OpenBSD cvs log

created 2025-02-05T23:41:34Z
begin 2025-01-13T17:50:54Z
end 2025-01-13T18:10:20Z
path src/sys
commits 2

date 2025-01-13T18:09:24Z
author mvs
files src/sys/kern/kern_sysctl.c log diff annotate
message Rework the INP tables walkthrough in the KERN_FILE_BYFILE case of
the sysctl_file().

I don't like how KERN_FILE_BYFILE case of the sysctl_file() delivers
sockets data to the userland. It not only takes exclusive netlock
around all except divert socket tables walkthrough, but also does
copyout() with mutex(9) held. Sounds strange, but the context switch
is avoided because userland pages are wired.

The table is protected with `inpt_mtx' mutex(9), so the socket lock or
netlock should be take outside. Since we have no socket pointer, we
can't use solock(). We can't use the only shared netlock because this
left socket unprotected against concurrent threads which rely on
solock_shared().

Now it is possible to rework sysctl_file(). We can use reference
counting for all socket types, and bluhm@ introduced `inp_sofree_mtx'
mutex(9) to protect `inp_socket'. The INP tables have the special
iterator to safely release `inpt_mtx' during table walkthrough.

The FILLSO() or FILLIT2() macros can't be used unrolled, because we
need to push mutexes re-locking and solock() deep within. Introduce
the FILLINPTABLE() macro which is the unrolling, but with the socket
related locking dances. FILLIT2() macro is not required anymore and
was merged with FILLIT().

Current implementation takes the reference on `inp_socket' and
releases `inpt_mtx' mutex(9). Now it's possible to fairly use
shared_solock() on socket instead of netlock while performing
fill_file(). The copyout() is external for fill_file() and touches
nothing required to be protected so it could be made lockless.

The KERN_FILE_BYFILE case became mp-safe, but the rest sysctl_file()
cases still not, so the sysctl_vslock() dances left as is.

ok bluhm

date 2025-01-13T18:10:20Z
author mvs
files src/sys/kern/uipc_socket.c log diff annotate
message Unlock the tcp(4) case of somove().

Tested and OK bluhm@.