[PATCH 2/2] drm/amdgpu: disable gfxoff when using register read interface
Yuan, Xiaojie
Xiaojie.Yuan at amd.com
Tue Nov 19 02:30:04 UTC 2019
Hi Chris,
The info ioctl code path seems already covered at libdrm initialization time:
amdgpu_device_initialize()
-> amdgpu_queury_gpu_info_init()
-> amdgpu_read_mm_registers()
-> drmCommandWrite(fd, DRM_AMDGPU_INFO, ...)
BR,
Xiaojie
________________________________________
From: Christian König <ckoenig.leichtzumerken at gmail.com>
Sent: Tuesday, November 19, 2019 12:18 AM
To: Yuan, Xiaojie; Alex Deucher
Cc: Deucher, Alexander; amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH 2/2] drm/amdgpu: disable gfxoff when using register read interface
Hi Xiaojie,
could you add that test to the unit tests we have in libdrm?
Would be rather nice to have,
Christian.
Am 18.11.19 um 15:09 schrieb Yuan, Xiaojie:
> Hi Alex,
>
> I tried on Navi14 with Gfxoff enabled (gfx in 'OFF' state when I run the test program) and used a test program to read GRBM_STATUS/CP_STAT/GB_ADDR_CONFIG via DRM_IOCTL_AMDGPU_INFO.
>
> All read out values are valid as below (in this read cycle, I saw gfx block is first awakened and then powered off again automatically):
>
> drm version: 3.36.0
> grbm_status: 0x00003028
> cp_stat: 0x00000000
> gb_addr_config: 0x00000043
>
> At least for Navi, reading register when gfxoff seems not a problem...
>
> Test program:
>
> #include <stdio.h>
> #include <stdlib.h>
> #include <stdint.h>
> #include <unistd.h>
> #include <string.h>
> #include <sys/types.h>
> #include <sys/stat.h>
> #include <sys/ioctl.h>
> #include <fcntl.h>
> #include <errno.h>
> #include <libdrm/drm.h>
> #include <libdrm/amdgpu_drm.h>
>
> int get_reg_val(int fd, int reg_offset, int *val)
> {
> struct drm_amdgpu_info request;
> int r;
>
> memset(&request, 0, sizeof(request));
> request.return_pointer = (uintptr_t)val;
> request.return_size = sizeof(uint32_t);
> request.query = AMDGPU_INFO_READ_MMR_REG;
> request.read_mmr_reg.dword_offset = reg_offset;
> request.read_mmr_reg.count = 1;
> request.read_mmr_reg.instance = 0xffffffff;
> request.read_mmr_reg.flags = 0;
>
> r = ioctl(fd, DRM_IOCTL_AMDGPU_INFO, &request);
> if (r < 0) {
> perror("failed to read register");
> return -errno;
> }
>
> return 0;
> }
>
> int main(int argc, const char *argv[])
> {
> const char *device = "/dev/dri/card0";
> int fd;
> int r;
> struct drm_version version;
> struct drm_auth auth;
> int cp_stat;
> int grbm_status;
> int gb_addr_config;
>
> fd = open(device, O_RDWR | O_CLOEXEC);
> if (fd < 0) {
> perror("failed to open device\n");
> return -errno;
> }
>
> memset(&version, 0, sizeof(version));
>
> r = ioctl(fd, DRM_IOCTL_VERSION, &version);
> if (r < 0) {
> perror("failed to get drm version");
> r = -errno;
> goto close;
> }
>
> printf("drm version: %d.%d.%d\n", version.version_major,
> version.version_minor,
> version.version_patchlevel);
>
> r = get_reg_val(fd, 0x2004, &grbm_status);
> if (r < 0)
> goto close;
> printf("grbm_status: 0x%08x\n", grbm_status);
>
> r = get_reg_val(fd, 0x21a0, &cp_stat);
> if (r < 0)
> goto close;
> printf("cp_stat: 0x%08x\n", cp_stat);
>
> r = get_reg_val(fd, 0x263e, &gb_addr_config);
> if (r < 0)
> goto close;
> printf("gb_addr_config: 0x%08x\n", gb_addr_config);
>
> close:
> close(fd);
>
> return r;
> }
>
> BR,
> Xiaojie
>
> ________________________________________
> From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> on behalf of Yuan, Xiaojie <Xiaojie.Yuan at amd.com>
> Sent: Saturday, November 16, 2019 1:02 AM
> To: Alex Deucher
> Cc: Deucher, Alexander; amd-gfx at lists.freedesktop.org
> Subject: Re: [PATCH 2/2] drm/amdgpu: disable gfxoff when using register read interface
>
> Yes. IIRC, some asics' amdgpu_gfx_ctrl() is implemented as synchronous (upon function returns, gfx block is guaranteed to be in power-up state). Anyway, let me confirm about that soon.
>
> BR,
> Xiaojie
>
>> On Nov 16, 2019, at 12:52 AM, Alex Deucher <alexdeucher at gmail.com> wrote:
>>
>>> On Fri, Nov 15, 2019 at 11:46 AM Yuan, Xiaojie <Xiaojie.Yuan at amd.com> wrote:
>>>
>>> Hi Alex,
>>>
>>> IMHO, driver sending Disallow_Gfxoff message to SMU doesn't mean gfx block will be immediately powered up, so I'm not sure MMIO register access will be successful within this time window(maybe GRBM access will be pending until gfx block is powered up?)
>>>
>>> If you are not in a hurry to commit this fix, I can verify on my Navi boards next Monday.
>> That would be great. Maybe we can add a delay in that function to
>> take that into account?
>>
>> Thanks!
>>
>> Alex
>>
>>> BR,
>>> Xiaojie
>>>
>>>> On Nov 15, 2019, at 12:44 AM, Alex Deucher <alexdeucher at gmail.com> wrote:
>>>>
>>>> When gfxoff is enabled, accessing gfx registers via MMIO
>>>> can lead to a hang.
>>>>
>>>> Bug: https://bugzilla.kernel.org/show_bug.cgi?id=205497
>>>> Signed-off-by: Alex Deucher <alexander.deucher at amd.com>
>>>> ---
>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 6 +++++-
>>>> 1 file changed, 5 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
>>>> index 6ddea7607ad0..5f3b3a705b29 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c
>>>> @@ -659,15 +659,19 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
>>>> return -ENOMEM;
>>>> alloc_size = info->read_mmr_reg.count * sizeof(*regs);
>>>>
>>>> - for (i = 0; i < info->read_mmr_reg.count; i++)
>>>> + amdgpu_gfx_off_ctrl(adev, false);
>>>> + for (i = 0; i < info->read_mmr_reg.count; i++) {
>>>> if (amdgpu_asic_read_register(adev, se_num, sh_num,
>>>> info->read_mmr_reg.dword_offset + i,
>>>> ®s[i])) {
>>>> DRM_DEBUG_KMS("unallowed offset %#x\n",
>>>> info->read_mmr_reg.dword_offset + i);
>>>> kfree(regs);
>>>> + amdgpu_gfx_off_ctrl(adev, true);
>>>> return -EFAULT;
>>>> }
>>>> + }
>>>> + amdgpu_gfx_off_ctrl(adev, true);
>>>> n = copy_to_user(out, regs, min(size, alloc_size));
>>>> kfree(regs);
>>>> return n ? -EFAULT : 0;
>>>> --
>>>> 2.23.0
>>>>
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
More information about the amd-gfx
mailing list