Re: [PATCH v2] mm: slab, with same context require fs_reclaim lock

[Vlastimil Babka]

On 10/12/22 13:23, Steven Rostedt wrote:
> On Tue, 27 Sep 2022 15:11:34 +0800
> eadavis@xxxxxxxx wrote:
>> --- a/mm/slab.c
>> +++ b/mm/slab.c
>> @@ -3275,15 +3275,19 @@ slab_alloc(struct kmem_cache *cachep, struct list_lru *lru, gfp_t flags,
>>  	bool init = false;
>>  
>>  	flags &= gfp_allowed_mask;
>> +	local_irq_save(save_flags);
> 
> Please do not do this. Open coding interrupt disabling due to locking
> issues is not the solution. You need to make the locks themselves
> disable interrupts if need be. This breaks PREEMPT_RT, and creates a
> "big kernel lock" situation where there's random interrupts being
> disabled for no apparent reason.
> 
> -- Steve

And FWIW the problem was in the calling context of io_uring, not slab code
itself in the first place, see:

https://lore.kernel.org/all/20220929135627.ykivmdks2w5vzrwg@quack3/