From 8dc1db3172ae2f17ae71e33b608a33411ce8a1aa Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Wed, 14 Sep 2022 16:39:48 +0800 Subject: drm/amdkfd: Introduce kfd_node struct (v5) Introduce a new structure, kfd_node, which will now represent a compute node. kfd_node is carved out of kfd_dev structure. kfd_dev struct now will become the parent of kfd_node, and will store common resources such as doorbells, GTT sub-alloctor etc. kfd_node struct will store all resources specific to a compute node, such as device queue manager, interrupt handling etc. This is the first step in adding compute partition support in KFD. v2: introduce kfd_node struct to gc v11 (Hawking) v3: make reference to kfd_dev struct through kfd_node (Morris) v4: use kfd_node instead for kfd isr/mqd functions (Morris) v5: rebase (Alex) Signed-off-by: Mukul Joshi Tested-by: Amber Lin Reviewed-by: Felix Kuehling Signed-off-by: Hawking Zhang Signed-off-by: Morris Zhang Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 18250845a989..54d7d4665ad2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -119,7 +119,7 @@ static int pm_runlist_v9(struct packet_manager *pm, uint32_t *buffer, struct pm4_mes_runlist *packet; int concurrent_proc_cnt = 0; - struct kfd_dev *kfd = pm->dqm->dev; + struct kfd_node *kfd = pm->dqm->dev; /* Determine the number of processes to map together to HW: * it can not exceed the number of VMIDs available to the @@ -220,7 +220,8 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, case KFD_QUEUE_TYPE_SDMA: case KFD_QUEUE_TYPE_SDMA_XGMI: use_static = false; /* no static queues under SDMA */ - if (q->properties.sdma_engine_id < 2 && !pm_use_ext_eng(q->device)) + if (q->properties.sdma_engine_id < 2 && + !pm_use_ext_eng(q->device->kfd)) packet->bitfields2.engine_sel = q->properties.sdma_engine_id + engine_sel__mes_map_queues__sdma0_vi; else { @@ -263,7 +264,8 @@ static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer, packet->header.u32All = pm_build_pm4_header(IT_UNMAP_QUEUES, sizeof(struct pm4_mes_unmap_queues)); - packet->bitfields2.extended_engine_sel = pm_use_ext_eng(pm->dqm->dev) ? + packet->bitfields2.extended_engine_sel = + pm_use_ext_eng(pm->dqm->dev->kfd) ? extended_engine_sel__mes_unmap_queues__sdma0_to_7_sel : extended_engine_sel__mes_unmap_queues__legacy_engine_sel; -- cgit v1.2.3 From 1794e9d7e78cb52605234d0ddc3f46084937f4e9 Mon Sep 17 00:00:00 2001 From: Mukul Joshi Date: Thu, 8 Dec 2022 12:08:17 -0500 Subject: drm/amdkfd: Update packet manager for GFX9.4.3 In GFX 9.4.3, there can be more than 8 SDMA engines. As a result, extended_engine_sel and engine_sel fields in MAP_QUEUES packet need to be updated to allow correct mapping of SDMA queues to these SDMA engines. Signed-off-by: Mukul Joshi Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 16 +++++++++++++--- drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h | 3 ++- 2 files changed, 15 insertions(+), 4 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 54d7d4665ad2..44cf3a5f6fdb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -225,9 +225,19 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, packet->bitfields2.engine_sel = q->properties.sdma_engine_id + engine_sel__mes_map_queues__sdma0_vi; else { - packet->bitfields2.extended_engine_sel = - extended_engine_sel__mes_map_queues__sdma0_to_7_sel; - packet->bitfields2.engine_sel = q->properties.sdma_engine_id; + /* + * For GFX9.4.3, SDMA engine id can be greater than 8. + * For such cases, set extended_engine_sel to 2 and + * ensure engine_sel lies between 0-7. + */ + if (q->properties.sdma_engine_id >= 8) + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_map_queues__sdma8_to_15_sel; + else + packet->bitfields2.extended_engine_sel = + extended_engine_sel__mes_map_queues__sdma0_to_7_sel; + + packet->bitfields2.engine_sel = q->properties.sdma_engine_id % 8; } break; default: diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index a666710ed403..2ad708c64012 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -263,7 +263,8 @@ enum mes_map_queues_engine_sel_enum { enum mes_map_queues_extended_engine_sel_enum { extended_engine_sel__mes_map_queues__legacy_engine_sel = 0, - extended_engine_sel__mes_map_queues__sdma0_to_7_sel = 1 + extended_engine_sel__mes_map_queues__sdma0_to_7_sel = 1, + extended_engine_sel__mes_map_queues__sdma8_to_15_sel = 2 }; struct pm4_mes_map_queues { -- cgit v1.2.3 From 7cee6a6824a0429a6255abe91b5af01b9a01cd03 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 23 Mar 2023 17:17:20 -0400 Subject: drm/amdgpu: add configurable grace period for unmap queues The HWS schedule allows a grace period for wave completion prior to preemption for better performance by avoiding CWSR on waves that can potentially complete quickly. The debugger, on the other hand, will want to inspect wave status immediately after it actively triggers preemption (a suspend function to be provided). To minimize latency between preemption and debugger wave inspection, allow immediate preemption by setting the grace period to 0. Note that setting the preepmtion grace period to 0 will result in an infinite grace period being set due to a CP FW bug so set it to 1 for now. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c | 2 + .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c | 43 ++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h | 6 ++ .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c | 43 ++++++++++++++ drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h | 8 ++- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 63 +++++++++++++++------ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 3 + drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c | 32 +++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 39 +++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c | 2 + drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h | 65 ++++++++++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 5 ++ 14 files changed, 295 insertions(+), 20 deletions(-) (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c index a6f98141c29c..b811a0985050 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c @@ -82,5 +82,7 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = { .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .enable_debug_trap = kgd_aldebaran_enable_debug_trap, .disable_debug_trap = kgd_aldebaran_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c index d2918e5c0dea..a62bd0068515 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c @@ -410,6 +410,8 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = { kgd_gfx_v9_set_vm_context_page_table_base, .enable_debug_trap = kgd_arcturus_enable_debug_trap, .disable_debug_trap = kgd_arcturus_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c index 240f5006e278..98006c7021dd 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c @@ -803,6 +803,47 @@ uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev, return 0; } +/* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values + * The values read are: + * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. + * atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads. + * wrm_offload_wait_time -- Wait Count for WAIT_REG_MEM Offloads. + * gws_wait_time -- Wait Count for Global Wave Syncs. + * que_sleep_wait_time -- Wait Count for Dequeue Retry. + * sch_wave_wait_time -- Wait Count for Scheduling Wave Message. + * sem_rearm_wait_time -- Wait Count for Semaphore re-arm. + * deq_retry_wait_time -- Wait Count for Global Wave Syncs. + */ +void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, + uint32_t *wait_times) + +{ + *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2)); +} + +void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data) +{ + *reg_data = wait_times; + + /* + * The CP cannont handle a 0 grace period input and will result in + * an infinite grace period being set so set to 1 to prevent this. + */ + if (grace_period == 0) + grace_period = 1; + + *reg_data = REG_SET_FIELD(*reg_data, + CP_IQ_WAIT_TIME2, + SCH_WAVE, + grace_period); + + *reg_offset = SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2); +} + static void program_trap_handler_settings(struct amdgpu_device *adev, uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, uint32_t inst) @@ -848,5 +889,7 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = { .set_vm_context_page_table_base = set_vm_context_page_table_base, .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, .disable_debug_trap = kgd_gfx_v10_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .program_trap_handler_settings = program_trap_handler_settings, }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h index 251d61fbde07..1e993a213646 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h @@ -26,3 +26,9 @@ uint32_t kgd_gfx_v10_enable_debug_trap(struct amdgpu_device *adev, uint32_t kgd_gfx_v10_disable_debug_trap(struct amdgpu_device *adev, bool keep_trap_enabled, uint32_t vmid); +void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); +void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data); diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c index 8b293f3dcbd2..387bdf4823c9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c @@ -672,6 +672,8 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = { .get_atc_vmid_pasid_mapping_info = get_atc_vmid_pasid_mapping_info_v10_3, .set_vm_context_page_table_base = set_vm_context_page_table_base_v10_3, .program_trap_handler_settings = program_trap_handler_settings_v10_3, + .get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info, .enable_debug_trap = kgd_gfx_v10_enable_debug_trap, .disable_debug_trap = kgd_gfx_v10_disable_debug_trap }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c index 8d7d04704b00..829ee720cc44 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c @@ -739,6 +739,24 @@ uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, return 0; } +/* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values + * The values read are: + * ib_offload_wait_time -- Wait Count for Indirect Buffer Offloads. + * atomic_offload_wait_time -- Wait Count for L2 and GDS Atomics Offloads. + * wrm_offload_wait_time -- Wait Count for WAIT_REG_MEM Offloads. + * gws_wait_time -- Wait Count for Global Wave Syncs. + * que_sleep_wait_time -- Wait Count for Dequeue Retry. + * sch_wave_wait_time -- Wait Count for Scheduling Wave Message. + * sem_rearm_wait_time -- Wait Count for Semaphore re-arm. + * deq_retry_wait_time -- Wait Count for Global Wave Syncs. + */ +void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, + uint32_t *wait_times) + +{ + *wait_times = RREG32(SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2)); +} + void kgd_gfx_v9_set_vm_context_page_table_base(struct amdgpu_device *adev, uint32_t vmid, uint64_t page_table_base) { @@ -926,6 +944,29 @@ void kgd_gfx_v9_get_cu_occupancy(struct amdgpu_device *adev, int pasid, adev->gfx.cu_info.max_waves_per_simd; } +void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data) +{ + *reg_data = wait_times; + + /* + * The CP cannont handle a 0 grace period input and will result in + * an infinite grace period being set so set to 1 to prevent this. + */ + if (grace_period == 0) + grace_period = 1; + + *reg_data = REG_SET_FIELD(*reg_data, + CP_IQ_WAIT_TIME2, + SCH_WAVE, + grace_period); + + *reg_offset = SOC15_REG_OFFSET(GC, 0, mmCP_IQ_WAIT_TIME2); +} + void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev, uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr, uint32_t inst) { @@ -969,6 +1010,8 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = { .set_vm_context_page_table_base = kgd_gfx_v9_set_vm_context_page_table_base, .enable_debug_trap = kgd_gfx_v9_enable_debug_trap, .disable_debug_trap = kgd_gfx_v9_disable_debug_trap, + .get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times, + .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info, .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy, .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings, }; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h index 9588ff055393..fed5b7f18b1a 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h @@ -20,8 +20,6 @@ * OTHER DEALINGS IN THE SOFTWARE. */ - - void kgd_gfx_v9_program_sh_mem_settings(struct amdgpu_device *adev, uint32_t vmid, uint32_t sh_mem_config, uint32_t sh_mem_ape1_base, uint32_t sh_mem_ape1_limit, @@ -73,3 +71,9 @@ uint32_t kgd_gfx_v9_enable_debug_trap(struct amdgpu_device *adev, uint32_t kgd_gfx_v9_disable_debug_trap(struct amdgpu_device *adev, bool keep_trap_enabled, uint32_t vmid); +void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times); +void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev, + uint32_t wait_times, + uint32_t grace_period, + uint32_t *reg_offset, + uint32_t *reg_data); diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 2baa0781eafc..0b88a64e61fe 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -46,10 +46,13 @@ static int set_pasid_vmid_mapping(struct device_queue_manager *dqm, static int execute_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param); + uint32_t filter_param, + uint32_t grace_period); static int unmap_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param, bool reset); + uint32_t filter_param, + uint32_t grace_period, + bool reset); static int map_queues_cpsch(struct device_queue_manager *dqm); @@ -866,7 +869,7 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q, if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) { if (!dqm->dev->kfd->shared_resources.enable_mes) retval = unmap_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); else if (prev_active) retval = remove_queue_mes(dqm, q, &pdd->qpd); @@ -1042,7 +1045,8 @@ static int evict_process_queues_cpsch(struct device_queue_manager *dqm, retval = execute_queues_cpsch(dqm, qpd->is_debug ? KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES : - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); out: dqm_unlock(dqm); @@ -1182,8 +1186,7 @@ static int restore_process_queues_cpsch(struct device_queue_manager *dqm, } if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); - + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp; atomic64_add(eviction_duration, &pdd->evict_duration_counter); vm_not_acquired: @@ -1525,6 +1528,9 @@ static int initialize_cpsch(struct device_queue_manager *dqm) init_sdma_bitmaps(dqm); + if (dqm->dev->kfd2kgd->get_iq_wait_times) + dqm->dev->kfd2kgd->get_iq_wait_times(dqm->dev->adev, + &dqm->wait_times); return 0; } @@ -1563,8 +1569,9 @@ static