drivers/gator/gator_events_mali_t6xx_hw.c

   1 /**
   2  * Copyright (C) ARM Limited 2012-2013. All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License version 2 as
   6  * published by the Free Software Foundation.
   7  *
   8  */
   9
  10 #include "gator.h"
  11
  12 #include <linux/module.h>
  13 #include <linux/time.h>
  14 #include <linux/math64.h>
  15 #include <linux/slab.h>
  16 #include <asm/io.h>
  17
  18 /* Mali T6xx DDK includes */
  19 #include "linux/mali_linux_trace.h"
  20 #include "kbase/src/common/mali_kbase.h"
  21 #include "kbase/src/linux/mali_kbase_mem_linux.h"
  22
  23 #include "gator_events_mali_common.h"
  24
  25 /* If API version is not specified then assume API version 1. */
  26 #ifndef MALI_DDK_GATOR_API_VERSION
  27 #define MALI_DDK_GATOR_API_VERSION 1
  28 #endif
  29
  30 #if (MALI_DDK_GATOR_API_VERSION != 1) && (MALI_DDK_GATOR_API_VERSION != 2)
  31 #error MALI_DDK_GATOR_API_VERSION is invalid (must be 1 for r1/r2 DDK, or 2 for r3 DDK).
  32 #endif
  33
  34 /*
  35  * Mali-T6xx
  36  */
  37 typedef struct kbase_device *kbase_find_device_type(int);
  38 typedef kbase_context *kbase_create_context_type(kbase_device *);
  39 typedef void kbase_destroy_context_type(kbase_context *);
  40
  41 #if MALI_DDK_GATOR_API_VERSION == 1
  42 typedef void *kbase_va_alloc_type(kbase_context *, u32);
  43 typedef void kbase_va_free_type(kbase_context *, void *);
  44 #elif MALI_DDK_GATOR_API_VERSION == 2
  45 typedef void *kbase_va_alloc_type(kbase_context *, u32, kbase_hwc_dma_mapping * handle);
  46 typedef void kbase_va_free_type(kbase_context *, kbase_hwc_dma_mapping * handle);
  47 #endif
  48
  49 typedef mali_error kbase_instr_hwcnt_enable_type(kbase_context *, kbase_uk_hwcnt_setup *);
  50 typedef mali_error kbase_instr_hwcnt_disable_type(kbase_context *);
  51 typedef mali_error kbase_instr_hwcnt_clear_type(kbase_context *);
  52 typedef mali_error kbase_instr_hwcnt_dump_irq_type(kbase_context *);
  53 typedef mali_bool kbase_instr_hwcnt_dump_complete_type(kbase_context *, mali_bool *);
  54
  55 static kbase_find_device_type *kbase_find_device_symbol;
  56 static kbase_create_context_type *kbase_create_context_symbol;
  57 static kbase_va_alloc_type *kbase_va_alloc_symbol;
  58 static kbase_instr_hwcnt_enable_type *kbase_instr_hwcnt_enable_symbol;
  59 static kbase_instr_hwcnt_clear_type *kbase_instr_hwcnt_clear_symbol;
  60 static kbase_instr_hwcnt_dump_irq_type *kbase_instr_hwcnt_dump_irq_symbol;
  61 static kbase_instr_hwcnt_dump_complete_type *kbase_instr_hwcnt_dump_complete_symbol;
  62 static kbase_instr_hwcnt_disable_type *kbase_instr_hwcnt_disable_symbol;
  63 static kbase_va_free_type *kbase_va_free_symbol;
  64 static kbase_destroy_context_type *kbase_destroy_context_symbol;
  65
  66 static long shader_present_low = 0;
  67
  68 /** The interval between reads, in ns.
  69  *
  70  * Earlier we introduced
  71  * a 'hold off for 1ms after last read' to resolve MIDBASE-2178 and MALINE-724.
  72  * However, the 1ms hold off is too long if no context switches occur as there is a race
  73  * between this value and the tick of the read clock in gator which is also 1ms. If we 'miss' the
  74  * current read, the counter values are effectively 'spread' over 2ms and the values seen are half
  75  * what they should be (since Streamline averages over sample time). In the presence of context switches
  76  * this spread can vary and markedly affect the counters.  Currently there is no 'proper' solution to
  77  * this, but empirically we have found that reducing the minimum read interval to 950us causes the
  78  * counts to be much more stable.
