?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
proc: Fix proc_sys_prune_dcache to hold a sb reference
[cris-mirror.git] / drivers / gpu / drm / amd / amdgpu / gfx_v9_0.c
blob0c16b7563b7317e2b63aef0a1800945f72430091
1 /*
2 * Copyright 2016 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/firmware.h>
24 #include "drmP.h"
25 #include "amdgpu.h"
26 #include "amdgpu_gfx.h"
27 #include "soc15.h"
28 #include "soc15d.h"
29
30 #include "vega10/soc15ip.h"
31 #include "vega10/GC/gc_9_0_offset.h"
32 #include "vega10/GC/gc_9_0_sh_mask.h"
33 #include "vega10/vega10_enum.h"
34 #include "vega10/HDP/hdp_4_0_offset.h"
35
36 #include "soc15_common.h"
37 #include "clearstate_gfx9.h"
38 #include "v9_structs.h"
39
40 #define GFX9_NUM_GFX_RINGS 1
41 #define GFX9_NUM_COMPUTE_RINGS 8
42 #define RLCG_UCODE_LOADING_START_ADDRESS 0x2000
43
44 MODULE_FIRMWARE("amdgpu/vega10_ce.bin");
45 MODULE_FIRMWARE("amdgpu/vega10_pfp.bin");
46 MODULE_FIRMWARE("amdgpu/vega10_me.bin");
47 MODULE_FIRMWARE("amdgpu/vega10_mec.bin");
48 MODULE_FIRMWARE("amdgpu/vega10_mec2.bin");
49 MODULE_FIRMWARE("amdgpu/vega10_rlc.bin");
50
51 static const struct amdgpu_gds_reg_offset amdgpu_gds_reg_offset[] =
52 {
53 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID0_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID0_SIZE),
54 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID0), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID0)},
55 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID1_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID1_SIZE),
56 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID1), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID1)},
57 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID2_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID2_SIZE),
58 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID2), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID2)},
59 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID3_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID3_SIZE),
60 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID3), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID3)},
61 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID4_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID4_SIZE),
62 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID4), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID4)},
63 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID5_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID5_SIZE),
64 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID5), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID5)},
65 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID6_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID6_SIZE),
66 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID6), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID6)},
67 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID7_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID7_SIZE),
68 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID7), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID7)},
69 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID8_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID8_SIZE),
70 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID8), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID8)},
71 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID9_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID9_SIZE),
72 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID9), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID9)},
73 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID10_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID10_SIZE),
74 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID10), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID10)},
75 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID11_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID11_SIZE),
76 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID11), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID11)},
77 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID12_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID12_SIZE),
78 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID12), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID12)},
79 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID13_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID13_SIZE),
80 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID13), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID13)},
81 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID14_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID14_SIZE),
82 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID14), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID14)},
83 {SOC15_REG_OFFSET(GC, 0, mmGDS_VMID15_BASE), SOC15_REG_OFFSET(GC, 0, mmGDS_VMID15_SIZE),
84 SOC15_REG_OFFSET(GC, 0, mmGDS_GWS_VMID15), SOC15_REG_OFFSET(GC, 0, mmGDS_OA_VMID15)}
85 };
86
87 static const u32 golden_settings_gc_9_0[] =
88 {
89 SOC15_REG_OFFSET(GC, 0, mmDB_DEBUG2), 0xf00ffeff, 0x00000400,
90 SOC15_REG_OFFSET(GC, 0, mmPA_SC_BINNER_EVENT_CNTL_3), 0x00000003, 0x82400024,
91 SOC15_REG_OFFSET(GC, 0, mmPA_SC_ENHANCE), 0x3fffffff, 0x00000001,
92 SOC15_REG_OFFSET(GC, 0, mmPA_SC_LINE_STIPPLE_STATE), 0x0000ff0f, 0x00000000,
93 SOC15_REG_OFFSET(GC, 0, mmTA_CNTL_AUX), 0xfffffeef, 0x010b0000,
94 SOC15_REG_OFFSET(GC, 0, mmTCP_CHAN_STEER_HI), 0xffffffff, 0x4a2c0e68,
95 SOC15_REG_OFFSET(GC, 0, mmTCP_CHAN_STEER_LO), 0xffffffff, 0xb5d3f197,
96 SOC15_REG_OFFSET(GC, 0, mmVGT_GS_MAX_WAVE_ID), 0x00000fff, 0x000003ff
97 };
98
99 static const u32 golden_settings_gc_9_0_vg10[] =
100 {
101 SOC15_REG_OFFSET(GC, 0, mmCB_HW_CONTROL), 0x0000f000, 0x00012107,
102 SOC15_REG_OFFSET(GC, 0, mmCB_HW_CONTROL_3), 0x30000000, 0x10000000,
103 SOC15_REG_OFFSET(GC, 0, mmGB_ADDR_CONFIG), 0xffff77ff, 0x2a114042,
104 SOC15_REG_OFFSET(GC, 0, mmGB_ADDR_CONFIG_READ), 0xffff77ff, 0x2a114042,
105 SOC15_REG_OFFSET(GC, 0, mmPA_SC_ENHANCE_1), 0x00008000, 0x00048000,
106 SOC15_REG_OFFSET(GC, 0, mmRMI_UTCL1_CNTL2), 0x00030000, 0x00020000,
107 SOC15_REG_OFFSET(GC, 0, mmTD_CNTL), 0x00001800, 0x00000800,
108 SOC15_REG_OFFSET(GC, 0, mmSPI_CONFIG_CNTL_1),0x0000000f, 0x00000007
109 };
110
111 #define VEGA10_GB_ADDR_CONFIG_GOLDEN 0x2a114042
112
113 static void gfx_v9_0_set_ring_funcs(struct amdgpu_device *adev);
114 static void gfx_v9_0_set_irq_funcs(struct amdgpu_device *adev);
115 static void gfx_v9_0_set_gds_init(struct amdgpu_device *adev);
116 static void gfx_v9_0_set_rlc_funcs(struct amdgpu_device *adev);
117 static int gfx_v9_0_get_cu_info(struct amdgpu_device *adev,
118 struct amdgpu_cu_info *cu_info);
119 static uint64_t gfx_v9_0_get_gpu_clock_counter(struct amdgpu_device *adev);
120 static void gfx_v9_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, u32 sh_num, u32 instance);
121
122 static void gfx_v9_0_init_golden_registers(struct amdgpu_device *adev)
123 {
124 switch (adev->asic_type) {
125 case CHIP_VEGA10:
126 amdgpu_program_register_sequence(adev,
127 golden_settings_gc_9_0,
128 (const u32)ARRAY_SIZE(golden_settings_gc_9_0));
129 amdgpu_program_register_sequence(adev,
130 golden_settings_gc_9_0_vg10,
131 (const u32)ARRAY_SIZE(golden_settings_gc_9_0_vg10));
132 break;
133 default:
134 break;
135 }
136 }
137
138 static void gfx_v9_0_scratch_init(struct amdgpu_device *adev)
139 {
140 adev->gfx.scratch.num_reg = 7;
141 adev->gfx.scratch.reg_base = SOC15_REG_OFFSET(GC, 0, mmSCRATCH_REG0);
142 adev->gfx.scratch.free_mask = (1u << adev->gfx.scratch.num_reg) - 1;
143 }
144
145 static void gfx_v9_0_write_data_to_reg(struct amdgpu_ring *ring, int eng_sel,
146 bool wc, uint32_t reg, uint32_t val)
147 {
148 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
149 amdgpu_ring_write(ring, WRITE_DATA_ENGINE_SEL(eng_sel) |
150 WRITE_DATA_DST_SEL(0) |
151 (wc ? WR_CONFIRM : 0));
152 amdgpu_ring_write(ring, reg);
153 amdgpu_ring_write(ring, 0);
154 amdgpu_ring_write(ring, val);
155 }
156
157 static void gfx_v9_0_wait_reg_mem(struct amdgpu_ring *ring, int eng_sel,
158 int mem_space, int opt, uint32_t addr0,
159 uint32_t addr1, uint32_t ref, uint32_t mask,
160 uint32_t inv)
161 {
162 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5));
163 amdgpu_ring_write(ring,
164 /* memory (1) or register (0) */
165 (WAIT_REG_MEM_MEM_SPACE(mem_space) |
166 WAIT_REG_MEM_OPERATION(opt) | /* wait */
167 WAIT_REG_MEM_FUNCTION(3) | /* equal */
168 WAIT_REG_MEM_ENGINE(eng_sel)));
169
170 if (mem_space)
171 BUG_ON(addr0 & 0x3); /* Dword align */
172 amdgpu_ring_write(ring, addr0);
173 amdgpu_ring_write(ring, addr1);
174 amdgpu_ring_write(ring, ref);
175 amdgpu_ring_write(ring, mask);
176 amdgpu_ring_write(ring, inv); /* poll interval */
177 }
178
179 static int gfx_v9_0_ring_test_ring(struct amdgpu_ring *ring)
180 {
181 struct amdgpu_device *adev = ring->adev;
182 uint32_t scratch;
183 uint32_t tmp = 0;
184 unsigned i;
185 int r;
186
187 r = amdgpu_gfx_scratch_get(adev, &scratch);
188 if (r) {
189 DRM_ERROR("amdgpu: cp failed to get scratch reg (%d).\n", r);
190 return r;
191 }
192 WREG32(scratch, 0xCAFEDEAD);
193 r = amdgpu_ring_alloc(ring, 3);
194 if (r) {
195 DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n",
196 ring->idx, r);
197 amdgpu_gfx_scratch_free(adev, scratch);
198 return r;
199 }
200 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1));
201 amdgpu_ring_write(ring, (scratch - PACKET3_SET_UCONFIG_REG_START));
202 amdgpu_ring_write(ring, 0xDEADBEEF);
203 amdgpu_ring_commit(ring);
204
205 for (i = 0; i < adev->usec_timeout; i++) {
206 tmp = RREG32(scratch);
207 if (tmp == 0xDEADBEEF)
208 break;
209 DRM_UDELAY(1);
210 }
211 if (i < adev->usec_timeout) {
212 DRM_INFO("ring test on %d succeeded in %d usecs\n",
213 ring->idx, i);
214 } else {
215 DRM_ERROR("amdgpu: ring %d test failed (scratch(0x%04X)=0x%08X)\n",
216 ring->idx, scratch, tmp);
217 r = -EINVAL;
218 }
219 amdgpu_gfx_scratch_free(adev, scratch);
220 return r;
221 }
222
223 static int gfx_v9_0_ring_test_ib(struct amdgpu_ring *ring, long timeout)
224 {
225 struct amdgpu_device *adev = ring->adev;
226 struct amdgpu_ib ib;
227 struct dma_fence *f = NULL;
228 uint32_t scratch;
229 uint32_t tmp = 0;
230 long r;
231
232 r = amdgpu_gfx_scratch_get(adev, &scratch);
233 if (r) {
234 DRM_ERROR("amdgpu: failed to get scratch reg (%ld).\n", r);
235 return r;
236 }
237 WREG32(scratch, 0xCAFEDEAD);
238 memset(&ib, 0, sizeof(ib));
239 r = amdgpu_ib_get(adev, NULL, 256, &ib);
240 if (r) {
241 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r);
242 goto err1;
243 }
244 ib.ptr[0] = PACKET3(PACKET3_SET_UCONFIG_REG, 1);
245 ib.ptr[1] = ((scratch - PACKET3_SET_UCONFIG_REG_START));
246 ib.ptr[2] = 0xDEADBEEF;
247 ib.length_dw = 3;
248
249 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f);
250 if (r)
251 goto err2;
252
253 r = dma_fence_wait_timeout(f, false, timeout);
254 if (r == 0) {
255 DRM_ERROR("amdgpu: IB test timed out.\n");
256 r = -ETIMEDOUT;
257 goto err2;
258 } else if (r < 0) {
259 DRM_ERROR("amdgpu: fence wait failed (%ld).\n", r);
260 goto err2;
261 }
262 tmp = RREG32(scratch);
263 if (tmp == 0xDEADBEEF) {
264 DRM_INFO("ib test on ring %d succeeded\n", ring->idx);
265 r = 0;
266 } else {
267 DRM_ERROR("amdgpu: ib test failed (scratch(0x%04X)=0x%08X)\n",
268 scratch, tmp);
269 r = -EINVAL;
270 }
271 err2:
272 amdgpu_ib_free(adev, &ib, NULL);
273 dma_fence_put(f);
274 err1:
275 amdgpu_gfx_scratch_free(adev, scratch);
276 return r;
277 }
278
279 static int gfx_v9_0_init_microcode(struct amdgpu_device *adev)
280 {
281 const char *chip_name;
282 char fw_name[30];
283 int err;
284 struct amdgpu_firmware_info *info = NULL;
285 const struct common_firmware_header *header = NULL;
286 const struct gfx_firmware_header_v1_0 *cp_hdr;
287
288 DRM_DEBUG("\n");
289
290 switch (adev->asic_type) {
291 case CHIP_VEGA10:
292 chip_name = "vega10";
293 break;
294 default:
295 BUG();
296 }
297
298 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_pfp.bin", chip_name);
299 err = request_firmware(&adev->gfx.pfp_fw, fw_name, adev->dev);
300 if (err)
301 goto out;
302 err = amdgpu_ucode_validate(adev->gfx.pfp_fw);
303 if (err)
304 goto out;
305 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.pfp_fw->data;
306 adev->gfx.pfp_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
307 adev->gfx.pfp_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
308
309 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_me.bin", chip_name);
310 err = request_firmware(&adev->gfx.me_fw, fw_name, adev->dev);
311 if (err)
312 goto out;
313 err = amdgpu_ucode_validate(adev->gfx.me_fw);
314 if (err)
315 goto out;
316 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.me_fw->data;
317 adev->gfx.me_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
318 adev->gfx.me_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
319
320 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_ce.bin", chip_name);
321 err = request_firmware(&adev->gfx.ce_fw, fw_name, adev->dev);
322 if (err)
323 goto out;
324 err = amdgpu_ucode_validate(adev->gfx.ce_fw);
325 if (err)
326 goto out;
327 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.ce_fw->data;
328 adev->gfx.ce_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
329 adev->gfx.ce_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
330
331 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_rlc.bin", chip_name);
332 err = request_firmware(&adev->gfx.rlc_fw, fw_name, adev->dev);
333 if (err)
334 goto out;
335 err = amdgpu_ucode_validate(adev->gfx.rlc_fw);
336 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.rlc_fw->data;
337 adev->gfx.rlc_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
338 adev->gfx.rlc_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
339
340 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mec.bin", chip_name);
341 err = request_firmware(&adev->gfx.mec_fw, fw_name, adev->dev);
342 if (err)
343 goto out;
344 err = amdgpu_ucode_validate(adev->gfx.mec_fw);
345 if (err)
346 goto out;
347 cp_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
348 adev->gfx.mec_fw_version = le32_to_cpu(cp_hdr->header.ucode_version);
349 adev->gfx.mec_feature_version = le32_to_cpu(cp_hdr->ucode_feature_version);
350
351
352 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mec2.bin", chip_name);
353 err = request_firmware(&adev->gfx.mec2_fw, fw_name, adev->dev);
354 if (!err) {
355 err = amdgpu_ucode_validate(adev->gfx.mec2_fw);
356 if (err)
357 goto out;
358 cp_hdr = (const struct gfx_firmware_header_v1_0 *)
359 adev->gfx.mec2_fw->data;
360 adev->gfx.mec2_fw_version =
361 le32_to_cpu(cp_hdr->header.ucode_version);
362 adev->gfx.mec2_feature_version =
363 le32_to_cpu(cp_hdr->ucode_feature_version);
364 } else {
365 err = 0;
366 adev->gfx.mec2_fw = NULL;
367 }
368
369 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) {
370 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_PFP];
371 info->ucode_id = AMDGPU_UCODE_ID_CP_PFP;
372 info->fw = adev->gfx.pfp_fw;
373 header = (const struct common_firmware_header *)info->fw->data;
374 adev->firmware.fw_size +=
375 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
376
377 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_ME];
378 info->ucode_id = AMDGPU_UCODE_ID_CP_ME;
379 info->fw = adev->gfx.me_fw;
380 header = (const struct common_firmware_header *)info->fw->data;
381 adev->firmware.fw_size +=
382 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
383
384 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_CE];
385 info->ucode_id = AMDGPU_UCODE_ID_CP_CE;
386 info->fw = adev->gfx.ce_fw;
387 header = (const struct common_firmware_header *)info->fw->data;
388 adev->firmware.fw_size +=
389 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
390
391 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_RLC_G];
392 info->ucode_id = AMDGPU_UCODE_ID_RLC_G;
393 info->fw = adev->gfx.rlc_fw;
394 header = (const struct common_firmware_header *)info->fw->data;
395 adev->firmware.fw_size +=
396 ALIGN(le32_to_cpu(header->ucode_size_bytes), PAGE_SIZE);
397
398 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC1];
399 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC1;
400 info->fw = adev->gfx.mec_fw;
401 header = (const struct common_firmware_header *)info->fw->data;
402 cp_hdr = (const struct gfx_firmware_header_v1_0 *)info->fw->data;
403 adev->firmware.fw_size +=
404 ALIGN(le32_to_cpu(header->ucode_size_bytes) - le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
405
406 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC1_JT];
407 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC1_JT;
408 info->fw = adev->gfx.mec_fw;
409 adev->firmware.fw_size +=
410 ALIGN(le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
411
412 if (adev->gfx.mec2_fw) {
413 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC2];
414 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC2;
415 info->fw = adev->gfx.mec2_fw;
