Re: Warning from AMD IOMMU performance counters

[Laura Abbott]

On 02/21/2016 05:52 PM, Wan Zongshun wrote:
> -------- Original Message --------
> > Hi,
> >
> > Since about 4.4, we've been seeing reports of this warning on every boot
> > from some users:
> >
> > WARNING: CPU: 2 PID: 1 at drivers/iommu/amd_iommu_init.c:2301
> > amd_iommu_pc_get_set_reg_val+0xa8/0xe0()
> > Modules linked in:
> > CPU: 2 PID: 1 Comm: swapper/0 Not tainted 4.4.2-300.fc23.x86_64 #1
> > Hardware name: Hewlett-Packard HP EliteBook 755 G2/221C, BIOS M84 Ver.
> > 01.10 10/20/2015
> >   0000000000000000 0000000026124b43 ffff88042d687d20 ffffffff813b0c9f
> >   0000000000000000 ffff88042d687d58 ffffffff810a2f12 ffff88042f014800
> >   0000000000000040 0000000000000000 ffffffff81d6c73d 0000000000000000
> > Call Trace:
> >   [<ffffffff813b0c9f>] dump_stack+0x44/0x55
> >   [<ffffffff810a2f12>] warn_slowpath_common+0x82/0xc0
> >   [<ffffffff81d6c73d>] ? memblock_find_dma_reserve+0x16a/0x16a
> >   [<ffffffff810a305a>] warn_slowpath_null+0x1a/0x20
> >   [<ffffffff814d3f48>] amd_iommu_pc_get_set_reg_val+0xa8/0xe0
> >   [<ffffffff81dafda7>] iommu_go_to_state+0x4d6/0x1384
> >   [<ffffffff813c02ea>] ? kvasprintf+0x7a/0xa0
> >   [<ffffffff81d6c73d>] ? memblock_find_dma_reserve+0x16a/0x16a
> >   [<ffffffff81db0cbd>] amd_iommu_init+0x13/0x201
> >   [<ffffffff81d6c74f>] pci_iommu_init+0x12/0x3c
> >   [<ffffffff81002123>] do_one_initcall+0xb3/0x200
> >   [<ffffffff810c0935>] ? parse_args+0x295/0x4b0
> >   [<ffffffff81d621c8>] kernel_init_freeable+0x189/0x223
> >   [<ffffffff8178dc00>] ? rest_init+0x80/0x80
> >   [<ffffffff8178dc0e>] kernel_init+0xe/0xe0
> >   [<ffffffff81799c8f>] ret_from_fork+0x3f/0x70
> >   [<ffffffff8178dc00>] ? rest_init+0x80/0x80
> >
> > Full bugzilla https://bugzilla.redhat.com/show_bug.cgi?id=1310258
> >
> > Any idea what might be causing this spew?
>
> I also see this warning.
>
> Suravee's the one of this patch series might be fix this issue, you can try it.
>
> [PATCH V4 0/6] perf/amd/iommu: Enable multi-IOMMU support

That patch had other dependencies. I brought what I thought were enough of
them in to at least compile for the reporter to test (perf_event changes
from Borislav Petkov). Apparently the kernel doesn't boot with the series
so suspect there is an integration issue somewhere. I'm going to see if
I can more information from the reporter about what exactly went wrong
to see if I need more dependencies.

It would be nice to get a single patch which could be applied to stable.
Tainting the kernel immediately on bootup makes sorting through problems
more difficult.

Thanks,
Laura