drivers/gpu/drm/radeon/radeon_fence.c

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   3  * All Rights Reserved.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the
   7  * "Software"), to deal in the Software without restriction, including
   8  * without limitation the rights to use, copy, modify, merge, publish,
   9  * distribute, sub license, and/or sell copies of the Software, and to
  10  * permit persons to whom the Software is furnished to do so, subject to
  11  * the following conditions:
  12  *
  13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
  17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  20  *
  21  * The above copyright notice and this permission notice (including the
  22  * next paragraph) shall be included in all copies or substantial portions
  23  * of the Software.
  24  *
  25  */
  26 /*
  27  * Authors:
  28  *    Jerome Glisse <glisse@freedesktop.org>
  29  *    Dave Airlie
  30  */
  31 #include <linux/seq_file.h>
  32 #include <asm/atomic.h>
  33 #include <linux/wait.h>
  34 #include <linux/list.h>
  35 #include <linux/kref.h>
  36 #include <linux/slab.h>
  37 #include "drmP.h"
  38 #include "drm.h"
  39 #include "radeon_reg.h"
  40 #include "radeon.h"
  41
  42 int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence *fence)
  43 {
  44         unsigned long irq_flags;
  45
  46         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
  47         if (fence->emited) {
  48                 write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
  49                 return 0;
  50         }
  51         fence->seq = atomic_add_return(1, &rdev->fence_drv.seq);
  52         if (!rdev->cp.ready) {
  53                 /* FIXME: cp is not running assume everythings is done right
  54                  * away
  55                  */
  56                 WREG32(rdev->fence_drv.scratch_reg, fence->seq);
  57         } else
  58                 radeon_fence_ring_emit(rdev, fence);
  59
  60         fence->emited = true;
  61         list_del(&fence->list);
  62         list_add_tail(&fence->list, &rdev->fence_drv.emited);
  63         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
  64         return 0;
  65 }
  66
  67 static bool radeon_fence_poll_locked(struct radeon_device *rdev)
  68 {
  69         struct radeon_fence *fence;
  70         struct list_head *i, *n;
  71         uint32_t seq;
  72         bool wake = false;
  73         unsigned long cjiffies;
  74
  75         if (rdev->wb.enabled) {
  76                 u32 scratch_index;
  77                 if (rdev->wb.use_event)
  78                         scratch_index = R600_WB_EVENT_OFFSET + rdev->fence_drv.scratch_reg - rdev->scratch.reg_base;
  79                 else
  80                         scratch_index = RADEON_WB_SCRATCH_OFFSET + rdev->fence_drv.scratch_reg - rdev->scratch.reg_base;
  81                 seq = rdev->wb.wb[scratch_index/4];
  82         } else
  83                 seq = RREG32(rdev->fence_drv.scratch_reg);
  84         if (seq != rdev->fence_drv.last_seq) {
  85                 rdev->fence_drv.last_seq = seq;
  86                 rdev->fence_drv.last_jiffies = jiffies;
  87                 rdev->fence_drv.last_timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
  88         } else {
  89                 cjiffies = jiffies;
  90                 if (time_after(cjiffies, rdev->fence_drv.last_jiffies)) {
  91                         cjiffies -= rdev->fence_drv.last_jiffies;
  92                         if (time_after(rdev->fence_drv.last_timeout, cjiffies)) {
  93                                 /* update the timeout */
  94                                 rdev->fence_drv.last_timeout -= cjiffies;
  95                         } else {
  96                                 /* the 500ms timeout is elapsed we should test
  97                                  * for GPU lockup
  98                                  */
  99                                 rdev->fence_drv.last_timeout = 1;
 100                         }
 101                 } else {
 102                         /* wrap around update last jiffies, we will just wait
 103                          * a little longer
 104                          */
 105                         rdev->fence_drv.last_jiffies = cjiffies;
 106                 }
 107                 return false;
 108         }
 109         n = NULL;
 110         list_for_each(i, &rdev->fence_drv.emited) {
 111                 fence = list_entry(i, struct radeon_fence, list);
 112                 if (fence->seq == seq) {
 113                         n = i;
 114                         break;
 115                 }
 116         }
 117         /* all fence previous to this one are considered as signaled */
 118         if (n) {
 119                 i = n;
 120                 do {
 121                         n = i->prev;
 122                         list_del(i);
 123                         list_add_tail(i, &rdev->fence_drv.signaled);
 124                         fence = list_entry(i, struct radeon_fence, list);
 125                         fence->signaled = true;
 126                         i = n;
 127                 } while (i != &rdev->fence_drv.emited);
 128                 wake = true;
 129         }
 130         return wake;
 131 }
 132
 133 static void radeon_fence_destroy(struct kref *kref)
 134 {
