panic: lock (sleep mutex) inp not locked

[Jon Noack]

On 11/28/04 17:07, Christian Brueffer wrote:
> Hi,
> 
> got this panic today on an SMP system.  Modifications to the kernel are
> polling and altq support for the sf driver and removed SMP ifdefs from
> sys/kern/kern_poll.c.
> 
> FreeBSD haakonia.hitnet.RWTH-Aachen.DE 5.3-STABLE FreeBSD 5.3-STABLE #61: Sat Nov 27 02:59:01 CET 2004
> chris at haakonia.hitnet.RWTH-Aachen.DE:/usr/home/build/usr/home/build/src/sys/LORIEN i386
> 
> Sources are from November 22nd.
> 
> Crashdump and debug kernel are available for further investigation.
> 
> panic: lock (sleep mutex) inp not locked @ /usr/home/build/src/sys/netinet/tcp_input.c:2521
> cpuid = 1
> KDB: enter: panic
> [thread 100034]
> Stopped at      kdb_enter+0x2b: nop
> db> tr
> kdb_enter(c06c6c54) at kdb_enter+0x2b
> panic(c06caa02,c06db3d0,c06d57ce,c06d2493,9d9) at panic+0x127
> witness_unlock(c3f4d57c,8,c06d2493,9d9) at witness_unlock+0xbc
> _mtx_unlock_flags(c3f4d57c,0,c06d2493,9d9,0) at _mtx_unlock_flags+0x5b
> tcp_input(c44bb000,14,5cb5e289,0,0) at tcp_input+0x2fb0
> ip_input(c44bb000) at ip_input+0x52d
> netisr_dispatch(2,c44bb000,0,c44b0800,c1a93c00) at netisr_dispatch+0x58
> ether_demux(c1a93c00,c44bb000,297,96,d74f6400) at ether_demux+0x292
> ether_input(c1a93c00,c44bb000,c1a93e84,0,c0833d13) at ether_input+0x269
> sf_rxeof(c1a93c00,c1a93c00,c1a93e84,c07276c0,d543ecbc) at sf_rxeof+0x183
> sf_poll_locked(c1a93c00,0,5,c1a93e84,0) at sf_poll_locked+0x6b
> sf_poll(c1a93c00,0,5) at sf_poll+0x2e
> netisr_poll(0) at netisr_poll+0x188
> swi_net(0) at swi_net+0x81
> ithread_loop(c19e1080,d543ed48,c19e1080,c051aa34,0) at
> ithread_loop+0x124
> fork_exit(c051aa34,c19e1080,d543ed48) at fork_exit+0xa4
> fork_trampoline() at fork_trampoline+0x8
> --- trap 0x1, eip = 0, esp = 0xd543ed7c, ebp = 0 ---
> db> show locks
> exclusive sleep mutex tcp r = 0 (0xc0755dac) locked @
> /usr/home/build/src/sys/netinet/tcp_input.c:617
> exclusive sleep mutex Giant r = 1 (0xc0728320) locked @
> /usr/home/build/src/sys/kern/kern_poll.c:332
> db> show pcpu
> cpuid        = 1
> curthread    = 0xc19e3640: pid 38 "swi1: net"
> curpcb       = 0xd543eda0
> fpcurthread  = none
> idlethread   = 0xc19a84b0: pid 11 "idle: cpu1"
> APIC ID      = 1
> currentldt   = 0x28
> spin locks held:
> db>

I successfully used polling on my SMP machine a while back, but it 
pessimized performance for my workloads so I disabled it.  All I did was 
to wrap the SMP error in /sys/kern/kern_poll.c with #ifdef 0...#endif 
and everything worked.

In any case, it seems rwatson@ is interested in polling on SMP; check 
out his netperf page (http://www.watson.org/~robert/freebsd/netperf/) 
for more info.  Specifically, the following entry in his 20041103 patch 
is of interest:

"A mutex to protect device polling state is introduced, pollmtx, and
   used to protect the global variables present in kern_poll.c.  As a
   result, Giant is removed from the polling netisr code, and the
   #ifdef causing polling build to fail on SMP is removed.  This has
   not been tested."

No idea whether or not it works, but I thought I'd give you a heads up.

Jon