  79  */
  80 static const int READ_INTERVAL_NSEC = 950000;
  81
  82 #if GATOR_TEST
  83 #include "gator_events_mali_t6xx_hw_test.c"
  84 #endif
  85
  86 /* Blocks for HW counters */
  87 enum {
  88         JM_BLOCK = 0,
  89         TILER_BLOCK,
  90         SHADER_BLOCK,
  91         MMU_BLOCK
  92 };
  93
  94 /* Counters for Mali-T6xx:
  95  *
  96  *  - HW counters, 4 blocks
  97  *    For HW counters we need strings to create /dev/gator/events files.
  98  *    Enums are not needed because the position of the HW name in the array is the same
  99  *    of the corresponding value in the received block of memory.
 100  *    HW counters are requested by calculating a bitmask, passed then to the driver.
 101  *    Every millisecond a HW counters dump is requested, and if the previous has been completed they are read.
 102  */
 103
 104 /* Hardware Counters */
 105 static const char *const hardware_counter_names[] = {
 106         /* Job Manager */
 107         "",
 108         "",
 109         "",
 110         "",
 111         "MESSAGES_SENT",
 112         "MESSAGES_RECEIVED",
 113         "GPU_ACTIVE",           /* 6 */
 114         "IRQ_ACTIVE",
 115         "JS0_JOBS",
 116         "JS0_TASKS",
 117         "JS0_ACTIVE",
 118         "",
 119         "JS0_WAIT_READ",
 120         "JS0_WAIT_ISSUE",
 121         "JS0_WAIT_DEPEND",
 122         "JS0_WAIT_FINISH",
 123         "JS1_JOBS",
 124         "JS1_TASKS",
 125         "JS1_ACTIVE",
 126         "",
 127         "JS1_WAIT_READ",
 128         "JS1_WAIT_ISSUE",
 129         "JS1_WAIT_DEPEND",
 130         "JS1_WAIT_FINISH",
 131         "JS2_JOBS",
 132         "JS2_TASKS",
 133         "JS2_ACTIVE",
 134         "",
 135         "JS2_WAIT_READ",
 136         "JS2_WAIT_ISSUE",
 137         "JS2_WAIT_DEPEND",
 138         "JS2_WAIT_FINISH",
 139         "JS3_JOBS",
 140         "JS3_TASKS",
 141         "JS3_ACTIVE",
 142         "",
 143         "JS3_WAIT_READ",
 144         "JS3_WAIT_ISSUE",
 145         "JS3_WAIT_DEPEND",
 146         "JS3_WAIT_FINISH",
 147         "JS4_JOBS",
 148         "JS4_TASKS",
 149         "JS4_ACTIVE",
 150         "",
 151         "JS4_WAIT_READ",
 152         "JS4_WAIT_ISSUE",
 153         "JS4_WAIT_DEPEND",
 154         "JS4_WAIT_FINISH",
 155         "JS5_JOBS",
 156         "JS5_TASKS",
 157         "JS5_ACTIVE",
 158         "",
 159         "JS5_WAIT_READ",
 160         "JS5_WAIT_ISSUE",
 161         "JS5_WAIT_DEPEND",
 162         "JS5_WAIT_FINISH",
 163         "JS6_JOBS",
 164         "JS6_TASKS",
 165         "JS6_ACTIVE",
 166         "",
 167         "JS6_WAIT_READ",
 168         "JS6_WAIT_ISSUE",
 169         "JS6_WAIT_DEPEND",
 170         "JS6_WAIT_FINISH",
 171
 172         /*Tiler */
 173         "",
 174         "",
 175         "",
 176         "JOBS_PROCESSED",
 177         "TRIANGLES",
 178         "QUADS",
 179         "POLYGONS",
 180         "POINTS",
 181         "LINES",
 182         "VCACHE_HIT",
