Re: [PATCH] kcmp: add separate Kconfig symbol for kcmp syscall

[Rasmus Villemoes <linux@rasmusvillemoes.dk>]

On 10/07/2020 17.57, Matthew Wilcox wrote:
> On Fri, Jul 10, 2020 at 09:56:31AM +0200, Rasmus Villemoes wrote:
>> The ability to check open file descriptions for equality (without
>> resorting to unreliable fstat() and fcntl(F_GETFL) comparisons) can be
>> useful outside of the checkpoint/restore use case - for example,
>> systemd uses kcmp() to deduplicate the per-service file descriptor
>> store.
>>
>> Make it possible to have the kcmp() syscall without the full
>> CONFIG_CHECKPOINT_RESTORE.
> 
> If systemd is using it, is it even worth making it conditional any more?
> Maybe for CONFIG_EXPERT builds, it could be de-selectable.
> 

[hm, I dropped the ball, sorry for the necromancy]

Well, first, I don't want to change any defaults here, if that is to be
done, it should be a separate patch.

Second, yes, systemd uses it for the de-duplication, and for that reason
recommends CONFIG_CHECKPOINT_RESTORE (at least, according to their
README) - but I'm not aware of any daemons that actually make use of
systemd's file descriptor store, so it's not really something essential
to every systemd-based system out there. It would be nice if systemd
could change its recommendation to just CONFIG_KCMP_SYSCALL.

But it's also useful for others, e.g. I have some code that wants to
temporarily replace stdin/stdout/stderr with some other file
descriptors, but needs to preserve the '0 is a dup of 1, or not' state
(i.e., is the same struct file) - that cannot reliably be determined
from fstat()/lseek(SEEK_CUR)/F_GETFL or whatever else one could throw at
an fd.

Rasmus