1 /*
2 * Copyright 2023 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23 #include <linux/delay.h>
24 #include <linux/kernel.h>
25 #include <linux/firmware.h>
26 #include <linux/module.h>
27 #include <linux/pci.h>
28 #include "amdgpu.h"
29 #include "amdgpu_gfx.h"
30 #include "amdgpu_psp.h"
31 #include "amdgpu_smu.h"
32 #include "amdgpu_atomfirmware.h"
33 #include "imu_v12_0.h"
34 #include "soc24.h"
35 #include "nvd.h"
36
37 #include "gc/gc_12_0_0_offset.h"
38 #include "gc/gc_12_0_0_sh_mask.h"
39 #include "soc24_enum.h"
40 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h"
41
42 #include "soc15.h"
43 #include "soc15d.h"
44 #include "clearstate_gfx12.h"
45 #include "v12_structs.h"
46 #include "gfx_v12_0.h"
47 #include "nbif_v6_3_1.h"
48 #include "mes_v12_0.h"
49
50 #define GFX12_NUM_GFX_RINGS 1
51 #define GFX12_MEC_HPD_SIZE 2048
52
53 #define RLCG_UCODE_LOADING_START_ADDRESS 0x00002000L
54
55 #define regCP_GFX_MQD_CONTROL_DEFAULT 0x00000100
56 #define regCP_GFX_HQD_VMID_DEFAULT 0x00000000
57 #define regCP_GFX_HQD_QUEUE_PRIORITY_DEFAULT 0x00000000
58 #define regCP_GFX_HQD_QUANTUM_DEFAULT 0x00000a01
59 #define regCP_GFX_HQD_CNTL_DEFAULT 0x00f00000
60 #define regCP_RB_DOORBELL_CONTROL_DEFAULT 0x00000000
61 #define regCP_GFX_HQD_RPTR_DEFAULT 0x00000000
62
63 #define regCP_HQD_EOP_CONTROL_DEFAULT 0x00000006
64 #define regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT 0x00000000
65 #define regCP_MQD_CONTROL_DEFAULT 0x00000100
66 #define regCP_HQD_PQ_CONTROL_DEFAULT 0x00308509
67 #define regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT 0x00000000
68 #define regCP_HQD_PQ_RPTR_DEFAULT 0x00000000
69 #define regCP_HQD_PERSISTENT_STATE_DEFAULT 0x0be05501
70 #define regCP_HQD_IB_CONTROL_DEFAULT 0x00300000
71
72
73 MODULE_FIRMWARE("amdgpu/gc_12_0_0_pfp.bin");
74 MODULE_FIRMWARE("amdgpu/gc_12_0_0_me.bin");
75 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mec.bin");
76 MODULE_FIRMWARE("amdgpu/gc_12_0_0_rlc.bin");
77 MODULE_FIRMWARE("amdgpu/gc_12_0_0_toc.bin");
78 MODULE_FIRMWARE("amdgpu/gc_12_0_1_pfp.bin");
79 MODULE_FIRMWARE("amdgpu/gc_12_0_1_me.bin");
80 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mec.bin");
81 MODULE_FIRMWARE("amdgpu/gc_12_0_1_rlc.bin");
82 MODULE_FIRMWARE("amdgpu/gc_12_0_1_toc.bin");
83
84 static const struct amdgpu_hwip_reg_entry gc_reg_list_12_0[] = {
85 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS),
86 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2),
87 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS3),
88 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT1),
89 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT2),
90 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT3),
91 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STALLED_STAT1),
92 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STALLED_STAT1),
93 SOC15_REG_ENTRY_STR(GC, 0, regCP_BUSY_STAT),
94 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT),
95 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT),
96 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT2),
97 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT2),
98 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STATUS),
99 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_ERROR),
100 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HPD_STATUS0),
101 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_BASE),
102 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR),
103 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR),
104 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_BASE),
105 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_RPTR),
106 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_WPTR),
107 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ),
108 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_CMD_BUFSZ),
109 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO),
110 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI),
111 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ),
112 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_LO),
113 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_HI),
114 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BUFSZ),
115 SOC15_REG_ENTRY_STR(GC, 0, regCPF_UTCL1_STATUS),
116 SOC15_REG_ENTRY_STR(GC, 0, regCPC_UTCL1_STATUS),
117 SOC15_REG_ENTRY_STR(GC, 0, regCPG_UTCL1_STATUS),
118 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS),
119 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS_2),
120 SOC15_REG_ENTRY_STR(GC, 0, regPA_CL_CNTL_STATUS),
121 SOC15_REG_ENTRY_STR(GC, 0, regRMI_UTCL1_STATUS),
122 SOC15_REG_ENTRY_STR(GC, 0, regSQC_CACHES),
123 SOC15_REG_ENTRY_STR(GC, 0, regSQG_STATUS),
124 SOC15_REG_ENTRY_STR(GC, 0, regWD_UTCL1_STATUS),
125 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_CNTL),
126 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_LO32),
127 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_HI32),
128 SOC15_REG_ENTRY_STR(GC, 0, regCP_DEBUG),
129 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_CNTL),
130 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_CNTL),
131 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_INSTR_PNTR),
132 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_INSTR_PNTR),
133 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_INSTR_PNTR),
134 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STATUS),
135 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR0),
136 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR1),
137 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_RS64_INSTR_PNTR),
138
139 /* cp header registers */
140 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP),
141 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP),
142 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP),
143 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP),
144 /* SE status registers */
145 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE0),
146 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE1),
147 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE2),
148 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE3)
149 };
150
151 static const struct amdgpu_hwip_reg_entry gc_cp_reg_list_12[] = {
152 /* compute registers */
153 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_VMID),
154 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PERSISTENT_STATE),
155 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PIPE_PRIORITY),
156 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUEUE_PRIORITY),
157 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUANTUM),
158 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE),
159 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE_HI),
160 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_RPTR),
161 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR),
162 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI),
163 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL),
164 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_CONTROL),
165 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR),
166 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR_HI),
167 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_RPTR),
168 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_CONTROL),
169 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_REQUEST),
170 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR),
171 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI),
172 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_CONTROL),
173 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_RPTR),
174 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR),
175 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_EVENTS),
176 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_LO),
177 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_HI),
178 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_CONTROL),
179 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_OFFSET),
180 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_SIZE),
181 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_WG_STATE_OFFSET),
182 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_SIZE),
183 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_GDS_RESOURCE_STATE),
184 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_ERROR),
185 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR_MEM),
186 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_LO),
187 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_HI),
188 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_OFFSET),
189 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_DW_CNT),
190 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_WG_STATE_OFFSET),
191 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_STATUS)
192 };
193
194 static const struct amdgpu_hwip_reg_entry gc_gfx_queue_reg_list_12[] = {
195 /* gfx queue registers */
196 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_ACTIVE),
197 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_VMID),
198 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY),
199 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUANTUM),
200 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE),
201 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE_HI),
202 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_OFFSET),
203 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CNTL),
204 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CSMD_RPTR),
205 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR),
206 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR_HI),
207 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_DEQUEUE_REQUEST),
208 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_MAPPED),
209 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUE_MGR_CONTROL),
210 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_CONTROL0),
211 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_STATUS0),
212 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR),
213 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR_HI),
214 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO),
215 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI),
216 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR),
217 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO),
218 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI),
219 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ),
220 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ)
221 };
222
223 static const struct soc15_reg_golden golden_settings_gc_12_0_rev0[] = {
224 SOC15_REG_GOLDEN_VALUE(GC, 0, regDB_MEM_CONFIG, 0x0000000f, 0x0000000f),
225 SOC15_REG_GOLDEN_VALUE(GC, 0, regCB_HW_CONTROL_1, 0x03000000, 0x03000000),
226 SOC15_REG_GOLDEN_VALUE(GC, 0, regGL2C_CTRL5, 0x00000070, 0x00000020)
227 };
228
229 static const struct soc15_reg_golden golden_settings_gc_12_0[] = {
230 SOC15_REG_GOLDEN_VALUE(GC, 0, regDB_MEM_CONFIG, 0x00008000, 0x00008000),
231 };
232
233 #define DEFAULT_SH_MEM_CONFIG \
234 ((SH_MEM_ADDRESS_MODE_64 << SH_MEM_CONFIG__ADDRESS_MODE__SHIFT) | \
235 (SH_MEM_ALIGNMENT_MODE_UNALIGNED << SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT) | \
236 (3 << SH_MEM_CONFIG__INITIAL_INST_PREFETCH__SHIFT))
237
238 static void gfx_v12_0_disable_gpa_mode(struct amdgpu_device *adev);
239 static void gfx_v12_0_set_ring_funcs(struct amdgpu_device *adev);
240 static void gfx_v12_0_set_irq_funcs(struct amdgpu_device *adev);
241 static void gfx_v12_0_set_rlc_funcs(struct amdgpu_device *adev);
242 static void gfx_v12_0_set_mqd_funcs(struct amdgpu_device *adev);
243 static void gfx_v12_0_set_imu_funcs(struct amdgpu_device *adev);
244 static int gfx_v12_0_get_cu_info(struct amdgpu_device *adev,
245 struct amdgpu_cu_info *cu_info);
246 static uint64_t gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device *adev);
247 static void gfx_v12_0_select_se_sh(struct amdgpu_device *adev, u32 se_num,
248 u32 sh_num, u32 instance, int xcc_id);
249 static u32 gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev);
250
251 static void gfx_v12_0_ring_emit_frame_cntl(struct amdgpu_ring *ring, bool start, bool secure);
252 static void gfx_v12_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg,
253 uint32_t val);
254 static int gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev);
255 static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring,
256 uint16_t pasid, uint32_t flush_type,
257 bool all_hub, uint8_t dst_sel);
258 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id);
259 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id);
260 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev,
261 bool enable);
262
gfx_v12_0_kiq_set_resources(struct amdgpu_ring * kiq_ring,uint64_t queue_mask)263 static void gfx_v12_0_kiq_set_resources(struct amdgpu_ring *kiq_ring,
264 uint64_t queue_mask)
265 {
266 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_RESOURCES, 6));
267 amdgpu_ring_write(kiq_ring, PACKET3_SET_RESOURCES_VMID_MASK(0) |
268 PACKET3_SET_RESOURCES_QUEUE_TYPE(0)); /* vmid_mask:0 queue_type:0 (KIQ) */
269 amdgpu_ring_write(kiq_ring, lower_32_bits(queue_mask)); /* queue mask lo */
270 amdgpu_ring_write(kiq_ring, upper_32_bits(queue_mask)); /* queue mask hi */
271 amdgpu_ring_write(kiq_ring, 0); /* gws mask lo */
272 amdgpu_ring_write(kiq_ring, 0); /* gws mask hi */
273 amdgpu_ring_write(kiq_ring, 0); /* oac mask */
274 amdgpu_ring_write(kiq_ring, 0);
275 }
276
gfx_v12_0_kiq_map_queues(struct amdgpu_ring * kiq_ring,struct amdgpu_ring * ring)277 static void gfx_v12_0_kiq_map_queues(struct amdgpu_ring *kiq_ring,
278 struct amdgpu_ring *ring)
279 {
280 uint64_t mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
281 uint64_t wptr_addr = ring->wptr_gpu_addr;
282 uint32_t me = 0, eng_sel = 0;
283
284 switch (ring->funcs->type) {
285 case AMDGPU_RING_TYPE_COMPUTE:
286 me = 1;
287 eng_sel = 0;
288 break;
289 case AMDGPU_RING_TYPE_GFX:
290 me = 0;
291 eng_sel = 4;
292 break;
293 case AMDGPU_RING_TYPE_MES:
294 me = 2;
295 eng_sel = 5;
296 break;
297 default:
298 WARN_ON(1);
299 }
300
301 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
302 /* Q_sel:0, vmid:0, vidmem: 1, engine:0, num_Q:1*/
303 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */
304 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */
305 PACKET3_MAP_QUEUES_VMID(0) | /* VMID */
306 PACKET3_MAP_QUEUES_QUEUE(ring->queue) |
307 PACKET3_MAP_QUEUES_PIPE(ring->pipe) |
308 PACKET3_MAP_QUEUES_ME((me)) |
309 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */
310 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */
311 PACKET3_MAP_QUEUES_ENGINE_SEL(eng_sel) |
312 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */
313 amdgpu_ring_write(kiq_ring, PACKET3_MAP_QUEUES_DOORBELL_OFFSET(ring->doorbell_index));
314 amdgpu_ring_write(kiq_ring, lower_32_bits(mqd_addr));
315 amdgpu_ring_write(kiq_ring, upper_32_bits(mqd_addr));
316 amdgpu_ring_write(kiq_ring, lower_32_bits(wptr_addr));
317 amdgpu_ring_write(kiq_ring, upper_32_bits(wptr_addr));
318 }
319
gfx_v12_0_kiq_unmap_queues(struct amdgpu_ring * kiq_ring,struct amdgpu_ring * ring,enum amdgpu_unmap_queues_action action,u64 gpu_addr,u64 seq)320 static void gfx_v12_0_kiq_unmap_queues(struct amdgpu_ring *kiq_ring,
321 struct amdgpu_ring *ring,
322 enum amdgpu_unmap_queues_action action,
323 u64 gpu_addr, u64 seq)
324 {
325 struct amdgpu_device *adev = kiq_ring->adev;
326 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0;
327
328 if (adev->enable_mes && !adev->gfx.kiq[0].ring.sched.ready) {
329 amdgpu_mes_unmap_legacy_queue(adev, ring, action, gpu_addr, seq);
330 return;
331 }
332
333 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_UNMAP_QUEUES, 4));
334 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */
335 PACKET3_UNMAP_QUEUES_ACTION(action) |
336 PACKET3_UNMAP_QUEUES_QUEUE_SEL(0) |
337 PACKET3_UNMAP_QUEUES_ENGINE_SEL(eng_sel) |
338 PACKET3_UNMAP_QUEUES_NUM_QUEUES(1));
339 amdgpu_ring_write(kiq_ring,
340 PACKET3_UNMAP_QUEUES_DOORBELL_OFFSET0(ring->doorbell_index));
341
342 if (action == PREEMPT_QUEUES_NO_UNMAP) {
343 amdgpu_ring_write(kiq_ring, lower_32_bits(gpu_addr));
344 amdgpu_ring_write(kiq_ring, upper_32_bits(gpu_addr));
345 amdgpu_ring_write(kiq_ring, seq);
346 } else {
347 amdgpu_ring_write(kiq_ring, 0);
348 amdgpu_ring_write(kiq_ring, 0);
349 amdgpu_ring_write(kiq_ring, 0);
350 }
351 }
352
gfx_v12_0_kiq_query_status(struct amdgpu_ring * kiq_ring,struct amdgpu_ring * ring,u64 addr,u64 seq)353 static void gfx_v12_0_kiq_query_status(struct amdgpu_ring *kiq_ring,
354 struct amdgpu_ring *ring,
355 u64 addr, u64 seq)
356 {
357 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0;
358
359 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_QUERY_STATUS, 5));
360 amdgpu_ring_write(kiq_ring,
361 PACKET3_QUERY_STATUS_CONTEXT_ID(0) |
362 PACKET3_QUERY_STATUS_INTERRUPT_SEL(0) |
363 PACKET3_QUERY_STATUS_COMMAND(2));
364 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */
365 PACKET3_QUERY_STATUS_DOORBELL_OFFSET(ring->doorbell_index) |
366 PACKET3_QUERY_STATUS_ENG_SEL(eng_sel));
367 amdgpu_ring_write(kiq_ring, lower_32_bits(addr));
368 amdgpu_ring_write(kiq_ring, upper_32_bits(addr));
369 amdgpu_ring_write(kiq_ring, lower_32_bits(seq));
370 amdgpu_ring_write(kiq_ring, upper_32_bits(seq));
371 }
372
gfx_v12_0_kiq_invalidate_tlbs(struct amdgpu_ring * kiq_ring,uint16_t pasid,uint32_t flush_type,bool all_hub)373 static void gfx_v12_0_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring,
374 uint16_t pasid,
375 uint32_t flush_type,
376 bool all_hub)
377 {
378 gfx_v12_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1);
379 }
380
381 static const struct kiq_pm4_funcs gfx_v12_0_kiq_pm4_funcs = {
382 .kiq_set_resources = gfx_v12_0_kiq_set_resources,
383 .kiq_map_queues = gfx_v12_0_kiq_map_queues,
384 .kiq_unmap_queues = gfx_v12_0_kiq_unmap_queues,
385 .kiq_query_status = gfx_v12_0_kiq_query_status,
386 .kiq_invalidate_tlbs = gfx_v12_0_kiq_invalidate_tlbs,
387 .set_resources_size = 8,
388 .map_queues_size = 7,
389 .unmap_queues_size = 6,
390 .query_status_size = 7,
391 .invalidate_tlbs_size = 2,
392 };
393
gfx_v12_0_set_kiq_pm4_funcs(struct amdgpu_device * adev)394 static void gfx_v12_0_set_kiq_pm4_funcs(struct amdgpu_device *adev)
395 {
396 adev->gfx.kiq[0].pmf = &gfx_v12_0_kiq_pm4_funcs;
397 }
398
gfx_v12_0_wait_reg_mem(struct amdgpu_ring * ring,int eng_sel,int mem_space,int opt,uint32_t addr0,uint32_t addr1,uint32_t ref,uint32_t mask,uint32_t inv)399 static void gfx_v12_0_wait_reg_mem(struct amdgpu_ring *ring, int eng_sel,
400 int mem_space, int opt, uint32_t addr0,
401 uint32_t addr1, uint32_t ref,
402 uint32_t mask, uint32_t inv)
403 {
404 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
405 amdgpu_ring_write(ring,
406 /* memory (1) or register (0) */
407 (WAIT_REG_MEM_MEM_SPACE(mem_space) |
408 WAIT_REG_MEM_OPERATION(opt) | /* wait */
409 WAIT_REG_MEM_FUNCTION(3) | /* equal */
410 WAIT_REG_MEM_ENGINE(eng_sel)));
411
412 if (mem_space)
413 BUG_ON(addr0 & 0x3); /* Dword align */
414 amdgpu_ring_write(ring, addr0);
415 amdgpu_ring_write(ring, addr1);
416 amdgpu_ring_write(ring, ref);
417 amdgpu_ring_write(ring, mask);
418 amdgpu_ring_write(ring, inv); /* poll interval */
419 }
420
gfx_v12_0_ring_test_ring(struct amdgpu_ring * ring)421 static int gfx_v12_0_ring_test_ring(struct amdgpu_ring *ring)
422 {
423 struct amdgpu_device *adev = ring->adev;
424 uint32_t scratch = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0);
425 uint32_t tmp = 0;
426 unsigned i;
427 int r;
428
429 WREG32(scratch, 0xCAFEDEAD);
430 r = amdgpu_ring_alloc(ring, 5);
431 if (r) {
432 dev_err(adev->dev,
433 "amdgpu: cp failed to lock ring %d (%d).\n",
434 ring->idx, r);
435 return r;
436 }
437
438 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ) {
439 gfx_v12_0_ring_emit_wreg(ring, scratch, 0xDEADBEEF);
440 } else {
441 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
442 amdgpu_ring_write(ring, scratch -
443 PACKET3_SET_UCONFIG_REG_START);
444 amdgpu_ring_write(ring, 0xDEADBEEF);
445 }
446 amdgpu_ring_commit(ring);
447
448 for (i = 0; i < adev->usec_timeout; i++) {
449 tmp = RREG32(scratch);
450 if (tmp == 0xDEADBEEF)
451 break;
452 if (amdgpu_emu_mode == 1)
453 msleep(1);
454 else
455 udelay(1);
456 }
457
458 if (i >= adev->usec_timeout)
459 r = -ETIMEDOUT;
460 return r;
461 }
462
gfx_v12_0_ring_test_ib(struct amdgpu_ring * ring,long timeout)463 static int gfx_v12_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
464 {
465 struct amdgpu_device *adev = ring->adev;
466 struct amdgpu_ib ib;
467 struct dma_fence *f = NULL;
468 unsigned index;
469 uint64_t gpu_addr;
470 volatile uint32_t *cpu_ptr;
471 long r;
472
473 /* MES KIQ fw hasn't indirect buffer support for now */
474 if (adev->enable_mes_kiq &&
475 ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
476 return 0;
477
478 memset(&ib, 0, sizeof(ib));
479
480 if (ring->is_mes_queue) {
481 uint32_t padding, offset;
482
483 offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS);
484 padding = amdgpu_mes_ctx_get_offs(ring,
485 AMDGPU_MES_CTX_PADDING_OFFS);
486
487 ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
488 ib.ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
489
490 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, padding);
491 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, padding);
492 *cpu_ptr = cpu_to_le32(0xCAFEDEAD);
493 } else {
494 r = amdgpu_device_wb_get(adev, &index);
495 if (r)
496 return r;
497
498 gpu_addr = adev->wb.gpu_addr + (index * 4);
499 adev->wb.wb[index] = cpu_to_le32(0xCAFEDEAD);
500 cpu_ptr = &adev->wb.wb[index];
501
502 r = amdgpu_ib_get(adev, NULL, 16, AMDGPU_IB_POOL_DIRECT, &ib);
503 if (r) {
504 dev_err(adev->dev, "amdgpu: failed to get ib (%ld).\n", r);
505 goto err1;
506 }
507 }
508
509 ib.ptr[0] = PACKET3(PACKET3_WRITE_DATA, 3);
510 ib.ptr[1] = WRITE_DATA_DST_SEL(5) | WR_CONFIRM;
511 ib.ptr[2] = lower_32_bits(gpu_addr);
512 ib.ptr[3] = upper_32_bits(gpu_addr);
513 ib.ptr[4] = 0xDEADBEEF;
514 ib.length_dw = 5;
515
516 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
517 if (r)
518 goto err2;
519
520 r = dma_fence_wait_timeout(f, false, timeout);
521 if (r == 0) {
522 r = -ETIMEDOUT;
523 goto err2;
524 } else if (r < 0) {
525 goto err2;
526 }
527
528 if (le32_to_cpu(*cpu_ptr) == 0xDEADBEEF)
529 r = 0;
530 else
531 r = -EINVAL;
532 err2:
533 if (!ring->is_mes_queue)
534 amdgpu_ib_free(adev, &ib, NULL);
535 dma_fence_put(f);
536 err1:
537 if (!ring->is_mes_queue)
538 amdgpu_device_wb_free(adev, index);
539 return r;
540 }
541
gfx_v12_0_free_microcode(struct amdgpu_device * adev)542 static void gfx_v12_0_free_microcode(struct amdgpu_device *adev)
543 {
544 amdgpu_ucode_release(&adev->gfx.pfp_fw);
545 amdgpu_ucode_release(&adev->gfx.me_fw);
546 amdgpu_ucode_release(&adev->gfx.rlc_fw);
547 amdgpu_ucode_release(&adev->gfx.mec_fw);
548
549 kfree(adev->gfx.rlc.register_list_format);
550 }
551
gfx_v12_0_init_toc_microcode(struct amdgpu_device * adev,const char * ucode_prefix)552 static int gfx_v12_0_init_toc_microcode(struct amdgpu_device *adev, const char *ucode_prefix)
553 {
554 const struct psp_firmware_header_v1_0 *toc_hdr;
555 int err = 0;
556
557 err = amdgpu_ucode_request(adev, &adev->psp.toc_fw,
558 "amdgpu/%s_toc.bin", ucode_prefix);
559 if (err)
560 goto out;
561
562 toc_hdr = (const struct psp_firmware_header_v1_0 *)adev->psp.toc_fw->data;
563 adev->psp.toc.fw_version = le32_to_cpu(toc_hdr->header.ucode_version);
564 adev->psp.toc.feature_version = le32_to_cpu(toc_hdr->sos.fw_version);
565 adev->psp.toc.size_bytes = le32_to_cpu(toc_hdr->header.ucode_size_bytes);
566 adev->psp.toc.start_addr = (uint8_t *)toc_hdr +
567 le32_to_cpu(toc_hdr->header.ucode_array_offset_bytes);
568 return 0;
569 out:
570 amdgpu_ucode_release(&adev->psp.toc_fw);
571 return err;
572 }
573
gfx_v12_0_init_microcode(struct amdgpu_device * adev)574 static int gfx_v12_0_init_microcode(struct amdgpu_device *adev)
575 {
576 char ucode_prefix[15];
577 int err;
578 const struct rlc_firmware_header_v2_0 *rlc_hdr;
579 uint16_t version_major;
580 uint16_t version_minor;
581
582 DRM_DEBUG("\n");
583
584 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix));
585
586 err = amdgpu_ucode_request(adev, &adev->gfx.pfp_fw,
587 "amdgpu/%s_pfp.bin", ucode_prefix);
588 if (err)
589 goto out;
590 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP);
591 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK);
592
593 err = amdgpu_ucode_request(adev, &adev->gfx.me_fw,
594 "amdgpu/%s_me.bin", ucode_prefix);
595 if (err)
596 goto out;
597 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME);
598 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK);
599
600 if (!amdgpu_sriov_vf(adev)) {
601 err = amdgpu_ucode_request(adev, &adev->gfx.rlc_fw,
602 "amdgpu/%s_rlc.bin", ucode_prefix);
603 if (err)
604 goto out;
605 rlc_hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data;
606 version_major = le16_to_cpu(rlc_hdr->header.header_version_major);
607 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor);
608 err = amdgpu_gfx_rlc_init_microcode(adev, version_major, version_minor);
609 if (err)
610 goto out;
611 }
612
613 err = amdgpu_ucode_request(adev, &adev->gfx.mec_fw,
614 "amdgpu/%s_mec.bin", ucode_prefix);
615 if (err)
616 goto out;
617 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC);
618 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK);
619 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK);
620
621 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO)
622 err = gfx_v12_0_init_toc_microcode(adev, ucode_prefix);
623
624 /* only one MEC for gfx 12 */
625 adev->gfx.mec2_fw = NULL;
626
627 if (adev->gfx.imu.funcs) {
628 if (adev->gfx.imu.funcs->init_microcode) {
629 err = adev->gfx.imu.funcs->init_microcode(adev);
630 if (err)
631 dev_err(adev->dev, "Failed to load imu firmware!\n");
632 }
633 }
634
635 out:
636 if (err) {
637 amdgpu_ucode_release(&adev->gfx.pfp_fw);
638 amdgpu_ucode_release(&adev->gfx.me_fw);
639 amdgpu_ucode_release(&adev->gfx.rlc_fw);
640 amdgpu_ucode_release(&adev->gfx.mec_fw);
641 }
642
643 return err;
644 }
645
gfx_v12_0_get_csb_size(struct amdgpu_device * adev)646 static u32 gfx_v12_0_get_csb_size(struct amdgpu_device *adev)
647 {
648 u32 count = 0;
649 const struct cs_section_def *sect = NULL;
