1 /* 2 * Tegra host1x Syncpoints 3 * 4 * Copyright (c) 2010-2013, NVIDIA Corporation. 5 * 6 * This program is free software; you can redistribute it and/or modify it 7 * under the terms and conditions of the GNU General Public License, 8 * version 2, as published by the Free Software Foundation. 9 * 10 * This program is distributed in the hope it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 13 * more details. 14 * 15 * You should have received a copy of the GNU General Public License 16 * along with this program. If not, see <http://www.gnu.org/licenses/>. 17 */ 18 19 #include <linux/module.h> 20 #include <linux/device.h> 21 #include <linux/slab.h> 22 23 #include <trace/events/host1x.h> 24 25 #include "syncpt.h" 26 #include "dev.h" 27 #include "intr.h" 28 #include "debug.h" 29 30 #define SYNCPT_CHECK_PERIOD (2 * HZ) 31 #define MAX_STUCK_CHECK_COUNT 15 32 33 static struct host1x_syncpt_base * 34 host1x_syncpt_base_request(struct host1x *host) 35 { 36 struct host1x_syncpt_base *bases = host->bases; 37 unsigned int i; 38 39 for (i = 0; i < host->info->nb_bases; i++) 40 if (!bases[i].requested) 41 break; 42 43 if (i >= host->info->nb_bases) 44 return NULL; 45 46 bases[i].requested = true; 47 return &bases[i]; 48 } 49 50 static void host1x_syncpt_base_free(struct host1x_syncpt_base *base) 51 { 52 if (base) 53 base->requested = false; 54 } 55 56 static struct host1x_syncpt *host1x_syncpt_alloc(struct host1x *host, 57 struct device *dev, 58 unsigned long flags) 59 { 60 int i; 61 struct host1x_syncpt *sp = host->syncpt; 62 char *name; 63 64 for (i = 0; i < host->info->nb_pts && sp->name; i++, sp++) 65 ; 66 67 if (i >= host->info->nb_pts) 68 return NULL; 69 70 if (flags & HOST1X_SYNCPT_HAS_BASE) { 71 sp->base = host1x_syncpt_base_request(host); 72 if (!sp->base) 73 return NULL; 74 } 75 76 name = kasprintf(GFP_KERNEL, "%02d-%s", sp->id, 77 dev ? dev_name(dev) : NULL); 78 if (!name) 79 return NULL; 80 81 sp->dev = dev; 82 sp->name = name; 83 84 if (flags & HOST1X_SYNCPT_CLIENT_MANAGED) 85 sp->client_managed = true; 86 else 87 sp->client_managed = false; 88 89 return sp; 90 } 91 92 u32 host1x_syncpt_id(struct host1x_syncpt *sp) 93 { 94 return sp->id; 95 } 96 EXPORT_SYMBOL(host1x_syncpt_id); 97 98 /* 99 * Updates the value sent to hardware. 100 */ 101 u32 host1x_syncpt_incr_max(struct host1x_syncpt *sp, u32 incrs) 102 { 103 return (u32)atomic_add_return(incrs, &sp->max_val); 104 } 105 106 /* 107 * Write cached syncpoint and waitbase values to hardware. 108 */ 109 void host1x_syncpt_restore(struct host1x *host) 110 { 111 struct host1x_syncpt *sp_base = host->syncpt; 112 u32 i; 113 114 for (i = 0; i < host1x_syncpt_nb_pts(host); i++) 115 host1x_hw_syncpt_restore(host, sp_base + i); 116 for (i = 0; i < host1x_syncpt_nb_bases(host); i++) 117 host1x_hw_syncpt_restore_wait_base(host, sp_base + i); 118 wmb(); 119 } 120 121 /* 122 * Update the cached syncpoint and waitbase values by reading them 123 * from the registers. 