[PATCH v2] super1.c: avoid useless sync when bitmap switches from clustered to none

[PATCH v2] super1.c: avoid useless sync when bitmap switches from clustered to none

[Zhao Heming]

With kernel commit 480523feae58 ("md: only call set_in_sync() when it
is expected to succeed."), mddev->in_sync in clustered array is always
zero. It makes metadata resync_offset to always zero.
When assembling a clusterd array with "-U no-bitmap" option, kernel
md layer "mddev->resync_offset == 0" and "mddev->bitmap == NULL" will
trigger raid1 do sync on every bitmap chunk. the sync action is useless,
we should avoid it.

Related kernel flow:
```
md_do_sync
 mddev->pers->sync_request
  raid1_sync_request
   md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1)
    __bitmap_start_sync(bitmap, offset,&blocks1, degraded)
      if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
        *blocks = 1024;
        return 1; /* always resync if no bitmap */
      }
```

Reprodusible steps:
```
node1 # mdadm -C /dev/md0 -b clustered -e 1.2 -n 2 -l mirror /dev/sd{a,b}
node1 # mdadm -Ss
(in another shell, executing & watching: watch -n 1 'cat /proc/mdstat')
node1 # mdadm -A -U no-bitmap /dev/md0 /dev/sd{a,b}
```

Signed-off-by: Zhao Heming <heming.zhao@suse.com>
---
v2: only set MaxSector on bitmap clean device
 
---
 super1.c | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/super1.c b/super1.c
index 19fe6f5..9470025 100644
--- a/super1.c
+++ b/super1.c
@@ -1318,6 +1318,8 @@ static int update_super1(struct supertype *st, struct mdinfo *info,
 			memcpy(bms->uuid, sb->set_uuid, 16);
 	} else if (strcmp(update, "no-bitmap") == 0) {
 		sb->feature_map &= ~__cpu_to_le32(MD_FEATURE_BITMAP_OFFSET);
+		if (bms->version == BITMAP_MAJOR_CLUSTERED && !IsBitmapDirty(devname))
+			sb->resync_offset = MaxSector;
 	} else if (strcmp(update, "bbl") == 0) {
 		/* only possible if there is room after the bitmap, or if
 		 * there is no bitmap
-- 
2.30.0

resend: [PATCH v2 mdadm] super1.c: avoid useless sync when bitmap switches from clustered to none

[heming.zhao]

With kernel commit 480523feae58 ("md: only call set_in_sync() when it
is expected to succeed."), mddev->in_sync in clustered array is always
zero. It makes metadata resync_offset to always zero.
When assembling a clusterd array with "-U no-bitmap" option, kernel
md layer "mddev->resync_offset == 0" and "mddev->bitmap == NULL" will
trigger raid1 do sync on every bitmap chunk. the sync action is useless,
it wastes user a lot of time, we should avoid it.

Related kernel flow:
```
md_do_sync
  mddev->pers->sync_request
   raid1_sync_request
    md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1)
     __bitmap_start_sync(bitmap, offset,&blocks1, degraded)
       if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
         *blocks = 1024;
         return 1; /* always resync if no bitmap */
       }
```

Reprodusible steps:
```
node1 # mdadm -C /dev/md0 -b clustered -e 1.2 -n 2 -l mirror /dev/sd{a,b}
node1 # mdadm -Ss
(in another shell, executing & watching: watch -n 1 'cat /proc/mdstat')
node1 # mdadm -A -U no-bitmap /dev/md0 /dev/sd{a,b}
```

Signed-off-by: Zhao Heming <heming.zhao@suse.com>
---
v2: only set MaxSector on bitmap clean device
  
---
  super1.c | 2 ++
  1 file changed, 2 insertions(+)

diff --git a/super1.c b/super1.c
index 19fe6f5..9470025 100644
--- a/super1.c
+++ b/super1.c
@@ -1318,6 +1318,8 @@ static int update_super1(struct supertype *st, struct mdinfo *info,
  			memcpy(bms->uuid, sb->set_uuid, 16);
  	} else if (strcmp(update, "no-bitmap") == 0) {
  		sb->feature_map &= ~__cpu_to_le32(MD_FEATURE_BITMAP_OFFSET);
+		if (bms->version == BITMAP_MAJOR_CLUSTERED && !IsBitmapDirty(devname))
+			sb->resync_offset = MaxSector;
  	} else if (strcmp(update, "bbl") == 0) {
  		/* only possible if there is room after the bitmap, or if
  		 * there is no bitmap
-- 
2.30.0

Re: resend: [PATCH v2 mdadm] super1.c: avoid useless sync when bitmap switches from clustered to none

[heming.zhao]

On 3/2/21 4:35 PM, heming.zhao@suse.com wrote:
> With kernel commit 480523feae58 ("md: only call set_in_sync() when it
> is expected to succeed."), mddev->in_sync in clustered array is always
> zero. It makes metadata resync_offset to always zero.
> When assembling a clusterd array with "-U no-bitmap" option, kernel
> md layer "mddev->resync_offset == 0" and "mddev->bitmap == NULL" will
> trigger raid1 do sync on every bitmap chunk. the sync action is useless,
> ... ...
> ---
> v2: only set MaxSector on bitmap clean device
> 
> ---
>   super1.c | 2 ++
>   1 file changed, 2 insertions(+)
> 
> diff --git a/super1.c b/super1.c


Hello all,

more info:
This patch had been verified & run in SUSE customer env for some time.

Thanks,
Heming

Re: [PATCH v2] super1.c: avoid useless sync when bitmap switches from clustered to none

[Jes Sorensen]

On 2/2/21 7:22 PM, Zhao Heming wrote:
> With kernel commit 480523feae58 ("md: only call set_in_sync() when it
> is expected to succeed."), mddev->in_sync in clustered array is always
> zero. It makes metadata resync_offset to always zero.
> When assembling a clusterd array with "-U no-bitmap" option, kernel
> md layer "mddev->resync_offset == 0" and "mddev->bitmap == NULL" will
> trigger raid1 do sync on every bitmap chunk. the sync action is useless,
> we should avoid it.
> 
> Related kernel flow:
> ```
> md_do_sync
>  mddev->pers->sync_request
>   raid1_sync_request
>    md_bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1)
>     __bitmap_start_sync(bitmap, offset,&blocks1, degraded)
>       if (bitmap == NULL) {/* FIXME or bitmap set as 'failed' */
>         *blocks = 1024;
>         return 1; /* always resync if no bitmap */
>       }
> ```
> 
> Reprodusible steps:
> ```
> node1 # mdadm -C /dev/md0 -b clustered -e 1.2 -n 2 -l mirror /dev/sd{a,b}
> node1 # mdadm -Ss
> (in another shell, executing & watching: watch -n 1 'cat /proc/mdstat')
> node1 # mdadm -A -U no-bitmap /dev/md0 /dev/sd{a,b}
> ```
> 
> Signed-off-by: Zhao Heming <heming.zhao@suse.com>
> ---
> v2: only set MaxSector on bitmap clean device

Applied!

Thanks,
Jes