pget: don't acquire allproc_lock when operating on curproc

[kaktus]

Summary

pget(9) in it's current form unconditionally acquire allproc_lock every time.
Test showed that in many cases requested pid == curproc->p_pid allowing this lock to be avoided:

make -j4 buildworld

saved about 5600 locks;

poudriere bulk -J 12 -C -t -j 111amd64 www/nginx-full

where this ports depends on 230 others (java, llvm40, etc..) saved about 27000 locks.

Test Plan

This patch was tested using varied mix of make buildworld, poudriere, ps, proctstat.

[kib]

You may check for curthread->td_tid as well.

[mjg]

I think checking if pfind may use a similar treatment would be nice.

[kaktus]

Address @kib comment and check for tid too.

[kaktus]

I'll check pfind and create new review if needed.

[kib]

sys/kern/kern_proc.c

393 ↗

(On Diff #34717)

Look below for PGET_NOTID. The check for tid should be conditional.

[mjg]

I would say scratch optimising for the _NOTID case. There is only one in-tree user of the flag and it is not expected to receive any tids.

[kaktus]

Shouldn't then CPUCLOCK_WHICH_PID case in kern_clock_getcpuclockid2 be amended?

[mjg]

As the name suggests CPUCLOCK_WHICH_PID only wants to match pids. Thus the _NOTID flag is passed to avoid the corner case of *tid* being passed and matching an existing thread.

p_pid wil never match a *tid* thus it can be tested unconditionally.

I don't see any use for extra treatment for the _NOTID flag.

[kaktus]

Roll back to the original version and optimise only for the pid == curproc->p_pid case.