On Fri, Oct 8, 2010 at 10:35 PM, Fred Isaman <iisaman@xxxxxxxxxx> wrote:
> On Fri, Oct 8, 2010 at 10:13 AM, P.B.Shelley <shelleypt@xxxxxxxxx> wrote:
>> On Thu, Oct 7, 2010 at 10:06 PM, Benny Halevy <bhalevy@xxxxxxxxxxx> wrote:
>>> On 2010-10-07 10:01, Fred Isaman wrote:
>>>> On Thu, Oct 7, 2010 at 9:34 AM, Benny Halevy <bhalevy@xxxxxxxxxxx> wrote:
>>>>> On 2010-10-06 16:35, Fred Isaman wrote:
>>>>>> Right now, when we set the stateid, we blindly overwrite the current
>>>>>> one, allowing the seqid to incorrectly roll backward.
>>>>>>
>>>>>> Signed-off-by: Fred Isaman <iisaman@xxxxxxxxxx>
>>>>>> ---
>>>>>> Âfs/nfs/pnfs.c | Â 38 ++++++++++++++++++++++++++++++++------
>>>>>> Â1 files changed, 32 insertions(+), 6 deletions(-)
>>>>>>
>>>>>> diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
>>>>>> index 39bce9b..555955b 100644
>>>>>> --- a/fs/nfs/pnfs.c
>>>>>> +++ b/fs/nfs/pnfs.c
>>>>>> @@ -459,16 +459,42 @@ pnfs_destroy_all_layouts(struct nfs_client *clp)
>>>>>> Â Â Â }
>>>>>> Â}
>>>>>>
>>>>>> +/* update lo->stateid with new if is more recent
>>>>>> + *
>>>>>> + * lo->stateid could be the open stateid, in which case we just use what given.
>>>>>> + */
>>>>>> Âstatic void
>>>>>> Âpnfs_set_layout_stateid(struct pnfs_layout_hdr *lo,
>>>>>> - Â Â Â Â Â Â Â Â Â Â const nfs4_stateid *stateid)
>>>>>> + Â Â Â Â Â Â Â Â Â Â const nfs4_stateid *new)
>>>>>> Â{
>>>>>> - Â Â /* TODO - should enforce that embedded seqid, in the case
>>>>>> - Â Â Â* that the two stateid.others are equal, Âonly increases.
>>>>>> - Â Â Â* Complicated by wrap-around.
>>>>>> - Â Â Â*/
>>>>>> + Â Â nfs4_stateid *old = &lo->stateid;
>>>>>> + Â Â bool overwrite = false;
>>>>>> +
>>>>>> Â Â Â write_seqlock(&lo->seqlock);
>>>>>> - Â Â memcpy(lo->stateid.data, stateid->data, sizeof(lo->stateid.data));
>>>>>> + Â Â if (!test_bit(NFS_LAYOUT_STATEID_SET, &lo->state) ||
>>>>>> + Â Â Â Â memcmp(old->stateid.other, new->stateid.other, sizeof(new->stateid.other)))
>>>>>> + Â Â Â Â Â Â overwrite = true;
>>>>>> + Â Â else {
>>>>>> + Â Â Â Â Â Â u32 oldseq, newseq, limit;
>>>>>> +
>>>>>> + Â Â Â Â Â Â oldseq = be32_to_cpu(old->stateid.seqid);
>>>>>> + Â Â Â Â Â Â newseq = be32_to_cpu(new->stateid.seqid);
>>>>>> + Â Â Â Â Â Â /* There are no good bounds on window size, so just
>>>>>> + Â Â Â Â Â Â Â* use a ridiculously large window of 2^31.
>>>>>> + Â Â Â Â Â Â Â*/
>>>>>> + Â Â Â Â Â Â limit = oldseq + (1 << 31);
>>>>>> + Â Â Â Â Â Â if (oldseq < limit) {
>>>>>> + Â Â Â Â Â Â Â Â Â Â /* The easy, non-wraparound case */
>>>>>> + Â Â Â Â Â Â Â Â Â Â if (oldseq < newseq && newseq < limit)
>>>>>> + Â Â Â Â Â Â Â Â Â Â Â Â Â Â overwrite = true;
>>>>>> + Â Â Â Â Â Â } else {
>>>>>> + Â Â Â Â Â Â Â Â Â Â /* Near wraparound edge */
>>>>>> + Â Â Â Â Â Â Â Â Â Â if (oldseq < newseq || newseq < limit)
>>>>>> + Â Â Â Â Â Â Â Â Â Â Â Â Â Â overwrite = true;
>>>>>> + Â Â Â Â Â Â }
>>>>>
>>>>> Wouldn't it be simpler to just look at (int32_t)(newseq - oldseq)?
>>>>>
>>>>
>>>> Why yes it would. ÂI'll send a new version of this patch shortly.
>>>>
>>>
>>> No need :)
>>> I'll just change this as follows:
>>>
>>> + Â Â Â else {
>>> + Â Â Â Â Â Â Â u32 oldseq, newseq, limit;
>>> +
>>> + Â Â Â Â Â Â Â oldseq = be32_to_cpu(old->stateid.seqid);
>>> + Â Â Â Â Â Â Â newseq = be32_to_cpu(new->stateid.seqid);
>>> + Â Â Â Â Â Â Â if ((int)(newseq - oldseq) > 0)
>>> + Â Â Â Â Â Â Â Â Â Â Â overwrite = true;
>> Do we also need to verify the other field of the stateid? Will there
>> be situations that server change the other field and reset the seqid?
>
> The server is going to use the "other" we sent, except in the case we
> sent an open stateid. ÂThe only potential for trouble I see is if a
> LAYOUTGET reply gets lost in the network for a long time and is
> received after the layout stateid has been reset for some reason.
> However, that implies an error elsewhere (which may well exist at the
> moment...careful stateid handling is next on the agenda), as we should
> have been waiting for that lseg to arrive before continuing.
Oops, I missed the other field comparing code. Thank you for pointing
it out, Benny.
Can client choose to send an open stateid for LATYOUTGET request, even
if client has a layout stateid for the file?
--
Thanks,
Shelley
--
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
