amdkfd_total_mem_size is the size of total GPUs vram plus system memory
to estimate page tables memory usage and leave enough VRAM room for page
tables allocation.
Calculate amdkfd_total_mem_size in amdgpu_amdkfd_device_probe is
incorrect because adev->gmc.real_vram_size is still 0 called from
amdgpu_device_ip_early_init. Move the calculation
to amdgpu_amdkfd_device_init to get the correct VRAM size.
Do reverse calculation in amdgpu_amdkfd_device_fini_sw to support
hot-unplugging GPUs.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Under VRAM usage pression, map to GPU may fail to create pt bo and
vmbo->shadow_list is not initialized, then ttm_bo_release calling
amdgpu_bo_vm_destroy to access vmbo->shadow_list generates below
dmesg and NULL pointer access backtrace:
Set vmbo destroy callback to amdgpu_bo_vm_destroy only after creating pt
bo successfully, otherwise use default callback amdgpu_bo_destroy.
amdgpu: amdgpu_vm_bo_update failed
amdgpu: update_gpuvm_pte() failed
amdgpu: Failed to map bo to gpuvm
amdgpu 0000:43:00.0: amdgpu: Failed to map peer:0000:43:00.0 mem_domain:2
BUG: kernel NULL pointer dereference, address:
RIP: 0010:amdgpu_bo_vm_destroy+0x4d/0x80 [amdgpu]
Call Trace:
<TASK>
ttm_bo_release+0x207/0x320 [amdttm]
amdttm_bo_init_reserved+0x1d6/0x210 [amdttm]
amdgpu_bo_create+0x1ba/0x520 [amdgpu]
amdgpu_bo_create_vm+0x3a/0x80 [amdgpu]
amdgpu_vm_pt_create+0xde/0x270 [amdgpu]
amdgpu_vm_ptes_update+0x63b/0x710 [amdgpu]
amdgpu_vm_update_range+0x2e7/0x6e0 [amdgpu]
amdgpu_vm_bo_update+0x2bd/0x600 [amdgpu]
update_gpuvm_pte+0x160/0x420 [amdgpu]
amdgpu_amdkfd_gpuvm_map_memory_to_gpu+0x313/0x1130 [amdgpu]
kfd_ioctl_map_memory_to_gpu+0x115/0x390 [amdgpu]
kfd_ioctl+0x24a/0x5b0 [amdgpu]
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
DRM buddy manager allocates the contiguous memory requests in
a single block or multiple blocks. So for the ttm move operation
(incase of low vram memory) we should consider all the blocks to
compute the total memory size which compared with the struct
ttm_resource num_pages in order to verify that the blocks are
contiguous for the eviction process.
v2: Added a Fixes tag
v3: Rewrite the code to save a bit of calculations and
variables (Christian)
Fixes: c9cad937c0 ("drm/amdgpu: add drm buddy support to amdgpu")
Signed-off-by: Arunpravin Paneer Selvam <Arunpravin.PaneerSelvam@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch is to fix the SDMA user queue doorbell missing issue on
SDMA 6.0. F32_WPTR_POLL_ENABLE has to be set if doorbell mode is
used. Otherwise ringing SDMA user queue doorbell can't wake up
system from gfxoff.
Signed-off-by: Ruili Ji <ruiliji2@amd.com>
Reviewed-by: Yifan Zhang <yifan1.zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org # 6.0.x
In some error path in amdgpu_sdma_init_microcode(), release_firmware() is
not called, the memory allocated in request_firmware() will be leaked,
calling amdgpu_sdma_destroy_inst_ctx() which calls release_firmware() to
avoid memory leak.
Fixes: 15aa13056d ("drm/amdgpu: add function to init SDMA microcode")
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For xgmi sriov, the reset is handled by host driver and hive->reset_domain
is not initialized so need to check if it exists before doing a put.