 135         unsigned long irq_flags;
 136         struct radeon_fence *fence;
 137
 138         fence = container_of(kref, struct radeon_fence, kref);
 139         write_lock_irqsave(&fence->rdev->fence_drv.lock, irq_flags);
 140         list_del(&fence->list);
 141         fence->emited = false;
 142         write_unlock_irqrestore(&fence->rdev->fence_drv.lock, irq_flags);
 143         kfree(fence);
 144 }
 145
 146 int radeon_fence_create(struct radeon_device *rdev, struct radeon_fence **fence)
 147 {
 148         unsigned long irq_flags;
 149
 150         *fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
 151         if ((*fence) == NULL) {
 152                 return -ENOMEM;
 153         }
 154         kref_init(&((*fence)->kref));
 155         (*fence)->rdev = rdev;
 156         (*fence)->emited = false;
 157         (*fence)->signaled = false;
 158         (*fence)->seq = 0;
 159         INIT_LIST_HEAD(&(*fence)->list);
 160
 161         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 162         list_add_tail(&(*fence)->list, &rdev->fence_drv.created);
 163         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 164         return 0;
 165 }
 166
 167
 168 bool radeon_fence_signaled(struct radeon_fence *fence)
 169 {
 170         unsigned long irq_flags;
 171         bool signaled = false;
 172
 173         if (!fence)
 174                 return true;
 175
 176         if (fence->rdev->gpu_lockup)
 177                 return true;
 178
 179         write_lock_irqsave(&fence->rdev->fence_drv.lock, irq_flags);
 180         signaled = fence->signaled;
 181         /* if we are shuting down report all fence as signaled */
 182         if (fence->rdev->shutdown) {
 183                 signaled = true;
 184         }
 185         if (!fence->emited) {
 186                 WARN(1, "Querying an unemited fence : %p !\n", fence);
 187                 signaled = true;
 188         }
 189         if (!signaled) {
 190                 radeon_fence_poll_locked(fence->rdev);
 191                 signaled = fence->signaled;
 192         }
 193         write_unlock_irqrestore(&fence->rdev->fence_drv.lock, irq_flags);
 194         return signaled;
 195 }
 196
 197 int radeon_fence_wait(struct radeon_fence *fence, bool intr)
 198 {
 199         struct radeon_device *rdev;
 200         unsigned long irq_flags, timeout;
 201         u32 seq;
 202         int r;
 203
 204         if (fence == NULL) {
 205                 WARN(1, "Querying an invalid fence : %p !\n", fence);
 206                 return 0;
 207         }
 208         rdev = fence->rdev;
 209         if (radeon_fence_signaled(fence)) {
 210                 return 0;
 211         }
 212         timeout = rdev->fence_drv.last_timeout;
 213 retry:
 214         /* save current sequence used to check for GPU lockup */
 215         seq = rdev->fence_drv.last_seq;
 216         if (intr) {
 217                 radeon_irq_kms_sw_irq_get(rdev);
 218                 r = wait_event_interruptible_timeout(rdev->fence_drv.queue,
 219                                 radeon_fence_signaled(fence), timeout);
 220                 radeon_irq_kms_sw_irq_put(rdev);
 221                 if (unlikely(r < 0)) {
 222                         return r;
 223                 }
 224         } else {
 225                 radeon_irq_kms_sw_irq_get(rdev);
 226                 r = wait_event_timeout(rdev->fence_drv.queue,
 227                          radeon_fence_signaled(fence), timeout);
 228                 radeon_irq_kms_sw_irq_put(rdev);
 229         }
 230         if (unlikely(!radeon_fence_signaled(fence))) {
 231                 /* we were interrupted for some reason and fence isn't
 232                  * isn't signaled yet, resume wait
 233                  */
 234                 if (r) {
 235                         timeout = r;
 236                         goto retry;
 237                 }
 238                 /* don't protect read access to rdev->fence_drv.last_seq
 239                  * if we experiencing a lockup the value doesn't change
 240                  */
 241                 if (seq == rdev->fence_drv.last_seq && radeon_gpu_is_lockup(rdev)) {
 242                         /* good news we believe it's a lockup */
 243                         WARN(1, "GPU lockup (waiting for 0x%08X last fence id 0x%08X)\n", fence->seq, seq);
 244                         /* FIXME: what should we do ? marking everyone
 245                          * as signaled for now
 246                          */
 247                         rdev->gpu_lockup = true;
 248                         r = radeon_gpu_reset(rdev);
 249                         if (r)
 250                                 return r;
 251                         WREG32(rdev->fence_drv.scratch_reg, fence->seq);
 252                         rdev->gpu_lockup = false;
 253                 }
 254                 timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
 255                 write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 256                 rdev->fence_drv.last_timeout = RADEON_FENCE_JIFFIES_TIMEOUT;
 257                 rdev->fence_drv.last_jiffies = jiffies;
 258                 write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 259                 goto retry;
 260         }
 261         return 0;
 262 }
 263
 264 int radeon_fence_wait_next(struct radeon_device *rdev)
 265 {
