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_device_queue_manager.h"
26 #include "gca/gfx_8_0_enum.h"
27 #include "gca/gfx_8_0_sh_mask.h"
28 #include "oss/oss_3_0_sh_mask.h"
29 
30 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
31 				   struct qcm_process_device *qpd,
32 				   enum cache_policy default_policy,
33 				   enum cache_policy alternate_policy,
34 				   void __user *alternate_aperture_base,
35 				   uint64_t alternate_aperture_size);
36 static bool set_cache_memory_policy_vi_tonga(struct device_queue_manager *dqm,
37 			struct qcm_process_device *qpd,
38 			enum cache_policy default_policy,
39 			enum cache_policy alternate_policy,
40 			void __user *alternate_aperture_base,
41 			uint64_t alternate_aperture_size);
42 static int update_qpd_vi(struct device_queue_manager *dqm,
43 					struct qcm_process_device *qpd);
44 static int update_qpd_vi_tonga(struct device_queue_manager *dqm,
45 			struct qcm_process_device *qpd);
46 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
47 				struct qcm_process_device *qpd);
48 static void init_sdma_vm_tonga(struct device_queue_manager *dqm,
49 			struct queue *q,
50 			struct qcm_process_device *qpd);
51 
52 void device_queue_manager_init_vi(
53 		struct device_queue_manager_asic_ops *asic_ops)
54 {
55 	asic_ops->set_cache_memory_policy = set_cache_memory_policy_vi;
56 	asic_ops->update_qpd = update_qpd_vi;
57 	asic_ops->init_sdma_vm = init_sdma_vm;
58 	asic_ops->mqd_manager_init = mqd_manager_init_vi;
59 }
60 
61 void device_queue_manager_init_vi_tonga(
62 		struct device_queue_manager_asic_ops *asic_ops)
63 {
64 	asic_ops->set_cache_memory_policy = set_cache_memory_policy_vi_tonga;
65 	asic_ops->update_qpd = update_qpd_vi_tonga;
66 	asic_ops->init_sdma_vm = init_sdma_vm_tonga;
67 	asic_ops->mqd_manager_init = mqd_manager_init_vi_tonga;
68 }
69 
70 static uint32_t compute_sh_mem_bases_64bit(unsigned int top_address_nybble)
71 {
72 	/* In 64-bit mode, we can only control the top 3 bits of the LDS,
73 	 * scratch and GPUVM apertures.
74 	 * The hardware fills in the remaining 59 bits according to the
75 	 * following pattern:
76 	 * LDS:		X0000000'00000000 - X0000001'00000000 (4GB)
77 	 * Scratch:	X0000001'00000000 - X0000002'00000000 (4GB)
78 	 * GPUVM:	Y0010000'00000000 - Y0020000'00000000 (1TB)
79 	 *
80 	 * (where X/Y is the configurable nybble with the low-bit 0)
81 	 *
82 	 * LDS and scratch will have the same top nybble programmed in the
83 	 * top 3 bits of SH_MEM_BASES.PRIVATE_BASE.
84 	 * GPUVM can have a different top nybble programmed in the
85 	 * top 3 bits of SH_MEM_BASES.SHARED_BASE.
86 	 * We don't bother to support different top nybbles
87 	 * for LDS/Scratch and GPUVM.
88 	 */
89 
90 	WARN_ON((top_address_nybble & 1) || top_address_nybble > 0xE ||
91 		top_address_nybble == 0);
92 
93 	return top_address_nybble << 12 |
94 			(top_address_nybble << 12) <<
95 			SH_MEM_BASES__SHARED_BASE__SHIFT;
96 }
97 
98 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
99 				   struct qcm_process_device *qpd,
100 				   enum cache_policy default_policy,
101 				   enum cache_policy alternate_policy,
102 				   void __user *alternate_aperture_base,
103 				   uint64_t alternate_aperture_size)
104 {
105 	uint32_t default_mtype;
106 	uint32_t ape1_mtype;
107 
108 	default_mtype = (default_policy == cache_policy_coherent) ?
109 			MTYPE_CC :
110 			MTYPE_NC;
111 
112 	ape1_mtype = (alternate_policy == cache_policy_coherent) ?
