Re: [PATCH] fs/netfs: Remove redundant use of smp_rmb()

[Akira Yokosawa <akiyks@gmail.com>]

Hi David,

On Sat, 07 Dec 2024 08:04:56 +0000, David Howells wrote:
> Akira Yokosawa <akiyks@gmail.com> wrote:
> 
>> Are you sure removing the smp_rmb() is realy the right thing to do?
> 
> The wait_on_bit*() class functions, e.g.:
> 
> 	wait_on_bit(unsigned long *word, int bit, unsigned mode)
> 	{
> 		might_sleep();
> 		if (!test_bit_acquire(bit, word))
> 			return 0;
> 		return out_of_line_wait_on_bit(word, bit,
> 					       bit_wait,
> 					       mode);
> 	}
> 
> now unconditionally includes an appropriate barrier on the test_bit(), so the
> smp_rmb() should be unnecessary, though netfslib should probably be using
> clear_and_wake_up_bit().
> 

Thank you for clarifying.

> Probably we need to update the doc to reflect this.

Agreed.

I see that wait_on_bit()'s kernel-doc comment mentions implicit ACQUIRE
semantics on success, and that of wake_up_bit() mentions the need of care
for memory ordering before calling it.

Unfortunately, neither of those comments is included into kernel
documentation build (Sphinx) at the moment.

I'm going to prepare a patch for including them somewhere under the
core-api doc.

WRT memory-barriers.txt, I'm not sure I can update it properly.

David, may I ask you doing that part?

Thanks, Akira