Re: page allocator regression on nommu

From: Paul Mundt
Date: Mon Aug 31 2009 - 06:26:54 EST


On Mon, Aug 31, 2009 at 01:08:19PM +0300, Pekka Enberg wrote:
> On Mon, Aug 31, 2009 at 10:48 AM, Paul Mundt<lethal@xxxxxxxxxxxx> wrote:
> > modprobe: page allocation failure. order:7, mode:0xd0
>
> OK, so we have order 7 page allocation here...
>
[snip]
> > Active_anon:0 active_file:0 inactive_anon:0
> > ?inactive_file:0 unevictable:323 dirty:0 writeback:0 unstable:0
> > ?free:2967 slab:0 mapped:0 pagetables:0 bounce:0
> > Normal free:11868kB min:0kB low:0kB high:0kB active_anon:0kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:1292kB present:16256kB pages_scanned:0 all_unreclaimable? no
> > lowmem_reserve[]: 0 0
> > Normal: 267*4kB 268*8kB 251*16kB 145*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB 0*8192kB 0*16384kB 0*32768kB = 11868kB
>
> ...but we seem to be all out of order > 3 pages. I'm not sure why
> commit 49255c619fbd482d704289b5eb2795f8e3b7ff2e changes any of this,
> though.
>
Nor am I, but it does. With it reverted, all of the order-7 allocations
succeed just fine. With some debugging printks added:

usbcore: registered new device driver usb
alloc order 7 for 49000: pages 0c21c000
alloc order 7 for 49000: pages 0c21c000
...

While with it applied:

alloc order 7 for 49000:
modprobe: page allocation failure. order:7, mode:0xd0
...
Mem-Info:
Normal per-cpu:
CPU 0: hi: 0, btch: 1 usd: 0
Active_anon:0 active_file:0 inactive_anon:0
inactive_file:0 unevictable:323 dirty:0 writeback:0 unstable:0
free:2911 slab:0 mapped:0 pagetables:0 bounce:0
Normal free:11644kB min:0kB low:0kB high:0kB active_anon:0kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:1292kB present:16256kB pages_scanned:0 all_unreclaimable? no
lowmem_reserve[]: 0 0
Normal: 259*4kB 264*8kB 247*16kB 142*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB 0*8192kB 0*16384kB 0*32768kB = 11644kB
323 total pagecache pages
4096 pages RAM
662 pages reserved
226 pages shared
288 pages non-shared
0 pages in pagetable cache
-ENOMEM
Allocation of length 299008 from process 50 (modprobe) failed
Normal per-cpu:
CPU 0: hi: 0, btch: 1 usd: 0
Active_anon:0 active_file:0 inactive_anon:0
inactive_file:0 unevictable:323 dirty:0 writeback:0 unstable:0
free:2911 slab:0 mapped:0 pagetables:0 bounce:0
Normal free:11644kB min:0kB low:0kB high:0kB active_anon:0kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:1292kB present:16256kB pages_scanned:0 all_unreclaimable? no
lowmem_reserve[]: 0 0
Normal: 259*4kB 264*8kB 247*16kB 142*32kB 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB 0*8192kB 0*16384kB 0*32768kB = 11644kB
323 total pagecache pages

the -ENOMEM printk() I've placed in the alloc_pages() error path.

> > ------------[ cut here ]------------
> > kernel BUG at mm/nommu.c:598!
> > Kernel BUG: 003e [#1]
> > Modules linked in:
> >
> > Pid : 51, Comm: ? ? ? ? ? ? ? ? modprobe
> > CPU : 0 ? ? ? ? ? ? ? ? Not tainted ?(2.6.31-rc7 #2835)
> >
> > PC is at __put_nommu_region+0xe/0xb0
> > PR is at do_mmap_pgoff+0x8dc/0xa68
>
> This looks to be a bug in nommu do_mmap_pgoff() error handling. I
> guess we shouldn't call __put_nommu_region() if add_nommu_region()
> hasn't been called?
>
Yeah, that looks a bit suspect. __put_nommu_region() is safe to be called
without a call to add_nommu_region(), but we happen to trip over the BUG_ON()
in this case because we've never made a single addition to the region tree.

We probably ought to just up_write() and return if nommu_region_tree == RB_ROOT,
which is what I'll do unless David objects.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/