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_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 
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 
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 
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 kfd_cu_info cu_info;
103 	uint32_t cu_per_sh[KFD_MAX_NUM_SE][KFD_MAX_NUM_SH_PER_SE] = {0};
104 	bool wgp_mode_req = KFD_GC_VERSION(mm->dev) >= IP_VERSION(10, 0, 0);
105 	uint32_t en_mask = wgp_mode_req ? 0x3 : 0x1;
106 	int i, se, sh, cu, cu_bitmap_sh_mul, cu_inc = wgp_mode_req ? 2 : 1;
107 	uint32_t cu_active_per_node;
108 	int inc = cu_inc * NUM_XCC(mm->dev->xcc_mask);
109 	int xcc_inst = inst + ffs(mm->dev->xcc_mask) - 1;
110 
111 	amdgpu_amdkfd_get_cu_info(mm->dev->adev, &cu_info);
112 
113 	cu_active_per_node = cu_info.cu_active_number / mm->dev->kfd->num_nodes;
114 	if (cu_mask_count > cu_active_per_node)
115 		cu_mask_count = cu_active_per_node;
116 
117 	/* Exceeding these bounds corrupts the stack and indicates a coding error.
118 	 * Returning with no CU's enabled will hang the queue, which should be
119 	 * attention grabbing.
120 	 */
121 	if (cu_info.num_shader_engines > KFD_MAX_NUM_SE) {
122 		pr_err("Exceeded KFD_MAX_NUM_SE, chip reports %d\n", cu_info.num_shader_engines);
123 		return;
124 	}
125 	if (cu_info.num_shader_arrays_per_engine > KFD_MAX_NUM_SH_PER_SE) {
126 		pr_err("Exceeded KFD_MAX_NUM_SH, chip reports %d\n",
127 			cu_info.num_shader_arrays_per_engine * cu_info.num_shader_engines);
128 		return;
129 	}
130 
131 	cu_bitmap_sh_mul = (KFD_GC_VERSION(mm->dev) >= IP_VERSION(11, 0, 0) &&
132 			    KFD_GC_VERSION(mm->dev) < IP_VERSION(12, 0, 0)) ? 2 : 1;
133 
134 	/* Count active CUs per SH.
135 	 *
136 	 * Some CUs in an SH may be disabled.	HW expects disabled CUs to be
137 	 * represented in the high bits of each SH's enable mask (the upper and lower
138 	 * 16 bits of se_mask) and will take care of the actual distribution of
139 	 * disabled CUs within each SH automatically.
140 	 * Each half of se_mask must be filled only on bits 0-cu_per_sh[se][sh]-1.
141 	 *
142 	 * See note on Arcturus cu_bitmap layout in gfx_v9_0_get_cu_info.
143 	 * See note on GFX11 cu_bitmap layout in gfx_v11_0_get_cu_info.
144 	 */
145 	for (se = 0; se < cu_info.num_shader_engines; se++)
146 		for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++)
147 			cu_per_sh[se][sh] = hweight32(
148 				cu_info.cu_bitmap[xcc_inst][se % 4][sh + (se / 4) *
149 				cu_bitmap_sh_mul]);
150 
151 	/* Symmetrically map cu_mask to all SEs & SHs:
152 	 * se_mask programs up to 2 SH in the upper and lower 16 bits.
153 	 *
154 	 * Examples
155 	 * Assuming 1 SH/SE, 4 SEs:
156 	 * cu_mask[0] bit0 -> se_mask[0] bit0
157 	 * cu_mask[0] bit1 -> se_mask[1] bit0
158 	 * ...
159 	 * cu_mask[0] bit4 -> se_mask[0] bit1
160 	 * ...
161 	 *
162 	 * Assuming 2 SH/SE, 4 SEs
163 	 * cu_mask[0] bit0 -> se_mask[0] bit0 (SE0,SH0,CU0)
164 	 * cu_mask[0] bit1 -> se_mask[1] bit0 (SE1,SH0,CU0)
165 	 * ...
166 	 * cu_mask[0] bit4 -> se_mask[0] bit16 (SE0,SH1,CU0)
167 	 * cu_mask[0] bit5 -> se_mask[1] bit16 (SE1,SH1,CU0)
168 	 * ...
169 	 * cu_mask[0] bit8 -> se_mask[0] bit1 (SE0,SH0,CU1)
170 	 * ...
171 	 *
172 	 * For GFX 9.4.3, the following code only looks at a
173 	 * subset of the cu_mask corresponding to the inst parameter.
174 	 * If we have n XCCs under one GPU node
175 	 * cu_mask[0] bit0 -> XCC0 se_mask[0] bit0 (XCC0,SE0,SH0,CU0)
176 	 * cu_mask[0] bit1 -> XCC1 se_mask[0] bit0 (XCC1,SE0,SH0,CU0)
177 	 * ..