 183         "VCACHE_MISS",
 184         "FRONT_FACING",
 185         "BACK_FACING",
 186         "PRIM_VISIBLE",
 187         "PRIM_CULLED",
 188         "PRIM_CLIPPED",
 189         "LEVEL0",
 190         "LEVEL1",
 191         "LEVEL2",
 192         "LEVEL3",
 193         "LEVEL4",
 194         "LEVEL5",
 195         "LEVEL6",
 196         "LEVEL7",
 197         "COMMAND_1",
 198         "COMMAND_2",
 199         "COMMAND_3",
 200         "COMMAND_4",
 201         "COMMAND_4_7",
 202         "COMMAND_8_15",
 203         "COMMAND_16_63",
 204         "COMMAND_64",
 205         "COMPRESS_IN",
 206         "COMPRESS_OUT",
 207         "COMPRESS_FLUSH",
 208         "TIMESTAMPS",
 209         "PCACHE_HIT",
 210         "PCACHE_MISS",
 211         "PCACHE_LINE",
 212         "PCACHE_STALL",
 213         "WRBUF_HIT",
 214         "WRBUF_MISS",
 215         "WRBUF_LINE",
 216         "WRBUF_PARTIAL",
 217         "WRBUF_STALL",
 218         "ACTIVE",
 219         "LOADING_DESC",
 220         "INDEX_WAIT",
 221         "INDEX_RANGE_WAIT",
 222         "VERTEX_WAIT",
 223         "PCACHE_WAIT",
 224         "WRBUF_WAIT",
 225         "BUS_READ",
 226         "BUS_WRITE",
 227         "",
 228         "",
 229         "",
 230         "",
 231         "",
 232         "UTLB_STALL",
 233         "UTLB_REPLAY_MISS",
 234         "UTLB_REPLAY_FULL",
 235         "UTLB_NEW_MISS",
 236         "UTLB_HIT",
 237
 238         /* Shader Core */
 239         "",
 240         "",
 241         "",
 242         "SHADER_CORE_ACTIVE",
 243         "FRAG_ACTIVE",
 244         "FRAG_PRIMATIVES",
 245         "FRAG_PRIMATIVES_DROPPED",
 246         "FRAG_CYCLE_DESC",
 247         "FRAG_CYCLES_PLR",
 248         "FRAG_CYCLES_VERT",
 249         "FRAG_CYCLES_TRISETUP",
 250         "FRAG_CYCLES_RAST",
 251         "FRAG_THREADS",
 252         "FRAG_DUMMY_THREADS",
 253         "FRAG_QUADS_RAST",
 254         "FRAG_QUADS_EZS_TEST",
 255         "FRAG_QUADS_EZS_KILLED",
 256         "FRAG_QUADS_LZS_TEST",
 257         "FRAG_QUADS_LZS_KILLED",
 258         "FRAG_CYCLE_NO_TILE",
 259         "FRAG_NUM_TILES",
 260         "FRAG_TRANS_ELIM",
 261         "COMPUTE_ACTIVE",
 262         "COMPUTE_TASKS",
 263         "COMPUTE_THREADS",
 264         "COMPUTE_CYCLES_DESC",
 265         "TRIPIPE_ACTIVE",
 266         "ARITH_WORDS",
 267         "ARITH_CYCLES_REG",
 268         "ARITH_CYCLES_L0",
 269         "ARITH_FRAG_DEPEND",
 270         "LS_WORDS",
 271         "LS_ISSUES",
 272         "LS_RESTARTS",
 273         "LS_REISSUES_MISS",
 274         "LS_REISSUES_VD",
 275         "LS_REISSUE_ATTRIB_MISS",
 276         "LS_NO_WB",
 277         "TEX_WORDS",
 278         "TEX_BUBBLES",
 279         "TEX_WORDS_L0",
 280         "TEX_WORDS_DESC",
 281         "TEX_THREADS",
 282         "TEX_RECIRC_FMISS",
 283         "TEX_RECIRC_DESC",
 284         "TEX_RECIRC_MULTI",
 285         "TEX_RECIRC_PMISS",
 286         "TEX_RECIRC_CONF",
 287         "LSC_READ_HITS",
