4BSD Scheduler Problem on 5.3
rwatson at FreeBSD.org
Fri Apr 1 11:07:51 PST 2005
On Fri, 1 Apr 2005, Poul-Henning Kamp wrote:
> In message <200504011120.27285.jhb at FreeBSD.org>, John Baldwin writes:
> >> I think we have a race in -STABLE relating to tty wakeups and
> >> open/close/device teardown. I've seen a panic relating to sio during a
> >> tty close on RELENG_5 about 5-6 months ago, but was unable to get a dump.
> >> Scott has since fixed dumps with twe, but I've not yet been able to get
> >> the bug to recur. I'll give it another try.
> >Sounds very plausible. Does Poul-Henning have any ideas?
> Is this before or after my tty changes ?
> There is a general nastyness about ttys/sessions/exit which I have never
> really felt comfortable about. My hope is that I have solved it by
> refcounting the tty structure.
> So if this is before my changes: "Yeah, known (but rare) issue)"
> If after my changes: "D**N!"
The instance of the panic I saw was in RELENG_5 in January when using a
serial console. Here's a copy of the e-mail I sent to stable@ when it
occurred. It's a little weak on the debugging side because I couldn't get
a dump and didn't have a kernel with symbols easily on hand, but my
reading was that there was a race relating to open/close and device
Robert N M Watson
Date: Sun, 23 Jan 2005 16:51:14 +0000 (GMT)
From: Robert Watson <rwatson at FreeBSD.org>
To: stable at FreeBSD.org
Cc: imp at FreeBSD.org
Subject: NULL pointer deref in sioopen() suggests a close/open race on sio
Ran into the following panic on a 5-STABLE box this morning, which
occurred after hitting Ctrl-D to close a login session on a serial console
(ttyd0 at 9600 bps):
login: Jan 23 10:43:27 fledge login: 2 LOGIN FAILURES ON ttyd0
Fatal trap 12: page fault while in kernel mode
fault virtual address = 0x1c
fault code = supervisor write, page not present
instruction pointer = 0x8:0xc051537b
stack pointer = 0x10:0xe7345988
frame pointer = 0x10:0xe7345994
code segment = base 0x0, limit 0xfffff, type 0x1b
= DPL 0, pres 1, def32 1, gran 1
processor eflags = interrupt enabled, resume, IOPL = 0
current process = 45092 (getty)
[thread pid 45092 tid 100201 ]
Stopped at knote+0x27: cmpxchgl %ecx,0x1c(%edx)
db> show pcpu
cpuid = 0
curthread = 0xc290d190: pid 45092 "getty"
curpcb = 0xe7345da0
fpcurthread = 0xc290d190: pid 45092 "getty"
idlethread = 0xc22644b0: pid 11 "idle"
APIC ID = 0
currentldt = 0x30
Tracing pid 45092 tid 100201 td 0xc290d190
knote(c264e098,0,0,c290d190,e73459c4) at knote+0x27
ttwwakeup(c264e000) at ttwwakeup+0xc8
comstart(c264e000) at comstart+0x385
comparam(c264e000,c264e0a4,c264e000,3,0) at comparam+0x253
sioopen(c079f060,3,2000,c290d190,c078e6a0) at sioopen+0x1df
spec_open(e7345a84,e7345b40,c058d585,e7345a84,180) at spec_open+0x2b6
spec_vnoperate(e7345a84) at spec_vnoperate+0x13
vn_open_cred(e7345be4,e7345ce4,c08,c2261d80,0) at vn_open_cred+0x419
vn_open(e7345be4,e7345ce4,c08,0,c4289b58) at vn_open+0x1e
kern_open(c290d190,804f8e0,0,3,bfbfee18) at kern_open+0xe3
open(c290d190,e7345d14,3,0,292) at open+0x18
syscall(2f,2f,2f,804f8e0,0) at syscall+0x27b
Xint0x80_syscall() at Xint0x80_syscall+0x1f
--- syscall (5, FreeBSD ELF32, open), eip = 0x280d155b, esp = 0xbfbfedec,
ebp = 0xbfbfee18 ---
The ps list is a bit boring, but the primary interesting thing is that it
looks like the close was going on in one thread just about when the sio
swi was scheduled to run also:
pid proc uarea uid ppid pgrp flag stat wmesg wchan cmd
45092 c6762388 e7387000 0 1 1 0004000 [CPU 0] getty ...
132 c235954c e4fbf000 0 0 0 000020c [RUNQ] swi5: clock sio
I didn't have a kernel with debugging symbols on-hand, but the above
address in knote() is a cmpxchg early in the function, which means it's
likely the conditional call to mtx_lock() hitting a NULL mutex pointer for
kl_lock. This in turn suggests that something has called ttyrel/tty_close
on the TTY in a race with the open, or otherwise NULL'd that pointer via
knlist_destroy(). Anyone have any pointers on this one? The TTY code is
not my forte...
Robert N M Watson
More information about the freebsd-hackers