Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Wed, 2 May 2018 19:21:05 +0100 (BST)
From:   James Simmons <jsimmons@infradead.org>
To:     "Dilger, Andreas" <andreas.dilger@intel.com>
cc:     NeilBrown <neilb@suse.com>, "Drokin, Oleg" <oleg.drokin@intel.com>,
        Greg Kroah-Hartman <gregkh@linuxfoundation.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Lustre Development List <lustre-devel@lists.lustre.org>
Subject: Re: [lustre-devel] [PATCH 04/10] staging: lustre: lu_object: move
 retry logic inside htable_lookup
In-Reply-To: <EFC6D018-5FB5-4902-AE7A-B91A5E0148FE@intel.com>
Message-ID: <alpine.LFD.2.21.1805021914550.24633@casper.infradead.org>
References: <152514658325.17843.11455067361317157487.stgit@noble> <152514675897.17843.15112214060540196720.stgit@noble> <EFC6D018-5FB5-4902-AE7A-B91A5E0148FE@intel.com>
User-Agent: Alpine 2.21 (LFD 202 2017-01-01)
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk


> On Apr 30, 2018, at 21:52, NeilBrown <neilb@suse.com> wrote:
> > 
> > The current retry logic, to wait when a 'dying' object is found,
> > spans multiple functions.  The process is attached to a waitqueue
> > and set TASK_UNINTERRUPTIBLE in htable_lookup, and this status
> > is passed back through lu_object_find_try() to lu_object_find_at()
> > where schedule() is called and the process is removed from the queue.
> > 
> > This can be simplified by moving all the logic (including
> > hashtable locking) inside htable_lookup(), which now never returns
> > EAGAIN.
> > 
> > Note that htable_lookup() is called with the hash bucket lock
> > held, and will drop and retake it if it needs to schedule.
> > 
> > I made this a 'goto' loop rather than a 'while(1)' loop as the
> > diff is easier to read.
> > 
> > Signed-off-by: NeilBrown <neilb@suse.com>
> > ---
> > drivers/staging/lustre/lustre/obdclass/lu_object.c |   73 +++++++-------------
> > 1 file changed, 27 insertions(+), 46 deletions(-)
> > 
> > diff --git a/drivers/staging/lustre/lustre/obdclass/lu_object.c b/drivers/staging/lustre/lustre/obdclass/lu_object.c
> > index 2bf089817157..93daa52e2535 100644
> > --- a/drivers/staging/lustre/lustre/obdclass/lu_object.c
> > +++ b/drivers/staging/lustre/lustre/obdclass/lu_object.c
> > @@ -586,16 +586,21 @@ EXPORT_SYMBOL(lu_object_print);
> > static struct lu_object *htable_lookup(struct lu_site *s,
> 
> It's probably a good idea to add a comment for this function that it may
> drop and re-acquire the hash bucket lock internally.
> 
> > 				       struct cfs_hash_bd *bd,
> > 				       const struct lu_fid *f,
> > -				       wait_queue_entry_t *waiter,
> > 				       __u64 *version)
> > {
> > +	struct cfs_hash		*hs = s->ls_obj_hash;
> > 	struct lu_site_bkt_data *bkt;
> > 	struct lu_object_header *h;
> > 	struct hlist_node	*hnode;
> > -	__u64  ver = cfs_hash_bd_version_get(bd);
> > +	__u64 ver;
> > +	wait_queue_entry_t waiter;
> > 
> > -	if (*version == ver)
> > +retry:
> > +	ver = cfs_hash_bd_version_get(bd);
> > +
> > +	if (*version == ver) {
> > 		return ERR_PTR(-ENOENT);
> > +	}
> 
> (style) we don't need the {} around a single-line if statement

I hate to be that guy but could you run checkpatch on your patches.
 
> > 	*version = ver;
> > 	bkt = cfs_hash_bd_extra_get(s->ls_obj_hash, bd);
> > @@ -625,11 +630,15 @@ static struct lu_object *htable_lookup(struct lu_site *s,
> > 	 * drained), and moreover, lookup has to wait until object is freed.
> > 	 */
> > 
> > -	init_waitqueue_entry(waiter, current);
> > -	add_wait_queue(&bkt->lsb_marche_funebre, waiter);
> > +	init_waitqueue_entry(&waiter, current);
> > +	add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
> > 	set_current_state(TASK_UNINTERRUPTIBLE);
> > 	lprocfs_counter_incr(s->ls_stats, LU_SS_CACHE_DEATH_RACE);
> > -	return ERR_PTR(-EAGAIN);
> > +	cfs_hash_bd_unlock(hs, bd, 1);
> 
> This looks like it isn't unlocking and locking the hash bucket in the same
> manner that it was done in the caller.  Here excl = 1, but in the caller
> you changed it to excl = 0?

This is very much like the work done by Lai. The difference is Lai remove 
the work queue handling complete in htable_lookup(). You can see the 
details at https://jira.hpdd.intel.com/browse/LU-9049. I will push the 
missing lu_object fixes including LU-9049 on top of your patch set so you
can see the approach Lai did. Form their we can figure out merge the 
lu_object work and fixing the issues Andreas and I pointed out.

