Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1 | /* Copyright (c) 2013-2017, The Linux Foundation. All rights reserved. |
| 2 | * |
| 3 | * This program is free software; you can redistribute it and/or modify |
| 4 | * it under the terms of the GNU General Public License version 2 and |
| 5 | * only version 2 as published by the Free Software Foundation. |
| 6 | * |
| 7 | * This program is distributed in the hope that it will be useful, |
| 8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 10 | * GNU General Public License for more details. |
| 11 | * |
| 12 | */ |
| 13 | #include <linux/fs.h> |
| 14 | #include <linux/kernel.h> |
| 15 | #include <linux/ctype.h> |
| 16 | #include <linux/slab.h> |
| 17 | #include <linux/delay.h> |
| 18 | #include <linux/uaccess.h> |
| 19 | #include <linux/vmalloc.h> |
| 20 | #include <linux/debugfs.h> |
| 21 | |
| 22 | #include "adreno.h" |
| 23 | #include "adreno_profile.h" |
| 24 | #include "kgsl_sharedmem.h" |
| 25 | #include "adreno_pm4types.h" |
| 26 | |
| 27 | #define ASSIGNS_STR_FORMAT "%.8s:%u " |
| 28 | |
| 29 | /* |
| 30 | * Raw Data for processing later: |
| 31 | * : 3 - timestamp, count, context id |
| 32 | * [per counter] - data for each counter |
| 33 | * : 1 - Register offset |
| 34 | * : 2 - Pre IB register hi/lo value |
| 35 | * : 2 - Post IB register hi/lo value |
| 36 | * [per counter end] |
| 37 | */ |
| 38 | #define SIZE_DATA(cnt) (6 + (cnt) * 5) |
| 39 | |
| 40 | /* |
| 41 | * Pre-IB command size (in dwords): |
| 42 | * : 2 - NOP start identifier |
| 43 | * : 4 - timestamp |
| 44 | * : 4 - count |
| 45 | * : 4 - context id |
| 46 | * : 4 - pid |
| 47 | * : 4 - tid |
| 48 | * : 4 - type |
| 49 | * [loop count start] - for each counter to watch |
| 50 | * : 4 - Register offset |
| 51 | * : 4 - Register read lo |
| 52 | * : 4 - Register read high |
| 53 | * [loop end] |
| 54 | * : 2 - NOP end identifier |
| 55 | */ |
| 56 | #define SIZE_PREIB(cnt) (28 + (cnt) * 12) |
| 57 | |
| 58 | /* |
| 59 | * Post-IB command size (in dwords): |
| 60 | * : 2 - NOP start identifier |
| 61 | * [loop count start] - for each counter to watch |
| 62 | * : 4 - Register read lo |
| 63 | * : 4 - Register read high |
| 64 | * [loop end] |
| 65 | * : 2 - NOP end identifier |
| 66 | */ |
| 67 | #define SIZE_POSTIB(cnt) (4 + (cnt) * 8) |
| 68 | |
| 69 | /* Counter data + Pre size + post size = total size */ |
| 70 | #define SIZE_SHARED_ENTRY(cnt) (SIZE_DATA(cnt) + SIZE_PREIB(cnt) \ |
| 71 | + SIZE_POSTIB(cnt)) |
| 72 | |
| 73 | /* |
| 74 | * Space for following string :"%u %u %u %.5s %u " |
| 75 | * [count iterations]: "%.8s:%u %llu %llu%c" |
| 76 | */ |
| 77 | #define SIZE_PIPE_ENTRY(cnt) (50 + (cnt) * 62) |
| 78 | #define SIZE_LOG_ENTRY(cnt) (6 + (cnt) * 5) |
| 79 | |
| 80 | static struct adreno_context_type ctxt_type_table[] = {KGSL_CONTEXT_TYPES}; |
| 81 | |
| 82 | static const char *get_api_type_str(unsigned int type) |
| 83 | { |
| 84 | int i; |
| 85 | |
| 86 | for (i = 0; i < ARRAY_SIZE(ctxt_type_table) - 1; i++) { |
| 87 | if (ctxt_type_table[i].type == type) |
| 88 | return ctxt_type_table[i].str; |
| 89 | } |
| 90 | return "UNKNOWN"; |
| 91 | } |
| 92 | |
| 93 | static inline uint _ib_start(struct adreno_device *adreno_dev, |
| 94 | unsigned int *cmds) |
| 95 | { |
| 96 | unsigned int *start = cmds; |
| 97 | |
| 98 | *cmds++ = cp_packet(adreno_dev, CP_NOP, 1); |
| 99 | *cmds++ = KGSL_START_OF_PROFILE_IDENTIFIER; |
| 100 | |
| 101 | return cmds - start; |
| 102 | } |
| 103 | |
| 104 | static inline uint _ib_end(struct adreno_device *adreno_dev, |
| 105 | unsigned int *cmds) |
| 106 | { |
| 107 | unsigned int *start = cmds; |
| 108 | |
| 109 | *cmds++ = cp_packet(adreno_dev, CP_NOP, 1); |
| 110 | *cmds++ = KGSL_END_OF_PROFILE_IDENTIFIER; |
| 111 | |
| 112 | return cmds - start; |
| 113 | } |
| 114 | |
| 115 | static inline uint _ib_cmd_mem_write(struct adreno_device *adreno_dev, |
| 116 | uint *cmds, uint64_t gpuaddr, uint val, uint *off) |
| 117 | { |
| 118 | unsigned int *start = cmds; |
| 119 | |
| 120 | *cmds++ = cp_mem_packet(adreno_dev, CP_MEM_WRITE, 2, 1); |
| 121 | cmds += cp_gpuaddr(adreno_dev, cmds, gpuaddr); |
| 122 | *cmds++ = val; |
| 123 | |
| 124 | *off += sizeof(unsigned int); |
| 125 | return cmds - start; |
| 126 | } |
| 127 | |
| 128 | static inline uint _ib_cmd_reg_to_mem(struct adreno_device *adreno_dev, |
| 129 | uint *cmds, uint64_t gpuaddr, uint val, uint *off) |
| 130 | { |
| 131 | unsigned int *start = cmds; |
| 132 | |
| 133 | *cmds++ = cp_mem_packet(adreno_dev, CP_REG_TO_MEM, 2, 1); |
| 134 | *cmds++ = val; |
| 135 | cmds += cp_gpuaddr(adreno_dev, cmds, gpuaddr); |
| 136 | |
| 137 | *off += sizeof(unsigned int); |
| 138 | return cmds - start; |
| 139 | } |
| 140 | |
| 141 | static inline int _create_ib_ref(struct adreno_device *adreno_dev, |
| 142 | struct kgsl_memdesc *memdesc, unsigned int *cmd, |
| 143 | unsigned int cnt, unsigned int off) |
| 144 | { |
| 145 | unsigned int *start = cmd; |
| 146 | |
| 147 | *cmd++ = cp_mem_packet(adreno_dev, CP_INDIRECT_BUFFER_PFE, 2, 1); |
| 148 | cmd += cp_gpuaddr(adreno_dev, cmd, (memdesc->gpuaddr + off)); |
| 149 | *cmd++ = cnt; |
| 150 | |
| 151 | return cmd - start; |
| 152 | } |
| 153 | |
| 154 | static int _build_pre_ib_cmds(struct adreno_device *adreno_dev, |
| 155 | struct adreno_profile *profile, |
| 156 | unsigned int *rbcmds, unsigned int head, |
| 157 | unsigned int timestamp, struct adreno_context *drawctxt) |
| 158 | { |
