Hello list,
When investigation livelock in RHEL3, i found possible badness in
prune_dcache, which exist in 2.6.24 also.
situation - system has ~6 mounted filesystem, at 5 FS do some io, and
one FS start umount.
shrink_dcache_parent - collect unused dentries after call
select_parent() and put these dentries into end of LRU for kill.
but between exit from select_parent() and enter to prune_dcache(,sb)
some processed add more unused dentries and put to end lru also.
prune_dcache start skip some dentiries in loop
>>>>
while (skip && tmp != &dentry_unused &&
list_entry(tmp, struct dentry,
d_lru)->d_sb != sb) {
skip--;
tmp = tmp->prev;
}
>>>
but not found correct dentry for superblock.
later condition
if (tmp == &dentry_unused)
break;
not hit - because LRU has additional dentry, and prune_dcache(,sb) kill
dentry not related to submitted superblock.
(this first stranges - for me)
but this not all, because count != 0 prune_dcache run in loop and kill
all dentries not related to submited sb, with can be need many time (in
my situation ~15min for ~200k dentries).
after exit from prune_dcache - shrink_dcache_parent() do loop and try
again destroy dentries - which also need some time.
my tests - show changing condition, from
if (tmp == &dentry_unused)
break;
to
if (tmp == &dentry_unused) || (sb && (list_entry(tmp,
struct dentry, d_lru)->d_sb != sb))
break;
help with live lock - because prune_dcache exit from loop early and move
required dentries to end of LRU - for easy kill.
Please comment this investigation.
PS. please send copy to me, i not subscribed to list.
--
Alex Lyashkov <Alexey.lyashkov@xxxxxxx>
Lustre Group, Sun Microsystems
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
