drivers/gpu/drm/i915/i915_gem_context.c

/*
 * Copyright © 2011-2012 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
 * IN THE SOFTWARE.
 *
 * Authors:
 *    Ben Widawsky <ben@bwidawsk.net>
 *
 */

/*
 * This file implements HW context support. On gen5+ a HW context consists of an
 * opaque GPU object which is referenced at times of context saves and restores.
 * With RC6 enabled, the context is also referenced as the GPU enters and exists
 * from RC6 (GPU has it's own internal power context, except on gen5). Though
 * something like a context does exist for the media ring, the code only
 * supports contexts for the render ring.
 *
 * In software, there is a distinction between contexts created by the user,
 * and the default HW context. The default HW context is used by GPU clients
 * that do not request setup of their own hardware context. The default
 * context's state is never restored to help prevent programming errors. This
 * would happen if a client ran and piggy-backed off another clients GPU state.
 * The default context only exists to give the GPU some offset to load as the
 * current to invoke a save of the context we actually care about. In fact, the
 * code could likely be constructed, albeit in a more complicated fashion, to
 * never use the default context, though that limits the driver's ability to
 * swap out, and/or destroy other contexts.
 *
 * All other contexts are created as a request by the GPU client. These contexts
 * store GPU state, and thus allow GPU clients to not re-emit state (and
 * potentially query certain state) at any time. The kernel driver makes
 * certain that the appropriate commands are inserted.
 *
 * The context life cycle is semi-complicated in that context BOs may live
 * longer than the context itself because of the way the hardware, and object
 * tracking works. Below is a very crude representation of the state machine
 * describing the context life.
 *                                         refcount     pincount     active
 * S0: initial state                          0            0           0
 * S1: context created                        1            0           0
 * S2: context is currently running           2            1           X
 * S3: GPU referenced, but not current        2            0           1
 * S4: context is current, but destroyed      1            1           0
 * S5: like S3, but destroyed                 1            0           1
 *
 * The most common (but not all) transitions:
 * S0->S1: client creates a context
 * S1->S2: client submits execbuf with context
 * S2->S3: other clients submits execbuf with context
 * S3->S1: context object was retired
 * S3->S2: clients submits another execbuf
 * S2->S4: context destroy called with current context
 * S3->S5->S0: destroy path
 * S4->S5->S0: destroy path on current context
 *
 * There are two confusing terms used above:
 *  The "current context" means the context which is currently running on the
 *  GPU. The GPU has loaded it's state already and has stored away the gtt
 *  offset of the BO. The GPU is not actively referencing the data at this
 *  offset, but it will on the next context switch. The only way to avoid this
 *  is to do a GPU reset.
 *
 *  An "active context' is one which was previously the "current context" and is
 *  on the active list waiting for the next context switch to occur. Until this
 *  happens, the object must remain at the same gtt offset. It is therefore
 *  possible to destroy a context, but it is still active.
 *
 */

#include "drmP.h"
#include "i915_drm.h"
#include "i915_drv.h"

/* This is a HW constraint. The value below is the largest known requirement
 * I've seen in a spec to date, and that was a workaround for a non-shipping
 * part. It should be safe to decrease this, but it's more future proof as is.
 */
#define CONTEXT_ALIGN (64<<10)

static struct i915_hw_context *
i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id);

static int get_context_size(struct drm_device *dev)
{
	struct drm_i915_private *dev_priv = dev->dev_private;
	int ret;
	u32 reg;

	switch (INTEL_INFO(dev)->gen) {
	case 6:
		reg = I915_READ(CXT_SIZE);
		ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
		break;
	case 7:
		reg = I915_READ(GEN7_CTX_SIZE);
		ret = GEN7_CTX_TOTAL_SIZE(reg) * 64;
		break;
	default:
		BUG();
	}

	return ret;
}

static void do_destroy(struct i915_hw_context *ctx)
{
	struct drm_device *dev = ctx->obj->base.dev;
	struct drm_i915_private *dev_priv = dev->dev_private;

	if (ctx->file_priv)
		idr_remove(&ctx->file_priv->context_idr, ctx->id);
	else
		BUG_ON(ctx != dev_priv->ring[RCS].default_context);

	drm_gem_object_unreference(&ctx->obj->base);
	kfree(ctx);
}

static int
create_hw_context(struct drm_device *dev,
		  struct drm_i915_file_private *file_priv,
		  struct i915_hw_context **ctx_out)
{
	struct drm_i915_private *dev_priv = dev->dev_private;
	int ret, id;

	*ctx_out = kzalloc(sizeof(struct drm_i915_file_private), GFP_KERNEL);
	if (*ctx_out == NULL)
		return -ENOMEM;

	(*ctx_out)->obj = i915_gem_alloc_object(dev,
						dev_priv->hw_context_size);
	if ((*ctx_out)->obj == NULL) {
		kfree(*ctx_out);
		DRM_DEBUG_DRIVER("Context object allocated failed\n");
		return -ENOMEM;
	}

