1 // SPDX-License-Identifier: GPL-2.0-only 2 /* Copyright (c) 2016-2017 The Linux Foundation. All rights reserved. 3 */ 4 5 #include <linux/kernel.h> 6 #include <linux/types.h> 7 #include <linux/cpumask.h> 8 #include <linux/firmware/qcom/qcom_scm.h> 9 #include <linux/pm_opp.h> 10 #include <linux/nvmem-consumer.h> 11 #include <linux/slab.h> 12 #include "msm_gem.h" 13 #include "msm_mmu.h" 14 #include "a5xx_gpu.h" 15 16 extern bool hang_debug; 17 static void a5xx_dump(struct msm_gpu *gpu); 18 19 #define GPU_PAS_ID 13 20 21 static void update_shadow_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring) 22 { 23 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 24 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 25 26 if (a5xx_gpu->has_whereami) { 27 OUT_PKT7(ring, CP_WHERE_AM_I, 2); 28 OUT_RING(ring, lower_32_bits(shadowptr(a5xx_gpu, ring))); 29 OUT_RING(ring, upper_32_bits(shadowptr(a5xx_gpu, ring))); 30 } 31 } 32 33 void a5xx_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring, 34 bool sync) 35 { 36 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 37 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 38 uint32_t wptr; 39 unsigned long flags; 40 41 /* 42 * Most flush operations need to issue a WHERE_AM_I opcode to sync up 43 * the rptr shadow 44 */ 45 if (sync) 46 update_shadow_rptr(gpu, ring); 47 48 spin_lock_irqsave(&ring->preempt_lock, flags); 49 50 /* Copy the shadow to the actual register */ 51 ring->cur = ring->next; 52 53 /* Make sure to wrap wptr if we need to */ 54 wptr = get_wptr(ring); 55 56 spin_unlock_irqrestore(&ring->preempt_lock, flags); 57 58 /* Make sure everything is posted before making a decision */ 59 mb(); 60 61 /* Update HW if this is the current ring and we are not in preempt */ 62 if (a5xx_gpu->cur_ring == ring && !a5xx_in_preempt(a5xx_gpu)) 63 gpu_write(gpu, REG_A5XX_CP_RB_WPTR, wptr); 64 } 65 66 static void a5xx_submit_in_rb(struct msm_gpu *gpu, struct msm_gem_submit *submit) 67 { 68 struct msm_ringbuffer *ring = submit->ring; 69 struct msm_gem_object *obj; 70 uint32_t *ptr, dwords; 71 unsigned int i; 72 73 for (i = 0; i < submit->nr_cmds; i++) { 74 switch (submit->cmd[i].type) { 75 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 76 break; 77 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 78 if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno) 79 break; 80 fallthrough; 81 case MSM_SUBMIT_CMD_BUF: 82 /* copy commands into RB: */ 83 obj = submit->bos[submit->cmd[i].idx].obj; 84 dwords = submit->cmd[i].size; 85 86 ptr = msm_gem_get_vaddr(&obj->base); 87 88 /* _get_vaddr() shouldn't fail at this point, 89 * since we've already mapped it once in 90 * submit_reloc() 91 */ 92 if (WARN_ON(!ptr)) 93 return; 94 95 for (i = 0; i < dwords; i++) { 96 /* normally the OUT_PKTn() would wait 97 * for space for the packet. But since 98 * we just OUT_RING() the whole thing, 99 * need to call adreno_wait_ring() 100 * ourself: 101 */ 102 adreno_wait_ring(ring, 1); 103 OUT_RING(ring, ptr[i]); 104 } 105 106 msm_gem_put_vaddr(&obj->base); 107 108 break; 109 } 110 } 111 112 a5xx_flush(gpu, ring, true); 113 a5xx_preempt_trigger(gpu); 114 115 /* we might not necessarily have a cmd from userspace to 116 * trigger an event to know that submit has completed, so 117 * do this manually: 118 */ 119 a5xx_idle(gpu, ring); 120 ring->memptrs->fence = submit->seqno; 121 msm_gpu_retire(gpu); 122 } 123 124 static void a5xx_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit) 125 { 126 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 127 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 128 struct msm_ringbuffer *ring = submit->ring; 129 unsigned int i, ibs = 0; 130 131 if (IS_ENABLED(CONFIG_DRM_MSM_GPU_SUDO) && submit->in_rb) { 132 gpu->cur_ctx_seqno = 0; 133 a5xx_submit_in_rb(gpu, submit); 134 return; 135 } 136 137 OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1); 138 OUT_RING(ring, 0x02); 139 140 /* Turn off protected mode to write to special registers */ 141 OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1); 142 OUT_RING(ring, 0); 143 144 /* Set the save preemption record for the ring/command */ 145 OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2); 146 OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[submit->ring->id])); 147 OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[submit->ring->id])); 148 149 /* Turn back on protected mode */ 150 OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1); 151 OUT_RING(ring, 1); 152 153 /* Enable local preemption for finegrain preemption */ 154 OUT_PKT7(ring, CP_PREEMPT_ENABLE_LOCAL, 1); 155 OUT_RING(ring, 0x1); 156 157 /* Allow CP_CONTEXT_SWITCH_YIELD packets in the IB2 */ 158 OUT_PKT7(ring, CP_YIELD_ENABLE, 1); 159 OUT_RING(ring, 0x02); 160 161 /* Submit the commands */ 162 for (i = 0; i < submit->nr_cmds; i++) { 163 switch (submit->cmd[i].type) { 164 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 165 break; 166 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 167 if (gpu->cur_ctx_seqno == submit->queue->ctx->seqno) 168 break; 169 fallthrough; 170 case MSM_SUBMIT_CMD_BUF: 171 OUT_PKT7(ring, CP_INDIRECT_BUFFER_PFE, 3); 172 OUT_RING(ring, lower_32_bits(submit->cmd[i].iova)); 173 OUT_RING(ring, upper_32_bits(submit->cmd[i].iova)); 174 OUT_RING(ring, submit->cmd[i].size); 175 ibs++; 176 break; 177 } 178 179 /* 180 * Periodically update shadow-wptr if needed, so that we 181 * can see partial progress of submits with large # of 182 * cmds.. otherwise we could needlessly stall waiting for 183 * ringbuffer state, simply due to looking at a shadow 184 * rptr value that has not been updated 185 */ 186 if ((ibs % 32) == 0) 187 update_shadow_rptr(gpu, ring); 188 } 189 190 /* 191 * Write the render mode to NULL (0) to indicate to the CP that the IBs 192 * are done rendering - otherwise a lucky preemption would start 193 * replaying from the last checkpoint 194 */ 195 OUT_PKT7(ring, CP_SET_RENDER_MODE, 5); 196 OUT_RING(ring, 0); 197 OUT_RING(ring, 0); 198 OUT_RING(ring, 0); 199 OUT_RING(ring, 0); 200 OUT_RING(ring, 0); 201 202 /* Turn off IB level preemptions */ 203 OUT_PKT7(ring, CP_YIELD_ENABLE, 1); 204 OUT_RING(ring, 0x01); 205 206 /* Write the fence to the scratch register */ 207 OUT_PKT4(ring, REG_A5XX_CP_SCRATCH_REG(2), 1); 208 OUT_RING(ring, submit->seqno); 209 210 /* 211 * Execute a CACHE_FLUSH_TS event. This will ensure that the 212 * timestamp is written to the memory and then triggers the interrupt 213 */ 214 OUT_PKT7(ring, CP_EVENT_WRITE, 4); 215 OUT_RING(ring, CP_EVENT_WRITE_0_EVENT(CACHE_FLUSH_TS) | 216 CP_EVENT_WRITE_0_IRQ); 217 OUT_RING(ring, lower_32_bits(rbmemptr(ring, fence))); 218 OUT_RING(ring, upper_32_bits(rbmemptr(ring, fence))); 219 OUT_RING(ring, submit->seqno); 220 221 /* Yield the floor on command completion */ 222 OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4); 223 /* 224 * If dword[2:1] are non zero, they specify an address for the CP to 225 * write the value of dword[3] to on preemption complete. Write 0 to 226 * skip the write 227 */ 228 OUT_RING(ring, 0x00); 229 OUT_RING(ring, 0x00); 230 /* Data value - not used if the address above is 0 */ 231 OUT_RING(ring, 0x01); 232 /* Set bit 0 to trigger an interrupt on preempt complete */ 233 OUT_RING(ring, 0x01); 234 235 /* A WHERE_AM_I packet is not needed after a YIELD */ 236 a5xx_flush(gpu, ring, false); 237 238 /* Check to see if we need to start preemption */ 239 a5xx_preempt_trigger(gpu); 240 } 241 242 static const struct adreno_five_hwcg_regs { 243 u32 offset; 244 u32 value; 245 } a5xx_hwcg[] = { 246 {REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222}, 247 {REG_A5XX_RBBM_CLOCK_CNTL_SP1, 0x02222222}, 248 {REG_A5XX_RBBM_CLOCK_CNTL_SP2, 0x02222222}, 249 {REG_A5XX_RBBM_CLOCK_CNTL_SP3, 0x02222222}, 250 {REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220}, 251 {REG_A5XX_RBBM_CLOCK_CNTL2_SP1, 0x02222220}, 252 {REG_A5XX_RBBM_CLOCK_CNTL2_SP2, 0x02222220}, 253 {REG_A5XX_RBBM_CLOCK_CNTL2_SP3, 0x02222220}, 254 {REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF}, 255 {REG_A5XX_RBBM_CLOCK_HYST_SP1, 0x0000F3CF}, 256 {REG_A5XX_RBBM_CLOCK_HYST_SP2, 0x0000F3CF}, 257 {REG_A5XX_RBBM_CLOCK_HYST_SP3, 0x0000F3CF}, 258 {REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080}, 259 {REG_A5XX_RBBM_CLOCK_DELAY_SP1, 0x00000080}, 260 {REG_A5XX_RBBM_CLOCK_DELAY_SP2, 0x00000080}, 261 {REG_A5XX_RBBM_CLOCK_DELAY_SP3, 0x00000080}, 262 {REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222}, 263 {REG_A5XX_RBBM_CLOCK_CNTL_TP1, 0x22222222}, 264 {REG_A5XX_RBBM_CLOCK_CNTL_TP2, 0x22222222}, 265 {REG_A5XX_RBBM_CLOCK_CNTL_TP3, 0x22222222}, 266 {REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222}, 267 {REG_A5XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222}, 268 {REG_A5XX_RBBM_CLOCK_CNTL2_TP2, 0x22222222}, 269 {REG_A5XX_RBBM_CLOCK_CNTL2_TP3, 0x22222222}, 270 {REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222}, 271 {REG_A5XX_RBBM_CLOCK_CNTL3_TP1, 0x00002222}, 272 {REG_A5XX_RBBM_CLOCK_CNTL3_TP2, 0x00002222}, 273 {REG_A5XX_RBBM_CLOCK_CNTL3_TP3, 0x00002222}, 274 {REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777}, 275 {REG_A5XX_RBBM_CLOCK_HYST_TP1, 0x77777777}, 276 {REG_A5XX_RBBM_CLOCK_HYST_TP2, 0x77777777}, 277 {REG_A5XX_RBBM_CLOCK_HYST_TP3, 0x77777777}, 278 {REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777}, 279 {REG_A5XX_RBBM_CLOCK_HYST2_TP1, 0x77777777}, 280 {REG_A5XX_RBBM_CLOCK_HYST2_TP2, 0x77777777}, 281 {REG_A5XX_RBBM_CLOCK_HYST2_TP3, 0x77777777}, 282 {REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777}, 283 {REG_A5XX_RBBM_CLOCK_HYST3_TP1, 0x00007777}, 284 {REG_A5XX_RBBM_CLOCK_HYST3_TP2, 0x00007777}, 285 {REG_A5XX_RBBM_CLOCK_HYST3_TP3, 0x00007777}, 286 {REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111}, 287 {REG_A5XX_RBBM_CLOCK_DELAY_TP1, 0x11111111}, 288 {REG_A5XX_RBBM_CLOCK_DELAY_TP2, 0x11111111}, 289 {REG_A5XX_RBBM_CLOCK_DELAY_TP3, 0x11111111}, 290 {REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111}, 291 {REG_A5XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111}, 292 {REG_A5XX_RBBM_CLOCK_DELAY2_TP2, 0x11111111}, 293 {REG_A5XX_RBBM_CLOCK_DELAY2_TP3, 0x11111111}, 294 {REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111}, 295 {REG_A5XX_RBBM_CLOCK_DELAY3_TP1, 0x00001111}, 296 {REG_A5XX_RBBM_CLOCK_DELAY3_TP2, 0x00001111}, 297 {REG_A5XX_RBBM_CLOCK_DELAY3_TP3, 0x00001111}, 298 {REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222}, 299 {REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222}, 300 {REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222}, 301 {REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222}, 302 {REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00444444}, 303 {REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002}, 304 {REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222}, 305 {REG_A5XX_RBBM_CLOCK_CNTL_RB1, 0x22222222}, 306 {REG_A5XX_RBBM_CLOCK_CNTL_RB2, 0x22222222}, 307 {REG_A5XX_RBBM_CLOCK_CNTL_RB3, 0x22222222}, 308 {REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222}, 309 {REG_A5XX_RBBM_CLOCK_CNTL2_RB1, 0x00222222}, 310 {REG_A5XX_RBBM_CLOCK_CNTL2_RB2, 0x00222222}, 311 {REG_A5XX_RBBM_CLOCK_CNTL2_RB3, 0x00222222}, 312 {REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220}, 313 {REG_A5XX_RBBM_CLOCK_CNTL_CCU1, 0x00022220}, 314 {REG_A5XX_RBBM_CLOCK_CNTL_CCU2, 0x00022220}, 315 {REG_A5XX_RBBM_CLOCK_CNTL_CCU3, 0x00022220}, 316 {REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222}, 317 {REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555}, 318 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404}, 319 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU1, 0x04040404}, 320 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU2, 0x04040404}, 321 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU3, 0x04040404}, 322 {REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044}, 323 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002}, 324 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_1, 0x00000002}, 325 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_2, 0x00000002}, 326 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_3, 0x00000002}, 327 {REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011}, 328 {REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222}, 329 {REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222}, 330 {REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222}, 331 {REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000}, 332 {REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004}, 333 {REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000}, 334 {REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000}, 335 {REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000}, 336 {REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200}, 337 {REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222} 338 }, a50x_hwcg[] = { 339 {REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222}, 340 {REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220}, 341 {REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF}, 342 {REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080}, 343 {REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222}, 344 {REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222}, 345 {REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222}, 346 {REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777}, 347 {REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777}, 348 {REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777}, 349 {REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111}, 350 {REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111}, 351 {REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111}, 352 {REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222}, 353 {REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222}, 354 {REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222}, 355 {REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222}, 356 {REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00FFFFF4}, 357 {REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002}, 358 {REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222}, 359 {REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222}, 360 {REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220}, 361 {REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222}, 362 {REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555}, 363 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404}, 364 {REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044}, 365 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002}, 366 {REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011}, 367 {REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222}, 368 {REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222}, 369 {REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222}, 370 {REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000}, 371 {REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004}, 372 {REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000}, 373 {REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000}, 374 {REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000}, 375 {REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200}, 376 {REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222}, 377 }, a512_hwcg[] = { 378 {REG_A5XX_RBBM_CLOCK_CNTL_SP0, 0x02222222}, 379 {REG_A5XX_RBBM_CLOCK_CNTL_SP1, 0x02222222}, 380 {REG_A5XX_RBBM_CLOCK_CNTL2_SP0, 0x02222220}, 381 {REG_A5XX_RBBM_CLOCK_CNTL2_SP1, 0x02222220}, 382 {REG_A5XX_RBBM_CLOCK_HYST_SP0, 0x0000F3CF}, 383 {REG_A5XX_RBBM_CLOCK_HYST_SP1, 0x0000F3CF}, 384 {REG_A5XX_RBBM_CLOCK_DELAY_SP0, 0x00000080}, 385 {REG_A5XX_RBBM_CLOCK_DELAY_SP1, 0x00000080}, 386 {REG_A5XX_RBBM_CLOCK_CNTL_TP0, 0x22222222}, 387 {REG_A5XX_RBBM_CLOCK_CNTL_TP1, 0x22222222}, 388 {REG_A5XX_RBBM_CLOCK_CNTL2_TP0, 0x22222222}, 389 {REG_A5XX_RBBM_CLOCK_CNTL2_TP1, 0x22222222}, 390 {REG_A5XX_RBBM_CLOCK_CNTL3_TP0, 0x00002222}, 391 {REG_A5XX_RBBM_CLOCK_CNTL3_TP1, 0x00002222}, 392 {REG_A5XX_RBBM_CLOCK_HYST_TP0, 0x77777777}, 393 {REG_A5XX_RBBM_CLOCK_HYST_TP1, 0x77777777}, 394 {REG_A5XX_RBBM_CLOCK_HYST2_TP0, 0x77777777}, 395 {REG_A5XX_RBBM_CLOCK_HYST2_TP1, 0x77777777}, 396 {REG_A5XX_RBBM_CLOCK_HYST3_TP0, 0x00007777}, 397 {REG_A5XX_RBBM_CLOCK_HYST3_TP1, 0x00007777}, 398 {REG_A5XX_RBBM_CLOCK_DELAY_TP0, 0x11111111}, 399 {REG_A5XX_RBBM_CLOCK_DELAY_TP1, 0x11111111}, 400 {REG_A5XX_RBBM_CLOCK_DELAY2_TP0, 0x11111111}, 401 {REG_A5XX_RBBM_CLOCK_DELAY2_TP1, 0x11111111}, 402 {REG_A5XX_RBBM_CLOCK_DELAY3_TP0, 0x00001111}, 403 {REG_A5XX_RBBM_CLOCK_DELAY3_TP1, 0x00001111}, 404 {REG_A5XX_RBBM_CLOCK_CNTL_UCHE, 0x22222222}, 405 {REG_A5XX_RBBM_CLOCK_CNTL2_UCHE, 0x22222222}, 406 {REG_A5XX_RBBM_CLOCK_CNTL3_UCHE, 0x22222222}, 407 {REG_A5XX_RBBM_CLOCK_CNTL4_UCHE, 0x00222222}, 408 {REG_A5XX_RBBM_CLOCK_HYST_UCHE, 0x00444444}, 409 {REG_A5XX_RBBM_CLOCK_DELAY_UCHE, 0x00000002}, 410 {REG_A5XX_RBBM_CLOCK_CNTL_RB0, 0x22222222}, 411 {REG_A5XX_RBBM_CLOCK_CNTL_RB1, 0x22222222}, 412 {REG_A5XX_RBBM_CLOCK_CNTL2_RB0, 0x00222222}, 413 {REG_A5XX_RBBM_CLOCK_CNTL2_RB1, 0x00222222}, 414 {REG_A5XX_RBBM_CLOCK_CNTL_CCU0, 0x00022220}, 415 {REG_A5XX_RBBM_CLOCK_CNTL_CCU1, 0x00022220}, 416 {REG_A5XX_RBBM_CLOCK_CNTL_RAC, 0x05522222}, 417 {REG_A5XX_RBBM_CLOCK_CNTL2_RAC, 0x00505555}, 418 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU0, 0x04040404}, 419 {REG_A5XX_RBBM_CLOCK_HYST_RB_CCU1, 0x04040404}, 420 {REG_A5XX_RBBM_CLOCK_HYST_RAC, 0x07444044}, 421 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_0, 0x00000002}, 422 {REG_A5XX_RBBM_CLOCK_DELAY_RB_CCU_L1_1, 0x00000002}, 423 {REG_A5XX_RBBM_CLOCK_DELAY_RAC, 0x00010011}, 424 {REG_A5XX_RBBM_CLOCK_CNTL_TSE_RAS_RBBM, 0x04222222}, 425 {REG_A5XX_RBBM_CLOCK_MODE_GPC, 0x02222222}, 426 {REG_A5XX_RBBM_CLOCK_MODE_VFD, 0x00002222}, 427 {REG_A5XX_RBBM_CLOCK_HYST_TSE_RAS_RBBM, 0x00000000}, 428 {REG_A5XX_RBBM_CLOCK_HYST_GPC, 0x04104004}, 429 {REG_A5XX_RBBM_CLOCK_HYST_VFD, 0x00000000}, 430 {REG_A5XX_RBBM_CLOCK_DELAY_HLSQ, 0x00000000}, 431 {REG_A5XX_RBBM_CLOCK_DELAY_TSE_RAS_RBBM, 0x00004000}, 432 {REG_A5XX_RBBM_CLOCK_DELAY_GPC, 0x00000200}, 433 {REG_A5XX_RBBM_CLOCK_DELAY_VFD, 0x00002222}, 434 }; 435 436 void a5xx_set_hwcg(struct msm_gpu *gpu, bool state) 437 { 438 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 439 const struct adreno_five_hwcg_regs *regs; 440 unsigned int i, sz; 441 442 if (adreno_is_a506(adreno_gpu) || adreno_is_a508(adreno_gpu)) { 443 regs = a50x_hwcg; 444 sz = ARRAY_SIZE(a50x_hwcg); 445 } else if (adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu)) { 446 regs = a512_hwcg; 447 sz = ARRAY_SIZE(a512_hwcg); 448 } else { 449 regs = a5xx_hwcg; 450 sz = ARRAY_SIZE(a5xx_hwcg); 451 } 452 453 for (i = 0; i < sz; i++) 454 gpu_write(gpu, regs[i].offset, 455 state ? regs[i].value : 0); 456 457 if (adreno_is_a540(adreno_gpu)) { 458 gpu_write(gpu, REG_A5XX_RBBM_CLOCK_DELAY_GPMU, state ? 0x00000770 : 0); 459 gpu_write(gpu, REG_A5XX_RBBM_CLOCK_HYST_GPMU, state ? 0x00000004 : 0); 460 } 461 462 gpu_write(gpu, REG_A5XX_RBBM_CLOCK_CNTL, state ? 0xAAA8AA00 : 0); 463 gpu_write(gpu, REG_A5XX_RBBM_ISDB_CNT, state ? 0x182 : 0x180); 464 } 465 466 static int a5xx_me_init(struct msm_gpu *gpu) 467 { 468 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 469 struct msm_ringbuffer *ring = gpu->rb[0]; 470 471 OUT_PKT7(ring, CP_ME_INIT, 8); 472 473 OUT_RING(ring, 0x0000002F); 474 475 /* Enable multiple hardware contexts */ 476 OUT_RING(ring, 0x00000003); 477 478 /* Enable error detection */ 479 OUT_RING(ring, 0x20000000); 480 481 /* Don't enable header dump */ 482 OUT_RING(ring, 0x00000000); 483 OUT_RING(ring, 0x00000000); 484 485 /* Specify workarounds for various microcode issues */ 486 if (adreno_is_a506(adreno_gpu) || adreno_is_a530(adreno_gpu)) { 487 /* Workaround for token end syncs 488 * Force a WFI after every direct-render 3D mode draw and every 489 * 2D mode 3 draw 490 */ 491 OUT_RING(ring, 0x0000000B); 492 } else if (adreno_is_a510(adreno_gpu)) { 493 /* Workaround for token and syncs */ 494 OUT_RING(ring, 0x00000001); 495 } else { 496 /* No workarounds enabled */ 497 OUT_RING(ring, 0x00000000); 498 } 499 500 OUT_RING(ring, 0x00000000); 501 OUT_RING(ring, 0x00000000); 502 503 a5xx_flush(gpu, ring, true); 504 return a5xx_idle(gpu, ring) ? 0 : -EINVAL; 505 } 506 507 static int a5xx_preempt_start(struct msm_gpu *gpu) 508 { 509 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 510 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 511 struct msm_ringbuffer *ring = gpu->rb[0]; 512 513 if (gpu->nr_rings == 1) 514 return 0; 515 516 /* Turn off protected mode to write to special registers */ 517 OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1); 518 OUT_RING(ring, 0); 519 520 /* Set the save preemption record for the ring/command */ 521 OUT_PKT4(ring, REG_A5XX_CP_CONTEXT_SWITCH_SAVE_ADDR_LO, 2); 522 OUT_RING(ring, lower_32_bits(a5xx_gpu->preempt_iova[ring->id])); 523 OUT_RING(ring, upper_32_bits(a5xx_gpu->preempt_iova[ring->id])); 524 525 /* Turn back on protected mode */ 526 OUT_PKT7(ring, CP_SET_PROTECTED_MODE, 1); 527 OUT_RING(ring, 1); 528 529 OUT_PKT7(ring, CP_PREEMPT_ENABLE_GLOBAL, 1); 530 OUT_RING(ring, 0x00); 531 532 OUT_PKT7(ring, CP_PREEMPT_ENABLE_LOCAL, 1); 533 OUT_RING(ring, 0x01); 534 535 OUT_PKT7(ring, CP_YIELD_ENABLE, 1); 536 OUT_RING(ring, 0x01); 537 538 /* Yield the floor on command completion */ 539 OUT_PKT7(ring, CP_CONTEXT_SWITCH_YIELD, 4); 540 OUT_RING(ring, 0x00); 541 OUT_RING(ring, 0x00); 542 OUT_RING(ring, 0x01); 543 OUT_RING(ring, 0x01); 544 545 /* The WHERE_AMI_I packet is not needed after a YIELD is issued */ 546 a5xx_flush(gpu, ring, false); 547 548 return a5xx_idle(gpu, ring) ? 0 : -EINVAL; 549 } 550 551 static void a5xx_ucode_check_version(struct a5xx_gpu *a5xx_gpu, 552 struct drm_gem_object *obj) 553 { 554 u32 *buf = msm_gem_get_vaddr(obj); 555 556 if (IS_ERR(buf)) 557 return; 558 559 /* 560 * If the lowest nibble is 0xa that is an indication that this microcode 561 * has been patched. The actual version is in dword [3] but we only care 562 * about the patchlevel which is the lowest nibble of dword [3] 563 */ 564 if (((buf[0] & 0xf) == 0xa) && (buf[2] & 0xf) >= 1) 565 a5xx_gpu->has_whereami = true; 566 567 msm_gem_put_vaddr(obj); 568 } 569 570 static int a5xx_ucode_load(struct msm_gpu *gpu) 571 { 572 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 573 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 574 int ret; 575 576 if (!a5xx_gpu->pm4_bo) { 577 a5xx_gpu->pm4_bo = adreno_fw_create_bo(gpu, 578 adreno_gpu->fw[ADRENO_FW_PM4], &a5xx_gpu->pm4_iova); 579 580 581 if (IS_ERR(a5xx_gpu->pm4_bo)) { 582 ret = PTR_ERR(a5xx_gpu->pm4_bo); 583 a5xx_gpu->pm4_bo = NULL; 584 DRM_DEV_ERROR(gpu->dev->dev, "could not allocate PM4: %d\n", 585 ret); 586 return ret; 587 } 588 589 msm_gem_object_set_name(a5xx_gpu->pm4_bo, "pm4fw"); 590 } 591 592 if (!a5xx_gpu->pfp_bo) { 593 a5xx_gpu->pfp_bo = adreno_fw_create_bo(gpu, 594 adreno_gpu->fw[ADRENO_FW_PFP], &a5xx_gpu->pfp_iova); 595 596 if (IS_ERR(a5xx_gpu->pfp_bo)) { 597 ret = PTR_ERR(a5xx_gpu->pfp_bo); 598 a5xx_gpu->pfp_bo = NULL; 599 DRM_DEV_ERROR(gpu->dev->dev, "could not allocate PFP: %d\n", 600 ret); 601 return ret; 602 } 603 604 msm_gem_object_set_name(a5xx_gpu->pfp_bo, "pfpfw"); 605 a5xx_ucode_check_version(a5xx_gpu, a5xx_gpu->pfp_bo); 606 } 607 608 if (a5xx_gpu->has_whereami) { 609 if (!a5xx_gpu->shadow_bo) { 610 a5xx_gpu->shadow = msm_gem_kernel_new(gpu->dev, 611 sizeof(u32) * gpu->nr_rings, 612 MSM_BO_WC | MSM_BO_MAP_PRIV, 613 gpu->aspace, &a5xx_gpu->shadow_bo, 614 &a5xx_gpu->shadow_iova); 615 616 if (IS_ERR(a5xx_gpu->shadow)) 617 return PTR_ERR(a5xx_gpu->shadow); 618 619 msm_gem_object_set_name(a5xx_gpu->shadow_bo, "shadow"); 620 } 621 } else if (gpu->nr_rings > 1) { 622 /* Disable preemption if WHERE_AM_I isn't available */ 623 a5xx_preempt_fini(gpu); 624 gpu->nr_rings = 1; 625 } 626 627 return 0; 628 } 629 630 #define SCM_GPU_ZAP_SHADER_RESUME 0 631 632 static int a5xx_zap_shader_resume(struct msm_gpu *gpu) 633 { 634 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 635 int ret; 636 637 /* 638 * Adreno 506 have CPZ Retention feature and doesn't require 639 * to resume zap shader 640 */ 641 if (adreno_is_a506(adreno_gpu)) 642 return 0; 643 644 ret = qcom_scm_set_remote_state(SCM_GPU_ZAP_SHADER_RESUME, GPU_PAS_ID); 645 if (ret) 646 DRM_ERROR("%s: zap-shader resume failed: %d\n", 647 gpu->name, ret); 648 649 return ret; 650 } 651 652 static int a5xx_zap_shader_init(struct msm_gpu *gpu) 653 { 654 static bool loaded; 655 int ret; 656 657 /* 658 * If the zap shader is already loaded into memory we just need to kick 659 * the remote processor to reinitialize it 660 */ 661 if (loaded) 662 return a5xx_zap_shader_resume(gpu); 663 664 ret = adreno_zap_shader_load(gpu, GPU_PAS_ID); 665 666 loaded = !ret; 667 return ret; 668 } 669 670 #define A5XX_INT_MASK (A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \ 671 A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \ 672 A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \ 673 A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \ 674 A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \ 675 A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW | \ 676 A5XX_RBBM_INT_0_MASK_CP_HW_ERROR | \ 677 A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT | \ 678 A5XX_RBBM_INT_0_MASK_CP_SW | \ 679 A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | \ 680 A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS | \ 681 A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP) 682 683 static int a5xx_hw_init(struct msm_gpu *gpu) 684 { 685 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 686 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 687 u32 regbit; 688 int ret; 689 690 gpu_write(gpu, REG_A5XX_VBIF_ROUND_ROBIN_QOS_ARB, 0x00000003); 691 692 if (adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu) || 693 adreno_is_a540(adreno_gpu)) 694 gpu_write(gpu, REG_A5XX_VBIF_GATE_OFF_WRREQ_EN, 0x00000009); 695 696 /* Make all blocks contribute to the GPU BUSY perf counter */ 697 gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_GPU_BUSY_MASKED, 0xFFFFFFFF); 698 699 /* Enable RBBM error reporting bits */ 700 gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL0, 0x00000001); 701 702 if (adreno_gpu->info->quirks & ADRENO_QUIRK_FAULT_DETECT_MASK) { 703 /* 704 * Mask out the activity signals from RB1-3 to avoid false 705 * positives 706 */ 707 708 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL11, 709 0xF0000000); 710 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL12, 711 0xFFFFFFFF); 712 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL13, 713 0xFFFFFFFF); 714 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL14, 715 0xFFFFFFFF); 716 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL15, 717 0xFFFFFFFF); 718 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL16, 719 0xFFFFFFFF); 720 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL17, 721 0xFFFFFFFF); 722 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_MASK_CNTL18, 723 0xFFFFFFFF); 724 } 725 726 /* Enable fault detection */ 727 gpu_write(gpu, REG_A5XX_RBBM_INTERFACE_HANG_INT_CNTL, 728 (1 << 30) | 0xFFFF); 729 730 /* Turn on performance counters */ 731 gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_CNTL, 0x01); 732 733 /* Select CP0 to always count cycles */ 734 gpu_write(gpu, REG_A5XX_CP_PERFCTR_CP_SEL_0, PERF_CP_ALWAYS_COUNT); 735 736 /* Select RBBM0 to countable 6 to get the busy status for devfreq */ 737 gpu_write(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_SEL_0, 6); 738 739 /* Increase VFD cache access so LRZ and other data gets evicted less */ 740 gpu_write(gpu, REG_A5XX_UCHE_CACHE_WAYS, 0x02); 741 742 /* Disable L2 bypass in the UCHE */ 743 gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_LO, 0xFFFF0000); 744 gpu_write(gpu, REG_A5XX_UCHE_TRAP_BASE_HI, 0x0001FFFF); 745 gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_LO, 0xFFFF0000); 746 gpu_write(gpu, REG_A5XX_UCHE_WRITE_THRU_BASE_HI, 0x0001FFFF); 747 748 /* Set the GMEM VA range (0 to gpu->gmem) */ 749 gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_LO, 0x00100000); 750 gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MIN_HI, 0x00000000); 751 gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_LO, 752 0x00100000 + adreno_gpu->gmem - 1); 753 gpu_write(gpu, REG_A5XX_UCHE_GMEM_RANGE_MAX_HI, 0x00000000); 754 755 if (adreno_is_a506(adreno_gpu) || adreno_is_a508(adreno_gpu) || 756 adreno_is_a510(adreno_gpu)) { 757 gpu_write(gpu, REG_A5XX_CP_MEQ_THRESHOLDS, 0x20); 758 if (adreno_is_a506(adreno_gpu) || adreno_is_a508(adreno_gpu)) 759 gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x400); 760 else 761 gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x20); 762 gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_2, 0x40000030); 763 gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_1, 0x20100D0A); 764 } else { 765 gpu_write(gpu, REG_A5XX_CP_MEQ_THRESHOLDS, 0x40); 766 if (adreno_is_a530(adreno_gpu)) 767 gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x40); 768 else 769 gpu_write(gpu, REG_A5XX_CP_MERCIU_SIZE, 0x400); 770 gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_2, 0x80000060); 771 gpu_write(gpu, REG_A5XX_CP_ROQ_THRESHOLDS_1, 0x40201B16); 772 } 773 774 if (adreno_is_a506(adreno_gpu) || adreno_is_a508(adreno_gpu)) 775 gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 776 (0x100 << 11 | 0x100 << 22)); 777 else if (adreno_is_a509(adreno_gpu) || adreno_is_a510(adreno_gpu) || 778 adreno_is_a512(adreno_gpu)) 779 gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 780 (0x200 << 11 | 0x200 << 22)); 781 else 782 gpu_write(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 783 (0x400 << 11 | 0x300 << 22)); 784 785 if (adreno_gpu->info->quirks & ADRENO_QUIRK_TWO_PASS_USE_WFI) 786 gpu_rmw(gpu, REG_A5XX_PC_DBG_ECO_CNTL, 0, (1 << 8)); 787 788 /* 789 * Disable the RB sampler datapath DP2 clock gating optimization 790 * for 1-SP GPUs, as it is enabled by default. 791 */ 792 if (adreno_is_a506(adreno_gpu) || adreno_is_a508(adreno_gpu) || 793 adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu)) 794 gpu_rmw(gpu, REG_A5XX_RB_DBG_ECO_CNTL, 0, (1 << 9)); 795 796 /* Disable UCHE global filter as SP can invalidate/flush independently */ 797 gpu_write(gpu, REG_A5XX_UCHE_MODE_CNTL, BIT(29)); 798 799 /* Enable USE_RETENTION_FLOPS */ 800 gpu_write(gpu, REG_A5XX_CP_CHICKEN_DBG, 0x02000000); 801 802 /* Enable ME/PFP split notification */ 803 gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL1, 0xA6FFFFFF); 804 805 /* 806 * In A5x, CCU can send context_done event of a particular context to 807 * UCHE which ultimately reaches CP even when there is valid 808 * transaction of that context inside CCU. This can let CP to program 809 * config registers, which will make the "valid transaction" inside 810 * CCU to be interpreted differently. This can cause gpu fault. This 811 * bug is fixed in latest A510 revision. To enable this bug fix - 812 * bit[11] of RB_DBG_ECO_CNTL need to be set to 0, default is 1 813 * (disable). For older A510 version this bit is unused. 