drivers/gpu/drm/etnaviv/etnaviv_buffer.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2014-2018 Etnaviv Project
   4  */
   5
   6 #include <drm/drm_drv.h>
   7
   8 #include "etnaviv_cmdbuf.h"
   9 #include "etnaviv_gpu.h"
  10 #include "etnaviv_gem.h"
  11 #include "etnaviv_mmu.h"
  12
  13 #include "common.xml.h"
  14 #include "state.xml.h"
  15 #include "state_hi.xml.h"
  16 #include "state_3d.xml.h"
  17 #include "cmdstream.xml.h"
  18
  19 /*
  20  * Command Buffer helper:
  21  */
  22
  23
  24 static inline void OUT(struct etnaviv_cmdbuf *buffer, u32 data)
  25 {
  26         u32 *vaddr = (u32 *)buffer->vaddr;
  27
  28         BUG_ON(buffer->user_size >= buffer->size);
  29
  30         vaddr[buffer->user_size / 4] = data;
  31         buffer->user_size += 4;
  32 }
  33
  34 static inline void CMD_LOAD_STATE(struct etnaviv_cmdbuf *buffer,
  35         u32 reg, u32 value)
  36 {
  37         u32 index = reg >> VIV_FE_LOAD_STATE_HEADER_OFFSET__SHR;
  38
  39         buffer->user_size = ALIGN(buffer->user_size, 8);
  40
  41         /* write a register via cmd stream */
  42         OUT(buffer, VIV_FE_LOAD_STATE_HEADER_OP_LOAD_STATE |
  43                     VIV_FE_LOAD_STATE_HEADER_COUNT(1) |
  44                     VIV_FE_LOAD_STATE_HEADER_OFFSET(index));
  45         OUT(buffer, value);
  46 }
  47
  48 static inline void CMD_END(struct etnaviv_cmdbuf *buffer)
  49 {
  50         buffer->user_size = ALIGN(buffer->user_size, 8);
  51
  52         OUT(buffer, VIV_FE_END_HEADER_OP_END);
  53 }
  54
  55 static inline void CMD_WAIT(struct etnaviv_cmdbuf *buffer)
  56 {
  57         buffer->user_size = ALIGN(buffer->user_size, 8);
  58
  59         OUT(buffer, VIV_FE_WAIT_HEADER_OP_WAIT | 200);
  60 }
  61
  62 static inline void CMD_LINK(struct etnaviv_cmdbuf *buffer,
  63         u16 prefetch, u32 address)
  64 {
  65         buffer->user_size = ALIGN(buffer->user_size, 8);
  66
  67         OUT(buffer, VIV_FE_LINK_HEADER_OP_LINK |
  68                     VIV_FE_LINK_HEADER_PREFETCH(prefetch));
  69         OUT(buffer, address);
  70 }
  71
  72 static inline void CMD_STALL(struct etnaviv_cmdbuf *buffer,
  73         u32 from, u32 to)
  74 {
  75         buffer->user_size = ALIGN(buffer->user_size, 8);
  76
  77         OUT(buffer, VIV_FE_STALL_HEADER_OP_STALL);
  78         OUT(buffer, VIV_FE_STALL_TOKEN_FROM(from) | VIV_FE_STALL_TOKEN_TO(to));
  79 }
  80
  81 static inline void CMD_SEM(struct etnaviv_cmdbuf *buffer, u32 from, u32 to)
  82 {
  83         CMD_LOAD_STATE(buffer, VIVS_GL_SEMAPHORE_TOKEN,
  84                        VIVS_GL_SEMAPHORE_TOKEN_FROM(from) |
  85                        VIVS_GL_SEMAPHORE_TOKEN_TO(to));
  86 }
  87
  88 static void etnaviv_cmd_select_pipe(struct etnaviv_gpu *gpu,
  89         struct etnaviv_cmdbuf *buffer, u8 pipe)
  90 {
  91         u32 flush = 0;
  92
  93         lockdep_assert_held(&gpu->lock);
  94
  95         /*
  96          * This assumes that if we're switching to 2D, we're switching
  97          * away from 3D, and vice versa.  Hence, if we're switching to
  98          * the 2D core, we need to flush the 3D depth and color caches,
  99          * otherwise we need to flush the 2D pixel engine cache.
