Linux commit 825cf206ed51 ("statx: add direct I/O alignment
information") added an interface to fetch O_DIRECT alignment values for
block devices and file systems.

Prefer STATX_DIOALIGN to older interfaces and probing, but keep them as
fallbacks in case STATX_DIOALIGN cannot provide the information.

Testing shows the status of STATX_DIOALIGN support in Linux 6.1-rc3
appears to be:
- btrfs: no
- ext4: yes
- XFS: yes
- NVMe block devices: yes
- dm-crypt: yes

Cc: Eric Biggers <ebiggers@google.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 block/file-posix.c | 54 +++++++++++++++++++++++++++++++---------------
 1 file changed, 37 insertions(+), 17 deletions(-)

diff --git a/block/file-posix.c b/block/file-posix.c
index b9d62f52fe..00d8191880 100644
--- a/block/file-posix.c
+++ b/block/file-posix.c
@@ -372,28 +372,48 @@ static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
 
     bs->bl.request_alignment = 0;
     s->buf_align = 0;
+
+#if defined(__linux__) && defined(STATX_DIOALIGN)
+    struct statx stx;
+
+    /*
+     * Linux 6.1 introduced an interface for both block devices and file
+     * systems. The system call returns with the STATX_DIOALIGN bit cleared
+     * when the information is unavailable.
+     */
+    if (statx(fd, "", AT_EMPTY_PATH, STATX_DIOALIGN, &stx) == 0 &&
+        (stx.stx_mask & STATX_DIOALIGN)) {
+        bs->bl.request_alignment = stx.stx_dio_offset_align;
+        s->buf_align = stx.stx_dio_mem_align;
+    }
+#endif /* defined(__linux__) && defined(STATX_DIOALIGN) */
+
     /* Let's try to use the logical blocksize for the alignment. */
-    if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
-        bs->bl.request_alignment = 0;
+    if (!bs->bl.request_alignment) {
+        if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
+            bs->bl.request_alignment = 0;
+        }
     }
 
 #ifdef __linux__
-    /*
-     * The XFS ioctl definitions are shipped in extra packages that might
-     * not always be available. Since we just need the XFS_IOC_DIOINFO ioctl
-     * here, we simply use our own definition instead:
-     */
-    struct xfs_dioattr {
-        uint32_t d_mem;
-        uint32_t d_miniosz;
-        uint32_t d_maxiosz;
-    } da;
-    if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
-        bs->bl.request_alignment = da.d_miniosz;
-        /* The kernel returns wrong information for d_mem */
-        /* s->buf_align = da.d_mem; */
+    if (!bs->bl.request_alignment) {
+        /*
+         * The XFS ioctl definitions are shipped in extra packages that might
+         * not always be available. Since we just need the XFS_IOC_DIOINFO ioctl
+         * here, we simply use our own definition instead:
+         */
+        struct xfs_dioattr {
+            uint32_t d_mem;
+            uint32_t d_miniosz;
+            uint32_t d_maxiosz;
+        } da;
+        if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
+            bs->bl.request_alignment = da.d_miniosz;
+            /* The kernel returns wrong information for d_mem */
+            /* s->buf_align = da.d_mem; */
+        }
     }
-#endif
+#endif /* __linux__ */
 
     /*
      * If we could not get the sizes so far, we can only guess them. First try
-- 
2.38.1

On Tue, Nov 01, 2022 at 03:00:31PM -0400, Stefan Hajnoczi wrote:
>      /* Let's try to use the logical blocksize for the alignment. */
> -    if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
> -        bs->bl.request_alignment = 0;
> +    if (!bs->bl.request_alignment) {
> +        if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
> +            bs->bl.request_alignment = 0;
> +        }
>      }
>  
>  #ifdef __linux__
> -    /*
> -     * The XFS ioctl definitions are shipped in extra packages that might
> -     * not always be available. Since we just need the XFS_IOC_DIOINFO ioctl
> -     * here, we simply use our own definition instead:
> -     */
> -    struct xfs_dioattr {
> -        uint32_t d_mem;
> -        uint32_t d_miniosz;
> -        uint32_t d_maxiosz;
> -    } da;
> -    if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
> -        bs->bl.request_alignment = da.d_miniosz;
> -        /* The kernel returns wrong information for d_mem */
> -        /* s->buf_align = da.d_mem; */
> +    if (!bs->bl.request_alignment) {

This patch changes the fallback code to make the request_alignment value from
probe_logical_blocksize() override the value from XFS_IOC_DIOINFO.  Is that
intentional?

> +        if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
> +            bs->bl.request_alignment = da.d_miniosz;
> +            /* The kernel returns wrong information for d_mem */
> +            /* s->buf_align = da.d_mem; */

Has this bug been reported to the XFS developers (linux-xfs@vger.kernel.org)?

- Eric

On Tue, Nov 01, 2022 at 08:32:30PM -0700, Eric Biggers wrote:
> On Tue, Nov 01, 2022 at 03:00:31PM -0400, Stefan Hajnoczi wrote:
> >      /* Let's try to use the logical blocksize for the alignment. */
> > -    if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
> > -        bs->bl.request_alignment = 0;
> > +    if (!bs->bl.request_alignment) {
> > +        if (probe_logical_blocksize(fd, &bs->bl.request_alignment) < 0) {
> > +            bs->bl.request_alignment = 0;
> > +        }
> >      }
> >  
> >  #ifdef __linux__
> > -    /*
> > -     * The XFS ioctl definitions are shipped in extra packages that might
> > -     * not always be available. Since we just need the XFS_IOC_DIOINFO ioctl
> > -     * here, we simply use our own definition instead:
> > -     */
> > -    struct xfs_dioattr {
> > -        uint32_t d_mem;
> > -        uint32_t d_miniosz;
> > -        uint32_t d_maxiosz;
> > -    } da;
> > -    if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
> > -        bs->bl.request_alignment = da.d_miniosz;
> > -        /* The kernel returns wrong information for d_mem */
> > -        /* s->buf_align = da.d_mem; */
> > +    if (!bs->bl.request_alignment) {
> 
> This patch changes the fallback code to make the request_alignment value from
> probe_logical_blocksize() override the value from XFS_IOC_DIOINFO.  Is that
> intentional?

Thanks for pointing out the bug. That was not intentional. Will fix.

> > +        if (ioctl(fd, _IOR('X', 30, struct xfs_dioattr), &da) >= 0) {
> > +            bs->bl.request_alignment = da.d_miniosz;
> > +            /* The kernel returns wrong information for d_mem */
> > +            /* s->buf_align = da.d_mem; */
> 
> Has this bug been reported to the XFS developers (linux-xfs@vger.kernel.org)?

Paolo: Do you remember if you reported this when you wrote commit
c25f53b06eba ("raw: Probe required direct I/O alignment")?

Stefan