416 header = (const struct common_firmware_header *)info->fw->data;
417 cp_hdr = (const struct gfx_firmware_header_v1_0 *)info->fw->data;
418 adev->firmware.fw_size +=
419 ALIGN(le32_to_cpu(header->ucode_size_bytes) - le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
420 info = &adev->firmware.ucode[AMDGPU_UCODE_ID_CP_MEC2_JT];
421 info->ucode_id = AMDGPU_UCODE_ID_CP_MEC2_JT;
422 info->fw = adev->gfx.mec2_fw;
423 adev->firmware.fw_size +=
424 ALIGN(le32_to_cpu(cp_hdr->jt_size) * 4, PAGE_SIZE);
425 }
426
427 }
428
429 out:
430 if (err) {
431 dev_err(adev->dev,
432 "gfx9: Failed to load firmware \"%s\"\n",
433 fw_name);
434 release_firmware(adev->gfx.pfp_fw);
435 adev->gfx.pfp_fw = NULL;
436 release_firmware(adev->gfx.me_fw);
437 adev->gfx.me_fw = NULL;
438 release_firmware(adev->gfx.ce_fw);
439 adev->gfx.ce_fw = NULL;
440 release_firmware(adev->gfx.rlc_fw);
441 adev->gfx.rlc_fw = NULL;
442 release_firmware(adev->gfx.mec_fw);
443 adev->gfx.mec_fw = NULL;
444 release_firmware(adev->gfx.mec2_fw);
445 adev->gfx.mec2_fw = NULL;
446 }
447 return err;
448 }
449
450 static void gfx_v9_0_mec_fini(struct amdgpu_device *adev)
451 {
452 int r;
453
454 if (adev->gfx.mec.hpd_eop_obj) {
455 r = amdgpu_bo_reserve(adev->gfx.mec.hpd_eop_obj, true);
456 if (unlikely(r != 0))
457 dev_warn(adev->dev, "(%d) reserve HPD EOP bo failed\n", r);
458 amdgpu_bo_unpin(adev->gfx.mec.hpd_eop_obj);
459 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
460
461 amdgpu_bo_unref(&adev->gfx.mec.hpd_eop_obj);
462 adev->gfx.mec.hpd_eop_obj = NULL;
463 }
464 if (adev->gfx.mec.mec_fw_obj) {
465 r = amdgpu_bo_reserve(adev->gfx.mec.mec_fw_obj, true);
466 if (unlikely(r != 0))
467 dev_warn(adev->dev, "(%d) reserve mec firmware bo failed\n", r);
468 amdgpu_bo_unpin(adev->gfx.mec.mec_fw_obj);
469 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj);
470
471 amdgpu_bo_unref(&adev->gfx.mec.mec_fw_obj);
472 adev->gfx.mec.mec_fw_obj = NULL;
473 }
474 }
475
476 #define MEC_HPD_SIZE 2048
477
478 static int gfx_v9_0_mec_init(struct amdgpu_device *adev)
479 {
480 int r;
481 u32 *hpd;
482 const __le32 *fw_data;
483 unsigned fw_size;
484 u32 *fw;
485
486 const struct gfx_firmware_header_v1_0 *mec_hdr;
487
488 /*
489 * we assign only 1 pipe because all other pipes will
490 * be handled by KFD
491 */
492 adev->gfx.mec.num_mec = 1;
493 adev->gfx.mec.num_pipe = 1;
494 adev->gfx.mec.num_queue = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe * 8;
495
496 if (adev->gfx.mec.hpd_eop_obj == NULL) {
497 r = amdgpu_bo_create(adev,
498 adev->gfx.mec.num_queue * MEC_HPD_SIZE,
499 PAGE_SIZE, true,
500 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL,
501 &adev->gfx.mec.hpd_eop_obj);
502 if (r) {
503 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r);
504 return r;
505 }
506 }
507
508 r = amdgpu_bo_reserve(adev->gfx.mec.hpd_eop_obj, false);
509 if (unlikely(r != 0)) {
510 gfx_v9_0_mec_fini(adev);
511 return r;
512 }
513 r = amdgpu_bo_pin(adev->gfx.mec.hpd_eop_obj, AMDGPU_GEM_DOMAIN_GTT,
514 &adev->gfx.mec.hpd_eop_gpu_addr);
515 if (r) {
516 dev_warn(adev->dev, "(%d) pin HDP EOP bo failed\n", r);
517 gfx_v9_0_mec_fini(adev);
518 return r;
519 }
520 r = amdgpu_bo_kmap(adev->gfx.mec.hpd_eop_obj, (void **)&hpd);
521 if (r) {
522 dev_warn(adev->dev, "(%d) map HDP EOP bo failed\n", r);
523 gfx_v9_0_mec_fini(adev);
524 return r;
525 }
526
527 memset(hpd, 0, adev->gfx.mec.hpd_eop_obj->tbo.mem.size);
528
529 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj);
530 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj);
531
532 mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
533
534 fw_data = (const __le32 *)
535 (adev->gfx.mec_fw->data +
536 le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes));
537 fw_size = le32_to_cpu(mec_hdr->header.ucode_size_bytes) / 4;
538
539 if (adev->gfx.mec.mec_fw_obj == NULL) {
540 r = amdgpu_bo_create(adev,
541 mec_hdr->header.ucode_size_bytes,
542 PAGE_SIZE, true,
543 AMDGPU_GEM_DOMAIN_GTT, 0, NULL, NULL,
544 &adev->gfx.mec.mec_fw_obj);
545 if (r) {
546 dev_warn(adev->dev, "(%d) create mec firmware bo failed\n", r);
547 return r;
548 }
549 }
550
551 r = amdgpu_bo_reserve(adev->gfx.mec.mec_fw_obj, false);
552 if (unlikely(r != 0)) {
553 gfx_v9_0_mec_fini(adev);
554 return r;
555 }
556 r = amdgpu_bo_pin(adev->gfx.mec.mec_fw_obj, AMDGPU_GEM_DOMAIN_GTT,
557 &adev->gfx.mec.mec_fw_gpu_addr);
558 if (r) {
559 dev_warn(adev->dev, "(%d) pin mec firmware bo failed\n", r);
560 gfx_v9_0_mec_fini(adev);
561 return r;
562 }
563 r = amdgpu_bo_kmap(adev->gfx.mec.mec_fw_obj, (void **)&fw);
564 if (r) {
565 dev_warn(adev->dev, "(%d) map firmware bo failed\n", r);
566 gfx_v9_0_mec_fini(adev);
567 return r;
568 }
569 memcpy(fw, fw_data, fw_size);
570
571 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj);
572 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj);
573
574
575 return 0;
576 }
577
578 static void gfx_v9_0_kiq_fini(struct amdgpu_device *adev)
579 {
580 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
581
582 amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
583 }
584
585 static int gfx_v9_0_kiq_init(struct amdgpu_device *adev)
586 {
587 int r;
588 u32 *hpd;
589 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
590
591 r = amdgpu_bo_create_kernel(adev, MEC_HPD_SIZE, PAGE_SIZE,
592 AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
593 &kiq->eop_gpu_addr, (void **)&hpd);
594 if (r) {
595 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
596 return r;
597 }
598
599 memset(hpd, 0, MEC_HPD_SIZE);
600
601 r = amdgpu_bo_reserve(kiq->eop_obj, true);
602 if (unlikely(r != 0))
603 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
604 amdgpu_bo_kunmap(kiq->eop_obj);
605 amdgpu_bo_unreserve(kiq->eop_obj);
606
607 return 0;
608 }
609
610 static int gfx_v9_0_kiq_init_ring(struct amdgpu_device *adev,
611 struct amdgpu_ring *ring,
612 struct amdgpu_irq_src *irq)
613 {
614 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
615 int r = 0;
616
617 r = amdgpu_wb_get(adev, &adev->virt.reg_val_offs);
618 if (r)
619 return r;
620
621 ring->adev = NULL;
622 ring->ring_obj = NULL;
623 ring->use_doorbell = true;
624 ring->doorbell_index = AMDGPU_DOORBELL_KIQ;
625 if (adev->gfx.mec2_fw) {
626 ring->me = 2;
627 ring->pipe = 0;
628 } else {
629 ring->me = 1;
630 ring->pipe = 1;
631 }
632
633 ring->queue = 0;
634 ring->eop_gpu_addr = kiq->eop_gpu_addr;
635 sprintf(ring->name, "kiq %d.%d.%d", ring->me, ring->pipe, ring->queue);
636 r = amdgpu_ring_init(adev, ring, 1024,
637 irq, AMDGPU_CP_KIQ_IRQ_DRIVER0);
638 if (r)
639 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
640
641 return r;
642 }
643 static void gfx_v9_0_kiq_free_ring(struct amdgpu_ring *ring,
644 struct amdgpu_irq_src *irq)
645 {
646 amdgpu_wb_free(ring->adev, ring->adev->virt.reg_val_offs);
647 amdgpu_ring_fini(ring);
648 }
649
650 /* create MQD for each compute queue */
651 static int gfx_v9_0_compute_mqd_sw_init(struct amdgpu_device *adev)
652 {
653 struct amdgpu_ring *ring = NULL;
654 int r, i;
655
656 /* create MQD for KIQ */
657 ring = &adev->gfx.kiq.ring;
658 if (!ring->mqd_obj) {
659 r = amdgpu_bo_create_kernel(adev, sizeof(struct v9_mqd), PAGE_SIZE,
660 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
661 &ring->mqd_gpu_addr, (void **)&ring->mqd_ptr);
662 if (r) {
663 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
664 return r;
665 }
666
667 /*TODO: prepare MQD backup */
668 }
669
670 /* create MQD for each KCQ */
671 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
672 ring = &adev->gfx.compute_ring[i];
673 if (!ring->mqd_obj) {
674 r = amdgpu_bo_create_kernel(adev, sizeof(struct v9_mqd), PAGE_SIZE,
675 AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
676 &ring->mqd_gpu_addr, (void **)&ring->mqd_ptr);
677 if (r) {
678 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
679 return r;
680 }
681
682 /* TODO: prepare MQD backup */
683 }
684 }
685
686 return 0;
687 }
688
689 static void gfx_v9_0_compute_mqd_sw_fini(struct amdgpu_device *adev)
690 {
691 struct amdgpu_ring *ring = NULL;
692 int i;
693
694 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
695 ring = &adev->gfx.compute_ring[i];
696 amdgpu_bo_free_kernel(&ring->mqd_obj, &ring->mqd_gpu_addr, (void **)&ring->mqd_ptr);
697 }
698
699 ring = &adev->gfx.kiq.ring;
700 amdgpu_bo_free_kernel(&ring->mqd_obj, &ring->mqd_gpu_addr, (void **)&ring->mqd_ptr);
701 }
702
703 static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t address)
704 {
705 WREG32_SOC15(GC, 0, mmSQ_IND_INDEX,
706 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
707 (simd << SQ_IND_INDEX__SIMD_ID__SHIFT) |
708 (address << SQ_IND_INDEX__INDEX__SHIFT) |
709 (SQ_IND_INDEX__FORCE_READ_MASK));
710 return RREG32_SOC15(GC, 0, mmSQ_IND_DATA);
711 }
712
713 static void wave_read_regs(struct amdgpu_device *adev, uint32_t simd,
714 uint32_t wave, uint32_t thread,
715 uint32_t regno, uint32_t num, uint32_t *out)
716 {
717 WREG32_SOC15(GC, 0, mmSQ_IND_INDEX,
718 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) |
719 (simd << SQ_IND_INDEX__SIMD_ID__SHIFT) |
720 (regno << SQ_IND_INDEX__INDEX__SHIFT) |
721 (thread << SQ_IND_INDEX__THREAD_ID__SHIFT) |
722 (SQ_IND_INDEX__FORCE_READ_MASK) |
723 (SQ_IND_INDEX__AUTO_INCR_MASK));
724 while (num--)
725 *(out++) = RREG32_SOC15(GC, 0, mmSQ_IND_DATA);
726 }
727
728 static void gfx_v9_0_read_wave_data(struct amdgpu_device *adev, uint32_t simd, uint32_t wave, uint32_t *dst, int *no_fields)
729 {
730 /* type 1 wave data */
731 dst[(*no_fields)++] = 1;
732 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_STATUS);
733 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_PC_LO);
734 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_PC_HI);
735 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_EXEC_LO);
736 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_EXEC_HI);
737 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_HW_ID);
738 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_INST_DW0);
739 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_INST_DW1);
740 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_GPR_ALLOC);
741 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_LDS_ALLOC);
742 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_TRAPSTS);
743 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_IB_STS);
744 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_IB_DBG0);
745 dst[(*no_fields)++] = wave_read_ind(adev, simd, wave, ixSQ_WAVE_M0);
746 }
747
748 static void gfx_v9_0_read_wave_sgprs(struct amdgpu_device *adev, uint32_t simd,
749 uint32_t wave, uint32_t start,
750 uint32_t size, uint32_t *dst)
751 {
752 wave_read_regs(
753 adev, simd, wave, 0,
754 start + SQIND_WAVE_SGPRS_OFFSET, size, dst);
755 }
756
757
758 static const struct amdgpu_gfx_funcs gfx_v9_0_gfx_funcs = {
759 .get_gpu_clock_counter = &gfx_v9_0_get_gpu_clock_counter,
760 .select_se_sh = &gfx_v9_0_select_se_sh,
761 .read_wave_data = &gfx_v9_0_read_wave_data,
762 .read_wave_sgprs = &gfx_v9_0_read_wave_sgprs,
763 };
764
765 static void gfx_v9_0_gpu_early_init(struct amdgpu_device *adev)
766 {
767 u32 gb_addr_config;
768
769 adev->gfx.funcs = &gfx_v9_0_gfx_funcs;
770
771 switch (adev->asic_type) {
772 case CHIP_VEGA10:
773 adev->gfx.config.max_shader_engines = 4;
774 adev->gfx.config.max_cu_per_sh = 16;
775 adev->gfx.config.max_sh_per_se = 1;
776 adev->gfx.config.max_backends_per_se = 4;
777 adev->gfx.config.max_texture_channel_caches = 16;
778 adev->gfx.config.max_gprs = 256;
779 adev->gfx.config.max_gs_threads = 32;
780 adev->gfx.config.max_hw_contexts = 8;
781
782 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20;
783 adev->gfx.config.sc_prim_fifo_size_backend = 0x100;
784 adev->gfx.config.sc_hiz_tile_fifo_size = 0x30;
785 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0;
786 adev->gfx.config.gs_vgt_table_depth = 32;
787 adev->gfx.config.gs_prim_buffer_depth = 1792;
788 gb_addr_config = VEGA10_GB_ADDR_CONFIG_GOLDEN;
789 break;
790 default:
791 BUG();
792 break;
793 }
794
795 adev->gfx.config.gb_addr_config = gb_addr_config;
796
797 adev->gfx.config.gb_addr_config_fields.num_pipes = 1 <<
798 REG_GET_FIELD(
799 adev->gfx.config.gb_addr_config,
800 GB_ADDR_CONFIG,
801 NUM_PIPES);
802
803 adev->gfx.config.max_tile_pipes =
804 adev->gfx.config.gb_addr_config_fields.num_pipes;
805
806 adev->gfx.config.gb_addr_config_fields.num_banks = 1 <<
807 REG_GET_FIELD(
808 adev->gfx.config.gb_addr_config,
809 GB_ADDR_CONFIG,
810 NUM_BANKS);
811 adev->gfx.config.gb_addr_config_fields.max_compress_frags = 1 <<
812 REG_GET_FIELD(
813 adev->gfx.config.gb_addr_config,
814 GB_ADDR_CONFIG,
815 MAX_COMPRESSED_FRAGS);
816 adev->gfx.config.gb_addr_config_fields.num_rb_per_se = 1 <<
817 REG_GET_FIELD(
818 adev->gfx.config.gb_addr_config,
819 GB_ADDR_CONFIG,
820 NUM_RB_PER_SE);
821 adev->gfx.config.gb_addr_config_fields.num_se = 1 <<
822 REG_GET_FIELD(
823 adev->gfx.config.gb_addr_config,
824 GB_ADDR_CONFIG,
825 NUM_SHADER_ENGINES);
826 adev->gfx.config.gb_addr_config_fields.pipe_interleave_size = 1 << (8 +
827 REG_GET_FIELD(
828 adev->gfx.config.gb_addr_config,
829 GB_ADDR_CONFIG,
830 PIPE_INTERLEAVE_SIZE));
831 }
832
833 static int gfx_v9_0_ngg_create_buf(struct amdgpu_device *adev,
834 struct amdgpu_ngg_buf *ngg_buf,
835 int size_se,
836 int default_size_se)
837 {
838 int r;
839
840 if (size_se < 0) {
841 dev_err(adev->dev, "Buffer size is invalid: %d\n", size_se);
842 return -EINVAL;
843 }
844 size_se = size_se ? size_se : default_size_se;
845
846 ngg_buf->size = size_se * adev->gfx.config.max_shader_engines;
847 r = amdgpu_bo_create_kernel(adev, ngg_buf->size,
848 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
849 &ngg_buf->bo,
850 &ngg_buf->gpu_addr,
851 NULL);
852 if (r) {
853 dev_err(adev->dev, "(%d) failed to create NGG buffer\n", r);
854 return r;
855 }
856 ngg_buf->bo_size = amdgpu_bo_size(ngg_buf->bo);
857
858 return r;
859 }
860
861 static int gfx_v9_0_ngg_fini(struct amdgpu_device *adev)
862 {
863 int i;
864
865 for (i = 0; i < NGG_BUF_MAX; i++)
866 amdgpu_bo_free_kernel(&adev->gfx.ngg.buf[i].bo,
867 &adev->gfx.ngg.buf[i].gpu_addr,
868 NULL);
869
870 memset(&adev->gfx.ngg.buf[0], 0,
871 sizeof(struct amdgpu_ngg_buf) * NGG_BUF_MAX);
872
873 adev->gfx.ngg.init = false;
874
875 return 0;
876 }
877
878 static int gfx_v9_0_ngg_init(struct amdgpu_device *adev)
879 {
880 int r;
881
882 if (!amdgpu_ngg || adev->gfx.ngg.init == true)
883 return 0;
884
885 /* GDS reserve memory: 64 bytes alignment */
886 adev->gfx.ngg.gds_reserve_size = ALIGN(5 * 4, 0x40);
887 adev->gds.mem.total_size -= adev->gfx.ngg.gds_reserve_size;
888 adev->gds.mem.gfx_partition_size -= adev->gfx.ngg.gds_reserve_size;
889 adev->gfx.ngg.gds_reserve_addr = amdgpu_gds_reg_offset[0].mem_base;
890 adev->gfx.ngg.gds_reserve_addr += adev->gds.mem.gfx_partition_size;
891
892 /* Primitive Buffer */
893 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_PRIM],
894 amdgpu_prim_buf_per_se,
895 64 * 1024);
896 if (r) {
897 dev_err(adev->dev, "Failed to create Primitive Buffer\n");
898 goto err;
899 }
900
901 /* Position Buffer */
902 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_POS],
903 amdgpu_pos_buf_per_se,
904 256 * 1024);
905 if (r) {
906 dev_err(adev->dev, "Failed to create Position Buffer\n");
907 goto err;
908 }
909
910 /* Control Sideband */
911 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_CNTL],
912 amdgpu_cntl_sb_buf_per_se,
913 256);
914 if (r) {
915 dev_err(adev->dev, "Failed to create Control Sideband Buffer\n");
916 goto err;
917 }
918
919 /* Parameter Cache, not created by default */
920 if (amdgpu_param_buf_per_se <= 0)
921 goto out;
922
923 r = gfx_v9_0_ngg_create_buf(adev, &adev->gfx.ngg.buf[NGG_PARAM],
924 amdgpu_param_buf_per_se,
925 512 * 1024);
926 if (r) {
927 dev_err(adev->dev, "Failed to create Parameter Cache\n");
928 goto err;
929 }
930
931 out:
932 adev->gfx.ngg.init = true;
933 return 0;
934 err:
935 gfx_v9_0_ngg_fini(adev);
936 return r;
937 }
938
939 static int gfx_v9_0_ngg_en(struct amdgpu_device *adev)
940 {
941 struct amdgpu_ring *ring = &adev->gfx.gfx_ring[0];
942 int r;
943 u32 data;
944 u32 size;
945 u32 base;
946
947 if (!amdgpu_ngg)
948 return 0;
949
950 /* Program buffer size */
951 data = 0;
952 size = adev->gfx.ngg.buf[NGG_PRIM].size / 256;
953 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_1, INDEX_BUF_SIZE, size);
954
955 size = adev->gfx.ngg.buf[NGG_POS].size / 256;
956 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_1, POS_BUF_SIZE, size);
957
958 WREG32_SOC15(GC, 0, mmWD_BUF_RESOURCE_1, data);
959
960 data = 0;
961 size = adev->gfx.ngg.buf[NGG_CNTL].size / 256;
962 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_2, CNTL_SB_BUF_SIZE, size);
963
964 size = adev->gfx.ngg.buf[NGG_PARAM].size / 1024;
965 data = REG_SET_FIELD(data, WD_BUF_RESOURCE_2, PARAM_BUF_SIZE, size);
966
967 WREG32_SOC15(GC, 0, mmWD_BUF_RESOURCE_2, data);
968
969 /* Program buffer base address */
970 base = lower_32_bits(adev->gfx.ngg.buf[NGG_PRIM].gpu_addr);
971 data = REG_SET_FIELD(0, WD_INDEX_BUF_BASE, BASE, base);