 100          */
 101         if (gpu->exec_state == ETNA_PIPE_2D)
 102                 flush = VIVS_GL_FLUSH_CACHE_PE2D;
 103         else if (gpu->exec_state == ETNA_PIPE_3D)
 104                 flush = VIVS_GL_FLUSH_CACHE_DEPTH | VIVS_GL_FLUSH_CACHE_COLOR;
 105
 106         CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
 107         CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 108         CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 109
 110         CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
 111                        VIVS_GL_PIPE_SELECT_PIPE(pipe));
 112 }
 113
 114 static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu,
 115         struct etnaviv_cmdbuf *buf, u32 off, u32 len)
 116 {
 117         u32 size = buf->size;
 118         u32 *ptr = buf->vaddr + off;
 119
 120         dev_info(gpu->dev, "virt %p phys 0x%08x free 0x%08x\n",
 121                         ptr, etnaviv_cmdbuf_get_va(buf,
 122                         &gpu->mmu_context->cmdbuf_mapping) +
 123                         off, size - len * 4 - off);
 124
 125         print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
 126                         ptr, len * 4, 0);
 127 }
 128
 129 /*
 130  * Safely replace the WAIT of a waitlink with a new command and argument.
 131  * The GPU may be executing this WAIT while we're modifying it, so we have
 132  * to write it in a specific order to avoid the GPU branching to somewhere
 133  * else.  'wl_offset' is the offset to the first byte of the WAIT command.
 134  */
 135 static void etnaviv_buffer_replace_wait(struct etnaviv_cmdbuf *buffer,
 136         unsigned int wl_offset, u32 cmd, u32 arg)
 137 {
 138         u32 *lw = buffer->vaddr + wl_offset;
 139
 140         lw[1] = arg;
 141         mb();
 142         lw[0] = cmd;
 143         mb();
 144 }
 145
 146 /*
 147  * Ensure that there is space in the command buffer to contiguously write
 148  * 'cmd_dwords' 64-bit words into the buffer, wrapping if necessary.
 149  */
 150 static u32 etnaviv_buffer_reserve(struct etnaviv_gpu *gpu,
 151         struct etnaviv_cmdbuf *buffer, unsigned int cmd_dwords)
 152 {
 153         if (buffer->user_size + cmd_dwords * sizeof(u64) > buffer->size)
 154                 buffer->user_size = 0;
 155
 156         return etnaviv_cmdbuf_get_va(buffer,
 157                                      &gpu->mmu_context->cmdbuf_mapping) +
 158                buffer->user_size;
 159 }
 160
 161 u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
 162 {
 163         struct etnaviv_cmdbuf *buffer = &gpu->buffer;
 164
 165         lockdep_assert_held(&gpu->lock);
 166
 167         /* initialize buffer */
 168         buffer->user_size = 0;
 169
 170         CMD_WAIT(buffer);
 171         CMD_LINK(buffer, 2,
 172                  etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
 173                  + buffer->user_size - 4);
 174
 175         return buffer->user_size / 8;
 176 }
 177
 178 u16 etnaviv_buffer_config_mmuv2(struct etnaviv_gpu *gpu, u32 mtlb_addr, u32 safe_addr)
 179 {
 180         struct etnaviv_cmdbuf *buffer = &gpu->buffer;
 181
 182         lockdep_assert_held(&gpu->lock);
 183
 184         buffer->user_size = 0;
 185
 186         if (gpu->identity.features & chipFeatures_PIPE_3D) {
 187                 CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
 188                                VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_3D));
 189                 CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
 190                         mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
 191                 CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
 192                 CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 193                 CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 194         }
 195
 196         if (gpu->identity.features & chipFeatures_PIPE_2D) {
 197                 CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
 198                                VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_2D));
 199                 CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
 200                         mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
