xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/amd/amdkfd/kfd_device_queue_manager_vi.c (revision 41ec02673d281bbb3d38e6c78504ce6e30c228c1)
1 /*	$NetBSD: kfd_device_queue_manager_vi.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2014 Advanced Micro Devices, Inc.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  */
25 
26 #include <sys/cdefs.h>
27 __KERNEL_RCSID(0, "$NetBSD: kfd_device_queue_manager_vi.c,v 1.3 2021/12/18 23:44:59 riastradh Exp $");
28 
29 #include "kfd_device_queue_manager.h"
30 #include "gca/gfx_8_0_enum.h"
31 #include "gca/gfx_8_0_sh_mask.h"
32 #include "oss/oss_3_0_sh_mask.h"
33 
34 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
35 				   struct qcm_process_device *qpd,
36 				   enum cache_policy default_policy,
37 				   enum cache_policy alternate_policy,
38 				   void __user *alternate_aperture_base,
39 				   uint64_t alternate_aperture_size);
40 static bool set_cache_memory_policy_vi_tonga(struct device_queue_manager *dqm,
41 			struct qcm_process_device *qpd,
42 			enum cache_policy default_policy,
43 			enum cache_policy alternate_policy,
44 			void __user *alternate_aperture_base,
45 			uint64_t alternate_aperture_size);
46 static int update_qpd_vi(struct device_queue_manager *dqm,
47 					struct qcm_process_device *qpd);
48 static int update_qpd_vi_tonga(struct device_queue_manager *dqm,
49 			struct qcm_process_device *qpd);
50 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
51 				struct qcm_process_device *qpd);
52 static void init_sdma_vm_tonga(struct device_queue_manager *dqm,
53 			struct queue *q,
54 			struct qcm_process_device *qpd);
55 
device_queue_manager_init_vi(struct device_queue_manager_asic_ops * asic_ops)56 void device_queue_manager_init_vi(
57 		struct device_queue_manager_asic_ops *asic_ops)
58 {
59 	asic_ops->set_cache_memory_policy = set_cache_memory_policy_vi;
60 	asic_ops->update_qpd = update_qpd_vi;
61 	asic_ops->init_sdma_vm = init_sdma_vm;
62 	asic_ops->mqd_manager_init = mqd_manager_init_vi;
63 }
64 
device_queue_manager_init_vi_tonga(struct device_queue_manager_asic_ops * asic_ops)65 void device_queue_manager_init_vi_tonga(
66 		struct device_queue_manager_asic_ops *asic_ops)
67 {
68 	asic_ops->set_cache_memory_policy = set_cache_memory_policy_vi_tonga;
69 	asic_ops->update_qpd = update_qpd_vi_tonga;
70 	asic_ops->init_sdma_vm = init_sdma_vm_tonga;
71 	asic_ops->mqd_manager_init = mqd_manager_init_vi_tonga;
72 }
73 
compute_sh_mem_bases_64bit(unsigned int top_address_nybble)74 static uint32_t compute_sh_mem_bases_64bit(unsigned int top_address_nybble)
75 {
76 	/* In 64-bit mode, we can only control the top 3 bits of the LDS,
77 	 * scratch and GPUVM apertures.
78 	 * The hardware fills in the remaining 59 bits according to the
79 	 * following pattern:
80 	 * LDS:		X0000000'00000000 - X0000001'00000000 (4GB)
81 	 * Scratch:	X0000001'00000000 - X0000002'00000000 (4GB)
82 	 * GPUVM:	Y0010000'00000000 - Y0020000'00000000 (1TB)
83 	 *
84 	 * (where X/Y is the configurable nybble with the low-bit 0)
85 	 *
86 	 * LDS and scratch will have the same top nybble programmed in the
87 	 * top 3 bits of SH_MEM_BASES.PRIVATE_BASE.
88 	 * GPUVM can have a different top nybble programmed in the
89 	 * top 3 bits of SH_MEM_BASES.SHARED_BASE.
