drm/amdgpu AMDgpu driver

The drm/amdgpu driver supports all AMD Radeon GPUs based on the Graphics Core Next (GCN) architecture.

Module Parameters

The amdgpu driver supports the following module parameters:

vramlimit (int)

Restrict the total amount of VRAM in MiB for testing. The default is 0 (Use full VRAM).

vis_vramlimit (int)

Restrict the amount of CPU visible VRAM in MiB for testing. The default is 0 (Use full CPU visible VRAM).

gartsize (uint)

Restrict the size of GART in Mib (32, 64, etc.) for testing. The default is -1 (The size depends on asic).

gttsize (int)

Restrict the size of GTT domain in MiB for testing. The default is -1 (It’s VRAM size if 3GB < VRAM < 3/4 RAM, otherwise 3/4 RAM size).

moverate (int)

Set maximum buffer migration rate in MB/s. The default is -1 (8 MB/s).

benchmark (int)

Run benchmarks. The default is 0 (Skip benchmarks).

test (int)

Test BO GTT->VRAM and VRAM->GTT GPU copies. The default is 0 (Skip test, only set 1 to run test).

audio (int)

Set HDMI/DPAudio. Only affects non-DC display handling. The default is -1 (Enabled), set 0 to disabled it.

disp_priority (int)

Set display Priority (1 = normal, 2 = high). Only affects non-DC display handling. The default is 0 (auto).

hw_i2c (int)

To enable hw i2c engine. Only affects non-DC display handling. The default is 0 (Disabled).

pcie_gen2 (int)

To disable PCIE Gen2/3 mode (0 = disable, 1 = enable). The default is -1 (auto, enabled).

msi (int)

To disable Message Signaled Interrupts (MSI) functionality (1 = enable, 0 = disable). The default is -1 (auto, enabled).

lockup_timeout (string)

Set GPU scheduler timeout value in ms.

The format can be [Non-Compute] or [GFX,Compute,SDMA,Video]. That is there can be one or multiple values specified. 0 and negative values are invalidated. They will be adjusted to the default timeout.

  • With one value specified, the setting will apply to all non-compute jobs.

  • With multiple values specified, the first one will be for GFX. The second one is for Compute. The third and fourth ones are for SDMA and Video.

By default(with no lockup_timeout settings), the timeout for all non-compute(GFX, SDMA and Video) jobs is 10000. And there is no timeout enforced on compute jobs.

dpm (int)

Override for dynamic power management setting (0 = disable, 1 = enable) The default is -1 (auto).

fw_load_type (int)

Set different firmware loading type for debugging (0 = direct, 1 = SMU, 2 = PSP). The default is -1 (auto).

aspm (int)

To disable ASPM (1 = enable, 0 = disable). The default is -1 (auto, enabled).

runpm (int)

Override for runtime power management control for dGPUs in PX/HG laptops. The amdgpu driver can dynamically power down the dGPU on PX/HG laptops when it is idle. The default is -1 (auto enable). Setting the value to 0 disables this functionality.

ip_block_mask (uint)

Override what IP blocks are enabled on the GPU. Each GPU is a collection of IP blocks (gfx, display, video, etc.). Use this parameter to disable specific blocks. Note that the IP blocks do not have a fixed index. Some asics may not have some IPs or may include multiple instances of an IP so the ordering various from asic to asic. See the driver output in the kernel log for the list of IPs on the asic. The default is 0xffffffff (enable all blocks on a device).

bapm (int)

Bidirectional Application Power Management (BAPM) used to dynamically share TDP between CPU and GPU. Set value 0 to disable it. The default -1 (auto, enabled)

deep_color (int)

Set 1 to enable Deep Color support. Only affects non-DC display handling. The default is 0 (disabled).

vm_size (int)

Override the size of the GPU’s per client virtual address space in GiB. The default is -1 (automatic for each asic).

vm_fragment_size (int)

Override VM fragment size in bits (4, 5, etc. 4 = 64K, 9 = 2M). The default is -1 (automatic for each asic).

vm_block_size (int)

Override VM page table size in bits (default depending on vm_size and hw setup). The default is -1 (automatic for each asic).

vm_fault_stop (int)

Stop on VM fault for debugging (0 = never, 1 = print first, 2 = always). The default is 0 (No stop).

vm_debug (int)

Debug VM handling (0 = disabled, 1 = enabled). The default is 0 (Disabled).

vm_update_mode (int)

Override VM update mode. VM updated by using CPU (0 = never, 1 = Graphics only, 2 = Compute only, 3 = Both). The default is -1 (Only in large BAR(LB) systems Compute VM tables will be updated by CPU, otherwise 0, never).

exp_hw_support (int)

Enable experimental hw support (1 = enable). The default is 0 (disabled).

dc (int)

Disable/Enable Display Core driver for debugging (1 = enable, 0 = disable). The default is -1 (automatic for each asic).

sched_jobs (int)

Override the max number of jobs supported in the sw queue. The default is 32.

sched_hw_submission (int)

Override the max number of HW submissions. The default is 2.

ppfeaturemask (hexint)

Override power features enabled. See enum PP_FEATURE_MASK in drivers/gpu/drm/amd/include/amd_shared.h. The default is the current set of stable power features.

forcelongtraining (uint)

Force long memory training in resume. The default is zero, indicates short training in resume.

pcie_gen_cap (uint)

Override PCIE gen speed capabilities. See the CAIL flags in drivers/gpu/drm/amd/include/amd_pcie.h. The default is 0 (automatic for each asic).

pcie_lane_cap (uint)

Override PCIE lanes capabilities. See the CAIL flags in drivers/gpu/drm/amd/include/amd_pcie.h. The default is 0 (automatic for each asic).

cg_mask (uint)

Override Clockgating features enabled on GPU (0 = disable clock gating). See the AMD_CG_SUPPORT flags in drivers/gpu/drm/amd/include/amd_shared.h. The default is 0xffffffff (all enabled).

pg_mask (uint)

Override Powergating features enabled on GPU (0 = disable power gating). See the AMD_PG_SUPPORT flags in drivers/gpu/drm/amd/include/amd_shared.h. The default is 0xffffffff (all enabled).

sdma_phase_quantum (uint)

Override SDMA context switch phase quantum (x 1K GPU clock cycles, 0 = no change). The default is 32.

disable_cu (charp)

Set to disable CUs (It’s set like se.sh.cu,…). The default is NULL.

virtual_display (charp)

Set to enable virtual display feature. This feature provides a virtual display hardware on headless boards or in virtualized environments. It will be set like xxxx:xx:xx.x,x;xxxx:xx:xx.x,x. It’s the pci address of the device, plus the number of crtcs to expose. E.g., 0000:26:00.0,4 would enable 4 virtual crtcs on the pci device at 26:00.0. The default is NULL.

job_hang_limit (int)

Set how much time allow a job hang and not drop it. The default is 0.

lbpw (int)

Override Load Balancing Per Watt (LBPW) support (1 = enable, 0 = disable). The default is -1 (auto, enabled).

gpu_recovery (int)

Set to enable GPU recovery mechanism (1 = enable, 0 = disable). The default is -1 (auto, disabled except SRIOV).

emu_mode (int)

Set value 1 to enable emulation mode. This is only needed when running on an emulator. The default is 0 (disabled).

ras_enable (int)

Enable RAS features on the GPU (0 = disable, 1 = enable, -1 = auto (default))

ras_mask (uint)

Mask of RAS features to enable (default 0xffffffff), only valid when ras_enable == 1 See the flags in drivers/gpu/drm/amd/amdgpu/amdgpu_ras.h

si_support (int)

Set SI support driver. This parameter works after set config CONFIG_DRM_AMDGPU_SI. For SI asic, when radeon driver is enabled, set value 0 to use radeon driver, while set value 1 to use amdgpu driver. The default is using radeon driver when it available, otherwise using amdgpu driver.

cik_support (int)

Set CIK support driver. This parameter works after set config CONFIG_DRM_AMDGPU_CIK. For CIK asic, when radeon driver is enabled, set value 0 to use radeon driver, while set value 1 to use amdgpu driver. The default is using radeon driver when it available, otherwise using amdgpu driver.

smu_memory_pool_size (uint)

It is used to reserve gtt for smu debug usage, setting value 0 to disable it. The actual size is value * 256MiB. E.g. 0x1 = 256Mbyte, 0x2 = 512Mbyte, 0x4 = 1 Gbyte, 0x8 = 2GByte. The default is 0 (disabled).

async_gfx_ring (int)

It is used to enable gfx rings that could be configured with different prioritites or equal priorities

mcbp (int)

It is used to enable mid command buffer preemption. (0 = disabled (default), 1 = enabled)

discovery (int)

Allow driver to discover hardware IP information from IP Discovery table at the top of VRAM. (-1 = auto (default), 0 = disabled, 1 = enabled)

mes (int)

Enable Micro Engine Scheduler. This is a new hw scheduling engine for gfx, sdma, and compute. (0 = disabled (default), 1 = enabled)

noretry (int)

Disable retry faults in the GPU memory controller. (0 = retry enabled, 1 = retry disabled, -1 auto (default))

force_asic_type (int)

A non negative value used to specify the asic type for all supported GPUs.

sched_policy (int)

Set scheduling policy. Default is HWS(hardware scheduling) with over-subscription. Setting 1 disables over-subscription. Setting 2 disables HWS and statically assigns queues to HQDs.

hws_max_conc_proc (int)

Maximum number of processes that HWS can schedule concurrently. The maximum is the number of VMIDs assigned to the HWS, which is also the default.

cwsr_enable (int)

CWSR(compute wave store and resume) allows the GPU to preempt shader execution in the middle of a compute wave. Default is 1 to enable this feature. Setting 0 disables it.

max_num_of_queues_per_device (int)

Maximum number of queues per device. Valid setting is between 1 and 4096. Default is 4096.

send_sigterm (int)

Send sigterm to HSA process on unhandled exceptions. Default is not to send sigterm but just print errors on dmesg. Setting 1 enables sending sigterm.

debug_largebar (int)

Set debug_largebar as 1 to enable simulating large-bar capability on non-large bar system. This limits the VRAM size reported to ROCm applications to the visible size, usually 256MB. Default value is 0, diabled.

ignore_crat (int)

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)

halt_if_hws_hang (int)

Halt if HWS hang is detected. Default value, 0, disables the halt on hang. Setting 1 enables halt on hang.

hws_gws_support(bool)

Assume that HWS supports GWS barriers regardless of what firmware version check says. Default value: false (rely on MEC2 firmware version check).

queue_preemption_timeout_ms (int)

queue preemption timeout in ms (1 = Minimum, 9000 = default)

debug_evictions(bool)

Enable extra debug messages to help determine the cause of evictions

no_system_mem_limit(bool)

Disable system memory limit, to support multiple process shared memory

dcfeaturemask (uint)

Override display features enabled. See enum DC_FEATURE_MASK in drivers/gpu/drm/amd/include/amd_shared.h. The default is the current set of stable display features.

dcdebugmask (uint)

Override display features enabled. See enum DC_DEBUG_MASK in drivers/gpu/drm/amd/include/amd_shared.h.

abmlevel (uint)

Override the default ABM (Adaptive Backlight Management) level used for DC enabled hardware. Requires DMCU to be supported and loaded. Valid levels are 0-4. A value of 0 indicates that ABM should be disabled by default. Values 1-4 control the maximum allowable brightness reduction via the ABM algorithm, with 1 being the least reduction and 4 being the most reduction.