	/* The ring associated with the context object is handled by the normal
	 * object tracking code. We give an initial ring value simple to pass an
	 * assertion in the context switch code.
	 */
	(*ctx_out)->ring = &dev_priv->ring[RCS];

	/* Default context will never have a file_priv */
	if (file_priv == NULL)
		return 0;

	(*ctx_out)->file_priv = file_priv;

again:
	if (idr_pre_get(&file_priv->context_idr, GFP_KERNEL) == 0) {
		ret = -ENOMEM;
		DRM_DEBUG_DRIVER("idr allocation failed\n");
		goto err_out;
	}

	ret = idr_get_new_above(&file_priv->context_idr, *ctx_out,
				DEFAULT_CONTEXT_ID + 1, &id);
	if (ret == 0)
		(*ctx_out)->id = id;

	if (ret == -EAGAIN)
		goto again;
	else if (ret)
		goto err_out;

	return 0;

err_out:
	do_destroy(*ctx_out);
	return ret;
}

static inline bool is_default_context(struct i915_hw_context *ctx)
{
	return (ctx == ctx->ring->default_context);
}

/**
 * The default context needs to exist per ring that uses contexts. It stores the
 * context state of the GPU for applications that don't utilize HW contexts, as
 * well as an idle case.
 */
static int create_default_context(struct drm_i915_private *dev_priv)
{
	struct i915_hw_context *ctx;
	int ret;

	BUG_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));

	ret = create_hw_context(dev_priv->dev, NULL,
				&dev_priv->ring[RCS].default_context);
	if (ret)
		return ret;

	/* We may need to do things with the shrinker which require us to
	 * immediately switch back to the default context. This can cause a
	 * problem as pinning the default context also requires GTT space which
	 * may not be available. To avoid this we always pin the
	 * default context.
	 */
	ctx = dev_priv->ring[RCS].default_context;
	ret = i915_gem_object_pin(ctx->obj, CONTEXT_ALIGN, false);
	if (ret) {
		do_destroy(ctx);
		return ret;
	}

	return ret;
}

void i915_gem_context_init(struct drm_device *dev)
{
	struct drm_i915_private *dev_priv = dev->dev_private;
	uint32_t ctx_size;

	if (!HAS_HW_CONTEXTS(dev))
		return;

	/* If called from reset, or thaw... we've been here already */
	if (dev_priv->hw_contexts_disabled ||
	    dev_priv->ring[RCS].default_context)
		return;

	ctx_size = get_context_size(dev);
	dev_priv->hw_context_size = get_context_size(dev);
	dev_priv->hw_context_size = round_up(dev_priv->hw_context_size, 4096);

	if (ctx_size <= 0 || ctx_size > (1<<20)) {
		dev_priv->hw_contexts_disabled = true;
		return;
	}

	if (create_default_context(dev_priv)) {
		dev_priv->hw_contexts_disabled = true;
		return;
	}

	DRM_DEBUG_DRIVER("HW context support initialized\n");
}

void i915_gem_context_fini(struct drm_device *dev)
{
	struct drm_i915_private *dev_priv = dev->dev_private;

	if (dev_priv->hw_contexts_disabled)
		return;

	i915_gem_object_unpin(dev_priv->ring[RCS].default_context->obj);

	do_destroy(dev_priv->ring[RCS].default_context);
}

void i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
{
	struct drm_i915_private *dev_priv = dev->dev_private;
	struct drm_i915_file_private *file_priv = file->driver_priv;

	if (dev_priv->hw_contexts_disabled)
		return;

	idr_init(&file_priv->context_idr);
}

static int context_idr_cleanup(int id, void *p, void *data)
{
	struct drm_file *file = (struct drm_file *)data;
	struct drm_i915_file_private *file_priv = file->driver_priv;
	struct i915_hw_context *ctx;

	BUG_ON(id == DEFAULT_CONTEXT_ID);
	ctx = i915_gem_context_get(file_priv, id);
	if (WARN_ON(ctx == NULL))
		return -ENXIO;

	do_destroy(ctx);

	return 0;
}

void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
{
	struct drm_i915_private *dev_priv = dev->dev_private;
	struct drm_i915_file_private *file_priv = file->driver_priv;

	if (dev_priv->hw_contexts_disabled)
		return;

	mutex_lock(&dev->struct_mutex);
	idr_for_each(&file_priv->context_idr, context_idr_cleanup, file);
	idr_destroy(&file_priv->context_idr);
	mutex_unlock(&dev->struct_mutex);
}

static struct i915_hw_context *
i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
{
	return (struct i915_hw_context *)idr_find(&file_priv->context_idr, id);
}

static inline int
mi_set_context(struct intel_ring_buffer *ring,
	       struct i915_hw_context *new_context,
	       u32 hw_flags)
{
	int ret;