> > +	schedule();
> > +	remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
> 
> Is it worthwhile to use your new helper function here to get the wq from "s"?
> 
> > +	cfs_hash_bd_lock(hs, bd, 1);
> > +	goto retry;
> > }
> > 
> > /**
> > @@ -693,13 +702,14 @@ static struct lu_object *lu_object_new(const struct lu_env *env,
> > }
> > 
> > /**
> > - * Core logic of lu_object_find*() functions.
> > + * Much like lu_object_find(), but top level device of object is specifically
> > + * \a dev rather than top level device of the site. This interface allows
> > + * objects of different "stacking" to be created within the same site.
> >  */
> > -static struct lu_object *lu_object_find_try(const struct lu_env *env,
> > -					    struct lu_device *dev,
> > -					    const struct lu_fid *f,
> > -					    const struct lu_object_conf *conf,
> > -					    wait_queue_entry_t *waiter)
> > +struct lu_object *lu_object_find_at(const struct lu_env *env,
> > +				    struct lu_device *dev,
> > +				    const struct lu_fid *f,
> > +				    const struct lu_object_conf *conf)
> > {
> > 	struct lu_object      *o;
> > 	struct lu_object      *shadow;
> > @@ -725,17 +735,16 @@ static struct lu_object *lu_object_find_try(const struct lu_env *env,
> > 	 * It is unnecessary to perform lookup-alloc-lookup-insert, instead,
> > 	 * just alloc and insert directly.
> > 	 *
> > -	 * If dying object is found during index search, add @waiter to the
> > -	 * site wait-queue and return ERR_PTR(-EAGAIN).
> > 	 */
> > 	if (conf && conf->loc_flags & LOC_F_NEW)
> > 		return lu_object_new(env, dev, f, conf);
> > 
> > 	s  = dev->ld_site;
> > 	hs = s->ls_obj_hash;
> > -	cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 1);
> > -	o = htable_lookup(s, &bd, f, waiter, &version);
> > -	cfs_hash_bd_unlock(hs, &bd, 1);
> > +	cfs_hash_bd_get_and_lock(hs, (void *)f, &bd, 0);
> > +	o = htable_lookup(s, &bd, f, &version);
> > +	cfs_hash_bd_unlock(hs, &bd, 0);
> 
> Here you changed the locking to a non-exclusive (read) lock instead of an
> exclusive (write) lock?  Why.

I have the same question.

> 
> > +
> > 	if (!IS_ERR(o) || PTR_ERR(o) != -ENOENT)
> > 		return o;
> > 
> > @@ -751,7 +760,7 @@ static struct lu_object *lu_object_find_try(const struct lu_env *env,
> > 
> > 	cfs_hash_bd_lock(hs, &bd, 1);
> > 
> > -	shadow = htable_lookup(s, &bd, f, waiter, &version);
> > +	shadow = htable_lookup(s, &bd, f, &version);
> > 	if (likely(PTR_ERR(shadow) == -ENOENT)) {
> > 		cfs_hash_bd_add_locked(hs, &bd, &o->lo_header->loh_hash);
> > 		cfs_hash_bd_unlock(hs, &bd, 1);
> > @@ -766,34 +775,6 @@ static struct lu_object *lu_object_find_try(const struct lu_env *env,
> > 	lu_object_free(env, o);
> > 	return shadow;
> > }
> > -
> > -/**
> > - * Much like lu_object_find(), but top level device of object is specifically
> > - * \a dev rather than top level device of the site. This interface allows
> > - * objects of different "stacking" to be created within the same site.
> > - */
> > -struct lu_object *lu_object_find_at(const struct lu_env *env,
> > -				    struct lu_device *dev,
> > -				    const struct lu_fid *f,
> > -				    const struct lu_object_conf *conf)
> > -{
> > -	wait_queue_head_t	*wq;
> > -	struct lu_object	*obj;
> > -	wait_queue_entry_t	   wait;
> > -
> > -	while (1) {
> > -		obj = lu_object_find_try(env, dev, f, conf, &wait);
> > -		if (obj != ERR_PTR(-EAGAIN))
> > -			return obj;
> > -		/*
> > -		 * lu_object_find_try() already added waiter into the
> > -		 * wait queue.
> > -		 */
> > -		schedule();
> > -		wq = lu_site_wq_from_fid(dev->ld_site, (void *)f);
> > -		remove_wait_queue(wq, &wait);
> > -	}
> > -}
> > EXPORT_SYMBOL(lu_object_find_at);
> > 
> > /**
> > 
> > 
> > _______________________________________________
> > lustre-devel mailing list
> > lustre-devel@lists.lustre.org
> > http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
> 
> Cheers, Andreas
> --
> Andreas Dilger
> Lustre Principal Architect
> Intel Corporation
> 
> 
> 
> 
> 
> 
> 
>