xen: add virtual block device driver.
authorJeremy Fitzhardinge <jeremy@xensource.com>
Wed, 18 Jul 2007 01:37:06 +0000 (18:37 -0700)
committerJeremy Fitzhardinge <jeremy@goop.org>
Wed, 18 Jul 2007 15:47:45 +0000 (08:47 -0700)
The block device frontend driver allows the kernel to access block
devices exported exported by a virtual machine containing a physical
block device driver.

Signed-off-by: Ian Pratt <ian.pratt@xensource.com>
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
Signed-off-by: Chris Wright <chrisw@sous-sol.org>
Cc: Arjan van de Ven <arjan@infradead.org>
Cc: Greg KH <greg@kroah.com>
Cc: Jens Axboe <axboe@kernel.dk>
drivers/block/Kconfig
drivers/block/Makefile
drivers/block/xen-blkfront.c [new file with mode: 0644]
include/linux/major.h

index 8f65b88cf7113bdcf0c29b67cdabc93c36091fc4..a4a311992408f5eda8ab093bebbca107dd04faea 100644 (file)
@@ -427,4 +427,13 @@ config XILINX_SYSACE
        help
          Include support for the Xilinx SystemACE CompactFlash interface
 
+config XEN_BLKDEV_FRONTEND
+       tristate "Xen virtual block device support"
+       depends on XEN
+       default y
+       help
+         This driver implements the front-end of the Xen virtual
+         block device driver.  It communicates with a back-end driver
+         in another domain which drives the actual block device.
+
 endif # BLK_DEV
index 9ee08ab4ffa8a474465564e8df17a99932a48059..3e31532df0eda9bfa6266ad5db9058d238220b68 100644 (file)
@@ -29,3 +29,4 @@ obj-$(CONFIG_VIODASD)         += viodasd.o
 obj-$(CONFIG_BLK_DEV_SX8)      += sx8.o
 obj-$(CONFIG_BLK_DEV_UB)       += ub.o
 
+obj-$(CONFIG_XEN_BLKDEV_FRONTEND)      += xen-blkfront.o
diff --git a/drivers/block/xen-blkfront.c b/drivers/block/xen-blkfront.c
new file mode 100644 (file)
index 0000000..6746c29
--- /dev/null
@@ -0,0 +1,988 @@
+/*
+ * blkfront.c
+ *
+ * XenLinux virtual block device driver.
+ *
+ * Copyright (c) 2003-2004, Keir Fraser & Steve Hand
+ * Modifications by Mark A. Williamson are (c) Intel Research Cambridge
+ * Copyright (c) 2004, Christian Limpach
+ * Copyright (c) 2004, Andrew Warfield
+ * Copyright (c) 2005, Christopher Clark
+ * Copyright (c) 2005, XenSource Ltd
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License version 2
+ * as published by the Free Software Foundation; or, when distributed
+ * separately from the Linux kernel or incorporated into other
+ * software packages, subject to the following license:
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this source file (the "Software"), to deal in the Software without
+ * restriction, including without limitation the rights to use, copy, modify,
+ * merge, publish, distribute, sublicense, and/or sell copies of the Software,
+ * and to permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <linux/interrupt.h>
+#include <linux/blkdev.h>
+#include <linux/module.h>
+
+#include <xen/xenbus.h>
+#include <xen/grant_table.h>
+#include <xen/events.h>
+#include <xen/page.h>
+
+#include <xen/interface/grant_table.h>
+#include <xen/interface/io/blkif.h>
+
+#include <asm/xen/hypervisor.h>
+
+enum blkif_state {
+       BLKIF_STATE_DISCONNECTED,
+       BLKIF_STATE_CONNECTED,
+       BLKIF_STATE_SUSPENDED,
+};
+
+struct blk_shadow {
+       struct blkif_request req;
+       unsigned long request;
+       unsigned long frame[BLKIF_MAX_SEGMENTS_PER_REQUEST];
+};
+
+static struct block_device_operations xlvbd_block_fops;
+
+#define BLK_RING_SIZE __RING_SIZE((struct blkif_sring *)0, PAGE_SIZE)
+
+/*
+ * We have one of these per vbd, whether ide, scsi or 'other'.  They
+ * hang in private_data off the gendisk structure. We may end up
+ * putting all kinds of interesting stuff here :-)
+ */
+struct blkfront_info
+{
+       struct xenbus_device *xbdev;
+       dev_t dev;
+       struct gendisk *gd;
+       int vdevice;
+       blkif_vdev_t handle;
+       enum blkif_state connected;
+       int ring_ref;
+       struct blkif_front_ring ring;
+       unsigned int evtchn, irq;
+       struct request_queue *rq;
+       struct work_struct work;
+       struct gnttab_free_callback callback;
+       struct blk_shadow shadow[BLK_RING_SIZE];
+       unsigned long shadow_free;
+       int feature_barrier;
+
+       /**
+        * The number of people holding this device open.  We won't allow a
+        * hot-unplug unless this is 0.
