[Qemu-devel] [PATCH 0/2] eliminate data write in bdrv_write_zeroes on Linux

[Qemu-devel] [PATCH 0/2] eliminate data write in bdrv_write_zeroes on Linux

[Denis V. Lunev]

These patches eliminate data writes completely on Linux if fallocate
FALLOC_FL_ZERO_RANGE or FALLOC_FL_PUNCH_HOLE are  supported on
underlying filesystem. This should seriously increase performance in
some cases.

Signed-off-by: Denis V. Lunev <den@openvz.org>
CC: Kevin Wolf <kwolf@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>

[Qemu-devel] [PATCH 1/2] block: use fallocate(FALLOC_FL_ZERO_RANGE) in handle_aiocb_write_zeroes

[Denis V. Lunev]

this efficiently writes zeroes in the middle of the file on Linux
systems if the kernel is capable enough.

Signed-off-by: Denis V. Lunev <den@openvz.org>
CC: Kevin Wolf <kwolf@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>
---
 block/raw-posix.c | 11 +++++++++++
 configure         | 19 +++++++++++++++++++
 2 files changed, 30 insertions(+)

diff --git a/block/raw-posix.c b/block/raw-posix.c
index e51293a..9e66cb7 100644
--- a/block/raw-posix.c
+++ b/block/raw-posix.c
@@ -919,6 +919,17 @@ static ssize_t handle_aiocb_write_zeroes(RawPosixAIOData *aiocb)
             return xfs_write_zeroes(s, aiocb->aio_offset, aiocb->aio_nbytes);
         }
 #endif
+
+#ifdef CONFIG_FALLOCATE_ZERO_RANGE
+        do {
+            if (fallocate(s->fd, CONFIG_FALLOCATE_ZERO_RANGE,
+                          aiocb->aio_offset, aiocb->aio_nbytes) == 0) {
+                return 0;
+            }
+        } while (errno == EINTR);
+
+        ret = -errno;
+#endif
     }
 
     if (ret == -ENODEV || ret == -ENOSYS || ret == -EOPNOTSUPP ||
diff --git a/configure b/configure
index cae588c..dfcf7b3 100755
--- a/configure
+++ b/configure
@@ -3309,6 +3309,22 @@ if compile_prog "" "" ; then
   fallocate_punch_hole=yes
 fi
 
+# check that fallocate supports range zeroing inside the file
+fallocate_zero_range=no
+cat > $TMPC << EOF
+#include <fcntl.h>
+#include <linux/falloc.h>
+
+int main(void)
+{
+    fallocate(0, FALLOC_FL_ZERO_RANGE, 0, 0);
+    return 0;
+}
+EOF
+if compile_prog "" "" ; then
+  fallocate_zero_range=yes
+fi
+
 # check for posix_fallocate
 posix_fallocate=no
 cat > $TMPC << EOF
@@ -4538,6 +4554,9 @@ fi
 if test "$fallocate_punch_hole" = "yes" ; then
   echo "CONFIG_FALLOCATE_PUNCH_HOLE=y" >> $config_host_mak
 fi
+if test "$fallocate_zero_range" = "yes" ; then
+  echo "CONFIG_FALLOCATE_ZERO_RANGE=y" >> $config_host_mak
+fi
 if test "$posix_fallocate" = "yes" ; then
   echo "CONFIG_POSIX_FALLOCATE=y" >> $config_host_mak
 fi
-- 
1.9.1

Re: [Qemu-devel] [PATCH 1/2] block: use fallocate(FALLOC_FL_ZERO_RANGE) in handle_aiocb_write_zeroes

[Roman Kagan]

On Thu, Dec 25, 2014 at 08:37:29AM +0300, Denis V. Lunev wrote:
> +#ifdef CONFIG_FALLOCATE_ZERO_RANGE
> +        do {
> +            if (fallocate(s->fd, CONFIG_FALLOCATE_ZERO_RANGE,

Must be a typo, FALLOC_FL_ZERO_RANGE is what you mean.

Roman.

[Qemu-devel] [PATCH 2/2] block: use fallocate(FALLOC_FL_PUNCH_HOLE) & fallocate(0) to write zeroes

[Denis V. Lunev]

This sequence works efficiently if FALLOC_FL_ZERO_RANGE is not supported.
The idea is that FALLOC_FL_PUNCH_HOLE could not increase file size
but it cleans already allocated blocks inside the file. If we have to
create something new, simple fallocate will do the job.

This should increase performance a bit for not-so-modern kernels or for
filesystems which do not support FALLOC_FL_ZERO_RANGE.

Signed-off-by: Denis V. Lunev <den@openvz.org>
CC: Kevin Wolf <kwolf@redhat.com>
CC: Stefan Hajnoczi <stefanha@redhat.com>
---
 block/raw-posix.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/block/raw-posix.c b/block/raw-posix.c
index 9e66cb7..60972a1 100644
--- a/block/raw-posix.c
+++ b/block/raw-posix.c
@@ -930,6 +930,18 @@ static ssize_t handle_aiocb_write_zeroes(RawPosixAIOData *aiocb)
 
         ret = -errno;
 #endif
+#ifdef CONFIG_FALLOCATE_PUNCH_HOLE
+        do {
+            if (fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
+                          aiocb->aio_offset, aiocb->aio_nbytes) == 0 &&
+                fallocate(s->fd, 0,
+                          aiocb->aio_offset, aiocb->aio_nbytes) == 0) {
+                return 0;
+            }
+        } while (errno == EINTR);
+
+        ret = -errno;
+#endif
     }
 
     if (ret == -ENODEV || ret == -ENOSYS || ret == -EOPNOTSUPP ||
-- 
1.9.1

Re: [Qemu-devel] [PATCH 2/2] block: use fallocate(FALLOC_FL_PUNCH_HOLE) & fallocate(0) to write zeroes

[Roman Kagan]

On Thu, Dec 25, 2014 at 08:37:30AM +0300, Denis V. Lunev wrote:
> This sequence works efficiently if FALLOC_FL_ZERO_RANGE is not supported.
> The idea is that FALLOC_FL_PUNCH_HOLE could not increase file size
> but it cleans already allocated blocks inside the file. If we have to
> create something new, simple fallocate will do the job.
> 
> This should increase performance a bit for not-so-modern kernels or for
> filesystems which do not support FALLOC_FL_ZERO_RANGE.
> 
> Signed-off-by: Denis V. Lunev <den@openvz.org>
> CC: Kevin Wolf <kwolf@redhat.com>
> CC: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  block/raw-posix.c | 12 ++++++++++++
>  1 file changed, 12 insertions(+)
> 
> diff --git a/block/raw-posix.c b/block/raw-posix.c
> index 9e66cb7..60972a1 100644
> --- a/block/raw-posix.c
> +++ b/block/raw-posix.c
> @@ -930,6 +930,18 @@ static ssize_t handle_aiocb_write_zeroes(RawPosixAIOData *aiocb)
>  
>          ret = -errno;
>  #endif
> +#ifdef CONFIG_FALLOCATE_PUNCH_HOLE
> +        do {
> +            if (fallocate(s->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
> +                          aiocb->aio_offset, aiocb->aio_nbytes) == 0 &&
> +                fallocate(s->fd, 0,
> +                          aiocb->aio_offset, aiocb->aio_nbytes) == 0) {
> +                return 0;
> +            }
> +        } while (errno == EINTR);
> +
> +        ret = -errno;
> +#endif

This is suboptimal in that fallocate(FALLOC_FL_ZERO_RANGE) would always
be called in vain for such systems.  Might be worth another flag in
BDRVRawState?

Roman.