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 | |
| 14 | #include <linux/wait.h> |
| 15 | #include <linux/delay.h> |
| 16 | #include <linux/sched.h> |
| 17 | #include <linux/jiffies.h> |
| 18 | #include <linux/err.h> |
| 19 | |
| 20 | #include "kgsl.h" |
| 21 | #include "kgsl_sharedmem.h" |
| 22 | #include "adreno.h" |
| 23 | #include "adreno_ringbuffer.h" |
| 24 | #include "adreno_trace.h" |
| 25 | #include "kgsl_sharedmem.h" |
| 26 | |
| 27 | #define DRAWQUEUE_NEXT(_i, _s) (((_i) + 1) % (_s)) |
| 28 | |
| 29 | /* Time in ms after which the dispatcher tries to schedule an unscheduled RB */ |
| 30 | unsigned int adreno_dispatch_starvation_time = 2000; |
| 31 | |
| 32 | /* Amount of time in ms that a starved RB is permitted to execute for */ |
| 33 | unsigned int adreno_dispatch_time_slice = 25; |
| 34 | |
| 35 | /* |
| 36 | * If set then dispatcher tries to schedule lower priority RB's after if they |
| 37 | * have commands in their pipe and have been inactive for |
| 38 | * _dispatch_starvation_time. Also, once an RB is schduled it will be allowed |
| 39 | * to run for _dispatch_time_slice unless it's commands complete before |
| 40 | * _dispatch_time_slice |
| 41 | */ |
| 42 | unsigned int adreno_disp_preempt_fair_sched; |
| 43 | |
| 44 | /* Number of commands that can be queued in a context before it sleeps */ |
| 45 | static unsigned int _context_drawqueue_size = 50; |
| 46 | |
| 47 | /* Number of milliseconds to wait for the context queue to clear */ |
| 48 | static unsigned int _context_queue_wait = 10000; |
| 49 | |
| 50 | /* Number of drawobjs sent at a time from a single context */ |
| 51 | static unsigned int _context_drawobj_burst = 5; |
| 52 | |
| 53 | /* |
| 54 | * GFT throttle parameters. If GFT recovered more than |
| 55 | * X times in Y ms invalidate the context and do not attempt recovery. |
| 56 | * X -> _fault_throttle_burst |
| 57 | * Y -> _fault_throttle_time |
| 58 | */ |
| 59 | static unsigned int _fault_throttle_time = 3000; |
| 60 | static unsigned int _fault_throttle_burst = 3; |
| 61 | |
| 62 | /* |
| 63 | * Maximum ringbuffer inflight for the single submitting context case - this |
| 64 | * should be sufficiently high to keep the GPU loaded |
| 65 | */ |
| 66 | static unsigned int _dispatcher_q_inflight_hi = 15; |
| 67 | |
| 68 | /* |
| 69 | * Minimum inflight for the multiple context case - this should sufficiently low |
| 70 | * to allow for lower latency context switching |
| 71 | */ |
| 72 | static unsigned int _dispatcher_q_inflight_lo = 4; |
| 73 | |
| 74 | /* Command batch timeout (in milliseconds) */ |
| 75 | unsigned int adreno_drawobj_timeout = 2000; |
| 76 | |
| 77 | /* Interval for reading and comparing fault detection registers */ |
| 78 | static unsigned int _fault_timer_interval = 200; |
| 79 | |
| 80 | #define DRAWQUEUE_RB(_drawqueue) \ |
| 81 | ((struct adreno_ringbuffer *) \ |
| 82 | container_of((_drawqueue),\ |
| 83 | struct adreno_ringbuffer, dispatch_q)) |
| 84 | |
| 85 | #define DRAWQUEUE(_ringbuffer) (&(_ringbuffer)->dispatch_q) |
| 86 | |
| 87 | static int adreno_dispatch_retire_drawqueue(struct adreno_device *adreno_dev, |
| 88 | struct adreno_dispatcher_drawqueue *drawqueue); |
| 89 | |
| 90 | static inline bool drawqueue_is_current( |
| 91 | struct adreno_dispatcher_drawqueue *drawqueue) |
| 92 | { |
| 93 | struct adreno_ringbuffer *rb = DRAWQUEUE_RB(drawqueue); |
| 94 | struct adreno_device *adreno_dev = ADRENO_RB_DEVICE(rb); |
| 95 | |
| 96 | return (adreno_dev->cur_rb == rb); |
| 97 | } |
| 98 | |
| 99 | static void _add_context(struct adreno_device *adreno_dev, |
| 100 | struct adreno_context *drawctxt) |
| 101 | { |
| 102 | /* Remove it from the list */ |
| 103 | list_del_init(&drawctxt->active_node); |
| 104 | |
| 105 | /* And push it to the front */ |
| 106 | drawctxt->active_time = jiffies; |
| 107 | list_add(&drawctxt->active_node, &adreno_dev->active_list); |
| 108 | } |
| 109 | |
| 110 | static int __count_context(struct adreno_context *drawctxt, void *data) |
| 111 | { |
| 112 | unsigned long expires = drawctxt->active_time + msecs_to_jiffies(100); |
| 113 | |
| 114 | return time_after(jiffies, expires) ? 0 : 1; |
| 115 | } |
| 116 | |
| 117 | static int __count_drawqueue_context(struct adreno_context *drawctxt, |
| 118 | void *data) |
| 119 | { |
| 120 | unsigned long expires = drawctxt->active_time + msecs_to_jiffies(100); |
| 121 | |
| 122 | if (time_after(jiffies, expires)) |
| 123 | return 0; |
| 124 | |
| 125 | return (&drawctxt->rb->dispatch_q == |
| 126 | (struct adreno_dispatcher_drawqueue *) data) ? 1 : 0; |
| 127 | } |
| 128 | |
| 129 | static int _adreno_count_active_contexts(struct adreno_device *adreno_dev, |
| 130 | int (*func)(struct adreno_context *, void *), void *data) |
| 131 | { |
| 132 | struct adreno_context *ctxt; |
| 133 | int count = 0; |
| 134 | |
| 135 | list_for_each_entry(ctxt, &adreno_dev->active_list, active_node) { |
| 136 | if (func(ctxt, data) == 0) |
| 137 | return count; |
| 138 | |
| 139 | count++; |
| 140 | } |
| 141 | |
| 142 | return count; |
| 143 | } |
| 144 | |
| 145 | static void _track_context(struct adreno_device *adreno_dev, |
| 146 | struct adreno_dispatcher_drawqueue *drawqueue, |
| 147 | struct adreno_context *drawctxt) |
| 148 | { |
| 149 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 150 | |
| 151 | spin_lock(&adreno_dev->active_list_lock); |
| 152 | |
| 153 | _add_context(adreno_dev, drawctxt); |
| 154 | |
| 155 | device->active_context_count = |
| 156 | _adreno_count_active_contexts(adreno_dev, |
| 157 | __count_context, NULL); |
| 158 | drawqueue->active_context_count = |
| 159 | _adreno_count_active_contexts(adreno_dev, |
| 160 | __count_drawqueue_context, drawqueue); |
| 161 | |
| 162 | spin_unlock(&adreno_dev->active_list_lock); |
| 163 | } |
| 164 | |
| 165 | /* |
| 166 | * If only one context has queued in the last 100 milliseconds increase |
| 167 | * inflight to a high number to load up the GPU. If multiple contexts |
| 168 | * have queued drop the inflight for better context switch latency. |
| 169 | * If no contexts have queued what are you even doing here? |
| 170 | */ |
| 171 | |
| 172 | static inline int |
| 173 | _drawqueue_inflight(struct adreno_dispatcher_drawqueue *drawqueue) |
| 174 | { |
| 175 | return (drawqueue->active_context_count > 1) |
| 176 | ? _dispatcher_q_inflight_lo : _dispatcher_q_inflight_hi; |
| 177 | } |
| 178 | |
| 179 | static void fault_detect_read(struct adreno_device *adreno_dev) |
| 180 | { |
| 181 | int i; |
| 182 | |
| 183 | if (!test_bit(ADRENO_DEVICE_SOFT_FAULT_DETECT, &adreno_dev->priv)) |
| 184 | return; |
| 185 | |
| 186 | for (i = 0; i < adreno_dev->num_ringbuffers; i++) { |
| 187 | struct adreno_ringbuffer *rb = &(adreno_dev->ringbuffers[i]); |
| 188 | |
| 189 | adreno_rb_readtimestamp(adreno_dev, rb, |
| 190 | KGSL_TIMESTAMP_RETIRED, &(rb->fault_detect_ts)); |
| 191 | } |
| 192 | |
| 193 | for (i = 0; i < adreno_ft_regs_num; i++) { |
| 194 | if (adreno_ft_regs[i] != 0) |
| 195 | kgsl_regread(KGSL_DEVICE(adreno_dev), adreno_ft_regs[i], |
| 196 | &adreno_ft_regs_val[i]); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | /* |
| 201 | * Check to see if the device is idle |
| 202 | */ |
| 203 | static inline bool _isidle(struct adreno_device *adreno_dev) |
| 204 | { |
| 205 | const struct adreno_gpu_core *gpucore = adreno_dev->gpucore; |
| 206 | unsigned int reg_rbbm_status; |
| 207 | |
| 208 | if (!kgsl_state_is_awake(KGSL_DEVICE(adreno_dev))) |
| 209 | goto ret; |
| 210 | |
Hareesh Gundu | a2fe6ec | 2017-03-06 14:53:36 +0530 | [diff] [blame] | 211 | if (adreno_rb_empty(adreno_dev->cur_rb)) |
| 212 | goto ret; |
| 213 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 214 | /* only check rbbm status to determine if GPU is idle */ |
| 215 | adreno_readreg(adreno_dev, ADRENO_REG_RBBM_STATUS, ®_rbbm_status); |
| 216 | |
| 217 | if (reg_rbbm_status & gpucore->busy_mask) |
| 218 | return false; |
| 219 | |
| 220 | ret: |
| 221 | /* Clear the existing register values */ |
| 222 | memset(adreno_ft_regs_val, 0, |
| 223 | adreno_ft_regs_num * sizeof(unsigned int)); |
| 224 | |
| 225 | return true; |
| 226 | } |
| 227 | |
| 228 | /** |
| 229 | * fault_detect_read_compare() - Read the fault detect registers and compare |
| 230 | * them to the current value |
| 231 | * @device: Pointer to the KGSL device struct |
| 232 | * |
| 233 | * Read the set of fault detect registers and compare them to the current set |
| 234 | * of registers. Return 1 if any of the register values changed. Also, compare |
| 235 | * if the current RB's timstamp has changed or not. |
| 236 | */ |
| 237 | static int fault_detect_read_compare(struct adreno_device *adreno_dev) |
| 238 | { |
| 239 | struct adreno_ringbuffer *rb = ADRENO_CURRENT_RINGBUFFER(adreno_dev); |
| 240 | int i, ret = 0; |
| 241 | unsigned int ts; |
| 242 | |
Shrenuj Bansal | ae67281 | 2016-02-24 14:17:30 -0800 | [diff] [blame] | 243 | if (!test_bit(ADRENO_DEVICE_SOFT_FAULT_DETECT, &adreno_dev->priv)) |
| 244 | return 1; |
| 245 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 246 | /* Check to see if the device is idle - if so report no hang */ |
| 247 | if (_isidle(adreno_dev) == true) |
| 248 | ret = 1; |
| 249 | |
| 250 | for (i = 0; i < adreno_ft_regs_num; i++) { |
| 251 | unsigned int val; |
| 252 | |
| 253 | if (adreno_ft_regs[i] == 0) |
| 254 | continue; |
| 255 | kgsl_regread(KGSL_DEVICE(adreno_dev), adreno_ft_regs[i], &val); |
| 256 | if (val != adreno_ft_regs_val[i]) |
| 257 | ret = 1; |
| 258 | adreno_ft_regs_val[i] = val; |
| 259 | } |
| 260 | |
| 261 | if (!adreno_rb_readtimestamp(adreno_dev, adreno_dev->cur_rb, |
| 262 | KGSL_TIMESTAMP_RETIRED, &ts)) { |
| 263 | if (ts != rb->fault_detect_ts) |
| 264 | ret = 1; |
| 265 | |
| 266 | rb->fault_detect_ts = ts; |
| 267 | } |
| 268 | |
| 269 | return ret; |
| 270 | } |
| 271 | |
| 272 | static void start_fault_timer(struct adreno_device *adreno_dev) |
| 273 | { |
| 274 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 275 | |
| 276 | if (adreno_soft_fault_detect(adreno_dev)) |
| 277 | mod_timer(&dispatcher->fault_timer, |
| 278 | jiffies + msecs_to_jiffies(_fault_timer_interval)); |
| 279 | } |
| 280 | |
| 281 | /** |
| 282 | * _retire_timestamp() - Retire object without sending it |
| 283 | * to the hardware |
| 284 | * @drawobj: Pointer to the object to retire |
| 285 | * |
| 286 | * In some cases ibs can be retired by the software |
| 287 | * without going to the GPU. In those cases, update the |
| 288 | * memstore from the CPU, kick off the event engine to handle |
| 289 | * expired events and destroy the ib. |
| 290 | */ |
| 291 | static void _retire_timestamp(struct kgsl_drawobj *drawobj) |
| 292 | { |
| 293 | struct kgsl_context *context = drawobj->context; |
| 294 | struct adreno_context *drawctxt = ADRENO_CONTEXT(context); |
| 295 | struct kgsl_device *device = context->device; |
| 296 | |
| 297 | /* |
| 298 | * Write the start and end timestamp to the memstore to keep the |
| 299 | * accounting sane |
| 300 | */ |
| 301 | kgsl_sharedmem_writel(device, &device->memstore, |
| 302 | KGSL_MEMSTORE_OFFSET(context->id, soptimestamp), |
| 303 | drawobj->timestamp); |
| 304 | |
| 305 | kgsl_sharedmem_writel(device, &device->memstore, |
| 306 | KGSL_MEMSTORE_OFFSET(context->id, eoptimestamp), |
| 307 | drawobj->timestamp); |
| 308 | |
| 309 | |
| 310 | /* Retire pending GPU events for the object */ |
| 311 | kgsl_process_event_group(device, &context->events); |
| 312 | |
| 313 | /* |
| 314 | * For A3xx we still get the rptr from the CP_RB_RPTR instead of |
| 315 | * rptr scratch out address. At this point GPU clocks turned off. |
| 316 | * So avoid reading GPU register directly for A3xx. |
| 317 | */ |
| 318 | if (adreno_is_a3xx(ADRENO_DEVICE(device))) |
| 319 | trace_adreno_cmdbatch_retired(drawobj, -1, 0, 0, drawctxt->rb, |
| 320 | 0, 0); |
| 321 | else |
| 322 | trace_adreno_cmdbatch_retired(drawobj, -1, 0, 0, drawctxt->rb, |
| 323 | adreno_get_rptr(drawctxt->rb), 0); |
| 324 | kgsl_drawobj_destroy(drawobj); |
| 325 | } |
| 326 | |
| 327 | static int _check_context_queue(struct adreno_context *drawctxt) |
| 328 | { |
| 329 | int ret; |
| 330 | |
| 331 | spin_lock(&drawctxt->lock); |
| 332 | |
| 333 | /* |
| 334 | * Wake up if there is room in the context or if the whole thing got |
| 335 | * invalidated while we were asleep |
| 336 | */ |
| 337 | |
| 338 | if (kgsl_context_invalid(&drawctxt->base)) |
| 339 | ret = 1; |
| 340 | else |
| 341 | ret = drawctxt->queued < _context_drawqueue_size ? 1 : 0; |
| 342 | |
| 343 | spin_unlock(&drawctxt->lock); |
| 344 | |
| 345 | return ret; |
| 346 | } |
| 347 | |
| 348 | /* |
| 349 | * return true if this is a marker command and the dependent timestamp has |
| 350 | * retired |
| 351 | */ |
| 352 | static bool _marker_expired(struct kgsl_drawobj_cmd *markerobj) |
| 353 | { |
| 354 | struct kgsl_drawobj *drawobj = DRAWOBJ(markerobj); |
| 355 | |
| 356 | return (drawobj->flags & KGSL_DRAWOBJ_MARKER) && |
| 357 | kgsl_check_timestamp(drawobj->device, drawobj->context, |
| 358 | markerobj->marker_timestamp); |
| 359 | } |
| 360 | |
| 361 | static inline void _pop_drawobj(struct adreno_context *drawctxt) |
| 362 | { |
| 363 | drawctxt->drawqueue_head = DRAWQUEUE_NEXT(drawctxt->drawqueue_head, |
| 364 | ADRENO_CONTEXT_DRAWQUEUE_SIZE); |
| 365 | drawctxt->queued--; |
| 366 | } |
| 367 | |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 368 | static void _retire_sparseobj(struct kgsl_drawobj_sparse *sparseobj, |
| 369 | struct adreno_context *drawctxt) |
| 370 | { |
| 371 | kgsl_sparse_bind(drawctxt->base.proc_priv, sparseobj); |
| 372 | _retire_timestamp(DRAWOBJ(sparseobj)); |
| 373 | } |
| 374 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 375 | static int _retire_markerobj(struct kgsl_drawobj_cmd *cmdobj, |
| 376 | struct adreno_context *drawctxt) |
| 377 | { |
| 378 | if (_marker_expired(cmdobj)) { |
| 379 | _pop_drawobj(drawctxt); |
| 380 | _retire_timestamp(DRAWOBJ(cmdobj)); |
| 381 | return 0; |
