On 2019/12/9 20:46, Jeff Layton wrote:
On Mon, 2019-12-09 at 19:54 +0800, Xiubo Li wrote:
On 2019/12/9 19:38, Jeff Layton wrote:
On Mon, 2019-12-09 at 04:28 -0500, xiubli@xxxxxxxxxx wrote:
From: Xiubo Li <xiubli@xxxxxxxxxx>
Try to queue writeback and invalidate the dirty pages when sessions
are closed, rejected or reconnect denied.
Signed-off-by: Xiubo Li <xiubli@xxxxxxxxxx>
---
fs/ceph/mds_client.c | 13 +++++++++++++
1 file changed, 13 insertions(+)
Can you explain a bit more about the problem you're fixing? In what
situation is this currently broken, and what are the effects of that
breakage?
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index be1ac9f8e0e6..68f3b5ed6ac8 100644
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -1385,9 +1385,11 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
{
struct ceph_fs_client *fsc = (struct ceph_fs_client *)arg;
struct ceph_inode_info *ci = ceph_inode(inode);
+ struct ceph_mds_session *session = cap->session;
LIST_HEAD(to_remove);
bool dirty_dropped = false;
bool invalidate = false;
+ bool writeback = false;
dout("removing cap %p, ci is %p, inode is %p\n",
cap, ci, &ci->vfs_inode);
@@ -1398,12 +1400,21 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
if (!ci->i_auth_cap) {
struct ceph_cap_flush *cf;
struct ceph_mds_client *mdsc = fsc->mdsc;
+ int s_state = session->s_state;
if (READ_ONCE(fsc->mount_state) == CEPH_MOUNT_SHUTDOWN) {
if (inode->i_data.nrpages > 0)
invalidate = true;
if (ci->i_wrbuffer_ref > 0)
mapping_set_error(&inode->i_data, -EIO);
+ } else if (s_state == CEPH_MDS_SESSION_CLOSED ||
+ s_state == CEPH_MDS_SESSION_REJECTED) {
+ /* reconnect denied or rejected */
+ if (!__ceph_is_any_real_caps(ci) &&
+ inode->i_data.nrpages > 0)
+ invalidate = true;
+ if (ci->i_wrbuffer_ref > 0)
+ writeback = true;
I don't know here. If the session is CLOSED/REJECTED, is kicking off
writeback the right thing to do? In principle, this means that the
client may have been blacklisted and none of the writes will succeed.
If the client was blacklisted, it will be not safe to still buffer the
data and flush it after the related sessions are reconnected without
remounting.
Maybe we need to throw it directly.
It depends. We have tunable behavior for this now.
I think you may want to look over the recover_session= patches that
were merged v5.4 ((131d7eb4faa1fc, and previous patches). That mount
option governs what happens if the client is blacklisted, and then the
blacklisting is later removed.
Yeah, I checked that patches before and from my test of the blacklist
stuff, the kclient's sessions won't recovery automatically even after
the kclient is removed from the blacklisting. And the kclient will force
remount after 30 minutes if recover_session=clean.
After a session is rejected due to blacklisting, the handle_session(case
CEPH_MDS_SESSION_REJECTED: ) will requeue some inflight requests of this
session, but these requests or the new requests couldn't use the
rejected session, because the session->blacklisted=true could only be
cleaned in the recovery_session=clean code until 30 minutes timedout, as
default with recovery_session=no the session maybe keep
session->blacklisted for ever (?), then I must do the remount manually.
I may not totally understand the purpose of the recovery_session=
patches and I may miss something important here.
From the https://docs.ceph.com/docs/master/cephfs/eviction/, maybe we
could just throw away the dirty pages of the sessions directly, so after
reconnected we can make sure it is safe without any stale buffered data.
Thanks.
Will this patch change either recover_session= behavior? If so, then you
should explain how and why.
}
while (!list_empty(&ci->i_cap_flush_list)) {
@@ -1472,6 +1483,8 @@ static int remove_session_caps_cb(struct inode *inode, struct ceph_cap *cap,
}
wake_up_all(&ci->i_cap_wq);
+ if (writeback)
+ ceph_queue_writeback(inode);
if (invalidate)
ceph_queue_invalidate(inode);
if (dirty_dropped)
Thanks,