| 159 | struct adreno_profile_assigns_list *entry; |
| 160 | unsigned int *start, *ibcmds; |
| 161 | unsigned int count = profile->assignment_count; |
| 162 | uint64_t gpuaddr = profile->shared_buffer.gpuaddr; |
| 163 | unsigned int ib_offset = head + SIZE_DATA(count); |
| 164 | unsigned int data_offset = head * sizeof(unsigned int); |
| 165 | |
| 166 | ibcmds = ib_offset + ((unsigned int *) profile->shared_buffer.hostptr); |
| 167 | start = ibcmds; |
| 168 | |
| 169 | /* start of profile identifier */ |
| 170 | ibcmds += _ib_start(adreno_dev, ibcmds); |
| 171 | |
| 172 | /* |
| 173 | * Write ringbuffer commands to save the following to memory: |
| 174 | * timestamp, count, context_id, pid, tid, context type |
| 175 | */ |
| 176 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, gpuaddr + data_offset, |
| 177 | timestamp, &data_offset); |
| 178 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, gpuaddr + data_offset, |
| 179 | profile->assignment_count, &data_offset); |
| 180 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, gpuaddr + data_offset, |
| 181 | drawctxt->base.id, &data_offset); |
| 182 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, gpuaddr + data_offset, |
| 183 | drawctxt->base.proc_priv->pid, &data_offset); |
| 184 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, gpuaddr + data_offset, |
| 185 | drawctxt->base.tid, &data_offset); |
| 186 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, gpuaddr + data_offset, |
| 187 | drawctxt->type, &data_offset); |
| 188 | |
| 189 | /* loop for each countable assigned */ |
| 190 | list_for_each_entry(entry, &profile->assignments_list, list) { |
| 191 | ibcmds += _ib_cmd_mem_write(adreno_dev, ibcmds, |
| 192 | gpuaddr + data_offset, entry->offset, |
| 193 | &data_offset); |
| 194 | ibcmds += _ib_cmd_reg_to_mem(adreno_dev, ibcmds, |
| 195 | gpuaddr + data_offset, entry->offset, |
| 196 | &data_offset); |
| 197 | ibcmds += _ib_cmd_reg_to_mem(adreno_dev, ibcmds, |
| 198 | gpuaddr + data_offset, entry->offset_hi, |
| 199 | &data_offset); |
| 200 | |
| 201 | /* skip over post_ib counter data */ |
| 202 | data_offset += sizeof(unsigned int) * 2; |
| 203 | } |
| 204 | |
| 205 | /* end of profile identifier */ |
| 206 | ibcmds += _ib_end(adreno_dev, ibcmds); |
| 207 | |
| 208 | return _create_ib_ref(adreno_dev, &profile->shared_buffer, rbcmds, |
| 209 | ibcmds - start, ib_offset * sizeof(unsigned int)); |
| 210 | } |
| 211 | |
| 212 | static int _build_post_ib_cmds(struct adreno_device *adreno_dev, |
| 213 | struct adreno_profile *profile, |
| 214 | unsigned int *rbcmds, unsigned int head) |
| 215 | { |
| 216 | struct adreno_profile_assigns_list *entry; |
| 217 | unsigned int *start, *ibcmds; |
| 218 | unsigned int count = profile->assignment_count; |
| 219 | uint64_t gpuaddr = profile->shared_buffer.gpuaddr; |
| 220 | unsigned int ib_offset = head + SIZE_DATA(count) + SIZE_PREIB(count); |
| 221 | unsigned int data_offset = head * sizeof(unsigned int); |
| 222 | |
| 223 | ibcmds = ib_offset + ((unsigned int *) profile->shared_buffer.hostptr); |
| 224 | start = ibcmds; |
| 225 | /* start of profile identifier */ |
| 226 | ibcmds += _ib_start(adreno_dev, ibcmds); |
| 227 | |
| 228 | /* skip over pre_ib preamble */ |
| 229 | data_offset += sizeof(unsigned int) * 6; |
| 230 | |
| 231 | /* loop for each countable assigned */ |
| 232 | list_for_each_entry(entry, &profile->assignments_list, list) { |
| 233 | /* skip over pre_ib counter data */ |
| 234 | data_offset += sizeof(unsigned int) * 3; |
| 235 | ibcmds += _ib_cmd_reg_to_mem(adreno_dev, ibcmds, |
| 236 | gpuaddr + data_offset, entry->offset, |
| 237 | &data_offset); |
| 238 | ibcmds += _ib_cmd_reg_to_mem(adreno_dev, ibcmds, |
| 239 | gpuaddr + data_offset, entry->offset_hi, |
| 240 | &data_offset); |
| 241 | } |
| 242 | |
| 243 | /* end of profile identifier */ |
| 244 | ibcmds += _ib_end(adreno_dev, ibcmds); |
| 245 | |
| 246 | return _create_ib_ref(adreno_dev, &profile->shared_buffer, rbcmds, |
| 247 | ibcmds - start, ib_offset * sizeof(unsigned int)); |
| 248 | } |
| 249 | |
| 250 | static bool shared_buf_empty(struct adreno_profile *profile) |
| 251 | { |
| 252 | if (profile->shared_buffer.hostptr == NULL || |
| 253 | profile->shared_buffer.size == 0) |
| 254 | return true; |
| 255 | |
| 256 | if (profile->shared_head == profile->shared_tail) |
| 257 | return true; |
| 258 | |
| 259 | return false; |
| 260 | } |
| 261 | |
| 262 | static inline void shared_buf_inc(unsigned int max_size, |
| 263 | unsigned int *offset, size_t inc) |
| 264 | { |
| 265 | *offset = (*offset + inc) % max_size; |
| 266 | } |
| 267 | |
| 268 | static inline void log_buf_wrapcnt(unsigned int cnt, uintptr_t *off) |
| 269 | { |
| 270 | *off = (*off + cnt) % ADRENO_PROFILE_LOG_BUF_SIZE_DWORDS; |
| 271 | } |
| 272 | |
| 273 | static inline void log_buf_wrapinc_len(unsigned int *profile_log_buffer, |
| 274 | unsigned int **ptr, unsigned int len) |
| 275 | { |
| 276 | *ptr += len; |
| 277 | if (*ptr >= (profile_log_buffer + |
| 278 | ADRENO_PROFILE_LOG_BUF_SIZE_DWORDS)) |
| 279 | *ptr -= ADRENO_PROFILE_LOG_BUF_SIZE_DWORDS; |
| 280 | } |
| 281 | |
| 282 | static inline void log_buf_wrapinc(unsigned int *profile_log_buffer, |
| 283 | unsigned int **ptr) |
| 284 | { |
| 285 | log_buf_wrapinc_len(profile_log_buffer, ptr, 1); |
| 286 | } |
| 287 | |
| 288 | static inline unsigned int log_buf_available(struct adreno_profile *profile, |
| 289 | unsigned int *head_ptr) |
| 290 | { |
