Return-Path: Received: from fieldses.org ([173.255.197.46]:59686 "EHLO fieldses.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751976AbeC1UKn (ORCPT ); Wed, 28 Mar 2018 16:10:43 -0400 Date: Wed, 28 Mar 2018 16:10:43 -0400 From: "J . Bruce Fields" To: Jeff Layton Cc: Trond Myklebust , linux-nfs@vger.kernel.org Subject: Re: [PATCH] nfsd: Do not refuse to serve out of cache Message-ID: <20180328201043.GA6091@fieldses.org> References: <20180328161801.8360-1-trond.myklebust@primarydata.com> <1522264845.5165.1.camel@kernel.org> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <1522264845.5165.1.camel@kernel.org> Sender: linux-nfs-owner@vger.kernel.org List-ID: Applying, thanks. On Wed, Mar 28, 2018 at 03:20:45PM -0400, Jeff Layton wrote: > On Wed, 2018-03-28 at 12:18 -0400, Trond Myklebust wrote: > > Currently the knfsd replay cache appears to try to refuse replying to > > retries that come within 200ms of the cache entry being created. That > > makes limited sense in today's world of high speed TCP. Trond gave me some helpful context in person, I may tag that onto the changelog: After a TCP disconnection, a client can very easily reconnect and retry an rpc in less than 200ms. If this logic drops that retry, however, the client may be quite slow to retry again. This logic is original to the first reply cache implementation in 2.1, and may have made more sense for UDP clients that retried much more frequently. We're still dropping on finding the original request still in progress, which can cause the same problem, though it's less likely. Note svc_check_conn_limits is often the cause of those disconnections. We may want to fix that some day. --b. > > > > Signed-off-by: Trond Myklebust > > --- > > fs/nfsd/cache.h | 5 ----- > > fs/nfsd/nfscache.c | 6 ++---- > > 2 files changed, 2 insertions(+), 9 deletions(-) > > > > diff --git a/fs/nfsd/cache.h b/fs/nfsd/cache.h > > index 046b3f048757..b7559c6f2b97 100644 > > --- a/fs/nfsd/cache.h > > +++ b/fs/nfsd/cache.h > > @@ -67,11 +67,6 @@ enum { > > RC_REPLBUFF, > > }; > > > > -/* > > - * If requests are retransmitted within this interval, they're > > dropped. > > - */ > > -#define RC_DELAY (HZ/5) > > - > > /* Cache entries expire after this time period */ > > #define RC_EXPIRE (120 * HZ) > > > > diff --git a/fs/nfsd/nfscache.c b/fs/nfsd/nfscache.c > > index 334f2ad60704..637f87c39183 100644 > > --- a/fs/nfsd/nfscache.c > > +++ b/fs/nfsd/nfscache.c > > @@ -394,7 +394,6 @@ nfsd_cache_lookup(struct svc_rqst *rqstp) > > __wsum csum; > > u32 hash = nfsd_cache_hash(xid); > > struct nfsd_drc_bucket *b = &drc_hashtbl[hash]; > > - unsigned long age; > > int type = rqstp->rq_cachetype; > > int rtn = RC_DOIT; > > > > @@ -461,12 +460,11 @@ nfsd_cache_lookup(struct svc_rqst *rqstp) > > found_entry: > > nfsdstats.rchits++; > > /* We found a matching entry which is either in progress or > > done. */ > > - age = jiffies - rp->c_timestamp; > > lru_put_end(b, rp); > > > > rtn = RC_DROPIT; > > - /* Request being processed or excessive rexmits */ > > - if (rp->c_state == RC_INPROG || age < RC_DELAY) > > + /* Request being processed */ > > + if (rp->c_state == RC_INPROG) > > goto out; > > > > /* From the hall of fame of impractical attacks: > > That condition always looked a bit suspicious to me. > > Acked-by: Jeff Layton