+        */
+       int users;
+};
+
+static DEFINE_SPINLOCK(blkif_io_lock);
+
+#define MAXIMUM_OUTSTANDING_BLOCK_REQS \
+       (BLKIF_MAX_SEGMENTS_PER_REQUEST * BLK_RING_SIZE)
+#define GRANT_INVALID_REF      0
+
+#define PARTS_PER_DISK         16
+
+#define BLKIF_MAJOR(dev) ((dev)>>8)
+#define BLKIF_MINOR(dev) ((dev) & 0xff)
+
+#define DEV_NAME       "xvd"   /* name in /dev */
+
+/* Information about our VBDs. */
+#define MAX_VBDS 64
+static LIST_HEAD(vbds_list);
+
+static int get_id_from_freelist(struct blkfront_info *info)
+{
+       unsigned long free = info->shadow_free;
+       BUG_ON(free > BLK_RING_SIZE);
+       info->shadow_free = info->shadow[free].req.id;
+       info->shadow[free].req.id = 0x0fffffee; /* debug */
+       return free;
+}
+
+static void add_id_to_freelist(struct blkfront_info *info,
+                              unsigned long id)
+{
+       info->shadow[id].req.id  = info->shadow_free;
+       info->shadow[id].request = 0;
+       info->shadow_free = id;
+}
+
+static void blkif_restart_queue_callback(void *arg)
+{
+       struct blkfront_info *info = (struct blkfront_info *)arg;
+       schedule_work(&info->work);
+}
+
+/*
+ * blkif_queue_request
+ *
+ * request block io
+ *
+ * id: for guest use only.
+ * operation: BLKIF_OP_{READ,WRITE,PROBE}
+ * buffer: buffer to read/write into. this should be a
+ *   virtual address in the guest os.
+ */
+static int blkif_queue_request(struct request *req)
+{
+       struct blkfront_info *info = req->rq_disk->private_data;
+       unsigned long buffer_mfn;
+       struct blkif_request *ring_req;
+       struct bio *bio;
+       struct bio_vec *bvec;
+       int idx;
+       unsigned long id;
+       unsigned int fsect, lsect;
+       int ref;
+       grant_ref_t gref_head;
+
+       if (unlikely(info->connected != BLKIF_STATE_CONNECTED))
+               return 1;
+
+       if (gnttab_alloc_grant_references(
+               BLKIF_MAX_SEGMENTS_PER_REQUEST, &gref_head) < 0) {
+               gnttab_request_free_callback(
+                       &info->callback,
+                       blkif_restart_queue_callback,
+                       info,
+                       BLKIF_MAX_SEGMENTS_PER_REQUEST);
+               return 1;
+       }
+
+       /* Fill out a communications ring structure. */
+       ring_req = RING_GET_REQUEST(&info->ring, info->ring.req_prod_pvt);
+       id = get_id_from_freelist(info);
+       info->shadow[id].request = (unsigned long)req;
+
+       ring_req->id = id;
+       ring_req->sector_number = (blkif_sector_t)req->sector;
+       ring_req->handle = info->handle;
+
+       ring_req->operation = rq_data_dir(req) ?
