2 * Copyright (C) 2007 Ben Skeggs.
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a 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, sublicense, 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:
13 * The above copyright notice and this permission notice (including the
14 * next paragraph) shall be included in all copies or substantial
15 * portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
19 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
20 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
21 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
22 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
23 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
30 #include <linux/ktime.h>
31 #include <linux/hrtimer.h>
33 #include "nouveau_drv.h"
34 #include "nouveau_ramht.h"
35 #include "nouveau_dma.h"
37 #define USE_REFCNT(dev) (nouveau_private(dev)->chipset >= 0x10)
38 #define USE_SEMA(dev) (nouveau_private(dev)->chipset >= 0x17)
40 struct nouveau_fence {
41 struct nouveau_channel *channel;
43 struct list_head entry;
48 void (*work)(void *priv, bool signalled);
52 struct nouveau_semaphore {
54 struct drm_device *dev;
55 struct drm_mm_node *mem;
58 static inline struct nouveau_fence *
59 nouveau_fence(void *sync_obj)
61 return (struct nouveau_fence *)sync_obj;
65 nouveau_fence_del(struct kref *ref)
67 struct nouveau_fence *fence =
68 container_of(ref, struct nouveau_fence, refcount);
70 nouveau_channel_ref(NULL, &fence->channel);
75 nouveau_fence_update(struct nouveau_channel *chan)
77 struct drm_device *dev = chan->dev;
78 struct nouveau_fence *tmp, *fence;
81 spin_lock(&chan->fence.lock);
83 /* Fetch the last sequence if the channel is still up and running */
84 if (likely(!list_empty(&chan->fence.pending))) {
86 sequence = nvchan_rd32(chan, 0x48);
88 sequence = atomic_read(&chan->fence.last_sequence_irq);
90 if (chan->fence.sequence_ack == sequence)
92 chan->fence.sequence_ack = sequence;
95 list_for_each_entry_safe(fence, tmp, &chan->fence.pending, entry) {
96 if (fence->sequence > chan->fence.sequence_ack)
99 fence->signalled = true;
100 list_del(&fence->entry);
102 fence->work(fence->priv, true);
104 kref_put(&fence->refcount, nouveau_fence_del);
108 spin_unlock(&chan->fence.lock);
112 nouveau_fence_new(struct nouveau_channel *chan, struct nouveau_fence **pfence,
115 struct nouveau_fence *fence;
118 fence = kzalloc(sizeof(*fence), GFP_KERNEL);
121 kref_init(&fence->refcount);
122 nouveau_channel_ref(chan, &fence->channel);
125 ret = nouveau_fence_emit(fence);
128 nouveau_fence_unref(&fence);
133 struct nouveau_channel *
134 nouveau_fence_channel(struct nouveau_fence *fence)
136 return fence ? nouveau_channel_get_unlocked(fence->channel) : NULL;
140 nouveau_fence_emit(struct nouveau_fence *fence)
142 struct nouveau_channel *chan = fence->channel;
143 struct drm_device *dev = chan->dev;
144 struct drm_nouveau_private *dev_priv = dev->dev_private;
147 ret = RING_SPACE(chan, 2);
151 if (unlikely(chan->fence.sequence == chan->fence.sequence_ack - 1)) {
152 nouveau_fence_update(chan);
154 BUG_ON(chan->fence.sequence ==
155 chan->fence.sequence_ack - 1);
158 fence->sequence = ++chan->fence.sequence;
160 kref_get(&fence->refcount);
161 spin_lock(&chan->fence.lock);
162 list_add_tail(&fence->entry, &chan->fence.pending);
163 spin_unlock(&chan->fence.lock);
165 if (USE_REFCNT(dev)) {
166 if (dev_priv->card_type < NV_C0)
167 BEGIN_RING(chan, 0, NV10_SUBCHAN_REF_CNT, 1);
169 BEGIN_NVC0(chan, 2, 0, NV10_SUBCHAN_REF_CNT, 1);
171 BEGIN_RING(chan, NvSubSw, 0x0150, 1);
173 OUT_RING (chan, fence->sequence);
180 nouveau_fence_work(struct nouveau_fence *fence,
181 void (*work)(void *priv, bool signalled),
186 spin_lock(&fence->channel->fence.lock);
188 if (fence->signalled) {
195 spin_unlock(&fence->channel->fence.lock);
199 __nouveau_fence_unref(void **sync_obj)
201 struct nouveau_fence *fence = nouveau_fence(*sync_obj);
204 kref_put(&fence->refcount, nouveau_fence_del);
209 __nouveau_fence_ref(void *sync_obj)
211 struct nouveau_fence *fence = nouveau_fence(sync_obj);
213 kref_get(&fence->refcount);
218 __nouveau_fence_signalled(void *sync_obj, void *sync_arg)
