LinuxLists.cc - AMD graphics performance regression in 4.15 and later

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Hi Jean,

yeah, that is a known problem. Using huge pages improves the performance
because of better TLB usage, but for the cost of higher allocation overhead.

What we found is that firefox is doing something rather strange by
allocating large textures and then just trowing them away again immediately.

We mitigated the problem by avoiding the slow coherent DMA code path on
almost all platforms on newer kernels, but essentially somebody needs to
figure out why firefox and/or the user space stack is doing this
constant allocation/freeing of memory.

There is also a bug tracker on bugs.kernel.org about this, but I can't
find it any more of hand.

Regards,
Christian.

Am 06.04.2018 um 02:30 schrieb Jean-Marc Valin:
> Hi,
>
> I noticed a serious graphics performance regression between 4.14 and
> 4.15. It is most noticeable with Firefox (tried FF57 through FF60) and
> causes scrolling to be really choppy/sluggish. I've confirmed that the
> problem is also there on 4.16, while 4.13 works fine.
>
> After a bisection, I've narrowed the regression down to this commit:
>
> commit 648bc3574716400acc06f99915815f80d9563783
> Author: Christian König <[email protected]>
> Date: Thu Jul 6 09:59:43 2017 +0200
>
> drm/ttm: add transparent huge page support for DMA allocations v2
>
>
> Some details about my system:
> Distro: Fedora 27 (up-to-date)
> Video: MSI Radeon RX 560 AERO
> CPU: Dual-socket Xeon E5-2640 v4 (20 cores total)
> RAM: 128 GB ECC
>
>
> As a comparison, when running Firefox with 4.15 on a Lenovo W540 laptop
> (with Intel graphics only) the responsiveness is much better then what
> I'm getting on the Xeon machine above with the Radeon card, so this
> really seems to be an AMD-only issue.
>
> Any way to fix the issue?
>
> Thanks,
>
> Jean-Marc
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

2018-04-06 08:12:53

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Hi Jean,

found the bug reports.

Here is the original bug report from the kernel:
https://bugzilla.kernel.org/show_bug.cgi?id=198511

And here is an fdo bug report where we tried to investigate the root
cause, but didn't had time for that yet:
https://bugs.freedesktop.org/show_bug.cgi?id=105038

Regards,
Christian.

Am 06.04.2018 um 10:03 schrieb Christian König:
> Hi Jean,
>
> yeah, that is a known problem. Using huge pages improves the
> performance because of better TLB usage, but for the cost of higher
> allocation overhead.
>
> What we found is that firefox is doing something rather strange by
> allocating large textures and then just trowing them away again
> immediately.
>
> We mitigated the problem by avoiding the slow coherent DMA code path
> on almost all platforms on newer kernels, but essentially somebody
> needs to figure out why firefox and/or the user space stack is doing
> this constant allocation/freeing of memory.
>
> There is also a bug tracker on bugs.kernel.org about this, but I can't
> find it any more of hand.
>
> Regards,
> Christian.
>
> Am 06.04.2018 um 02:30 schrieb Jean-Marc Valin:
>> Hi,
>>
>> I noticed a serious graphics performance regression between 4.14 and
>> 4.15. It is most noticeable with Firefox (tried FF57 through FF60) and
>> causes scrolling to be really choppy/sluggish. I've confirmed that the
>> problem is also there on 4.16, while 4.13 works fine.
>>
>> After a bisection, I've narrowed the regression down to this commit:
>>
>> commit 648bc3574716400acc06f99915815f80d9563783
>> Author: Christian König <[email protected]>
>> Date:   Thu Jul 6 09:59:43 2017 +0200
>>
>>      drm/ttm: add transparent huge page support for DMA allocations v2
>>
>>
>> Some details about my system:
>> Distro: Fedora 27 (up-to-date)
>> Video: MSI Radeon RX 560 AERO
>> CPU: Dual-socket Xeon E5-2640 v4 (20 cores total)
>> RAM: 128 GB ECC
>>
>>
>> As a comparison, when running Firefox with 4.15 on a Lenovo W540 laptop
>> (with Intel graphics only) the responsiveness is much better then what
>> I'm getting on the Xeon machine above with the Radeon card, so this
>> really seems to be an AMD-only issue.
>>
>> Any way to fix the issue?
>>
>> Thanks,
>>
>>     Jean-Marc
>> _______________________________________________
>> dri-devel mailing list
>> [email protected]
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>

2018-04-06 15:31:35

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Hi Christian,

Is there a way to turn off these huge pages at boot-time/run-time? Right
now the recent kernels are making Firefox pretty much unusable for me.
I've been able to revert the patch from 4.15 but it's not really a
long-term solution.

You mention that the purpose of the patch is to improve performance, but
I haven't actually noticed anything running faster on my system. Is
there any particular test where I'm supposed to see an improvement
compared to 4.14?

I'm not sure what you mean by "We mitigated the problem by avoiding the
slow coherent DMA code path on almost all platforms on newer kernels". I
tested up to 4.16 and the performance regression is just as bad as it is
for 4.15.

Unlike the older hardware reported on kernel bug 198511, the hardware I
have is quite recent (RX 560) and still being sold. I've also confirmed
that neither nvidia (on the same machine) nor intel GPUs (on a less
powerful machine) are affected, so it seems like there's a way to avoid
that slow performance. I'm not saying that what Firefox is doing is
ideal (I don't know what it does and why), but it still seems like
something that should still be avoided in the kernel.

Cheers,

Jean-Marc

On 04/06/2018 04:03 AM, Christian König wrote:
> Hi Jean,
>
> yeah, that is a known problem. Using huge pages improves the performance
> because of better TLB usage, but for the cost of higher allocation
> overhead.
>
> What we found is that firefox is doing something rather strange by
> allocating large textures and then just trowing them away again
> immediately.
>
> We mitigated the problem by avoiding the slow coherent DMA code path on
> almost all platforms on newer kernels, but essentially somebody needs to
> figure out why firefox and/or the user space stack is doing this
> constant allocation/freeing of memory.
>
> There is also a bug tracker on bugs.kernel.org about this, but I can't
> find it any more of hand.
>
> Regards,
> Christian.
>
> Am 06.04.2018 um 02:30 schrieb Jean-Marc Valin:
>> Hi,
>>
>> I noticed a serious graphics performance regression between 4.14 and
>> 4.15. It is most noticeable with Firefox (tried FF57 through FF60) and
>> causes scrolling to be really choppy/sluggish. I've confirmed that the
>> problem is also there on 4.16, while 4.13 works fine.
>>
>> After a bisection, I've narrowed the regression down to this commit:
>>
>> commit 648bc3574716400acc06f99915815f80d9563783
>> Author: Christian König <[email protected]>
>> Date:   Thu Jul 6 09:59:43 2017 +0200
>>
>>      drm/ttm: add transparent huge page support for DMA allocations v2
>>
>>
>> Some details about my system:
>> Distro: Fedora 27 (up-to-date)
>> Video: MSI Radeon RX 560 AERO
>> CPU: Dual-socket Xeon E5-2640 v4 (20 cores total)
>> RAM: 128 GB ECC
>>
>>
>> As a comparison, when running Firefox with 4.15 on a Lenovo W540 laptop
>> (with Intel graphics only) the responsiveness is much better then what
>> I'm getting on the Xeon machine above with the Radeon card, so this
>> really seems to be an AMD-only issue.
>>
>> Any way to fix the issue?
>>
>> Thanks,
>>
>>     Jean-Marc
>> _______________________________________________
>> dri-devel mailing list
>> [email protected]
>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>

2018-04-06 15:56:35

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 06.04.2018 um 17:30 schrieb Jean-Marc Valin:
> Hi Christian,
>
> Is there a way to turn off these huge pages at boot-time/run-time?

