reproducible panic with mount_smbfs

[John Baldwin]

On Sunday 02 November 2008 11:17:18 am Attilio Rao wrote:
> 2008/11/2, Attilio Rao <attilio at freebsd.org>:
> > 2008/11/2, Yuri Pankov <yuri.pankov at gmail.com>:
> >
> > > Hi,
> >  >
> >  >  Trying to mount nonexistent smb share with mount_smbfs leads to
> >  >  following panic:
> >  >
> >  >  # mount_smbfs //yuri at lifebane/blahblah /mnt
> >  >
> >  >  Unread portion of the kernel message buffer:
> >  >  smb_co_lock: recursive lock for object 1
> >  >  panic: Lock (lockmgr) smb_vc not locked @
> >  >  /usr/src/sys/modules/smbfs/../../netsmb/smb_conn.c:329.
> >  >  cpuid = 0
> >  >  KDB: stack backtrace:
> >  >  db_trace_self_wrapper() at db_trace_self_wrapper+0x2a
> >  >  panic() at panic+0x182
> >  >  witness_assert() at witness_assert+0x21a
> >  >  __lockmgr_args() at __lockmgr_args+0x17a
> >  >  smb_co_put() at smb_co_put+0x76
> >  >  smb_sm_lookup() at smb_sm_lookup+0xfe
> >  >  smb_usr_lookup() at smb_usr_lookup+0xcd
> >  >  nsmb_dev_ioctl() at nsmb_dev_ioctl+0x1f6
> >  >  giant_ioctl() at giant_ioctl+0x75
> >  >  devfs_ioctl_f() at devfs_ioctl_f+0x76
> >  >  kern_ioctl() at kern_ioctl+0x92
> >  >  ioctl() at ioctl+0xfd
> >  >  syscall() at syscall+0x1bf
> >  >  Xfast_syscall() at Xfast_syscall+0xab
> >  >  --- syscall (54, FreeBSD ELF64, ioctl), rip = 0x800939aec, rsp =
> >  >  0x7fffffffe038, rbp = 0x7fffffffe450 ---
> >  >  Uptime: 6m46s
> >  >  Physical memory: 2032 MB
> >
> >
> > So, what is happening here is that smb_co_lock() is AFU.
> >  Infact looking at the code:
> >  int
> >  smb_co_lock(struct smb_connobj *cp, int flags, struct thread *td)
> >  {
> >  ...
> >         if (smb_co_lockstatus(cp, td) == LK_EXCLUSIVE &&
> >             (flags & LK_CANRECURSE) == 0) {
> >                 SMBERROR("recursive lock for object %d\n", cp->co_level);
> >                 return 0;
> >         }
> >  ...
> 
> Yuri,
> could you please test this fix:
> http://www.freebsd.org/~attilio/netsmb.diff
> 
> and report if it works?
> You could get a KASSERT running but this is expected as I want to
> identify on the callers who passes a malformed request and fix it.

This allows all smb locks to recurse unlike the original code I think.  It may 
be better if smb_vclist was initialized with LK_RECURSE, but not all the 
other smb locks.  Also, in smb_co_addchild() I think you should just replace 
the existing asserts with appropriate lockmgr_assert() (you could add a 
smb_co_assert() to preserve the layering) rather than removing assertions 
altogether.

-- 
John Baldwin