Re: [PATCH] io_uring: introduce inline reqs for IORING_SETUP_IOPOLL & direct_io

["jianchao.wang" <jianchao.w.wang@oracle.com>]

Hi Jens

On 4/2/19 11:47 AM, Jens Axboe wrote:
> On 4/1/19 9:10 PM, Jianchao Wang wrote:
>> For the IORING_SETUP_IOPOLL & direct_io case, all of the submission
>> and completion are handled under ctx->uring_lock or in SQ poll thread
>> context, so io_get_req and io_put_req has been serialized well.
>>
>> Based on this, we introduce the preallocated reqs ring per ctx and
>> needn't to provide any lock to serialize the updating of the head
>> and tail. The performacne benefits from this. The test result of
>> following fio command
>>
>> fio --name=io_uring_test --ioengine=io_uring --hipri --fixedbufs
>> --iodepth=16 --direct=1 --numjobs=1 --filename=/dev/nvme0n1 --bs=4k
>> --group_reporting --runtime=10
>>
>> shows IOPS upgrade from 197K to 206K.
> 
> I like this idea, but not a fan of the execution of it. See below.
> 
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 6aaa3058..40837e4 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -104,11 +104,17 @@ struct async_list {
>>  	size_t			io_pages;
>>  };
>>  
>> +#define INLINE_REQS_TOTAL 128
>> +
>>  struct io_ring_ctx {
>>  	struct {
>>  		struct percpu_ref	refs;
>>  	} ____cacheline_aligned_in_smp;
>>  
>> +	struct io_kiocb *inline_reqs[INLINE_REQS_TOTAL];
>> +	struct io_kiocb *inline_req_array;
>> +	unsigned long inline_reqs_h, inline_reqs_t;
> 
> Why not just use a list? The req has a list member anyway. Then you don't
> need a huge array, just a count.

Yes, indeed.

> 
>> +
>>  	struct {
>>  		unsigned int		flags;
>>  		bool			compat;
>> @@ -183,7 +189,9 @@ struct io_ring_ctx {
>>  
>>  struct sqe_submit {
>>  	const struct io_uring_sqe	*sqe;
>> +	struct file 			*file;
>>  	unsigned short			index;
>> +	bool 				is_fixed;
>>  	bool				has_user;
>>  	bool				needs_lock;
>>  	bool				needs_fixed_file;
> 
> Not sure why you're moving these to the sqe_submit?

Just want to get the file before io_get_req to know whether it is direct_io.
This is unnecessary if eliminate the direct io limitation.

> 
>> @@ -228,7 +236,7 @@ struct io_kiocb {
>>  #define REQ_F_PREPPED		16	/* prep already done */
>>  	u64			user_data;
>>  	u64			error;
>> -
>> +	bool 			ctx_inline;
>>  	struct work_struct	work;
>>  };
> 
> ctx_inline should just be a req flag.

Yes.

> 
>>  
>> @@ -397,7 +405,8 @@ static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
>>  }
>>  
>>  static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
>> -				   struct io_submit_state *state)
>> +				   struct io_submit_state *state,
>> +				   bool direct_io)
>>  {
>>  	gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
>>  	struct io_kiocb *req;
>> @@ -405,10 +414,19 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
>>  	if (!percpu_ref_tryget(&ctx->refs))
>>  		return NULL;
>>  
>> -	if (!state) {
>> +	/*
>> +	 * Avoid race with workqueue context that handle buffered IO.
>> +	 */
>> +	if (direct_io &&
>> +	    ctx->inline_reqs_h - ctx->inline_reqs_t < INLINE_REQS_TOTAL) {
>> +	    req = ctx->inline_reqs[ctx->inline_reqs_h % INLINE_REQS_TOTAL];
>> +	    ctx->inline_reqs_h++;
>> +	    req->ctx_inline = true;
>> +	} else if (!state) {
> 
> What happens for O_DIRECT that ends up being punted to async context?

I misunderstand that only buffered io would be punted to async workqueue context.

> We need a clearer indication of whether or not we're under the lock or
> not, and then get rid of the direct_io "limitation" for this. Arguably,
> cached buffered IO needs this even more than O_DIRECT does, since that
> is much faster.
> 

Before punt the IO to async workqueue context, a sqe_copy will be allocated.
How about allocating a structure with both a sqe and a io_kiocb ?
Then use the newly allocated io_kiocb to replace the preallocated io_kiocb and
release the latter one. Then we could eliminate the wrong direct_io limitation.

Thanks
Jianchao