Only at compile time by not setting CONFIG_TRANSPARENT_HUGEPAGE.

Alternatively you can avoid enabling CONFIG_SWIOTLB which will avoid the
slow DMA path as well.

> Right now the recent kernels are making Firefox pretty much unusable for me.
> I've been able to revert the patch from 4.15 but it's not really a
> long-term solution.
>
> You mention that the purpose of the patch is to improve performance, but
> I haven't actually noticed anything running faster on my system. Is
> there any particular test where I'm supposed to see an improvement
> compared to 4.14?

Mostly crypto mining, maybe some games as well.

> I'm not sure what you mean by "We mitigated the problem by avoiding the
> slow coherent DMA code path on almost all platforms on newer kernels". I
> tested up to 4.16 and the performance regression is just as bad as it is
> for 4.15.

Indeed 4.16 still doesn't have that. You could use the
amd-staging-drm-next branch or wait for 4.17.

> Unlike the older hardware reported on kernel bug 198511, the hardware I
> have is quite recent (RX 560) and still being sold.

That isn't related to the GFX hardware, but to your CPU/motherboard and
whatever else you have in the system.

Some part of your system needs SWIOTLB and that makes allocating memory
much slower.

> I've also confirmed that neither nvidia (on the same machine) nor intel GPUs (on a less
> powerful machine) are affected, so it seems like there's a way to avoid
> that slow performance.

Intel doesn't use TTM because they don't have dedicated VRAM, but the
open source nvidia driver should be affected as well.

> I'm not saying that what Firefox is doing is
> ideal (I don't know what it does and why), but it still seems like
> something that should still be avoided in the kernel.

We already mitigated that problem and I don't see any solution which
will arrive faster than 4.17.

The only quick workaround I can see is to avoid firefox, chrome for
example is reported to work perfectly fine.

Christian.

>
> Cheers,
>
> Jean-Marc
>
> On 04/06/2018 04:03 AM, Christian König wrote:
>> Hi Jean,
>>
>> yeah, that is a known problem. Using huge pages improves the performance
>> because of better TLB usage, but for the cost of higher allocation
>> overhead.
>>
>> What we found is that firefox is doing something rather strange by
>> allocating large textures and then just trowing them away again
>> immediately.
>>
>> We mitigated the problem by avoiding the slow coherent DMA code path on
>> almost all platforms on newer kernels, but essentially somebody needs to
>> figure out why firefox and/or the user space stack is doing this
>> constant allocation/freeing of memory.
>>
>> There is also a bug tracker on bugs.kernel.org about this, but I can't
>> find it any more of hand.
>>
>> Regards,
>> Christian.
>>
>> Am 06.04.2018 um 02:30 schrieb Jean-Marc Valin:
>>> Hi,
>>>
>>> I noticed a serious graphics performance regression between 4.14 and
>>> 4.15. It is most noticeable with Firefox (tried FF57 through FF60) and
>>> causes scrolling to be really choppy/sluggish. I've confirmed that the
>>> problem is also there on 4.16, while 4.13 works fine.
>>>
>>> After a bisection, I've narrowed the regression down to this commit:
>>>
>>> commit 648bc3574716400acc06f99915815f80d9563783
>>> Author: Christian König <[email protected]>
>>> Date:   Thu Jul 6 09:59:43 2017 +0200
>>>
>>>      drm/ttm: add transparent huge page support for DMA allocations v2
>>>
>>>
>>> Some details about my system:
>>> Distro: Fedora 27 (up-to-date)
>>> Video: MSI Radeon RX 560 AERO
>>> CPU: Dual-socket Xeon E5-2640 v4 (20 cores total)
>>> RAM: 128 GB ECC
>>>
>>>
>>> As a comparison, when running Firefox with 4.15 on a Lenovo W540 laptop
>>> (with Intel graphics only) the responsiveness is much better then what
>>> I'm getting on the Xeon machine above with the Radeon card, so this
>>> really seems to be an AMD-only issue.
>>>
>>> Any way to fix the issue?
>>>
>>> Thanks,
>>>
>>>     Jean-Marc
>>> _______________________________________________
>>> dri-devel mailing list
>>> [email protected]
>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel

2018-04-06 16:43:52

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Hi Christian,

On 04/09/2018 07:48 AM, Christian König wrote:
> Am 06.04.2018 um 17:30 schrieb Jean-Marc Valin:
>> Hi Christian,
>>
>> Is there a way to turn off these huge pages at boot-time/run-time?
>
> Only at compile time by not setting CONFIG_TRANSPARENT_HUGEPAGE.

Any reason why
echo never > /sys/kernel/mm/transparent_hugepage/enabled
doesn't solve the problem?

Also, I assume that disabling CONFIG_TRANSPARENT_HUGEPAGE will disable
them for everything and not just what your patch added, right?

>> I'm not sure what you mean by "We mitigated the problem by avoiding the
>> slow coherent DMA code path on almost all platforms on newer kernels". I
>> tested up to 4.16 and the performance regression is just as bad as it is
>> for 4.15.
>
> Indeed 4.16 still doesn't have that. You could use the
> amd-staging-drm-next branch or wait for 4.17.

Is there a way to pull just that change or is there too much
interactions with other changes?

> That isn't related to the GFX hardware, but to your CPU/motherboard and
> whatever else you have in the system.

Well, I have an nvidia GPU in the same system (normally only used for
CUDA) and if I use it instead of my RX 560 then I'm not seeing any
performance issue with 4.15.

> Some part of your system needs SWIOTLB and that makes allocating memory
> much slower.

What would that part be? FTR, I have a complete description of my system
at https://jmvalin.dreamwidth.org/15583.html

I don't know if it's related, but I can maybe see one thing in common
between my machine and the Core 2 Quad from the other bug report and
that's the "NUMA part". I have a dual-socket Xeon and (AFAIK) the Core 2
Quad is made of two two-core CPUs glued together with little
communication between them.

> Intel doesn't use TTM because they don't have dedicated VRAM, but the
> open source nvidia driver should be affected as well.

I'm using the proprietary nvidia driver (because CUDA). Is that supposed
to be affected as well?

> We already mitigated that problem and I don't see any solution which
> will arrive faster than 4.17.

Is that supposed to make the slowdown unnoticeable or just slightly better?

> The only quick workaround I can see is to avoid firefox, chrome for
> example is reported to work perfectly fine.

Or use an unaffected GPU/driver ;-)

Cheers,

Jean-Marc

2018-04-06 17:24:13

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 06.04.2018 um 18:42 schrieb Jean-Marc Valin:
> Hi Christian,
>
> On 04/09/2018 07:48 AM, Christian König wrote:
>> Am 06.04.2018 um 17:30 schrieb Jean-Marc Valin:
>>> Hi Christian,
>>>
>>> Is there a way to turn off these huge pages at boot-time/run-time?
>> Only at compile time by not setting CONFIG_TRANSPARENT_HUGEPAGE.
> Any reason why
> echo never > /sys/kernel/mm/transparent_hugepage/enabled
> doesn't solve the problem?

Because we unfortunately try to allocate huge pages anyway, we
unfortunately just fail in 100% of all cases.

That basically gives you both, the extra allocation overhead and the
still bad throughput.

