Since commit a6b257a08e3 ('file-posix: Handle undetectable alignment'),
we assume that if we open a file with O_DIRECT and alignment probing
returns 1, we just couldn't find out the real alignment requirement
because some filesystems make the requirement only for allocated blocks.
In this case, a safe default of 4k is used.

This is too strict NFS, which does actually allow byte-aligned requests
even with O_DIRECT. Because we can't distinguish both cases with generic
code, let's just look at the file system magic and disable
s->needs_alignment for NFS. This way, O_DIRECT can still be used on NFS
for images that are not aligned to 4k.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
---
 block/file-posix.c | 26 +++++++++++++++++++++++++-
 1 file changed, 25 insertions(+), 1 deletion(-)

diff --git a/block/file-posix.c b/block/file-posix.c
index 0c4e07c415..4e9dac461b 100644
--- a/block/file-posix.c
+++ b/block/file-posix.c
@@ -62,10 +62,12 @@
 #include <sys/ioctl.h>
 #include <sys/param.h>
 #include <sys/syscall.h>
+#include <sys/vfs.h>
 #include <linux/cdrom.h>
 #include <linux/fd.h>
 #include <linux/fs.h>
 #include <linux/hdreg.h>
+#include <linux/magic.h>
 #include <scsi/sg.h>
 #ifdef __s390__
 #include <asm/dasd.h>
@@ -300,6 +302,28 @@ static int probe_physical_blocksize(int fd, unsigned int *blk_size)
 #endif
 }
 
+/*
+ * Returns true if no alignment restrictions are necessary even for files
+ * opened with O_DIRECT.
+ *
+ * raw_probe_alignment() probes the required alignment and assume that 1 means
+ * the probing failed, so it falls back to a safe default of 4k. This can be
+ * avoided if we know that byte alignment is okay for the file.
+ */
+static bool dio_byte_aligned(int fd)
+{
+#ifdef __linux__
+    struct statfs buf;
+    int ret;
+
+    ret = fstatfs(fd, &buf);
+    if (ret == 0 && buf.f_type == NFS_SUPER_MAGIC) {
+        return true;
+    }
+#endif
+    return false;
+}
+
 /* Check if read is allowed with given memory buffer and length.
  *
  * This function is used to check O_DIRECT memory buffer and request alignment.
@@ -631,7 +655,7 @@ static int raw_open_common(BlockDriverState *bs, QDict *options,
 
     s->has_discard = true;
     s->has_write_zeroes = true;
-    if ((bs->open_flags & BDRV_O_NOCACHE) != 0) {
+    if ((bs->open_flags & BDRV_O_NOCACHE) != 0 && !dio_byte_aligned(s->fd)) {
         s->needs_alignment = true;
     }
 
-- 
2.25.4

On Fri, Jul 10, 2020 at 5:22 PM Kevin Wolf <kwolf@redhat.com> wrote:
>
> Since commit a6b257a08e3 ('file-posix: Handle undetectable alignment'),
> we assume that if we open a file with O_DIRECT and alignment probing
> returns 1, we just couldn't find out the real alignment requirement
> because some filesystems make the requirement only for allocated blocks.
> In this case, a safe default of 4k is used.
>
> This is too strict NFS, which does actually allow byte-aligned requests
> even with O_DIRECT. Because we can't distinguish both cases with generic
> code, let's just look at the file system magic and disable
> s->needs_alignment for NFS. This way, O_DIRECT can still be used on NFS
> for images that are not aligned to 4k.
>
> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> ---
>  block/file-posix.c | 26 +++++++++++++++++++++++++-
>  1 file changed, 25 insertions(+), 1 deletion(-)
>
> diff --git a/block/file-posix.c b/block/file-posix.c
> index 0c4e07c415..4e9dac461b 100644
> --- a/block/file-posix.c
> +++ b/block/file-posix.c
> @@ -62,10 +62,12 @@
>  #include <sys/ioctl.h>
>  #include <sys/param.h>
>  #include <sys/syscall.h>
> +#include <sys/vfs.h>
>  #include <linux/cdrom.h>
>  #include <linux/fd.h>
>  #include <linux/fs.h>
>  #include <linux/hdreg.h>
> +#include <linux/magic.h>
>  #include <scsi/sg.h>
>  #ifdef __s390__
>  #include <asm/dasd.h>
> @@ -300,6 +302,28 @@ static int probe_physical_blocksize(int fd, unsigned int *blk_size)
>  #endif
>  }
>
> +/*
> + * Returns true if no alignment restrictions are necessary even for files
> + * opened with O_DIRECT.
> + *
> + * raw_probe_alignment() probes the required alignment and assume that 1 means
> + * the probing failed, so it falls back to a safe default of 4k. This can be
> + * avoided if we know that byte alignment is okay for the file.
> + */
> +static bool dio_byte_aligned(int fd)
> +{
> +#ifdef __linux__
> +    struct statfs buf;
> +    int ret;
> +
> +    ret = fstatfs(fd, &buf);
> +    if (ret == 0 && buf.f_type == NFS_SUPER_MAGIC) {
> +        return true;
> +    }
> +#endif
> +    return false;
> +}
> +
>  /* Check if read is allowed with given memory buffer and length.
>   *
>   * This function is used to check O_DIRECT memory buffer and request alignment.
> @@ -631,7 +655,7 @@ static int raw_open_common(BlockDriverState *bs, QDict *options,
>
>      s->has_discard = true;
>      s->has_write_zeroes = true;
> -    if ((bs->open_flags & BDRV_O_NOCACHE) != 0) {
> +    if ((bs->open_flags & BDRV_O_NOCACHE) != 0 && !dio_byte_aligned(s->fd)) {
>          s->needs_alignment = true;

I did not know we have needs_alignment. Isn't this the same as using
request_alignment = 1?

For example we can check if we are on NFS and avoid the fallback to max_align:

    if (!bs->bl.request_alignment) {
        int i;
        size_t align;
        buf = qemu_memalign(max_align, max_align);
        for (i = 0; i < ARRAY_SIZE(alignments); i++) {
            align = alignments[i];
            if (raw_is_io_aligned(fd, buf, align)) {
                /* Fallback to safe value. */
                bs->bl.request_alignment = (align != 1) ? align : max_align;
                break;
            }
        }
        qemu_vfree(buf);
    }

After this we will have correct bl.request_alignment and buf_align.
Hopefully this will not break code expecting request_alignment >= 512.

Assuming that needs_alignment is well tested, this patch may be safer.

Nir

>      }
>
> --
> 2.25.4
>

On 7/10/20 9:21 AM, Kevin Wolf wrote:
> Since commit a6b257a08e3 ('file-posix: Handle undetectable alignment'),
> we assume that if we open a file with O_DIRECT and alignment probing
> returns 1, we just couldn't find out the real alignment requirement
> because some filesystems make the requirement only for allocated blocks.
> In this case, a safe default of 4k is used.
> 
> This is too strict NFS, which does actually allow byte-aligned requests

strict for

> even with O_DIRECT. Because we can't distinguish both cases with generic
> code, let's just look at the file system magic and disable
> s->needs_alignment for NFS. This way, O_DIRECT can still be used on NFS
> for images that are not aligned to 4k.
> 
> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
> ---
>   block/file-posix.c | 26 +++++++++++++++++++++++++-
>   1 file changed, 25 insertions(+), 1 deletion(-)
> 

Reviewed-by: Eric Blake <eblake@redhat.com>

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3226
Virtualization:  qemu.org | libvirt.org