On 2011-09-13 10:32, tao.peng@xxxxxxx wrote: >> -----Original Message----- >> From: Benny Halevy [mailto:bhalevy@xxxxxxxxxx] >> Sent: Tuesday, September 13, 2011 3:51 PM >> To: Trond Myklebust >> Cc: Peng Tao; Peng, Tao; gusev.vitaliy@xxxxxxxxxxx; gusev.vitaliy@xxxxxxxxx; >> linux-nfs@xxxxxxxxxxxxxxx >> Subject: Re: [PATCH] nfs: fix inifinite loop at nfs4_layoutcommit_release >> >> On 2011-09-12 14:10, Trond Myklebust wrote: >>> On Mon, 2011-09-12 at 13:31 -0700, Benny Halevy wrote: >>>> On 2011-09-12 07:56, Peng Tao wrote: >>>>>> The layout segments are not really in use while in LAYOUTCOMMIT. >>>>>> We only need to get the stateid right with respect to concurrent layout recalls. >>>>> LAYOUTCOMMIT takes lseg reference to mark them as in use so that >>>>> layoutrecall cannot free them. >>>>> >>>> >>>> And if layoutrecall would have freed layout segments during layoutcommit, >>>> what is your specific concern? >>> >>> That layoutcommit is supposed to return NFS4ERR_BAD_LAYOUT in that case >>> according to section 18.42.3 of RFC5661. I can't find anything in the >>> errata that changes that requirement. >>> >> >> Right. That tells me there no need to strictly serialize LAYOUTCOMMITs >> with CB_LAYOUTRECALL, as long as the layout stateid sent with LAYOUTCOMMIT >> atomically represents the state when the operation was prepared. >> >> That said, since we do want the LAYOUTCOMMIT to succeed, it would be beneficial >> for the client to reply to a CB_LAYOUTRECALL received while a conflicting >> LAYOUTCOMMIT is in progress with NFS4ERR_DELAY. > I agree. How about adding a new flag to nfsi->flags for this? We can use the same flag on to ensure serialization of multiple layoutcommit. nfs_commit_set_lock/nfs_commit_clear_lock may not fit for this. > Sounds good in principle. Can you take a stab at a patch that does this? Benny >> >> The server, on its side, should prevent a distributed deadlock by avoiding >> blocking of a LAYOUTCOMMIT on an outstanding CB_LAYOUTRECALL for the same >> client that sent the LAYOUTCOMMIT. I'm not sure what error would be best to >> return. Maybe NFS4ERR_RECALL_CONFLICT if it would be allowed (it isn't listed >> for LAYOUTCOMMIT at the moment). Just returning NFS4ER_DELAY might lead to >> a live lock situation where neither the LAYOUTCOMMIT not the CB_LAYOUTRECALL >> complete. >> >> Benny > -- 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