| 291 | uintptr_t tail, head; |
| 292 | |
| 293 | tail = (uintptr_t) profile->log_tail - |
| 294 | (uintptr_t) profile->log_buffer; |
| 295 | head = (uintptr_t)head_ptr - (uintptr_t) profile->log_buffer; |
| 296 | if (tail > head) |
| 297 | return (tail - head) / sizeof(uintptr_t); |
| 298 | else |
| 299 | return ADRENO_PROFILE_LOG_BUF_SIZE_DWORDS - ((head - tail) / |
| 300 | sizeof(uintptr_t)); |
| 301 | } |
| 302 | |
| 303 | static inline unsigned int shared_buf_available(struct adreno_profile *profile) |
| 304 | { |
| 305 | if (profile->shared_tail > profile->shared_head) |
| 306 | return profile->shared_tail - profile->shared_head; |
| 307 | else |
| 308 | return profile->shared_size - |
| 309 | (profile->shared_head - profile->shared_tail); |
| 310 | } |
| 311 | |
| 312 | static struct adreno_profile_assigns_list *_find_assignment_by_offset( |
| 313 | struct adreno_profile *profile, unsigned int offset) |
| 314 | { |
| 315 | struct adreno_profile_assigns_list *entry; |
| 316 | |
| 317 | list_for_each_entry(entry, &profile->assignments_list, list) { |
| 318 | if (entry->offset == offset) |
| 319 | return entry; |
| 320 | } |
| 321 | |
| 322 | return NULL; |
| 323 | } |
| 324 | |
| 325 | static bool _in_assignments_list(struct adreno_profile *profile, |
| 326 | unsigned int groupid, unsigned int countable) |
| 327 | { |
| 328 | struct adreno_profile_assigns_list *entry; |
| 329 | |
| 330 | list_for_each_entry(entry, &profile->assignments_list, list) { |
| 331 | if (entry->groupid == groupid && entry->countable == |
| 332 | countable) |
| 333 | return true; |
| 334 | } |
| 335 | |
| 336 | return false; |
| 337 | } |
| 338 | |
| 339 | static bool _add_to_assignments_list(struct adreno_profile *profile, |
| 340 | const char *str, unsigned int groupid, unsigned int countable, |
| 341 | unsigned int offset, unsigned int offset_hi) |
| 342 | { |
| 343 | struct adreno_profile_assigns_list *entry; |
| 344 | |
| 345 | /* first make sure we can alloc memory */ |
| 346 | entry = kmalloc(sizeof(struct adreno_profile_assigns_list), GFP_KERNEL); |
| 347 | if (!entry) |
| 348 | return false; |
| 349 | |
| 350 | list_add_tail(&entry->list, &profile->assignments_list); |
| 351 | |
| 352 | entry->countable = countable; |
| 353 | entry->groupid = groupid; |
| 354 | entry->offset = offset; |
| 355 | entry->offset_hi = offset_hi; |
| 356 | |
| 357 | strlcpy(entry->name, str, sizeof(entry->name)); |
| 358 | |
| 359 | profile->assignment_count++; |
| 360 | |
| 361 | return true; |
| 362 | } |
| 363 | |
| 364 | static bool results_available(struct adreno_device *adreno_dev, |
| 365 | struct adreno_profile *profile, unsigned int *shared_buf_tail) |
| 366 | { |
| 367 | unsigned int global_eop; |
| 368 | unsigned int off = profile->shared_tail; |
| 369 | unsigned int *shared_ptr = (unsigned int *) |
| 370 | profile->shared_buffer.hostptr; |
| 371 | unsigned int ts, cnt; |
| 372 | int ts_cmp; |
| 373 | |
| 374 | /* |
| 375 | * If shared_buffer empty or Memstore EOP timestamp is less than |
| 376 | * outstanding counter buffer timestamps then no results available |
| 377 | */ |
| 378 | if (shared_buf_empty(profile)) |
| 379 | return false; |
| 380 | |
| 381 | if (adreno_rb_readtimestamp(adreno_dev, |
| 382 | adreno_dev->cur_rb, |
| 383 | KGSL_TIMESTAMP_RETIRED, &global_eop)) |
| 384 | return false; |
| 385 | do { |
| 386 | cnt = *(shared_ptr + off + 1); |
| 387 | if (cnt == 0) |
| 388 | return false; |
| 389 | |
| 390 | ts = *(shared_ptr + off); |
| 391 | ts_cmp = timestamp_cmp(ts, global_eop); |
| 392 | if (ts_cmp >= 0) { |
| 393 | *shared_buf_tail = off; |
| 394 | if (off == profile->shared_tail) |
| 395 | return false; |
| 396 | else |
| 397 | return true; |
| 398 | } |
| 399 | shared_buf_inc(profile->shared_size, &off, |
| 400 | SIZE_SHARED_ENTRY(cnt)); |
| 401 | } while (off != profile->shared_head); |
| 402 | |
| 403 | *shared_buf_tail = profile->shared_head; |
| 404 | |
| 405 | return true; |
| 406 | } |
| 407 | |
| 408 | static void transfer_results(struct adreno_profile *profile, |
| 409 | unsigned int shared_buf_tail) |
| 410 | { |
| 411 | unsigned int buf_off; |
| 412 | unsigned int ts, cnt, ctxt_id, pid, tid, client_type; |
| 413 | unsigned int *ptr = (unsigned int *) profile->shared_buffer.hostptr; |
| 414 | unsigned int *log_ptr, *log_base; |
| 415 | struct adreno_profile_assigns_list *assigns_list; |
| 416 | int i, tmp_tail; |
| 417 | |
| 418 | log_ptr = profile->log_head; |
| 419 | log_base = profile->log_buffer; |
| 420 | if (log_ptr == NULL) |
| 421 | return; |
| 422 | |
| 423 | /* |
| 424 | * go through counter buffers and format for write into log_buffer |
| 425 | * if log buffer doesn't have space just overwrite it circularly |
| 426 | * shared_buf is guaranteed to not wrap within an entry so can use |
| 427 | * ptr increment |
| 428 | */ |
| 429 | while (profile->shared_tail != shared_buf_tail) { |
| 430 | buf_off = profile->shared_tail; |
| 431 | /* |
| 432 | * format: timestamp, count, context_id |
| 433 | * count entries: pc_off, pc_start, pc_end |
| 434 | */ |
| 435 | ts = *(ptr + buf_off++); |
| 436 | cnt = *(ptr + buf_off++); |
| 437 | ctxt_id = *(ptr + buf_off++); |
| 438 | pid = *(ptr + buf_off++); |
| 439 | tid = *(ptr + buf_off++); |
| 440 | client_type = *(ptr + buf_off++); |
| 441 | |
| 442 | /* |
| 443 | * if entry overwrites the tail of log_buffer then adjust tail |
| 444 | * ptr to make room for the new entry, discarding old entry |
| 445 | */ |