 201                 CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
 202                 CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 203                 CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 204         }
 205
 206         CMD_END(buffer);
 207
 208         buffer->user_size = ALIGN(buffer->user_size, 8);
 209
 210         return buffer->user_size / 8;
 211 }
 212
 213 u16 etnaviv_buffer_config_pta(struct etnaviv_gpu *gpu, unsigned short id)
 214 {
 215         struct etnaviv_cmdbuf *buffer = &gpu->buffer;
 216
 217         lockdep_assert_held(&gpu->lock);
 218
 219         buffer->user_size = 0;
 220
 221         CMD_LOAD_STATE(buffer, VIVS_MMUv2_PTA_CONFIG,
 222                        VIVS_MMUv2_PTA_CONFIG_INDEX(id));
 223
 224         CMD_END(buffer);
 225
 226         buffer->user_size = ALIGN(buffer->user_size, 8);
 227
 228         return buffer->user_size / 8;
 229 }
 230
 231 void etnaviv_buffer_end(struct etnaviv_gpu *gpu)
 232 {
 233         struct etnaviv_cmdbuf *buffer = &gpu->buffer;
 234         unsigned int waitlink_offset = buffer->user_size - 16;
 235         u32 link_target, flush = 0;
 236
 237         lockdep_assert_held(&gpu->lock);
 238
 239         if (gpu->exec_state == ETNA_PIPE_2D)
 240                 flush = VIVS_GL_FLUSH_CACHE_PE2D;
 241         else if (gpu->exec_state == ETNA_PIPE_3D)
 242                 flush = VIVS_GL_FLUSH_CACHE_DEPTH |
 243                         VIVS_GL_FLUSH_CACHE_COLOR |
 244                         VIVS_GL_FLUSH_CACHE_TEXTURE |
 245                         VIVS_GL_FLUSH_CACHE_TEXTUREVS |
 246                         VIVS_GL_FLUSH_CACHE_SHADER_L2;
 247
 248         if (flush) {
 249                 unsigned int dwords = 7;
 250
 251                 link_target = etnaviv_buffer_reserve(gpu, buffer, dwords);
 252
 253                 CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 254                 CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 255                 CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
 256                 if (gpu->exec_state == ETNA_PIPE_3D)
 257                         CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
 258                                        VIVS_TS_FLUSH_CACHE_FLUSH);
 259                 CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 260                 CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 261                 CMD_END(buffer);
 262
 263                 etnaviv_buffer_replace_wait(buffer, waitlink_offset,
 264                                             VIV_FE_LINK_HEADER_OP_LINK |
 265                                             VIV_FE_LINK_HEADER_PREFETCH(dwords),
 266                                             link_target);
 267         } else {
 268                 /* Replace the last link-wait with an "END" command */
 269                 etnaviv_buffer_replace_wait(buffer, waitlink_offset,
 270                                             VIV_FE_END_HEADER_OP_END, 0);
 271         }
 272 }
 273
 274 /* Append a 'sync point' to the ring buffer. */
 275 void etnaviv_sync_point_queue(struct etnaviv_gpu *gpu, unsigned int event)
 276 {
 277         struct etnaviv_cmdbuf *buffer = &gpu->buffer;
 278         unsigned int waitlink_offset = buffer->user_size - 16;
 279         u32 dwords, target;
 280
 281         lockdep_assert_held(&gpu->lock);
 282
 283         /*
 284          * We need at most 3 dwords in the return target:
 285          * 1 event + 1 end + 1 wait + 1 link.
 286          */
 287         dwords = 4;
 288         target = etnaviv_buffer_reserve(gpu, buffer, dwords);
 289
 290         /* Signal sync point event */
 291         CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
 292                        VIVS_GL_EVENT_FROM_PE);
 293
 294         /* Stop the FE to 'pause' the GPU */
 295         CMD_END(buffer);
 296
 297         /* Append waitlink */
 298         CMD_WAIT(buffer);
 299         CMD_LINK(buffer, 2,
 300                  etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
 301                  + buffer->user_size - 4);
 302
 303         /*
 304          * Kick off the 'sync point' command by replacing the previous
 305          * WAIT with a link to the address in the ring buffer.