650 const struct cs_extent_def *ext = NULL;
651
652 count += 1;
653
654 for (sect = gfx12_cs_data; sect->section != NULL; ++sect) {
655 if (sect->id == SECT_CONTEXT) {
656 for (ext = sect->section; ext->extent != NULL; ++ext)
657 count += 2 + ext->reg_count;
658 } else
659 return 0;
660 }
661
662 return count;
663 }
664
gfx_v12_0_get_csb_buffer(struct amdgpu_device * adev,volatile u32 * buffer)665 static void gfx_v12_0_get_csb_buffer(struct amdgpu_device *adev,
666 volatile u32 *buffer)
667 {
668 u32 count = 0, clustercount = 0, i;
669 const struct cs_section_def *sect = NULL;
670 const struct cs_extent_def *ext = NULL;
671
672 if (adev->gfx.rlc.cs_data == NULL)
673 return;
674 if (buffer == NULL)
675 return;
676
677 count += 1;
678
679 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) {
680 if (sect->id == SECT_CONTEXT) {
681 for (ext = sect->section; ext->extent != NULL; ++ext) {
682 clustercount++;
683 buffer[count++] = ext->reg_count;
684 buffer[count++] = ext->reg_index;
685
686 for (i = 0; i < ext->reg_count; i++)
687 buffer[count++] = cpu_to_le32(ext->extent[i]);
688 }
689 } else
690 return;
691 }
692
693 buffer[0] = clustercount;
694 }
695
gfx_v12_0_rlc_fini(struct amdgpu_device * adev)696 static void gfx_v12_0_rlc_fini(struct amdgpu_device *adev)
697 {
698 /* clear state block */
699 amdgpu_bo_free_kernel(&adev->gfx.rlc.clear_state_obj,
700 &adev->gfx.rlc.clear_state_gpu_addr,
701 (void **)&adev->gfx.rlc.cs_ptr);
702
703 /* jump table block */
704 amdgpu_bo_free_kernel(&adev->gfx.rlc.cp_table_obj,
705 &adev->gfx.rlc.cp_table_gpu_addr,
706 (void **)&adev->gfx.rlc.cp_table_ptr);
707 }
708
gfx_v12_0_init_rlcg_reg_access_ctrl(struct amdgpu_device * adev)709 static void gfx_v12_0_init_rlcg_reg_access_ctrl(struct amdgpu_device *adev)
710 {
711 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl;
712
713 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[0];
714 reg_access_ctrl->scratch_reg0 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0);
715 reg_access_ctrl->scratch_reg1 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG1);
716 reg_access_ctrl->scratch_reg2 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG2);
717 reg_access_ctrl->scratch_reg3 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG3);
718 reg_access_ctrl->grbm_cntl = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_CNTL);
719 reg_access_ctrl->grbm_idx = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX);
720 reg_access_ctrl->spare_int = SOC15_REG_OFFSET(GC, 0, regRLC_SPARE_INT_0);
721 adev->gfx.rlc.rlcg_reg_access_supported = true;
722 }
723
gfx_v12_0_rlc_init(struct amdgpu_device * adev)724 static int gfx_v12_0_rlc_init(struct amdgpu_device *adev)
725 {
726 const struct cs_section_def *cs_data;
727 int r;
728
729 adev->gfx.rlc.cs_data = gfx12_cs_data;
730
731 cs_data = adev->gfx.rlc.cs_data;
732
733 if (cs_data) {
734 /* init clear state block */
735 r = amdgpu_gfx_rlc_init_csb(adev);
736 if (r)
737 return r;
738 }
739
740 /* init spm vmid with 0xf */
741 if (adev->gfx.rlc.funcs->update_spm_vmid)
742 adev->gfx.rlc.funcs->update_spm_vmid(adev, NULL, 0xf);
743
744 return 0;
745 }
746
gfx_v12_0_mec_fini(struct amdgpu_device * adev)747 static void gfx_v12_0_mec_fini(struct amdgpu_device *adev)
748 {
749 amdgpu_bo_free_kernel(&adev->gfx.mec.hpd_eop_obj, NULL, NULL);
750 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_obj, NULL, NULL);
751 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_data_obj, NULL, NULL);
752 }
753
gfx_v12_0_me_init(struct amdgpu_device * adev)754 static void gfx_v12_0_me_init(struct amdgpu_device *adev)
755 {
756 bitmap_zero(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES);
757
758 amdgpu_gfx_graphics_queue_acquire(adev);
759 }
760
gfx_v12_0_mec_init(struct amdgpu_device * adev)761 static int gfx_v12_0_mec_init(struct amdgpu_device *adev)
762 {
763 int r;
764 u32 *hpd;
765 size_t mec_hpd_size;
766
767 bitmap_zero(adev->gfx.mec_bitmap[0].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
768
769 /* take ownership of the relevant compute queues */
770 amdgpu_gfx_compute_queue_acquire(adev);
771 mec_hpd_size = adev->gfx.num_compute_rings * GFX12_MEC_HPD_SIZE;
772
773 if (mec_hpd_size) {
774 r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE,
775 AMDGPU_GEM_DOMAIN_GTT,
776 &adev->gfx.mec.hpd_eop_obj,
777 &adev->gfx.mec.hpd_eop_gpu_addr,
778 (void **)&hpd);
779 if (r) {
780 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
781 gfx_v12_0_mec_fini(adev);
782 return r;
783 }
784
785 memset(hpd, 0, mec_hpd_size);
786
787 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
788 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
789 }
790
791 return 0;
792 }
793
wave_read_ind(struct amdgpu_device * adev,uint32_t wave,uint32_t address)794 static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t wave, uint32_t address)
795 {
796 WREG32_SOC15(GC, 0, regSQ_IND_INDEX,
797 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
798 (address << SQ_IND_INDEX__INDEX__SHIFT));
799 return RREG32_SOC15(GC, 0, regSQ_IND_DATA);
800 }
801
wave_read_regs(struct amdgpu_device * adev,uint32_t wave,uint32_t thread,uint32_t regno,uint32_t num,uint32_t * out)802 static void wave_read_regs(struct amdgpu_device *adev, uint32_t wave,
803 uint32_t thread, uint32_t regno,
804 uint32_t num, uint32_t *out)
805 {
806 WREG32_SOC15(GC, 0, regSQ_IND_INDEX,
807 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
808 (regno << SQ_IND_INDEX__INDEX__SHIFT) |
809 (thread << SQ_IND_INDEX__WORKITEM_ID__SHIFT) |
810 (SQ_IND_INDEX__AUTO_INCR_MASK));
811 while (num--)
812 *(out++) = RREG32_SOC15(GC, 0, regSQ_IND_DATA);
813 }
814
gfx_v12_0_read_wave_data(struct amdgpu_device * adev,uint32_t xcc_id,uint32_t simd,uint32_t wave,uint32_t * dst,int * no_fields)815 static void gfx_v12_0_read_wave_data(struct amdgpu_device *adev,
816 uint32_t xcc_id,
817 uint32_t simd, uint32_t wave,
818 uint32_t *dst, int *no_fields)
819 {
820 /* in gfx12 the SIMD_ID is specified as part of the INSTANCE
821 * field when performing a select_se_sh so it should be
822 * zero here */
823 WARN_ON(simd != 0);
824
825 /* type 4 wave data */
826 dst[(*no_fields)++] = 4;
827 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATUS);
828 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_LO);
829 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_HI);
830 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_LO);
831 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_HI);
832 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID1);
833 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID2);
834 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_GPR_ALLOC);
835 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_LDS_ALLOC);
836 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS);
837 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS2);
838 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_DBG1);
839 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_M0);
840 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_MODE);
841 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATE_PRIV);
842 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXCP_FLAG_PRIV);
843 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXCP_FLAG_USER);
844 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_TRAP_CTRL);
845 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_ACTIVE);
846 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_VALID_AND_IDLE);
847 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_DVGPR_ALLOC_LO);
848 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_DVGPR_ALLOC_HI);
849 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_SCHED_MODE);
850 }
851
gfx_v12_0_read_wave_sgprs(struct amdgpu_device * adev,uint32_t xcc_id,uint32_t simd,uint32_t wave,uint32_t start,uint32_t size,uint32_t * dst)852 static void gfx_v12_0_read_wave_sgprs(struct amdgpu_device *adev,
853 uint32_t xcc_id, uint32_t simd,
854 uint32_t wave, uint32_t start,
855 uint32_t size, uint32_t *dst)
856 {
857 WARN_ON(simd != 0);
858
859 wave_read_regs(
860 adev, wave, 0, start + SQIND_WAVE_SGPRS_OFFSET, size,
861 dst);
862 }
863
gfx_v12_0_read_wave_vgprs(struct amdgpu_device * adev,uint32_t xcc_id,uint32_t simd,uint32_t wave,uint32_t thread,uint32_t start,uint32_t size,uint32_t * dst)864 static void gfx_v12_0_read_wave_vgprs(struct amdgpu_device *adev,
865 uint32_t xcc_id, uint32_t simd,
866 uint32_t wave, uint32_t thread,
867 uint32_t start, uint32_t size,
868 uint32_t *dst)
869 {
870 wave_read_regs(
871 adev, wave, thread,
872 start + SQIND_WAVE_VGPRS_OFFSET, size, dst);
873 }
874
gfx_v12_0_select_me_pipe_q(struct amdgpu_device * adev,u32 me,u32 pipe,u32 q,u32 vm,u32 xcc_id)875 static void gfx_v12_0_select_me_pipe_q(struct amdgpu_device *adev,
876 u32 me, u32 pipe, u32 q, u32 vm, u32 xcc_id)
877 {
878 soc24_grbm_select(adev, me, pipe, q, vm);
879 }
880
881 static const struct amdgpu_gfx_funcs gfx_v12_0_gfx_funcs = {
882 .get_gpu_clock_counter = &gfx_v12_0_get_gpu_clock_counter,
883 .select_se_sh = &gfx_v12_0_select_se_sh,
884 .read_wave_data = &gfx_v12_0_read_wave_data,
885 .read_wave_sgprs = &gfx_v12_0_read_wave_sgprs,
886 .read_wave_vgprs = &gfx_v12_0_read_wave_vgprs,
887 .select_me_pipe_q = &gfx_v12_0_select_me_pipe_q,
888 .update_perfmon_mgcg = &gfx_v12_0_update_perf_clk,
889 };
890
gfx_v12_0_gpu_early_init(struct amdgpu_device * adev)891 static int gfx_v12_0_gpu_early_init(struct amdgpu_device *adev)
892 {
893
894 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
895 case IP_VERSION(12, 0, 0):
896 case IP_VERSION(12, 0, 1):
897 adev->gfx.config.max_hw_contexts = 8;
898 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
899 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
900 adev->gfx.config.sc_hiz_tile_fifo_size = 0;
901 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0;
902 break;
903 default:
904 BUG();
905 break;
906 }
907
908 return 0;
909 }
910
gfx_v12_0_gfx_ring_init(struct amdgpu_device * adev,int ring_id,int me,int pipe,int queue)911 static int gfx_v12_0_gfx_ring_init(struct amdgpu_device *adev, int ring_id,
912 int me, int pipe, int queue)
913 {
914 int r;
915 struct amdgpu_ring *ring;
916 unsigned int irq_type;
917
918 ring = &adev->gfx.gfx_ring[ring_id];
919
920 ring->me = me;
921 ring->pipe = pipe;
922 ring->queue = queue;
923
924 ring->ring_obj = NULL;
925 ring->use_doorbell = true;
926
927 if (!ring_id)
928 ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1;
929 else
930 ring->doorbell_index = adev->doorbell_index.gfx_ring1 << 1;
931 ring->vm_hub = AMDGPU_GFXHUB(0);
932 sprintf(ring->name, "gfx_%d.%d.%d", ring->me, ring->pipe, ring->queue);
933
934 irq_type = AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP + ring->pipe;
935 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type,
936 AMDGPU_RING_PRIO_DEFAULT, NULL);
937 if (r)
938 return r;
939 return 0;
940 }
941
gfx_v12_0_compute_ring_init(struct amdgpu_device * adev,int ring_id,int mec,int pipe,int queue)942 static int gfx_v12_0_compute_ring_init(struct amdgpu_device *adev, int ring_id,
943 int mec, int pipe, int queue)
944 {
945 int r;
946 unsigned irq_type;
947 struct amdgpu_ring *ring;
948 unsigned int hw_prio;
949
950 ring = &adev->gfx.compute_ring[ring_id];
951
952 /* mec0 is me1 */
953 ring->me = mec + 1;
954 ring->pipe = pipe;
955 ring->queue = queue;
956
957 ring->ring_obj = NULL;
958 ring->use_doorbell = true;
959 ring->doorbell_index = (adev->doorbell_index.mec_ring0 + ring_id) << 1;
960 ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr
961 + (ring_id * GFX12_MEC_HPD_SIZE);
962 ring->vm_hub = AMDGPU_GFXHUB(0);
963 sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue);
964
965 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP
966 + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec)
967 + ring->pipe;
968 hw_prio = amdgpu_gfx_is_high_priority_compute_queue(adev, ring) ?
969 AMDGPU_GFX_PIPE_PRIO_HIGH : AMDGPU_GFX_PIPE_PRIO_NORMAL;
970 /* type-2 packets are deprecated on MEC, use type-3 instead */
971 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type,
972 hw_prio, NULL);
973 if (r)
974 return r;
975
976 return 0;
977 }
978
979 static struct {
980 SOC24_FIRMWARE_ID id;
981 unsigned int offset;
982 unsigned int size;
983 unsigned int size_x16;
984 } rlc_autoload_info[SOC24_FIRMWARE_ID_MAX];
985
986 #define RLC_TOC_OFFSET_DWUNIT 8
987 #define RLC_SIZE_MULTIPLE 1024
988 #define RLC_TOC_UMF_SIZE_inM 23ULL
989 #define RLC_TOC_FORMAT_API 165ULL
990
gfx_v12_0_parse_rlc_toc(struct amdgpu_device * adev,void * rlc_toc)991 static void gfx_v12_0_parse_rlc_toc(struct amdgpu_device *adev, void *rlc_toc)
992 {
993 RLC_TABLE_OF_CONTENT_V2 *ucode = rlc_toc;
994
995 while (ucode && (ucode->id > SOC24_FIRMWARE_ID_INVALID)) {
996 rlc_autoload_info[ucode->id].id = ucode->id;
997 rlc_autoload_info[ucode->id].offset =
998 ucode->offset * RLC_TOC_OFFSET_DWUNIT * 4;
999 rlc_autoload_info[ucode->id].size =
1000 ucode->size_x16 ? ucode->size * RLC_SIZE_MULTIPLE * 4 :
1001 ucode->size * 4;
1002 ucode++;
1003 }
1004 }
1005
gfx_v12_0_calc_toc_total_size(struct amdgpu_device * adev)1006 static uint32_t gfx_v12_0_calc_toc_total_size(struct amdgpu_device *adev)
1007 {
1008 uint32_t total_size = 0;
1009 SOC24_FIRMWARE_ID id;
1010
1011 gfx_v12_0_parse_rlc_toc(adev, adev->psp.toc.start_addr);
1012
1013 for (id = SOC24_FIRMWARE_ID_RLC_G_UCODE; id < SOC24_FIRMWARE_ID_MAX; id++)
1014 total_size += rlc_autoload_info[id].size;
1015
1016 /* In case the offset in rlc toc ucode is aligned */
1017 if (total_size < rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].offset)
1018 total_size = rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].offset +
1019 rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].size;
1020 if (total_size < (RLC_TOC_UMF_SIZE_inM << 20))
1021 total_size = RLC_TOC_UMF_SIZE_inM << 20;
1022
1023 return total_size;
1024 }
1025
gfx_v12_0_rlc_autoload_buffer_init(struct amdgpu_device * adev)1026 static int gfx_v12_0_rlc_autoload_buffer_init(struct amdgpu_device *adev)
1027 {
1028 int r;
1029 uint32_t total_size;
1030
1031 total_size = gfx_v12_0_calc_toc_total_size(adev);
1032
1033 r = amdgpu_bo_create_reserved(adev, total_size, 64 * 1024,
1034 AMDGPU_GEM_DOMAIN_VRAM,
1035 &adev->gfx.rlc.rlc_autoload_bo,
1036 &adev->gfx.rlc.rlc_autoload_gpu_addr,
1037 (void **)&adev->gfx.rlc.rlc_autoload_ptr);
1038
1039 if (r) {
1040 dev_err(adev->dev, "(%d) failed to create fw autoload bo\n", r);
1041 return r;
1042 }
1043
1044 return 0;
1045 }
1046
gfx_v12_0_rlc_backdoor_autoload_copy_ucode(struct amdgpu_device * adev,SOC24_FIRMWARE_ID id,const void * fw_data,uint32_t fw_size)1047 static void gfx_v12_0_rlc_backdoor_autoload_copy_ucode(struct amdgpu_device *adev,
1048 SOC24_FIRMWARE_ID id,
1049 const void *fw_data,
1050 uint32_t fw_size)
1051 {
1052 uint32_t toc_offset;
1053 uint32_t toc_fw_size;
1054 char *ptr = adev->gfx.rlc.rlc_autoload_ptr;
1055
1056 if (id <= SOC24_FIRMWARE_ID_INVALID || id >= SOC24_FIRMWARE_ID_MAX)
1057 return;
1058
1059 toc_offset = rlc_autoload_info[id].offset;
1060 toc_fw_size = rlc_autoload_info[id].size;
1061
1062 if (fw_size == 0)
1063 fw_size = toc_fw_size;
1064
1065 if (fw_size > toc_fw_size)
1066 fw_size = toc_fw_size;
1067
1068 memcpy(ptr + toc_offset, fw_data, fw_size);
1069
1070 if (fw_size < toc_fw_size)
1071 memset(ptr + toc_offset + fw_size, 0, toc_fw_size - fw_size);
1072 }
1073
1074 static void
gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(struct amdgpu_device * adev)1075 gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(struct amdgpu_device *adev)
1076 {
1077 void *data;
1078 uint32_t size;
1079 uint32_t *toc_ptr;
1080
1081 data = adev->psp.toc.start_addr;
1082 size = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_TOC].size;
1083
1084 toc_ptr = (uint32_t *)data + size / 4 - 2;
1085 *toc_ptr = (RLC_TOC_FORMAT_API << 24) | 0x1;
1086
1087 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_TOC,
1088 data, size);
1089 }
1090
1091 static void
gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(struct amdgpu_device * adev)1092 gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(struct amdgpu_device *adev)
1093 {
1094 const __le32 *fw_data;
1095 uint32_t fw_size;
1096 const struct gfx_firmware_header_v2_0 *cpv2_hdr;
1097 const struct rlc_firmware_header_v2_0 *rlc_hdr;
1098 const struct rlc_firmware_header_v2_1 *rlcv21_hdr;
1099 const struct rlc_firmware_header_v2_2 *rlcv22_hdr;
1100 uint16_t version_major, version_minor;
1101
1102 /* pfp ucode */
1103 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *)
1104 adev->gfx.pfp_fw->data;
1105 /* instruction */
1106 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data +
1107 le32_to_cpu(cpv2_hdr->ucode_offset_bytes));
1108 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes);
1109 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP,
1110 fw_data, fw_size);
1111 /* data */
1112 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data +
1113 le32_to_cpu(cpv2_hdr->data_offset_bytes));
1114 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes);
1115 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP_P0_STACK,
1116 fw_data, fw_size);
1117 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP_P1_STACK,
1118 fw_data, fw_size);
1119 /* me ucode */
1120 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *)
1121 adev->gfx.me_fw->data;
1122 /* instruction */
1123 fw_data = (const __le32 *)(adev->gfx.me_fw->data +
1124 le32_to_cpu(cpv2_hdr->ucode_offset_bytes));
1125 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes);
1126 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME,
1127 fw_data, fw_size);
1128 /* data */
1129 fw_data = (const __le32 *)(adev->gfx.me_fw->data +
1130 le32_to_cpu(cpv2_hdr->data_offset_bytes));
1131 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes);
1132 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME_P0_STACK,
1133 fw_data, fw_size);
1134 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME_P1_STACK,
1135 fw_data, fw_size);
1136 /* mec ucode */
1137 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *)
1138 adev->gfx.mec_fw->data;
1139 /* instruction */
1140 fw_data = (const __le32 *) (adev->gfx.mec_fw->data +
1141 le32_to_cpu(cpv2_hdr->ucode_offset_bytes));
1142 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes);
1143 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC,
1144 fw_data, fw_size);
1145 /* data */
1146 fw_data = (const __le32 *) (adev->gfx.mec_fw->data +
1147 le32_to_cpu(cpv2_hdr->data_offset_bytes));
1148 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes);
1149 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P0_STACK,
1150 fw_data, fw_size);
1151 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P1_STACK,
1152 fw_data, fw_size);
1153 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P2_STACK,
1154 fw_data, fw_size);
1155 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P3_STACK,
1156 fw_data, fw_size);
1157
1158 /* rlc ucode */
1159 rlc_hdr = (const struct rlc_firmware_header_v2_0 *)
1160 adev->gfx.rlc_fw->data;
1161 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1162 le32_to_cpu(rlc_hdr->header.ucode_array_offset_bytes));
1163 fw_size = le32_to_cpu(rlc_hdr->header.ucode_size_bytes);
1164 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_G_UCODE,
1165 fw_data, fw_size);
1166
1167 version_major = le16_to_cpu(rlc_hdr->header.header_version_major);
1168 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor);
1169 if (version_major == 2) {
1170 if (version_minor >= 1) {
1171 rlcv21_hdr = (const struct rlc_firmware_header_v2_1 *)adev->gfx.rlc_fw->data;
1172
1173 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1174 le32_to_cpu(rlcv21_hdr->save_restore_list_gpm_offset_bytes));
1175 fw_size = le32_to_cpu(rlcv21_hdr->save_restore_list_gpm_size_bytes);
1176 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLCG_SCRATCH,
1177 fw_data, fw_size);
1178
1179 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1180 le32_to_cpu(rlcv21_hdr->save_restore_list_srm_offset_bytes));
1181 fw_size = le32_to_cpu(rlcv21_hdr->save_restore_list_srm_size_bytes);
1182 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_SRM_ARAM,
1183 fw_data, fw_size);
1184 }
1185 if (version_minor >= 2) {
1186 rlcv22_hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data;
1187
1188 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1189 le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_offset_bytes));
1190 fw_size = le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_size_bytes);
1191 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLX6_UCODE,
1192 fw_data, fw_size);
1193
1194 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1195 le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_offset_bytes));
1196 fw_size = le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_size_bytes);
1197 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLX6_DRAM_BOOT,
1198 fw_data, fw_size);
1199 }
1200 }
1201 }
1202
1203 static void
gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(struct amdgpu_device * adev)1204 gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(struct amdgpu_device *adev)
1205 {
1206 const __le32 *fw_data;
1207 uint32_t fw_size;
1208 const struct sdma_firmware_header_v3_0 *sdma_hdr;
1209
1210 sdma_hdr = (const struct sdma_firmware_header_v3_0 *)
1211 adev->sdma.instance[0].fw->data;
1212 fw_data = (const __le32 *) (adev->sdma.instance[0].fw->data +
1213 le32_to_cpu(sdma_hdr->ucode_offset_bytes));
1214 fw_size = le32_to_cpu(sdma_hdr->ucode_size_bytes);
1215
1216 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_SDMA_UCODE_TH0,
1217 fw_data, fw_size);
1218 }
1219
1220 static void
gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(struct amdgpu_device * adev)1221 gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(struct amdgpu_device *adev)
1222 {
1223 const __le32 *fw_data;
1224 unsigned fw_size;
1225 const struct mes_firmware_header_v1_0 *mes_hdr;
1226 int pipe, ucode_id, data_id;
1227
1228 for (pipe = 0; pipe < 2; pipe++) {
1229 if (pipe == 0) {
1230 ucode_id = SOC24_FIRMWARE_ID_RS64_MES_P0;
1231 data_id = SOC24_FIRMWARE_ID_RS64_MES_P0_STACK;
1232 } else {
1233 ucode_id = SOC24_FIRMWARE_ID_RS64_MES_P1;
1234 data_id = SOC24_FIRMWARE_ID_RS64_MES_P1_STACK;
1235 }
1236
1237 mes_hdr = (const struct mes_firmware_header_v1_0 *)
1238 adev->mes.fw[pipe]->data;
1239
1240 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
1241 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes));
1242 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes);
1243
1244 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, ucode_id, fw_data, fw_size);
1245
1246 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data +
1247 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes));
1248 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes);
1249
1250 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, data_id, fw_data, fw_size);
1251 }
1252 }
1253
gfx_v12_0_rlc_backdoor_autoload_enable(struct amdgpu_device * adev)1254 static int gfx_v12_0_rlc_backdoor_autoload_enable(struct amdgpu_device *adev)
1255 {
1256 uint32_t rlc_g_offset, rlc_g_size;
1257 uint64_t gpu_addr;
1258 uint32_t data;
1259
1260 /* RLC autoload sequence 2: copy ucode */
1261 gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(adev);
1262 gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(adev);
1263 gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(adev);
1264 gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(adev);
1265
1266 rlc_g_offset = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_G_UCODE].offset;
1267 rlc_g_size = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_G_UCODE].size;
1268 gpu_addr = adev->gfx.rlc.rlc_autoload_gpu_addr + rlc_g_offset - adev->gmc.vram_start;
1269
1270 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_HI, upper_32_bits(gpu_addr));
1271 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_LO, lower_32_bits(gpu_addr));
1272
1273 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_SIZE, rlc_g_size);
1274
1275 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) {
1276 /* RLC autoload sequence 3: load IMU fw */
1277 if (adev->gfx.imu.funcs->load_microcode)
1278 adev->gfx.imu.funcs->load_microcode(adev);
1279 /* RLC autoload sequence 4 init IMU fw */
1280 if (adev->gfx.imu.funcs->setup_imu)
1281 adev->gfx.imu.funcs->setup_imu(adev);
1282 if (adev->gfx.imu.funcs->start_imu)
1283 adev->gfx.imu.funcs->start_imu(adev);
1284
1285 /* RLC autoload sequence 5 disable gpa mode */
1286 gfx_v12_0_disable_gpa_mode(adev);
1287 } else {
1288 /* unhalt rlc to start autoload without imu */
1289 data = RREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE);
1290 data = REG_SET_FIELD(data, RLC_GPM_THREAD_ENABLE, THREAD0_ENABLE, 1);
1291 data = REG_SET_FIELD(data, RLC_GPM_THREAD_ENABLE, THREAD1_ENABLE, 1);
1292 WREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE, data);
1293 WREG32_SOC15(GC, 0, regRLC_CNTL, RLC_CNTL__RLC_ENABLE_F32_MASK);
1294 }
1295
1296 return 0;
1297 }
1298
gfx_v12_0_alloc_ip_dump(struct amdgpu_device * adev)1299 static void gfx_v12_0_alloc_ip_dump(struct amdgpu_device *adev)
1300 {
1301 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0);
1302 uint32_t *ptr;
1303 uint32_t inst;
1304
1305 ptr = kcalloc(reg_count, sizeof(uint32_t), GFP_KERNEL);
1306 if (!ptr) {
1307 DRM_ERROR("Failed to allocate memory for GFX IP Dump\n");
1308 adev->gfx.ip_dump_core = NULL;
1309 } else {
1310 adev->gfx.ip_dump_core = ptr;
1311 }
1312
1313 /* Allocate memory for compute queue registers for all the instances */
1314 reg_count = ARRAY_SIZE(gc_cp_reg_list_12);
1315 inst = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe_per_mec *
1316 adev->gfx.mec.num_queue_per_pipe;
1317
1318 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL);
1319 if (!ptr) {
1320 DRM_ERROR("Failed to allocate memory for Compute Queues IP Dump\n");
1321 adev->gfx.ip_dump_compute_queues = NULL;
1322 } else {
1323 adev->gfx.ip_dump_compute_queues = ptr;
1324 }
1325
1326 /* Allocate memory for gfx queue registers for all the instances */
1327 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12);
1328 inst = adev->gfx.me.num_me * adev->gfx.me.num_pipe_per_me *
1329 adev->gfx.me.num_queue_per_pipe;
1330
