+{
+       struct nvme_rw_command *cmnd;
+       struct nvme_req_info *info;
+       enum dma_data_direction dma_dir;
+       int cmdid;
+       u16 control;
+       u32 dsmgmt;
+       unsigned long flags;
+       int psegs = bio_phys_segments(ns->queue, bio);
+
+       info = alloc_info(psegs, GFP_NOIO);
+       if (!info)
+               goto congestion;
+       info->bio = bio;
+
+       cmdid = alloc_cmdid(nvmeq, info, bio_completion_id);
+       if (unlikely(cmdid < 0))
+               goto free_info;
+
+       control = 0;
+       if (bio->bi_rw & REQ_FUA)
+               control |= NVME_RW_FUA;
+       if (bio->bi_rw & (REQ_FAILFAST_DEV | REQ_RAHEAD))
+               control |= NVME_RW_LR;
+
+       dsmgmt = 0;
+       if (bio->bi_rw & REQ_RAHEAD)
+               dsmgmt |= NVME_RW_DSM_FREQ_PREFETCH;
+
+       spin_lock_irqsave(&nvmeq->q_lock, flags);
+       cmnd = &nvmeq->sq_cmds[nvmeq->sq_tail].rw;
+
+       if (bio_data_dir(bio)) {
+               cmnd->opcode = nvme_cmd_write;
+               dma_dir = DMA_TO_DEVICE;
+       } else {
+               cmnd->opcode = nvme_cmd_read;
+               dma_dir = DMA_FROM_DEVICE;
+       }
+
+       nvme_map_bio(nvmeq->q_dmadev, info, bio, dma_dir, psegs);
+
+       cmnd->flags = 1;
+       cmnd->command_id = cmdid;
+       cmnd->nsid = cpu_to_le32(ns->ns_id);
+       cmnd->prp1 = cpu_to_le64(sg_phys(info->sg));
+       /* XXX: Support more than one PRP */
+       cmnd->slba = cpu_to_le64(bio->bi_sector >> (ns->lba_shift - 9));
+       cmnd->length = cpu_to_le16((bio->bi_size >> ns->lba_shift) - 1);
+       cmnd->control = cpu_to_le16(control);
+       cmnd->dsmgmt = cpu_to_le32(dsmgmt);
+
+       writel(nvmeq->sq_tail, nvmeq->q_db);
+       if (++nvmeq->sq_tail == nvmeq->q_depth)
+               nvmeq->sq_tail = 0;
+
+       spin_unlock_irqrestore(&nvmeq->q_lock, flags);
+
+       return 0;
+
+ free_info:
+       free_info(info);
+ congestion:
+       return -EBUSY;
+}
+
+/*
+ * NB: return value of non-zero would mean that we were a stacking driver.
+ * make_request must always succeed.
+ */
+static int nvme_make_request(struct request_queue *q, struct bio *bio)
+{
+       struct nvme_ns *ns = q->queuedata;
+       struct nvme_queue *nvmeq = get_nvmeq(ns);
+
+       if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
+               blk_set_queue_congested(q, rw_is_sync(bio->bi_rw));
+               bio_list_add(&nvmeq->sq_cong, bio);
+       }
+       put_nvmeq(nvmeq);
+
+       return 0;
+}
+
+struct sync_cmd_info {
+       struct task_struct *task;
+       u32 result;
+       int status;
+};
+
+static void sync_completion(struct nvme_queue *nvmeq, void *ctx,
+                                               struct nvme_completion *cqe)
+{
+       struct sync_cmd_info *cmdinfo = ctx;
+       cmdinfo->result = le32_to_cpup(&cqe->result);
+       cmdinfo->status = le16_to_cpup(&cqe->status) >> 1;
+       wake_up_process(cmdinfo->task);
+}
+
+typedef void (*completion_fn)(struct nvme_queue *, void *,
+                                               struct nvme_completion *);
+
+static irqreturn_t nvme_process_cq(struct nvme_queue *nvmeq)
+{
+       u16 head, cycle;
+
+       static const completion_fn completions[4] = {
+               [sync_completion_id] = sync_completion,
+               [bio_completion_id]  = bio_completion,
+       };
+
+       head = nvmeq->cq_head;
+       cycle = nvmeq->cq_cycle;
+
+       for (;;) {
+               unsigned long data;
+               void *ptr;
+               unsigned char handler;
+               struct nvme_completion cqe = nvmeq->cqes[head];
+               if ((le16_to_cpu(cqe.status) & 1) != cycle)
+                       break;
+               nvmeq->sq_head = le16_to_cpu(cqe.sq_head);
+               if (++head == nvmeq->q_depth) {
+                       head = 0;
+                       cycle = !cycle;
+               }
+
+               data = free_cmdid(nvmeq, cqe.command_id);
+               handler = data & 3;
+               ptr = (void *)(data & ~3UL);
+               completions[handler](nvmeq, ptr, &cqe);
+       }
+
+       /* If the controller ignores the cq head doorbell and continuously
+        * writes to the queue, it is theoretically possible to wrap around
+        * the queue twice and mistakenly return IRQ_NONE.  Linux only
+        * requires that 0.1% of your interrupts are handled, so this isn't
+        * a big problem.