972 WREG32_SOC15(GC, 0, mmWD_INDEX_BUF_BASE, data);
973
974 base = upper_32_bits(adev->gfx.ngg.buf[NGG_PRIM].gpu_addr);
975 data = REG_SET_FIELD(0, WD_INDEX_BUF_BASE_HI, BASE_HI, base);
976 WREG32_SOC15(GC, 0, mmWD_INDEX_BUF_BASE_HI, data);
977
978 base = lower_32_bits(adev->gfx.ngg.buf[NGG_POS].gpu_addr);
979 data = REG_SET_FIELD(0, WD_POS_BUF_BASE, BASE, base);
980 WREG32_SOC15(GC, 0, mmWD_POS_BUF_BASE, data);
981
982 base = upper_32_bits(adev->gfx.ngg.buf[NGG_POS].gpu_addr);
983 data = REG_SET_FIELD(0, WD_POS_BUF_BASE_HI, BASE_HI, base);
984 WREG32_SOC15(GC, 0, mmWD_POS_BUF_BASE_HI, data);
985
986 base = lower_32_bits(adev->gfx.ngg.buf[NGG_CNTL].gpu_addr);
987 data = REG_SET_FIELD(0, WD_CNTL_SB_BUF_BASE, BASE, base);
988 WREG32_SOC15(GC, 0, mmWD_CNTL_SB_BUF_BASE, data);
989
990 base = upper_32_bits(adev->gfx.ngg.buf[NGG_CNTL].gpu_addr);
991 data = REG_SET_FIELD(0, WD_CNTL_SB_BUF_BASE_HI, BASE_HI, base);
992 WREG32_SOC15(GC, 0, mmWD_CNTL_SB_BUF_BASE_HI, data);
993
994 /* Clear GDS reserved memory */
995 r = amdgpu_ring_alloc(ring, 17);
996 if (r) {
997 DRM_ERROR("amdgpu: NGG failed to lock ring %d (%d).\n",
998 ring->idx, r);
999 return r;
1000 }
1001
1002 gfx_v9_0_write_data_to_reg(ring, 0, false,
1003 amdgpu_gds_reg_offset[0].mem_size,
1004 (adev->gds.mem.total_size +
1005 adev->gfx.ngg.gds_reserve_size) >>
1006 AMDGPU_GDS_SHIFT);
1007
1008 amdgpu_ring_write(ring, PACKET3(PACKET3_DMA_DATA, 5));
1009 amdgpu_ring_write(ring, (PACKET3_DMA_DATA_CP_SYNC |
1010 PACKET3_DMA_DATA_SRC_SEL(2)));
1011 amdgpu_ring_write(ring, 0);
1012 amdgpu_ring_write(ring, 0);
1013 amdgpu_ring_write(ring, adev->gfx.ngg.gds_reserve_addr);
1014 amdgpu_ring_write(ring, 0);
1015 amdgpu_ring_write(ring, adev->gfx.ngg.gds_reserve_size);
1016
1017
1018 gfx_v9_0_write_data_to_reg(ring, 0, false,
1019 amdgpu_gds_reg_offset[0].mem_size, 0);
1020
1021 amdgpu_ring_commit(ring);
1022
1023 return 0;
1024 }
1025
1026 static int gfx_v9_0_sw_init(void *handle)
1027 {
1028 int i, r;
1029 struct amdgpu_ring *ring;
1030 struct amdgpu_kiq *kiq;
1031 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1032
1033 /* KIQ event */
1034 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 178, &adev->gfx.kiq.irq);
1035 if (r)
1036 return r;
1037
1038 /* EOP Event */
1039 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 181, &adev->gfx.eop_irq);
1040 if (r)
1041 return r;
1042
1043 /* Privileged reg */
1044 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 184,
1045 &adev->gfx.priv_reg_irq);
1046 if (r)
1047 return r;
1048
1049 /* Privileged inst */
1050 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_GRBM_CP, 185,
1051 &adev->gfx.priv_inst_irq);
1052 if (r)
1053 return r;
1054
1055 adev->gfx.gfx_current_status = AMDGPU_GFX_NORMAL_MODE;
1056
1057 gfx_v9_0_scratch_init(adev);
1058
1059 r = gfx_v9_0_init_microcode(adev);
1060 if (r) {
1061 DRM_ERROR("Failed to load gfx firmware!\n");
1062 return r;
1063 }
1064
1065 r = gfx_v9_0_mec_init(adev);
1066 if (r) {
1067 DRM_ERROR("Failed to init MEC BOs!\n");
1068 return r;
1069 }
1070
1071 /* set up the gfx ring */
1072 for (i = 0; i < adev->gfx.num_gfx_rings; i++) {
1073 ring = &adev->gfx.gfx_ring[i];
1074 ring->ring_obj = NULL;
1075 sprintf(ring->name, "gfx");
1076 ring->use_doorbell = true;
1077 ring->doorbell_index = AMDGPU_DOORBELL64_GFX_RING0 << 1;
1078 r = amdgpu_ring_init(adev, ring, 1024,
1079 &adev->gfx.eop_irq, AMDGPU_CP_IRQ_GFX_EOP);
1080 if (r)
1081 return r;
1082 }
1083
1084 /* set up the compute queues */
1085 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
1086 unsigned irq_type;
1087
1088 /* max 32 queues per MEC */
1089 if ((i >= 32) || (i >= AMDGPU_MAX_COMPUTE_RINGS)) {
1090 DRM_ERROR("Too many (%d) compute rings!\n", i);
1091 break;
1092 }
1093 ring = &adev->gfx.compute_ring[i];
1094 ring->ring_obj = NULL;
1095 ring->use_doorbell = true;
1096 ring->doorbell_index = (AMDGPU_DOORBELL64_MEC_RING0 + i) << 1;
1097 ring->me = 1; /* first MEC */
1098 ring->pipe = i / 8;
1099 ring->queue = i % 8;
1100 ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr + (i * MEC_HPD_SIZE);
1101 sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue);
1102 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP + ring->pipe;
1103 /* type-2 packets are deprecated on MEC, use type-3 instead */
1104 r = amdgpu_ring_init(adev, ring, 1024,
1105 &adev->gfx.eop_irq, irq_type);
1106 if (r)
1107 return r;
1108 }
1109
1110 if (amdgpu_sriov_vf(adev)) {
1111 r = gfx_v9_0_kiq_init(adev);
1112 if (r) {
1113 DRM_ERROR("Failed to init KIQ BOs!\n");
1114 return r;
1115 }
1116
1117 kiq = &adev->gfx.kiq;
1118 r = gfx_v9_0_kiq_init_ring(adev, &kiq->ring, &kiq->irq);
1119 if (r)
1120 return r;
1121
1122 /* create MQD for all compute queues as wel as KIQ for SRIOV case */
1123 r = gfx_v9_0_compute_mqd_sw_init(adev);
1124 if (r)
1125 return r;
1126 }
1127
1128 /* reserve GDS, GWS and OA resource for gfx */
1129 r = amdgpu_bo_create_kernel(adev, adev->gds.mem.gfx_partition_size,
1130 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GDS,
1131 &adev->gds.gds_gfx_bo, NULL, NULL);
1132 if (r)
1133 return r;
1134
1135 r = amdgpu_bo_create_kernel(adev, adev->gds.gws.gfx_partition_size,
1136 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GWS,
1137 &adev->gds.gws_gfx_bo, NULL, NULL);
1138 if (r)
1139 return r;
1140
1141 r = amdgpu_bo_create_kernel(adev, adev->gds.oa.gfx_partition_size,
1142 PAGE_SIZE, AMDGPU_GEM_DOMAIN_OA,
1143 &adev->gds.oa_gfx_bo, NULL, NULL);
1144 if (r)
1145 return r;
1146
1147 adev->gfx.ce_ram_size = 0x8000;
1148
1149 gfx_v9_0_gpu_early_init(adev);
1150
1151 r = gfx_v9_0_ngg_init(adev);
1152 if (r)
1153 return r;
1154
1155 return 0;
1156 }
1157
1158
1159 static int gfx_v9_0_sw_fini(void *handle)
1160 {
1161 int i;
1162 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1163
1164 amdgpu_bo_free_kernel(&adev->gds.oa_gfx_bo, NULL, NULL);
1165 amdgpu_bo_free_kernel(&adev->gds.gws_gfx_bo, NULL, NULL);
1166 amdgpu_bo_free_kernel(&adev->gds.gds_gfx_bo, NULL, NULL);
1167
1168 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
1169 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]);
1170 for (i = 0; i < adev->gfx.num_compute_rings; i++)
1171 amdgpu_ring_fini(&adev->gfx.compute_ring[i]);
1172
1173 if (amdgpu_sriov_vf(adev)) {
1174 gfx_v9_0_compute_mqd_sw_fini(adev);
1175 gfx_v9_0_kiq_free_ring(&adev->gfx.kiq.ring, &adev->gfx.kiq.irq);
1176 gfx_v9_0_kiq_fini(adev);
1177 }
1178
1179 gfx_v9_0_mec_fini(adev);
1180 gfx_v9_0_ngg_fini(adev);
1181
1182 return 0;
1183 }
1184
1185
1186 static void gfx_v9_0_tiling_mode_table_init(struct amdgpu_device *adev)
1187 {
1188 /* TODO */
1189 }
1190
1191 static void gfx_v9_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, u32 sh_num, u32 instance)
1192 {
1193 u32 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_BROADCAST_WRITES, 1);
1194
1195 if ((se_num == 0xffffffff) && (sh_num == 0xffffffff)) {
1196 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SH_BROADCAST_WRITES, 1);
1197 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1);
1198 } else if (se_num == 0xffffffff) {
1199 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SH_INDEX, sh_num);
1200 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1);
1201 } else if (sh_num == 0xffffffff) {
1202 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SH_BROADCAST_WRITES, 1);
1203 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num);
1204 } else {
1205 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SH_INDEX, sh_num);
1206 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num);
1207 }
1208 WREG32_SOC15(GC, 0, mmGRBM_GFX_INDEX, data);
1209 }
1210
1211 static u32 gfx_v9_0_create_bitmask(u32 bit_width)
1212 {
1213 return (u32)((1ULL << bit_width) - 1);
1214 }
1215
1216 static u32 gfx_v9_0_get_rb_active_bitmap(struct amdgpu_device *adev)
1217 {
1218 u32 data, mask;
1219
1220 data = RREG32_SOC15(GC, 0, mmCC_RB_BACKEND_DISABLE);
1221 data |= RREG32_SOC15(GC, 0, mmGC_USER_RB_BACKEND_DISABLE);
1222
1223 data &= CC_RB_BACKEND_DISABLE__BACKEND_DISABLE_MASK;
1224 data >>= GC_USER_RB_BACKEND_DISABLE__BACKEND_DISABLE__SHIFT;
1225
1226 mask = gfx_v9_0_create_bitmask(adev->gfx.config.max_backends_per_se /
1227 adev->gfx.config.max_sh_per_se);
1228
1229 return (~data) & mask;
1230 }
1231
1232 static void gfx_v9_0_setup_rb(struct amdgpu_device *adev)
1233 {
1234 int i, j;
1235 u32 data;
1236 u32 active_rbs = 0;
1237 u32 rb_bitmap_width_per_sh = adev->gfx.config.max_backends_per_se /
1238 adev->gfx.config.max_sh_per_se;
1239
1240 mutex_lock(&adev->grbm_idx_mutex);
1241 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
1242 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
1243 gfx_v9_0_select_se_sh(adev, i, j, 0xffffffff);
1244 data = gfx_v9_0_get_rb_active_bitmap(adev);
1245 active_rbs |= data << ((i * adev->gfx.config.max_sh_per_se + j) *
1246 rb_bitmap_width_per_sh);
1247 }
1248 }
1249 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1250 mutex_unlock(&adev->grbm_idx_mutex);
1251
1252 adev->gfx.config.backend_enable_mask = active_rbs;
1253 adev->gfx.config.num_rbs = hweight32(active_rbs);
1254 }
1255
1256 #define DEFAULT_SH_MEM_BASES (0x6000)
1257 #define FIRST_COMPUTE_VMID (8)
1258 #define LAST_COMPUTE_VMID (16)
1259 static void gfx_v9_0_init_compute_vmid(struct amdgpu_device *adev)
1260 {
1261 int i;
1262 uint32_t sh_mem_config;
1263 uint32_t sh_mem_bases;
1264
1265 /*
1266 * Configure apertures:
1267 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB)
1268 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB)
1269 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB)
1270 */
1271 sh_mem_bases = DEFAULT_SH_MEM_BASES | (DEFAULT_SH_MEM_BASES << 16);
1272
1273 sh_mem_config = SH_MEM_ADDRESS_MODE_64 |
1274 SH_MEM_ALIGNMENT_MODE_UNALIGNED <<
1275 SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT;
1276
1277 mutex_lock(&adev->srbm_mutex);
1278 for (i = FIRST_COMPUTE_VMID; i < LAST_COMPUTE_VMID; i++) {
1279 soc15_grbm_select(adev, 0, 0, 0, i);
1280 /* CP and shaders */
1281 WREG32_SOC15(GC, 0, mmSH_MEM_CONFIG, sh_mem_config);
1282 WREG32_SOC15(GC, 0, mmSH_MEM_BASES, sh_mem_bases);
1283 }
1284 soc15_grbm_select(adev, 0, 0, 0, 0);
1285 mutex_unlock(&adev->srbm_mutex);
1286 }
1287
1288 static void gfx_v9_0_gpu_init(struct amdgpu_device *adev)
1289 {
1290 u32 tmp;
1291 int i;
1292
1293 WREG32_FIELD15(GC, 0, GRBM_CNTL, READ_TIMEOUT, 0xff);
1294
1295 gfx_v9_0_tiling_mode_table_init(adev);
1296
1297 gfx_v9_0_setup_rb(adev);
1298 gfx_v9_0_get_cu_info(adev, &adev->gfx.cu_info);
1299
1300 /* XXX SH_MEM regs */
1301 /* where to put LDS, scratch, GPUVM in FSA64 space */
1302 mutex_lock(&adev->srbm_mutex);
1303 for (i = 0; i < 16; i++) {
1304 soc15_grbm_select(adev, 0, 0, 0, i);
1305 /* CP and shaders */
1306 tmp = 0;
1307 tmp = REG_SET_FIELD(tmp, SH_MEM_CONFIG, ALIGNMENT_MODE,
1308 SH_MEM_ALIGNMENT_MODE_UNALIGNED);
1309 WREG32_SOC15(GC, 0, mmSH_MEM_CONFIG, tmp);
1310 WREG32_SOC15(GC, 0, mmSH_MEM_BASES, 0);
1311 }
1312 soc15_grbm_select(adev, 0, 0, 0, 0);
1313
1314 mutex_unlock(&adev->srbm_mutex);
1315
1316 gfx_v9_0_init_compute_vmid(adev);
1317
1318 mutex_lock(&adev->grbm_idx_mutex);
1319 /*
1320 * making sure that the following register writes will be broadcasted
1321 * to all the shaders
1322 */
1323 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1324
1325 WREG32_SOC15(GC, 0, mmPA_SC_FIFO_SIZE,
1326 (adev->gfx.config.sc_prim_fifo_size_frontend <<
1327 PA_SC_FIFO_SIZE__SC_FRONTEND_PRIM_FIFO_SIZE__SHIFT) |
1328 (adev->gfx.config.sc_prim_fifo_size_backend <<
1329 PA_SC_FIFO_SIZE__SC_BACKEND_PRIM_FIFO_SIZE__SHIFT) |
1330 (adev->gfx.config.sc_hiz_tile_fifo_size <<
1331 PA_SC_FIFO_SIZE__SC_HIZ_TILE_FIFO_SIZE__SHIFT) |
1332 (adev->gfx.config.sc_earlyz_tile_fifo_size <<
1333 PA_SC_FIFO_SIZE__SC_EARLYZ_TILE_FIFO_SIZE__SHIFT));
1334 mutex_unlock(&adev->grbm_idx_mutex);
1335
1336 }
1337
1338 static void gfx_v9_0_wait_for_rlc_serdes(struct amdgpu_device *adev)
1339 {
1340 u32 i, j, k;
1341 u32 mask;
1342
1343 mutex_lock(&adev->grbm_idx_mutex);
1344 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
1345 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
1346 gfx_v9_0_select_se_sh(adev, i, j, 0xffffffff);
1347 for (k = 0; k < adev->usec_timeout; k++) {
1348 if (RREG32_SOC15(GC, 0, mmRLC_SERDES_CU_MASTER_BUSY) == 0)
1349 break;
1350 udelay(1);
1351 }
1352 }
1353 }
1354 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
1355 mutex_unlock(&adev->grbm_idx_mutex);
1356
1357 mask = RLC_SERDES_NONCU_MASTER_BUSY__SE_MASTER_BUSY_MASK |
1358 RLC_SERDES_NONCU_MASTER_BUSY__GC_MASTER_BUSY_MASK |
1359 RLC_SERDES_NONCU_MASTER_BUSY__TC0_MASTER_BUSY_MASK |
1360 RLC_SERDES_NONCU_MASTER_BUSY__TC1_MASTER_BUSY_MASK;
1361 for (k = 0; k < adev->usec_timeout; k++) {
1362 if ((RREG32_SOC15(GC, 0, mmRLC_SERDES_NONCU_MASTER_BUSY) & mask) == 0)
1363 break;
1364 udelay(1);
1365 }
1366 }
1367
1368 static void gfx_v9_0_enable_gui_idle_interrupt(struct amdgpu_device *adev,
1369 bool enable)
1370 {
1371 u32 tmp = RREG32_SOC15(GC, 0, mmCP_INT_CNTL_RING0);
1372
1373 if (enable)
1374 return;
1375
1376 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE, enable ? 1 : 0);
1377 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_EMPTY_INT_ENABLE, enable ? 1 : 0);
1378 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CMP_BUSY_INT_ENABLE, enable ? 1 : 0);
1379 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, GFX_IDLE_INT_ENABLE, enable ? 1 : 0);
1380
1381 WREG32_SOC15(GC, 0, mmCP_INT_CNTL_RING0, tmp);
1382 }
1383
1384 void gfx_v9_0_rlc_stop(struct amdgpu_device *adev)
1385 {
1386 u32 tmp = RREG32_SOC15(GC, 0, mmRLC_CNTL);
1387
1388 tmp = REG_SET_FIELD(tmp, RLC_CNTL, RLC_ENABLE_F32, 0);
1389 WREG32_SOC15(GC, 0, mmRLC_CNTL, tmp);
1390
1391 gfx_v9_0_enable_gui_idle_interrupt(adev, false);
1392
1393 gfx_v9_0_wait_for_rlc_serdes(adev);
1394 }
1395
1396 static void gfx_v9_0_rlc_reset(struct amdgpu_device *adev)
1397 {
1398 WREG32_FIELD15(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1);
1399 udelay(50);
1400 WREG32_FIELD15(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0);
1401 udelay(50);
1402 }
1403
1404 static void gfx_v9_0_rlc_start(struct amdgpu_device *adev)
1405 {
1406 #ifdef AMDGPU_RLC_DEBUG_RETRY
1407 u32 rlc_ucode_ver;
1408 #endif
1409
1410 WREG32_FIELD15(GC, 0, RLC_CNTL, RLC_ENABLE_F32, 1);
1411
1412 /* carrizo do enable cp interrupt after cp inited */
1413 if (!(adev->flags & AMD_IS_APU))
1414 gfx_v9_0_enable_gui_idle_interrupt(adev, true);
1415
1416 udelay(50);
1417
1418 #ifdef AMDGPU_RLC_DEBUG_RETRY
1419 /* RLC_GPM_GENERAL_6 : RLC Ucode version */
1420 rlc_ucode_ver = RREG32_SOC15(GC, 0, mmRLC_GPM_GENERAL_6);
1421 if(rlc_ucode_ver == 0x108) {
1422 DRM_INFO("Using rlc debug ucode. mmRLC_GPM_GENERAL_6 ==0x08%x / fw_ver == %i \n",
1423 rlc_ucode_ver, adev->gfx.rlc_fw_version);
1424 /* RLC_GPM_TIMER_INT_3 : Timer interval in RefCLK cycles,
1425 * default is 0x9C4 to create a 100us interval */
1426 WREG32_SOC15(GC, 0, mmRLC_GPM_TIMER_INT_3, 0x9C4);
1427 /* RLC_GPM_GENERAL_12 : Minimum gap between wptr and rptr
1428 * to disable the page fault retry interrupts, default is
1429 * 0x100 (256) */
1430 WREG32_SOC15(GC, 0, mmRLC_GPM_GENERAL_12, 0x100);
1431 }
1432 #endif
1433 }
1434
1435 static int gfx_v9_0_rlc_load_microcode(struct amdgpu_device *adev)
1436 {
1437 const struct rlc_firmware_header_v2_0 *hdr;
1438 const __le32 *fw_data;
1439 unsigned i, fw_size;
1440
1441 if (!adev->gfx.rlc_fw)
1442 return -EINVAL;
1443
1444 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data;
1445 amdgpu_ucode_print_rlc_hdr(&hdr->header);
1446
1447 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data +
1448 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1449 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
1450
1451 WREG32_SOC15(GC, 0, mmRLC_GPM_UCODE_ADDR,
1452 RLCG_UCODE_LOADING_START_ADDRESS);
1453 for (i = 0; i < fw_size; i++)
1454 WREG32_SOC15(GC, 0, mmRLC_GPM_UCODE_DATA, le32_to_cpup(fw_data++));
1455 WREG32_SOC15(GC, 0, mmRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version);
1456
1457 return 0;
1458 }
1459
1460 static int gfx_v9_0_rlc_resume(struct amdgpu_device *adev)
1461 {
1462 int r;
1463
1464 if (amdgpu_sriov_vf(adev))
1465 return 0;
1466
1467 gfx_v9_0_rlc_stop(adev);
1468
1469 /* disable CG */
1470 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL, 0);
1471
1472 /* disable PG */
1473 WREG32_SOC15(GC, 0, mmRLC_PG_CNTL, 0);
1474
1475 gfx_v9_0_rlc_reset(adev);
1476
1477 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
1478 /* legacy rlc firmware loading */
1479 r = gfx_v9_0_rlc_load_microcode(adev);
1480 if (r)
1481 return r;
1482 }
1483
1484 gfx_v9_0_rlc_start(adev);
1485
1486 return 0;
1487 }
1488
1489 static void gfx_v9_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable)
1490 {
1491 int i;
1492 u32 tmp = RREG32_SOC15(GC, 0, mmCP_ME_CNTL);
1493
1494 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_HALT, enable ? 0 : 1);
1495 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_HALT, enable ? 0 : 1);
1496 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, CE_HALT, enable ? 0 : 1);