| 446 | while (log_buf_available(profile, log_ptr) <= |
| 447 | SIZE_LOG_ENTRY(cnt)) { |
| 448 | unsigned int size_tail; |
| 449 | uintptr_t boff; |
| 450 | |
| 451 | size_tail = SIZE_LOG_ENTRY(0xffff & |
| 452 | *(profile->log_tail)); |
| 453 | boff = ((uintptr_t) profile->log_tail - |
| 454 | (uintptr_t) log_base) / sizeof(uintptr_t); |
| 455 | log_buf_wrapcnt(size_tail, &boff); |
| 456 | profile->log_tail = log_base + boff; |
| 457 | } |
| 458 | |
| 459 | *log_ptr = cnt; |
| 460 | log_buf_wrapinc(log_base, &log_ptr); |
| 461 | *log_ptr = client_type; |
| 462 | log_buf_wrapinc(log_base, &log_ptr); |
| 463 | *log_ptr = pid; |
| 464 | log_buf_wrapinc(log_base, &log_ptr); |
| 465 | *log_ptr = tid; |
| 466 | log_buf_wrapinc(log_base, &log_ptr); |
| 467 | *log_ptr = ctxt_id; |
| 468 | log_buf_wrapinc(log_base, &log_ptr); |
| 469 | *log_ptr = ts; |
| 470 | log_buf_wrapinc(log_base, &log_ptr); |
| 471 | |
| 472 | for (i = 0; i < cnt; i++) { |
| 473 | assigns_list = _find_assignment_by_offset( |
| 474 | profile, *(ptr + buf_off++)); |
| 475 | if (assigns_list == NULL) { |
| 476 | *log_ptr = (unsigned int) -1; |
| 477 | |
| 478 | shared_buf_inc(profile->shared_size, |
| 479 | &profile->shared_tail, |
| 480 | SIZE_SHARED_ENTRY(cnt)); |
| 481 | goto err; |
| 482 | } else { |
| 483 | *log_ptr = assigns_list->groupid << 16 | |
| 484 | (assigns_list->countable & 0xffff); |
| 485 | } |
| 486 | log_buf_wrapinc(log_base, &log_ptr); |
| 487 | *log_ptr = *(ptr + buf_off++); /* perf cntr start hi */ |
| 488 | log_buf_wrapinc(log_base, &log_ptr); |
| 489 | *log_ptr = *(ptr + buf_off++); /* perf cntr start lo */ |
| 490 | log_buf_wrapinc(log_base, &log_ptr); |
| 491 | *log_ptr = *(ptr + buf_off++); /* perf cntr end hi */ |
| 492 | log_buf_wrapinc(log_base, &log_ptr); |
| 493 | *log_ptr = *(ptr + buf_off++); /* perf cntr end lo */ |
| 494 | log_buf_wrapinc(log_base, &log_ptr); |
| 495 | |
| 496 | } |
| 497 | |
| 498 | tmp_tail = profile->shared_tail; |
| 499 | shared_buf_inc(profile->shared_size, |
| 500 | &profile->shared_tail, |
| 501 | SIZE_SHARED_ENTRY(cnt)); |
| 502 | /* |
| 503 | * Possibly lost some room as we cycled around, so it's safe to |
| 504 | * reset the max size |
| 505 | */ |
| 506 | if (profile->shared_tail < tmp_tail) |
| 507 | profile->shared_size = |
| 508 | ADRENO_PROFILE_SHARED_BUF_SIZE_DWORDS; |
| 509 | |
| 510 | } |
| 511 | profile->log_head = log_ptr; |
| 512 | return; |
| 513 | err: |
| 514 | /* reset head/tail to same on error in hopes we work correctly later */ |
| 515 | profile->log_head = profile->log_tail; |
| 516 | } |
| 517 | |
| 518 | static int profile_enable_get(void *data, u64 *val) |
| 519 | { |
| 520 | struct kgsl_device *device = data; |
| 521 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 522 | |
| 523 | mutex_lock(&device->mutex); |
| 524 | *val = adreno_profile_enabled(&adreno_dev->profile); |
| 525 | mutex_unlock(&device->mutex); |
| 526 | |
| 527 | return 0; |
| 528 | } |
| 529 | |
| 530 | static int profile_enable_set(void *data, u64 val) |
| 531 | { |
| 532 | struct kgsl_device *device = data; |
| 533 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 534 | struct adreno_profile *profile = &adreno_dev->profile; |
| 535 | |
| 536 | mutex_lock(&device->mutex); |
| 537 | |
| 538 | if (val && profile->log_buffer == NULL) { |
| 539 | /* allocate profile_log_buffer the first time enabled */ |
| 540 | profile->log_buffer = vmalloc(ADRENO_PROFILE_LOG_BUF_SIZE); |
| 541 | if (profile->log_buffer == NULL) { |
| 542 | mutex_unlock(&device->mutex); |
| 543 | return -ENOMEM; |
| 544 | } |
| 545 | profile->log_tail = profile->log_buffer; |
| 546 | profile->log_head = profile->log_buffer; |
| 547 | } |
| 548 | |
| 549 | profile->enabled = val; |
| 550 | |
| 551 | mutex_unlock(&device->mutex); |
| 552 | |
| 553 | return 0; |
| 554 | } |
| 555 | |
| 556 | static ssize_t profile_assignments_read(struct file *filep, |
| 557 | char __user *ubuf, size_t max, loff_t *ppos) |
| 558 | { |
| 559 | struct kgsl_device *device = (struct kgsl_device *) filep->private_data; |
| 560 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 561 | struct adreno_profile *profile = &adreno_dev->profile; |
| 562 | struct adreno_profile_assigns_list *entry; |
| 563 | int len = 0, max_size = PAGE_SIZE; |
| 564 | char *buf, *pos; |
| 565 | ssize_t size = 0; |
| 566 | |
| 567 | mutex_lock(&device->mutex); |
| 568 | |
| 569 | if (profile->assignment_count == 0) { |
| 570 | mutex_unlock(&device->mutex); |
| 571 | return 0; |
| 572 | } |
| 573 | |
| 574 | buf = kmalloc(max_size, GFP_KERNEL); |
| 575 | if (!buf) { |
| 576 | mutex_unlock(&device->mutex); |
| 577 | return -ENOMEM; |
| 578 | } |
| 579 | |
| 580 | pos = buf; |
| 581 | |
| 582 | /* copy all assingments from list to str */ |
| 583 | list_for_each_entry(entry, &profile->assignments_list, list) { |
| 584 | len = snprintf(pos, max_size, ASSIGNS_STR_FORMAT, |
| 585 | entry->name, entry->countable); |
| 586 | |
| 587 | max_size -= len; |
| 588 | pos += len; |
| 589 | } |
| 590 | |
| 591 | size = simple_read_from_buffer(ubuf, max, ppos, buf, |
| 592 | strlen(buf)); |
| 593 | |
| 594 | kfree(buf); |
| 595 | |
| 596 | mutex_unlock(&device->mutex); |
| 597 | return size; |
| 598 | } |
| 599 | |
| 600 | static void _remove_assignment(struct adreno_device *adreno_dev, |
| 601 | unsigned int groupid, unsigned int countable) |
| 602 | { |
| 603 | struct adreno_profile *profile = &adreno_dev->profile; |
| 604 | struct adreno_profile_assigns_list *entry, *tmp; |
