https://jira.whamcloud.com This file is an XML representation of an issue en-us 9.4.14 940014 05-12-2023 https://jira.whamcloud.com/browse/LU-7875 Lustre <p>In mdt_reint_unlink we have suspect locking cleanup on error:</p> <div class="code panel" style="border-width: 1px;"><div class="codeContent panelContent"> <pre class="code-java"> rc = mdt_reint_object_lock(info, mc, child_lh, lock_ibits, cos_incompat); <span class="code-keyword">if</span> (rc != 0) GOTO(unlock_child, rc); ... unlock_child: mdt_unlock_slaves(info, mc, MDS_INODELOCK_UPDATE, s0_lh, s0_obj, einfo, rc); mdt_object_unlock(info, mc, child_lh, rc); unlock_parent: mdt_object_unlock(info, mp, parent_lh, rc); put_child: mdt_object_put(info-&gt;mti_env, mc); put_parent: mdt_object_put(info-&gt;mti_env, mp); </pre> </div></div> <p>So it looks like we failed a child lock and then jus jump to unlock it anyway?<br/> need to make that goto to unlock_parent?<br/> Also need to check if the below lock_slaves call to unlock slaves would be fine too. Not even sure why the unlock order is reversed.</p> LU-7875

Suspect locking cleanup in mdt_reint_unlink()

Bug Minor Resolved Not a Bug Lai Siyao Oleg Drokin Mon, 14 Mar 2016 16:13:59 +0000 Sun, 14 Jan 2018 03:11:09 +0000 Sun, 14 Jan 2018 03:11:09 +0000 Lustre 2.11.0 0 3 <p>Lai</p> <p>Can you please look into this?</p> <p>Peter</p> <p>This is not a bug IMO, because unlink locks in below order:<br/> 1. lock parent UPDATE bit.<br/> 2. if parent is remote, lock child LOOKUP bit on remote mdt, then lock child UPDATE bit on current mdt,<br/> else lock child LOOKUP | UPDATE bits on current mdt.<br/> 3. if child is striped dir, lock all its slave stripes.</p> <p>So we can see if parent is remote, child inodebits are locked on different MDTs, so if the second lock fails, it needs to unlock the first lock.</p> <p>And slave locks should be unlocked first, mdt_unlock_slaves() is safe to call at any time, because if no slave locks taken, nothing will be done.</p> <p>It's a bit confusing that in step 2, if parent is remote, mdt_remote_object_lock() is called on parent object, but actually it's used to find the parent MDT, the real locked resource is child.</p> Development Rank 1|hzy4db: Rank (Obsolete) 9223372036854775807 Severity