]> git.proxmox.com Git - pve-qemu.git/commitdiff
backup: improve error when copy-before-write fails for fleecing
authorFiona Ebner <f.ebner@proxmox.com>
Mon, 29 Apr 2024 15:20:22 +0000 (17:20 +0200)
committerThomas Lamprecht <t.lamprecht@proxmox.com>
Mon, 29 Apr 2024 15:25:20 +0000 (17:25 +0200)
With fleecing, failure for copy-before-write does not fail the guest
write, but only sets the snapshot error that is associated to the
copy-before-write filter, making further requests to the snapshot
access fail with EACCES, which then also fails the job. But that error
code is not the root cause of why the backup failed, so bubble up the
original snapshot error instead.

Reported-by: Friedrich Weber <f.weber@proxmox.com>
Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
debian/patches/pve/0051-PVE-backup-improve-error-when-copy-before-write-fail.patch [new file with mode: 0644]
debian/patches/series

diff --git a/debian/patches/pve/0051-PVE-backup-improve-error-when-copy-before-write-fail.patch b/debian/patches/pve/0051-PVE-backup-improve-error-when-copy-before-write-fail.patch
new file mode 100644 (file)
index 0000000..c7f2ccb
--- /dev/null
@@ -0,0 +1,117 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Fiona Ebner <f.ebner@proxmox.com>
+Date: Mon, 29 Apr 2024 14:43:58 +0200
+Subject: [PATCH] PVE backup: improve error when copy-before-write fails for
+ fleecing
+
+With fleecing, failure for copy-before-write does not fail the guest
+write, but only sets the snapshot error that is associated to the
+copy-before-write filter, making further requests to the snapshot
+access fail with EACCES, which then also fails the job. But that error
+code is not the root cause of why the backup failed, so bubble up the
+original snapshot error instead.
+
+Reported-by: Friedrich Weber <f.weber@proxmox.com>
+Signed-off-by: Fiona Ebner <f.ebner@proxmox.com>
+Tested-by: Friedrich Weber <f.weber@proxmox.com>
+---
+ block/copy-before-write.c | 18 ++++++++++++------
+ block/copy-before-write.h |  1 +
+ pve-backup.c              |  9 +++++++++
+ 3 files changed, 22 insertions(+), 6 deletions(-)
+
+diff --git a/block/copy-before-write.c b/block/copy-before-write.c
+index bba58326d7..50cc4c7aae 100644
+--- a/block/copy-before-write.c
++++ b/block/copy-before-write.c
+@@ -27,6 +27,7 @@
+ #include "qapi/qmp/qjson.h"
+ #include "sysemu/block-backend.h"
++#include "qemu/atomic.h"
+ #include "qemu/cutils.h"
+ #include "qapi/error.h"
+ #include "block/block_int.h"
+@@ -74,7 +75,8 @@ typedef struct BDRVCopyBeforeWriteState {
+      * @snapshot_error is normally zero. But on first copy-before-write failure
+      * when @on_cbw_error == ON_CBW_ERROR_BREAK_SNAPSHOT, @snapshot_error takes
+      * value of this error (<0). After that all in-flight and further
+-     * snapshot-API requests will fail with that error.
++     * snapshot-API requests will fail with that error. To be accessed with
++     * atomics.
+      */
+     int snapshot_error;
+ } BDRVCopyBeforeWriteState;
+@@ -114,7 +116,7 @@ static coroutine_fn int cbw_do_copy_before_write(BlockDriverState *bs,
+         return 0;
+     }
+-    if (s->snapshot_error) {
++    if (qatomic_read(&s->snapshot_error)) {
+         return 0;
+     }
+@@ -138,9 +140,7 @@ static coroutine_fn int cbw_do_copy_before_write(BlockDriverState *bs,
+     WITH_QEMU_LOCK_GUARD(&s->lock) {
+         if (ret < 0) {
+             assert(s->on_cbw_error == ON_CBW_ERROR_BREAK_SNAPSHOT);
+-            if (!s->snapshot_error) {
+-                s->snapshot_error = ret;
+-            }
++            qatomic_cmpxchg(&s->snapshot_error, 0, ret);
+         } else {
+             bdrv_set_dirty_bitmap(s->done_bitmap, off, end - off);
+         }
+@@ -214,7 +214,7 @@ cbw_snapshot_read_lock(BlockDriverState *bs, int64_t offset, int64_t bytes,
+     QEMU_LOCK_GUARD(&s->lock);
+-    if (s->snapshot_error) {
++    if (qatomic_read(&s->snapshot_error)) {
+         g_free(req);
+         return NULL;
+     }
+@@ -585,6 +585,12 @@ void bdrv_cbw_drop(BlockDriverState *bs)
+     bdrv_unref(bs);
+ }
++int bdrv_cbw_snapshot_error(BlockDriverState *bs)
++{
++    BDRVCopyBeforeWriteState *s = bs->opaque;
++    return qatomic_read(&s->snapshot_error);
++}
++
+ static void cbw_init(void)
+ {
+     bdrv_register(&bdrv_cbw_filter);
+diff --git a/block/copy-before-write.h b/block/copy-before-write.h
+index dc6cafe7fa..a27d2d7d9f 100644
+--- a/block/copy-before-write.h
++++ b/block/copy-before-write.h
+@@ -44,5 +44,6 @@ BlockDriverState *bdrv_cbw_append(BlockDriverState *source,
+                                   BlockCopyState **bcs,
+                                   Error **errp);
+ void bdrv_cbw_drop(BlockDriverState *bs);
++int bdrv_cbw_snapshot_error(BlockDriverState *bs);
+ #endif /* COPY_BEFORE_WRITE_H */
+diff --git a/pve-backup.c b/pve-backup.c
+index 7cc1dd3724..07709aa350 100644
+--- a/pve-backup.c
++++ b/pve-backup.c
+@@ -379,6 +379,15 @@ static void pvebackup_complete_cb(void *opaque, int ret)
+         di->fleecing.snapshot_access = NULL;
+     }
+     if (di->fleecing.cbw) {
++        /*
++         * With fleecing, failure for cbw does not fail the guest write, but only sets the snapshot
++         * error, making further requests to the snapshot fail with EACCES, which then also fail the
++         * job. But that code is not the root cause and just confusing, so update it.
++         */
++        int snapshot_error = bdrv_cbw_snapshot_error(di->fleecing.cbw);
++        if (di->completed_ret == -EACCES && snapshot_error) {
++            di->completed_ret = snapshot_error;
++        }
+         bdrv_cbw_drop(di->fleecing.cbw);
+         di->fleecing.cbw = NULL;
+     }
index 47e83070bf875d4150536d6a751644da86e13631..b97881e0e695b43a39ea878422091795d43b2775 100644 (file)
@@ -59,3 +59,4 @@ pve/0047-qapi-blockdev-backup-add-discard-source-parameter.patch
 pve/0048-copy-before-write-allow-specifying-minimum-cluster-s.patch
 pve/0049-backup-add-minimum-cluster-size-to-performance-optio.patch
 pve/0050-PVE-backup-add-fleecing-option.patch
+pve/0051-PVE-backup-improve-error-when-copy-before-write-fail.patch