From fa2213e7be9818edd2a9ce4c17ac8572ef291282 Mon Sep 17 00:00:00 2001 From: Yong Zhao Date: Wed, 13 Feb 2019 13:15:05 -0500 Subject: [PATCH 2738/2940] drm/amdkfd: Fix bugs regarding CP queue doorbell mask on SOC15 Reserved doorbells for SDMA IH and VCN were not properly masked out when allocating doorbells for CP user queues. This patch fixed that. Change-Id: I670adfc3fd7725d2ed0bd9665cb7f69f8b9023c2 Signed-off-by: Yong Zhao Reviewed-by: Felix Kuehling --- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c | 34 +++++++------------ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 11 ++++++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 14 +++++--- .../gpu/drm/amd/include/kgd_kfd_interface.h | 15 ++++---- 4 files changed, 39 insertions(+), 35 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c index 1372d6dcdb6a..ea15e27f61f9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c @@ -174,28 +174,20 @@ void amdgpu_amdkfd_device_init(struct amdgpu_device *adev) &gpu_resources.doorbell_physical_address, &gpu_resources.doorbell_aperture_size, &gpu_resources.doorbell_start_offset); + /* Since SOC15, BIF starts to statically use the + * lower 12 bits of doorbell addresses for routing + * based on settings in registers like + * SDMA0_DOORBELL_RANGE etc.. + * In order to route a doorbell to CP engine, the lower + * 12 bits of its address has to be outside the range + * set for SDMA, VCN, and IH blocks. + */ if (adev->asic_type >= CHIP_VEGA10) { - /* On SOC15 the BIF is involved in routing - * doorbells using the low 12 bits of the - * address. Communicate the assignments to - * KFD. KFD uses two doorbell pages per - * process in case of 64-bit doorbells so we - * can use each doorbell assignment twice. - */ - gpu_resources.sdma_doorbell[0][i] = - adev->doorbell_index.sdma_engine[0] + (i >> 1); - gpu_resources.sdma_doorbell[0][i+1] = - adev->doorbell_index.sdma_engine[0] + 0x200 + (i >> 1); - gpu_resources.sdma_doorbell[1][i] = - adev->doorbell_index.sdma_engine[1] + (i >> 1); - gpu_resources.sdma_doorbell[1][i+1] = - adev->doorbell_index.sdma_engine[1] + 0x200 + (i >> 1); - } - /* Doorbells 0x0f0-0ff and 0x2f0-2ff are reserved for - * SDMA, IH and VCN. So don't use them for the CP. - */ - gpu_resources.reserved_doorbell_mask = 0x1f0; - gpu_resources.reserved_doorbell_val = 0x0f0; + gpu_resources.non_cp_doorbells_start = + adev->doorbell_index.first_non_cp; + gpu_resources.non_cp_doorbells_end = + adev->doorbell_index.last_non_cp; + } kgd2kfd_device_init(adev->kfd.dev, &gpu_resources); } diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 8758175d1fd4..b0102b9426bf 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -103,6 +103,17 @@ #define KFD_KERNEL_QUEUE_SIZE 2048 +/* + * 512 = 0x200 + * The doorbell index distance between SDMA RLC (2*i) and (2*i+1) in the + * same SDMA engine on SOC15, which has 8-byte doorbells for SDMA. + * 512 8-byte doorbell distance (i.e. one page away) ensures that SDMA RLC + * (2*i+1) doorbells (in terms of the lower 12 bit address) lie exactly in + * the OFFSET and SIZE set in registers like BIF_SDMA0_DOORBELL_RANGE. + */ +#define KFD_QUEUE_DOORBELL_MIRROR_OFFSET 512 + + /* * Kernel module parameter to specify maximum number of supported queues per * device diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index 6b99f6b9fdc8..7632d66cb218 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -607,13 +607,17 @@ static int init_doorbell_bitmap(struct qcm_process_device *qpd, if (!qpd->doorbell_bitmap) return -ENOMEM; - /* Mask out any reserved doorbells */ - for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS; i++) - if ((dev->shared_resources.reserved_doorbell_mask & i) == - dev->shared_resources.reserved_doorbell_val) { + /* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */ + for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) { + if (i >= dev->shared_resources.non_cp_doorbells_start + && i <= dev->shared_resources.non_cp_doorbells_end) { set_bit(i, qpd->doorbell_bitmap); - pr_debug("reserved doorbell 0x%03x\n", i); + set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET, + qpd->doorbell_bitmap); + pr_debug("reserved doorbell 0x%03x and 0x%03x\n", i, + i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET); } + } return 0; } diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h index ec2757b058cd..7e44dcd15cdc 100644 --- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h +++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h @@ -113,17 +113,14 @@ struct kgd2kfd_shared_resources { /* Doorbell assignments (SOC15 and later chips only). Only * specific doorbells are routed to each SDMA engine. Others * are routed to IH and VCN. They are not usable by the CP. - * - * Any doorbell number D that satisfies the following condition - * is reserved: (D & reserved_doorbell_mask) == reserved_doorbell_val - * - * KFD currently uses 1024 (= 0x3ff) doorbells per process. If - * doorbells 0x0f0-0x0f7 and 0x2f-0x2f7 are reserved, that means - * mask would be set to 0x1f8 and val set to 0x0f0. */ unsigned int sdma_doorbell[2][2]; - unsigned int reserved_doorbell_mask; - unsigned int reserved_doorbell_val; + + /* From SOC15 onward, the doorbell index range not usable for CP + * queues. + */ + uint32_t non_cp_doorbells_start; + uint32_t non_cp_doorbells_end; /* Base address of doorbell aperture. */ phys_addr_t doorbell_physical_address; -- 2.17.1