render / rpms / qemu

Forked from rpms/qemu 9 months ago
Clone

Blame 0001-file-posix-Handle-undetectable-alignment.patch

9db63c
From: Nir Soffer <nirsof@gmail.com>
9db63c
Date: Tue, 13 Aug 2019 21:21:03 +0300
9db63c
Subject: [PATCH] file-posix: Handle undetectable alignment
9db63c
9db63c
In some cases buf_align or request_alignment cannot be detected:
9db63c
9db63c
1. With Gluster, buf_align cannot be detected since the actual I/O is
9db63c
   done on Gluster server, and qemu buffer alignment does not matter.
9db63c
   Since we don't have alignment requirement, buf_align=1 is the best
9db63c
   value.
9db63c
9db63c
2. With local XFS filesystem, buf_align cannot be detected if reading
9db63c
   from unallocated area. In this we must align the buffer, but we don't
9db63c
   know what is the correct size. Using the wrong alignment results in
9db63c
   I/O error.
9db63c
9db63c
3. With Gluster backed by XFS, request_alignment cannot be detected if
9db63c
   reading from unallocated area. In this case we need to use the
9db63c
   correct alignment, and failing to do so results in I/O errors.
9db63c
9db63c
4. With NFS, the server does not use direct I/O, so both buf_align cannot
9db63c
   be detected. In this case we don't need any alignment so we can use
9db63c
   buf_align=1 and request_alignment=1.
9db63c
9db63c
These cases seems to work when storage sector size is 512 bytes, because
9db63c
the current code starts checking align=512. If the check succeeds
9db63c
because alignment cannot be detected we use 512. But this does not work
9db63c
for storage with 4k sector size.
9db63c
9db63c
To determine if we can detect the alignment, we probe first with
9db63c
align=1. If probing succeeds, maybe there are no alignment requirement
9db63c
(cases 1, 4) or we are probing unallocated area (cases 2, 3). Since we
9db63c
don't have any way to tell, we treat this as undetectable alignment. If
9db63c
probing with align=1 fails with EINVAL, but probing with one of the
9db63c
expected alignments succeeds, we know that we found a working alignment.
9db63c
9db63c
Practically the alignment requirements are the same for buffer
9db63c
alignment, buffer length, and offset in file. So in case we cannot
9db63c
detect buf_align, we can use request alignment. If we cannot detect
9db63c
request alignment, we can fallback to a safe value. To use this logic,
9db63c
we probe first request alignment instead of buf_align.
9db63c
9db63c
Here is a table showing the behaviour with current code (the value in
9db63c
parenthesis is the optimal value).
9db63c
9db63c
Case    Sector    buf_align (opt)   request_alignment (opt)     result
9db63c
======================================================================
9db63c
1       512       512   (1)          512   (512)                 OK
9db63c
1       4096      512   (1)          4096  (4096)                FAIL
9db63c
----------------------------------------------------------------------
9db63c
2       512       512   (512)        512   (512)                 OK
9db63c
2       4096      512   (4096)       4096  (4096)                FAIL
9db63c
----------------------------------------------------------------------
9db63c
3       512       512   (1)          512   (512)                 OK
9db63c
3       4096      512   (1)          512   (4096)                FAIL
9db63c
----------------------------------------------------------------------
9db63c
4       512       512   (1)          512   (1)                   OK
9db63c
4       4096      512   (1)          512   (1)                   OK
9db63c
9db63c
Same cases with this change:
9db63c
9db63c
Case    Sector    buf_align (opt)   request_alignment (opt)     result
9db63c
======================================================================
9db63c
1       512       512   (1)          512   (512)                 OK
9db63c
1       4096      4096  (1)          4096  (4096)                OK
9db63c
----------------------------------------------------------------------
9db63c
2       512       512   (512)        512   (512)                 OK
9db63c
2       4096      4096  (4096)       4096  (4096)                OK
9db63c
----------------------------------------------------------------------
9db63c
3       512       4096  (1)          4096  (512)                 OK
9db63c
3       4096      4096  (1)          4096  (4096)                OK
9db63c
----------------------------------------------------------------------
9db63c
4       512       4096  (1)          4096  (1)                   OK
9db63c
4       4096      4096  (1)          4096  (1)                   OK
9db63c
9db63c
I tested that provisioning VMs and copying disks on local XFS and
9db63c
Gluster with 4k bytes sector size work now, resolving bugs [1],[2].
9db63c
I tested also on XFS, NFS, Gluster with 512 bytes sector size.
9db63c
9db63c
[1] https://bugzilla.redhat.com/1737256
9db63c
[2] https://bugzilla.redhat.com/1738657
9db63c
9db63c
Signed-off-by: Nir Soffer <nsoffer@redhat.com>
9db63c
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
9db63c
(cherry picked from commit a6b257a08e3d72219f03e461a52152672fec0612)
9db63c
---
9db63c
 block/file-posix.c | 36 +++++++++++++++++++++++++-----------
9db63c
 1 file changed, 25 insertions(+), 11 deletions(-)
9db63c
9db63c
diff --git a/block/file-posix.c b/block/file-posix.c
9db63c
index 4479cc7ab4..b8b4dad553 100644
9db63c
--- a/block/file-posix.c
9db63c
+++ b/block/file-posix.c
9db63c
@@ -323,6 +323,7 @@ static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
9db63c
     BDRVRawState *s = bs->opaque;
9db63c
     char *buf;
9db63c
     size_t max_align = MAX(MAX_BLOCKSIZE, getpagesize());
9db63c
+    size_t alignments[] = {1, 512, 1024, 2048, 4096};
9db63c
 