+        */
+       if (head == nvmeq->cq_head && cycle == nvmeq->cq_cycle)
+               return IRQ_NONE;
+
+       writel(head, nvmeq->q_db + 1);
+       nvmeq->cq_head = head;
+       nvmeq->cq_cycle = cycle;
+
+       return IRQ_HANDLED;
+}
+
+static irqreturn_t nvme_irq(int irq, void *data)
+{
+       return nvme_process_cq(data);
+}
+
+/*
+ * Returns 0 on success.  If the result is negative, it's a Linux error code;
+ * if the result is positive, it's an NVM Express status code
+ */
+static int nvme_submit_sync_cmd(struct nvme_queue *q, struct nvme_command *cmd,
+                                                               u32 *result)
+{
+       int cmdid;
+       struct sync_cmd_info cmdinfo;
+
+       cmdinfo.task = current;
+       cmdinfo.status = -EINTR;
+
+       cmdid = alloc_cmdid_killable(q, &cmdinfo, sync_completion_id);
+       if (cmdid < 0)
+               return cmdid;
+       cmd->common.command_id = cmdid;
+
+       set_current_state(TASK_UNINTERRUPTIBLE);
+       nvme_submit_cmd(q, cmd);
+       schedule();
+
+       if (result)
+               *result = cmdinfo.result;
+
+       return cmdinfo.status;
+}
+
+static int nvme_submit_admin_cmd(struct nvme_dev *dev, struct nvme_command *cmd,
+                                                               u32 *result)
+{
+       return nvme_submit_sync_cmd(dev->queues[0], cmd, result);
+}
+
+static int adapter_delete_queue(struct nvme_dev *dev, u8 opcode, u16 id)
+{
+       int status;
+       struct nvme_command c;
+
+       memset(&c, 0, sizeof(c));
+       c.delete_queue.opcode = opcode;
+       c.delete_queue.qid = cpu_to_le16(id);
+
+       status = nvme_submit_admin_cmd(dev, &c, NULL);
+       if (status)
+               return -EIO;
+       return 0;
+}
+
+static int adapter_alloc_cq(struct nvme_dev *dev, u16 qid,
+                                               struct nvme_queue *nvmeq)
+{
+       int status;
+       struct nvme_command c;
+       int flags = NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED;
+
+       memset(&c, 0, sizeof(c));
+       c.create_cq.opcode = nvme_admin_create_cq;
+       c.create_cq.prp1 = cpu_to_le64(nvmeq->cq_dma_addr);
+       c.create_cq.cqid = cpu_to_le16(qid);
+       c.create_cq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
+       c.create_cq.cq_flags = cpu_to_le16(flags);
+       c.create_cq.irq_vector = cpu_to_le16(nvmeq->cq_vector);
+
+       status = nvme_submit_admin_cmd(dev, &c, NULL);
+       if (status)
+               return -EIO;
+       return 0;
+}
+
+static int adapter_alloc_sq(struct nvme_dev *dev, u16 qid,
+                                               struct nvme_queue *nvmeq)
+{
+       int status;
+       struct nvme_command c;
+       int flags = NVME_QUEUE_PHYS_CONTIG | NVME_SQ_PRIO_MEDIUM;
+
+       memset(&c, 0, sizeof(c));
+       c.create_sq.opcode = nvme_admin_create_sq;
+       c.create_sq.prp1 = cpu_to_le64(nvmeq->sq_dma_addr);
+       c.create_sq.sqid = cpu_to_le16(qid);
+       c.create_sq.qsize = cpu_to_le16(nvmeq->q_depth - 1);
+       c.create_sq.sq_flags = cpu_to_le16(flags);
+       c.create_sq.cqid = cpu_to_le16(qid);
+
+       status = nvme_submit_admin_cmd(dev, &c, NULL);
+       if (status)
+               return -EIO;
+       return 0;
+}
+
+static int adapter_delete_cq(struct nvme_dev *dev, u16 cqid)
+{
+       return adapter_delete_queue(dev, nvme_admin_delete_cq, cqid);
+}
+
+static int adapter_delete_sq(struct nvme_dev *dev, u16 sqid)
+{
+       return adapter_delete_queue(dev, nvme_admin_delete_sq, sqid);
+}
+
+static void nvme_free_queue(struct nvme_dev *dev, int qid)
+{
+       struct nvme_queue *nvmeq = dev->queues[qid];
+
+       free_irq(dev->entry[nvmeq->cq_vector].vector, nvmeq);
+
+       /* Don't tell the adapter to delete the admin queue */
+       if (qid) {
+               adapter_delete_sq(dev, qid);
+               adapter_delete_cq(dev, qid);
+       }
+
+       dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
+                               (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
+       dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
+                                       nvmeq->sq_cmds, nvmeq->sq_dma_addr);
+       kfree(nvmeq);
+}
+
+static struct nvme_queue *nvme_alloc_queue(struct nvme_dev *dev, int qid,
+                                                       int depth, int vector)
+{
+       struct device *dmadev = &dev->pci_dev->dev;