220 struct nouveau_fence *fence = nouveau_fence(sync_obj);
221 struct nouveau_channel *chan = fence->channel;
223 if (fence->signalled)
226 nouveau_fence_update(chan);
227 return fence->signalled;
231 __nouveau_fence_wait(void *sync_obj, void *sync_arg, bool lazy, bool intr)
233 unsigned long timeout = jiffies + (3 * DRM_HZ);
234 unsigned long sleep_time = NSEC_PER_MSEC / 1000;
239 if (__nouveau_fence_signalled(sync_obj, sync_arg))
242 if (time_after_eq(jiffies, timeout)) {
247 __set_current_state(intr ? TASK_INTERRUPTIBLE
248 : TASK_UNINTERRUPTIBLE);
250 t = ktime_set(0, sleep_time);
251 schedule_hrtimeout(&t, HRTIMER_MODE_REL);
253 if (sleep_time > NSEC_PER_MSEC)
254 sleep_time = NSEC_PER_MSEC;
257 if (intr && signal_pending(current)) {
263 __set_current_state(TASK_RUNNING);
268 static struct nouveau_semaphore *
269 semaphore_alloc(struct drm_device *dev)
271 struct drm_nouveau_private *dev_priv = dev->dev_private;
272 struct nouveau_semaphore *sema;
273 int size = (dev_priv->chipset < 0x84) ? 4 : 16;
279 sema = kmalloc(sizeof(*sema), GFP_KERNEL);
283 ret = drm_mm_pre_get(&dev_priv->fence.heap);
287 spin_lock(&dev_priv->fence.lock);
288 sema->mem = drm_mm_search_free(&dev_priv->fence.heap, size, 0, 0);
290 sema->mem = drm_mm_get_block_atomic(sema->mem, size, 0);
291 spin_unlock(&dev_priv->fence.lock);
296 kref_init(&sema->ref);
298 for (i = sema->mem->start; i < sema->mem->start + size; i += 4)
299 nouveau_bo_wr32(dev_priv->fence.bo, i / 4, 0);
308 semaphore_free(struct kref *ref)
310 struct nouveau_semaphore *sema =
311 container_of(ref, struct nouveau_semaphore, ref);
312 struct drm_nouveau_private *dev_priv = sema->dev->dev_private;
314 spin_lock(&dev_priv->fence.lock);
315 drm_mm_put_block(sema->mem);
316 spin_unlock(&dev_priv->fence.lock);
322 semaphore_work(void *priv, bool signalled)
324 struct nouveau_semaphore *sema = priv;
325 struct drm_nouveau_private *dev_priv = sema->dev->dev_private;
327 if (unlikely(!signalled))
328 nouveau_bo_wr32(dev_priv->fence.bo, sema->mem->start / 4, 1);
330 kref_put(&sema->ref, semaphore_free);
334 semaphore_acquire(struct nouveau_channel *chan, struct nouveau_semaphore *sema)
336 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
337 struct nouveau_fence *fence = NULL;
338 u64 offset = chan->fence.vma.offset + sema->mem->start;
341 if (dev_priv->chipset < 0x84) {
342 ret = RING_SPACE(chan, 4);
346 BEGIN_RING(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 3);
347 OUT_RING (chan, NvSema);
348 OUT_RING (chan, offset);
351 if (dev_priv->chipset < 0xc0) {
352 ret = RING_SPACE(chan, 7);
356 BEGIN_RING(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 1);
357 OUT_RING (chan, chan->vram_handle);
358 BEGIN_RING(chan, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
359 OUT_RING (chan, upper_32_bits(offset));
360 OUT_RING (chan, lower_32_bits(offset));
362 OUT_RING (chan, 1); /* ACQUIRE_EQ */
364 ret = RING_SPACE(chan, 5);
368 BEGIN_NVC0(chan, 2, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
369 OUT_RING (chan, upper_32_bits(offset));
370 OUT_RING (chan, lower_32_bits(offset));
372 OUT_RING (chan, 0x1001); /* ACQUIRE_EQ */
375 /* Delay semaphore destruction until its work is done */
376 ret = nouveau_fence_new(chan, &fence, true);
380 kref_get(&sema->ref);
381 nouveau_fence_work(fence, semaphore_work, sema);
382 nouveau_fence_unref(&fence);
387 semaphore_release(struct nouveau_channel *chan, struct nouveau_semaphore *sema)
389 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
390 struct nouveau_fence *fence = NULL;
391 u64 offset = chan->fence.vma.offset + sema->mem->start;
394 if (dev_priv->chipset < 0x84) {
395 ret = RING_SPACE(chan, 5);
399 BEGIN_RING(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 2);
400 OUT_RING (chan, NvSema);
401 OUT_RING (chan, offset);
402 BEGIN_RING(chan, 0, NV11_SUBCHAN_SEMAPHORE_RELEASE, 1);
405 if (dev_priv->chipset < 0xc0) {
406 ret = RING_SPACE(chan, 7);
410 BEGIN_RING(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 1);
411 OUT_RING (chan, chan->vram_handle);
412 BEGIN_RING(chan, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
