Message-ID: <20210723130554.GA38923@rowland.harvard.edu>
Date:   Fri, 23 Jul 2021 09:05:54 -0400
From:   Alan Stern <stern@...land.harvard.edu>
To:     Manfred Spraul <manfred@...orfullife.com>
Cc:     "Paul E. McKenney" <paulmck@...nel.org>,
        linux-kernel@...r.kernel.org, linux-arch@...r.kernel.org,
        kernel-team@...com, mingo@...nel.org, parri.andrea@...il.com,
        will@...nel.org, peterz@...radead.org, boqun.feng@...il.com,
        npiggin@...il.com, dhowells@...hat.com, j.alglave@....ac.uk,
        luc.maranget@...ia.fr, akiyks@...il.com
Subject: Re: [PATCH memory-model 2/4] tools/memory-model: Add example for
 heuristic lockless reads

On Fri, Jul 23, 2021 at 08:52:50AM +0200, Manfred Spraul wrote:
> Hi Alan,

Hi.

> On 7/23/21 4:08 AM, Alan Stern wrote:
> > On Wed, Jul 21, 2021 at 02:10:01PM -0700, Paul E. McKenney wrote:
> > > This commit adds example code for heuristic lockless reads, based loosely
> > > on the sem_lock() and sem_unlock() functions.
> > > 
> > > Reported-by: Manfred Spraul <manfred@...orfullife.com>
> > > [ paulmck: Update per Manfred Spraul and Hillf Danton feedback. ]
> > > Signed-off-by: Paul E. McKenney <paulmck@...nel.org>
> > > ---
> > >   .../Documentation/access-marking.txt          | 94 +++++++++++++++++++
> > >   1 file changed, 94 insertions(+)
> > > 
> > > diff --git a/tools/memory-model/Documentation/access-marking.txt b/tools/memory-model/Documentation/access-marking.txt
> > > index 58bff26198767..be7d507997cf8 100644
> > > --- a/tools/memory-model/Documentation/access-marking.txt
> > > +++ b/tools/memory-model/Documentation/access-marking.txt
> > > @@ -319,6 +319,100 @@ of the ASSERT_EXCLUSIVE_WRITER() is to allow KCSAN to check for a buggy
> > >   concurrent lockless write.
> > > +Lock-Protected Writes With Heuristic Lockless Reads
> > > +---------------------------------------------------
> > > +
> > > +For another example, suppose that the code can normally make use of
> > > +a per-data-structure lock, but there are times when a global lock
> > > +is required.  These times are indicated via a global flag.  The code
> > > +might look as follows, and is based loosely on nf_conntrack_lock(),
> > > +nf_conntrack_all_lock(), and nf_conntrack_all_unlock():
> > > +
> > > +	bool global_flag;
> > > +	DEFINE_SPINLOCK(global_lock);
> > > +	struct foo {
> > > +		spinlock_t f_lock;
> > > +		int f_data;
> > > +	};
> > > +
> > > +	/* All foo structures are in the following array. */
> > > +	int nfoo;
> > > +	struct foo *foo_array;
> > > +
> > > +	void do_something_locked(struct foo *fp)
> > > +	{
> > > +		bool gf = true;
> > > +
> > > +		/* IMPORTANT: Heuristic plus spin_lock()! */
> > > +		if (!data_race(global_flag)) {
> > > +			spin_lock(&fp->f_lock);
> > > +			if (!smp_load_acquire(&global_flag)) {
> > > +				do_something(fp);
> > > +				spin_unlock(&fp->f_lock);
> > > +				return;
> > > +			}
> > > +			spin_unlock(&fp->f_lock);
> > > +		}
> > > +		spin_lock(&global_lock);
> > > +		/* Lock held, thus global flag cannot change. */
> > > +		if (!global_flag) {
> > How can global_flag ever be true at this point?  The only line of code
> > that sets it is in begin_global() below, it only runs while global_lock
> > is held, and global_flag is set back to false before the lock is
> > released.
> 
> It can't be true. The code is a simplified version of the algorithm in
> ipc/sem.c.
> 
> For the ipc/sem.c, global_flag can remain true even after dropping
> global_lock.
> 
> When transferring the approach to nf_conntrack_core, I didn't notice that
> nf_conntrack doesn't need a persistent global_flag.
> 
> Thus the recheck after spin_lock(&global_lock) is not needed.

