On Wed, Oct 12, 2016 at 10:18:18AM +0200, Julien Charbon wrote:

> 
>  Hi Slawa,
> 
> On 10/11/16 2:11 PM, Slawa Olhovchenkov wrote:
> > On Tue, Oct 11, 2016 at 09:20:17AM +0200, Julien Charbon wrote:
> >>  Then threads are competing for the INP_WLOCK lock.  For the example,
> >> let's say the thread A wants to run tcp_input()/in_pcblookup_mbuf() and
> >> racing for this INP_WLOCK:
> >>
> >> https://github.com/freebsd/freebsd/blob/release/11.0.0/sys/netinet/in_pcb.c#L1964
> >>
> >>  And thread B wants to run tcp_timer_2msl()/tcp_close()/in_pcbdrop() and
> >> racing for this INP_WLOCK:
> >>
> >> https://github.com/freebsd/freebsd/blob/release/11.0.0/sys/netinet/tcp_timer.c#L323
> >>
> >>  That leads to two cases:
> >>
> >>  o Thread A wins the race:
> >>
> >>   Thread A will continue tcp_input() as usal and INP_DROPPED flags is
> >> not set and inp is still in TCP hash table.
> >>   Thread B is waiting on thread A to release INP_WLOCK after finishing
> >> tcp_input() processing, and thread B will continue
> >> tcp_timer_2msl()/tcp_close()/in_pcbdrop() processing.
> >>
> >>  o Thread B wins the race:
> >>
> >>   Thread B runs tcp_timer_2msl()/tcp_close()/in_pcbdrop() and inp
> >> INP_DROPPED is set and inp being removed from TCP hash table.
> >>   In parallel, thread A has found the inp in TCP hash before is was
> >> removed, and waiting on the found inp INP_WLOCK lock.
> >>   Once thread B has released the INP_WLOCK lock, thread A gets this lock
> >> and sees the INP_DROPPED flag and do "goto findpcb" but here because the
> >> inp is not more in TCP hash table and it will not be find again by
> >> in_pcblookup_mbuf().
> >>
> >>  Hopefully I am clear enough here.
> > 
> > Thanks, very clear.
> > Small qeustion: when both thread run on same CPU core, INP_WLOCK will
> > be re-schedule?
> 
>  Hmm, a thread can re-scheduled but not a lock. Thus no sure I
> understand your question here.  :)

I am don't know how work INP_WLOCK in this case (all on same cpu):

thread1: INP_WLOCK
-interrupt--
thread2: INP_WLOCK

if INP_WLOCK is like spinlock -- this is dead lock.
if INP_WLOCK is like mutex -- thread1 resheduled.

> > As I remeber race created by call tcp_twstart() at time of end
> > tcp_close(), at path sofree()-tcp_usr_detach() and unexpected
> > INP_TIMEWAIT state in the tcp_usr_detach(). INP_TIMEWAIT set in 
> > tcp_twstart()
> 
>  Exactly, thus the current fix is:  If you already have the INP_DROPPED
> flag set you are not allowed to call tcp_twstart(), actually it is a
> good candidate for a new INVARIANT.  Let me add that.
> 
> > After check source code I am found invocation of tcp_twstart() in
> > sys/netinet/tcp_stacks/fastpath.c, sys/netinet/tcp_input.c,
> > sys/dev/cxgb/ulp/tom/cxgb_cpl_io.c, sys/dev/cxgbe/tom/t4_cpl_io.c.
> > 
> > Invocation from sys/netinet/tcp_stacks/fastpath.c and
> > sys/netinet/tcp_input.c guarded by INP_WLOCK in tcp_input(), and now
> > will be OK.
> > 
> > Invocation from sys/dev/cxgb/ulp/tom/cxgb_cpl_io.c and
> > sys/dev/cxgbe/tom/t4_cpl_io.c is not clear to me, I am see independed
> > INP_WLOCK. Is this OK?
> > 
> > Can be thread A wants do_peer_close() directed from chelsio IRQ
> > handler, bypass tcp_input()?
> 
>  If you look carefully INP_WLOCK is used in cxgb_cpl_io.c and
> t4_cpl_io.c before calling tcp_twstart().

Yes, and you remeber: sys/netinet/tcp_subr.c

  1535  struct tcpcb *
  1536  tcp_close(struct tcpcb *tp)
  1537  {
...
  1569                  INP_WUNLOCK(inp);
  1570                  ACCEPT_LOCK();
  1571                  SOCK_LOCK(so);
  1572                  so->so_state &= ~SS_PROTOREF;
  1573                  sofree(so);
  1574                  return (NULL);

sofree() call tcp_usr_detach() and in tcp_usr_detach() we have
unexpected INP_TIMEWAIT.
_______________________________________________
freebsd-stable@freebsd.org mailing list
https://lists.freebsd.org/mailman/listinfo/freebsd-stable
To unsubscribe, send any mail to "freebsd-stable-unsubscr...@freebsd.org"

Reply via email to