Linux 2.4 quota (accounting?) bug ...
Hi Honza!
a friend of mine, made me aware of the following imbalance, which looks like a minor accounting bug to me, but might be a quota bug ...
fs/dquot.c : 394 vfs_quota_sync()
-----------------------------------------------------
/* Get reference to quota so it won't be invalidated. get_dquot_ref()
* is enough since if dquot is locked/modified it can't be
* on the free list */
Quoted 1 line
get_dquot_ref(dquot);
if (dquot->dq_flags & DQ_LOCKED)
wait_on_dquot(dquot);
if (dquot_dirty(dquot))
sb->dq_op->sync_dquot(dquot);
Quoted 1 line
dqput(dquot);
goto restart;
}
best,
Herbert
Jan Kara
reviewerHi,
Quoted 3 lines
a friend of mine, made me aware of the following imbalance, which looks like a minor accounting bug to me, but might be a quota bug ...
Sorry but the code seems correct to me - we get reference to dquot byget_dquot_ref() and than we put the reference by dqput(). dqput() is correct because something nasty might happen in the mean time and so we might be the last holders of the dquot. What do you think is wrong?
Honza
Quoted 21 lines
fs/dquot.c: 394vfs_quota_sync()
-----------------------------------------------------
/* Get reference to quota so it won't be invalidated.get_dquot_ref()
* is enough since if dquot is locked/modified it can't be
* on the free list */Quoted 1 line
get_dquot_ref(dquot);(quoted diff) +0
-0if (dquot->dq_flags & DQ_LOCKED) wait_on_dquot(dquot); if (dquot_dirty(dquot)) sb->dq_op->sync_dquot(dquot);Quoted 1 line
dqput(dquot);(quoted diff) +0
-0goto restart; }best,
Herbert
Signature · 2 lines
Jan Kara <[email protected]> SuSE CR Labs
Quoted 9 lines · Jan Kara
Hi,
Quoted 3 lines
a friend of mine, made me aware of the following imbalance, which looks like a minor accounting bug to me, but might be a quota bug ...
Sorry but the code seems correct to me - we get reference to dquot by
get_dquot_ref()and than we put the reference bydqput().dqput()is correct because something nasty might happen in the mean time and so we might be the last holders of the dquot. What do you think is wrong?
dqput() does dqstats.drops++;
which isn't correct if this should be the same as put_dquot_ref(), but maybe I'm just irritated by strange statistics on some kernels showing more drops than lookups+allocated after sync/quotaoff
best,
Herbert
Quoted 31 lines
Honza
Quoted 21 lines
fs/dquot.c: 394vfs_quota_sync()
-----------------------------------------------------
/* Get reference to quota so it won't be invalidated.get_dquot_ref()
* is enough since if dquot is locked/modified it can't be
* on the free list */Quoted 1 line
get_dquot_ref(dquot);(quoted diff) +0
-0if (dquot->dq_flags & DQ_LOCKED) wait_on_dquot(dquot); if (dquot_dirty(dquot)) sb->dq_op->sync_dquot(dquot);Quoted 1 line
dqput(dquot);(quoted diff) +0
-0goto restart; }best,
HerbertSignature · 7 lines
Jan Kara <[email protected]> SuSE CR Labs - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Jan Kara
reviewerHi,
Quoted 11 lines · Jan Kara
Quoted 9 lines
Hi,
Quoted 3 lines
a friend of mine, made me aware of the following imbalance, which looks like a minor accounting bug to me, but might be a quota bug ...
Sorry but the code seems correct to me - we get reference to dquot by
get_dquot_ref()and than we put the reference bydqput().dqput()is correct because something nasty might happen in the mean time and so we might be the last holders of the dquot. What do you think is wrong?
Quoted 5 lines
dqput()does dqstats.drops++;
which isn't correct if this should be the same asput_dquot_ref(), but maybe I'm just irritated by strange statistics on some kernels showing more drops than lookups+allocated after sync/quotaoff
Oh, now I see... At first I didn't understand that the problem is in
the quota statistics. I'll fix that.
Honza
Jan Kara
reviewerHi,
Quoted 11 lines · Jan Kara
Quoted 9 lines
Hi,
Quoted 3 lines
a friend of mine, made me aware of the following imbalance, which looks like a minor accounting bug to me, but might be a quota bug ...
Sorry but the code seems correct to me - we get reference to dquot by
get_dquot_ref()and than we put the reference bydqput().dqput()is correct because something nasty might happen in the mean time and so we might be the last holders of the dquot. What do you think is wrong?
Quoted 5 lines
dqput()does dqstats.drops++;
which isn't correct if this should be the same asput_dquot_ref(), but maybe I'm just irritated by strange statistics on some kernels showing more drops than lookups+allocated after sync/quotaoff
So I finally got to fixing the bug. The patch (also with a deletition
of unnecessary variable) is attached. Marcello please apply.
Honza
---------------------cut here-------------------------------------------
fs/dquot.c +1 -2
@@ -392,6 +392,7 @@* is enough since if dquot is locked/modified it can't be * on the free list */ get_dquot_ref (dquot); + dqstats.lookups++;if (dquot->dq_flags &DQ_LOCKED )wait_on_dquot (dquot);if (dquot_dirty (dquot))
@@ -622,7 +623,6 @@dqput (dquot);return NODQUOT ; } - ++dquot->dq_referenced; dqstats.lookups++;return dquot;
@@ -642,7 +642,6 @@if (dquot->dq_flags &DQ_LOCKED )printk (KERN_ERR "VFS: dqduplicate(): Locked quota to be duplicated!\n" );get_dquot_dup_ref (dquot); - dquot->dq_referenced++; dqstats.lookups++;return dquot;
Only in linux-2.4.22-fixstat/include/linux: modules
include/linux/quota.h +0 -2
@@ -222,8 +222,6 @@loff_t dq_off;/* Offset of dquot on disk */ short dq_type;/* Type of quota */ short dq_flags;/* See DQ_* */ -unsigned long dq_referenced;/* Number of times this dquot was -referenced during its lifetime */ struct mem_dqblk dq_dqb;/* Diskquota usage */ };