| 382 | } |
| 383 | |
| 384 | /* |
| 385 | * If the marker isn't expired but the SKIP bit |
| 386 | * is set then there are real commands following |
| 387 | * this one in the queue. This means that we |
| 388 | * need to dispatch the command so that we can |
| 389 | * keep the timestamp accounting correct. If |
| 390 | * skip isn't set then we block this queue |
| 391 | * until the dependent timestamp expires |
| 392 | */ |
| 393 | return test_bit(CMDOBJ_SKIP, &cmdobj->priv) ? 1 : -EAGAIN; |
| 394 | } |
| 395 | |
| 396 | static int _retire_syncobj(struct kgsl_drawobj_sync *syncobj, |
| 397 | struct adreno_context *drawctxt) |
| 398 | { |
| 399 | if (!kgsl_drawobj_events_pending(syncobj)) { |
| 400 | _pop_drawobj(drawctxt); |
| 401 | kgsl_drawobj_destroy(DRAWOBJ(syncobj)); |
| 402 | return 0; |
| 403 | } |
| 404 | |
| 405 | /* |
| 406 | * If we got here, there are pending events for sync object. |
| 407 | * Start the canary timer if it hasnt been started already. |
| 408 | */ |
| 409 | if (!syncobj->timeout_jiffies) { |
| 410 | syncobj->timeout_jiffies = jiffies + msecs_to_jiffies(5000); |
| 411 | mod_timer(&syncobj->timer, syncobj->timeout_jiffies); |
| 412 | } |
| 413 | |
| 414 | return -EAGAIN; |
| 415 | } |
| 416 | |
| 417 | /* |
| 418 | * Retires all expired marker and sync objs from the context |
| 419 | * queue and returns one of the below |
| 420 | * a) next drawobj that needs to be sent to ringbuffer |
| 421 | * b) -EAGAIN for syncobj with syncpoints pending. |
| 422 | * c) -EAGAIN for markerobj whose marker timestamp has not expired yet. |
| 423 | * c) NULL for no commands remaining in drawqueue. |
| 424 | */ |
| 425 | static struct kgsl_drawobj *_process_drawqueue_get_next_drawobj( |
| 426 | struct adreno_context *drawctxt) |
| 427 | { |
| 428 | struct kgsl_drawobj *drawobj; |
| 429 | unsigned int i = drawctxt->drawqueue_head; |
| 430 | int ret = 0; |
| 431 | |
| 432 | if (drawctxt->drawqueue_head == drawctxt->drawqueue_tail) |
| 433 | return NULL; |
| 434 | |
| 435 | for (i = drawctxt->drawqueue_head; i != drawctxt->drawqueue_tail; |
| 436 | i = DRAWQUEUE_NEXT(i, ADRENO_CONTEXT_DRAWQUEUE_SIZE)) { |
| 437 | |
| 438 | drawobj = drawctxt->drawqueue[i]; |
| 439 | |
| 440 | if (drawobj == NULL) |
| 441 | return NULL; |
| 442 | |
| 443 | if (drawobj->type == CMDOBJ_TYPE) |
| 444 | return drawobj; |
| 445 | else if (drawobj->type == MARKEROBJ_TYPE) { |
| 446 | ret = _retire_markerobj(CMDOBJ(drawobj), drawctxt); |
| 447 | /* Special case where marker needs to be sent to GPU */ |
| 448 | if (ret == 1) |
| 449 | return drawobj; |
| 450 | } else if (drawobj->type == SYNCOBJ_TYPE) |
| 451 | ret = _retire_syncobj(SYNCOBJ(drawobj), drawctxt); |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 452 | else |
| 453 | return ERR_PTR(-EINVAL); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 454 | |
| 455 | if (ret == -EAGAIN) |
| 456 | return ERR_PTR(-EAGAIN); |
| 457 | |
| 458 | continue; |
| 459 | } |
| 460 | |
| 461 | return NULL; |
| 462 | } |
| 463 | |
| 464 | /** |
| 465 | * adreno_dispatcher_requeue_cmdobj() - Put a command back on the context |
| 466 | * queue |
| 467 | * @drawctxt: Pointer to the adreno draw context |
| 468 | * @cmdobj: Pointer to the KGSL command object to requeue |
| 469 | * |
| 470 | * Failure to submit a command to the ringbuffer isn't the fault of the command |
| 471 | * being submitted so if a failure happens, push it back on the head of the the |
| 472 | * context queue to be reconsidered again unless the context got detached. |
| 473 | */ |
| 474 | static inline int adreno_dispatcher_requeue_cmdobj( |
| 475 | struct adreno_context *drawctxt, |
| 476 | struct kgsl_drawobj_cmd *cmdobj) |
| 477 | { |
| 478 | unsigned int prev; |
| 479 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 480 | |
| 481 | spin_lock(&drawctxt->lock); |
| 482 | |
| 483 | if (kgsl_context_detached(&drawctxt->base) || |
| 484 | kgsl_context_invalid(&drawctxt->base)) { |
| 485 | spin_unlock(&drawctxt->lock); |
| 486 | /* get rid of this drawobj since the context is bad */ |
| 487 | kgsl_drawobj_destroy(drawobj); |
| 488 | return -ENOENT; |
| 489 | } |
| 490 | |
| 491 | prev = drawctxt->drawqueue_head == 0 ? |
| 492 | (ADRENO_CONTEXT_DRAWQUEUE_SIZE - 1) : |
| 493 | (drawctxt->drawqueue_head - 1); |
| 494 | |
| 495 | /* |
| 496 | * The maximum queue size always needs to be one less then the size of |
| 497 | * the ringbuffer queue so there is "room" to put the drawobj back in |
| 498 | */ |
| 499 | |
| 500 | WARN_ON(prev == drawctxt->drawqueue_tail); |
| 501 | |
| 502 | drawctxt->drawqueue[prev] = drawobj; |
| 503 | drawctxt->queued++; |
| 504 | |
| 505 | /* Reset the command queue head to reflect the newly requeued change */ |
| 506 | drawctxt->drawqueue_head = prev; |
| 507 | spin_unlock(&drawctxt->lock); |
| 508 | return 0; |
| 509 | } |
| 510 | |
| 511 | /** |
| 512 | * dispatcher_queue_context() - Queue a context in the dispatcher pending list |
| 513 | * @dispatcher: Pointer to the adreno dispatcher struct |
| 514 | * @drawctxt: Pointer to the adreno draw context |
| 515 | * |
| 516 | * Add a context to the dispatcher pending list. |
| 517 | */ |
| 518 | static void dispatcher_queue_context(struct adreno_device *adreno_dev, |
| 519 | struct adreno_context *drawctxt) |
| 520 | { |
| 521 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 522 | |
| 523 | /* Refuse to queue a detached context */ |
| 524 | if (kgsl_context_detached(&drawctxt->base)) |
| 525 | return; |
| 526 | |
| 527 | spin_lock(&dispatcher->plist_lock); |
| 528 | |
| 529 | if (plist_node_empty(&drawctxt->pending)) { |
| 530 | /* Get a reference to the context while it sits on the list */ |
| 531 | if (_kgsl_context_get(&drawctxt->base)) { |
| 532 | trace_dispatch_queue_context(drawctxt); |
| 533 | plist_add(&drawctxt->pending, &dispatcher->pending); |
| 534 | } |
| 535 | } |
| 536 | |
| 537 | spin_unlock(&dispatcher->plist_lock); |
| 538 | } |
| 539 | |
| 540 | /** |
| 541 | * sendcmd() - Send a drawobj to the GPU hardware |
| 542 | * @dispatcher: Pointer to the adreno dispatcher struct |
| 543 | * @drawobj: Pointer to the KGSL drawobj being sent |
| 544 | * |
| 545 | * Send a KGSL drawobj to the GPU hardware |
| 546 | */ |
| 547 | static int sendcmd(struct adreno_device *adreno_dev, |
| 548 | struct kgsl_drawobj_cmd *cmdobj) |
| 549 | { |
| 550 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 551 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 552 | struct adreno_gpudev *gpudev = ADRENO_GPU_DEVICE(adreno_dev); |
| 553 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 554 | struct adreno_context *drawctxt = ADRENO_CONTEXT(drawobj->context); |
| 555 | struct adreno_dispatcher_drawqueue *dispatch_q = |
| 556 | ADRENO_DRAWOBJ_DISPATCH_DRAWQUEUE(drawobj); |
| 557 | struct adreno_submit_time time; |
| 558 | uint64_t secs = 0; |
| 559 | unsigned long nsecs = 0; |
| 560 | int ret; |
| 561 | |
| 562 | mutex_lock(&device->mutex); |
| 563 | if (adreno_gpu_halt(adreno_dev) != 0) { |
| 564 | mutex_unlock(&device->mutex); |
| 565 | return -EBUSY; |
| 566 | } |
| 567 | |
| 568 | dispatcher->inflight++; |
| 569 | dispatch_q->inflight++; |
| 570 | |
| 571 | if (dispatcher->inflight == 1 && |
| 572 | !test_bit(ADRENO_DISPATCHER_POWER, &dispatcher->priv)) { |
| 573 | /* Time to make the donuts. Turn on the GPU */ |
| 574 | ret = kgsl_active_count_get(device); |
| 575 | if (ret) { |
| 576 | dispatcher->inflight--; |
| 577 | dispatch_q->inflight--; |
| 578 | mutex_unlock(&device->mutex); |
| 579 | return ret; |
| 580 | } |
| 581 | |
| 582 | set_bit(ADRENO_DISPATCHER_POWER, &dispatcher->priv); |
| 583 | } |
| 584 | |
| 585 | if (test_bit(ADRENO_DEVICE_DRAWOBJ_PROFILE, &adreno_dev->priv)) { |
| 586 | set_bit(CMDOBJ_PROFILE, &cmdobj->priv); |
| 587 | cmdobj->profile_index = adreno_dev->profile_index; |
| 588 | adreno_dev->profile_index = |
| 589 | (adreno_dev->profile_index + 1) % |
| 590 | ADRENO_DRAWOBJ_PROFILE_COUNT; |
| 591 | } |
| 592 | |
| 593 | ret = adreno_ringbuffer_submitcmd(adreno_dev, cmdobj, &time); |
| 594 | |
| 595 | /* |
| 596 | * On the first command, if the submission was successful, then read the |
| 597 | * fault registers. If it failed then turn off the GPU. Sad face. |
| 598 | */ |
| 599 | |
| 600 | if (dispatcher->inflight == 1) { |
| 601 | if (ret == 0) { |
| 602 | |
| 603 | /* Stop fault timer before reading fault registers */ |
| 604 | del_timer_sync(&dispatcher->fault_timer); |
| 605 | |
| 606 | fault_detect_read(adreno_dev); |
| 607 | |
| 608 | /* Start the fault timer on first submission */ |
| 609 | start_fault_timer(adreno_dev); |
| 610 | |
| 611 | if (!test_and_set_bit(ADRENO_DISPATCHER_ACTIVE, |
| 612 | &dispatcher->priv)) |
| 613 | reinit_completion(&dispatcher->idle_gate); |
Prakash Kamliya | c9b7da0 | 2016-12-20 17:45:41 +0530 | [diff] [blame] | 614 | |
| 615 | /* |
| 616 | * We update power stats generally at the expire of |
| 617 | * cmdbatch. In cases where the cmdbatch takes a long |
| 618 | * time to finish, it will delay power stats update, |
| 619 | * in effect it will delay DCVS decision. Start a |
| 620 | * timer to update power state on expire of this timer. |
| 621 | */ |
| 622 | kgsl_pwrscale_midframe_timer_restart(device); |
| 623 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 624 | } else { |
| 625 | kgsl_active_count_put(device); |
| 626 | clear_bit(ADRENO_DISPATCHER_POWER, &dispatcher->priv); |
| 627 | } |
| 628 | } |
| 629 | |
| 630 | |
| 631 | if (ret) { |
| 632 | dispatcher->inflight--; |
| 633 | dispatch_q->inflight--; |
| 634 | |
| 635 | mutex_unlock(&device->mutex); |
| 636 | |
| 637 | /* |
| 638 | * Don't log a message in case of: |
| 639 | * -ENOENT means that the context was detached before the |
| 640 | * command was submitted |
| 641 | * -ENOSPC means that there temporarily isn't any room in the |
| 642 | * ringbuffer |
| 643 | * -PROTO means that a fault is currently being worked |
| 644 | */ |
| 645 | |
| 646 | if (ret != -ENOENT && ret != -ENOSPC && ret != -EPROTO) |
| 647 | KGSL_DRV_ERR(device, |
| 648 | "Unable to submit command to the ringbuffer %d\n", |
| 649 | ret); |
| 650 | return ret; |
| 651 | } |
| 652 | |
| 653 | secs = time.ktime; |
| 654 | nsecs = do_div(secs, 1000000000); |
| 655 | |
| 656 | trace_adreno_cmdbatch_submitted(drawobj, (int) dispatcher->inflight, |
| 657 | time.ticks, (unsigned long) secs, nsecs / 1000, drawctxt->rb, |
| 658 | adreno_get_rptr(drawctxt->rb)); |
| 659 | |
| 660 | mutex_unlock(&device->mutex); |
| 661 | |
| 662 | cmdobj->submit_ticks = time.ticks; |
| 663 | |
| 664 | dispatch_q->cmd_q[dispatch_q->tail] = cmdobj; |
| 665 | dispatch_q->tail = (dispatch_q->tail + 1) % |
| 666 | ADRENO_DISPATCH_DRAWQUEUE_SIZE; |
| 667 | |
| 668 | /* |
| 669 | * For the first submission in any given command queue update the |
| 670 | * expected expire time - this won't actually be used / updated until |
| 671 | * the command queue in question goes current, but universally setting |
| 672 | * it here avoids the possibilty of some race conditions with preempt |
| 673 | */ |
| 674 | |
| 675 | if (dispatch_q->inflight == 1) |
| 676 | dispatch_q->expires = jiffies + |
| 677 | msecs_to_jiffies(adreno_drawobj_timeout); |
| 678 | |
| 679 | /* |
| 680 | * If we believe ourselves to be current and preemption isn't a thing, |
| 681 | * then set up the timer. If this misses, then preemption is indeed a |
| 682 | * thing and the timer will be set up in due time |
| 683 | */ |
| 684 | if (!adreno_in_preempt_state(adreno_dev, ADRENO_PREEMPT_NONE)) { |
| 685 | if (drawqueue_is_current(dispatch_q)) |
| 686 | mod_timer(&dispatcher->timer, dispatch_q->expires); |
| 687 | } |
| 688 | |
| 689 | |
| 690 | /* |
| 691 | * we just submitted something, readjust ringbuffer |
| 692 | * execution level |
| 693 | */ |
| 694 | if (gpudev->preemption_schedule) |
| 695 | gpudev->preemption_schedule(adreno_dev); |
| 696 | return 0; |
| 697 | } |
| 698 | |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 699 | |
| 700 | /* |
| 701 | * Retires all sync objs from the sparse context |
| 702 | * queue and returns one of the below |
| 703 | * a) next sparseobj |
| 704 | * b) -EAGAIN for syncobj with syncpoints pending |
| 705 | * c) -EINVAL for unexpected drawobj |
| 706 | * d) NULL for no sparseobj |
| 707 | */ |
| 708 | static struct kgsl_drawobj_sparse *_get_next_sparseobj( |
| 709 | struct adreno_context *drawctxt) |
| 710 | { |
| 711 | struct kgsl_drawobj *drawobj; |
| 712 | unsigned int i = drawctxt->drawqueue_head; |
| 713 | int ret = 0; |
| 714 | |
| 715 | if (drawctxt->drawqueue_head == drawctxt->drawqueue_tail) |
| 716 | return NULL; |
| 717 | |
| 718 | for (i = drawctxt->drawqueue_head; i != drawctxt->drawqueue_tail; |
| 719 | i = DRAWQUEUE_NEXT(i, ADRENO_CONTEXT_DRAWQUEUE_SIZE)) { |
| 720 | |
| 721 | drawobj = drawctxt->drawqueue[i]; |
| 722 | |
| 723 | if (drawobj == NULL) |
| 724 | return NULL; |
| 725 | |
| 726 | if (drawobj->type == SYNCOBJ_TYPE) |
| 727 | ret = _retire_syncobj(SYNCOBJ(drawobj), drawctxt); |
| 728 | else if (drawobj->type == SPARSEOBJ_TYPE) |
| 729 | return SPARSEOBJ(drawobj); |
| 730 | else |
| 731 | return ERR_PTR(-EINVAL); |
| 732 | |
| 733 | if (ret == -EAGAIN) |
| 734 | return ERR_PTR(-EAGAIN); |
| 735 | |
| 736 | continue; |
| 737 | } |
| 738 | |
| 739 | return NULL; |
| 740 | } |
| 741 | |
| 742 | static int _process_drawqueue_sparse( |
| 743 | struct adreno_context *drawctxt) |
| 744 | { |
| 745 | struct kgsl_drawobj_sparse *sparseobj; |
| 746 | int ret = 0; |
| 747 | unsigned int i; |
| 748 | |
| 749 | for (i = 0; i < ADRENO_CONTEXT_DRAWQUEUE_SIZE; i++) { |
| 750 | |
| 751 | spin_lock(&drawctxt->lock); |
| 752 | sparseobj = _get_next_sparseobj(drawctxt); |
| 753 | if (IS_ERR_OR_NULL(sparseobj)) { |
| 754 | if (IS_ERR(sparseobj)) |
| 755 | ret = PTR_ERR(sparseobj); |
| 756 | spin_unlock(&drawctxt->lock); |
| 757 | return ret; |
| 758 | } |
| 759 | |
| 760 | _pop_drawobj(drawctxt); |
| 761 | spin_unlock(&drawctxt->lock); |
| 762 | |
| 763 | _retire_sparseobj(sparseobj, drawctxt); |
