On 8/19/2015 12:50, Al Viro wrote:
> On Tue, Aug 18, 2015 at 03:23:59PM +0800, Kinglong Mee wrote:
>> @@ -181,7 +191,11 @@ static int expkey_show(struct seq_file *m,
>> if (test_bit(CACHE_VALID, &h->flags) &&
>> !test_bit(CACHE_NEGATIVE, &h->flags)) {
>> seq_printf(m, " ");
>> - seq_path(m, &ek->ek_path, "\\ \t\n");
>> + if (legitimize_mntget(ek->ek_path.mnt)) {
>> + seq_path(m, &ek->ek_path, "\\ \t\n");
>> + mntput(ek->ek_path.mnt);
>> + } else
>> + seq_printf(m, "Dir-unmounting");
>
> IDGI... What locking environment do you have here? Note that use
> of mnt_add_count(mnt, -1) in MNT_SYNC_UMOUNT case in __legitimize_mnt()
> is OK only because we
> a) are under rcu_read_lock() and
> b) have synchronize_rcu() in namespace_unlock().
There are not any locking exist is this patch site.
Thanks for your comments about the mnt_add_count() following.
I want add rcu_read_lock in legitimize_mntget as,
struct vfsmount *legitimize_mntget(struct vfsmount *vfsmnt)
{
rcu_read_lock();
......
rcu_read_unlock();
return vfsmnt;
}
Is it OK?
>
> You don't seem to be under rcu_read_lock() here, so what's to stop that
> from racing with the final mntput()? IOW, suppose that on the first
> pass through legitimize_mntget() you read mount_lock, bump refcount,
> recheck the lock and notice that it has been touched. You proceed to
> decrement refcount. Fine, except that what would've been the final
> mntput() has just noticed that refcount hadn't reached 0 and buggered
> off. And in the meanwhile, MNT_UMOUNT had been set. Now you decrement
> the refcount to zero, notice that MNT_UMOUNT and go away. Have a nice
> leak...
>
> The only reason why we are able to get away with mnt_add_count(mnt, -1) in
> that specific case in legitimize_mnt() is that MNT_SYNC_UMOUNT must have
> been set after we'd got rcu_read_lock() (otherwise we would've either hit
> a mismatch on mount_lock before incrementing refcount or wouldn't have
> run into that vfsmount at all) and thus we _know_ that the process that
> has set MNT_SYNC_UMOUNT couldn't have passed through synchronize_rcu()
> in namespace_unlock(), so it couldn't reach mntput_no_expire() until our
> caller does rcu_read_unlock() and we are free to decrement the refcount and
> leave - we won't be dropping the last reference here.
>
> Without MNT_SYNC_UMOUNT the callers of __legitimize_mnt() must use mntput()
> to drop the mistakenly acquired reference. Exactly because they can't
> rely on that syncronize_rcu() delaying the final mntput().
>
>> @@ -664,7 +696,12 @@ static int svc_export_show(struct seq_file *m,
>> return 0;
>> }
>> exp = container_of(h, struct svc_export, h);
>> - seq_path(m, &exp->ex_path, " \t\n\\");
>> + if (legitimize_mntget(exp->ex_path.mnt)) {
>> + seq_path(m, &exp->ex_path, " \t\n\\");
>> + mntput(exp->ex_path.mnt);
>> + } else
>> + seq_printf(m, "Dir-unmounting");
>
> Ditto. And grabbing/dropping references here seems to be an overkill...
Do you mean that calling seq_path without legitimize_mntget here ?
But the mnt is without reference to vfsmnt, only a fs_pin many times.
>
>> @@ -819,6 +867,12 @@ exp_find_key(struct cache_detail *cd, struct auth_domain *clp, int fsid_type,
>> err = cache_check(cd, &ek->h, reqp);
>> if (err)
>> return ERR_PTR(err);
>> +
>> + if (!legitimize_mntget(ek->ek_path.mnt)) {
>> + cache_put(&ek->h, ek->cd);
>> + return ERR_PTR(-ENOENT);
>> + }
>
> Ditto.
>
>> @@ -842,6 +896,8 @@ exp_get_by_name(struct cache_detail *cd, struct auth_domain *clp,
>> err = cache_check(cd, &exp->h, reqp);
>> if (err)
>> return ERR_PTR(err);
>> +
>> + mntget(exp->ex_path.mnt);
>
> What's to make that mntget() legitimate?
The mnt has the reference to vfsmnt here, so just using mntget is safe.
>
>> static inline struct svc_export *exp_get(struct svc_export *exp)
>> {
>> cache_get(&exp->h);
>> + mntget(exp->ex_path.mnt);
>
> Ditto.
>
The mnt has the reference, Same as above.
thanks,
Kinglong Mee
--
To unsubscribe from this list: send the line "unsubscribe linux-nfs" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