 288         "LSC_READ_MISSES",
 289         "LSC_WRITE_HITS",
 290         "LSC_WRITE_MISSES",
 291         "LSC_ATOMIC_HITS",
 292         "LSC_ATOMIC_MISSES",
 293         "LSC_LINE_FETCHES",
 294         "LSC_DIRTY_LINE",
 295         "LSC_SNOOPS",
 296         "AXI_TLB_STALL",
 297         "AXI_TLB_MIESS",
 298         "AXI_TLB_TRANSACTION",
 299         "LS_TLB_MISS",
 300         "LS_TLB_HIT",
 301         "AXI_BEATS_READ",
 302         "AXI_BEATS_WRITTEN",
 303
 304         /*L2 and MMU */
 305         "",
 306         "",
 307         "",
 308         "",
 309         "MMU_HIT",
 310         "MMU_NEW_MISS",
 311         "MMU_REPLAY_FULL",
 312         "MMU_REPLAY_MISS",
 313         "MMU_TABLE_WALK",
 314         "",
 315         "",
 316         "",
 317         "",
 318         "",
 319         "",
 320         "",
 321         "UTLB_HIT",
 322         "UTLB_NEW_MISS",
 323         "UTLB_REPLAY_FULL",
 324         "UTLB_REPLAY_MISS",
 325         "UTLB_STALL",
 326         "",
 327         "",
 328         "",
 329         "",
 330         "",
 331         "",
 332         "",
 333         "",
 334         "",
 335         "L2_WRITE_BEATS",
 336         "L2_READ_BEATS",
 337         "L2_ANY_LOOKUP",
 338         "L2_READ_LOOKUP",
 339         "L2_SREAD_LOOKUP",
 340         "L2_READ_REPLAY",
 341         "L2_READ_SNOOP",
 342         "L2_READ_HIT",
 343         "L2_CLEAN_MISS",
 344         "L2_WRITE_LOOKUP",
 345         "L2_SWRITE_LOOKUP",
 346         "L2_WRITE_REPLAY",
 347         "L2_WRITE_SNOOP",
 348         "L2_WRITE_HIT",
 349         "L2_EXT_READ_FULL",
 350         "L2_EXT_READ_HALF",
 351         "L2_EXT_WRITE_FULL",
 352         "L2_EXT_WRITE_HALF",
 353         "L2_EXT_READ",
 354         "L2_EXT_READ_LINE",
 355         "L2_EXT_WRITE",
 356         "L2_EXT_WRITE_LINE",
 357         "L2_EXT_WRITE_SMALL",
 358         "L2_EXT_BARRIER",
 359         "L2_EXT_AR_STALL",
 360         "L2_EXT_R_BUF_FULL",
 361         "L2_EXT_RD_BUF_FULL",
 362         "L2_EXT_R_RAW",
 363         "L2_EXT_W_STALL",
 364         "L2_EXT_W_BUF_FULL",
 365         "L2_EXT_R_W_HAZARD",
 366         "L2_TAG_HAZARD",
 367         "L2_SNOOP_FULL",
 368         "L2_REPLAY_FULL"
 369 };
 370
 371 #define NUMBER_OF_HARDWARE_COUNTERS (sizeof(hardware_counter_names) / sizeof(hardware_counter_names[0]))
 372
 373 #define GET_HW_BLOCK(c) (((c) >> 6) & 0x3)
 374 #define GET_COUNTER_OFFSET(c) ((c) & 0x3f)
 375
 376 /* Memory to dump hardware counters into */
 377 static void *kernel_dump_buffer;
 378
 379 #if MALI_DDK_GATOR_API_VERSION == 2
 380 /* DMA state used to manage lifetime of the buffer */
 381 kbase_hwc_dma_mapping kernel_dump_buffer_handle;
 382 #endif
 383
 384 /* kbase context and device */
 385 static kbase_context *kbcontext = NULL;
 386 static struct kbase_device *kbdevice = NULL;
 387
 388 /*
 389  * The following function has no external prototype in older DDK revisions.  When the DDK
 390  * is updated then this should be removed.
