Re: OOPS (kswapd) in 2.4.5 and 2.4.6

[Andrew Morton]

Henry wrote:
> 
> ...
> Dual-cpu pentium 233 (intel) with 128MB RAM and more than double that swap.
> 
> ...
> Unable to handle kernel NULL pointer dereference at virtual address 00000008
> c01b4227
> *pde = 00000000
> Oops: 0000
> CPU:    0
> EIP:    0010:[<c01b4227>]
> Using defaults from ksymoops -t elf32-i386 -a i386
> EFLAGS: 00010207
> eax: 00000001   ebx: 00000000   ecx: 000000c0   edx: c12c49c0
> esi: c12d3f4c   edi: 00000001   ebp: c0d0f2a0   esp: c12d3ee0
> ds: 0018   es: 0018   ss: 0018
> Process kswapd (pid: 3, stackpage=c12d3000)
> Stack: 00000000 c12d3f4c c12d3f4c c01330cb 00000001 00000000 001c4300 c1203048
>        00000000 00000028 c0129752 00000001 c1203048 00000305 c12d3f48 00001000
>        001c4300 c1203048 00000000 00000028 c12d3f48 00000000 00001000 00001c43
> Call Trace: [<c01330cb>] [<c0129752>] [<c0106cec>] [<c012981f>] [<c012a4e8>] [<c
> 0128b1d>] [<c01293f5>]
>        [<c0129486>] [<c01054cc>]
> Code: 0f b7 43 08 66 c1 e8 09 0f b7 f0 8b 43 18 a8 04 75 19 68 a7
> 
> >>EIP; c01b4227 <submit_bh+b/74>   <=====
> Trace; c01330cb <brw_page+8f/a0>
> Trace; c0129752 <rw_swap_page_base+152/1b0>
> Trace; c0106cec <ret_from_intr+0/7>
> Trace; c012981f <rw_swap_page+6f/b8>
> Trace; c012a4e8 <swap_writepage+78/80>
> Trace; c0128b1d <page_launder+285/874>
> Trace; c01293f5 <do_try_to_free_pages+1d/58>
> Trace; c0129486 <kswapd+56/e8>
> Trace; c01054cc <kernel_thread+28/38>

There does appear to be an SMP race in brw_page() which can cause
this - end_buffer_io_async() unlocks the page, try_to_free_buffers()
zaps the buffer_head ring and brw_page() gets a null pointer.  But
gee, it's unlikely unless you have super-fast disks and/or something
which has a super-slow interrupt routine.

Could you please provide a description of your hardware lineup?

And could you please test 2.4.6 with this patch?

--- linux-2.4.6/fs/buffer.c	Wed Jul  4 18:21:31 2001
+++ lk-ext3/fs/buffer.c	Fri Jul  6 18:25:00 2001
@@ -2181,8 +2181,9 @@ int brw_page(int rw, struct page *page, 
 
 	/* Stage 2: start the IO */
 	do {
+		struct buffer_head *next = bh->b_this_page;
 		submit_bh(rw, bh);
-		bh = bh->b_this_page;
+		bh = next;
 	} while (bh != head);
 	return 0;
 }

-