9db63c
     /* For SCSI generic devices the alignment is not really used.
9db63c
        With buffered I/O, we don't have any restrictions. */
9db63c
@@ -349,25 +350,38 @@ static void raw_probe_alignment(BlockDriverState *bs, int fd, Error **errp)
9db63c
     }
9db63c
 #endif
9db63c
 
9db63c
-    /* If we could not get the sizes so far, we can only guess them */
9db63c
-    if (!s->buf_align) {
9db63c
+    /*
9db63c
+     * If we could not get the sizes so far, we can only guess them. First try
9db63c
+     * to detect request alignment, since it is more likely to succeed. Then
9db63c
+     * try to detect buf_align, which cannot be detected in some cases (e.g.
9db63c
+     * Gluster). If buf_align cannot be detected, we fallback to the value of
9db63c
+     * request_alignment.
9db63c
+     */
9db63c
+
9db63c
+    if (!bs->bl.request_alignment) {
9db63c
+        int i;
9db63c
         size_t align;
9db63c
-        buf = qemu_memalign(max_align, 2 * max_align);
9db63c
-        for (align = 512; align <= max_align; align <<= 1) {
9db63c
-            if (raw_is_io_aligned(fd, buf + align, max_align)) {
9db63c
-                s->buf_align = align;
9db63c
+        buf = qemu_memalign(max_align, max_align);
9db63c
+        for (i = 0; i < ARRAY_SIZE(alignments); i++) {
9db63c
+            align = alignments[i];
9db63c
+            if (raw_is_io_aligned(fd, buf, align)) {
9db63c
+                /* Fallback to safe value. */
9db63c
+                bs->bl.request_alignment = (align != 1) ? align : max_align;
9db63c
                 break;
9db63c
             }
9db63c
         }
9db63c
         qemu_vfree(buf);
9db63c
     }
9db63c
 
9db63c
-    if (!bs->bl.request_alignment) {
9db63c
+    if (!s->buf_align) {
9db63c
+        int i;
9db63c
         size_t align;
9db63c
-        buf = qemu_memalign(s->buf_align, max_align);
9db63c
-        for (align = 512; align <= max_align; align <<= 1) {
9db63c
-            if (raw_is_io_aligned(fd, buf, align)) {
9db63c
-                bs->bl.request_alignment = align;
9db63c
+        buf = qemu_memalign(max_align, 2 * max_align);
9db63c
+        for (i = 0; i < ARRAY_SIZE(alignments); i++) {
9db63c
+            align = alignments[i];
9db63c
+            if (raw_is_io_aligned(fd, buf + align, max_align)) {
9db63c
+                /* Fallback to request_aligment. */
9db63c
+                s->buf_align = (align != 1) ? align : bs->bl.request_alignment;
9db63c
                 break;
9db63c
             }
9db63c
         }