| 764 | } |
| 765 | |
| 766 | return 0; |
| 767 | } |
| 768 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 769 | /** |
| 770 | * dispatcher_context_sendcmds() - Send commands from a context to the GPU |
| 771 | * @adreno_dev: Pointer to the adreno device struct |
| 772 | * @drawctxt: Pointer to the adreno context to dispatch commands from |
| 773 | * |
| 774 | * Dequeue and send a burst of commands from the specified context to the GPU |
| 775 | * Returns postive if the context needs to be put back on the pending queue |
| 776 | * 0 if the context is empty or detached and negative on error |
| 777 | */ |
| 778 | static int dispatcher_context_sendcmds(struct adreno_device *adreno_dev, |
| 779 | struct adreno_context *drawctxt) |
| 780 | { |
| 781 | struct adreno_dispatcher_drawqueue *dispatch_q = |
| 782 | &(drawctxt->rb->dispatch_q); |
| 783 | int count = 0; |
| 784 | int ret = 0; |
| 785 | int inflight = _drawqueue_inflight(dispatch_q); |
| 786 | unsigned int timestamp; |
| 787 | |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 788 | if (drawctxt->base.flags & KGSL_CONTEXT_SPARSE) |
| 789 | return _process_drawqueue_sparse(drawctxt); |
| 790 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 791 | if (dispatch_q->inflight >= inflight) { |
| 792 | spin_lock(&drawctxt->lock); |
| 793 | _process_drawqueue_get_next_drawobj(drawctxt); |
| 794 | spin_unlock(&drawctxt->lock); |
| 795 | return -EBUSY; |
| 796 | } |
| 797 | |
| 798 | /* |
| 799 | * Each context can send a specific number of drawobjs per cycle |
| 800 | */ |
| 801 | while ((count < _context_drawobj_burst) && |
| 802 | (dispatch_q->inflight < inflight)) { |
| 803 | struct kgsl_drawobj *drawobj; |
| 804 | struct kgsl_drawobj_cmd *cmdobj; |
| 805 | |
| 806 | if (adreno_gpu_fault(adreno_dev) != 0) |
| 807 | break; |
| 808 | |
| 809 | spin_lock(&drawctxt->lock); |
| 810 | drawobj = _process_drawqueue_get_next_drawobj(drawctxt); |
| 811 | |
| 812 | /* |
| 813 | * adreno_context_get_drawobj returns -EAGAIN if the current |
| 814 | * drawobj has pending sync points so no more to do here. |
| 815 | * When the sync points are satisfied then the context will get |
| 816 | * reqeueued |
| 817 | */ |
| 818 | |
| 819 | if (IS_ERR_OR_NULL(drawobj)) { |
| 820 | if (IS_ERR(drawobj)) |
| 821 | ret = PTR_ERR(drawobj); |
| 822 | spin_unlock(&drawctxt->lock); |
| 823 | break; |
| 824 | } |
| 825 | _pop_drawobj(drawctxt); |
| 826 | spin_unlock(&drawctxt->lock); |
| 827 | |
| 828 | timestamp = drawobj->timestamp; |
| 829 | cmdobj = CMDOBJ(drawobj); |
| 830 | ret = sendcmd(adreno_dev, cmdobj); |
| 831 | |
| 832 | /* |
| 833 | * On error from sendcmd() try to requeue the cmdobj |
| 834 | * unless we got back -ENOENT which means that the context has |
| 835 | * been detached and there will be no more deliveries from here |
| 836 | */ |
| 837 | if (ret != 0) { |
| 838 | /* Destroy the cmdobj on -ENOENT */ |
| 839 | if (ret == -ENOENT) |
| 840 | kgsl_drawobj_destroy(drawobj); |
| 841 | else { |
| 842 | /* |
| 843 | * If the requeue returns an error, return that |
| 844 | * instead of whatever sendcmd() sent us |
| 845 | */ |
| 846 | int r = adreno_dispatcher_requeue_cmdobj( |
| 847 | drawctxt, cmdobj); |
| 848 | if (r) |
| 849 | ret = r; |
| 850 | } |
| 851 | |
| 852 | break; |
| 853 | } |
| 854 | |
| 855 | drawctxt->submitted_timestamp = timestamp; |
| 856 | |
| 857 | count++; |
| 858 | } |
| 859 | |
| 860 | /* |
| 861 | * Wake up any snoozing threads if we have consumed any real commands |
| 862 | * or marker commands and we have room in the context queue. |
| 863 | */ |
| 864 | |
| 865 | if (_check_context_queue(drawctxt)) |
| 866 | wake_up_all(&drawctxt->wq); |
| 867 | |
| 868 | if (!ret) |
| 869 | ret = count; |
| 870 | |
| 871 | /* Return error or the number of commands queued */ |
| 872 | return ret; |
| 873 | } |
| 874 | |
| 875 | /** |
| 876 | * _adreno_dispatcher_issuecmds() - Issue commmands from pending contexts |
| 877 | * @adreno_dev: Pointer to the adreno device struct |
| 878 | * |
| 879 | * Issue as many commands as possible (up to inflight) from the pending contexts |
| 880 | * This function assumes the dispatcher mutex has been locked. |
| 881 | */ |
| 882 | static void _adreno_dispatcher_issuecmds(struct adreno_device *adreno_dev) |
| 883 | { |
| 884 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 885 | struct adreno_context *drawctxt, *next; |
| 886 | struct plist_head requeue, busy_list; |
| 887 | int ret; |
| 888 | |
| 889 | /* Leave early if the dispatcher isn't in a happy state */ |
| 890 | if (adreno_gpu_fault(adreno_dev) != 0) |
| 891 | return; |
| 892 | |
| 893 | plist_head_init(&requeue); |
| 894 | plist_head_init(&busy_list); |
| 895 | |
| 896 | /* Try to fill the ringbuffers as much as possible */ |
| 897 | while (1) { |
| 898 | |
| 899 | /* Stop doing things if the dispatcher is paused or faulted */ |
| 900 | if (adreno_gpu_fault(adreno_dev) != 0) |
| 901 | break; |
| 902 | |
| 903 | if (adreno_gpu_halt(adreno_dev) != 0) |
| 904 | break; |
| 905 | |
| 906 | spin_lock(&dispatcher->plist_lock); |
| 907 | |
| 908 | if (plist_head_empty(&dispatcher->pending)) { |
| 909 | spin_unlock(&dispatcher->plist_lock); |
| 910 | break; |
| 911 | } |
| 912 | |
| 913 | /* Get the next entry on the list */ |
| 914 | drawctxt = plist_first_entry(&dispatcher->pending, |
| 915 | struct adreno_context, pending); |
| 916 | |
| 917 | plist_del(&drawctxt->pending, &dispatcher->pending); |
| 918 | |
| 919 | spin_unlock(&dispatcher->plist_lock); |
| 920 | |
| 921 | if (kgsl_context_detached(&drawctxt->base) || |
| 922 | kgsl_context_invalid(&drawctxt->base)) { |
| 923 | kgsl_context_put(&drawctxt->base); |
| 924 | continue; |
| 925 | } |
| 926 | |
| 927 | ret = dispatcher_context_sendcmds(adreno_dev, drawctxt); |
| 928 | |
| 929 | /* Don't bother requeuing on -ENOENT - context is detached */ |
| 930 | if (ret != 0 && ret != -ENOENT) { |
| 931 | spin_lock(&dispatcher->plist_lock); |
| 932 | |
| 933 | /* |
| 934 | * Check to seen if the context had been requeued while |
| 935 | * we were processing it (probably by another thread |
| 936 | * pushing commands). If it has then shift it to the |
| 937 | * requeue list if it was not able to submit commands |
| 938 | * due to the dispatch_q being full. Also, do a put to |
| 939 | * make sure the reference counting stays accurate. |
| 940 | * If the node is empty then we will put it on the |
| 941 | * requeue list and not touch the refcount since we |
| 942 | * already hold it from the first time it went on the |
| 943 | * list. |
| 944 | */ |
| 945 | |
| 946 | if (!plist_node_empty(&drawctxt->pending)) { |
| 947 | plist_del(&drawctxt->pending, |
| 948 | &dispatcher->pending); |
| 949 | kgsl_context_put(&drawctxt->base); |
| 950 | } |
| 951 | |
| 952 | if (ret == -EBUSY) |
| 953 | /* Inflight queue is full */ |
| 954 | plist_add(&drawctxt->pending, &busy_list); |
| 955 | else |
| 956 | plist_add(&drawctxt->pending, &requeue); |
| 957 | |
| 958 | spin_unlock(&dispatcher->plist_lock); |
| 959 | } else { |
| 960 | /* |
| 961 | * If the context doesn't need be requeued put back the |
| 962 | * refcount |
| 963 | */ |
| 964 | |
| 965 | kgsl_context_put(&drawctxt->base); |
| 966 | } |
| 967 | } |
| 968 | |
| 969 | spin_lock(&dispatcher->plist_lock); |
| 970 | |
| 971 | /* Put the contexts that couldn't submit back on the pending list */ |
| 972 | plist_for_each_entry_safe(drawctxt, next, &busy_list, pending) { |
| 973 | plist_del(&drawctxt->pending, &busy_list); |
| 974 | plist_add(&drawctxt->pending, &dispatcher->pending); |
| 975 | } |
| 976 | |
| 977 | /* Now put the contexts that need to be requeued back on the list */ |
| 978 | plist_for_each_entry_safe(drawctxt, next, &requeue, pending) { |
| 979 | plist_del(&drawctxt->pending, &requeue); |
| 980 | plist_add(&drawctxt->pending, &dispatcher->pending); |
| 981 | } |
| 982 | |
| 983 | spin_unlock(&dispatcher->plist_lock); |
| 984 | } |
| 985 | |
Hareesh Gundu | 2eb74d7 | 2017-06-07 14:50:15 +0530 | [diff] [blame] | 986 | static inline void _decrement_submit_now(struct kgsl_device *device) |
| 987 | { |
| 988 | spin_lock(&device->submit_lock); |
| 989 | device->submit_now--; |
| 990 | spin_unlock(&device->submit_lock); |
| 991 | } |
| 992 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 993 | /** |
| 994 | * adreno_dispatcher_issuecmds() - Issue commmands from pending contexts |
| 995 | * @adreno_dev: Pointer to the adreno device struct |
| 996 | * |
| 997 | * Lock the dispatcher and call _adreno_dispatcher_issueibcmds |
| 998 | */ |
| 999 | static void adreno_dispatcher_issuecmds(struct adreno_device *adreno_dev) |
| 1000 | { |
| 1001 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
Hareesh Gundu | 2eb74d7 | 2017-06-07 14:50:15 +0530 | [diff] [blame] | 1002 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 1003 | |
| 1004 | spin_lock(&device->submit_lock); |
| 1005 | /* If state transition to SLUMBER, schedule the work for later */ |
| 1006 | if (device->slumber == true) { |
| 1007 | spin_unlock(&device->submit_lock); |
| 1008 | goto done; |
| 1009 | } |
| 1010 | device->submit_now++; |
| 1011 | spin_unlock(&device->submit_lock); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1012 | |
| 1013 | /* If the dispatcher is busy then schedule the work for later */ |
| 1014 | if (!mutex_trylock(&dispatcher->mutex)) { |
Hareesh Gundu | 2eb74d7 | 2017-06-07 14:50:15 +0530 | [diff] [blame] | 1015 | _decrement_submit_now(device); |
| 1016 | goto done; |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1017 | } |
| 1018 | |
| 1019 | _adreno_dispatcher_issuecmds(adreno_dev); |
| 1020 | mutex_unlock(&dispatcher->mutex); |
Hareesh Gundu | 2eb74d7 | 2017-06-07 14:50:15 +0530 | [diff] [blame] | 1021 | _decrement_submit_now(device); |
| 1022 | return; |
| 1023 | done: |
| 1024 | adreno_dispatcher_schedule(device); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1025 | } |
| 1026 | |
| 1027 | /** |
| 1028 | * get_timestamp() - Return the next timestamp for the context |
| 1029 | * @drawctxt - Pointer to an adreno draw context struct |
| 1030 | * @drawobj - Pointer to a drawobj |
| 1031 | * @timestamp - Pointer to a timestamp value possibly passed from the user |
| 1032 | * @user_ts - user generated timestamp |
| 1033 | * |
| 1034 | * Assign a timestamp based on the settings of the draw context and the command |
| 1035 | * batch. |
| 1036 | */ |
| 1037 | static int get_timestamp(struct adreno_context *drawctxt, |
| 1038 | struct kgsl_drawobj *drawobj, unsigned int *timestamp, |
| 1039 | unsigned int user_ts) |
| 1040 | { |
| 1041 | |
| 1042 | if (drawctxt->base.flags & KGSL_CONTEXT_USER_GENERATED_TS) { |
| 1043 | /* |
| 1044 | * User specified timestamps need to be greater than the last |
| 1045 | * issued timestamp in the context |
| 1046 | */ |
| 1047 | if (timestamp_cmp(drawctxt->timestamp, user_ts) >= 0) |
| 1048 | return -ERANGE; |
| 1049 | |
| 1050 | drawctxt->timestamp = user_ts; |
| 1051 | } else |
| 1052 | drawctxt->timestamp++; |
| 1053 | |
| 1054 | *timestamp = drawctxt->timestamp; |
| 1055 | drawobj->timestamp = *timestamp; |
| 1056 | return 0; |
| 1057 | } |
| 1058 | |
| 1059 | static void _set_ft_policy(struct adreno_device *adreno_dev, |
| 1060 | struct adreno_context *drawctxt, |
| 1061 | struct kgsl_drawobj_cmd *cmdobj) |
| 1062 | { |
| 1063 | /* |
| 1064 | * Set the fault tolerance policy for the command batch - assuming the |
| 1065 | * context hasn't disabled FT use the current device policy |
| 1066 | */ |
| 1067 | if (drawctxt->base.flags & KGSL_CONTEXT_NO_FAULT_TOLERANCE) |
| 1068 | set_bit(KGSL_FT_DISABLE, &cmdobj->fault_policy); |
Hareesh Gundu | ccfb89b | 2017-04-14 18:36:20 +0530 | [diff] [blame] | 1069 | /* |
| 1070 | * Set the fault tolerance policy to FT_REPLAY - As context wants |
| 1071 | * to invalidate it after a replay attempt fails. This doesn't |
| 1072 | * require to execute the default FT policy. |
| 1073 | */ |
| 1074 | else if (drawctxt->base.flags & KGSL_CONTEXT_INVALIDATE_ON_FAULT) |
| 1075 | set_bit(KGSL_FT_REPLAY, &cmdobj->fault_policy); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1076 | else |
| 1077 | cmdobj->fault_policy = adreno_dev->ft_policy; |
| 1078 | } |
| 1079 | |
| 1080 | static void _cmdobj_set_flags(struct adreno_context *drawctxt, |
| 1081 | struct kgsl_drawobj_cmd *cmdobj) |
| 1082 | { |
| 1083 | /* |
| 1084 | * Force the preamble for this submission only - this is usually |
| 1085 | * requested by the dispatcher as part of fault recovery |
| 1086 | */ |
| 1087 | if (test_and_clear_bit(ADRENO_CONTEXT_FORCE_PREAMBLE, |
| 1088 | &drawctxt->base.priv)) |
| 1089 | set_bit(CMDOBJ_FORCE_PREAMBLE, &cmdobj->priv); |
| 1090 | |
| 1091 | /* |
| 1092 | * Force the premable if set from userspace in the context or |
| 1093 | * command obj flags |
| 1094 | */ |
| 1095 | if ((drawctxt->base.flags & KGSL_CONTEXT_CTX_SWITCH) || |
| 1096 | (cmdobj->base.flags & KGSL_DRAWOBJ_CTX_SWITCH)) |
| 1097 | set_bit(CMDOBJ_FORCE_PREAMBLE, &cmdobj->priv); |
| 1098 | |
| 1099 | /* Skip this ib if IFH_NOP is enabled */ |
| 1100 | if (drawctxt->base.flags & KGSL_CONTEXT_IFH_NOP) |
| 1101 | set_bit(CMDOBJ_SKIP, &cmdobj->priv); |
| 1102 | |
| 1103 | /* |
| 1104 | * If we are waiting for the end of frame and it hasn't appeared yet, |
| 1105 | * then mark the command obj as skipped. It will still progress |
| 1106 | * through the pipeline but it won't actually send any commands |
| 1107 | */ |
| 1108 | |
| 1109 | if (test_bit(ADRENO_CONTEXT_SKIP_EOF, &drawctxt->base.priv)) { |
| 1110 | set_bit(CMDOBJ_SKIP, &cmdobj->priv); |
| 1111 | |
| 1112 | /* |
| 1113 | * If this command obj represents the EOF then clear the way |
| 1114 | * for the dispatcher to continue submitting |
| 1115 | */ |
| 1116 | |
| 1117 | if (cmdobj->base.flags & KGSL_DRAWOBJ_END_OF_FRAME) { |
| 1118 | clear_bit(ADRENO_CONTEXT_SKIP_EOF, |
| 1119 | &drawctxt->base.priv); |
| 1120 | |
| 1121 | /* |
| 1122 | * Force the preamble on the next command to ensure that |
| 1123 | * the state is correct |