1331 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL);
1332 if (!ptr) {
1333 DRM_ERROR("Failed to allocate memory for GFX Queues IP Dump\n");
1334 adev->gfx.ip_dump_gfx_queues = NULL;
1335 } else {
1336 adev->gfx.ip_dump_gfx_queues = ptr;
1337 }
1338 }
1339
gfx_v12_0_sw_init(void * handle)1340 static int gfx_v12_0_sw_init(void *handle)
1341 {
1342 int i, j, k, r, ring_id = 0;
1343 unsigned num_compute_rings;
1344 int xcc_id = 0;
1345 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1346
1347 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1348 case IP_VERSION(12, 0, 0):
1349 case IP_VERSION(12, 0, 1):
1350 adev->gfx.me.num_me = 1;
1351 adev->gfx.me.num_pipe_per_me = 1;
1352 adev->gfx.me.num_queue_per_pipe = 1;
1353 adev->gfx.mec.num_mec = 1;
1354 adev->gfx.mec.num_pipe_per_mec = 2;
1355 adev->gfx.mec.num_queue_per_pipe = 4;
1356 break;
1357 default:
1358 adev->gfx.me.num_me = 1;
1359 adev->gfx.me.num_pipe_per_me = 1;
1360 adev->gfx.me.num_queue_per_pipe = 1;
1361 adev->gfx.mec.num_mec = 1;
1362 adev->gfx.mec.num_pipe_per_mec = 4;
1363 adev->gfx.mec.num_queue_per_pipe = 8;
1364 break;
1365 }
1366
1367 /* recalculate compute rings to use based on hardware configuration */
1368 num_compute_rings = (adev->gfx.mec.num_pipe_per_mec *
1369 adev->gfx.mec.num_queue_per_pipe) / 2;
1370 adev->gfx.num_compute_rings = min(adev->gfx.num_compute_rings,
1371 num_compute_rings);
1372
1373 /* EOP Event */
1374 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP,
1375 GFX_11_0_0__SRCID__CP_EOP_INTERRUPT,
1376 &adev->gfx.eop_irq);
1377 if (r)
1378 return r;
1379
1380 /* Bad opcode Event */
1381 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP,
1382 GFX_11_0_0__SRCID__CP_BAD_OPCODE_ERROR,
1383 &adev->gfx.bad_op_irq);
1384 if (r)
1385 return r;
1386
1387 /* Privileged reg */
1388 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP,
1389 GFX_11_0_0__SRCID__CP_PRIV_REG_FAULT,
1390 &adev->gfx.priv_reg_irq);
1391 if (r)
1392 return r;
1393
1394 /* Privileged inst */
1395 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP,
1396 GFX_11_0_0__SRCID__CP_PRIV_INSTR_FAULT,
1397 &adev->gfx.priv_inst_irq);
1398 if (r)
1399 return r;
1400
1401 adev->gfx.gfx_current_status = AMDGPU_GFX_NORMAL_MODE;
1402
1403 gfx_v12_0_me_init(adev);
1404
1405 r = gfx_v12_0_rlc_init(adev);
1406 if (r) {
1407 dev_err(adev->dev, "Failed to init rlc BOs!\n");
1408 return r;
1409 }
1410
1411 r = gfx_v12_0_mec_init(adev);
1412 if (r) {
1413 dev_err(adev->dev, "Failed to init MEC BOs!\n");
1414 return r;
1415 }
1416
1417 /* set up the gfx ring */
1418 for (i = 0; i < adev->gfx.me.num_me; i++) {
1419 for (j = 0; j < adev->gfx.me.num_queue_per_pipe; j++) {
1420 for (k = 0; k < adev->gfx.me.num_pipe_per_me; k++) {
1421 if (!amdgpu_gfx_is_me_queue_enabled(adev, i, k, j))
1422 continue;
1423
1424 r = gfx_v12_0_gfx_ring_init(adev, ring_id,
1425 i, k, j);
1426 if (r)
1427 return r;
1428 ring_id++;
1429 }
1430 }
1431 }
1432
1433 ring_id = 0;
1434 /* set up the compute queues - allocate horizontally across pipes */
1435 for (i = 0; i < adev->gfx.mec.num_mec; ++i) {
1436 for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) {
1437 for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) {
1438 if (!amdgpu_gfx_is_mec_queue_enabled(adev,
1439 0, i, k, j))
1440 continue;
1441
1442 r = gfx_v12_0_compute_ring_init(adev, ring_id,
1443 i, k, j);
1444 if (r)
1445 return r;
1446
1447 ring_id++;
1448 }
1449 }
1450 }
1451
1452 if (!adev->enable_mes_kiq) {
1453 r = amdgpu_gfx_kiq_init(adev, GFX12_MEC_HPD_SIZE, 0);
1454 if (r) {
1455 dev_err(adev->dev, "Failed to init KIQ BOs!\n");
1456 return r;
1457 }
1458
1459 r = amdgpu_gfx_kiq_init_ring(adev, xcc_id);
1460 if (r)
1461 return r;
1462 }
1463
1464 r = amdgpu_gfx_mqd_sw_init(adev, sizeof(struct v12_compute_mqd), 0);
1465 if (r)
1466 return r;
1467
1468 /* allocate visible FB for rlc auto-loading fw */
1469 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
1470 r = gfx_v12_0_rlc_autoload_buffer_init(adev);
1471 if (r)
1472 return r;
1473 }
1474
1475 r = gfx_v12_0_gpu_early_init(adev);
1476 if (r)
1477 return r;
1478
1479 gfx_v12_0_alloc_ip_dump(adev);
1480
1481 return 0;
1482 }
1483
gfx_v12_0_pfp_fini(struct amdgpu_device * adev)1484 static void gfx_v12_0_pfp_fini(struct amdgpu_device *adev)
1485 {
1486 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_obj,
1487 &adev->gfx.pfp.pfp_fw_gpu_addr,
1488 (void **)&adev->gfx.pfp.pfp_fw_ptr);
1489
1490 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_data_obj,
1491 &adev->gfx.pfp.pfp_fw_data_gpu_addr,
1492 (void **)&adev->gfx.pfp.pfp_fw_data_ptr);
1493 }
1494
gfx_v12_0_me_fini(struct amdgpu_device * adev)1495 static void gfx_v12_0_me_fini(struct amdgpu_device *adev)
1496 {
1497 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_obj,
1498 &adev->gfx.me.me_fw_gpu_addr,
1499 (void **)&adev->gfx.me.me_fw_ptr);
1500
1501 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_data_obj,
1502 &adev->gfx.me.me_fw_data_gpu_addr,
1503 (void **)&adev->gfx.me.me_fw_data_ptr);
1504 }
1505
gfx_v12_0_rlc_autoload_buffer_fini(struct amdgpu_device * adev)1506 static void gfx_v12_0_rlc_autoload_buffer_fini(struct amdgpu_device *adev)
1507 {
1508 amdgpu_bo_free_kernel(&adev->gfx.rlc.rlc_autoload_bo,
1509 &adev->gfx.rlc.rlc_autoload_gpu_addr,
1510 (void **)&adev->gfx.rlc.rlc_autoload_ptr);
1511 }
1512
gfx_v12_0_sw_fini(void * handle)1513 static int gfx_v12_0_sw_fini(void *handle)
1514 {
1515 int i;
1516 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1517
1518 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
1519 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
1520 for (i = 0; i < adev->gfx.num_compute_rings; i++)
1521 amdgpu_ring_fini(&adev->gfx.compute_ring[i]);
1522
1523 amdgpu_gfx_mqd_sw_fini(adev, 0);
1524
1525 if (!adev->enable_mes_kiq) {
1526 amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq[0].ring);
1527 amdgpu_gfx_kiq_fini(adev, 0);
1528 }
1529
1530 gfx_v12_0_pfp_fini(adev);
1531 gfx_v12_0_me_fini(adev);
1532 gfx_v12_0_rlc_fini(adev);
1533 gfx_v12_0_mec_fini(adev);
1534
1535 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO)
1536 gfx_v12_0_rlc_autoload_buffer_fini(adev);
1537
1538 gfx_v12_0_free_microcode(adev);
1539
1540 kfree(adev->gfx.ip_dump_core);
1541 kfree(adev->gfx.ip_dump_compute_queues);
1542 kfree(adev->gfx.ip_dump_gfx_queues);
1543
1544 return 0;
1545 }
1546
gfx_v12_0_select_se_sh(struct amdgpu_device * adev,u32 se_num,u32 sh_num,u32 instance,int xcc_id)1547 static void gfx_v12_0_select_se_sh(struct amdgpu_device *adev, u32 se_num,
1548 u32 sh_num, u32 instance, int xcc_id)
1549 {
1550 u32 data;
1551
1552 if (instance == 0xffffffff)
1553 data = REG_SET_FIELD(0, GRBM_GFX_INDEX,
1554 INSTANCE_BROADCAST_WRITES, 1);
1555 else
1556 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_INDEX,
1557 instance);
1558
1559 if (se_num == 0xffffffff)
1560 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES,
1561 1);
1562 else
1563 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num);
1564
1565 if (sh_num == 0xffffffff)
1566 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_BROADCAST_WRITES,
1567 1);
1568 else
1569 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_INDEX, sh_num);
1570
1571 WREG32_SOC15(GC, 0, regGRBM_GFX_INDEX, data);
1572 }
1573
gfx_v12_0_get_sa_active_bitmap(struct amdgpu_device * adev)1574 static u32 gfx_v12_0_get_sa_active_bitmap(struct amdgpu_device *adev)
1575 {
1576 u32 gc_disabled_sa_mask, gc_user_disabled_sa_mask, sa_mask;
1577
1578 gc_disabled_sa_mask = RREG32_SOC15(GC, 0, regGRBM_CC_GC_SA_UNIT_DISABLE);
1579 gc_disabled_sa_mask = REG_GET_FIELD(gc_disabled_sa_mask,
1580 GRBM_CC_GC_SA_UNIT_DISABLE,
1581 SA_DISABLE);
1582 gc_user_disabled_sa_mask = RREG32_SOC15(GC, 0, regGRBM_GC_USER_SA_UNIT_DISABLE);
1583 gc_user_disabled_sa_mask = REG_GET_FIELD(gc_user_disabled_sa_mask,
1584 GRBM_GC_USER_SA_UNIT_DISABLE,
1585 SA_DISABLE);
1586 sa_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_sh_per_se *
1587 adev->gfx.config.max_shader_engines);
1588
1589 return sa_mask & (~(gc_disabled_sa_mask | gc_user_disabled_sa_mask));
1590 }
1591
gfx_v12_0_get_rb_active_bitmap(struct amdgpu_device * adev)1592 static u32 gfx_v12_0_get_rb_active_bitmap(struct amdgpu_device *adev)
1593 {
1594 u32 gc_disabled_rb_mask, gc_user_disabled_rb_mask;
1595 u32 rb_mask;
1596
1597 gc_disabled_rb_mask = RREG32_SOC15(GC, 0, regCC_RB_BACKEND_DISABLE);
1598 gc_disabled_rb_mask = REG_GET_FIELD(gc_disabled_rb_mask,
1599 CC_RB_BACKEND_DISABLE,
1600 BACKEND_DISABLE);
1601 gc_user_disabled_rb_mask = RREG32_SOC15(GC, 0, regGC_USER_RB_BACKEND_DISABLE);
1602 gc_user_disabled_rb_mask = REG_GET_FIELD(gc_user_disabled_rb_mask,
1603 GC_USER_RB_BACKEND_DISABLE,
1604 BACKEND_DISABLE);
1605 rb_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_backends_per_se *
1606 adev->gfx.config.max_shader_engines);
1607
1608 return rb_mask & (~(gc_disabled_rb_mask | gc_user_disabled_rb_mask));
1609 }
1610
gfx_v12_0_setup_rb(struct amdgpu_device * adev)1611 static void gfx_v12_0_setup_rb(struct amdgpu_device *adev)
1612 {
1613 u32 rb_bitmap_width_per_sa;
1614 u32 max_sa;
1615 u32 active_sa_bitmap;
1616 u32 global_active_rb_bitmap;
1617 u32 active_rb_bitmap = 0;
1618 u32 i;
1619
1620 /* query sa bitmap from SA_UNIT_DISABLE registers */
1621 active_sa_bitmap = gfx_v12_0_get_sa_active_bitmap(adev);
1622 /* query rb bitmap from RB_BACKEND_DISABLE registers */
1623 global_active_rb_bitmap = gfx_v12_0_get_rb_active_bitmap(adev);
1624
1625 /* generate active rb bitmap according to active sa bitmap */
1626 max_sa = adev->gfx.config.max_shader_engines *
1627 adev->gfx.config.max_sh_per_se;
1628 rb_bitmap_width_per_sa = adev->gfx.config.max_backends_per_se /
1629 adev->gfx.config.max_sh_per_se;
1630 for (i = 0; i < max_sa; i++) {
1631 if (active_sa_bitmap & (1 << i))
1632 active_rb_bitmap |= (0x3 << (i * rb_bitmap_width_per_sa));
1633 }
1634
1635 active_rb_bitmap |= global_active_rb_bitmap;
1636 adev->gfx.config.backend_enable_mask = active_rb_bitmap;
1637 adev->gfx.config.num_rbs = hweight32(active_rb_bitmap);
1638 }
1639
1640 #define LDS_APP_BASE 0x1
1641 #define SCRATCH_APP_BASE 0x2
1642
gfx_v12_0_init_compute_vmid(struct amdgpu_device * adev)1643 static void gfx_v12_0_init_compute_vmid(struct amdgpu_device *adev)
1644 {
1645 int i;
1646 uint32_t sh_mem_bases;
1647 uint32_t data;
1648
1649 /*
1650 * Configure apertures:
1651 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB)
1652 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB)
1653 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB)
1654 */
1655 sh_mem_bases = (LDS_APP_BASE << SH_MEM_BASES__SHARED_BASE__SHIFT) |
1656 SCRATCH_APP_BASE;
1657
1658 mutex_lock(&adev->srbm_mutex);
1659 for (i = adev->vm_manager.first_kfd_vmid; i < AMDGPU_NUM_VMID; i++) {
1660 soc24_grbm_select(adev, 0, 0, 0, i);
1661 /* CP and shaders */
1662 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG);
1663 WREG32_SOC15(GC, 0, regSH_MEM_BASES, sh_mem_bases);
1664
1665 /* Enable trap for each kfd vmid. */
1666 data = RREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL);
1667 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1);
1668 WREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL, data);
1669 }
1670 soc24_grbm_select(adev, 0, 0, 0, 0);
1671 mutex_unlock(&adev->srbm_mutex);
1672 }
1673
gfx_v12_0_tcp_harvest(struct amdgpu_device * adev)1674 static void gfx_v12_0_tcp_harvest(struct amdgpu_device *adev)
1675 {
1676 /* TODO: harvest feature to be added later. */
1677 }
1678
gfx_v12_0_get_tcc_info(struct amdgpu_device * adev)1679 static void gfx_v12_0_get_tcc_info(struct amdgpu_device *adev)
1680 {
1681 }
1682
gfx_v12_0_constants_init(struct amdgpu_device * adev)1683 static void gfx_v12_0_constants_init(struct amdgpu_device *adev)
1684 {
1685 u32 tmp;
1686 int i;
1687
1688 if (!amdgpu_sriov_vf(adev))
1689 WREG32_FIELD15_PREREG(GC, 0, GRBM_CNTL, READ_TIMEOUT, 0xff);
1690
1691 gfx_v12_0_setup_rb(adev);
1692 gfx_v12_0_get_cu_info(adev, &adev->gfx.cu_info);
1693 gfx_v12_0_get_tcc_info(adev);
1694 adev->gfx.config.pa_sc_tile_steering_override = 0;
1695
1696 /* XXX SH_MEM regs */
1697 /* where to put LDS, scratch, GPUVM in FSA64 space */
1698 mutex_lock(&adev->srbm_mutex);
1699 for (i = 0; i < adev->vm_manager.id_mgr[AMDGPU_GFXHUB(0)].num_ids; i++) {
1700 soc24_grbm_select(adev, 0, 0, 0, i);
1701 /* CP and shaders */
1702 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG);
1703 if (i != 0) {
1704 tmp = REG_SET_FIELD(0, SH_MEM_BASES, PRIVATE_BASE,
1705 (adev->gmc.private_aperture_start >> 48));
1706 tmp = REG_SET_FIELD(tmp, SH_MEM_BASES, SHARED_BASE,
1707 (adev->gmc.shared_aperture_start >> 48));
1708 WREG32_SOC15(GC, 0, regSH_MEM_BASES, tmp);
1709 }
1710 }
1711 soc24_grbm_select(adev, 0, 0, 0, 0);
1712
1713 mutex_unlock(&adev->srbm_mutex);
1714
1715 gfx_v12_0_init_compute_vmid(adev);
1716 }
1717
gfx_v12_0_get_cpg_int_cntl(struct amdgpu_device * adev,int me,int pipe)1718 static u32 gfx_v12_0_get_cpg_int_cntl(struct amdgpu_device *adev,
1719 int me, int pipe)
1720 {
1721 if (me != 0)
1722 return 0;
1723
1724 switch (pipe) {
1725 case 0:
1726 return SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0);
1727 default:
1728 return 0;
1729 }
1730 }
1731
gfx_v12_0_get_cpc_int_cntl(struct amdgpu_device * adev,int me,int pipe)1732 static u32 gfx_v12_0_get_cpc_int_cntl(struct amdgpu_device *adev,
1733 int me, int pipe)
1734 {
1735 /*
1736 * amdgpu controls only the first MEC. That's why this function only
1737 * handles the setting of interrupts for this specific MEC. All other
1738 * pipes' interrupts are set by amdkfd.
1739 */
1740 if (me != 1)
1741 return 0;
1742
1743 switch (pipe) {
1744 case 0:
1745 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL);
1746 case 1:
1747 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL);
1748 default:
1749 return 0;
1750 }
1751 }
1752
gfx_v12_0_enable_gui_idle_interrupt(struct amdgpu_device * adev,bool enable)1753 static void gfx_v12_0_enable_gui_idle_interrupt(struct amdgpu_device *adev,
1754 bool enable)
1755 {
1756 u32 tmp, cp_int_cntl_reg;
1757 int i, j;
1758
1759 if (amdgpu_sriov_vf(adev))
1760 return;
1761
1762 for (i = 0; i < adev->gfx.me.num_me; i++) {
1763 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) {
1764 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j);
1765
1766 if (cp_int_cntl_reg) {
1767 tmp = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
1768 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE,
1769 enable ? 1 : 0);
1770 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_EMPTY_INT_ENABLE,
1771 enable ? 1 : 0);
1772 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CMP_BUSY_INT_ENABLE,
1773 enable ? 1 : 0);
1774 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, GFX_IDLE_INT_ENABLE,
1775 enable ? 1 : 0);
1776 WREG32_SOC15_IP(GC, cp_int_cntl_reg, tmp);
1777 }
1778 }
1779 }
1780 }
1781
gfx_v12_0_init_csb(struct amdgpu_device * adev)1782 static int gfx_v12_0_init_csb(struct amdgpu_device *adev)
1783 {
1784 adev->gfx.rlc.funcs->get_csb_buffer(adev, adev->gfx.rlc.cs_ptr);
1785
1786 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_HI,
1787 adev->gfx.rlc.clear_state_gpu_addr >> 32);
1788 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_LO,
1789 adev->gfx.rlc.clear_state_gpu_addr & 0xfffffffc);
1790 WREG32_SOC15(GC, 0, regRLC_CSIB_LENGTH, adev->gfx.rlc.clear_state_size);
1791
1792 return 0;
1793 }
1794
gfx_v12_0_rlc_stop(struct amdgpu_device * adev)1795 static void gfx_v12_0_rlc_stop(struct amdgpu_device *adev)
1796 {
1797 u32 tmp = RREG32_SOC15(GC, 0, regRLC_CNTL);
1798
1799 tmp = REG_SET_FIELD(tmp, RLC_CNTL, RLC_ENABLE_F32, 0);
1800 WREG32_SOC15(GC, 0, regRLC_CNTL, tmp);
1801 }
1802
gfx_v12_0_rlc_reset(struct amdgpu_device * adev)1803 static void gfx_v12_0_rlc_reset(struct amdgpu_device *adev)
1804 {
1805 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1);
1806 udelay(50);
1807 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0);
1808 udelay(50);
1809 }
1810
gfx_v12_0_rlc_smu_handshake_cntl(struct amdgpu_device * adev,bool enable)1811 static void gfx_v12_0_rlc_smu_handshake_cntl(struct amdgpu_device *adev,
1812 bool enable)
1813 {
1814 uint32_t rlc_pg_cntl;
1815
1816 rlc_pg_cntl = RREG32_SOC15(GC, 0, regRLC_PG_CNTL);
1817
1818 if (!enable) {
1819 /* RLC_PG_CNTL[23] = 0 (default)
1820 * RLC will wait for handshake acks with SMU
1821 * GFXOFF will be enabled
1822 * RLC_PG_CNTL[23] = 1
1823 * RLC will not issue any message to SMU
1824 * hence no handshake between SMU & RLC
1825 * GFXOFF will be disabled
1826 */
1827 rlc_pg_cntl |= RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK;
1828 } else
1829 rlc_pg_cntl &= ~RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK;
1830 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, rlc_pg_cntl);
1831 }
1832
gfx_v12_0_rlc_start(struct amdgpu_device * adev)1833 static void gfx_v12_0_rlc_start(struct amdgpu_device *adev)
1834 {
1835 /* TODO: enable rlc & smu handshake until smu
1836 * and gfxoff feature works as expected */
1837 if (!(amdgpu_pp_feature_mask & PP_GFXOFF_MASK))
1838 gfx_v12_0_rlc_smu_handshake_cntl(adev, false);
1839
1840 WREG32_FIELD15_PREREG(GC, 0, RLC_CNTL, RLC_ENABLE_F32, 1);
1841 udelay(50);
1842 }
1843
gfx_v12_0_rlc_enable_srm(struct amdgpu_device * adev)1844 static void gfx_v12_0_rlc_enable_srm(struct amdgpu_device *adev)
1845 {
1846 uint32_t tmp;
1847
1848 /* enable Save Restore Machine */
1849 tmp = RREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL));
1850 tmp |= RLC_SRM_CNTL__AUTO_INCR_ADDR_MASK;
1851 tmp |= RLC_SRM_CNTL__SRM_ENABLE_MASK;
1852 WREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL), tmp);
1853 }
1854
gfx_v12_0_load_rlcg_microcode(struct amdgpu_device * adev)1855 static void gfx_v12_0_load_rlcg_microcode(struct amdgpu_device *adev)
1856 {
1857 const struct rlc_firmware_header_v2_0 *hdr;
1858 const __le32 *fw_data;
1859 unsigned i, fw_size;
1860
1861 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data;
1862 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1863 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1864 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
1865
1866 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR,
1867 RLCG_UCODE_LOADING_START_ADDRESS);
1868
1869 for (i = 0; i < fw_size; i++)
1870 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_DATA,
1871 le32_to_cpup(fw_data++));
1872
1873 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version);
1874 }
1875
gfx_v12_0_load_rlc_iram_dram_microcode(struct amdgpu_device * adev)1876 static void gfx_v12_0_load_rlc_iram_dram_microcode(struct amdgpu_device *adev)
1877 {
1878 const struct rlc_firmware_header_v2_2 *hdr;
1879 const __le32 *fw_data;
1880 unsigned i, fw_size;
1881 u32 tmp;
1882
1883 hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data;
1884
1885 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1886 le32_to_cpu(hdr->rlc_iram_ucode_offset_bytes));
1887 fw_size = le32_to_cpu(hdr->rlc_iram_ucode_size_bytes) / 4;
1888
1889 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, 0);
1890
1891 for (i = 0; i < fw_size; i++) {
1892 if ((amdgpu_emu_mode == 1) && (i % 100 == 99))
1893 msleep(1);
1894 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_DATA,
1895 le32_to_cpup(fw_data++));
1896 }
1897
1898 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version);
1899
1900 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1901 le32_to_cpu(hdr->rlc_dram_ucode_offset_bytes));
1902 fw_size = le32_to_cpu(hdr->rlc_dram_ucode_size_bytes) / 4;
1903
1904 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_ADDR, 0);
1905 for (i = 0; i < fw_size; i++) {
1906 if ((amdgpu_emu_mode == 1) && (i % 100 == 99))
1907 msleep(1);
1908 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_DATA,
1909 le32_to_cpup(fw_data++));
1910 }
1911
1912 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version);
1913
1914 tmp = RREG32_SOC15(GC, 0, regRLC_LX6_CNTL);
1915 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, PDEBUG_ENABLE, 1);
1916 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, BRESET, 0);
1917 WREG32_SOC15(GC, 0, regRLC_LX6_CNTL, tmp);
1918 }
1919
gfx_v12_0_rlc_load_microcode(struct amdgpu_device * adev)1920 static int gfx_v12_0_rlc_load_microcode(struct amdgpu_device *adev)
1921 {
1922 const struct rlc_firmware_header_v2_0 *hdr;
1923 uint16_t version_major;
1924 uint16_t version_minor;
1925
1926 if (!adev->gfx.rlc_fw)
1927 return -EINVAL;
1928
1929 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data;
1930 amdgpu_ucode_print_rlc_hdr(&hdr->header);
1931
1932 version_major = le16_to_cpu(hdr->header.header_version_major);
1933 version_minor = le16_to_cpu(hdr->header.header_version_minor);
1934
1935 if (version_major == 2) {
1936 gfx_v12_0_load_rlcg_microcode(adev);
1937 if (amdgpu_dpm == 1) {
1938 if (version_minor >= 2)
1939 gfx_v12_0_load_rlc_iram_dram_microcode(adev);
1940 }
1941
1942 return 0;
1943 }
1944
1945 return -EINVAL;
1946 }
1947
gfx_v12_0_rlc_resume(struct amdgpu_device * adev)1948 static int gfx_v12_0_rlc_resume(struct amdgpu_device *adev)
1949 {
1950 int r;
1951
1952 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
1953 gfx_v12_0_init_csb(adev);
1954
1955 if (!amdgpu_sriov_vf(adev)) /* enable RLC SRM */
1956 gfx_v12_0_rlc_enable_srm(adev);
1957 } else {
1958 if (amdgpu_sriov_vf(adev)) {
1959 gfx_v12_0_init_csb(adev);
1960 return 0;
1961 }
1962
1963 adev->gfx.rlc.funcs->stop(adev);
1964
1965 /* disable CG */
1966 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, 0);
1967
1968 /* disable PG */
1969 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, 0);
1970
1971 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
1972 /* legacy rlc firmware loading */
1973 r = gfx_v12_0_rlc_load_microcode(adev);
1974 if (r)
1975 return r;
1976 }
1977
1978 gfx_v12_0_init_csb(adev);
1979
1980 adev->gfx.rlc.funcs->start(adev);
1981 }
1982
1983 return 0;
1984 }
1985
gfx_v12_0_config_gfx_rs64(struct amdgpu_device * adev)1986 static void gfx_v12_0_config_gfx_rs64(struct amdgpu_device *adev)
1987 {
1988 const struct gfx_firmware_header_v2_0 *pfp_hdr;
1989 const struct gfx_firmware_header_v2_0 *me_hdr;
1990 const struct gfx_firmware_header_v2_0 *mec_hdr;
1991 uint32_t pipe_id, tmp;
1992
1993 mec_hdr = (const struct gfx_firmware_header_v2_0 *)
1994 adev->gfx.mec_fw->data;
1995 me_hdr = (const struct gfx_firmware_header_v2_0 *)
1996 adev->gfx.me_fw->data;
1997 pfp_hdr = (const struct gfx_firmware_header_v2_0 *)
1998 adev->gfx.pfp_fw->data;
1999
2000 /* config pfp program start addr */
2001 for (pipe_id = 0; pipe_id < 2; pipe_id++) {
2002 soc24_grbm_select(adev, 0, pipe_id, 0, 0);
2003 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START,
2004 (pfp_hdr->ucode_start_addr_hi << 30) |
2005 (pfp_hdr->ucode_start_addr_lo >> 2));
2006 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI,
2007 pfp_hdr->ucode_start_addr_hi >> 2);
2008 }
2009 soc24_grbm_select(adev, 0, 0, 0, 0);
2010
2011 /* reset pfp pipe */
2012 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL);
2013 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 1);
2014 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 1);
2015 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2016
2017 /* clear pfp pipe reset */
2018 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 0);
2019 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 0);
2020 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2021
2022 /* config me program start addr */
2023 for (pipe_id = 0; pipe_id < 2; pipe_id++) {
2024 soc24_grbm_select(adev, 0, pipe_id, 0, 0);
2025 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START,
2026 (me_hdr->ucode_start_addr_hi << 30) |
2027 (me_hdr->ucode_start_addr_lo >> 2));
2028 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI,
2029 me_hdr->ucode_start_addr_hi>>2);
2030 }
2031 soc24_grbm_select(adev, 0, 0, 0, 0);
2032
2033 /* reset me pipe */
2034 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL);
2035 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 1);
2036 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 1);
2037 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2038
2039 /* clear me pipe reset */
2040 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 0);
2041 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 0);
2042 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2043
2044 /* config mec program start addr */
2045 for (pipe_id = 0; pipe_id < 4; pipe_id++) {
2046 soc24_grbm_select(adev, 1, pipe_id, 0, 0);
2047 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START,
2048 mec_hdr->ucode_start_addr_lo >> 2 |
2049 mec_hdr->ucode_start_addr_hi << 30);
2050 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI,
2051 mec_hdr->ucode_start_addr_hi >> 2);
2052 }
2053 soc24_grbm_select(adev, 0, 0, 0, 0);
2054
2055 /* reset mec pipe */
2056 tmp = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL);
2057 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 1);
2058 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 1);
2059 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 1);
2060 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 1);
2061 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp);
2062
2063 /* clear mec pipe reset */
2064 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 0);
2065 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 0);
2066 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 0);
2067 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 0);
2068 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp);
2069 }
2070
gfx_v12_0_set_pfp_ucode_start_addr(struct amdgpu_device * adev)2071 static void gfx_v12_0_set_pfp_ucode_start_addr(struct amdgpu_device *adev)
2072 {
2073 const struct gfx_firmware_header_v2_0 *cp_hdr;
2074 unsigned pipe_id, tmp;
2075
2076 cp_hdr = (const struct gfx_firmware_header_v2_0 *)
2077 adev->gfx.pfp_fw->data;
2078 mutex_lock(&adev->srbm_mutex);
2079 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) {
2080 soc24_grbm_select(adev, 0, pipe_id, 0, 0);
2081 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START,
2082 (cp_hdr->ucode_start_addr_hi << 30) |
2083 (cp_hdr->ucode_start_addr_lo >> 2));
2084 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI,
2085 cp_hdr->ucode_start_addr_hi>>2);
2086
2087 /*
2088 * Program CP_ME_CNTL to reset given PIPE to take
2089 * effect of CP_PFP_PRGRM_CNTR_START.