 306          */
 307         etnaviv_buffer_replace_wait(buffer, waitlink_offset,
 308                                     VIV_FE_LINK_HEADER_OP_LINK |
 309                                     VIV_FE_LINK_HEADER_PREFETCH(dwords),
 310                                     target);
 311 }
 312
 313 /* Append a command buffer to the ring buffer. */
 314 void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, u32 exec_state,
 315         struct etnaviv_iommu_context *mmu_context, unsigned int event,
 316         struct etnaviv_cmdbuf *cmdbuf)
 317 {
 318         struct etnaviv_cmdbuf *buffer = &gpu->buffer;
 319         unsigned int waitlink_offset = buffer->user_size - 16;
 320         u32 return_target, return_dwords;
 321         u32 link_target, link_dwords;
 322         bool switch_context = gpu->exec_state != exec_state;
 323         bool switch_mmu_context = gpu->mmu_context != mmu_context;
 324         unsigned int new_flush_seq = READ_ONCE(gpu->mmu_context->flush_seq);
 325         bool need_flush = switch_mmu_context || gpu->flush_seq != new_flush_seq;
 326
 327         lockdep_assert_held(&gpu->lock);
 328
 329         if (drm_debug & DRM_UT_DRIVER)
 330                 etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
 331
 332         link_target = etnaviv_cmdbuf_get_va(cmdbuf,
 333                                             &gpu->mmu_context->cmdbuf_mapping);
 334         link_dwords = cmdbuf->size / 8;
 335
 336         /*
 337          * If we need maintenance prior to submitting this buffer, we will
 338          * need to append a mmu flush load state, followed by a new
 339          * link to this buffer - a total of four additional words.
 340          */
 341         if (need_flush || switch_context) {
 342                 u32 target, extra_dwords;
 343
 344                 /* link command */
 345                 extra_dwords = 1;
 346
 347                 /* flush command */
 348                 if (need_flush) {
 349                         if (gpu->mmu_context->global->version == ETNAVIV_IOMMU_V1)
 350                                 extra_dwords += 1;
 351                         else
 352                                 extra_dwords += 3;
 353                 }
 354
 355                 /* pipe switch commands */
 356                 if (switch_context)
 357                         extra_dwords += 4;
 358
 359                 /* PTA load command */
 360                 if (switch_mmu_context && gpu->sec_mode == ETNA_SEC_KERNEL)
 361                         extra_dwords += 1;
 362
 363                 target = etnaviv_buffer_reserve(gpu, buffer, extra_dwords);
 364                 /*
 365                  * Switch MMU context if necessary. Must be done after the
 366                  * link target has been calculated, as the jump forward in the
 367                  * kernel ring still uses the last active MMU context before
 368                  * the switch.
 369                  */
 370                 if (switch_mmu_context) {
 371                         struct etnaviv_iommu_context *old_context = gpu->mmu_context;
 372
 373                         etnaviv_iommu_context_get(mmu_context);
 374                         gpu->mmu_context = mmu_context;
 375                         etnaviv_iommu_context_put(old_context);
 376                 }
 377
 378                 if (need_flush) {
 379                         /* Add the MMU flush */
 380                         if (gpu->mmu_context->global->version == ETNAVIV_IOMMU_V1) {
 381                                 CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_MMU,
 382                                                VIVS_GL_FLUSH_MMU_FLUSH_FEMMU |
 383                                                VIVS_GL_FLUSH_MMU_FLUSH_UNK1 |
 384                                                VIVS_GL_FLUSH_MMU_FLUSH_UNK2 |
 385                                                VIVS_GL_FLUSH_MMU_FLUSH_PEMMU |
 386                                                VIVS_GL_FLUSH_MMU_FLUSH_UNK4);
 387                         } else {
 388                                 u32 flush = VIVS_MMUv2_CONFIGURATION_MODE_MASK |
 389                                             VIVS_MMUv2_CONFIGURATION_FLUSH_FLUSH;
 390
 391                                 if (switch_mmu_context &&
 392                                     gpu->sec_mode == ETNA_SEC_KERNEL) {
 393                                         unsigned short id =
 394                                                 etnaviv_iommuv2_get_pta_id(gpu->mmu_context);
 395                                         CMD_LOAD_STATE(buffer,
 396                                                 VIVS_MMUv2_PTA_CONFIG,
