1 // SPDX-License-Identifier: GPL-2.0 OR MIT 2 /* 3 * Copyright 2014-2022 Advanced Micro Devices, Inc. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice shall be included in 13 * all copies or substantial portions of the Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 21 * OTHER DEALINGS IN THE SOFTWARE. 22 * 23 */ 24 25 #include "kfd_mqd_manager.h" 26 #include "amdgpu_amdkfd.h" 27 #include "kfd_device_queue_manager.h" 28 29 /* Mapping queue priority to pipe priority, indexed by queue priority */ 30 int pipe_priority_map[] = { 31 KFD_PIPE_PRIORITY_CS_LOW, 32 KFD_PIPE_PRIORITY_CS_LOW, 33 KFD_PIPE_PRIORITY_CS_LOW, 34 KFD_PIPE_PRIORITY_CS_LOW, 35 KFD_PIPE_PRIORITY_CS_LOW, 36 KFD_PIPE_PRIORITY_CS_LOW, 37 KFD_PIPE_PRIORITY_CS_LOW, 38 KFD_PIPE_PRIORITY_CS_MEDIUM, 39 KFD_PIPE_PRIORITY_CS_MEDIUM, 40 KFD_PIPE_PRIORITY_CS_MEDIUM, 41 KFD_PIPE_PRIORITY_CS_MEDIUM, 42 KFD_PIPE_PRIORITY_CS_HIGH, 43 KFD_PIPE_PRIORITY_CS_HIGH, 44 KFD_PIPE_PRIORITY_CS_HIGH, 45 KFD_PIPE_PRIORITY_CS_HIGH, 46 KFD_PIPE_PRIORITY_CS_HIGH 47 }; 48 49 struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_dev *dev, struct queue_properties *q) 50 { 51 struct kfd_mem_obj *mqd_mem_obj = NULL; 52 53 mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); 54 if (!mqd_mem_obj) 55 return NULL; 56 57 mqd_mem_obj->gtt_mem = dev->dqm->hiq_sdma_mqd.gtt_mem; 58 mqd_mem_obj->gpu_addr = dev->dqm->hiq_sdma_mqd.gpu_addr; 59 mqd_mem_obj->cpu_ptr = dev->dqm->hiq_sdma_mqd.cpu_ptr; 60 61 return mqd_mem_obj; 62 } 63 64 struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_dev *dev, 65 struct queue_properties *q) 66 { 67 struct kfd_mem_obj *mqd_mem_obj = NULL; 68 uint64_t offset; 69 70 mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL); 71 if (!mqd_mem_obj) 72 return NULL; 73 74 offset = (q->sdma_engine_id * 75 dev->device_info.num_sdma_queues_per_engine + 76 q->sdma_queue_id) * 77 dev->dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size; 78 79 offset += dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size; 80 81 mqd_mem_obj->gtt_mem = (void *)((uint64_t)dev->dqm->hiq_sdma_mqd.gtt_mem 82 + offset); 83 mqd_mem_obj->gpu_addr = dev->dqm->hiq_sdma_mqd.gpu_addr + offset; 84 mqd_mem_obj->cpu_ptr = (uint32_t *)((uint64_t) 85 dev->dqm->hiq_sdma_mqd.cpu_ptr + offset); 86 87 return mqd_mem_obj; 88 } 89 90 void free_mqd_hiq_sdma(struct mqd_manager *mm, void *mqd, 91 struct kfd_mem_obj *mqd_mem_obj) 92 { 93 WARN_ON(!mqd_mem_obj->gtt_mem); 94 kfree(mqd_mem_obj); 95 } 96 97 void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm, 98 const uint32_t *cu_mask, uint32_t cu_mask_count, 99 uint32_t *se_mask) 100 { 101 struct kfd_cu_info cu_info; 102 uint32_t cu_per_sh[KFD_MAX_NUM_SE][KFD_MAX_NUM_SH_PER_SE] = {0}; 103 int i, se, sh, cu; 104 105 amdgpu_amdkfd_get_cu_info(mm->dev->adev, &cu_info); 106 107 if (cu_mask_count > cu_info.cu_active_number) 108 cu_mask_count = cu_info.cu_active_number; 109 110 /* Exceeding these bounds corrupts the stack and indicates a coding error. 111 * Returning with no CU's enabled will hang the queue, which should be 112 * attention grabbing. 113 */ 114 if (cu_info.num_shader_engines > KFD_MAX_NUM_SE) { 115 pr_err("Exceeded KFD_MAX_NUM_SE, chip reports %d\n", cu_info.num_shader_engines); 116 return; 117 } 118 if (cu_info.num_shader_arrays_per_engine > KFD_MAX_NUM_SH_PER_SE) { 119 pr_err("Exceeded KFD_MAX_NUM_SH, chip reports %d\n", 120 cu_info.num_shader_arrays_per_engine * cu_info.num_shader_engines); 121 return; 122 } 123 /* Count active CUs per SH. 124 * 125 * Some CUs in an SH may be disabled. HW expects disabled CUs to be 126 * represented in the high bits of each SH's enable mask (the upper and lower 127 * 16 bits of se_mask) and will take care of the actual distribution of 128 * disabled CUs within each SH automatically. 