+       unsigned extra = (depth + BITS_TO_LONGS(depth)) * sizeof(long);
+       struct nvme_queue *nvmeq = kzalloc(sizeof(*nvmeq) + extra, GFP_KERNEL);
+       if (!nvmeq)
+               return NULL;
+
+       nvmeq->cqes = dma_alloc_coherent(dmadev, CQ_SIZE(depth),
+                                       &nvmeq->cq_dma_addr, GFP_KERNEL);
+       if (!nvmeq->cqes)
+               goto free_nvmeq;
+       memset((void *)nvmeq->cqes, 0, CQ_SIZE(depth));
+
+       nvmeq->sq_cmds = dma_alloc_coherent(dmadev, SQ_SIZE(depth),
+                                       &nvmeq->sq_dma_addr, GFP_KERNEL);
+       if (!nvmeq->sq_cmds)
+               goto free_cqdma;
+
+       nvmeq->q_dmadev = dmadev;
+       spin_lock_init(&nvmeq->q_lock);
+       nvmeq->cq_head = 0;
+       nvmeq->cq_cycle = 1;
+       init_waitqueue_head(&nvmeq->sq_full);
+       bio_list_init(&nvmeq->sq_cong);
+       nvmeq->q_db = &dev->dbs[qid * 2];
+       nvmeq->q_depth = depth;
+       nvmeq->cq_vector = vector;
+
+       return nvmeq;
+
+ free_cqdma:
+       dma_free_coherent(dmadev, CQ_SIZE(nvmeq->q_depth), (void *)nvmeq->cqes,
+                                                       nvmeq->cq_dma_addr);
+ free_nvmeq:
+       kfree(nvmeq);
+       return NULL;
+}
+
+static __devinit struct nvme_queue *nvme_create_queue(struct nvme_dev *dev,
+                                       int qid, int cq_size, int vector)
+{
+       int result;
+       struct nvme_queue *nvmeq = nvme_alloc_queue(dev, qid, cq_size, vector);
+
+       result = adapter_alloc_cq(dev, qid, nvmeq);
+       if (result < 0)
+               goto free_nvmeq;
+
+       result = adapter_alloc_sq(dev, qid, nvmeq);
+       if (result < 0)
+               goto release_cq;
+
+       result = request_irq(dev->entry[vector].vector, nvme_irq,
+                               IRQF_DISABLED | IRQF_SHARED, "nvme", nvmeq);
+       if (result < 0)
+               goto release_sq;
+
+       return nvmeq;
+
+ release_sq:
+       adapter_delete_sq(dev, qid);
+ release_cq:
+       adapter_delete_cq(dev, qid);
+ free_nvmeq:
+       dma_free_coherent(nvmeq->q_dmadev, CQ_SIZE(nvmeq->q_depth),
+                               (void *)nvmeq->cqes, nvmeq->cq_dma_addr);
+       dma_free_coherent(nvmeq->q_dmadev, SQ_SIZE(nvmeq->q_depth),
+                                       nvmeq->sq_cmds, nvmeq->sq_dma_addr);
+       kfree(nvmeq);
+       return NULL;
+}
+
+static int __devinit nvme_configure_admin_queue(struct nvme_dev *dev)
+{
+       int result;
+       u32 aqa;
+       struct nvme_queue *nvmeq;
+
+       dev->dbs = ((void __iomem *)dev->bar) + 4096;
+
+       nvmeq = nvme_alloc_queue(dev, 0, 64, 0);
+
+       aqa = nvmeq->q_depth - 1;
+       aqa |= aqa << 16;
+
+       dev->ctrl_config = NVME_CC_ENABLE | NVME_CC_CSS_NVM;
+       dev->ctrl_config |= (PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
+       dev->ctrl_config |= NVME_CC_ARB_RR | NVME_CC_SHN_NONE;
+
+       writel(aqa, &dev->bar->aqa);
+       writeq(nvmeq->sq_dma_addr, &dev->bar->asq);
+       writeq(nvmeq->cq_dma_addr, &dev->bar->acq);
+       writel(dev->ctrl_config, &dev->bar->cc);
+
+       while (!(readl(&dev->bar->csts) & NVME_CSTS_RDY)) {
+               msleep(100);
+               if (fatal_signal_pending(current))
+                       return -EINTR;
+       }
+
+       result = request_irq(dev->entry[0].vector, nvme_irq,
+                       IRQF_DISABLED | IRQF_SHARED, "nvme admin", nvmeq);
+       dev->queues[0] = nvmeq;
+       return result;
+}
+
+static int nvme_identify(struct nvme_ns *ns, void __user *addr, int cns)
+{
+       struct nvme_dev *dev = ns->dev;
+       int status;
+       struct nvme_command c;
+       void *page;
+       dma_addr_t dma_addr;
+
+       page = dma_alloc_coherent(&dev->pci_dev->dev, 4096, &dma_addr,
+                                                               GFP_KERNEL);
+
+       memset(&c, 0, sizeof(c));
+       c.identify.opcode = nvme_admin_identify;
+       c.identify.nsid = cns ? 0 : cpu_to_le32(ns->ns_id);
+       c.identify.prp1 = cpu_to_le64(dma_addr);
+       c.identify.cns = cpu_to_le32(cns);
+
+       status = nvme_submit_admin_cmd(dev, &c, NULL);
+
+       if (status)
+               status = -EIO;
+       else if (copy_to_user(addr, page, 4096))
+               status = -EFAULT;
+
+       dma_free_coherent(&dev->pci_dev->dev, 4096, page, dma_addr);
+
+       return status;
+}
+
+static int nvme_get_range_type(struct nvme_ns *ns, void __user *addr)
