Jeff Layton wrote:
> On Wed, Apr 25, 2007 at 02:09:34PM -0400, Jeff Layton wrote:
>> On Wed, Apr 25, 2007 at 01:39:20PM -0400, Jeff Layton wrote:
>>> Currently, when auth_reload is called, it only looks at the tv_sec field
>>> of the mtime when deciding whether to invalidate the exports cache. It's
>>> fairly simple to fool this by doing something like:
>>>
>>> # exportfs -rv && rpc.mountd && exportfs -uva && exportfs -iv -o no_root_squash,rw 127.0.0.1:/foo
>>>
>>> With this, exportfs will show the export for /foo, but mountd will still have
>>> whatever contents were in /etc/exports. The issue is that the etab is being
>>> updated twice during the same second, and mountd is reading in the file in
>>> between updates. When it goes to look at the file again, its timestamp matches
>>> the timestamp of the cache, and it ends up keeping the cached contents even
>>> though the file has changed.
>>>
>>> While not all local filesystems provide sub-second timestamps, we might as
>>> well fix this problem on those that do. The following patch changes
>>> auth_reload to consider the tv_nsec field of the mtime when deciding whether
>>> to invalidate the export cache. It also fixes up the callers to pass it a
>>> pointer to a struct timespec for it to fill out.
>>>
>>> I've not yet tested this on a filesystem that provides sub-second timestamps,
>>> so I'm not clear on how well this works yet, but it seems to not break
>>> anything on ext3 in some cursory testing.
>>>
>> No sooner than I post than I see a (minor) problem. The check for a NULL
>> pointer in auth_reload should be (ts != NULL) instead of (ts). Respun patch
>> follows:
>>
> 
> I tested this on a filesystem that does nanosecond timestamps (xfs), and it
> seems to correct the original problem.

Hi Jeff-

I don't know much about the export logic, but why not use inode change 
notification instead of time stamps?