RE: [PATCH 09/17] drm/radeon: use common fence implementation for fences

[Deucher, Alexander]

> -----Original Message-----
> From: Maarten Lankhorst [mailto:maarten.lankhorst@xxxxxxxxxxxxx]
> Sent: Wednesday, July 09, 2014 8:30 AM
> To: airlied@xxxxxxxx
> Cc: thellstrom@xxxxxxxxxx; nouveau@xxxxxxxxxxxxxxxxxxxxx; linux-
> kernel@xxxxxxxxxxxxxxx; dri-devel@xxxxxxxxxxxxxxxxxxxxx;
> bskeggs@xxxxxxxxxx; Deucher, Alexander; Koenig, Christian
> Subject: [PATCH 09/17] drm/radeon: use common fence implementation for
> fences
> 
> Signed-off-by: Maarten Lankhorst <maarten.lankhorst@xxxxxxxxxxxxx>
> ---
>  drivers/gpu/drm/radeon/radeon.h        |   15 +-
>  drivers/gpu/drm/radeon/radeon_device.c |   60 ++++++++-
>  drivers/gpu/drm/radeon/radeon_fence.c  |  223
> ++++++++++++++++++++++++++------
>  3 files changed, 248 insertions(+), 50 deletions(-)
> 
> diff --git a/drivers/gpu/drm/radeon/radeon.h
> b/drivers/gpu/drm/radeon/radeon.h
> index 29d9cc04c04e..03a5567f2c2f 100644
> --- a/drivers/gpu/drm/radeon/radeon.h
> +++ b/drivers/gpu/drm/radeon/radeon.h
> @@ -64,6 +64,7 @@
>  #include <linux/wait.h>
>  #include <linux/list.h>
>  #include <linux/kref.h>
> +#include <linux/fence.h>
> 
>  #include <ttm/ttm_bo_api.h>
>  #include <ttm/ttm_bo_driver.h>
> @@ -116,9 +117,6 @@ extern int radeon_deep_color;
>  #define RADEONFB_CONN_LIMIT			4
>  #define RADEON_BIOS_NUM_SCRATCH			8
> 
> -/* fence seq are set to this number when signaled */
> -#define RADEON_FENCE_SIGNALED_SEQ		0LL
> -
>  /* internal ring indices */
>  /* r1xx+ has gfx CP ring */
>  #define RADEON_RING_TYPE_GFX_INDEX		0
> @@ -350,12 +348,15 @@ struct radeon_fence_driver {
>  };
> 
>  struct radeon_fence {
> +	struct fence base;
> +
>  	struct radeon_device		*rdev;
> -	struct kref			kref;
>  	/* protected by radeon_fence.lock */
>  	uint64_t			seq;
>  	/* RB, DMA, etc. */
>  	unsigned			ring;
> +
> +	wait_queue_t fence_wake;
>  };
> 
>  int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring);
> @@ -2268,6 +2269,7 @@ struct radeon_device {
>  	struct radeon_mman		mman;
>  	struct radeon_fence_driver	fence_drv[RADEON_NUM_RINGS];
>  	wait_queue_head_t		fence_queue;
> +	unsigned			fence_context;
>  	struct mutex			ring_lock;
>  	struct radeon_ring		ring[RADEON_NUM_RINGS];
>  	bool				ib_pool_ready;
> @@ -2358,11 +2360,6 @@ u32 cik_mm_rdoorbell(struct radeon_device
> *rdev, u32 index);
>  void cik_mm_wdoorbell(struct radeon_device *rdev, u32 index, u32 v);
> 
>  /*
> - * Cast helper
> - */
> -#define to_radeon_fence(p) ((struct radeon_fence *)(p))
> -
> -/*
>   * Registers read & write functions.