178 	 * cu_mask[0] bitn -> XCCn se_mask[0] bit0 (XCCn,SE0,SH0,CU0)
179 	 * cu_mask[0] bit n+1 -> XCC0 se_mask[1] bit0 (XCC0,SE1,SH0,CU0)
180 	 *
181 	 * For example, if there are 6 XCCs under 1 KFD node, this code
182 	 * running for each inst, will look at the bits as:
183 	 * inst, inst + 6, inst + 12...
184 	 *
185 	 * First ensure all CUs are disabled, then enable user specified CUs.
186 	 */
187 	for (i = 0; i < cu_info.num_shader_engines; i++)
188 		se_mask[i] = 0;
189 
190 	i = inst;
191 	for (cu = 0; cu < 16; cu += cu_inc) {
192 		for (sh = 0; sh < cu_info.num_shader_arrays_per_engine; sh++) {
193 			for (se = 0; se < cu_info.num_shader_engines; se++) {
194 				if (cu_per_sh[se][sh] > cu) {
195 					if (cu_mask[i / 32] & (en_mask << (i % 32)))
196 						se_mask[se] |= en_mask << (cu + sh * 16);
197 					i += inc;
198 					if (i >= cu_mask_count)
199 						return;
200 				}
201 			}
202 		}
203 	}
204 }
205 
206 int kfd_hiq_load_mqd_kiq(struct mqd_manager *mm, void *mqd,
207 		     uint32_t pipe_id, uint32_t queue_id,
208 		     struct queue_properties *p, struct mm_struct *mms)
209 {
210 	return mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, mqd, pipe_id,
211 					      queue_id, p->doorbell_off, 0);
212 }
213 
214 int kfd_destroy_mqd_cp(struct mqd_manager *mm, void *mqd,
215 		enum kfd_preempt_type type, unsigned int timeout,
216 		uint32_t pipe_id, uint32_t queue_id)
217 {
218 	return mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, mqd, type, timeout,
219 						pipe_id, queue_id, 0);
220 }
221 
222 void kfd_free_mqd_cp(struct mqd_manager *mm, void *mqd,
223 	      struct kfd_mem_obj *mqd_mem_obj)
224 {
225 	if (mqd_mem_obj->gtt_mem) {
226 		amdgpu_amdkfd_free_gtt_mem(mm->dev->adev, mqd_mem_obj->gtt_mem);
227 		kfree(mqd_mem_obj);
228 	} else {
229 		kfd_gtt_sa_free(mm->dev, mqd_mem_obj);
230 	}
231 }
232 
233 bool kfd_is_occupied_cp(struct mqd_manager *mm, void *mqd,
234 		 uint64_t queue_address, uint32_t pipe_id,
235 		 uint32_t queue_id)
236 {
237 	return mm->dev->kfd2kgd->hqd_is_occupied(mm->dev->adev, queue_address,
238 						pipe_id, queue_id, 0);
239 }
240 
241 int kfd_load_mqd_sdma(struct mqd_manager *mm, void *mqd,
242 		  uint32_t pipe_id, uint32_t queue_id,
243 		  struct queue_properties *p, struct mm_struct *mms)
244 {
245 	return mm->dev->kfd2kgd->hqd_sdma_load(mm->dev->adev, mqd,
246 						(uint32_t __user *)p->write_ptr,
247 						mms);
248 }
249 
250 /*
251  * preempt type here is ignored because there is only one way
252  * to preempt sdma queue
253  */
254 int kfd_destroy_mqd_sdma(struct mqd_manager *mm, void *mqd,
255 		     enum kfd_preempt_type type,
256 		     unsigned int timeout, uint32_t pipe_id,
257 		     uint32_t queue_id)
258 {
259 	return mm->dev->kfd2kgd->hqd_sdma_destroy(mm->dev->adev, mqd, timeout);
260 }
261 
262 bool kfd_is_occupied_sdma(struct mqd_manager *mm, void *mqd,
263 		      uint64_t queue_address, uint32_t pipe_id,
264 		      uint32_t queue_id)
265 {
266 	return mm->dev->kfd2kgd->hqd_sdma_is_occupied(mm->dev->adev, mqd);
267 }
268 
269 uint64_t kfd_hiq_mqd_stride(struct kfd_node *dev)
270 {
271 	return dev->dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size;
272 }
273 
274 void kfd_get_hiq_xcc_mqd(struct kfd_node *dev, struct kfd_mem_obj *mqd_mem_obj,
275 		     uint32_t virtual_xcc_id)
276 {
277 	uint64_t offset;
278 
279 	offset = kfd_hiq_mqd_stride(dev) * virtual_xcc_id;
280 
281 	mqd_mem_obj->gtt_mem = (virtual_xcc_id == 0) ?
282 			dev->dqm->hiq_sdma_mqd.gtt_mem : NULL;
283 	mqd_mem_obj->gpu_addr = dev->dqm->hiq_sdma_mqd.gpu_addr + offset;
284 	mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t)
285 				dev->dqm->hiq_sdma_mqd.cpu_ptr + offset);
286 }
287 
288 uint64_t kfd_mqd_stride(struct mqd_manager *mm,
289 			struct queue_properties *q)
290 {
291 	return mm->mqd_size;
292 }
293