> Also, I assume that disabling CONFIG_TRANSPARENT_HUGEPAGE will disable
> them for everything and not just what your patch added, right?

Correct, that's why I wrote that disabling SWIOTLBs might be better.

>>> I'm not sure what you mean by "We mitigated the problem by avoiding the
>>> slow coherent DMA code path on almost all platforms on newer kernels". I
>>> tested up to 4.16 and the performance regression is just as bad as it is
>>> for 4.15.
>> Indeed 4.16 still doesn't have that. You could use the
>> amd-staging-drm-next branch or wait for 4.17.
> Is there a way to pull just that change or is there too much
> interactions with other changes?

It adds a new detection if memory allocation needs to be coherent or
not, that is not something you can easily pull into older versions.

>> That isn't related to the GFX hardware, but to your CPU/motherboard and
>> whatever else you have in the system.
> Well, I have an nvidia GPU in the same system (normally only used for
> CUDA) and if I use it instead of my RX 560 then I'm not seeing any
> performance issue with 4.15.

That's because you are probably using the Nvidia binary driver which has
a completely separate code base.

>> Some part of your system needs SWIOTLB and that makes allocating memory
>> much slower.
> What would that part be? FTR, I have a complete description of my system
> at https://jmvalin.dreamwidth.org/15583.html
>
> I don't know if it's related, but I can maybe see one thing in common
> between my machine and the Core 2 Quad from the other bug report and
> that's the "NUMA part". I have a dual-socket Xeon and (AFAIK) the Core 2
> Quad is made of two two-core CPUs glued together with little
> communication between them.

Yeah, that is probably the reason.

>> Intel doesn't use TTM because they don't have dedicated VRAM, but the
>> open source nvidia driver should be affected as well.
> I'm using the proprietary nvidia driver (because CUDA). Is that supposed
> to be affected as well?

No.

>> We already mitigated that problem and I don't see any solution which
>> will arrive faster than 4.17.
> Is that supposed to make the slowdown unnoticeable or just slightly better?

It completely goes away. The issue with the coherent path is that it
tries to always allocate the lowest possible memory to make sure that it
fits into the DMA constrains of all devices in the system.

But since AMD GPU can handle 40bits of addresses you would need at least
1TB of memory in the system to trigger that (or a NUMA where some system
is low and some in a high area).

Christian.

>> The only quick workaround I can see is to avoid firefox, chrome for
>> example is reported to work perfectly fine.
> Or use an unaffected GPU/driver ;-)
>
> Cheers,
>
> Jean-Marc
>

2018-04-06 22:04:30

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Hi Christian,

Thanks for the info. FYI, I've also opened a Firefox bug for that at:
https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
Feel free to comment since you have a better understanding of what's
going on.

One last question: right now I'm running 4.15.0 with the "offending"
patch reverted. Is that safe to run or are there possible bad
interactions with other changes.

Cheers,

Jean-Marc

On 04/06/2018 01:20 PM, Christian König wrote:
> Am 06.04.2018 um 18:42 schrieb Jean-Marc Valin:
>> Hi Christian,
>>
>> On 04/09/2018 07:48 AM, Christian König wrote:
>>> Am 06.04.2018 um 17:30 schrieb Jean-Marc Valin:
>>>> Hi Christian,
>>>>
>>>> Is there a way to turn off these huge pages at boot-time/run-time?
>>> Only at compile time by not setting CONFIG_TRANSPARENT_HUGEPAGE.
>> Any reason why
>> echo never > /sys/kernel/mm/transparent_hugepage/enabled
>> doesn't solve the problem?
>
> Because we unfortunately try to allocate huge pages anyway, we
> unfortunately just fail in 100% of all cases.
>
> That basically gives you both, the extra allocation overhead and the
> still bad throughput.
>
>> Also, I assume that disabling CONFIG_TRANSPARENT_HUGEPAGE will disable
>> them for everything and not just what your patch added, right?
>
> Correct, that's why I wrote that disabling SWIOTLBs might be better.
>
>>>> I'm not sure what you mean by "We mitigated the problem by avoiding the
>>>> slow coherent DMA code path on almost all platforms on newer
>>>> kernels". I
>>>> tested up to 4.16 and the performance regression is just as bad as
>>>> it is
>>>> for 4.15.
>>> Indeed 4.16 still doesn't have that. You could use the
>>> amd-staging-drm-next branch or wait for 4.17.
>> Is there a way to pull just that change or is there too much
>> interactions with other changes?
>
> It adds a new detection if memory allocation needs to be coherent or
> not, that is not something you can easily pull into older versions.
>
>>> That isn't related to the GFX hardware, but to your CPU/motherboard and
>>> whatever else you have in the system.
>> Well, I have an nvidia GPU in the same system (normally only used for
>> CUDA) and if I use it instead of my RX 560 then I'm not seeing any
>> performance issue with 4.15.
>
> That's because you are probably using the Nvidia binary driver which has
> a completely separate code base.
>
>>> Some part of your system needs SWIOTLB and that makes allocating memory
>>> much slower.
>> What would that part be? FTR, I have a complete description of my system
>> at https://jmvalin.dreamwidth.org/15583.html
>>
>> I don't know if it's related, but I can maybe see one thing in common
>> between my machine and the Core 2 Quad from the other bug report and
>> that's the "NUMA part". I have a dual-socket Xeon and (AFAIK) the Core 2
>> Quad is made of two two-core CPUs glued together with little
>> communication between them.
>
> Yeah, that is probably the reason.
>
>>> Intel doesn't use TTM because they don't have dedicated VRAM, but the
>>> open source nvidia driver should be affected as well.
>> I'm using the proprietary nvidia driver (because CUDA). Is that supposed
>> to be affected as well?
>
> No.
>
>>> We already mitigated that problem and I don't see any solution which
>>> will arrive faster than 4.17.
>> Is that supposed to make the slowdown unnoticeable or just slightly
>> better?
>
> It completely goes away. The issue with the coherent path is that it
> tries to always allocate the lowest possible memory to make sure that it
> fits into the DMA constrains of all devices in the system.
>
> But since AMD GPU can handle 40bits of addresses you would need at least
> 1TB of memory in the system to trigger that (or a NUMA where some system
> is low and some in a high area).
>
> Christian.
>
>>> The only quick workaround I can see is to avoid firefox, chrome for
>>> example is reported to work perfectly fine.
>> Or use an unaffected GPU/driver ;-)
>>
>> Cheers,
>>
>> Jean-Marc
>>
>

2018-04-09 09:46:37

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
> Hi Christian,
>
> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
> Feel free to comment since you have a better understanding of what's
> going on.
>
> One last question: right now I'm running 4.15.0 with the "offending"
> patch reverted. Is that safe to run or are there possible bad
> interactions with other changes.

That should work without problems.

But I just had another idea as well, if you want you could still test
the new code path which will be using in 4.17.

Backporting all the detection logic is to invasive, but you could just
go into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the
other code path.

Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.

Regards,
Christian.