| 1124 | */ |
| 1125 | set_bit(ADRENO_CONTEXT_FORCE_PREAMBLE, |
| 1126 | &drawctxt->base.priv); |
| 1127 | } |
| 1128 | } |
| 1129 | } |
| 1130 | |
| 1131 | static inline int _check_context_state(struct kgsl_context *context) |
| 1132 | { |
| 1133 | if (kgsl_context_invalid(context)) |
| 1134 | return -EDEADLK; |
| 1135 | |
| 1136 | if (kgsl_context_detached(context)) |
| 1137 | return -ENOENT; |
| 1138 | |
| 1139 | return 0; |
| 1140 | } |
| 1141 | |
| 1142 | static inline bool _verify_ib(struct kgsl_device_private *dev_priv, |
| 1143 | struct kgsl_context *context, struct kgsl_memobj_node *ib) |
| 1144 | { |
| 1145 | struct kgsl_device *device = dev_priv->device; |
| 1146 | struct kgsl_process_private *private = dev_priv->process_priv; |
| 1147 | |
| 1148 | /* The maximum allowable size for an IB in the CP is 0xFFFFF dwords */ |
| 1149 | if (ib->size == 0 || ((ib->size >> 2) > 0xFFFFF)) { |
| 1150 | pr_context(device, context, "ctxt %d invalid ib size %lld\n", |
| 1151 | context->id, ib->size); |
| 1152 | return false; |
| 1153 | } |
| 1154 | |
| 1155 | /* Make sure that the address is mapped */ |
| 1156 | if (!kgsl_mmu_gpuaddr_in_range(private->pagetable, ib->gpuaddr)) { |
| 1157 | pr_context(device, context, "ctxt %d invalid ib gpuaddr %llX\n", |
| 1158 | context->id, ib->gpuaddr); |
| 1159 | return false; |
| 1160 | } |
| 1161 | |
| 1162 | return true; |
| 1163 | } |
| 1164 | |
| 1165 | static inline int _verify_cmdobj(struct kgsl_device_private *dev_priv, |
| 1166 | struct kgsl_context *context, struct kgsl_drawobj *drawobj[], |
| 1167 | uint32_t count) |
| 1168 | { |
| 1169 | struct kgsl_device *device = dev_priv->device; |
| 1170 | struct kgsl_memobj_node *ib; |
| 1171 | unsigned int i; |
| 1172 | |
| 1173 | for (i = 0; i < count; i++) { |
| 1174 | /* Verify the IBs before they get queued */ |
| 1175 | if (drawobj[i]->type == CMDOBJ_TYPE) { |
| 1176 | struct kgsl_drawobj_cmd *cmdobj = CMDOBJ(drawobj[i]); |
| 1177 | |
| 1178 | list_for_each_entry(ib, &cmdobj->cmdlist, node) |
| 1179 | if (_verify_ib(dev_priv, |
| 1180 | &ADRENO_CONTEXT(context)->base, ib) |
| 1181 | == false) |
| 1182 | return -EINVAL; |
| 1183 | /* |
| 1184 | * Clear the wake on touch bit to indicate an IB has |
| 1185 | * been submitted since the last time we set it. |
| 1186 | * But only clear it when we have rendering commands. |
| 1187 | */ |
| 1188 | device->flags &= ~KGSL_FLAG_WAKE_ON_TOUCH; |
| 1189 | } |
| 1190 | |
| 1191 | /* A3XX does not have support for drawobj profiling */ |
| 1192 | if (adreno_is_a3xx(ADRENO_DEVICE(device)) && |
| 1193 | (drawobj[i]->flags & KGSL_DRAWOBJ_PROFILING)) |
| 1194 | return -EOPNOTSUPP; |
| 1195 | } |
| 1196 | |
| 1197 | return 0; |
| 1198 | } |
| 1199 | |
| 1200 | static inline int _wait_for_room_in_context_queue( |
| 1201 | struct adreno_context *drawctxt) |
| 1202 | { |
| 1203 | int ret = 0; |
| 1204 | |
| 1205 | /* Wait for room in the context queue */ |
| 1206 | while (drawctxt->queued >= _context_drawqueue_size) { |
| 1207 | trace_adreno_drawctxt_sleep(drawctxt); |
| 1208 | spin_unlock(&drawctxt->lock); |
| 1209 | |
| 1210 | ret = wait_event_interruptible_timeout(drawctxt->wq, |
| 1211 | _check_context_queue(drawctxt), |
| 1212 | msecs_to_jiffies(_context_queue_wait)); |
| 1213 | |
| 1214 | spin_lock(&drawctxt->lock); |
| 1215 | trace_adreno_drawctxt_wake(drawctxt); |
| 1216 | |
| 1217 | if (ret <= 0) |
| 1218 | return (ret == 0) ? -ETIMEDOUT : (int) ret; |
| 1219 | } |
| 1220 | |
| 1221 | return 0; |
| 1222 | } |
| 1223 | |
| 1224 | static unsigned int _check_context_state_to_queue_cmds( |
| 1225 | struct adreno_context *drawctxt) |
| 1226 | { |
| 1227 | int ret = _check_context_state(&drawctxt->base); |
| 1228 | |
| 1229 | if (ret) |
| 1230 | return ret; |
| 1231 | |
| 1232 | ret = _wait_for_room_in_context_queue(drawctxt); |
| 1233 | if (ret) |
| 1234 | return ret; |
| 1235 | |
| 1236 | /* |
| 1237 | * Account for the possiblity that the context got invalidated |
| 1238 | * while we were sleeping |
| 1239 | */ |
| 1240 | return _check_context_state(&drawctxt->base); |
| 1241 | } |
| 1242 | |
| 1243 | static void _queue_drawobj(struct adreno_context *drawctxt, |
| 1244 | struct kgsl_drawobj *drawobj) |
| 1245 | { |
| 1246 | /* Put the command into the queue */ |
| 1247 | drawctxt->drawqueue[drawctxt->drawqueue_tail] = drawobj; |
| 1248 | drawctxt->drawqueue_tail = (drawctxt->drawqueue_tail + 1) % |
| 1249 | ADRENO_CONTEXT_DRAWQUEUE_SIZE; |
| 1250 | drawctxt->queued++; |
| 1251 | trace_adreno_cmdbatch_queued(drawobj, drawctxt->queued); |
| 1252 | } |
| 1253 | |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 1254 | static int _queue_sparseobj(struct adreno_device *adreno_dev, |
| 1255 | struct adreno_context *drawctxt, struct kgsl_drawobj_sparse *sparseobj, |
| 1256 | uint32_t *timestamp, unsigned int user_ts) |
| 1257 | { |
| 1258 | struct kgsl_drawobj *drawobj = DRAWOBJ(sparseobj); |
| 1259 | int ret; |
| 1260 | |
| 1261 | ret = get_timestamp(drawctxt, drawobj, timestamp, user_ts); |
| 1262 | if (ret) |
| 1263 | return ret; |
| 1264 | |
| 1265 | /* |
| 1266 | * See if we can fastpath this thing - if nothing is |
| 1267 | * queued bind/unbind without queueing the context |
| 1268 | */ |
| 1269 | if (!drawctxt->queued) |
| 1270 | return 1; |
| 1271 | |
| 1272 | drawctxt->queued_timestamp = *timestamp; |
| 1273 | _queue_drawobj(drawctxt, drawobj); |
| 1274 | |
| 1275 | return 0; |
| 1276 | } |
| 1277 | |
| 1278 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1279 | static int _queue_markerobj(struct adreno_device *adreno_dev, |
| 1280 | struct adreno_context *drawctxt, struct kgsl_drawobj_cmd *markerobj, |
| 1281 | uint32_t *timestamp, unsigned int user_ts) |
| 1282 | { |
| 1283 | struct kgsl_drawobj *drawobj = DRAWOBJ(markerobj); |
| 1284 | int ret; |
| 1285 | |
| 1286 | ret = get_timestamp(drawctxt, drawobj, timestamp, user_ts); |
| 1287 | if (ret) |
| 1288 | return ret; |
| 1289 | |
| 1290 | /* |
| 1291 | * See if we can fastpath this thing - if nothing is queued |
| 1292 | * and nothing is inflight retire without bothering the GPU |
| 1293 | */ |
| 1294 | if (!drawctxt->queued && kgsl_check_timestamp(drawobj->device, |
| 1295 | drawobj->context, drawctxt->queued_timestamp)) { |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1296 | _retire_timestamp(drawobj); |
| 1297 | return 1; |
| 1298 | } |
| 1299 | |
| 1300 | /* |
| 1301 | * Remember the last queued timestamp - the marker will block |
| 1302 | * until that timestamp is expired (unless another command |
| 1303 | * comes along and forces the marker to execute) |
| 1304 | */ |
| 1305 | |
| 1306 | markerobj->marker_timestamp = drawctxt->queued_timestamp; |
| 1307 | drawctxt->queued_timestamp = *timestamp; |
| 1308 | _set_ft_policy(adreno_dev, drawctxt, markerobj); |
| 1309 | _cmdobj_set_flags(drawctxt, markerobj); |
| 1310 | |
| 1311 | _queue_drawobj(drawctxt, drawobj); |
| 1312 | |
| 1313 | return 0; |
| 1314 | } |
| 1315 | |
| 1316 | static int _queue_cmdobj(struct adreno_device *adreno_dev, |
| 1317 | struct adreno_context *drawctxt, struct kgsl_drawobj_cmd *cmdobj, |
| 1318 | uint32_t *timestamp, unsigned int user_ts) |
| 1319 | { |
| 1320 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1321 | unsigned int j; |
| 1322 | int ret; |
| 1323 | |
| 1324 | ret = get_timestamp(drawctxt, drawobj, timestamp, user_ts); |
| 1325 | if (ret) |
| 1326 | return ret; |
| 1327 | |
| 1328 | /* |
| 1329 | * If this is a real command then we need to force any markers |
| 1330 | * queued before it to dispatch to keep time linear - set the |
| 1331 | * skip bit so the commands get NOPed. |
| 1332 | */ |
| 1333 | j = drawctxt->drawqueue_head; |
| 1334 | |
| 1335 | while (j != drawctxt->drawqueue_tail) { |
| 1336 | if (drawctxt->drawqueue[j]->type == MARKEROBJ_TYPE) { |
| 1337 | struct kgsl_drawobj_cmd *markerobj = |
| 1338 | CMDOBJ(drawctxt->drawqueue[j]); |
| 1339 | set_bit(CMDOBJ_SKIP, &markerobj->priv); |
| 1340 | } |
| 1341 | |
| 1342 | j = DRAWQUEUE_NEXT(j, ADRENO_CONTEXT_DRAWQUEUE_SIZE); |
| 1343 | } |
| 1344 | |
| 1345 | drawctxt->queued_timestamp = *timestamp; |
| 1346 | _set_ft_policy(adreno_dev, drawctxt, cmdobj); |
| 1347 | _cmdobj_set_flags(drawctxt, cmdobj); |
| 1348 | |
| 1349 | _queue_drawobj(drawctxt, drawobj); |
| 1350 | |
| 1351 | return 0; |
| 1352 | } |
| 1353 | |
| 1354 | static void _queue_syncobj(struct adreno_context *drawctxt, |
| 1355 | struct kgsl_drawobj_sync *syncobj, uint32_t *timestamp) |
| 1356 | { |
| 1357 | struct kgsl_drawobj *drawobj = DRAWOBJ(syncobj); |
| 1358 | |
| 1359 | *timestamp = 0; |
| 1360 | drawobj->timestamp = 0; |
| 1361 | |
| 1362 | _queue_drawobj(drawctxt, drawobj); |
| 1363 | } |
| 1364 | |
| 1365 | /** |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 1366 | * adreno_dispactcher_queue_cmds() - Queue a new draw object in the context |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1367 | * @dev_priv: Pointer to the device private struct |
| 1368 | * @context: Pointer to the kgsl draw context |
| 1369 | * @drawobj: Pointer to the array of drawobj's being submitted |
| 1370 | * @count: Number of drawobj's being submitted |
| 1371 | * @timestamp: Pointer to the requested timestamp |
| 1372 | * |
| 1373 | * Queue a command in the context - if there isn't any room in the queue, then |
| 1374 | * block until there is |
| 1375 | */ |
| 1376 | int adreno_dispatcher_queue_cmds(struct kgsl_device_private *dev_priv, |
| 1377 | struct kgsl_context *context, struct kgsl_drawobj *drawobj[], |
| 1378 | uint32_t count, uint32_t *timestamp) |
| 1379 | |
| 1380 | { |
| 1381 | struct kgsl_device *device = dev_priv->device; |
| 1382 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 1383 | struct adreno_context *drawctxt = ADRENO_CONTEXT(context); |
| 1384 | struct adreno_dispatcher_drawqueue *dispatch_q; |
| 1385 | int ret; |
| 1386 | unsigned int i, user_ts; |
| 1387 | |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 1388 | if (!count) |
| 1389 | return -EINVAL; |
| 1390 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1391 | ret = _check_context_state(&drawctxt->base); |
| 1392 | if (ret) |
| 1393 | return ret; |
| 1394 | |
| 1395 | ret = _verify_cmdobj(dev_priv, context, drawobj, count); |
| 1396 | if (ret) |
| 1397 | return ret; |
| 1398 | |
| 1399 | /* wait for the suspend gate */ |
| 1400 | wait_for_completion(&device->halt_gate); |
| 1401 | |
| 1402 | spin_lock(&drawctxt->lock); |
| 1403 | |
| 1404 | ret = _check_context_state_to_queue_cmds(drawctxt); |
| 1405 | if (ret) { |
| 1406 | spin_unlock(&drawctxt->lock); |
| 1407 | return ret; |
| 1408 | } |
| 1409 | |
| 1410 | user_ts = *timestamp; |
| 1411 | |
| 1412 | for (i = 0; i < count; i++) { |
| 1413 | |
| 1414 | switch (drawobj[i]->type) { |
| 1415 | case MARKEROBJ_TYPE: |
| 1416 | ret = _queue_markerobj(adreno_dev, drawctxt, |
| 1417 | CMDOBJ(drawobj[i]), |
| 1418 | timestamp, user_ts); |
| 1419 | if (ret == 1) { |
| 1420 | spin_unlock(&drawctxt->lock); |
| 1421 | goto done; |
| 1422 | } else if (ret) { |
| 1423 | spin_unlock(&drawctxt->lock); |
| 1424 | return ret; |
| 1425 | } |
| 1426 | break; |
| 1427 | case CMDOBJ_TYPE: |
| 1428 | ret = _queue_cmdobj(adreno_dev, drawctxt, |
| 1429 | CMDOBJ(drawobj[i]), |
| 1430 | timestamp, user_ts); |
| 1431 | if (ret) { |
| 1432 | spin_unlock(&drawctxt->lock); |
| 1433 | return ret; |
| 1434 | } |
| 1435 | break; |
| 1436 | case SYNCOBJ_TYPE: |
| 1437 | _queue_syncobj(drawctxt, SYNCOBJ(drawobj[i]), |
| 1438 | timestamp); |
| 1439 | break; |
Tarun Karra | 2b8b363 | 2016-11-14 16:38:27 -0800 | [diff] [blame] | 1440 | case SPARSEOBJ_TYPE: |
| 1441 | ret = _queue_sparseobj(adreno_dev, drawctxt, |
| 1442 | SPARSEOBJ(drawobj[i]), |
| 1443 | timestamp, user_ts); |
| 1444 | if (ret == 1) { |
| 1445 | spin_unlock(&drawctxt->lock); |
| 1446 | _retire_sparseobj(SPARSEOBJ(drawobj[i]), |
| 1447 | drawctxt); |
| 1448 | return 0; |
| 1449 | } else if (ret) { |
| 1450 | spin_unlock(&drawctxt->lock); |
| 1451 | return ret; |
| 1452 | } |
| 1453 | break; |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1454 | default: |
| 1455 | spin_unlock(&drawctxt->lock); |
| 1456 | return -EINVAL; |
| 1457 | } |
| 1458 | |
| 1459 | } |
| 1460 | |
| 1461 | dispatch_q = ADRENO_DRAWOBJ_DISPATCH_DRAWQUEUE(drawobj[0]); |
| 1462 | |
| 1463 | _track_context(adreno_dev, dispatch_q, drawctxt); |
| 1464 | |
| 1465 | spin_unlock(&drawctxt->lock); |
| 1466 | |
Gaurav Sonwani | c169c32 | 2017-06-15 14:11:23 +0530 | [diff] [blame] | 1467 | if (device->pwrctrl.l2pc_update_queue) |
| 1468 | kgsl_pwrctrl_update_l2pc(&adreno_dev->dev, |
| 1469 | KGSL_L2PC_QUEUE_TIMEOUT); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 1470 | |
| 1471 | /* Add the context to the dispatcher pending list */ |
| 1472 | dispatcher_queue_context(adreno_dev, drawctxt); |
| 1473 | |
| 1474 | /* |
| 1475 | * Only issue commands if inflight is less than burst -this prevents us |
| 1476 | * from sitting around waiting for the mutex on a busy system - the work |
| 1477 | * loop will schedule it for us. Inflight is mutex protected but the |
| 1478 | * worse that can happen is that it will go to 0 after we check and if |
| 1479 | * it goes to 0 it is because the work loop decremented it and the work |
| 1480 | * queue will try to schedule new commands anyway. |
| 1481 | */ |
| 1482 | |
| 1483 | if (dispatch_q->inflight < _context_drawobj_burst) |
| 1484 | adreno_dispatcher_issuecmds(adreno_dev); |
| 1485 | done: |
| 1486 | if (test_and_clear_bit(ADRENO_CONTEXT_FAULT, &context->priv)) |
| 1487 | return -EPROTO; |
| 1488 | |
| 1489 | return 0; |
| 1490 | } |
| 1491 | |
| 1492 | static int _mark_context(int id, void *ptr, void *data) |
| 1493 | { |
| 1494 | unsigned int guilty = *((unsigned int *) data); |
| 1495 | struct kgsl_context *context = ptr; |
| 1496 | |
| 1497 | /* |
| 1498 | * If the context is guilty mark it as such. Otherwise mark it as |
| 1499 | * innocent if it had not already been marked as guilty. If id is |
| 1500 | * passed as 0 then mark EVERYBODY guilty (recovery failed) |
