1 /* SPDX-License-Identifier: GPL-2.0-only */ 2 /* 3 * Copyright (C) 2013 Red Hat 4 * Author: Rob Clark <robdclark@gmail.com> 5 * 6 * Copyright (c) 2014,2017, 2019 The Linux Foundation. All rights reserved. 7 */ 8 9 #ifndef __ADRENO_GPU_H__ 10 #define __ADRENO_GPU_H__ 11 12 #include <linux/firmware.h> 13 #include <linux/iopoll.h> 14 15 #include "msm_gpu.h" 16 17 #include "adreno_common.xml.h" 18 #include "adreno_pm4.xml.h" 19 20 #define REG_ADRENO_DEFINE(_offset, _reg) [_offset] = (_reg) + 1 21 #define REG_SKIP ~0 22 #define REG_ADRENO_SKIP(_offset) [_offset] = REG_SKIP 23 24 /** 25 * adreno_regs: List of registers that are used in across all 26 * 3D devices. Each device type has different offset value for the same 27 * register, so an array of register offsets are declared for every device 28 * and are indexed by the enumeration values defined in this enum 29 */ 30 enum adreno_regs { 31 REG_ADRENO_CP_RB_BASE, 32 REG_ADRENO_CP_RB_BASE_HI, 33 REG_ADRENO_CP_RB_RPTR_ADDR, 34 REG_ADRENO_CP_RB_RPTR_ADDR_HI, 35 REG_ADRENO_CP_RB_RPTR, 36 REG_ADRENO_CP_RB_WPTR, 37 REG_ADRENO_CP_RB_CNTL, 38 REG_ADRENO_REGISTER_MAX, 39 }; 40 41 enum { 42 ADRENO_FW_PM4 = 0, 43 ADRENO_FW_SQE = 0, /* a6xx */ 44 ADRENO_FW_PFP = 1, 45 ADRENO_FW_GMU = 1, /* a6xx */ 46 ADRENO_FW_GPMU = 2, 47 ADRENO_FW_MAX, 48 }; 49 50 enum adreno_quirks { 51 ADRENO_QUIRK_TWO_PASS_USE_WFI = 1, 52 ADRENO_QUIRK_FAULT_DETECT_MASK = 2, 53 ADRENO_QUIRK_LMLOADKILL_DISABLE = 3, 54 }; 55 56 struct adreno_rev { 57 uint8_t core; 58 uint8_t major; 59 uint8_t minor; 60 uint8_t patchid; 61 }; 62 63 #define ADRENO_REV(core, major, minor, patchid) \ 64 ((struct adreno_rev){ core, major, minor, patchid }) 65 66 struct adreno_gpu_funcs { 67 struct msm_gpu_funcs base; 68 int (*get_timestamp)(struct msm_gpu *gpu, uint64_t *value); 69 }; 70 71 struct adreno_info { 72 struct adreno_rev rev; 73 uint32_t revn; 74 const char *name; 75 const char *fw[ADRENO_FW_MAX]; 76 uint32_t gmem; 77 enum adreno_quirks quirks; 78 struct msm_gpu *(*init)(struct drm_device *dev); 79 const char *zapfw; 80 u32 inactive_period; 81 }; 82 83 const struct adreno_info *adreno_info(struct adreno_rev rev); 84 85 struct adreno_gpu { 86 struct msm_gpu base; 87 struct adreno_rev rev; 88 const struct adreno_info *info; 89 uint32_t gmem; /* actual gmem size */ 90 uint32_t revn; /* numeric revision name */ 91 const struct adreno_gpu_funcs *funcs; 92 93 /* interesting register offsets to dump: */ 94 const unsigned int *registers; 95 96 /* 97 * Are we loading fw from legacy path? Prior to addition 98 * of gpu firmware to linux-firmware, the fw files were 99 * placed in toplevel firmware directory, following qcom's 100 * android kernel. But linux-firmware preferred they be 101 * placed in a 'qcom' subdirectory. 102 * 103 * For backwards compatibility, we try first to load from 104 * the new path, using request_firmware_direct() to avoid 105 * any potential timeout waiting for usermode helper, then 106 * fall back to the old path (with direct load). And 107 * finally fall back to request_firmware() with the new 108 * path to allow the usermode helper. 