Defaults to 0, or disabled. Userspace can still override this level later after boot.

tmz (int)

Trusted Memory Zone (TMZ) is a method to protect data being written to or read from memory.

The default value: 0 (off). TODO: change to auto till it is completed.

reset_method (int)

GPU reset method (-1 = auto (default), 0 = legacy, 1 = mode0, 2 = mode1, 3 = mode2, 4 = baco, 5 = pci)

bad_page_threshold (int)

Bad page threshold is to specify the threshold value of faulty pages detected by RAS ECC, that may result in GPU entering bad status if total faulty pages by ECC exceed threshold value and leave it for user’s further check.

Core Driver Infrastructure

This section covers core driver infrastructure.

Memory Domains

AMDGPU_GEM_DOMAIN_CPU System memory that is not GPU accessible. Memory in this pool could be swapped out to disk if there is pressure.

AMDGPU_GEM_DOMAIN_GTT GPU accessible system memory, mapped into the GPU’s virtual address space via gart. Gart memory linearizes non-contiguous pages of system memory, allows GPU access system memory in a linezrized fashion.

AMDGPU_GEM_DOMAIN_VRAM Local video memory. For APUs, it is memory carved out by the BIOS.

AMDGPU_GEM_DOMAIN_GDS Global on-chip data storage used to share data across shader threads.

AMDGPU_GEM_DOMAIN_GWS Global wave sync, used to synchronize the execution of all the waves on a device.

AMDGPU_GEM_DOMAIN_OA Ordered append, used by 3D or Compute engines for appending data.

Buffer Objects

This defines the interfaces to operate on an amdgpu_bo buffer object which represents memory used by driver (VRAM, system memory, etc.). The driver provides DRM/GEM APIs to userspace. DRM/GEM APIs then use these interfaces to create/destroy/set buffer object which are then managed by the kernel TTM memory manager. The interfaces are also used internally by kernel clients, including gfx, uvd, etc. for kernel managed allocations used by the GPU.

void amdgpu_bo_subtract_pin_size(struct amdgpu_bo *bo)

Remove BO from pin_size accounting

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

Description

This function is called when a BO stops being pinned, and updates the amdgpu_device pin_size values accordingly.

bool amdgpu_bo_is_amdgpu_bo(struct ttm_buffer_object *bo)

check if the buffer object is an amdgpu_bo

Parameters

struct ttm_buffer_object *bo

buffer object to be checked

Description

Uses destroy function associated with the object to determine if this is an amdgpu_bo.

Return

true if the object belongs to amdgpu_bo, false if not.

void amdgpu_bo_placement_from_domain(struct amdgpu_bo *abo, u32 domain)

set buffer’s placement

Parameters

struct amdgpu_bo *abo

amdgpu_bo buffer object whose placement is to be set

u32 domain

requested domain

Description

Sets buffer’s placement according to requested domain and the buffer’s flags.

int amdgpu_bo_create_reserved(struct amdgpu_device *adev, unsigned long size, int align, u32 domain, struct amdgpu_bo **bo_ptr, u64 *gpu_addr, void **cpu_addr)

create reserved BO for kernel use

Parameters

struct amdgpu_device *adev

amdgpu device object

unsigned long size

size for the new BO

int align

alignment for the new BO

u32 domain

where to place it

struct amdgpu_bo **bo_ptr

used to initialize BOs in structures

u64 *gpu_addr

GPU addr of the pinned BO

void **cpu_addr

optional CPU address mapping

Description

Allocates and pins a BO for kernel internal use, and returns it still reserved.

Note

For bo_ptr new BO is only created if bo_ptr points to NULL.

Return

0 on success, negative error code otherwise.

int amdgpu_bo_create_kernel(struct amdgpu_device *adev, unsigned long size, int align, u32 domain, struct amdgpu_bo **bo_ptr, u64 *gpu_addr, void **cpu_addr)

create BO for kernel use

Parameters

struct amdgpu_device *adev

amdgpu device object

unsigned long size

size for the new BO

int align

alignment for the new BO

u32 domain

where to place it

struct amdgpu_bo **bo_ptr

used to initialize BOs in structures

u64 *gpu_addr

GPU addr of the pinned BO

void **cpu_addr

optional CPU address mapping

Description

Allocates and pins a BO for kernel internal use.

Note

For bo_ptr new BO is only created if bo_ptr points to NULL.

Return

0 on success, negative error code otherwise.

int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev, uint64_t offset, uint64_t size, uint32_t domain, struct amdgpu_bo **bo_ptr, void **cpu_addr)

create BO for kernel use at specific location

Parameters

struct amdgpu_device *adev

amdgpu device object

uint64_t offset

offset of the BO

uint64_t size

size of the BO

uint32_t domain

where to place it

struct amdgpu_bo **bo_ptr

used to initialize BOs in structures

void **cpu_addr

optional CPU address mapping

Description

Creates a kernel BO at a specific offset in the address space of the domain.

Return

0 on success, negative error code otherwise.

void amdgpu_bo_free_kernel(struct amdgpu_bo **bo, u64 *gpu_addr, void **cpu_addr)

free BO for kernel use

Parameters

struct amdgpu_bo **bo

amdgpu BO to free

u64 *gpu_addr

pointer to where the BO’s GPU memory space address was stored

void **cpu_addr

pointer to where the BO’s CPU memory space address was stored

Description

unmaps and unpin a BO for kernel internal use.

int amdgpu_bo_create(struct amdgpu_device *adev, struct amdgpu_bo_param *bp, struct amdgpu_bo **bo_ptr)

create an amdgpu_bo buffer object

Parameters

struct amdgpu_device *adev

amdgpu device object

struct amdgpu_bo_param *bp

parameters to be used for the buffer object

struct amdgpu_bo **bo_ptr

pointer to the buffer object pointer

Description

Creates an amdgpu_bo buffer object; and if requested, also creates a shadow object. Shadow object is used to backup the original buffer object, and is always in GTT.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_validate(struct amdgpu_bo *bo)

validate an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

pointer to the buffer object

Description

Sets placement according to domain; and changes placement and caching policy of the buffer object according to the placement. This is used for validating shadow bos. It calls ttm_bo_validate() to make sure the buffer is resident where it needs to be.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_restore_shadow(struct amdgpu_bo *shadow, struct dma_fence **fence)

restore an amdgpu_bo shadow

Parameters

struct amdgpu_bo *shadow

amdgpu_bo shadow to be restored

struct dma_fence **fence

dma_fence associated with the operation

Description

Copies a buffer object’s shadow content back to the object. This is used for recovering a buffer from its shadow in case of a gpu reset where vram context may be lost.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_kmap(struct amdgpu_bo *bo, void **ptr)

map an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object to be mapped

void **ptr

kernel virtual address to be returned

Description

Calls ttm_bo_kmap() to set up the kernel virtual mapping; calls amdgpu_bo_kptr() to get the kernel virtual address.

Return

0 for success or a negative error code on failure.

void * amdgpu_bo_kptr(struct amdgpu_bo *bo)

returns a kernel virtual address of the buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

Description

Calls ttm_kmap_obj_virtual() to get the kernel virtual address

Return

the virtual address of a buffer object area.

void amdgpu_bo_kunmap(struct amdgpu_bo *bo)

unmap an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object to be unmapped

Description

Unmaps a kernel map set up by amdgpu_bo_kmap().

struct amdgpu_bo * amdgpu_bo_ref(struct amdgpu_bo *bo)

reference an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

Description

References the contained ttm_buffer_object.

Return

a refcounted pointer to the amdgpu_bo buffer object.

void amdgpu_bo_unref(struct amdgpu_bo **bo)

unreference an amdgpu_bo buffer object

Parameters

struct amdgpu_bo **bo

amdgpu_bo buffer object

Description

Unreferences the contained ttm_buffer_object and clear the pointer

int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain, u64 min_offset, u64 max_offset)

pin an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object to be pinned

u32 domain

domain to be pinned to

u64 min_offset

the start of requested address range

u64 max_offset

the end of requested address range

Description

Pins the buffer object according to requested domain and address range. If the memory is unbound gart memory, binds the pages into gart table. Adjusts pin_count and pin_size accordingly.

Pinning means to lock pages in memory along with keeping them at a fixed offset. It is required when a buffer can not be moved, for example, when a display buffer is being scanned out.

Compared with amdgpu_bo_pin(), this function gives more flexibility on where to pin a buffer if there are specific restrictions on where a buffer must be located.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_pin(struct amdgpu_bo *bo, u32 domain)

pin an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object to be pinned

u32 domain

domain to be pinned to

Description

A simple wrapper to amdgpu_bo_pin_restricted(). Provides a simpler API for buffers that do not have any strict restrictions on where a buffer must be located.

Return

0 for success or a negative error code on failure.

void amdgpu_bo_unpin(struct amdgpu_bo *bo)

unpin an amdgpu_bo buffer object

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object to be unpinned

Description

Decreases the pin_count, and clears the flags if pin_count reaches 0. Changes placement and pin size accordingly.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_evict_vram(struct amdgpu_device *adev)

evict VRAM buffers

Parameters

struct amdgpu_device *adev

amdgpu device object

Description

Evicts all VRAM buffers on the lru list of the memory type. Mainly used for evicting vram at suspend time.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_init(struct amdgpu_device *adev)

initialize memory manager

Parameters

struct amdgpu_device *adev

amdgpu device object

Description

Calls amdgpu_ttm_init() to initialize amdgpu memory manager.

Return

0 for success or a negative error code on failure.

void amdgpu_bo_fini(struct amdgpu_device *adev)

tear down memory manager

Parameters

struct amdgpu_device *adev

amdgpu device object

Description

Reverses amdgpu_bo_init() to tear down memory manager.

int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo, struct vm_area_struct *vma)

mmap fbdev memory

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

struct vm_area_struct *vma

vma as input from the fbdev mmap method

Description

Calls ttm_fbdev_mmap() to mmap fbdev memory if it is backed by a bo.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_set_tiling_flags(struct amdgpu_bo *bo, u64 tiling_flags)

set tiling flags

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

u64 tiling_flags

new flags

Description

Sets buffer object’s tiling flags with the new one. Used by GEM ioctl or kernel driver to set the tiling flags on a buffer.