+               BLKIF_OP_WRITE : BLKIF_OP_READ;
+       if (blk_barrier_rq(req))
+               ring_req->operation = BLKIF_OP_WRITE_BARRIER;
+
+       ring_req->nr_segments = 0;
+       rq_for_each_bio (bio, req) {
+               bio_for_each_segment (bvec, bio, idx) {
+                       BUG_ON(ring_req->nr_segments
+                              == BLKIF_MAX_SEGMENTS_PER_REQUEST);
+                       buffer_mfn = pfn_to_mfn(page_to_pfn(bvec->bv_page));
+                       fsect = bvec->bv_offset >> 9;
+                       lsect = fsect + (bvec->bv_len >> 9) - 1;
+                       /* install a grant reference. */
+                       ref = gnttab_claim_grant_reference(&gref_head);
+                       BUG_ON(ref == -ENOSPC);
+
+                       gnttab_grant_foreign_access_ref(
+                               ref,
+                               info->xbdev->otherend_id,
+                               buffer_mfn,
+                               rq_data_dir(req) );
+
+                       info->shadow[id].frame[ring_req->nr_segments] =
+                               mfn_to_pfn(buffer_mfn);
+
+                       ring_req->seg[ring_req->nr_segments] =
+                               (struct blkif_request_segment) {
+                                       .gref       = ref,
+                                       .first_sect = fsect,
+                                       .last_sect  = lsect };
+
+                       ring_req->nr_segments++;
+               }
+       }
+
+       info->ring.req_prod_pvt++;
+
+       /* Keep a private copy so we can reissue requests when recovering. */
+       info->shadow[id].req = *ring_req;
+
+       gnttab_free_grant_references(gref_head);
+
+       return 0;
+}
+
+
+static inline void flush_requests(struct blkfront_info *info)
+{
+       int notify;
+
+       RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(&info->ring, notify);
+
+       if (notify)
+               notify_remote_via_irq(info->irq);
+}
+
+/*
+ * do_blkif_request
+ *  read a block; request is in a request queue
+ */
+static void do_blkif_request(request_queue_t *rq)
+{
+       struct blkfront_info *info = NULL;
+       struct request *req;
+       int queued;
+
+       pr_debug("Entered do_blkif_request\n");
+
+       queued = 0;
+
+       while ((req = elv_next_request(rq)) != NULL) {
+               info = req->rq_disk->private_data;
+               if (!blk_fs_request(req)) {
+                       end_request(req, 0);
+                       continue;
+               }
+
+               if (RING_FULL(&info->ring))
+                       goto wait;
+
+               pr_debug("do_blk_req %p: cmd %p, sec %lx, "
+                        "(%u/%li) buffer:%p [%s]\n",
+                        req, req->cmd, (unsigned long)req->sector,
+                        req->current_nr_sectors,
+                        req->nr_sectors, req->buffer,
+                        rq_data_dir(req) ? "write" : "read");
+
+
+               blkdev_dequeue_request(req);
+               if (blkif_queue_request(req)) {
+                       blk_requeue_request(rq, req);
+wait:
+                       /* Avoid pointless unplugs. */
+                       blk_stop_queue(rq);
+                       break;
+               }
+
+               queued++;
+       }
+
+       if (queued != 0)
+               flush_requests(info);
+}
+
+static int xlvbd_init_blk_queue(struct gendisk *gd, u16 sector_size)
+{
+       request_queue_t *rq;
+
+       rq = blk_init_queue(do_blkif_request, &blkif_io_lock);
+       if (rq == NULL)
+               return -1;
+
+       elevator_init(rq, "noop");
+
+       /* Hard sector size and max sectors impersonate the equiv. hardware. */
+       blk_queue_hardsect_size(rq, sector_size);
+       blk_queue_max_sectors(rq, 512);
+
+       /* Each segment in a request is up to an aligned page in size. */
+       blk_queue_segment_boundary(rq, PAGE_SIZE - 1);
+       blk_queue_max_segment_size(rq, PAGE_SIZE);
+
+       /* Ensure a merged request will fit in a single I/O ring slot. */
+       blk_queue_max_phys_segments(rq, BLKIF_MAX_SEGMENTS_PER_REQUEST);
+       blk_queue_max_hw_segments(rq, BLKIF_MAX_SEGMENTS_PER_REQUEST);
+
+       /* Make sure buffer addresses are sector-aligned. */
+       blk_queue_dma_alignment(rq, 511);
+
+       gd->queue = rq;
+
+       return 0;
+}
+
+
+static int xlvbd_barrier(struct blkfront_info *info)
+{
+       int err;
+
+       err = blk_queue_ordered(info->rq,
+                               info->feature_barrier ? QUEUE_ORDERED_DRAIN : QUEUE_ORDERED_NONE,
+                               NULL);
+
+       if (err)
+               return err;
+
+       printk(KERN_INFO "blkfront: %s: barriers %s\n",
+              info->gd->disk_name,
