RE: Regression introduced by 0a2b9a6ea93650b8a00f9fd5ee8fdd25671e2df6 ("X86: integrate CMA with DMA-mapping subsystem" Re: Bug in BUG: Bad page state in process work_for_cpu pfn:cf800

[Marek Szyprowski <m.szyprowski@samsung.com>]

Hi Konrad,

On Tuesday, June 05, 2012 7:04 PM Konrad Rzeszutek Wilk wrote:

> On Sat, Jun 2, 2012 at 7:36 AM, Konrad Rzeszutek Wilk <konrad@darnok.org> wrote:
> > On Thu, May 31, 2012 at 3:19 AM, Marek Szyprowski
> > <m.szyprowski@samsung.com> wrote:
> >> Hi Konrad,
> >>
> >> On Thursday, May 31, 2012 2:45 AM Konrad Rzeszutek Wilk wrote:
> >>
> >>> About two-three days ago I started getting this on one of the AMD
> >>> machines I run nighly bootup test (full bootup log attached):
> >>> [Note: This is baremetal]
> >>>
> >>> ehci_hcd 0000:00:02.1: reset hcc_params a086 caching frame 256/512/1024 park
> >>> BUG: Bad page state in process work_for_cpu  pfn:cf800
> >>> page:ffffea0002d64000 count:-1 mapcount:0 ing:          (null) index:0x0
> >>> page flags: 0x100000000000000()
> >>> Modules linked in:
> >>> Pid: 1207, comm: work_for_cpu Not tainted 3.4.0upstream-09208-gaf56e0a #1
> >>> Call Trace:
> >>>  [<ffffffff81103eb7>] ? dump_page+0x97/0xf0
> >>>  [<ffffffff811050bd>] bad_page+0xad/0x100
> >>>  [<ffffffff811067a2>] get_page_from_freelist+0x712/0x850
> >>>  [<ffffffff812916d8>] ? __const_udelay+0x28/0x30
> >>>  [<ffffffff81107a82>] __alloc_pages_nodemask+0x162/0x900
> >>>  [<ffffffff810a2975>] ? dequeue_task_fair+0xa5/0x330
> >>>  [<ffffffff810367e2>] ? __switch_to+0x152/0x440
> >>>  [<ffffffff8107ee37>] ? lock_timer_base+0x37/0x70
> >>>  [<ffffffff8103c7ff>] dma_generic_alloc_coherent+0x10f/0x170
> >>>  [<ffffffff81062e7e>] gart_alloc_coherent+0xee/0x120
> >>>  [<ffffffff81137542>] dma_pool_alloc+0x102/0x2e0
> >>>  [<ffffffff8109f240>] ? try_to_wake_up+0x310/0x310
> >>>  [<ffffffff813f3dc7>] ehci_qh_alloc+0x47/0xf0
> >>>  [<ffffffff813f81e7>] ehci_pci_setup+0x367/0xea0
> >>>  [<ffffffff81389213>] ? device_pm_init+0x43/0x80
> >>>  [<ffffffff813d3065>] ? usb_alloc_dev+0x2d5/0x330
> >>>  [<ffffffff81002030>] ? do_one_initcall+0x30/0x170
> >>>  [<ffffffff813db6a9>] usb_add_hcd+0x1e9/0x7a0
> >>>  [<ffffffff813ea0fa>] usb_hcd_pci_probe+0x1ba/0x3a0
> >>>  [<ffffffff81088890>] ? cwq_dec_nr_in_flight+0x90/0x90
> >>>  [<ffffffff812ad3f2>] local_pci_probe+0x12/0x20
> >>>  [<ffffffff810888a3>] do_work_for_cpu+0x13/0x30
> >>>  [<ffffffff810906e6>] kthread+0x96/0xa0
> >>>  [<ffffffff815b61e4>] kernel_thread_helper+0x4/0x10
> >>>  [<ffffffff81090650>] ? kthread_freezable_should_stop+0x70/0x70
> >>>  [<ffffffff815b61e0>] ? gs_change+0x13/0x13
> >>> Disabling lock debugging due to kernel taint
> >>> BUG: Bad page state in process work_for_cpu  pfn:cf801
> >>>
> >>> I haven't actually run a git bisection, but the last git commit
> >>> that does something in the gart code looks to be this one:
> ..snip..
> > Doing a git bisection points it to this one:
> > is the first bad commit
> 
> I pulled todays linus's tree and if I revert the commit the bug disappears.
> This is on an Dell T105 AMD box running native x86_64.
> Any thoughts on what the bug might be?
 
I've read that patch three times line by line and I really have no idea what might cause such 
weird effect. When CMA is disabled this patch should not change anything in the code flow and 
the called functions... I assume that the CMA has not been enabled in your test config?

Best regards
-- 
Marek Szyprowski
Samsung Poland R&D Center