>   */
>  #define RREG8(reg) readb((rdev->rmmio) + (reg))
> diff --git a/drivers/gpu/drm/radeon/radeon_device.c
> b/drivers/gpu/drm/radeon/radeon_device.c
> index 03686fab842d..86699df7c8f3 100644
> --- a/drivers/gpu/drm/radeon/radeon_device.c
> +++ b/drivers/gpu/drm/radeon/radeon_device.c
> @@ -1213,6 +1213,7 @@ int radeon_device_init(struct radeon_device
> *rdev,
>  	for (i = 0; i < RADEON_NUM_RINGS; i++) {
>  		rdev->ring[i].idx = i;
>  	}
> +	rdev->fence_context =
> fence_context_alloc(RADEON_NUM_RINGS);
> 
>  	DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X
> 0x%04X:0x%04X).\n",
>  		radeon_family_name[rdev->family], pdev->vendor, pdev-
> >device,
> @@ -1607,6 +1608,54 @@ int radeon_resume_kms(struct drm_device *dev,
> bool resume, bool fbcon)
>  	return 0;
>  }
> 
> +static uint32_t radeon_gpu_mask_sw_irq(struct radeon_device *rdev)
> +{
> +	uint32_t mask = 0;
> +	int i;
> +
> +	if (!rdev->ddev->irq_enabled)
> +		return mask;
> +
> +	/*
> +	 * increase refcount on sw interrupts for all rings to stop
> +	 * enabling interrupts in radeon_fence_enable_signaling during
> +	 * gpu reset.
> +	 */
> +
> +	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> +		if (!rdev->ring[i].ready)
> +			continue;
> +
> +		atomic_inc(&rdev->irq.ring_int[i]);
> +		mask |= 1 << i;
> +	}
> +	return mask;
> +}
> +
> +static void radeon_gpu_unmask_sw_irq(struct radeon_device *rdev,
> uint32_t mask)
> +{
> +	unsigned long irqflags;
> +	int i;
> +
> +	if (!mask)
> +		return;
> +
> +	/*
> +	 * undo refcount increase, and reset irqs to correct value.
> +	 */
> +
> +	for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> +		if (!(mask & (1 << i)))
> +			continue;
> +
> +		atomic_dec(&rdev->irq.ring_int[i]);
> +	}
> +
> +	spin_lock_irqsave(&rdev->irq.lock, irqflags);
> +	radeon_irq_set(rdev);
> +	spin_unlock_irqrestore(&rdev->irq.lock, irqflags);
> +}
> +
>  /**
>   * radeon_gpu_reset - reset the asic
>   *
> @@ -1624,6 +1673,7 @@ int radeon_gpu_reset(struct radeon_device *rdev)
> 
>  	int i, r;
>  	int resched;
> +	uint32_t sw_mask;
> 
>  	down_write(&rdev->exclusive_lock);
> 
> @@ -1637,6 +1687,7 @@ int radeon_gpu_reset(struct radeon_device *rdev)
>  	radeon_save_bios_scratch_regs(rdev);
>  	/* block TTM */
>  	resched = ttm_bo_lock_delayed_workqueue(&rdev->mman.bdev);
> +	sw_mask = radeon_gpu_mask_sw_irq(rdev);
>  	radeon_pm_suspend(rdev);
>  	radeon_suspend(rdev);
> 
> @@ -1686,13 +1737,20 @@ retry:
>  	radeon_pm_resume(rdev);
>  	drm_helper_resume_force_mode(rdev->ddev);
> 
> +	radeon_gpu_unmask_sw_irq(rdev, sw_mask);
>  	ttm_bo_unlock_delayed_workqueue(&rdev->mman.bdev,
> resched);
>  	if (r) {
>  		/* bad news, how to tell it to userspace ? */
>  		dev_info(rdev->dev, "GPU reset failed\n");
>  	}
> 
> -	up_write(&rdev->exclusive_lock);
> +	/*
> +	 * force all waiters to recheck, some may have been
> +	 * added while the exclusive_lock was unavailable
> +	 */
> +	downgrade_write(&rdev->exclusive_lock);
> +	wake_up_all(&rdev->fence_queue);
> +	up_read(&rdev->exclusive_lock);
>  	return r;
>  }
> 
> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c
> b/drivers/gpu/drm/radeon/radeon_fence.c
> index 6435719fd45b..81c98f6ff0ca 100644
> --- a/drivers/gpu/drm/radeon/radeon_fence.c
> +++ b/drivers/gpu/drm/radeon/radeon_fence.c
> @@ -39,6 +39,15 @@
>  #include "radeon.h"
>  #include "radeon_trace.h"
