block: add .bdrv_co_write_zeroes() interface
authorStefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Tue, 7 Feb 2012 13:27:25 +0000 (13:27 +0000)
committerKevin Wolf <kwolf@redhat.com>
Thu, 9 Feb 2012 15:17:50 +0000 (16:17 +0100)
The ability to zero regions of an image file is a useful primitive for
higher-level features such as image streaming or zero write detection.

Image formats may support an optimized metadata representation instead
of writing zeroes into the image file.  This allows zero writes to be
potentially faster than regular write operations and also preserve
sparseness of the image file.

The .bdrv_co_write_zeroes() interface should be implemented by block
drivers that wish to provide efficient zeroing.

Note that this operation is different from the discard operation, which
may leave the contents of the region indeterminate.  That means
discarded blocks are not guaranteed to contain zeroes and may contain
junk data instead.

Signed-off-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
block.c
block.h
block_int.h
trace-events

diff --git a/block.c b/block.c
index 3621d11de0ab8b012e9e0b2dd229cc0ded5c38aa..c9fa5c17fff47d4c4e0e30204e390b715de0ad13 100644 (file)
--- a/block.c
+++ b/block.c
@@ -50,6 +50,7 @@
 
 typedef enum {
     BDRV_REQ_COPY_ON_READ = 0x1,
+    BDRV_REQ_ZERO_WRITE   = 0x2,
 } BdrvRequestFlags;
 
 static void bdrv_dev_change_media_cb(BlockDriverState *bs, bool load);
@@ -69,7 +70,8 @@ static int coroutine_fn bdrv_co_do_readv(BlockDriverState *bs,
     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
     BdrvRequestFlags flags);
 static int coroutine_fn bdrv_co_do_writev(BlockDriverState *bs,
-    int64_t sector_num, int nb_sectors, QEMUIOVector *qiov);
+    int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
+    BdrvRequestFlags flags);
 static BlockDriverAIOCB *bdrv_co_aio_rw_vector(BlockDriverState *bs,
                                                int64_t sector_num,
                                                QEMUIOVector *qiov,
@@ -1300,7 +1302,7 @@ static void coroutine_fn bdrv_rw_co_entry(void *opaque)
                                      rwco->nb_sectors, rwco->qiov, 0);
     } else {
         rwco->ret = bdrv_co_do_writev(rwco->bs, rwco->sector_num,
-                                      rwco->nb_sectors, rwco->qiov);
+                                      rwco->nb_sectors, rwco->qiov, 0);
     }
 }
 
@@ -1639,11 +1641,37 @@ int coroutine_fn bdrv_co_copy_on_readv(BlockDriverState *bs,
                             BDRV_REQ_COPY_ON_READ);
 }
 
+static int coroutine_fn bdrv_co_do_write_zeroes(BlockDriverState *bs,
+    int64_t sector_num, int nb_sectors)
+{
+    BlockDriver *drv = bs->drv;
+    QEMUIOVector qiov;
+    struct iovec iov;
+    int ret;
+
+    /* First try the efficient write zeroes operation */
+    if (drv->bdrv_co_write_zeroes) {
+        return drv->bdrv_co_write_zeroes(bs, sector_num, nb_sectors);
+    }
+
+    /* Fall back to bounce buffer if write zeroes is unsupported */
+    iov.iov_len  = nb_sectors * BDRV_SECTOR_SIZE;
+    iov.iov_base = qemu_blockalign(bs, iov.iov_len);
+    memset(iov.iov_base, 0, iov.iov_len);
+    qemu_iovec_init_external(&qiov, &iov, 1);
+
+    ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, &qiov);
+
+    qemu_vfree(iov.iov_base);
+    return ret;
+}
+
 /*
  * Handle a write request in coroutine context
  */
 static int coroutine_fn bdrv_co_do_writev(BlockDriverState *bs,
-    int64_t sector_num, int nb_sectors, QEMUIOVector *qiov)
+    int64_t sector_num, int nb_sectors, QEMUIOVector *qiov,
+    BdrvRequestFlags flags)
 {
     BlockDriver *drv = bs->drv;
     BdrvTrackedRequest req;
@@ -1670,7 +1698,11 @@ static int coroutine_fn bdrv_co_do_writev(BlockDriverState *bs,
 
     tracked_request_begin(&req, bs, sector_num, nb_sectors, true);
 
-    ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov);
+    if (flags & BDRV_REQ_ZERO_WRITE) {
+        ret = bdrv_co_do_write_zeroes(bs, sector_num, nb_sectors);
+    } else {
+        ret = drv->bdrv_co_writev(bs, sector_num, nb_sectors, qiov);
+    }
 
     if (bs->dirty_bitmap) {
         set_dirty_bitmap(bs, sector_num, nb_sectors, 1);
@@ -1690,7 +1722,16 @@ int coroutine_fn bdrv_co_writev(BlockDriverState *bs, int64_t sector_num,
 {
     trace_bdrv_co_writev(bs, sector_num, nb_sectors);
 
