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
allocate_hiq_mqd(struct kfd_node * dev,struct queue_properties * q)49 struct kfd_mem_obj *allocate_hiq_mqd(struct kfd_node *dev, struct queue_properties *q)
50 {
51 struct kfd_mem_obj *mqd_mem_obj;
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
allocate_sdma_mqd(struct kfd_node * dev,struct queue_properties * q)64 struct kfd_mem_obj *allocate_sdma_mqd(struct kfd_node *dev,
65 struct queue_properties *q)
66 {
67 struct kfd_mem_obj *mqd_mem_obj;
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->kfd->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 NUM_XCC(dev->xcc_mask);
81
82 mqd_mem_obj->gtt_mem = (void *)((uint64_t)dev->dqm->hiq_sdma_mqd.gtt_mem
83 + offset);
84 mqd_mem_obj->gpu_addr = dev->dqm->hiq_sdma_mqd.gpu_addr + offset;
85 mqd_mem_obj->cpu_ptr = (uint32_t *)((uint64_t)
86 dev->dqm->hiq_sdma_mqd.cpu_ptr + offset);
87
88 return mqd_mem_obj;
89 }
90
free_mqd_hiq_sdma(struct mqd_manager * mm,void * mqd,struct kfd_mem_obj * mqd_mem_obj)91 void free_mqd_hiq_sdma(struct mqd_manager *mm, void *mqd,
92 struct kfd_mem_obj *mqd_mem_obj)
93 {
94 WARN_ON(!mqd_mem_obj->gtt_mem);
95 kfree(mqd_mem_obj);
96 }
97
mqd_symmetrically_map_cu_mask(struct mqd_manager * mm,const uint32_t * cu_mask,uint32_t cu_mask_count,uint32_t * se_mask,uint32_t inst)98 void mqd_symmetrically_map_cu_mask(struct mqd_manager *mm,
99 const uint32_t *cu_mask, uint32_t cu_mask_count,
100 uint32_t *se_mask, uint32_t inst)
101 {
102 struct amdgpu_cu_info *cu_info = &mm->dev->adev->gfx.cu_info;
103 struct amdgpu_gfx_config *gfx_info = &mm->dev->adev->gfx.config;
104 uint32_t cu_per_sh[KFD_MAX_NUM_SE][KFD_MAX_NUM_SH_PER_SE] = {0};
105 bool wgp_mode_req = KFD_GC_VERSION(mm->dev) >= IP_VERSION(10, 0, 0);
106 uint32_t en_mask = wgp_mode_req ? 0x3 : 0x1;
107 int i, se, sh, cu, cu_bitmap_sh_mul, cu_inc = wgp_mode_req ? 2 : 1;
108 uint32_t cu_active_per_node;
109 int inc = cu_inc * NUM_XCC(mm->dev->xcc_mask);
110 int xcc_inst = inst + ffs(mm->dev->xcc_mask) - 1;
111
112 cu_active_per_node = cu_info->number / mm->dev->kfd->num_nodes;
113 if (cu_mask_count > cu_active_per_node)
114 cu_mask_count = cu_active_per_node;
115
116 /* Exceeding these bounds corrupts the stack and indicates a coding error.
117 * Returning with no CU's enabled will hang the queue, which should be
118 * attention grabbing.
119 */
120 if (gfx_info->max_shader_engines > KFD_MAX_NUM_SE) {
121 dev_err(mm->dev->adev->dev,
122 "Exceeded KFD_MAX_NUM_SE, chip reports %d\n",
123 gfx_info->max_shader_engines);
124 return;
125 }
126 if (gfx_info->max_sh_per_se > KFD_MAX_NUM_SH_PER_SE) {
127 dev_err(mm->dev->adev->dev,
128 "Exceeded KFD_MAX_NUM_SH, chip reports %d\n",
129 gfx_info->max_sh_per_se * gfx_info->max_shader_engines);
130 return;
131 }
132
133 cu_bitmap_sh_mul = (KFD_GC_VERSION(mm->dev) >= IP_VERSION(11, 0, 0) &&
134 KFD_GC_VERSION(mm->dev) < IP_VERSION(13, 0, 0)) ? 2 : 1;
135
136 /* Count active CUs per SH.
137 *
138 * Some CUs in an SH may be disabled. HW expects disabled CUs to be
139 * represented in the high bits of each SH's enable mask (the upper and lower
140 * 16 bits of se_mask) and will take care of the actual distribution of
141 * disabled CUs within each SH automatically.