124 */ 125 void host1x_syncpt_save(struct host1x *host) 126 { 127 struct host1x_syncpt *sp_base = host->syncpt; 128 u32 i; 129 130 for (i = 0; i < host1x_syncpt_nb_pts(host); i++) { 131 if (host1x_syncpt_client_managed(sp_base + i)) 132 host1x_hw_syncpt_load(host, sp_base + i); 133 else 134 WARN_ON(!host1x_syncpt_idle(sp_base + i)); 135 } 136 137 for (i = 0; i < host1x_syncpt_nb_bases(host); i++) 138 host1x_hw_syncpt_load_wait_base(host, sp_base + i); 139 } 140 141 /* 142 * Updates the cached syncpoint value by reading a new value from the hardware 143 * register 144 */ 145 u32 host1x_syncpt_load(struct host1x_syncpt *sp) 146 { 147 u32 val; 148 val = host1x_hw_syncpt_load(sp->host, sp); 149 trace_host1x_syncpt_load_min(sp->id, val); 150 151 return val; 152 } 153 154 /* 155 * Get the current syncpoint base 156 */ 157 u32 host1x_syncpt_load_wait_base(struct host1x_syncpt *sp) 158 { 159 u32 val; 160 host1x_hw_syncpt_load_wait_base(sp->host, sp); 161 val = sp->base_val; 162 return val; 163 } 164 165 /* 166 * Increment syncpoint value from cpu, updating cache 167 */ 168 int host1x_syncpt_incr(struct host1x_syncpt *sp) 169 { 170 return host1x_hw_syncpt_cpu_incr(sp->host, sp); 171 } 172 EXPORT_SYMBOL(host1x_syncpt_incr); 173 174 /* 175 * Updated sync point form hardware, and returns true if syncpoint is expired, 176 * false if we may need to wait 177 */ 178 static bool syncpt_load_min_is_expired(struct host1x_syncpt *sp, u32 thresh) 179 { 180 host1x_hw_syncpt_load(sp->host, sp); 181 return host1x_syncpt_is_expired(sp, thresh); 182 } 183 184 /* 185 * Main entrypoint for syncpoint value waits. 186 */ 187 int host1x_syncpt_wait(struct host1x_syncpt *sp, u32 thresh, long timeout, 188 u32 *value) 189 { 190 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(wq); 191 void *ref; 192 struct host1x_waitlist *waiter; 193 int err = 0, check_count = 0; 194 u32 val; 195 196 if (value) 197 *value = 0; 198 199 /* first check cache */ 200 if (host1x_syncpt_is_expired(sp, thresh)) { 201 if (value) 202 *value = host1x_syncpt_load(sp); 203 return 0; 204 } 205 206 /* try to read from register */ 207 val = host1x_hw_syncpt_load(sp->host, sp); 208 if (host1x_syncpt_is_expired(sp, thresh)) { 209 if (value) 210 *value = val; 211 goto done; 212 } 213 214 if (!timeout) { 215 err = -EAGAIN; 216 goto done; 217 } 218 219 /* allocate a waiter */ 220 waiter = kzalloc(sizeof(*waiter), GFP_KERNEL); 221 if (!waiter) { 222 err = -ENOMEM; 223 goto done; 224 } 225 226 /* schedule a wakeup when the syncpoint value is reached */ 227 err = host1x_intr_add_action(sp->host, sp->id, thresh, 228 HOST1X_INTR_ACTION_WAKEUP_INTERRUPTIBLE, 229 &wq, waiter, &ref); 230 if (err) 231 goto done; 232 233 err = -EAGAIN; 234 /* Caller-specified timeout may be impractically low */ 235 if (timeout < 0) 236 timeout = LONG_MAX; 237 238 /* wait for the syncpoint, or timeout, or signal */ 239 while (timeout) { 240 long check = min_t(long, SYNCPT_CHECK_PERIOD, timeout); 241 int remain = wait_event_interruptible_timeout(wq, 