| 1501 | */ |
| 1502 | |
| 1503 | if (guilty == 0 || guilty == context->id) |
| 1504 | context->reset_status = |
| 1505 | KGSL_CTX_STAT_GUILTY_CONTEXT_RESET_EXT; |
| 1506 | else if (context->reset_status != |
| 1507 | KGSL_CTX_STAT_GUILTY_CONTEXT_RESET_EXT) |
| 1508 | context->reset_status = |
| 1509 | KGSL_CTX_STAT_INNOCENT_CONTEXT_RESET_EXT; |
| 1510 | |
| 1511 | return 0; |
| 1512 | } |
| 1513 | |
| 1514 | /** |
| 1515 | * mark_guilty_context() - Mark the given context as guilty (failed recovery) |
| 1516 | * @device: Pointer to a KGSL device structure |
| 1517 | * @id: Context ID of the guilty context (or 0 to mark all as guilty) |
| 1518 | * |
| 1519 | * Mark the given (or all) context(s) as guilty (failed recovery) |
| 1520 | */ |
| 1521 | static void mark_guilty_context(struct kgsl_device *device, unsigned int id) |
| 1522 | { |
| 1523 | /* Mark the status for all the contexts in the device */ |
| 1524 | |
| 1525 | read_lock(&device->context_lock); |
| 1526 | idr_for_each(&device->context_idr, _mark_context, &id); |
| 1527 | read_unlock(&device->context_lock); |
| 1528 | } |
| 1529 | |
| 1530 | /* |
| 1531 | * If an IB inside of the drawobj has a gpuaddr that matches the base |
| 1532 | * passed in then zero the size which effectively skips it when it is submitted |
| 1533 | * in the ringbuffer. |
| 1534 | */ |
| 1535 | static void _skip_ib(struct kgsl_drawobj_cmd *cmdobj, uint64_t base) |
| 1536 | { |
| 1537 | struct kgsl_memobj_node *ib; |
| 1538 | |
| 1539 | list_for_each_entry(ib, &cmdobj->cmdlist, node) { |
| 1540 | if (ib->gpuaddr == base) { |
| 1541 | ib->priv |= MEMOBJ_SKIP; |
| 1542 | if (base) |
| 1543 | return; |
| 1544 | } |
| 1545 | } |
| 1546 | } |
| 1547 | |
| 1548 | static void _skip_cmd(struct kgsl_drawobj_cmd *cmdobj, |
| 1549 | struct kgsl_drawobj_cmd **replay, int count) |
| 1550 | { |
| 1551 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1552 | struct adreno_context *drawctxt = ADRENO_CONTEXT(drawobj->context); |
| 1553 | int i; |
| 1554 | |
| 1555 | /* |
| 1556 | * SKIPCMD policy: next IB issued for this context is tentative |
| 1557 | * if it fails we assume that GFT failed and if it succeeds |
| 1558 | * we mark GFT as a success. |
| 1559 | * |
| 1560 | * Find next commandbatch for the faulting context |
| 1561 | * If commandbatch is found |
| 1562 | * a) store the current commandbatch fault_policy in context's next |
| 1563 | * commandbatch fault_policy |
| 1564 | * b) force preamble for next commandbatch |
| 1565 | */ |
| 1566 | for (i = 1; i < count; i++) { |
| 1567 | if (DRAWOBJ(replay[i])->context->id == drawobj->context->id) { |
| 1568 | replay[i]->fault_policy = replay[0]->fault_policy; |
| 1569 | set_bit(CMDOBJ_FORCE_PREAMBLE, &replay[i]->priv); |
| 1570 | set_bit(KGSL_FT_SKIPCMD, &replay[i]->fault_recovery); |
| 1571 | break; |
| 1572 | } |
| 1573 | } |
| 1574 | |
| 1575 | /* |
| 1576 | * If we did not find the next cmd then |
| 1577 | * a) set a flag for next command issued in this context |
| 1578 | * b) store the fault_policy, this fault_policy becomes the policy of |
| 1579 | * next command issued in this context |
| 1580 | */ |
| 1581 | if ((i == count) && drawctxt) { |
| 1582 | set_bit(ADRENO_CONTEXT_SKIP_CMD, &drawctxt->base.priv); |
| 1583 | drawctxt->fault_policy = replay[0]->fault_policy; |
| 1584 | } |
| 1585 | |
| 1586 | /* set the flags to skip this cmdobj */ |
| 1587 | set_bit(CMDOBJ_SKIP, &cmdobj->priv); |
| 1588 | cmdobj->fault_recovery = 0; |
| 1589 | } |
| 1590 | |
| 1591 | static void _skip_frame(struct kgsl_drawobj_cmd *cmdobj, |
| 1592 | struct kgsl_drawobj_cmd **replay, int count) |
| 1593 | { |
| 1594 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1595 | struct adreno_context *drawctxt = ADRENO_CONTEXT(drawobj->context); |
| 1596 | int skip = 1; |
| 1597 | int i; |
| 1598 | |
| 1599 | for (i = 0; i < count; i++) { |
| 1600 | |
| 1601 | struct kgsl_drawobj *replay_obj = DRAWOBJ(replay[i]); |
| 1602 | |
| 1603 | /* |
| 1604 | * Only operate on drawobj's that belong to the |
| 1605 | * faulting context |
| 1606 | */ |
| 1607 | |
| 1608 | if (replay_obj->context->id != drawobj->context->id) |
| 1609 | continue; |
| 1610 | |
| 1611 | /* |
| 1612 | * Skip all the drawobjs in this context until |
| 1613 | * the EOF flag is seen. If the EOF flag is seen then |
| 1614 | * force the preamble for the next command. |
| 1615 | */ |
| 1616 | |
| 1617 | if (skip) { |
| 1618 | set_bit(CMDOBJ_SKIP, &replay[i]->priv); |
| 1619 | |
| 1620 | if (replay_obj->flags & KGSL_DRAWOBJ_END_OF_FRAME) |
| 1621 | skip = 0; |
| 1622 | } else { |
| 1623 | set_bit(CMDOBJ_FORCE_PREAMBLE, &replay[i]->priv); |
| 1624 | return; |
| 1625 | } |
| 1626 | } |
| 1627 | |
| 1628 | /* |
| 1629 | * If the EOF flag hasn't been seen yet then set the flag in the |
| 1630 | * drawctxt to keep looking for it |
| 1631 | */ |
| 1632 | |
| 1633 | if (skip && drawctxt) |
| 1634 | set_bit(ADRENO_CONTEXT_SKIP_EOF, &drawctxt->base.priv); |
| 1635 | |
| 1636 | /* |
| 1637 | * If we did see the EOF flag then force the preamble on for the |
| 1638 | * next command issued on this context |
| 1639 | */ |
| 1640 | |
| 1641 | if (!skip && drawctxt) |
| 1642 | set_bit(ADRENO_CONTEXT_FORCE_PREAMBLE, &drawctxt->base.priv); |
| 1643 | } |
| 1644 | |
| 1645 | static void remove_invalidated_cmdobjs(struct kgsl_device *device, |
| 1646 | struct kgsl_drawobj_cmd **replay, int count) |
| 1647 | { |
| 1648 | int i; |
| 1649 | |
| 1650 | for (i = 0; i < count; i++) { |
| 1651 | struct kgsl_drawobj_cmd *cmdobj = replay[i]; |
| 1652 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1653 | |
| 1654 | if (cmdobj == NULL) |
| 1655 | continue; |
| 1656 | |
| 1657 | if (kgsl_context_detached(drawobj->context) || |
| 1658 | kgsl_context_invalid(drawobj->context)) { |
| 1659 | replay[i] = NULL; |
| 1660 | |
| 1661 | mutex_lock(&device->mutex); |
| 1662 | kgsl_cancel_events_timestamp(device, |
| 1663 | &drawobj->context->events, drawobj->timestamp); |
| 1664 | mutex_unlock(&device->mutex); |
| 1665 | |
| 1666 | kgsl_drawobj_destroy(drawobj); |
| 1667 | } |
| 1668 | } |
| 1669 | } |
| 1670 | |
| 1671 | static char _pidname[TASK_COMM_LEN]; |
| 1672 | |
| 1673 | static inline const char *_kgsl_context_comm(struct kgsl_context *context) |
| 1674 | { |
| 1675 | if (context && context->proc_priv) |
| 1676 | strlcpy(_pidname, context->proc_priv->comm, sizeof(_pidname)); |
| 1677 | else |
| 1678 | snprintf(_pidname, TASK_COMM_LEN, "unknown"); |
| 1679 | |
| 1680 | return _pidname; |
| 1681 | } |
| 1682 | |
| 1683 | #define pr_fault(_d, _c, fmt, args...) \ |
| 1684 | dev_err((_d)->dev, "%s[%d]: " fmt, \ |
| 1685 | _kgsl_context_comm((_c)->context), \ |
| 1686 | (_c)->context->proc_priv->pid, ##args) |
| 1687 | |
| 1688 | |
| 1689 | static void adreno_fault_header(struct kgsl_device *device, |
| 1690 | struct adreno_ringbuffer *rb, struct kgsl_drawobj_cmd *cmdobj) |
| 1691 | { |
| 1692 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 1693 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1694 | unsigned int status, rptr, wptr, ib1sz, ib2sz; |
| 1695 | uint64_t ib1base, ib2base; |
| 1696 | |
| 1697 | adreno_readreg(adreno_dev, ADRENO_REG_RBBM_STATUS, &status); |
| 1698 | adreno_readreg(adreno_dev, ADRENO_REG_CP_RB_RPTR, &rptr); |
| 1699 | adreno_readreg(adreno_dev, ADRENO_REG_CP_RB_WPTR, &wptr); |
| 1700 | adreno_readreg64(adreno_dev, ADRENO_REG_CP_IB1_BASE, |
| 1701 | ADRENO_REG_CP_IB1_BASE_HI, &ib1base); |
| 1702 | adreno_readreg(adreno_dev, ADRENO_REG_CP_IB1_BUFSZ, &ib1sz); |
| 1703 | adreno_readreg64(adreno_dev, ADRENO_REG_CP_IB2_BASE, |
| 1704 | ADRENO_REG_CP_IB2_BASE_HI, &ib2base); |
| 1705 | adreno_readreg(adreno_dev, ADRENO_REG_CP_IB2_BUFSZ, &ib2sz); |
| 1706 | |
| 1707 | if (drawobj != NULL) { |
| 1708 | struct adreno_context *drawctxt = |
| 1709 | ADRENO_CONTEXT(drawobj->context); |
| 1710 | |
| 1711 | trace_adreno_gpu_fault(drawobj->context->id, |
| 1712 | drawobj->timestamp, |
| 1713 | status, rptr, wptr, ib1base, ib1sz, |
| 1714 | ib2base, ib2sz, drawctxt->rb->id); |
| 1715 | |
| 1716 | pr_fault(device, drawobj, |
| 1717 | "gpu fault ctx %d ts %d status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n", |
| 1718 | drawobj->context->id, drawobj->timestamp, status, |
| 1719 | rptr, wptr, ib1base, ib1sz, ib2base, ib2sz); |
| 1720 | |
| 1721 | if (rb != NULL) |
| 1722 | pr_fault(device, drawobj, |
| 1723 | "gpu fault rb %d rb sw r/w %4.4x/%4.4x\n", |
| 1724 | rb->id, rptr, rb->wptr); |
| 1725 | } else { |
| 1726 | int id = (rb != NULL) ? rb->id : -1; |
| 1727 | |
| 1728 | dev_err(device->dev, |
| 1729 | "RB[%d]: gpu fault status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n", |
| 1730 | id, status, rptr, wptr, ib1base, ib1sz, ib2base, |
| 1731 | ib2sz); |
| 1732 | if (rb != NULL) |
| 1733 | dev_err(device->dev, |
| 1734 | "RB[%d] gpu fault rb sw r/w %4.4x/%4.4x\n", |
| 1735 | rb->id, rptr, rb->wptr); |
| 1736 | } |
| 1737 | } |
| 1738 | |
| 1739 | void adreno_fault_skipcmd_detached(struct adreno_device *adreno_dev, |
| 1740 | struct adreno_context *drawctxt, |
| 1741 | struct kgsl_drawobj *drawobj) |
| 1742 | { |
| 1743 | if (test_bit(ADRENO_CONTEXT_SKIP_CMD, &drawctxt->base.priv) && |
| 1744 | kgsl_context_detached(&drawctxt->base)) { |
| 1745 | pr_context(KGSL_DEVICE(adreno_dev), drawobj->context, |
| 1746 | "gpu detached context %d\n", drawobj->context->id); |
| 1747 | clear_bit(ADRENO_CONTEXT_SKIP_CMD, &drawctxt->base.priv); |
| 1748 | } |
| 1749 | } |
| 1750 | |
| 1751 | /** |
| 1752 | * process_cmdobj_fault() - Process a cmdobj for fault policies |
| 1753 | * @device: Device on which the cmdobj caused a fault |
| 1754 | * @replay: List of cmdobj's that are to be replayed on the device. The |
| 1755 | * first command in the replay list is the faulting command and the remaining |
| 1756 | * cmdobj's in the list are commands that were submitted to the same queue |
| 1757 | * as the faulting one. |
| 1758 | * @count: Number of cmdobj's in replay |
| 1759 | * @base: The IB1 base at the time of fault |
| 1760 | * @fault: The fault type |
| 1761 | */ |
| 1762 | static void process_cmdobj_fault(struct kgsl_device *device, |
| 1763 | struct kgsl_drawobj_cmd **replay, int count, |
| 1764 | unsigned int base, |
| 1765 | int fault) |
| 1766 | { |
| 1767 | struct kgsl_drawobj_cmd *cmdobj = replay[0]; |
| 1768 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1769 | int i; |
| 1770 | char *state = "failed"; |
| 1771 | |
| 1772 | /* |
| 1773 | * If GFT recovered more than X times in Y ms invalidate the context |
| 1774 | * and do not attempt recovery. |
| 1775 | * Example: X==3 and Y==3000 ms, GPU hung at 500ms, 1700ms, 25000ms and |
| 1776 | * 3000ms for the same context, we will not try FT and invalidate the |
| 1777 | * context @3000ms because context triggered GFT more than 3 times in |
| 1778 | * last 3 seconds. If a context caused recoverable GPU hangs |
| 1779 | * where 1st and 4th gpu hang are more than 3 seconds apart we |
| 1780 | * won't disable GFT and invalidate the context. |
| 1781 | */ |
| 1782 | if (test_bit(KGSL_FT_THROTTLE, &cmdobj->fault_policy)) { |
| 1783 | if (time_after(jiffies, (drawobj->context->fault_time |
| 1784 | + msecs_to_jiffies(_fault_throttle_time)))) { |
| 1785 | drawobj->context->fault_time = jiffies; |
| 1786 | drawobj->context->fault_count = 1; |
| 1787 | } else { |
| 1788 | drawobj->context->fault_count++; |
| 1789 | if (drawobj->context->fault_count > |
| 1790 | _fault_throttle_burst) { |
| 1791 | set_bit(KGSL_FT_DISABLE, |
| 1792 | &cmdobj->fault_policy); |
| 1793 | pr_context(device, drawobj->context, |
| 1794 | "gpu fault threshold exceeded %d faults in %d msecs\n", |
| 1795 | _fault_throttle_burst, |
| 1796 | _fault_throttle_time); |
| 1797 | } |
| 1798 | } |
| 1799 | } |
| 1800 | |
| 1801 | /* |
| 1802 | * If FT is disabled for this cmdobj invalidate immediately |
| 1803 | */ |
| 1804 | |
| 1805 | if (test_bit(KGSL_FT_DISABLE, &cmdobj->fault_policy) || |
| 1806 | test_bit(KGSL_FT_TEMP_DISABLE, &cmdobj->fault_policy)) { |
| 1807 | state = "skipped"; |
| 1808 | bitmap_zero(&cmdobj->fault_policy, BITS_PER_LONG); |
| 1809 | } |
| 1810 | |
| 1811 | /* If the context is detached do not run FT on context */ |
| 1812 | if (kgsl_context_detached(drawobj->context)) { |
| 1813 | state = "detached"; |
| 1814 | bitmap_zero(&cmdobj->fault_policy, BITS_PER_LONG); |
| 1815 | } |
| 1816 | |
| 1817 | /* |
| 1818 | * Set a flag so we don't print another PM dump if the cmdobj fails |
| 1819 | * again on replay |
| 1820 | */ |
| 1821 | |
| 1822 | set_bit(KGSL_FT_SKIP_PMDUMP, &cmdobj->fault_policy); |
| 1823 | |
| 1824 | /* |
| 1825 | * A hardware fault generally means something was deterministically |
| 1826 | * wrong with the cmdobj - no point in trying to replay it |
| 1827 | * Clear the replay bit and move on to the next policy level |
| 1828 | */ |
| 1829 | |
| 1830 | if (fault & ADRENO_HARD_FAULT) |
| 1831 | clear_bit(KGSL_FT_REPLAY, &(cmdobj->fault_policy)); |
| 1832 | |
| 1833 | /* |
| 1834 | * A timeout fault means the IB timed out - clear the policy and |
| 1835 | * invalidate - this will clear the FT_SKIP_PMDUMP bit but that is okay |
| 1836 | * because we won't see this cmdobj again |
| 1837 | */ |
| 1838 | |
| 1839 | if (fault & ADRENO_TIMEOUT_FAULT) |
| 1840 | bitmap_zero(&cmdobj->fault_policy, BITS_PER_LONG); |
| 1841 | |
| 1842 | /* |
| 1843 | * If the context had a GPU page fault then it is likely it would fault |
| 1844 | * again if replayed |
| 1845 | */ |
| 1846 | |
| 1847 | if (test_bit(KGSL_CONTEXT_PRIV_PAGEFAULT, |
| 1848 | &drawobj->context->priv)) { |
| 1849 | /* we'll need to resume the mmu later... */ |
| 1850 | clear_bit(KGSL_FT_REPLAY, &cmdobj->fault_policy); |
| 1851 | clear_bit(KGSL_CONTEXT_PRIV_PAGEFAULT, |
| 1852 | &drawobj->context->priv); |
| 1853 | } |
| 1854 | |
| 1855 | /* |
| 1856 | * Execute the fault tolerance policy. Each cmdobj stores the |
| 1857 | * current fault policy that was set when it was queued. |
| 1858 | * As the options are tried in descending priority |