| 605 | |
| 606 | list_for_each_entry_safe(entry, tmp, &profile->assignments_list, list) { |
| 607 | if (entry->groupid == groupid && |
| 608 | entry->countable == countable) { |
| 609 | list_del(&entry->list); |
| 610 | |
| 611 | profile->assignment_count--; |
| 612 | |
| 613 | kfree(entry); |
| 614 | |
| 615 | /* remove from perf counter allocation */ |
| 616 | adreno_perfcounter_put(adreno_dev, groupid, countable, |
| 617 | PERFCOUNTER_FLAG_KERNEL); |
| 618 | } |
| 619 | } |
| 620 | } |
| 621 | |
| 622 | static void _add_assignment(struct adreno_device *adreno_dev, |
| 623 | unsigned int groupid, unsigned int countable) |
| 624 | { |
| 625 | struct adreno_profile *profile = &adreno_dev->profile; |
| 626 | unsigned int offset, offset_hi; |
| 627 | const char *name = NULL; |
| 628 | |
| 629 | name = adreno_perfcounter_get_name(adreno_dev, groupid); |
| 630 | if (!name) |
| 631 | return; |
| 632 | |
| 633 | /* if already in assigned list skip it */ |
| 634 | if (_in_assignments_list(profile, groupid, countable)) |
| 635 | return; |
| 636 | |
| 637 | /* add to perf counter allocation, if fail skip it */ |
| 638 | if (adreno_perfcounter_get(adreno_dev, groupid, countable, |
| 639 | &offset, &offset_hi, PERFCOUNTER_FLAG_NONE)) |
| 640 | return; |
| 641 | |
| 642 | /* add to assignments list, put counter back if error */ |
| 643 | if (!_add_to_assignments_list(profile, name, groupid, |
| 644 | countable, offset, offset_hi)) |
| 645 | adreno_perfcounter_put(adreno_dev, groupid, |
| 646 | countable, PERFCOUNTER_FLAG_KERNEL); |
| 647 | } |
| 648 | |
| 649 | static char *_parse_next_assignment(struct adreno_device *adreno_dev, |
| 650 | char *str, int *groupid, int *countable, bool *remove) |
| 651 | { |
| 652 | char *groupid_str, *countable_str, *next_str = NULL; |
| 653 | int ret; |
| 654 | |
| 655 | *groupid = -EINVAL; |
| 656 | *countable = -EINVAL; |
| 657 | *remove = false; |
| 658 | |
| 659 | /* remove spaces */ |
| 660 | while (*str == ' ') |
| 661 | str++; |
| 662 | |
| 663 | /* check if it's a remove assignment */ |
| 664 | if (*str == '-') { |
| 665 | *remove = true; |
| 666 | str++; |
| 667 | } |
| 668 | |
| 669 | /* get the groupid string */ |
| 670 | groupid_str = str; |
| 671 | while (*str != ':') { |
| 672 | if (*str == '\0') |
| 673 | return NULL; |
| 674 | *str = tolower(*str); |
| 675 | str++; |
| 676 | } |
| 677 | if (groupid_str == str) |
| 678 | return NULL; |
| 679 | |
| 680 | *str = '\0'; |
| 681 | str++; |
| 682 | |
| 683 | /* get the countable string */ |
| 684 | countable_str = str; |
| 685 | while (*str != ' ' && *str != '\0') |
| 686 | str++; |
| 687 | if (countable_str == str) |
| 688 | return NULL; |
| 689 | |
| 690 | /* |
| 691 | * If we have reached the end of the original string then make sure we |
| 692 | * return NULL from this function or we could accidently overrun |
| 693 | */ |
| 694 | |
| 695 | if (*str != '\0') { |
| 696 | *str = '\0'; |
| 697 | next_str = str + 1; |
| 698 | } |
| 699 | |
| 700 | /* set results */ |
| 701 | *groupid = adreno_perfcounter_get_groupid(adreno_dev, |
| 702 | groupid_str); |
| 703 | if (*groupid < 0) |
| 704 | return NULL; |
| 705 | ret = kstrtou32(countable_str, 10, countable); |
| 706 | if (ret) |
| 707 | return NULL; |
| 708 | |
| 709 | return next_str; |
| 710 | } |
| 711 | |
| 712 | static ssize_t profile_assignments_write(struct file *filep, |
| 713 | const char __user *user_buf, size_t len, loff_t *off) |
| 714 | { |
| 715 | struct kgsl_device *device = (struct kgsl_device *) filep->private_data; |
| 716 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 717 | struct adreno_profile *profile = &adreno_dev->profile; |
| 718 | size_t size = 0; |
| 719 | char *buf, *pbuf; |
| 720 | bool remove_assignment = false; |
| 721 | int groupid, countable, ret; |
| 722 | |
| 723 | if (len >= PAGE_SIZE || len == 0) |
| 724 | return -EINVAL; |
| 725 | |
| 726 | buf = kmalloc(len + 1, GFP_KERNEL); |
| 727 | if (buf == NULL) |
| 728 | return -ENOMEM; |
| 729 | |
| 730 | if (copy_from_user(buf, user_buf, len)) { |
| 731 | size = -EFAULT; |
| 732 | goto error_free; |
| 733 | } |
| 734 | |
| 735 | mutex_lock(&device->mutex); |
| 736 | |
| 737 | if (adreno_profile_enabled(profile)) { |
| 738 | size = -EINVAL; |
| 739 | goto error_unlock; |
| 740 | } |
| 741 | |
Carter Cooper | 05f2a6b | 2017-03-20 11:43:11 -0600 | [diff] [blame] | 742 | ret = adreno_perfcntr_active_oob_get(adreno_dev); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 743 | if (ret) { |
| 744 | size = ret; |
| 745 | goto error_unlock; |
| 746 | } |
| 747 | |
| 748 | /* |
| 749 | * When adding/removing assignments, ensure that the GPU is done with |
| 750 | * all it's work. This helps to synchronize the work flow to the |
| 751 | * GPU and avoid racey conditions. |
| 752 | */ |
| 753 | if (adreno_idle(device)) { |
| 754 | size = -ETIMEDOUT; |
| 755 | goto error_put; |
| 756 | } |
| 757 | |
| 758 | /* clear all shared buffer results */ |
| 759 | adreno_profile_process_results(adreno_dev); |
| 760 | |
| 761 | pbuf = buf; |
| 762 | |
| 763 | /* clear the log buffer */ |
| 764 | if (profile->log_buffer != NULL) { |
| 765 | profile->log_head = profile->log_buffer; |
| 766 | profile->log_tail = profile->log_buffer; |
| 767 | } |
| 768 | |
| 769 | |
| 770 | /* for sanity and parsing, ensure it is null terminated */ |
| 771 | buf[len] = '\0'; |
| 772 | |
| 773 | /* parse file buf and add(remove) to(from) appropriate lists */ |