1497 if (!enable) {
1498 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
1499 adev->gfx.gfx_ring[i].ready = false;
1500 }
1501 WREG32_SOC15(GC, 0, mmCP_ME_CNTL, tmp);
1502 udelay(50);
1503 }
1504
1505 static int gfx_v9_0_cp_gfx_load_microcode(struct amdgpu_device *adev)
1506 {
1507 const struct gfx_firmware_header_v1_0 *pfp_hdr;
1508 const struct gfx_firmware_header_v1_0 *ce_hdr;
1509 const struct gfx_firmware_header_v1_0 *me_hdr;
1510 const __le32 *fw_data;
1511 unsigned i, fw_size;
1512
1513 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw || !adev->gfx.ce_fw)
1514 return -EINVAL;
1515
1516 pfp_hdr = (const struct gfx_firmware_header_v1_0 *)
1517 adev->gfx.pfp_fw->data;
1518 ce_hdr = (const struct gfx_firmware_header_v1_0 *)
1519 adev->gfx.ce_fw->data;
1520 me_hdr = (const struct gfx_firmware_header_v1_0 *)
1521 adev->gfx.me_fw->data;
1522
1523 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header);
1524 amdgpu_ucode_print_gfx_hdr(&ce_hdr->header);
1525 amdgpu_ucode_print_gfx_hdr(&me_hdr->header);
1526
1527 gfx_v9_0_cp_gfx_enable(adev, false);
1528
1529 /* PFP */
1530 fw_data = (const __le32 *)
1531 (adev->gfx.pfp_fw->data +
1532 le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes));
1533 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes) / 4;
1534 WREG32_SOC15(GC, 0, mmCP_PFP_UCODE_ADDR, 0);
1535 for (i = 0; i < fw_size; i++)
1536 WREG32_SOC15(GC, 0, mmCP_PFP_UCODE_DATA, le32_to_cpup(fw_data++));
1537 WREG32_SOC15(GC, 0, mmCP_PFP_UCODE_ADDR, adev->gfx.pfp_fw_version);
1538
1539 /* CE */
1540 fw_data = (const __le32 *)
1541 (adev->gfx.ce_fw->data +
1542 le32_to_cpu(ce_hdr->header.ucode_array_offset_bytes));
1543 fw_size = le32_to_cpu(ce_hdr->header.ucode_size_bytes) / 4;
1544 WREG32_SOC15(GC, 0, mmCP_CE_UCODE_ADDR, 0);
1545 for (i = 0; i < fw_size; i++)
1546 WREG32_SOC15(GC, 0, mmCP_CE_UCODE_DATA, le32_to_cpup(fw_data++));
1547 WREG32_SOC15(GC, 0, mmCP_CE_UCODE_ADDR, adev->gfx.ce_fw_version);
1548
1549 /* ME */
1550 fw_data = (const __le32 *)
1551 (adev->gfx.me_fw->data +
1552 le32_to_cpu(me_hdr->header.ucode_array_offset_bytes));
1553 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes) / 4;
1554 WREG32_SOC15(GC, 0, mmCP_ME_RAM_WADDR, 0);
1555 for (i = 0; i < fw_size; i++)
1556 WREG32_SOC15(GC, 0, mmCP_ME_RAM_DATA, le32_to_cpup(fw_data++));
1557 WREG32_SOC15(GC, 0, mmCP_ME_RAM_WADDR, adev->gfx.me_fw_version);
1558
1559 return 0;
1560 }
1561
1562 static u32 gfx_v9_0_get_csb_size(struct amdgpu_device *adev)
1563 {
1564 u32 count = 0;
1565 const struct cs_section_def *sect = NULL;
1566 const struct cs_extent_def *ext = NULL;
1567
1568 /* begin clear state */
1569 count += 2;
1570 /* context control state */
1571 count += 3;
1572
1573 for (sect = gfx9_cs_data; sect->section != NULL; ++sect) {
1574 for (ext = sect->section; ext->extent != NULL; ++ext) {
1575 if (sect->id == SECT_CONTEXT)
1576 count += 2 + ext->reg_count;
1577 else
1578 return 0;
1579 }
1580 }
1581 /* pa_sc_raster_config/pa_sc_raster_config1 */
1582 count += 4;
1583 /* end clear state */
1584 count += 2;
1585 /* clear state */
1586 count += 2;
1587
1588 return count;
1589 }
1590
1591 static int gfx_v9_0_cp_gfx_start(struct amdgpu_device *adev)
1592 {
1593 struct amdgpu_ring *ring = &adev->gfx.gfx_ring[0];
1594 const struct cs_section_def *sect = NULL;
1595 const struct cs_extent_def *ext = NULL;
1596 int r, i;
1597
1598 /* init the CP */
1599 WREG32_SOC15(GC, 0, mmCP_MAX_CONTEXT, adev->gfx.config.max_hw_contexts - 1);
1600 WREG32_SOC15(GC, 0, mmCP_DEVICE_ID, 1);
1601
1602 gfx_v9_0_cp_gfx_enable(adev, true);
1603
1604 r = amdgpu_ring_alloc(ring, gfx_v9_0_get_csb_size(adev) + 4);
1605 if (r) {
1606 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r);
1607 return r;
1608 }
1609
1610 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
1611 amdgpu_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE);
1612
1613 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
1614 amdgpu_ring_write(ring, 0x80000000);
1615 amdgpu_ring_write(ring, 0x80000000);
1616
1617 for (sect = gfx9_cs_data; sect->section != NULL; ++sect) {
1618 for (ext = sect->section; ext->extent != NULL; ++ext) {
1619 if (sect->id == SECT_CONTEXT) {
1620 amdgpu_ring_write(ring,
1621 PACKET3(PACKET3_SET_CONTEXT_REG,
1622 ext->reg_count));
1623 amdgpu_ring_write(ring,
1624 ext->reg_index - PACKET3_SET_CONTEXT_REG_START);
1625 for (i = 0; i < ext->reg_count; i++)
1626 amdgpu_ring_write(ring, ext->extent[i]);
1627 }
1628 }
1629 }
1630
1631 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0));
1632 amdgpu_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE);
1633
1634 amdgpu_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0));
1635 amdgpu_ring_write(ring, 0);
1636
1637 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_BASE, 2));
1638 amdgpu_ring_write(ring, PACKET3_BASE_INDEX(CE_PARTITION_BASE));
1639 amdgpu_ring_write(ring, 0x8000);
1640 amdgpu_ring_write(ring, 0x8000);
1641
1642 amdgpu_ring_commit(ring);
1643
1644 return 0;
1645 }
1646
1647 static int gfx_v9_0_cp_gfx_resume(struct amdgpu_device *adev)
1648 {
1649 struct amdgpu_ring *ring;
1650 u32 tmp;
1651 u32 rb_bufsz;
1652 u64 rb_addr, rptr_addr, wptr_gpu_addr;
1653
1654 /* Set the write pointer delay */
1655 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_DELAY, 0);
1656
1657 /* set the RB to use vmid 0 */
1658 WREG32_SOC15(GC, 0, mmCP_RB_VMID, 0);
1659
1660 /* Set ring buffer size */
1661 ring = &adev->gfx.gfx_ring[0];
1662 rb_bufsz = order_base_2(ring->ring_size / 8);
1663 tmp = REG_SET_FIELD(0, CP_RB0_CNTL, RB_BUFSZ, rb_bufsz);
1664 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, RB_BLKSZ, rb_bufsz - 2);
1665 #ifdef __BIG_ENDIAN
1666 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, BUF_SWAP, 1);
1667 #endif
1668 WREG32_SOC15(GC, 0, mmCP_RB0_CNTL, tmp);
1669
1670 /* Initialize the ring buffer's write pointers */
1671 ring->wptr = 0;
1672 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR, lower_32_bits(ring->wptr));
1673 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR_HI, upper_32_bits(ring->wptr));
1674
1675 /* set the wb address wether it's enabled or not */
1676 rptr_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
1677 WREG32_SOC15(GC, 0, mmCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr));
1678 WREG32_SOC15(GC, 0, mmCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & CP_RB_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK);
1679
1680 wptr_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
1681 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_LO, lower_32_bits(wptr_gpu_addr));
1682 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_ADDR_HI, upper_32_bits(wptr_gpu_addr));
1683
1684 mdelay(1);
1685 WREG32_SOC15(GC, 0, mmCP_RB0_CNTL, tmp);
1686
1687 rb_addr = ring->gpu_addr >> 8;
1688 WREG32_SOC15(GC, 0, mmCP_RB0_BASE, rb_addr);
1689 WREG32_SOC15(GC, 0, mmCP_RB0_BASE_HI, upper_32_bits(rb_addr));
1690
1691 tmp = RREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_CONTROL);
1692 if (ring->use_doorbell) {
1693 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
1694 DOORBELL_OFFSET, ring->doorbell_index);
1695 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL,
1696 DOORBELL_EN, 1);
1697 } else {
1698 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, DOORBELL_EN, 0);
1699 }
1700 WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_CONTROL, tmp);
1701
1702 tmp = REG_SET_FIELD(0, CP_RB_DOORBELL_RANGE_LOWER,
1703 DOORBELL_RANGE_LOWER, ring->doorbell_index);
1704 WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_RANGE_LOWER, tmp);
1705
1706 WREG32_SOC15(GC, 0, mmCP_RB_DOORBELL_RANGE_UPPER,
1707 CP_RB_DOORBELL_RANGE_UPPER__DOORBELL_RANGE_UPPER_MASK);
1708
1709
1710 /* start the ring */
1711 gfx_v9_0_cp_gfx_start(adev);
1712 ring->ready = true;
1713
1714 return 0;
1715 }
1716
1717 static void gfx_v9_0_cp_compute_enable(struct amdgpu_device *adev, bool enable)
1718 {
1719 int i;
1720
1721 if (enable) {
1722 WREG32_SOC15(GC, 0, mmCP_MEC_CNTL, 0);
1723 } else {
1724 WREG32_SOC15(GC, 0, mmCP_MEC_CNTL,
1725 (CP_MEC_CNTL__MEC_ME1_HALT_MASK | CP_MEC_CNTL__MEC_ME2_HALT_MASK));
1726 for (i = 0; i < adev->gfx.num_compute_rings; i++)
1727 adev->gfx.compute_ring[i].ready = false;
1728 adev->gfx.kiq.ring.ready = false;
1729 }
1730 udelay(50);
1731 }
1732
1733 static int gfx_v9_0_cp_compute_start(struct amdgpu_device *adev)
1734 {
1735 gfx_v9_0_cp_compute_enable(adev, true);
1736
1737 return 0;
1738 }
1739
1740 static int gfx_v9_0_cp_compute_load_microcode(struct amdgpu_device *adev)
1741 {
1742 const struct gfx_firmware_header_v1_0 *mec_hdr;
1743 const __le32 *fw_data;
1744 unsigned i;
1745 u32 tmp;
1746
1747 if (!adev->gfx.mec_fw)
1748 return -EINVAL;
1749
1750 gfx_v9_0_cp_compute_enable(adev, false);
1751
1752 mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data;
1753 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header);
1754
1755 fw_data = (const __le32 *)
1756 (adev->gfx.mec_fw->data +
1757 le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes));
1758 tmp = 0;
1759 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0);
1760 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0);
1761 WREG32_SOC15(GC, 0, mmCP_CPC_IC_BASE_CNTL, tmp);
1762
1763 WREG32_SOC15(GC, 0, mmCP_CPC_IC_BASE_LO,
1764 adev->gfx.mec.mec_fw_gpu_addr & 0xFFFFF000);
1765 WREG32_SOC15(GC, 0, mmCP_CPC_IC_BASE_HI,
1766 upper_32_bits(adev->gfx.mec.mec_fw_gpu_addr));
1767
1768 /* MEC1 */
1769 WREG32_SOC15(GC, 0, mmCP_MEC_ME1_UCODE_ADDR,
1770 mec_hdr->jt_offset);
1771 for (i = 0; i < mec_hdr->jt_size; i++)
1772 WREG32_SOC15(GC, 0, mmCP_MEC_ME1_UCODE_DATA,
1773 le32_to_cpup(fw_data + mec_hdr->jt_offset + i));
1774
1775 WREG32_SOC15(GC, 0, mmCP_MEC_ME1_UCODE_ADDR,
1776 adev->gfx.mec_fw_version);
1777 /* Todo : Loading MEC2 firmware is only necessary if MEC2 should run different microcode than MEC1. */
1778
1779 return 0;
1780 }
1781
1782 static void gfx_v9_0_cp_compute_fini(struct amdgpu_device *adev)
1783 {
1784 int i, r;
1785
1786 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
1787 struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
1788
1789 if (ring->mqd_obj) {
1790 r = amdgpu_bo_reserve(ring->mqd_obj, true);
1791 if (unlikely(r != 0))
1792 dev_warn(adev->dev, "(%d) reserve MQD bo failed\n", r);
1793
1794 amdgpu_bo_unpin(ring->mqd_obj);
1795 amdgpu_bo_unreserve(ring->mqd_obj);
1796
1797 amdgpu_bo_unref(&ring->mqd_obj);
1798 ring->mqd_obj = NULL;
1799 }
1800 }
1801 }
1802
1803 static int gfx_v9_0_init_queue(struct amdgpu_ring *ring);
1804
1805 static int gfx_v9_0_cp_compute_resume(struct amdgpu_device *adev)
1806 {
1807 int i, r;
1808 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
1809 struct amdgpu_ring *ring = &adev->gfx.compute_ring[i];
1810 if (gfx_v9_0_init_queue(ring))
1811 dev_warn(adev->dev, "compute queue %d init failed!\n", i);
1812 }
1813
1814 r = gfx_v9_0_cp_compute_start(adev);
1815 if (r)
1816 return r;
1817
1818 return 0;
1819 }
1820
1821 /* KIQ functions */
1822 static void gfx_v9_0_kiq_setting(struct amdgpu_ring *ring)
1823 {
1824 uint32_t tmp;
1825 struct amdgpu_device *adev = ring->adev;
1826
1827 /* tell RLC which is KIQ queue */
1828 tmp = RREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS);
1829 tmp &= 0xffffff00;
1830 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue);
1831 WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
1832 tmp |= 0x80;
1833 WREG32_SOC15(GC, 0, mmRLC_CP_SCHEDULERS, tmp);
1834 }
1835
1836 static void gfx_v9_0_kiq_enable(struct amdgpu_ring *ring)
1837 {
1838 amdgpu_ring_alloc(ring, 8);
1839 /* set resources */
1840 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_RESOURCES, 6));
1841 amdgpu_ring_write(ring, 0); /* vmid_mask:0 queue_type:0 (KIQ) */
1842 amdgpu_ring_write(ring, 0x000000FF); /* queue mask lo */
1843 amdgpu_ring_write(ring, 0); /* queue mask hi */
1844 amdgpu_ring_write(ring, 0); /* gws mask lo */
1845 amdgpu_ring_write(ring, 0); /* gws mask hi */
1846 amdgpu_ring_write(ring, 0); /* oac mask */
1847 amdgpu_ring_write(ring, 0); /* gds heap base:0, gds heap size:0 */
1848 amdgpu_ring_commit(ring);
1849 udelay(50);
1850 }
1851
1852 static void gfx_v9_0_map_queue_enable(struct amdgpu_ring *kiq_ring,
1853 struct amdgpu_ring *ring)
1854 {
1855 struct amdgpu_device *adev = kiq_ring->adev;
1856 uint64_t mqd_addr, wptr_addr;
1857
1858 mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj);
1859 wptr_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
1860 amdgpu_ring_alloc(kiq_ring, 8);
1861
1862 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5));
1863 /* Q_sel:0, vmid:0, vidmem: 1, engine:0, num_Q:1*/
1864 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */
1865 (0 << 4) | /* Queue_Sel */
1866 (0 << 8) | /* VMID */
1867 (ring->queue << 13 ) |
1868 (ring->pipe << 16) |
1869 ((ring->me == 1 ? 0 : 1) << 18) |
1870 (0 << 21) | /*queue_type: normal compute queue */
1871 (1 << 24) | /* alloc format: all_on_one_pipe */
1872 (0 << 26) | /* engine_sel: compute */
1873 (1 << 29)); /* num_queues: must be 1 */
1874 amdgpu_ring_write(kiq_ring, (ring->doorbell_index << 2));
1875 amdgpu_ring_write(kiq_ring, lower_32_bits(mqd_addr));
1876 amdgpu_ring_write(kiq_ring, upper_32_bits(mqd_addr));
1877 amdgpu_ring_write(kiq_ring, lower_32_bits(wptr_addr));
1878 amdgpu_ring_write(kiq_ring, upper_32_bits(wptr_addr));
1879 amdgpu_ring_commit(kiq_ring);
1880 udelay(50);
1881 }
1882
1883 static int gfx_v9_0_mqd_init(struct amdgpu_ring *ring)
1884 {
1885 struct amdgpu_device *adev = ring->adev;
1886 struct v9_mqd *mqd = ring->mqd_ptr;
1887 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr;
1888 uint32_t tmp;
1889
1890 mqd->header = 0xC0310800;
1891 mqd->compute_pipelinestat_enable = 0x00000001;
1892 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
1893 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
1894 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
1895 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
1896 mqd->compute_misc_reserved = 0x00000003;
1897
1898 eop_base_addr = ring->eop_gpu_addr >> 8;
1899 mqd->cp_hqd_eop_base_addr_lo = eop_base_addr;
1900 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr);
1901
1902 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
1903 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_EOP_CONTROL);
1904 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
1905 (order_base_2(MEC_HPD_SIZE / 4) - 1));
1906
1907 mqd->cp_hqd_eop_control = tmp;
1908
1909 /* enable doorbell? */
1910 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
1911
1912 if (ring->use_doorbell) {
1913 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1914 DOORBELL_OFFSET, ring->doorbell_index);
1915 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1916 DOORBELL_EN, 1);
1917 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1918 DOORBELL_SOURCE, 0);
1919 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1920 DOORBELL_HIT, 0);
1921 }
1922 else
1923 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1924 DOORBELL_EN, 0);
1925
1926 mqd->cp_hqd_pq_doorbell_control = tmp;
1927
1928 /* disable the queue if it's active */
1929 ring->wptr = 0;
1930 mqd->cp_hqd_dequeue_request = 0;
1931 mqd->cp_hqd_pq_rptr = 0;
1932 mqd->cp_hqd_pq_wptr_lo = 0;
1933 mqd->cp_hqd_pq_wptr_hi = 0;
1934
1935 /* set the pointer to the MQD */
1936 mqd->cp_mqd_base_addr_lo = ring->mqd_gpu_addr & 0xfffffffc;
1937 mqd->cp_mqd_base_addr_hi = upper_32_bits(ring->mqd_gpu_addr);
1938
1939 /* set MQD vmid to 0 */
1940 tmp = RREG32_SOC15(GC, 0, mmCP_MQD_CONTROL);
1941 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
1942 mqd->cp_mqd_control = tmp;
1943
1944 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
1945 hqd_gpu_addr = ring->gpu_addr >> 8;
1946 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr;
1947 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
1948
1949 /* set up the HQD, this is similar to CP_RB0_CNTL */
1950 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL);
1951 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
1952 (order_base_2(ring->ring_size / 4) - 1));
1953 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
1954 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
1955 #ifdef __BIG_ENDIAN
1956 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ENDIAN_SWAP, 1);
1957 #endif
1958 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 0);
1959 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ROQ_PQ_IB_FLIP, 0);
1960 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
1961 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
1962 mqd->cp_hqd_pq_control = tmp;
1963
1964 /* set the wb address whether it's enabled or not */
1965 wb_gpu_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
1966 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
1967 mqd->cp_hqd_pq_rptr_report_addr_hi =
1968 upper_32_bits(wb_gpu_addr) & 0xffff;
1969
1970 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
1971 wb_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
