[PATCH v3 2/3] drm/amdkfd: force raven as "dgpu" path (v3)
Felix Kuehling
felix.kuehling at amd.com
Thu Aug 20 04:05:56 UTC 2020
Am 2020-08-19 um 11:09 p.m. schrieb Huang Rui:
> On Thu, Aug 20, 2020 at 08:18:57AM +0800, Kuehling, Felix wrote:
>> On 2020-08-19 7:56 p.m., Huang Rui wrote:
>>> On Wed, Aug 19, 2020 at 11:38:34PM +0800, Kuehling, Felix wrote:
>>>> Am 2020-08-19 um 7:06 a.m. schrieb Huang Rui:
>>>>> We still have a few iommu issues which need to address, so force raven
>>>>> as "dgpu" path for the moment.
>>>>>
>>>>> This is to add the fallback path to bypass IOMMU if IOMMU v2 is disabled
>>>>> or ACPI CRAT table not correct.
>>>>>
>>>>> v2: Use ignore_crat parameter to decide whether it will go with IOMMUv2.
>>>>> v3: Align with existed thunk, don't change the way of raven, only renoir
>>>>> will use "dgpu" path by default.
>>>>>
>>>>> Signed-off-by: Huang Rui <ray.huang at amd.com>
>>>>> ---
>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 5 +++-
>>>>> drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 28 ++++++++++++++++++++++-
>>>>> drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +-
>>>>> drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +-
>>>>> drivers/gpu/drm/amd/amdkfd/kfd_topology.c | 1 +
>>>>> 5 files changed, 34 insertions(+), 4 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>> index a9a4319c24ae..189f9d7e190d 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>>>> @@ -684,11 +684,14 @@ MODULE_PARM_DESC(debug_largebar,
>>>>> * Ignore CRAT table during KFD initialization. By default, KFD uses the ACPI CRAT
>>>>> * table to get information about AMD APUs. This option can serve as a workaround on
>>>>> * systems with a broken CRAT table.
>>>>> + *
>>>>> + * Default is auto (according to asic type, iommu_v2, and crat table, to decide
>>>>> + * whehter use CRAT)
>>>>> */
>>>>> int ignore_crat;
>>>>> module_param(ignore_crat, int, 0444);
>>>>> MODULE_PARM_DESC(ignore_crat,
>>>>> - "Ignore CRAT table during KFD initialization (0 = use CRAT (default), 1 = ignore CRAT)");
>>>>> + "Ignore CRAT table during KFD initialization (0 = auto (default), 1 = ignore CRAT)");
>>>>>
>>>>> /**
>>>>> * DOC: halt_if_hws_hang (int)
>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
>>>>> index 59557e3e206a..f8346d4402e2 100644
>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
>>>>> @@ -22,6 +22,7 @@
>>>>>
>>>>> #include <linux/pci.h>
>>>>> #include <linux/acpi.h>
>>>>> +#include <asm/processor.h>
>>>>> #include "kfd_crat.h"
>>>>> #include "kfd_priv.h"
>>>>> #include "kfd_topology.h"
>>>>> @@ -740,6 +741,30 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
>>>>> return 0;
>>>>> }
>>>>>
>>>>> +
>>>>> +#ifdef CONFIG_ACPI
>>>>> +static void kfd_setup_ignore_crat_option(void)
>>>>> +{
>>>>> +
>>>>> + if (ignore_crat)
>>>>> + return;
>>>>> +
>>>>> +#ifndef KFD_SUPPORT_IOMMU_V2
>>>>> + ignore_crat = 1;
>>>>> +#else
>>>>> + ignore_crat = 0;
>>>>> +#endif
>>>>> +
>>>>> + /* Renoir use the fallback path to align with existed thunk */
>>>> Are you sure you need special code for Renoir here? For Renoir the
>>>> dev->device_info already treats it as a dGPU and always has.
>>> Renoir also is an APU, in other words, we might have got the correct CRAT
>>> table from SBIOS (the CRAT table in SBIOS for renoir is broken so far). If
>>> we had got CRAT table, the kfd would create an APU node. That's not
>>> expected.
>> kfd_assign_gpu will not assign a Renoir GPU as the APU from the CRAT
>> table because gpu->device_info->needs_iommu_device is False for Renoir.
>> So Renoir will always show up in the topology as its own discrete GPU node.
>>
>> How does this work today? Renoir is already treated as a dGPU. But the
>> CPU node info (/sys/class/kfd/kfd/topology/nodes/0/properties) from the
>> CRAT table still shows GPU cores?
