diff options
Diffstat (limited to 'drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c')
-rw-r--r-- | drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c | 294 |
1 files changed, 97 insertions, 197 deletions
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c index 7131998848d7..1d312603de9f 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c @@ -26,7 +26,6 @@ #include "kfd_device_queue_manager.h" #include "kfd_kernel_queue.h" #include "kfd_priv.h" -#include "kfd_pm4_headers.h" #include "kfd_pm4_headers_vi.h" #include "kfd_pm4_opcodes.h" @@ -35,7 +34,8 @@ static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes, { unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t); - BUG_ON((temp * sizeof(uint32_t)) > buffer_size_bytes); + WARN((temp * sizeof(uint32_t)) > buffer_size_bytes, + "Runlist IB overflow"); *wptr = temp; } @@ -43,12 +43,12 @@ static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size) { union PM4_MES_TYPE_3_HEADER header; - header.u32all = 0; + header.u32All = 0; header.opcode = opcode; header.count = packet_size/sizeof(uint32_t) - 2; header.type = PM4_TYPE_3; - return header.u32all; + return header.u32All; } static void pm_calc_rlib_size(struct packet_manager *pm, @@ -58,8 +58,6 @@ static void pm_calc_rlib_size(struct packet_manager *pm, unsigned int process_count, queue_count; unsigned int map_queue_size; - BUG_ON(!pm || !rlib_size || !over_subscription); - process_count = pm->dqm->processes_count; queue_count = pm->dqm->queue_count; @@ -67,15 +65,12 @@ static void pm_calc_rlib_size(struct packet_manager *pm, *over_subscription = false; if ((process_count > 1) || queue_count > get_queues_num(pm->dqm)) { *over_subscription = true; - pr_debug("kfd: over subscribed runlist\n"); + pr_debug("Over subscribed runlist\n"); } - map_queue_size = - (pm->dqm->dev->device_info->asic_family == CHIP_CARRIZO) ? - sizeof(struct pm4_mes_map_queues) : - sizeof(struct pm4_map_queues); + map_queue_size = sizeof(struct pm4_mes_map_queues); /* calculate run list ib allocation size */ - *rlib_size = process_count * sizeof(struct pm4_map_process) + + *rlib_size = process_count * sizeof(struct pm4_mes_map_process) + queue_count * map_queue_size; /* @@ -83,9 +78,9 @@ static void pm_calc_rlib_size(struct packet_manager *pm, * when over subscription */ if (*over_subscription) - *rlib_size += sizeof(struct pm4_runlist); + *rlib_size += sizeof(struct pm4_mes_runlist); - pr_debug("kfd: runlist ib size %d\n", *rlib_size); + pr_debug("runlist ib size %d\n", *rlib_size); } static int pm_allocate_runlist_ib(struct packet_manager *pm, @@ -96,17 +91,16 @@ static int pm_allocate_runlist_ib(struct packet_manager *pm, { int retval; - BUG_ON(!pm); - BUG_ON(pm->allocated); - BUG_ON(is_over_subscription == NULL); + if (WARN_ON(pm->allocated)) + return -EINVAL; pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription); retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size, &pm->ib_buffer_obj); - if (retval != 0) { - pr_err("kfd: failed to allocate runlist IB\n"); + if (retval) { + pr_err("Failed to allocate runlist IB\n"); return retval; } @@ -121,15 +115,16 @@ static int pm_allocate_runlist_ib(struct packet_manager *pm, static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer, uint64_t ib, size_t ib_size_in_dwords, bool chain) { - struct pm4_runlist *packet; + struct pm4_mes_runlist *packet; - BUG_ON(!pm || !buffer || !ib); + if (WARN_ON(!ib)) + return -EFAULT; - packet = (struct pm4_runlist *)buffer; + packet = (struct pm4_mes_runlist *)buffer; - memset(buffer, 0, sizeof(struct pm4_runlist)); - packet->header.u32all = build_pm4_header(IT_RUN_LIST, - sizeof(struct pm4_runlist)); + memset(buffer, 0, sizeof(struct pm4_mes_runlist)); + packet->header.u32All = build_pm4_header(IT_RUN_LIST, + sizeof(struct pm4_mes_runlist)); packet->bitfields4.ib_size = ib_size_in_dwords; packet->bitfields4.chain = chain ? 