413 OUT_RING (chan, upper_32_bits(offset));
414 OUT_RING (chan, lower_32_bits(offset));
416 OUT_RING (chan, 2); /* RELEASE */
418 ret = RING_SPACE(chan, 5);
422 BEGIN_NVC0(chan, 2, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
423 OUT_RING (chan, upper_32_bits(offset));
424 OUT_RING (chan, lower_32_bits(offset));
426 OUT_RING (chan, 0x1002); /* RELEASE */
429 /* Delay semaphore destruction until its work is done */
430 ret = nouveau_fence_new(chan, &fence, true);
434 kref_get(&sema->ref);
435 nouveau_fence_work(fence, semaphore_work, sema);
436 nouveau_fence_unref(&fence);
441 nouveau_fence_sync(struct nouveau_fence *fence,
442 struct nouveau_channel *wchan)
444 struct nouveau_channel *chan = nouveau_fence_channel(fence);
445 struct drm_device *dev = wchan->dev;
446 struct nouveau_semaphore *sema;
449 if (likely(!chan || chan == wchan ||
450 nouveau_fence_signalled(fence)))
453 sema = semaphore_alloc(dev);
455 /* Early card or broken userspace, fall back to
457 ret = nouveau_fence_wait(fence, true, false);
461 /* try to take chan's mutex, if we can't take it right away
462 * we have to fallback to software sync to prevent locking
465 if (!mutex_trylock(&chan->mutex)) {
466 ret = nouveau_fence_wait(fence, true, false);
470 /* Make wchan wait until it gets signalled */
471 ret = semaphore_acquire(wchan, sema);
475 /* Signal the semaphore from chan */
476 ret = semaphore_release(chan, sema);
479 mutex_unlock(&chan->mutex);
481 kref_put(&sema->ref, semaphore_free);
484 nouveau_channel_put_unlocked(&chan);
489 __nouveau_fence_flush(void *sync_obj, void *sync_arg)
495 nouveau_fence_channel_init(struct nouveau_channel *chan)
497 struct drm_device *dev = chan->dev;
498 struct drm_nouveau_private *dev_priv = dev->dev_private;
499 struct nouveau_gpuobj *obj = NULL;
502 if (dev_priv->card_type < NV_C0) {
503 /* Create an NV_SW object for various sync purposes */
504 ret = nouveau_gpuobj_gr_new(chan, NvSw, NV_SW);
508 ret = RING_SPACE(chan, 2);
512 BEGIN_RING(chan, NvSubSw, NV01_SUBCHAN_OBJECT, 1);
513 OUT_RING (chan, NvSw);
517 /* Setup area of memory shared between all channels for x-chan sync */
518 if (USE_SEMA(dev) && dev_priv->chipset < 0x84) {
519 struct ttm_mem_reg *mem = &dev_priv->fence.bo->bo.mem;
521 ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_FROM_MEMORY,
522 mem->start << PAGE_SHIFT,
523 mem->size, NV_MEM_ACCESS_RW,
524 NV_MEM_TARGET_VRAM, &obj);
528 ret = nouveau_ramht_insert(chan, NvSema, obj);
529 nouveau_gpuobj_ref(NULL, &obj);
534 /* map fence bo into channel's vm */
535 ret = nouveau_bo_vma_add(dev_priv->fence.bo, chan->vm,
541 atomic_set(&chan->fence.last_sequence_irq, 0);
546 nouveau_fence_channel_fini(struct nouveau_channel *chan)
548 struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
549 struct nouveau_fence *tmp, *fence;
551 spin_lock(&chan->fence.lock);
552 list_for_each_entry_safe(fence, tmp, &chan->fence.pending, entry) {
553 fence->signalled = true;
554 list_del(&fence->entry);
556 if (unlikely(fence->work))
557 fence->work(fence->priv, false);
559 kref_put(&fence->refcount, nouveau_fence_del);
561 spin_unlock(&chan->fence.lock);
563 nouveau_bo_vma_del(dev_priv->fence.bo, &chan->fence.vma);
567 nouveau_fence_init(struct drm_device *dev)
569 struct drm_nouveau_private *dev_priv = dev->dev_private;
570 int size = (dev_priv->chipset < 0x84) ? 4096 : 16384;
573 /* Create a shared VRAM heap for cross-channel sync. */
575 ret = nouveau_bo_new(dev, size, 0, TTM_PL_FLAG_VRAM,
576 0, 0, &dev_priv->fence.bo);
580 ret = nouveau_bo_pin(dev_priv->fence.bo, TTM_PL_FLAG_VRAM);
584 ret = nouveau_bo_map(dev_priv->fence.bo);
588 ret = drm_mm_init(&dev_priv->fence.heap, 0,
589 dev_priv->fence.bo->bo.mem.size);
593 spin_lock_init(&dev_priv->fence.lock);
598 nouveau_bo_unmap(dev_priv->fence.bo);
599 nouveau_bo_ref(NULL, &dev_priv->fence.bo);
604 nouveau_fence_fini(struct drm_device *dev)
606 struct drm_nouveau_private *dev_priv = dev->dev_private;
609 drm_mm_takedown(&dev_priv->fence.heap);
610 nouveau_bo_unmap(dev_priv->fence.bo);
611 nouveau_bo_unpin(dev_priv->fence.bo);
612 nouveau_bo_ref(NULL, &dev_priv->fence.bo);