242 syncpt_load_min_is_expired(sp, thresh), 243 check); 244 if (remain > 0 || host1x_syncpt_is_expired(sp, thresh)) { 245 if (value) 246 *value = host1x_syncpt_load(sp); 247 err = 0; 248 break; 249 } 250 if (remain < 0) { 251 err = remain; 252 break; 253 } 254 timeout -= check; 255 if (timeout && check_count <= MAX_STUCK_CHECK_COUNT) { 256 dev_warn(sp->host->dev, 257 "%s: syncpoint id %d (%s) stuck waiting %d, timeout=%ld\n", 258 current->comm, sp->id, sp->name, 259 thresh, timeout); 260 261 host1x_debug_dump_syncpts(sp->host); 262 if (check_count == MAX_STUCK_CHECK_COUNT) 263 host1x_debug_dump(sp->host); 264 check_count++; 265 } 266 } 267 host1x_intr_put_ref(sp->host, sp->id, ref); 268 269 done: 270 return err; 271 } 272 EXPORT_SYMBOL(host1x_syncpt_wait); 273 274 /* 275 * Returns true if syncpoint is expired, false if we may need to wait 276 */ 277 bool host1x_syncpt_is_expired(struct host1x_syncpt *sp, u32 thresh) 278 { 279 u32 current_val; 280 u32 future_val; 281 smp_rmb(); 282 current_val = (u32)atomic_read(&sp->min_val); 283 future_val = (u32)atomic_read(&sp->max_val); 284 285 /* Note the use of unsigned arithmetic here (mod 1<<32). 286 * 287 * c = current_val = min_val = the current value of the syncpoint. 288 * t = thresh = the value we are checking 289 * f = future_val = max_val = the value c will reach when all 290 * outstanding increments have completed. 291 * 292 * Note that c always chases f until it reaches f. 293 * 294 * Dtf = (f - t) 295 * Dtc = (c - t) 296 * 297 * Consider all cases: 298 * 299 * A) .....c..t..f..... Dtf < Dtc need to wait 300 * B) .....c.....f..t.. Dtf > Dtc expired 301 * C) ..t..c.....f..... Dtf > Dtc expired (Dct very large) 302 * 303 * Any case where f==c: always expired (for any t). Dtf == Dcf 304 * Any case where t==c: always expired (for any f). Dtf >= Dtc (because Dtc==0) 305 * Any case where t==f!=c: always wait. Dtf < Dtc (because Dtf==0, 306 * Dtc!=0) 307 * 308 * Other cases: 309 * 310 * A) .....t..f..c..... Dtf < Dtc need to wait 311 * A) .....f..c..t..... Dtf < Dtc need to wait 312 * A) .....f..t..c..... Dtf > Dtc expired 313 * 314 * So: 315 * Dtf >= Dtc implies EXPIRED (return true) 316 * Dtf < Dtc implies WAIT (return false) 317 * 318 * Note: If t is expired then we *cannot* wait on it. We would wait 319 * forever (hang the system). 320 * 321 * Note: do NOT get clever and remove the -thresh from both sides. It 322 * is NOT the same. 323 * 324 * If future valueis zero, we have a client managed sync point. In that 325 * case we do a direct comparison. 