int start_cpsch(struct device_queue_manager *dqm) dqm->is_hws_hang = false; dqm->is_resetting = false; dqm->sched_running = true; + if (!dqm->dev->kfd->shared_resources.enable_mes) - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); dqm_unlock(dqm); return 0; @@ -1589,7 +1596,7 @@ static int stop_cpsch(struct device_queue_manager *dqm) if (!dqm->is_hws_hang) { if (!dqm->dev->kfd->shared_resources.enable_mes) - unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, false); + unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false); else remove_all_queues_mes(dqm); } @@ -1631,7 +1638,8 @@ static int create_kernel_queue_cpsch(struct device_queue_manager *dqm, list_add(&kq->list, &qpd->priv_queue_list); increment_queue_count(dqm, qpd, kq->queue); qpd->is_debug = true; - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); dqm_unlock(dqm); return 0; @@ -1645,7 +1653,8 @@ static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm, list_del(&kq->list); decrement_queue_count(dqm, qpd, kq->queue); qpd->is_debug = false; - execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); + execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); /* * Unconditionally decrement this counter, regardless of the queue's * type. @@ -1722,7 +1731,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, if (!dqm->dev->kfd->shared_resources.enable_mes) retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD); else retval = add_queue_mes(dqm, q, qpd); if (retval) @@ -1811,7 +1820,9 @@ static int map_queues_cpsch(struct device_queue_manager *dqm) /* dqm->lock mutex has to be locked before calling this function */ static int unmap_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param, bool reset) + uint32_t filter_param, + uint32_t grace_period, + bool reset) { int retval = 0; struct mqd_manager *mqd_mgr; @@ -1823,6 +1834,12 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, if (!dqm->active_runlist) return retval; + if (grace_period != USE_DEFAULT_GRACE_PERIOD) { + retval = pm_update_grace_period(&dqm->packet_mgr, grace_period); + if (retval) + return retval; + } + retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset); if (retval) return retval; @@ -1855,6 +1872,13 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm, return -ETIME; } + /* We need to reset the grace period value for this device */ + if (grace_period != USE_DEFAULT_GRACE_PERIOD) { + if (pm_update_grace_period(&dqm->packet_mgr, + USE_DEFAULT_GRACE_PERIOD)) + pr_err("Failed to reset grace period\n"); + } + pm_release_ib(&dqm->packet_mgr); dqm->active_runlist = false; @@ -1870,7 +1894,7 @@ static int reset_queues_cpsch(struct device_queue_manager *dqm, dqm_lock(dqm); retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID, - pasid, true); + pasid, USE_DEFAULT_GRACE_PERIOD, true); dqm_unlock(dqm); return retval; @@ -1879,13 +1903,14 @@ static int reset_queues_cpsch(struct device_queue_manager *dqm, /* dqm->lock mutex has to be locked before calling this function */ static int execute_queues_cpsch(struct device_queue_manager *dqm, enum kfd_unmap_queues_filter filter, - uint32_t filter_param) + uint32_t filter_param, + uint32_t grace_period) { int retval; if (dqm->is_hws_hang) return -EIO; - retval = unmap_queues_cpsch(dqm, filter, filter_param, false); + retval = unmap_queues_cpsch(dqm, filter, filter_param, grace_period, false); if (retval) return retval; @@ -1943,7 +1968,8 @@ static int destroy_queue_cpsch(struct device_queue_manager *dqm, if (!dqm->dev->kfd->shared_resources.enable_mes) { decrement_queue_count(dqm, qpd, q); retval = execute_queues_cpsch(dqm, - KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); + KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD); if (retval == -ETIME) qpd->reset_wavefronts = true; } else { @@ -2228,7 +2254,7 @@ static int process_termination_cpsch(struct device_queue_manager *dqm, } if (!dqm->dev->kfd->shared_resources.enable_mes) - retval = execute_queues_cpsch(dqm, filter, 0); + retval = execute_queues_cpsch(dqm, filter, 0, USE_DEFAULT_GRACE_PERIOD); if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) { pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev); @@ -2589,7 +2615,8 @@ int dqm_debugfs_hang_hws(struct device_queue_manager *dqm) return r; } dqm->active_runlist = true; - r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0); + r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, + 0, USE_DEFAULT_GRACE_PERIOD); dqm_unlock(dqm); return r; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index cd4383bb207f..d4dd3b4acbf0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -37,6 +37,7 @@ #define KFD_MES_PROCESS_QUANTUM 100000 #define KFD_MES_GANG_QUANTUM 10000 +#define USE_DEFAULT_GRACE_PERIOD 0xffffffff struct device_process_node { struct qcm_process_device *qpd; @@ -259,6 +260,8 @@ struct device_queue_manager { /* used