Signed-off-by: Vignesh Chander <Vignesh.Chander@amd.com>
Reviewed-by: Shaoyun Liu <Shaoyun.Liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Simplify the logic in amdgpu_gmc_noretry_set by getting rid of the
switch. Also set noretry=1 as default for GFX 10.3.0 and greater since
retry faults are not supported.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On ChromeOS clang build, the following warning is seen:
/mnt/host/source/src/third_party/kernel/v5.15/drivers/gpu/drm/amd/amdgpu/umc_v6_7.c:463:6: error: variable 'mc_umc_status' is used uninitialized whenever 'if' condition is false [-Werror,-Wsometimes-uninitialized]
if (mca_addr == UMC_INVALID_ADDR) {
^~~~~~~~~~~~~~~~~~~~~~~~~~~~
/mnt/host/source/src/third_party/kernel/v5.15/drivers/gpu/drm/amd/amdgpu/umc_v6_7.c:485:21: note: uninitialized use occurs here
if ((REG_GET_FIELD(mc_umc_status, MCA_UMC_UMC0_MCUMC_STATUST0, Val) == 1 &&
^~~~~~~~~~~~~
/mnt/host/source/src/third_party/kernel/v5.15/drivers/gpu/drm/amd/amdgpu/../amdgpu/amdgpu.h:1208:5: note: expanded from macro 'REG_GET_FIELD'
(((value) & REG_FIELD_MASK(reg, field)) >> REG_FIELD_SHIFT(reg, field))
^~~~~
/mnt/host/source/src/third_party/kernel/v5.15/drivers/gpu/drm/amd/amdgpu/umc_v6_7.c:463:2: note: remove the 'if' if its condition is always true
if (mca_addr == UMC_INVALID_ADDR) {
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/mnt/host/source/src/third_party/kernel/v5.15/drivers/gpu/drm/amd/amdgpu/umc_v6_7.c:460:24: note: initialize the variable 'mc_umc_status' to silence this warning
uint64_t mc_umc_status, mc_umc_addrt0;
^
= 0
1 error generated.
make[5]: *** [/mnt/host/source/src/third_party/kernel/v5.15/scripts/Makefile.build:289: drivers/gpu/drm/amd/amdgpu/umc_v6_7.o] Error 1
Fix by initializing mc_umc_status = 0.
Fixes: 1014bd1cb3 ("drm/amdgpu: support to convert dedicated umc mca address")
Reviewed-by: Hamza Mahfooz <hamza.mahfooz@amd.com>
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Update umc error address query interface, the mca address can be read
from register or input from parameter.
TODO: define a common address conversion function to simplify the code.
Signed-off-by: Tao Zhou <tao.zhou1@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Make it global so we can convert specific mca address.
v2: rename query_error_address_per_channel to
convert_ras_error_address
Signed-off-by: Tao Zhou <tao.zhou1@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Fix init SDMA microcode error for sdma v4, which caused by mistake when
rearch sdma init microcode function (coding 4.2.2 to 4.2.0).
Signed-off-by: Likun Gao <Likun.Gao@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
- Under SRIOV, we need to send REQ_GPU_FINI to the hypervisor
during the suspend time. Furthermore, we cannot request a
mode 1 reset under SRIOV as VF. Therefore, we will skip it
as it is called in suspend_noirq() function.
- In the resume code path, we need to send REQ_GPU_INIT to the
hypervisor and also resume PSP IP block under SRIOV.
Signed-off-by: Bokun Zhang <Bokun.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The function amdgpu_fence_count_emitted used in work_hander should not call
amdgpu_fence_process which must be used in irq handler.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Jiadong.Zhu <Jiadong.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The current position calulated in gfx_v9_0_ring_emit_patch_cond_exec
underflows when the wptr is divisible by ring->buf_mask + 1.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Jiadong.Zhu <Jiadong.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Update mes_v11_api_def.h add_queue API with is_aql_queue parameter. Also
re-use gds_size for the queue size (unused for KFD). MES requires the
queue size in order to compute the actual wptr offset within the queue
RB since it increases monotonically for AQL queues.
v2: Make is_aql_queue assign clearer
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Disable verbose while getting p2p distance. With verbose, it shows
warning if ACS redirect is set between the devices. Adds noise
to dmesg logs when a few GPU devices are on the same platform.
Example log:
amdgpu 0000:34:00.0: ACS redirect is set between the client and provider (0000:31:00.0)
amdgpu 0000:34:00.0: to disable ACS redirect for this path, add the kernel parameter:
pci=disable_acs_redir=0000:30:00.0;0000:2e:00.0;0000:33:00.0;0000:2e:10.0
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Guchun Chen <guchun.chen@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For the asic using smu v13_0_2, there is the following
warning when uninstalling amdgpu:
amdgpu: ras disable gfx failed poison:1 ret:-22.
[Why]:
For the asic using smu v13_0_2, the psp .suspend and
mode1reset is called before amdgpu_ras_pre_fini during
amdgpu uninstall, it has disabled all ras features and
reset the psp. Since the psp is reset, calling
amdgpu_ras_disable_all_features in amdgpu_ras_pre_fini
to disable ras features will fail.
[How]:
If all ras features are disabled, amdgpu_ras_disable_all_features
will not be called to disable all ras features again.
Signed-off-by: YiPeng Chai <YiPeng.Chai@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