+{
+       struct nvme_dev *dev = ns->dev;
+       int status;
+       struct nvme_command c;
+       void *page;
+       dma_addr_t dma_addr;
+
+       page = dma_alloc_coherent(&dev->pci_dev->dev, 4096, &dma_addr,
+                                                               GFP_KERNEL);
+
+       memset(&c, 0, sizeof(c));
+       c.features.opcode = nvme_admin_get_features;
+       c.features.nsid = cpu_to_le32(ns->ns_id);
+       c.features.prp1 = cpu_to_le64(dma_addr);
+       c.features.fid = cpu_to_le32(NVME_FEAT_LBA_RANGE);
+
+       status = nvme_submit_admin_cmd(dev, &c, NULL);
+
+       /* XXX: Assuming first range for now */
+       if (status)
+               status = -EIO;
+       else if (copy_to_user(addr, page, 64))
+               status = -EFAULT;
+
+       dma_free_coherent(&dev->pci_dev->dev, 4096, page, dma_addr);
+
+       return status;
+}
+
+static int nvme_ioctl(struct block_device *bdev, fmode_t mode, unsigned int cmd,
+                                                       unsigned long arg)
+{
+       struct nvme_ns *ns = bdev->bd_disk->private_data;
+
+       switch (cmd) {
+       case NVME_IOCTL_IDENTIFY_NS:
+               return nvme_identify(ns, (void __user *)arg, 0);
+       case NVME_IOCTL_IDENTIFY_CTRL:
+               return nvme_identify(ns, (void __user *)arg, 1);
+       case NVME_IOCTL_GET_RANGE_TYPE:
+               return nvme_get_range_type(ns, (void __user *)arg);
+       default:
+               return -ENOTTY;
+       }
+}
+
+static const struct block_device_operations nvme_fops = {
+       .owner          = THIS_MODULE,
+       .ioctl          = nvme_ioctl,
+};
+
+static struct nvme_ns *nvme_alloc_ns(struct nvme_dev *dev, int index,
+                       struct nvme_id_ns *id, struct nvme_lba_range_type *rt)
+{
+       struct nvme_ns *ns;
+       struct gendisk *disk;
+       int lbaf;
+
+       if (rt->attributes & NVME_LBART_ATTRIB_HIDE)
+               return NULL;
+
+       ns = kzalloc(sizeof(*ns), GFP_KERNEL);
+       if (!ns)
+               return NULL;
+       ns->queue = blk_alloc_queue(GFP_KERNEL);
+       if (!ns->queue)
+               goto out_free_ns;
+       ns->queue->queue_flags = QUEUE_FLAG_DEFAULT | QUEUE_FLAG_NOMERGES |
+                               QUEUE_FLAG_NONROT | QUEUE_FLAG_DISCARD;
+       blk_queue_make_request(ns->queue, nvme_make_request);
+       ns->dev = dev;
+       ns->queue->queuedata = ns;
+
+       disk = alloc_disk(NVME_MINORS);
+       if (!disk)
+               goto out_free_queue;
+       ns->ns_id = index;
+       ns->disk = disk;
+       lbaf = id->flbas & 0xf;
+       ns->lba_shift = id->lbaf[lbaf].ds;
+
+       disk->major = nvme_major;
+       disk->minors = NVME_MINORS;
+       disk->first_minor = NVME_MINORS * index;
+       disk->fops = &nvme_fops;
+       disk->private_data = ns;
+       disk->queue = ns->queue;
+       sprintf(disk->disk_name, "nvme%dn%d", dev->instance, index);
+       set_capacity(disk, le64_to_cpup(&id->nsze) << (ns->lba_shift - 9));
+
+       return ns;
+
+ out_free_queue:
+       blk_cleanup_queue(ns->queue);
+ out_free_ns:
+       kfree(ns);
+       return NULL;
+}
+
+static void nvme_ns_free(struct nvme_ns *ns)
+{
+       put_disk(ns->disk);
+       blk_cleanup_queue(ns->queue);
+       kfree(ns);
+}
+
+static int set_queue_count(struct nvme_dev *dev, int sq_count, int cq_count)
+{
+       int status;
+       u32 result;
+       struct nvme_command c;
+       u32 q_count = (sq_count - 1) | ((cq_count - 1) << 16);
+
+       memset(&c, 0, sizeof(c));
+       c.features.opcode = nvme_admin_get_features;
+       c.features.fid = cpu_to_le32(NVME_FEAT_NUM_QUEUES);
+       c.features.dword11 = cpu_to_le32(q_count);
+
+       status = nvme_submit_admin_cmd(dev, &c, &result);
+       if (status)
+               return -EIO;
+       return min(result & 0xffff, result >> 16) + 1;
+}
+
+/* XXX: Create per-CPU queues */
+static int __devinit nvme_setup_io_queues(struct nvme_dev *dev)
+{
+       int this_cpu;
+
+       set_queue_count(dev, 1, 1);
+
+       this_cpu = get_cpu();
+       dev->queues[1] = nvme_create_queue(dev, 1, NVME_Q_DEPTH, this_cpu);
+       put_cpu();
+       if (!dev->queues[1])
+               return -ENOMEM;
+       dev->queue_count++;
+
+       return 0;
+}
+
+static void nvme_free_queues(struct nvme_dev *dev)
+{
+       int i;
+
+       for (i = dev->queue_count - 1; i >= 0; i--)
+               nvme_free_queue(dev, i);
+}
+