for GFX 9.4.3 only */ uint32_t current_logical_xcc_start; + + uint32_t wait_times; }; void device_queue_manager_init_cik( diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c index 2f54172e9175..401096c103b2 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c @@ -370,6 +370,38 @@ out: return retval; } +int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period) +{ + int retval = 0; + uint32_t *buffer, size; + + size = pm->pmf->set_grace_period_size; + + mutex_lock(&pm->lock); + + if (size) { + kq_acquire_packet_buffer(pm->priv_queue, + size / sizeof(uint32_t), + (unsigned int **)&buffer); + + if (!buffer) { + pr_err("Failed to allocate buffer on kernel queue\n"); + retval = -ENOMEM; + goto out; + } + + retval = pm->pmf->set_grace_period(pm, buffer, grace_period); + if (!retval) + kq_submit_packet(pm->priv_queue); + else + kq_rollback_packet(pm->priv_queue); + } + +out: + mutex_unlock(&pm->lock); + return retval; +} + int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_unmap_queues_filter filter, uint32_t filter_param, bool reset) diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 44cf3a5f6fdb..1fda6dcf84b1 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -262,6 +262,41 @@ static int pm_map_queues_v9(struct packet_manager *pm, uint32_t *buffer, return 0; } +static int pm_set_grace_period_v9(struct packet_manager *pm, + uint32_t *buffer, + uint32_t grace_period) +{ + struct pm4_mec_write_data_mmio *packet; + uint32_t reg_offset = 0; + uint32_t reg_data = 0; + + pm->dqm->dev->kfd2kgd->build_grace_period_packet_info( + pm->dqm->dev->adev, + pm->dqm->wait_times, + grace_period, + ®_offset, + ®_data); + + if (grace_period == USE_DEFAULT_GRACE_PERIOD) + reg_data = pm->dqm->wait_times; + + packet = (struct pm4_mec_write_data_mmio *)buffer; + memset(buffer, 0, sizeof(struct pm4_mec_write_data_mmio)); + + packet->header.u32All = pm_build_pm4_header(IT_WRITE_DATA, + sizeof(struct pm4_mec_write_data_mmio)); + + packet->bitfields2.dst_sel = dst_sel___write_data__mem_mapped_register; + packet->bitfields2.addr_incr = + addr_incr___write_data__do_not_increment_address; + + packet->bitfields3.dst_mmreg_addr = reg_offset; + + packet->data = reg_data; + + return 0; +} + static int pm_unmap_queues_v9(struct packet_manager *pm, uint32_t *buffer, enum kfd_unmap_queues_filter filter, uint32_t filter_param, bool reset) @@ -345,6 +380,7 @@ const struct packet_manager_funcs kfd_v9_pm_funcs = { .set_resources = pm_set_resources_v9, .map_queues = pm_map_queues_v9, .unmap_queues = pm_unmap_queues_v9, + .set_grace_period = pm_set_grace_period_v9, .query_status = pm_query_status_v9, .release_mem = NULL, .map_process_size = sizeof(struct pm4_mes_map_process), @@ -352,6 +388,7 @@ const struct packet_manager_funcs kfd_v9_pm_funcs = { .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), + .set_grace_period_size = sizeof(struct pm4_mec_write_data_mmio), .query_status_size = sizeof(struct pm4_mes_query_status), .release_mem_size = 0, }; @@ -362,6 +399,7 @@ const struct packet_manager_funcs kfd_aldebaran_pm_funcs = { .set_resources = pm_set_resources_v9, .map_queues = pm_map_queues_v9, .unmap_queues = pm_unmap_queues_v9, + .set_grace_period = pm_set_grace_period_v9, .query_status = pm_query_status_v9, .release_mem = NULL, .map_process_size = sizeof(struct pm4_mes_map_process_aldebaran), @@ -369,6 +407,7 @@ const struct packet_manager_funcs kfd_aldebaran_pm_funcs = { .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), + .set_grace_period_size = sizeof(struct pm4_mec_write_data_mmio), .query_status_size = sizeof(struct pm4_mes_query_status), .release_mem_size = 0, }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c index faf4772ed317..c1199d06d131 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_vi.c @@ -303,6 +303,7 @@ const struct packet_manager_funcs kfd_vi_pm_funcs = { .set_resources = pm_set_resources_vi, .map_queues = pm_map_queues_vi, .unmap_queues = pm_unmap_queues_vi, + .set_grace_period = NULL, .query_status = pm_query_status_vi, .release_mem = pm_release_mem_vi, .map_process_size = sizeof(struct pm4_mes_map_process), @@ -310,6 +311,7 @@ const struct packet_manager_funcs kfd_vi_pm_funcs = { .set_resources_size = sizeof(struct pm4_mes_set_resources), .map_queues_size = sizeof(struct pm4_mes_map_queues), .unmap_queues_size = sizeof(struct pm4_mes_unmap_queues), + .set_grace_period_size = 0, .query_status_size = sizeof(struct pm4_mes_query_status), .release_mem_size = sizeof(struct pm4_mec_release_mem) }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index 2ad708c64012..206f1960857f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -584,6 +584,71 @@ struct pm4_mec_release_mem { #endif +#ifndef PM4_MEC_WRITE_DATA_DEFINED +#define PM4_MEC_WRITE_DATA_DEFINED + +enum WRITE_DATA_dst_sel_enum { + dst_sel___write_data__mem_mapped_register = 