1 : 0; @@ -144,20 +139,16 @@ static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer, static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer, struct qcm_process_device *qpd) { - struct pm4_map_process *packet; + struct pm4_mes_map_process *packet; struct queue *cur; uint32_t num_queues; - BUG_ON(!pm || !buffer || !qpd); - - packet = (struct pm4_map_process *)buffer; - - pr_debug("kfd: In func %s\n", __func__); + packet = (struct pm4_mes_map_process *)buffer; - memset(buffer, 0, sizeof(struct pm4_map_process)); + memset(buffer, 0, sizeof(struct pm4_mes_map_process)); - packet->header.u32all = build_pm4_header(IT_MAP_PROCESS, - sizeof(struct pm4_map_process)); + packet->header.u32All = build_pm4_header(IT_MAP_PROCESS, + sizeof(struct pm4_mes_map_process)); packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0; packet->bitfields2.process_quantum = 1; packet->bitfields2.pasid = qpd->pqm->process->pasid; @@ -175,27 +166,26 @@ static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer, packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base; packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit; + /* TODO: scratch support */ + packet->sh_hidden_private_base_vmid = 0; + packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area); packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area); return 0; } -static int pm_create_map_queue_vi(struct packet_manager *pm, uint32_t *buffer, +static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer, struct queue *q, bool is_static) { struct pm4_mes_map_queues *packet; bool use_static = is_static; - BUG_ON(!pm || !buffer || !q); - - pr_debug("kfd: In func %s\n", __func__); - packet = (struct pm4_mes_map_queues *)buffer; - memset(buffer, 0, sizeof(struct pm4_map_queues)); + memset(buffer, 0, sizeof(struct pm4_mes_map_queues)); - packet->header.u32all = build_pm4_header(IT_MAP_QUEUES, - sizeof(struct pm4_map_queues)); + packet->header.u32All = build_pm4_header(IT_MAP_QUEUES, + sizeof(struct pm4_mes_map_queues)); packet->bitfields2.alloc_format = alloc_format__mes_map_queues__one_per_pipe_vi; packet->bitfields2.num_queues = 1; @@ -223,10 +213,8 @@ static int pm_create_map_queue_vi(struct packet_manager *pm, uint32_t *buffer, use_static = false; /* no static queues under SDMA */ break; default: - pr_err("kfd: in %s queue type %d\n", __func__, - q->properties.type); - BUG(); - break; + WARN(1, "queue type %d", q->properties.type); + return -EINVAL; } packet->bitfields3.doorbell_offset = q->properties.doorbell_off; @@ -246,68 +234,6 @@ static int pm_create_map_queue_vi(struct packet_manager *pm, uint32_t *buffer, return 0; } -static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer, - struct queue *q, bool is_static) -{ - struct pm4_map_queues *packet; - bool use_static = is_static; - - BUG_ON(!pm || !buffer || !q); - - pr_debug("kfd: In func %s\n", __func__); - - packet = (struct pm4_map_queues *)buffer; - memset(buffer, 0, sizeof(struct pm4_map_queues)); - - packet->header.u32all = build_pm4_header(IT_MAP_QUEUES, - sizeof(struct pm4_map_queues)); - packet->bitfields2.alloc_format = - alloc_format__mes_map_queues__one_per_pipe; - packet->bitfields2.num_queues = 1; - packet->bitfields2.queue_sel = - queue_sel__mes_map_queues__map_to_hws_determined_queue_slots; - - packet->bitfields2.vidmem = (q->properties.is_interop) ? - vidmem__mes_map_queues__uses_video_memory : - vidmem__mes_map_queues__uses_no_video_memory; - - switch (q->properties.type) { - case KFD_QUEUE_TYPE_COMPUTE: - case KFD_QUEUE_TYPE_DIQ: - packet->bitfields2.engine_sel = - engine_sel__mes_map_queues__compute; - break; - case KFD_QUEUE_TYPE_SDMA: - packet->bitfields2.engine_sel = - engine_sel__mes_map_queues__sdma0; - use_static = false; /* no static queues under SDMA */ - break; - default: - BUG(); - break; - } - - packet->mes_map_queues_ordinals[0].bitfields3.doorbell_offset = - q->properties.doorbell_off; - - packet->mes_map_queues_ordinals[0].bitfields3.is_static = - (use_static) ? 