Re: [PATCH v11 06/10] io_uring: introduce attributes for read/write and PI support

[Pavel Begunkov <asml.silence@gmail.com>]

On 12/3/24 06:56, Anuj Gupta wrote:
> On Mon, Dec 02, 2024 at 09:13:14PM -0500, Martin K. Petersen wrote:
>>
>> I have things running on my end on top of Jens' tree (without error
>> injection, that's to come).
>>
>> One question, though: How am I to determine that the kernel supports
>> attr_ptr and IORING_RW_ATTR_FLAG_PI? Now that we no longer have separate
>> IORING_OP_{READ,WRITE}_META commands I can't use IO_URING_OP_SUPPORTED
>> to find out whether the running kernel supports PI passthrough.
> 
> Martin, right currently there is no way to probe whether the kernel
> supports read/write attributes or not.
> 
> Jens, Pavel how about introducing a new IO_URING_OP_* flag (something
> like IO_URING_OP_RW_ATTR_SUPPORTED) to probe whether read/write attributes
> are supported or not. Something like this [*]

An IORING_FEAT_ flag might be simpler for now. Or is it in plans to
somehow support IORING_OP_READ_MULTISHOT as well?

I have to say Adding a good probing infra is long overdue. For example a
user might want to know which attributes are supported. And beyond
io_uring it might be interesting to probe whether a particular file
supports it.


> [*]
> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
> index 38f0d6b10eaf..787a2df8037f 100644
> --- a/include/uapi/linux/io_uring.h
> +++ b/include/uapi/linux/io_uring.h
> @@ -723,6 +723,7 @@ struct io_uring_rsrc_update2 {
>   #define IORING_REGISTER_FILES_SKIP	(-2)
>   
>   #define IO_URING_OP_SUPPORTED	(1U << 0)
> +#define IO_URING_OP_RW_ATTR_SUPPORTED	(1U << 1)
>   
>   struct io_uring_probe_op {
>   	__u8 op;
> diff --git a/io_uring/opdef.c b/io_uring/opdef.c
> index 3de75eca1c92..64e1e5d48dec 100644
> --- a/io_uring/opdef.c
> +++ b/io_uring/opdef.c
> @@ -67,6 +67,7 @@ const struct io_issue_def io_issue_defs[] = {

io_cold_defs would be better

-- 
Pavel Begunkov