Return

0 for success or a negative error code on failure.

void amdgpu_bo_get_tiling_flags(struct amdgpu_bo *bo, u64 *tiling_flags)

get tiling flags

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

u64 *tiling_flags

returned flags

Description

Gets buffer object’s tiling flags. Used by GEM ioctl or kernel driver to set the tiling flags on a buffer.

int amdgpu_bo_set_metadata(struct amdgpu_bo *bo, void *metadata, uint32_t metadata_size, uint64_t flags)

set metadata

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

void *metadata

new metadata

uint32_t metadata_size

size of the new metadata

uint64_t flags

flags of the new metadata

Description

Sets buffer object’s metadata, its size and flags. Used via GEM ioctl.

Return

0 for success or a negative error code on failure.

int amdgpu_bo_get_metadata(struct amdgpu_bo *bo, void *buffer, size_t buffer_size, uint32_t *metadata_size, uint64_t *flags)

get metadata

Parameters

struct amdgpu_bo *bo

amdgpu_bo buffer object

void *buffer

returned metadata

size_t buffer_size

size of the buffer

uint32_t *metadata_size

size of the returned metadata

uint64_t *flags

flags of the returned metadata

Description

Gets buffer object’s metadata, its size and flags. buffer_size shall not be less than metadata_size. Used via GEM ioctl.

Return

0 for success or a negative error code on failure.

void amdgpu_bo_move_notify(struct ttm_buffer_object *bo, bool evict, struct ttm_resource *new_mem)

notification about a memory move

Parameters

struct ttm_buffer_object *bo

pointer to a buffer object

bool evict

if this move is evicting the buffer from the graphics address space

struct ttm_resource *new_mem

new information of the bufer object

Description

Marks the corresponding amdgpu_bo buffer object as invalid, also performs bookkeeping. TTM driver callback which is called when ttm moves a buffer.

void amdgpu_bo_release_notify(struct ttm_buffer_object *bo)

notification about a BO being released

Parameters

struct ttm_buffer_object *bo

pointer to a buffer object

Description

Wipes VRAM buffers whose contents should not be leaked before the memory is released.

vm_fault_t amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)

notification about a memory fault

Parameters

struct ttm_buffer_object *bo

pointer to a buffer object

Description

Notifies the driver we are taking a fault on this BO and have reserved it, also performs bookkeeping. TTM driver callback for dealing with vm faults.

Return

0 for success or a negative error code on failure.

void amdgpu_bo_fence(struct amdgpu_bo *bo, struct dma_fence *fence, bool shared)

add fence to buffer object

Parameters

struct amdgpu_bo *bo

buffer object in question

struct dma_fence *fence

fence to add

bool shared

true if fence should be added shared

int amdgpu_bo_sync_wait_resv(struct amdgpu_device *adev, struct dma_resv *resv, enum amdgpu_sync_mode sync_mode, void *owner, bool intr)

Wait for BO reservation fences

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct dma_resv *resv

reservation object to sync to

enum amdgpu_sync_mode sync_mode

synchronization mode

void *owner

fence owner

bool intr

Whether the wait is interruptible

Description

Extract the fences from the reservation object and waits for them to finish.

Return

0 on success, errno otherwise.

int amdgpu_bo_sync_wait(struct amdgpu_bo *bo, void *owner, bool intr)

Wrapper for amdgpu_bo_sync_wait_resv

Parameters

struct amdgpu_bo *bo

buffer object to wait for

void *owner

fence owner

bool intr

Whether the wait is interruptible

Description

Wrapper to wait for fences in a BO.

Return

0 on success, errno otherwise.

u64 amdgpu_bo_gpu_offset(struct amdgpu_bo *bo)

return GPU offset of bo

Parameters

struct amdgpu_bo *bo

amdgpu object for which we query the offset

Note

object should either be pinned or reserved when calling this function, it might be useful to add check for this for debugging.

Return

current GPU offset of the object.

u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)

return GPU offset of bo

Parameters

struct amdgpu_bo *bo

amdgpu object for which we query the offset

Return

current GPU offset of the object without raising warnings.

uint32_t amdgpu_bo_get_preferred_pin_domain(struct amdgpu_device *adev, uint32_t domain)

get preferred domain for scanout

Parameters

struct amdgpu_device *adev

amdgpu device object

uint32_t domain

allowed memory domains

Return

Which of the allowed domains is preferred for pinning the BO for scanout.

u64 amdgpu_bo_print_info(int id, struct amdgpu_bo *bo, struct seq_file *m)

print BO info in debugfs file

Parameters

int id

Index or Id of the BO

struct amdgpu_bo *bo

Requested BO for printing info

struct seq_file *m

debugfs file

Description

Print BO information in debugfs file

Return

Size of the BO in bytes.

PRIME Buffer Sharing

The following callback implementations are used for sharing GEM buffer objects between different devices via PRIME.

int amdgpu_gem_prime_mmap(struct drm_gem_object *obj, struct vm_area_struct *vma)

drm_driver.gem_prime_mmap implementation

Parameters

struct drm_gem_object *obj

GEM BO

struct vm_area_struct *vma

Virtual memory area

Description

Sets up a userspace mapping of the BO’s memory in the given virtual memory area.

Return

0 on success or a negative error code on failure.

int amdgpu_dma_buf_attach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)

dma_buf_ops.attach implementation

Parameters

struct dma_buf *dmabuf

DMA-buf where we attach to

struct dma_buf_attachment *attach

attachment to add

Description

Add the attachment as user to the exported DMA-buf.

void amdgpu_dma_buf_detach(struct dma_buf *dmabuf, struct dma_buf_attachment *attach)

dma_buf_ops.detach implementation

Parameters

struct dma_buf *dmabuf

DMA-buf where we remove the attachment from

struct dma_buf_attachment *attach

the attachment to remove

Description

Called when an attachment is removed from the DMA-buf.

int amdgpu_dma_buf_pin(struct dma_buf_attachment *attach)

dma_buf_ops.pin implementation

Parameters

struct dma_buf_attachment *attach

attachment to pin down

Description

Pin the BO which is backing the DMA-buf so that it can’t move any more.

void amdgpu_dma_buf_unpin(struct dma_buf_attachment *attach)

dma_buf_ops.unpin implementation

Parameters

struct dma_buf_attachment *attach

attachment to unpin

Description

Unpin a previously pinned BO to make it movable again.

struct sg_table * amdgpu_dma_buf_map(struct dma_buf_attachment *attach, enum dma_data_direction dir)

dma_buf_ops.map_dma_buf implementation

Parameters

struct dma_buf_attachment *attach

DMA-buf attachment

enum dma_data_direction dir

DMA direction

Description

Makes sure that the shared DMA buffer can be accessed by the target device. For now, simply pins it to the GTT domain, where it should be accessible by all DMA devices.

Return

sg_table filled with the DMA addresses to use or ERR_PRT with negative error code.

void amdgpu_dma_buf_unmap(struct dma_buf_attachment *attach, struct sg_table *sgt, enum dma_data_direction dir)

dma_buf_ops.unmap_dma_buf implementation

Parameters

struct dma_buf_attachment *attach

DMA-buf attachment

struct sg_table *sgt

sg_table to unmap

enum dma_data_direction dir

DMA direction

Description

This is called when a shared DMA buffer no longer needs to be accessible by another device. For now, simply unpins the buffer from GTT.

int amdgpu_dma_buf_begin_cpu_access(struct dma_buf *dma_buf, enum dma_data_direction direction)

dma_buf_ops.begin_cpu_access implementation

Parameters

struct dma_buf *dma_buf

Shared DMA buffer

enum dma_data_direction direction

Direction of DMA transfer

Description

This is called before CPU access to the shared DMA buffer’s memory. If it’s a read access, the buffer is moved to the GTT domain if possible, for optimal CPU read performance.

Return

0 on success or a negative error code on failure.

struct dma_buf * amdgpu_gem_prime_export(struct drm_gem_object *gobj, int flags)

drm_driver.gem_prime_export implementation

Parameters

struct drm_gem_object *gobj

GEM BO

int flags

Flags such as DRM_CLOEXEC and DRM_RDWR.

Description

The main work is done by the drm_gem_prime_export helper.

Return

Shared DMA buffer representing the GEM BO from the given device.

struct drm_gem_object * amdgpu_dma_buf_create_obj(struct drm_device *dev, struct dma_buf *dma_buf)

create BO for DMA-buf import

Parameters

struct drm_device *dev

DRM device

struct dma_buf *dma_buf

DMA-buf

Description

Creates an empty SG BO for DMA-buf import.

Return

A new GEM BO of the given DRM device, representing the memory described by the given DMA-buf attachment and scatter/gather table.

void amdgpu_dma_buf_move_notify(struct dma_buf_attachment *attach)

attach.move_notify implementation

Parameters

struct dma_buf_attachment *attach

the DMA-buf attachment

Description

Invalidate the DMA-buf attachment, making sure that the we re-create the mapping before the next use.

struct drm_gem_object * amdgpu_gem_prime_import(struct drm_device *dev, struct dma_buf *dma_buf)

drm_driver.gem_prime_import implementation

Parameters

struct drm_device *dev

DRM device

struct dma_buf *dma_buf

Shared DMA buffer

Description

Import a dma_buf into a the driver and potentially create a new GEM object.

Return

GEM BO representing the shared DMA buffer for the given device.

bool amdgpu_dmabuf_is_xgmi_accessible(struct amdgpu_device *adev, struct amdgpu_bo *bo)

Check if xgmi available for P2P transfer

Parameters

struct amdgpu_device *adev

amdgpu_device pointer of the importer

struct amdgpu_bo *bo

amdgpu buffer object

Return

True if dmabuf accessible over xgmi, false otherwise.

MMU Notifier

For coherent userptr handling registers an MMU notifier to inform the driver about updates on the page tables of a process.

When somebody tries to invalidate the page tables we block the update until all operations on the pages in question are completed, then those pages are marked as accessed and also dirty if it wasn’t a read only access.

New command submissions using the userptrs in question are delayed until all page table invalidation are completed and we once more see a coherent process address space.

bool amdgpu_mn_invalidate_gfx(struct mmu_interval_notifier *mni, const struct mmu_notifier_range *range, unsigned long cur_seq)

callback to notify about mm change

Parameters

struct mmu_interval_notifier *mni

the range (mm) is about to update

const struct mmu_notifier_range *range

details on the invalidation

unsigned long cur_seq

Value to pass to mmu_interval_set_seq()

Description

Block for operations on BOs to finish and mark pages as accessed and potentially dirty.

bool amdgpu_mn_invalidate_hsa(struct mmu_interval_notifier *mni, const struct mmu_notifier_range *range, unsigned long cur_seq)

callback to notify about mm change

Parameters

struct mmu_interval_notifier *mni

the range (mm) is about to update

const struct mmu_notifier_range *range

details on the invalidation

unsigned long cur_seq

Value to pass to mmu_interval_set_seq()

Description

We temporarily evict the BO attached to this range. This necessitates evicting all user-mode queues of the process.

int amdgpu_mn_register(struct amdgpu_bo *bo, unsigned long addr)

register a BO for notifier updates

Parameters

struct amdgpu_bo *bo

amdgpu buffer object

unsigned long addr

userptr addr we should monitor

Description

Registers a mmu_notifier for the given BO at the specified address. Returns 0 on success, -ERRNO if anything goes wrong.

void amdgpu_mn_unregister(struct amdgpu_bo *bo)

unregister a BO for notifier updates

Parameters

struct amdgpu_bo *bo

amdgpu buffer object

Description

Remove any registration of mmu notifier updates from the buffer object.