> 
> +static const struct fence_ops radeon_fence_ops;
> +
> +#define to_radeon_fence(p) \
> +	({								\
> +		struct radeon_fence *__f;				\
> +		__f = container_of((p), struct radeon_fence, base);	\
> +		__f->base.ops == &radeon_fence_ops ? __f : NULL;	\
> +	})
> +
>  /*
>   * Fences
>   * Fences mark an event in the GPUs pipeline and are used
> @@ -111,30 +120,55 @@ int radeon_fence_emit(struct radeon_device
> *rdev,
>  		      struct radeon_fence **fence,
>  		      int ring)
>  {
> +	u64 seq = ++rdev->fence_drv[ring].sync_seq[ring];
> +
>  	/* we are protected by the ring emission mutex */
>  	*fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL);
>  	if ((*fence) == NULL) {
>  		return -ENOMEM;
>  	}
> -	kref_init(&((*fence)->kref));
> -	(*fence)->rdev = rdev;
> -	(*fence)->seq = ++rdev->fence_drv[ring].sync_seq[ring];
>  	(*fence)->ring = ring;
> +	fence_init(&(*fence)->base, &radeon_fence_ops,
> +		   &rdev->fence_queue.lock, rdev->fence_context + ring,
> seq);
> +	(*fence)->rdev = rdev;
> +	(*fence)->seq = seq;
>  	radeon_fence_ring_emit(rdev, ring, *fence);
>  	trace_radeon_fence_emit(rdev->ddev, ring, (*fence)->seq);
>  	return 0;
>  }
> 
>  /**
> - * radeon_fence_process - process a fence
> + * radeon_fence_check_signaled - callback from fence_queue
>   *
> - * @rdev: radeon_device pointer
> - * @ring: ring index the fence is associated with
> - *
> - * Checks the current fence value and wakes the fence queue
> - * if the sequence number has increased (all asics).
> + * this function is called with fence_queue lock held, which is also used
> + * for the fence locking itself, so unlocked variants are used for
> + * fence_signal, and remove_wait_queue.
>   */
> -void radeon_fence_process(struct radeon_device *rdev, int ring)
> +static int radeon_fence_check_signaled(wait_queue_t *wait, unsigned
> mode, int flags, void *key)
> +{
> +	struct radeon_fence *fence;
> +	u64 seq;
> +
> +	fence = container_of(wait, struct radeon_fence, fence_wake);
> +
> +	seq = atomic64_read(&fence->rdev->fence_drv[fence-
> >ring].last_seq);
> +	if (seq >= fence->seq) {
> +		int ret = fence_signal_locked(&fence->base);
> +
> +		if (!ret)
> +			FENCE_TRACE(&fence->base, "signaled from irq
> context\n");
> +		else
> +			FENCE_TRACE(&fence->base, "was already
> signaled\n");
> +
> +		radeon_irq_kms_sw_irq_put(fence->rdev, fence->ring);
> +		__remove_wait_queue(&fence->rdev->fence_queue,
> &fence->fence_wake);
> +		fence_put(&fence->base);
> +	} else
> +		FENCE_TRACE(&fence->base, "pending\n");
> +	return 0;
> +}
> +
> +static bool __radeon_fence_process(struct radeon_device *rdev, int ring)
>  {
>  	uint64_t seq, last_seq, last_emitted;
>  	unsigned count_loop = 0;
> @@ -190,23 +224,22 @@ void radeon_fence_process(struct radeon_device
> *rdev, int ring)
>  		}
>  	} while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) >
> seq);
> 
> -	if (wake)
> -		wake_up_all(&rdev->fence_queue);
> +	return wake;
>  }
> 
>  /**
> - * radeon_fence_destroy - destroy a fence
> + * radeon_fence_process - process a fence
>   *
> - * @kref: fence kref
> + * @rdev: radeon_device pointer
> + * @ring: ring index the fence is associated with
>   *
> - * Frees the fence object (all asics).
> + * Checks the current fence value and wakes the fence queue
> + * if the sequence number has increased (all asics).