+              info->feature_barrier ? "enabled" : "disabled");
+       return 0;
+}
+
+
+static int xlvbd_alloc_gendisk(int minor, blkif_sector_t capacity,
+                              int vdevice, u16 vdisk_info, u16 sector_size,
+                              struct blkfront_info *info)
+{
+       struct gendisk *gd;
+       int nr_minors = 1;
+       int err = -ENODEV;
+
+       BUG_ON(info->gd != NULL);
+       BUG_ON(info->rq != NULL);
+
+       if ((minor % PARTS_PER_DISK) == 0)
+               nr_minors = PARTS_PER_DISK;
+
+       gd = alloc_disk(nr_minors);
+       if (gd == NULL)
+               goto out;
+
+       if (nr_minors > 1)
+               sprintf(gd->disk_name, "%s%c", DEV_NAME,
+                       'a' + minor / PARTS_PER_DISK);
+       else
+               sprintf(gd->disk_name, "%s%c%d", DEV_NAME,
+                       'a' + minor / PARTS_PER_DISK,
+                       minor % PARTS_PER_DISK);
+
+       gd->major = XENVBD_MAJOR;
+       gd->first_minor = minor;
+       gd->fops = &xlvbd_block_fops;
+       gd->private_data = info;
+       gd->driverfs_dev = &(info->xbdev->dev);
+       set_capacity(gd, capacity);
+
+       if (xlvbd_init_blk_queue(gd, sector_size)) {
+               del_gendisk(gd);
+               goto out;
+       }
+
+       info->rq = gd->queue;
+       info->gd = gd;
+
+       if (info->feature_barrier)
+               xlvbd_barrier(info);
+
+       if (vdisk_info & VDISK_READONLY)
+               set_disk_ro(gd, 1);
+
+       if (vdisk_info & VDISK_REMOVABLE)
+               gd->flags |= GENHD_FL_REMOVABLE;
+
+       if (vdisk_info & VDISK_CDROM)
+               gd->flags |= GENHD_FL_CD;
+
+       return 0;
+
+ out:
+       return err;
+}
+
+static void kick_pending_request_queues(struct blkfront_info *info)
+{
+       if (!RING_FULL(&info->ring)) {
+               /* Re-enable calldowns. */
+               blk_start_queue(info->rq);
+               /* Kick things off immediately. */
+               do_blkif_request(info->rq);
+       }
+}
+
+static void blkif_restart_queue(struct work_struct *work)
+{
+       struct blkfront_info *info = container_of(work, struct blkfront_info, work);
+
+       spin_lock_irq(&blkif_io_lock);
+       if (info->connected == BLKIF_STATE_CONNECTED)
+               kick_pending_request_queues(info);
+       spin_unlock_irq(&blkif_io_lock);
+}
+
+static void blkif_free(struct blkfront_info *info, int suspend)
+{
+       /* Prevent new requests being issued until we fix things up. */
+       spin_lock_irq(&blkif_io_lock);
+       info->connected = suspend ?
+               BLKIF_STATE_SUSPENDED : BLKIF_STATE_DISCONNECTED;
+       /* No more blkif_request(). */
+       if (info->rq)
+               blk_stop_queue(info->rq);
+       /* No more gnttab callback work. */
+       gnttab_cancel_free_callback(&info->callback);
+       spin_unlock_irq(&blkif_io_lock);
+
+       /* Flush gnttab callback work. Must be done with no locks held. */
+       flush_scheduled_work();
+
+       /* Free resources associated with old device channel. */
+       if (info->ring_ref != GRANT_INVALID_REF) {
+               gnttab_end_foreign_access(info->ring_ref, 0,
+                                         (unsigned long)info->ring.sring);
+               info->ring_ref = GRANT_INVALID_REF;
+               info->ring.sring = NULL;
+       }
+       if (info->irq)
+               unbind_from_irqhandler(info->irq, info);
+       info->evtchn = info->irq = 0;
+
+}
+
+static void blkif_completion(struct blk_shadow *s)
+{
+       int i;
+       for (i = 0; i < s->req.nr_segments; i++)
+               gnttab_end_foreign_access(s->req.seg[i].gref, 0, 0UL);
+}
+
+static irqreturn_t blkif_interrupt(int irq, void *dev_id)
+{
+       struct request *req;
+       struct blkif_response *bret;
+       RING_IDX i, rp;
+       unsigned long flags;
+       struct blkfront_info *info = (struct blkfront_info *)dev_id;
+       int uptodate;
+
+       spin_lock_irqsave(&blkif_io_lock, flags);
+
+       if (unlikely(info->connected != BLKIF_STATE_CONNECTED)) {
+               spin_unlock_irqrestore(&blkif_io_lock, flags);
+               return IRQ_HANDLED;
+       }
+
+ again:
+       rp = info->ring.sring->rsp_prod;
+       rmb(); /* Ensure we see queued responses up to 'rp'. */
+
+       for (i = info->ring.rsp_cons; i != rp; i++) {
+               unsigned long id;
+               int ret;
+
+               bret = RING_GET_RESPONSE(&info->ring, i);
+               id   = bret->id;
+               req  = (struct request *)info->shadow[id].request;
+
+               blkif_completion(&info->shadow[id]);
+
+               add_id_to_freelist(info, id);
+
+               uptodate = (bret->status == BLKIF_RSP_OKAY);
+               switch (bret->operation) {
+               case BLKIF_OP_WRITE_BARRIER:
+                       if (unlikely(bret->status == BLKIF_RSP_EOPNOTSUPP)) {
+                               printk(KERN_WARNING "blkfront: %s: write barrier op failed\n",
+                                      info->gd->disk_name);
+                               uptodate = -EOPNOTSUPP;
+                               info->feature_barrier = 0;
+                               xlvbd_barrier(info);
+                       }
+                       /* fall through */
+               case BLKIF_OP_READ:
+               case BLKIF_OP_WRITE:
+                       if (unlikely(bret->status != BLKIF_RSP_OKAY))
+                               dev_dbg(&info->xbdev->dev, "Bad return from blkdev data "
+                                       "request: %x\n", bret->status);
+
+                       ret = end_that_request_first(req, uptodate,
+                               req->hard_nr_sectors);
+                       BUG_ON(ret);
+                       end_that_request_last(req, uptodate);
+                       break;
+               default:
+                       BUG();
+               }
+       }
+
+       info->ring.rsp_cons = i;
+
+       if (i != info->ring.req_prod_pvt) {
+               int more_to_do;
+               RING_FINAL_CHECK_FOR_RESPONSES(&info->ring, more_to_do);
+               if (more_to_do)
+                       goto again;
+       } else
+               info->ring.sring->rsp_event = i + 1;
+
+       kick_pending_request_queues(info);
+
+       spin_unlock_irqrestore(&blkif_io_lock, flags);
+
+       return IRQ_HANDLED;
+}
+
+
+static int setup_blkring(struct xenbus_device *dev,
+                        struct blkfront_info *info)
+{
+       struct blkif_sring *sring;
+       int err;
+
+       info->ring_ref = GRANT_INVALID_REF;
+
+       sring = (struct blkif_sring *)__get_free_page(GFP_KERNEL);
+       if (!sring) {
+               xenbus_dev_fatal(dev, -ENOMEM, "allocating shared ring");
+               return -ENOMEM;
+       }
+       SHARED_RING_INIT(sring);
+       FRONT_RING_INIT(&info->ring, sring, PAGE_SIZE);
+
+       err = xenbus_grant_ring(dev, virt_to_mfn(info->ring.sring));
+       if (err < 0) {
+               free_page((unsigned long)sring);
+               info->ring.sring = NULL;
+               goto fail;
+       }
+       info->ring_ref = err;
+
+       err = xenbus_alloc_evtchn(dev, &info->evtchn);
+       if (err)
+               goto fail;
+
+       err = bind_evtchn_to_irqhandler(info->evtchn,
+                                       blkif_interrupt,
+                                       IRQF_SAMPLE_RANDOM, "blkif", info);
+       if (err <= 0) {
+               xenbus_dev_fatal(dev, err,
+                                "bind_evtchn_to_irqhandler failed");
+               goto fail;
+       }
+       info->irq = err;
+
+       return 0;
+fail:
+       blkif_free(info, 0);
+       return err;
+}
+
+
+/* Common code used when first setting up, and when resuming. */
+static int talk_to_backend(struct xenbus_device *dev,
+                          struct blkfront_info *info)
+{
+       const char *message = NULL;
+       struct xenbus_transaction xbt;
+       int err;
+
+       /* Create shared ring, alloc event channel. */
+       err = setup_blkring(dev, info);
+       if (err)
+               goto out;
+
+again:
+       err = xenbus_transaction_start(&xbt);
+       if (err) {
+               xenbus_dev_fatal(dev, err, "starting transaction");
+               goto destroy_blkring;
+       }
+
+       err = xenbus_printf(xbt, dev->nodename,
+                           "ring-ref", "%u", info->ring_ref);
+       if (err) {
+               message = "writing ring-ref";
+               goto abort_transaction;
+       }
+       err = xenbus_printf(xbt, dev->nodename,
+                           "event-channel", "%u", info->evtchn);
+       if (err) {
+               message = "writing event-channel";
+               goto abort_transaction;
+       }
+
+       err = xenbus_transaction_end(xbt, 0);
+       if (err) {
+               if (err == -EAGAIN)
+                       goto again;
+               xenbus_dev_fatal(dev, err, "completing transaction");
+               goto destroy_blkring;
+       }
+
+       xenbus_switch_state(dev, XenbusStateInitialised);
+
+       return 0;
+
+ abort_transaction:
+       xenbus_transaction_end(xbt, 1);
+       if (message)
+               xenbus_dev_fatal(dev, err, "%s", message);
+ destroy_blkring:
+       blkif_free(info, 0);
+ out:
+       return err;
+}
+
+
+/**
+ * Entry point to this code when a new device is created.  Allocate the basic
+ * structures and the ring buffer for communication with the backend, and
+ * inform the backend of the appropriate details for those.  Switch to
+ * Initialised state.