1 : 0; - - packet->mes_map_queues_ordinals[0].mqd_addr_lo = - lower_32_bits(q->gart_mqd_addr); - - packet->mes_map_queues_ordinals[0].mqd_addr_hi = - upper_32_bits(q->gart_mqd_addr); - - packet->mes_map_queues_ordinals[0].wptr_addr_lo = - lower_32_bits((uint64_t)q->properties.write_ptr); - - packet->mes_map_queues_ordinals[0].wptr_addr_hi = - upper_32_bits((uint64_t)q->properties.write_ptr); - - return 0; -} - static int pm_create_runlist_ib(struct packet_manager *pm, struct list_head *queues, uint64_t *rl_gpu_addr, @@ -322,19 +248,16 @@ static int pm_create_runlist_ib(struct packet_manager *pm, struct kernel_queue *kq; bool is_over_subscription; - BUG_ON(!pm || !queues || !rl_size_bytes || !rl_gpu_addr); - rl_wptr = retval = proccesses_mapped = 0; retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr, &alloc_size_bytes, &is_over_subscription); - if (retval != 0) + if (retval) return retval; *rl_size_bytes = alloc_size_bytes; - pr_debug("kfd: In func %s\n", __func__); - pr_debug("kfd: building runlist ib process count: %d queues count %d\n", + pr_debug("Building runlist ib process count: %d queues count %d\n", pm->dqm->processes_count, pm->dqm->queue_count); /* build the run list ib packet */ @@ -342,42 +265,35 @@ static int pm_create_runlist_ib(struct packet_manager *pm, qpd = cur->qpd; /* build map process packet */ if (proccesses_mapped >= pm->dqm->processes_count) { - pr_debug("kfd: not enough space left in runlist IB\n"); + pr_debug("Not enough space left in runlist IB\n"); pm_release_ib(pm); return -ENOMEM; } retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd); - if (retval != 0) + if (retval) return retval; proccesses_mapped++; - inc_wptr(&rl_wptr, sizeof(struct pm4_map_process), + inc_wptr(&rl_wptr, sizeof(struct pm4_mes_map_process), alloc_size_bytes); list_for_each_entry(kq, &qpd->priv_queue_list, list) { if (!kq->queue->properties.is_active) continue; - pr_debug("kfd: static_queue, mapping kernel q %d, is debug status %d\n", + pr_debug("static_queue, mapping kernel q %d, is debug status %d\n", kq->queue->queue, qpd->is_debug); - if (pm->dqm->dev->device_info->asic_family == - CHIP_CARRIZO) - retval = pm_create_map_queue_vi(pm, - &rl_buffer[rl_wptr], - kq->queue, - qpd->is_debug); - else - retval = pm_create_map_queue(pm, + retval = pm_create_map_queue(pm, &rl_buffer[rl_wptr], kq->queue, qpd->is_debug); - if (retval != 0) + if (retval) return retval; inc_wptr(&rl_wptr, - sizeof(struct pm4_map_queues), + sizeof(struct pm4_mes_map_queues), alloc_size_bytes); } @@ -385,51 +301,44 @@ static int pm_create_runlist_ib(struct packet_manager *pm, if (!q->properties.is_active) continue; - pr_debug("kfd: static_queue, mapping user queue %d, is debug status %d\n", + pr_debug("static_queue, mapping user queue %d, is debug status %d\n", q->queue, qpd->is_debug); - if (pm->dqm->dev->device_info->asic_family == - CHIP_CARRIZO) - retval = pm_create_map_queue_vi(pm, - &rl_buffer[rl_wptr], - q, - qpd->is_debug); - else - retval = pm_create_map_queue(pm, + retval = pm_create_map_queue(pm, &rl_buffer[rl_wptr], q, qpd->is_debug); - if (retval != 0) + if (retval) return