linux-kernel - Re: [PATCH v2] locking/rwbase: Prevent indefinite writer starvation

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20230119174101.rddtxk5xlamlnquh@techsingularity.net>
Date:   Thu, 19 Jan 2023 17:41:01 +0000
From:   Mel Gorman <mgorman@...hsingularity.net>
To:     Sebastian Andrzej Siewior <bigeasy@...utronix.de>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Thomas Gleixner <tglx@...utronix.de>,
        Ingo Molnar <mingo@...nel.org>,
        Davidlohr Bueso <dave@...olabs.net>,
        Linux-RT <linux-rt-users@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2] locking/rwbase: Prevent indefinite writer starvation

On Thu, Jan 19, 2023 at 05:28:48PM +0100, Sebastian Andrzej Siewior wrote:
> On 2023-01-19 11:02:20 [+0000], Mel Gorman wrote:
> > > - Once the writer removes READER_BIAS, it forces the reader into the
> > >   slowpath.
> > 
> > Removed in __rwbase_write_trylock IIUC
> 
> And added back in case try trylock failed via __rwbase_write_unlock().
> The RTmutex is unlocked and the READER_BIAS is "returned".
> 

Indeed.

> > >   At that time the writer does not own the wait_lock meaning
> > >   the reader _could_ check the timeout before writer had a chance to set
> > >   it. The worst thing is probably that if jiffies does not have the
> > >   highest bit set then it will always disable the reader bias here.
> > >   The easiest thing is probably to check timeout vs 0 and ensure on the
> > >   writer side that the lowest bit is always set (in the unlikely case it
> > >   will end up as zero).
> > > 
> > 
> > I am missing something important. On the read side, we have
> > 
> 
> Look at this side by side:
> 
>                 writer                                                       reader
> 
> | static int __sched rwbase_write_lock(struct rwbase_rt *rwb,
> |                                      unsigned int state)
> | {
> |         /* Force readers into slow path */
> |         atomic_sub(READER_BIAS, &rwb->readers);
> 
> 
> |                                                               static int __sched __rwbase_read_lock(struct rwbase_rt *rwb,
> |                                                                                                     unsigned int state)
> |                                                               {       
> |                                                                       struct rt_mutex_base *rtm = &rwb->rtmutex;
> |                                                                       int ret;                         
> |                                                               
> |                                                                       raw_spin_lock_irq(&rtm->wait_lock);
> 
> Reader has the lock, writer will wait.
> 
> |                                                                       /*
> |                                                                        * Allow readers, as long as the writer has not completely
> |                                                                        * acquired the semaphore for write.
> |                                                                        */
> |                                                                       if (atomic_read(&rwb->readers) != WRITER_BIAS) {
> 
> here, the timeout value is not yet populated by the writer so the reader
> compares vs 0.
> 
> |                                                                               atomic_inc(&rwb->readers);
> |                                                                               raw_spin_unlock_irq(&rtm->wait_lock);
> |                                                                               return 0;
> |                                                                       }
> |                                                              
> 
> |         raw_spin_lock_irqsave(&rtm->wait_lock, flags);
> |         if (__rwbase_write_trylock(rwb))
> |                 goto out_unlock;
> |
> 
> Hope this makes it easier.
> 

Yes, it makes your concern much clearer but I'm not sure it actually matters
in terms of preventing write starvation or in terms of correctness. At
worst, a writer is blocked that could have acquired the lock during a tiny
race but that's a timing issue rather than a correctness issue.

Lets say the race hits

									reader sees waiter_timeout == 0
	writer acquires wait_lock
	__rwbase_write_trylock fails
	update waiter_timeout
	rwbase_schedule

Each reader that hits the race goes ahead at a point in time but anything
readers after that observe the timeout and eventually the writer goes ahead.

If the waiter_timeout was updated before atomic_sub(READER_BIAS),
it doesn't close the race as atomic_sub is unordered so barriers would
also be needed and clearing of waiter_timeout moves to out_unlock in case
__rwbase_write_trylock succeeds. That's possible but the need for barriers
makes it more complicated than is necessary.

The race could be closed by moving wait_lock acquisition before the
atomic_sub in rwbase_write_lock() but it expands the scope of the wait_lock
and I'm not sure that's necessary for either correctness or preventing
writer starvation. It's a more straight-forward fix but expanding the
scope of a lock unnecessarily has been unpopular in the past.

I think we can close the race that concerns you but I'm not convinced we
need to and changing the scope of wait_lock would need a big comment and
probably deserves a separate patch.

Sorry if I'm still missing something stupid and thanks for your patience
reviewing this.

-- 
Mel Gorman
SUSE Labs