| 1859 | * (REPLAY -> SKIPIBS -> SKIPFRAME -> NOTHING) the bits are cleared |
| 1860 | * from the cmdobj policy so the next thing can be tried if the |
| 1861 | * change comes around again |
| 1862 | */ |
| 1863 | |
| 1864 | /* Replay the hanging cmdobj again */ |
| 1865 | if (test_and_clear_bit(KGSL_FT_REPLAY, &cmdobj->fault_policy)) { |
| 1866 | trace_adreno_cmdbatch_recovery(cmdobj, BIT(KGSL_FT_REPLAY)); |
| 1867 | set_bit(KGSL_FT_REPLAY, &cmdobj->fault_recovery); |
| 1868 | return; |
| 1869 | } |
| 1870 | |
| 1871 | /* |
| 1872 | * Skip the last IB1 that was played but replay everything else. |
| 1873 | * Note that the last IB1 might not be in the "hung" cmdobj |
| 1874 | * because the CP may have caused a page-fault while it was prefetching |
| 1875 | * the next IB1/IB2. walk all outstanding commands and zap the |
| 1876 | * supposedly bad IB1 where ever it lurks. |
| 1877 | */ |
| 1878 | |
| 1879 | if (test_and_clear_bit(KGSL_FT_SKIPIB, &cmdobj->fault_policy)) { |
| 1880 | trace_adreno_cmdbatch_recovery(cmdobj, BIT(KGSL_FT_SKIPIB)); |
| 1881 | set_bit(KGSL_FT_SKIPIB, &cmdobj->fault_recovery); |
| 1882 | |
| 1883 | for (i = 0; i < count; i++) { |
| 1884 | if (replay[i] != NULL && |
| 1885 | DRAWOBJ(replay[i])->context->id == |
| 1886 | drawobj->context->id) |
| 1887 | _skip_ib(replay[i], base); |
| 1888 | } |
| 1889 | |
| 1890 | return; |
| 1891 | } |
| 1892 | |
| 1893 | /* Skip the faulted cmdobj submission */ |
| 1894 | if (test_and_clear_bit(KGSL_FT_SKIPCMD, &cmdobj->fault_policy)) { |
| 1895 | trace_adreno_cmdbatch_recovery(cmdobj, BIT(KGSL_FT_SKIPCMD)); |
| 1896 | |
| 1897 | /* Skip faulting cmdobj */ |
| 1898 | _skip_cmd(cmdobj, replay, count); |
| 1899 | |
| 1900 | return; |
| 1901 | } |
| 1902 | |
| 1903 | if (test_and_clear_bit(KGSL_FT_SKIPFRAME, &cmdobj->fault_policy)) { |
| 1904 | trace_adreno_cmdbatch_recovery(cmdobj, |
| 1905 | BIT(KGSL_FT_SKIPFRAME)); |
| 1906 | set_bit(KGSL_FT_SKIPFRAME, &cmdobj->fault_recovery); |
| 1907 | |
| 1908 | /* |
| 1909 | * Skip all the pending cmdobj's for this context until |
| 1910 | * the EOF frame is seen |
| 1911 | */ |
| 1912 | _skip_frame(cmdobj, replay, count); |
| 1913 | return; |
| 1914 | } |
| 1915 | |
| 1916 | /* If we get here then all the policies failed */ |
| 1917 | |
| 1918 | pr_context(device, drawobj->context, "gpu %s ctx %d ts %d\n", |
| 1919 | state, drawobj->context->id, drawobj->timestamp); |
| 1920 | |
| 1921 | /* Mark the context as failed */ |
| 1922 | mark_guilty_context(device, drawobj->context->id); |
| 1923 | |
| 1924 | /* Invalidate the context */ |
| 1925 | adreno_drawctxt_invalidate(device, drawobj->context); |
| 1926 | } |
| 1927 | |
| 1928 | /** |
| 1929 | * recover_dispatch_q() - Recover all commands in a dispatch queue by |
| 1930 | * resubmitting the commands |
| 1931 | * @device: Device on which recovery is performed |
| 1932 | * @dispatch_q: The command queue to recover |
| 1933 | * @fault: Faults caused by the command in the dispatch q |
| 1934 | * @base: The IB1 base during the fault |
| 1935 | */ |
| 1936 | static void recover_dispatch_q(struct kgsl_device *device, |
| 1937 | struct adreno_dispatcher_drawqueue *dispatch_q, |
| 1938 | int fault, |
| 1939 | unsigned int base) |
| 1940 | { |
| 1941 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 1942 | struct kgsl_drawobj_cmd **replay; |
| 1943 | unsigned int ptr; |
| 1944 | int first = 0; |
| 1945 | int count = 0; |
| 1946 | int i; |
| 1947 | |
| 1948 | /* Allocate memory to store the inflight commands */ |
| 1949 | replay = kcalloc(dispatch_q->inflight, sizeof(*replay), GFP_KERNEL); |
| 1950 | |
| 1951 | if (replay == NULL) { |
| 1952 | unsigned int ptr = dispatch_q->head; |
| 1953 | |
| 1954 | /* Recovery failed - mark everybody on this q guilty */ |
| 1955 | while (ptr != dispatch_q->tail) { |
| 1956 | struct kgsl_drawobj_cmd *cmdobj = |
| 1957 | dispatch_q->cmd_q[ptr]; |
| 1958 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 1959 | |
| 1960 | mark_guilty_context(device, drawobj->context->id); |
| 1961 | adreno_drawctxt_invalidate(device, drawobj->context); |
| 1962 | kgsl_drawobj_destroy(drawobj); |
| 1963 | |
| 1964 | ptr = DRAWQUEUE_NEXT(ptr, |
| 1965 | ADRENO_DISPATCH_DRAWQUEUE_SIZE); |
| 1966 | } |
| 1967 | |
| 1968 | /* |
| 1969 | * Set the replay count to zero - this will ensure that the |
| 1970 | * hardware gets reset but nothing else gets played |
| 1971 | */ |
| 1972 | |
| 1973 | count = 0; |
| 1974 | goto replay; |
| 1975 | } |
| 1976 | |
| 1977 | /* Copy the inflight cmdobj's into the temporary storage */ |
| 1978 | ptr = dispatch_q->head; |
| 1979 | |
| 1980 | while (ptr != dispatch_q->tail) { |
| 1981 | replay[count++] = dispatch_q->cmd_q[ptr]; |
| 1982 | ptr = DRAWQUEUE_NEXT(ptr, ADRENO_DISPATCH_DRAWQUEUE_SIZE); |
| 1983 | } |
| 1984 | |
| 1985 | if (fault && count) |
| 1986 | process_cmdobj_fault(device, replay, |
| 1987 | count, base, fault); |
| 1988 | replay: |
| 1989 | dispatch_q->inflight = 0; |
| 1990 | dispatch_q->head = dispatch_q->tail = 0; |
| 1991 | /* Remove any pending cmdobj's that have been invalidated */ |
| 1992 | remove_invalidated_cmdobjs(device, replay, count); |
| 1993 | |
| 1994 | /* Replay the pending command buffers */ |
| 1995 | for (i = 0; i < count; i++) { |
| 1996 | |
| 1997 | int ret; |
| 1998 | |
| 1999 | if (replay[i] == NULL) |
| 2000 | continue; |
| 2001 | |
| 2002 | /* |
| 2003 | * Force the preamble on the first command (if applicable) to |
| 2004 | * avoid any strange stage issues |
| 2005 | */ |
| 2006 | |
| 2007 | if (first == 0) { |
| 2008 | set_bit(CMDOBJ_FORCE_PREAMBLE, &replay[i]->priv); |
| 2009 | first = 1; |
| 2010 | } |
| 2011 | |
| 2012 | /* |
| 2013 | * Force each cmdobj to wait for idle - this avoids weird |
| 2014 | * CP parse issues |
| 2015 | */ |
| 2016 | |
| 2017 | set_bit(CMDOBJ_WFI, &replay[i]->priv); |
| 2018 | |
| 2019 | ret = sendcmd(adreno_dev, replay[i]); |
| 2020 | |
| 2021 | /* |
| 2022 | * If sending the command fails, then try to recover by |
| 2023 | * invalidating the context |
| 2024 | */ |
| 2025 | |
| 2026 | if (ret) { |
| 2027 | pr_context(device, replay[i]->base.context, |
| 2028 | "gpu reset failed ctx %d ts %d\n", |
| 2029 | replay[i]->base.context->id, |
| 2030 | replay[i]->base.timestamp); |
| 2031 | |
| 2032 | /* Mark this context as guilty (failed recovery) */ |
| 2033 | mark_guilty_context(device, |
| 2034 | replay[i]->base.context->id); |
| 2035 | |
| 2036 | adreno_drawctxt_invalidate(device, |
| 2037 | replay[i]->base.context); |
| 2038 | remove_invalidated_cmdobjs(device, &replay[i], |
| 2039 | count - i); |
| 2040 | } |
| 2041 | } |
| 2042 | |
| 2043 | /* Clear the fault bit */ |
| 2044 | clear_bit(ADRENO_DEVICE_FAULT, &adreno_dev->priv); |
| 2045 | |
| 2046 | kfree(replay); |
| 2047 | } |
| 2048 | |
| 2049 | static void do_header_and_snapshot(struct kgsl_device *device, |
| 2050 | struct adreno_ringbuffer *rb, struct kgsl_drawobj_cmd *cmdobj) |
| 2051 | { |
| 2052 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 2053 | |
| 2054 | /* Always dump the snapshot on a non-drawobj failure */ |
| 2055 | if (cmdobj == NULL) { |
| 2056 | adreno_fault_header(device, rb, NULL); |
| 2057 | kgsl_device_snapshot(device, NULL); |
| 2058 | return; |
| 2059 | } |
| 2060 | |
| 2061 | /* Skip everything if the PMDUMP flag is set */ |
| 2062 | if (test_bit(KGSL_FT_SKIP_PMDUMP, &cmdobj->fault_policy)) |
| 2063 | return; |
| 2064 | |
| 2065 | /* Print the fault header */ |
| 2066 | adreno_fault_header(device, rb, cmdobj); |
| 2067 | |
| 2068 | if (!(drawobj->context->flags & KGSL_CONTEXT_NO_SNAPSHOT)) |
| 2069 | kgsl_device_snapshot(device, drawobj->context); |
| 2070 | } |
| 2071 | |
| 2072 | static int dispatcher_do_fault(struct adreno_device *adreno_dev) |
| 2073 | { |
| 2074 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
Carter Cooper | df7ba70 | 2017-03-20 11:28:04 -0600 | [diff] [blame] | 2075 | struct adreno_gpudev *gpudev = ADRENO_GPU_DEVICE(adreno_dev); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2076 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2077 | struct adreno_dispatcher_drawqueue *dispatch_q = NULL, *dispatch_q_temp; |
| 2078 | struct adreno_ringbuffer *rb; |
| 2079 | struct adreno_ringbuffer *hung_rb = NULL; |
| 2080 | unsigned int reg; |
| 2081 | uint64_t base; |
| 2082 | struct kgsl_drawobj_cmd *cmdobj = NULL; |
| 2083 | int ret, i; |
| 2084 | int fault; |
| 2085 | int halt; |
Shrenuj Bansal | d197bf6 | 2017-04-07 11:00:09 -0700 | [diff] [blame] | 2086 | bool gx_on = true; |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2087 | |
| 2088 | fault = atomic_xchg(&dispatcher->fault, 0); |
| 2089 | if (fault == 0) |
| 2090 | return 0; |
| 2091 | |
Shrenuj Bansal | d197bf6 | 2017-04-07 11:00:09 -0700 | [diff] [blame] | 2092 | /* Mask all GMU interrupts */ |
| 2093 | if (kgsl_gmu_isenabled(device)) { |
| 2094 | adreno_write_gmureg(adreno_dev, |
| 2095 | ADRENO_REG_GMU_AO_HOST_INTERRUPT_MASK, |
| 2096 | 0xFFFFFFFF); |
| 2097 | adreno_write_gmureg(adreno_dev, |
| 2098 | ADRENO_REG_GMU_GMU2HOST_INTR_MASK, |
| 2099 | 0xFFFFFFFF); |
| 2100 | } |
| 2101 | |
| 2102 | if (gpudev->gx_is_on) |
| 2103 | gx_on = gpudev->gx_is_on(adreno_dev); |
| 2104 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2105 | /* |
Hareesh Gundu | a2fe6ec | 2017-03-06 14:53:36 +0530 | [diff] [blame] | 2106 | * In the very unlikely case that the power is off, do nothing - the |
| 2107 | * state will be reset on power up and everybody will be happy |
| 2108 | */ |
| 2109 | |
| 2110 | if (!kgsl_state_is_awake(device) && (fault & ADRENO_SOFT_FAULT)) { |
| 2111 | /* Clear the existing register values */ |
| 2112 | memset(adreno_ft_regs_val, 0, |
| 2113 | adreno_ft_regs_num * sizeof(unsigned int)); |
| 2114 | return 0; |
| 2115 | } |
| 2116 | |
| 2117 | /* |
Lynus Vaz | 8402073 | 2017-03-27 18:12:24 +0530 | [diff] [blame] | 2118 | * On A5xx and A6xx, read RBBM_STATUS3:SMMU_STALLED_ON_FAULT (BIT 24) |
| 2119 | * to tell if this function was entered after a pagefault. If so, only |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2120 | * proceed if the fault handler has already run in the IRQ thread, |
| 2121 | * else return early to give the fault handler a chance to run. |
| 2122 | */ |
Lynus Vaz | 8402073 | 2017-03-27 18:12:24 +0530 | [diff] [blame] | 2123 | if (!(fault & ADRENO_IOMMU_PAGE_FAULT) && |
Shrenuj Bansal | d197bf6 | 2017-04-07 11:00:09 -0700 | [diff] [blame] | 2124 | (adreno_is_a5xx(adreno_dev) || adreno_is_a6xx(adreno_dev)) && |
| 2125 | gx_on) { |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2126 | unsigned int val; |
| 2127 | |
| 2128 | mutex_lock(&device->mutex); |
| 2129 | adreno_readreg(adreno_dev, ADRENO_REG_RBBM_STATUS3, &val); |
| 2130 | mutex_unlock(&device->mutex); |
| 2131 | if (val & BIT(24)) |
| 2132 | return 0; |
| 2133 | } |
| 2134 | |
| 2135 | /* Turn off all the timers */ |
| 2136 | del_timer_sync(&dispatcher->timer); |
| 2137 | del_timer_sync(&dispatcher->fault_timer); |
Wenbin Wang | 9361513 | 2017-03-30 14:10:19 +0800 | [diff] [blame] | 2138 | /* |
| 2139 | * Deleting uninitialized timer will block for ever on kernel debug |
| 2140 | * disable build. Hence skip del timer if it is not initialized. |
| 2141 | */ |
| 2142 | if (adreno_is_preemption_enabled(adreno_dev)) |
| 2143 | del_timer_sync(&adreno_dev->preempt.timer); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2144 | |
| 2145 | mutex_lock(&device->mutex); |
| 2146 | |
Shrenuj Bansal | d197bf6 | 2017-04-07 11:00:09 -0700 | [diff] [blame] | 2147 | if (gx_on) |
| 2148 | adreno_readreg64(adreno_dev, ADRENO_REG_CP_RB_BASE, |
| 2149 | ADRENO_REG_CP_RB_BASE_HI, &base); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2150 | |
| 2151 | /* |
| 2152 | * Force the CP off for anything but a hard fault to make sure it is |
| 2153 | * good and stopped |
| 2154 | */ |
Shrenuj Bansal | d197bf6 | 2017-04-07 11:00:09 -0700 | [diff] [blame] | 2155 | if (!(fault & ADRENO_HARD_FAULT) && gx_on) { |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2156 | adreno_readreg(adreno_dev, ADRENO_REG_CP_ME_CNTL, ®); |
Lynus Vaz | 8402073 | 2017-03-27 18:12:24 +0530 | [diff] [blame] | 2157 | if (adreno_is_a5xx(adreno_dev) || adreno_is_a6xx(adreno_dev)) |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2158 | reg |= 1 | (1 << 1); |
| 2159 | else |
| 2160 | reg |= (1 << 27) | (1 << 28); |
| 2161 | adreno_writereg(adreno_dev, ADRENO_REG_CP_ME_CNTL, reg); |
| 2162 | } |
| 2163 | /* |
| 2164 | * retire cmdobj's from all the dispatch_q's before starting recovery |
| 2165 | */ |
| 2166 | FOR_EACH_RINGBUFFER(adreno_dev, rb, i) { |
| 2167 | adreno_dispatch_retire_drawqueue(adreno_dev, |
| 2168 | &(rb->dispatch_q)); |
| 2169 | /* Select the active dispatch_q */ |
| 2170 | if (base == rb->buffer_desc.gpuaddr) { |
| 2171 | dispatch_q = &(rb->dispatch_q); |
| 2172 | hung_rb = rb; |
| 2173 | if (adreno_dev->cur_rb != hung_rb) { |
| 2174 | adreno_dev->prev_rb = adreno_dev->cur_rb; |
| 2175 | adreno_dev->cur_rb = hung_rb; |
| 2176 | } |
| 2177 | } |
| 2178 | if (ADRENO_DISPATCHER_RB_STARVE_TIMER_ELAPSED == |
| 2179 | rb->starve_timer_state) { |
| 2180 | adreno_put_gpu_halt(adreno_dev); |
| 2181 | rb->starve_timer_state = |
| 2182 | ADRENO_DISPATCHER_RB_STARVE_TIMER_UNINIT; |
| 2183 | } |
| 2184 | } |
| 2185 | |
| 2186 | if (dispatch_q && !adreno_drawqueue_is_empty(dispatch_q)) { |
| 2187 | cmdobj = dispatch_q->cmd_q[dispatch_q->head]; |
| 2188 | trace_adreno_cmdbatch_fault(cmdobj, fault); |
| 2189 | } |
| 2190 | |
Shrenuj Bansal | d197bf6 | 2017-04-07 11:00:09 -0700 | [diff] [blame] | 2191 | if (gx_on) |
| 2192 | adreno_readreg64(adreno_dev, ADRENO_REG_CP_IB1_BASE, |
| 2193 | ADRENO_REG_CP_IB1_BASE_HI, &base); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2194 | |
| 2195 | do_header_and_snapshot(device, hung_rb, cmdobj); |
| 2196 | |
Carter Cooper | df7ba70 | 2017-03-20 11:28:04 -0600 | [diff] [blame] | 2197 | /* Turn off the KEEPALIVE vote from the ISR for hard fault */ |
| 2198 | if (gpudev->gpu_keepalive && fault & ADRENO_HARD_FAULT) |