109 */ 110 enum { 111 FW_LOCATION_UNKNOWN = 0, 112 FW_LOCATION_NEW, /* /lib/firmware/qcom/$fwfile */ 113 FW_LOCATION_LEGACY, /* /lib/firmware/$fwfile */ 114 FW_LOCATION_HELPER, 115 } fwloc; 116 117 /* firmware: */ 118 const struct firmware *fw[ADRENO_FW_MAX]; 119 120 /* 121 * Register offsets are different between some GPUs. 122 * GPU specific offsets will be exported by GPU specific 123 * code (a3xx_gpu.c) and stored in this common location. 124 */ 125 const unsigned int *reg_offsets; 126 }; 127 #define to_adreno_gpu(x) container_of(x, struct adreno_gpu, base) 128 129 struct adreno_ocmem { 130 struct ocmem *ocmem; 131 unsigned long base; 132 void *hdl; 133 }; 134 135 /* platform config data (ie. from DT, or pdata) */ 136 struct adreno_platform_config { 137 struct adreno_rev rev; 138 }; 139 140 #define ADRENO_IDLE_TIMEOUT msecs_to_jiffies(1000) 141 142 #define spin_until(X) ({ \ 143 int __ret = -ETIMEDOUT; \ 144 unsigned long __t = jiffies + ADRENO_IDLE_TIMEOUT; \ 145 do { \ 146 if (X) { \ 147 __ret = 0; \ 148 break; \ 149 } \ 150 } while (time_before(jiffies, __t)); \ 151 __ret; \ 152 }) 153 154 static inline bool adreno_is_a2xx(struct adreno_gpu *gpu) 155 { 156 return (gpu->revn < 300); 157 } 158 159 static inline bool adreno_is_a20x(struct adreno_gpu *gpu) 160 { 161 return (gpu->revn < 210); 162 } 163 164 static inline bool adreno_is_a225(struct adreno_gpu *gpu) 165 { 166 return gpu->revn == 225; 167 } 168 169 static inline bool adreno_is_a3xx(struct adreno_gpu *gpu) 170 { 171 return (gpu->revn >= 300) && (gpu->revn < 400); 172 } 173 174 static inline bool adreno_is_a305(struct adreno_gpu *gpu) 175 { 176 return gpu->revn == 305; 177 } 178 179 static inline bool adreno_is_a306(struct adreno_gpu *gpu) 180 { 181 /* yes, 307, because a305c is 306 */ 182 return gpu->revn == 307; 183 } 184 185 static inline bool adreno_is_a320(struct adreno_gpu *gpu) 186 { 187 return gpu->revn == 320; 188 } 189 190 static inline bool adreno_is_a330(struct adreno_gpu *gpu) 191 { 192 return gpu->revn == 330; 193 } 194 195 static inline bool adreno_is_a330v2(struct adreno_gpu *gpu) 196 { 197 return adreno_is_a330(gpu) && (gpu->rev.patchid > 0); 198 } 199 200 static inline bool adreno_is_a4xx(struct adreno_gpu *gpu) 201 { 202 return (gpu->revn >= 400) && (gpu->revn < 500); 203 } 204 205 static inline int adreno_is_a420(struct adreno_gpu *gpu) 206 { 207 return gpu->revn == 420; 208 } 209 210 static inline int adreno_is_a430(struct adreno_gpu *gpu) 211 { 212 return gpu->revn == 430; 213 } 214 215 static inline int adreno_is_a510(struct adreno_gpu *gpu) 216 { 217 return gpu->revn == 510; 218 } 219 220 static inline int adreno_is_a530(struct adreno_gpu *gpu) 221 { 222 return gpu->revn == 530; 223 } 224 225 static inline int adreno_is_a540(struct adreno_gpu *gpu) 226 { 227 return gpu->revn == 540; 228 } 229 230 static inline int adreno_is_a618(struct adreno_gpu *gpu) 231 { 232 return gpu->revn == 618; 233 } 234 235 static inline int adreno_is_a630(struct adreno_gpu *gpu) 236 { 237 return gpu->revn == 630; 238 } 239 240 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value); 241 const struct firmware *adreno_request_fw(struct adreno_gpu *adreno_gpu, 242 const char *fwname); 243 struct drm_gem_object *adreno_fw_create_bo(struct msm_gpu *gpu, 244 const struct firmware *fw, u64 *iova); 245 int adreno_hw_init(struct msm_gpu *gpu); 246 void adreno_recover(struct msm_gpu *gpu); 247 void adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit, 248 struct msm_file_private *ctx); 249 void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring); 250 bool adreno_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring); 251 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP) 252 void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state, 253 struct drm_printer *p); 254 #endif 255 void adreno_dump_info(struct msm_gpu *gpu); 256 void adreno_dump(struct msm_gpu *gpu); 257 void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords); 258 struct msm_ringbuffer *adreno_active_ring(struct msm_gpu *gpu); 259 260 int adreno_gpu_ocmem_init(struct device *dev, struct adreno_gpu *adreno_gpu, 261 struct adreno_ocmem *ocmem); 262 void adreno_gpu_ocmem_cleanup(struct adreno_ocmem *ocmem); 263 264 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev, 265 struct adreno_gpu *gpu, const struct adreno_gpu_funcs *funcs, 266 int nr_rings); 267 void adreno_gpu_cleanup(struct adreno_gpu *gpu); 268 int adreno_load_fw(struct adreno_gpu *adreno_gpu); 269 270 void adreno_gpu_state_destroy(struct msm_gpu_state *state); 271 272 int adreno_gpu_state_get(struct msm_gpu *gpu, struct msm_gpu_state *state); 273 int adreno_gpu_state_put(struct msm_gpu_state *state); 274 275 /* 276 * For a5xx and a6xx targets load the zap shader that is used to pull the GPU 277 * out of secure mode 278 */ 279 int adreno_zap_shader_load(struct msm_gpu *gpu, u32 pasid); 280 281 /* ringbuffer helpers (the parts that are adreno specific) */ 282 283 static inline void 284 OUT_PKT0(struct msm_ringbuffer *ring, uint16_t regindx, uint16_t cnt) 285 { 286 adreno_wait_ring(ring, cnt+1); 287 OUT_RING(ring, CP_TYPE0_PKT | ((cnt-1) << 16) | (regindx & 0x7FFF)); 288 } 289 290 /* no-op packet: */ 291 static inline void 292 OUT_PKT2(struct msm_ringbuffer *ring) 293 { 294 adreno_wait_ring(ring, 1); 295 OUT_RING(ring, CP_TYPE2_PKT); 296 } 297 298 static inline void 299 OUT_PKT3(struct msm_ringbuffer *ring, uint8_t opcode, uint16_t cnt) 300 { 301 adreno_wait_ring(ring, cnt+1); 302 OUT_RING(ring, CP_TYPE3_PKT | ((cnt-1) << 16) | ((opcode & 0xFF) << 8)); 303 } 304 305 static inline u32 PM4_PARITY(u32 val) 306 { 307 return (0x9669 >> (0xF & (val ^ 308 (val >> 4) ^ (val >> 8) ^ (val >> 12) ^ 309 (val >> 16) ^ ((val) >> 20) ^ (val >> 24) ^ 310 (val >> 28)))) & 1; 311 } 312 313 /* Maximum number of values that can be executed for one opcode */ 314 #define TYPE4_MAX_PAYLOAD 127 315 316 #define PKT4(_reg, _cnt) \ 317 (CP_TYPE4_PKT | ((_cnt) << 0) | (PM4_PARITY((_cnt)) << 7) | \ 318 (((_reg) & 0x3FFFF) << 8) | (PM4_PARITY((_reg)) << 27)) 319 320 static inline void 321 OUT_PKT4(struct msm_ringbuffer *ring, uint16_t regindx, uint16_t cnt) 322 { 323 adreno_wait_ring(ring, cnt + 1); 324 OUT_RING(ring, PKT4(regindx, cnt)); 325 } 326 327 static inline void 328 OUT_PKT7(struct msm_ringbuffer *ring, uint8_t opcode, uint16_t cnt) 329 { 330 adreno_wait_ring(ring, cnt + 1); 331 OUT_RING(ring, CP_TYPE7_PKT | (cnt << 0) | (PM4_PARITY(cnt) << 15) | 332 ((opcode & 0x7F) << 16) | (PM4_PARITY(opcode) << 23)); 