2090 */
2091 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL);
2092 if (pipe_id == 0)
2093 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2094 PFP_PIPE0_RESET, 1);
2095 else
2096 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2097 PFP_PIPE1_RESET, 1);
2098 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2099
2100 /* Clear pfp pipe0 reset bit. */
2101 if (pipe_id == 0)
2102 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2103 PFP_PIPE0_RESET, 0);
2104 else
2105 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2106 PFP_PIPE1_RESET, 0);
2107 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2108 }
2109 soc24_grbm_select(adev, 0, 0, 0, 0);
2110 mutex_unlock(&adev->srbm_mutex);
2111 }
2112
gfx_v12_0_set_me_ucode_start_addr(struct amdgpu_device * adev)2113 static void gfx_v12_0_set_me_ucode_start_addr(struct amdgpu_device *adev)
2114 {
2115 const struct gfx_firmware_header_v2_0 *cp_hdr;
2116 unsigned pipe_id, tmp;
2117
2118 cp_hdr = (const struct gfx_firmware_header_v2_0 *)
2119 adev->gfx.me_fw->data;
2120 mutex_lock(&adev->srbm_mutex);
2121 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) {
2122 soc24_grbm_select(adev, 0, pipe_id, 0, 0);
2123 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START,
2124 (cp_hdr->ucode_start_addr_hi << 30) |
2125 (cp_hdr->ucode_start_addr_lo >> 2) );
2126 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI,
2127 cp_hdr->ucode_start_addr_hi>>2);
2128
2129 /*
2130 * Program CP_ME_CNTL to reset given PIPE to take
2131 * effect of CP_ME_PRGRM_CNTR_START.
2132 */
2133 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL);
2134 if (pipe_id == 0)
2135 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2136 ME_PIPE0_RESET, 1);
2137 else
2138 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2139 ME_PIPE1_RESET, 1);
2140 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2141
2142 /* Clear pfp pipe0 reset bit. */
2143 if (pipe_id == 0)
2144 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2145 ME_PIPE0_RESET, 0);
2146 else
2147 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL,
2148 ME_PIPE1_RESET, 0);
2149 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2150 }
2151 soc24_grbm_select(adev, 0, 0, 0, 0);
2152 mutex_unlock(&adev->srbm_mutex);
2153 }
2154
gfx_v12_0_set_mec_ucode_start_addr(struct amdgpu_device * adev)2155 static void gfx_v12_0_set_mec_ucode_start_addr(struct amdgpu_device *adev)
2156 {
2157 const struct gfx_firmware_header_v2_0 *cp_hdr;
2158 unsigned pipe_id;
2159
2160 cp_hdr = (const struct gfx_firmware_header_v2_0 *)
2161 adev->gfx.mec_fw->data;
2162 mutex_lock(&adev->srbm_mutex);
2163 for (pipe_id = 0; pipe_id < adev->gfx.mec.num_pipe_per_mec; pipe_id++) {
2164 soc24_grbm_select(adev, 1, pipe_id, 0, 0);
2165 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START,
2166 cp_hdr->ucode_start_addr_lo >> 2 |
2167 cp_hdr->ucode_start_addr_hi << 30);
2168 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI,
2169 cp_hdr->ucode_start_addr_hi >> 2);
2170 }
2171 soc24_grbm_select(adev, 0, 0, 0, 0);
2172 mutex_unlock(&adev->srbm_mutex);
2173 }
2174
gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device * adev)2175 static int gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev)
2176 {
2177 uint32_t cp_status;
2178 uint32_t bootload_status;
2179 int i;
2180
2181 for (i = 0; i < adev->usec_timeout; i++) {
2182 cp_status = RREG32_SOC15(GC, 0, regCP_STAT);
2183 bootload_status = RREG32_SOC15(GC, 0, regRLC_RLCS_BOOTLOAD_STATUS);
2184
2185 if ((cp_status == 0) &&
2186 (REG_GET_FIELD(bootload_status,
2187 RLC_RLCS_BOOTLOAD_STATUS, BOOTLOAD_COMPLETE) == 1)) {
2188 break;
2189 }
2190 udelay(1);
2191 if (amdgpu_emu_mode)
2192 msleep(10);
2193 }
2194
2195 if (i >= adev->usec_timeout) {
2196 dev_err(adev->dev, "rlc autoload: gc ucode autoload timeout\n");
2197 return -ETIMEDOUT;
2198 }
2199
2200 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
2201 gfx_v12_0_set_pfp_ucode_start_addr(adev);
2202 gfx_v12_0_set_me_ucode_start_addr(adev);
2203 gfx_v12_0_set_mec_ucode_start_addr(adev);
2204 }
2205
2206 return 0;
2207 }
2208
gfx_v12_0_cp_gfx_enable(struct amdgpu_device * adev,bool enable)2209 static int gfx_v12_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable)
2210 {
2211 int i;
2212 u32 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL);
2213
2214 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_HALT, enable ? 0 : 1);
2215 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_HALT, enable ? 0 : 1);
2216 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp);
2217
2218 for (i = 0; i < adev->usec_timeout; i++) {
2219 if (RREG32_SOC15(GC, 0, regCP_STAT) == 0)
2220 break;
2221 udelay(1);
2222 }
2223
2224 if (i >= adev->usec_timeout)
2225 DRM_ERROR("failed to %s cp gfx\n", enable ? "unhalt" : "halt");
2226
2227 return 0;
2228 }
2229
gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(struct amdgpu_device * adev)2230 static int gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(struct amdgpu_device *adev)
2231 {
2232 int r;
2233 const struct gfx_firmware_header_v2_0 *pfp_hdr;
2234 const __le32 *fw_ucode, *fw_data;
2235 unsigned i, pipe_id, fw_ucode_size, fw_data_size;
2236 uint32_t tmp;
2237 uint32_t usec_timeout = 50000; /* wait for 50ms */
2238
2239 pfp_hdr = (const struct gfx_firmware_header_v2_0 *)
2240 adev->gfx.pfp_fw->data;
2241
2242 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header);
2243
2244 /* instruction */
2245 fw_ucode = (const __le32 *)(adev->gfx.pfp_fw->data +
2246 le32_to_cpu(pfp_hdr->ucode_offset_bytes));
2247 fw_ucode_size = le32_to_cpu(pfp_hdr->ucode_size_bytes);
2248 /* data */
2249 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data +
2250 le32_to_cpu(pfp_hdr->data_offset_bytes));
2251 fw_data_size = le32_to_cpu(pfp_hdr->data_size_bytes);
2252
2253 /* 64kb align */
2254 r = amdgpu_bo_create_reserved(adev, fw_ucode_size,
2255 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
2256 &adev->gfx.pfp.pfp_fw_obj,
2257 &adev->gfx.pfp.pfp_fw_gpu_addr,
2258 (void **)&adev->gfx.pfp.pfp_fw_ptr);
2259 if (r) {
2260 dev_err(adev->dev, "(%d) failed to create pfp ucode fw bo\n", r);
2261 gfx_v12_0_pfp_fini(adev);
2262 return r;
2263 }
2264
2265 r = amdgpu_bo_create_reserved(adev, fw_data_size,
2266 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
2267 &adev->gfx.pfp.pfp_fw_data_obj,
2268 &adev->gfx.pfp.pfp_fw_data_gpu_addr,
2269 (void **)&adev->gfx.pfp.pfp_fw_data_ptr);
2270 if (r) {
2271 dev_err(adev->dev, "(%d) failed to create pfp data fw bo\n", r);
2272 gfx_v12_0_pfp_fini(adev);
2273 return r;
2274 }
2275
2276 memcpy(adev->gfx.pfp.pfp_fw_ptr, fw_ucode, fw_ucode_size);
2277 memcpy(adev->gfx.pfp.pfp_fw_data_ptr, fw_data, fw_data_size);
2278
2279 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_obj);
2280 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_data_obj);
2281 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_obj);
2282 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_data_obj);
2283
2284 if (amdgpu_emu_mode == 1)
2285 amdgpu_device_flush_hdp(adev, NULL);
2286
2287 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_LO,
2288 lower_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr));
2289 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_HI,
2290 upper_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr));
2291
2292 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL);
2293 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, VMID, 0);
2294 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, CACHE_POLICY, 0);
2295 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, EXE_DISABLE, 0);
2296 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL, tmp);
2297
2298 /*
2299 * Programming any of the CP_PFP_IC_BASE registers
2300 * forces invalidation of the ME L1 I$. Wait for the
2301 * invalidation complete
2302 */
2303 for (i = 0; i < usec_timeout; i++) {
2304 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL);
2305 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL,
2306 INVALIDATE_CACHE_COMPLETE))
2307 break;
2308 udelay(1);
2309 }
2310
2311 if (i >= usec_timeout) {
2312 dev_err(adev->dev, "failed to invalidate instruction cache\n");
2313 return -EINVAL;
2314 }
2315
2316 /* Prime the L1 instruction caches */
2317 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL);
2318 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_OP_CNTL, PRIME_ICACHE, 1);
2319 WREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL, tmp);
2320 /* Waiting for cache primed*/
2321 for (i = 0; i < usec_timeout; i++) {
2322 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL);
2323 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL,
2324 ICACHE_PRIMED))
2325 break;
2326 udelay(1);
2327 }
2328
2329 if (i >= usec_timeout) {
2330 dev_err(adev->dev, "failed to prime instruction cache\n");
2331 return -EINVAL;
2332 }
2333
2334 mutex_lock(&adev->srbm_mutex);
2335 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) {
2336 soc24_grbm_select(adev, 0, pipe_id, 0, 0);
2337
2338 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_LO,
2339 lower_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr));
2340 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_HI,
2341 upper_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr));
2342 }
2343 soc24_grbm_select(adev, 0, 0, 0, 0);
2344 mutex_unlock(&adev->srbm_mutex);
2345
2346 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL);
2347 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0);
2348 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0);
2349 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp);
2350
2351 /* Invalidate the data caches */
2352 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL);
2353 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1);
2354 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp);
2355
2356 for (i = 0; i < usec_timeout; i++) {
2357 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL);
2358 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL,
2359 INVALIDATE_DCACHE_COMPLETE))
2360 break;
2361 udelay(1);
2362 }
2363
2364 if (i >= usec_timeout) {
2365 dev_err(adev->dev, "failed to invalidate RS64 data cache\n");
2366 return -EINVAL;
2367 }
2368
2369 gfx_v12_0_set_pfp_ucode_start_addr(adev);
2370
2371 return 0;
2372 }
2373
gfx_v12_0_cp_gfx_load_me_microcode_rs64(struct amdgpu_device * adev)2374 static int gfx_v12_0_cp_gfx_load_me_microcode_rs64(struct amdgpu_device *adev)
2375 {
2376 int r;
2377 const struct gfx_firmware_header_v2_0 *me_hdr;
2378 const __le32 *fw_ucode, *fw_data;
2379 unsigned i, pipe_id, fw_ucode_size, fw_data_size;
2380 uint32_t tmp;
2381 uint32_t usec_timeout = 50000; /* wait for 50ms */
2382
2383 me_hdr = (const struct gfx_firmware_header_v2_0 *)
2384 adev->gfx.me_fw->data;
2385
2386 amdgpu_ucode_print_gfx_hdr(&me_hdr->header);
2387
2388 /* instruction */
2389 fw_ucode = (const __le32 *)(adev->gfx.me_fw->data +
2390 le32_to_cpu(me_hdr->ucode_offset_bytes));
2391 fw_ucode_size = le32_to_cpu(me_hdr->ucode_size_bytes);
2392 /* data */
2393 fw_data = (const __le32 *)(adev->gfx.me_fw->data +
2394 le32_to_cpu(me_hdr->data_offset_bytes));
2395 fw_data_size = le32_to_cpu(me_hdr->data_size_bytes);
2396
2397 /* 64kb align*/
2398 r = amdgpu_bo_create_reserved(adev, fw_ucode_size,
2399 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
2400 &adev->gfx.me.me_fw_obj,
2401 &adev->gfx.me.me_fw_gpu_addr,
2402 (void **)&adev->gfx.me.me_fw_ptr);
2403 if (r) {
2404 dev_err(adev->dev, "(%d) failed to create me ucode bo\n", r);
2405 gfx_v12_0_me_fini(adev);
2406 return r;
2407 }
2408
2409 r = amdgpu_bo_create_reserved(adev, fw_data_size,
2410 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
2411 &adev->gfx.me.me_fw_data_obj,
2412 &adev->gfx.me.me_fw_data_gpu_addr,
2413 (void **)&adev->gfx.me.me_fw_data_ptr);
2414 if (r) {
2415 dev_err(adev->dev, "(%d) failed to create me data bo\n", r);
2416 gfx_v12_0_me_fini(adev);
2417 return r;
2418 }
2419
2420 memcpy(adev->gfx.me.me_fw_ptr, fw_ucode, fw_ucode_size);
2421 memcpy(adev->gfx.me.me_fw_data_ptr, fw_data, fw_data_size);
2422
2423 amdgpu_bo_kunmap(adev->gfx.me.me_fw_obj);
2424 amdgpu_bo_kunmap(adev->gfx.me.me_fw_data_obj);
2425 amdgpu_bo_unreserve(adev->gfx.me.me_fw_obj);
2426 amdgpu_bo_unreserve(adev->gfx.me.me_fw_data_obj);
2427
2428 if (amdgpu_emu_mode == 1)
2429 amdgpu_device_flush_hdp(adev, NULL);
2430
2431 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_LO,
2432 lower_32_bits(adev->gfx.me.me_fw_gpu_addr));
2433 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_HI,
2434 upper_32_bits(adev->gfx.me.me_fw_gpu_addr));
2435
2436 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL);
2437 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, VMID, 0);
2438 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, CACHE_POLICY, 0);
2439 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, EXE_DISABLE, 0);
2440 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL, tmp);
2441
2442 /*
2443 * Programming any of the CP_ME_IC_BASE registers
2444 * forces invalidation of the ME L1 I$. Wait for the
2445 * invalidation complete
2446 */
2447 for (i = 0; i < usec_timeout; i++) {
2448 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL);
2449 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL,
2450 INVALIDATE_CACHE_COMPLETE))
2451 break;
2452 udelay(1);
2453 }
2454
2455 if (i >= usec_timeout) {
2456 dev_err(adev->dev, "failed to invalidate instruction cache\n");
2457 return -EINVAL;
2458 }
2459
2460 /* Prime the instruction caches */
2461 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL);
2462 tmp = REG_SET_FIELD(tmp, CP_ME_IC_OP_CNTL, PRIME_ICACHE, 1);
2463 WREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL, tmp);
2464
2465 /* Waiting for instruction cache primed*/
2466 for (i = 0; i < usec_timeout; i++) {
2467 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL);
2468 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL,
2469 ICACHE_PRIMED))
2470 break;
2471 udelay(1);
2472 }
2473
2474 if (i >= usec_timeout) {
2475 dev_err(adev->dev, "failed to prime instruction cache\n");
2476 return -EINVAL;
2477 }
2478
2479 mutex_lock(&adev->srbm_mutex);
2480 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) {
2481 soc24_grbm_select(adev, 0, pipe_id, 0, 0);
2482
2483 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_LO,
2484 lower_32_bits(adev->gfx.me.me_fw_data_gpu_addr));
2485 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_HI,
2486 upper_32_bits(adev->gfx.me.me_fw_data_gpu_addr));
2487 }
2488 soc24_grbm_select(adev, 0, 0, 0, 0);
2489 mutex_unlock(&adev->srbm_mutex);
2490
2491 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL);
2492 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0);
2493 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0);
2494 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp);
2495
2496 /* Invalidate the data caches */
2497 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL);
2498 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1);
2499 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp);
2500
2501 for (i = 0; i < usec_timeout; i++) {
2502 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL);
2503 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL,
2504 INVALIDATE_DCACHE_COMPLETE))
2505 break;
2506 udelay(1);
2507 }
2508
2509 if (i >= usec_timeout) {
2510 dev_err(adev->dev, "failed to invalidate RS64 data cache\n");
2511 return -EINVAL;
2512 }
2513
2514 gfx_v12_0_set_me_ucode_start_addr(adev);
2515
2516 return 0;
2517 }
2518
gfx_v12_0_cp_gfx_load_microcode(struct amdgpu_device * adev)2519 static int gfx_v12_0_cp_gfx_load_microcode(struct amdgpu_device *adev)
2520 {
2521 int r;
2522
2523 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw)
2524 return -EINVAL;
2525
2526 gfx_v12_0_cp_gfx_enable(adev, false);
2527
2528 r = gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(adev);
2529 if (r) {
2530 dev_err(adev->dev, "(%d) failed to load pfp fw\n", r);
2531 return r;
2532 }
2533
2534 r = gfx_v12_0_cp_gfx_load_me_microcode_rs64(adev);
2535 if (r) {
2536 dev_err(adev->dev, "(%d) failed to load me fw\n", r);
2537 return r;
2538 }
2539
2540 return 0;
2541 }
2542
gfx_v12_0_cp_gfx_start(struct amdgpu_device * adev)2543 static int gfx_v12_0_cp_gfx_start(struct amdgpu_device *adev)
2544 {
2545 /* init the CP */
2546 WREG32_SOC15(GC, 0, regCP_MAX_CONTEXT,
2547 adev->gfx.config.max_hw_contexts - 1);
2548 WREG32_SOC15(GC, 0, regCP_DEVICE_ID, 1);
2549
2550 if (!amdgpu_async_gfx_ring)
2551 gfx_v12_0_cp_gfx_enable(adev, true);
2552
2553 return 0;
2554 }
2555
gfx_v12_0_cp_gfx_switch_pipe(struct amdgpu_device * adev,CP_PIPE_ID pipe)2556 static void gfx_v12_0_cp_gfx_switch_pipe(struct amdgpu_device *adev,
2557 CP_PIPE_ID pipe)
2558 {
2559 u32 tmp;
2560
2561 tmp = RREG32_SOC15(GC, 0, regGRBM_GFX_CNTL);
2562 tmp = REG_SET_FIELD(tmp, GRBM_GFX_CNTL, PIPEID, pipe);
2563
2564 WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, tmp);
2565 }
2566
gfx_v12_0_cp_gfx_set_doorbell(struct amdgpu_device * adev,struct amdgpu_ring * ring)2567 static void gfx_v12_0_cp_gfx_set_doorbell(struct amdgpu_device *adev,
2568 struct amdgpu_ring *ring)
2569 {
2570 u32 tmp;
2571
2572 tmp = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL);
2573 if (ring->use_doorbell) {
2574 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2575 DOORBELL_OFFSET, ring->doorbell_index);
2576 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2577 DOORBELL_EN, 1);
2578 } else {
2579 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2580 DOORBELL_EN, 0);
2581 }
2582 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL, tmp);
2583
2584 tmp = REG_SET_FIELD(0, CP_RB_DOORBELL_RANGE_LOWER,
2585 DOORBELL_RANGE_LOWER, ring->doorbell_index);
2586 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, tmp);
2587
2588 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER,
2589 CP_RB_DOORBELL_RANGE_UPPER__DOORBELL_RANGE_UPPER_MASK);
2590 }
2591
gfx_v12_0_cp_gfx_resume(struct amdgpu_device * adev)2592 static int gfx_v12_0_cp_gfx_resume(struct amdgpu_device *adev)
2593 {
2594 struct amdgpu_ring *ring;
2595 u32 tmp;
2596 u32 rb_bufsz;
2597 u64 rb_addr, rptr_addr, wptr_gpu_addr;
2598 u32 i;
2599
2600 /* Set the write pointer delay */
2601 WREG32_SOC15(GC, 0, regCP_RB_WPTR_DELAY, 0);
2602
2603 /* set the RB to use vmid 0 */
2604 WREG32_SOC15(GC, 0, regCP_RB_VMID, 0);
2605
2606 /* Init gfx ring 0 for pipe 0 */
2607 mutex_lock(&adev->srbm_mutex);
2608 gfx_v12_0_cp_gfx_switch_pipe(adev, PIPE_ID0);
2609
2610 /* Set ring buffer size */
2611 ring = &adev->gfx.gfx_ring[0];
2612 rb_bufsz = order_base_2(ring->ring_size / 8);
2613 tmp = REG_SET_FIELD(0, CP_RB0_CNTL, RB_BUFSZ, rb_bufsz);
2614 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, RB_BLKSZ, rb_bufsz - 2);
2615 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp);
2616
2617 /* Initialize the ring buffer's write pointers */
2618 ring->wptr = 0;
2619 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, lower_32_bits(ring->wptr));
2620 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, upper_32_bits(ring->wptr));
2621
2622 /* set the wb address wether it's enabled or not */
2623 rptr_addr = ring->rptr_gpu_addr;
2624 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr));
2625 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) &
2626 CP_RB_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK);
2627
2628 wptr_gpu_addr = ring->wptr_gpu_addr;
2629 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO,
2630 lower_32_bits(wptr_gpu_addr));
2631 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI,
2632 upper_32_bits(wptr_gpu_addr));
2633
2634 mdelay(1);
2635 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp);
2636
2637 rb_addr = ring->gpu_addr >> 8;
2638 WREG32_SOC15(GC, 0, regCP_RB0_BASE, rb_addr);
2639 WREG32_SOC15(GC, 0, regCP_RB0_BASE_HI, upper_32_bits(rb_addr));
2640
2641 WREG32_SOC15(GC, 0, regCP_RB_ACTIVE, 1);
2642
2643 gfx_v12_0_cp_gfx_set_doorbell(adev, ring);
2644 mutex_unlock(&adev->srbm_mutex);
2645
2646 /* Switch to pipe 0 */
2647 mutex_lock(&adev->srbm_mutex);
2648 gfx_v12_0_cp_gfx_switch_pipe(adev, PIPE_ID0);
2649 mutex_unlock(&adev->srbm_mutex);
2650
2651 /* start the ring */
2652 gfx_v12_0_cp_gfx_start(adev);
2653
2654 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
2655 ring = &adev->gfx.gfx_ring[i];
2656 ring->sched.ready = true;
2657 }
2658
2659 return 0;
2660 }
2661
gfx_v12_0_cp_compute_enable(struct amdgpu_device * adev,bool enable)2662 static void gfx_v12_0_cp_compute_enable(struct amdgpu_device *adev, bool enable)
2663 {
2664 u32 data;
2665
2666 data = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL);
2667 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_INVALIDATE_ICACHE,
2668 enable ? 0 : 1);
2669 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET,
2670 enable ? 0 : 1);
2671 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET,
2672 enable ? 0 : 1);
2673 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET,
2674 enable ? 0 : 1);
2675 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET,
2676 enable ? 0 : 1);
2677 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_ACTIVE,
2678 enable ? 1 : 0);
2679 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_ACTIVE,
2680 enable ? 1 : 0);
2681 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_ACTIVE,
2682 enable ? 1 : 0);
2683 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_ACTIVE,
2684 enable ? 1 : 0);
2685 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_HALT,
2686 enable ? 0 : 1);
2687 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, data);
2688
2689 adev->gfx.kiq[0].ring.sched.ready = enable;
2690
2691 udelay(50);
2692 }
2693
gfx_v12_0_cp_compute_load_microcode_rs64(struct amdgpu_device * adev)2694 static int gfx_v12_0_cp_compute_load_microcode_rs64(struct amdgpu_device *adev)
2695 {
2696 const struct gfx_firmware_header_v2_0 *mec_hdr;
2697 const __le32 *fw_ucode, *fw_data;
2698 u32 tmp, fw_ucode_size, fw_data_size;
2699 u32 i, usec_timeout = 50000; /* Wait for 50 ms */
2700 u32 *fw_ucode_ptr, *fw_data_ptr;
2701 int r;
2702
2703 if (!adev->gfx.mec_fw)
2704 return -EINVAL;
2705
2706 gfx_v12_0_cp_compute_enable(adev, false);
2707
2708 mec_hdr = (const struct gfx_firmware_header_v2_0 *)adev->gfx.mec_fw->data;
2709 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header);
2710
2711 fw_ucode = (const __le32 *) (adev->gfx.mec_fw->data +
2712 le32_to_cpu(mec_hdr->ucode_offset_bytes));
2713 fw_ucode_size = le32_to_cpu(mec_hdr->ucode_size_bytes);
2714
2715 fw_data = (const __le32 *) (adev->gfx.mec_fw->data +
2716 le32_to_cpu(mec_hdr->data_offset_bytes));
2717 fw_data_size = le32_to_cpu(mec_hdr->data_size_bytes);
2718
2719 r = amdgpu_bo_create_reserved(adev, fw_ucode_size,
2720 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
2721 &adev->gfx.mec.mec_fw_obj,
2722 &adev->gfx.mec.mec_fw_gpu_addr,
2723 (void **)&fw_ucode_ptr);
2724 if (r) {
2725 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r);
2726 gfx_v12_0_mec_fini(adev);
2727 return r;
2728 }
2729
2730 r = amdgpu_bo_create_reserved(adev,
2731 ALIGN(fw_data_size, 64 * 1024) *
2732 adev->gfx.mec.num_pipe_per_mec,
2733 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM,
2734 &adev->gfx.mec.mec_fw_data_obj,
2735 &adev->gfx.mec.mec_fw_data_gpu_addr,
2736 (void **)&fw_data_ptr);
2737 if (r) {
2738 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r);
2739 gfx_v12_0_mec_fini(adev);
2740 return r;
2741 }
2742
2743 memcpy(fw_ucode_ptr, fw_ucode, fw_ucode_size);
2744 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) {
2745 memcpy(fw_data_ptr + i * ALIGN(fw_data_size, 64 * 1024) / 4, fw_data, fw_data_size);
2746 }
2747
2748 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj);
2749 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_data_obj);
2750 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj);
2751 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_data_obj);
2752
2753 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL);
2754 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0);
2755 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, EXE_DISABLE, 0);
2756 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0);
2757 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL, tmp);
2758
2759 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL);
2760 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, VMID, 0);
2761 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, CACHE_POLICY, 0);
2762 WREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL, tmp);
2763
2764 mutex_lock(&adev->srbm_mutex);
2765 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) {
2766 soc24_grbm_select(adev, 1, i, 0, 0);
2767
2768 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_LO,
2769 lower_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr +
2770 i * ALIGN(fw_data_size, 64 * 1024)));
2771 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_HI,
2772 upper_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr +
2773 i * ALIGN(fw_data_size, 64 * 1024)));
2774
2775 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_LO,
2776 lower_32_bits(adev->gfx.mec.mec_fw_gpu_addr));
2777 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_HI,
2778 upper_32_bits(adev->gfx.mec.mec_fw_gpu_addr));
2779 }
2780 mutex_unlock(&adev->srbm_mutex);
2781 soc24_grbm_select(adev, 0, 0, 0, 0);
2782
2783 /* Trigger an invalidation of the L1 instruction caches */
2784 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL);
2785 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_OP_CNTL, INVALIDATE_DCACHE, 1);
2786 WREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL, tmp);
2787
2788 /* Wait for invalidation complete */
2789 for (i = 0; i < usec_timeout; i++) {
2790 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL);
2791 if (1 == REG_GET_FIELD(tmp, CP_MEC_DC_OP_CNTL,
2792 INVALIDATE_DCACHE_COMPLETE))
2793 break;
2794 udelay(1);
2795 }
2796
2797 if (i >= usec_timeout) {
2798 dev_err(adev->dev, "failed to invalidate instruction cache\n");
2799 return -EINVAL;
2800 }
2801
2802 /* Trigger an invalidation of the L1 instruction caches */
2803 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL);
2804 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_OP_CNTL, INVALIDATE_CACHE, 1);
2805 WREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL, tmp);
2806
2807 /* Wait for invalidation complete */
2808 for (i = 0; i < usec_timeout; i++) {
2809 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL);
2810 if (1 == REG_GET_FIELD(tmp, CP_CPC_IC_OP_CNTL,
2811 INVALIDATE_CACHE_COMPLETE))
2812 break;
2813 udelay(1);
2814 }
2815
2816 if (i >= usec_timeout) {
2817 dev_err(adev->dev, "failed to invalidate instruction cache\n");
2818 return -EINVAL;
2819 }
2820