 266         unsigned long irq_flags;
 267         struct radeon_fence *fence;
 268         int r;
 269
 270         if (rdev->gpu_lockup) {
 271                 return 0;
 272         }
 273         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 274         if (list_empty(&rdev->fence_drv.emited)) {
 275                 write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 276                 return 0;
 277         }
 278         fence = list_entry(rdev->fence_drv.emited.next,
 279                            struct radeon_fence, list);
 280         radeon_fence_ref(fence);
 281         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 282         r = radeon_fence_wait(fence, false);
 283         radeon_fence_unref(&fence);
 284         return r;
 285 }
 286
 287 int radeon_fence_wait_last(struct radeon_device *rdev)
 288 {
 289         unsigned long irq_flags;
 290         struct radeon_fence *fence;
 291         int r;
 292
 293         if (rdev->gpu_lockup) {
 294                 return 0;
 295         }
 296         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 297         if (list_empty(&rdev->fence_drv.emited)) {
 298                 write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 299                 return 0;
 300         }
 301         fence = list_entry(rdev->fence_drv.emited.prev,
 302                            struct radeon_fence, list);
 303         radeon_fence_ref(fence);
 304         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 305         r = radeon_fence_wait(fence, false);
 306         radeon_fence_unref(&fence);
 307         return r;
 308 }
 309
 310 struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
 311 {
 312         kref_get(&fence->kref);
 313         return fence;
 314 }
 315
 316 void radeon_fence_unref(struct radeon_fence **fence)
 317 {
 318         struct radeon_fence *tmp = *fence;
 319
 320         *fence = NULL;
 321         if (tmp) {
 322                 kref_put(&tmp->kref, &radeon_fence_destroy);
 323         }
 324 }
 325
 326 void radeon_fence_process(struct radeon_device *rdev)
 327 {
 328         unsigned long irq_flags;
 329         bool wake;
 330
 331         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 332         wake = radeon_fence_poll_locked(rdev);
 333         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 334         if (wake) {
 335                 wake_up_all(&rdev->fence_drv.queue);
 336         }
 337 }
 338
 339 int radeon_fence_driver_init(struct radeon_device *rdev)
 340 {
 341         unsigned long irq_flags;
 342         int r;
 343
 344         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 345         r = radeon_scratch_get(rdev, &rdev->fence_drv.scratch_reg);
 346         if (r) {
 347                 dev_err(rdev->dev, "fence failed to get scratch register\n");
 348                 write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 349                 return r;
 350         }
 351         WREG32(rdev->fence_drv.scratch_reg, 0);
 352         atomic_set(&rdev->fence_drv.seq, 0);
 353         INIT_LIST_HEAD(&rdev->fence_drv.created);
 354         INIT_LIST_HEAD(&rdev->fence_drv.emited);
 355         INIT_LIST_HEAD(&rdev->fence_drv.signaled);
 356         init_waitqueue_head(&rdev->fence_drv.queue);
 357         rdev->fence_drv.initialized = true;
 358         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 359         if (radeon_debugfs_fence_init(rdev)) {
 360                 dev_err(rdev->dev, "fence debugfs file creation failed\n");
 361         }
 362         return 0;
 363 }
 364
 365 void radeon_fence_driver_fini(struct radeon_device *rdev)
 366 {
 367         unsigned long irq_flags;
 368
 369         if (!rdev->fence_drv.initialized)
 370                 return;
 371         wake_up_all(&rdev->fence_drv.queue);
 372         write_lock_irqsave(&rdev->fence_drv.lock, irq_flags);
 373         radeon_scratch_free(rdev, rdev->fence_drv.scratch_reg);
 374         write_unlock_irqrestore(&rdev->fence_drv.lock, irq_flags);
 375         rdev->fence_drv.initialized = false;
 376 }
 377
 378
 379 /*
 380  * Fence debugfs
 381  */
 382 #if defined(CONFIG_DEBUG_FS)
 383 static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
 384 {
 385         struct drm_info_node *node = (struct drm_info_node *)m->private;
 386         struct drm_device *dev = node->minor->dev;
 387         struct radeon_device *rdev = dev->dev_private;
 388         struct radeon_fence *fence;
 389
 390         seq_printf(m, "Last signaled fence 0x%08X\n",
 391                    RREG32(rdev->fence_drv.scratch_reg));
 392         if (!list_empty(&rdev->fence_drv.emited)) {
 393                    fence = list_entry(rdev->fence_drv.emited.prev,
 394                                       struct radeon_fence, list);
 395                    seq_printf(m, "Last emited fence %p with 0x%08X\n",
 396                               fence,  fence->seq);
 397         }
 398         return 0;
 399 }
 400
 401 static struct drm_info_list radeon_debugfs_fence_list[] = {
 402         {"radeon_fence_info", &radeon_debugfs_fence_info, 0, NULL},
 403 };
 404 #endif
 405
 406 int radeon_debugfs_fence_init(struct radeon_device *rdev)
 407 {
 408 #if defined(CONFIG_DEBUG_FS)
 409         return radeon_debugfs_add_files(rdev, radeon_debugfs_fence_list, 1);
 410 #else
 411         return 0;
 412 #endif
 413 }