326 */ 327 if (!host1x_syncpt_client_managed(sp)) 328 return future_val - thresh >= current_val - thresh; 329 else 330 return (s32)(current_val - thresh) >= 0; 331 } 332 333 /* remove a wait pointed to by patch_addr */ 334 int host1x_syncpt_patch_wait(struct host1x_syncpt *sp, void *patch_addr) 335 { 336 return host1x_hw_syncpt_patch_wait(sp->host, sp, patch_addr); 337 } 338 339 int host1x_syncpt_init(struct host1x *host) 340 { 341 struct host1x_syncpt_base *bases; 342 struct host1x_syncpt *syncpt; 343 int i; 344 345 syncpt = devm_kzalloc(host->dev, sizeof(*syncpt) * host->info->nb_pts, 346 GFP_KERNEL); 347 if (!syncpt) 348 return -ENOMEM; 349 350 bases = devm_kzalloc(host->dev, sizeof(*bases) * host->info->nb_bases, 351 GFP_KERNEL); 352 if (!bases) 353 return -ENOMEM; 354 355 for (i = 0; i < host->info->nb_pts; i++) { 356 syncpt[i].id = i; 357 syncpt[i].host = host; 358 } 359 360 for (i = 0; i < host->info->nb_bases; i++) 361 bases[i].id = i; 362 363 host->syncpt = syncpt; 364 host->bases = bases; 365 366 host1x_syncpt_restore(host); 367 368 /* Allocate sync point to use for clearing waits for expired fences */ 369 host->nop_sp = host1x_syncpt_alloc(host, NULL, 0); 370 if (!host->nop_sp) 371 return -ENOMEM; 372 373 return 0; 374 } 375 376 struct host1x_syncpt *host1x_syncpt_request(struct device *dev, 377 unsigned long flags) 378 { 379 struct host1x *host = dev_get_drvdata(dev->parent); 380 return host1x_syncpt_alloc(host, dev, flags); 381 } 382 EXPORT_SYMBOL(host1x_syncpt_request); 383 384 void host1x_syncpt_free(struct host1x_syncpt *sp) 385 { 386 if (!sp) 387 return; 388 389 host1x_syncpt_base_free(sp->base); 390 kfree(sp->name); 391 sp->base = NULL; 392 sp->dev = NULL; 393 sp->name = NULL; 394 sp->client_managed = false; 395 } 396 EXPORT_SYMBOL(host1x_syncpt_free); 397 398 void host1x_syncpt_deinit(struct host1x *host) 399 { 400 int i; 401 struct host1x_syncpt *sp = host->syncpt; 402 for (i = 0; i < host->info->nb_pts; i++, sp++) 403 kfree(sp->name); 404 } 405 406 /* 407 * Read max. It indicates how many operations there are in queue, either in 408 * channel or in a software thread. 409 * */ 410 u32 host1x_syncpt_read_max(struct host1x_syncpt *sp) 411 { 412 smp_rmb(); 413 return (u32)atomic_read(&sp->max_val); 414 } 415 EXPORT_SYMBOL(host1x_syncpt_read_max); 416 417 /* 418 * Read min, which is a shadow of the current sync point value in hardware. 419 */ 420 u32 host1x_syncpt_read_min(struct host1x_syncpt *sp) 421 { 422 smp_rmb(); 423 return (u32)atomic_read(&sp->min_val); 424 } 425 EXPORT_SYMBOL(host1x_syncpt_read_min); 426 427 int host1x_syncpt_nb_pts(struct host1x *host) 428 { 429 return host->info->nb_pts; 430 } 431 432 int host1x_syncpt_nb_bases(struct host1x *host) 433 { 434 return host->info->nb_bases; 435 } 436 437 int host1x_syncpt_nb_mlocks(struct host1x *host) 438 { 439 return host->info->nb_mlocks; 440 } 441 442 struct host1x_syncpt *host1x_syncpt_get(struct host1x *host, u32 id) 443 { 444 if (host->info->nb_pts < id) 445 return NULL; 446 return host->syncpt + id; 447 } 448 EXPORT_SYMBOL(host1x_syncpt_get); 449 450 struct host1x_syncpt_base *host1x_syncpt_get_base(struct host1x_syncpt *sp) 451 { 452 return sp ? sp->base : NULL; 453 } 454 EXPORT_SYMBOL(host1x_syncpt_get_base); 455 456 u32 host1x_syncpt_base_id(struct host1x_syncpt_base *base) 457 { 458 return base->id; 459 } 460 EXPORT_SYMBOL(host1x_syncpt_base_id); 461