+ */
+static int blkfront_probe(struct xenbus_device *dev,
+                         const struct xenbus_device_id *id)
+{
+       int err, vdevice, i;
+       struct blkfront_info *info;
+
+       /* FIXME: Use dynamic device id if this is not set. */
+       err = xenbus_scanf(XBT_NIL, dev->nodename,
+                          "virtual-device", "%i", &vdevice);
+       if (err != 1) {
+               xenbus_dev_fatal(dev, err, "reading virtual-device");
+               return err;
+       }
+
+       info = kzalloc(sizeof(*info), GFP_KERNEL);
+       if (!info) {
+               xenbus_dev_fatal(dev, -ENOMEM, "allocating info structure");
+               return -ENOMEM;
+       }
+
+       info->xbdev = dev;
+       info->vdevice = vdevice;
+       info->connected = BLKIF_STATE_DISCONNECTED;
+       INIT_WORK(&info->work, blkif_restart_queue);
+
+       for (i = 0; i < BLK_RING_SIZE; i++)
+               info->shadow[i].req.id = i+1;
+       info->shadow[BLK_RING_SIZE-1].req.id = 0x0fffffff;
+
+       /* Front end dir is a number, which is used as the id. */
+       info->handle = simple_strtoul(strrchr(dev->nodename, '/')+1, NULL, 0);
+       dev->dev.driver_data = info;
+
+       err = talk_to_backend(dev, info);
+       if (err) {
+               kfree(info);
+               dev->dev.driver_data = NULL;
+               return err;
+       }
+
+       return 0;
+}
+
+
+static int blkif_recover(struct blkfront_info *info)
+{
+       int i;
+       struct blkif_request *req;
+       struct blk_shadow *copy;
+       int j;
+
+       /* Stage 1: Make a safe copy of the shadow state. */
+       copy = kmalloc(sizeof(info->shadow), GFP_KERNEL);
+       if (!copy)
+               return -ENOMEM;
+       memcpy(copy, info->shadow, sizeof(info->shadow));
+
+       /* Stage 2: Set up free list. */
+       memset(&info->shadow, 0, sizeof(info->shadow));
+       for (i = 0; i < BLK_RING_SIZE; i++)
+               info->shadow[i].req.id = i+1;
+       info->shadow_free = info->ring.req_prod_pvt;
+       info->shadow[BLK_RING_SIZE-1].req.id = 0x0fffffff;
+
+       /* Stage 3: Find pending requests and requeue them. */
+       for (i = 0; i < BLK_RING_SIZE; i++) {
+               /* Not in use? */
+               if (copy[i].request == 0)
+                       continue;
+
+               /* Grab a request slot and copy shadow state into it. */
+               req = RING_GET_REQUEST(&info->ring, info->ring.req_prod_pvt);
+               *req = copy[i].req;
+
+               /* We get a new request id, and must reset the shadow state. */
+               req->id = get_id_from_freelist(info);
+               memcpy(&info->shadow[req->id], &copy[i], sizeof(copy[i]));
+
+               /* Rewrite any grant references invalidated by susp/resume. */
+               for (j = 0; j < req->nr_segments; j++)
+                       gnttab_grant_foreign_access_ref(
+                               req->seg[j].gref,
+                               info->xbdev->otherend_id,
+                               pfn_to_mfn(info->shadow[req->id].frame[j]),
+                               rq_data_dir(
+                                       (struct request *)
+                                       info->shadow[req->id].request));
+               info->shadow[req->id].req = *req;
+
+               info->ring.req_prod_pvt++;
+       }
+
+       kfree(copy);
+
+       xenbus_switch_state(info->xbdev, XenbusStateConnected);
+
+       spin_lock_irq(&blkif_io_lock);
+
+       /* Now safe for us to use the shared ring */
+       info->connected = BLKIF_STATE_CONNECTED;
+
+       /* Send off requeued requests */
+       flush_requests(info);
+
+       /* Kick any other new requests queued since we resumed */
+       kick_pending_request_queues(info);
+
+       spin_unlock_irq(&blkif_io_lock);
+
+       return 0;
+}
+
+/**
+ * We are reconnecting to the backend, due to a suspend/resume, or a backend
+ * driver restart.  We tear down our blkif structure and recreate it, but
+ * leave the device-layer structures intact so that this is transparent to the
+ * rest of the kernel.
