Re: [PATCH 3/5] io_uring: add support for getdents

[Hao Xu <hao.xu@linux.dev>]

On 8/8/23 12:34, Hao Xu wrote:
> On 8/1/23 08:28, Jens Axboe wrote:
>> On 7/31/23 9:26?AM, Darrick J. Wong wrote:
>>> I've watched quite a bit of NOWAIT whackamole going on over the past 
>>> few
>>> years (i_rwsem, the ILOCK, the IO layer, memory allocations...).  IIRC
>>> these filesystem ios all have to run in process context, right?  If so,
>>> why don't we capture the NOWAIT state in a PF flag?  We already do that
>>> for NOFS/NOIO memory allocations to make sure that /all/ reclaim
>>> attempts cannot recurse into the fs/io stacks.
>>
>> I would greatly prefer passing down the context rather than capitulating
>> and adding a task_struct flag for this. I think it _kind of_ makes sense
>> for things like allocations, as you cannot easily track that all the way
>> down, but it's a really ugly solution. It certainly creates more churn
>> passing it down, but it also reveals the parts that need to check it.
>> WHen new code is added, it's much more likely you'll spot the fact that
>> there's passed in context. For allocation, you end up in the allocator
>> anyway, which can augment the gfp mask with whatever is set in the task.
>> The same is not true for locking and other bits, as they don't return a
>> value to begin with. When we know they are sane, we can flag the fs as
>> supporting it (like we've done for async buffered reads, for example).
>>
>> It's also not an absolute thing, like memory allocations are. It's
>> perfectly fine to grab a mutex under NOWAIT issue. What you should not
>
> Hi Jens,
> To make sure, I'd like to ask, for memory allocation, GFP_NOIO semantics
> is all we need in NOWAIT issue, GFP_NOWAIT is not necessary, do I
> understand it right?
>
> Thanks,
> Hao


Trying to find a lock in mem allocation process that GFP_NOIO holds it while

other normal GFP_* like GFP_KERNEL also holds it and does IO.

Failed to find one such lock.  So I guess though GFP_NOIO may cause sleep

but won't wait on IO.