AMDGPU Virtual Memory

GPUVM is similar to the legacy gart on older asics, however rather than there being a single global gart table for the entire GPU, there are multiple VM page tables active at any given time. The VM page tables can contain a mix vram pages and system memory pages and system memory pages can be mapped as snooped (cached system pages) or unsnooped (uncached system pages). Each VM has an ID associated with it and there is a page table associated with each VMID. When execting a command buffer, the kernel tells the the ring what VMID to use for that command buffer. VMIDs are allocated dynamically as commands are submitted. The userspace drivers maintain their own address space and the kernel sets up their pages tables accordingly when they submit their command buffers and a VMID is assigned. Cayman/Trinity support up to 8 active VMs at any given time; SI supports 16.

struct amdgpu_prt_cb

Helper to disable partial resident texture feature from a fence callback

Definition

struct amdgpu_prt_cb {
  struct amdgpu_device *adev;
  struct dma_fence_cb cb;
};

Members

adev

amdgpu device

cb

callback

unsigned amdgpu_vm_level_shift(struct amdgpu_device *adev, unsigned level)

return the addr shift for each level

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

unsigned level

VMPT level

Return

The number of bits the pfn needs to be right shifted for a level.

unsigned amdgpu_vm_num_entries(struct amdgpu_device *adev, unsigned level)

return the number of entries in a PD/PT

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

unsigned level

VMPT level

Return

The number of entries in a page directory or page table.

unsigned amdgpu_vm_num_ats_entries(struct amdgpu_device *adev)

return the number of ATS entries in the root PD

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

Return

The number of entries in the root page directory which needs the ATS setting.

uint32_t amdgpu_vm_entries_mask(struct amdgpu_device *adev, unsigned int level)

the mask to get the entry number of a PD/PT

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

unsigned int level

VMPT level

Return

The mask to extract the entry number of a PD/PT from an address.

unsigned amdgpu_vm_bo_size(struct amdgpu_device *adev, unsigned level)

returns the size of the BOs in bytes

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

unsigned level

VMPT level

Return

The size of the BO for a page directory or page table in bytes.

void amdgpu_vm_bo_evicted(struct amdgpu_vm_bo_base *vm_bo)

vm_bo is evicted

Parameters

struct amdgpu_vm_bo_base *vm_bo

vm_bo which is evicted

Description

State for PDs/PTs and per VM BOs which are not at the location they should be.

void amdgpu_vm_bo_moved(struct amdgpu_vm_bo_base *vm_bo)

vm_bo is moved

Parameters

struct amdgpu_vm_bo_base *vm_bo

vm_bo which is moved

Description

State for per VM BOs which are moved, but that change is not yet reflected in the page tables.

void amdgpu_vm_bo_idle(struct amdgpu_vm_bo_base *vm_bo)

vm_bo is idle

Parameters

struct amdgpu_vm_bo_base *vm_bo

vm_bo which is now idle

Description

State for PDs/PTs and per VM BOs which have gone through the state machine and are now idle.

void amdgpu_vm_bo_invalidated(struct amdgpu_vm_bo_base *vm_bo)

vm_bo is invalidated

Parameters

struct amdgpu_vm_bo_base *vm_bo

vm_bo which is now invalidated

Description

State for normal BOs which are invalidated and that change not yet reflected in the PTs.

void amdgpu_vm_bo_relocated(struct amdgpu_vm_bo_base *vm_bo)

vm_bo is reloacted

Parameters

struct amdgpu_vm_bo_base *vm_bo

vm_bo which is relocated

Description

State for PDs/PTs which needs to update their parent PD. For the root PD, just move to idle state.

void amdgpu_vm_bo_done(struct amdgpu_vm_bo_base *vm_bo)

vm_bo is done

Parameters

struct amdgpu_vm_bo_base *vm_bo

vm_bo which is now done

Description

State for normal BOs which are invalidated and that change has been updated in the PTs.

void amdgpu_vm_bo_base_init(struct amdgpu_vm_bo_base *base, struct amdgpu_vm *vm, struct amdgpu_bo *bo)

Adds bo to the list of bos associated with the vm

Parameters

struct amdgpu_vm_bo_base *base

base structure for tracking BO usage in a VM

struct amdgpu_vm *vm

vm to which bo is to be added

struct amdgpu_bo *bo

amdgpu buffer object

Description

Initialize a bo_va_base structure and add it to the appropriate lists

struct amdgpu_vm_pt * amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt)

get the parent page directory

Parameters

struct amdgpu_vm_pt *pt

child page table

Description

Helper to get the parent entry for the child page table. NULL if we are at the root page directory.

void amdgpu_vm_pt_start(struct amdgpu_device *adev, struct amdgpu_vm *vm, uint64_t start, struct amdgpu_vm_pt_cursor *cursor)

start PD/PT walk

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

amdgpu_vm structure

uint64_t start

start address of the walk

struct amdgpu_vm_pt_cursor *cursor

state to initialize

Description

Initialize a amdgpu_vm_pt_cursor to start a walk.

bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev, struct amdgpu_vm_pt_cursor *cursor)

go to child node

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm_pt_cursor *cursor

current state

Description

Walk to the child node of the current node.

Return

True if the walk was possible, false otherwise.

bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev, struct amdgpu_vm_pt_cursor *cursor)

go to sibling node

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm_pt_cursor *cursor

current state

Description

Walk to the sibling node of the current node.

Return

True if the walk was possible, false otherwise.

bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)

go to parent node

Parameters

struct amdgpu_vm_pt_cursor *cursor

current state

Description

Walk to the parent node of the current node.

Return

True if the walk was possible, false otherwise.

void amdgpu_vm_pt_next(struct amdgpu_device *adev, struct amdgpu_vm_pt_cursor *cursor)

get next PD/PT in hieratchy

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm_pt_cursor *cursor

current state

Description

Walk the PD/PT tree to the next node.

void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_vm_pt_cursor *start, struct amdgpu_vm_pt_cursor *cursor)

start a deep first search

Parameters

struct amdgpu_device *adev

amdgpu_device structure

struct amdgpu_vm *vm

amdgpu_vm structure

struct amdgpu_vm_pt_cursor *start

optional cursor to start with

struct amdgpu_vm_pt_cursor *cursor

state to initialize

Description

Starts a deep first traversal of the PD/PT tree.

bool amdgpu_vm_pt_continue_dfs(struct amdgpu_vm_pt_cursor *start, struct amdgpu_vm_pt *entry)

check if the deep first search should continue

Parameters

struct amdgpu_vm_pt_cursor *start

starting point for the search

struct amdgpu_vm_pt *entry

current entry

Return

True when the search should continue, false otherwise.

void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev, struct amdgpu_vm_pt_cursor *cursor)

get the next node for a deep first search

Parameters

struct amdgpu_device *adev

amdgpu_device structure

struct amdgpu_vm_pt_cursor *cursor

current state

Description

Move the cursor to the next node in a deep first search.

void amdgpu_vm_get_pd_bo(struct amdgpu_vm *vm, struct list_head *validated, struct amdgpu_bo_list_entry *entry)

add the VM PD to a validation list

Parameters

struct amdgpu_vm *vm

vm providing the BOs

struct list_head *validated

head of validation list

struct amdgpu_bo_list_entry *entry

entry to add

Description

Add the page directory to the list of BOs to validate for command submission.

void amdgpu_vm_del_from_lru_notify(struct ttm_buffer_object *bo)

update bulk_moveable flag

Parameters

struct ttm_buffer_object *bo

BO which was removed from the LRU

Description

Make sure the bulk_moveable flag is updated when a BO is removed from the LRU.

void amdgpu_vm_move_to_lru_tail(struct amdgpu_device *adev, struct amdgpu_vm *vm)

move all BOs to the end of LRU

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_vm *vm

vm providing the BOs

Description

Move all BOs to the end of LRU and remember their positions to put them together.

int amdgpu_vm_validate_pt_bos(struct amdgpu_device *adev, struct amdgpu_vm *vm, int (*validate)(void *p, struct amdgpu_bo *bo), void *param)

validate the page table BOs

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_vm *vm

vm providing the BOs

int (*validate)(void *p, struct amdgpu_bo *bo)

callback to do the validation

void *param

parameter for the validation callback

Description

Validate the page table BOs on command submission if neccessary.

Return

Validation result.

bool amdgpu_vm_ready(struct amdgpu_vm *vm)

check VM is ready for updates

Parameters

struct amdgpu_vm *vm

VM to check

Description

Check if all VM PDs/PTs are ready for updates

Return

True if eviction list is empty.

int amdgpu_vm_clear_bo(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_bo *bo, bool immediate)

initially clear the PDs/PTs

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

VM to clear BO from

struct amdgpu_bo *bo

BO to clear

bool immediate

use an immediate update

Description

Root PD needs to be reserved when calling this.

Return

0 on success, errno otherwise.

void amdgpu_vm_bo_param(struct amdgpu_device *adev, struct amdgpu_vm *vm, int level, bool immediate, struct amdgpu_bo_param *bp)

fill in parameters for PD/PT allocation

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requesting vm

int level

the page table level

bool immediate

use a immediate update

struct amdgpu_bo_param *bp

resulting BO allocation parameters

int amdgpu_vm_alloc_pts(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_vm_pt_cursor *cursor, bool immediate)

Allocate a specific page table

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

VM to allocate page tables for

struct amdgpu_vm_pt_cursor *cursor

Which page table to allocate

bool immediate

use an immediate update

Description

Make sure a specific page table or directory is allocated.

Return

1 if page table needed to be allocated, 0 if page table was already allocated, negative errno if an error occurred.

void amdgpu_vm_free_table(struct amdgpu_vm_pt *entry)

fre one PD/PT

Parameters

struct amdgpu_vm_pt *entry

PDE to free

void amdgpu_vm_free_pts(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_vm_pt_cursor *start)

free PD/PT levels

Parameters

struct amdgpu_device *adev

amdgpu device structure

struct amdgpu_vm *vm

amdgpu vm structure

struct amdgpu_vm_pt_cursor *start

optional cursor where to start freeing PDs/PTs

Description

Free the page directory or page table level and all sub levels.

void amdgpu_vm_check_compute_bug(struct amdgpu_device *adev)

check whether asic has compute vm bug

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

bool amdgpu_vm_need_pipeline_sync(struct amdgpu_ring *ring, struct amdgpu_job *job)

Check if pipe sync is needed for job.

Parameters

struct amdgpu_ring *ring

ring on which the job will be submitted

struct amdgpu_job *job

job to submit

Return

True if sync is needed.

int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job, bool need_pipe_sync)

hardware flush the vm

Parameters

struct amdgpu_ring *ring

ring to use for flush

struct amdgpu_job *job

related job

bool need_pipe_sync

is pipe sync needed

Description

Emit a VM flush when it is necessary.

