summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/nouveau/nouveau_fence.c
diff options
context:
space:
mode:
authorBen Skeggs <bskeggs@redhat.com>2012-04-30 13:55:29 +1000
committerBen Skeggs <bskeggs@redhat.com>2012-05-24 16:55:53 +1000
commit5e120f6e4b3f35b741c5445dfc755f50128c3c44 (patch)
tree210b2bb8f5dccfcb4a6c134341fa31a633ba5243 /drivers/gpu/drm/nouveau/nouveau_fence.c
parentd375e7d56dffa564a6c337d2ed3217fb94826100 (diff)
drm/nouveau/fence: convert to exec engine, and improve channel sync
Now have a somewhat simpler semaphore sync implementation for nv17:nv84, and a switched to using semaphores as fences on nv84+ and making use of the hardware's >= acquire operation. Signed-off-by: Ben Skeggs <bskeggs@redhat.com>
Diffstat (limited to 'drivers/gpu/drm/nouveau/nouveau_fence.c')
-rw-r--r--drivers/gpu/drm/nouveau/nouveau_fence.c474
1 files changed, 62 insertions, 412 deletions
diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
index 2c10d54fc493..4ba41a45114f 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -36,85 +36,71 @@
#include "nouveau_software.h"
#include "nouveau_dma.h"
-#define USE_REFCNT(dev) (nouveau_private(dev)->chipset >= 0x10)
-#define USE_SEMA(dev) (nouveau_private(dev)->chipset >= 0x17)
+void
+nouveau_fence_context_del(struct nouveau_fence_chan *fctx)
+{
+ struct nouveau_fence *fence, *fnext;
+ spin_lock(&fctx->lock);
+ list_for_each_entry_safe(fence, fnext, &fctx->pending, head) {
+ if (fence->work)
+ fence->work(fence->priv, false);
+ fence->channel = NULL;
+ list_del(&fence->head);
+ nouveau_fence_unref(&fence);
+ }
+ spin_unlock(&fctx->lock);
+}
+
+void
+nouveau_fence_context_new(struct nouveau_fence_chan *fctx)
+{
+ INIT_LIST_HEAD(&fctx->pending);
+ spin_lock_init(&fctx->lock);
+}
void
nouveau_fence_update(struct nouveau_channel *chan)
{
struct drm_device *dev = chan->dev;
- struct nouveau_fence *tmp, *fence;
- uint32_t sequence;
-
- spin_lock(&chan->fence.lock);
-
- /* Fetch the last sequence if the channel is still up and running */
- if (likely(!list_empty(&chan->fence.pending))) {
- if (USE_REFCNT(dev))
- sequence = nvchan_rd32(chan, 0x48);
- else
- sequence = atomic_read(&chan->fence.last_sequence_irq);
-
- if (chan->fence.sequence_ack == sequence)
- goto out;
- chan->fence.sequence_ack = sequence;
- }
+ struct nouveau_fence_priv *priv = nv_engine(dev, NVOBJ_ENGINE_FENCE);
+ struct nouveau_fence_chan *fctx = chan->engctx[NVOBJ_ENGINE_FENCE];
+ struct nouveau_fence *fence, *fnext;
- list_for_each_entry_safe(fence, tmp, &chan->fence.pending, head) {
- if (fence->sequence > chan->fence.sequence_ack)
+ spin_lock(&fctx->lock);
+ list_for_each_entry_safe(fence, fnext, &fctx->pending, head) {
+ if (priv->read(chan) < fence->sequence)
break;
- fence->channel = NULL;
- list_del(&fence->head);
if (fence->work)
fence->work(fence->priv, true);
-
+ fence->channel = NULL;
+ list_del(&fence->head);
nouveau_fence_unref(&fence);
}
-
-out:
- spin_unlock(&chan->fence.lock);
+ spin_unlock(&fctx->lock);
}
int
nouveau_fence_emit(struct nouveau_fence *fence, struct nouveau_channel *chan)
{
struct drm_device *dev = chan->dev;
- struct drm_nouveau_private *dev_priv = dev->dev_private;
+ struct nouveau_fence_priv *priv = nv_engine(dev, NVOBJ_ENGINE_FENCE);
+ struct nouveau_fence_chan *fctx = chan->engctx[NVOBJ_ENGINE_FENCE];
int ret;
- ret = RING_SPACE(chan, 2);
- if (ret)
- return ret;
-
- if (unlikely(chan->fence.sequence == chan->fence.sequence_ack - 1)) {
- nouveau_fence_update(chan);
-
- BUG_ON(chan->fence.sequence ==
- chan->fence.sequence_ack - 1);
- }