90 	 * We don't bother to support different top nybbles
91 	 * for LDS/Scratch and GPUVM.
92 	 */
93 
94 	WARN_ON((top_address_nybble & 1) || top_address_nybble > 0xE ||
95 		top_address_nybble == 0);
96 
97 	return top_address_nybble << 12 |
98 			(top_address_nybble << 12) <<
99 			SH_MEM_BASES__SHARED_BASE__SHIFT;
100 }
101 
set_cache_memory_policy_vi(struct device_queue_manager * dqm,struct qcm_process_device * qpd,enum cache_policy default_policy,enum cache_policy alternate_policy,void __user * alternate_aperture_base,uint64_t alternate_aperture_size)102 static bool set_cache_memory_policy_vi(struct device_queue_manager *dqm,
103 				   struct qcm_process_device *qpd,
104 				   enum cache_policy default_policy,
105 				   enum cache_policy alternate_policy,
106 				   void __user *alternate_aperture_base,
107 				   uint64_t alternate_aperture_size)
108 {
109 	uint32_t default_mtype;
110 	uint32_t ape1_mtype;
111 
112 	default_mtype = (default_policy == cache_policy_coherent) ?
113 			MTYPE_CC :
114 			MTYPE_NC;
115 
116 	ape1_mtype = (alternate_policy == cache_policy_coherent) ?
117 			MTYPE_CC :
118 			MTYPE_NC;
119 
120 	qpd->sh_mem_config = (qpd->sh_mem_config &
121 			SH_MEM_CONFIG__ADDRESS_MODE_MASK) |
122 		SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
123 				SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
124 		default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
125 		ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
126 		SH_MEM_CONFIG__PRIVATE_ATC_MASK;
127 
128 	return true;
129 }
130 
set_cache_memory_policy_vi_tonga(struct device_queue_manager * dqm,struct qcm_process_device * qpd,enum cache_policy default_policy,enum cache_policy alternate_policy,void __user * alternate_aperture_base,uint64_t alternate_aperture_size)131 static bool set_cache_memory_policy_vi_tonga(struct device_queue_manager *dqm,
132 		struct qcm_process_device *qpd,
133 		enum cache_policy default_policy,
134 		enum cache_policy alternate_policy,
135 		void __user *alternate_aperture_base,
136 		uint64_t alternate_aperture_size)
137 {
138 	uint32_t default_mtype;
139 	uint32_t ape1_mtype;
140 
141 	default_mtype = (default_policy == cache_policy_coherent) ?
142 			MTYPE_UC :
143 			MTYPE_NC;
144 
145 	ape1_mtype = (alternate_policy == cache_policy_coherent) ?
146 			MTYPE_UC :
147 			MTYPE_NC;
148 
149 	qpd->sh_mem_config =
150 			SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
151 				   SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
152 			default_mtype << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
153 			ape1_mtype << SH_MEM_CONFIG__APE1_MTYPE__SHIFT;
154 
155 	return true;
156 }
157 
update_qpd_vi(struct device_queue_manager * dqm,struct qcm_process_device * qpd)158 static int update_qpd_vi(struct device_queue_manager *dqm,
159 					struct qcm_process_device *qpd)
160 {
161 	struct kfd_process_device *pdd;
162 	unsigned int temp;
163 
164 	pdd = qpd_to_pdd(qpd);
165 
166 	/* check if sh_mem_config register already configured */
167 	if (qpd->sh_mem_config == 0) {
168 		qpd->sh_mem_config =
169 			SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
170 				SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
171 			MTYPE_CC << SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