814 */ 815 if (adreno_is_a510(adreno_gpu)) 816 gpu_rmw(gpu, REG_A5XX_RB_DBG_ECO_CNTL, (1 << 11), 0); 817 818 /* Enable HWCG */ 819 a5xx_set_hwcg(gpu, true); 820 821 gpu_write(gpu, REG_A5XX_RBBM_AHB_CNTL2, 0x0000003F); 822 823 /* Set the highest bank bit */ 824 if (adreno_is_a540(adreno_gpu) || adreno_is_a530(adreno_gpu)) 825 regbit = 2; 826 else 827 regbit = 1; 828 829 gpu_write(gpu, REG_A5XX_TPL1_MODE_CNTL, regbit << 7); 830 gpu_write(gpu, REG_A5XX_RB_MODE_CNTL, regbit << 1); 831 832 if (adreno_is_a509(adreno_gpu) || adreno_is_a512(adreno_gpu) || 833 adreno_is_a540(adreno_gpu)) 834 gpu_write(gpu, REG_A5XX_UCHE_DBG_ECO_CNTL_2, regbit); 835 836 /* Disable All flat shading optimization (ALLFLATOPTDIS) */ 837 gpu_rmw(gpu, REG_A5XX_VPC_DBG_ECO_CNTL, 0, (1 << 10)); 838 839 /* Protect registers from the CP */ 840 gpu_write(gpu, REG_A5XX_CP_PROTECT_CNTL, 0x00000007); 841 842 /* RBBM */ 843 gpu_write(gpu, REG_A5XX_CP_PROTECT(0), ADRENO_PROTECT_RW(0x04, 4)); 844 gpu_write(gpu, REG_A5XX_CP_PROTECT(1), ADRENO_PROTECT_RW(0x08, 8)); 845 gpu_write(gpu, REG_A5XX_CP_PROTECT(2), ADRENO_PROTECT_RW(0x10, 16)); 846 gpu_write(gpu, REG_A5XX_CP_PROTECT(3), ADRENO_PROTECT_RW(0x20, 32)); 847 gpu_write(gpu, REG_A5XX_CP_PROTECT(4), ADRENO_PROTECT_RW(0x40, 64)); 848 gpu_write(gpu, REG_A5XX_CP_PROTECT(5), ADRENO_PROTECT_RW(0x80, 64)); 849 850 /* Content protect */ 851 gpu_write(gpu, REG_A5XX_CP_PROTECT(6), 852 ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO, 853 16)); 854 gpu_write(gpu, REG_A5XX_CP_PROTECT(7), 855 ADRENO_PROTECT_RW(REG_A5XX_RBBM_SECVID_TRUST_CNTL, 2)); 856 857 /* CP */ 858 gpu_write(gpu, REG_A5XX_CP_PROTECT(8), ADRENO_PROTECT_RW(0x800, 64)); 859 gpu_write(gpu, REG_A5XX_CP_PROTECT(9), ADRENO_PROTECT_RW(0x840, 8)); 860 gpu_write(gpu, REG_A5XX_CP_PROTECT(10), ADRENO_PROTECT_RW(0x880, 32)); 861 gpu_write(gpu, REG_A5XX_CP_PROTECT(11), ADRENO_PROTECT_RW(0xAA0, 1)); 862 863 /* RB */ 864 gpu_write(gpu, REG_A5XX_CP_PROTECT(12), ADRENO_PROTECT_RW(0xCC0, 1)); 865 gpu_write(gpu, REG_A5XX_CP_PROTECT(13), ADRENO_PROTECT_RW(0xCF0, 2)); 866 867 /* VPC */ 868 gpu_write(gpu, REG_A5XX_CP_PROTECT(14), ADRENO_PROTECT_RW(0xE68, 8)); 869 gpu_write(gpu, REG_A5XX_CP_PROTECT(15), ADRENO_PROTECT_RW(0xE70, 16)); 870 871 /* UCHE */ 872 gpu_write(gpu, REG_A5XX_CP_PROTECT(16), ADRENO_PROTECT_RW(0xE80, 16)); 873 874 /* SMMU */ 875 gpu_write(gpu, REG_A5XX_CP_PROTECT(17), 876 ADRENO_PROTECT_RW(0x10000, 0x8000)); 877 878 gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_CNTL, 0); 879 /* 880 * Disable the trusted memory range - we don't actually supported secure 881 * memory rendering at this point in time and we don't want to block off 882 * part of the virtual memory space. 883 */ 884 gpu_write64(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_BASE_LO, 0x00000000); 885 gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_TRUSTED_SIZE, 0x00000000); 886 887 /* Put the GPU into 64 bit by default */ 888 gpu_write(gpu, REG_A5XX_CP_ADDR_MODE_CNTL, 0x1); 889 gpu_write(gpu, REG_A5XX_VSC_ADDR_MODE_CNTL, 0x1); 890 gpu_write(gpu, REG_A5XX_GRAS_ADDR_MODE_CNTL, 0x1); 891 gpu_write(gpu, REG_A5XX_RB_ADDR_MODE_CNTL, 0x1); 892 gpu_write(gpu, REG_A5XX_PC_ADDR_MODE_CNTL, 0x1); 893 gpu_write(gpu, REG_A5XX_HLSQ_ADDR_MODE_CNTL, 0x1); 894 gpu_write(gpu, REG_A5XX_VFD_ADDR_MODE_CNTL, 0x1); 895 gpu_write(gpu, REG_A5XX_VPC_ADDR_MODE_CNTL, 0x1); 896 gpu_write(gpu, REG_A5XX_UCHE_ADDR_MODE_CNTL, 0x1); 897 gpu_write(gpu, REG_A5XX_SP_ADDR_MODE_CNTL, 0x1); 898 gpu_write(gpu, REG_A5XX_TPL1_ADDR_MODE_CNTL, 0x1); 899 gpu_write(gpu, REG_A5XX_RBBM_SECVID_TSB_ADDR_MODE_CNTL, 0x1); 900 901 /* 902 * VPC corner case with local memory load kill leads to corrupt 903 * internal state. Normal Disable does not work for all a5x chips. 904 * So do the following setting to disable it. 905 */ 906 if (adreno_gpu->info->quirks & ADRENO_QUIRK_LMLOADKILL_DISABLE) { 907 gpu_rmw(gpu, REG_A5XX_VPC_DBG_ECO_CNTL, 0, BIT(23)); 908 gpu_rmw(gpu, REG_A5XX_HLSQ_DBG_ECO_CNTL, BIT(18), 0); 909 } 910 911 ret = adreno_hw_init(gpu); 912 if (ret) 913 return ret; 914 915 if (adreno_is_a530(adreno_gpu) || adreno_is_a540(adreno_gpu)) 916 a5xx_gpmu_ucode_init(gpu); 917 918 gpu_write64(gpu, REG_A5XX_CP_ME_INSTR_BASE_LO, a5xx_gpu->pm4_iova); 919 gpu_write64(gpu, REG_A5XX_CP_PFP_INSTR_BASE_LO, a5xx_gpu->pfp_iova); 920 921 /* Set the ringbuffer address */ 922 gpu_write64(gpu, REG_A5XX_CP_RB_BASE, gpu->rb[0]->iova); 923 924 /* 925 * If the microcode supports the WHERE_AM_I opcode then we can use that 926 * in lieu of the RPTR shadow and enable preemption. Otherwise, we 927 * can't safely use the RPTR shadow or preemption. In either case, the 928 * RPTR shadow should be disabled in hardware. 929 */ 930 gpu_write(gpu, REG_A5XX_CP_RB_CNTL, 931 MSM_GPU_RB_CNTL_DEFAULT | AXXX_CP_RB_CNTL_NO_UPDATE); 932 933 /* Configure the RPTR shadow if needed: */ 934 if (a5xx_gpu->shadow_bo) { 935 gpu_write64(gpu, REG_A5XX_CP_RB_RPTR_ADDR, 936 shadowptr(a5xx_gpu, gpu->rb[0])); 937 } 938 939 a5xx_preempt_hw_init(gpu); 940 941 /* Disable the interrupts through the initial bringup stage */ 942 gpu_write(gpu, REG_A5XX_RBBM_INT_0_MASK, A5XX_INT_MASK); 943 944 /* Clear ME_HALT to start the micro engine */ 945 gpu_write(gpu, REG_A5XX_CP_PFP_ME_CNTL, 0); 946 ret = a5xx_me_init(gpu); 947 if (ret) 948 return ret; 949 950 ret = a5xx_power_init(gpu); 951 if (ret) 952 return ret; 953 954 /* 955 * Send a pipeline event stat to get misbehaving counters to start 956 * ticking correctly 957 */ 958 if (adreno_is_a530(adreno_gpu)) { 959 OUT_PKT7(gpu->rb[0], CP_EVENT_WRITE, 1); 960 OUT_RING(gpu->rb[0], CP_EVENT_WRITE_0_EVENT(STAT_EVENT)); 961 962 a5xx_flush(gpu, gpu->rb[0], true); 963 if (!a5xx_idle(gpu, gpu->rb[0])) 964 return -EINVAL; 965 } 966 967 /* 968 * If the chip that we are using does support loading one, then 969 * try to load a zap shader into the secure world. If successful 970 * we can use the CP to switch out of secure mode. If not then we 971 * have no resource but to try to switch ourselves out manually. If we 972 * guessed wrong then access to the RBBM_SECVID_TRUST_CNTL register will 973 * be blocked and a permissions violation will soon follow. 974 */ 975 ret = a5xx_zap_shader_init(gpu); 976 if (!ret) { 977 OUT_PKT7(gpu->rb[0], CP_SET_SECURE_MODE, 1); 978 OUT_RING(gpu->rb[0], 0x00000000); 979 980 a5xx_flush(gpu, gpu->rb[0], true); 981 if (!a5xx_idle(gpu, gpu->rb[0])) 982 return -EINVAL; 983 } else if (ret == -ENODEV) { 984 /* 985 * This device does not use zap shader (but print a warning 986 * just in case someone got their dt wrong.. hopefully they 987 * have a debug UART to realize the error of their ways... 988 * if you mess this up you are about to crash horribly) 989 */ 990 dev_warn_once(gpu->dev->dev, 991 "Zap shader not enabled - using SECVID_TRUST_CNTL instead\n"); 992 gpu_write(gpu, REG_A5XX_RBBM_SECVID_TRUST_CNTL, 0x0); 993 } else { 994 return ret; 995 } 996 997 /* Last step - yield the ringbuffer */ 998 a5xx_preempt_start(gpu); 999 1000 return 0; 1001 } 1002 1003 static void a5xx_recover(struct msm_gpu *gpu) 1004 { 1005 int i; 1006 1007 adreno_dump_info(gpu); 1008 1009 for (i = 0; i < 8; i++) { 1010 printk("CP_SCRATCH_REG%d: %u\n", i, 1011 gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(i))); 1012 } 1013 1014 if (hang_debug) 1015 a5xx_dump(gpu); 1016 1017 gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 1); 1018 gpu_read(gpu, REG_A5XX_RBBM_SW_RESET_CMD); 1019 gpu_write(gpu, REG_A5XX_RBBM_SW_RESET_CMD, 0); 1020 adreno_recover(gpu); 1021 } 1022 1023 static void a5xx_destroy(struct msm_gpu *gpu) 1024 { 1025 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 1026 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 1027 1028 DBG("%s", gpu->name); 1029 1030 a5xx_preempt_fini(gpu); 1031 1032 if (a5xx_gpu->pm4_bo) { 1033 msm_gem_unpin_iova(a5xx_gpu->pm4_bo, gpu->aspace); 1034 drm_gem_object_put(a5xx_gpu->pm4_bo); 1035 } 1036 1037 if (a5xx_gpu->pfp_bo) { 1038 msm_gem_unpin_iova(a5xx_gpu->pfp_bo, gpu->aspace); 1039 drm_gem_object_put(a5xx_gpu->pfp_bo); 1040 } 1041 1042 if (a5xx_gpu->gpmu_bo) { 1043 msm_gem_unpin_iova(a5xx_gpu->gpmu_bo, gpu->aspace); 1044 drm_gem_object_put(a5xx_gpu->gpmu_bo); 1045 } 1046 1047 if (a5xx_gpu->shadow_bo) { 1048 msm_gem_unpin_iova(a5xx_gpu->shadow_bo, gpu->aspace); 1049 drm_gem_object_put(a5xx_gpu->shadow_bo); 1050 } 1051 1052 adreno_gpu_cleanup(adreno_gpu); 1053 kfree(a5xx_gpu); 1054 } 1055 1056 static inline bool _a5xx_check_idle(struct msm_gpu *gpu) 1057 { 1058 if (gpu_read(gpu, REG_A5XX_RBBM_STATUS) & ~A5XX_RBBM_STATUS_HI_BUSY) 1059 return false; 1060 1061 /* 1062 * Nearly every abnormality ends up pausing the GPU and triggering a 1063 * fault so we can safely just watch for this one interrupt to fire 1064 */ 1065 return !(gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS) & 1066 A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT); 1067 } 1068 1069 bool a5xx_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring) 1070 { 1071 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 1072 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 1073 1074 if (ring != a5xx_gpu->cur_ring) { 1075 WARN(1, "Tried to idle a non-current ringbuffer\n"); 1076 return false; 1077 } 1078 1079 /* wait for CP to drain ringbuffer: */ 1080 if (!adreno_idle(gpu, ring)) 1081 return false; 1082 1083 if (spin_until(_a5xx_check_idle(gpu))) { 1084 DRM_ERROR("%s: %ps: timeout waiting for GPU to idle: status %8.8X irq %8.8X rptr/wptr %d/%d\n", 1085 gpu->name, __builtin_return_address(0), 1086 gpu_read(gpu, REG_A5XX_RBBM_STATUS), 1087 gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS), 1088 gpu_read(gpu, REG_A5XX_CP_RB_RPTR), 1089 gpu_read(gpu, REG_A5XX_CP_RB_WPTR)); 1090 return false; 1091 } 1092 1093 return true; 1094 } 1095 1096 static int a5xx_fault_handler(void *arg, unsigned long iova, int flags, void *data) 1097 { 1098 struct msm_gpu *gpu = arg; 1099 struct adreno_smmu_fault_info *info = data; 1100 char block[12] = "unknown"; 1101 u32 scratch[] = { 1102 gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(4)), 1103 gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(5)), 1104 gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(6)), 1105 gpu_read(gpu, REG_A5XX_CP_SCRATCH_REG(7)), 1106 }; 1107 1108 if (info) 1109 snprintf(block, sizeof(block), "%x", info->fsynr1); 1110 1111 return adreno_fault_handler(gpu, iova, flags, info, block, scratch); 1112 } 1113 1114 static void a5xx_cp_err_irq(struct msm_gpu *gpu) 1115 { 1116 u32 status = gpu_read(gpu, REG_A5XX_CP_INTERRUPT_STATUS); 1117 1118 if (status & A5XX_CP_INT_CP_OPCODE_ERROR) { 1119 u32 val; 1120 1121 gpu_write(gpu, REG_A5XX_CP_PFP_STAT_ADDR, 0); 1122 1123 /* 1124 * REG_A5XX_CP_PFP_STAT_DATA is indexed, and we want index 1 so 1125 * read it twice 1126 */ 1127 1128 gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA); 1129 val = gpu_read(gpu, REG_A5XX_CP_PFP_STAT_DATA); 1130 1131 dev_err_ratelimited(gpu->dev->dev, "CP | opcode error | possible opcode=0x%8.8X\n", 1132 val); 1133 } 1134 1135 if (status & A5XX_CP_INT_CP_HW_FAULT_ERROR) 1136 dev_err_ratelimited(gpu->dev->dev, "CP | HW fault | status=0x%8.8X\n", 1137 gpu_read(gpu, REG_A5XX_CP_HW_FAULT)); 1138 1139 if (status & A5XX_CP_INT_CP_DMA_ERROR) 1140 dev_err_ratelimited(gpu->dev->dev, "CP | DMA error\n"); 1141 1142 if (status & A5XX_CP_INT_CP_REGISTER_PROTECTION_ERROR) { 1143 u32 val = gpu_read(gpu, REG_A5XX_CP_PROTECT_STATUS); 1144 1145 dev_err_ratelimited(gpu->dev->dev, 1146 "CP | protected mode error | %s | addr=0x%8.8X | status=0x%8.8X\n", 1147 val & (1 << 24) ? "WRITE" : "READ", 1148 (val & 0xFFFFF) >> 2, val); 1149 } 1150 1151 if (status & A5XX_CP_INT_CP_AHB_ERROR) { 1152 u32 status = gpu_read(gpu, REG_A5XX_CP_AHB_FAULT); 1153 const char *access[16] = { "reserved", "reserved", 1154 "timestamp lo", "timestamp hi", "pfp read", "pfp write", 1155 "", "", "me read", "me write", "", "", "crashdump read", 1156 "crashdump write" }; 1157 1158 dev_err_ratelimited(gpu->dev->dev, 1159 "CP | AHB error | addr=%X access=%s error=%d | status=0x%8.8X\n", 1160 status & 0xFFFFF, access[(status >> 24) & 0xF], 1161 (status & (1 << 31)), status); 1162 } 1163 } 1164 1165 static void a5xx_rbbm_err_irq(struct msm_gpu *gpu, u32 status) 1166 { 1167 if (status & A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR) { 1168 u32 val = gpu_read(gpu, REG_A5XX_RBBM_AHB_ERROR_STATUS); 1169 1170 dev_err_ratelimited(gpu->dev->dev, 1171 "RBBM | AHB bus error | %s | addr=0x%X | ports=0x%X:0x%X\n", 1172 val & (1 << 28) ? "WRITE" : "READ", 1173 (val & 0xFFFFF) >> 2, (val >> 20) & 0x3, 1174 (val >> 24) & 0xF); 1175 1176 /* Clear the error */ 1177 gpu_write(gpu, REG_A5XX_RBBM_AHB_CMD, (1 << 4)); 1178 1179 /* Clear the interrupt */ 1180 gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD, 1181 A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR); 1182 } 1183 1184 if (status & A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT) 1185 dev_err_ratelimited(gpu->dev->dev, "RBBM | AHB transfer timeout\n"); 1186 1187 if (status & A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT) 1188 dev_err_ratelimited(gpu->dev->dev, "RBBM | ME master split | status=0x%X\n", 1189 gpu_read(gpu, REG_A5XX_RBBM_AHB_ME_SPLIT_STATUS)); 1190 1191 if (status & A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT) 1192 dev_err_ratelimited(gpu->dev->dev, "RBBM | PFP master split | status=0x%X\n", 1193 gpu_read(gpu, REG_A5XX_RBBM_AHB_PFP_SPLIT_STATUS)); 1194 1195 if (status & A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT) 1196 dev_err_ratelimited(gpu->dev->dev, "RBBM | ETS master split | status=0x%X\n", 1197 gpu_read(gpu, REG_A5XX_RBBM_AHB_ETS_SPLIT_STATUS)); 1198 1199 if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW) 1200 dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB ASYNC overflow\n"); 1201 1202 if (status & A5XX_RBBM_INT_0_MASK_RBBM_ATB_BUS_OVERFLOW) 1203 dev_err_ratelimited(gpu->dev->dev, "RBBM | ATB bus overflow\n"); 1204 } 1205 1206 static void a5xx_uche_err_irq(struct msm_gpu *gpu) 1207 { 1208 uint64_t addr = (uint64_t) gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_HI); 1209 1210 addr |= gpu_read(gpu, REG_A5XX_UCHE_TRAP_LOG_LO); 1211 1212 dev_err_ratelimited(gpu->dev->dev, "UCHE | Out of bounds access | addr=0x%llX\n", 1213 addr); 1214 } 1215 1216 static void a5xx_gpmu_err_irq(struct msm_gpu *gpu) 1217 { 1218 dev_err_ratelimited(gpu->dev->dev, "GPMU | voltage droop\n"); 1219 } 1220 1221 static void a5xx_fault_detect_irq(struct msm_gpu *gpu) 1222 { 1223 struct drm_device *dev = gpu->dev; 1224 struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu); 1225 1226 /* 1227 * If stalled on SMMU fault, we could trip the GPU's hang detection, 1228 * but the fault handler will trigger the devcore dump, and we want 1229 * to otherwise resume normally rather than killing the submit, so 1230 * just bail. 1231 */ 1232 if (gpu_read(gpu, REG_A5XX_RBBM_STATUS3) & BIT(24)) 1233 return; 1234 1235 DRM_DEV_ERROR(dev->dev, "gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n", 1236 ring ? ring->id : -1, ring ? ring->fctx->last_fence : 0, 1237 gpu_read(gpu, REG_A5XX_RBBM_STATUS), 1238 gpu_read(gpu, REG_A5XX_CP_RB_RPTR), 1239 gpu_read(gpu, REG_A5XX_CP_RB_WPTR), 1240 gpu_read64(gpu, REG_A5XX_CP_IB1_BASE), 1241 gpu_read(gpu, REG_A5XX_CP_IB1_BUFSZ), 1242 gpu_read64(gpu, REG_A5XX_CP_IB2_BASE), 1243 gpu_read(gpu, REG_A5XX_CP_IB2_BUFSZ)); 1244 1245 /* Turn off the hangcheck timer to keep it from bothering us */ 1246 del_timer(&gpu->hangcheck_timer); 1247 1248 kthread_queue_work(gpu->worker, &gpu->recover_work); 1249 } 1250 1251 #define RBBM_ERROR_MASK \ 1252 (A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR | \ 1253 A5XX_RBBM_INT_0_MASK_RBBM_TRANSFER_TIMEOUT | \ 1254 A5XX_RBBM_INT_0_MASK_RBBM_ME_MS_TIMEOUT | \ 1255 A5XX_RBBM_INT_0_MASK_RBBM_PFP_MS_TIMEOUT | \ 1256 A5XX_RBBM_INT_0_MASK_RBBM_ETS_MS_TIMEOUT | \ 1257 A5XX_RBBM_INT_0_MASK_RBBM_ATB_ASYNC_OVERFLOW) 1258 1259 static irqreturn_t a5xx_irq(struct msm_gpu *gpu) 1260 { 1261 struct msm_drm_private *priv = gpu->dev->dev_private; 1262 u32 status = gpu_read(gpu, REG_A5XX_RBBM_INT_0_STATUS); 1263 1264 /* 1265 * Clear all the interrupts except RBBM_AHB_ERROR - if we clear it 1266 * before the source is cleared the interrupt will storm. 