| 774 | while (pbuf) { |
| 775 | pbuf = _parse_next_assignment(adreno_dev, pbuf, &groupid, |
| 776 | &countable, &remove_assignment); |
| 777 | if (groupid < 0 || countable < 0) |
| 778 | break; |
| 779 | |
| 780 | if (remove_assignment) |
| 781 | _remove_assignment(adreno_dev, groupid, countable); |
| 782 | else |
| 783 | _add_assignment(adreno_dev, groupid, countable); |
| 784 | } |
| 785 | |
| 786 | size = len; |
| 787 | |
| 788 | error_put: |
Carter Cooper | 05f2a6b | 2017-03-20 11:43:11 -0600 | [diff] [blame] | 789 | adreno_perfcntr_active_oob_put(adreno_dev); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 790 | error_unlock: |
| 791 | mutex_unlock(&device->mutex); |
| 792 | error_free: |
| 793 | kfree(buf); |
| 794 | return size; |
| 795 | } |
| 796 | |
| 797 | static int _pipe_print_pending(char __user *ubuf, size_t max) |
| 798 | { |
| 799 | loff_t unused = 0; |
| 800 | char str[] = "Operation Would Block!"; |
| 801 | |
| 802 | return simple_read_from_buffer(ubuf, max, |
| 803 | &unused, str, strlen(str)); |
| 804 | } |
| 805 | |
| 806 | static int _pipe_print_results(struct adreno_device *adreno_dev, |
| 807 | char __user *ubuf, size_t max) |
| 808 | { |
| 809 | struct adreno_profile *profile = &adreno_dev->profile; |
| 810 | const char *grp_name; |
| 811 | char __user *usr_buf = ubuf; |
| 812 | unsigned int *log_ptr = NULL, *tmp_log_ptr = NULL; |
| 813 | int len, i; |
| 814 | int status = 0; |
| 815 | ssize_t size, total_size = 0; |
| 816 | unsigned int cnt, api_type, ctxt_id, pid, tid, ts, cnt_reg; |
| 817 | unsigned long long pc_start, pc_end; |
| 818 | const char *api_str; |
| 819 | char format_space; |
| 820 | loff_t unused = 0; |
| 821 | char pipe_hdr_buf[51]; /* 4 uint32 + 5 space + 5 API type + '\0' */ |
| 822 | char pipe_cntr_buf[63]; /* 2 uint64 + 1 uint32 + 4 spaces + 8 group */ |
| 823 | |
| 824 | /* convert unread entries to ASCII, copy to user-space */ |
| 825 | log_ptr = profile->log_tail; |
| 826 | |
| 827 | do { |
| 828 | /* store the tmp var for error cases so we can skip */ |
| 829 | tmp_log_ptr = log_ptr; |
| 830 | |
| 831 | /* Too many to output to pipe, so skip this data */ |
| 832 | cnt = *log_ptr; |
| 833 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 834 | |
| 835 | if (SIZE_PIPE_ENTRY(cnt) > max) { |
| 836 | log_buf_wrapinc_len(profile->log_buffer, |
| 837 | &tmp_log_ptr, SIZE_PIPE_ENTRY(cnt)); |
| 838 | log_ptr = tmp_log_ptr; |
| 839 | goto done; |
| 840 | } |
| 841 | |
| 842 | /* |
| 843 | * Not enough space left in pipe, return without doing |
| 844 | * anything |
| 845 | */ |
| 846 | if ((max - (usr_buf - ubuf)) < SIZE_PIPE_ENTRY(cnt)) { |
| 847 | log_ptr = tmp_log_ptr; |
| 848 | goto done; |
| 849 | } |
| 850 | |
| 851 | api_type = *log_ptr; |
| 852 | api_str = get_api_type_str(api_type); |
| 853 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 854 | pid = *log_ptr; |
| 855 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 856 | tid = *log_ptr; |
| 857 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 858 | ctxt_id = *log_ptr; |
| 859 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 860 | ts = *log_ptr; |
| 861 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 862 | len = snprintf(pipe_hdr_buf, sizeof(pipe_hdr_buf) - 1, |
| 863 | "%u %u %u %.5s %u ", |
| 864 | pid, tid, ctxt_id, api_str, ts); |
| 865 | size = simple_read_from_buffer(usr_buf, |
| 866 | max - (usr_buf - ubuf), |
| 867 | &unused, pipe_hdr_buf, len); |
| 868 | |
| 869 | /* non-fatal error, so skip rest of entry and return */ |
| 870 | if (size < 0) { |
| 871 | log_buf_wrapinc_len(profile->log_buffer, |
| 872 | &tmp_log_ptr, SIZE_PIPE_ENTRY(cnt)); |
| 873 | log_ptr = tmp_log_ptr; |
| 874 | goto done; |
| 875 | } |
| 876 | |
| 877 | unused = 0; |
| 878 | usr_buf += size; |
| 879 | total_size += size; |
| 880 | |
| 881 | for (i = 0; i < cnt; i++) { |
| 882 | unsigned int start_lo, start_hi; |
| 883 | unsigned int end_lo, end_hi; |
| 884 | |
| 885 | grp_name = adreno_perfcounter_get_name( |
| 886 | adreno_dev, (*log_ptr >> 16) & 0xffff); |
| 887 | |
| 888 | /* non-fatal error, so skip rest of entry and return */ |
| 889 | if (grp_name == NULL) { |
| 890 | log_buf_wrapinc_len(profile->log_buffer, |
| 891 | &tmp_log_ptr, SIZE_PIPE_ENTRY(cnt)); |
| 892 | log_ptr = tmp_log_ptr; |
| 893 | goto done; |
| 894 | } |
| 895 | |
| 896 | if (i == cnt - 1) |
| 897 | format_space = '\n'; |
| 898 | else |
| 899 | format_space = ' '; |
| 900 | |
| 901 | cnt_reg = *log_ptr & 0xffff; |
| 902 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 903 | start_lo = *log_ptr; |
| 904 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 905 | start_hi = *log_ptr; |
| 906 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 907 | end_lo = *log_ptr; |
| 908 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 909 | end_hi = *log_ptr; |
| 910 | log_buf_wrapinc(profile->log_buffer, &log_ptr); |
| 911 | |
| 912 | pc_start = (((uint64_t) start_hi) << 32) | start_lo; |
| 913 | pc_end = (((uint64_t) end_hi) << 32) | end_lo; |
| 914 | |
| 915 | len = snprintf(pipe_cntr_buf, |
| 916 | sizeof(pipe_cntr_buf) - 1, |
| 917 | "%.8s:%u %llu %llu%c", |
| 918 | grp_name, cnt_reg, pc_start, |
| 919 | pc_end, format_space); |
| 920 | |
| 921 | size = simple_read_from_buffer(usr_buf, |
| 922 | max - (usr_buf - ubuf), |
| 923 | &unused, pipe_cntr_buf, len); |
| 924 | |
| 925 | /* non-fatal error, so skip rest of entry and return */ |
| 926 | if (size < 0) { |
| 927 | log_buf_wrapinc_len(profile->log_buffer, |