0, + dst_sel___write_data__tc_l2 = 2, + dst_sel___write_data__gds = 3, + dst_sel___write_data__memory = 5, + dst_sel___write_data__memory_mapped_adc_persistent_state = 6, +}; + +enum WRITE_DATA_addr_incr_enum { + addr_incr___write_data__increment_address = 0, + addr_incr___write_data__do_not_increment_address = 1 +}; + +enum WRITE_DATA_wr_confirm_enum { + wr_confirm___write_data__do_not_wait_for_write_confirmation = 0, + wr_confirm___write_data__wait_for_write_confirmation = 1 +}; + +enum WRITE_DATA_cache_policy_enum { + cache_policy___write_data__lru = 0, + cache_policy___write_data__stream = 1 +}; + + +struct pm4_mec_write_data_mmio { + union { + union PM4_MES_TYPE_3_HEADER header; /*header */ + unsigned int ordinal1; + }; + + union { + struct { + unsigned int reserved1:8; + unsigned int dst_sel:4; + unsigned int reserved2:4; + unsigned int addr_incr:1; + unsigned int reserved3:2; + unsigned int resume_vf:1; + unsigned int wr_confirm:1; + unsigned int reserved4:4; + unsigned int cache_policy:2; + unsigned int reserved5:5; + } bitfields2; + unsigned int ordinal2; + }; + + union { + struct { + unsigned int dst_mmreg_addr:18; + unsigned int reserved6:14; + } bitfields3; + unsigned int ordinal3; + }; + + uint32_t reserved7; + + uint32_t data; + +}; + +#endif + enum { CACHE_FLUSH_AND_INV_TS_EVENT = 0x00000014 }; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 1b272f879b4c..4c912b7735b5 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -1350,6 +1350,8 @@ struct packet_manager_funcs { int (*unmap_queues)(struct packet_manager *pm, uint32_t *buffer, enum kfd_unmap_queues_filter mode, uint32_t filter_param, bool reset); + int (*set_grace_period)(struct packet_manager *pm, uint32_t *buffer, + uint32_t grace_period); int (*query_status)(struct packet_manager *pm, uint32_t *buffer, uint64_t fence_address, uint64_t fence_value); int (*release_mem)(uint64_t gpu_addr, uint32_t *buffer); @@ -1360,6 +1362,7 @@ struct packet_manager_funcs { int set_resources_size; int map_queues_size; int unmap_queues_size; + int set_grace_period_size; int query_status_size; int release_mem_size; }; @@ -1382,6 +1385,8 @@ int pm_send_unmap_queue(struct packet_manager *pm, void pm_release_ib(struct packet_manager *pm); +int pm_update_grace_period(struct packet_manager *pm, uint32_t grace_period); + /* Following PM funcs can be shared among VI and AI */ unsigned int pm_build_pm4_header(unsigned int opcode, size_t packet_size); -- cgit v1.2.3 From 97ae3c8cce96f3bebf883d0812cef5d3fdbe3e64 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 4 Apr 2022 12:27:43 -0400 Subject: drm/amdkfd: prepare map process for single process debug devices Older HW only supports debugging on a single process because the SPI debug mode setting registers are device global. The HWS has supplied a single pinned VMID (0xf) for MAP_PROCESS for debug purposes. To pin the VMID, the KFD will remove the VMID from the HWS dynamic VMID allocation via SET_RESOUCES so that a debugged process will never migrate away from its pinned VMID. The KFD is responsible for reserving and releasing this pinned VMID accordingly whenever the debugger attaches and detaches respectively. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 93 ++++++++++++++++++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 5 ++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 9 +++ drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h | 5 +- 4 files changed, 111 insertions(+), 1 deletion(-) (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 0b88a64e61fe..cfe5bd59070e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -1525,6 +1525,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm) dqm->gws_queue_count = 0; dqm->active_runlist = false; INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception); + dqm->trap_debug_vmid = 0; init_sdma_bitmaps(dqm); @@ -2501,6 +2502,98 @@ static void kfd_process_hw_exception(struct work_struct *work) amdgpu_amdkfd_gpu_reset(dqm->dev->adev); } +int reserve_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd) +{ + int r; + int updated_vmid_mask; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + dqm_lock(dqm); + + if (dqm->trap_debug_vmid != 0) { + pr_err("Trap debug id already reserved\n"); + r = -EBUSY; + goto out_unlock; + } + + r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD, false); + if (r) + goto out_unlock; + + updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; + updated_vmid_mask &= ~(1 << dqm->dev->vm_info.last_vmid_kfd); + + dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask; + dqm->trap_debug_vmid = dqm->dev->vm_info.last_vmid_kfd; + r = set_sched_resources(dqm); + if (r) + goto out_unlock; + + r = map_queues_cpsch(dqm); + if (r) + goto out_unlock; + + pr_debug("Reserved VMID for trap debug: %i\n", dqm->trap_debug_vmid); + +out_unlock: + dqm_unlock(dqm); + return r; +} + +/* + * Releases