Re: [PATCH 2/2] io_uring/kbuf: fix infinite loop in io_kbuf_inc_commit()

[Jens Axboe <axboe@kernel.dk>]

On 8/27/25 3:59 PM, Keith Busch wrote:
> On Wed, Aug 27, 2025 at 03:45:28PM -0600, Jens Axboe wrote:
>>> +		buf_len = READ_ONCE(buf->len);
>>> +		this_len = min_t(int, len, buf_len);
>>> +		buf_len -= this_len;
>>> +		if (buf_len) {
>>>  			buf->addr += this_len;
>>> +			buf->len = buf_len;
>>>  			return false;
>>>  		}
>>> +		buf->len = 0;
> 
> Purely for symmetry, assigning buf->len ought to be a WRITE_ONCE.

I did think about that, perhaps I should've mentioned it in the commit
message. While the reader side is important for the reasons stated, the
updating of buf->len isn't really as only the serialized kernel side
will do it. Hence the WRITE_ONCE() should not be needed on the write
side, outside of perhaps documenting that this is a shared buffer.

-- 
Jens Axboe