| 2199 | gpudev->gpu_keepalive(adreno_dev, false); |
| 2200 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2201 | /* Terminate the stalled transaction and resume the IOMMU */ |
| 2202 | if (fault & ADRENO_IOMMU_PAGE_FAULT) |
| 2203 | kgsl_mmu_pagefault_resume(&device->mmu); |
| 2204 | |
| 2205 | /* Reset the dispatcher queue */ |
| 2206 | dispatcher->inflight = 0; |
| 2207 | |
| 2208 | /* Reset the GPU and make sure halt is not set during recovery */ |
| 2209 | halt = adreno_gpu_halt(adreno_dev); |
| 2210 | adreno_clear_gpu_halt(adreno_dev); |
| 2211 | |
| 2212 | /* |
| 2213 | * If there is a stall in the ringbuffer after all commands have been |
| 2214 | * retired then we could hit problems if contexts are waiting for |
| 2215 | * internal timestamps that will never retire |
| 2216 | */ |
| 2217 | |
| 2218 | if (hung_rb != NULL) { |
| 2219 | kgsl_sharedmem_writel(device, &device->memstore, |
| 2220 | MEMSTORE_RB_OFFSET(hung_rb, soptimestamp), |
| 2221 | hung_rb->timestamp); |
| 2222 | |
| 2223 | kgsl_sharedmem_writel(device, &device->memstore, |
| 2224 | MEMSTORE_RB_OFFSET(hung_rb, eoptimestamp), |
| 2225 | hung_rb->timestamp); |
| 2226 | |
| 2227 | /* Schedule any pending events to be run */ |
| 2228 | kgsl_process_event_group(device, &hung_rb->events); |
| 2229 | } |
| 2230 | |
Shrenuj Bansal | d0fe746 | 2017-05-08 16:11:19 -0700 | [diff] [blame] | 2231 | if (gpudev->reset) |
| 2232 | ret = gpudev->reset(device, fault); |
| 2233 | else |
| 2234 | ret = adreno_reset(device, fault); |
| 2235 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2236 | mutex_unlock(&device->mutex); |
| 2237 | /* if any other fault got in until reset then ignore */ |
| 2238 | atomic_set(&dispatcher->fault, 0); |
| 2239 | |
| 2240 | /* If adreno_reset() fails then what hope do we have for the future? */ |
| 2241 | BUG_ON(ret); |
| 2242 | |
| 2243 | /* recover all the dispatch_q's starting with the one that hung */ |
| 2244 | if (dispatch_q) |
| 2245 | recover_dispatch_q(device, dispatch_q, fault, base); |
| 2246 | FOR_EACH_RINGBUFFER(adreno_dev, rb, i) { |
| 2247 | dispatch_q_temp = &(rb->dispatch_q); |
| 2248 | if (dispatch_q_temp != dispatch_q) |
| 2249 | recover_dispatch_q(device, dispatch_q_temp, 0, base); |
| 2250 | } |
| 2251 | |
| 2252 | atomic_add(halt, &adreno_dev->halt); |
| 2253 | |
| 2254 | return 1; |
| 2255 | } |
| 2256 | |
| 2257 | static inline int drawobj_consumed(struct kgsl_drawobj *drawobj, |
| 2258 | unsigned int consumed, unsigned int retired) |
| 2259 | { |
| 2260 | return ((timestamp_cmp(drawobj->timestamp, consumed) >= 0) && |
| 2261 | (timestamp_cmp(retired, drawobj->timestamp) < 0)); |
| 2262 | } |
| 2263 | |
| 2264 | static void _print_recovery(struct kgsl_device *device, |
| 2265 | struct kgsl_drawobj_cmd *cmdobj) |
| 2266 | { |
| 2267 | static struct { |
| 2268 | unsigned int mask; |
| 2269 | const char *str; |
| 2270 | } flags[] = { ADRENO_FT_TYPES }; |
| 2271 | |
| 2272 | int i, nr = find_first_bit(&cmdobj->fault_recovery, BITS_PER_LONG); |
| 2273 | char *result = "unknown"; |
| 2274 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 2275 | |
| 2276 | for (i = 0; i < ARRAY_SIZE(flags); i++) { |
| 2277 | if (flags[i].mask == BIT(nr)) { |
| 2278 | result = (char *) flags[i].str; |
| 2279 | break; |
| 2280 | } |
| 2281 | } |
| 2282 | |
| 2283 | pr_context(device, drawobj->context, |
| 2284 | "gpu %s ctx %d ts %d policy %lX\n", |
| 2285 | result, drawobj->context->id, drawobj->timestamp, |
| 2286 | cmdobj->fault_recovery); |
| 2287 | } |
| 2288 | |
| 2289 | static void cmdobj_profile_ticks(struct adreno_device *adreno_dev, |
| 2290 | struct kgsl_drawobj_cmd *cmdobj, uint64_t *start, uint64_t *retire) |
| 2291 | { |
| 2292 | void *ptr = adreno_dev->profile_buffer.hostptr; |
| 2293 | struct adreno_drawobj_profile_entry *entry; |
| 2294 | |
| 2295 | entry = (struct adreno_drawobj_profile_entry *) |
| 2296 | (ptr + (cmdobj->profile_index * sizeof(*entry))); |
| 2297 | |
| 2298 | /* get updated values of started and retired */ |
| 2299 | rmb(); |
| 2300 | *start = entry->started; |
| 2301 | *retire = entry->retired; |
| 2302 | } |
| 2303 | |
| 2304 | static void retire_cmdobj(struct adreno_device *adreno_dev, |
| 2305 | struct kgsl_drawobj_cmd *cmdobj) |
| 2306 | { |
| 2307 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2308 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 2309 | struct adreno_context *drawctxt = ADRENO_CONTEXT(drawobj->context); |
| 2310 | uint64_t start = 0, end = 0; |
| 2311 | |
| 2312 | if (cmdobj->fault_recovery != 0) { |
| 2313 | set_bit(ADRENO_CONTEXT_FAULT, &drawobj->context->priv); |
| 2314 | _print_recovery(KGSL_DEVICE(adreno_dev), cmdobj); |
| 2315 | } |
| 2316 | |
| 2317 | if (test_bit(CMDOBJ_PROFILE, &cmdobj->priv)) |
| 2318 | cmdobj_profile_ticks(adreno_dev, cmdobj, &start, &end); |
| 2319 | |
| 2320 | /* |
| 2321 | * For A3xx we still get the rptr from the CP_RB_RPTR instead of |
| 2322 | * rptr scratch out address. At this point GPU clocks turned off. |
| 2323 | * So avoid reading GPU register directly for A3xx. |
| 2324 | */ |
| 2325 | if (adreno_is_a3xx(adreno_dev)) |
| 2326 | trace_adreno_cmdbatch_retired(drawobj, |
| 2327 | (int) dispatcher->inflight, start, end, |
| 2328 | ADRENO_DRAWOBJ_RB(drawobj), 0, cmdobj->fault_recovery); |
| 2329 | else |
| 2330 | trace_adreno_cmdbatch_retired(drawobj, |
| 2331 | (int) dispatcher->inflight, start, end, |
| 2332 | ADRENO_DRAWOBJ_RB(drawobj), |
| 2333 | adreno_get_rptr(drawctxt->rb), cmdobj->fault_recovery); |
| 2334 | |
| 2335 | drawctxt->submit_retire_ticks[drawctxt->ticks_index] = |
| 2336 | end - cmdobj->submit_ticks; |
| 2337 | |
| 2338 | drawctxt->ticks_index = (drawctxt->ticks_index + 1) % |
| 2339 | SUBMIT_RETIRE_TICKS_SIZE; |
| 2340 | |
| 2341 | kgsl_drawobj_destroy(drawobj); |
| 2342 | } |
| 2343 | |
| 2344 | static int adreno_dispatch_retire_drawqueue(struct adreno_device *adreno_dev, |
| 2345 | struct adreno_dispatcher_drawqueue *drawqueue) |
| 2346 | { |
| 2347 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2348 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2349 | int count = 0; |
| 2350 | |
| 2351 | while (!adreno_drawqueue_is_empty(drawqueue)) { |
| 2352 | struct kgsl_drawobj_cmd *cmdobj = |
| 2353 | drawqueue->cmd_q[drawqueue->head]; |
| 2354 | struct kgsl_drawobj *drawobj = DRAWOBJ(cmdobj); |
| 2355 | |
| 2356 | if (!kgsl_check_timestamp(device, drawobj->context, |
| 2357 | drawobj->timestamp)) |
| 2358 | break; |
| 2359 | |
| 2360 | retire_cmdobj(adreno_dev, cmdobj); |
| 2361 | |
| 2362 | dispatcher->inflight--; |
| 2363 | drawqueue->inflight--; |
| 2364 | |
| 2365 | drawqueue->cmd_q[drawqueue->head] = NULL; |
| 2366 | |
| 2367 | drawqueue->head = DRAWQUEUE_NEXT(drawqueue->head, |
| 2368 | ADRENO_DISPATCH_DRAWQUEUE_SIZE); |
| 2369 | |
| 2370 | count++; |
| 2371 | } |
| 2372 | |
| 2373 | return count; |
| 2374 | } |
| 2375 | |
| 2376 | static void _adreno_dispatch_check_timeout(struct adreno_device *adreno_dev, |
| 2377 | struct adreno_dispatcher_drawqueue *drawqueue) |
| 2378 | { |
| 2379 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2380 | struct kgsl_drawobj *drawobj = |
| 2381 | DRAWOBJ(drawqueue->cmd_q[drawqueue->head]); |
| 2382 | |
| 2383 | /* Don't timeout if the timer hasn't expired yet (duh) */ |
| 2384 | if (time_is_after_jiffies(drawqueue->expires)) |
| 2385 | return; |
| 2386 | |
| 2387 | /* Don't timeout if the IB timeout is disabled globally */ |
| 2388 | if (!adreno_long_ib_detect(adreno_dev)) |
| 2389 | return; |
| 2390 | |
| 2391 | /* Don't time out if the context has disabled it */ |
| 2392 | if (drawobj->context->flags & KGSL_CONTEXT_NO_FAULT_TOLERANCE) |
| 2393 | return; |
| 2394 | |
| 2395 | pr_context(device, drawobj->context, "gpu timeout ctx %d ts %d\n", |
| 2396 | drawobj->context->id, drawobj->timestamp); |
| 2397 | |
| 2398 | adreno_set_gpu_fault(adreno_dev, ADRENO_TIMEOUT_FAULT); |
| 2399 | } |
| 2400 | |
| 2401 | static int adreno_dispatch_process_drawqueue(struct adreno_device *adreno_dev, |
| 2402 | struct adreno_dispatcher_drawqueue *drawqueue) |
| 2403 | { |
| 2404 | int count = adreno_dispatch_retire_drawqueue(adreno_dev, drawqueue); |
| 2405 | |
| 2406 | /* Nothing to do if there are no pending commands */ |
| 2407 | if (adreno_drawqueue_is_empty(drawqueue)) |
| 2408 | return count; |
| 2409 | |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2410 | /* Don't update the drawqueue timeout if it isn't active */ |
| 2411 | if (!drawqueue_is_current(drawqueue)) |
| 2412 | return count; |
| 2413 | |
| 2414 | /* |
| 2415 | * If the current ringbuffer retired any commands then universally |
| 2416 | * reset the timeout |
| 2417 | */ |
| 2418 | |
| 2419 | if (count) { |
| 2420 | drawqueue->expires = jiffies + |
| 2421 | msecs_to_jiffies(adreno_drawobj_timeout); |
| 2422 | return count; |
| 2423 | } |
| 2424 | |
| 2425 | /* |
| 2426 | * If we get here then 1) the ringbuffer is current and 2) we haven't |
| 2427 | * retired anything. Check to see if the timeout if valid for the |
| 2428 | * current drawobj and fault if it has expired |
| 2429 | */ |
| 2430 | _adreno_dispatch_check_timeout(adreno_dev, drawqueue); |
| 2431 | return 0; |
| 2432 | } |
| 2433 | |
| 2434 | /* Update the dispatcher timers */ |
| 2435 | static void _dispatcher_update_timers(struct adreno_device *adreno_dev) |
| 2436 | { |
| 2437 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2438 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2439 | |
| 2440 | /* Kick the idle timer */ |
| 2441 | mutex_lock(&device->mutex); |
| 2442 | kgsl_pwrscale_update(device); |
| 2443 | mod_timer(&device->idle_timer, |
| 2444 | jiffies + device->pwrctrl.interval_timeout); |
| 2445 | mutex_unlock(&device->mutex); |
| 2446 | |
| 2447 | /* Check to see if we need to update the command timer */ |
| 2448 | if (adreno_in_preempt_state(adreno_dev, ADRENO_PREEMPT_NONE)) { |
| 2449 | struct adreno_dispatcher_drawqueue *drawqueue = |
| 2450 | DRAWQUEUE(adreno_dev->cur_rb); |
| 2451 | |
| 2452 | if (!adreno_drawqueue_is_empty(drawqueue)) |
| 2453 | mod_timer(&dispatcher->timer, drawqueue->expires); |
| 2454 | } |
| 2455 | } |
| 2456 | |
| 2457 | /* Take down the dispatcher and release any power states */ |
| 2458 | static void _dispatcher_power_down(struct adreno_device *adreno_dev) |
| 2459 | { |
| 2460 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2461 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2462 | |
| 2463 | mutex_lock(&device->mutex); |
| 2464 | |
| 2465 | if (test_and_clear_bit(ADRENO_DISPATCHER_ACTIVE, &dispatcher->priv)) |
| 2466 | complete_all(&dispatcher->idle_gate); |
| 2467 | |
| 2468 | del_timer_sync(&dispatcher->fault_timer); |
| 2469 | |
| 2470 | if (test_bit(ADRENO_DISPATCHER_POWER, &dispatcher->priv)) { |
| 2471 | kgsl_active_count_put(device); |
| 2472 | clear_bit(ADRENO_DISPATCHER_POWER, &dispatcher->priv); |
| 2473 | } |
| 2474 | |
| 2475 | mutex_unlock(&device->mutex); |
| 2476 | } |
| 2477 | |
Tim Murray | 8504043 | 2017-02-20 15:59:32 +0530 | [diff] [blame] | 2478 | static void adreno_dispatcher_work(struct kthread_work *work) |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2479 | { |
| 2480 | struct adreno_dispatcher *dispatcher = |
| 2481 | container_of(work, struct adreno_dispatcher, work); |
| 2482 | struct adreno_device *adreno_dev = |
| 2483 | container_of(dispatcher, struct adreno_device, dispatcher); |
| 2484 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2485 | struct adreno_gpudev *gpudev = ADRENO_GPU_DEVICE(adreno_dev); |
| 2486 | int count = 0; |
| 2487 | unsigned int i = 0; |
| 2488 | |
| 2489 | mutex_lock(&dispatcher->mutex); |
| 2490 | |
| 2491 | /* |
| 2492 | * As long as there are inflight commands, process retired comamnds from |
| 2493 | * all drawqueues |
| 2494 | */ |
| 2495 | for (i = 0; i < adreno_dev->num_ringbuffers; i++) { |
| 2496 | struct adreno_dispatcher_drawqueue *drawqueue = |
| 2497 | DRAWQUEUE(&adreno_dev->ringbuffers[i]); |
| 2498 | |
| 2499 | count += adreno_dispatch_process_drawqueue(adreno_dev, |
| 2500 | drawqueue); |
| 2501 | if (dispatcher->inflight == 0) |
| 2502 | break; |
| 2503 | } |
| 2504 | |
| 2505 | kgsl_process_event_groups(device); |
| 2506 | |
| 2507 | /* |
| 2508 | * dispatcher_do_fault() returns 0 if no faults occurred. If that is the |
| 2509 | * case, then clean up preemption and try to schedule more work |
| 2510 | */ |
| 2511 | if (dispatcher_do_fault(adreno_dev) == 0) { |
| 2512 | |
| 2513 | /* Clean up after preemption */ |
| 2514 | if (gpudev->preemption_schedule) |
| 2515 | gpudev->preemption_schedule(adreno_dev); |
| 2516 | |
| 2517 | /* Run the scheduler for to dispatch new commands */ |
| 2518 | _adreno_dispatcher_issuecmds(adreno_dev); |
| 2519 | } |
| 2520 | |
| 2521 | /* |
| 2522 | * If there are commands pending, update the timers, otherwise release |
| 2523 | * the power state to prepare for power down |
| 2524 | */ |
| 2525 | if (dispatcher->inflight > 0) |
| 2526 | _dispatcher_update_timers(adreno_dev); |
| 2527 | else |
| 2528 | _dispatcher_power_down(adreno_dev); |
| 2529 | |
| 2530 | mutex_unlock(&dispatcher->mutex); |
| 2531 | } |
| 2532 | |
| 2533 | void adreno_dispatcher_schedule(struct kgsl_device *device) |
| 2534 | { |
| 2535 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 2536 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2537 | |
Tim Murray | 8504043 | 2017-02-20 15:59:32 +0530 | [diff] [blame] | 2538 | kthread_queue_work(&kgsl_driver.worker, &dispatcher->work); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2539 | } |
| 2540 | |
| 2541 | /** |
| 2542 | * adreno_dispatcher_queue_context() - schedule a drawctxt in the dispatcher |
| 2543 | * device: pointer to the KGSL device |
| 2544 | * drawctxt: pointer to the drawctxt to schedule |
| 2545 | * |
| 2546 | * Put a draw context on the dispatcher pending queue and schedule the |
| 2547 | * dispatcher. This is used to reschedule changes that might have been blocked |
| 2548 | * for sync points or other concerns |
