Return-Path: Received: from fieldses.org ([173.255.197.46]:54296 "EHLO fieldses.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750822AbdADUhD (ORCPT ); Wed, 4 Jan 2017 15:37:03 -0500 Date: Wed, 4 Jan 2017 15:26:34 -0500 From: "J. Bruce Fields" To: andros@netapp.com Cc: neilb@suse.com, linux-nfs@vger.kernel.org Subject: Re: [PATCH Version 4] SVCAUTH reap the rsc cache entry on RPC_SS_PROC_DESTROY Message-ID: <20170104202634.GA21562@fieldses.org> References: <1482428286-33744-1-git-send-email-andros@netapp.com> <1482428286-33744-2-git-send-email-andros@netapp.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <1482428286-33744-2-git-send-email-andros@netapp.com> Sender: linux-nfs-owner@vger.kernel.org List-ID: I'm not against the patch, but I'm still not convinced by the explanation: On Thu, Dec 22, 2016 at 12:38:06PM -0500, andros@netapp.com wrote: > From: Neil Brown > > The rsc cache code operates in a read_lock/write_lock environment. > Changes to a cache entry should use the provided rsc_update > routine which takes the write_lock. It looks pretty suspicious to be setting CACHE_NEGATIVE without the cache_lock for write, but I'm not actually convinced there's a bug there either. In any case not one that you'd be hitting reliably. > The current code sets the expiry_time and the CACHE_NEGATIVE flag > without taking the write_lock as it does not call rsc_update. > Without this patch, while cache_clean sees the entries to be > removed, it does not remove the rsc_entries. This is because > rsc_update updates other fields such as flush_time and last_refresh > in the entry to trigger cache_clean to reap the entry. I think the root cause of the particular behavior you were seeing was actually an oversight from Neil's c5b29f885afe "sunrpc: use seconds since boot in expiry cache", which missed this one occurrence of get_seconds(). So it's setting the item's entry to something decades in the future. And that's probably not been a huge deal since these entries aren't so big, and they will eventually get cleaned up by cache_purge when the cache is destroyed. Still, I can imagine it slowing down cache lookups on a long-lived server. The one-liner: - rsci->h.expiry_time = get_seconds(); + rsci->h.expiry_time = seconds_since_boot(); would probably also do the job. Am I missing something? But, OK, I think Neil's patch will ensure entries get cleaned up more quickly than that would, and might also fix a rare race. --b. > > Add a new sunrpc_cache_unhash() function (by Neil Brown) designed > to directly unhash and reap the to be destroyed cache entry. > > Signed-off-by: Neil Brown > Reported-by: Andy Adamson > Signed-off-by: Andy Adamson > --- > include/linux/sunrpc/cache.h | 1 + > net/sunrpc/auth_gss/svcauth_gss.c | 4 ++-- > net/sunrpc/cache.c | 13 +++++++++++++ > 3 files changed, 16 insertions(+), 2 deletions(-) > > diff --git a/include/linux/sunrpc/cache.h b/include/linux/sunrpc/cache.h > index 62a60ee..9dcf2c8 100644 > --- a/include/linux/sunrpc/cache.h > +++ b/include/linux/sunrpc/cache.h > @@ -227,6 +227,7 @@ extern int cache_check(struct cache_detail *detail, > extern int sunrpc_cache_register_pipefs(struct dentry *parent, const char *, > umode_t, struct cache_detail *); > extern void sunrpc_cache_unregister_pipefs(struct cache_detail *); > +extern void sunrpc_cache_unhash(struct cache_detail *, struct cache_head *); > > /* Must store cache_detail in seq_file->private if using next three functions */ > extern void *cache_seq_start(struct seq_file *file, loff_t *pos); > diff --git a/net/sunrpc/auth_gss/svcauth_gss.c b/net/sunrpc/auth_gss/svcauth_gss.c > index 45662d7..78f8a9c 100644 > --- a/net/sunrpc/auth_gss/svcauth_gss.c > +++ b/net/sunrpc/auth_gss/svcauth_gss.c > @@ -1489,8 +1489,8 @@ static void destroy_use_gss_proxy_proc_entry(struct net *net) {} > case RPC_GSS_PROC_DESTROY: > if (gss_write_verf(rqstp, rsci->mechctx, gc->gc_seq)) > goto auth_err; > - rsci->h.expiry_time = get_seconds(); > - set_bit(CACHE_NEGATIVE, &rsci->h.flags); > + /* Delete the entry from the cache_list and call cache_put */ > + sunrpc_cache_unhash(sn->rsc_cache, &rsci->h); > if (resv->iov_len + 4 > PAGE_SIZE) > goto drop; > svc_putnl(resv, RPC_SUCCESS); > diff --git a/net/sunrpc/cache.c b/net/sunrpc/cache.c > index 8aabe12..153e254 100644 > --- a/net/sunrpc/cache.c > +++ b/net/sunrpc/cache.c > @@ -1855,3 +1855,16 @@ void sunrpc_cache_unregister_pipefs(struct cache_detail *cd) > } > EXPORT_SYMBOL_GPL(sunrpc_cache_unregister_pipefs); > > +void sunrpc_cache_unhash(struct cache_detail *cd, struct cache_head *h) > +{ > + write_lock(&cd->hash_lock); > + if (!hlist_unhashed(&h->cache_list)){ > + hlist_del_init(&h->cache_list); > + cd->entries--; > + write_unlock(&cd->hash_lock); > + cache_put(h, cd); > + } else > + write_unlock(&cd->hash_lock); > +} > +EXPORT_SYMBOL_GPL(sunrpc_cache_unhash); > + > -- > 1.8.3.1