NFS PANIC and FIX in 2.4
Hi All,
I get this panic running RedHat 2.4.3-6:
Unable to handle kernel NULL pointer dereference at virtual address 00000004 c014b13d Oops: 0002 CPU: 0 EIP: 0010:[<c014b13d>]
Using defaults from ksymoops -t elf32-i386 -a i386
EFLAGS: 00010246
eax: 00000000 ebx: c4a6dcc0 ecx: c4a6dce8 edx: 00000000
esi: c0256ed8 edi: c59197a0 ebp: c4ba79e8 esp: c4ba78a0
ds: 0018 es: 0018 ss: 0018
Process nfsd (pid: 3230, stackpage=c4ba7000)
Stack: c1239148 c4a6d9e8 c4a6dcc0 c4a6d9c0 c4a6dcc0 c4a6d9c0 c68af335 c4a6dcc0
c4a6d9c0 00000000 c4ba79e8 c4a6d9c0 c59197a0 c68af59a c4a6d9c0 c59197a0
c4ba79e8 00000000 00706d74 c5e07d80 c1e9ec00 c014d972 c1e9ec00 00007458
Call Trace: [<c68af335>] [<c68af59a>] [<c014d972>] [<c016bf17>] [<c015fa28>]
[<c02275a0>] [<c015fa51>] [<c014d972>] [<c015fbfd>] [<c014b0b4>] [<c68af887>]
[...]
Quoted 1 line
Quoted 1 line
EIP;
c014b13d<dput+ad/180> <=====
Trace; c68af335 <[nfsd]d_splice+e5/160>
Trace; c68af59a <[nfsd]splice+ea/140>
Trace; c014d972 <iget4+52/100>
Trace; c016bf17 <ll_rw_block+167/1b0>
[...]
Code; c014b13d <dput+ad/180>
00000000 <_EIP>:
Code; c014b13d <dput+ad/180> <=====
0: 89 50 04 mov %edx,0x4(%eax) <=====
Code; c014b140 <dput+b0/180>
3: 89 02 mov %eax,(%edx)
Code; c014b142 <dput+b2/180>
5: c7 43 28 00 00 00 00 movl $0x0,0x28(%ebx)
Code; c014b149 <dput+b9/180>
c: c7 41 04 00 00 00 00 movl $0x0,0x4(%ecx)
Code; c014b150 <dput+c0/180>
13: 8b 00 mov (%eax),%eax
This is in dcache.c here:
kill_it: {
struct dentry *parent;
list_del(&dentry->d_child);
^^^^^^^^^^^^^^^^^^^^^^^^^^
/* drops the lock, at that point nobody can reach this dentry */
dentry_iput(dentry);The problem is pretty specific to 2.4.3-6 because it has code in list_del() to null out the prev and next pointers (I don't know where they picked it up, but
its gone in 2.4.5). However, it exposes a bug in NFS, namely that d_splice()
also calls list_del() on tdentry->d_child. This shouldn't be done because it makes the d_child list invalid, so any subsequent call to list_del on d_child could panic. I think the correct fix (attached below) is to change the NFS list_del to list_del_init.
James Bottomley
Steeleye technology.
Index: linux/2.4/fs/nfsd/nfsfh.c
2.4/fs/nfsd/nfsfh.c renamed from 2.4/fs/nfsd/nfsfh.c:1.1.1.8.4.1 +1 -1
@@ -238,7 +238,7 @@* make it an IS_ROOT instead */ spin_lock (&dcache_lock); -list_del (&tdentry->d_child); +list_del_init (&tdentry->d_child); tdentry->d_parent = tdentry;spin_unlock (&dcache_lock);d_rehash (target);