 391  */
 392 struct kbase_device *kbase_find_device(int minor);
 393
 394 static volatile bool kbase_device_busy = false;
 395 static unsigned int num_hardware_counters_enabled;
 396
 397 /*
 398  * gatorfs variables for counter enable state
 399  */
 400 static mali_counter counters[NUMBER_OF_HARDWARE_COUNTERS];
 401
 402 /* An array used to return the data we recorded
 403  * as key,value pairs hence the *2
 404  */
 405 static unsigned long counter_dump[NUMBER_OF_HARDWARE_COUNTERS * 2];
 406
 407 #define SYMBOL_GET(FUNCTION, ERROR_COUNT) \
 408         if(FUNCTION ## _symbol) \
 409         { \
 410                 printk("gator: mali " #FUNCTION " symbol was already registered\n"); \
 411                 (ERROR_COUNT)++; \
 412         } \
 413         else \
 414         { \
 415                 FUNCTION ## _symbol = symbol_get(FUNCTION); \
 416                 if(! FUNCTION ## _symbol) \
 417                 { \
 418                         printk("gator: mali online " #FUNCTION " symbol not found\n"); \
 419                         (ERROR_COUNT)++; \
 420                 } \
 421         }
 422
 423 #define SYMBOL_CLEANUP(FUNCTION) \
 424         if(FUNCTION ## _symbol) \
 425         { \
 426         symbol_put(FUNCTION); \
 427         FUNCTION ## _symbol = NULL; \
 428         }
 429
 430 /**
 431  * Execute symbol_get for all the Mali symbols and check for success.
 432  * @return the number of symbols not loaded.
 433  */
 434 static int init_symbols(void)
 435 {
 436         int error_count = 0;
 437         SYMBOL_GET(kbase_find_device, error_count);
 438         SYMBOL_GET(kbase_create_context, error_count);
 439         SYMBOL_GET(kbase_va_alloc, error_count);
 440         SYMBOL_GET(kbase_instr_hwcnt_enable, error_count);
 441         SYMBOL_GET(kbase_instr_hwcnt_clear, error_count);
 442         SYMBOL_GET(kbase_instr_hwcnt_dump_irq, error_count);
 443         SYMBOL_GET(kbase_instr_hwcnt_dump_complete, error_count);
 444         SYMBOL_GET(kbase_instr_hwcnt_disable, error_count);
 445         SYMBOL_GET(kbase_va_free, error_count);
 446         SYMBOL_GET(kbase_destroy_context, error_count);
 447
 448         return error_count;
 449 }
 450
 451 /**
 452  * Execute symbol_put for all the registered Mali symbols.
 453  */
 454 static void clean_symbols(void)
 455 {
 456         SYMBOL_CLEANUP(kbase_find_device);
 457         SYMBOL_CLEANUP(kbase_create_context);
 458         SYMBOL_CLEANUP(kbase_va_alloc);
 459         SYMBOL_CLEANUP(kbase_instr_hwcnt_enable);
 460         SYMBOL_CLEANUP(kbase_instr_hwcnt_clear);
 461         SYMBOL_CLEANUP(kbase_instr_hwcnt_dump_irq);
 462         SYMBOL_CLEANUP(kbase_instr_hwcnt_dump_complete);
 463         SYMBOL_CLEANUP(kbase_instr_hwcnt_disable);
 464         SYMBOL_CLEANUP(kbase_va_free);
 465         SYMBOL_CLEANUP(kbase_destroy_context);
 466 }
 467
 468 /**
 469  * Determines whether a read should take place
 470  * @param current_time The current time, obtained from getnstimeofday()
 471  * @param prev_time_s The number of seconds at the previous read attempt.
 472  * @param next_read_time_ns The time (in ns) when the next read should be allowed.
 473  *
 474  * Note that this function has been separated out here to allow it to be tested.