>
> Cheers,
>
> Jean-Marc
>
> On 04/06/2018 01:20 PM, Christian König wrote:
>> Am 06.04.2018 um 18:42 schrieb Jean-Marc Valin:
>>> Hi Christian,
>>>
>>> On 04/09/2018 07:48 AM, Christian König wrote:
>>>> Am 06.04.2018 um 17:30 schrieb Jean-Marc Valin:
>>>>> Hi Christian,
>>>>>
>>>>> Is there a way to turn off these huge pages at boot-time/run-time?
>>>> Only at compile time by not setting CONFIG_TRANSPARENT_HUGEPAGE.
>>> Any reason why
>>> echo never > /sys/kernel/mm/transparent_hugepage/enabled
>>> doesn't solve the problem?
>> Because we unfortunately try to allocate huge pages anyway, we
>> unfortunately just fail in 100% of all cases.
>>
>> That basically gives you both, the extra allocation overhead and the
>> still bad throughput.
>>
>>> Also, I assume that disabling CONFIG_TRANSPARENT_HUGEPAGE will disable
>>> them for everything and not just what your patch added, right?
>> Correct, that's why I wrote that disabling SWIOTLBs might be better.
>>
>>>>> I'm not sure what you mean by "We mitigated the problem by avoiding the
>>>>> slow coherent DMA code path on almost all platforms on newer
>>>>> kernels". I
>>>>> tested up to 4.16 and the performance regression is just as bad as
>>>>> it is
>>>>> for 4.15.
>>>> Indeed 4.16 still doesn't have that. You could use the
>>>> amd-staging-drm-next branch or wait for 4.17.
>>> Is there a way to pull just that change or is there too much
>>> interactions with other changes?
>> It adds a new detection if memory allocation needs to be coherent or
>> not, that is not something you can easily pull into older versions.
>>
>>>> That isn't related to the GFX hardware, but to your CPU/motherboard and
>>>> whatever else you have in the system.
>>> Well, I have an nvidia GPU in the same system (normally only used for
>>> CUDA) and if I use it instead of my RX 560 then I'm not seeing any
>>> performance issue with 4.15.
>> That's because you are probably using the Nvidia binary driver which has
>> a completely separate code base.
>>
>>>> Some part of your system needs SWIOTLB and that makes allocating memory
>>>> much slower.
>>> What would that part be? FTR, I have a complete description of my system
>>> at https://jmvalin.dreamwidth.org/15583.html
>>>
>>> I don't know if it's related, but I can maybe see one thing in common
>>> between my machine and the Core 2 Quad from the other bug report and
>>> that's the "NUMA part". I have a dual-socket Xeon and (AFAIK) the Core 2
>>> Quad is made of two two-core CPUs glued together with little
>>> communication between them.
>> Yeah, that is probably the reason.
>>
>>>> Intel doesn't use TTM because they don't have dedicated VRAM, but the
>>>> open source nvidia driver should be affected as well.
>>> I'm using the proprietary nvidia driver (because CUDA). Is that supposed
>>> to be affected as well?
>> No.
>>
>>>> We already mitigated that problem and I don't see any solution which
>>>> will arrive faster than 4.17.
>>> Is that supposed to make the slowdown unnoticeable or just slightly
>>> better?
>> It completely goes away. The issue with the coherent path is that it
>> tries to always allocate the lowest possible memory to make sure that it
>> fits into the DMA constrains of all devices in the system.
>>
>> But since AMD GPU can handle 40bits of addresses you would need at least
>> 1TB of memory in the system to trigger that (or a NUMA where some system
>> is low and some in a high area).
>>
>> Christian.
>>
>>>> The only quick workaround I can see is to avoid firefox, chrome for
>>>> example is reported to work perfectly fine.
>>> Or use an unaffected GPU/driver ;-)
>>>
>>> Cheers,
>>>
>>> Jean-Marc
>>>
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

2018-04-09 15:21:51

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

On 04/09/2018 05:42 AM, Christian König wrote:
> Backporting all the detection logic is to invasive, but you could just
> go into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the
> other code path.
>
> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.

Do you mean just taking the 4.15 code as is and replacing
"#ifdef CONFIG_SWIOTLB" with "#if 0" in
drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c or are you talking about using a
different version of drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c ?

Jean-Marc

2018-04-10 06:51:52

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 09.04.2018 um 17:17 schrieb Jean-Marc Valin:
> On 04/09/2018 05:42 AM, Christian König wrote:
>> Backporting all the detection logic is to invasive, but you could just
>> go into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the
>> other code path.
>>
>> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
> Do you mean just taking the 4.15 code as is and replacing
> "#ifdef CONFIG_SWIOTLB" with "#if 0" in
> drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c or are you talking about using a
> different version of drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c ?

Yes, exactly. The code then won't work any more on some ARMs or systems
with more than 1TB of memory, but I don't think you care about that :)

Christian.

>
> Jean-Marc
> _______________________________________________
> dri-devel mailing list
> [email protected]
> https://lists.freedesktop.org/mailman/listinfo/dri-devel

2018-04-11 04:06:08

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-04-09 11:42 GMT+02:00 Christian König <[email protected]>:
> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>
>> Hi Christian,
>>
>> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
>> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
>> Feel free to comment since you have a better understanding of what's
>> going on.
>>
>> One last question: right now I'm running 4.15.0 with the "offending"
>> patch reverted. Is that safe to run or are there possible bad
>> interactions with other changes.
>
>
> That should work without problems.
>
> But I just had another idea as well, if you want you could still test the
> new code path which will be using in 4.17.
>

While Firefox may do some strange things is not about only Firefox.

With your patches my EPYC box is unusable with 4.15++ kernels.
The whole Desktop is acting weird. This one is using
an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.

Box is 2 * EPYC 7281 with 128 GB ECC RAM

Also a 14C Xeon box with a HD7700 is broken same way.

Everything breaks in X .. scrolling , moving windows , flickering etc.

reverting f4c809914a7c3e4a59cf543da6c2a15d0f75ee38 and
648bc3574716400acc06f99915815f80d9563783
from an 4.15 kernel makes things work again.

> Backporting all the detection logic is to invasive, but you could just go
> into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the other
> code path.
>
> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
>

Well you really can't be serious about these suggestions ? Are you ?

Telling peoples to #if 0 random code is not a solution.

You broke existsing working userland with your patches and at least
please fix that for 4.16.

I can help testing code for 4.17/++ if you wish but that is *different* storry.

Regards,

Gabriel C

2018-04-11 05:06:56

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

>2018-04-11 6:00 GMT+02:00 Gabriel C <[email protected]>:
> 2018-04-09 11:42 GMT+02:00 Christian König <[email protected]>:
>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
...
> I can help testing code for 4.17/++ if you wish but that is *different* storry.
>

Quick tested an 4.16.0-11490-gb284d4d5a678 , amdgpu and radeon driver
are broken now in this one.

radeon tells:

...

[ 6.337838] [drm] PCIE GART of 2048M enabled (table at 0x00000000001D6000).
[ 6.338210] radeon 0000:21:00.0: (-12) create WB bo failed
[ 6.338214] radeon 0000:21:00.0: disabling GPU acceleration

...

And no way to start X .. flickering and hangs.

amdgpu hits an bug:

http://ftp.frugalware.org/pub/other/people/crazy/trace.txt

Do you have some git tree I can test from ?

Also if you need full , logs or any other infos just let me know.

Regards

2018-04-11 09:42:07

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 11.04.2018 um 06:00 schrieb Gabriel C:
> 2018-04-09 11:42 GMT+02:00 Christian König <[email protected]>:
>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>> Hi Christian,
>>>
>>> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
>>> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
>>> Feel free to comment since you have a better understanding of what's
>>> going on.
>>>
>>> One last question: right now I'm running 4.15.0 with the "offending"
>>> patch reverted. Is that safe to run or are there possible bad
>>> interactions with other changes.
>>
>> That should work without problems.
>>
>> But I just had another idea as well, if you want you could still test the
>> new code path which will be using in 4.17.
>>
> While Firefox may do some strange things is not about only Firefox.
>
> With your patches my EPYC box is unusable with 4.15++ kernels.
> The whole Desktop is acting weird. This one is using
> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>
> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>
> Also a 14C Xeon box with a HD7700 is broken same way.