>>
>> Regards,
>> Felix
>>
>>
>>>> I don't like the whole idea of changing the value of a module parameter,
>>>> because it is global and visible to the user through sysfs. Instead, if
>>>> you need to override the value of ignore_crat to consider other
>>>> conditions, I think kfd_device_use_iommu_v2 and
>>>> kfd_create_crat_image_acpi would be the right place to do it.
>>>>
>>>> To avoid duplicating the conditions, you could add a helper function
>>>> bool kfd_ignore_crat(void) that can be called instead of using the
>>>> ignore_crat parameter directly. This function, changing the global
>>>> module parameter, should be removed.
>>> That makes sense. Will update it in next version.
>>>
>>>>> + if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
>>>>> + boot_cpu_data.x86 == 0x17 &&
>>>>> + boot_cpu_data.x86_model >= 0x60 && boot_cpu_data.x86_model < 0x70) {
>>>>> + ignore_crat = 1;
>>>>> + }
>>>>> +
>>>>> + return;
>>>>> +}
>>>>> +
>>>>> /*
>>>>> * kfd_create_crat_image_acpi - Allocates memory for CRAT image and
>>>>> * copies CRAT from ACPI (if available).
>>>>> @@ -751,7 +776,6 @@ static int kfd_fill_gpu_cache_info(struct kfd_dev *kdev,
>>>>> *
>>>>> * Return 0 if successful else return error code
>>>>> */
>>>>> -#ifdef CONFIG_ACPI
>>>>> int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
>>>>> {
>>>>> struct acpi_table_header *crat_table;
>>>>> @@ -775,6 +799,8 @@ int kfd_create_crat_image_acpi(void **crat_image, size_t *size)
>>>>> return -EINVAL;
>>>>> }
>>>>>
>>>>> + kfd_setup_ignore_crat_option();
>>>>> +
>>>>> if (ignore_crat) {
>>>>> pr_info("CRAT table disabled by module option\n");
>>>>> return -ENODATA;
>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
>>>>> index 2c030c2b5b8d..dab44951c4d8 100644
>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
>>>>> @@ -112,6 +112,7 @@ static const struct kfd_device_info carrizo_device_info = {
>>>>> .num_xgmi_sdma_engines = 0,
>>>>> .num_sdma_queues_per_engine = 2,
>>>>> };
>>>>> +#endif
>>>>>
>>>>> static const struct kfd_device_info raven_device_info = {
>>>>> .asic_family = CHIP_RAVEN,
>>>>> @@ -130,7 +131,6 @@ static const struct kfd_device_info raven_device_info = {
>>>>> .num_xgmi_sdma_engines = 0,
>>>>> .num_sdma_queues_per_engine = 2,
>>>>> };
>>>>> -#endif
>>>>>
>>>>> static const struct kfd_device_info hawaii_device_info = {
>>>>> .asic_family = CHIP_HAWAII,
>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>>>> index 82f955750e75..4b6e7ef7a71c 100644
>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
>>>>> @@ -1234,7 +1234,7 @@ static inline int kfd_devcgroup_check_permission(struct kfd_dev *kfd)
>>>>>
>>>>> static inline bool kfd_device_use_iommu_v2(const struct kfd_dev *dev)
>>>>> {
>>>>> - return dev && dev->device_info->needs_iommu_device;
>>>>> + return !ignore_crat && dev && dev->device_info->needs_iommu_device;
>>>>> }
>>>>>
>>>>> /* Debugfs */
>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
>>>>> index 4b29815e9205..b92ce75a4c53 100644
>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_topology.c
>>>>> @@ -1090,6 +1090,7 @@ int kfd_topology_init(void)
>>>>> COMPUTE_UNIT_CPU, NULL,
>>>>> proximity_domain);
>>>>> cpu_only_node = 1;
>>>>> + ignore_crat = 1;
>>>> Don't change the global variable. I think you're doing this here in case
>>>> the CRAT table is broken and contains no GPU info. Maybe we need to add
>>>> a new flag "use_iommu_v2" into the kfd_dev structure to handle this.
>>>>
> Find it just now, kfd_dev is not initialized here. So we may be unable to
> use flag in kfd_dev.
I see. This is very early during module init. When you get here, you
already failed to read the ACPI CRAT table and created a VCRAT for the
CPU with no GPU cores.
If you wanted to add a per device "use_iommu_v2" flag, you could
probably set that in kfd_assign_gpu when it assigns a KFD device to a
node with CPU cores.
Regards,
Felix
>
> Thanks,
> Ray
More information about the amd-gfx
mailing list