2821 gfx_v12_0_set_mec_ucode_start_addr(adev);
2822
2823 return 0;
2824 }
2825
gfx_v12_0_kiq_setting(struct amdgpu_ring * ring)2826 static void gfx_v12_0_kiq_setting(struct amdgpu_ring *ring)
2827 {
2828 uint32_t tmp;
2829 struct amdgpu_device *adev = ring->adev;
2830
2831 /* tell RLC which is KIQ queue */
2832 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
2833 tmp &= 0xffffff00;
2834 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
2835 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
2836 tmp |= 0x80;
2837 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
2838 }
2839
gfx_v12_0_cp_set_doorbell_range(struct amdgpu_device * adev)2840 static void gfx_v12_0_cp_set_doorbell_range(struct amdgpu_device *adev)
2841 {
2842 /* set graphics engine doorbell range */
2843 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER,
2844 (adev->doorbell_index.gfx_ring0 * 2) << 2);
2845 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER,
2846 (adev->doorbell_index.gfx_userqueue_end * 2) << 2);
2847
2848 /* set compute engine doorbell range */
2849 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER,
2850 (adev->doorbell_index.kiq * 2) << 2);
2851 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER,
2852 (adev->doorbell_index.userqueue_end * 2) << 2);
2853 }
2854
gfx_v12_0_gfx_mqd_init(struct amdgpu_device * adev,void * m,struct amdgpu_mqd_prop * prop)2855 static int gfx_v12_0_gfx_mqd_init(struct amdgpu_device *adev, void *m,
2856 struct amdgpu_mqd_prop *prop)
2857 {
2858 struct v12_gfx_mqd *mqd = m;
2859 uint64_t hqd_gpu_addr, wb_gpu_addr;
2860 uint32_t tmp;
2861 uint32_t rb_bufsz;
2862
2863 /* set up gfx hqd wptr */
2864 mqd->cp_gfx_hqd_wptr = 0;
2865 mqd->cp_gfx_hqd_wptr_hi = 0;
2866
2867 /* set the pointer to the MQD */
2868 mqd->cp_mqd_base_addr = prop->mqd_gpu_addr & 0xfffffffc;
2869 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr);
2870
2871 /* set up mqd control */
2872 tmp = regCP_GFX_MQD_CONTROL_DEFAULT;
2873 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, VMID, 0);
2874 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, PRIV_STATE, 1);
2875 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, CACHE_POLICY, 0);
2876 mqd->cp_gfx_mqd_control = tmp;
2877
2878 /* set up gfx_hqd_vimd with 0x0 to indicate the ring buffer's vmid */
2879 tmp = regCP_GFX_HQD_VMID_DEFAULT;
2880 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_VMID, VMID, 0);
2881 mqd->cp_gfx_hqd_vmid = 0;
2882
2883 /* set up default queue priority level
2884 * 0x0 = low priority, 0x1 = high priority */
2885 tmp = regCP_GFX_HQD_QUEUE_PRIORITY_DEFAULT;
2886 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUEUE_PRIORITY, PRIORITY_LEVEL, 0);
2887 mqd->cp_gfx_hqd_queue_priority = tmp;
2888
2889 /* set up time quantum */
2890 tmp = regCP_GFX_HQD_QUANTUM_DEFAULT;
2891 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUANTUM, QUANTUM_EN, 1);
2892 mqd->cp_gfx_hqd_quantum = tmp;
2893
2894 /* set up gfx hqd base. this is similar as CP_RB_BASE */
2895 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8;
2896 mqd->cp_gfx_hqd_base = hqd_gpu_addr;
2897 mqd->cp_gfx_hqd_base_hi = upper_32_bits(hqd_gpu_addr);
2898
2899 /* set up hqd_rptr_addr/_hi, similar as CP_RB_RPTR */
2900 wb_gpu_addr = prop->rptr_gpu_addr;
2901 mqd->cp_gfx_hqd_rptr_addr = wb_gpu_addr & 0xfffffffc;
2902 mqd->cp_gfx_hqd_rptr_addr_hi =
2903 upper_32_bits(wb_gpu_addr) & 0xffff;
2904
2905 /* set up rb_wptr_poll addr */
2906 wb_gpu_addr = prop->wptr_gpu_addr;
2907 mqd->cp_rb_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc;
2908 mqd->cp_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
2909
2910 /* set up the gfx_hqd_control, similar as CP_RB0_CNTL */
2911 rb_bufsz = order_base_2(prop->queue_size / 4) - 1;
2912 tmp = regCP_GFX_HQD_CNTL_DEFAULT;
2913 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BUFSZ, rb_bufsz);
2914 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BLKSZ, rb_bufsz - 2);
2915 #ifdef __BIG_ENDIAN
2916 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, BUF_SWAP, 1);
2917 #endif
2918 mqd->cp_gfx_hqd_cntl = tmp;
2919
2920 /* set up cp_doorbell_control */
2921 tmp = regCP_RB_DOORBELL_CONTROL_DEFAULT;
2922 if (prop->use_doorbell) {
2923 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2924 DOORBELL_OFFSET, prop->doorbell_index);
2925 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2926 DOORBELL_EN, 1);
2927 } else
2928 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
2929 DOORBELL_EN, 0);
2930 mqd->cp_rb_doorbell_control = tmp;
2931
2932 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
2933 mqd->cp_gfx_hqd_rptr = regCP_GFX_HQD_RPTR_DEFAULT;
2934
2935 /* active the queue */
2936 mqd->cp_gfx_hqd_active = 1;
2937
2938 return 0;
2939 }
2940
gfx_v12_0_kgq_init_queue(struct amdgpu_ring * ring,bool reset)2941 static int gfx_v12_0_kgq_init_queue(struct amdgpu_ring *ring, bool reset)
2942 {
2943 struct amdgpu_device *adev = ring->adev;
2944 struct v12_gfx_mqd *mqd = ring->mqd_ptr;
2945 int mqd_idx = ring - &adev->gfx.gfx_ring[0];
2946
2947 if (!reset && !amdgpu_in_reset(adev) && !adev->in_suspend) {
2948 memset((void *)mqd, 0, sizeof(*mqd));
2949 mutex_lock(&adev->srbm_mutex);
2950 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
2951 amdgpu_ring_init_mqd(ring);
2952 soc24_grbm_select(adev, 0, 0, 0, 0);
2953 mutex_unlock(&adev->srbm_mutex);
2954 if (adev->gfx.me.mqd_backup[mqd_idx])
2955 memcpy_fromio(adev->gfx.me.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
2956 } else {
2957 /* restore mqd with the backup copy */
2958 if (adev->gfx.me.mqd_backup[mqd_idx])
2959 memcpy_toio(mqd, adev->gfx.me.mqd_backup[mqd_idx], sizeof(*mqd));
2960 /* reset the ring */
2961 ring->wptr = 0;
2962 *ring->wptr_cpu_addr = 0;
2963 amdgpu_ring_clear_ring(ring);
2964 }
2965
2966 return 0;
2967 }
2968
gfx_v12_0_cp_async_gfx_ring_resume(struct amdgpu_device * adev)2969 static int gfx_v12_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev)
2970 {
2971 int r, i;
2972 struct amdgpu_ring *ring;
2973
2974 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
2975 ring = &adev->gfx.gfx_ring[i];
2976
2977 r = amdgpu_bo_reserve(ring->mqd_obj, false);
2978 if (unlikely(r != 0))
2979 goto done;
2980
2981 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
2982 if (!r) {
2983 r = gfx_v12_0_kgq_init_queue(ring, false);
2984 amdgpu_bo_kunmap(ring->mqd_obj);
2985 ring->mqd_ptr = NULL;
2986 }
2987 amdgpu_bo_unreserve(ring->mqd_obj);
2988 if (r)
2989 goto done;
2990 }
2991
2992 r = amdgpu_gfx_enable_kgq(adev, 0);
2993 if (r)
2994 goto done;
2995
2996 r = gfx_v12_0_cp_gfx_start(adev);
2997 if (r)
2998 goto done;
2999
3000 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
3001 ring = &adev->gfx.gfx_ring[i];
3002 ring->sched.ready = true;
3003 }
3004 done:
3005 return r;
3006 }
3007
gfx_v12_0_compute_mqd_init(struct amdgpu_device * adev,void * m,struct amdgpu_mqd_prop * prop)3008 static int gfx_v12_0_compute_mqd_init(struct amdgpu_device *adev, void *m,
3009 struct amdgpu_mqd_prop *prop)
3010 {
3011 struct v12_compute_mqd *mqd = m;
3012 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
3013 uint32_t tmp;
3014
3015 mqd->header = 0xC0310800;
3016 mqd->compute_pipelinestat_enable = 0x00000001;
3017 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
3018 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
3019 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
3020 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
3021 mqd->compute_misc_reserved = 0x00000007;
3022
3023 eop_base_addr = prop->eop_gpu_addr >> 8;
3024 mqd->cp_hqd_eop_base_addr_lo = eop_base_addr;
3025 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
3026
3027 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
3028 tmp = regCP_HQD_EOP_CONTROL_DEFAULT;
3029 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
3030 (order_base_2(GFX12_MEC_HPD_SIZE / 4) - 1));
3031
3032 mqd->cp_hqd_eop_control = tmp;
3033
3034 /* enable doorbell? */
3035 tmp = regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT;
3036
3037 if (prop->use_doorbell) {
3038 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3039 DOORBELL_OFFSET, prop->doorbell_index);
3040 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3041 DOORBELL_EN, 1);
3042 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3043 DOORBELL_SOURCE, 0);
3044 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3045 DOORBELL_HIT, 0);
3046 } else {
3047 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3048 DOORBELL_EN, 0);
3049 }
3050
3051 mqd->cp_hqd_pq_doorbell_control = tmp;
3052
3053 /* disable the queue if it's active */
3054 mqd->cp_hqd_dequeue_request = 0;
3055 mqd->cp_hqd_pq_rptr = 0;
3056 mqd->cp_hqd_pq_wptr_lo = 0;
3057 mqd->cp_hqd_pq_wptr_hi = 0;
3058
3059 /* set the pointer to the MQD */
3060 mqd->cp_mqd_base_addr_lo = prop->mqd_gpu_addr & 0xfffffffc;
3061 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr);
3062
3063 /* set MQD vmid to 0 */
3064 tmp = regCP_MQD_CONTROL_DEFAULT;
3065 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
3066 mqd->cp_mqd_control = tmp;
3067
3068 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
3069 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8;
3070 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr;
3071 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
3072
3073 /* set up the HQD, this is similar to CP_RB0_CNTL */
3074 tmp = regCP_HQD_PQ_CONTROL_DEFAULT;
3075 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
3076 (order_base_2(prop->queue_size / 4) - 1));
3077 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
3078 (order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1));
3079 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1);
3080 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0);
3081 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
3082 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
3083 mqd->cp_hqd_pq_control = tmp;
3084
3085 /* set the wb address whether it's enabled or not */
3086 wb_gpu_addr = prop->rptr_gpu_addr;
3087 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
3088 mqd->cp_hqd_pq_rptr_report_addr_hi =
3089 upper_32_bits(wb_gpu_addr) & 0xffff;
3090
3091 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
3092 wb_gpu_addr = prop->wptr_gpu_addr;
3093 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc;
3094 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
3095
3096 tmp = 0;
3097 /* enable the doorbell if requested */
3098 if (prop->use_doorbell) {
3099 tmp = regCP_HQD_PQ_DOORBELL_CONTROL_DEFAULT;
3100 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3101 DOORBELL_OFFSET, prop->doorbell_index);
3102
3103 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3104 DOORBELL_EN, 1);
3105 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3106 DOORBELL_SOURCE, 0);
3107 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3108 DOORBELL_HIT, 0);
3109 }
3110
3111 mqd->cp_hqd_pq_doorbell_control = tmp;
3112
3113 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
3114 mqd->cp_hqd_pq_rptr = regCP_HQD_PQ_RPTR_DEFAULT;
3115
3116 /* set the vmid for the queue */
3117 mqd->cp_hqd_vmid = 0;
3118
3119 tmp = regCP_HQD_PERSISTENT_STATE_DEFAULT;
3120 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x55);
3121 mqd->cp_hqd_persistent_state = tmp;
3122
3123 /* set MIN_IB_AVAIL_SIZE */
3124 tmp = regCP_HQD_IB_CONTROL_DEFAULT;
3125 tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3);
3126 mqd->cp_hqd_ib_control = tmp;
3127
3128 /* set static priority for a compute queue/ring */
3129 mqd->cp_hqd_pipe_priority = prop->hqd_pipe_priority;
3130 mqd->cp_hqd_queue_priority = prop->hqd_queue_priority;
3131
3132 mqd->cp_hqd_active = prop->hqd_active;
3133
3134 return 0;
3135 }
3136
gfx_v12_0_kiq_init_register(struct amdgpu_ring * ring)3137 static int gfx_v12_0_kiq_init_register(struct amdgpu_ring *ring)
3138 {
3139 struct amdgpu_device *adev = ring->adev;
3140 struct v12_compute_mqd *mqd = ring->mqd_ptr;
3141 int j;
3142
3143 /* inactivate the queue */
3144 if (amdgpu_sriov_vf(adev))
3145 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 0);
3146
3147 /* disable wptr polling */
3148 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0);
3149
3150 /* write the EOP addr */
3151 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR,
3152 mqd->cp_hqd_eop_base_addr_lo);
3153 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI,
3154 mqd->cp_hqd_eop_base_addr_hi);
3155
3156 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
3157 WREG32_SOC15(GC, 0, regCP_HQD_EOP_CONTROL,
3158 mqd->cp_hqd_eop_control);
3159
3160 /* enable doorbell? */
3161 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
3162 mqd->cp_hqd_pq_doorbell_control);
3163
3164 /* disable the queue if it's active */
3165 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) {
3166 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1);
3167 for (j = 0; j < adev->usec_timeout; j++) {
3168 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
3169 break;
3170 udelay(1);
3171 }
3172 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST,
3173 mqd->cp_hqd_dequeue_request);
3174 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR,
3175 mqd->cp_hqd_pq_rptr);
3176 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO,
3177 mqd->cp_hqd_pq_wptr_lo);
3178 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI,
3179 mqd->cp_hqd_pq_wptr_hi);
3180 }
3181
3182 /* set the pointer to the MQD */
3183 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR,
3184 mqd->cp_mqd_base_addr_lo);
3185 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI,
3186 mqd->cp_mqd_base_addr_hi);
3187
3188 /* set MQD vmid to 0 */
3189 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL,
3190 mqd->cp_mqd_control);
3191
3192 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
3193 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE,
3194 mqd->cp_hqd_pq_base_lo);
3195 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI,
3196 mqd->cp_hqd_pq_base_hi);
3197
3198 /* set up the HQD, this is similar to CP_RB0_CNTL */
3199 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL,
3200 mqd->cp_hqd_pq_control);
3201
3202 /* set the wb address whether it's enabled or not */
3203 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR,
3204 mqd->cp_hqd_pq_rptr_report_addr_lo);
3205 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
3206 mqd->cp_hqd_pq_rptr_report_addr_hi);
3207
3208 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
3209 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR,
3210 mqd->cp_hqd_pq_wptr_poll_addr_lo);
3211 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI,
3212 mqd->cp_hqd_pq_wptr_poll_addr_hi);
3213
3214 /* enable the doorbell if requested */
3215 if (ring->use_doorbell) {
3216 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER,
3217 (adev->doorbell_index.kiq * 2) << 2);
3218 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER,
3219 (adev->doorbell_index.userqueue_end * 2) << 2);
3220 }
3221
3222 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL,
3223 mqd->cp_hqd_pq_doorbell_control);
3224
3225 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
3226 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO,
3227 mqd->cp_hqd_pq_wptr_lo);
3228 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI,
3229 mqd->cp_hqd_pq_wptr_hi);
3230
3231 /* set the vmid for the queue */
3232 WREG32_SOC15(GC, 0, regCP_HQD_VMID, mqd->cp_hqd_vmid);
3233
3234 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE,
3235 mqd->cp_hqd_persistent_state);
3236
3237 /* activate the queue */
3238 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE,
3239 mqd->cp_hqd_active);
3240
3241 if (ring->use_doorbell)
3242 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1);
3243
3244 return 0;
3245 }
3246
gfx_v12_0_kiq_init_queue(struct amdgpu_ring * ring)3247 static int gfx_v12_0_kiq_init_queue(struct amdgpu_ring *ring)
3248 {
3249 struct amdgpu_device *adev = ring->adev;
3250 struct v12_compute_mqd *mqd = ring->mqd_ptr;
3251 int mqd_idx = AMDGPU_MAX_COMPUTE_RINGS;
3252
3253 gfx_v12_0_kiq_setting(ring);
3254
3255 if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */
3256 /* reset MQD to a clean status */
3257 if (adev->gfx.mec.mqd_backup[mqd_idx])
3258 memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd));
3259
3260 /* reset ring buffer */
3261 ring->wptr = 0;
3262 amdgpu_ring_clear_ring(ring);
3263
3264 mutex_lock(&adev->srbm_mutex);
3265 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
3266 gfx_v12_0_kiq_init_register(ring);
3267 soc24_grbm_select(adev, 0, 0, 0, 0);
3268 mutex_unlock(&adev->srbm_mutex);
3269 } else {
3270 memset((void *)mqd, 0, sizeof(*mqd));
3271 if (amdgpu_sriov_vf(adev) && adev->in_suspend)
3272 amdgpu_ring_clear_ring(ring);
3273 mutex_lock(&adev->srbm_mutex);
3274 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
3275 amdgpu_ring_init_mqd(ring);
3276 gfx_v12_0_kiq_init_register(ring);
3277 soc24_grbm_select(adev, 0, 0, 0, 0);
3278 mutex_unlock(&adev->srbm_mutex);
3279
3280 if (adev->gfx.mec.mqd_backup[mqd_idx])
3281 memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
3282 }
3283
3284 return 0;
3285 }
3286
gfx_v12_0_kcq_init_queue(struct amdgpu_ring * ring,bool reset)3287 static int gfx_v12_0_kcq_init_queue(struct amdgpu_ring *ring, bool reset)
3288 {
3289 struct amdgpu_device *adev = ring->adev;
3290 struct v12_compute_mqd *mqd = ring->mqd_ptr;
3291 int mqd_idx = ring - &adev->gfx.compute_ring[0];
3292
3293 if (!reset && !amdgpu_in_reset(adev) && !adev->in_suspend) {
3294 memset((void *)mqd, 0, sizeof(*mqd));
3295 mutex_lock(&adev->srbm_mutex);
3296 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
3297 amdgpu_ring_init_mqd(ring);
3298 soc24_grbm_select(adev, 0, 0, 0, 0);
3299 mutex_unlock(&adev->srbm_mutex);
3300
3301 if (adev->gfx.mec.mqd_backup[mqd_idx])
3302 memcpy_fromio(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd));
3303 } else {
3304 /* restore MQD to a clean status */
3305 if (adev->gfx.mec.mqd_backup[mqd_idx])
3306 memcpy_toio(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd));
3307 /* reset ring buffer */
3308 ring->wptr = 0;
3309 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0);
3310 amdgpu_ring_clear_ring(ring);
3311 }
3312
3313 return 0;
3314 }
3315
gfx_v12_0_kiq_resume(struct amdgpu_device * adev)3316 static int gfx_v12_0_kiq_resume(struct amdgpu_device *adev)
3317 {
3318 struct amdgpu_ring *ring;
3319 int r;
3320
3321 ring = &adev->gfx.kiq[0].ring;
3322
3323 r = amdgpu_bo_reserve(ring->mqd_obj, false);
3324 if (unlikely(r != 0))
3325 return r;
3326
3327 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
3328 if (unlikely(r != 0)) {
3329 amdgpu_bo_unreserve(ring->mqd_obj);
3330 return r;
3331 }
3332
3333 gfx_v12_0_kiq_init_queue(ring);
3334 amdgpu_bo_kunmap(ring->mqd_obj);
3335 ring->mqd_ptr = NULL;
3336 amdgpu_bo_unreserve(ring->mqd_obj);
3337 ring->sched.ready = true;
3338 return 0;
3339 }
3340
gfx_v12_0_kcq_resume(struct amdgpu_device * adev)3341 static int gfx_v12_0_kcq_resume(struct amdgpu_device *adev)
3342 {
3343 struct amdgpu_ring *ring = NULL;
3344 int r = 0, i;
3345
3346 if (!amdgpu_async_gfx_ring)
3347 gfx_v12_0_cp_compute_enable(adev, true);
3348
3349 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
3350 ring = &adev->gfx.compute_ring[i];
3351
3352 r = amdgpu_bo_reserve(ring->mqd_obj, false);
3353 if (unlikely(r != 0))
3354 goto done;
3355 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
3356 if (!r) {
3357 r = gfx_v12_0_kcq_init_queue(ring, false);
3358 amdgpu_bo_kunmap(ring->mqd_obj);
3359 ring->mqd_ptr = NULL;
3360 }
3361 amdgpu_bo_unreserve(ring->mqd_obj);
3362 if (r)
3363 goto done;
3364 }
3365
3366 r = amdgpu_gfx_enable_kcq(adev, 0);
3367 done:
3368 return r;
3369 }
3370
gfx_v12_0_cp_resume(struct amdgpu_device * adev)3371 static int gfx_v12_0_cp_resume(struct amdgpu_device *adev)
3372 {
3373 int r, i;
3374 struct amdgpu_ring *ring;
3375
3376 if (!(adev->flags & AMD_IS_APU))
3377 gfx_v12_0_enable_gui_idle_interrupt(adev, false);
3378
3379 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
3380 /* legacy firmware loading */
3381 r = gfx_v12_0_cp_gfx_load_microcode(adev);
3382 if (r)
3383 return r;
3384
3385 r = gfx_v12_0_cp_compute_load_microcode_rs64(adev);
3386 if (r)
3387 return r;
3388 }
3389
3390 gfx_v12_0_cp_set_doorbell_range(adev);
3391
3392 if (amdgpu_async_gfx_ring) {
3393 gfx_v12_0_cp_compute_enable(adev, true);
3394 gfx_v12_0_cp_gfx_enable(adev, true);
3395 }
3396
3397 if (adev->enable_mes_kiq && adev->mes.kiq_hw_init)
3398 r = amdgpu_mes_kiq_hw_init(adev);
3399 else
3400 r = gfx_v12_0_kiq_resume(adev);
3401 if (r)
3402 return r;
3403
3404 r = gfx_v12_0_kcq_resume(adev);
3405 if (r)
3406 return r;
3407
3408 if (!amdgpu_async_gfx_ring) {
3409 r = gfx_v12_0_cp_gfx_resume(adev);
3410 if (r)
3411 return r;
3412 } else {
3413 r = gfx_v12_0_cp_async_gfx_ring_resume(adev);
3414 if (r)
3415 return r;
3416 }
3417
3418 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
3419 ring = &adev->gfx.gfx_ring[i];
3420 r = amdgpu_ring_test_helper(ring);
3421 if (r)
3422 return r;
3423 }
3424
3425 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
3426 ring = &adev->gfx.compute_ring[i];
3427 r = amdgpu_ring_test_helper(ring);
3428 if (r)
3429 return r;
3430 }
3431
3432 return 0;
3433 }
3434
gfx_v12_0_cp_enable(struct amdgpu_device * adev,bool enable)3435 static void gfx_v12_0_cp_enable(struct amdgpu_device *adev, bool enable)
3436 {
3437 gfx_v12_0_cp_gfx_enable(adev, enable);
3438 gfx_v12_0_cp_compute_enable(adev, enable);
3439 }
3440
gfx_v12_0_gfxhub_enable(struct amdgpu_device * adev)3441 static int gfx_v12_0_gfxhub_enable(struct amdgpu_device *adev)
3442 {
3443 int r;
3444 bool value;
3445
3446 r = adev->gfxhub.funcs->gart_enable(adev);
3447 if (r)
3448 return r;
3449
3450 amdgpu_device_flush_hdp(adev, NULL);
3451
3452 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
3453 false : true;
3454
3455 adev->gfxhub.funcs->set_fault_enable_default(adev, value);
3456 /* TODO investigate why this and the hdp flush above is needed,
3457 * are we missing a flush somewhere else? */
3458 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB(0), 0);
3459
3460 return 0;
3461 }
3462
get_gb_addr_config(struct amdgpu_device * adev)3463 static int get_gb_addr_config(struct amdgpu_device *adev)
3464 {
3465 u32 gb_addr_config;
3466
3467 gb_addr_config = RREG32_SOC15(GC, 0, regGB_ADDR_CONFIG);
3468 if (gb_addr_config == 0)
3469 return -EINVAL;
3470
3471 adev->gfx.config.gb_addr_config_fields.num_pkrs =
3472 1 << REG_GET_FIELD(gb_addr_config, GB_ADDR_CONFIG, NUM_PKRS);
3473
3474 adev->gfx.config.gb_addr_config = gb_addr_config;
3475
3476 adev->gfx.config.gb_addr_config_fields.num_pipes = 1 <<
3477 REG_GET_FIELD(adev->gfx.config.gb_addr_config,
3478 GB_ADDR_CONFIG, NUM_PIPES);
3479
3480 adev->gfx.config.max_tile_pipes =
3481 adev->gfx.config.gb_addr_config_fields.num_pipes;
3482
3483 adev->gfx.config.gb_addr_config_fields.max_compress_frags = 1 <<
3484 REG_GET_FIELD(adev->gfx.config.gb_addr_config,
3485 GB_ADDR_CONFIG, MAX_COMPRESSED_FRAGS);
3486 adev->gfx.config.gb_addr_config_fields.num_rb_per_se = 1 <<
3487 REG_GET_FIELD(adev->gfx.config.gb_addr_config,
3488 GB_ADDR_CONFIG, NUM_RB_PER_SE);
3489 adev->gfx.config.gb_addr_config_fields.num_se = 1 <<
3490 REG_GET_FIELD(adev->gfx.config.gb_addr_config,
3491 GB_ADDR_CONFIG, NUM_SHADER_ENGINES);
3492 adev->gfx.config.gb_addr_config_fields.pipe_interleave_size = 1 << (8 +
3493 REG_GET_FIELD(adev->gfx.config.gb_addr_config,
3494 GB_ADDR_CONFIG, PIPE_INTERLEAVE_SIZE));
3495
3496 return 0;
3497 }
3498
gfx_v12_0_disable_gpa_mode(struct amdgpu_device * adev)3499 static void gfx_v12_0_disable_gpa_mode(struct amdgpu_device *adev)
3500 {
3501 uint32_t data;
3502
3503 data = RREG32_SOC15(GC, 0, regCPC_PSP_DEBUG);
3504 data |= CPC_PSP_DEBUG__GPA_OVERRIDE_MASK;
3505 WREG32_SOC15(GC, 0, regCPC_PSP_DEBUG, data);
3506
3507 data = RREG32_SOC15(GC, 0, regCPG_PSP_DEBUG);
3508 data |= CPG_PSP_DEBUG__GPA_OVERRIDE_MASK;
3509 WREG32_SOC15(GC, 0, regCPG_PSP_DEBUG, data);
3510 }
3511
gfx_v12_0_init_golden_registers(struct amdgpu_device * adev)3512 static void gfx_v12_0_init_golden_registers(struct amdgpu_device *adev)
3513 {
3514 if (amdgpu_sriov_vf(adev))
3515 return;
3516
3517 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
3518 case IP_VERSION(12, 0, 0):
3519 case IP_VERSION(12, 0, 1):
3520 soc15_program_register_sequence(adev,
3521 golden_settings_gc_12_0,
3522 (const u32)ARRAY_SIZE(golden_settings_gc_12_0));
3523
3524 if (adev->rev_id == 0)
3525 soc15_program_register_sequence(adev,
3526 golden_settings_gc_12_0_rev0,
3527 (const u32)ARRAY_SIZE(golden_settings_gc_12_0_rev0));
3528 break;
3529 default:
3530 break;
3531 }
3532 }
3533
gfx_v12_0_hw_init(void * handle)3534 static int gfx_v12_0_hw_init(void *handle)
3535 {
3536 int r;
3537 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3538
3539 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) {
3540 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) {
3541 /* RLC autoload sequence 1: Program rlc ram */
3542 if (adev->gfx.imu.funcs->program_rlc_ram)
3543 adev->gfx.imu.funcs->program_rlc_ram(adev);
3544 }
3545 /* rlc autoload firmware */
3546 r = gfx_v12_0_rlc_backdoor_autoload_enable(adev);
3547 if (r)
3548 return r;
3549 } else {
3550 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) {
3551 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) {
3552 if (adev->gfx.imu.funcs->load_microcode)
3553 adev->gfx.imu.funcs->load_microcode(adev);
3554 if (adev->gfx.imu.funcs->setup_imu)
3555 adev->gfx.imu.funcs->setup_imu(adev);
3556 if (adev->gfx.imu.funcs->start_imu)
3557 adev->gfx.imu.funcs->start_imu(adev);
3558 }
3559
3560 /* disable gpa mode in backdoor loading */
3561 gfx_v12_0_disable_gpa_mode(adev);
3562 }
3563 }
3564
3565 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) ||
3566 (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP)) {
3567 r = gfx_v12_0_wait_for_rlc_autoload_complete(adev);
3568 if (r) {
3569 dev_err(adev->dev, "(%d) failed to wait rlc autoload complete\n", r);
3570 return r;
3571 }
3572 }
3573
3574 if (!amdgpu_emu_mode)
3575 gfx_v12_0_init_golden_registers(adev);
3576
3577 adev->gfx.is_poweron = true;
3578
3579 if (get_gb_addr_config(adev))
3580 DRM_WARN("Invalid gb_addr_config !\n");
3581
3582 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP)
3583 gfx_v12_0_config_gfx_rs64(adev);
3584
3585 r = gfx_v12_0_gfxhub_enable(adev);
3586 if (r)
3587 return r;
3588
3589 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT ||
3590 adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) &&
3591 (amdgpu_dpm == 1)) {
3592 /**
3593 * For gfx 12, rlc firmware loading relies on smu firmware is
3594 * loaded firstly, so in direct type, it has to load smc ucode
3595 * here before rlc.