+static int __devinit nvme_dev_add(struct nvme_dev *dev)
+{
+       int res, nn, i;
+       struct nvme_ns *ns, *next;
+       void *id;
+       dma_addr_t dma_addr;
+       struct nvme_command cid, crt;
+
+       res = nvme_setup_io_queues(dev);
+       if (res)
+               return res;
+
+       /* XXX: Switch to a SG list once prp2 works */
+       id = dma_alloc_coherent(&dev->pci_dev->dev, 8192, &dma_addr,
+                                                               GFP_KERNEL);
+
+       memset(&cid, 0, sizeof(cid));
+       cid.identify.opcode = nvme_admin_identify;
+       cid.identify.nsid = 0;
+       cid.identify.prp1 = cpu_to_le64(dma_addr);
+       cid.identify.cns = cpu_to_le32(1);
+
+       res = nvme_submit_admin_cmd(dev, &cid, NULL);
+       if (res) {
+               res = -EIO;
+               goto out_free;
+       }
+
+       nn = le32_to_cpup(&((struct nvme_id_ctrl *)id)->nn);
+
+       cid.identify.cns = 0;
+       memset(&crt, 0, sizeof(crt));
+       crt.features.opcode = nvme_admin_get_features;
+       crt.features.prp1 = cpu_to_le64(dma_addr + 4096);
+       crt.features.fid = cpu_to_le32(NVME_FEAT_LBA_RANGE);
+
+       for (i = 0; i < nn; i++) {
+               cid.identify.nsid = cpu_to_le32(i);
+               res = nvme_submit_admin_cmd(dev, &cid, NULL);
+               if (res)
+                       continue;
+
+               if (((struct nvme_id_ns *)id)->ncap == 0)
+                       continue;
+
+               crt.features.nsid = cpu_to_le32(i);
+               res = nvme_submit_admin_cmd(dev, &crt, NULL);
+               if (res)
+                       continue;
+
+               ns = nvme_alloc_ns(dev, i, id, id + 4096);
+               if (ns)
+                       list_add_tail(&ns->list, &dev->namespaces);
+       }
+       list_for_each_entry(ns, &dev->namespaces, list)
+               add_disk(ns->disk);
+
+       dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
+       return 0;
+
+ out_free:
+       list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
+               list_del(&ns->list);
+               nvme_ns_free(ns);
+       }
+
+       dma_free_coherent(&dev->pci_dev->dev, 4096, id, dma_addr);
+       return res;
+}
+
+static int nvme_dev_remove(struct nvme_dev *dev)
+{
+       struct nvme_ns *ns, *next;
+
+       /* TODO: wait all I/O finished or cancel them */
+
+       list_for_each_entry_safe(ns, next, &dev->namespaces, list) {
+               list_del(&ns->list);
+               del_gendisk(ns->disk);
+               nvme_ns_free(ns);
+       }
+
+       nvme_free_queues(dev);
+
+       return 0;
+}
+
+/* XXX: Use an ida or something to let remove / add work correctly */
+static void nvme_set_instance(struct nvme_dev *dev)
+{
+       static int instance;
+       dev->instance = instance++;
+}
+
+static void nvme_release_instance(struct nvme_dev *dev)
+{
+}
+
+static int __devinit nvme_probe(struct pci_dev *pdev,
+                                               const struct pci_device_id *id)
+{
+       int result = -ENOMEM;
+       struct nvme_dev *dev;
+
+       dev = kzalloc(sizeof(*dev), GFP_KERNEL);
+       if (!dev)
+               return -ENOMEM;
+       dev->entry = kcalloc(num_possible_cpus(), sizeof(*dev->entry),
+                                                               GFP_KERNEL);
+       if (!dev->entry)
+               goto free;
+       dev->queues = kcalloc(2, sizeof(void *), GFP_KERNEL);
+       if (!dev->queues)
+               goto free;
+
+       INIT_LIST_HEAD(&dev->namespaces);
+       dev->pci_dev = pdev;
+       pci_set_drvdata(pdev, dev);
+       dma_set_mask(&dev->pci_dev->dev, DMA_BIT_MASK(64));
+       nvme_set_instance(dev);
+
+       dev->bar = ioremap(pci_resource_start(pdev, 0), 8192);
+       if (!dev->bar) {
+               result = -ENOMEM;
+               goto disable;
+       }
+
+       result = nvme_configure_admin_queue(dev);
+       if (result)
+               goto unmap;
+       dev->queue_count++;
+
+       result = nvme_dev_add(dev);
+       if (result)
+               goto delete;
+       return 0;
+
+ delete:
+       nvme_free_queues(dev);
+ unmap:
+       iounmap(dev->bar);
+ disable:
+       pci_disable_msix(pdev);
+       nvme_release_instance(dev);
+ free:
+       kfree(dev->queues);
+       kfree(dev->entry);
+       kfree(dev);
+       return result;
+}
+
+static void __devexit nvme_remove(struct pci_dev *pdev)
+{
+       struct nvme_dev *dev = pci_get_drvdata(pdev);