+ */
+static int blkfront_resume(struct xenbus_device *dev)
+{
+       struct blkfront_info *info = dev->dev.driver_data;
+       int err;
+
+       dev_dbg(&dev->dev, "blkfront_resume: %s\n", dev->nodename);
+
+       blkif_free(info, info->connected == BLKIF_STATE_CONNECTED);
+
+       err = talk_to_backend(dev, info);
+       if (info->connected == BLKIF_STATE_SUSPENDED && !err)
+               err = blkif_recover(info);
+
+       return err;
+}
+
+
+/*
+ * Invoked when the backend is finally 'ready' (and has told produced
+ * the details about the physical device - #sectors, size, etc).
+ */
+static void blkfront_connect(struct blkfront_info *info)
+{
+       unsigned long long sectors;
+       unsigned long sector_size;
+       unsigned int binfo;
+       int err;
+
+       if ((info->connected == BLKIF_STATE_CONNECTED) ||
+           (info->connected == BLKIF_STATE_SUSPENDED) )
+               return;
+
+       dev_dbg(&info->xbdev->dev, "%s:%s.\n",
+               __func__, info->xbdev->otherend);
+
+       err = xenbus_gather(XBT_NIL, info->xbdev->otherend,
+                           "sectors", "%llu", &sectors,
+                           "info", "%u", &binfo,
+                           "sector-size", "%lu", &sector_size,
+                           NULL);
+       if (err) {
+               xenbus_dev_fatal(info->xbdev, err,
+                                "reading backend fields at %s",
+                                info->xbdev->otherend);
+               return;
+       }
+
+       err = xenbus_gather(XBT_NIL, info->xbdev->otherend,
+                           "feature-barrier", "%lu", &info->feature_barrier,
+                           NULL);
+       if (err)
+               info->feature_barrier = 0;
+
+       err = xlvbd_alloc_gendisk(BLKIF_MINOR(info->vdevice),
+                                 sectors, info->vdevice,
+                                 binfo, sector_size, info);
+       if (err) {
+               xenbus_dev_fatal(info->xbdev, err, "xlvbd_add at %s",
+                                info->xbdev->otherend);
+               return;
+       }
+
+       xenbus_switch_state(info->xbdev, XenbusStateConnected);
+
+       /* Kick pending requests. */
+       spin_lock_irq(&blkif_io_lock);
+       info->connected = BLKIF_STATE_CONNECTED;
+       kick_pending_request_queues(info);
+       spin_unlock_irq(&blkif_io_lock);
+
+       add_disk(info->gd);
+}
+
+/**
+ * Handle the change of state of the backend to Closing.  We must delete our
+ * device-layer structures now, to ensure that writes are flushed through to
+ * the backend.  Once is this done, we can switch to Closed in
+ * acknowledgement.
+ */
+static void blkfront_closing(struct xenbus_device *dev)
+{
+       struct blkfront_info *info = dev->dev.driver_data;
+       unsigned long flags;
+
+       dev_dbg(&dev->dev, "blkfront_closing: %s removed\n", dev->nodename);
+
+       if (info->rq == NULL)
+               goto out;
+
+       spin_lock_irqsave(&blkif_io_lock, flags);
+
+       del_gendisk(info->gd);
+
+       /* No more blkif_request(). */
+       blk_stop_queue(info->rq);
+
+       /* No more gnttab callback work. */
+       gnttab_cancel_free_callback(&info->callback);
+       spin_unlock_irqrestore(&blkif_io_lock, flags);
+
+       /* Flush gnttab callback work. Must be done with no locks held. */
+       flush_scheduled_work();
+
+       blk_cleanup_queue(info->rq);
+       info->rq = NULL;
+
+ out:
+       xenbus_frontend_closed(dev);
+}
+
+/**
+ * Callback received when the backend's state changes.
