Re: [PATCH linux-next] mm: swap: get rid of deadloop in swapin readahead

[Hugh Dickins <hughd@google.com>]

On Fri, 25 Feb 2022, Andrew Morton wrote:
> On Mon, 21 Feb 2022 11:17:49 +0000 cgel.zte@gmail.com wrote:
> > From: Guo Ziliang <guo.ziliang@zte.com.cn>
> > 
> > In our testing, a deadloop task was found. Through sysrq printing, same 
> > stack was found every time, as follows:
> > __swap_duplicate+0x58/0x1a0
> > swapcache_prepare+0x24/0x30
> > __read_swap_cache_async+0xac/0x220
> > read_swap_cache_async+0x58/0xa0
> > swapin_readahead+0x24c/0x628
> > do_swap_page+0x374/0x8a0
> > __handle_mm_fault+0x598/0xd60
> > handle_mm_fault+0x114/0x200
> > do_page_fault+0x148/0x4d0
> > do_translation_fault+0xb0/0xd4
> > do_mem_abort+0x50/0xb0
> > 
> > The reason for the deadloop is that swapcache_prepare() always returns
> > EEXIST, indicating that SWAP_HAS_CACHE has not been cleared, so that
> > it cannot jump out of the loop. We suspect that the task that clears
> > the SWAP_HAS_CACHE flag never gets a chance to run. We try to lower
> > the priority of the task stuck in a deadloop so that the task that
> > clears the SWAP_HAS_CACHE flag will run. The results show that the
> > system returns to normal after the priority is lowered.
> > 
> > In our testing, multiple real-time tasks are bound to the same core,
> > and the task in the deadloop is the highest priority task of the
> > core, so the deadloop task cannot be preempted.
> > 
> > Although cond_resched() is used by __read_swap_cache_async, it is an
> > empty function in the preemptive system and cannot achieve the purpose
> > of releasing the CPU. A high-priority task cannot release the CPU
> > unless preempted by a higher-priority task. But when this task
> > is already the highest priority task on this core, other tasks
> > will not be able to be scheduled. So we think we should replace
> > cond_resched() with schedule_timeout_uninterruptible(1),
> > schedule_timeout_interruptible will call set_current_state
> > first to set the task state, so the task will be removed
> > from the running queue, so as to achieve the purpose of
> > giving up the CPU and prevent it from running in kernel
> > mode for too long.
> > 
> > ...
> >
> > --- a/mm/swap_state.c
> > +++ b/mm/swap_state.c
> > @@ -478,7 +478,7 @@ struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
> >  		 * __read_swap_cache_async(), which has set SWAP_HAS_CACHE
> >  		 * in swap_map, but not yet added its page to swap cache.
> >  		 */
> > -		cond_resched();
> > +		schedule_timeout_uninterruptible(1);
> >  	}
> >  
> >  	/*
> 
> Sigh.  I guess yes, we should do this, at least in a short-term,
> backportable-to-stable way.
> 
> But busy-waiting while hoping that someone else will save us isn't an
> attractive design.  Hugh, have you ever thought about something more
> deterministic in there?

Not something more deterministic, no: I think that would entail
heavier locking, perhaps slowing down hotter paths, just to avoid
this swap oddity.

This loop was written long before there was a preemptive kernel:
it was appropriate then, and almost never needed more than one retry
to complete; but preemption changed the story without us realizing.

Sigh here too.  I commend the thread on it from July 2018:
https://lore.kernel.org/linux-mm/2018072514403228778860@wingtech.com/

There the 4.9-stable user proposed preempt_disable(), I agreed but
found the patch provided insufficient, and offered another 4.9 patch
further down the thread.  Your preference at the time was msleep(1).

I was working on a similar patch for 4.18, but have not completed it
yet ;) and don't remember how satisfied or not I was with that one;
and wonder if I'm any more likely to get it finished by 2026.  It's
clear that I put much more thought into it back then than just now.

Maybe someone else would have a go: my 4.9 patch in that thread
shows most of it, but might need a lot of work to update to 5.17.

And it also gathered some temporary debug stats on how often this
happens: I'm not conscious of using RT at all, but was disturbed to see
how long an ordinary preemptive kernel was sometimes spinning there.
So I think I agree with you more than Michal on that: RT just makes
the bad behaviour more obvious.

Hugh