Messages in this thread |  | | | From | (Eric W. Biederman) | | Subject | Re: [Bluez-devel] Oops involving RFCOMM and sysfs | | Date | Mon, 07 Jan 2008 01:21:40 -0700 |
Tejun Heo <htejun@gmail.com> writes:
> Hello,
>
> Tejun Heo wrote:
>> Al Viro wrote:
>>> On Sat, Jan 05, 2008 at 11:30:25PM +0900, Tejun Heo wrote:
>>>>> Assuming that this is what we get, everything looks explainable - we
>>>>> have sysfs_rename_dir() calling sysfs_get_dentry() while the parent
>>>>> gets evicted. We don't have any exclusion, so while we are playing
>>>>> silly buggers with lookups in sysfs_get_dentry() we have parent become
>>>>> negative; the rest is obvious...
>>>> That part of code is walking down the sysfs tree from the s_root of
>>>> sysfs hierarchy and on each step parent is held using dget() while being
>>>> referenced, so I don't think they can turn negative there.
>>> Turn? Just what stops you from getting a negative (and unhashed) from
>>> lookup_one_noperm() and on the next iteration being buggered on mutex_lock()?
>>
>> Right, I haven't thought about that. When sysfs_get_dentry() is called,
>> @sd is always valid so unless there was existing negative dentry, lookup
>> is guaranteed to return positive dentry, but by populating dcache with
>> negative dentry before a node is created, things can go wrong. I don't
>> think that's what's going on here tho. If that was the case, the
>> while() loop looking up the next sd to lookup (@cur) should have blown
>> up as negative dentry will have NULL d_fsdata which doesn't match any sd.
>>
>> I guess what's needed here is d_revalidate() as other distributed
>> filesystems do. I'll test whether this can be actually triggered and
>> prepare a fix. Thanks a lot for pointing out the problem.
>
> This can't happen because lookup of non-existent entry doesn't create a
> negative dentry. The new dentry is never hashed and killed after lookup
> failure, the above scenario can't happen.
>
> That said, the mechanism is a bit too fragile. sysfs currently ensures
> that dentry/inode point to the associated sysfs_dirent. This is mainly
> remanent of conversion from previous VFS based implementation. I think
> the right thing to do here is to make sysfs behave like other proper
> distributed filesystems using d_revalidate.
Huh? We still need something like sysfs_get_dentry to find the dentries
for the rename or move operation. So we can call d_move.
Further we should be talking about sysfs_move_dir not sysfs_rename_dir
as only the networking code calls sysfs_rename_dir. Not that it is
significantly different.
I think I saw a change in lock ordering recently to help distributed
filesystems, and maybe that will simplify some of this.
Anyway I figure if I am to understand this I better go look and see if
I can find the start of this thread. I don't have a clue
where we are blowing up.
Eric
|  |