 475  */
 476 static int is_read_scheduled(const struct timespec *current_time, u32 *prev_time_s, s32 *next_read_time_ns)
 477 {
 478         /* If the current ns count rolls over a second, roll the next read time too. */
 479         if (current_time->tv_sec != *prev_time_s) {
 480                 *next_read_time_ns = *next_read_time_ns - NSEC_PER_SEC;
 481         }
 482
 483         /* Abort the read if the next read time has not arrived. */
 484         if (current_time->tv_nsec < *next_read_time_ns) {
 485                 return 0;
 486         }
 487
 488         /* Set the next read some fixed time after this one, and update the read timestamp. */
 489         *next_read_time_ns = current_time->tv_nsec + READ_INTERVAL_NSEC;
 490
 491         *prev_time_s = current_time->tv_sec;
 492         return 1;
 493 }
 494
 495 static int start(void)
 496 {
 497         kbase_uk_hwcnt_setup setup;
 498         mali_error err;
 499         int cnt;
 500         u16 bitmask[] = { 0, 0, 0, 0 };
 501         unsigned long long shadersPresent = 0;
 502
 503         /* Setup HW counters */
 504         num_hardware_counters_enabled = 0;
 505
 506         if (NUMBER_OF_HARDWARE_COUNTERS != 256) {
 507                 pr_debug("Unexpected number of hardware counters defined: expecting 256, got %d\n", NUMBER_OF_HARDWARE_COUNTERS);
 508         }
 509
 510         /* Calculate enable bitmasks based on counters_enabled array */
 511         for (cnt = 0; cnt < NUMBER_OF_HARDWARE_COUNTERS; cnt++) {
 512                 const mali_counter *counter = &counters[cnt];
 513                 if (counter->enabled) {
 514                         int block = GET_HW_BLOCK(cnt);
 515                         int enable_bit = GET_COUNTER_OFFSET(cnt) / 4;
 516                         bitmask[block] |= (1 << enable_bit);
 517                         pr_debug("gator: Mali-T6xx: hardware counter %s selected [%d]\n", hardware_counter_names[cnt], cnt);
 518                         num_hardware_counters_enabled++;
 519                 }
 520         }
 521
 522         /* Create a kbase context for HW counters */
 523         if (num_hardware_counters_enabled > 0) {
 524                 if (init_symbols() > 0) {
 525                         clean_symbols();
 526                         /* No Mali driver code entrypoints found - not a fault. */
 527                         return 0;
 528                 }
 529
 530                 kbdevice = kbase_find_device_symbol(-1);
 531
 532                 /* If we already got a context, fail */
 533                 if (kbcontext) {
 534                         pr_debug("gator: Mali-T6xx: error context already present\n");
 535                         goto out;
 536                 }
 537
 538                 /* kbcontext will only be valid after all the Mali symbols are loaded successfully */
 539                 kbcontext = kbase_create_context_symbol(kbdevice);
 540                 if (!kbcontext) {
 541                         pr_debug("gator: Mali-T6xx: error creating kbase context\n");
 542                         goto out;
 543                 }
 544
 545
 546                 /* See if we can get the number of shader cores */
 547                 shadersPresent = kbdevice->shader_present_bitmap;
 548                 shader_present_low = (unsigned long)shadersPresent;
 549
 550                 /*
 551                  * The amount of memory needed to store the dump (bytes)
 552                  * DUMP_SIZE = number of core groups
 553                  *             * number of blocks (always 8 for midgard)
 554                  *             * number of counters per block (always 64 for midgard)
 555                  *             * number of bytes per counter (always 4 in midgard)
 556                  * For a Mali-T6xx with a single core group = 1 * 8 * 64 * 4 = 2048
 557                  * For a Mali-T6xx with a dual core group   = 2 * 8 * 64 * 4 = 4096
 558                  */
 559 #if MALI_DDK_GATOR_API_VERSION == 1
 560                 kernel_dump_buffer = kbase_va_alloc_symbol(kbcontext, 4096);
 561 #elif MALI_DDK_GATOR_API_VERSION == 2