In fact, since global_flag is true if and only if global_lock is locked, 
perhaps it can be removed entirely and replaced with 
spin_is_locked(&global_lock).

> > > +			spin_lock(&fp->f_lock);
> > > +			spin_unlock(&global_lock);
> > > +			gf = false;
> > > +		}
> > > +		do_something(fp);
> > > +		if (fg)
> > Should be gf, not fg.
> > 
> > > +			spin_unlock(&global_lock);
> > > +		else
> > > +			spin_lock(&fp->f_lock);
> > > +	}
> > > +
> > > +	void begin_global(void)
> > > +	{
> > > +		int i;
> > > +
> > > +		spin_lock(&global_lock);
> > > +		WRITE_ONCE(global_flag, true);
> > Why does this need to be WRITE_ONCE?  It still races with the first read
> > of global_flag above.
> > 
> > > +		for (i = 0; i < nfoo; i++) {
> > > +			/* Wait for pre-existing local locks. */
> > > +			spin_lock(&fp->f_lock);
> > > +			spin_unlock(&fp->f_lock);
> > Why not acquire all the locks here and release all of them in
> > end_global()?  Then global_flag wouldn't need acquire-release
> > sychronization.
> 
> From my understanding:
> spin_lock contains preempt_count_add, thus you can't acquire more than 255
> spinlocks (actually 245, the warning limit is 10 below 255)

It might be worth mentioning this in a code comment.  Or in the 
accompanying text.

> > > +		}
> > > +	}
> > > +
> > > +	void end_global(void)
> > > +	{
> > > +		smp_store_release(&global_flag, false);
> > > +		/* Pre-existing global lock acquisitions will recheck. */
> > What does that comment mean?  How can there be any pre-existing global
> > lock acquisitions when we hold the lock right now?
> 
> > > +		spin_unlock(&global_lock);
> > > +	}
> > > +
> > > +All code paths leading from the do_something_locked() function's first
> > > +read from global_flag acquire a lock, so endless load fusing cannot
> > > +happen.
> > > +
> > > +If the value read from global_flag is true, then global_flag is rechecked
> > > +while holding global_lock, which prevents global_flag from changing.
> > > +If this recheck finds that global_flag is now false, the acquisition
> > Again, how can't global_flag be false now?
> > 
> > Did you originally have in mind some sort of scheme in which
> > begin_global() would release global_lock before returning and
> > end_global() would acquire global_lock before clearing global_flag?  But
> > I don't see how that could work without changes to do_something_locked().
> > 
> > > +of ->f_lock prior to the release of global_lock will result in any subsequent
> > > +begin_global() invocation waiting to acquire ->f_lock.
> > > +
> > > +On the other hand, if the value read from global_flag is false, then
> > > +global_flag, then rechecking under ->f_lock combined with synchronization
> > ---^^^^^^^^^^^^^^^^^^
> > 
> > Typo?
> > 
> > > +with begin_global() guarantees than any erroneous read will cause the
> > > +do_something_locked() function's first do_something() invocation to happen
> > > +before begin_global() returns.  The combination of the smp_load_acquire()
> > > +in do_something_locked() and the smp_store_release() in end_global()
> > > +guarantees that either the do_something_locked() function's first
> > > +do_something() invocation happens after the call to end_global() or that
> > > +do_something_locked() acquires global_lock() and rechecks under the lock.
> > This last sentence also makes no sense unless you imagine dropping
> > global_lock between begin_global() and end_global().
> 
> ipc/sem.c does that and needs that, nf_conntrack doesn't use this.

Given all these issues, it seems like this patch needs to be re-written.

Alan Stern

Powered by blists - more mailing lists