LOR: kern_descrip.c:2277,2278

[John Baldwin]

On Monday 17 January 2005 05:07 pm, Bjoern A. Zeeb wrote:
> Hi,
>
> I have added the following as LOR #055 to "the LOR page"[1].
>
> lock order reversal
>  1st 0xffffffff80c3a9e8 sleep mtxpool (sleep mtxpool) @
> sys/kern/kern_descrip.c:2277 2nd 0xffffff00222d8a48 filedesc structure
> (filedesc structure) @ sys/kern/kern_descrip.c:2278 KDB: stack backtrace:
> witness_checkorder() at witness_checkorder+0x5f1
> _mtx_lock_flags() at _mtx_lock_flags+0x4a
> dupfdopen() at dupfdopen+0x320
> kern_open() at kern_open+0x5de
> syscall() at syscall+0x4ab
> Xfast_syscall() at Xfast_syscall+0xa8
> --- syscall (5, FreeBSD ELF64, open), rip = 0x80077e7d0, rsp =
> 0x7fffffffe6f8, rbp = 0x7fffffffec70 ---
>
> I can easily reproduce it every boot. It seems to be triggered by
> Capi4BSD[2] framework which I incorporated in my private tree.
>
> Can someone please comment on this ?

The problem is that FILEDESC_UNLOCK() actually includes an entirely separate 
lock of its own (it's like an sx lock sort of).  One possible fix might be to 
change struct file to either use a dedicated mutex pool (instead of the more 
generic mtxpool_sleep one that is intended only for leaf-lock usage) or to 
have each struct file include its own mutex rather than using a pool lock.

-- 
John Baldwin <jhb at FreeBSD.org>  <><  http://www.FreeBSD.org/~jhb/
"Power Users Use the Power to Serve"  =  http://www.FreeBSD.org