On 01/05/2018 08:19 AM, Kamil Dudka wrote:

> I am only fixing the case where the destination file is created after the
> lstat() call.  In that case, the only result is ENOENT, which is harmless.
>

Ah, you're right. Sorry, I misread your patch. It should work.

> On Friday, January 5, 2018 4:29:55 PM CET Pádraig Brady wrote:
>
>> Paul's also avoids a stat() in the common case
>> where the initial renameat2() succeeds.
> At the cost of _not_ avoiding the renameat2() call in the most common case.

I expect that the most common case is 'mv A B' where B does not already 
exist. This is the case that avoids the stat of B.

Come to think of it, we don't need to stat A either, when the initial 
renameat2 succeeds. Attached is a revised proposed patchset to do that. 
The first is the same as before; the second causes 'mv A B' to issue 
just a renameat2 syscall (with no calls to stat) in the common case 
where A exists and B does not.

> I will go with my conservative (or even the
> documentation-only) patch for RHEL-7, which was the trigger of this effort,
> because Paul's patch comes with changes of behavior observable beyond the
Sounds good.