| 928 | &tmp_log_ptr, SIZE_PIPE_ENTRY(cnt)); |
| 929 | log_ptr = tmp_log_ptr; |
| 930 | goto done; |
| 931 | } |
| 932 | unused = 0; |
| 933 | usr_buf += size; |
| 934 | total_size += size; |
| 935 | } |
| 936 | } while (log_ptr != profile->log_head); |
| 937 | |
| 938 | done: |
| 939 | status = total_size; |
| 940 | profile->log_tail = log_ptr; |
| 941 | |
| 942 | return status; |
| 943 | } |
| 944 | |
| 945 | static ssize_t profile_pipe_print(struct file *filep, char __user *ubuf, |
| 946 | size_t max, loff_t *ppos) |
| 947 | { |
| 948 | struct kgsl_device *device = (struct kgsl_device *) filep->private_data; |
| 949 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 950 | struct adreno_profile *profile = &adreno_dev->profile; |
| 951 | char __user *usr_buf = ubuf; |
| 952 | int status = 0; |
| 953 | |
| 954 | /* |
| 955 | * this file not seekable since it only supports streaming, ignore |
| 956 | * ppos <> 0 |
| 957 | */ |
| 958 | /* |
| 959 | * format <pid> <tid> <context id> <cnt<<16 | client type> <timestamp> |
| 960 | * for each perf counter <cntr_reg_off> <start hi & lo> <end hi & low> |
| 961 | */ |
| 962 | |
| 963 | mutex_lock(&device->mutex); |
| 964 | |
| 965 | while (1) { |
| 966 | /* process any results that are available into the log_buffer */ |
| 967 | status = adreno_profile_process_results(adreno_dev); |
| 968 | if (status > 0) { |
| 969 | /* if we have results, print them and exit */ |
| 970 | status = _pipe_print_results(adreno_dev, usr_buf, max); |
| 971 | break; |
| 972 | } |
| 973 | |
| 974 | /* there are no unread results, act accordingly */ |
| 975 | if (filep->f_flags & O_NONBLOCK) { |
| 976 | if (profile->shared_tail != profile->shared_head) { |
| 977 | status = _pipe_print_pending(usr_buf, max); |
| 978 | break; |
| 979 | } |
| 980 | |
| 981 | status = 0; |
| 982 | break; |
| 983 | } |
| 984 | |
| 985 | mutex_unlock(&device->mutex); |
| 986 | set_current_state(TASK_INTERRUPTIBLE); |
| 987 | schedule_timeout(msecs_to_jiffies(100)); |
| 988 | mutex_lock(&device->mutex); |
| 989 | |
| 990 | if (signal_pending(current)) { |
| 991 | status = 0; |
| 992 | break; |
| 993 | } |
| 994 | } |
| 995 | |
| 996 | mutex_unlock(&device->mutex); |
| 997 | |
| 998 | return status; |
| 999 | } |
| 1000 | |
| 1001 | static int profile_groups_print(struct seq_file *s, void *unused) |
| 1002 | { |
| 1003 | struct kgsl_device *device = (struct kgsl_device *) s->private; |
| 1004 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 1005 | struct adreno_gpudev *gpudev = ADRENO_GPU_DEVICE(adreno_dev); |
| 1006 | struct adreno_perfcounters *counters = gpudev->perfcounters; |
| 1007 | struct adreno_perfcount_group *group; |
| 1008 | int i, j, used; |
| 1009 | |
| 1010 | mutex_lock(&device->mutex); |
| 1011 | |
| 1012 | for (i = 0; i < counters->group_count; ++i) { |
| 1013 | group = &(counters->groups[i]); |
| 1014 | /* get number of counters used for this group */ |
| 1015 | used = 0; |
| 1016 | for (j = 0; j < group->reg_count; j++) { |
| 1017 | if (group->regs[j].countable != |
| 1018 | KGSL_PERFCOUNTER_NOT_USED) |
| 1019 | used++; |
| 1020 | } |
| 1021 | |
| 1022 | seq_printf(s, "%s %d %d\n", group->name, |
| 1023 | group->reg_count, used); |
| 1024 | } |
| 1025 | |
| 1026 | mutex_unlock(&device->mutex); |
| 1027 | |
| 1028 | return 0; |
| 1029 | } |
| 1030 | |
| 1031 | static int profile_groups_open(struct inode *inode, struct file *file) |
| 1032 | { |
| 1033 | return single_open(file, profile_groups_print, inode->i_private); |
| 1034 | } |
| 1035 | |
| 1036 | static const struct file_operations profile_groups_fops = { |
| 1037 | .owner = THIS_MODULE, |
| 1038 | .open = profile_groups_open, |
| 1039 | .read = seq_read, |
| 1040 | .llseek = noop_llseek, |
| 1041 | .release = single_release, |
| 1042 | }; |
| 1043 | |
| 1044 | static const struct file_operations profile_pipe_fops = { |
| 1045 | .owner = THIS_MODULE, |
| 1046 | .open = simple_open, |
| 1047 | .read = profile_pipe_print, |
| 1048 | .llseek = noop_llseek, |
| 1049 | }; |
| 1050 | |
| 1051 | static const struct file_operations profile_assignments_fops = { |
| 1052 | .owner = THIS_MODULE, |
| 1053 | .open = simple_open, |
| 1054 | .read = profile_assignments_read, |
| 1055 | .write = profile_assignments_write, |
| 1056 | .llseek = noop_llseek, |
| 1057 | }; |
| 1058 | |
| 1059 | DEFINE_SIMPLE_ATTRIBUTE(profile_enable_fops, |
| 1060 | profile_enable_get, |
| 1061 | profile_enable_set, "%llu\n"); |
| 1062 | |
| 1063 | void adreno_profile_init(struct adreno_device *adreno_dev) |
| 1064 | { |
| 1065 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 1066 | struct adreno_profile *profile = &adreno_dev->profile; |
| 1067 | struct dentry *profile_dir; |
| 1068 | int ret; |
| 1069 | |
| 1070 | profile->enabled = false; |
| 1071 | |
| 1072 | /* allocate shared_buffer, which includes pre_ib and post_ib */ |
| 1073 | profile->shared_size = ADRENO_PROFILE_SHARED_BUF_SIZE_DWORDS; |
| 1074 | ret = kgsl_allocate_global(device, &profile->shared_buffer, |
| 1075 | profile->shared_size * sizeof(unsigned int), |
| 1076 | 0, 0, "profile"); |
| 1077 | |
| 1078 | if (ret) { |
| 1079 | profile->shared_size = 0; |
| 1080 | return; |
| 1081 | } |
| 1082 | |
| 1083 | INIT_LIST_HEAD(&profile->assignments_list); |
| 1084 | |
| 1085 | /* Create perf counter debugfs */ |
| 1086 | profile_dir = debugfs_create_dir("profiling", device->d_debugfs); |
| 1087 | if (IS_ERR(profile_dir)) |
| 1088 | return; |