142 * Each half of se_mask must be filled only on bits 0-cu_per_sh[se][sh]-1.
143 *
144 * See note on Arcturus cu_bitmap layout in gfx_v9_0_get_cu_info.
145 * See note on GFX11 cu_bitmap layout in gfx_v11_0_get_cu_info.
146 */
147 for (se = 0; se < gfx_info->max_shader_engines; se++)
148 for (sh = 0; sh < gfx_info->max_sh_per_se; sh++)
149 cu_per_sh[se][sh] = hweight32(
150 cu_info->bitmap[xcc_inst][se % 4][sh + (se / 4) *
151 cu_bitmap_sh_mul]);
152
153 /* Symmetrically map cu_mask to all SEs & SHs:
154 * se_mask programs up to 2 SH in the upper and lower 16 bits.
155 *
156 * Examples
157 * Assuming 1 SH/SE, 4 SEs:
158 * cu_mask[0] bit0 -> se_mask[0] bit0
159 * cu_mask[0] bit1 -> se_mask[1] bit0
160 * ...
161 * cu_mask[0] bit4 -> se_mask[0] bit1
162 * ...
163 *
164 * Assuming 2 SH/SE, 4 SEs
165 * cu_mask[0] bit0 -> se_mask[0] bit0 (SE0,SH0,CU0)
166 * cu_mask[0] bit1 -> se_mask[1] bit0 (SE1,SH0,CU0)
167 * ...
168 * cu_mask[0] bit4 -> se_mask[0] bit16 (SE0,SH1,CU0)
169 * cu_mask[0] bit5 -> se_mask[1] bit16 (SE1,SH1,CU0)
170 * ...
171 * cu_mask[0] bit8 -> se_mask[0] bit1 (SE0,SH0,CU1)
172 * ...
173 *
174 * For GFX 9.4.3, the following code only looks at a
175 * subset of the cu_mask corresponding to the inst parameter.
176 * If we have n XCCs under one GPU node
177 * cu_mask[0] bit0 -> XCC0 se_mask[0] bit0 (XCC0,SE0,SH0,CU0)
178 * cu_mask[0] bit1 -> XCC1 se_mask[0] bit0 (XCC1,SE0,SH0,CU0)
179 * ..
180 * cu_mask[0] bitn -> XCCn se_mask[0] bit0 (XCCn,SE0,SH0,CU0)
181 * cu_mask[0] bit n+1 -> XCC0 se_mask[1] bit0 (XCC0,SE1,SH0,CU0)
182 *
183 * For example, if there are 6 XCCs under 1 KFD node, this code
184 * running for each inst, will look at the bits as:
185 * inst, inst + 6, inst + 12...
186 *
187 * First ensure all CUs are disabled, then enable user specified CUs.
188 */
189 for (i = 0; i < gfx_info->max_shader_engines; i++)
190 se_mask[i] = 0;
191
192 i = inst;
193 for (cu = 0; cu < 16; cu += cu_inc) {
194 for (sh = 0; sh < gfx_info->max_sh_per_se; sh++) {
195 for (se = 0; se < gfx_info->max_shader_engines; se++) {
196 if (cu_per_sh[se][sh] > cu) {
197 if (cu_mask[i / 32] & (en_mask << (i % 32)))
198 se_mask[se] |= en_mask << (cu + sh * 16);
199 i += inc;
200 if (i >= cu_mask_count)
201 return;
202 }
203 }
204 }
205 }
206 }
207
kfd_hiq_load_mqd_kiq(struct mqd_manager * mm,void * mqd,uint32_t pipe_id,uint32_t queue_id,struct queue_properties * p,struct mm_struct * mms)208 int kfd_hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd,
209 uint32_t pipe_id, uint32_t queue_id,
210 struct queue_properties *p, struct mm_struct *mms)
211 {
212 return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, mqd, pipe_id,
213 queue_id, p->doorbell_off, 0);
214 }
215
kfd_destroy_mqd_cp(struct mqd_manager * mm,void * mqd,enum kfd_preempt_type type,unsigned int timeout,uint32_t pipe_id,uint32_t queue_id)216 int kfd_destroy_mqd_cp(struct mqd_manager *mm, void *mqd,
217 enum kfd_preempt_type type, unsigned int timeout,
218 uint32_t pipe_id, uint32_t queue_id)
219 {
220 return mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, mqd, type, timeout,
221 pipe_id, queue_id, 0);
222 }
223
kfd_free_mqd_cp(struct mqd_manager * mm,void * mqd,struct kfd_mem_obj * mqd_mem_obj)224 void kfd_free_mqd_cp(struct mqd_manager *mm, void *mqd,
225 struct kfd_mem_obj *mqd_mem_obj)
226 {
