From: Vitaliy Makkoveev <mvs@openbsd.org>
Subject: Re: unlock tcp drop sysctl
To: Alexander Bluhm <alexander.bluhm@gmx.net>
Cc: OpenBSD tech <tech@openbsd.org>
Date: Thu, 9 Jan 2025 23:43:35 +0300

> On 9 Jan 2025, at 21:59, Alexander Bluhm <alexander.bluhm@gmx.net> wrote:
> 
> On Thu, Jan 09, 2025 at 07:00:42PM +0300, Vitaliy Makkoveev wrote:
>> Would you like to split tcp_sysctl() with tcp_sysctl_locked() and
>> start slightly moving some paths out of sysctl lock? 
> 
> If I understand correctly, PR_MPSYSCTL prevents kernel lock, does
> not serialize sysctl calls, and does not wire the memory for copyin().
> 

Yes, PR_MPSYSCTL exists for that purpose.

> TCP sysctl does not care about kernel lock.  It uses either atomic
> integers, mutex protection or exclusive net lock.  Calling copyin()
> or copyout() without wired memory while holding the net lock is bad
> due to NFS.
> 

Memory wiring requires kernel lock, because uvm_vslock() acquires
kernel lock with mutex held deep within, but the kernel lock taken
before mutex acquisition hides this locking issue. I triggered this
long time ago during sys_sysctl() unlocking, may be this is not
actual anymore.

So, yes we wire memory to prevent possible context switch during
copy{in,out}(), but uvm_vslock() requires complicated locking around.

> I think is easier to remove net lock sysctl per sysctl and then
> switch the whole tcp_sysctl() instead of introducing tcp_sysctl_locked().
> 
> bluhm