+       nvme_dev_remove(dev);
+       pci_disable_msix(pdev);
+       iounmap(dev->bar);
+       nvme_release_instance(dev);
+       kfree(dev->queues);
+       kfree(dev->entry);
+       kfree(dev);
+}
+
+/* These functions are yet to be implemented */
+#define nvme_error_detected NULL
+#define nvme_dump_registers NULL
+#define nvme_link_reset NULL
+#define nvme_slot_reset NULL
+#define nvme_error_resume NULL
+#define nvme_suspend NULL
+#define nvme_resume NULL
+
+static struct pci_error_handlers nvme_err_handler = {
+       .error_detected = nvme_error_detected,
+       .mmio_enabled   = nvme_dump_registers,
+       .link_reset     = nvme_link_reset,
+       .slot_reset     = nvme_slot_reset,
+       .resume         = nvme_error_resume,
+};
+
+/* Move to pci_ids.h later */
+#define PCI_CLASS_STORAGE_EXPRESS      0x010802
+
+static DEFINE_PCI_DEVICE_TABLE(nvme_id_table) = {
+       { PCI_DEVICE_CLASS(PCI_CLASS_STORAGE_EXPRESS, 0xffffff) },
+       { 0, }
+};
+MODULE_DEVICE_TABLE(pci, nvme_id_table);
+
+static struct pci_driver nvme_driver = {
+       .name           = "nvme",
+       .id_table       = nvme_id_table,
+       .probe          = nvme_probe,
+       .remove         = __devexit_p(nvme_remove),
+       .suspend        = nvme_suspend,
+       .resume         = nvme_resume,
+       .err_handler    = &nvme_err_handler,
+};
+
+static int __init nvme_init(void)
+{
+       int result;
+
+       nvme_major = register_blkdev(nvme_major, "nvme");
+       if (nvme_major <= 0)
+               return -EBUSY;
+
+       result = pci_register_driver(&nvme_driver);
+       if (!result)
+               return 0;
+
+       unregister_blkdev(nvme_major, "nvme");
+       return result;
+}
+
+static void __exit nvme_exit(void)
+{
+       pci_unregister_driver(&nvme_driver);
+       unregister_blkdev(nvme_major, "nvme");
+}
+
+MODULE_AUTHOR("Matthew Wilcox <willy@linux.intel.com>");
+MODULE_LICENSE("GPL");
+MODULE_VERSION("0.1");
+module_init(nvme_init);
+module_exit(nvme_exit);
diff --git a/include/linux/nvme.h b/include/linux/nvme.h
new file mode 100644 (file)
index 0000000..9ba5358
--- /dev/null
@@ -0,0 +1,343 @@
+/*
+ * Definitions for the NVM Express interface
+ * Copyright (c) 2011, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 
+ * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
+ */
+
+#ifndef _LINUX_NVME_H
+#define _LINUX_NVME_H
+
+#include <linux/types.h>
+
+struct nvme_bar {
+       __u64                   cap;    /* Controller Capabilities */
+       __u32                   vs;     /* Version */
+       __u32                   ims;    /* Interrupt Mask Set */
+       __u32                   imc;    /* Interrupt Mask Clear */
+       __u32                   cc;     /* Controller Configuration */
+       __u32                   csts;   /* Controller Status */
+       __u32                   aqa;    /* Admin Queue Attributes */
+       __u64                   asq;    /* Admin SQ Base Address */
+       __u64                   acq;    /* Admin CQ Base Address */
+};
+
+enum {
+       NVME_CC_ENABLE          = 1 << 0,
+       NVME_CC_CSS_NVM         = 0 << 4,
+       NVME_CC_MPS_SHIFT       = 7,
+       NVME_CC_ARB_RR          = 0 << 11,
+       NVME_CC_ARB_WRRU        = 1 << 11,
+       NVME_CC_ARB_VS          = 3 << 11,
+       NVME_CC_SHN_NONE        = 0 << 13,
+       NVME_CC_SHN_NORMAL      = 1 << 13,
+       NVME_CC_SHN_ABRUPT      = 2 << 13,
+       NVME_CSTS_RDY           = 1 << 0,
+       NVME_CSTS_CFS           = 1 << 1,
+       NVME_CSTS_SHST_NORMAL   = 0 << 2,
+       NVME_CSTS_SHST_OCCUR    = 1 << 2,
+       NVME_CSTS_SHST_CMPLT    = 2 << 2,
+};
+
+#define NVME_VS(major, minor)  (major << 16 | minor)
+
+struct nvme_id_ctrl {
+       __le16                  vid;
+       __le16                  ssvid;
+       char                    sn[20];
+       char                    mn[40];
+       char                    fr[8];
+       __le32                  nn;
+       __u8                    rab;
+       __u8                    rsvd77[178];
+       __le16                  oacs;
+       __u8                    acl;
+       __u8                    aerl;
+       __u8                    frmw;
+       __u8                    lpa;
+       __u8                    elpe;
+       __u8                    npss;
+       __u8                    rsvd264[248];
+       __le64                  psd[32];
+       __le16                  oncs;
+       __le16                  fuses;