3596 */
3597 r = amdgpu_pm_load_smu_firmware(adev, NULL);
3598 if (r)
3599 return r;
3600 }
3601
3602 gfx_v12_0_constants_init(adev);
3603
3604 if (adev->nbio.funcs->gc_doorbell_init)
3605 adev->nbio.funcs->gc_doorbell_init(adev);
3606
3607 r = gfx_v12_0_rlc_resume(adev);
3608 if (r)
3609 return r;
3610
3611 /*
3612 * init golden registers and rlc resume may override some registers,
3613 * reconfig them here
3614 */
3615 gfx_v12_0_tcp_harvest(adev);
3616
3617 r = gfx_v12_0_cp_resume(adev);
3618 if (r)
3619 return r;
3620
3621 return r;
3622 }
3623
gfx_v12_0_hw_fini(void * handle)3624 static int gfx_v12_0_hw_fini(void *handle)
3625 {
3626 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3627 uint32_t tmp;
3628
3629 amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
3630 amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
3631 amdgpu_irq_put(adev, &adev->gfx.bad_op_irq, 0);
3632
3633 if (!adev->no_hw_access) {
3634 if (amdgpu_async_gfx_ring) {
3635 if (amdgpu_gfx_disable_kgq(adev, 0))
3636 DRM_ERROR("KGQ disable failed\n");
3637 }
3638
3639 if (amdgpu_gfx_disable_kcq(adev, 0))
3640 DRM_ERROR("KCQ disable failed\n");
3641
3642 amdgpu_mes_kiq_hw_fini(adev);
3643 }
3644
3645 if (amdgpu_sriov_vf(adev)) {
3646 gfx_v12_0_cp_gfx_enable(adev, false);
3647 /* Program KIQ position of RLC_CP_SCHEDULERS during destroy */
3648 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS);
3649 tmp &= 0xffffff00;
3650 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp);
3651
3652 return 0;
3653 }
3654 gfx_v12_0_cp_enable(adev, false);
3655 gfx_v12_0_enable_gui_idle_interrupt(adev, false);
3656
3657 adev->gfxhub.funcs->gart_disable(adev);
3658
3659 adev->gfx.is_poweron = false;
3660
3661 return 0;
3662 }
3663
gfx_v12_0_suspend(void * handle)3664 static int gfx_v12_0_suspend(void *handle)
3665 {
3666 return gfx_v12_0_hw_fini(handle);
3667 }
3668
gfx_v12_0_resume(void * handle)3669 static int gfx_v12_0_resume(void *handle)
3670 {
3671 return gfx_v12_0_hw_init(handle);
3672 }
3673
gfx_v12_0_is_idle(void * handle)3674 static bool gfx_v12_0_is_idle(void *handle)
3675 {
3676 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3677
3678 if (REG_GET_FIELD(RREG32_SOC15(GC, 0, regGRBM_STATUS),
3679 GRBM_STATUS, GUI_ACTIVE))
3680 return false;
3681 else
3682 return true;
3683 }
3684
gfx_v12_0_wait_for_idle(void * handle)3685 static int gfx_v12_0_wait_for_idle(void *handle)
3686 {
3687 unsigned i;
3688 u32 tmp;
3689 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3690
3691 for (i = 0; i < adev->usec_timeout; i++) {
3692 /* read MC_STATUS */
3693 tmp = RREG32_SOC15(GC, 0, regGRBM_STATUS) &
3694 GRBM_STATUS__GUI_ACTIVE_MASK;
3695
3696 if (!REG_GET_FIELD(tmp, GRBM_STATUS, GUI_ACTIVE))
3697 return 0;
3698 udelay(1);
3699 }
3700 return -ETIMEDOUT;
3701 }
3702
gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device * adev)3703 static uint64_t gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device *adev)
3704 {
3705 uint64_t clock = 0;
3706
3707 if (adev->smuio.funcs &&
3708 adev->smuio.funcs->get_gpu_clock_counter)
3709 clock = adev->smuio.funcs->get_gpu_clock_counter(adev);
3710 else
3711 dev_warn(adev->dev, "query gpu clock counter is not supported\n");
3712
3713 return clock;
3714 }
3715
gfx_v12_0_early_init(void * handle)3716 static int gfx_v12_0_early_init(void *handle)
3717 {
3718 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3719
3720 adev->gfx.funcs = &gfx_v12_0_gfx_funcs;
3721
3722 adev->gfx.num_gfx_rings = GFX12_NUM_GFX_RINGS;
3723 adev->gfx.num_compute_rings = min(amdgpu_gfx_get_num_kcq(adev),
3724 AMDGPU_MAX_COMPUTE_RINGS);
3725
3726 gfx_v12_0_set_kiq_pm4_funcs(adev);
3727 gfx_v12_0_set_ring_funcs(adev);
3728 gfx_v12_0_set_irq_funcs(adev);
3729 gfx_v12_0_set_rlc_funcs(adev);
3730 gfx_v12_0_set_mqd_funcs(adev);
3731 gfx_v12_0_set_imu_funcs(adev);
3732
3733 gfx_v12_0_init_rlcg_reg_access_ctrl(adev);
3734
3735 return gfx_v12_0_init_microcode(adev);
3736 }
3737
gfx_v12_0_late_init(void * handle)3738 static int gfx_v12_0_late_init(void *handle)
3739 {
3740 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3741 int r;
3742
3743 r = amdgpu_irq_get(adev, &adev->gfx.priv_reg_irq, 0);
3744 if (r)
3745 return r;
3746
3747 r = amdgpu_irq_get(adev, &adev->gfx.priv_inst_irq, 0);
3748 if (r)
3749 return r;
3750
3751 r = amdgpu_irq_get(adev, &adev->gfx.bad_op_irq, 0);
3752 if (r)
3753 return r;
3754
3755 return 0;
3756 }
3757
gfx_v12_0_is_rlc_enabled(struct amdgpu_device * adev)3758 static bool gfx_v12_0_is_rlc_enabled(struct amdgpu_device *adev)
3759 {
3760 uint32_t rlc_cntl;
3761
3762 /* if RLC is not enabled, do nothing */
3763 rlc_cntl = RREG32_SOC15(GC, 0, regRLC_CNTL);
3764 return (REG_GET_FIELD(rlc_cntl, RLC_CNTL, RLC_ENABLE_F32)) ? true : false;
3765 }
3766
gfx_v12_0_set_safe_mode(struct amdgpu_device * adev,int xcc_id)3767 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev,
3768 int xcc_id)
3769 {
3770 uint32_t data;
3771 unsigned i;
3772
3773 data = RLC_SAFE_MODE__CMD_MASK;
3774 data |= (1 << RLC_SAFE_MODE__MESSAGE__SHIFT);
3775
3776 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, data);
3777
3778 /* wait for RLC_SAFE_MODE */
3779 for (i = 0; i < adev->usec_timeout; i++) {
3780 if (!REG_GET_FIELD(RREG32_SOC15(GC, 0, regRLC_SAFE_MODE),
3781 RLC_SAFE_MODE, CMD))
3782 break;
3783 udelay(1);
3784 }
3785 }
3786
gfx_v12_0_unset_safe_mode(struct amdgpu_device * adev,int xcc_id)3787 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev,
3788 int xcc_id)
3789 {
3790 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, RLC_SAFE_MODE__CMD_MASK);
3791 }
3792
gfx_v12_0_update_perf_clk(struct amdgpu_device * adev,bool enable)3793 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev,
3794 bool enable)
3795 {
3796 uint32_t def, data;
3797
3798 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_PERF_CLK))
3799 return;
3800
3801 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
3802
3803 if (enable)
3804 data &= ~RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK;
3805 else
3806 data |= RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK;
3807
3808 if (def != data)
3809 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data);
3810 }
3811
gfx_v12_0_update_spm_vmid(struct amdgpu_device * adev,struct amdgpu_ring * ring,unsigned vmid)3812 static void gfx_v12_0_update_spm_vmid(struct amdgpu_device *adev,
3813 struct amdgpu_ring *ring,
3814 unsigned vmid)
3815 {
3816 u32 reg, data;
3817
3818 reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL);
3819 if (amdgpu_sriov_is_pp_one_vf(adev))
3820 data = RREG32_NO_KIQ(reg);
3821 else
3822 data = RREG32(reg);
3823
3824 data &= ~RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK;
3825 data |= (vmid & RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK) << RLC_SPM_MC_CNTL__RLC_SPM_VMID__SHIFT;
3826
3827 if (amdgpu_sriov_is_pp_one_vf(adev))
3828 WREG32_SOC15_NO_KIQ(GC, 0, regRLC_SPM_MC_CNTL, data);
3829 else
3830 WREG32_SOC15(GC, 0, regRLC_SPM_MC_CNTL, data);
3831
3832 if (ring
3833 && amdgpu_sriov_is_pp_one_vf(adev)
3834 && ((ring->funcs->type == AMDGPU_RING_TYPE_GFX)
3835 || (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE))) {
3836 uint32_t reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL);
3837 amdgpu_ring_emit_wreg(ring, reg, data);
3838 }
3839 }
3840
3841 static const struct amdgpu_rlc_funcs gfx_v12_0_rlc_funcs = {
3842 .is_rlc_enabled = gfx_v12_0_is_rlc_enabled,
3843 .set_safe_mode = gfx_v12_0_set_safe_mode,
3844 .unset_safe_mode = gfx_v12_0_unset_safe_mode,
3845 .init = gfx_v12_0_rlc_init,
3846 .get_csb_size = gfx_v12_0_get_csb_size,
3847 .get_csb_buffer = gfx_v12_0_get_csb_buffer,
3848 .resume = gfx_v12_0_rlc_resume,
3849 .stop = gfx_v12_0_rlc_stop,
3850 .reset = gfx_v12_0_rlc_reset,
3851 .start = gfx_v12_0_rlc_start,
3852 .update_spm_vmid = gfx_v12_0_update_spm_vmid,
3853 };
3854
3855 #if 0
3856 static void gfx_v12_cntl_power_gating(struct amdgpu_device *adev, bool enable)
3857 {
3858 /* TODO */
3859 }
3860
3861 static void gfx_v12_cntl_pg(struct amdgpu_device *adev, bool enable)
3862 {
3863 /* TODO */
3864 }
3865 #endif
3866
gfx_v12_0_set_powergating_state(void * handle,enum amd_powergating_state state)3867 static int gfx_v12_0_set_powergating_state(void *handle,
3868 enum amd_powergating_state state)
3869 {
3870 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
3871 bool enable = (state == AMD_PG_STATE_GATE);
3872
3873 if (amdgpu_sriov_vf(adev))
3874 return 0;
3875
3876 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
3877 case IP_VERSION(12, 0, 0):
3878 case IP_VERSION(12, 0, 1):
3879 amdgpu_gfx_off_ctrl(adev, enable);
3880 break;
3881 default:
3882 break;
3883 }
3884
3885 return 0;
3886 }
3887
gfx_v12_0_update_coarse_grain_clock_gating(struct amdgpu_device * adev,bool enable)3888 static void gfx_v12_0_update_coarse_grain_clock_gating(struct amdgpu_device *adev,
3889 bool enable)
3890 {
3891 uint32_t def, data;
3892
3893 if (!(adev->cg_flags &
3894 (AMD_CG_SUPPORT_GFX_CGCG |
3895 AMD_CG_SUPPORT_GFX_CGLS |
3896 AMD_CG_SUPPORT_GFX_3D_CGCG |
3897 AMD_CG_SUPPORT_GFX_3D_CGLS)))
3898 return;
3899
3900 if (enable) {
3901 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
3902
3903 /* unset CGCG override */
3904 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG)
3905 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGCG_OVERRIDE_MASK;
3906 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS)
3907 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK;
3908 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG ||
3909 adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS)
3910 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_GFX3D_CG_OVERRIDE_MASK;
3911
3912 /* update CGCG override bits */
3913 if (def != data)
3914 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data);
3915
3916 /* enable cgcg FSM(0x0000363F) */
3917 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL);
3918
3919 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) {
3920 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD_MASK;
3921 data |= (0x36 << RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD__SHIFT) |
3922 RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK;
3923 }
3924
3925 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) {
3926 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY_MASK;
3927 data |= (0x000F << RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY__SHIFT) |
3928 RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK;
3929 }
3930
3931 if (def != data)
3932 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data);
3933
3934 /* Program RLC_CGCG_CGLS_CTRL_3D */
3935 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D);
3936
3937 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) {
3938 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD_MASK;
3939 data |= (0x36 << RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD__SHIFT) |
3940 RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK;
3941 }
3942
3943 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) {
3944 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY_MASK;
3945 data |= (0xf << RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY__SHIFT) |
3946 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK;
3947 }
3948
3949 if (def != data)
3950 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data);
3951
3952 /* set IDLE_POLL_COUNT(0x00900100) */
3953 def = data = RREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL);
3954
3955 data &= ~(CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY_MASK | CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT_MASK);
3956 data |= (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) |
3957 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
3958
3959 if (def != data)
3960 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL, data);
3961
3962 data = RREG32_SOC15(GC, 0, regCP_INT_CNTL);
3963 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_BUSY_INT_ENABLE, 1);
3964 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_EMPTY_INT_ENABLE, 1);
3965 data = REG_SET_FIELD(data, CP_INT_CNTL, CMP_BUSY_INT_ENABLE, 1);
3966 data = REG_SET_FIELD(data, CP_INT_CNTL, GFX_IDLE_INT_ENABLE, 1);
3967 WREG32_SOC15(GC, 0, regCP_INT_CNTL, data);
3968
3969 data = RREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL);
3970 data = REG_SET_FIELD(data, SDMA0_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1);
3971 WREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL, data);
3972
3973 /* Some ASICs only have one SDMA instance, not need to configure SDMA1 */
3974 if (adev->sdma.num_instances > 1) {
3975 data = RREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL);
3976 data = REG_SET_FIELD(data, SDMA1_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1);
3977 WREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL, data);
3978 }
3979 } else {
3980 /* Program RLC_CGCG_CGLS_CTRL */
3981 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL);
3982
3983 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG)
3984 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK;
3985
3986 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS)
3987 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK;
3988
3989 if (def != data)
3990 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data);
3991
3992 /* Program RLC_CGCG_CGLS_CTRL_3D */
3993 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D);
3994
3995 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG)
3996 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK;
3997 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS)
3998 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK;
3999
4000 if (def != data)
4001 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data);
4002 }
4003 }
4004
gfx_v12_0_update_medium_grain_clock_gating(struct amdgpu_device * adev,bool enable)4005 static void gfx_v12_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
4006 bool enable)
4007 {
4008 uint32_t data, def;
4009 if (!(adev->cg_flags & (AMD_CG_SUPPORT_GFX_MGCG | AMD_CG_SUPPORT_GFX_MGLS)))
4010 return;
4011
4012 /* It is disabled by HW by default */
4013 if (enable) {
4014 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) {
4015 /* 1 - RLC_CGTT_MGCG_OVERRIDE */
4016 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
4017
4018 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK |
4019 RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK |
4020 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK);
4021
4022 if (def != data)
4023 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data);
4024 }
4025 } else {
4026 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) {
4027 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
4028
4029 data |= (RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK |
4030 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK |
4031 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK);
4032
4033 if (def != data)
4034 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data);
4035 }
4036 }
4037 }
4038
gfx_v12_0_update_repeater_fgcg(struct amdgpu_device * adev,bool enable)4039 static void gfx_v12_0_update_repeater_fgcg(struct amdgpu_device *adev,
4040 bool enable)
4041 {
4042 uint32_t def, data;
4043
4044 if (!(adev->cg_flags & AMD_CG_SUPPORT_REPEATER_FGCG))
4045 return;
4046
4047 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
4048
4049 if (enable)
4050 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK |
4051 RLC_CGTT_MGCG_OVERRIDE__RLC_REPEATER_FGCG_OVERRIDE_MASK);
4052 else
4053 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK |
4054 RLC_CGTT_MGCG_OVERRIDE__RLC_REPEATER_FGCG_OVERRIDE_MASK;
4055
4056 if (def != data)
4057 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data);
4058 }
4059
gfx_v12_0_update_sram_fgcg(struct amdgpu_device * adev,bool enable)4060 static void gfx_v12_0_update_sram_fgcg(struct amdgpu_device *adev,
4061 bool enable)
4062 {
4063 uint32_t def, data;
4064
4065 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_FGCG))
4066 return;
4067
4068 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
4069
4070 if (enable)
4071 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK;
4072 else
4073 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK;
4074
4075 if (def != data)
4076 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data);
4077 }
4078
gfx_v12_0_update_gfx_clock_gating(struct amdgpu_device * adev,bool enable)4079 static int gfx_v12_0_update_gfx_clock_gating(struct amdgpu_device *adev,
4080 bool enable)
4081 {
4082 amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
4083
4084 gfx_v12_0_update_coarse_grain_clock_gating(adev, enable);
4085
4086 gfx_v12_0_update_medium_grain_clock_gating(adev, enable);
4087
4088 gfx_v12_0_update_repeater_fgcg(adev, enable);
4089
4090 gfx_v12_0_update_sram_fgcg(adev, enable);
4091
4092 gfx_v12_0_update_perf_clk(adev, enable);
4093
4094 if (adev->cg_flags &
4095 (AMD_CG_SUPPORT_GFX_MGCG |
4096 AMD_CG_SUPPORT_GFX_CGLS |
4097 AMD_CG_SUPPORT_GFX_CGCG |
4098 AMD_CG_SUPPORT_GFX_3D_CGCG |
4099 AMD_CG_SUPPORT_GFX_3D_CGLS))
4100 gfx_v12_0_enable_gui_idle_interrupt(adev, enable);
4101
4102 amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
4103
4104 return 0;
4105 }
4106
gfx_v12_0_set_clockgating_state(void * handle,enum amd_clockgating_state state)4107 static int gfx_v12_0_set_clockgating_state(void *handle,
4108 enum amd_clockgating_state state)
4109 {
4110 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4111
4112 if (amdgpu_sriov_vf(adev))
4113 return 0;
4114
4115 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
4116 case IP_VERSION(12, 0, 0):
4117 case IP_VERSION(12, 0, 1):
4118 gfx_v12_0_update_gfx_clock_gating(adev,
4119 state == AMD_CG_STATE_GATE);
4120 break;
4121 default:
4122 break;
4123 }
4124
4125 return 0;
4126 }
4127
gfx_v12_0_get_clockgating_state(void * handle,u64 * flags)4128 static void gfx_v12_0_get_clockgating_state(void *handle, u64 *flags)
4129 {
4130 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
4131 int data;
4132
4133 /* AMD_CG_SUPPORT_GFX_MGCG */
4134 data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE);
4135 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK))
4136 *flags |= AMD_CG_SUPPORT_GFX_MGCG;
4137
4138 /* AMD_CG_SUPPORT_REPEATER_FGCG */
4139 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK))
4140 *flags |= AMD_CG_SUPPORT_REPEATER_FGCG;
4141
4142 /* AMD_CG_SUPPORT_GFX_FGCG */
4143 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK))
4144 *flags |= AMD_CG_SUPPORT_GFX_FGCG;
4145
4146 /* AMD_CG_SUPPORT_GFX_PERF_CLK */
4147 if (!(data & RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK))
4148 *flags |= AMD_CG_SUPPORT_GFX_PERF_CLK;
4149
4150 /* AMD_CG_SUPPORT_GFX_CGCG */
4151 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL);
4152 if (data & RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK)
4153 *flags |= AMD_CG_SUPPORT_GFX_CGCG;
4154
4155 /* AMD_CG_SUPPORT_GFX_CGLS */
4156 if (data & RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK)
4157 *flags |= AMD_CG_SUPPORT_GFX_CGLS;
4158
4159 /* AMD_CG_SUPPORT_GFX_3D_CGCG */
4160 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D);
4161 if (data & RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK)
4162 *flags |= AMD_CG_SUPPORT_GFX_3D_CGCG;
4163
4164 /* AMD_CG_SUPPORT_GFX_3D_CGLS */
4165 if (data & RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK)
4166 *flags |= AMD_CG_SUPPORT_GFX_3D_CGLS;
4167 }
4168
gfx_v12_0_ring_get_rptr_gfx(struct amdgpu_ring * ring)4169 static u64 gfx_v12_0_ring_get_rptr_gfx(struct amdgpu_ring *ring)
4170 {
4171 /* gfx12 is 32bit rptr*/
4172 return *(uint32_t *)ring->rptr_cpu_addr;
4173 }
4174
gfx_v12_0_ring_get_wptr_gfx(struct amdgpu_ring * ring)4175 static u64 gfx_v12_0_ring_get_wptr_gfx(struct amdgpu_ring *ring)
4176 {
4177 struct amdgpu_device *adev = ring->adev;
4178 u64 wptr;
4179
4180 /* XXX check if swapping is necessary on BE */
4181 if (ring->use_doorbell) {
4182 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
4183 } else {
4184 wptr = RREG32_SOC15(GC, 0, regCP_RB0_WPTR);
4185 wptr += (u64)RREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI) << 32;
4186 }
4187
4188 return wptr;
4189 }
4190
gfx_v12_0_ring_set_wptr_gfx(struct amdgpu_ring * ring)4191 static void gfx_v12_0_ring_set_wptr_gfx(struct amdgpu_ring *ring)
4192 {
4193 struct amdgpu_device *adev = ring->adev;
4194 uint32_t *wptr_saved;
4195 uint32_t *is_queue_unmap;
4196 uint64_t aggregated_db_index;
4197 uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_GFX].mqd_size;
4198 uint64_t wptr_tmp;
4199
4200 if (ring->is_mes_queue) {
4201 wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size);
4202 is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size +
4203 sizeof(uint32_t));
4204 aggregated_db_index =
4205 amdgpu_mes_get_aggregated_doorbell_index(adev,
4206 ring->hw_prio);
4207
4208 wptr_tmp = ring->wptr & ring->buf_mask;
4209 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, wptr_tmp);
4210 *wptr_saved = wptr_tmp;
4211 /* assume doorbell always being used by mes mapped queue */
4212 if (*is_queue_unmap) {
4213 WDOORBELL64(aggregated_db_index, wptr_tmp);
4214 WDOORBELL64(ring->doorbell_index, wptr_tmp);
4215 } else {
4216 WDOORBELL64(ring->doorbell_index, wptr_tmp);
4217
4218 if (*is_queue_unmap)
4219 WDOORBELL64(aggregated_db_index, wptr_tmp);
4220 }
4221 } else {
4222 if (ring->use_doorbell) {
4223 /* XXX check if swapping is necessary on BE */
4224 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
4225 ring->wptr);
4226 WDOORBELL64(ring->doorbell_index, ring->wptr);
4227 } else {
4228 WREG32_SOC15(GC, 0, regCP_RB0_WPTR,
4229 lower_32_bits(ring->wptr));
4230 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI,
4231 upper_32_bits(ring->wptr));
4232 }
4233 }
4234 }
4235
gfx_v12_0_ring_get_rptr_compute(struct amdgpu_ring * ring)4236 static u64 gfx_v12_0_ring_get_rptr_compute(struct amdgpu_ring *ring)
4237 {
4238 /* gfx12 hardware is 32bit rptr */
4239 return *(uint32_t *)ring->rptr_cpu_addr;
4240 }
4241
gfx_v12_0_ring_get_wptr_compute(struct amdgpu_ring * ring)4242 static u64 gfx_v12_0_ring_get_wptr_compute(struct amdgpu_ring *ring)
4243 {
4244 u64 wptr;
4245
4246 /* XXX check if swapping is necessary on BE */
4247 if (ring->use_doorbell)
4248 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr);
4249 else
4250 BUG();
4251 return wptr;
4252 }
4253
gfx_v12_0_ring_set_wptr_compute(struct amdgpu_ring * ring)4254 static void gfx_v12_0_ring_set_wptr_compute(struct amdgpu_ring *ring)
4255 {
4256 struct amdgpu_device *adev = ring->adev;
4257 uint32_t *wptr_saved;
4258 uint32_t *is_queue_unmap;
4259 uint64_t aggregated_db_index;
4260 uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size;
4261 uint64_t wptr_tmp;
4262
4263 if (ring->is_mes_queue) {
4264 wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size);
4265 is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size +
4266 sizeof(uint32_t));
4267 aggregated_db_index =
4268 amdgpu_mes_get_aggregated_doorbell_index(adev,
4269 ring->hw_prio);
4270
4271 wptr_tmp = ring->wptr & ring->buf_mask;
4272 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, wptr_tmp);
4273 *wptr_saved = wptr_tmp;
4274 /* assume doorbell always used by mes mapped queue */
4275 if (*is_queue_unmap) {
4276 WDOORBELL64(aggregated_db_index, wptr_tmp);
4277 WDOORBELL64(ring->doorbell_index, wptr_tmp);
4278 } else {
4279 WDOORBELL64(ring->doorbell_index, wptr_tmp);
4280
4281 if (*is_queue_unmap)
4282 WDOORBELL64(aggregated_db_index, wptr_tmp);
4283 }
4284 } else {
4285 /* XXX check if swapping is necessary on BE */
4286 if (ring->use_doorbell) {
4287 atomic64_set((atomic64_t *)ring->wptr_cpu_addr,
4288 ring->wptr);
4289 WDOORBELL64(ring->doorbell_index, ring->wptr);
4290 } else {
4291 BUG(); /* only DOORBELL method supported on gfx12 now */
4292 }
4293 }
4294 }
4295
gfx_v12_0_ring_emit_hdp_flush(struct amdgpu_ring * ring)4296 static void gfx_v12_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
4297 {
4298 struct amdgpu_device *adev = ring->adev;
4299 u32 ref_and_mask, reg_mem_engine;
4300 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg;
4301
4302 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) {
4303 switch (ring->me) {
4304 case 1:
4305 ref_and_mask = nbio_hf_reg->ref_and_mask_cp2 << ring->pipe;
4306 break;
4307 case 2:
4308 ref_and_mask = nbio_hf_reg->ref_and_mask_cp6 << ring->pipe;
4309 break;
4310 default:
4311 return;
4312 }
4313 reg_mem_engine = 0;
4314 } else {
4315 ref_and_mask = nbio_hf_reg->ref_and_mask_cp0;
4316 reg_mem_engine = 1; /* pfp */
4317 }
4318
4319 gfx_v12_0_wait_reg_mem(ring, reg_mem_engine, 0, 1,
4320 adev->nbio.funcs->get_hdp_flush_req_offset(adev),
4321 adev->nbio.funcs->get_hdp_flush_done_offset(adev),
4322 ref_and_mask, ref_and_mask, 0x20);
4323 }
4324
gfx_v12_0_ring_emit_ib_gfx(struct amdgpu_ring * ring,struct amdgpu_job * job,struct amdgpu_ib * ib,uint32_t flags)4325 static void gfx_v12_0_ring_emit_ib_gfx(struct amdgpu_ring *ring,
4326 struct amdgpu_job *job,
4327 struct amdgpu_ib *ib,
4328 uint32_t flags)
4329 {
4330 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
4331 u32 header, control = 0;
4332
4333 BUG_ON(ib->flags & AMDGPU_IB_FLAG_CE);
4334
4335 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
4336
4337 control |= ib->length_dw | (vmid << 24);
4338
4339 if (ring->is_mes_queue)
4340 /* inherit vmid from mqd */
4341 control |= 0x400000;
4342
4343 amdgpu_ring_write(ring, header);
4344 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */
4345 amdgpu_ring_write(ring,
4346 #ifdef __BIG_ENDIAN
4347 (2 << 0) |
4348 #endif
4349 lower_32_bits(ib->gpu_addr));
4350 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
4351 amdgpu_ring_write(ring, control);
4352 }
4353
gfx_v12_0_ring_emit_ib_compute(struct amdgpu_ring * ring,struct amdgpu_job * job,struct amdgpu_ib * ib,uint32_t flags)4354 static void gfx_v12_0_ring_emit_ib_compute(struct amdgpu_ring *ring,
4355 struct amdgpu_job *job,
4356 struct amdgpu_ib *ib,
4357 uint32_t flags)
4358 {
4359 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
4360 u32 control = INDIRECT_BUFFER_VALID | ib->length_dw | (vmid << 24);
4361
4362 if (ring->is_mes_queue)
4363 /* inherit vmid from mqd */
4364 control |= 0x40000000;
4365
4366 amdgpu_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
4367 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */
4368 amdgpu_ring_write(ring,
4369 #ifdef __BIG_ENDIAN
4370 (2 << 0) |
4371 #endif
4372 lower_32_bits(ib->gpu_addr));
4373 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
4374 amdgpu_ring_write(ring, control);
4375 }
4376
gfx_v12_0_ring_emit_fence(struct amdgpu_ring * ring,u64 addr,u64 seq,unsigned flags)4377 static void gfx_v12_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
4378 u64 seq, unsigned flags)
4379 {
4380 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
4381 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
4382
4383 /* RELEASE_MEM - flush caches, send int */
4384 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
4385 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ |
4386 PACKET3_RELEASE_MEM_GCR_GL2_WB |
4387 PACKET3_RELEASE_MEM_CACHE_POLICY(3) |
4388 PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
4389 PACKET3_RELEASE_MEM_EVENT_INDEX(5)));
4390 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ? 2 : 1) |
4391 PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 : 0)));
4392
4393 /*
4394 * the address should be Qword aligned if 64bit write, Dword
4395 * aligned if only send 32bit data low (discard data high)
4396 */
4397 if (write64bit)
4398 BUG_ON(addr & 0x7);
4399 else
4400 BUG_ON(addr & 0x3);
4401 amdgpu_ring_write(ring, lower_32_bits(addr));
4402 amdgpu_ring_write(ring, upper_32_bits(addr));
4403 amdgpu_ring_write(ring, lower_32_bits(seq));
4404 amdgpu_ring_write(ring, upper_32_bits(seq));
4405 amdgpu_ring_write(ring, ring->is_mes_queue ?