The hardware is irrelevant for this. We need to know what software stack
you use on top of it.

E.g. desktop environment/Mesa and DDX version etc...

>
> Everything breaks in X .. scrolling , moving windows , flickering etc.
>
>
> reverting f4c809914a7c3e4a59cf543da6c2a15d0f75ee38 and
> 648bc3574716400acc06f99915815f80d9563783
> from an 4.15 kernel makes things work again.
>
>
>> Backporting all the detection logic is to invasive, but you could just go
>> into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the other
>> code path.
>>
>> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
>>
> Well you really can't be serious about these suggestions ? Are you ?
>
> Telling peoples to #if 0 random code is not a solution.

That is for testing and not a permanent solution.

> You broke existsing working userland with your patches and at least
> please fix that for 4.16.
>
> I can help testing code for 4.17/++ if you wish but that is *different* storry.

Please test Alex's amd-staging-drm-next branch from
git://people.freedesktop.org/~agd5f/linux.

Regards,
Christian.

>
> Regards,
>
> Gabriel C

2018-04-11 14:31:04

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-04-11 11:37 GMT+02:00 Christian König <[email protected]>:
> Am 11.04.2018 um 06:00 schrieb Gabriel C:
>>
>> 2018-04-09 11:42 GMT+02:00 Christian König
>> <[email protected]>:
>>>
>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>>>
>>>> Hi Christian,
>>>>
>>>> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
>>>> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
>>>> Feel free to comment since you have a better understanding of what's
>>>> going on.
>>>>
>>>> One last question: right now I'm running 4.15.0 with the "offending"
>>>> patch reverted. Is that safe to run or are there possible bad
>>>> interactions with other changes.
>>>
>>>
>>> That should work without problems.
>>>
>>> But I just had another idea as well, if you want you could still test the
>>> new code path which will be using in 4.17.
>>>
>> While Firefox may do some strange things is not about only Firefox.
>>
>> With your patches my EPYC box is unusable with 4.15++ kernels.
>> The whole Desktop is acting weird. This one is using
>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>
>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>
>> Also a 14C Xeon box with a HD7700 is broken same way.
>
>
> The hardware is irrelevant for this. We need to know what software stack you
> use on top of it.
>
> E.g. desktop environment/Mesa and DDX version etc...

Plasma 5.12.4 compiled wth frameworks 5.44.0 , Qt5 5.10.1
mesa 18.0.0 and mesa 17.3.7 on the other box
Xorg is 1.19.6
xf86-video-amdgpu and xf86-video-ati both 18.0.1

>
>>
>> Everything breaks in X .. scrolling , moving windows , flickering etc.
>>
>>
>> reverting f4c809914a7c3e4a59cf543da6c2a15d0f75ee38 and
>> 648bc3574716400acc06f99915815f80d9563783
>> from an 4.15 kernel makes things work again.
>>
>>
>>> Backporting all the detection logic is to invasive, but you could just go
>>> into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the other
>>> code path.
>>>
>>> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
>>>
>> Well you really can't be serious about these suggestions ? Are you ?
>>
>> Telling peoples to #if 0 random code is not a solution.
>
>
> That is for testing and not a permanent solution.
>
>> You broke existsing working userland with your patches and at least
>> please fix that for 4.16.
>>
>> I can help testing code for 4.17/++ if you wish but that is *different*
>> storry.
>
>
> Please test Alex's amd-staging-drm-next branch from
> git://people.freedesktop.org/~agd5f/linux.

I'm on it just the connection to freedesktop.org is slow as hell.
Will take a while to get that branch with 62KiB/s :)

Regards

2018-04-11 17:25:41

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-04-11 16:26 GMT+02:00 Gabriel C <[email protected]>:
> 2018-04-11 11:37 GMT+02:00 Christian König <[email protected]>:
>> Am 11.04.2018 um 06:00 schrieb Gabriel C:

...
>>
>> Please test Alex's amd-staging-drm-next branch from
>> git://people.freedesktop.org/~agd5f/linux.
>
> I'm on it just the connection to freedesktop.org is slow as hell.
> Will take a while to get that branch with 62KiB/s :)
>

Testing done on that branch on commit 24110c70630998dc83da23cae910a9538f54ef64.

On default Plasma OpenGL 2.0 profiles things are still laggy but a lot better.
On OpenGL 3.1 things are working much better just minor gliches on
maximzing/minimizing windows.

Firefox is still broken , frames drops , video stops etc
Cromium-browser works fine
Otter-browser does not work at all
Qupzilla/Falkon has Firefox like issues too

Things I noticed while testing Firefox or Qupzilla..
Once these start acting up it does affect the whole Desktop,
for some secons scrolling lags , mouse is slow , etc.
Once these are closed the Desktop start working again after few seconds.

Do you want me to test any mesa/xorg-server/drivers git/branches too ?

If so just let me know.

Regards

2018-04-11 18:38:50

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

On 04/11/2018 05:37 AM, Christian König wrote:
>> With your patches my EPYC box is unusable with 4.15++ kernels.
>> The whole Desktop is acting weird. This one is using
>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>
>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>
>> Also a 14C Xeon box with a HD7700 is broken same way.
>
> The hardware is irrelevant for this. We need to know what software stack
> you use on top of it.

Well, the hardware appears to be part of the issue too. I don't think
it's a coincidence that Gabriel has the problem on 2xEPYC, I have it on
2xXeon and the previous reported had it on a Core 2 Quad that internally
has two dies.

I've not yet tested your disable CONFIG_SWIOTLB fix yet -- might try it
over the weekend and report what happens.

Cheers,

Jean-Marc

2018-04-11 22:25:18

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-04-11 20:35 GMT+02:00 Jean-Marc Valin <[email protected]>:
> On 04/11/2018 05:37 AM, Christian König wrote:
>>> With your patches my EPYC box is unusable with 4.15++ kernels.
>>> The whole Desktop is acting weird. This one is using
>>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>>
>>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>>
>>> Also a 14C Xeon box with a HD7700 is broken same way.
>>
>> The hardware is irrelevant for this. We need to know what software stack
>> you use on top of it.
>
> Well, the hardware appears to be part of the issue too. I don't think
> it's a coincidence that Gabriel has the problem on 2xEPYC, I have it on
> 2xXeon and the previous reported had it on a Core 2 Quad that internally
> has two dies.
>
> I've not yet tested your disable CONFIG_SWIOTLB fix yet -- might try it
> over the weekend and report what happens.
>

To get that right .. is only a matter of disabling SWIOTLB *code*
while CONFIG_SWIOTLB is still set ?

2018-04-12 01:50:49

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-04-12 0:20 GMT+02:00 Gabriel C <[email protected]>:
> 2018-04-11 20:35 GMT+02:00 Jean-Marc Valin <[email protected]>:
>> On 04/11/2018 05:37 AM, Christian König wrote:
>>>> With your patches my EPYC box is unusable with 4.15++ kernels.
>>>> The whole Desktop is acting weird. This one is using
>>>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>>>
>>>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>>>
>>>> Also a 14C Xeon box with a HD7700 is broken same way.
>>>
>>> The hardware is irrelevant for this. We need to know what software stack
>>> you use on top of it.
>>
>> Well, the hardware appears to be part of the issue too. I don't think
>> it's a coincidence that Gabriel has the problem on 2xEPYC, I have it on
>> 2xXeon and the previous reported had it on a Core 2 Quad that internally
>> has two dies.
>>
>> I've not yet tested your disable CONFIG_SWIOTLB fix yet -- might try it
>> over the weekend and report what happens.
>>
>
> To get that right .. is only a matter of disabling SWIOTLB *code*
> while CONFIG_SWIOTLB is still set ?