+       __u8                    fna;
+       __u8                    vwc;
+       __le16                  awun;
+       __le16                  awupf;
+       __u8                    rsvd778[246];
+       __u8                    cmdset[2048];
+       __u8                    vs[1024];
+};
+
+struct nvme_lbaf {
+       __le16                  ms;
+       __u8                    ds;
+       __u8                    rp;
+};
+
+struct nvme_id_ns {
+       __le64                  nsze;
+       __le64                  ncap;
+       __le64                  nuse;
+       __u8                    nsfeat;
+       __u8                    nlbaf;
+       __u8                    flbas;
+       __u8                    mc;
+       __u8                    dpc;
+       __u8                    dps;
+       __u8                    rsvd30[98];
+       struct nvme_lbaf        lbaf[16];
+       __u8                    rsvd192[192];
+       __u8                    vs[3712];
+};
+
+enum {
+       NVME_NS_FEAT_THIN       = 1 << 0,
+       NVME_LBAF_RP_BEST       = 0,
+       NVME_LBAF_RP_BETTER     = 1,
+       NVME_LBAF_RP_GOOD       = 2,
+       NVME_LBAF_RP_DEGRADED   = 3,
+};
+
+struct nvme_lba_range_type {
+       __u8                    type;
+       __u8                    attributes;
+       __u8                    rsvd2[14];
+       __u64                   slba;
+       __u64                   nlb;
+       __u8                    guid[16];
+       __u8                    rsvd48[16];
+};
+
+enum {
+       NVME_LBART_TYPE_FS      = 0x01,
+       NVME_LBART_TYPE_RAID    = 0x02,
+       NVME_LBART_TYPE_CACHE   = 0x03,
+       NVME_LBART_TYPE_SWAP    = 0x04,
+
+       NVME_LBART_ATTRIB_TEMP  = 1 << 0,
+       NVME_LBART_ATTRIB_HIDE  = 1 << 1,
+};
+
+/* I/O commands */
+
+enum nvme_opcode {
+       nvme_cmd_flush          = 0x00,
+       nvme_cmd_write          = 0x01,
+       nvme_cmd_read           = 0x02,
+       nvme_cmd_write_uncor    = 0x04,
+       nvme_cmd_compare        = 0x05,
+       nvme_cmd_dsm            = 0x09,
+};
+
+struct nvme_rw_command {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __le32                  nsid;
+       __u64                   rsvd2;
+       __le64                  metadata;
+       __le64                  prp1;
+       __le64                  prp2;
+       __le64                  slba;
+       __le16                  length;
+       __le16                  control;
+       __le32                  dsmgmt;
+       __le32                  reftag;
+       __le16                  apptag;
+       __le16                  appmask;
+};
+
+enum {
+       NVME_RW_LR                      = 1 << 15,
+       NVME_RW_FUA                     = 1 << 14,
+       NVME_RW_DSM_FREQ_UNSPEC         = 0,
+       NVME_RW_DSM_FREQ_TYPICAL        = 1,
+       NVME_RW_DSM_FREQ_RARE           = 2,
+       NVME_RW_DSM_FREQ_READS          = 3,
+       NVME_RW_DSM_FREQ_WRITES         = 4,
+       NVME_RW_DSM_FREQ_RW             = 5,
+       NVME_RW_DSM_FREQ_ONCE           = 6,
+       NVME_RW_DSM_FREQ_PREFETCH       = 7,
+       NVME_RW_DSM_FREQ_TEMP           = 8,
+       NVME_RW_DSM_LATENCY_NONE        = 0 << 4,
+       NVME_RW_DSM_LATENCY_IDLE        = 1 << 4,
+       NVME_RW_DSM_LATENCY_NORM        = 2 << 4,
+       NVME_RW_DSM_LATENCY_LOW         = 3 << 4,
+       NVME_RW_DSM_SEQ_REQ             = 1 << 6,
+       NVME_RW_DSM_COMPRESSED          = 1 << 7,
+};
+
+/* Admin commands */
+
+enum nvme_admin_opcode {
+       nvme_admin_delete_sq            = 0x00,
+       nvme_admin_create_sq            = 0x01,
+       nvme_admin_get_features         = 0x02,
+       nvme_admin_delete_cq            = 0x04,
+       nvme_admin_create_cq            = 0x05,
+       nvme_admin_identify             = 0x06,
+       nvme_admin_abort_cmd            = 0x08,
+       nvme_admin_set_features         = 0x09,
+       nvme_admin_get_log_page         = 0x0a,
+       nvme_admin_async_event          = 0x0c,
+       nvme_admin_download_fw          = 0x0d,
+       nvme_admin_security_recv        = 0x0e,
+       nvme_admin_format_nvm           = 0x10,
+       nvme_admin_security_send        = 0x11,
+       nvme_admin_activate_fw          = 0x14,
+};
+
+enum {
+       NVME_QUEUE_PHYS_CONTIG  = (1 << 0),
+       NVME_CQ_IRQ_ENABLED     = (1 << 1),
+       NVME_SQ_PRIO_URGENT     = (0 << 1),
+       NVME_SQ_PRIO_HIGH       = (1 << 1),
+       NVME_SQ_PRIO_MEDIUM     = (2 << 1),
+       NVME_SQ_PRIO_LOW        = (3 << 1),
