Re: [RESEND PATCH V4] pidns: introduce syscall translate_pid

[Konstantin Khlebnikov]

On 04.04.2018 00:51, Nagarathnam Muthusamy wrote:
> On 04/03/2018 02:52 PM, Andrew Morton wrote:
> > On Tue, 3 Apr 2018 14:45:28 -0700 Nagarathnam Muthusamy <nagarathnam.muthusamy@xxxxxxxxxx> wrote:
> >
> > > > This changelog doesn't explain what the value is to our users. I
> > > > assume it is a performance optimization because "backward translation
> > > > requires scanning all tasks"? If so, please show us real-world
> > > > examples of the performance benefit from this patch, and please go to
> > > > great lengths to explain to us why this optimisation is needed by our
> > > > users.
> > > One of the usecase by Oracle database involves multiple levels of
> > > nested pid namespaces and we require pid translation between the
> > > levels. Discussions on the particular usecase, why any of the existing
> > > methods was not usable happened in the following thread.
> > >
> > > https://patchwork.kernel.org/patch/10276785/
> > >
> > > At the end, it was agreed that this patch along with flocks will solve the
> > > issue.
> > Nobody who reads this patch's changelog will know any of this. Please
> > let's get all this information into the proper place.
> Sure! Will resend the patch with updated change log.

I have v5 version of this proposal in work.

I've redesigned interface to be more convenient for cases where
strict race-protection isn't required and pid-ns could be referenced pid.

It has 5 arguments rather than 3 because types of references are
defined explicitly rather than magic like -1, >0, <0.
This more verbose but protects against errors like passing -1 from
failed previous syscall as argument.

kind of
translate_pid(pid, TRANSLATE_PID_FD_PIDNS, ns_fd, TRANSLATE_PID_CURRENT_PIDNS, 0)

I'll send it today with including more detailed motivation for patch.