Return-Path: Received: from mail-ig0-f173.google.com ([209.85.213.173]:37858 "EHLO mail-ig0-f173.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753043AbcCaP1G (ORCPT ); Thu, 31 Mar 2016 11:27:06 -0400 Received: by mail-ig0-f173.google.com with SMTP id l20so70421918igf.0 for ; Thu, 31 Mar 2016 08:27:05 -0700 (PDT) MIME-Version: 1.0 In-Reply-To: References: Date: Thu, 31 Mar 2016 11:27:05 -0400 Message-ID: Subject: Re: commit 7c2dad99d6 "Don't let the ctime override attribute barriers" From: Olga Kornievskaia To: Trond Myklebust Cc: linux-nfs Content-Type: text/plain; charset=UTF-8 Sender: linux-nfs-owner@vger.kernel.org List-ID: On Thu, Mar 31, 2016 at 11:16 AM, Trond Myklebust wrote: > On Thu, Mar 31, 2016 at 10:36 AM, Olga Kornievskaia wrote: >> On Thu, Mar 31, 2016 at 10:21 AM, Trond Myklebust >> wrote: >>> On Thu, Mar 31, 2016 at 10:15 AM, Olga Kornievskaia wrote: >>>> On Wed, Mar 30, 2016 at 7:32 PM, Trond Myklebust >>>> wrote: >>>>> On Wed, Mar 30, 2016 at 5:51 PM, Olga Kornievskaia wrote: >>>>>> On Wed, Mar 30, 2016 at 5:45 PM, Trond Myklebust >>>>>> wrote: >>>>>>> On Wed, Mar 30, 2016 at 5:02 PM, Olga Kornievskaia wrote: >>>>>>>> On Tue, Mar 29, 2016 at 3:47 PM, Olga Kornievskaia wrote: >>>>>>>>> I think that patch introduces a problem. Since the checking for the >>>>>>>>> change in ctime was removed by the commit it leads to (improper) cache >>>>>>>>> invalidation in NFSv3. >>>>>>>>> >>>>>>>>> Test is write 10240bytes to the server then read it. Expectation is >>>>>>>>> not to see read on the wire. In the test the write is spread over >>>>>>>>> 3rpcs. >>>>>>>>> >>>>>>>>> On the 1nd reply >>>>>>>>> fattr->gencount=33 nfsi->gencount=32 generation_counter=35 >>>>>>>>> On the 2nd reply >>>>>>>>> fattr->gencount=34 nfsi->gencount=36 generation_counter=36 >>>>>>>>> >>>>>>>>> In the code when processing 2nd reply, >>>>>>>>> nfs_post_op_update_inode_force_wcc_locked() calls into >>>>>>>>> nfs_inode_attrs_need_update() it determines that it doesn't need to >>>>>>>>> update them (even though the size and the time have changed). so it >>>>>>>>> doesn't call nfs_wcc_update_inode() so the inode->i_version doesn't >>>>>>>>> get set to the ctime that was received in the 2nd reply. >>>>>>>>> >>>>>>>>> On the 3rd reply >>>>>>>>> fattr->gencount=37 nfsi->gencount=36 generation_counter=37 >>>>>>>>> >>>>>>>>> It leads to nfs_inode_attrs_need_update() returns 1 and in the >>>>>>>>> nfs_update_inode() the difference in the ctimes leads to invalidation. >>>>>>>>> fattr->gencount was update from nfs_writeback_update_node() -> >>>>>>>>> nfs_post_op_update_inode_force_wcc() calling nfs_fattr_set_barrier(). >>>>>>>>> >>>>>>>>> I'm not sure what appropriate values for "gencount" should have been. >>>>>>>>> But if the check for nfs_ctime_need_update() was still there in >>>>>>>>> nfs_inode_attrs_need_update() then the 2nd reply would have >>>>>>>>> appropriately updated the i_version and not lead to invalidation. >>>>>>>> >>>>>>>> Would like to add that this problem is not seen against the Linux >>>>>>>> server because it doesn't send "before" attributes. So code doesn't >>>>>>>> set the "pre_change_attr" which later doesn't make what's stored in >>>>>>>> inode->i_version. >>>>>>>> >>>>>>>> The problem also not seen for v4 because pre_change_attr is not gotten >>>>>>>> from the "before" attributes but instead from the previous value in >>>>>>>> inode->i_version which is then compared to the itself. >>>>>>>> >>>>>>>> If reverting the problematic commit is not the solution, then how >>>>>>>> about ignoring the "before" ctime attributes sent by the server. This >>>>>>>> also helps with the out-of-order RPCs. >>>>>>> >>>>>>> Why bother doing that on the client? These attributes aren't mandatory >>>>>>> to send... >>>>>>> >>>>>> >>>>>> Leads to poor client performances. Every large enough read invalidates >>>>>> the cache so all the reads go to the server always. >>>>> >>>>> I'm saying why not just turn off the WCC functionality on the server then. >>>> >>>> One reasoning could be that providing cache consistency is client's >>>> responsibility. Server only provides needed information. >>> >>> Sure, but the server is the single point of control for all clients. >> >> What about other client implementations that might use before attributes? >> >>>> While I can see that handling out-of-order RPCs might not be worth it >>>> because hopefully it doesn't happen often but handling in order RPCs >>>> and always invalidating the cache is a bug. >>> >>> We don't invalidate on in-order RPCs. >> >> Yes we do and that was the point of my initial post. The commit >> introduced a regression. > > No we don't. If the RPC replies were received in the order in which > they were processed on the server, then the WCC information would > match what is currently in the client cache. Your patch turns off WCC > processing on the client in order to avoid that ordering problem by > having the client fill in WCC info that matches what is in its cache. The RPCs are returned in order. The problem is when the 2nd reply is gotten the attributes are not updated. When the 3rd reply is processed it finds the attributes no longer match. The attributes should have been updated. They would have been updated if the patch was still present. Enabling the check lead to a problem Chuck reported. I'm reporting that disabling it also leads to a problem. My proposal to turn off WCC was in attempt to solve the problem by having something that also solves Chuck problem.