2004-09-19 07:51:06

by Leandro Santi

Subject: [PATCH 2.4] fix dcache nr_dentry race

Hi Marcelo,

The dentry_stat.nr_dentry counter isn't being properly protected against
concurrent access. We've been observing a drift of about 8000 units per
day on some large MP Maildir++ mailstore nodes.

The following (trivial) patch is pretty much a backport from 2.6.

diff -ru linux-2.4.27.orig/fs/dcache.c linux-2.4.27.patched/fs/dcache.c
--- linux-2.4.27.orig/fs/dcache.c Fri Jun 13 11:51:37 2003
+++ linux-2.4.27.patched/fs/dcache.c Thu Sep 9 17:37:56 2004
@@ -55,7 +55,10 @@
/* Statistics gathering. */
struct dentry_stat_t dentry_stat = {0, 0, 45, 0,};

-/* no dcache_lock, please */
+/*
+ * no dcache_lock, please. The caller must decrement dentry_stat.nr_dentry
+ * inside dcache_lock.
+ */
static inline void d_free(struct dentry *dentry)
{
if (dentry->d_op && dentry->d_op->d_release)
@@ -63,7 +66,6 @@
if (dname_external(dentry))
kfree(dentry->d_name.name);
kmem_cache_free(dentry_cache, dentry);
- dentry_stat.nr_dentry--;
}

/*
@@ -148,6 +150,7 @@
kill_it: {
struct dentry *parent;
list_del(&dentry->d_child);
+ dentry_stat.nr_dentry--; /* For d_free, below */
/* drops the lock, at that point nobody can reach this dentry */
dentry_iput(dentry);
parent = dentry->d_parent;
@@ -297,6 +300,7 @@

list_del_init(&dentry->d_hash);
list_del(&dentry->d_child);
+ dentry_stat.nr_dentry--; /* For d_free, below */
dentry_iput(dentry);
parent = dentry->d_parent;
d_free(dentry);
@@ -623,13 +627,15 @@
if (parent) {
dentry->d_parent = dget(parent);
dentry->d_sb = parent->d_sb;
- spin_lock(&dcache_lock);
- list_add(&dentry->d_child, &parent->d_subdirs);
- spin_unlock(&dcache_lock);
} else
INIT_LIST_HEAD(&dentry->d_child);

+ spin_lock(&dcache_lock);
+ if (parent)
+ list_add(&dentry->d_child, &parent->d_subdirs);
dentry_stat.nr_dentry++;
+ spin_unlock(&dcache_lock);
+
return dentry;
}

Bye,
Leandro.

2004-09-19 13:07:56

by Marc-Christian Petersen

Subject: Re: [PATCH 2.4] fix dcache nr_dentry race

On Sunday 19 September 2004 09:50, Leandro Santi wrote:

Hi Leandro,

> The dentry_stat.nr_dentry counter isn't being properly protected against
> concurrent access. We've been observing a drift of about 8000 units per
> day on some large MP Maildir++ mailstore nodes.
> The following (trivial) patch is pretty much a backport from 2.6.

Almost the same fix was posted in August 2003.

http://marc.theaimsgroup.com/?l=linux-kernel&m=106203778320540&w=2


ciao, Marc

2004-09-20 18:23:02

by Marcelo Tosatti

Subject: Re: [PATCH 2.4] fix dcache nr_dentry race

This one should have been applied long time ago, as Marc noted.

Thanks Leandro, committed to 2.4 BK repo.

On Sun, Sep 19, 2004 at 04:50:57AM -0300, Leandro Santi wrote:
>
> Hi Marcelo,
>
> The dentry_stat.nr_dentry counter isn't being properly protected against
> concurrent access. We've been observing a drift of about 8000 units per
> day on some large MP Maildir++ mailstore nodes.
>
> The following (trivial) patch is pretty much a backport from 2.6.