129 * Each half of se_mask must be filled only on bits 0-cu_per_sh[se][sh]-1. 130 * 131 * See note on Arcturus cu_bitmap layout in gfx_v9_0_get_cu_info. 132 */ 133 for (se = 0; se < cu_info.num_shader_engines; se++) 134 for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++) 135 cu_per_sh[se][sh] = hweight32(cu_info.cu_bitmap[se % 4][sh + (se / 4)]); 136 137 /* Symmetrically map cu_mask to all SEs & SHs: 138 * se_mask programs up to 2 SH in the upper and lower 16 bits. 139 * 140 * Examples 141 * Assuming 1 SH/SE, 4 SEs: 142 * cu_mask[0] bit0 -> se_mask[0] bit0 143 * cu_mask[0] bit1 -> se_mask[1] bit0 144 * ... 145 * cu_mask[0] bit4 -> se_mask[0] bit1 146 * ... 147 * 148 * Assuming 2 SH/SE, 4 SEs 149 * cu_mask[0] bit0 -> se_mask[0] bit0 (SE0,SH0,CU0) 150 * cu_mask[0] bit1 -> se_mask[1] bit0 (SE1,SH0,CU0) 151 * ... 152 * cu_mask[0] bit4 -> se_mask[0] bit16 (SE0,SH1,CU0) 153 * cu_mask[0] bit5 -> se_mask[1] bit16 (SE1,SH1,CU0) 154 * ... 155 * cu_mask[0] bit8 -> se_mask[0] bit1 (SE0,SH0,CU1) 156 * ... 157 * 158 * First ensure all CUs are disabled, then enable user specified CUs. 159 */ 160 for (i = 0; i < cu_info.num_shader_engines; i++) 161 se_mask[i] = 0; 162 163 i = 0; 164 for (cu = 0; cu < 16; cu++) { 165 for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++) { 166 for (se = 0; se < cu_info.num_shader_engines; se++) { 167 if (cu_per_sh[se][sh] > cu) { 168 if (cu_mask[i / 32] & (1 << (i % 32))) 169 se_mask[se] |= 1 << (cu + sh * 16); 170 i++; 171 if (i == cu_mask_count) 172 return; 173 } 174 } 175 } 176 } 177 } 178 179 int kfd_hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd, 180 uint32_t pipe_id, uint32_t queue_id, 181 struct queue_properties *p, struct mm_struct *mms) 182 { 183 return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, mqd, pipe_id, 184 queue_id, p->doorbell_off); 185 } 186 187 int kfd_destroy_mqd_cp(struct mqd_manager *mm, void *mqd, 188 enum kfd_preempt_type type, unsigned int timeout, 189 uint32_t pipe_id, uint32_t queue_id) 190 { 191 return mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, mqd, type, timeout, 192 pipe_id, queue_id); 193 } 194 195 void kfd_free_mqd_cp(struct mqd_manager *mm, void *mqd, 196 struct kfd_mem_obj *mqd_mem_obj) 197 { 198 if (mqd_mem_obj->gtt_mem) { 199 amdgpu_amdkfd_free_gtt_mem(mm->dev->adev, mqd_mem_obj->gtt_mem); 200 kfree(mqd_mem_obj); 201 } else { 202 kfd_gtt_sa_free(mm->dev, mqd_mem_obj); 203 } 204 } 205 206 bool kfd_is_occupied_cp(struct mqd_manager *mm, void *mqd, 207 uint64_t queue_address, uint32_t pipe_id, 208 uint32_t queue_id) 209 { 210 return mm->dev->kfd2kgd->hqd_is_occupied(mm->dev->adev, queue_address, 211 pipe_id, queue_id); 212 } 213 214 int kfd_load_mqd_sdma(struct mqd_manager *mm, void *mqd, 215 uint32_t pipe_id, uint32_t queue_id, 216 struct queue_properties *p, struct mm_struct *mms) 217 { 218 return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->adev, mqd, 219 (uint32_t __user *)p->write_ptr, 220 mms); 221 } 222 223 /* 224 * preempt type here is ignored because there is only one way 225 * to preempt sdma queue 226 */ 227 int kfd_destroy_mqd_sdma(struct mqd_manager *mm, void *mqd, 228 enum kfd_preempt_type type, 229 unsigned int timeout, uint32_t pipe_id, 230 uint32_t queue_id) 231 { 232 return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->adev, mqd, timeout); 233 } 234 235 bool kfd_is_occupied_sdma(struct mqd_manager *mm, void *mqd, 236 uint64_t queue_address, uint32_t pipe_id, 237 uint32_t queue_id) 238 { 239 return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->adev, mqd); 240 } 241