Return

0 on success, errno otherwise.

struct amdgpu_bo_va * amdgpu_vm_bo_find(struct amdgpu_vm *vm, struct amdgpu_bo *bo)

find the bo_va for a specific vm & bo

Parameters

struct amdgpu_vm *vm

requested vm

struct amdgpu_bo *bo

requested buffer object

Description

Find bo inside the requested vm. Search inside the bos vm list for the requested vm Returns the found bo_va or NULL if none is found

Object has to be reserved!

Return

Found bo_va or NULL.

uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)

Resolve gart mapping of addr

Parameters

const dma_addr_t *pages_addr

optional DMA address to use for lookup

uint64_t addr

the unmapped addr

Description

Look up the physical address of the page that the pte resolves to.

Return

The pointer for the page table entry.

int amdgpu_vm_update_pde(struct amdgpu_vm_update_params *params, struct amdgpu_vm *vm, struct amdgpu_vm_pt *entry)

update a single level in the hierarchy

Parameters

struct amdgpu_vm_update_params *params

parameters for the update

struct amdgpu_vm *vm

requested vm

struct amdgpu_vm_pt *entry

entry to update

Description

Makes sure the requested entry in parent is up to date.

void amdgpu_vm_invalidate_pds(struct amdgpu_device *adev, struct amdgpu_vm *vm)

mark all PDs as invalid

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

related vm

Description

Mark all PD level as invalid after an error.

int amdgpu_vm_update_pdes(struct amdgpu_device *adev, struct amdgpu_vm *vm, bool immediate)

make sure that all directories are valid

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

bool immediate

submit immediately to the paging queue

Description

Makes sure all directories are up to date.

Return

0 for success, error for failure.

void amdgpu_vm_fragment(struct amdgpu_vm_update_params *params, uint64_t start, uint64_t end, uint64_t flags, unsigned int *frag, uint64_t *frag_end)

get fragment for PTEs

Parameters

struct amdgpu_vm_update_params *params

see amdgpu_vm_update_params definition

uint64_t start

first PTE to handle

uint64_t end

last PTE to handle

uint64_t flags

hw mapping flags

unsigned int *frag

resulting fragment size

uint64_t *frag_end

end of this fragment

Description

Returns the first possible fragment for the start and end address.

int amdgpu_vm_update_ptes(struct amdgpu_vm_update_params *params, uint64_t start, uint64_t end, uint64_t dst, uint64_t flags)

make sure that page tables are valid

Parameters

struct amdgpu_vm_update_params *params

see amdgpu_vm_update_params definition

uint64_t start

start of GPU address range

uint64_t end

end of GPU address range

uint64_t dst

destination address to map to, the next dst inside the function

uint64_t flags

mapping flags

Description

Update the page tables in the range start - end.

Return

0 for success, -EINVAL for failure.

int amdgpu_vm_bo_update_mapping(struct amdgpu_device *adev, struct amdgpu_device *bo_adev, struct amdgpu_vm *vm, bool immediate, bool unlocked, struct dma_resv *resv, uint64_t start, uint64_t last, uint64_t flags, uint64_t offset, struct drm_mm_node *nodes, dma_addr_t *pages_addr, struct dma_fence **fence)

update a mapping in the vm page table

Parameters

struct amdgpu_device *adev

amdgpu_device pointer of the VM

struct amdgpu_device *bo_adev

amdgpu_device pointer of the mapped BO

struct amdgpu_vm *vm

requested vm

bool immediate

immediate submission in a page fault

bool unlocked

unlocked invalidation during MM callback

struct dma_resv *resv

fences we need to sync to

uint64_t start

start of mapped range

uint64_t last

last mapped entry

uint64_t flags

flags for the entries

uint64_t offset

offset into nodes and pages_addr

struct drm_mm_node *nodes

array of drm_mm_nodes with the MC addresses

dma_addr_t *pages_addr

DMA addresses to use for mapping

struct dma_fence **fence

optional resulting fence

Description

Fill in the page table entries between start and last.

Return

0 for success, -EINVAL for failure.

int amdgpu_vm_bo_update(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va, bool clear)

update all BO mappings in the vm page table

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo_va *bo_va

requested BO and VM object

bool clear

if true clear the entries

Description

Fill in the page table entries for bo_va.

Return

0 for success, -EINVAL for failure.

void amdgpu_vm_update_prt_state(struct amdgpu_device *adev)

update the global PRT state

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

void amdgpu_vm_prt_get(struct amdgpu_device *adev)

add a PRT user

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

void amdgpu_vm_prt_put(struct amdgpu_device *adev)

drop a PRT user

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

void amdgpu_vm_prt_cb(struct dma_fence *fence, struct dma_fence_cb *_cb)

callback for updating the PRT status

Parameters

struct dma_fence *fence

fence for the callback

struct dma_fence_cb *_cb

the callback function

void amdgpu_vm_add_prt_cb(struct amdgpu_device *adev, struct dma_fence *fence)

add callback for updating the PRT status

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct dma_fence *fence

fence for the callback

void amdgpu_vm_free_mapping(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_bo_va_mapping *mapping, struct dma_fence *fence)

free a mapping

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

struct amdgpu_bo_va_mapping *mapping

mapping to be freed

struct dma_fence *fence

fence of the unmap operation

Description

Free a mapping and make sure we decrease the PRT usage count if applicable.

void amdgpu_vm_prt_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)

finish all prt mappings

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

Description

Register a cleanup callback to disable PRT support after VM dies.

int amdgpu_vm_clear_freed(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct dma_fence **fence)

clear freed BOs in the PT

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

struct dma_fence **fence

optional resulting fence (unchanged if no work needed to be done or if an error occurred)

Description

Make sure all freed BOs are cleared in the PT. PTs have to be reserved and mutex must be locked!

Return

0 for success.

int amdgpu_vm_handle_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm)

handle moved BOs in the PT

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

Description

Make sure all BOs which are moved are updated in the PTs.

PTs have to be reserved!

Return

0 for success.

struct amdgpu_bo_va * amdgpu_vm_bo_add(struct amdgpu_device *adev, struct amdgpu_vm *vm, struct amdgpu_bo *bo)

add a bo to a specific vm

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

struct amdgpu_bo *bo

amdgpu buffer object

Description

Add bo into the requested vm. Add bo to the list of bos associated with the vm

Object has to be reserved!

Return

Newly added bo_va or NULL for failure

void amdgpu_vm_bo_insert_map(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va, struct amdgpu_bo_va_mapping *mapping)

insert a new mapping

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo_va *bo_va

bo_va to store the address

struct amdgpu_bo_va_mapping *mapping

the mapping to insert

Description

Insert a new mapping into all structures.

int amdgpu_vm_bo_map(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va, uint64_t saddr, uint64_t offset, uint64_t size, uint64_t flags)

map bo inside a vm

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo_va *bo_va

bo_va to store the address

uint64_t saddr

where to map the BO

uint64_t offset

requested offset in the BO

uint64_t size

BO size in bytes

uint64_t flags

attributes of pages (read/write/valid/etc.)

Description

Add a mapping of the BO at the specefied addr into the VM.

Object has to be reserved and unreserved outside!

Return

0 for success, error for failure.

int amdgpu_vm_bo_replace_map(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va, uint64_t saddr, uint64_t offset, uint64_t size, uint64_t flags)

map bo inside a vm, replacing existing mappings

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo_va *bo_va

bo_va to store the address

uint64_t saddr

where to map the BO

uint64_t offset

requested offset in the BO

uint64_t size

BO size in bytes

uint64_t flags

attributes of pages (read/write/valid/etc.)

Description

Add a mapping of the BO at the specefied addr into the VM. Replace existing mappings as we do so.

Object has to be reserved and unreserved outside!

Return

0 for success, error for failure.

int amdgpu_vm_bo_unmap(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va, uint64_t saddr)

remove bo mapping from vm

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo_va *bo_va

bo_va to remove the address from

uint64_t saddr

where to the BO is mapped

Description

Remove a mapping of the BO at the specefied addr from the VM.

Object has to be reserved and unreserved outside!

Return

0 for success, error for failure.

int amdgpu_vm_bo_clear_mappings(struct amdgpu_device *adev, struct amdgpu_vm *vm, uint64_t saddr, uint64_t size)

remove all mappings in a specific range

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

VM structure to use

uint64_t saddr

start of the range

uint64_t size

size of the range

Description

Remove all mappings in a range, split them as appropriate.

Return

0 for success, error for failure.

struct amdgpu_bo_va_mapping * amdgpu_vm_bo_lookup_mapping(struct amdgpu_vm *vm, uint64_t addr)

find mapping by address

Parameters

struct amdgpu_vm *vm

the requested VM

uint64_t addr

the address

Description

Find a mapping by it’s address.

Return

The amdgpu_bo_va_mapping matching for addr or NULL

void amdgpu_vm_bo_trace_cs(struct amdgpu_vm *vm, struct ww_acquire_ctx *ticket)

trace all reserved mappings

Parameters

struct amdgpu_vm *vm

the requested vm

struct ww_acquire_ctx *ticket

CS ticket

Description

Trace all mappings of BOs reserved during a command submission.

void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, struct amdgpu_bo_va *bo_va)

remove a bo to a specific vm

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo_va *bo_va

requested bo_va

Description

Remove bo_va->bo from the requested vm.

Object have to be reserved!

bool amdgpu_vm_evictable(struct amdgpu_bo *bo)

check if we can evict a VM

Parameters

struct amdgpu_bo *bo

A page table of the VM.

Description

Check if it is possible to evict a VM.

void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, struct amdgpu_bo *bo, bool evicted)

mark the bo as invalid

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_bo *bo

amdgpu buffer object

bool evicted

is the BO evicted

Description

Mark bo as invalid.

uint32_t amdgpu_vm_get_block_size(uint64_t vm_size)

calculate VM page table size as power of two

Parameters

uint64_t vm_size

VM size

Return

VM page table as power of two

void amdgpu_vm_adjust_size(struct amdgpu_device *adev, uint32_t min_vm_size, uint32_t fragment_size_default, unsigned max_level, unsigned max_bits)

adjust vm size, block size and fragment size

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

uint32_t min_vm_size

the minimum vm size in GB if it’s set auto

uint32_t fragment_size_default

Default PTE fragment size

unsigned max_level

max VMPT level

unsigned max_bits

max address space size in bits

long amdgpu_vm_wait_idle(struct amdgpu_vm *vm, long timeout)

wait for the VM to become idle

Parameters

struct amdgpu_vm *vm

VM object to wait for

long timeout

timeout to wait for VM to become idle

int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, int vm_context, u32 pasid)

initialize a vm instance

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

int vm_context

Indicates if it GFX or Compute context

u32 pasid

Process address space identifier

Description

Init vm fields.

