Re: Buffered IO async context overhead

[Jens Axboe <axboe@kernel.dk>]

On 2/14/20 1:31 PM, Andres Freund wrote:
> Hi,
> 
> On 2020-02-14 13:13:35 -0700, Jens Axboe wrote:
>> On 2/14/20 12:50 PM, Andres Freund wrote:
>>> which I think is pretty clear evidence we're hitting fairly significant
>>> contention on the queue lock.
>>>
>>>
>>> I am hitting this in postgres originally, not fio, but I thought it's
>>> easier to reproduce this way.  There's obviously benefit to doing things
>>> in the background - but it requires odd logic around deciding when to
>>> use io_uring, and when not.
>>>
>>> To be clear, none of this happens with DIO, but I don't forsee switching
>>> to DIO for all IO by default ever (too high demands on accurate
>>> configuration).
>>
>> Can you try with this added?
>>
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 76cbf474c184..207daf83f209 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -620,6 +620,7 @@ static const struct io_op_def io_op_defs[] = {
>>  		.async_ctx		= 1,
>>  		.needs_mm		= 1,
>>  		.needs_file		= 1,
>> +		.hash_reg_file		= 1,
>>  		.unbound_nonreg_file	= 1,
>>  	},
>>  	[IORING_OP_WRITEV] = {
>> @@ -634,6 +635,7 @@ static const struct io_op_def io_op_defs[] = {
>>  	},
>>  	[IORING_OP_READ_FIXED] = {
>>  		.needs_file		= 1,
>> +		.hash_reg_file		= 1,
>>  		.unbound_nonreg_file	= 1,
>>  	},
>>  	[IORING_OP_WRITE_FIXED] = {
>> @@ -711,11 +713,13 @@ static const struct io_op_def io_op_defs[] = {
>>  	[IORING_OP_READ] = {
>>  		.needs_mm		= 1,
>>  		.needs_file		= 1,
>> +		.hash_reg_file		= 1,
>>  		.unbound_nonreg_file	= 1,
>>  	},
>>  	[IORING_OP_WRITE] = {
>>  		.needs_mm		= 1,
>>  		.needs_file		= 1,
>> +		.hash_reg_file		= 1,
>>  		.unbound_nonreg_file	= 1,
>>  	},
>>  	[IORING_OP_FADVISE] = {
>> @@ -955,7 +959,7 @@ static inline bool io_prep_async_work(struct io_kiocb *req,
>>  	bool do_hashed = false;
>>  
>>  	if (req->flags & REQ_F_ISREG) {
>> -		if (def->hash_reg_file)
>> +		if (!(req->kiocb->ki_flags & IOCB_DIRECT) && def->hash_reg_file)
>>  			do_hashed = true;
>>  	} else {
>>  		if (def->unbound_nonreg_file)
> 
> I can (will do Sunday, on the road till then). But I'm a bit doubtful
> it'll help. This is using WRITEV after all, and I only see a single
> worker?

Because I'm working on other items, I didn't read carefully enough. Yes
this won't change the situation for writes. I'll take a look at this when
I get time, maybe there's something we can do to improve the situation.

-- 
Jens Axboe