227 if (mqd_mem_obj->gtt_mem) {
228 amdgpu_amdkfd_free_gtt_mem(mm->dev->adev, &mqd_mem_obj->gtt_mem);
229 kfree(mqd_mem_obj);
230 } else {
231 kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
232 }
233 }
234
kfd_is_occupied_cp(struct mqd_manager * mm,void * mqd,uint64_t queue_address,uint32_t pipe_id,uint32_t queue_id)235 bool kfd_is_occupied_cp(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_is_occupied(mm->dev->adev, queue_address,
240 pipe_id, queue_id, 0);
241 }
242
kfd_load_mqd_sdma(struct mqd_manager * mm,void * mqd,uint32_t pipe_id,uint32_t queue_id,struct queue_properties * p,struct mm_struct * mms)243 int kfd_load_mqd_sdma(struct mqd_manager *mm, void *mqd,
244 uint32_t pipe_id, uint32_t queue_id,
245 struct queue_properties *p, struct mm_struct *mms)
246 {
247 return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->adev, mqd,
248 (uint32_t __user *)p->write_ptr,
249 mms);
250 }
251
252 /*
253 * preempt type here is ignored because there is only one way
254 * to preempt sdma queue
255 */
kfd_destroy_mqd_sdma(struct mqd_manager * mm,void * mqd,enum kfd_preempt_type type,unsigned int timeout,uint32_t pipe_id,uint32_t queue_id)256 int kfd_destroy_mqd_sdma(struct mqd_manager *mm, void *mqd,
257 enum kfd_preempt_type type,
258 unsigned int timeout, uint32_t pipe_id,
259 uint32_t queue_id)
260 {
261 return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->adev, mqd, timeout);
262 }
263
kfd_is_occupied_sdma(struct mqd_manager * mm,void * mqd,uint64_t queue_address,uint32_t pipe_id,uint32_t queue_id)264 bool kfd_is_occupied_sdma(struct mqd_manager *mm, void *mqd,
265 uint64_t queue_address, uint32_t pipe_id,
266 uint32_t queue_id)
267 {
268 return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->adev, mqd);
269 }
270
kfd_hiq_mqd_stride(struct kfd_node * dev)271 uint64_t kfd_hiq_mqd_stride(struct kfd_node *dev)
272 {
273 return dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size;
274 }
275
kfd_get_hiq_xcc_mqd(struct kfd_node * dev,struct kfd_mem_obj * mqd_mem_obj,uint32_t virtual_xcc_id)276 void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd_mem_obj,
277 uint32_t virtual_xcc_id)
278 {
279 uint64_t offset;
280
281 offset = kfd_hiq_mqd_stride(dev) * virtual_xcc_id;
282
283 mqd_mem_obj->gtt_mem = (virtual_xcc_id == 0) ?
284 dev->dqm->hiq_sdma_mqd.gtt_mem : NULL;
285 mqd_mem_obj->gpu_addr = dev->dqm->hiq_sdma_mqd.gpu_addr + offset;
286 mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t)
287 dev->dqm->hiq_sdma_mqd.cpu_ptr + offset);
288 }
289
kfd_mqd_stride(struct mqd_manager * mm,struct queue_properties * q)290 uint64_t kfd_mqd_stride(struct mqd_manager *mm,
291 struct queue_properties *q)
292 {
293 return mm->mqd_size;
294 }
295
kfd_check_hiq_mqd_doorbell_id(struct kfd_node * node,uint32_t doorbell_id,uint32_t inst)296 bool kfd_check_hiq_mqd_doorbell_id(struct kfd_node *node, uint32_t doorbell_id,
297 uint32_t inst)
298 {
299 if (doorbell_id) {
300 struct device *dev = node->adev->dev;
301
302 if (node->adev->xcp_mgr && node->adev->xcp_mgr->num_xcps > 0)
303 dev_err(dev, "XCC %d: Queue preemption failed for queue with doorbell_id: %x\n",
304 inst, doorbell_id);
305 else
306 dev_err(dev, "Queue preemption failed for queue with doorbell_id: %x\n",
307 doorbell_id);
308 return true;
309 }
310
311 return false;
312 }
313