333 } 334 335 /* 336 * adreno_reg_check() - Checks the validity of a register enum 337 * @gpu: Pointer to struct adreno_gpu 338 * @offset_name: The register enum that is checked 339 */ 340 static inline bool adreno_reg_check(struct adreno_gpu *gpu, 341 enum adreno_regs offset_name) 342 { 343 BUG_ON(offset_name >= REG_ADRENO_REGISTER_MAX || !gpu->reg_offsets[offset_name]); 344 345 /* 346 * REG_SKIP is a special value that tell us that the register in 347 * question isn't implemented on target but don't trigger a BUG(). This 348 * is used to cleanly implement adreno_gpu_write64() and 349 * adreno_gpu_read64() in a generic fashion 350 */ 351 if (gpu->reg_offsets[offset_name] == REG_SKIP) 352 return false; 353 354 return true; 355 } 356 357 static inline u32 adreno_gpu_read(struct adreno_gpu *gpu, 358 enum adreno_regs offset_name) 359 { 360 u32 reg = gpu->reg_offsets[offset_name]; 361 u32 val = 0; 362 if(adreno_reg_check(gpu,offset_name)) 363 val = gpu_read(&gpu->base, reg - 1); 364 return val; 365 } 366 367 static inline void adreno_gpu_write(struct adreno_gpu *gpu, 368 enum adreno_regs offset_name, u32 data) 369 { 370 u32 reg = gpu->reg_offsets[offset_name]; 371 if(adreno_reg_check(gpu, offset_name)) 372 gpu_write(&gpu->base, reg - 1, data); 373 } 374 375 struct msm_gpu *a2xx_gpu_init(struct drm_device *dev); 376 struct msm_gpu *a3xx_gpu_init(struct drm_device *dev); 377 struct msm_gpu *a4xx_gpu_init(struct drm_device *dev); 378 struct msm_gpu *a5xx_gpu_init(struct drm_device *dev); 379 struct msm_gpu *a6xx_gpu_init(struct drm_device *dev); 380 381 static inline void adreno_gpu_write64(struct adreno_gpu *gpu, 382 enum adreno_regs lo, enum adreno_regs hi, u64 data) 383 { 384 adreno_gpu_write(gpu, lo, lower_32_bits(data)); 385 adreno_gpu_write(gpu, hi, upper_32_bits(data)); 386 } 387 388 static inline uint32_t get_wptr(struct msm_ringbuffer *ring) 389 { 390 return (ring->cur - ring->start) % (MSM_GPU_RINGBUFFER_SZ >> 2); 391 } 392 393 /* 394 * Given a register and a count, return a value to program into 395 * REG_CP_PROTECT_REG(n) - this will block both reads and writes for _len 396 * registers starting at _reg. 397 * 398 * The register base needs to be a multiple of the length. If it is not, the 399 * hardware will quietly mask off the bits for you and shift the size. For 400 * example, if you intend the protection to start at 0x07 for a length of 4 401 * (0x07-0x0A) the hardware will actually protect (0x04-0x07) which might 402 * expose registers you intended to protect! 403 */ 404 #define ADRENO_PROTECT_RW(_reg, _len) \ 405 ((1 << 30) | (1 << 29) | \ 406 ((ilog2((_len)) & 0x1F) << 24) | (((_reg) << 2) & 0xFFFFF)) 407 408 /* 409 * Same as above, but allow reads over the range. For areas of mixed use (such 410 * as performance counters) this allows us to protect a much larger range with a 411 * single register 412 */ 413 #define ADRENO_PROTECT_RDONLY(_reg, _len) \ 414 ((1 << 29) \ 415 ((ilog2((_len)) & 0x1F) << 24) | (((_reg) << 2) & 0xFFFFF)) 416 417 418 #define gpu_poll_timeout(gpu, addr, val, cond, interval, timeout) \ 419 readl_poll_timeout((gpu)->mmio + ((addr) << 2), val, cond, \ 420 interval, timeout) 421 422 #endif /* __ADRENO_GPU_H__ */ 423