 562                 kernel_dump_buffer = kbase_va_alloc_symbol(kbcontext, 4096, &kernel_dump_buffer_handle);
 563 #endif
 564                 if (!kernel_dump_buffer) {
 565                         pr_debug("gator: Mali-T6xx: error trying to allocate va\n");
 566                         goto destroy_context;
 567                 }
 568
 569                 setup.dump_buffer = (uintptr_t)kernel_dump_buffer;
 570                 setup.jm_bm = bitmask[JM_BLOCK];
 571                 setup.tiler_bm = bitmask[TILER_BLOCK];
 572                 setup.shader_bm = bitmask[SHADER_BLOCK];
 573                 setup.mmu_l2_bm = bitmask[MMU_BLOCK];
 574                 /* These counters do not exist on Mali-T60x */
 575                 setup.l3_cache_bm = 0;
 576
 577                 /* Use kbase API to enable hardware counters and provide dump buffer */
 578                 err = kbase_instr_hwcnt_enable_symbol(kbcontext, &setup);
 579                 if (err != MALI_ERROR_NONE) {
 580                         pr_debug("gator: Mali-T6xx: can't setup hardware counters\n");
 581                         goto free_buffer;
 582                 }
 583                 pr_debug("gator: Mali-T6xx: hardware counters enabled\n");
 584                 kbase_instr_hwcnt_clear_symbol(kbcontext);
 585                 pr_debug("gator: Mali-T6xx: hardware counters cleared \n");
 586
 587                 kbase_device_busy = false;
 588         }
 589
 590         return 0;
 591
 592 free_buffer:
 593 #if MALI_DDK_GATOR_API_VERSION == 1
 594         kbase_va_free_symbol(kbcontext, kernel_dump_buffer);
 595 #elif MALI_DDK_GATOR_API_VERSION == 2
 596         kbase_va_free_symbol(kbcontext, &kernel_dump_buffer_handle);
 597 #endif
 598
 599 destroy_context:
 600         kbase_destroy_context_symbol(kbcontext);
 601
 602 out:
 603         clean_symbols();
 604         return -1;
 605 }
 606
 607 static void stop(void)
 608 {
 609         unsigned int cnt;
 610         kbase_context *temp_kbcontext;
 611
 612         pr_debug("gator: Mali-T6xx: stop\n");
 613
 614         /* Set all counters as disabled */
 615         for (cnt = 0; cnt < NUMBER_OF_HARDWARE_COUNTERS; cnt++) {
 616                 counters[cnt].enabled = 0;
 617         }
 618
 619         /* Destroy the context for HW counters */
 620         if (num_hardware_counters_enabled > 0 && kbcontext != NULL) {
 621                 /*
 622                  * Set the global variable to NULL before destroying it, because
 623                  * other function will check this before using it.
 624                  */
 625                 temp_kbcontext = kbcontext;
 626                 kbcontext = NULL;
 627
 628                 kbase_instr_hwcnt_disable_symbol(temp_kbcontext);
 629
 630 #if MALI_DDK_GATOR_API_VERSION == 1
 631                 kbase_va_free_symbol(temp_kbcontext, kernel_dump_buffer);
 632 #elif MALI_DDK_GATOR_API_VERSION == 2
 633                 kbase_va_free_symbol(temp_kbcontext, &kernel_dump_buffer_handle);
 634 #endif
 635
 636                 kbase_destroy_context_symbol(temp_kbcontext);
 637
 638                 pr_debug("gator: Mali-T6xx: hardware counters stopped\n");
 639
 640                 clean_symbols();
 641         }
 642 }
 643
 644 static int read(int **buffer)
 645 {
 646         int cnt;
 647         int len = 0;
 648         u32 value = 0;
 649         mali_bool success;
 650
 651         struct timespec current_time;
 652         static u32 prev_time_s = 0;
 653         static s32 next_read_time_ns = 0;
 654
 655         if (!on_primary_core()) {
 656                 return 0;
 657         }
 658
 659         getnstimeofday(&current_time);
 660
 661         /*
 662          * Discard reads unless a respectable time has passed.  This reduces the load on the GPU without sacrificing
 663          * accuracy on the Streamline display.
 664          */
 665         if (!is_read_scheduled(&current_time, &prev_time_s, &next_read_time_ns)) {
 666                 return 0;
 667         }
 668
 669         /*
 670          * Report the HW counters
 671          * Only process hardware counters if at least one of the hardware counters is enabled.