>   */
> -static void radeon_fence_destroy(struct kref *kref)
> +void radeon_fence_process(struct radeon_device *rdev, int ring)
>  {
> -	struct radeon_fence *fence;
> -
> -	fence = container_of(kref, struct radeon_fence, kref);
> -	kfree(fence);
> +	if (__radeon_fence_process(rdev, ring))
> +		wake_up_all(&rdev->fence_queue);
>  }
> 
>  /**
> @@ -237,6 +270,69 @@ static bool radeon_fence_seq_signaled(struct
> radeon_device *rdev,
>  	return false;
>  }
> 
> +static bool __radeon_fence_signaled(struct fence *f)
> +{
> +	struct radeon_fence *fence = to_radeon_fence(f);
> +	struct radeon_device *rdev = fence->rdev;
> +	unsigned ring = fence->ring;
> +	u64 seq = fence->seq;
> +
> +	if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) {
> +		return true;
> +	}
> +
> +	if (down_read_trylock(&rdev->exclusive_lock)) {
> +		radeon_fence_process(rdev, ring);
> +		up_read(&rdev->exclusive_lock);
> +
> +		if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq)
> {
> +			return true;
> +		}
> +	}
> +	return false;
> +}
> +
> +/**
> + * radeon_fence_enable_signaling - enable signalling on fence
> + * @fence: fence
> + *
> + * This function is called with fence_queue lock held, and adds a callback
> + * to fence_queue that checks if this fence is signaled, and if so it
> + * signals the fence and removes itself.
> + */
> +static bool radeon_fence_enable_signaling(struct fence *f)
> +{
> +	struct radeon_fence *fence = to_radeon_fence(f);
> +	struct radeon_device *rdev = fence->rdev;
> +
> +	if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >=
> fence->seq ||
> +	    !rdev->ddev->irq_enabled)
> +		return false;
> +
> +	radeon_irq_kms_sw_irq_get(rdev, fence->ring);
> +
> +	if (down_read_trylock(&rdev->exclusive_lock)) {
> +		if (__radeon_fence_process(rdev, fence->ring))
> +			wake_up_all_locked(&rdev->fence_queue);
> +
> +		up_read(&rdev->exclusive_lock);
> +	}
> +
> +	/* did fence get signaled after we enabled the sw irq? */
> +	if (atomic64_read(&rdev->fence_drv[fence->ring].last_seq) >=
> fence->seq) {
> +		radeon_irq_kms_sw_irq_put(rdev, fence->ring);
> +		return false;
> +	}
> +
> +	fence->fence_wake.flags = 0;
> +	fence->fence_wake.private = NULL;
> +	fence->fence_wake.func = radeon_fence_check_signaled;
> +	__add_wait_queue(&rdev->fence_queue, &fence->fence_wake);
> +	fence_get(f);
> +
> +	return true;
> +}
> +
>  /**
>   * radeon_fence_signaled - check if a fence has signaled
>   *
> @@ -250,11 +346,13 @@ bool radeon_fence_signaled(struct radeon_fence
> *fence)
>  	if (!fence) {
>  		return true;
>  	}
> -	if (fence->seq == RADEON_FENCE_SIGNALED_SEQ) {
> -		return true;
> -	}
> +
>  	if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence-
> >ring)) {
> -		fence->seq = RADEON_FENCE_SIGNALED_SEQ;
> +		int ret;
> +
> +		ret = fence_signal(&fence->base);
> +		if (!ret)
> +			FENCE_TRACE(&fence->base, "signaled from
> radeon_fence_signaled\n");
>  		return true;
>  	}
>  	return false;
> @@ -413,21 +511,18 @@ int radeon_fence_wait(struct radeon_fence
> *fence, bool intr)
>  	uint64_t seq[RADEON_NUM_RINGS] = {};
>  	long r;
> 
> -	if (fence == NULL) {
> -		WARN(1, "Querying an invalid fence : %p !\n", fence);
> -		return -EINVAL;
> -	}
> -
> -	seq[fence->ring] = fence->seq;
> -	if (seq[fence->ring] == RADEON_FENCE_SIGNALED_SEQ)