172 			MTYPE_CC << SH_MEM_CONFIG__APE1_MTYPE__SHIFT |
173 			SH_MEM_CONFIG__PRIVATE_ATC_MASK;
174 
175 		qpd->sh_mem_ape1_limit = 0;
176 		qpd->sh_mem_ape1_base = 0;
177 	}
178 
179 	if (qpd->pqm->process->is_32bit_user_mode) {
180 		temp = get_sh_mem_bases_32(pdd);
181 		qpd->sh_mem_bases = temp << SH_MEM_BASES__SHARED_BASE__SHIFT;
182 		qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA32 <<
183 					SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
184 	} else {
185 		temp = get_sh_mem_bases_nybble_64(pdd);
186 		qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
187 		qpd->sh_mem_config |= SH_MEM_ADDRESS_MODE_HSA64 <<
188 			SH_MEM_CONFIG__ADDRESS_MODE__SHIFT;
189 		qpd->sh_mem_config |= 1  <<
190 			SH_MEM_CONFIG__PRIVATE_ATC__SHIFT;
191 	}
192 
193 	pr_debug("is32bit process: %d sh_mem_bases nybble: 0x%X and register 0x%X\n",
194 		qpd->pqm->process->is_32bit_user_mode, temp, qpd->sh_mem_bases);
195 
196 	return 0;
197 }
198 
update_qpd_vi_tonga(struct device_queue_manager * dqm,struct qcm_process_device * qpd)199 static int update_qpd_vi_tonga(struct device_queue_manager *dqm,
200 			struct qcm_process_device *qpd)
201 {
202 	struct kfd_process_device *pdd;
203 	unsigned int temp;
204 
205 	pdd = qpd_to_pdd(qpd);
206 
207 	/* check if sh_mem_config register already configured */
208 	if (qpd->sh_mem_config == 0) {
209 		qpd->sh_mem_config =
210 				SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
211 					SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT |
212 				MTYPE_UC <<
213 					SH_MEM_CONFIG__DEFAULT_MTYPE__SHIFT |
214 				MTYPE_UC <<
215 					SH_MEM_CONFIG__APE1_MTYPE__SHIFT;
216 
217 		qpd->sh_mem_ape1_limit = 0;
218 		qpd->sh_mem_ape1_base = 0;
219 	}
220 
221 	/* On dGPU we're always in GPUVM64 addressing mode with 64-bit
222 	 * aperture addresses.
223 	 */
224 	temp = get_sh_mem_bases_nybble_64(pdd);
225 	qpd->sh_mem_bases = compute_sh_mem_bases_64bit(temp);
226 
227 	pr_debug("sh_mem_bases nybble: 0x%X and register 0x%X\n",
228 		temp, qpd->sh_mem_bases);
229 
230 	return 0;
231 }
232 
init_sdma_vm(struct device_queue_manager * dqm,struct queue * q,struct qcm_process_device * qpd)233 static void init_sdma_vm(struct device_queue_manager *dqm, struct queue *q,
234 				struct qcm_process_device *qpd)
235 {
236 	uint32_t value = (1 << SDMA0_RLC0_VIRTUAL_ADDR__ATC__SHIFT);
237 
238 	if (q->process->is_32bit_user_mode)
239 		value |= (1 << SDMA0_RLC0_VIRTUAL_ADDR__PTR32__SHIFT) |
240 				get_sh_mem_bases_32(qpd_to_pdd(qpd));
241 	else
242 		value |= ((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
243 				SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
244 				SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
245 
246 	q->properties.sdma_vm_addr = value;
247 }
248 
init_sdma_vm_tonga(struct device_queue_manager * dqm,struct queue * q,struct qcm_process_device * qpd)249 static void init_sdma_vm_tonga(struct device_queue_manager *dqm,
250 			struct queue *q,
251 			struct qcm_process_device *qpd)
252 {
253 	/* On dGPU we're always in GPUVM64 addressing mode with 64-bit
254 	 * aperture addresses.
255 	 */
256 	q->properties.sdma_vm_addr =
257 		((get_sh_mem_bases_nybble_64(qpd_to_pdd(qpd))) <<
258 		 SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE__SHIFT) &
259 		SDMA0_RLC0_VIRTUAL_ADDR__SHARED_BASE_MASK;
260 }
261