1972 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc;
1973 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
1974
1975 tmp = 0;
1976 /* enable the doorbell if requested */
1977 if (ring->use_doorbell) {
1978 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
1979 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1980 DOORBELL_OFFSET, ring->doorbell_index);
1981
1982 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1983 DOORBELL_EN, 1);
1984 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1985 DOORBELL_SOURCE, 0);
1986 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
1987 DOORBELL_HIT, 0);
1988 }
1989
1990 mqd->cp_hqd_pq_doorbell_control = tmp;
1991
1992 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
1993 ring->wptr = 0;
1994 mqd->cp_hqd_pq_rptr = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR);
1995
1996 /* set the vmid for the queue */
1997 mqd->cp_hqd_vmid = 0;
1998
1999 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE);
2000 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x53);
2001 mqd->cp_hqd_persistent_state = tmp;
2002
2003 /* set MIN_IB_AVAIL_SIZE */
2004 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_IB_CONTROL);
2005 tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3);
2006 mqd->cp_hqd_ib_control = tmp;
2007
2008 /* activate the queue */
2009 mqd->cp_hqd_active = 1;
2010
2011 return 0;
2012 }
2013
2014 static int gfx_v9_0_kiq_init_register(struct amdgpu_ring *ring)
2015 {
2016 struct amdgpu_device *adev = ring->adev;
2017 struct v9_mqd *mqd = ring->mqd_ptr;
2018 int j;
2019
2020 /* disable wptr polling */
2021 WREG32_FIELD15(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0);
2022
2023 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_BASE_ADDR,
2024 mqd->cp_hqd_eop_base_addr_lo);
2025 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_BASE_ADDR_HI,
2026 mqd->cp_hqd_eop_base_addr_hi);
2027
2028 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
2029 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_CONTROL,
2030 mqd->cp_hqd_eop_control);
2031
2032 /* enable doorbell? */
2033 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL,
2034 mqd->cp_hqd_pq_doorbell_control);
2035
2036 /* disable the queue if it's active */
2037 if (RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1) {
2038 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, 1);
2039 for (j = 0; j < adev->usec_timeout; j++) {
2040 if (!(RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1))
2041 break;
2042 udelay(1);
2043 }
2044 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST,
2045 mqd->cp_hqd_dequeue_request);
2046 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR,
2047 mqd->cp_hqd_pq_rptr);
2048 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO,
2049 mqd->cp_hqd_pq_wptr_lo);
2050 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI,
2051 mqd->cp_hqd_pq_wptr_hi);
2052 }
2053
2054 /* set the pointer to the MQD */
2055 WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR,
2056 mqd->cp_mqd_base_addr_lo);
2057 WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR_HI,
2058 mqd->cp_mqd_base_addr_hi);
2059
2060 /* set MQD vmid to 0 */
2061 WREG32_SOC15(GC, 0, mmCP_MQD_CONTROL,
2062 mqd->cp_mqd_control);
2063
2064 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
2065 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE,
2066 mqd->cp_hqd_pq_base_lo);
2067 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI,
2068 mqd->cp_hqd_pq_base_hi);
2069
2070 /* set up the HQD, this is similar to CP_RB0_CNTL */
2071 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL,
2072 mqd->cp_hqd_pq_control);
2073
2074 /* set the wb address whether it's enabled or not */
2075 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR,
2076 mqd->cp_hqd_pq_rptr_report_addr_lo);
2077 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
2078 mqd->cp_hqd_pq_rptr_report_addr_hi);
2079
2080 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
2081 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR,
2082 mqd->cp_hqd_pq_wptr_poll_addr_lo);
2083 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
2084 mqd->cp_hqd_pq_wptr_poll_addr_hi);
2085
2086 /* enable the doorbell if requested */
2087 if (ring->use_doorbell) {
2088 WREG32_SOC15(GC, 0, mmCP_MEC_DOORBELL_RANGE_LOWER,
2089 (AMDGPU_DOORBELL64_KIQ *2) << 2);
2090 WREG32_SOC15(GC, 0, mmCP_MEC_DOORBELL_RANGE_UPPER,
2091 (AMDGPU_DOORBELL64_USERQUEUE_END * 2) << 2);
2092 }
2093
2094 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL,
2095 mqd->cp_hqd_pq_doorbell_control);
2096
2097 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
2098 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO,
2099 mqd->cp_hqd_pq_wptr_lo);
2100 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI,
2101 mqd->cp_hqd_pq_wptr_hi);
2102
2103 /* set the vmid for the queue */
2104 WREG32_SOC15(GC, 0, mmCP_HQD_VMID, mqd->cp_hqd_vmid);
2105
2106 WREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE,
2107 mqd->cp_hqd_persistent_state);
2108
2109 /* activate the queue */
2110 WREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE,
2111 mqd->cp_hqd_active);
2112
2113 if (ring->use_doorbell)
2114 WREG32_FIELD15(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1);
2115
2116 return 0;
2117 }
2118
2119 static int gfx_v9_0_kiq_init_queue(struct amdgpu_ring *ring)
2120 {
2121 struct amdgpu_device *adev = ring->adev;
2122 struct amdgpu_kiq *kiq = &adev->gfx.kiq;
2123 struct v9_mqd *mqd = ring->mqd_ptr;
2124 bool is_kiq = (ring->funcs->type == AMDGPU_RING_TYPE_KIQ);
2125 int mqd_idx = AMDGPU_MAX_COMPUTE_RINGS;
2126
2127 if (is_kiq) {
2128 gfx_v9_0_kiq_setting(&kiq->ring);
2129 } else {
2130 mqd_idx = ring - &adev->gfx.compute_ring[0];
2131 }
2132
2133 if (!adev->gfx.in_reset) {
2134 memset((void *)mqd, 0, sizeof(*mqd));
2135 mutex_lock(&adev->srbm_mutex);
2136 soc15_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
2137 gfx_v9_0_mqd_init(ring);
2138 if (is_kiq)
2139 gfx_v9_0_kiq_init_register(ring);
2140 soc15_grbm_select(adev, 0, 0, 0, 0);
2141 mutex_unlock(&adev->srbm_mutex);
2142
2143 } else { /* for GPU_RESET case */
2144 /* reset MQD to a clean status */
2145
2146 /* reset ring buffer */
2147 ring->wptr = 0;
2148
2149 if (is_kiq) {
2150 mutex_lock(&adev->srbm_mutex);
2151 soc15_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0);
2152 gfx_v9_0_kiq_init_register(ring);
2153 soc15_grbm_select(adev, 0, 0, 0, 0);
2154 mutex_unlock(&adev->srbm_mutex);
2155 }
2156 }
2157
2158 if (is_kiq)
2159 gfx_v9_0_kiq_enable(ring);
2160 else
2161 gfx_v9_0_map_queue_enable(&kiq->ring, ring);
2162
2163 return 0;
2164 }
2165
2166 static int gfx_v9_0_kiq_resume(struct amdgpu_device *adev)
2167 {
2168 struct amdgpu_ring *ring = NULL;
2169 int r = 0, i;
2170
2171 gfx_v9_0_cp_compute_enable(adev, true);
2172
2173 ring = &adev->gfx.kiq.ring;
2174
2175 r = amdgpu_bo_reserve(ring->mqd_obj, false);
2176 if (unlikely(r != 0))
2177 goto done;
2178
2179 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
2180 if (!r) {
2181 r = gfx_v9_0_kiq_init_queue(ring);
2182 amdgpu_bo_kunmap(ring->mqd_obj);
2183 ring->mqd_ptr = NULL;
2184 }
2185 amdgpu_bo_unreserve(ring->mqd_obj);
2186 if (r)
2187 goto done;
2188
2189 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
2190 ring = &adev->gfx.compute_ring[i];
2191
2192 r = amdgpu_bo_reserve(ring->mqd_obj, false);
2193 if (unlikely(r != 0))
2194 goto done;
2195 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr);
2196 if (!r) {
2197 r = gfx_v9_0_kiq_init_queue(ring);
2198 amdgpu_bo_kunmap(ring->mqd_obj);
2199 ring->mqd_ptr = NULL;
2200 }
2201 amdgpu_bo_unreserve(ring->mqd_obj);
2202 if (r)
2203 goto done;
2204 }
2205
2206 done:
2207 return r;
2208 }
2209
2210 static int gfx_v9_0_cp_resume(struct amdgpu_device *adev)
2211 {
2212 int r,i;
2213 struct amdgpu_ring *ring;
2214
2215 if (!(adev->flags & AMD_IS_APU))
2216 gfx_v9_0_enable_gui_idle_interrupt(adev, false);
2217
2218 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) {
2219 /* legacy firmware loading */
2220 r = gfx_v9_0_cp_gfx_load_microcode(adev);
2221 if (r)
2222 return r;
2223
2224 r = gfx_v9_0_cp_compute_load_microcode(adev);
2225 if (r)
2226 return r;
2227 }
2228
2229 r = gfx_v9_0_cp_gfx_resume(adev);
2230 if (r)
2231 return r;
2232
2233 if (amdgpu_sriov_vf(adev))
2234 r = gfx_v9_0_kiq_resume(adev);
2235 else
2236 r = gfx_v9_0_cp_compute_resume(adev);
2237 if (r)
2238 return r;
2239
2240 ring = &adev->gfx.gfx_ring[0];
2241 r = amdgpu_ring_test_ring(ring);
2242 if (r) {
2243 ring->ready = false;
2244 return r;
2245 }
2246 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
2247 ring = &adev->gfx.compute_ring[i];
2248
2249 ring->ready = true;
2250 r = amdgpu_ring_test_ring(ring);
2251 if (r)
2252 ring->ready = false;
2253 }
2254
2255 if (amdgpu_sriov_vf(adev)) {
2256 ring = &adev->gfx.kiq.ring;
2257 ring->ready = true;
2258 r = amdgpu_ring_test_ring(ring);
2259 if (r)
2260 ring->ready = false;
2261 }
2262
2263 gfx_v9_0_enable_gui_idle_interrupt(adev, true);
2264
2265 return 0;
2266 }
2267
2268 static void gfx_v9_0_cp_enable(struct amdgpu_device *adev, bool enable)
2269 {
2270 gfx_v9_0_cp_gfx_enable(adev, enable);
2271 gfx_v9_0_cp_compute_enable(adev, enable);
2272 }
2273
2274 static int gfx_v9_0_hw_init(void *handle)
2275 {
2276 int r;
2277 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2278
2279 gfx_v9_0_init_golden_registers(adev);
2280
2281 gfx_v9_0_gpu_init(adev);
2282
2283 r = gfx_v9_0_rlc_resume(adev);
2284 if (r)
2285 return r;
2286
2287 r = gfx_v9_0_cp_resume(adev);
2288 if (r)
2289 return r;
2290
2291 r = gfx_v9_0_ngg_en(adev);
2292 if (r)
2293 return r;
2294
2295 return r;
2296 }
2297
2298 static int gfx_v9_0_hw_fini(void *handle)
2299 {
2300 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2301
2302 amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0);
2303 amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0);
2304 if (amdgpu_sriov_vf(adev)) {
2305 pr_debug("For SRIOV client, shouldn't do anything.\n");
2306 return 0;
2307 }
2308 gfx_v9_0_cp_enable(adev, false);
2309 gfx_v9_0_rlc_stop(adev);
2310 gfx_v9_0_cp_compute_fini(adev);
2311
2312 return 0;
2313 }
2314
2315 static int gfx_v9_0_suspend(void *handle)
2316 {
2317 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2318
2319 return gfx_v9_0_hw_fini(adev);
2320 }
2321
2322 static int gfx_v9_0_resume(void *handle)
2323 {
2324 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2325
2326 return gfx_v9_0_hw_init(adev);
2327 }
2328
2329 static bool gfx_v9_0_is_idle(void *handle)
2330 {
2331 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2332
2333 if (REG_GET_FIELD(RREG32_SOC15(GC, 0, mmGRBM_STATUS),
2334 GRBM_STATUS, GUI_ACTIVE))
2335 return false;
2336 else
2337 return true;
2338 }
2339
2340 static int gfx_v9_0_wait_for_idle(void *handle)
2341 {
2342 unsigned i;
2343 u32 tmp;
2344 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2345
2346 for (i = 0; i < adev->usec_timeout; i++) {
2347 /* read MC_STATUS */
2348 tmp = RREG32_SOC15(GC, 0, mmGRBM_STATUS) &
2349 GRBM_STATUS__GUI_ACTIVE_MASK;
2350
2351 if (!REG_GET_FIELD(tmp, GRBM_STATUS, GUI_ACTIVE))
2352 return 0;
2353 udelay(1);
2354 }
2355 return -ETIMEDOUT;
2356 }
2357
2358 static int gfx_v9_0_soft_reset(void *handle)
2359 {
2360 u32 grbm_soft_reset = 0;
2361 u32 tmp;
2362 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2363
2364 /* GRBM_STATUS */
2365 tmp = RREG32_SOC15(GC, 0, mmGRBM_STATUS);
2366 if (tmp & (GRBM_STATUS__PA_BUSY_MASK | GRBM_STATUS__SC_BUSY_MASK |
2367 GRBM_STATUS__BCI_BUSY_MASK | GRBM_STATUS__SX_BUSY_MASK |
2368 GRBM_STATUS__TA_BUSY_MASK | GRBM_STATUS__VGT_BUSY_MASK |
2369 GRBM_STATUS__DB_BUSY_MASK | GRBM_STATUS__CB_BUSY_MASK |
2370 GRBM_STATUS__GDS_BUSY_MASK | GRBM_STATUS__SPI_BUSY_MASK |
2371 GRBM_STATUS__IA_BUSY_MASK | GRBM_STATUS__IA_BUSY_NO_DMA_MASK)) {
2372 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
2373 GRBM_SOFT_RESET, SOFT_RESET_CP, 1);
2374 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
2375 GRBM_SOFT_RESET, SOFT_RESET_GFX, 1);
2376 }
2377
2378 if (tmp & (GRBM_STATUS__CP_BUSY_MASK | GRBM_STATUS__CP_COHERENCY_BUSY_MASK)) {
2379 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
2380 GRBM_SOFT_RESET, SOFT_RESET_CP, 1);
2381 }
2382
2383 /* GRBM_STATUS2 */
2384 tmp = RREG32_SOC15(GC, 0, mmGRBM_STATUS2);
2385 if (REG_GET_FIELD(tmp, GRBM_STATUS2, RLC_BUSY))
2386 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset,
2387 GRBM_SOFT_RESET, SOFT_RESET_RLC, 1);
2388
2389
2390 if (grbm_soft_reset) {
2391 /* stop the rlc */
2392 gfx_v9_0_rlc_stop(adev);
2393
2394 /* Disable GFX parsing/prefetching */
2395 gfx_v9_0_cp_gfx_enable(adev, false);
2396
2397 /* Disable MEC parsing/prefetching */
2398 gfx_v9_0_cp_compute_enable(adev, false);
2399
2400 if (grbm_soft_reset) {
2401 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
2402 tmp |= grbm_soft_reset;
2403 dev_info(adev->dev, "GRBM_SOFT_RESET=0x%08X\n", tmp);
2404 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
2405 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
2406
2407 udelay(50);
2408
2409 tmp &= ~grbm_soft_reset;
2410 WREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET, tmp);
2411 tmp = RREG32_SOC15(GC, 0, mmGRBM_SOFT_RESET);
2412 }
2413
2414 /* Wait a little for things to settle down */
2415 udelay(50);
2416 }
2417 return 0;
2418 }
2419
2420 static uint64_t gfx_v9_0_get_gpu_clock_counter(struct amdgpu_device *adev)
2421 {
2422 uint64_t clock;
2423
2424 mutex_lock(&adev->gfx.gpu_clock_mutex);
2425 WREG32_SOC15(GC, 0, mmRLC_CAPTURE_GPU_CLOCK_COUNT, 1);
2426 clock = (uint64_t)RREG32_SOC15(GC, 0, mmRLC_GPU_CLOCK_COUNT_LSB) |
2427 ((uint64_t)RREG32_SOC15(GC, 0, mmRLC_GPU_CLOCK_COUNT_MSB) << 32ULL);
2428 mutex_unlock(&adev->gfx.gpu_clock_mutex);
2429 return clock;
2430 }
2431
2432 static void gfx_v9_0_ring_emit_gds_switch(struct amdgpu_ring *ring,
2433 uint32_t vmid,
2434 uint32_t gds_base, uint32_t gds_size,
2435 uint32_t gws_base, uint32_t gws_size,
2436 uint32_t oa_base, uint32_t oa_size)
2437 {
2438 gds_base = gds_base >> AMDGPU_GDS_SHIFT;
2439 gds_size = gds_size >> AMDGPU_GDS_SHIFT;
2440
2441 gws_base = gws_base >> AMDGPU_GWS_SHIFT;
2442 gws_size = gws_size >> AMDGPU_GWS_SHIFT;
2443
2444 oa_base = oa_base >> AMDGPU_OA_SHIFT;
2445 oa_size = oa_size >> AMDGPU_OA_SHIFT;
2446
2447 /* GDS Base */
2448 gfx_v9_0_write_data_to_reg(ring, 0, false,
2449 amdgpu_gds_reg_offset[vmid].mem_base,
2450 gds_base);
2451
2452 /* GDS Size */
2453 gfx_v9_0_write_data_to_reg(ring, 0, false,
2454 amdgpu_gds_reg_offset[vmid].mem_size,
2455 gds_size);
2456
2457 /* GWS */
2458 gfx_v9_0_write_data_to_reg(ring, 0, false,
2459 amdgpu_gds_reg_offset[vmid].gws,
2460 gws_size << GDS_GWS_VMID0__SIZE__SHIFT | gws_base);
2461
2462 /* OA */
2463 gfx_v9_0_write_data_to_reg(ring, 0, false,
2464 amdgpu_gds_reg_offset[vmid].oa,
2465 (1 << (oa_size + oa_base)) - (1 << oa_base));
2466 }
2467
2468 static int gfx_v9_0_early_init(void *handle)
2469 {
2470 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2471
2472 adev->gfx.num_gfx_rings = GFX9_NUM_GFX_RINGS;
2473 adev->gfx.num_compute_rings = GFX9_NUM_COMPUTE_RINGS;
2474 gfx_v9_0_set_ring_funcs(adev);
2475 gfx_v9_0_set_irq_funcs(adev);
2476 gfx_v9_0_set_gds_init(adev);
2477 gfx_v9_0_set_rlc_funcs(adev);
2478
2479 return 0;
2480 }
2481
2482 static int gfx_v9_0_late_init(void *handle)
2483 {
2484 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2485 int r;
2486
2487 r = amdgpu_irq_get(adev, &adev->gfx.priv_reg_irq, 0);
2488 if (r)
2489 return r;
2490
2491 r = amdgpu_irq_get(adev, &adev->gfx.priv_inst_irq, 0);
2492 if (r)
2493 return r;
2494
2495 return 0;
2496 }
2497
2498 static void gfx_v9_0_enter_rlc_safe_mode(struct amdgpu_device *adev)
2499 {
2500 uint32_t rlc_setting, data;
2501 unsigned i;
2502
2503 if (adev->gfx.rlc.in_safe_mode)
2504 return;
2505
2506 /* if RLC is not enabled, do nothing */
2507 rlc_setting = RREG32_SOC15(GC, 0, mmRLC_CNTL);
2508 if (!(rlc_setting & RLC_CNTL__RLC_ENABLE_F32_MASK))
2509 return;
2510
2511 if (adev->cg_flags &
2512 (AMD_CG_SUPPORT_GFX_CGCG | AMD_CG_SUPPORT_GFX_MGCG |
2513 AMD_CG_SUPPORT_GFX_3D_CGCG)) {
2514 data = RLC_SAFE_MODE__CMD_MASK;
2515 data |= (1 << RLC_SAFE_MODE__MESSAGE__SHIFT);
2516 WREG32_SOC15(GC, 0, mmRLC_SAFE_MODE, data);
2517
2518 /* wait for RLC_SAFE_MODE */
2519 for (i = 0; i < adev->usec_timeout; i++) {
2520 if (!REG_GET_FIELD(SOC15_REG_OFFSET(GC, 0, mmRLC_SAFE_MODE), RLC_SAFE_MODE, CMD))
2521 break;
2522 udelay(1);
2523 }
2524 adev->gfx.rlc.in_safe_mode = true;
2525 }
2526 }
2527
2528 static void gfx_v9_0_exit_rlc_safe_mode(struct amdgpu_device *adev)
2529 {
2530 uint32_t rlc_setting, data;
2531
2532 if (!adev->gfx.rlc.in_safe_mode)
2533 return;
2534
2535 /* if RLC is not enabled, do nothing */
2536 rlc_setting = RREG32_SOC15(GC, 0, mmRLC_CNTL);
2537 if (!(rlc_setting & RLC_CNTL__RLC_ENABLE_F32_MASK))
2538 return;
2539
2540 if (adev->cg_flags &
2541 (AMD_CG_SUPPORT_GFX_CGCG | AMD_CG_SUPPORT_GFX_MGCG)) {
2542 /*
2543 * Try to exit safe mode only if it is already in safe
2544 * mode.