+ */
+static void backend_changed(struct xenbus_device *dev,
+                           enum xenbus_state backend_state)
+{
+       struct blkfront_info *info = dev->dev.driver_data;
+       struct block_device *bd;
+
+       dev_dbg(&dev->dev, "blkfront:backend_changed.\n");
+
+       switch (backend_state) {
+       case XenbusStateInitialising:
+       case XenbusStateInitWait:
+       case XenbusStateInitialised:
+       case XenbusStateUnknown:
+       case XenbusStateClosed:
+               break;
+
+       case XenbusStateConnected:
+               blkfront_connect(info);
+               break;
+
+       case XenbusStateClosing:
+               bd = bdget(info->dev);
+               if (bd == NULL)
+                       xenbus_dev_fatal(dev, -ENODEV, "bdget failed");
+
+               mutex_lock(&bd->bd_mutex);
+               if (info->users > 0)
+                       xenbus_dev_error(dev, -EBUSY,
+                                        "Device in use; refusing to close");
+               else
+                       blkfront_closing(dev);
+               mutex_unlock(&bd->bd_mutex);
+               bdput(bd);
+               break;
+       }
+}
+
+static int blkfront_remove(struct xenbus_device *dev)
+{
+       struct blkfront_info *info = dev->dev.driver_data;
+
+       dev_dbg(&dev->dev, "blkfront_remove: %s removed\n", dev->nodename);
+
+       blkif_free(info, 0);
+
+       kfree(info);
+
+       return 0;
+}
+
+static int blkif_open(struct inode *inode, struct file *filep)
+{
+       struct blkfront_info *info = inode->i_bdev->bd_disk->private_data;
+       info->users++;
+       return 0;
+}
+
+static int blkif_release(struct inode *inode, struct file *filep)
+{
+       struct blkfront_info *info = inode->i_bdev->bd_disk->private_data;
+       info->users--;
+       if (info->users == 0) {
+               /* Check whether we have been instructed to close.  We will
+                  have ignored this request initially, as the device was
+                  still mounted. */
+               struct xenbus_device *dev = info->xbdev;
+               enum xenbus_state state = xenbus_read_driver_state(dev->otherend);
+
+               if (state == XenbusStateClosing)
+                       blkfront_closing(dev);
+       }
+       return 0;
+}
+
+static struct block_device_operations xlvbd_block_fops =
+{
+       .owner = THIS_MODULE,
+       .open = blkif_open,
+       .release = blkif_release,
+};
+
+
+static struct xenbus_device_id blkfront_ids[] = {
+       { "vbd" },
+       { "" }
+};
+
+static struct xenbus_driver blkfront = {
+       .name = "vbd",
+       .owner = THIS_MODULE,
+       .ids = blkfront_ids,
+       .probe = blkfront_probe,
+       .remove = blkfront_remove,
+       .resume = blkfront_resume,
+       .otherend_changed = backend_changed,
+};
+
+static int __init xlblk_init(void)
+{
+       if (!is_running_on_xen())
+               return -ENODEV;
+
+       if (register_blkdev(XENVBD_MAJOR, DEV_NAME)) {
+               printk(KERN_WARNING "xen_blk: can't get major %d with name %s\n",
+                      XENVBD_MAJOR, DEV_NAME);
+               return -ENODEV;
+       }
+
+       return xenbus_register_frontend(&blkfront);
+}
+module_init(xlblk_init);
+
+
+static void xlblk_exit(void)
+{
+       return xenbus_unregister_driver(&blkfront);
+}
+module_exit(xlblk_exit);
+
+MODULE_DESCRIPTION("Xen virtual block device frontend");
+MODULE_LICENSE("GPL");
+MODULE_ALIAS_BLOCKDEV_MAJOR(XENVBD_MAJOR);
index 7e7c9093919af3ab51461e7c14c849008ca582d6..0cb98053537af4874ef83cfc149c6af629758694 100644 (file)
 #define VXSPEC_MAJOR           200     /* VERITAS volume config driver */
 #define VXDMP_MAJOR            201     /* VERITAS volume multipath driver */
 
+#define XENVBD_MAJOR           202     /* Xen virtual block device */
+
 #define MSR_MAJOR              202
 #define CPUID_MAJOR            203