 397                                                 VIVS_MMUv2_PTA_CONFIG_INDEX(id));
 398                                 }
 399
 400                                 if (gpu->sec_mode == ETNA_SEC_NONE)
 401                                         flush |= etnaviv_iommuv2_get_mtlb_addr(gpu->mmu_context);
 402
 403                                 CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
 404                                                flush);
 405                                 CMD_SEM(buffer, SYNC_RECIPIENT_FE,
 406                                         SYNC_RECIPIENT_PE);
 407                                 CMD_STALL(buffer, SYNC_RECIPIENT_FE,
 408                                         SYNC_RECIPIENT_PE);
 409                         }
 410
 411                         gpu->flush_seq = new_flush_seq;
 412                 }
 413
 414                 if (switch_context) {
 415                         etnaviv_cmd_select_pipe(gpu, buffer, exec_state);
 416                         gpu->exec_state = exec_state;
 417                 }
 418
 419                 /* And the link to the submitted buffer */
 420                 link_target = etnaviv_cmdbuf_get_va(cmdbuf,
 421                                         &gpu->mmu_context->cmdbuf_mapping);
 422                 CMD_LINK(buffer, link_dwords, link_target);
 423
 424                 /* Update the link target to point to above instructions */
 425                 link_target = target;
 426                 link_dwords = extra_dwords;
 427         }
 428
 429         /*
 430          * Append a LINK to the submitted command buffer to return to
 431          * the ring buffer.  return_target is the ring target address.
 432          * We need at most 7 dwords in the return target: 2 cache flush +
 433          * 2 semaphore stall + 1 event + 1 wait + 1 link.
 434          */
 435         return_dwords = 7;
 436         return_target = etnaviv_buffer_reserve(gpu, buffer, return_dwords);
 437         CMD_LINK(cmdbuf, return_dwords, return_target);
 438
 439         /*
 440          * Append a cache flush, stall, event, wait and link pointing back to
 441          * the wait command to the ring buffer.
 442          */
 443         if (gpu->exec_state == ETNA_PIPE_2D) {
 444                 CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
 445                                        VIVS_GL_FLUSH_CACHE_PE2D);
 446         } else {
 447                 CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
 448                                        VIVS_GL_FLUSH_CACHE_DEPTH |
 449                                        VIVS_GL_FLUSH_CACHE_COLOR);
 450                 CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
 451                                        VIVS_TS_FLUSH_CACHE_FLUSH);
 452         }
 453         CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 454         CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
 455         CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
 456                        VIVS_GL_EVENT_FROM_PE);
 457         CMD_WAIT(buffer);
 458         CMD_LINK(buffer, 2,
 459                  etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
 460                  + buffer->user_size - 4);
 461
 462         if (drm_debug & DRM_UT_DRIVER)
 463                 pr_info("stream link to 0x%08x @ 0x%08x %p\n",
 464                         return_target,
 465                         etnaviv_cmdbuf_get_va(cmdbuf, &gpu->mmu_context->cmdbuf_mapping),
 466                         cmdbuf->vaddr);
 467
 468         if (drm_debug & DRM_UT_DRIVER) {
 469                 print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
 470                                cmdbuf->vaddr, cmdbuf->size, 0);
 471
 472                 pr_info("link op: %p\n", buffer->vaddr + waitlink_offset);
 473                 pr_info("addr: 0x%08x\n", link_target);
 474                 pr_info("back: 0x%08x\n", return_target);
 475                 pr_info("event: %d\n", event);
 476         }
 477
 478         /*
 479          * Kick off the submitted command by replacing the previous
 480          * WAIT with a link to the address in the ring buffer.
 481          */
 482         etnaviv_buffer_replace_wait(buffer, waitlink_offset,
 483                                     VIV_FE_LINK_HEADER_OP_LINK |
 484                                     VIV_FE_LINK_HEADER_PREFETCH(link_dwords),
 485                                     link_target);
 486
 487         if (drm_debug & DRM_UT_DRIVER)
 488                 etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
 489 }