2545 */
2546 data = RLC_SAFE_MODE__CMD_MASK;
2547 WREG32_SOC15(GC, 0, mmRLC_SAFE_MODE, data);
2548 adev->gfx.rlc.in_safe_mode = false;
2549 }
2550 }
2551
2552 static void gfx_v9_0_update_medium_grain_clock_gating(struct amdgpu_device *adev,
2553 bool enable)
2554 {
2555 uint32_t data, def;
2556
2557 /* It is disabled by HW by default */
2558 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG)) {
2559 /* 1 - RLC_CGTT_MGCG_OVERRIDE */
2560 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
2561 data &= ~(RLC_CGTT_MGCG_OVERRIDE__CPF_CGTT_SCLK_OVERRIDE_MASK |
2562 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK |
2563 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK |
2564 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGLS_OVERRIDE_MASK);
2565
2566 /* only for Vega10 & Raven1 */
2567 data |= RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK;
2568
2569 if (def != data)
2570 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
2571
2572 /* MGLS is a global flag to control all MGLS in GFX */
2573 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGLS) {
2574 /* 2 - RLC memory Light sleep */
2575 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_RLC_LS) {
2576 def = data = RREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL);
2577 data |= RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK;
2578 if (def != data)
2579 WREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL, data);
2580 }
2581 /* 3 - CP memory Light sleep */
2582 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CP_LS) {
2583 def = data = RREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL);
2584 data |= CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
2585 if (def != data)
2586 WREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL, data);
2587 }
2588 }
2589 } else {
2590 /* 1 - MGCG_OVERRIDE */
2591 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
2592 data |= (RLC_CGTT_MGCG_OVERRIDE__CPF_CGTT_SCLK_OVERRIDE_MASK |
2593 RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK |
2594 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK |
2595 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK |
2596 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGLS_OVERRIDE_MASK);
2597 if (def != data)
2598 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
2599
2600 /* 2 - disable MGLS in RLC */
2601 data = RREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL);
2602 if (data & RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK) {
2603 data &= ~RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK;
2604 WREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL, data);
2605 }
2606
2607 /* 3 - disable MGLS in CP */
2608 data = RREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL);
2609 if (data & CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK) {
2610 data &= ~CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK;
2611 WREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL, data);
2612 }
2613 }
2614 }
2615
2616 static void gfx_v9_0_update_3d_clock_gating(struct amdgpu_device *adev,
2617 bool enable)
2618 {
2619 uint32_t data, def;
2620
2621 adev->gfx.rlc.funcs->enter_safe_mode(adev);
2622
2623 /* Enable 3D CGCG/CGLS */
2624 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG)) {
2625 /* write cmd to clear cgcg/cgls ov */
2626 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
2627 /* unset CGCG override */
2628 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_GFX3D_CG_OVERRIDE_MASK;
2629 /* update CGCG and CGLS override bits */
2630 if (def != data)
2631 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
2632 /* enable 3Dcgcg FSM(0x0020003f) */
2633 def = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D);
2634 data = (0x2000 << RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD__SHIFT) |
2635 RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK;
2636 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS)
2637 data |= (0x000F << RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY__SHIFT) |
2638 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK;
2639 if (def != data)
2640 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D, data);
2641
2642 /* set IDLE_POLL_COUNT(0x00900100) */
2643 def = RREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL);
2644 data = (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) |
2645 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
2646 if (def != data)
2647 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL, data);
2648 } else {
2649 /* Disable CGCG/CGLS */
2650 def = data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D);
2651 /* disable cgcg, cgls should be disabled */
2652 data &= ~(RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK |
2653 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK);
2654 /* disable cgcg and cgls in FSM */
2655 if (def != data)
2656 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D, data);
2657 }
2658
2659 adev->gfx.rlc.funcs->exit_safe_mode(adev);
2660 }
2661
2662 static void gfx_v9_0_update_coarse_grain_clock_gating(struct amdgpu_device *adev,
2663 bool enable)
2664 {
2665 uint32_t def, data;
2666
2667 adev->gfx.rlc.funcs->enter_safe_mode(adev);
2668
2669 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG)) {
2670 def = data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
2671 /* unset CGCG override */
2672 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGCG_OVERRIDE_MASK;
2673 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS)
2674 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK;
2675 else
2676 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK;
2677 /* update CGCG and CGLS override bits */
2678 if (def != data)
2679 WREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE, data);
2680
2681 /* enable cgcg FSM(0x0020003F) */
2682 def = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL);
2683 data = (0x2000 << RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD__SHIFT) |
2684 RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK;
2685 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS)
2686 data |= (0x000F << RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY__SHIFT) |
2687 RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK;
2688 if (def != data)
2689 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL, data);
2690
2691 /* set IDLE_POLL_COUNT(0x00900100) */
2692 def = RREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL);
2693 data = (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) |
2694 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT);
2695 if (def != data)
2696 WREG32_SOC15(GC, 0, mmCP_RB_WPTR_POLL_CNTL, data);
2697 } else {
2698 def = data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL);
2699 /* reset CGCG/CGLS bits */
2700 data &= ~(RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK | RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK);
2701 /* disable cgcg and cgls in FSM */
2702 if (def != data)
2703 WREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL, data);
2704 }
2705
2706 adev->gfx.rlc.funcs->exit_safe_mode(adev);
2707 }
2708
2709 static int gfx_v9_0_update_gfx_clock_gating(struct amdgpu_device *adev,
2710 bool enable)
2711 {
2712 if (enable) {
2713 /* CGCG/CGLS should be enabled after MGCG/MGLS
2714 * === MGCG + MGLS ===
2715 */
2716 gfx_v9_0_update_medium_grain_clock_gating(adev, enable);
2717 /* === CGCG /CGLS for GFX 3D Only === */
2718 gfx_v9_0_update_3d_clock_gating(adev, enable);
2719 /* === CGCG + CGLS === */
2720 gfx_v9_0_update_coarse_grain_clock_gating(adev, enable);
2721 } else {
2722 /* CGCG/CGLS should be disabled before MGCG/MGLS
2723 * === CGCG + CGLS ===
2724 */
2725 gfx_v9_0_update_coarse_grain_clock_gating(adev, enable);
2726 /* === CGCG /CGLS for GFX 3D Only === */
2727 gfx_v9_0_update_3d_clock_gating(adev, enable);
2728 /* === MGCG + MGLS === */
2729 gfx_v9_0_update_medium_grain_clock_gating(adev, enable);
2730 }
2731 return 0;
2732 }
2733
2734 static const struct amdgpu_rlc_funcs gfx_v9_0_rlc_funcs = {
2735 .enter_safe_mode = gfx_v9_0_enter_rlc_safe_mode,
2736 .exit_safe_mode = gfx_v9_0_exit_rlc_safe_mode
2737 };
2738
2739 static int gfx_v9_0_set_powergating_state(void *handle,
2740 enum amd_powergating_state state)
2741 {
2742 return 0;
2743 }
2744
2745 static int gfx_v9_0_set_clockgating_state(void *handle,
2746 enum amd_clockgating_state state)
2747 {
2748 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2749
2750 if (amdgpu_sriov_vf(adev))
2751 return 0;
2752
2753 switch (adev->asic_type) {
2754 case CHIP_VEGA10:
2755 gfx_v9_0_update_gfx_clock_gating(adev,
2756 state == AMD_CG_STATE_GATE ? true : false);
2757 break;
2758 default:
2759 break;
2760 }
2761 return 0;
2762 }
2763
2764 static void gfx_v9_0_get_clockgating_state(void *handle, u32 *flags)
2765 {
2766 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
2767 int data;
2768
2769 if (amdgpu_sriov_vf(adev))
2770 *flags = 0;
2771
2772 /* AMD_CG_SUPPORT_GFX_MGCG */
2773 data = RREG32_SOC15(GC, 0, mmRLC_CGTT_MGCG_OVERRIDE);
2774 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK))
2775 *flags |= AMD_CG_SUPPORT_GFX_MGCG;
2776
2777 /* AMD_CG_SUPPORT_GFX_CGCG */
2778 data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL);
2779 if (data & RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK)
2780 *flags |= AMD_CG_SUPPORT_GFX_CGCG;
2781
2782 /* AMD_CG_SUPPORT_GFX_CGLS */
2783 if (data & RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK)
2784 *flags |= AMD_CG_SUPPORT_GFX_CGLS;
2785
2786 /* AMD_CG_SUPPORT_GFX_RLC_LS */
2787 data = RREG32_SOC15(GC, 0, mmRLC_MEM_SLP_CNTL);
2788 if (data & RLC_MEM_SLP_CNTL__RLC_MEM_LS_EN_MASK)
2789 *flags |= AMD_CG_SUPPORT_GFX_RLC_LS | AMD_CG_SUPPORT_GFX_MGLS;
2790
2791 /* AMD_CG_SUPPORT_GFX_CP_LS */
2792 data = RREG32_SOC15(GC, 0, mmCP_MEM_SLP_CNTL);
2793 if (data & CP_MEM_SLP_CNTL__CP_MEM_LS_EN_MASK)
2794 *flags |= AMD_CG_SUPPORT_GFX_CP_LS | AMD_CG_SUPPORT_GFX_MGLS;
2795
2796 /* AMD_CG_SUPPORT_GFX_3D_CGCG */
2797 data = RREG32_SOC15(GC, 0, mmRLC_CGCG_CGLS_CTRL_3D);
2798 if (data & RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK)
2799 *flags |= AMD_CG_SUPPORT_GFX_3D_CGCG;
2800
2801 /* AMD_CG_SUPPORT_GFX_3D_CGLS */
2802 if (data & RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK)
2803 *flags |= AMD_CG_SUPPORT_GFX_3D_CGLS;
2804 }
2805
2806 static u64 gfx_v9_0_ring_get_rptr_gfx(struct amdgpu_ring *ring)
2807 {
2808 return ring->adev->wb.wb[ring->rptr_offs]; /* gfx9 is 32bit rptr*/
2809 }
2810
2811 static u64 gfx_v9_0_ring_get_wptr_gfx(struct amdgpu_ring *ring)
2812 {
2813 struct amdgpu_device *adev = ring->adev;
2814 u64 wptr;
2815
2816 /* XXX check if swapping is necessary on BE */
2817 if (ring->use_doorbell) {
2818 wptr = atomic64_read((atomic64_t *)&adev->wb.wb[ring->wptr_offs]);
2819 } else {
2820 wptr = RREG32_SOC15(GC, 0, mmCP_RB0_WPTR);
2821 wptr += (u64)RREG32_SOC15(GC, 0, mmCP_RB0_WPTR_HI) << 32;
2822 }
2823
2824 return wptr;
2825 }
2826
2827 static void gfx_v9_0_ring_set_wptr_gfx(struct amdgpu_ring *ring)
2828 {
2829 struct amdgpu_device *adev = ring->adev;
2830
2831 if (ring->use_doorbell) {
2832 /* XXX check if swapping is necessary on BE */
2833 atomic64_set((atomic64_t*)&adev->wb.wb[ring->wptr_offs], ring->wptr);
2834 WDOORBELL64(ring->doorbell_index, ring->wptr);
2835 } else {
2836 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR, lower_32_bits(ring->wptr));
2837 WREG32_SOC15(GC, 0, mmCP_RB0_WPTR_HI, upper_32_bits(ring->wptr));
2838 }
2839 }
2840
2841 static void gfx_v9_0_ring_emit_hdp_flush(struct amdgpu_ring *ring)
2842 {
2843 u32 ref_and_mask, reg_mem_engine;
2844 struct nbio_hdp_flush_reg *nbio_hf_reg;
2845
2846 if (ring->adev->asic_type == CHIP_VEGA10)
2847 nbio_hf_reg = &nbio_v6_1_hdp_flush_reg;
2848
2849 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) {
2850 switch (ring->me) {
2851 case 1:
2852 ref_and_mask = nbio_hf_reg->ref_and_mask_cp2 << ring->pipe;
2853 break;
2854 case 2:
2855 ref_and_mask = nbio_hf_reg->ref_and_mask_cp6 << ring->pipe;
2856 break;
2857 default:
2858 return;
2859 }
2860 reg_mem_engine = 0;
2861 } else {
2862 ref_and_mask = nbio_hf_reg->ref_and_mask_cp0;
2863 reg_mem_engine = 1; /* pfp */
2864 }
2865
2866 gfx_v9_0_wait_reg_mem(ring, reg_mem_engine, 0, 1,
2867 nbio_hf_reg->hdp_flush_req_offset,
2868 nbio_hf_reg->hdp_flush_done_offset,
2869 ref_and_mask, ref_and_mask, 0x20);
2870 }
2871
2872 static void gfx_v9_0_ring_emit_hdp_invalidate(struct amdgpu_ring *ring)
2873 {
2874 gfx_v9_0_write_data_to_reg(ring, 0, true,
2875 SOC15_REG_OFFSET(HDP, 0, mmHDP_DEBUG0), 1);
2876 }
2877
2878 static void gfx_v9_0_ring_emit_ib_gfx(struct amdgpu_ring *ring,
2879 struct amdgpu_ib *ib,
2880 unsigned vm_id, bool ctx_switch)
2881 {
2882 u32 header, control = 0;
2883
2884 if (ib->flags & AMDGPU_IB_FLAG_CE)
2885 header = PACKET3(PACKET3_INDIRECT_BUFFER_CONST, 2);
2886 else
2887 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2);
2888
2889 control |= ib->length_dw | (vm_id << 24);
2890
2891 if (amdgpu_sriov_vf(ring->adev) && (ib->flags & AMDGPU_IB_FLAG_PREEMPT))
2892 control |= INDIRECT_BUFFER_PRE_ENB(1);
2893
2894 amdgpu_ring_write(ring, header);
2895 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */
2896 amdgpu_ring_write(ring,
2897 #ifdef __BIG_ENDIAN
2898 (2 << 0) |
2899 #endif
2900 lower_32_bits(ib->gpu_addr));
2901 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
2902 amdgpu_ring_write(ring, control);
2903 }
2904
2905 #define INDIRECT_BUFFER_VALID (1 << 23)
2906
2907 static void gfx_v9_0_ring_emit_ib_compute(struct amdgpu_ring *ring,
2908 struct amdgpu_ib *ib,
2909 unsigned vm_id, bool ctx_switch)
2910 {
2911 u32 control = INDIRECT_BUFFER_VALID | ib->length_dw | (vm_id << 24);
2912
2913 amdgpu_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2));
2914 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */
2915 amdgpu_ring_write(ring,
2916 #ifdef __BIG_ENDIAN
2917 (2 << 0) |
2918 #endif
2919 lower_32_bits(ib->gpu_addr));
2920 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
2921 amdgpu_ring_write(ring, control);
2922 }
2923
2924 static void gfx_v9_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
2925 u64 seq, unsigned flags)
2926 {
2927 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT;
2928 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
2929
2930 /* RELEASE_MEM - flush caches, send int */
2931 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
2932 amdgpu_ring_write(ring, (EOP_TCL1_ACTION_EN |
2933 EOP_TC_ACTION_EN |
2934 EOP_TC_WB_ACTION_EN |
2935 EOP_TC_MD_ACTION_EN |
2936 EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
2937 EVENT_INDEX(5)));
2938 amdgpu_ring_write(ring, DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0));
2939
2940 /*
2941 * the address should be Qword aligned if 64bit write, Dword
2942 * aligned if only send 32bit data low (discard data high)
2943 */
2944 if (write64bit)
2945 BUG_ON(addr & 0x7);
2946 else
2947 BUG_ON(addr & 0x3);
2948 amdgpu_ring_write(ring, lower_32_bits(addr));
2949 amdgpu_ring_write(ring, upper_32_bits(addr));
2950 amdgpu_ring_write(ring, lower_32_bits(seq));
2951 amdgpu_ring_write(ring, upper_32_bits(seq));
2952 amdgpu_ring_write(ring, 0);
2953 }
2954
2955 static void gfx_v9_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
2956 {
2957 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
2958 uint32_t seq = ring->fence_drv.sync_seq;
2959 uint64_t addr = ring->fence_drv.gpu_addr;
2960
2961 gfx_v9_0_wait_reg_mem(ring, usepfp, 1, 0,
2962 lower_32_bits(addr), upper_32_bits(addr),
2963 seq, 0xffffffff, 4);
2964 }
2965
2966 static void gfx_v9_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
2967 unsigned vm_id, uint64_t pd_addr)
2968 {
2969 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
2970 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX);
2971 uint32_t req = ring->adev->gart.gart_funcs->get_invalidate_req(vm_id);
2972 unsigned eng = ring->vm_inv_eng;
2973
2974 pd_addr = pd_addr | 0x1; /* valid bit */
2975 /* now only use physical base address of PDE and valid */
2976 BUG_ON(pd_addr & 0xFFFF00000000003EULL);
2977
2978 gfx_v9_0_write_data_to_reg(ring, usepfp, true,
2979 hub->ctx0_ptb_addr_lo32 + (2 * vm_id),
2980 lower_32_bits(pd_addr));
2981
2982 gfx_v9_0_write_data_to_reg(ring, usepfp, true,
2983 hub->ctx0_ptb_addr_hi32 + (2 * vm_id),
2984 upper_32_bits(pd_addr));
2985
2986 gfx_v9_0_write_data_to_reg(ring, usepfp, true,
2987 hub->vm_inv_eng0_req + eng, req);
2988
2989 /* wait for the invalidate to complete */
2990 gfx_v9_0_wait_reg_mem(ring, 0, 0, 0, hub->vm_inv_eng0_ack +
2991 eng, 0, 1 << vm_id, 1 << vm_id, 0x20);
2992
2993 /* compute doesn't have PFP */
2994 if (usepfp) {
2995 /* sync PFP to ME, otherwise we might get invalid PFP reads */
2996 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
2997 amdgpu_ring_write(ring, 0x0);
2998 }
2999 }
3000
3001 static u64 gfx_v9_0_ring_get_rptr_compute(struct amdgpu_ring *ring)
3002 {
3003 return ring->adev->wb.wb[ring->rptr_offs]; /* gfx9 hardware is 32bit rptr */
3004 }
3005
3006 static u64 gfx_v9_0_ring_get_wptr_compute(struct amdgpu_ring *ring)
3007 {
3008 u64 wptr;
3009
3010 /* XXX check if swapping is necessary on BE */
3011 if (ring->use_doorbell)
3012 wptr = atomic64_read((atomic64_t *)&ring->adev->wb.wb[ring->wptr_offs]);
3013 else
3014 BUG();
3015 return wptr;
3016 }
3017
3018 static void gfx_v9_0_ring_set_wptr_compute(struct amdgpu_ring *ring)
3019 {
3020 struct amdgpu_device *adev = ring->adev;
3021
3022 /* XXX check if swapping is necessary on BE */
3023 if (ring->use_doorbell) {
3024 atomic64_set((atomic64_t*)&adev->wb.wb[ring->wptr_offs], ring->wptr);
3025 WDOORBELL64(ring->doorbell_index, ring->wptr);
3026 } else{
3027 BUG(); /* only DOORBELL method supported on gfx9 now */
3028 }
3029 }
3030
3031 static void gfx_v9_0_ring_emit_fence_kiq(struct amdgpu_ring *ring, u64 addr,
3032 u64 seq, unsigned int flags)
3033 {
3034 /* we only allocate 32bit for each seq wb address */
3035 BUG_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
3036
3037 /* write fence seq to the "addr" */
3038 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3039 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3040 WRITE_DATA_DST_SEL(5) | WR_CONFIRM));
3041 amdgpu_ring_write(ring, lower_32_bits(addr));
3042 amdgpu_ring_write(ring, upper_32_bits(addr));
3043 amdgpu_ring_write(ring, lower_32_bits(seq));
3044
3045 if (flags & AMDGPU_FENCE_FLAG_INT) {
3046 /* set register to trigger INT */
3047 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3048 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) |
3049 WRITE_DATA_DST_SEL(0) | WR_CONFIRM));
3050 amdgpu_ring_write(ring, SOC15_REG_OFFSET(GC, 0, mmCPC_INT_STATUS));
3051 amdgpu_ring_write(ring, 0);
3052 amdgpu_ring_write(ring, 0x20000000); /* src_id is 178 */
3053 }
3054 }
3055
3056 static void gfx_v9_ring_emit_sb(struct amdgpu_ring *ring)
3057 {
3058 amdgpu_ring_write(ring, PACKET3(PACKET3_SWITCH_BUFFER, 0));
3059 amdgpu_ring_write(ring, 0);
3060 }
3061
3062 static void gfx_v9_0_ring_emit_ce_meta(struct amdgpu_ring *ring)
3063 {
3064 static struct v9_ce_ib_state ce_payload = {0};
3065 uint64_t csa_addr;
3066 int cnt;
3067
3068 cnt = (sizeof(ce_payload) >> 2) + 4 - 2;
3069 csa_addr = AMDGPU_VA_RESERVED_SIZE - 2 * 4096;
3070
3071 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, cnt));
3072 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(2) |
3073 WRITE_DATA_DST_SEL(8) |
3074 WR_CONFIRM) |
3075 WRITE_DATA_CACHE_POLICY(0));
3076 amdgpu_ring_write(ring, lower_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, ce_payload)));
3077 amdgpu_ring_write(ring, upper_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, ce_payload)));
3078 amdgpu_ring_write_multiple(ring, (void *)&ce_payload, sizeof(ce_payload) >> 2);
3079 }
3080
3081 static void gfx_v9_0_ring_emit_de_meta(struct amdgpu_ring *ring)
3082 {
3083 static struct v9_de_ib_state de_payload = {0};
3084 uint64_t csa_addr, gds_addr;
3085 int cnt;
3086
3087 csa_addr = AMDGPU_VA_RESERVED_SIZE - 2 * 4096;
3088 gds_addr = csa_addr + 4096;
3089 de_payload.gds_backup_addrlo = lower_32_bits(gds_addr);
3090 de_payload.gds_backup_addrhi = upper_32_bits(gds_addr);
3091
3092 cnt = (sizeof(de_payload) >> 2) + 4 - 2;
3093 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, cnt));
3094 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) |
3095 WRITE_DATA_DST_SEL(8) |
3096 WR_CONFIRM) |
3097 WRITE_DATA_CACHE_POLICY(0));
3098 amdgpu_ring_write(ring, lower_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, de_payload)));
3099 amdgpu_ring_write(ring, upper_32_bits(csa_addr + offsetof(struct v9_gfx_meta_data, de_payload)));
3100 amdgpu_ring_write_multiple(ring, (void *)&de_payload, sizeof(de_payload) >> 2);
3101 }
3102
3103 static void gfx_v9_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)
3104 {
3105 uint32_t dw2 = 0;
3106
3107 if (amdgpu_sriov_vf(ring->adev))
3108 gfx_v9_0_ring_emit_ce_meta(ring);
3109
3110 dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */
3111 if (flags & AMDGPU_HAVE_CTX_SWITCH) {
3112 /* set load_global_config & load_global_uconfig */
3113 dw2 |= 0x8001;
3114 /* set load_cs_sh_regs */
3115 dw2 |= 0x01000000;
3116 /* set load_per_context_state & load_gfx_sh_regs for GFX */
3117 dw2 |= 0x10002;
3118
3119 /* set load_ce_ram if preamble presented */
3120 if (AMDGPU_PREAMBLE_IB_PRESENT & flags)
3121 dw2 |= 0x10000000;
3122 } else {
3123 /* still load_ce_ram if this is the first time preamble presented
3124 * although there is no context switch happens.