-
- fence->sequence = ++chan->fence.sequence;
- fence->channel = chan;
-
- kref_get(&fence->kref);
- spin_lock(&chan->fence.lock);
- list_add_tail(&fence->head, &chan->fence.pending);
- spin_unlock(&chan->fence.lock);
+ fence->channel = chan;
+ fence->timeout = jiffies + (3 * DRM_HZ);
+ fence->sequence = ++fctx->sequence;
- if (USE_REFCNT(dev)) {
- if (dev_priv->card_type < NV_C0)
- BEGIN_NV04(chan, 0, NV10_SUBCHAN_REF_CNT, 1);
- else
- BEGIN_NVC0(chan, 0, NV10_SUBCHAN_REF_CNT, 1);
- } else {
- BEGIN_NV04(chan, NvSubSw, 0x0150, 1);
+ ret = priv->emit(fence);
+ if (!ret) {
+ kref_get(&fence->kref);
+ spin_lock(&fctx->lock);
+ list_add_tail(&fence->head, &fctx->pending);
+ spin_unlock(&fctx->lock);
}
- OUT_RING (chan, fence->sequence);
- FIRE_RING(chan);
- fence->timeout = jiffies + 3 * DRM_HZ;
- return 0;
+ return ret;
}
bool
@@ -158,6 +144,23 @@ nouveau_fence_wait(struct nouveau_fence *fence, bool lazy, bool intr)
return ret;
}
+int
+nouveau_fence_sync(struct nouveau_fence *fence, struct nouveau_channel *chan)
+{
+ struct nouveau_channel *prev = fence ? fence->channel : NULL;
+ struct drm_device *dev = chan->dev;
+ struct nouveau_fence_priv *priv = nv_engine(dev, NVOBJ_ENGINE_FENCE);
+ int ret = 0;
+
+ if (unlikely(prev && prev != chan && !nouveau_fence_done(fence))) {
+ ret = priv->sync(fence, chan);
+ if (unlikely(ret))
+ ret = nouveau_fence_wait(fence, true, false);
+ }
+
+ return ret;
+}
+
static void
nouveau_fence_del(struct kref *kref)
{
@@ -186,6 +189,9 @@ nouveau_fence_new(struct nouveau_channel *chan, struct nouveau_fence **pfence)
struct nouveau_fence *fence;
int ret = 0;
+ if (unlikely(!chan->engctx[NVOBJ_ENGINE_FENCE]))
+ return -ENODEV;
+
fence = kzalloc(sizeof(*fence), GFP_KERNEL);
if (!fence)
return -ENOMEM;
@@ -200,359 +206,3 @@ nouveau_fence_new(struct nouveau_channel *chan, struct nouveau_fence **pfence)
*pfence = fence;
return ret;
}
-
-struct nouveau_semaphore {
- struct kref ref;
- struct drm_device *dev;
- struct drm_mm_node *mem;
-};
-
-void
-nouveau_fence_work(struct nouveau_fence *fence,
- void (*work)(void *priv, bool signalled),
- void *priv)
-{
- if (!fence->channel) {
- work(priv, true);
- } else {
- fence->work = work;
- fence->priv = priv;
- }
-}
-
-static struct nouveau_semaphore *
-semaphore_alloc(struct drm_device *dev)
-{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nouveau_semaphore *sema;
- int size = (dev_priv->chipset < 0x84) ? 4 : 16;
- int ret, i;
-
- if (!USE_SEMA(dev))
- return NULL;
-
- sema = kmalloc(sizeof(*sema), GFP_KERNEL);
- if (!sema)
- goto fail;
-
- ret = drm_mm_pre_get(&dev_priv->fence.heap);
- if (ret)
- goto fail;
-
- spin_lock(&dev_priv->fence.lock);
- sema->mem = drm_mm_search_free(&dev_priv->fence.heap, size, 0, 0);
- if (sema->mem)
- sema->mem = drm_mm_get_block_atomic(sema->mem, size, 0);
- spin_unlock(&dev_priv->fence.lock);
-
- if (!sema->mem)
- goto fail;
-
- kref_init(&sema->ref);
- sema->dev = dev;
- for (i = sema->mem->start; i < sema->mem->start + size; i += 4)
- nouveau_bo_wr32(dev_priv->fence.bo, i / 4, 0);
-
- return sema;
-fail:
- kfree(sema);
- return NULL;
-}
-
-static void
-semaphore_free(struct kref *ref)
-{
- struct nouveau_semaphore *sema =
- container_of(ref, struct nouveau_semaphore, ref);
- struct drm_nouveau_private *dev_priv = sema->dev->dev_private;
-
- spin_lock(&dev_priv->fence.lock);
- drm_mm_put_block(sema->mem);
- spin_unlock(&dev_priv->fence.lock);
-
- kfree(sema);
-}
-
-static void
-semaphore_work(void *priv, bool signalled)
-{
- struct nouveau_semaphore *sema = priv;