vmid for the trap debugger + */ +int release_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd) +{ + int r; + int updated_vmid_mask; + uint32_t trap_debug_vmid; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + dqm_lock(dqm); + trap_debug_vmid = dqm->trap_debug_vmid; + if (dqm->trap_debug_vmid == 0) { + pr_err("Trap debug id is not reserved\n"); + r = -EINVAL; + goto out_unlock; + } + + r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, + USE_DEFAULT_GRACE_PERIOD, false); + if (r) + goto out_unlock; + + updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap; + updated_vmid_mask |= (1 << dqm->dev->vm_info.last_vmid_kfd); + + dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask; + dqm->trap_debug_vmid = 0; + r = set_sched_resources(dqm); + if (r) + goto out_unlock; + + r = map_queues_cpsch(dqm); + if (r) + goto out_unlock; + + pr_debug("Released VMID for trap debug: %i\n", trap_debug_vmid); + +out_unlock: + dqm_unlock(dqm); + return r; +} + #if defined(CONFIG_DEBUG_FS) static void seq_reg_dump(struct seq_file *m, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index d4dd3b4acbf0..bf7aa3f84182 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -250,6 +250,7 @@ struct device_queue_manager { struct kfd_mem_obj *fence_mem; bool active_runlist; int sched_policy; + uint32_t trap_debug_vmid; /* hw exception */ bool is_hws_hang; @@ -285,6 +286,10 @@ unsigned int get_queues_per_pipe(struct device_queue_manager *dqm); unsigned int get_pipes_per_mec(struct device_queue_manager *dqm); unsigned int get_num_sdma_queues(struct device_queue_manager *dqm); unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm); +int reserve_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd); +int release_debug_trap_vmid(struct device_queue_manager *dqm, + struct qcm_process_device *qpd); static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 1fda6dcf84b1..0fe73dbd28af 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -34,6 +34,9 @@ static int pm_map_process_v9(struct packet_manager *pm, { struct pm4_mes_map_process *packet; uint64_t vm_page_table_base_addr = qpd->page_table_base; + struct kfd_node *kfd = pm->dqm->dev; + struct kfd_process_device *pdd = + container_of(qpd, struct kfd_process_device, qpd); packet = (struct pm4_mes_map_process *)buffer; memset(buffer, 0, sizeof(struct pm4_mes_map_process)); @@ -49,6 +52,12 @@ static int pm_map_process_v9(struct packet_manager *pm, packet->bitfields14.sdma_enable = 1; packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; + if (kfd->dqm->trap_debug_vmid && pdd->process->debug_trap_enabled && + pdd->process->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) { + packet->bitfields2.debug_vmid = kfd->dqm->trap_debug_vmid; + packet->bitfields2.new_debug = 1; + } + packet->sh_mem_config = qpd->sh_mem_config; packet->sh_mem_bases = qpd->sh_mem_bases; if (qpd->tba_addr) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h index 206f1960857f..8b6b2bd5c148 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_pm4_headers_ai.h @@ -146,7 +146,10 @@ struct pm4_mes_map_process { union { struct { uint32_t pasid:16; - uint32_t reserved1:8; + uint32_t reserved1:2; + uint32_t debug_vmid:4; + uint32_t new_debug:1; + uint32_t reserved2:1; uint32_t diq_enable:1; uint32_t process_quantum:7; } bitfields2; -- cgit v1.2.3 From 0de4ec9a03537bd2b189b5afbf83acd6b72b0258 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Mon, 4 Apr 2022 13:38:11 -0400 Subject: drm/amdgpu: prepare map process for multi-process debug devices Unlike single process debug devices, multi-process debug devices allow debug mode setting per-VMID (non-device-global). Because the HWS manages PASID-VMID mapping, the new MAP_PROCESS API allows the KFD to forward the required SPI debug register write requests. To request a new debug mode setting change, the KFD must be able to preempt all queues then remap all queues with these new setting requests for MAP_PROCESS to take effect. Note that by default, trap enablement in non-debug mode must be disabled for performance reasons for multi-process debug devices due to setup overhead in FW. Signed-off-by: Jonathan Kim Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher --- drivers/gpu/drm/amd/amdkfd/kfd_debug.h | 5 +++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 51 ++++++++++++++++++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h | 3 ++ drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c | 14 ++++++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 9 ++++ drivers/gpu/drm/amd/amdkfd/kfd_process.c | 5 +++ 6 files changed, 87 insertions(+) (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h index a8abfe2a0a14..db6d72e7930f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_debug.