3125 */
3126 if (AMDGPU_PREAMBLE_IB_PRESENT_FIRST & flags)
3127 dw2 |= 0x10000000;
3128 }
3129
3130 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
3131 amdgpu_ring_write(ring, dw2);
3132 amdgpu_ring_write(ring, 0);
3133
3134 if (amdgpu_sriov_vf(ring->adev))
3135 gfx_v9_0_ring_emit_de_meta(ring);
3136 }
3137
3138 static unsigned gfx_v9_0_ring_emit_init_cond_exec(struct amdgpu_ring *ring)
3139 {
3140 unsigned ret;
3141 amdgpu_ring_write(ring, PACKET3(PACKET3_COND_EXEC, 3));
3142 amdgpu_ring_write(ring, lower_32_bits(ring->cond_exe_gpu_addr));
3143 amdgpu_ring_write(ring, upper_32_bits(ring->cond_exe_gpu_addr));
3144 amdgpu_ring_write(ring, 0); /* discard following DWs if *cond_exec_gpu_addr==0 */
3145 ret = ring->wptr & ring->buf_mask;
3146 amdgpu_ring_write(ring, 0x55aa55aa); /* patch dummy value later */
3147 return ret;
3148 }
3149
3150 static void gfx_v9_0_ring_emit_patch_cond_exec(struct amdgpu_ring *ring, unsigned offset)
3151 {
3152 unsigned cur;
3153 BUG_ON(offset > ring->buf_mask);
3154 BUG_ON(ring->ring[offset] != 0x55aa55aa);
3155
3156 cur = (ring->wptr & ring->buf_mask) - 1;
3157 if (likely(cur > offset))
3158 ring->ring[offset] = cur - offset;
3159 else
3160 ring->ring[offset] = (ring->ring_size>>2) - offset + cur;
3161 }
3162
3163 static void gfx_v9_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg)
3164 {
3165 struct amdgpu_device *adev = ring->adev;
3166
3167 amdgpu_ring_write(ring, PACKET3(PACKET3_COPY_DATA, 4));
3168 amdgpu_ring_write(ring, 0 | /* src: register*/
3169 (5 << 8) | /* dst: memory */
3170 (1 << 20)); /* write confirm */
3171 amdgpu_ring_write(ring, reg);
3172 amdgpu_ring_write(ring, 0);
3173 amdgpu_ring_write(ring, lower_32_bits(adev->wb.gpu_addr +
3174 adev->virt.reg_val_offs * 4));
3175 amdgpu_ring_write(ring, upper_32_bits(adev->wb.gpu_addr +
3176 adev->virt.reg_val_offs * 4));
3177 }
3178
3179 static void gfx_v9_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg,
3180 uint32_t val)
3181 {
3182 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3));
3183 amdgpu_ring_write(ring, (1 << 16)); /* no inc addr */
3184 amdgpu_ring_write(ring, reg);
3185 amdgpu_ring_write(ring, 0);
3186 amdgpu_ring_write(ring, val);
3187 }
3188
3189 static void gfx_v9_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev,
3190 enum amdgpu_interrupt_state state)
3191 {
3192 switch (state) {
3193 case AMDGPU_IRQ_STATE_DISABLE:
3194 case AMDGPU_IRQ_STATE_ENABLE:
3195 WREG32_FIELD15(GC, 0, CP_INT_CNTL_RING0,
3196 TIME_STAMP_INT_ENABLE,
3197 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
3198 break;
3199 default:
3200 break;
3201 }
3202 }
3203
3204 static void gfx_v9_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev,
3205 int me, int pipe,
3206 enum amdgpu_interrupt_state state)
3207 {
3208 u32 mec_int_cntl, mec_int_cntl_reg;
3209
3210 /*
3211 * amdgpu controls only pipe 0 of MEC1. That's why this function only
3212 * handles the setting of interrupts for this specific pipe. All other
3213 * pipes' interrupts are set by amdkfd.
3214 */
3215
3216 if (me == 1) {
3217 switch (pipe) {
3218 case 0:
3219 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE0_INT_CNTL);
3220 break;
3221 default:
3222 DRM_DEBUG("invalid pipe %d\n", pipe);
3223 return;
3224 }
3225 } else {
3226 DRM_DEBUG("invalid me %d\n", me);
3227 return;
3228 }
3229
3230 switch (state) {
3231 case AMDGPU_IRQ_STATE_DISABLE:
3232 mec_int_cntl = RREG32(mec_int_cntl_reg);
3233 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
3234 TIME_STAMP_INT_ENABLE, 0);
3235 WREG32(mec_int_cntl_reg, mec_int_cntl);
3236 break;
3237 case AMDGPU_IRQ_STATE_ENABLE:
3238 mec_int_cntl = RREG32(mec_int_cntl_reg);
3239 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL,
3240 TIME_STAMP_INT_ENABLE, 1);
3241 WREG32(mec_int_cntl_reg, mec_int_cntl);
3242 break;
3243 default:
3244 break;
3245 }
3246 }
3247
3248 static int gfx_v9_0_set_priv_reg_fault_state(struct amdgpu_device *adev,
3249 struct amdgpu_irq_src *source,
3250 unsigned type,
3251 enum amdgpu_interrupt_state state)
3252 {
3253 switch (state) {
3254 case AMDGPU_IRQ_STATE_DISABLE:
3255 case AMDGPU_IRQ_STATE_ENABLE:
3256 WREG32_FIELD15(GC, 0, CP_INT_CNTL_RING0,
3257 PRIV_REG_INT_ENABLE,
3258 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
3259 break;
3260 default:
3261 break;
3262 }
3263
3264 return 0;
3265 }
3266
3267 static int gfx_v9_0_set_priv_inst_fault_state(struct amdgpu_device *adev,
3268 struct amdgpu_irq_src *source,
3269 unsigned type,
3270 enum amdgpu_interrupt_state state)
3271 {
3272 switch (state) {
3273 case AMDGPU_IRQ_STATE_DISABLE:
3274 case AMDGPU_IRQ_STATE_ENABLE:
3275 WREG32_FIELD15(GC, 0, CP_INT_CNTL_RING0,
3276 PRIV_INSTR_INT_ENABLE,
3277 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0);
3278 default:
3279 break;
3280 }
3281
3282 return 0;
3283 }
3284
3285 static int gfx_v9_0_set_eop_interrupt_state(struct amdgpu_device *adev,
3286 struct amdgpu_irq_src *src,
3287 unsigned type,
3288 enum amdgpu_interrupt_state state)
3289 {
3290 switch (type) {
3291 case AMDGPU_CP_IRQ_GFX_EOP:
3292 gfx_v9_0_set_gfx_eop_interrupt_state(adev, state);
3293 break;
3294 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP:
3295 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 0, state);
3296 break;
3297 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP:
3298 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 1, state);
3299 break;
3300 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP:
3301 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 2, state);
3302 break;
3303 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP:
3304 gfx_v9_0_set_compute_eop_interrupt_state(adev, 1, 3, state);
3305 break;
3306 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE0_EOP:
3307 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 0, state);
3308 break;
3309 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE1_EOP:
3310 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 1, state);
3311 break;
3312 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE2_EOP:
3313 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 2, state);
3314 break;
3315 case AMDGPU_CP_IRQ_COMPUTE_MEC2_PIPE3_EOP:
3316 gfx_v9_0_set_compute_eop_interrupt_state(adev, 2, 3, state);
3317 break;
3318 default:
3319 break;
3320 }
3321 return 0;
3322 }
3323
3324 static int gfx_v9_0_eop_irq(struct amdgpu_device *adev,
3325 struct amdgpu_irq_src *source,
3326 struct amdgpu_iv_entry *entry)
3327 {
3328 int i;
3329 u8 me_id, pipe_id, queue_id;
3330 struct amdgpu_ring *ring;
3331
3332 DRM_DEBUG("IH: CP EOP\n");
3333 me_id = (entry->ring_id & 0x0c) >> 2;
3334 pipe_id = (entry->ring_id & 0x03) >> 0;
3335 queue_id = (entry->ring_id & 0x70) >> 4;
3336
3337 switch (me_id) {
3338 case 0:
3339 amdgpu_fence_process(&adev->gfx.gfx_ring[0]);
3340 break;
3341 case 1:
3342 case 2:
3343 for (i = 0; i < adev->gfx.num_compute_rings; i++) {
3344 ring = &adev->gfx.compute_ring[i];
3345 /* Per-queue interrupt is supported for MEC starting from VI.
3346 * The interrupt can only be enabled/disabled per pipe instead of per queue.
3347 */
3348 if ((ring->me == me_id) && (ring->pipe == pipe_id) && (ring->queue == queue_id))
3349 amdgpu_fence_process(ring);
3350 }
3351 break;
3352 }
3353 return 0;
3354 }
3355
3356 static int gfx_v9_0_priv_reg_irq(struct amdgpu_device *adev,
3357 struct amdgpu_irq_src *source,
3358 struct amdgpu_iv_entry *entry)
3359 {
3360 DRM_ERROR("Illegal register access in command stream\n");
3361 schedule_work(&adev->reset_work);
3362 return 0;
3363 }
3364
3365 static int gfx_v9_0_priv_inst_irq(struct amdgpu_device *adev,
3366 struct amdgpu_irq_src *source,
3367 struct amdgpu_iv_entry *entry)
3368 {
3369 DRM_ERROR("Illegal instruction in command stream\n");
3370 schedule_work(&adev->reset_work);
3371 return 0;
3372 }
3373
3374 static int gfx_v9_0_kiq_set_interrupt_state(struct amdgpu_device *adev,
3375 struct amdgpu_irq_src *src,
3376 unsigned int type,
3377 enum amdgpu_interrupt_state state)
3378 {
3379 uint32_t tmp, target;
3380 struct amdgpu_ring *ring = &(adev->gfx.kiq.ring);
3381
3382 if (ring->me == 1)
3383 target = SOC15_REG_OFFSET(GC, 0, mmCP_ME1_PIPE0_INT_CNTL);
3384 else
3385 target = SOC15_REG_OFFSET(GC, 0, mmCP_ME2_PIPE0_INT_CNTL);
3386 target += ring->pipe;
3387
3388 switch (type) {
3389 case AMDGPU_CP_KIQ_IRQ_DRIVER0:
3390 if (state == AMDGPU_IRQ_STATE_DISABLE) {
3391 tmp = RREG32_SOC15(GC, 0, mmCPC_INT_CNTL);
3392 tmp = REG_SET_FIELD(tmp, CPC_INT_CNTL,
3393 GENERIC2_INT_ENABLE, 0);
3394 WREG32_SOC15(GC, 0, mmCPC_INT_CNTL, tmp);
3395
3396 tmp = RREG32(target);
3397 tmp = REG_SET_FIELD(tmp, CP_ME2_PIPE0_INT_CNTL,
3398 GENERIC2_INT_ENABLE, 0);
3399 WREG32(target, tmp);
3400 } else {
3401 tmp = RREG32_SOC15(GC, 0, mmCPC_INT_CNTL);
3402 tmp = REG_SET_FIELD(tmp, CPC_INT_CNTL,
3403 GENERIC2_INT_ENABLE, 1);
3404 WREG32_SOC15(GC, 0, mmCPC_INT_CNTL, tmp);
3405
3406 tmp = RREG32(target);
3407 tmp = REG_SET_FIELD(tmp, CP_ME2_PIPE0_INT_CNTL,
3408 GENERIC2_INT_ENABLE, 1);
3409 WREG32(target, tmp);
3410 }
3411 break;
3412 default:
3413 BUG(); /* kiq only support GENERIC2_INT now */
3414 break;
3415 }
3416 return 0;
3417 }
3418
3419 static int gfx_v9_0_kiq_irq(struct amdgpu_device *adev,
3420 struct amdgpu_irq_src *source,
3421 struct amdgpu_iv_entry *entry)
3422 {
3423 u8 me_id, pipe_id, queue_id;
3424 struct amdgpu_ring *ring = &(adev->gfx.kiq.ring);
3425
3426 me_id = (entry->ring_id & 0x0c) >> 2;
3427 pipe_id = (entry->ring_id & 0x03) >> 0;
3428 queue_id = (entry->ring_id & 0x70) >> 4;
3429 DRM_DEBUG("IH: CPC GENERIC2_INT, me:%d, pipe:%d, queue:%d\n",
3430 me_id, pipe_id, queue_id);
3431
3432 amdgpu_fence_process(ring);
3433 return 0;
3434 }
3435
3436 const struct amd_ip_funcs gfx_v9_0_ip_funcs = {
3437 .name = "gfx_v9_0",
3438 .early_init = gfx_v9_0_early_init,
3439 .late_init = gfx_v9_0_late_init,
3440 .sw_init = gfx_v9_0_sw_init,
3441 .sw_fini = gfx_v9_0_sw_fini,
3442 .hw_init = gfx_v9_0_hw_init,
3443 .hw_fini = gfx_v9_0_hw_fini,
3444 .suspend = gfx_v9_0_suspend,
3445 .resume = gfx_v9_0_resume,
3446 .is_idle = gfx_v9_0_is_idle,
3447 .wait_for_idle = gfx_v9_0_wait_for_idle,
3448 .soft_reset = gfx_v9_0_soft_reset,
3449 .set_clockgating_state = gfx_v9_0_set_clockgating_state,
3450 .set_powergating_state = gfx_v9_0_set_powergating_state,
3451 .get_clockgating_state = gfx_v9_0_get_clockgating_state,
3452 };
3453
3454 static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_gfx = {
3455 .type = AMDGPU_RING_TYPE_GFX,
3456 .align_mask = 0xff,
3457 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
3458 .support_64bit_ptrs = true,
3459 .vmhub = AMDGPU_GFXHUB,
3460 .get_rptr = gfx_v9_0_ring_get_rptr_gfx,
3461 .get_wptr = gfx_v9_0_ring_get_wptr_gfx,
3462 .set_wptr = gfx_v9_0_ring_set_wptr_gfx,
3463 .emit_frame_size = /* totally 242 maximum if 16 IBs */
3464 5 + /* COND_EXEC */
3465 7 + /* PIPELINE_SYNC */
3466 24 + /* VM_FLUSH */
3467 8 + /* FENCE for VM_FLUSH */
3468 20 + /* GDS switch */
3469 4 + /* double SWITCH_BUFFER,
3470 the first COND_EXEC jump to the place just
3471 prior to this double SWITCH_BUFFER */
3472 5 + /* COND_EXEC */
3473 7 + /* HDP_flush */
3474 4 + /* VGT_flush */
3475 14 + /* CE_META */
3476 31 + /* DE_META */
3477 3 + /* CNTX_CTRL */
3478 5 + /* HDP_INVL */
3479 8 + 8 + /* FENCE x2 */
3480 2, /* SWITCH_BUFFER */
3481 .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_gfx */
3482 .emit_ib = gfx_v9_0_ring_emit_ib_gfx,
3483 .emit_fence = gfx_v9_0_ring_emit_fence,
3484 .emit_pipeline_sync = gfx_v9_0_ring_emit_pipeline_sync,
3485 .emit_vm_flush = gfx_v9_0_ring_emit_vm_flush,
3486 .emit_gds_switch = gfx_v9_0_ring_emit_gds_switch,
3487 .emit_hdp_flush = gfx_v9_0_ring_emit_hdp_flush,
3488 .emit_hdp_invalidate = gfx_v9_0_ring_emit_hdp_invalidate,
3489 .test_ring = gfx_v9_0_ring_test_ring,
3490 .test_ib = gfx_v9_0_ring_test_ib,
3491 .insert_nop = amdgpu_ring_insert_nop,
3492 .pad_ib = amdgpu_ring_generic_pad_ib,
3493 .emit_switch_buffer = gfx_v9_ring_emit_sb,
3494 .emit_cntxcntl = gfx_v9_ring_emit_cntxcntl,
3495 .init_cond_exec = gfx_v9_0_ring_emit_init_cond_exec,
3496 .patch_cond_exec = gfx_v9_0_ring_emit_patch_cond_exec,
3497 };
3498
3499 static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_compute = {
3500 .type = AMDGPU_RING_TYPE_COMPUTE,
3501 .align_mask = 0xff,
3502 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
3503 .support_64bit_ptrs = true,
3504 .vmhub = AMDGPU_GFXHUB,
3505 .get_rptr = gfx_v9_0_ring_get_rptr_compute,
3506 .get_wptr = gfx_v9_0_ring_get_wptr_compute,
3507 .set_wptr = gfx_v9_0_ring_set_wptr_compute,
3508 .emit_frame_size =
3509 20 + /* gfx_v9_0_ring_emit_gds_switch */
3510 7 + /* gfx_v9_0_ring_emit_hdp_flush */
3511 5 + /* gfx_v9_0_ring_emit_hdp_invalidate */
3512 7 + /* gfx_v9_0_ring_emit_pipeline_sync */
3513 24 + /* gfx_v9_0_ring_emit_vm_flush */
3514 8 + 8 + 8, /* gfx_v9_0_ring_emit_fence x3 for user fence, vm fence */
3515 .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_compute */
3516 .emit_ib = gfx_v9_0_ring_emit_ib_compute,
3517 .emit_fence = gfx_v9_0_ring_emit_fence,
3518 .emit_pipeline_sync = gfx_v9_0_ring_emit_pipeline_sync,
3519 .emit_vm_flush = gfx_v9_0_ring_emit_vm_flush,
3520 .emit_gds_switch = gfx_v9_0_ring_emit_gds_switch,
3521 .emit_hdp_flush = gfx_v9_0_ring_emit_hdp_flush,
3522 .emit_hdp_invalidate = gfx_v9_0_ring_emit_hdp_invalidate,
3523 .test_ring = gfx_v9_0_ring_test_ring,
3524 .test_ib = gfx_v9_0_ring_test_ib,
3525 .insert_nop = amdgpu_ring_insert_nop,
3526 .pad_ib = amdgpu_ring_generic_pad_ib,
3527 };
3528
3529 static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_kiq = {
3530 .type = AMDGPU_RING_TYPE_KIQ,
3531 .align_mask = 0xff,
3532 .nop = PACKET3(PACKET3_NOP, 0x3FFF),
3533 .support_64bit_ptrs = true,
3534 .vmhub = AMDGPU_GFXHUB,
3535 .get_rptr = gfx_v9_0_ring_get_rptr_compute,
3536 .get_wptr = gfx_v9_0_ring_get_wptr_compute,
3537 .set_wptr = gfx_v9_0_ring_set_wptr_compute,
3538 .emit_frame_size =
3539 20 + /* gfx_v9_0_ring_emit_gds_switch */
