Re: [PATCH] ocfs2: ocfs2_downconvert_lock failure results in deadlock

From: Gang He
Date: Fri Aug 27 2021 - 01:51:04 EST




On 2021/8/26 17:55, Joseph Qi wrote:


On 8/26/21 4:45 PM, Gang He wrote:
Hi Joseph,

On 2021/8/26 16:23, Joseph Qi wrote:


On 8/26/21 2:10 PM, Gang He wrote:
Usually, ocfs2_downconvert_lock() function always downconverts
dlm lock to the expected level for satisfy dlm bast requests

s/for/to

from the other nodes.
But there is a rare situation. When dlm lock conversion is being
canceled, ocfs2_downconvert_lock() function will return -EBUSY.
You need to be aware that ocfs2_cancel_convert() function is
asynchronous in fsdlm implementation.
If we does not requeue this lockres entry, ocfs2 downconvert
thread no longer handles this dlm lock bast request. Then, the
other nodes will not get the dlm lock again, the current node's
process will be blocked when acquire this dlm lock again.

Signed-off-by: Gang He <ghe@xxxxxxxx>
---
  fs/ocfs2/dlmglue.c | 13 +++++++++----
  1 file changed, 9 insertions(+), 4 deletions(-)

diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index 48fd369c29a4..c454c218fbfe 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -3671,13 +3671,11 @@ static int ocfs2_downconvert_lock(struct ocfs2_super *osb,
                   OCFS2_LOCK_ID_MAX_LEN - 1);
      lockres_clear_pending(lockres, generation, osb);
      if (ret) {
-        ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);
+        if (ret != -EBUSY)
+            ocfs2_log_dlm_error("ocfs2_dlm_lock", ret, lockres);

Do we have to treat EBUSY as a normal case?
Yes, this return code is expected when call dlm_lock() to convert a dlm lock to another level, but this dlm lock is being cancelled.
As I said in another mail, for fsdlm implementation,ocfs2_cancel_convert
will return immediately, but the related dlm lock will(is) be cancelled in background. For o2dlm implementation,ocfs2_cancel_convert will return after the dlm lock is cancelled and it's ast is invoked, that is why o2cb based ocfs2 does not encounter -EBUSY return code in my test script, of course, this kind of implementation will block other lockres entries to be handled for a little time in down-convert thread.

Better to leave this log for later issue tracking.
I'm worrying about if there are other cases here.
OK




          ocfs2_recover_from_dlm_error(lockres, 1);
-        goto bail;
      }
  -    ret = 0;
-bail:
      return ret;
  }
  @@ -3912,6 +3910,13 @@ static int ocfs2_unblock_lock(struct ocfs2_super *osb,
      spin_unlock_irqrestore(&lockres->l_lock, flags);
      ret = ocfs2_downconvert_lock(osb, lockres, new_level, set_lvb,
                       gen);
+    /* ocfs2_cancel_convert() is in progress, try again later */
+    if (ret == -EBUSY) {
+        ctl->requeue = 1;
+        mlog(ML_BASTS, "lockres %s, ReQ: Downconvert busy\n",
+             lockres->l_name);
+        ret = 0;

Ditto. If EBUSY is not a normal case, I'd like just requeue it but not
change it to normal return code.
You know ML_BASTS is always switched off in production environment.
Since this case should be considered as a normal case, although it's rare.
We should not print any error message to kernel journal, but if the user
turn on the BASTS trace, he should watch this trace for debugging.


Okay, since we leave an error message above, we can return normal to
caller. And now caller only print a simple error which doesn't have
much meaning.

BTW, could we change it like:

ret = ocfs2_downconvert_lock();
if (ret == -EBUSY) {
mlog(ML_BASTS, ...);
/* Describe the case why we have to requeue */
goto requeue;
}

According to the current code, it is easy to
goto leave label with set ctl->requeue = 1 for requeue;


Thanks
Gang


...
requeue:
ctl->requeue = 1;
return 0;