| 1089 | |
| 1090 | debugfs_create_file("enable", 0644, profile_dir, device, |
| 1091 | &profile_enable_fops); |
| 1092 | debugfs_create_file("blocks", 0444, profile_dir, device, |
| 1093 | &profile_groups_fops); |
| 1094 | debugfs_create_file("pipe", 0444, profile_dir, device, |
| 1095 | &profile_pipe_fops); |
| 1096 | debugfs_create_file("assignments", 0644, profile_dir, device, |
| 1097 | &profile_assignments_fops); |
| 1098 | } |
| 1099 | |
| 1100 | void adreno_profile_close(struct adreno_device *adreno_dev) |
| 1101 | { |
| 1102 | struct adreno_profile *profile = &adreno_dev->profile; |
| 1103 | struct adreno_profile_assigns_list *entry, *tmp; |
| 1104 | |
| 1105 | profile->enabled = false; |
| 1106 | vfree(profile->log_buffer); |
| 1107 | profile->log_buffer = NULL; |
| 1108 | profile->log_head = NULL; |
| 1109 | profile->log_tail = NULL; |
| 1110 | profile->shared_head = 0; |
| 1111 | profile->shared_tail = 0; |
| 1112 | kgsl_free_global(KGSL_DEVICE(adreno_dev), &profile->shared_buffer); |
| 1113 | profile->shared_size = 0; |
| 1114 | |
| 1115 | profile->assignment_count = 0; |
| 1116 | |
| 1117 | list_for_each_entry_safe(entry, tmp, &profile->assignments_list, list) { |
| 1118 | list_del(&entry->list); |
| 1119 | kfree(entry); |
| 1120 | } |
| 1121 | } |
| 1122 | |
| 1123 | int adreno_profile_process_results(struct adreno_device *adreno_dev) |
| 1124 | { |
| 1125 | struct adreno_profile *profile = &adreno_dev->profile; |
| 1126 | unsigned int shared_buf_tail = profile->shared_tail; |
| 1127 | |
| 1128 | if (!results_available(adreno_dev, profile, &shared_buf_tail)) |
| 1129 | return 0; |
| 1130 | |
| 1131 | /* |
| 1132 | * transfer retired results to log_buffer |
| 1133 | * update shared_buffer tail ptr |
| 1134 | */ |
| 1135 | transfer_results(profile, shared_buf_tail); |
| 1136 | |
| 1137 | return 1; |
| 1138 | } |
| 1139 | |
| 1140 | void adreno_profile_preib_processing(struct adreno_device *adreno_dev, |
| 1141 | struct adreno_context *drawctxt, unsigned int *cmd_flags, |
| 1142 | unsigned int **rbptr) |
| 1143 | { |
| 1144 | struct adreno_profile *profile = &adreno_dev->profile; |
| 1145 | int count = profile->assignment_count; |
| 1146 | unsigned int entry_head = profile->shared_head; |
| 1147 | unsigned int *shared_ptr; |
| 1148 | struct adreno_ringbuffer *rb = ADRENO_CURRENT_RINGBUFFER(adreno_dev); |
| 1149 | unsigned int rbcmds[4]; |
| 1150 | unsigned int *ptr = *rbptr; |
| 1151 | unsigned int i, ret = 0; |
| 1152 | |
| 1153 | *cmd_flags &= ~KGSL_CMD_FLAGS_PROFILE; |
| 1154 | |
| 1155 | if (!adreno_profile_assignments_ready(profile)) |
| 1156 | goto done; |
| 1157 | |
| 1158 | /* |
| 1159 | * check if space available, include the post_ib in space available |
| 1160 | * check so don't have to handle trying to undo the pre_ib insertion in |
| 1161 | * ringbuffer in the case where only the post_ib fails enough space |
| 1162 | */ |
| 1163 | if (SIZE_SHARED_ENTRY(count) >= shared_buf_available(profile)) |
| 1164 | goto done; |
| 1165 | |
| 1166 | if (entry_head + SIZE_SHARED_ENTRY(count) >= profile->shared_size) { |
| 1167 | /* entry_head would wrap, start entry_head at 0 in buffer */ |
| 1168 | entry_head = 0; |
| 1169 | profile->shared_size = profile->shared_head; |
| 1170 | profile->shared_head = 0; |
| 1171 | |
| 1172 | /* recheck space available */ |
| 1173 | if (SIZE_SHARED_ENTRY(count) >= shared_buf_available(profile)) |
| 1174 | goto done; |
| 1175 | } |
| 1176 | |
| 1177 | /* zero out the counter area of shared_buffer entry_head */ |
| 1178 | shared_ptr = entry_head + ((unsigned int *) |
| 1179 | profile->shared_buffer.hostptr); |
| 1180 | memset(shared_ptr, 0, SIZE_SHARED_ENTRY(count) * sizeof(unsigned int)); |
| 1181 | |
| 1182 | /* reserve space for the pre ib shared buffer */ |
| 1183 | shared_buf_inc(profile->shared_size, &profile->shared_head, |
| 1184 | SIZE_SHARED_ENTRY(count)); |
| 1185 | |
| 1186 | /* create the shared ibdesc */ |
| 1187 | ret = _build_pre_ib_cmds(adreno_dev, profile, rbcmds, entry_head, |
| 1188 | rb->timestamp + 1, drawctxt); |
| 1189 | |
| 1190 | /* set flag to sync with post ib commands */ |
| 1191 | *cmd_flags |= KGSL_CMD_FLAGS_PROFILE; |
| 1192 | |
| 1193 | done: |
| 1194 | /* write the ibdesc to the ringbuffer */ |
| 1195 | for (i = 0; i < ret; i++) |
| 1196 | *ptr++ = rbcmds[i]; |
| 1197 | |
| 1198 | *rbptr = ptr; |
| 1199 | } |
| 1200 | |
| 1201 | void adreno_profile_postib_processing(struct adreno_device *adreno_dev, |
| 1202 | unsigned int *cmd_flags, unsigned int **rbptr) |
| 1203 | { |
| 1204 | struct adreno_profile *profile = &adreno_dev->profile; |
| 1205 | int count = profile->assignment_count; |
| 1206 | unsigned int entry_head = profile->shared_head - |
| 1207 | SIZE_SHARED_ENTRY(count); |
| 1208 | unsigned int *ptr = *rbptr; |
| 1209 | unsigned int rbcmds[4]; |
| 1210 | int ret = 0, i; |
| 1211 | |
| 1212 | if (!adreno_profile_assignments_ready(profile)) |
| 1213 | goto done; |
| 1214 | |
| 1215 | if (!(*cmd_flags & KGSL_CMD_FLAGS_PROFILE)) |
| 1216 | goto done; |
| 1217 | |
| 1218 | /* create the shared ibdesc */ |
| 1219 | ret = _build_post_ib_cmds(adreno_dev, profile, rbcmds, entry_head); |
| 1220 | |
| 1221 | done: |
| 1222 | /* write the ibdesc to the ringbuffer */ |
| 1223 | for (i = 0; i < ret; i++) |
| 1224 | *ptr++ = rbcmds[i]; |
| 1225 | |
| 1226 | *rbptr = ptr; |
| 1227 | |
| 1228 | /* reset the sync flag */ |
| 1229 | *cmd_flags &= ~KGSL_CMD_FLAGS_PROFILE; |
| 1230 | } |
| 1231 | |