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_debug.h @@ -29,4 +29,9 @@ int kfd_dbg_trap_disable(struct kfd_process *target); int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, void __user *runtime_info, uint32_t *runtime_info_size); +static inline bool kfd_dbg_is_per_vmid_supported(struct kfd_node *dev) +{ + return KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2); +} + #endif diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index cfe5bd59070e..495c9238254e 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c @@ -36,6 +36,7 @@ #include "kfd_kernel_queue.h" #include "amdgpu_amdkfd.h" #include "mes_api_def.h" +#include "kfd_debug.h" /* Size of the per-pipe EOP queue */ #define CIK_HPD_EOP_BYTES_LOG2 11 @@ -2594,6 +2595,56 @@ out_unlock: return r; } +int debug_lock_and_unmap(struct device_queue_manager *dqm) +{ + int r; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + if (!kfd_dbg_is_per_vmid_supported(dqm->dev)) + return 0; + + dqm_lock(dqm); + + r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 0, false); + if (r) + dqm_unlock(dqm); + + return r; +} + +int debug_map_and_unlock(struct device_queue_manager *dqm) +{ + int r; + + if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) { + pr_err("Unsupported on sched_policy: %i\n", dqm->sched_policy); + return -EINVAL; + } + + if (!kfd_dbg_is_per_vmid_supported(dqm->dev)) + return 0; + + r = map_queues_cpsch(dqm); + + dqm_unlock(dqm); + + return r; +} + +int debug_refresh_runlist(struct device_queue_manager *dqm) +{ + int r = debug_lock_and_unmap(dqm); + + if (r) + return r; + + return debug_map_and_unlock(dqm); +} + #if defined(CONFIG_DEBUG_FS) static void seq_reg_dump(struct seq_file *m, diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h index bf7aa3f84182..bb75d93712eb 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h @@ -290,6 +290,9 @@ int reserve_debug_trap_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd); int release_debug_trap_vmid(struct device_queue_manager *dqm, struct qcm_process_device *qpd); +int debug_lock_and_unmap(struct device_queue_manager *dqm); +int debug_map_and_unlock(struct device_queue_manager *dqm); +int debug_refresh_runlist(struct device_queue_manager *dqm); static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd) { diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c index 0fe73dbd28af..29a2d0499b67 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager_v9.c @@ -88,6 +88,10 @@ static int pm_map_process_aldebaran(struct packet_manager *pm, { struct pm4_mes_map_process_aldebaran *packet; uint64_t vm_page_table_base_addr = qpd->page_table_base; + struct kfd_dev *kfd = pm->dqm->dev->kfd; + struct kfd_process_device *pdd = + container_of(qpd, struct kfd_process_device, qpd); + int i; packet = (struct pm4_mes_map_process_aldebaran *)buffer; memset(buffer, 0, sizeof(struct pm4_mes_map_process_aldebaran)); @@ -102,6 +106,16 @@ static int pm_map_process_aldebaran(struct packet_manager *pm, packet->bitfields14.num_oac = qpd->num_oac; packet->bitfields14.sdma_enable = 1; packet->bitfields14.num_queues = (qpd->is_debug) ? 0 : qpd->queue_count; + packet->spi_gdbg_per_vmid_cntl = pdd->spi_dbg_override | + pdd->spi_dbg_launch_mode; + + if (pdd->process->debug_trap_enabled) { + for (i = 0; i < kfd->device_info.num_of_watch_points; i++) + packet->tcp_watch_cntl[i] = pdd->watch_points[i]; + + packet->bitfields2.single_memops = + !!(pdd->process->dbg_flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP); + } packet->sh_mem_config = qpd->sh_mem_config; packet->sh_mem_bases = qpd->sh_mem_bases; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 4c912b7735b5..8fca7175daab 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -816,6 +816,12 @@ struct kfd_process_device { uint64_t faults; uint64_t page_in; uint64_t page_out; + + /* Tracks debug per-vmid request settings */ + uint32_t spi_dbg_override; + uint32_t spi_dbg_launch_mode; + uint32_t watch_points[4]; + /* * If this process has been checkpointed before, then the user * application will use the original gpu_id on the @@ -952,6 +958,9 @@ struct kfd_process { bool xnack_enabled; + /* Tracks debug per-vmid request for debug flags */ + bool dbg_flags; + atomic_t poison; /* Queues are in paused stated because we are in the process of doing a CRIU checkpoint */ bool queues_paused; diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_process.c b/drivers/gpu/drm/amd/amdkfd/kfd_process.c index d75dac92775c..725d936b2cc7 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_process.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_process.c @@ -1612,6 +1612,11 @@ struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev, } p->pdds[p->n_pdds++] = pdd; + if (kfd_dbg_is_per_vmid_supported(pdd->dev)) + pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap( + pdd->dev->adev, + false, + 0); /* Init idr used for memory handle translation */ idr_init(&pdd->alloc_idr); -- cgit v1.2.3