4406 (ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0);
4407 }
4408
gfx_v12_0_ring_emit_pipeline_sync(struct amdgpu_ring * ring)4409 static void gfx_v12_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
4410 {
4411 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
4412 uint32_t seq = ring->fence_drv.sync_seq;
4413 uint64_t addr = ring->fence_drv.gpu_addr;
4414
4415 gfx_v12_0_wait_reg_mem(ring, usepfp, 1, 0, lower_32_bits(addr),
4416 upper_32_bits(addr), seq, 0xffffffff, 4);
4417 }
4418
gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring * ring,uint16_t pasid,uint32_t flush_type,bool all_hub,uint8_t dst_sel)4419 static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring,
4420 uint16_t pasid, uint32_t flush_type,
4421 bool all_hub, uint8_t dst_sel)
4422 {
4423 amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
4424 amdgpu_ring_write(ring,
4425 PACKET3_INVALIDATE_TLBS_DST_SEL(dst_sel) |
4426 PACKET3_INVALIDATE_TLBS_ALL_HUB(all_hub) |
4427 PACKET3_INVALIDATE_TLBS_PASID(pasid) |
4428 PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type));
4429 }
4430
gfx_v12_0_ring_emit_vm_flush(struct amdgpu_ring * ring,unsigned vmid,uint64_t pd_addr)4431 static void gfx_v12_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
4432 unsigned vmid, uint64_t pd_addr)
4433 {
4434 if (ring->is_mes_queue)
4435 gfx_v12_0_ring_invalidate_tlbs(ring, 0, 0, false, 0);
4436 else
4437 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr);
4438
4439 /* compute doesn't have PFP */
4440 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX) {
4441 /* sync PFP to ME, otherwise we might get invalid PFP reads */
4442 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
4443 amdgpu_ring_write(ring, 0x0);
4444 }
4445 }
4446
gfx_v12_0_ring_emit_fence_kiq(struct amdgpu_ring * ring,u64 addr,u64 seq,unsigned int flags)4447 static void gfx_v12_0_ring_emit_fence_kiq(struct amdgpu_ring *ring, u64 addr,
4448 u64 seq, unsigned int flags)
4449 {
4450 struct amdgpu_device *adev = ring->adev;
4451
4452 /* we only allocate 32bit for each seq wb address */
4453 BUG_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
4454
4455 /* write fence seq to the "addr" */
4456 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4457 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4458 WRITE_DATA_DST_SEL(5) | WR_CONFIRM));
4459 amdgpu_ring_write(ring, lower_32_bits(addr));
4460 amdgpu_ring_write(ring, upper_32_bits(addr));
4461 amdgpu_ring_write(ring, lower_32_bits(seq));
4462
4463 if (flags & AMDGPU_FENCE_FLAG_INT) {
4464 /* set register to trigger INT */
4465 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4466 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
4467 WRITE_DATA_DST_SEL(0) | WR_CONFIRM));
4468 amdgpu_ring_write(ring, SOC15_REG_OFFSET(GC, 0, regCPC_INT_STATUS));
4469 amdgpu_ring_write(ring, 0);
4470 amdgpu_ring_write(ring, 0x20000000); /* src_id is 178 */
4471 }
4472 }
4473
gfx_v12_0_ring_emit_cntxcntl(struct amdgpu_ring * ring,uint32_t flags)4474 static void gfx_v12_0_ring_emit_cntxcntl(struct amdgpu_ring *ring,
4475 uint32_t flags)
4476 {
4477 uint32_t dw2 = 0;
4478
4479 dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */
4480 if (flags & AMDGPU_HAVE_CTX_SWITCH) {
4481 /* set load_global_config & load_global_uconfig */
4482 dw2 |= 0x8001;
4483 /* set load_cs_sh_regs */
4484 dw2 |= 0x01000000;
4485 /* set load_per_context_state & load_gfx_sh_regs for GFX */
4486 dw2 |= 0x10002;
4487 }
4488
4489 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
4490 amdgpu_ring_write(ring, dw2);
4491 amdgpu_ring_write(ring, 0);
4492 }
4493
gfx_v12_0_ring_emit_init_cond_exec(struct amdgpu_ring * ring,uint64_t addr)4494 static unsigned gfx_v12_0_ring_emit_init_cond_exec(struct amdgpu_ring *ring,
4495 uint64_t addr)
4496 {
4497 unsigned ret;
4498
4499 amdgpu_ring_write(ring, PACKET3(PACKET3_COND_EXEC, 3));
4500 amdgpu_ring_write(ring, lower_32_bits(addr));
4501 amdgpu_ring_write(ring, upper_32_bits(addr));
4502 /* discard following DWs if *cond_exec_gpu_addr==0 */
4503 amdgpu_ring_write(ring, 0);
4504 ret = ring->wptr & ring->buf_mask;
4505 /* patch dummy value later */
4506 amdgpu_ring_write(ring, 0);
4507
4508 return ret;
4509 }
4510
gfx_v12_0_ring_preempt_ib(struct amdgpu_ring * ring)4511 static int gfx_v12_0_ring_preempt_ib(struct amdgpu_ring *ring)
4512 {
4513 int i, r = 0;
4514 struct amdgpu_device *adev = ring->adev;
4515 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
4516 struct amdgpu_ring *kiq_ring = &kiq->ring;
4517 unsigned long flags;
4518
4519 if (adev->enable_mes)
4520 return -EINVAL;
4521
4522 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues)
4523 return -EINVAL;
4524
4525 spin_lock_irqsave(&kiq->ring_lock, flags);
4526
4527 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) {
4528 spin_unlock_irqrestore(&kiq->ring_lock, flags);
4529 return -ENOMEM;
4530 }
4531
4532 /* assert preemption condition */
4533 amdgpu_ring_set_preempt_cond_exec(ring, false);
4534
4535 /* assert IB preemption, emit the trailing fence */
4536 kiq->pmf->kiq_unmap_queues(kiq_ring, ring, PREEMPT_QUEUES_NO_UNMAP,
4537 ring->trail_fence_gpu_addr,
4538 ++ring->trail_seq);
4539 amdgpu_ring_commit(kiq_ring);
4540
4541 spin_unlock_irqrestore(&kiq->ring_lock, flags);
4542
4543 /* poll the trailing fence */
4544 for (i = 0; i < adev->usec_timeout; i++) {
4545 if (ring->trail_seq ==
4546 le32_to_cpu(*(ring->trail_fence_cpu_addr)))
4547 break;
4548 udelay(1);
4549 }
4550
4551 if (i >= adev->usec_timeout) {
4552 r = -EINVAL;
4553 DRM_ERROR("ring %d failed to preempt ib\n", ring->idx);
4554 }
4555
4556 /* deassert preemption condition */
4557 amdgpu_ring_set_preempt_cond_exec(ring, true);
4558 return r;
4559 }
4560
gfx_v12_0_ring_emit_frame_cntl(struct amdgpu_ring * ring,bool start,bool secure)4561 static void gfx_v12_0_ring_emit_frame_cntl(struct amdgpu_ring *ring,
4562 bool start,
4563 bool secure)
4564 {
4565 uint32_t v = secure ? FRAME_TMZ : 0;
4566
4567 amdgpu_ring_write(ring, PACKET3(PACKET3_FRAME_CONTROL, 0));
4568 amdgpu_ring_write(ring, v | FRAME_CMD(start ? 0 : 1));
4569 }
4570
gfx_v12_0_ring_emit_rreg(struct amdgpu_ring * ring,uint32_t reg,uint32_t reg_val_offs)4571 static void gfx_v12_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg,
4572 uint32_t reg_val_offs)
4573 {
4574 struct amdgpu_device *adev = ring->adev;
4575
4576 amdgpu_ring_write(ring, PACKET3(PACKET3_COPY_DATA, 4));
4577 amdgpu_ring_write(ring, 0 | /* src: register*/
4578 (5 << 8) | /* dst: memory */
4579 (1 << 20)); /* write confirm */
4580 amdgpu_ring_write(ring, reg);
4581 amdgpu_ring_write(ring, 0);
4582 amdgpu_ring_write(ring, lower_32_bits(adev->wb.gpu_addr +
4583 reg_val_offs * 4));
4584 amdgpu_ring_write(ring, upper_32_bits(adev->wb.gpu_addr +
4585 reg_val_offs * 4));
4586 }
4587
gfx_v12_0_ring_emit_wreg(struct amdgpu_ring * ring,uint32_t reg,uint32_t val)4588 static void gfx_v12_0_ring_emit_wreg(struct amdgpu_ring *ring,
4589 uint32_t reg,
4590 uint32_t val)
4591 {
4592 uint32_t cmd = 0;
4593
4594 switch (ring->funcs->type) {
4595 case AMDGPU_RING_TYPE_GFX:
4596 cmd = WRITE_DATA_ENGINE_SEL(1) | WR_CONFIRM;
4597 break;
4598 case AMDGPU_RING_TYPE_KIQ:
4599 cmd = (1 << 16); /* no inc addr */
4600 break;
4601 default:
4602 cmd = WR_CONFIRM;
4603 break;
4604 }
4605 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
4606 amdgpu_ring_write(ring, cmd);
4607 amdgpu_ring_write(ring, reg);
4608 amdgpu_ring_write(ring, 0);
4609 amdgpu_ring_write(ring, val);
4610 }
4611
gfx_v12_0_ring_emit_reg_wait(struct amdgpu_ring * ring,uint32_t reg,uint32_t val,uint32_t mask)4612 static void gfx_v12_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg,
4613 uint32_t val, uint32_t mask)
4614 {
4615 gfx_v12_0_wait_reg_mem(ring, 0, 0, 0, reg, 0, val, mask, 0x20);
4616 }
4617
gfx_v12_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring * ring,uint32_t reg0,uint32_t reg1,uint32_t ref,uint32_t mask)4618 static void gfx_v12_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
4619 uint32_t reg0, uint32_t reg1,
4620 uint32_t ref, uint32_t mask)
4621 {
4622 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
4623
4624 gfx_v12_0_wait_reg_mem(ring, usepfp, 0, 1, reg0, reg1,
4625 ref, mask, 0x20);
4626 }
4627
gfx_v12_0_ring_soft_recovery(struct amdgpu_ring * ring,unsigned vmid)4628 static void gfx_v12_0_ring_soft_recovery(struct amdgpu_ring *ring,
4629 unsigned vmid)
4630 {
4631 struct amdgpu_device *adev = ring->adev;
4632 uint32_t value = 0;
4633
4634 value = REG_SET_FIELD(value, SQ_CMD, CMD, 0x03);
4635 value = REG_SET_FIELD(value, SQ_CMD, MODE, 0x01);
4636 value = REG_SET_FIELD(value, SQ_CMD, CHECK_VMID, 1);
4637 value = REG_SET_FIELD(value, SQ_CMD, VM_ID, vmid);
4638 amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
4639 WREG32_SOC15(GC, 0, regSQ_CMD, value);
4640 amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
4641 }
4642
4643 static void
gfx_v12_0_set_gfx_eop_interrupt_state(struct amdgpu_device * adev,uint32_t me,uint32_t pipe,enum amdgpu_interrupt_state state)4644 gfx_v12_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev,
4645 uint32_t me, uint32_t pipe,
4646 enum amdgpu_interrupt_state state)
4647 {
4648 uint32_t cp_int_cntl, cp_int_cntl_reg;
4649
4650 if (!me) {
4651 switch (pipe) {
4652 case 0:
4653 cp_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0);
4654 break;
4655 default:
4656 DRM_DEBUG("invalid pipe %d\n", pipe);
4657 return;
4658 }
4659 } else {
4660 DRM_DEBUG("invalid me %d\n", me);
4661 return;
4662 }
4663
4664 switch (state) {
4665 case AMDGPU_IRQ_STATE_DISABLE:
4666 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4667 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4668 TIME_STAMP_INT_ENABLE, 0);
4669 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4670 GENERIC0_INT_ENABLE, 0);
4671 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4672 break;
4673 case AMDGPU_IRQ_STATE_ENABLE:
4674 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4675 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4676 TIME_STAMP_INT_ENABLE, 1);
4677 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4678 GENERIC0_INT_ENABLE, 1);
4679 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4680 break;
4681 default:
4682 break;
4683 }
4684 }
4685
gfx_v12_0_set_compute_eop_interrupt_state(struct amdgpu_device * adev,int me,int pipe,enum amdgpu_interrupt_state state)4686 static void gfx_v12_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev,
4687 int me, int pipe,
4688 enum amdgpu_interrupt_state state)
4689 {
4690 u32 mec_int_cntl, mec_int_cntl_reg;
4691
4692 /*
4693 * amdgpu controls only the first MEC. That's why this function only
4694 * handles the setting of interrupts for this specific MEC. All other
4695 * pipes' interrupts are set by amdkfd.
4696 */
4697
4698 if (me == 1) {
4699 switch (pipe) {
4700 case 0:
4701 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL);
4702 break;
4703 case 1:
4704 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL);
4705 break;
4706 default:
4707 DRM_DEBUG("invalid pipe %d\n", pipe);
4708 return;
4709 }
4710 } else {
4711 DRM_DEBUG("invalid me %d\n", me);
4712 return;
4713 }
4714
4715 switch (state) {
4716 case AMDGPU_IRQ_STATE_DISABLE:
4717 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg);
4718 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
4719 TIME_STAMP_INT_ENABLE, 0);
4720 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
4721 GENERIC0_INT_ENABLE, 0);
4722 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl);
4723 break;
4724 case AMDGPU_IRQ_STATE_ENABLE:
4725 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg);
4726 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
4727 TIME_STAMP_INT_ENABLE, 1);
4728 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
4729 GENERIC0_INT_ENABLE, 1);
4730 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl);
4731 break;
4732 default:
4733 break;
4734 }
4735 }
4736
gfx_v12_0_set_eop_interrupt_state(struct amdgpu_device * adev,struct amdgpu_irq_src * src,unsigned type,enum amdgpu_interrupt_state state)4737 static int gfx_v12_0_set_eop_interrupt_state(struct amdgpu_device *adev,
4738 struct amdgpu_irq_src *src,
4739 unsigned type,
4740 enum amdgpu_interrupt_state state)
4741 {
4742 switch (type) {
4743 case AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP:
4744 gfx_v12_0_set_gfx_eop_interrupt_state(adev, 0, 0, state);
4745 break;
4746 case AMDGPU_CP_IRQ_GFX_ME0_PIPE1_EOP:
4747 gfx_v12_0_set_gfx_eop_interrupt_state(adev, 0, 1, state);
4748 break;
4749 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP:
4750 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 0, state);
4751 break;
4752 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP:
4753 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 1, state);
4754 break;
4755 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP:
4756 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 2, state);
4757 break;
4758 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP:
4759 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 3, state);
4760 break;
4761 default:
4762 break;
4763 }
4764 return 0;
4765 }
4766
gfx_v12_0_eop_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)4767 static int gfx_v12_0_eop_irq(struct amdgpu_device *adev,
4768 struct amdgpu_irq_src *source,
4769 struct amdgpu_iv_entry *entry)
4770 {
4771 int i;
4772 u8 me_id, pipe_id, queue_id;
4773 struct amdgpu_ring *ring;
4774 uint32_t mes_queue_id = entry->src_data[0];
4775
4776 DRM_DEBUG("IH: CP EOP\n");
4777
4778 if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) {
4779 struct amdgpu_mes_queue *queue;
4780
4781 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK;
4782
4783 spin_lock(&adev->mes.queue_id_lock);
4784 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id);
4785 if (queue) {
4786 DRM_DEBUG("process mes queue id = %d\n", mes_queue_id);
4787 amdgpu_fence_process(queue->ring);
4788 }
4789 spin_unlock(&adev->mes.queue_id_lock);
4790 } else {
4791 me_id = (entry->ring_id & 0x0c) >> 2;
4792 pipe_id = (entry->ring_id & 0x03) >> 0;
4793 queue_id = (entry->ring_id & 0x70) >> 4;
4794
4795 switch (me_id) {
4796 case 0:
4797 if (pipe_id == 0)
4798 amdgpu_fence_process(&adev->gfx.gfx_ring[0]);
4799 else
4800 amdgpu_fence_process(&adev->gfx.gfx_ring[1]);
4801 break;
4802 case 1:
4803 case 2:
4804 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
4805 ring = &adev->gfx.compute_ring[i];
4806 /* Per-queue interrupt is supported for MEC starting from VI.
4807 * The interrupt can only be enabled/disabled per pipe instead
4808 * of per queue.