-    return bdrv_co_do_writev(bs, sector_num, nb_sectors, qiov);
+    return bdrv_co_do_writev(bs, sector_num, nb_sectors, qiov, 0);
+}
+
+int coroutine_fn bdrv_co_write_zeroes(BlockDriverState *bs,
+                                      int64_t sector_num, int nb_sectors)
+{
+    trace_bdrv_co_write_zeroes(bs, sector_num, nb_sectors);
+
+    return bdrv_co_do_writev(bs, sector_num, nb_sectors, NULL,
+                             BDRV_REQ_ZERO_WRITE);
 }
 
 /**
@@ -3192,7 +3233,7 @@ static void coroutine_fn bdrv_co_do_rw(void *opaque)
             acb->req.nb_sectors, acb->req.qiov, 0);
     } else {
         acb->req.error = bdrv_co_do_writev(bs, acb->req.sector,
-            acb->req.nb_sectors, acb->req.qiov);
+            acb->req.nb_sectors, acb->req.qiov, 0);
     }
 
     acb->bh = qemu_bh_new(bdrv_co_em_bh, acb);
diff --git a/block.h b/block.h
index cae289b2fbf7a217bb693db075e7a5872b404c2f..60ea7308c6c95a1c8b68d0e1d21c53a6d05a3992 100644 (file)
--- a/block.h
+++ b/block.h
@@ -146,6 +146,14 @@ int coroutine_fn bdrv_co_copy_on_readv(BlockDriverState *bs,
     int64_t sector_num, int nb_sectors, QEMUIOVector *qiov);
 int coroutine_fn bdrv_co_writev(BlockDriverState *bs, int64_t sector_num,
     int nb_sectors, QEMUIOVector *qiov);
+/*
+ * Efficiently zero a region of the disk image.  Note that this is a regular
+ * I/O request like read or write and should have a reasonable size.  This
+ * function is not suitable for zeroing the entire image in a single request
+ * because it may allocate memory for the entire region.
+ */
+int coroutine_fn bdrv_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
+    int nb_sectors);
 int coroutine_fn bdrv_co_is_allocated(BlockDriverState *bs, int64_t sector_num,
     int nb_sectors, int *pnum);
 BlockDriverState *bdrv_find_backing_image(BlockDriverState *bs,
index 7be2988ca7098f3ca16bf24ff69dfa957690fb7a..7946cf66a1d6375e83bc7ad37e08406c561a71b8 100644 (file)
@@ -131,6 +131,14 @@ struct BlockDriver {
         int64_t sector_num, int nb_sectors, QEMUIOVector *qiov);
     int coroutine_fn (*bdrv_co_writev)(BlockDriverState *bs,
         int64_t sector_num, int nb_sectors, QEMUIOVector *qiov);
+    /*
+     * Efficiently zero a region of the disk image.  Typically an image format
+     * would use a compact metadata representation to implement this.  This
+     * function pointer may be NULL and .bdrv_co_writev() will be called
+     * instead.
+     */
+    int coroutine_fn (*bdrv_co_write_zeroes)(BlockDriverState *bs,
+        int64_t sector_num, int nb_sectors);
     int coroutine_fn (*bdrv_co_discard)(BlockDriverState *bs,
         int64_t sector_num, int nb_sectors);
     int coroutine_fn (*bdrv_co_is_allocated)(BlockDriverState *bs,
index 5d057496436dc00e33418059b99893958777badc..3372087ed9669ddbea487210672864c5d7d080e2 100644 (file)
@@ -67,6 +67,7 @@ bdrv_lock_medium(void *bs, bool locked) "bs %p locked %d"
 bdrv_co_readv(void *bs, int64_t sector_num, int nb_sector) "bs %p sector_num %"PRId64" nb_sectors %d"
 bdrv_co_copy_on_readv(void *bs, int64_t sector_num, int nb_sector) "bs %p sector_num %"PRId64" nb_sectors %d"
 bdrv_co_writev(void *bs, int64_t sector_num, int nb_sector) "bs %p sector_num %"PRId64" nb_sectors %d"
+bdrv_co_write_zeroes(void *bs, int64_t sector_num, int nb_sector) "bs %p sector_num %"PRId64" nb_sectors %d"
 bdrv_co_io_em(void *bs, int64_t sector_num, int nb_sectors, int is_write, void *acb) "bs %p sector_num %"PRId64" nb_sectors %d is_write %d acb %p"
 bdrv_co_do_copy_on_readv(void *bs, int64_t sector_num, int nb_sectors, int64_t cluster_sector_num, int cluster_nb_sectors) "bs %p sector_num %"PRId64" nb_sectors %d cluster_sector_num %"PRId64" cluster_nb_sectors %d"