Re: LOR between nfs and proctree

On Thu, Dec 18, 2008 at 01:43:38AM +0300, pluknet wrote:
> I didn't find this LOR in archives and don't know if this is known
> one, so I'm posting it here.
> 
> lock order reversal:
>  1st 0xc5ea37ac nfs (nfs) _at_ /usr/src/sys/kern/vfs_vnops.c:531
>  2nd 0xc0889090 proctree (proctree) _at_ /usr/src/sys/kern/kern_fork.c:311
> KDB: stack backtrace:
> db_trace_self_wrapper(c07fb0cf,e7d1f8cc,c05d79e5,4,c07f662d,...) at
> db_trace_self_wrapper+0x26
> kdb_backtrace(4,c07f662d,c54fcad8,c54f6068,e7d1f928,...) at kdb_backtrace+0x29
> _witness_debugger(c07fdd7f,c0889090,c07f702d,c54f6068,c07f3e44,...) at
> _witness_debugger+0x25
> witness_checkorder(c0889090,1,c07f3e44,137,0,...) at witness_checkorder+0x839
> _sx_slock(c0889090,0,c07f3e44,137,0,...) at _sx_slock+0x85
> fork1(c0887530,60034,0,e7d1f9c0,c07f662d,...) at fork1+0x2cd
> kproc_create(c06dc780,c09f8920,0,40000,0,...) at kproc_create+0x47
> nfs_nfsiodnew(c09f8f70,0,c080e39a,553,0,...) at nfs_nfsiodnew+0xc5
> nfs_asyncio(c5c024e8,d952b670,c5af0000,c5abdd80,0,...) at nfs_asyncio+0x3a5
> nfs_bioread(c5ea3754,e7d1fc58,20000,c5af0000,e7d1fba0,...) at nfs_bioread+0x788
> nfs_read(e7d1fbc8,c5abdd80,c5e63968,c5abdd80,e7d1fbe8,...) at nfs_read+0x44
> VOP_READ_APV(c085d360,e7d1fbc8,c0805cb8,213,c5abde24,...) at VOP_READ_APV+0xa5
> vn_read(c5e63968,e7d1fc58,c5af0000,0,c5abdd80,...) at vn_read+0x1ee
> dofileread(e7d1fc58,ffffffff,ffffffff,0,c5e63968,...) at dofileread+0x96
> kern_readv(c5abdd80,4,e7d1fc58,e7d1fc78,1,...) at kern_readv+0x58
> read(c5abdd80,e7d1fcf8,c,c07fe5a8,c083f5c8,...) at read+0x4f
> syscall(e7d1fd38) at syscall+0x2a3
> Xint0x80_syscall() at Xint0x80_syscall+0x20
> --- syscall (3, FreeBSD ELF32, read), eip = 0x281c7f93, esp =
> 0xbfbfeaec, ebp = 0xbfbfec28 ---

It is more or less known, nfs may create new nfsiod while holding vnode
lock. Severity of the issue lowered with removal of the single-threading
from the fork(), but it is still real. Workaround is to create required
amount of nfsiod in advance.