[PATCH v2] btrfs-progs: zoned: open host-managed devices with O_DIRECT for read-only

[PATCH v2] btrfs-progs: zoned: open host-managed devices with O_DIRECT for read-only

[Johannes Thumshirn]

On host-managed zoned block devices the page cache cannot be used for
writing to the device because it does not guarantee ordering, therefore
direct I/O is used to write any on-disk data structures.

But reading in btrfs-progs is still performed using buffered I/O. This
mix of direct and buffered I/O can lead to stale page cache entries
being handed off to btrfs-progs and tools like "btrfs check" can report
spurious errors like:

	checksum verify failed on <bytenr> wanted 0x00000000 found 0xb6bde3e4

even though both copies are intact on disk.

Drop the O_RDWR condition so host-managed zoned devices always use
O_DIRECT, matching btrfstune.

Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>
---
Changes to v1:
- reworded commit message

 kernel-shared/disk-io.c | 2 +-
 kernel-shared/volumes.c | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel-shared/disk-io.c b/kernel-shared/disk-io.c
index 7f87e567f0b1..8b9c6dcac967 100644
--- a/kernel-shared/disk-io.c
+++ b/kernel-shared/disk-io.c
@@ -1722,7 +1722,7 @@ struct btrfs_fs_info *open_ctree_fs_info(struct open_ctree_args *oca)
 	if (!(oca->flags & OPEN_CTREE_WRITES))
 		oflags = O_RDONLY;

-	if ((oflags & O_RDWR) && zoned_model(oca->filename) == ZONED_HOST_MANAGED)
+	if (zoned_model(oca->filename) == ZONED_HOST_MANAGED)
 		oflags |= O_DIRECT;

 	fp = open(oca->filename, oflags);
diff --git a/kernel-shared/volumes.c b/kernel-shared/volumes.c
index f35b59bcc20e..e2b6708ad872 100644
--- a/kernel-shared/volumes.c
+++ b/kernel-shared/volumes.c
@@ -697,7 +697,7 @@ int btrfs_open_devices(struct btrfs_fs_info *fs_info,
 			continue;
 		}

-		if ((flags & O_RDWR) && zoned_model(device->name) == ZONED_HOST_MANAGED)
+		if (zoned_model(device->name) == ZONED_HOST_MANAGED)
 			flags |= O_DIRECT;

 		fd = open(device->name, flags);
--
2.54.0

Re: [PATCH v2] btrfs-progs: zoned: open host-managed devices with O_DIRECT for read-only

[Qu Wenruo]

在 2026/7/16 20:58, Johannes Thumshirn 写道:
> On host-managed zoned block devices the page cache cannot be used for
> writing to the device because it does not guarantee ordering, therefore
> direct I/O is used to write any on-disk data structures.
> 
> But reading in btrfs-progs is still performed using buffered I/O. This
> mix of direct and buffered I/O can lead to stale page cache entries
> being handed off to btrfs-progs and tools like "btrfs check" can report
> spurious errors like:
> 
> 	checksum verify failed on <bytenr> wanted 0x00000000 found 0xb6bde3e4
> 
> even though both copies are intact on disk.
> 
> Drop the O_RDWR condition so host-managed zoned devices always use
> O_DIRECT, matching btrfstune.
> 
> Signed-off-by: Johannes Thumshirn <johannes.thumshirn@wdc.com>

Reviewed-by: Qu Wenruo <wqu@suse.com>

Thanks,
Qu

> ---
> Changes to v1:
> - reworded commit message
> 
>   kernel-shared/disk-io.c | 2 +-
>   kernel-shared/volumes.c | 2 +-
>   2 files changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel-shared/disk-io.c b/kernel-shared/disk-io.c
> index 7f87e567f0b1..8b9c6dcac967 100644
> --- a/kernel-shared/disk-io.c
> +++ b/kernel-shared/disk-io.c
> @@ -1722,7 +1722,7 @@ struct btrfs_fs_info *open_ctree_fs_info(struct open_ctree_args *oca)
>   	if (!(oca->flags & OPEN_CTREE_WRITES))
>   		oflags = O_RDONLY;
> 
> -	if ((oflags & O_RDWR) && zoned_model(oca->filename) == ZONED_HOST_MANAGED)
> +	if (zoned_model(oca->filename) == ZONED_HOST_MANAGED)
>   		oflags |= O_DIRECT;
> 
>   	fp = open(oca->filename, oflags);
> diff --git a/kernel-shared/volumes.c b/kernel-shared/volumes.c
> index f35b59bcc20e..e2b6708ad872 100644
> --- a/kernel-shared/volumes.c
> +++ b/kernel-shared/volumes.c
> @@ -697,7 +697,7 @@ int btrfs_open_devices(struct btrfs_fs_info *fs_info,
>   			continue;
>   		}
> 
> -		if ((flags & O_RDWR) && zoned_model(device->name) == ZONED_HOST_MANAGED)
> +		if (zoned_model(device->name) == ZONED_HOST_MANAGED)
>   			flags |= O_DIRECT;
> 
>   		fd = open(device->name, flags);
> --
> 2.54.0
> 
>