1267 */ 1268 gpu_write(gpu, REG_A5XX_RBBM_INT_CLEAR_CMD, 1269 status & ~A5XX_RBBM_INT_0_MASK_RBBM_AHB_ERROR); 1270 1271 if (priv->disable_err_irq) { 1272 status &= A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS | 1273 A5XX_RBBM_INT_0_MASK_CP_SW; 1274 } 1275 1276 /* Pass status to a5xx_rbbm_err_irq because we've already cleared it */ 1277 if (status & RBBM_ERROR_MASK) 1278 a5xx_rbbm_err_irq(gpu, status); 1279 1280 if (status & A5XX_RBBM_INT_0_MASK_CP_HW_ERROR) 1281 a5xx_cp_err_irq(gpu); 1282 1283 if (status & A5XX_RBBM_INT_0_MASK_MISC_HANG_DETECT) 1284 a5xx_fault_detect_irq(gpu); 1285 1286 if (status & A5XX_RBBM_INT_0_MASK_UCHE_OOB_ACCESS) 1287 a5xx_uche_err_irq(gpu); 1288 1289 if (status & A5XX_RBBM_INT_0_MASK_GPMU_VOLTAGE_DROOP) 1290 a5xx_gpmu_err_irq(gpu); 1291 1292 if (status & A5XX_RBBM_INT_0_MASK_CP_CACHE_FLUSH_TS) { 1293 a5xx_preempt_trigger(gpu); 1294 msm_gpu_retire(gpu); 1295 } 1296 1297 if (status & A5XX_RBBM_INT_0_MASK_CP_SW) 1298 a5xx_preempt_irq(gpu); 1299 1300 return IRQ_HANDLED; 1301 } 1302 1303 static const u32 a5xx_registers[] = { 1304 0x0000, 0x0002, 0x0004, 0x0020, 0x0022, 0x0026, 0x0029, 0x002B, 1305 0x002E, 0x0035, 0x0038, 0x0042, 0x0044, 0x0044, 0x0047, 0x0095, 1306 0x0097, 0x00BB, 0x03A0, 0x0464, 0x0469, 0x046F, 0x04D2, 0x04D3, 1307 0x04E0, 0x0533, 0x0540, 0x0555, 0x0800, 0x081A, 0x081F, 0x0841, 1308 0x0860, 0x0860, 0x0880, 0x08A0, 0x0B00, 0x0B12, 0x0B15, 0x0B28, 1309 0x0B78, 0x0B7F, 0x0BB0, 0x0BBD, 0x0BC0, 0x0BC6, 0x0BD0, 0x0C53, 1310 0x0C60, 0x0C61, 0x0C80, 0x0C82, 0x0C84, 0x0C85, 0x0C90, 0x0C98, 1311 0x0CA0, 0x0CA0, 0x0CB0, 0x0CB2, 0x2180, 0x2185, 0x2580, 0x2585, 1312 0x0CC1, 0x0CC1, 0x0CC4, 0x0CC7, 0x0CCC, 0x0CCC, 0x0CD0, 0x0CD8, 1313 0x0CE0, 0x0CE5, 0x0CE8, 0x0CE8, 0x0CEC, 0x0CF1, 0x0CFB, 0x0D0E, 1314 0x2100, 0x211E, 0x2140, 0x2145, 0x2500, 0x251E, 0x2540, 0x2545, 1315 0x0D10, 0x0D17, 0x0D20, 0x0D23, 0x0D30, 0x0D30, 0x20C0, 0x20C0, 1316 0x24C0, 0x24C0, 0x0E40, 0x0E43, 0x0E4A, 0x0E4A, 0x0E50, 0x0E57, 1317 0x0E60, 0x0E7C, 0x0E80, 0x0E8E, 0x0E90, 0x0E96, 0x0EA0, 0x0EA8, 1318 0x0EB0, 0x0EB2, 0xE140, 0xE147, 0xE150, 0xE187, 0xE1A0, 0xE1A9, 1319 0xE1B0, 0xE1B6, 0xE1C0, 0xE1C7, 0xE1D0, 0xE1D1, 0xE200, 0xE201, 1320 0xE210, 0xE21C, 0xE240, 0xE268, 0xE000, 0xE006, 0xE010, 0xE09A, 1321 0xE0A0, 0xE0A4, 0xE0AA, 0xE0EB, 0xE100, 0xE105, 0xE380, 0xE38F, 1322 0xE3B0, 0xE3B0, 0xE400, 0xE405, 0xE408, 0xE4E9, 0xE4F0, 0xE4F0, 1323 0xE280, 0xE280, 0xE282, 0xE2A3, 0xE2A5, 0xE2C2, 0xE940, 0xE947, 1324 0xE950, 0xE987, 0xE9A0, 0xE9A9, 0xE9B0, 0xE9B6, 0xE9C0, 0xE9C7, 1325 0xE9D0, 0xE9D1, 0xEA00, 0xEA01, 0xEA10, 0xEA1C, 0xEA40, 0xEA68, 1326 0xE800, 0xE806, 0xE810, 0xE89A, 0xE8A0, 0xE8A4, 0xE8AA, 0xE8EB, 1327 0xE900, 0xE905, 0xEB80, 0xEB8F, 0xEBB0, 0xEBB0, 0xEC00, 0xEC05, 1328 0xEC08, 0xECE9, 0xECF0, 0xECF0, 0xEA80, 0xEA80, 0xEA82, 0xEAA3, 1329 0xEAA5, 0xEAC2, 0xA800, 0xA800, 0xA820, 0xA828, 0xA840, 0xA87D, 1330 0XA880, 0xA88D, 0xA890, 0xA8A3, 0xA8D0, 0xA8D8, 0xA8E0, 0xA8F5, 1331 0xAC60, 0xAC60, ~0, 1332 }; 1333 1334 static void a5xx_dump(struct msm_gpu *gpu) 1335 { 1336 DRM_DEV_INFO(gpu->dev->dev, "status: %08x\n", 1337 gpu_read(gpu, REG_A5XX_RBBM_STATUS)); 1338 adreno_dump(gpu); 1339 } 1340 1341 static int a5xx_pm_resume(struct msm_gpu *gpu) 1342 { 1343 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 1344 int ret; 1345 1346 /* Turn on the core power */ 1347 ret = msm_gpu_pm_resume(gpu); 1348 if (ret) 1349 return ret; 1350 1351 /* Adreno 506, 508, 509, 510, 512 needs manual RBBM sus/res control */ 1352 if (!(adreno_is_a530(adreno_gpu) || adreno_is_a540(adreno_gpu))) { 1353 /* Halt the sp_input_clk at HM level */ 1354 gpu_write(gpu, REG_A5XX_RBBM_CLOCK_CNTL, 0x00000055); 1355 a5xx_set_hwcg(gpu, true); 1356 /* Turn on sp_input_clk at HM level */ 1357 gpu_rmw(gpu, REG_A5XX_RBBM_CLOCK_CNTL, 0xff, 0); 1358 return 0; 1359 } 1360 1361 /* Turn the RBCCU domain first to limit the chances of voltage droop */ 1362 gpu_write(gpu, REG_A5XX_GPMU_RBCCU_POWER_CNTL, 0x778000); 1363 1364 /* Wait 3 usecs before polling */ 1365 udelay(3); 1366 1367 ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS, 1368 (1 << 20), (1 << 20)); 1369 if (ret) { 1370 DRM_ERROR("%s: timeout waiting for RBCCU GDSC enable: %X\n", 1371 gpu->name, 1372 gpu_read(gpu, REG_A5XX_GPMU_RBCCU_PWR_CLK_STATUS)); 1373 return ret; 1374 } 1375 1376 /* Turn on the SP domain */ 1377 gpu_write(gpu, REG_A5XX_GPMU_SP_POWER_CNTL, 0x778000); 1378 ret = spin_usecs(gpu, 20, REG_A5XX_GPMU_SP_PWR_CLK_STATUS, 1379 (1 << 20), (1 << 20)); 1380 if (ret) 1381 DRM_ERROR("%s: timeout waiting for SP GDSC enable\n", 1382 gpu->name); 1383 1384 return ret; 1385 } 1386 1387 static int a5xx_pm_suspend(struct msm_gpu *gpu) 1388 { 1389 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 1390 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 1391 u32 mask = 0xf; 1392 int i, ret; 1393 1394 /* A506, A508, A510 have 3 XIN ports in VBIF */ 1395 if (adreno_is_a506(adreno_gpu) || adreno_is_a508(adreno_gpu) || 1396 adreno_is_a510(adreno_gpu)) 1397 mask = 0x7; 1398 1399 /* Clear the VBIF pipe before shutting down */ 1400 gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, mask); 1401 spin_until((gpu_read(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL1) & 1402 mask) == mask); 1403 1404 gpu_write(gpu, REG_A5XX_VBIF_XIN_HALT_CTRL0, 0); 1405 1406 /* 1407 * Reset the VBIF before power collapse to avoid issue with FIFO 1408 * entries on Adreno A510 and A530 (the others will tend to lock up) 1409 */ 1410 if (adreno_is_a510(adreno_gpu) || adreno_is_a530(adreno_gpu)) { 1411 gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x003C0000); 1412 gpu_write(gpu, REG_A5XX_RBBM_BLOCK_SW_RESET_CMD, 0x00000000); 1413 } 1414 1415 ret = msm_gpu_pm_suspend(gpu); 1416 if (ret) 1417 return ret; 1418 1419 if (a5xx_gpu->has_whereami) 1420 for (i = 0; i < gpu->nr_rings; i++) 1421 a5xx_gpu->shadow[i] = 0; 1422 1423 return 0; 1424 } 1425 1426 static int a5xx_get_timestamp(struct msm_gpu *gpu, uint64_t *value) 1427 { 1428 *value = gpu_read64(gpu, REG_A5XX_RBBM_ALWAYSON_COUNTER_LO); 1429 1430 return 0; 1431 } 1432 1433 struct a5xx_crashdumper { 1434 void *ptr; 1435 struct drm_gem_object *bo; 1436 u64 iova; 1437 }; 1438 1439 struct a5xx_gpu_state { 1440 struct msm_gpu_state base; 1441 u32 *hlsqregs; 1442 }; 1443 1444 static int a5xx_crashdumper_init(struct msm_gpu *gpu, 1445 struct a5xx_crashdumper *dumper) 1446 { 1447 dumper->ptr = msm_gem_kernel_new(gpu->dev, 1448 SZ_1M, MSM_BO_WC, gpu->aspace, 1449 &dumper->bo, &dumper->iova); 1450 1451 if (!IS_ERR(dumper->ptr)) 1452 msm_gem_object_set_name(dumper->bo, "crashdump"); 1453 1454 return PTR_ERR_OR_ZERO(dumper->ptr); 1455 } 1456 1457 static int a5xx_crashdumper_run(struct msm_gpu *gpu, 1458 struct a5xx_crashdumper *dumper) 1459 { 1460 u32 val; 1461 1462 if (IS_ERR_OR_NULL(dumper->ptr)) 1463 return -EINVAL; 1464 1465 gpu_write64(gpu, REG_A5XX_CP_CRASH_SCRIPT_BASE_LO, dumper->iova); 1466 1467 gpu_write(gpu, REG_A5XX_CP_CRASH_DUMP_CNTL, 1); 1468 1469 return gpu_poll_timeout(gpu, REG_A5XX_CP_CRASH_DUMP_CNTL, val, 1470 val & 0x04, 100, 10000); 1471 } 1472 1473 /* 1474 * These are a list of the registers that need to be read through the HLSQ 1475 * aperture through the crashdumper. These are not nominally accessible from 1476 * the CPU on a secure platform. 1477 */ 1478 static const struct { 1479 u32 type; 1480 u32 regoffset; 1481 u32 count; 1482 } a5xx_hlsq_aperture_regs[] = { 1483 { 0x35, 0xe00, 0x32 }, /* HSLQ non-context */ 1484 { 0x31, 0x2080, 0x1 }, /* HLSQ 2D context 0 */ 1485 { 0x33, 0x2480, 0x1 }, /* HLSQ 2D context 1 */ 1486 { 0x32, 0xe780, 0x62 }, /* HLSQ 3D context 0 */ 1487 { 0x34, 0xef80, 0x62 }, /* HLSQ 3D context 1 */ 1488 { 0x3f, 0x0ec0, 0x40 }, /* SP non-context */ 1489 { 0x3d, 0x2040, 0x1 }, /* SP 2D context 0 */ 1490 { 0x3b, 0x2440, 0x1 }, /* SP 2D context 1 */ 1491 { 0x3e, 0xe580, 0x170 }, /* SP 3D context 0 */ 1492 { 0x3c, 0xed80, 0x170 }, /* SP 3D context 1 */ 1493 { 0x3a, 0x0f00, 0x1c }, /* TP non-context */ 1494 { 0x38, 0x2000, 0xa }, /* TP 2D context 0 */ 1495 { 0x36, 0x2400, 0xa }, /* TP 2D context 1 */ 1496 { 0x39, 0xe700, 0x80 }, /* TP 3D context 0 */ 1497 { 0x37, 0xef00, 0x80 }, /* TP 3D context 1 */ 1498 }; 1499 1500 static void a5xx_gpu_state_get_hlsq_regs(struct msm_gpu *gpu, 1501 struct a5xx_gpu_state *a5xx_state) 1502 { 1503 struct a5xx_crashdumper dumper = { 0 }; 1504 u32 offset, count = 0; 1505 u64 *ptr; 1506 int i; 1507 1508 if (a5xx_crashdumper_init(gpu, &dumper)) 1509 return; 1510 1511 /* The script will be written at offset 0 */ 1512 ptr = dumper.ptr; 1513 1514 /* Start writing the data at offset 256k */ 1515 offset = dumper.iova + (256 * SZ_1K); 1516 1517 /* Count how many additional registers to get from the HLSQ aperture */ 1518 for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) 1519 count += a5xx_hlsq_aperture_regs[i].count; 1520 1521 a5xx_state->hlsqregs = kcalloc(count, sizeof(u32), GFP_KERNEL); 1522 if (!a5xx_state->hlsqregs) 1523 return; 1524 1525 /* Build the crashdump script */ 1526 for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) { 1527 u32 type = a5xx_hlsq_aperture_regs[i].type; 1528 u32 c = a5xx_hlsq_aperture_regs[i].count; 1529 1530 /* Write the register to select the desired bank */ 1531 *ptr++ = ((u64) type << 8); 1532 *ptr++ = (((u64) REG_A5XX_HLSQ_DBG_READ_SEL) << 44) | 1533 (1 << 21) | 1; 1534 1535 *ptr++ = offset; 1536 *ptr++ = (((u64) REG_A5XX_HLSQ_DBG_AHB_READ_APERTURE) << 44) 1537 | c; 1538 1539 offset += c * sizeof(u32); 1540 } 1541 1542 /* Write two zeros to close off the script */ 1543 *ptr++ = 0; 1544 *ptr++ = 0; 1545 1546 if (a5xx_crashdumper_run(gpu, &dumper)) { 1547 kfree(a5xx_state->hlsqregs); 1548 msm_gem_kernel_put(dumper.bo, gpu->aspace); 1549 return; 1550 } 1551 1552 /* Copy the data from the crashdumper to the state */ 1553 memcpy(a5xx_state->hlsqregs, dumper.ptr + (256 * SZ_1K), 1554 count * sizeof(u32)); 1555 1556 msm_gem_kernel_put(dumper.bo, gpu->aspace); 1557 } 1558 1559 static struct msm_gpu_state *a5xx_gpu_state_get(struct msm_gpu *gpu) 1560 { 1561 struct a5xx_gpu_state *a5xx_state = kzalloc(sizeof(*a5xx_state), 1562 GFP_KERNEL); 1563 bool stalled = !!(gpu_read(gpu, REG_A5XX_RBBM_STATUS3) & BIT(24)); 1564 1565 if (!a5xx_state) 1566 return ERR_PTR(-ENOMEM); 1567 1568 /* Temporarily disable hardware clock gating before reading the hw */ 1569 a5xx_set_hwcg(gpu, false); 1570 1571 /* First get the generic state from the adreno core */ 1572 adreno_gpu_state_get(gpu, &(a5xx_state->base)); 1573 1574 a5xx_state->base.rbbm_status = gpu_read(gpu, REG_A5XX_RBBM_STATUS); 1575 1576 /* 1577 * Get the HLSQ regs with the help of the crashdumper, but only if 1578 * we are not stalled in an iommu fault (in which case the crashdumper 1579 * would not have access to memory) 1580 */ 1581 if (!stalled) 1582 a5xx_gpu_state_get_hlsq_regs(gpu, a5xx_state); 1583 1584 a5xx_set_hwcg(gpu, true); 1585 1586 return &a5xx_state->base; 1587 } 1588 1589 static void a5xx_gpu_state_destroy(struct kref *kref) 1590 { 1591 struct msm_gpu_state *state = container_of(kref, 1592 struct msm_gpu_state, ref); 1593 struct a5xx_gpu_state *a5xx_state = container_of(state, 1594 struct a5xx_gpu_state, base); 1595 1596 kfree(a5xx_state->hlsqregs); 1597 1598 adreno_gpu_state_destroy(state); 1599 kfree(a5xx_state); 1600 } 1601 1602 static int a5xx_gpu_state_put(struct msm_gpu_state *state) 1603 { 1604 if (IS_ERR_OR_NULL(state)) 1605 return 1; 1606 1607 return kref_put(&state->ref, a5xx_gpu_state_destroy); 1608 } 1609 1610 1611 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP) 1612 static void a5xx_show(struct msm_gpu *gpu, struct msm_gpu_state *state, 1613 struct drm_printer *p) 1614 { 1615 int i, j; 1616 u32 pos = 0; 1617 struct a5xx_gpu_state *a5xx_state = container_of(state, 1618 struct a5xx_gpu_state, base); 1619 1620 if (IS_ERR_OR_NULL(state)) 1621 return; 1622 1623 adreno_show(gpu, state, p); 1624 1625 /* Dump the additional a5xx HLSQ registers */ 1626 if (!a5xx_state->hlsqregs) 1627 return; 1628 1629 drm_printf(p, "registers-hlsq:\n"); 1630 1631 for (i = 0; i < ARRAY_SIZE(a5xx_hlsq_aperture_regs); i++) { 1632 u32 o = a5xx_hlsq_aperture_regs[i].regoffset; 1633 u32 c = a5xx_hlsq_aperture_regs[i].count; 1634 1635 for (j = 0; j < c; j++, pos++, o++) { 1636 /* 1637 * To keep the crashdump simple we pull the entire range 1638 * for each register type but not all of the registers 1639 * in the range are valid. Fortunately invalid registers 1640 * stick out like a sore thumb with a value of 1641 * 0xdeadbeef 1642 */ 1643 if (a5xx_state->hlsqregs[pos] == 0xdeadbeef) 1644 continue; 1645 1646 drm_printf(p, " - { offset: 0x%04x, value: 0x%08x }\n", 1647 o << 2, a5xx_state->hlsqregs[pos]); 1648 } 1649 } 1650 } 1651 #endif 1652 1653 static struct msm_ringbuffer *a5xx_active_ring(struct msm_gpu *gpu) 1654 { 1655 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 1656 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 1657 1658 return a5xx_gpu->cur_ring; 1659 } 1660 1661 static u64 a5xx_gpu_busy(struct msm_gpu *gpu, unsigned long *out_sample_rate) 1662 { 1663 u64 busy_cycles; 1664 1665 busy_cycles = gpu_read64(gpu, REG_A5XX_RBBM_PERFCTR_RBBM_0_LO); 1666 *out_sample_rate = clk_get_rate(gpu->core_clk); 1667 1668 return busy_cycles; 1669 } 1670 1671 static uint32_t a5xx_get_rptr(struct msm_gpu *gpu, struct msm_ringbuffer *ring) 1672 { 1673 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 1674 struct a5xx_gpu *a5xx_gpu = to_a5xx_gpu(adreno_gpu); 1675 1676 if (a5xx_gpu->has_whereami) 1677 return a5xx_gpu->shadow[ring->id]; 1678 1679 return ring->memptrs->rptr = gpu_read(gpu, REG_A5XX_CP_RB_RPTR); 1680 } 1681 1682 static const struct adreno_gpu_funcs funcs = { 1683 .base = { 1684 .get_param = adreno_get_param, 1685 .set_param = adreno_set_param, 1686 .hw_init = a5xx_hw_init, 1687 .ucode_load = a5xx_ucode_load, 1688 .pm_suspend = a5xx_pm_suspend, 1689 .pm_resume = a5xx_pm_resume, 1690 .recover = a5xx_recover, 1691 .submit = a5xx_submit, 1692 .active_ring = a5xx_active_ring, 1693 .irq = a5xx_irq, 1694 .destroy = a5xx_destroy, 1695 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP) 1696 .show = a5xx_show, 1697 #endif 1698 #if defined(CONFIG_DEBUG_FS) 1699 .debugfs_init = a5xx_debugfs_init, 1700 #endif 1701 .gpu_busy = a5xx_gpu_busy, 1702 .gpu_state_get = a5xx_gpu_state_get, 1703 .gpu_state_put = a5xx_gpu_state_put, 1704 .create_address_space = adreno_create_address_space, 1705 .get_rptr = a5xx_get_rptr, 1706 }, 1707 .get_timestamp = a5xx_get_timestamp, 1708 }; 1709 1710 static void check_speed_bin(struct device *dev) 1711 { 1712 struct nvmem_cell *cell; 1713 u32 val; 1714 1715 /* 1716 * If the OPP table specifies a opp-supported-hw property then we have 1717 * to set something with dev_pm_opp_set_supported_hw() or the table 1718 * doesn't get populated so pick an arbitrary value that should 1719 * ensure the default frequencies are selected but not conflict with any 1720 * actual bins 1721 */ 1722 val = 0x80; 1723 1724 cell = nvmem_cell_get(dev, "speed_bin"); 1725 1726 if (!IS_ERR(cell)) { 1727 void *buf = nvmem_cell_read(cell, NULL); 1728 1729 if (!IS_ERR(buf)) { 1730 u8 bin = *((u8 *) buf); 1731 1732 val = (1 << bin); 1733 kfree(buf); 1734 } 1735 1736 nvmem_cell_put(cell); 1737 } 1738 1739 devm_pm_opp_set_supported_hw(dev, &val, 1); 1740 } 1741 1742 struct msm_gpu *a5xx_gpu_init(struct drm_device *dev) 1743 { 1744 struct msm_drm_private *priv = dev->dev_private; 1745 struct platform_device *pdev = priv->gpu_pdev; 1746 struct a5xx_gpu *a5xx_gpu = NULL; 1747 struct adreno_gpu *adreno_gpu; 1748 struct msm_gpu *gpu; 1749 unsigned int nr_rings; 1750 int ret; 1751 1752 if (!pdev) { 1753 DRM_DEV_ERROR(dev->dev, "No A5XX device is defined\n"); 1754 return ERR_PTR(-ENXIO); 1755 } 1756 1757 a5xx_gpu = kzalloc(sizeof(*a5xx_gpu), GFP_KERNEL); 1758 if (!a5xx_gpu) 1759 return ERR_PTR(-ENOMEM); 1760 1761 adreno_gpu = &a5xx_gpu->base; 1762 gpu = &adreno_gpu->base; 1763 1764 adreno_gpu->registers = a5xx_registers; 1765 1766 a5xx_gpu->lm_leakage = 0x4E001A; 1767 1768 check_speed_bin(&pdev->dev); 1769 1770 nr_rings = 4; 1771 1772 if (adreno_is_a510(adreno_gpu)) 1773 nr_rings = 1; 1774 1775 ret = adreno_gpu_init(dev, pdev, adreno_gpu, &funcs, nr_rings); 1776 if (ret) { 1777 a5xx_destroy(&(a5xx_gpu->base.base)); 1778 return ERR_PTR(ret); 1779 } 1780 1781 if (gpu->aspace) 1782 msm_mmu_set_fault_handler(gpu->aspace->mmu, gpu, a5xx_fault_handler); 1783 1784 /* Set up the preemption specific bits and pieces for each ringbuffer */ 1785 a5xx_preempt_init(gpu); 1786 1787 return gpu; 1788 } 1789