- struct drm_nouveau_private *dev_priv = sema->dev->dev_private;
-
- if (unlikely(!signalled))
- nouveau_bo_wr32(dev_priv->fence.bo, sema->mem->start / 4, 1);
-
- kref_put(&sema->ref, semaphore_free);
-}
-
-static int
-semaphore_acquire(struct nouveau_channel *chan, struct nouveau_semaphore *sema)
-{
- struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
- struct nouveau_fence *fence = NULL;
- u64 offset = chan->fence.vma.offset + sema->mem->start;
- int ret;
-
- if (dev_priv->chipset < 0x84) {
- ret = RING_SPACE(chan, 4);
- if (ret)
- return ret;
-
- BEGIN_NV04(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 3);
- OUT_RING (chan, NvSema);
- OUT_RING (chan, offset);
- OUT_RING (chan, 1);
- } else
- if (dev_priv->chipset < 0xc0) {
- ret = RING_SPACE(chan, 7);
- if (ret)
- return ret;
-
- BEGIN_NV04(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 1);
- OUT_RING (chan, chan->vram_handle);
- BEGIN_NV04(chan, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
- OUT_RING (chan, upper_32_bits(offset));
- OUT_RING (chan, lower_32_bits(offset));
- OUT_RING (chan, 1);
- OUT_RING (chan, 1); /* ACQUIRE_EQ */
- } else {
- ret = RING_SPACE(chan, 5);
- if (ret)
- return ret;
-
- BEGIN_NVC0(chan, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
- OUT_RING (chan, upper_32_bits(offset));
- OUT_RING (chan, lower_32_bits(offset));
- OUT_RING (chan, 1);
- OUT_RING (chan, 0x1001); /* ACQUIRE_EQ */
- }
-
- /* Delay semaphore destruction until its work is done */
- ret = nouveau_fence_new(chan, &fence);
- if (ret)
- return ret;
-
- kref_get(&sema->ref);
- nouveau_fence_work(fence, semaphore_work, sema);
- nouveau_fence_unref(&fence);
- return 0;
-}
-
-static int
-semaphore_release(struct nouveau_channel *chan, struct nouveau_semaphore *sema)
-{
- struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
- struct nouveau_fence *fence = NULL;
- u64 offset = chan->fence.vma.offset + sema->mem->start;
- int ret;
-
- if (dev_priv->chipset < 0x84) {
- ret = RING_SPACE(chan, 5);
- if (ret)
- return ret;
-
- BEGIN_NV04(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 2);
- OUT_RING (chan, NvSema);
- OUT_RING (chan, offset);
- BEGIN_NV04(chan, 0, NV11_SUBCHAN_SEMAPHORE_RELEASE, 1);
- OUT_RING (chan, 1);
- } else
- if (dev_priv->chipset < 0xc0) {
- ret = RING_SPACE(chan, 7);
- if (ret)
- return ret;
-
- BEGIN_NV04(chan, 0, NV11_SUBCHAN_DMA_SEMAPHORE, 1);
- OUT_RING (chan, chan->vram_handle);
- BEGIN_NV04(chan, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
- OUT_RING (chan, upper_32_bits(offset));
- OUT_RING (chan, lower_32_bits(offset));
- OUT_RING (chan, 1);
- OUT_RING (chan, 2); /* RELEASE */
- } else {
- ret = RING_SPACE(chan, 5);
- if (ret)
- return ret;
-
- BEGIN_NVC0(chan, 0, NV84_SUBCHAN_SEMAPHORE_ADDRESS_HIGH, 4);
- OUT_RING (chan, upper_32_bits(offset));
- OUT_RING (chan, lower_32_bits(offset));
- OUT_RING (chan, 1);
- OUT_RING (chan, 0x1002); /* RELEASE */
- }
-
- /* Delay semaphore destruction until its work is done */
- ret = nouveau_fence_new(chan, &fence);
- if (ret)
- return ret;
-
- kref_get(&sema->ref);
- nouveau_fence_work(fence, semaphore_work, sema);
- nouveau_fence_unref(&fence);
- return 0;
-}
-
-int
-nouveau_fence_sync(struct nouveau_fence *fence,
- struct nouveau_channel *wchan)
-{
- struct nouveau_channel *chan;
- struct drm_device *dev = wchan->dev;
- struct nouveau_semaphore *sema;
- int ret = 0;
-
- chan = fence ? nouveau_channel_get_unlocked(fence->channel) : NULL;
- if (likely(!chan || chan == wchan || nouveau_fence_done(fence)))
- goto out;
-
- sema = semaphore_alloc(dev);