3540 7 + /* gfx_v9_0_ring_emit_hdp_flush */
3541 5 + /* gfx_v9_0_ring_emit_hdp_invalidate */
3542 7 + /* gfx_v9_0_ring_emit_pipeline_sync */
3543 24 + /* gfx_v9_0_ring_emit_vm_flush */
3544 8 + 8 + 8, /* gfx_v9_0_ring_emit_fence_kiq x3 for user fence, vm fence */
3545 .emit_ib_size = 4, /* gfx_v9_0_ring_emit_ib_compute */
3546 .emit_ib = gfx_v9_0_ring_emit_ib_compute,
3547 .emit_fence = gfx_v9_0_ring_emit_fence_kiq,
3548 .test_ring = gfx_v9_0_ring_test_ring,
3549 .test_ib = gfx_v9_0_ring_test_ib,
3550 .insert_nop = amdgpu_ring_insert_nop,
3551 .pad_ib = amdgpu_ring_generic_pad_ib,
3552 .emit_rreg = gfx_v9_0_ring_emit_rreg,
3553 .emit_wreg = gfx_v9_0_ring_emit_wreg,
3554 };
3555
3556 static void gfx_v9_0_set_ring_funcs(struct amdgpu_device *adev)
3557 {
3558 int i;
3559
3560 adev->gfx.kiq.ring.funcs = &gfx_v9_0_ring_funcs_kiq;
3561
3562 for (i = 0; i < adev->gfx.num_gfx_rings; i++)
3563 adev->gfx.gfx_ring[i].funcs = &gfx_v9_0_ring_funcs_gfx;
3564
3565 for (i = 0; i < adev->gfx.num_compute_rings; i++)
3566 adev->gfx.compute_ring[i].funcs = &gfx_v9_0_ring_funcs_compute;
3567 }
3568
3569 static const struct amdgpu_irq_src_funcs gfx_v9_0_kiq_irq_funcs = {
3570 .set = gfx_v9_0_kiq_set_interrupt_state,
3571 .process = gfx_v9_0_kiq_irq,
3572 };
3573
3574 static const struct amdgpu_irq_src_funcs gfx_v9_0_eop_irq_funcs = {
3575 .set = gfx_v9_0_set_eop_interrupt_state,
3576 .process = gfx_v9_0_eop_irq,
3577 };
3578
3579 static const struct amdgpu_irq_src_funcs gfx_v9_0_priv_reg_irq_funcs = {
3580 .set = gfx_v9_0_set_priv_reg_fault_state,
3581 .process = gfx_v9_0_priv_reg_irq,
3582 };
3583
3584 static const struct amdgpu_irq_src_funcs gfx_v9_0_priv_inst_irq_funcs = {
3585 .set = gfx_v9_0_set_priv_inst_fault_state,
3586 .process = gfx_v9_0_priv_inst_irq,
3587 };
3588
3589 static void gfx_v9_0_set_irq_funcs(struct amdgpu_device *adev)
3590 {
3591 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST;
3592 adev->gfx.eop_irq.funcs = &gfx_v9_0_eop_irq_funcs;
3593
3594 adev->gfx.priv_reg_irq.num_types = 1;
3595 adev->gfx.priv_reg_irq.funcs = &gfx_v9_0_priv_reg_irq_funcs;
3596
3597 adev->gfx.priv_inst_irq.num_types = 1;
3598 adev->gfx.priv_inst_irq.funcs = &gfx_v9_0_priv_inst_irq_funcs;
3599
3600 adev->gfx.kiq.irq.num_types = AMDGPU_CP_KIQ_IRQ_LAST;
3601 adev->gfx.kiq.irq.funcs = &gfx_v9_0_kiq_irq_funcs;
3602 }
3603
3604 static void gfx_v9_0_set_rlc_funcs(struct amdgpu_device *adev)
3605 {
3606 switch (adev->asic_type) {
3607 case CHIP_VEGA10:
3608 adev->gfx.rlc.funcs = &gfx_v9_0_rlc_funcs;
3609 break;
3610 default:
3611 break;
3612 }
3613 }
3614
3615 static void gfx_v9_0_set_gds_init(struct amdgpu_device *adev)
3616 {
3617 /* init asci gds info */
3618 adev->gds.mem.total_size = RREG32_SOC15(GC, 0, mmGDS_VMID0_SIZE);
3619 adev->gds.gws.total_size = 64;
3620 adev->gds.oa.total_size = 16;
3621
3622 if (adev->gds.mem.total_size == 64 * 1024) {
3623 adev->gds.mem.gfx_partition_size = 4096;
3624 adev->gds.mem.cs_partition_size = 4096;
3625
3626 adev->gds.gws.gfx_partition_size = 4;
3627 adev->gds.gws.cs_partition_size = 4;
3628
3629 adev->gds.oa.gfx_partition_size = 4;
3630 adev->gds.oa.cs_partition_size = 1;
3631 } else {
3632 adev->gds.mem.gfx_partition_size = 1024;
3633 adev->gds.mem.cs_partition_size = 1024;
3634
3635 adev->gds.gws.gfx_partition_size = 16;
3636 adev->gds.gws.cs_partition_size = 16;
3637
3638 adev->gds.oa.gfx_partition_size = 4;
3639 adev->gds.oa.cs_partition_size = 4;
3640 }
3641 }
3642
3643 static u32 gfx_v9_0_get_cu_active_bitmap(struct amdgpu_device *adev)
3644 {
3645 u32 data, mask;
3646
3647 data = RREG32_SOC15(GC, 0, mmCC_GC_SHADER_ARRAY_CONFIG);
3648 data |= RREG32_SOC15(GC, 0, mmGC_USER_SHADER_ARRAY_CONFIG);
3649
3650 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_CUS_MASK;
3651 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_CUS__SHIFT;
3652
3653 mask = gfx_v9_0_create_bitmask(adev->gfx.config.max_cu_per_sh);
3654
3655 return (~data) & mask;
3656 }
3657
3658 static int gfx_v9_0_get_cu_info(struct amdgpu_device *adev,
3659 struct amdgpu_cu_info *cu_info)
3660 {
3661 int i, j, k, counter, active_cu_number = 0;
3662 u32 mask, bitmap, ao_bitmap, ao_cu_mask = 0;
3663
3664 if (!adev || !cu_info)
3665 return -EINVAL;
3666
3667 memset(cu_info, 0, sizeof(*cu_info));
3668
3669 mutex_lock(&adev->grbm_idx_mutex);
3670 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) {
3671 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) {
3672 mask = 1;
3673 ao_bitmap = 0;
3674 counter = 0;
3675 gfx_v9_0_select_se_sh(adev, i, j, 0xffffffff);
3676 bitmap = gfx_v9_0_get_cu_active_bitmap(adev);
3677 cu_info->bitmap[i][j] = bitmap;
3678
3679 for (k = 0; k < 16; k ++) {
3680 if (bitmap & mask) {
3681 if (counter < 2)
3682 ao_bitmap |= mask;
3683 counter ++;
3684 }
3685 mask <<= 1;
3686 }
3687 active_cu_number += counter;
3688 ao_cu_mask |= (ao_bitmap << (i * 16 + j * 8));
3689 }
3690 }
3691 gfx_v9_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
3692 mutex_unlock(&adev->grbm_idx_mutex);
3693
3694 cu_info->number = active_cu_number;
3695 cu_info->ao_cu_mask = ao_cu_mask;
3696
3697 return 0;
3698 }
3699
3700 static int gfx_v9_0_init_queue(struct amdgpu_ring *ring)
3701 {
3702 int r, j;
3703 u32 tmp;
3704 bool use_doorbell = true;
3705 u64 hqd_gpu_addr;
3706 u64 mqd_gpu_addr;
3707 u64 eop_gpu_addr;
3708 u64 wb_gpu_addr;
3709 u32 *buf;
3710 struct v9_mqd *mqd;
3711 struct amdgpu_device *adev;
3712
3713 adev = ring->adev;
3714 if (ring->mqd_obj == NULL) {
3715 r = amdgpu_bo_create(adev,
3716 sizeof(struct v9_mqd),
3717 PAGE_SIZE,true,
3718 AMDGPU_GEM_DOMAIN_GTT, 0, NULL,
3719 NULL, &ring->mqd_obj);
3720 if (r) {
3721 dev_warn(adev->dev, "(%d) create MQD bo failed\n", r);
3722 return r;
3723 }
3724 }
3725
3726 r = amdgpu_bo_reserve(ring->mqd_obj, false);
3727 if (unlikely(r != 0)) {
3728 gfx_v9_0_cp_compute_fini(adev);
3729 return r;
3730 }
3731
3732 r = amdgpu_bo_pin(ring->mqd_obj, AMDGPU_GEM_DOMAIN_GTT,
3733 &mqd_gpu_addr);
3734 if (r) {
3735 dev_warn(adev->dev, "(%d) pin MQD bo failed\n", r);
3736 gfx_v9_0_cp_compute_fini(adev);
3737 return r;
3738 }
3739 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&buf);
3740 if (r) {
3741 dev_warn(adev->dev, "(%d) map MQD bo failed\n", r);
3742 gfx_v9_0_cp_compute_fini(adev);
3743 return r;
3744 }
3745
3746 /* init the mqd struct */
3747 memset(buf, 0, sizeof(struct v9_mqd));
3748
3749 mqd = (struct v9_mqd *)buf;
3750 mqd->header = 0xC0310800;
3751 mqd->compute_pipelinestat_enable = 0x00000001;
3752 mqd->compute_static_thread_mgmt_se0 = 0xffffffff;
3753 mqd->compute_static_thread_mgmt_se1 = 0xffffffff;
3754 mqd->compute_static_thread_mgmt_se2 = 0xffffffff;
3755 mqd->compute_static_thread_mgmt_se3 = 0xffffffff;
3756 mqd->compute_misc_reserved = 0x00000003;
3757 mutex_lock(&adev->srbm_mutex);
3758 soc15_grbm_select(adev, ring->me,
3759 ring->pipe,
3760 ring->queue, 0);
3761 /* disable wptr polling */
3762 WREG32_FIELD15(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0);
3763
3764 /* write the EOP addr */
3765 BUG_ON(ring->me != 1 || ring->pipe != 0); /* can't handle other cases eop address */
3766 eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr + (ring->queue * MEC_HPD_SIZE);
3767 eop_gpu_addr >>= 8;
3768
3769 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_BASE_ADDR, lower_32_bits(eop_gpu_addr));
3770 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_BASE_ADDR_HI, upper_32_bits(eop_gpu_addr));
3771 mqd->cp_hqd_eop_base_addr_lo = lower_32_bits(eop_gpu_addr);
3772 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_gpu_addr);
3773
3774 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */
3775 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_EOP_CONTROL);
3776 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE,
3777 (order_base_2(MEC_HPD_SIZE / 4) - 1));
3778 WREG32_SOC15(GC, 0, mmCP_HQD_EOP_CONTROL, tmp);
3779
3780 /* enable doorbell? */
3781 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
3782 if (use_doorbell)
3783 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
3784 else
3785 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 0);
3786
3787 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL, tmp);
3788 mqd->cp_hqd_pq_doorbell_control = tmp;
3789
3790 /* disable the queue if it's active */
3791 ring->wptr = 0;
3792 mqd->cp_hqd_dequeue_request = 0;
3793 mqd->cp_hqd_pq_rptr = 0;
3794 mqd->cp_hqd_pq_wptr_lo = 0;
3795 mqd->cp_hqd_pq_wptr_hi = 0;
3796 if (RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1) {
3797 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, 1);
3798 for (j = 0; j < adev->usec_timeout; j++) {
3799 if (!(RREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE) & 1))
3800 break;
3801 udelay(1);
3802 }
3803 WREG32_SOC15(GC, 0, mmCP_HQD_DEQUEUE_REQUEST, mqd->cp_hqd_dequeue_request);
3804 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR, mqd->cp_hqd_pq_rptr);
3805 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO, mqd->cp_hqd_pq_wptr_lo);
3806 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI, mqd->cp_hqd_pq_wptr_hi);
3807 }
3808
3809 /* set the pointer to the MQD */
3810 mqd->cp_mqd_base_addr_lo = mqd_gpu_addr & 0xfffffffc;
3811 mqd->cp_mqd_base_addr_hi = upper_32_bits(mqd_gpu_addr);
3812 WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR, mqd->cp_mqd_base_addr_lo);
3813 WREG32_SOC15(GC, 0, mmCP_MQD_BASE_ADDR_HI, mqd->cp_mqd_base_addr_hi);
3814
3815 /* set MQD vmid to 0 */
3816 tmp = RREG32_SOC15(GC, 0, mmCP_MQD_CONTROL);
3817 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0);
3818 WREG32_SOC15(GC, 0, mmCP_MQD_CONTROL, tmp);
3819 mqd->cp_mqd_control = tmp;
3820
3821 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */
3822 hqd_gpu_addr = ring->gpu_addr >> 8;
3823 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr;
3824 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr);
3825 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE, mqd->cp_hqd_pq_base_lo);
3826 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_BASE_HI, mqd->cp_hqd_pq_base_hi);
3827
3828 /* set up the HQD, this is similar to CP_RB0_CNTL */
3829 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL);
3830 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE,
3831 (order_base_2(ring->ring_size / 4) - 1));
3832 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE,
3833 ((order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1) << 8));
3834 #ifdef __BIG_ENDIAN
3835 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ENDIAN_SWAP, 1);
3836 #endif
3837 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 0);
3838 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, ROQ_PQ_IB_FLIP, 0);
3839 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1);
3840 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1);
3841 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_CONTROL, tmp);
3842 mqd->cp_hqd_pq_control = tmp;
3843
3844 /* set the wb address wether it's enabled or not */
3845 wb_gpu_addr = adev->wb.gpu_addr + (ring->rptr_offs * 4);
3846 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc;
3847 mqd->cp_hqd_pq_rptr_report_addr_hi =
3848 upper_32_bits(wb_gpu_addr) & 0xffff;
3849 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR,
3850 mqd->cp_hqd_pq_rptr_report_addr_lo);
3851 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_RPTR_REPORT_ADDR_HI,
3852 mqd->cp_hqd_pq_rptr_report_addr_hi);
3853
3854 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */
3855 wb_gpu_addr = adev->wb.gpu_addr + (ring->wptr_offs * 4);
3856 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc;
3857 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff;
3858 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR,
3859 mqd->cp_hqd_pq_wptr_poll_addr_lo);
3860 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_POLL_ADDR_HI,
3861 mqd->cp_hqd_pq_wptr_poll_addr_hi);
3862
3863 /* enable the doorbell if requested */
3864 if (use_doorbell) {
3865 WREG32_SOC15(GC, 0, mmCP_MEC_DOORBELL_RANGE_LOWER,
3866 (AMDGPU_DOORBELL64_KIQ * 2) << 2);
3867 WREG32_SOC15(GC, 0, mmCP_MEC_DOORBELL_RANGE_UPPER,
3868 (AMDGPU_DOORBELL64_MEC_RING7 * 2) << 2);
3869 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL);
3870 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL,
3871 DOORBELL_OFFSET, ring->doorbell_index);
3872 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_EN, 1);
3873 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_SOURCE, 0);
3874 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, DOORBELL_HIT, 0);
3875 mqd->cp_hqd_pq_doorbell_control = tmp;
3876
3877 } else {
3878 mqd->cp_hqd_pq_doorbell_control = 0;
3879 }
3880 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_DOORBELL_CONTROL,
3881 mqd->cp_hqd_pq_doorbell_control);
3882
3883 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */
3884 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_LO, mqd->cp_hqd_pq_wptr_lo);
3885 WREG32_SOC15(GC, 0, mmCP_HQD_PQ_WPTR_HI, mqd->cp_hqd_pq_wptr_hi);
3886
3887 /* set the vmid for the queue */
3888 mqd->cp_hqd_vmid = 0;
3889 WREG32_SOC15(GC, 0, mmCP_HQD_VMID, mqd->cp_hqd_vmid);
3890
3891 tmp = RREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE);
3892 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x53);
3893 WREG32_SOC15(GC, 0, mmCP_HQD_PERSISTENT_STATE, tmp);
3894 mqd->cp_hqd_persistent_state = tmp;
3895
3896 /* activate the queue */
3897 mqd->cp_hqd_active = 1;
3898 WREG32_SOC15(GC, 0, mmCP_HQD_ACTIVE, mqd->cp_hqd_active);
3899
3900 soc15_grbm_select(adev, 0, 0, 0, 0);
3901 mutex_unlock(&adev->srbm_mutex);
3902
3903 amdgpu_bo_kunmap(ring->mqd_obj);
3904 amdgpu_bo_unreserve(ring->mqd_obj);
3905
3906 if (use_doorbell)
3907 WREG32_FIELD15(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1);
3908
3909 return 0;
3910 }
3911
3912 const struct amdgpu_ip_block_version gfx_v9_0_ip_block =
3913 {
3914 .type = AMD_IP_BLOCK_TYPE_GFX,
3915 .major = 9,
3916 .minor = 0,
3917 .rev = 0,
3918 .funcs = &gfx_v9_0_ip_funcs,
3919 };
CRIS linux kernel tree