Return

0 for success, error for failure.

int amdgpu_vm_check_clean_reserved(struct amdgpu_device *adev, struct amdgpu_vm *vm)

check if a VM is clean

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

the VM to check

Description

check all entries of the root PD, if any subsequent PDs are allocated, it means there are page table creating and filling, and is no a clean VM

Return

0 if this VM is clean

int amdgpu_vm_make_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm, u32 pasid)

Turn a GFX VM into a compute VM

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

u32 pasid

pasid to use

Description

This only works on GFX VMs that don’t have any BOs added and no page tables allocated yet.

Changes the following VM parameters: - use_cpu_for_update - pte_supports_ats - pasid (old PASID is released, because compute manages its own PASIDs)

Reinitializes the page directory to reflect the changed ATS setting.

Return

0 for success, -errno for errors.

void amdgpu_vm_release_compute(struct amdgpu_device *adev, struct amdgpu_vm *vm)

release a compute vm

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

a vm turned into compute vm by calling amdgpu_vm_make_compute

Description

This is a correspondant of amdgpu_vm_make_compute. It decouples compute pasid from vm. Compute should stop use of vm after this call.

void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)

tear down a vm instance

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

struct amdgpu_vm *vm

requested vm

Description

Tear down vm. Unbind the VM and remove all bos from the vm bo list

void amdgpu_vm_manager_init(struct amdgpu_device *adev)

init the VM manager

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

Description

Initialize the VM manager structures

void amdgpu_vm_manager_fini(struct amdgpu_device *adev)

cleanup VM manager

Parameters

struct amdgpu_device *adev

amdgpu_device pointer

Description

Cleanup the VM manager and free resources.

int amdgpu_vm_ioctl(struct drm_device *dev, void *data, struct drm_file *filp)

Manages VMID reservation for vm hubs.

Parameters

struct drm_device *dev

drm device pointer

void *data

drm_amdgpu_vm

struct drm_file *filp

drm file pointer

Return

0 for success, -errno for errors.

void amdgpu_vm_get_task_info(struct amdgpu_device *adev, u32 pasid, struct amdgpu_task_info *task_info)

Extracts task info for a PASID.

Parameters

struct amdgpu_device *adev

drm device pointer

u32 pasid

PASID identifier for VM

struct amdgpu_task_info *task_info

task_info to fill.

void amdgpu_vm_set_task_info(struct amdgpu_vm *vm)

Sets VMs task info.

Parameters

struct amdgpu_vm *vm

vm for which to set the info

bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid, uint64_t addr)

graceful handling of VM faults.

Parameters

struct amdgpu_device *adev

amdgpu device pointer

u32 pasid

PASID of the VM

uint64_t addr

Address of the fault

Description

Try to gracefully handle a VM fault. Return true if the fault was handled and shouldn’t be reported any more.

void amdgpu_debugfs_vm_bo_info(struct amdgpu_vm *vm, struct seq_file *m)

print BO info for the VM

Parameters

struct amdgpu_vm *vm

Requested VM for printing BO info

struct seq_file *m

debugfs file

Description

Print BO information in debugfs file for the VM

Interrupt Handling

Interrupts generated within GPU hardware raise interrupt requests that are passed to amdgpu IRQ handler which is responsible for detecting source and type of the interrupt and dispatching matching handlers. If handling an interrupt requires calling kernel functions that may sleep processing is dispatched to work handlers.

If MSI functionality is not disabled by module parameter then MSI support will be enabled.

For GPU interrupt sources that may be driven by another driver, IRQ domain support is used (with mapping between virtual and hardware IRQs).

void amdgpu_hotplug_work_func(struct work_struct *work)

work handler for display hotplug event

Parameters

struct work_struct *work

work struct pointer

Description

This is the hotplug event work handler (all ASICs). The work gets scheduled from the IRQ handler if there was a hotplug interrupt. It walks through the connector table and calls hotplug handler for each connector. After this, it sends a DRM hotplug event to alert userspace.

This design approach is required in order to defer hotplug event handling from the IRQ handler to a work handler because hotplug handler has to use mutexes which cannot be locked in an IRQ handler (since mutex_lock may sleep).

void amdgpu_irq_disable_all(struct amdgpu_device *adev)

disable all interrupts

Parameters

struct amdgpu_device *adev

amdgpu device pointer

Description

Disable all types of interrupts from all sources.

irqreturn_t amdgpu_irq_handler(int irq, void *arg)

IRQ handler

Parameters

int irq

IRQ number (unused)

void *arg

pointer to DRM device

Description

IRQ handler for amdgpu driver (all ASICs).

Return

result of handling the IRQ, as defined by irqreturn_t

void amdgpu_irq_handle_ih1(struct work_struct *work)

kick of processing for IH1

Parameters

struct work_struct *work

work structure in struct amdgpu_irq

Description

Kick of processing IH ring 1.

void amdgpu_irq_handle_ih2(struct work_struct *work)

kick of processing for IH2

Parameters

struct work_struct *work

work structure in struct amdgpu_irq

Description

Kick of processing IH ring 2.

void amdgpu_irq_handle_ih_soft(struct work_struct *work)

kick of processing for ih_soft

Parameters

struct work_struct *work

work structure in struct amdgpu_irq

Description

Kick of processing IH soft ring.

bool amdgpu_msi_ok(struct amdgpu_device *adev)

check whether MSI functionality is enabled

Parameters

struct amdgpu_device *adev

amdgpu device pointer (unused)

Description

Checks whether MSI functionality has been disabled via module parameter (all ASICs).

Return

true if MSIs are allowed to be enabled or false otherwise

int amdgpu_irq_init(struct amdgpu_device *adev)

initialize interrupt handling

Parameters

struct amdgpu_device *adev

amdgpu device pointer

Description

Sets up work functions for hotplug and reset interrupts, enables MSI functionality, initializes vblank, hotplug and reset interrupt handling.

Return

0 on success or error code on failure

void amdgpu_irq_fini(struct amdgpu_device *adev)

shut down interrupt handling

Parameters

struct amdgpu_device *adev

amdgpu device pointer

Description

Tears down work functions for hotplug and reset interrupts, disables MSI functionality, shuts down vblank, hotplug and reset interrupt handling, turns off interrupts from all sources (all ASICs).

int amdgpu_irq_add_id(struct amdgpu_device *adev, unsigned client_id, unsigned src_id, struct amdgpu_irq_src *source)

register IRQ source

Parameters

struct amdgpu_device *adev

amdgpu device pointer

unsigned client_id

client id

unsigned src_id

source id

struct amdgpu_irq_src *source

IRQ source pointer

Description

Registers IRQ source on a client.

Return

0 on success or error code otherwise

void amdgpu_irq_dispatch(struct amdgpu_device *adev, struct amdgpu_ih_ring *ih)

dispatch IRQ to IP blocks

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_ih_ring *ih

interrupt ring instance

Description

Dispatches IRQ to IP blocks.

void amdgpu_irq_delegate(struct amdgpu_device *adev, struct amdgpu_iv_entry *entry, unsigned int num_dw)

delegate IV to soft IH ring

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_iv_entry *entry

IV entry

unsigned int num_dw

size of IV

Description

Delegate the IV to the soft IH ring and schedule processing of it. Used if the hardware delegation to IH1 or IH2 doesn’t work for some reason.

int amdgpu_irq_update(struct amdgpu_device *adev, struct amdgpu_irq_src *src, unsigned type)

update hardware interrupt state

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_irq_src *src

interrupt source pointer

unsigned type

type of interrupt

Description

Updates interrupt state for the specific source (all ASICs).

void amdgpu_irq_gpu_reset_resume_helper(struct amdgpu_device *adev)

update interrupt states on all sources

Parameters

struct amdgpu_device *adev

amdgpu device pointer

Description

Updates state of all types of interrupts on all sources on resume after reset.

int amdgpu_irq_get(struct amdgpu_device *adev, struct amdgpu_irq_src *src, unsigned type)

enable interrupt

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_irq_src *src

interrupt source pointer

unsigned type

type of interrupt

Description

Enables specified type of interrupt on the specified source (all ASICs).

Return

0 on success or error code otherwise

int amdgpu_irq_put(struct amdgpu_device *adev, struct amdgpu_irq_src *src, unsigned type)

disable interrupt

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_irq_src *src

interrupt source pointer

unsigned type

type of interrupt

Description

Enables specified type of interrupt on the specified source (all ASICs).

Return

0 on success or error code otherwise

bool amdgpu_irq_enabled(struct amdgpu_device *adev, struct amdgpu_irq_src *src, unsigned type)

check whether interrupt is enabled or not

Parameters

struct amdgpu_device *adev

amdgpu device pointer

struct amdgpu_irq_src *src

interrupt source pointer

unsigned type

type of interrupt

Description

Checks whether the given type of interrupt is enabled on the given source.

Return

true if interrupt is enabled, false if interrupt is disabled or on invalid parameters

int amdgpu_irqdomain_map(struct irq_domain *d, unsigned int irq, irq_hw_number_t hwirq)

create mapping between virtual and hardware IRQ numbers

Parameters

struct irq_domain *d

amdgpu IRQ domain pointer (unused)

unsigned int irq

virtual IRQ number

irq_hw_number_t hwirq

hardware irq number

Description

Current implementation assigns simple interrupt handler to the given virtual IRQ.

Return

0 on success or error code otherwise

int amdgpu_irq_add_domain(struct amdgpu_device *adev)

create a linear IRQ domain

Parameters

struct amdgpu_device *adev

amdgpu device pointer

Description

Creates an IRQ domain for GPU interrupt sources that may be driven by another driver (e.g., ACP).

Return

0 on success or error code otherwise

void amdgpu_irq_remove_domain(struct amdgpu_device *adev)

remove the IRQ domain

Parameters

struct amdgpu_device *adev

amdgpu device pointer

Description

Removes the IRQ domain for GPU interrupt sources that may be driven by another driver (e.g., ACP).

unsigned amdgpu_irq_create_mapping(struct amdgpu_device *adev, unsigned src_id)

create mapping between domain Linux IRQs

Parameters

struct amdgpu_device *adev

amdgpu device pointer

unsigned src_id

IH source id

Description

Creates mapping between a domain IRQ (GPU IH src id) and a Linux IRQ Use this for components that generate a GPU interrupt, but are driven by a different driver (e.g., ACP).

Return

Linux IRQ

IP Blocks

GPUs are composed of IP (intellectual property) blocks. These IP blocks provide various functionalities: display, graphics, video decode, etc. The IP blocks that comprise a particular GPU are listed in the GPU’s respective SoC file. amdgpu_device.c acquires the list of IP blocks for the GPU in use on initialization. It can then operate on this list to perform standard driver operations such as: init, fini, suspend, resume, etc.

IP block implementations are named using the following convention: <functionality>_v<version> (E.g.: gfx_v6_0).

enum amd_ip_block_type

Used to classify IP blocks by functionality.