Ok I tested that on 4.16.1 and yes it does work. However I didn't like the
#if 0 method and so compile an kernel twice just to compare an test.

I created an small patch and added swiotlb option for amdgpu and radeon
so I can boot and compare / test with and without SWIOTLB code.

( not meant for upstream )

http://ftp.frugalware.org/pub/other/people/crazy/0001-Make-it-possible-to-disable-SWIOTLB-code-on-admgpu-a.patch

With SWIOTLB code off all works fine , while hell breaks when turning on.

Maybe similar options should be added upstream until code is more
stable in 4.17/4.18

Regards

2018-04-20 14:50:07

by Michel Dänzer

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

On 2018-04-11 11:37 AM, Christian König wrote:
> Am 11.04.2018 um 06:00 schrieb Gabriel C:
>> 2018-04-09 11:42 GMT+02:00 Christian König
>> <[email protected]>:
>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>>> Hi Christian,
>>>>
>>>> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
>>>> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
>>>> Feel free to comment since you have a better understanding of what's
>>>> going on.
>>>>
>>>> One last question: right now I'm running 4.15.0 with the "offending"
>>>> patch reverted. Is that safe to run or are there possible bad
>>>> interactions with other changes.
>>>
>>> That should work without problems.
>>>
>>> But I just had another idea as well, if you want you could still test
>>> the
>>> new code path which will be using in 4.17.
>>>
>> While Firefox may do some strange things is not about only Firefox.
>>
>> With your patches my EPYC box is unusable with 4.15++ kernels.
>> The whole Desktop is acting weird. This one is using
>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>
>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>
>> Also a 14C Xeon box with a HD7700 is broken same way.
>
> The hardware is irrelevant for this. We need to know what software stack
> you use on top of it.
>
> E.g. desktop environment/Mesa and DDX version etc...
>
>>
>> Everything breaks in X .. scrolling , moving windows , flickering etc.
>>
>>
>> reverting f4c809914a7c3e4a59cf543da6c2a15d0f75ee38 and
>> 648bc3574716400acc06f99915815f80d9563783
>> from an 4.15 kernel makes things work again.
>>
>>
>>> Backporting all the detection logic is to invasive, but you could
>>> just go
>>> into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the other
>>> code path.
>>>
>>> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
>>>
>> Well you really can't be serious about these suggestions ? Are you ?
>>
>> Telling peoples to #if 0 random code is not a solution.
>
> That is for testing and not a permanent solution.
>
>> You broke existsing working userland with your patches and at least
>> please fix that for 4.16.
>>
>> I can help testing code for 4.17/++ if you wish but that is
>> *different* storry.
>
> Please test Alex's amd-staging-drm-next branch from
> git://people.freedesktop.org/~agd5f/linux.

I think we're still missing something here.

I'm currently running 4.16.2 + the DRM subsystem changes which are going
into 4.17 (so I have the changes Christian is referring to) with a
Kaveri APU, and I'm seeing similar symptoms as described by Jean-Marc.
Some observations:

Firefox, Thunderbird, or worst, gnome-shell, can freeze for up to on the
order of a minute, during which the kernel is spending most of one
core's cycles inside alloc_pages (__alloc_pages_nodemask to be more
precise), called from ttm_alloc_new_pages.

At least in the case of Firefox, this happens due to Mesa internal BO
allocations for glTex(Sub)Image, so it's not obvious that Firefox is
doing something wrong.

I never noticed this before this week. Before, I was running 4.15.y +
DRM subsystem changes from 4.16. Maybe something has changed in core
code, trying harder to allocate huge pages.

Maybe TTM should only try to use any huge pages that happen to be
available, not spend any (/ "too much"?) additional effort trying to
free up huge pages?

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

2018-04-20 19:47:28

by Felix Kuehling

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

[+Philip]

On 2018-04-20 10:47 AM, Michel Dänzer wrote:
> On 2018-04-11 11:37 AM, Christian König wrote:
>> Am 11.04.2018 um 06:00 schrieb Gabriel C:
>>> 2018-04-09 11:42 GMT+02:00 Christian König
>>> <[email protected]>:
>>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>>>> Hi Christian,
>>>>>
>>>>> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
>>>>> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
>>>>> Feel free to comment since you have a better understanding of what's
>>>>> going on.
>>>>>
>>>>> One last question: right now I'm running 4.15.0 with the "offending"
>>>>> patch reverted. Is that safe to run or are there possible bad
>>>>> interactions with other changes.
>>>> That should work without problems.
>>>>
>>>> But I just had another idea as well, if you want you could still test
>>>> the
>>>> new code path which will be using in 4.17.
>>>>
>>> While Firefox may do some strange things is not about only Firefox.
>>>
>>> With your patches my EPYC box is unusable with 4.15++ kernels.
>>> The whole Desktop is acting weird. This one is using
>>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>>
>>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>>
>>> Also a 14C Xeon box with a HD7700 is broken same way.
>> The hardware is irrelevant for this. We need to know what software stack
>> you use on top of it.
>>
>> E.g. desktop environment/Mesa and DDX version etc...
>>
>>> Everything breaks in X .. scrolling , moving windows , flickering etc.
>>>
>>>
>>> reverting f4c809914a7c3e4a59cf543da6c2a15d0f75ee38 and
>>> 648bc3574716400acc06f99915815f80d9563783
>>> from an 4.15 kernel makes things work again.
>>>
>>>
>>>> Backporting all the detection logic is to invasive, but you could
>>>> just go
>>>> into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the other
>>>> code path.
>>>>
>>>> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
>>>>
>>> Well you really can't be serious about these suggestions ? Are you ?
>>>
>>> Telling peoples to #if 0 random code is not a solution.
>> That is for testing and not a permanent solution.
>>
>>> You broke existsing working userland with your patches and at least
>>> please fix that for 4.16.
>>>
>>> I can help testing code for 4.17/++ if you wish but that is
>>> *different* storry.
>> Please test Alex's amd-staging-drm-next branch from
>> git://people.freedesktop.org/~agd5f/linux.
> I think we're still missing something here.
>
> I'm currently running 4.16.2 + the DRM subsystem changes which are going
> into 4.17 (so I have the changes Christian is referring to) with a
> Kaveri APU, and I'm seeing similar symptoms as described by Jean-Marc.
> Some observations:
>
> Firefox, Thunderbird, or worst, gnome-shell, can freeze for up to on the
> order of a minute, during which the kernel is spending most of one
> core's cycles inside alloc_pages (__alloc_pages_nodemask to be more
> precise), called from ttm_alloc_new_pages.
Philip debugged a similar problem with a KFD memory stress test about
two weeks ago, where the kernel was seemingly stuck in an infinite loop
trying to allocate huge pages. I'm pasting his analysis for the record:

> [...] it uses huge_flags GFP_TRANSHUGE to call alloc_pages(), this
> seems a corner case inside __alloc_pages_slowpath(), it never exits
> but goes to retry path every time. It can reclaim pages and
> did_some_progress (as a result, no_progress_loops is reset to 0 every
> loop, never reach MAX_RECLAIM_RETRIES) but cannot finish huge page
> allocations under this specific memory pressure.
As a workaround to unblock our release branch testing we removed
transparent huge page allocation from ttm_get_pages. We're seeing this
as far back as 4.13 on our release branch.