 672          */
 673         if (num_hardware_counters_enabled > 0) {
 674                 const unsigned int vithar_blocks[] = {
 675                         0x700,  /* VITHAR_JOB_MANAGER,     Block 0 */
 676                         0x400,  /* VITHAR_TILER,           Block 1 */
 677                         0x000,  /* VITHAR_SHADER_CORE,     Block 2 */
 678                         0x500   /* VITHAR_MEMORY_SYSTEM,   Block 3 */
 679                 };
 680
 681                 if (!kbcontext) {
 682                         return -1;
 683                 }
 684
 685                 /* Mali symbols can be called safely since a kbcontext is valid */
 686                 if (kbase_instr_hwcnt_dump_complete_symbol(kbcontext, &success) == MALI_TRUE) {
 687                         kbase_device_busy = false;
 688
 689                         if (success == MALI_TRUE) {
 690                                 /* Cycle through hardware counters and accumulate totals */
 691                                 for (cnt = 0; cnt < NUMBER_OF_HARDWARE_COUNTERS; cnt++) {
 692                                         const mali_counter *counter = &counters[cnt];
 693                                         if (counter->enabled) {
 694                                                 const int block = GET_HW_BLOCK(cnt);
 695                                                 const int counter_offset = GET_COUNTER_OFFSET(cnt);
 696
 697                                                 const char* block_base_address = (char*)kernel_dump_buffer + vithar_blocks[block];
 698
 699                                                 /* If counter belongs to shader block need to take into account all cores */
 700                                                 if (block == SHADER_BLOCK) {
 701                                                         int i = 0;
 702                                                         int shader_core_count = 0;
 703                                                         value = 0;
 704
 705                                                         for (i = 0; i < 4; i++) {
 706                                                                 if ((shader_present_low >> i) & 1) {
 707                                                                         value += *((u32*) (block_base_address + (0x100 * i)) + counter_offset);
 708                                                                         shader_core_count++;
 709                                                                 }
 710                                                         }
 711
 712                                                         for (i = 0; i < 4; i++) {
 713                                                                 if((shader_present_low >> (i+4)) & 1) {
 714                                                                         value += *((u32*)(block_base_address + (0x100 * i) + 0x800) + counter_offset);
 715                                                                         shader_core_count++;
 716                                                                 }
 717                                                         }
 718
 719                                                         /* Need to total by number of cores to produce an average */
 720                                                         if (shader_core_count != 0) {
 721                                                                 value /= shader_core_count;
 722                                                         }
 723                                                 } else {
 724                                                         value = *((u32*)block_base_address + counter_offset);
 725                                                 }
 726
 727                                                 counter_dump[len++] = counter->key;
 728                                                 counter_dump[len++] = value;
 729                                         }
 730                                 }
 731                         }
 732                 }
 733
 734                 if (!kbase_device_busy) {
 735                         kbase_device_busy = true;
 736                         kbase_instr_hwcnt_dump_irq_symbol(kbcontext);
 737                 }
 738         }
 739
 740         /* Update the buffer */
 741         if (buffer) {
 742                 *buffer = (int *)counter_dump;
 743         }
 744
 745         return len;
 746 }
 747
 748 static int create_files(struct super_block *sb, struct dentry *root)
 749 {
 750         unsigned int event;
 751         /*
 752          * Create the filesystem for all events
 753          */
 754         int counter_index = 0;
 755         const char *mali_name = gator_mali_get_mali_name();
 756
 757         for (event = 0; event < NUMBER_OF_HARDWARE_COUNTERS; event++) {
 758                 if (gator_mali_create_file_system(mali_name, hardware_counter_names[counter_index], sb, root, &counters[event], NULL) != 0)
 759                         return -1;
 760                 counter_index++;
 761         }
 762
 763         return 0;
 764 }
 765
 766 static struct gator_interface gator_events_mali_t6xx_interface = {
 767         .create_files = create_files,
 768         .start = start,
 769         .stop = stop,
 770         .read = read
 771 };
 772
 773 int gator_events_mali_t6xx_hw_init(void)
 774 {
 775         pr_debug("gator: Mali-T6xx: sw_counters init\n");
 776
 777 #if GATOR_TEST
 778         test_all_is_read_scheduled();
 779 #endif
 780
 781         gator_mali_initialise_counters(counters, NUMBER_OF_HARDWARE_COUNTERS);
 782
 783         return gator_events_install(&gator_events_mali_t6xx_interface);
 784 }