> +	if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fence->base.flags))
>  		return 0;
> 
> +	seq[fence->ring] = fence->seq;
>  	r = radeon_fence_wait_seq_timeout(fence->rdev, seq, intr,
> MAX_SCHEDULE_TIMEOUT);
>  	if (r < 0) {
>  		return r;
>  	}
> 
> -	fence->seq = RADEON_FENCE_SIGNALED_SEQ;
> +	r = fence_signal(&fence->base);
> +	if (!r)
> +		FENCE_TRACE(&fence->base, "signaled from
> fence_wait\n");
>  	return 0;
>  }
> 
> @@ -459,12 +554,13 @@ int radeon_fence_wait_any(struct radeon_device
> *rdev,
>  			continue;
>  		}
> 
> +		if (test_bit(FENCE_FLAG_SIGNALED_BIT, &fences[i]-
> >base.flags)) {
> +			/* already signaled */
> +			return 0;
> +		}
> +
>  		seq[i] = fences[i]->seq;
>  		++num_rings;
> -
> -		/* test if something was allready signaled */
> -		if (seq[i] == RADEON_FENCE_SIGNALED_SEQ)
> -			return 0;
>  	}
> 
>  	/* nothing to wait for ? */
> @@ -545,7 +641,7 @@ int radeon_fence_wait_empty(struct radeon_device
> *rdev, int ring)
>   */
>  struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence)
>  {
> -	kref_get(&fence->kref);
> +	fence_get(&fence->base);
>  	return fence;
>  }
> 
> @@ -561,9 +657,8 @@ void radeon_fence_unref(struct radeon_fence
> **fence)
>  	struct radeon_fence *tmp = *fence;
> 
>  	*fence = NULL;
> -	if (tmp) {
> -		kref_put(&tmp->kref, radeon_fence_destroy);
> -	}
> +	if (tmp)
> +		fence_put(&tmp->base);
>  }
> 
>  /**
> @@ -872,3 +967,51 @@ int radeon_debugfs_fence_init(struct
> radeon_device *rdev)
>  	return 0;
>  #endif
>  }
> +
> +static long __radeon_fence_wait(struct fence *f, bool intr, long timeout)
> +{
> +	struct radeon_fence *fence = to_radeon_fence(f);
> +	u64 target_seq[RADEON_NUM_RINGS] = {};
> +	struct radeon_device *rdev = fence->rdev;
> +	long r;
> +
> +	target_seq[fence->ring] = fence->seq;
> +
> +	down_read(&rdev->exclusive_lock);
> +	r = radeon_fence_wait_seq_timeout(fence->rdev, target_seq, intr,
> timeout);
> +
> +	if (r > 0 && !fence_signal(&fence->base))
> +		FENCE_TRACE(&fence->base, "signaled from
> __radeon_fence_wait\n");
> +
> +	up_read(&rdev->exclusive_lock);
> +	return r;
> +
> +}
> +
> +static const char *radeon_fence_get_driver_name(struct fence *fence)
> +{
> +	return "radeon";
> +}
> +
> +static const char *radeon_fence_get_timeline_name(struct fence *f)
> +{
> +	struct radeon_fence *fence = to_radeon_fence(f);
> +	switch (fence->ring) {
> +	case RADEON_RING_TYPE_GFX_INDEX: return "radeon.gfx";
> +	case CAYMAN_RING_TYPE_CP1_INDEX: return "radeon.cp1";
> +	case CAYMAN_RING_TYPE_CP2_INDEX: return "radeon.cp2";
> +	case R600_RING_TYPE_DMA_INDEX: return "radeon.dma";
> +	case CAYMAN_RING_TYPE_DMA1_INDEX: return "radeon.dma1";
> +	case R600_RING_TYPE_UVD_INDEX: return "radeon.uvd";

Radeon supports vce rings on newer ascis.  Probably want to add the case for those here too.

Alex

> +	default: WARN_ON_ONCE(1); return "radeon.unk";
> +	}
> +}
> +
> +static const struct fence_ops radeon_fence_ops = {
> +	.get_driver_name = radeon_fence_get_driver_name,
> +	.get_timeline_name = radeon_fence_get_timeline_name,
> +	.enable_signaling = radeon_fence_enable_signaling,
> +	.signaled = __radeon_fence_signaled,
> +	.wait = __radeon_fence_wait,
> +	.release = NULL,
> +};

N‹§²æ¸›yú²X¬¶ÇvØ–)Þ{.nlj·¥Š{±‘êX§¶›¡Ü}©ž²ÆzÚj:+v‰¨¾«‘êZ+€Êzf£¢·hšˆ§~†­†Ûÿû®w¥¢¸?™¨è&¢)ßf”ùy§m…á«a¶Úÿ0¶ìå