4809 */
4810 if ((ring->me == me_id) &&
4811 (ring->pipe == pipe_id) &&
4812 (ring->queue == queue_id))
4813 amdgpu_fence_process(ring);
4814 }
4815 break;
4816 }
4817 }
4818
4819 return 0;
4820 }
4821
gfx_v12_0_set_priv_reg_fault_state(struct amdgpu_device * adev,struct amdgpu_irq_src * source,unsigned int type,enum amdgpu_interrupt_state state)4822 static int gfx_v12_0_set_priv_reg_fault_state(struct amdgpu_device *adev,
4823 struct amdgpu_irq_src *source,
4824 unsigned int type,
4825 enum amdgpu_interrupt_state state)
4826 {
4827 u32 cp_int_cntl_reg, cp_int_cntl;
4828 int i, j;
4829
4830 switch (state) {
4831 case AMDGPU_IRQ_STATE_DISABLE:
4832 case AMDGPU_IRQ_STATE_ENABLE:
4833 for (i = 0; i < adev->gfx.me.num_me; i++) {
4834 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) {
4835 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j);
4836
4837 if (cp_int_cntl_reg) {
4838 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4839 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4840 PRIV_REG_INT_ENABLE,
4841 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
4842 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4843 }
4844 }
4845 }
4846 for (i = 0; i < adev->gfx.mec.num_mec; i++) {
4847 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) {
4848 /* MECs start at 1 */
4849 cp_int_cntl_reg = gfx_v12_0_get_cpc_int_cntl(adev, i + 1, j);
4850
4851 if (cp_int_cntl_reg) {
4852 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4853 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL,
4854 PRIV_REG_INT_ENABLE,
4855 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
4856 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4857 }
4858 }
4859 }
4860 break;
4861 default:
4862 break;
4863 }
4864
4865 return 0;
4866 }
4867
gfx_v12_0_set_bad_op_fault_state(struct amdgpu_device * adev,struct amdgpu_irq_src * source,unsigned type,enum amdgpu_interrupt_state state)4868 static int gfx_v12_0_set_bad_op_fault_state(struct amdgpu_device *adev,
4869 struct amdgpu_irq_src *source,
4870 unsigned type,
4871 enum amdgpu_interrupt_state state)
4872 {
4873 u32 cp_int_cntl_reg, cp_int_cntl;
4874 int i, j;
4875
4876 switch (state) {
4877 case AMDGPU_IRQ_STATE_DISABLE:
4878 case AMDGPU_IRQ_STATE_ENABLE:
4879 for (i = 0; i < adev->gfx.me.num_me; i++) {
4880 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) {
4881 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j);
4882
4883 if (cp_int_cntl_reg) {
4884 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4885 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4886 OPCODE_ERROR_INT_ENABLE,
4887 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
4888 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4889 }
4890 }
4891 }
4892 for (i = 0; i < adev->gfx.mec.num_mec; i++) {
4893 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) {
4894 /* MECs start at 1 */
4895 cp_int_cntl_reg = gfx_v12_0_get_cpc_int_cntl(adev, i + 1, j);
4896
4897 if (cp_int_cntl_reg) {
4898 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4899 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL,
4900 OPCODE_ERROR_INT_ENABLE,
4901 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
4902 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4903 }
4904 }
4905 }
4906 break;
4907 default:
4908 break;
4909 }
4910 return 0;
4911 }
4912
gfx_v12_0_set_priv_inst_fault_state(struct amdgpu_device * adev,struct amdgpu_irq_src * source,unsigned int type,enum amdgpu_interrupt_state state)4913 static int gfx_v12_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
4914 struct amdgpu_irq_src *source,
4915 unsigned int type,
4916 enum amdgpu_interrupt_state state)
4917 {
4918 u32 cp_int_cntl_reg, cp_int_cntl;
4919 int i, j;
4920
4921 switch (state) {
4922 case AMDGPU_IRQ_STATE_DISABLE:
4923 case AMDGPU_IRQ_STATE_ENABLE:
4924 for (i = 0; i < adev->gfx.me.num_me; i++) {
4925 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) {
4926 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j);
4927
4928 if (cp_int_cntl_reg) {
4929 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg);
4930 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0,
4931 PRIV_INSTR_INT_ENABLE,
4932 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
4933 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl);
4934 }
4935 }
4936 }
4937 break;
4938 default:
4939 break;
4940 }
4941
4942 return 0;
4943 }
4944
gfx_v12_0_handle_priv_fault(struct amdgpu_device * adev,struct amdgpu_iv_entry * entry)4945 static void gfx_v12_0_handle_priv_fault(struct amdgpu_device *adev,
4946 struct amdgpu_iv_entry *entry)
4947 {
4948 u8 me_id, pipe_id, queue_id;
4949 struct amdgpu_ring *ring;
4950 int i;
4951
4952 me_id = (entry->ring_id & 0x0c) >> 2;
4953 pipe_id = (entry->ring_id & 0x03) >> 0;
4954 queue_id = (entry->ring_id & 0x70) >> 4;
4955
4956 switch (me_id) {
4957 case 0:
4958 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
4959 ring = &adev->gfx.gfx_ring[i];
4960 if (ring->me == me_id && ring->pipe == pipe_id &&
4961 ring->queue == queue_id)
4962 drm_sched_fault(&ring->sched);
4963 }
4964 break;
4965 case 1:
4966 case 2:
4967 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
4968 ring = &adev->gfx.compute_ring[i];
4969 if (ring->me == me_id && ring->pipe == pipe_id &&
4970 ring->queue == queue_id)
4971 drm_sched_fault(&ring->sched);
4972 }
4973 break;
4974 default:
4975 BUG();
4976 break;
4977 }
4978 }
4979
gfx_v12_0_priv_reg_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)4980 static int gfx_v12_0_priv_reg_irq(struct amdgpu_device *adev,
4981 struct amdgpu_irq_src *source,
4982 struct amdgpu_iv_entry *entry)
4983 {
4984 DRM_ERROR("Illegal register access in command stream\n");
4985 gfx_v12_0_handle_priv_fault(adev, entry);
4986 return 0;
4987 }
4988
gfx_v12_0_bad_op_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)4989 static int gfx_v12_0_bad_op_irq(struct amdgpu_device *adev,
4990 struct amdgpu_irq_src *source,
4991 struct amdgpu_iv_entry *entry)
4992 {
4993 DRM_ERROR("Illegal opcode in command stream \n");
4994 gfx_v12_0_handle_priv_fault(adev, entry);
4995 return 0;
4996 }
4997
gfx_v12_0_priv_inst_irq(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)4998 static int gfx_v12_0_priv_inst_irq(struct amdgpu_device *adev,
4999 struct amdgpu_irq_src *source,
5000 struct amdgpu_iv_entry *entry)
5001 {
5002 DRM_ERROR("Illegal instruction in command stream\n");
5003 gfx_v12_0_handle_priv_fault(adev, entry);
5004 return 0;
5005 }
5006
gfx_v12_0_emit_mem_sync(struct amdgpu_ring * ring)5007 static void gfx_v12_0_emit_mem_sync(struct amdgpu_ring *ring)
5008 {
5009 const unsigned int gcr_cntl =
5010 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_INV(1) |
5011 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_WB(1) |
5012 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_INV(1) |
5013 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_WB(1) |
5014 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL1_INV(1) |
5015 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLV_INV(1) |
5016 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLK_INV(1) |
5017 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLI_INV(1);
5018
5019 /* ACQUIRE_MEM - make one or more surfaces valid for use by the subsequent operations */
5020 amdgpu_ring_write(ring, PACKET3(PACKET3_ACQUIRE_MEM, 6));
5021 amdgpu_ring_write(ring, 0); /* CP_COHER_CNTL */
5022 amdgpu_ring_write(ring, 0xffffffff); /* CP_COHER_SIZE */
5023 amdgpu_ring_write(ring, 0xffffff); /* CP_COHER_SIZE_HI */
5024 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE */
5025 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE_HI */
5026 amdgpu_ring_write(ring, 0x0000000A); /* POLL_INTERVAL */
5027 amdgpu_ring_write(ring, gcr_cntl); /* GCR_CNTL */
5028 }
5029
gfx_v12_ring_insert_nop(struct amdgpu_ring * ring,uint32_t num_nop)5030 static void gfx_v12_ring_insert_nop(struct amdgpu_ring *ring, uint32_t num_nop)
5031 {
5032 int i;
5033
5034 /* Header itself is a NOP packet */
5035 if (num_nop == 1) {
5036 amdgpu_ring_write(ring, ring->funcs->nop);
5037 return;
5038 }
5039
5040 /* Max HW optimization till 0x3ffe, followed by remaining one NOP at a time*/
5041 amdgpu_ring_write(ring, PACKET3(PACKET3_NOP, min(num_nop - 2, 0x3ffe)));
5042
5043 /* Header is at index 0, followed by num_nops - 1 NOP packet's */
5044 for (i = 1; i < num_nop; i++)
5045 amdgpu_ring_write(ring, ring->funcs->nop);
5046 }
5047
gfx_v12_ip_print(void * handle,struct drm_printer * p)5048 static void gfx_v12_ip_print(void *handle, struct drm_printer *p)
5049 {
5050 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
5051 uint32_t i, j, k, reg, index = 0;
5052 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0);
5053
5054 if (!adev->gfx.ip_dump_core)
5055 return;
5056
5057 for (i = 0; i < reg_count; i++)
5058 drm_printf(p, "%-50s \t 0x%08x\n",
5059 gc_reg_list_12_0[i].reg_name,
5060 adev->gfx.ip_dump_core[i]);
5061
5062 /* print compute queue registers for all instances */
5063 if (!adev->gfx.ip_dump_compute_queues)
5064 return;
5065
5066 reg_count = ARRAY_SIZE(gc_cp_reg_list_12);
5067 drm_printf(p, "\nnum_mec: %d num_pipe: %d num_queue: %d\n",
5068 adev->gfx.mec.num_mec,
5069 adev->gfx.mec.num_pipe_per_mec,
5070 adev->gfx.mec.num_queue_per_pipe);
5071
5072 for (i = 0; i < adev->gfx.mec.num_mec; i++) {
5073 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) {
5074 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) {
5075 drm_printf(p, "\nmec %d, pipe %d, queue %d\n", i, j, k);
5076 for (reg = 0; reg < reg_count; reg++) {
5077 drm_printf(p, "%-50s \t 0x%08x\n",
5078 gc_cp_reg_list_12[reg].reg_name,
5079 adev->gfx.ip_dump_compute_queues[index + reg]);
5080 }
5081 index += reg_count;
5082 }
5083 }
5084 }
5085
5086 /* print gfx queue registers for all instances */
5087 if (!adev->gfx.ip_dump_gfx_queues)
5088 return;
5089
5090 index = 0;
5091 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12);
5092 drm_printf(p, "\nnum_me: %d num_pipe: %d num_queue: %d\n",
5093 adev->gfx.me.num_me,
5094 adev->gfx.me.num_pipe_per_me,
5095 adev->gfx.me.num_queue_per_pipe);
5096
5097 for (i = 0; i < adev->gfx.me.num_me; i++) {
5098 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) {
5099 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) {
5100 drm_printf(p, "\nme %d, pipe %d, queue %d\n", i, j, k);
5101 for (reg = 0; reg < reg_count; reg++) {
5102 drm_printf(p, "%-50s \t 0x%08x\n",
5103 gc_gfx_queue_reg_list_12[reg].reg_name,
5104 adev->gfx.ip_dump_gfx_queues[index + reg]);
5105 }
5106 index += reg_count;
5107 }
5108 }
5109 }
5110 }
5111
gfx_v12_ip_dump(void * handle)5112 static void gfx_v12_ip_dump(void *handle)
5113 {
5114 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
5115 uint32_t i, j, k, reg, index = 0;
5116 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0);
5117
5118 if (!adev->gfx.ip_dump_core)
5119 return;
5120
5121 amdgpu_gfx_off_ctrl(adev, false);
5122 for (i = 0; i < reg_count; i++)
5123 adev->gfx.ip_dump_core[i] = RREG32(SOC15_REG_ENTRY_OFFSET(gc_reg_list_12_0[i]));
5124 amdgpu_gfx_off_ctrl(adev, true);
5125
5126 /* dump compute queue registers for all instances */
5127 if (!adev->gfx.ip_dump_compute_queues)
5128 return;
5129
5130 reg_count = ARRAY_SIZE(gc_cp_reg_list_12);
5131 amdgpu_gfx_off_ctrl(adev, false);
5132 mutex_lock(&adev->srbm_mutex);
5133 for (i = 0; i < adev->gfx.mec.num_mec; i++) {
5134 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) {
5135 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) {
5136 /* ME0 is for GFX so start from 1 for CP */
5137 soc24_grbm_select(adev, adev->gfx.me.num_me + i, j, k, 0);
5138 for (reg = 0; reg < reg_count; reg++) {
5139 adev->gfx.ip_dump_compute_queues[index + reg] =
5140 RREG32(SOC15_REG_ENTRY_OFFSET(
5141 gc_cp_reg_list_12[reg]));
5142 }
5143 index += reg_count;
5144 }
5145 }
5146 }
5147 soc24_grbm_select(adev, 0, 0, 0, 0);
5148 mutex_unlock(&adev->srbm_mutex);
5149 amdgpu_gfx_off_ctrl(adev, true);
5150
5151 /* dump gfx queue registers for all instances */
5152 if (!adev->gfx.ip_dump_gfx_queues)
5153 return;
5154
5155 index = 0;
5156 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12);
5157 amdgpu_gfx_off_ctrl(adev, false);
5158 mutex_lock(&adev->srbm_mutex);
5159 for (i = 0; i < adev->gfx.me.num_me; i++) {
5160 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) {
5161 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) {
5162 soc24_grbm_select(adev, i, j, k, 0);
5163
5164 for (reg = 0; reg < reg_count; reg++) {
5165 adev->gfx.ip_dump_gfx_queues[index + reg] =
5166 RREG32(SOC15_REG_ENTRY_OFFSET(
5167 gc_gfx_queue_reg_list_12[reg]));
5168 }
5169 index += reg_count;
5170 }
5171 }
5172 }
5173 soc24_grbm_select(adev, 0, 0, 0, 0);
5174 mutex_unlock(&adev->srbm_mutex);
5175 amdgpu_gfx_off_ctrl(adev, true);
5176 }
5177
gfx_v12_0_reset_kgq(struct amdgpu_ring * ring,unsigned int vmid)5178 static int gfx_v12_0_reset_kgq(struct amdgpu_ring *ring, unsigned int vmid)
5179 {
5180 struct amdgpu_device *adev = ring->adev;
5181 int r;
5182
5183 if (amdgpu_sriov_vf(adev))
5184 return -EINVAL;
5185
5186 r = amdgpu_mes_reset_legacy_queue(ring->adev, ring, vmid, false);
5187 if (r) {
5188 dev_err(adev->dev, "reset via MES failed %d\n", r);
5189 return r;
5190 }
5191
5192 r = amdgpu_bo_reserve(ring->mqd_obj, false);
5193 if (unlikely(r != 0)) {
5194 dev_err(adev->dev, "fail to resv mqd_obj\n");
5195 return r;
5196 }
5197 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
5198 if (!r) {
5199 r = gfx_v12_0_kgq_init_queue(ring, true);
5200 amdgpu_bo_kunmap(ring->mqd_obj);
5201 ring->mqd_ptr = NULL;
5202 }
5203 amdgpu_bo_unreserve(ring->mqd_obj);
5204 if (r) {
5205 DRM_ERROR("fail to unresv mqd_obj\n");
5206 return r;
5207 }
5208
5209 r = amdgpu_mes_map_legacy_queue(adev, ring);
5210 if (r) {
5211 dev_err(adev->dev, "failed to remap kgq\n");
5212 return r;
5213 }
5214
5215 return amdgpu_ring_test_ring(ring);
5216 }
5217
gfx_v12_0_reset_kcq(struct amdgpu_ring * ring,unsigned int vmid)5218 static int gfx_v12_0_reset_kcq(struct amdgpu_ring *ring, unsigned int vmid)
5219 {
5220 struct amdgpu_device *adev = ring->adev;
5221 int r, i;
5222
5223 if (amdgpu_sriov_vf(adev))
5224 return -EINVAL;
5225
5226 amdgpu_gfx_rlc_enter_safe_mode(adev, 0);
5227 mutex_lock(&adev->srbm_mutex);
5228 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
5229 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 0x2);
5230 WREG32_SOC15(GC, 0, regSPI_COMPUTE_QUEUE_RESET, 0x1);
5231 for (i = 0; i < adev->usec_timeout; i++) {
5232 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1))
5233 break;
5234 udelay(1);
5235 }
5236 soc24_grbm_select(adev, 0, 0, 0, 0);
5237 mutex_unlock(&adev->srbm_mutex);
5238 amdgpu_gfx_rlc_exit_safe_mode(adev, 0);
5239
5240 r = amdgpu_bo_reserve(ring->mqd_obj, false);
5241 if (unlikely(r != 0)) {
5242 DRM_ERROR("fail to resv mqd_obj\n");
5243 return r;
5244 }
5245 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
5246 if (!r) {
5247 r = gfx_v12_0_kcq_init_queue(ring, true);
5248 amdgpu_bo_kunmap(ring->mqd_obj);
5249 ring->mqd_ptr = NULL;
5250 }
5251 amdgpu_bo_unreserve(ring->mqd_obj);
5252 if (r) {
5253 DRM_ERROR("fail to unresv mqd_obj\n");
5254 return r;
5255 }
5256 r = amdgpu_mes_map_legacy_queue(adev, ring);
5257 if (r) {
5258 dev_err(adev->dev, "failed to remap kcq\n");
5259 return r;
5260 }
5261
5262 return amdgpu_ring_test_ring(ring);
5263 }
5264
5265 static const struct amd_ip_funcs gfx_v12_0_ip_funcs = {
5266 .name = "gfx_v12_0",
5267 .early_init = gfx_v12_0_early_init,
5268 .late_init = gfx_v12_0_late_init,
5269 .sw_init = gfx_v12_0_sw_init,
5270 .sw_fini = gfx_v12_0_sw_fini,
5271 .hw_init = gfx_v12_0_hw_init,
5272 .hw_fini = gfx_v12_0_hw_fini,
5273 .suspend = gfx_v12_0_suspend,
5274 .resume = gfx_v12_0_resume,
5275 .is_idle = gfx_v12_0_is_idle,
5276 .wait_for_idle = gfx_v12_0_wait_for_idle,
5277 .set_clockgating_state = gfx_v12_0_set_clockgating_state,
5278 .set_powergating_state = gfx_v12_0_set_powergating_state,
5279 .get_clockgating_state = gfx_v12_0_get_clockgating_state,
5280 .dump_ip_state = gfx_v12_ip_dump,
5281 .print_ip_state = gfx_v12_ip_print,
5282 };
5283
5284 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_gfx = {
5285 .type = AMDGPU_RING_TYPE_GFX,
5286 .align_mask = 0xff,
5287 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
5288 .support_64bit_ptrs = true,
5289 .secure_submission_supported = true,
5290 .get_rptr = gfx_v12_0_ring_get_rptr_gfx,
5291 .get_wptr = gfx_v12_0_ring_get_wptr_gfx,
5292 .set_wptr = gfx_v12_0_ring_set_wptr_gfx,
5293 .emit_frame_size = /* totally 242 maximum if 16 IBs */
5294 5 + /* COND_EXEC */
5295 7 + /* PIPELINE_SYNC */
5296 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 +
5297 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 +
5298 2 + /* VM_FLUSH */
5299 8 + /* FENCE for VM_FLUSH */
5300 5 + /* COND_EXEC */
5301 7 + /* HDP_flush */
5302 4 + /* VGT_flush */
5303 31 + /* DE_META */
5304 3 + /* CNTX_CTRL */
5305 5 + /* HDP_INVL */
5306 8 + 8 + /* FENCE x2 */
5307 8, /* gfx_v12_0_emit_mem_sync */
5308 .emit_ib_size = 4, /* gfx_v12_0_ring_emit_ib_gfx */
5309 .emit_ib = gfx_v12_0_ring_emit_ib_gfx,
5310 .emit_fence = gfx_v12_0_ring_emit_fence,
5311 .emit_pipeline_sync = gfx_v12_0_ring_emit_pipeline_sync,
5312 .emit_vm_flush = gfx_v12_0_ring_emit_vm_flush,
5313 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush,
5314 .test_ring = gfx_v12_0_ring_test_ring,
5315 .test_ib = gfx_v12_0_ring_test_ib,
5316 .insert_nop = gfx_v12_ring_insert_nop,
5317 .pad_ib = amdgpu_ring_generic_pad_ib,
5318 .emit_cntxcntl = gfx_v12_0_ring_emit_cntxcntl,
5319 .init_cond_exec = gfx_v12_0_ring_emit_init_cond_exec,
5320 .preempt_ib = gfx_v12_0_ring_preempt_ib,
5321 .emit_frame_cntl = gfx_v12_0_ring_emit_frame_cntl,
5322 .emit_wreg = gfx_v12_0_ring_emit_wreg,
5323 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait,
5324 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait,
5325 .soft_recovery = gfx_v12_0_ring_soft_recovery,
5326 .emit_mem_sync = gfx_v12_0_emit_mem_sync,
5327 .reset = gfx_v12_0_reset_kgq,
5328 };
5329
5330 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_compute = {
5331 .type = AMDGPU_RING_TYPE_COMPUTE,
5332 .align_mask = 0xff,
5333 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
5334 .support_64bit_ptrs = true,
5335 .get_rptr = gfx_v12_0_ring_get_rptr_compute,
5336 .get_wptr = gfx_v12_0_ring_get_wptr_compute,
5337 .set_wptr = gfx_v12_0_ring_set_wptr_compute,
5338 .emit_frame_size =
5339 7 + /* gfx_v12_0_ring_emit_hdp_flush */
5340 5 + /* hdp invalidate */
5341 7 + /* gfx_v12_0_ring_emit_pipeline_sync */
5342 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 +
5343 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 +
5344 2 + /* gfx_v12_0_ring_emit_vm_flush */
5345 8 + 8 + 8 + /* gfx_v12_0_ring_emit_fence x3 for user fence, vm fence */
5346 8, /* gfx_v12_0_emit_mem_sync */
5347 .emit_ib_size = 7, /* gfx_v12_0_ring_emit_ib_compute */
5348 .emit_ib = gfx_v12_0_ring_emit_ib_compute,
5349 .emit_fence = gfx_v12_0_ring_emit_fence,
5350 .emit_pipeline_sync = gfx_v12_0_ring_emit_pipeline_sync,
5351 .emit_vm_flush = gfx_v12_0_ring_emit_vm_flush,
5352 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush,
5353 .test_ring = gfx_v12_0_ring_test_ring,
5354 .test_ib = gfx_v12_0_ring_test_ib,
5355 .insert_nop = gfx_v12_ring_insert_nop,
5356 .pad_ib = amdgpu_ring_generic_pad_ib,
5357 .emit_wreg = gfx_v12_0_ring_emit_wreg,
5358 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait,
5359 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait,
5360 .soft_recovery = gfx_v12_0_ring_soft_recovery,
5361 .emit_mem_sync = gfx_v12_0_emit_mem_sync,
5362 .reset = gfx_v12_0_reset_kcq,
5363 };
5364
5365 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_kiq = {
5366 .type = AMDGPU_RING_TYPE_KIQ,
5367 .align_mask = 0xff,
5368 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
5369 .support_64bit_ptrs = true,
5370 .get_rptr = gfx_v12_0_ring_get_rptr_compute,
5371 .get_wptr = gfx_v12_0_ring_get_wptr_compute,
5372 .set_wptr = gfx_v12_0_ring_set_wptr_compute,
5373 .emit_frame_size =
5374 7 + /* gfx_v12_0_ring_emit_hdp_flush */
5375 5 + /*hdp invalidate */
5376 7 + /* gfx_v12_0_ring_emit_pipeline_sync */
5377 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 +
5378 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 +
5379 2 + /* gfx_v12_0_ring_emit_vm_flush */
5380 8 + 8 + 8, /* gfx_v12_0_ring_emit_fence_kiq x3 for user fence, vm fence */
5381 .emit_ib_size = 7, /* gfx_v12_0_ring_emit_ib_compute */
5382 .emit_ib = gfx_v12_0_ring_emit_ib_compute,
5383 .emit_fence = gfx_v12_0_ring_emit_fence_kiq,
5384 .test_ring = gfx_v12_0_ring_test_ring,
5385 .test_ib = gfx_v12_0_ring_test_ib,
5386 .insert_nop = amdgpu_ring_insert_nop,
5387 .pad_ib = amdgpu_ring_generic_pad_ib,
5388 .emit_rreg = gfx_v12_0_ring_emit_rreg,
5389 .emit_wreg = gfx_v12_0_ring_emit_wreg,
5390 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait,
5391 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait,
5392 };
5393
gfx_v12_0_set_ring_funcs(struct amdgpu_device * adev)5394 static void gfx_v12_0_set_ring_funcs(struct amdgpu_device *adev)
5395 {
5396 int i;
5397
5398 adev->gfx.kiq[0].ring.funcs = &gfx_v12_0_ring_funcs_kiq;
5399
5400 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
5401 adev->gfx.gfx_ring[i].funcs = &gfx_v12_0_ring_funcs_gfx;
5402
5403 for (i = 0; i < adev->gfx.num_compute_rings; i++)
5404 adev->gfx.compute_ring[i].funcs = &gfx_v12_0_ring_funcs_compute;
5405 }
5406
5407 static const struct amdgpu_irq_src_funcs gfx_v12_0_eop_irq_funcs = {
5408 .set = gfx_v12_0_set_eop_interrupt_state,
5409 .process = gfx_v12_0_eop_irq,
5410 };
5411
5412 static const struct amdgpu_irq_src_funcs gfx_v12_0_priv_reg_irq_funcs = {
5413 .set = gfx_v12_0_set_priv_reg_fault_state,
5414 .process = gfx_v12_0_priv_reg_irq,
5415 };
5416
5417 static const struct amdgpu_irq_src_funcs gfx_v12_0_bad_op_irq_funcs = {
5418 .set = gfx_v12_0_set_bad_op_fault_state,
5419 .process = gfx_v12_0_bad_op_irq,
5420 };
5421
5422 static const struct amdgpu_irq_src_funcs gfx_v12_0_priv_inst_irq_funcs = {
5423 .set = gfx_v12_0_set_priv_inst_fault_state,
5424 .process = gfx_v12_0_priv_inst_irq,
5425 };
5426
gfx_v12_0_set_irq_funcs(struct amdgpu_device * adev)5427 static void gfx_v12_0_set_irq_funcs(struct amdgpu_device *adev)
5428 {
5429 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST;
5430 adev->gfx.eop_irq.funcs = &gfx_v12_0_eop_irq_funcs;
5431
5432 adev->gfx.priv_reg_irq.num_types = 1;
5433 adev->gfx.priv_reg_irq.funcs = &gfx_v12_0_priv_reg_irq_funcs;
5434
5435 adev->gfx.bad_op_irq.num_types = 1;
5436 adev->gfx.bad_op_irq.funcs = &gfx_v12_0_bad_op_irq_funcs;
5437
5438 adev->gfx.priv_inst_irq.num_types = 1;
5439 adev->gfx.priv_inst_irq.funcs = &gfx_v12_0_priv_inst_irq_funcs;
5440 }
5441
gfx_v12_0_set_imu_funcs(struct amdgpu_device * adev)5442 static void gfx_v12_0_set_imu_funcs(struct amdgpu_device *adev)
5443 {
5444 if (adev->flags & AMD_IS_APU)
5445 adev->gfx.imu.mode = MISSION_MODE;
5446 else
5447 adev->gfx.imu.mode = DEBUG_MODE;
5448
5449 adev->gfx.imu.funcs = &gfx_v12_0_imu_funcs;
5450 }
5451
gfx_v12_0_set_rlc_funcs(struct amdgpu_device * adev)5452 static void gfx_v12_0_set_rlc_funcs(struct amdgpu_device *adev)
5453 {
5454 adev->gfx.rlc.funcs = &gfx_v12_0_rlc_funcs;
5455 }
5456
gfx_v12_0_set_mqd_funcs(struct amdgpu_device * adev)5457 static void gfx_v12_0_set_mqd_funcs(struct amdgpu_device *adev)
5458 {
5459 /* set gfx eng mqd */
5460 adev->mqds[AMDGPU_HW_IP_GFX].mqd_size =
5461 sizeof(struct v12_gfx_mqd);
5462 adev->mqds[AMDGPU_HW_IP_GFX].init_mqd =
5463 gfx_v12_0_gfx_mqd_init;
5464 /* set compute eng mqd */
5465 adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size =
5466 sizeof(struct v12_compute_mqd);
5467 adev->mqds[AMDGPU_HW_IP_COMPUTE].init_mqd =
5468 gfx_v12_0_compute_mqd_init;
5469 }
5470
gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh(struct amdgpu_device * adev,u32 bitmap)5471 static void gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh(struct amdgpu_device *adev,
5472 u32 bitmap)
5473 {
5474 u32 data;
5475
5476 if (!bitmap)
5477 return;
5478
5479 data = bitmap << GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT;
5480 data &= GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK;
5481
5482 WREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG, data);
5483 }
5484
gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device * adev)5485 static u32 gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev)
5486 {
5487 u32 data, wgp_bitmask;
5488 data = RREG32_SOC15(GC, 0, regCC_GC_SHADER_ARRAY_CONFIG);
5489 data |= RREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG);
5490
5491 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK;
5492 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT;
5493
5494 wgp_bitmask =
5495 amdgpu_gfx_create_bitmask(adev->gfx.config.max_cu_per_sh >> 1);
5496
5497 return (~data) & wgp_bitmask;
5498 }
5499
gfx_v12_0_get_cu_active_bitmap_per_sh(struct amdgpu_device * adev)5500 static u32 gfx_v12_0_get_cu_active_bitmap_per_sh(struct amdgpu_device *adev)
5501 {
5502 u32 wgp_idx, wgp_active_bitmap;
5503 u32 cu_bitmap_per_wgp, cu_active_bitmap;
5504
5505 wgp_active_bitmap = gfx_v12_0_get_wgp_active_bitmap_per_sh(adev);
5506 cu_active_bitmap = 0;
5507
5508 for (wgp_idx = 0; wgp_idx < 16; wgp_idx++) {
5509 /* if there is one WGP enabled, it means 2 CUs will be enabled */
5510 cu_bitmap_per_wgp = 3 << (2 * wgp_idx);
5511 if (wgp_active_bitmap & (1 << wgp_idx))
5512 cu_active_bitmap |= cu_bitmap_per_wgp;
5513 }
5514
5515 return cu_active_bitmap;
5516 }
5517
gfx_v12_0_get_cu_info(struct amdgpu_device * adev,struct amdgpu_cu_info * cu_info)5518 static int gfx_v12_0_get_cu_info(struct amdgpu_device *adev,
5519 struct amdgpu_cu_info *cu_info)
5520 {
5521 int i, j, k, counter, active_cu_number = 0;
5522 u32 mask, bitmap;
5523 unsigned disable_masks[8 * 2];
5524
5525 if (!adev || !cu_info)
5526 return -EINVAL;
5527
5528 amdgpu_gfx_parse_disable_cu(disable_masks, 8, 2);
5529
5530 mutex_lock(&adev->grbm_idx_mutex);
5531 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
5532 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
5533 bitmap = i * adev->gfx.config.max_sh_per_se + j;
5534 if (!((gfx_v12_0_get_sa_active_bitmap(adev) >> bitmap) & 1))
5535 continue;
5536 mask = 1;
5537 counter = 0;
5538 gfx_v12_0_select_se_sh(adev, i, j, 0xffffffff, 0);
5539 if (i < 8 && j < 2)
5540 gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh(
5541 adev, disable_masks[i * 2 + j]);
5542 bitmap = gfx_v12_0_get_cu_active_bitmap_per_sh(adev);
5543
5544 /**
5545 * GFX12 could support more than 4 SEs, while the bitmap
5546 * in cu_info struct is 4x4 and ioctl interface struct
5547 * drm_amdgpu_info_device should keep stable.
5548 * So we use last two columns of bitmap to store cu mask for
5549 * SEs 4 to 7, the layout of the bitmap is as below:
5550 * SE0: {SH0,SH1} --> {bitmap[0][0], bitmap[0][1]}
5551 * SE1: {SH0,SH1} --> {bitmap[1][0], bitmap[1][1]}
5552 * SE2: {SH0,SH1} --> {bitmap[2][0], bitmap[2][1]}
5553 * SE3: {SH0,SH1} --> {bitmap[3][0], bitmap[3][1]}
5554 * SE4: {SH0,SH1} --> {bitmap[0][2], bitmap[0][3]}
5555 * SE5: {SH0,SH1} --> {bitmap[1][2], bitmap[1][3]}
5556 * SE6: {SH0,SH1} --> {bitmap[2][2], bitmap[2][3]}
5557 * SE7: {SH0,SH1} --> {bitmap[3][2], bitmap[3][3]}
5558 */
5559 cu_info->bitmap[0][i % 4][j + (i / 4) * 2] = bitmap;
5560
5561 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k++) {
5562 if (bitmap & mask)
5563 counter++;
5564
5565 mask <<= 1;
5566 }
5567 active_cu_number += counter;
5568 }
5569 }
5570 gfx_v12_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, 0);
5571 mutex_unlock(&adev->grbm_idx_mutex);
5572
5573 cu_info->number = active_cu_number;
5574 cu_info->simd_per_cu = NUM_SIMD_PER_CU;
5575
5576 return 0;
5577 }
5578
5579 const struct amdgpu_ip_block_version gfx_v12_0_ip_block = {
5580 .type = AMD_IP_BLOCK_TYPE_GFX,
5581 .major = 12,
5582 .minor = 0,
5583 .rev = 0,
5584 .funcs = &gfx_v12_0_ip_funcs,
5585 };
5586