- if (!sema) {
- /* Early card or broken userspace, fall back to
- * software sync. */
- ret = nouveau_fence_wait(fence, true, false);
- goto out;
- }
-
- /* try to take chan's mutex, if we can't take it right away
- * we have to fallback to software sync to prevent locking
- * order issues
- */
- if (!mutex_trylock(&chan->mutex)) {
- ret = nouveau_fence_wait(fence, true, false);
- goto out_unref;
- }
-
- /* Make wchan wait until it gets signalled */
- ret = semaphore_acquire(wchan, sema);
- if (ret)
- goto out_unlock;
-
- /* Signal the semaphore from chan */
- ret = semaphore_release(chan, sema);
-
-out_unlock:
- mutex_unlock(&chan->mutex);
-out_unref:
- kref_put(&sema->ref, semaphore_free);
-out:
- if (chan)
- nouveau_channel_put_unlocked(&chan);
- return ret;
-}
-
-int
-nouveau_fence_channel_init(struct nouveau_channel *chan)
-{
- struct drm_device *dev = chan->dev;
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nouveau_gpuobj *obj = NULL;
- int ret;
-
- if (dev_priv->card_type < NV_C0) {
- ret = RING_SPACE(chan, 2);
- if (ret)
- return ret;
-
- BEGIN_NV04(chan, NvSubSw, NV01_SUBCHAN_OBJECT, 1);
- OUT_RING (chan, NvSw);
- FIRE_RING (chan);
- }
-
- /* Setup area of memory shared between all channels for x-chan sync */
- if (USE_SEMA(dev) && dev_priv->chipset < 0x84) {
- struct ttm_mem_reg *mem = &dev_priv->fence.bo->bo.mem;
-
- ret = nouveau_gpuobj_dma_new(chan, NV_CLASS_DMA_FROM_MEMORY,
- mem->start << PAGE_SHIFT,
- mem->size, NV_MEM_ACCESS_RW,
- NV_MEM_TARGET_VRAM, &obj);
- if (ret)
- return ret;
-
- ret = nouveau_ramht_insert(chan, NvSema, obj);
- nouveau_gpuobj_ref(NULL, &obj);
- if (ret)
- return ret;
- } else
- if (USE_SEMA(dev)) {
- /* map fence bo into channel's vm */
- ret = nouveau_bo_vma_add(dev_priv->fence.bo, chan->vm,
- &chan->fence.vma);
- if (ret)
- return ret;
- }
-
- atomic_set(&chan->fence.last_sequence_irq, 0);
- return 0;
-}
-
-void
-nouveau_fence_channel_fini(struct nouveau_channel *chan)
-{
- struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
- struct nouveau_fence *tmp, *fence;
-
- spin_lock(&chan->fence.lock);
- list_for_each_entry_safe(fence, tmp, &chan->fence.pending, head) {
- fence->channel = NULL;
- list_del(&fence->head);
-
- if (unlikely(fence->work))
- fence->work(fence->priv, false);
-
- kref_put(&fence->kref, nouveau_fence_del);
- }
- spin_unlock(&chan->fence.lock);
-
- nouveau_bo_vma_del(dev_priv->fence.bo, &chan->fence.vma);
-}
-
-int
-nouveau_fence_init(struct drm_device *dev)
-{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- int size = (dev_priv->chipset < 0x84) ? 4096 : 16384;
- int ret;
-
- /* Create a shared VRAM heap for cross-channel sync. */
- if (USE_SEMA(dev)) {
- ret = nouveau_bo_new(dev, size, 0, TTM_PL_FLAG_VRAM,
- 0, 0, NULL, &dev_priv->fence.bo);
- if (ret)
- return ret;
-
- ret = nouveau_bo_pin(dev_priv->fence.bo, TTM_PL_FLAG_VRAM);
- if (ret)
- goto fail;
-
- ret = nouveau_bo_map(dev_priv->fence.bo);
- if (ret)
- goto fail;
-
- ret = drm_mm_init(&dev_priv->fence.heap, 0,
- dev_priv->fence.bo->bo.mem.size);
- if (ret)
- goto fail;
-
- spin_lock_init(&dev_priv->fence.lock);
- }
-
- return 0;
-fail:
- nouveau_bo_unmap(dev_priv->fence.bo);
- nouveau_bo_ref(NULL, &dev_priv->fence.bo);
- return ret;
-}
-
-void
-nouveau_fence_fini(struct drm_device *dev)
-{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
-
- if (USE_SEMA(dev)) {
- drm_mm_takedown(&dev_priv->fence.heap);
- nouveau_bo_unmap(dev_priv->fence.bo);
- nouveau_bo_unpin(dev_priv->fence.bo);
- nouveau_bo_ref(NULL, &dev_priv->fence.bo);
- }
-}