retval; inc_wptr(&rl_wptr, - sizeof(struct pm4_map_queues), + sizeof(struct pm4_mes_map_queues), alloc_size_bytes); } } - pr_debug("kfd: finished map process and queues to runlist\n"); + pr_debug("Finished map process and queues to runlist\n"); if (is_over_subscription) - pm_create_runlist(pm, &rl_buffer[rl_wptr], *rl_gpu_addr, - alloc_size_bytes / sizeof(uint32_t), true); + retval = pm_create_runlist(pm, &rl_buffer[rl_wptr], + *rl_gpu_addr, + alloc_size_bytes / sizeof(uint32_t), + true); for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++) pr_debug("0x%2X ", rl_buffer[i]); pr_debug("\n"); - return 0; + return retval; } int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) { - BUG_ON(!dqm); - pm->dqm = dqm; mutex_init(&pm->lock); pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ); - if (pm->priv_queue == NULL) { + if (!pm->priv_queue) { mutex_destroy(&pm->lock); return -ENOMEM; } @@ -440,8 +349,6 @@ int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm) void pm_uninit(struct packet_manager *pm) { - BUG_ON(!pm); - mutex_destroy(&pm->lock); kernel_queue_uninit(pm->priv_queue); } @@ -449,25 +356,22 @@ void pm_uninit(struct packet_manager *pm) int pm_send_set_resources(struct packet_manager *pm, struct scheduling_resources *res) { - struct pm4_set_resources *packet; - - BUG_ON(!pm || !res); - - pr_debug("kfd: In func %s\n", __func__); + struct pm4_mes_set_resources *packet; + int retval = 0; mutex_lock(&pm->lock); pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, sizeof(*packet) / sizeof(uint32_t), - (unsigned int **)&packet); - if (packet == NULL) { - mutex_unlock(&pm->lock); - pr_err("kfd: failed to allocate buffer on kernel queue\n"); - return -ENOMEM; + (unsigned int **)&packet); + if (!packet) { + pr_err("Failed to allocate buffer on kernel queue\n"); + retval = -ENOMEM; + goto out; } - memset(packet, 0, sizeof(struct pm4_set_resources)); - packet->header.u32all = build_pm4_header(IT_SET_RESOURCES, - sizeof(struct pm4_set_resources)); + memset(packet, 0, sizeof(struct pm4_mes_set_resources)); + packet->header.u32All = build_pm4_header(IT_SET_RESOURCES, + sizeof(struct pm4_mes_set_resources)); packet->bitfields2.queue_type = queue_type__mes_set_resources__hsa_interface_queue_hiq; @@ -485,9 +389,10 @@ int pm_send_set_resources(struct packet_manager *pm, pm->priv_queue->ops.submit_packet(pm->priv_queue); +out: mutex_unlock(&pm->lock); - return 0; + return retval; } int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) @@ -497,26 +402,24 @@ int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues) size_t rl_ib_size, packet_size_dwords; int retval; - BUG_ON(!pm || !dqm_queues); - retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr, &rl_ib_size); - if (retval != 0) + if (retval) goto fail_create_runlist_ib; - pr_debug("kfd: runlist IB address: 0x%llX\n", rl_gpu_ib_addr); + pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr); - packet_size_dwords = sizeof(struct pm4_runlist) / sizeof(uint32_t); + packet_size_dwords = sizeof(struct pm4_mes_runlist) / sizeof(uint32_t); mutex_lock(&pm->lock); retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue, packet_size_dwords, &rl_buffer); - if (retval != 0) + if (retval) goto fail_acquire_packet_buffer; retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr, rl_ib_size / sizeof(uint32_t), false); - if (retval != 0) + if (retval) goto fail_create_runlist; pm->priv_queue->ops.submit_packet(pm->priv_queue); @@ -530,8 +433,7 @@ fail_create_runlist: fail_acquire_packet_buffer: mutex_unlock(&pm->lock); fail_create_runlist_ib: - if (pm->allocated) - pm_release_ib(pm); + pm_release_ib(pm); return retval; } @@ -539,20 +441,21 @@ int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, uint32_t fence_value) { int retval; - struct pm4_query_status *packet; + struct pm4_mes_query_status *packet; - BUG_ON(!pm || !fence_address); + if (WARN_ON(!fence_address)) + return -EFAULT; mutex_lock(&pm->lock); retval = pm->priv_queue->ops.acquire_packet_buffer( pm->priv_queue, - sizeof(struct pm4_query_status) / sizeof(uint32_t), + sizeof(struct pm4_mes_query_status) / sizeof(uint32_t), (unsigned int **)&packet); - if (retval != 0) + if (retval) goto fail_acquire_packet_buffer; - packet->header.u32all = build_pm4_header(IT_QUERY_STATUS, - sizeof(struct pm4_query_status)); + packet->header.u32All = build_pm4_header(IT_QUERY_STATUS, + sizeof(struct pm4_mes_query_status)); packet->bitfields2.context_id = 0; packet->bitfields2.interrupt_sel = @@ -566,9 +469,6 @@ int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address, packet->data_lo = lower_32_bits((uint64_t)fence_value); pm->priv_queue->ops.submit_packet(pm->priv_queue); - mutex_unlock(&pm->lock); - - return 0; fail_acquire_packet_buffer: mutex_unlock(&pm->lock); @@ -582,24 +482,22 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, { int retval; uint32_t *buffer; - struct pm4_unmap_queues *packet; - - BUG_ON(!pm); + struct pm4_mes_unmap_queues *packet; mutex_lock(&pm->lock); retval = pm->priv_queue->ops.acquire_packet_buffer( pm->priv_queue, - sizeof(struct pm4_unmap_queues) / sizeof(uint32_t), + sizeof(struct pm4_mes_unmap_queues) / sizeof(uint32_t), &buffer); - if (retval != 0) + if (retval) goto err_acquire_packet_buffer; - packet = (struct pm4_unmap_queues *)buffer; - memset(buffer, 0, sizeof(struct pm4_unmap_queues)); - pr_debug("kfd: static_queue: unmapping queues: mode is %d , reset is %d , type is %d\n", + packet = (struct pm4_mes_unmap_queues *)buffer; + memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues)); + pr_debug("static_queue: unmapping queues: mode is %d , reset is %d , type is %d\n", mode, reset, type); - packet->header.u32all = build_pm4_header(IT_UNMAP_QUEUES, - sizeof(struct pm4_unmap_queues)); + packet->header.u32All = build_pm4_header(IT_UNMAP_QUEUES, + sizeof(struct pm4_mes_unmap_queues)); switch (type) { case KFD_QUEUE_TYPE_COMPUTE: case KFD_QUEUE_TYPE_DIQ: @@ -611,8 +509,9 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, engine_sel__mes_unmap_queues__sdma0 + sdma_engine; break; default: - BUG(); - break; + WARN(1, "queue type %d", type); + retval = -EINVAL; + goto err_invalid; } if (reset) @@ -636,16 +535,17 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, break; case KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES: packet->bitfields2.queue_sel = - queue_sel__mes_unmap_queues__perform_request_on_all_active_queues; + queue_sel__mes_unmap_queues__unmap_all_queues; break; case KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES: /* in this case, we do not preempt static queues */ packet->bitfields2.queue_sel = - queue_sel__mes_unmap_queues__perform_request_on_dynamic_queues_only; + queue_sel__mes_unmap_queues__unmap_all_non_static_queues; break; default: - BUG(); - break; + WARN(1, "filter %d", mode); + retval = -EINVAL; + goto err_invalid; } pm->priv_queue->ops.submit_packet(pm->priv_queue); @@ -653,6 +553,8 @@ int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type, mutex_unlock(&pm->lock); return 0; +err_invalid: + pm->priv_queue->ops.rollback_packet(pm->priv_queue); err_acquire_packet_buffer: mutex_unlock(&pm->lock); return retval; @@ -660,8 +562,6 @@ err_acquire_packet_buffer: void pm_release_ib(struct packet_manager *pm) { - BUG_ON(!pm); - mutex_lock(&pm->lock); if (pm->allocated) { kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj); |