If we're really talking about the same problem, I don't think it's
caused by recent page allocator changes, but rather exposed by recent
TTM changes.

Regards,
Felix

>
> At least in the case of Firefox, this happens due to Mesa internal BO
> allocations for glTex(Sub)Image, so it's not obvious that Firefox is
> doing something wrong.
>
> I never noticed this before this week. Before, I was running 4.15.y +
> DRM subsystem changes from 4.16. Maybe something has changed in core
> code, trying harder to allocate huge pages.
>
>
> Maybe TTM should only try to use any huge pages that happen to be
> available, not spend any (/ "too much"?) additional effort trying to
> free up huge pages?
>
>

2018-04-23 10:25:18

by Michel Dänzer

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

On 2018-04-20 09:40 PM, Felix Kuehling wrote:
> On 2018-04-20 10:47 AM, Michel Dänzer wrote:
>> On 2018-04-11 11:37 AM, Christian König wrote:
>>> Am 11.04.2018 um 06:00 schrieb Gabriel C:
>>>> 2018-04-09 11:42 GMT+02:00 Christian König
>>>> <[email protected]>:
>>>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>>>>> Hi Christian,
>>>>>>
>>>>>> Thanks for the info. FYI, I've also opened a Firefox bug for that at:
>>>>>> https://bugzilla.mozilla.org/show_bug.cgi?id=1448778
>>>>>> Feel free to comment since you have a better understanding of what's
>>>>>> going on.
>>>>>>
>>>>>> One last question: right now I'm running 4.15.0 with the "offending"
>>>>>> patch reverted. Is that safe to run or are there possible bad
>>>>>> interactions with other changes.
>>>>> That should work without problems.
>>>>>
>>>>> But I just had another idea as well, if you want you could still test
>>>>> the
>>>>> new code path which will be using in 4.17.
>>>>>
>>>> While Firefox may do some strange things is not about only Firefox.
>>>>
>>>> With your patches my EPYC box is unusable with 4.15++ kernels.
>>>> The whole Desktop is acting weird. This one is using
>>>> an Cape Verde PRO [Radeon HD 7750/8740 / R7 250E] GPU.
>>>>
>>>> Box is 2 * EPYC 7281 with 128 GB ECC RAM
>>>>
>>>> Also a 14C Xeon box with a HD7700 is broken same way.
>>> The hardware is irrelevant for this. We need to know what software stack
>>> you use on top of it.
>>>
>>> E.g. desktop environment/Mesa and DDX version etc...
>>>
>>>> Everything breaks in X .. scrolling , moving windows , flickering etc.
>>>>
>>>>
>>>> reverting f4c809914a7c3e4a59cf543da6c2a15d0f75ee38 and
>>>> 648bc3574716400acc06f99915815f80d9563783
>>>> from an 4.15 kernel makes things work again.
>>>>
>>>>
>>>>> Backporting all the detection logic is to invasive, but you could
>>>>> just go
>>>>> into drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c and forcefull use the other
>>>>> code path.
>>>>>
>>>>> Just look out for "#ifdef CONFIG_SWIOTLB" checks and disable those.
>>>>>
>>>> Well you really can't be serious about these suggestions ? Are you ?
>>>>
>>>> Telling peoples to #if 0 random code is not a solution.
>>> That is for testing and not a permanent solution.
>>>
>>>> You broke existsing working userland with your patches and at least
>>>> please fix that for 4.16.
>>>>
>>>> I can help testing code for 4.17/++ if you wish but that is
>>>> *different* storry.
>>> Please test Alex's amd-staging-drm-next branch from
>>> git://people.freedesktop.org/~agd5f/linux.
>> I think we're still missing something here.
>>
>> I'm currently running 4.16.2 + the DRM subsystem changes which are going
>> into 4.17 (so I have the changes Christian is referring to) with a
>> Kaveri APU, and I'm seeing similar symptoms as described by Jean-Marc.
>> Some observations:
>>
>> Firefox, Thunderbird, or worst, gnome-shell, can freeze for up to on the
>> order of a minute, during which the kernel is spending most of one
>> core's cycles inside alloc_pages (__alloc_pages_nodemask to be more
>> precise), called from ttm_alloc_new_pages.
> Philip debugged a similar problem with a KFD memory stress test about
> two weeks ago, where the kernel was seemingly stuck in an infinite loop
> trying to allocate huge pages. I'm pasting his analysis for the record:
>
>> [...] it uses huge_flags GFP_TRANSHUGE to call alloc_pages(), this
>> seems a corner case inside __alloc_pages_slowpath(), it never exits
>> but goes to retry path every time. It can reclaim pages and
>> did_some_progress (as a result, no_progress_loops is reset to 0 every
>> loop, never reach MAX_RECLAIM_RETRIES) but cannot finish huge page
>> allocations under this specific memory pressure.
> As a workaround to unblock our release branch testing we removed
> transparent huge page allocation from ttm_get_pages. We're seeing this
> as far back as 4.13 on our release branch.

Thanks for sharing this. In the future, please raise issues like this on
the public mailing lists from the beginning.

> If we're really talking about the same problem, I don't think it's
> caused by recent page allocator changes, but rather exposed by recent
> TTM changes.

It sounds related, but probably not exactly the same problem. I already
had the TTM code using GFP_TRANSHUGE before I ran into the issue. Also,
__alloc_pages_slowpath eventually succeeds for me, it can just take up
to about a minute.

I'm currently testing using (GFP_TRANSHUGE_LIGHT | __GFP_NORETRY)
instead of GFP_TRANSHUGE in TTM.

--
Earthling Michel Dänzer | http://www.amd.com
Libre software enthusiast | Mesa and X developer

2018-06-06 11:30:04

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-04-11 7:02 GMT+02:00 Gabriel C <[email protected]>:
>>2018-04-11 6:00 GMT+02:00 Gabriel C <[email protected]>:
>> 2018-04-09 11:42 GMT+02:00 Christian König <[email protected]>:
>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
> ...
>> I can help testing code for 4.17/++ if you wish but that is *different* storry.
>>
>
> Quick tested an 4.16.0-11490-gb284d4d5a678 , amdgpu and radeon driver
> are broken now in this one.
>
> radeon tells:
>
> ...
>
> [ 6.337838] [drm] PCIE GART of 2048M enabled (table at 0x00000000001D6000).
> [ 6.338210] radeon 0000:21:00.0: (-12) create WB bo failed
> [ 6.338214] radeon 0000:21:00.0: disabling GPU acceleration
>
> ...
>

I have the same Issue now on final 4.17.

Also I played with BIOS options also which does not fix anything but
changes the error message.