Constants

AMD_IP_BLOCK_TYPE_COMMON

GPU Family

AMD_IP_BLOCK_TYPE_GMC

Graphics Memory Controller

AMD_IP_BLOCK_TYPE_IH

Interrupt Handler

AMD_IP_BLOCK_TYPE_SMC

System Management Controller

AMD_IP_BLOCK_TYPE_PSP

Platform Security Processor

AMD_IP_BLOCK_TYPE_DCE

Display and Compositing Engine

AMD_IP_BLOCK_TYPE_GFX

Graphics and Compute Engine

AMD_IP_BLOCK_TYPE_SDMA

System DMA Engine

AMD_IP_BLOCK_TYPE_UVD

Unified Video Decoder

AMD_IP_BLOCK_TYPE_VCE

Video Compression Engine

AMD_IP_BLOCK_TYPE_ACP

Audio Co-Processor

AMD_IP_BLOCK_TYPE_VCN

Video Core/Codec Next

AMD_IP_BLOCK_TYPE_MES

Micro-Engine Scheduler

AMD_IP_BLOCK_TYPE_JPEG

JPEG Engine

struct amd_ip_funcs

general hooks for managing amdgpu IP Blocks

Definition

struct amd_ip_funcs {
  char *name;
  int (*early_init)(void *handle);
  int (*late_init)(void *handle);
  int (*sw_init)(void *handle);
  int (*sw_fini)(void *handle);
  int (*hw_init)(void *handle);
  int (*hw_fini)(void *handle);
  void (*late_fini)(void *handle);
  int (*suspend)(void *handle);
  int (*resume)(void *handle);
  bool (*is_idle)(void *handle);
  int (*wait_for_idle)(void *handle);
  bool (*check_soft_reset)(void *handle);
  int (*pre_soft_reset)(void *handle);
  int (*soft_reset)(void *handle);
  int (*post_soft_reset)(void *handle);
  int (*set_clockgating_state)(void *handle, enum amd_clockgating_state state);
  int (*set_powergating_state)(void *handle, enum amd_powergating_state state);
  void (*get_clockgating_state)(void *handle, u32 *flags);
  int (*enable_umd_pstate)(void *handle, enum amd_dpm_forced_level *level);
};

Members

name

Name of IP block

early_init

sets up early driver state (pre sw_init), does not configure hw - Optional

late_init

sets up late driver/hw state (post hw_init) - Optional

sw_init

sets up driver state, does not configure hw

sw_fini

tears down driver state, does not configure hw

hw_init

sets up the hw state

hw_fini

tears down the hw state

late_fini

final cleanup

suspend

handles IP specific hw/sw changes for suspend

resume

handles IP specific hw/sw changes for resume

is_idle

returns current IP block idle status

wait_for_idle

poll for idle

check_soft_reset

check soft reset the IP block

pre_soft_reset

pre soft reset the IP block

soft_reset

soft reset the IP block

post_soft_reset

post soft reset the IP block

set_clockgating_state

enable/disable cg for the IP block

set_powergating_state

enable/disable pg for the IP block

get_clockgating_state

get current clockgating status

enable_umd_pstate

enable UMD powerstate

Description

These hooks provide an interface for controlling the operational state of IP blocks. After acquiring a list of IP blocks for the GPU in use, the driver can make chip-wide state changes by walking this list and making calls to hooks from each IP block. This list is ordered to ensure that the driver initializes the IP blocks in a safe sequence.

AMDGPU XGMI Support

AMDGPU XGMI Support

XGMI is a high speed interconnect that joins multiple GPU cards into a homogeneous memory space that is organized by a collective hive ID and individual node IDs, both of which are 64-bit numbers.

The file xgmi_device_id contains the unique per GPU device ID and is stored in the /sys/class/drm/card${cardno}/device/ directory.

Inside the device directory a sub-directory ‘xgmi_hive_info’ is created which contains the hive ID and the list of nodes.

The hive ID is stored in:

/sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id

The node information is stored in numbered directories:

/sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id

Each device has their own xgmi_hive_info direction with a mirror set of node sub-directories.

The XGMI memory space is built by contiguously adding the power of two padded VRAM space from each node to each other.

AMDGPU RAS Support

The AMDGPU RAS interfaces are exposed via sysfs (for informational queries) and debugfs (for error injection).

RAS debugfs/sysfs Control and Error Injection Interfaces

It accepts struct ras_debug_if who has two members.

First member: ras_debug_if::head or ras_debug_if::inject.

head is used to indicate which IP block will be under control.

head has four members, they are block, type, sub_block_index, name. block: which IP will be under control. type: what kind of error will be enabled/disabled/injected. sub_block_index: some IPs have subcomponets. say, GFX, sDMA. name: the name of IP.

inject has two more members than head, they are address, value. As their names indicate, inject operation will write the value to the address.

The second member: struct ras_debug_if::op. It has three kinds of operations.

  • 0: disable RAS on the block. Take ::head as its data.

  • 1: enable RAS on the block. Take ::head as its data.

  • 2: inject errors on the block. Take ::inject as its data.

How to use the interface?

Programs

Copy the struct ras_debug_if in your codes and initialize it. Write the struct to the control node.

Shells

echo op block [error [sub_block address value]] > .../ras/ras_ctrl

Parameters:

op: disable, enable, inject

disable: only block is needed enable: block and error are needed inject: error, address, value are needed

block: umc, sdma, gfx, ………

see ras_block_string[] for details

error: ue, ce

ue: multi_uncorrectable ce: single_correctable

sub_block:

sub block index, pass 0 if there is no sub block

here are some examples for bash commands:

echo inject umc ue 0x0 0x0 0x0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
echo inject umc ce 0 0 0 > /sys/kernel/debug/dri/0/ras/ras_ctrl
echo disable umc > /sys/kernel/debug/dri/0/ras/ras_ctrl

How to check the result?

For disable/enable, please check ras features at /sys/class/drm/card[0/1/2…]/device/ras/features

For inject, please check corresponding err count at /sys/class/drm/card[0/1/2…]/device/ras/[gfx/sdma/…]_err_count

Note

Operations are only allowed on blocks which are supported. Please check ras mask at /sys/module/amdgpu/parameters/ras_mask to see which blocks support RAS on a particular asic.

RAS Reboot Behavior for Unrecoverable Errors

Normally when there is an uncorrectable error, the driver will reset the GPU to recover. However, in the event of an unrecoverable error, the driver provides an interface to reboot the system automatically in that event.

The following file in debugfs provides that interface: /sys/kernel/debug/dri/[0/1/2…]/ras/auto_reboot

Usage:

echo true > .../ras/auto_reboot

RAS Error Count sysfs Interface

It allows the user to read the error count for each IP block on the gpu through /sys/class/drm/card[0/1/2…]/device/ras/[gfx/sdma/…]_err_count

It outputs the multiple lines which report the uncorrected (ue) and corrected (ce) error counts.

The format of one line is below,

[ce|ue]: count

Example:

ue: 0
ce: 1

RAS EEPROM debugfs Interface

Some boards contain an EEPROM which is used to persistently store a list of bad pages which experiences ECC errors in vram. This interface provides a way to reset the EEPROM, e.g., after testing error injection.

Usage:

echo 1 > ../ras/ras_eeprom_reset

will reset EEPROM table to 0 entries.

RAS VRAM Bad Pages sysfs Interface

It allows user to read the bad pages of vram on the gpu through /sys/class/drm/card[0/1/2…]/device/ras/gpu_vram_bad_pages

It outputs multiple lines, and each line stands for one gpu page.

The format of one line is below, gpu pfn : gpu page size : flags

gpu pfn and gpu page size are printed in hex format. flags can be one of below character,

R: reserved, this gpu page is reserved and not able to use.

P: pending for reserve, this gpu page is marked as bad, will be reserved in next window of page_reserve.

F: unable to reserve. this gpu page can’t be reserved due to some reasons.

Examples:

0x00000001 : 0x00001000 : R
0x00000002 : 0x00001000 : P

Sample Code

Sample code for testing error injection can be found here: https://cgit.freedesktop.org/mesa/drm/tree/tests/amdgpu/ras_tests.c

This is part of the libdrm amdgpu unit tests which cover several areas of the GPU. There are four sets of tests:

RAS Basic Test

The test verifies the RAS feature enabled status and makes sure the necessary sysfs and debugfs files are present.

RAS Query Test

This test checks the RAS availability and enablement status for each supported IP block as well as the error counts.

RAS Inject Test

This test injects errors for each IP.

RAS Disable Test

This test tests disabling of RAS features for each IP block.

GPU Power/Thermal Controls and Monitoring

This section covers hwmon and power/thermal controls.

HWMON Interfaces

The amdgpu driver exposes the following sensor interfaces:

  • GPU temperature (via the on-die sensor)

  • GPU voltage

  • Northbridge voltage (APUs only)

  • GPU power

  • GPU fan

  • GPU gfx/compute engine clock

  • GPU memory clock (dGPU only)

hwmon interfaces for GPU temperature:

  • temp[1-3]_input: the on die GPU temperature in millidegrees Celsius - temp2_input and temp3_input are supported on SOC15 dGPUs only

  • temp[1-3]_label: temperature channel label - temp2_label and temp3_label are supported on SOC15 dGPUs only

  • temp[1-3]_crit: temperature critical max value in millidegrees Celsius - temp2_crit and temp3_crit are supported on SOC15 dGPUs only

  • temp[1-3]_crit_hyst: temperature hysteresis for critical limit in millidegrees Celsius - temp2_crit_hyst and temp3_crit_hyst are supported on SOC15 dGPUs only

  • temp[1-3]_emergency: temperature emergency max value(asic shutdown) in millidegrees Celsius - these are supported on SOC15 dGPUs only

hwmon interfaces for GPU voltage:

  • in0_input: the voltage on the GPU in millivolts

  • in1_input: the voltage on the Northbridge in millivolts

hwmon interfaces for GPU power:

  • power1_average: average power used by the GPU in microWatts

  • power1_cap_min: minimum cap supported in microWatts

  • power1_cap_max: maximum cap supported in microWatts

  • power1_cap: selected power cap in microWatts

hwmon interfaces for GPU fan:

  • pwm1: pulse width modulation fan level (0-255)

  • pwm1_enable: pulse width modulation fan control method (0: no fan speed control, 1: manual fan speed control using pwm interface, 2: automatic fan speed control)

  • pwm1_min: pulse width modulation fan control minimum level (0)

  • pwm1_max: pulse width modulation fan control maximum level (255)

  • fan1_min: an minimum value Unit: revolution/min (RPM)

  • fan1_max: an maxmum value Unit: revolution/max (RPM)

  • fan1_input: fan speed in RPM

  • fan[1-*]_target: Desired fan speed Unit: revolution/min (RPM)

  • fan[1-*]_enable: Enable or disable the sensors.1: Enable 0: Disable

hwmon interfaces for GPU clocks:

  • freq1_input: the gfx/compute clock in hertz

  • freq2_input: the memory clock in hertz

You can use hwmon tools like sensors to view this information on your system.

