Hi Luis,

On 2017/9/6 8:52, Luis R. Rodriguez wrote:

On Mon, Sep 04, 2017 at 02:50:59PM +0800, Hou Tao wrote:

quoted

It will be useful if there is a corresponding online uevent after
a XFS filesystem has been mounted. A typical usage of the uevent
is setting the error configuration for a specific XFS filesystem
or all XFS filesystems by using udevd.

The following is an example of udevd rule which will shutdown
any XFS filesystem (except the one with the matched UUID) after
the filesystem gets any IO error and the filesystem with the matched
UUID will retry 5 times before its shutdown:

    ACTION=="online", SUBSYSTEM=="xfs", \
    ENV{ID_FS_UUID}=="6c1eebfd-d1af-4b69-a0f1-c9b4663df44d", \
    RUN+="/bin/sh -c 'echo 5 > /sys%p/error/metadata/EIO/max_retries'", \
    GOTO="end"

    ACTION=="online", SUBSYSTEM=="xfs", DEVPATH=="/fs/xfs/*", \
    RUN+="/bin/sh -c 'echo 0 > /sys%p/error/metadata/default/max_retries; \
	echo 0 > /sys%p/error/metadata/EIO/max_retries; \
	echo 0 > /sys%p/error/metadata/ENOSPC/max_retries; \
	echo 0 > /sys%p/error/metadata/ENODEV/max_retries'"

    LABEL="end"

Suggested-by: Dave Chinner <david@fromorbit.com>
Signed-off-by: Hou Tao <redacted>
---
v3:
    * code style fixes
    * use "ID_FS_UUID" instead of "UUID" as the name of uuid environment
v2:
    * add UUID property for mount uevent
    * add an udev example for UUID filtering
v1:
    * http://www.spinics.net/lists/linux-xfs/msg09484.html
---
 fs/xfs/xfs_super.c | 24 ++++++++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c
index 3a3812b4..1f0d895 100644
--- a/fs/xfs/xfs_super.c
+++ b/fs/xfs/xfs_super.c

@@ -1530,6 +1530,28 @@ xfs_destroy_percpu_counters(
 	percpu_counter_destroy(&mp->m_fdblocks);
 }
 
+static void
+xfs_fs_uevent(
+	struct xfs_mount	*mp,
+	enum kobject_action	action)
+{
+#define XFS_UEVENT_MAX_ENV_COUNT 1
+	/* "+ 1" for the trailing NULL pointer */
+	char			*envp[XFS_UEVENT_MAX_ENV_COUNT + 1];
+	const char		*prefix = "ID_FS_UUID=";
+	char			buf[strlen(prefix) + UUID_STRING_LEN + 1];
+	int			i = 0;
+	int			err;
+
+	snprintf(buf, sizeof(buf), "%s%pUb", prefix, &mp->m_super->s_uuid);
+	envp[i++] = buf;
+	envp[i] = NULL;
+	err = kobject_uevent_env(&mp->m_kobj.kobject, action, envp);
+	if (err)
+		xfs_notice(mp, "Sending XFS uevent %d got error %d",

kobject_uevent_env() can fail for a few reasons, most commonly it can fail for
when we're out of memory. I've seen quite a bit of use cases these days where
tons of remounts can happen, one example is actually is when there is not
enough space dockers instances can get restarted. There are many reasons for
restarts of docker instance, but as stupid as it is, since -ENOMEM could
actually be common, I think we should consider treating it as fatal and not
mount. Otherwise the assumption that userspace will configure the filesystem
correctly may be false.

I understand and agree your opinion on error handler, but i don't follow the
example about docker instances. Do you mean the docker instances will be restarted
and the filesystem will be unmounted and mounted again when there is not enough
memory for the cgroup where the docker instance residents in ? If there is not
enough memory, the mount may abort before the uevent sending.

Note that kobject_uevent_env() can also fail during
netlink_broadcast_filtered(),  so perhaps we should consider all errors well
here.

Yes, to deliver the uevent reliably we need to handle the error returned by
kobject_uevent_evn(), and abort the filesystem mount if any error occurs.

Tao

  Luis

.