Change queue/pipe split between amdkfd and amdgpu

Thu Feb 9 20:11:36 UTC 2017

 Andres,

I tried your patches on Kaveri with airlied's drm-next branch.
I used radeon+amdkfd

The following test failed: KFDQMTest.CreateMultipleCpQueues
However, I can't debug it because I don't have the sources of kfdtest.

In dmesg, I saw the following warning during boot:
WARNING: CPU: 0 PID: 150 at
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c:670
start_cpsch+0xc5/0x220 [amdkfd]
[    4.393796] Modules linked in: hid_logitech_hidpp hid_logitech_dj
hid_generic usbhid hid uas usb_storage amdkfd amd_iommu_v2 radeon(+)
i2c_algo_bit ttm drm_kms_helper syscopyarea ahci sysfillrect sysimgblt
libahci fb_sys_fops drm r8169 mii fjes video
[    4.393811] CPU: 0 PID: 150 Comm: systemd-udevd Not tainted 4.10.0-rc5+ #1
[    4.393811] Hardware name: Gigabyte Technology Co., Ltd. To be
filled by O.E.M./F2A88XM-D3H, BIOS F5 01/09/2014
[    4.393812] Call Trace:
[    4.393818]  dump_stack+0x63/0x90
[    4.393822]  __warn+0xcb/0xf0
[    4.393823]  warn_slowpath_null+0x1d/0x20
[    4.393830]  start_cpsch+0xc5/0x220 [amdkfd]
[    4.393836]  ? initialize_cpsch+0xa0/0xb0 [amdkfd]
[    4.393841]  kgd2kfd_device_init+0x375/0x490 [amdkfd]
[    4.393883]  radeon_kfd_device_init+0xaf/0xd0 [radeon]
[    4.393911]  radeon_driver_load_kms+0x11e/0x1f0 [radeon]
[    4.393933]  drm_dev_register+0x14a/0x200 [drm]
[    4.393946]  drm_get_pci_dev+0x9d/0x160 [drm]
[    4.393974]  radeon_pci_probe+0xb8/0xe0 [radeon]
[    4.393976]  local_pci_probe+0x45/0xa0
[    4.393978]  pci_device_probe+0x103/0x150
[    4.393981]  driver_probe_device+0x2bf/0x460
[    4.393982]  __driver_attach+0xdf/0xf0
[    4.393984]  ? driver_probe_device+0x460/0x460
[    4.393985]  bus_for_each_dev+0x6c/0xc0
[    4.393987]  driver_attach+0x1e/0x20
[    4.393988]  bus_add_driver+0x1fd/0x270
[    4.393989]  ? 0xffffffffc05c8000
[    4.393991]  driver_register+0x60/0xe0
[    4.393992]  ? 0xffffffffc05c8000
[    4.393993]  __pci_register_driver+0x4c/0x50
[    4.394007]  drm_pci_init+0xeb/0x100 [drm]
[    4.394008]  ? 0xffffffffc05c8000
[    4.394031]  radeon_init+0x98/0xb6 [radeon]
[    4.394034]  do_one_initcall+0x53/0x1a0
[    4.394037]  ? __vunmap+0x81/0xd0
[    4.394039]  ? kmem_cache_alloc_trace+0x152/0x1c0
[    4.394041]  ? vfree+0x2e/0x70
[    4.394044]  do_init_module+0x5f/0x1ff
[    4.394046]  load_module+0x24cc/0x29f0
[    4.394047]  ? __symbol_put+0x60/0x60
[    4.394050]  ? security_kernel_post_read_file+0x6b/0x80
[    4.394052]  SYSC_finit_module+0xdf/0x110
[    4.394054]  SyS_finit_module+0xe/0x10
[    4.394056]  entry_SYSCALL_64_fastpath+0x1e/0xad
[    4.394058] RIP: 0033:0x7f9cda77c8e9
[    4.394059] RSP: 002b:00007ffe195d3378 EFLAGS: 00000246 ORIG_RAX:
0000000000000139
[    4.394060] RAX: ffffffffffffffda RBX: 00007f9cdb8dda7e RCX: 00007f9cda77c8e9
[    4.394061] RDX: 0000000000000000 RSI: 00007f9cdac7ce2a RDI: 0000000000000013
[    4.394062] RBP: 00007ffe195d2450 R08: 0000000000000000 R09: 0000000000000000
[    4.394063] R10: 0000000000000013 R11: 0000000000000246 R12: 00007ffe195d245a
[    4.394063] R13: 00007ffe195d1378 R14: 0000563f70cc93b0 R15: 0000563f70cba4d0
[    4.394091] ---[ end trace 9c5af17304d998bb ]---
[    4.394092] Invalid queue enabled by amdgpu: 9

I suggest you get a Kaveri/Carrizo machine to debug these issues.

Until that, I don't think we should merge this patch-set.

Oded

On Wed, Feb 8, 2017 at 9:47 PM, Andres Rodriguez <andresx7 at gmail.com> wrote:
> Thank you Oded.
>
> - Andres
>
>
> On 2017-02-08 02:32 PM, Oded Gabbay wrote:
>>
>> On Wed, Feb 8, 2017 at 6:23 PM, Andres Rodriguez <andresx7 at gmail.com>
>> wrote:
>>>
>>> Hey Felix,
>>>
>>> Thanks for the pointer to the ROCm mqd commit. I like that the
>>> workarounds
>>> are easy to spot. I'll add that to a new patch series I'm working on for
>>> some bug-fixes for perf being lower on pipes other than pipe 0.
>>>
>>> I haven't tested this yet on kaveri/carrizo. I'm hoping someone with the
>>> HW
>>> will be able to give it a go. I put in a few small hacks to get KFD to
>>> boot
>>> but do nothing on polaris10.
>>>
>>> Regards,
>>> Andres
>>>
>>>
>>> On 2017-02-06 03:20 PM, Felix Kuehling wrote:
>>>>
>>>> Hi Andres,
>>>>
>>>> Thank you for tackling this task. It's more involved than I expected,
>>>> mostly because I didn't have much awareness of the MQD management in
>>>> amdgpu.
>>>>
>>>> I made one comment in a separate message about the unified MQD commit
>>>> function, if you want to bring that more in line with our latest ROCm
>>>> release on github.
>>>>
>>>> Also, were you able to test the upstream KFD with your changes on a
>>>> Kaveri or Carrizo?
>>>>
>>>> Regards,
>>>>    Felix
>>>>
>>>>
>>>> On 17-02-03 11:51 PM, Andres Rodriguez wrote:
>>>>>
>>>>> The current queue/pipe split policy is for amdgpu to take the first
>>>>> pipe
>>>>> of
>>>>> MEC0 and leave the rest for amdkfd to use. This policy is taken as an
>>>>> assumption in a few areas of the implementation.
>>>>>
>>>>> This patch series aims to allow for flexible/tunable queue/pipe split
>>>>> policies
>>>>> between kgd and kfd. It also updates the queue/pipe split policy to one
>>>>> that
>>>>> allows better compute app concurrency for both drivers.
>>>>>
>>>>> In the process some duplicate code and hardcoded constants were
>>>>> removed.
>>>>>
>>>>> Any suggestions or feedback on improvements welcome.
>>>>>
>>> _______________________________________________
>>> amd-gfx mailing list
>>> amd-gfx at lists.freedesktop.org
>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>
>> Hi Andres,
>> I will try to find sometime to test it on my Kaveri machine.
>>
>> Oded
>
>