GPU sysfs Power State Interfaces

GPU power controls are exposed via sysfs files.

power_dpm_state

The power_dpm_state file is a legacy interface and is only provided for backwards compatibility. The amdgpu driver provides a sysfs API for adjusting certain power related parameters. The file power_dpm_state is used for this. It accepts the following arguments:

  • battery

  • balanced

  • performance

battery

On older GPUs, the vbios provided a special power state for battery operation. Selecting battery switched to this state. This is no longer provided on newer GPUs so the option does nothing in that case.

balanced

On older GPUs, the vbios provided a special power state for balanced operation. Selecting balanced switched to this state. This is no longer provided on newer GPUs so the option does nothing in that case.

performance

On older GPUs, the vbios provided a special power state for performance operation. Selecting performance switched to this state. This is no longer provided on newer GPUs so the option does nothing in that case.

power_dpm_force_performance_level

The amdgpu driver provides a sysfs API for adjusting certain power related parameters. The file power_dpm_force_performance_level is used for this. It accepts the following arguments:

  • auto

  • low

  • high

  • manual

  • profile_standard

  • profile_min_sclk

  • profile_min_mclk

  • profile_peak

auto

When auto is selected, the driver will attempt to dynamically select the optimal power profile for current conditions in the driver.

low

When low is selected, the clocks are forced to the lowest power state.

high

When high is selected, the clocks are forced to the highest power state.

manual

When manual is selected, the user can manually adjust which power states are enabled for each clock domain via the sysfs pp_dpm_mclk, pp_dpm_sclk, and pp_dpm_pcie files and adjust the power state transition heuristics via the pp_power_profile_mode sysfs file.

profile_standard profile_min_sclk profile_min_mclk profile_peak

When the profiling modes are selected, clock and power gating are disabled and the clocks are set for different profiling cases. This mode is recommended for profiling specific work loads where you do not want clock or power gating for clock fluctuation to interfere with your results. profile_standard sets the clocks to a fixed clock level which varies from asic to asic. profile_min_sclk forces the sclk to the lowest level. profile_min_mclk forces the mclk to the lowest level. profile_peak sets all clocks (mclk, sclk, pcie) to the highest levels.

pp_table

The amdgpu driver provides a sysfs API for uploading new powerplay tables. The file pp_table is used for this. Reading the file will dump the current power play table. Writing to the file will attempt to upload a new powerplay table and re-initialize powerplay using that new table.

pp_od_clk_voltage

The amdgpu driver provides a sysfs API for adjusting the clocks and voltages in each power level within a power state. The pp_od_clk_voltage is used for this.

Note that the actual memory controller clock rate are exposed, not the effective memory clock of the DRAMs. To translate it, use the following formula:

Clock conversion (Mhz):

HBM: effective_memory_clock = memory_controller_clock * 1

G5: effective_memory_clock = memory_controller_clock * 1

G6: effective_memory_clock = memory_controller_clock * 2

DRAM data rate (MT/s):

HBM: effective_memory_clock * 2 = data_rate

G5: effective_memory_clock * 4 = data_rate

G6: effective_memory_clock * 8 = data_rate

Bandwidth (MB/s):

data_rate * vram_bit_width / 8 = memory_bandwidth

Some examples:

G5 on RX460:

memory_controller_clock = 1750 Mhz

effective_memory_clock = 1750 Mhz * 1 = 1750 Mhz

data rate = 1750 * 4 = 7000 MT/s

memory_bandwidth = 7000 * 128 bits / 8 = 112000 MB/s

G6 on RX5700:

memory_controller_clock = 875 Mhz

effective_memory_clock = 875 Mhz * 2 = 1750 Mhz

data rate = 1750 * 8 = 14000 MT/s

memory_bandwidth = 14000 * 256 bits / 8 = 448000 MB/s

< For Vega10 and previous ASICs >

Reading the file will display:

  • a list of engine clock levels and voltages labeled OD_SCLK

  • a list of memory clock levels and voltages labeled OD_MCLK

  • a list of valid ranges for sclk, mclk, and voltage labeled OD_RANGE

To manually adjust these settings, first select manual using power_dpm_force_performance_level. Enter a new value for each level by writing a string that contains “s/m level clock voltage” to the file. E.g., “s 1 500 820” will update sclk level 1 to be 500 MHz at 820 mV; “m 0 350 810” will update mclk level 0 to be 350 MHz at 810 mV. When you have edited all of the states as needed, write “c” (commit) to the file to commit your changes. If you want to reset to the default power levels, write “r” (reset) to the file to reset them.

< For Vega20 and newer ASICs >

Reading the file will display:

  • minimum and maximum engine clock labeled OD_SCLK

  • minimum(not available for Vega20 and Navi1x) and maximum memory clock labeled OD_MCLK

  • three <frequency, voltage> points labeled OD_VDDC_CURVE. They can be used to calibrate the sclk voltage curve.

  • voltage offset(in mV) applied on target voltage calculation. This is available for Sienna Cichlid, Navy Flounder and Dimgrey Cavefish. For these ASICs, the target voltage calculation can be illustrated by “voltage = voltage calculated from v/f curve + overdrive vddgfx offset”

  • a list of valid ranges for sclk, mclk, and voltage curve points labeled OD_RANGE

To manually adjust these settings:

  • First select manual using power_dpm_force_performance_level

  • For clock frequency setting, enter a new value by writing a string that contains “s/m index clock” to the file. The index should be 0 if to set minimum clock. And 1 if to set maximum clock. E.g., “s 0 500” will update minimum sclk to be 500 MHz. “m 1 800” will update maximum mclk to be 800Mhz.

    For sclk voltage curve, enter the new values by writing a string that contains “vc point clock voltage” to the file. The points are indexed by 0, 1 and 2. E.g., “vc 0 300 600” will update point1 with clock set as 300Mhz and voltage as 600mV. “vc 2 1000 1000” will update point3 with clock set as 1000Mhz and voltage 1000mV.

    To update the voltage offset applied for gfxclk/voltage calculation, enter the new value by writing a string that contains “vo offset”. This is supported by Sienna Cichlid, Navy Flounder and Dimgrey Cavefish. And the offset can be a positive or negative value.

  • When you have edited all of the states as needed, write “c” (commit) to the file to commit your changes

  • If you want to reset to the default power levels, write “r” (reset) to the file to reset them

pp_dpm_*

The amdgpu driver provides a sysfs API for adjusting what power levels are enabled for a given power state. The files pp_dpm_sclk, pp_dpm_mclk, pp_dpm_socclk, pp_dpm_fclk, pp_dpm_dcefclk and pp_dpm_pcie are used for this.

pp_dpm_socclk and pp_dpm_dcefclk interfaces are only available for Vega10 and later ASICs. pp_dpm_fclk interface is only available for Vega20 and later ASICs.

Reading back the files will show you the available power levels within the power state and the clock information for those levels.

To manually adjust these states, first select manual using power_dpm_force_performance_level. Secondly, enter a new value for each level by inputing a string that contains ” echo xx xx xx > pp_dpm_sclk/mclk/pcie” E.g.,

echo "4 5 6" > pp_dpm_sclk

will enable sclk levels 4, 5, and 6.

NOTE: change to the dcefclk max dpm level is not supported now

pp_power_profile_mode

The amdgpu driver provides a sysfs API for adjusting the heuristics related to switching between power levels in a power state. The file pp_power_profile_mode is used for this.

Reading this file outputs a list of all of the predefined power profiles and the relevant heuristics settings for that profile.

To select a profile or create a custom profile, first select manual using power_dpm_force_performance_level. Writing the number of a predefined profile to pp_power_profile_mode will enable those heuristics. To create a custom set of heuristics, write a string of numbers to the file starting with the number of the custom profile along with a setting for each heuristic parameter. Due to differences across asic families the heuristic parameters vary from family to family.

*_busy_percent

The amdgpu driver provides a sysfs API for reading how busy the GPU is as a percentage. The file gpu_busy_percent is used for this. The SMU firmware computes a percentage of load based on the aggregate activity level in the IP cores.

The amdgpu driver provides a sysfs API for reading how busy the VRAM is as a percentage. The file mem_busy_percent is used for this. The SMU firmware computes a percentage of load based on the aggregate activity level in the IP cores.

gpu_metrics

The amdgpu driver provides a sysfs API for retrieving current gpu metrics data. The file gpu_metrics is used for this. Reading the file will dump all the current gpu metrics data.

These data include temperature, frequency, engines utilization, power consume, throttler status, fan speed and cpu core statistics( available for APU only). That’s it will give a snapshot of all sensors at the same time.

GPU Product Information

Information about the GPU can be obtained on certain cards via sysfs

product_name

The amdgpu driver provides a sysfs API for reporting the product name for the device The file serial_number is used for this and returns the product name as returned from the FRU. NOTE: This is only available for certain server cards

product_number

The amdgpu driver provides a sysfs API for reporting the product name for the device The file serial_number is used for this and returns the product name as returned from the FRU. NOTE: This is only available for certain server cards

serial_number

The amdgpu driver provides a sysfs API for reporting the serial number for the device The file serial_number is used for this and returns the serial number as returned from the FRU. NOTE: This is only available for certain server cards

unique_id

The amdgpu driver provides a sysfs API for providing a unique ID for the GPU The file unique_id is used for this. This will provide a Unique ID that will persist from machine to machine

NOTE: This will only work for GFX9 and newer. This file will be absent on unsupported ASICs (GFX8 and older)

GPU Memory Usage Information

Various memory accounting can be accessed via sysfs

mem_info_vram_total

The amdgpu driver provides a sysfs API for reporting current total VRAM available on the device The file mem_info_vram_total is used for this and returns the total amount of VRAM in bytes

mem_info_vram_used

The amdgpu driver provides a sysfs API for reporting current total VRAM available on the device The file mem_info_vram_used is used for this and returns the total amount of currently used VRAM in bytes

mem_info_vis_vram_total

The amdgpu driver provides a sysfs API for reporting current total visible VRAM available on the device The file mem_info_vis_vram_total is used for this and returns the total amount of visible VRAM in bytes

mem_info_vis_vram_used

The amdgpu driver provides a sysfs API for reporting current total of used visible VRAM The file mem_info_vis_vram_used is used for this and returns the total amount of currently used visible VRAM in bytes

mem_info_gtt_total

The amdgpu driver provides a sysfs API for reporting current total size of the GTT. The file mem_info_gtt_total is used for this, and returns the total size of the GTT block, in bytes

mem_info_gtt_used

The amdgpu driver provides a sysfs API for reporting current total amount of used GTT. The file mem_info_gtt_used is used for this, and returns the current used size of the GTT block, in bytes

PCIe Accounting Information

pcie_bw

The amdgpu driver provides a sysfs API for estimating how much data has been received and sent by the GPU in the last second through PCIe. The file pcie_bw is used for this. The Perf counters count the number of received and sent messages and return those values, as well as the maximum payload size of a PCIe packet (mps). Note that it is not possible to easily and quickly obtain the size of each packet transmitted, so we output the max payload size (mps) to allow for quick estimation of the PCIe bandwidth usage

pcie_replay_count

The amdgpu driver provides a sysfs API for reporting the total number of PCIe replays (NAKs) The file pcie_replay_count is used for this and returns the total number of replays as a sum of the NAKs generated and NAKs received