If amdgpu_ttm_init() fails after ttm_device_init() succeeded, the error paths just return. Nothing later calls ttm_device_fini() because the GMC IP block never reaches sw_fini, so the freed amdgpu_device stays linked on ttm_glob.device_list. The next ttm_device_init() or ttm_global_swapout() then walks freed memory. The MMIO_REMAP BO created in amdgpu_ttm_init() is also leaked, which shows up as "Memory manager not clean" on the following bind.
Route every error after ttm_device_init() through amdgpu_ttm_fini(), and mark the manager initialized as soon as ttm_device_init() succeeds so the unwind also covers a pool init failure. Make amdgpu_ttm_fini() skip the VRAM, GTT and PREEMPT managers when they were never registered, as it can now run on a partially initialized device. Signed-off-by: Mario Limonciello <[email protected]> --- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 43 ++++++++++++++++++++------------- 1 file changed, 26 insertions(+), 17 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c index 8dfa92967b630..7828caf8aba0d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c @@ -2153,20 +2153,21 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) "failed initializing buffer object driver(%d).\n", r); return r; } + /* from here on amdgpu_ttm_fini() is responsible for unwinding */ + adev->mman.initialized = true; r = amdgpu_ttm_pools_init(adev); if (r) { dev_err(adev->dev, "failed to init ttm pools(%d).\n", r); - return r; + goto err_fini; } - adev->mman.initialized = true; if (!adev->gmc.is_app_apu) { /* Initialize VRAM pool with all of VRAM divided into pages */ r = amdgpu_vram_mgr_init(adev); if (r) { dev_err(adev->dev, "Failed initializing VRAM heap.\n"); - return r; + goto err_fini; } } @@ -2190,7 +2191,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) r = amdgpu_ttm_alloc_vram_resv_regions(adev); if (r) - return r; + goto err_fini; /* RAS loads and reserves bad pages */ amdgpu_ras_mgr_early_init_service(adev); @@ -2239,7 +2240,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) r = amdgpu_gtt_mgr_init(adev, gtt_size); if (r) { dev_err(adev->dev, "Failed initializing GTT heap.\n"); - return r; + goto err_fini; } dev_info(adev->dev, " %uM of GTT memory ready.\n", (unsigned int)(gtt_size / (1024 * 1024))); @@ -2253,52 +2254,52 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_DOORBELL, adev->doorbell.size / PAGE_SIZE); if (r) { dev_err(adev->dev, "Failed initializing doorbell heap.\n"); - return r; + goto err_fini; } /* Create a doorbell page for kernel usages */ r = amdgpu_doorbell_create_kernel_doorbells(adev); if (r) { dev_err(adev->dev, "Failed to initialize kernel doorbells.\n"); - return r; + goto err_fini; } /* Initialize MMIO-remap pool (single page 4K) */ r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_MMIO_REMAP, 1); if (r) { dev_err(adev->dev, "Failed initializing MMIO-remap heap.\n"); - return r; + goto err_fini; } /* Allocate the singleton MMIO_REMAP BO if supported */ r = amdgpu_ttm_alloc_mmio_remap_bo(adev); if (r) - return r; + goto err_fini; /* Initialize preemptible memory pool */ r = amdgpu_preempt_mgr_init(adev); if (r) { dev_err(adev->dev, "Failed initializing PREEMPT heap.\n"); - return r; + goto err_fini; } /* Initialize various on-chip memory pools */ r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GDS, adev->gds.gds_size); if (r) { dev_err(adev->dev, "Failed initializing GDS heap.\n"); - return r; + goto err_fini; } r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_GWS, adev->gds.gws_size); if (r) { dev_err(adev->dev, "Failed initializing gws heap.\n"); - return r; + goto err_fini; } r = amdgpu_ttm_init_on_chip(adev, AMDGPU_PL_OA, adev->gds.oa_size); if (r) { dev_err(adev->dev, "Failed initializing oa heap.\n"); - return r; + goto err_fini; } if (adev->ualink.npa_size) { @@ -2306,7 +2307,7 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) adev->ualink.npa_size); if (r) { dev_err(adev->dev, "Failed initializing NPA heap.\n"); - return r; + goto err_fini; } } @@ -2318,6 +2319,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev) "Debug VRAM access will use slowpath MM access\n"); return 0; + +err_fini: + amdgpu_ttm_fini(adev); + return r; } /* @@ -2352,10 +2357,14 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev) adev->mman.aper_base_kaddr = NULL; } - if (!adev->gmc.is_app_apu) + /* may be called on a partially initialized TTM, skip unregistered managers */ + if (!adev->gmc.is_app_apu && + ttm_manager_type(&adev->mman.bdev, TTM_PL_VRAM)) amdgpu_vram_mgr_fini(adev); - amdgpu_gtt_mgr_fini(adev); - amdgpu_preempt_mgr_fini(adev); + if (ttm_manager_type(&adev->mman.bdev, TTM_PL_TT)) + amdgpu_gtt_mgr_fini(adev); + if (ttm_manager_type(&adev->mman.bdev, AMDGPU_PL_PREEMPT)) + amdgpu_preempt_mgr_fini(adev); amdgpu_doorbell_fini(adev); ttm_range_man_fini(&adev->mman.bdev, AMDGPU_PL_GDS); -- 2.43.0