IOMMU && SR-IOV disabled the error changes to this :

[ 7.092044] [drm:r600_ring_test [radeon]] *ERROR* radeon: ring 0
test failed (scratch(0x850C)=0xCAFEDEAD)
[ 7.092059] radeon 0000:21:00.0: disabling GPU acceleration

While I could workaround SWIOTLB bugs in 4.15 and 4.16 , 4.17 seems to
kill the GPU with no way
for me to make it work ( at least I could not find any workaround by now )

BR

2018-06-06 11:35:23

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 06.06.2018 um 13:28 schrieb Gabriel C:
> 2018-04-11 7:02 GMT+02:00 Gabriel C <[email protected]>:
>>> 2018-04-11 6:00 GMT+02:00 Gabriel C <[email protected]>:
>>> 2018-04-09 11:42 GMT+02:00 Christian König <[email protected]>:
>>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>> ...
>>> I can help testing code for 4.17/++ if you wish but that is *different* storry.
>>>
>> Quick tested an 4.16.0-11490-gb284d4d5a678 , amdgpu and radeon driver
>> are broken now in this one.
>>
>> radeon tells:
>>
>> ...
>>
>> [ 6.337838] [drm] PCIE GART of 2048M enabled (table at 0x00000000001D6000).
>> [ 6.338210] radeon 0000:21:00.0: (-12) create WB bo failed
>> [ 6.338214] radeon 0000:21:00.0: disabling GPU acceleration
>>
>> ...
>>
> I have the same Issue now on final 4.17.

Actually Michel came up with a fix for the performance regression which
is now backported to older kernels as well.

So the original issue of this mail thread should be fixed by now.

> Also I played with BIOS options also which does not fix anything but
> changes the error message.
>
> IOMMU && SR-IOV disabled the error changes to this :
>
> [ 7.092044] [drm:r600_ring_test [radeon]] *ERROR* radeon: ring 0
> test failed (scratch(0x850C)=0xCAFEDEAD)
> [ 7.092059] radeon 0000:21:00.0: disabling GPU acceleration
>
>
> While I could workaround SWIOTLB bugs in 4.15 and 4.16 , 4.17 seems to
> kill the GPU with no way
> for me to make it work ( at least I could not find any workaround by now )

That actually sounds like something completely different. Can you
provide a full dmesg of radeon and/or amdgpu?

Thanks,
Christian.

>
> BR

2018-06-06 12:10:03

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

2018-06-06 13:33 GMT+02:00 Christian König <[email protected]>:
> Am 06.06.2018 um 13:28 schrieb Gabriel C:
>>
>> 2018-04-11 7:02 GMT+02:00 Gabriel C <[email protected]>:
>>>>
>>>> 2018-04-11 6:00 GMT+02:00 Gabriel C <[email protected]>:
>>>> 2018-04-09 11:42 GMT+02:00 Christian König
>>>> <[email protected]>:
>>>>>
>>>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>>
>>> ...
>>>>
>>>> I can help testing code for 4.17/++ if you wish but that is *different*
>>>> storry.
>>>>
>>> Quick tested an 4.16.0-11490-gb284d4d5a678 , amdgpu and radeon driver
>>> are broken now in this one.
>>>
>>> radeon tells:
>>>
>>> ...
>>>
>>> [ 6.337838] [drm] PCIE GART of 2048M enabled (table at
>>> 0x00000000001D6000).
>>> [ 6.338210] radeon 0000:21:00.0: (-12) create WB bo failed
>>> [ 6.338214] radeon 0000:21:00.0: disabling GPU acceleration
>>>
>>> ...
>>>
>> I have the same Issue now on final 4.17.
>
>
> Actually Michel came up with a fix for the performance regression which is
> now backported to older kernels as well.
>
> So the original issue of this mail thread should be fixed by now.

Ok , will test as soon I get the GPU to work :))

>
>> Also I played with BIOS options also which does not fix anything but
>> changes the error message.
>>
>> IOMMU && SR-IOV disabled the error changes to this :
>>
>> [ 7.092044] [drm:r600_ring_test [radeon]] *ERROR* radeon: ring 0
>> test failed (scratch(0x850C)=0xCAFEDEAD)
>> [ 7.092059] radeon 0000:21:00.0: disabling GPU acceleration
>>
>>
>> While I could workaround SWIOTLB bugs in 4.15 and 4.16 , 4.17 seems to
>> kill the GPU with no way
>> for me to make it work ( at least I could not find any workaround by now )
>
>
> That actually sounds like something completely different. Can you provide a
> full dmesg of radeon and/or amdgpu?

Sure here from boot with IOMMU/SR-IOV ON/OFF in BIOS :

http://ftp.frugalware.org/pub/other/people/crazy/radeon/dmesg-iommu-sr-iov-off.txt
http://ftp.frugalware.org/pub/other/people/crazy/radeon/dmesg-iommu-sr-iov-on.txt

Also nothing else changed in that setup just testing kernel 4.17.

I can force the GPU to use amdgpu if you wish and post dmesg's too.
Just let me know

2018-06-06 12:20:40

[permalink] [raw]

Subject: Re: AMD graphics performance regression in 4.15 and later

Am 06.06.2018 um 14:08 schrieb Gabriel C:
> 2018-06-06 13:33 GMT+02:00 Christian König <[email protected]>:
>> Am 06.06.2018 um 13:28 schrieb Gabriel C:
>>> 2018-04-11 7:02 GMT+02:00 Gabriel C <[email protected]>:
>>>>> 2018-04-11 6:00 GMT+02:00 Gabriel C <[email protected]>:
>>>>> 2018-04-09 11:42 GMT+02:00 Christian König
>>>>> <[email protected]>:
>>>>>> Am 07.04.2018 um 00:00 schrieb Jean-Marc Valin:
>>>> ...
>>>>> I can help testing code for 4.17/++ if you wish but that is *different*
>>>>> storry.
>>>>>
>>>> Quick tested an 4.16.0-11490-gb284d4d5a678 , amdgpu and radeon driver
>>>> are broken now in this one.
>>>>
>>>> radeon tells:
>>>>
>>>> ...
>>>>
>>>> [ 6.337838] [drm] PCIE GART of 2048M enabled (table at
>>>> 0x00000000001D6000).
>>>> [ 6.338210] radeon 0000:21:00.0: (-12) create WB bo failed
>>>> [ 6.338214] radeon 0000:21:00.0: disabling GPU acceleration
>>>>
>>>> ...
>>>>
>>> I have the same Issue now on final 4.17.
>>
>> Actually Michel came up with a fix for the performance regression which is
>> now backported to older kernels as well.
>>
>> So the original issue of this mail thread should be fixed by now.
> Ok , will test as soon I get the GPU to work :))
>
>>> Also I played with BIOS options also which does not fix anything but
>>> changes the error message.
>>>
>>> IOMMU && SR-IOV disabled the error changes to this :
>>>
>>> [ 7.092044] [drm:r600_ring_test [radeon]] *ERROR* radeon: ring 0
>>> test failed (scratch(0x850C)=0xCAFEDEAD)
>>> [ 7.092059] radeon 0000:21:00.0: disabling GPU acceleration
>>>
>>>
>>> While I could workaround SWIOTLB bugs in 4.15 and 4.16 , 4.17 seems to
>>> kill the GPU with no way
>>> for me to make it work ( at least I could not find any workaround by now )
>>
>> That actually sounds like something completely different. Can you provide a
>> full dmesg of radeon and/or amdgpu?
> Sure here from boot with IOMMU/SR-IOV ON/OFF in BIOS :
>
> http://ftp.frugalware.org/pub/other/people/crazy/radeon/dmesg-iommu-sr-iov-off.txt
> http://ftp.frugalware.org/pub/other/people/crazy/radeon/dmesg-iommu-sr-iov-on.txt
>
> Also nothing else changed in that setup just testing kernel 4.17.

That has nothing TODO with the driver nor the original bug you reported.
The problem is that SME is active and that is currently not supported at
all with a that hardware.

Try to disable SME either in the BIOS or on the kernel command line.

Regards,
Christian.

>
> I can force the GPU to use amdgpu if you wish and post dmesg's too.
> Just let me know

2018-07-13 02:24:56