| 2549 | */ |
| 2550 | void adreno_dispatcher_queue_context(struct kgsl_device *device, |
| 2551 | struct adreno_context *drawctxt) |
| 2552 | { |
| 2553 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 2554 | |
| 2555 | dispatcher_queue_context(adreno_dev, drawctxt); |
| 2556 | adreno_dispatcher_schedule(device); |
| 2557 | } |
| 2558 | |
| 2559 | /* |
| 2560 | * This is called on a regular basis while cmdobj's are inflight. Fault |
| 2561 | * detection registers are read and compared to the existing values - if they |
| 2562 | * changed then the GPU is still running. If they are the same between |
| 2563 | * subsequent calls then the GPU may have faulted |
| 2564 | */ |
| 2565 | |
| 2566 | static void adreno_dispatcher_fault_timer(unsigned long data) |
| 2567 | { |
| 2568 | struct adreno_device *adreno_dev = (struct adreno_device *) data; |
| 2569 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2570 | |
| 2571 | /* Leave if the user decided to turn off fast hang detection */ |
| 2572 | if (!adreno_soft_fault_detect(adreno_dev)) |
| 2573 | return; |
| 2574 | |
| 2575 | if (adreno_gpu_fault(adreno_dev)) { |
| 2576 | adreno_dispatcher_schedule(KGSL_DEVICE(adreno_dev)); |
| 2577 | return; |
| 2578 | } |
| 2579 | |
| 2580 | /* |
| 2581 | * Read the fault registers - if it returns 0 then they haven't changed |
| 2582 | * so mark the dispatcher as faulted and schedule the work loop. |
| 2583 | */ |
| 2584 | |
| 2585 | if (!fault_detect_read_compare(adreno_dev)) { |
| 2586 | adreno_set_gpu_fault(adreno_dev, ADRENO_SOFT_FAULT); |
| 2587 | adreno_dispatcher_schedule(KGSL_DEVICE(adreno_dev)); |
Hareesh Gundu | a2fe6ec | 2017-03-06 14:53:36 +0530 | [diff] [blame] | 2588 | } else if (dispatcher->inflight > 0) { |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2589 | mod_timer(&dispatcher->fault_timer, |
| 2590 | jiffies + msecs_to_jiffies(_fault_timer_interval)); |
| 2591 | } |
| 2592 | } |
| 2593 | |
| 2594 | /* |
| 2595 | * This is called when the timer expires - it either means the GPU is hung or |
| 2596 | * the IB is taking too long to execute |
| 2597 | */ |
| 2598 | static void adreno_dispatcher_timer(unsigned long data) |
| 2599 | { |
| 2600 | struct adreno_device *adreno_dev = (struct adreno_device *) data; |
| 2601 | |
| 2602 | adreno_dispatcher_schedule(KGSL_DEVICE(adreno_dev)); |
| 2603 | } |
| 2604 | |
| 2605 | /** |
| 2606 | * adreno_dispatcher_start() - activate the dispatcher |
| 2607 | * @adreno_dev: pointer to the adreno device structure |
| 2608 | * |
| 2609 | */ |
| 2610 | void adreno_dispatcher_start(struct kgsl_device *device) |
| 2611 | { |
| 2612 | complete_all(&device->halt_gate); |
| 2613 | |
| 2614 | /* Schedule the work loop to get things going */ |
| 2615 | adreno_dispatcher_schedule(device); |
| 2616 | } |
| 2617 | |
| 2618 | /** |
| 2619 | * adreno_dispatcher_stop() - stop the dispatcher |
| 2620 | * @adreno_dev: pointer to the adreno device structure |
| 2621 | * |
| 2622 | * Stop the dispatcher and close all the timers |
| 2623 | */ |
| 2624 | void adreno_dispatcher_stop(struct adreno_device *adreno_dev) |
| 2625 | { |
| 2626 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2627 | |
| 2628 | del_timer_sync(&dispatcher->timer); |
| 2629 | del_timer_sync(&dispatcher->fault_timer); |
| 2630 | } |
| 2631 | |
| 2632 | /** |
Hareesh Gundu | a2fe6ec | 2017-03-06 14:53:36 +0530 | [diff] [blame] | 2633 | * adreno_dispatcher_stop_fault_timer() - stop the dispatcher fault timer |
| 2634 | * @device: pointer to the KGSL device structure |
| 2635 | * |
| 2636 | * Stop the dispatcher fault timer |
| 2637 | */ |
| 2638 | void adreno_dispatcher_stop_fault_timer(struct kgsl_device *device) |
| 2639 | { |
| 2640 | struct adreno_device *adreno_dev = ADRENO_DEVICE(device); |
| 2641 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2642 | |
| 2643 | del_timer_sync(&dispatcher->fault_timer); |
| 2644 | } |
| 2645 | |
| 2646 | /** |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2647 | * adreno_dispatcher_close() - close the dispatcher |
| 2648 | * @adreno_dev: pointer to the adreno device structure |
| 2649 | * |
| 2650 | * Close the dispatcher and free all the oustanding commands and memory |
| 2651 | */ |
| 2652 | void adreno_dispatcher_close(struct adreno_device *adreno_dev) |
| 2653 | { |
| 2654 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2655 | int i; |
| 2656 | struct adreno_ringbuffer *rb; |
| 2657 | |
| 2658 | mutex_lock(&dispatcher->mutex); |
| 2659 | del_timer_sync(&dispatcher->timer); |
| 2660 | del_timer_sync(&dispatcher->fault_timer); |
| 2661 | |
| 2662 | FOR_EACH_RINGBUFFER(adreno_dev, rb, i) { |
| 2663 | struct adreno_dispatcher_drawqueue *dispatch_q = |
| 2664 | &(rb->dispatch_q); |
| 2665 | while (!adreno_drawqueue_is_empty(dispatch_q)) { |
| 2666 | kgsl_drawobj_destroy( |
| 2667 | DRAWOBJ(dispatch_q->cmd_q[dispatch_q->head])); |
| 2668 | dispatch_q->head = (dispatch_q->head + 1) |
| 2669 | % ADRENO_DISPATCH_DRAWQUEUE_SIZE; |
| 2670 | } |
| 2671 | } |
| 2672 | |
| 2673 | mutex_unlock(&dispatcher->mutex); |
| 2674 | |
| 2675 | kobject_put(&dispatcher->kobj); |
| 2676 | } |
| 2677 | |
| 2678 | struct dispatcher_attribute { |
| 2679 | struct attribute attr; |
| 2680 | ssize_t (*show)(struct adreno_dispatcher *, |
| 2681 | struct dispatcher_attribute *, char *); |
| 2682 | ssize_t (*store)(struct adreno_dispatcher *, |
| 2683 | struct dispatcher_attribute *, const char *buf, |
| 2684 | size_t count); |
| 2685 | unsigned int max; |
| 2686 | unsigned int *value; |
| 2687 | }; |
| 2688 | |
| 2689 | #define DISPATCHER_UINT_ATTR(_name, _mode, _max, _value) \ |
| 2690 | struct dispatcher_attribute dispatcher_attr_##_name = { \ |
| 2691 | .attr = { .name = __stringify(_name), .mode = _mode }, \ |
| 2692 | .show = _show_uint, \ |
| 2693 | .store = _store_uint, \ |
| 2694 | .max = _max, \ |
| 2695 | .value = &(_value), \ |
| 2696 | } |
| 2697 | |
| 2698 | #define to_dispatcher_attr(_a) \ |
| 2699 | container_of((_a), struct dispatcher_attribute, attr) |
| 2700 | #define to_dispatcher(k) container_of(k, struct adreno_dispatcher, kobj) |
| 2701 | |
| 2702 | static ssize_t _store_uint(struct adreno_dispatcher *dispatcher, |
| 2703 | struct dispatcher_attribute *attr, |
| 2704 | const char *buf, size_t size) |
| 2705 | { |
| 2706 | unsigned int val = 0; |
| 2707 | int ret; |
| 2708 | |
| 2709 | ret = kgsl_sysfs_store(buf, &val); |
| 2710 | if (ret) |
| 2711 | return ret; |
| 2712 | |
| 2713 | if (!val || (attr->max && (val > attr->max))) |
| 2714 | return -EINVAL; |
| 2715 | |
| 2716 | *((unsigned int *) attr->value) = val; |
| 2717 | return size; |
| 2718 | } |
| 2719 | |
| 2720 | static ssize_t _show_uint(struct adreno_dispatcher *dispatcher, |
| 2721 | struct dispatcher_attribute *attr, |
| 2722 | char *buf) |
| 2723 | { |
| 2724 | return snprintf(buf, PAGE_SIZE, "%u\n", |
| 2725 | *((unsigned int *) attr->value)); |
| 2726 | } |
| 2727 | |
| 2728 | static DISPATCHER_UINT_ATTR(inflight, 0644, ADRENO_DISPATCH_DRAWQUEUE_SIZE, |
| 2729 | _dispatcher_q_inflight_hi); |
| 2730 | |
| 2731 | static DISPATCHER_UINT_ATTR(inflight_low_latency, 0644, |
| 2732 | ADRENO_DISPATCH_DRAWQUEUE_SIZE, _dispatcher_q_inflight_lo); |
| 2733 | /* |
| 2734 | * Our code that "puts back" a command from the context is much cleaner |
| 2735 | * if we are sure that there will always be enough room in the |
| 2736 | * ringbuffer so restrict the maximum size of the context queue to |
| 2737 | * ADRENO_CONTEXT_DRAWQUEUE_SIZE - 1 |
| 2738 | */ |
| 2739 | static DISPATCHER_UINT_ATTR(context_drawqueue_size, 0644, |
| 2740 | ADRENO_CONTEXT_DRAWQUEUE_SIZE - 1, _context_drawqueue_size); |
| 2741 | static DISPATCHER_UINT_ATTR(context_burst_count, 0644, 0, |
| 2742 | _context_drawobj_burst); |
| 2743 | static DISPATCHER_UINT_ATTR(drawobj_timeout, 0644, 0, |
| 2744 | adreno_drawobj_timeout); |
| 2745 | static DISPATCHER_UINT_ATTR(context_queue_wait, 0644, 0, _context_queue_wait); |
| 2746 | static DISPATCHER_UINT_ATTR(fault_detect_interval, 0644, 0, |
| 2747 | _fault_timer_interval); |
| 2748 | static DISPATCHER_UINT_ATTR(fault_throttle_time, 0644, 0, |
| 2749 | _fault_throttle_time); |
| 2750 | static DISPATCHER_UINT_ATTR(fault_throttle_burst, 0644, 0, |
| 2751 | _fault_throttle_burst); |
| 2752 | static DISPATCHER_UINT_ATTR(disp_preempt_fair_sched, 0644, 0, |
| 2753 | adreno_disp_preempt_fair_sched); |
| 2754 | static DISPATCHER_UINT_ATTR(dispatch_time_slice, 0644, 0, |
| 2755 | adreno_dispatch_time_slice); |
| 2756 | static DISPATCHER_UINT_ATTR(dispatch_starvation_time, 0644, 0, |
| 2757 | adreno_dispatch_starvation_time); |
| 2758 | |
| 2759 | static struct attribute *dispatcher_attrs[] = { |
| 2760 | &dispatcher_attr_inflight.attr, |
| 2761 | &dispatcher_attr_inflight_low_latency.attr, |
| 2762 | &dispatcher_attr_context_drawqueue_size.attr, |
| 2763 | &dispatcher_attr_context_burst_count.attr, |
| 2764 | &dispatcher_attr_drawobj_timeout.attr, |
| 2765 | &dispatcher_attr_context_queue_wait.attr, |
| 2766 | &dispatcher_attr_fault_detect_interval.attr, |
| 2767 | &dispatcher_attr_fault_throttle_time.attr, |
| 2768 | &dispatcher_attr_fault_throttle_burst.attr, |
| 2769 | &dispatcher_attr_disp_preempt_fair_sched.attr, |
| 2770 | &dispatcher_attr_dispatch_time_slice.attr, |
| 2771 | &dispatcher_attr_dispatch_starvation_time.attr, |
| 2772 | NULL, |
| 2773 | }; |
| 2774 | |
| 2775 | static ssize_t dispatcher_sysfs_show(struct kobject *kobj, |
| 2776 | struct attribute *attr, char *buf) |
| 2777 | { |
| 2778 | struct adreno_dispatcher *dispatcher = to_dispatcher(kobj); |
| 2779 | struct dispatcher_attribute *pattr = to_dispatcher_attr(attr); |
| 2780 | ssize_t ret = -EIO; |
| 2781 | |
| 2782 | if (pattr->show) |
| 2783 | ret = pattr->show(dispatcher, pattr, buf); |
| 2784 | |
| 2785 | return ret; |
| 2786 | } |
| 2787 | |
| 2788 | static ssize_t dispatcher_sysfs_store(struct kobject *kobj, |
| 2789 | struct attribute *attr, |
| 2790 | const char *buf, size_t count) |
| 2791 | { |
| 2792 | struct adreno_dispatcher *dispatcher = to_dispatcher(kobj); |
| 2793 | struct dispatcher_attribute *pattr = to_dispatcher_attr(attr); |
| 2794 | ssize_t ret = -EIO; |
| 2795 | |
| 2796 | if (pattr->store) |
| 2797 | ret = pattr->store(dispatcher, pattr, buf, count); |
| 2798 | |
| 2799 | return ret; |
| 2800 | } |
| 2801 | |
| 2802 | static const struct sysfs_ops dispatcher_sysfs_ops = { |
| 2803 | .show = dispatcher_sysfs_show, |
| 2804 | .store = dispatcher_sysfs_store |
| 2805 | }; |
| 2806 | |
| 2807 | static struct kobj_type ktype_dispatcher = { |
| 2808 | .sysfs_ops = &dispatcher_sysfs_ops, |
| 2809 | .default_attrs = dispatcher_attrs, |
| 2810 | }; |
| 2811 | |
| 2812 | /** |
| 2813 | * adreno_dispatcher_init() - Initialize the dispatcher |
| 2814 | * @adreno_dev: pointer to the adreno device structure |
| 2815 | * |
| 2816 | * Initialize the dispatcher |
| 2817 | */ |
| 2818 | int adreno_dispatcher_init(struct adreno_device *adreno_dev) |
| 2819 | { |
| 2820 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2821 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2822 | int ret; |
| 2823 | |
| 2824 | memset(dispatcher, 0, sizeof(*dispatcher)); |
| 2825 | |
| 2826 | mutex_init(&dispatcher->mutex); |
| 2827 | |
| 2828 | setup_timer(&dispatcher->timer, adreno_dispatcher_timer, |
| 2829 | (unsigned long) adreno_dev); |
| 2830 | |
| 2831 | setup_timer(&dispatcher->fault_timer, adreno_dispatcher_fault_timer, |
| 2832 | (unsigned long) adreno_dev); |
| 2833 | |
Tim Murray | 8504043 | 2017-02-20 15:59:32 +0530 | [diff] [blame] | 2834 | kthread_init_work(&dispatcher->work, adreno_dispatcher_work); |
Shrenuj Bansal | a419c79 | 2016-10-20 14:05:11 -0700 | [diff] [blame] | 2835 | |
| 2836 | init_completion(&dispatcher->idle_gate); |
| 2837 | complete_all(&dispatcher->idle_gate); |
| 2838 | |
| 2839 | plist_head_init(&dispatcher->pending); |
| 2840 | spin_lock_init(&dispatcher->plist_lock); |
| 2841 | |
| 2842 | ret = kobject_init_and_add(&dispatcher->kobj, &ktype_dispatcher, |
| 2843 | &device->dev->kobj, "dispatch"); |
| 2844 | |
| 2845 | return ret; |
| 2846 | } |
| 2847 | |
| 2848 | /* |
| 2849 | * adreno_dispatcher_idle() - Wait for dispatcher to idle |
| 2850 | * @adreno_dev: Adreno device whose dispatcher needs to idle |
| 2851 | * |
| 2852 | * Signal dispatcher to stop sending more commands and complete |
| 2853 | * the commands that have already been submitted. This function |
| 2854 | * should not be called when dispatcher mutex is held. |
| 2855 | * The caller must hold the device mutex. |
| 2856 | */ |
| 2857 | int adreno_dispatcher_idle(struct adreno_device *adreno_dev) |
| 2858 | { |
| 2859 | struct kgsl_device *device = KGSL_DEVICE(adreno_dev); |
| 2860 | struct adreno_dispatcher *dispatcher = &adreno_dev->dispatcher; |
| 2861 | int ret; |
| 2862 | |
| 2863 | if (!test_bit(ADRENO_DEVICE_STARTED, &adreno_dev->priv)) |
| 2864 | return 0; |
| 2865 | |
| 2866 | /* |
| 2867 | * Ensure that this function is not called when dispatcher |
| 2868 | * mutex is held and device is started |
| 2869 | */ |
| 2870 | if (mutex_is_locked(&dispatcher->mutex) && |
| 2871 | dispatcher->mutex.owner == current) |
| 2872 | return -EDEADLK; |
| 2873 | |
| 2874 | adreno_get_gpu_halt(adreno_dev); |
| 2875 | |
| 2876 | mutex_unlock(&device->mutex); |
| 2877 | |
| 2878 | ret = wait_for_completion_timeout(&dispatcher->idle_gate, |
| 2879 | msecs_to_jiffies(ADRENO_IDLE_TIMEOUT)); |
| 2880 | if (ret == 0) { |
| 2881 | ret = -ETIMEDOUT; |
| 2882 | WARN(1, "Dispatcher halt timeout "); |
| 2883 | } else if (ret < 0) { |
| 2884 | KGSL_DRV_ERR(device, "Dispatcher halt failed %d\n", ret); |
| 2885 | } else { |
| 2886 | ret = 0; |
| 2887 | } |
| 2888 | |
| 2889 | mutex_lock(&device->mutex); |
| 2890 | adreno_put_gpu_halt(adreno_dev); |
| 2891 | /* |
| 2892 | * requeue dispatcher work to resubmit pending commands |
| 2893 | * that may have been blocked due to this idling request |
| 2894 | */ |
| 2895 | adreno_dispatcher_schedule(device); |
| 2896 | return ret; |
| 2897 | } |