113 			MTYPE_CC :
114 			MTYPE_NC;
115 
116 	qpd->sh_mem_config = (qpd->sh_mem_config &
117 			SH_MEM_CONFIG__ADDRESS_MODE_MASK) |
118 		SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
119 				SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
120 		default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
121 		ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
122 		SH_MEM_CONFIG__PRIVATE_ATC_MASK;
123 
124 	return true;
125 }
126 
127 static bool set_cache_memory_policy_vi_tonga(struct device_queue_manager *dqm,
128 		struct qcm_process_device *qpd,
129 		enum cache_policy default_policy,
130 		enum cache_policy alternate_policy,
131 		void __user *alternate_aperture_base,
132 		uint64_t alternate_aperture_size)
133 {
134 	uint32_t default_mtype;
135 	uint32_t ape1_mtype;
136 
137 	default_mtype = (default_policy == cache_policy_coherent) ?
138 			MTYPE_UC :
139 			MTYPE_NC;
140 
141 	ape1_mtype = (alternate_policy == cache_policy_coherent) ?
142 			MTYPE_UC :
143 			MTYPE_NC;
144 
145 	qpd->sh_mem_config =
146 			SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
147 				   SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
148 			default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
149 			ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT;
150 
151 	return true;
152 }
153 
154 static int update_qpd_vi(struct device_queue_manager *dqm,
155 					struct qcm_process_device *qpd)
156 {
157 	struct kfd_process_device *pdd;
158 	unsigned int temp;
159 
160 	pdd = qpd_to_pdd(qpd);
161 
162 	/* check if sh_mem_config register already configured */
163 	if (qpd->sh_mem_config == 0) {
164 		qpd->sh_mem_config =
165 			SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
166 				SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
167 			MTYPE_CC << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
168 			MTYPE_CC << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
169 			SH_MEM_CONFIG__PRIVATE_ATC_MASK;
170 
171 		qpd->sh_mem_ape1_limit = 0;
172 		qpd->sh_mem_ape1_base = 0;
173 	}
174 
175 	if (qpd->pqm->process->is_32bit_user_mode) {
176 		temp = get_sh_mem_bases_32(pdd);
177 		qpd->sh_mem_bases = temp << SH_MEM_BASES__SHARED_BASE__SHIFT;
178 		qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA32 <<
179 					SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
180 	} else {
181 		temp = get_sh_mem_bases_nybble_64(pdd);
182 		qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
183 		qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA64 <<
184 			SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
185 		qpd->sh_mem_config |= 1  <<
186 			SH_MEM_CONFIG__PRIVATE_ATC__SHIFT;
187 	}
188 
189 	pr_debug("is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
190 		qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
191 
192 	return 0;
193 }
194 
195 static int update_qpd_vi_tonga(struct device_queue_manager *dqm,
196 			struct qcm_process_device *qpd)
197 {
198 	struct kfd_process_device *pdd;
199 	unsigned int temp;
200 
201 	pdd = qpd_to_pdd(qpd);
202 
203 	/* check if sh_mem_config register already configured */
204 	if (qpd->sh_mem_config == 0) {
205 		qpd->sh_mem_config =
206 				SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
207 					SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
208 				MTYPE_UC <<
209 					SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
210 				MTYPE_UC <<
211 					SH_MEM_CONFIG__APE1_MTYPE__SHIFT;
212 
213 		qpd->sh_mem_ape1_limit = 0;
214 		qpd->sh_mem_ape1_base = 0;
215 	}
216 
217 	/* On dGPU we're always in GPUVM64 addressing mode with 64-bit
218 	 * aperture addresses.
219 	 */
220 	temp = get_sh_mem_bases_nybble_64(pdd);
221 	qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
222 
223 	pr_debug("sh_mem_bases nybble: 0x%X and register 0x%X\n",
224 		temp, qpd->sh_mem_bases);
225 
226 	return 0;
227 }
228 
229 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
230 				struct qcm_process_device *qpd)
231 {
232 	uint32_t value = (1 << SDMA0_RLC0_VIRTUAL_ADDR__ATC__SHIFT);
233 
234 	if (q->process->is_32bit_user_mode)
235 		value |= (1 << SDMA0_RLC0_VIRTUAL_ADDR__PTR32__SHIFT) |
236 				get_sh_mem_bases_32(qpd_to_pdd(qpd));
237 	else
238 		value |= ((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
239 				SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
240 				SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
241 
242 	q->properties.sdma_vm_addr = value;
243 }
244 
245 static void init_sdma_vm_tonga(struct device_queue_manager *dqm,
246 			struct queue *q,
247 			struct qcm_process_device *qpd)
248 {
249 	/* On dGPU we're always in GPUVM64 addressing mode with 64-bit
250 	 * aperture addresses.
251 	 */
252 	q->properties.sdma_vm_addr =
253 		((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
254 		 SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
255 		SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
256 }
257