Re: [PATCH v2] mm/page_alloc: don't wake kswapd from rmqueue() unless __GFP_KSWAPD_RECLAIM is specified

["Huang, Ying" <ying.huang@intel.com>]

Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> writes:

> On 2023/05/15 16:38, Mel Gorman wrote:
>> On Sun, May 14, 2023 at 09:28:56AM +0900, Tetsuo Handa wrote:
>>> Commit 73444bc4d8f9 ("mm, page_alloc: do not wake kswapd with zone lock
>>> held") moved wakeup_kswapd() from steal_suitable_fallback() to rmqueue()
>>> using ZONE_BOOSTED_WATERMARK flag.
>>>
>>> Only allocation contexts that include ALLOC_KSWAPD (which corresponds to
>>> __GFP_KSWAPD_RECLAIM) should wake kswapd, for callers are supposed to
>>> remove __GFP_KSWAPD_RECLAIM if trying to hold pgdat->kswapd_wait has a
>>> risk of deadlock.
>> 
>> kswapd_wait is a waitqueue so what is being held? It's safe for kswapd
>> to try wake itself as the waitqueue will be active when wakeup_kswapd()
>> is called so no wakeup occurs. If there is a deadlock, it needs a better
>> explanation. I believe I already stated why this patch is fixing a bug
>> but it wasn't deadlock related.
>> 
>
> I noticed this problem ( pgdat->kswapd_wait might be held without
> __GFP_KSWAPD_RECLAIM ) when analyzing a different problem ( debugobject code
> is holding pgdat->kswapd_wait due to __GFP_KSWAPD_RECLAIM ) reported at
> https://syzkaller.appspot.com/bug?extid=fe0c72f0ccbb93786380 .
>
> I'm calling pgdat->kswapd_wait a lock, as lockdep uses it as a lock name.

This has confused me much before.  IIUC, the deadlock is unrelated with
kswapd wakeup itself.  pgdat->kswapd_wait is the lock name and the lock
in fact is the spinlock: pgdat->kswapd_wait.lock.  So the deadlock is,

pgdat->kswapd_wait.lock holders take the pi lock
pi lock holders take the rq lock
rq lock holders take the timer base lock
timer base lock holders take the pgdat->kswapd_wait.lock (missing __GFP_KSWAPD_RECLAIM check)

The above is based on analysis in

https://lore.kernel.org/all/20190107204627.GA25526@cmpxchg.org/
https://lore.kernel.org/linux-mm/d642e597-cf7d-b410-16ce-22dff483fd8e@I-love.SAKURA.ne.jp/

Tetsuo's patch avoids to take pgdat->kswapd_wait.lock when timer base
lock is held via adding check for __GFP_KSWAPD_RECLAIM, so breaks the
circular dependency chain.

> The latter was explained at
> https://lore.kernel.org/linux-mm/d642e597-cf7d-b410-16ce-22dff483fd8e@I-love.SAKURA.ne.jp/
> and we agreed that debugobject code needs to drop __GFP_KSWAPD_RECLAIM at
> https://lore.kernel.org/linux-mm/87fs809fwk.ffs@tglx/ .
>
> This patch is for making sure that debugobject code will not try to hold
> pgdat->kswapd_wait after debugobject code dropped __GFP_KSWAPD_RECLAIM.
>
> Thus, the problem this patch will fix is a deadlock related, isn't it?

Best Regards,
Huang, Ying