Skip to contentSkip to conversation

page allocation failures with 2.6.5 on s390

DISCUSSION Arnd Bergmann 1 participant 1 message Apr 25, 2004

lore
Share & tools
lore
#0 lore raw
Quoted 10 lines · Bastian Blank
Quoted 8 lines · Andrew Morton
Quoted 2 lines · Bastian Blank

Today I see the following error message in the logs of two machines with 2.6.5 on s390, both 31 and 64 bit.

This is pretty much unavoidable - the kernel is looking for four physically-contiguous free pages at interrupt time, and there aren't any. You can reduce its frequency by increasing /proc/sys/vm/min_free_kbytes. The best fix would be to alter the driver to use order-0 pages if that's possible.

Quoted 3 lines

The pages are 4KiB large, so it allocates 16KiB at once. So the problem is = rather clear:
| 3: eth0: <BROADCAST,MULTICAST,UP> mtu 8192 qdisc pfifo_fast qlen 100

Is this the vanilla kernel? Note that 2.6.6-rc2 contains a different qeth driver (the same that you get if you apply the 2.6.5 patches from
http://www10.software.ibm.com/developerworks/opensource/linux390), which might help you here.
Getting the mtu size below 7500 could also give you relief if that causes order-1 allocations instead of order-2

Quoted 4 lines

I'm currently not sure if this was the problem causing D process. I reread the logs and found the following:
| Unable to handle kernel pointer dereference at virtual kernel address 000= 0000000000000

Since this is the pdflush thread that was killed, it's just a matter of time until everything hangs.

Quoted 9 lines
| Oops: 0004 [#1]
| CPU:    0    Not tainted
| Process pdflush (pid: 33, task: 000000000fa58e90, ksp: 000000000fa53b98)
| Krnl PSW : 0700200180000000 00000000001022b0 (nfs3_request_init+0x34/0x50)
| Krnl GPRS: 0000000000000000 0000000000000001 0000000000000000 0000000000000028
|            0000000000000001 0000000000000000 0000000000001000 000000000672f548
|            0000000000403620 000000000dceb600 0000000000000000 0000000000823020
|            00000000061bfd00 000000000029afb8 00000000001022a0 000000000fa53178
| Krnl Code: ba 54 30 00 a7 44 ff fc e3 20 c0 30 00 24 e3 40 f1 10 00 04

ba 54 30 00 means 'cs %r5,%r4,0(%r3)' and accessing %r3 failed (it contains 0x28, as you can see).

In the corresponding code (from fs/nfs/nfs3proc.c), you can see what this means:
void nfs3_request_init(struct nfs_page *req, struct file *filp)

{
	req->wb_cred = get_rpccred(nfs_cred(req->wb_inode, filp));
}

nfs_cread(...) returns NULL, and get_rpccred(...) accesses this at offset 0x28 with atomic_inc (aka cs).
Someone familiar with nfs needs to look into how this can happen, but it does not appear to be s390 specific.

Quoted 9 lines
| Call Trace:
|  [nfs_create_request+216/312] nfs_create_request+0xd8/0x138
|  [nfs_update_request+678/1236] nfs_update_request+0x2a6/0x4d4
|  [nfs_writepage_async+44/264] nfs_writepage_async+0x2c/0x108
|  [nfs_writepage+466/476] nfs_writepage+0x1d2/0x1dc
|  [mpage_writepages+796/1084] mpage_writepages+0x31c/0x43c
|  [nfs_writepages+56/196] nfs_writepages+0x38/0xc4
|  [do_writepages+46/76] do_writepages+0x2e/0x4c
|  [__sync_single_inode+200/744] __sync_single_inode+0xc8/0x2e8
Quoted 8 lines
|  [__writeback_single_inode+246/260] __writeback_single_inode+0xf6/0x104
|  [sync_sb_inodes+472/792] sync_sb_inodes+0x1d8/0x318
|  [writeback_inodes+144/248] writeback_inodes+0x90/0xf8
|  [wb_kupdate+182/324] wb_kupdate+0xb6/0x144
|  [__pdflush+430/700] __pdflush+0x1ae/0x2bc
|  [pdflush+46/60] pdflush+0x2e/0x3c
|  [kthread+228/236] kthread+0xe4/0xec
|  [kernel_thread_starter+20/28] kernel_thread_starter+0x14/0x1c
	Arnd <><