linux-kernel - Re: [PATCH 0/3] Introduce a light-weight queue close feature

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]

Date:   Thu, 6 Sep 2018 20:57:01 +0800
From:   Ming Lei <ming.lei@...hat.com>
To:     "jianchao.wang" <jianchao.w.wang@...cle.com>
Cc:     axboe@...nel.dk, linux-block@...r.kernel.org, jsmart2021@...il.com,
        sagi@...mberg.me, linux-kernel@...r.kernel.org,
        linux-nvme@...ts.infradead.org, keith.busch@...el.com,
        jthumshirn@...e.de, bart.vanassche@....com
Subject: Re: [PATCH 0/3] Introduce a light-weight queue close feature

On Thu, Sep 06, 2018 at 09:51:43AM +0800, jianchao.wang wrote:
> Hi Ming
> 
> On 09/06/2018 05:27 AM, Ming Lei wrote:
> > On Wed, Sep 05, 2018 at 12:09:43PM +0800, Jianchao Wang wrote:
> >> Dear all
> >>
> >> As we know, queue freeze is used to stop new IO comming in and drain
> >> the request queue. And the draining queue here is necessary, because
> >> queue freeze kills the percpu-ref q_usage_counter and need to drain
> >> the q_usage_counter before switch it back to percpu mode. This could
> >> be a trouble when we just want to prevent new IO.
> >>
> >> In nvme-pci, nvme_dev_disable freezes queues to prevent new IO.
> >> nvme_reset_work will unfreeze and wait to drain the queues. However,
> >> if IO timeout at the moment, no body could do recovery as nvme_reset_work
> >> is waiting. We will encounter IO hang.
> > 
> > As we discussed this nvme time issue before, I have pointed out that
> > this is because of blk_mq_unfreeze_queue()'s limit which requires that
> > unfreeze can only be done when this queue ref counter drops to zero.
> > 
> > For this nvme timeout case, we may relax the limit, for example,
> > introducing another API of blk_freeze_queue_stop() as counter-pair of
> > blk_freeze_queue_start(), and simply switch the percpu-ref to percpu mode
> > from atomic mode inside the new API.
> 
> Looks like we cannot switch a percpu-ref to percpu mode directly w/o drain it.
> Some references maybe lost.
> 
> static void __percpu_ref_switch_to_percpu(struct percpu_ref *ref)
> {
> 	unsigned long __percpu *percpu_count = percpu_count_ptr(ref);
> 	int cpu;
> 
> 	BUG_ON(!percpu_count);
> 
> 	if (!(ref->percpu_count_ptr & __PERCPU_REF_ATOMIC))
> 		return;
> 
> 	atomic_long_add(PERCPU_COUNT_BIAS, &ref->count);
> 
> 	/*
> 	 * Restore per-cpu operation.  smp_store_release() is paired
> 	 * with READ_ONCE() in __ref_is_percpu() and guarantees that the
> 	 * zeroing is visible to all percpu accesses which can see the
> 	 * following __PERCPU_REF_ATOMIC clearing.
> 	 */
> 	for_each_possible_cpu(cpu)
> 		*per_cpu_ptr(percpu_count, cpu) = 0;
> 
> 	smp_store_release(&ref->percpu_count_ptr,
> 			  ref->percpu_count_ptr & ~__PERCPU_REF_ATOMIC);

Before REF_ATOMIC is cleared, all counting is done on the atomic type
of &ref->count, and it is easy to keep the reference counter at
ATOMIC mode. Also the reference counter can only be READ at atomic mode.

So could you explain a bit how the lost may happen? And it is lost at
atomic mode or percpu mode?

> }
> 
> > 
> >>
> >> So introduce a light-weight queue close feature in this patch set
> >> which could prevent new IO and needn't drain the queue.
> > 
> > Frankly speaking, IMO, it may not be an good idea to mess up the fast path
> > just for handling the extremely unusual timeout event. The same is true
> > for doing the preemp only stuff, as you saw I have posted patchset for
> > killing it.
> > 
> 
> In normal case, it is just a judgment like 
> 
> 	if (unlikely(READ_ONCE(q->queue_gate))
> 
> It should not be a big deal.

Adding this stuff in fast path is quite difficult to verify its correctness
because it is really lockless, or even barrier-less.

Not to mention, READ_ONCE() implies one barrier of smp_read_barrier_depends().

Thanks,
Ming