	ret = intel_ring_begin(ring, 6);
	if (ret)
		return ret;

	if (IS_GEN7(ring->dev))
		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
	else
		intel_ring_emit(ring, MI_NOOP);

	intel_ring_emit(ring, MI_NOOP);
	intel_ring_emit(ring, MI_SET_CONTEXT);
	intel_ring_emit(ring, new_context->obj->gtt_offset |
			MI_MM_SPACE_GTT |
			MI_SAVE_EXT_STATE_EN |
			MI_RESTORE_EXT_STATE_EN |
			hw_flags);
	/* w/a: MI_SET_CONTEXT must always be followed by MI_NOOP */
	intel_ring_emit(ring, MI_NOOP);

	if (IS_GEN7(ring->dev))
		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
	else
		intel_ring_emit(ring, MI_NOOP);

	intel_ring_advance(ring);

	return ret;
}

static int do_switch(struct drm_i915_gem_object *from_obj,
		     struct i915_hw_context *to,
		     u32 seqno)
{
	struct intel_ring_buffer *ring = NULL;
	u32 hw_flags = 0;
	int ret;

	BUG_ON(to == NULL);
	BUG_ON(from_obj != NULL && from_obj->pin_count == 0);

	ret = i915_gem_object_pin(to->obj, CONTEXT_ALIGN, false);
	if (ret)
		return ret;

	if (!to->obj->has_global_gtt_mapping)
		i915_gem_gtt_bind_object(to->obj, to->obj->cache_level);

	if (!to->is_initialized || is_default_context(to))
		hw_flags |= MI_RESTORE_INHIBIT;
	else if (WARN_ON_ONCE(from_obj == to->obj)) /* not yet expected */
		hw_flags |= MI_FORCE_RESTORE;

	ring = to->ring;
	ret = mi_set_context(ring, to, hw_flags);
	if (ret) {
		i915_gem_object_unpin(to->obj);
		return ret;
	}

	/* The backing object for the context is done after switching to the
	 * *next* context. Therefore we cannot retire the previous context until
	 * the next context has already started running. In fact, the below code
	 * is a bit suboptimal because the retiring can occur simply after the
	 * MI_SET_CONTEXT instead of when the next seqno has completed.
	 */
	if (from_obj != NULL) {
		from_obj->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
		i915_gem_object_move_to_active(from_obj, ring, seqno);
		/* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
		 * whole damn pipeline, we don't need to explicitly mark the
		 * object dirty. The only exception is that the context must be
		 * correct in case the object gets swapped out. Ideally we'd be
		 * able to defer doing this until we know the object would be
		 * swapped, but there is no way to do that yet.
		 */
		from_obj->dirty = 1;
		BUG_ON(from_obj->ring != to->ring);
		i915_gem_object_unpin(from_obj);
	}

	ring->last_context_obj = to->obj;
	to->is_initialized = true;

	return 0;
}

/**
 * i915_switch_context() - perform a GPU context switch.
 * @ring: ring for which we'll execute the context switch
 * @file_priv: file_priv associated with the context, may be NULL
 * @id: context id number
 * @seqno: sequence number by which the new context will be switched to
 * @flags:
 *
 * The context life cycle is simple. The context refcount is incremented and
 * decremented by 1 and create and destroy. If the context is in use by the GPU,
 * it will have a refoucnt > 1. This allows us to destroy the context abstract
 * object while letting the normal object tracking destroy the backing BO.
 */
int i915_switch_context(struct intel_ring_buffer *ring,
			struct drm_file *file,
			int to_id)
{
	struct drm_i915_private *dev_priv = ring->dev->dev_private;
	struct drm_i915_file_private *file_priv = NULL;
	struct i915_hw_context *to;
	struct drm_i915_gem_object *from_obj = ring->last_context_obj;
	int ret;

	if (dev_priv->hw_contexts_disabled)
		return 0;

	if (ring != &dev_priv->ring[RCS])
		return 0;

	if (file)
		file_priv = file->driver_priv;

	if (to_id == DEFAULT_CONTEXT_ID) {
		to = ring->default_context;
	} else {
		to = i915_gem_context_get(file_priv, to_id);
		if (to == NULL)
			return -EINVAL;
	}

	if (from_obj == to->obj)
		return 0;

	ret = do_switch(from_obj, to, i915_gem_next_request_seqno(to->ring));
	if (ret)
		return ret;

	/* Just to make the code a little cleaner we take the object reference
	 * after the switch was successful. It would be more intuitive to ref
	 * the 'to' object before the switch but we know the refcount must be >0
	 * if context_get() succeeded, and we hold struct mutex. So it's safe to
	 * do this here/now
	 */
	drm_gem_object_reference(&to->obj->base);
	if (from_obj != NULL)
		drm_gem_object_unreference(&from_obj->base);
	return ret;
}