+       NVME_FEAT_ARBITRATION   = 0x01,
+       NVME_FEAT_POWER_MGMT    = 0x02,
+       NVME_FEAT_LBA_RANGE     = 0x03,
+       NVME_FEAT_TEMP_THRESH   = 0x04,
+       NVME_FEAT_ERR_RECOVERY  = 0x05,
+       NVME_FEAT_VOLATILE_WC   = 0x06,
+       NVME_FEAT_NUM_QUEUES    = 0x07,
+       NVME_FEAT_IRQ_COALESCE  = 0x08,
+       NVME_FEAT_IRQ_CONFIG    = 0x09,
+       NVME_FEAT_WRITE_ATOMIC  = 0x0a,
+       NVME_FEAT_ASYNC_EVENT   = 0x0b,
+       NVME_FEAT_SW_PROGRESS   = 0x0c,
+};
+
+struct nvme_identify {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __le32                  nsid;
+       __u64                   rsvd2[2];
+       __le64                  prp1;
+       __le64                  prp2;
+       __le32                  cns;
+       __u32                   rsvd11[5];
+};
+
+struct nvme_features {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __le32                  nsid;
+       __u64                   rsvd2[2];
+       __le64                  prp1;
+       __le64                  prp2;
+       __le32                  fid;
+       __le32                  dword11;
+       __u32                   rsvd12[4];
+};
+
+struct nvme_create_cq {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __le32                  rsvd1[5];
+       __le64                  prp1;
+       __u64                   rsvd8;
+       __le16                  cqid;
+       __le16                  qsize;
+       __le16                  cq_flags;
+       __le16                  irq_vector;
+       __u32                   rsvd12[4];
+};
+
+struct nvme_create_sq {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __le32                  rsvd1[5];
+       __le64                  prp1;
+       __u64                   rsvd8;
+       __le16                  sqid;
+       __le16                  qsize;
+       __le16                  sq_flags;
+       __le16                  cqid;
+       __le32                  rsvd12[4];
+};
+
+struct nvme_delete_queue {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __u32                   rsvd1[9];
+       __le16                  qid;
+       __le16                  rsvd10;
+       __le32                  rsvd11[5];
+};
+
+struct nvme_common_command {
+       __u8                    opcode;
+       __u8                    flags;
+       __u16                   command_id;
+       __le32                  nsid;
+       __u32                   rsvd2[14];
+};
+
+struct nvme_command {
+       union {
+               struct nvme_common_command common;
+               struct nvme_rw_command rw;
+               struct nvme_identify identify;
+               struct nvme_features features;
+               struct nvme_create_cq create_cq;
+               struct nvme_create_sq create_sq;
+               struct nvme_delete_queue delete_queue;
+       };
+};
+
+/* XXX: Sync with spec */
+enum {
+       NVME_SC_SUCCESS                 = 0x0,
+       NVME_SC_INVALID_OPCODE          = 0x1,
+       NVME_SC_INVALID_FIELD           = 0x2,
+       NVME_SC_CMDID_CONFLICT          = 0x3,
+       NVME_SC_DATA_XFER_ERROR         = 0x4,
+       NVME_SC_POWER_LOSS              = 0x5,
+       NVME_SC_INTERNAL                = 0x6,
+       NVME_SC_ABORT_REQ               = 0x7,
+       NVME_SC_ABORT_QUEUE             = 0x8,
+       NVME_SC_FUSED_FAIL              = 0x9,
+       NVME_SC_FUSED_MISSING           = 0xa,
+       NVME_SC_LBA_RANGE               = 0x80,
+       NVME_SC_CAP_EXCEEDED            = 0x81,
+       NVME_SC_NS_NOT_READY            = 0x82,
+       NVME_SC_CQ_INVALID              = 0x100,
+       NVME_SC_QID_INVALID             = 0x101,
+       NVME_SC_QUEUE_SIZE              = 0x102,
+       NVME_SC_WRITE_FAULT             = 0x280,
+       NVME_SC_READ_ERROR              = 0x281,
+};
+
+struct nvme_completion {
+       __le32  result;         /* Used by admin commands to return data */
+       __le32  rsvd;
+       __le16  sq_head;        /* how much of this queue may be reclaimed */
+       __le16  sq_id;          /* submission queue that generated this entry */
+       __u16   command_id;     /* of the command which completed */
+       __le16  status;         /* did the command fail, and if so, why? */
+};
+
+#define NVME_IOCTL_IDENTIFY_NS _IOW('N', 0x40, struct nvme_id_ns)
+#define NVME_IOCTL_IDENTIFY_CTRL _IOW('N', 0x41, struct nvme_id_ctrl)
+#define NVME_IOCTL_GET_RANGE_TYPE _IOW('N', 0x42, struct nvme_lba_range_type)
+
+#endif /* _LINUX_NVME_H */