linux/drivers/gpu/drm/i915/intel_fbc.c

1282 lines
35 KiB
C
Raw Normal View History

/*
* Copyright © 2014 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.
*/
/**
* DOC: Frame Buffer Compression (FBC)
*
* FBC tries to save memory bandwidth (and so power consumption) by
* compressing the amount of memory used by the display. It is total
* transparent to user space and completely handled in the kernel.
*
* The benefits of FBC are mostly visible with solid backgrounds and
* variation-less patterns. It comes from keeping the memory footprint small
* and having fewer memory pages opened and accessed for refreshing the display.
*
* i915 is responsible to reserve stolen memory for FBC and configure its
* offset on proper registers. The hardware takes care of all
* compress/decompress. However there are many known cases where we have to
* forcibly disable it to allow proper screen updates.
*/
#include "intel_drv.h"
#include "i915_drv.h"
static inline bool fbc_supported(struct drm_i915_private *dev_priv)
{
return HAS_FBC(dev_priv);
}
static inline bool fbc_on_pipe_a_only(struct drm_i915_private *dev_priv)
{
return IS_HASWELL(dev_priv) || INTEL_INFO(dev_priv)->gen >= 8;
}
static inline bool fbc_on_plane_a_only(struct drm_i915_private *dev_priv)
{
return INTEL_INFO(dev_priv)->gen < 4;
}
static inline bool no_fbc_on_multiple_pipes(struct drm_i915_private *dev_priv)
{
return INTEL_INFO(dev_priv)->gen <= 3;
}
/*
* In some platforms where the CRTC's x:0/y:0 coordinates doesn't match the
* frontbuffer's x:0/y:0 coordinates we lie to the hardware about the plane's
* origin so the x and y offsets can actually fit the registers. As a
* consequence, the fence doesn't really start exactly at the display plane
* address we program because it starts at the real start of the buffer, so we
* have to take this into consideration here.
*/
static unsigned int get_crtc_fence_y_offset(struct intel_crtc *crtc)
{
return crtc->base.y - crtc->adjusted_y;
}
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
/*
* For SKL+, the plane source size used by the hardware is based on the value we
* write to the PLANE_SIZE register. For BDW-, the hardware looks at the value
* we wrote to PIPESRC.
*/
static void intel_fbc_get_plane_source_size(struct intel_fbc_state_cache *cache,
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
int *width, int *height)
{
int w, h;
if (intel_rotation_90_or_270(cache->plane.rotation)) {
w = cache->plane.src_h;
h = cache->plane.src_w;
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
} else {
w = cache->plane.src_w;
h = cache->plane.src_h;
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
}
if (width)
*width = w;
if (height)
*height = h;
}
static int intel_fbc_calculate_cfb_size(struct drm_i915_private *dev_priv,
struct intel_fbc_state_cache *cache)
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
{
int lines;
intel_fbc_get_plane_source_size(cache, NULL, &lines);
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
if (INTEL_INFO(dev_priv)->gen >= 7)
lines = min(lines, 2048);
/* Hardware needs the full buffer stride, not just the active area. */
return lines * cache->fb.stride;
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
}
static void i8xx_fbc_deactivate(struct drm_i915_private *dev_priv)
{
u32 fbc_ctl;
/* Disable compression */
fbc_ctl = I915_READ(FBC_CONTROL);
if ((fbc_ctl & FBC_CTL_EN) == 0)
return;
fbc_ctl &= ~FBC_CTL_EN;
I915_WRITE(FBC_CONTROL, fbc_ctl);
/* Wait for compressing bit to clear */
if (intel_wait_for_register(dev_priv,
FBC_STATUS, FBC_STAT_COMPRESSING, 0,
10)) {
DRM_DEBUG_KMS("FBC idle timed out\n");
return;
}
}
static void i8xx_fbc_activate(struct drm_i915_private *dev_priv)
{
struct intel_fbc_reg_params *params = &dev_priv->fbc.params;
int cfb_pitch;
int i;
u32 fbc_ctl;
/* Note: fbc.threshold == 1 for i8xx */
cfb_pitch = params->cfb_size / FBC_LL_SIZE;
if (params->fb.stride < cfb_pitch)
cfb_pitch = params->fb.stride;
/* FBC_CTL wants 32B or 64B units */
if (IS_GEN2(dev_priv))
cfb_pitch = (cfb_pitch / 32) - 1;
else
cfb_pitch = (cfb_pitch / 64) - 1;
/* Clear old tags */
for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
I915_WRITE(FBC_TAG(i), 0);
if (IS_GEN4(dev_priv)) {
u32 fbc_ctl2;
/* Set it up... */
fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
fbc_ctl2 |= FBC_CTL_PLANE(params->crtc.plane);
I915_WRITE(FBC_CONTROL2, fbc_ctl2);
I915_WRITE(FBC_FENCE_OFF, params->crtc.fence_y_offset);
}
/* enable it... */
fbc_ctl = I915_READ(FBC_CONTROL);
fbc_ctl &= 0x3fff << FBC_CTL_INTERVAL_SHIFT;
fbc_ctl |= FBC_CTL_EN | FBC_CTL_PERIODIC;
if (IS_I945GM(dev_priv))
fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
fbc_ctl |= params->fb.fence_reg;
I915_WRITE(FBC_CONTROL, fbc_ctl);
}
static bool i8xx_fbc_is_active(struct drm_i915_private *dev_priv)
{
return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
}
static void g4x_fbc_activate(struct drm_i915_private *dev_priv)
{
struct intel_fbc_reg_params *params = &dev_priv->fbc.params;
u32 dpfc_ctl;
dpfc_ctl = DPFC_CTL_PLANE(params->crtc.plane) | DPFC_SR_EN;
if (drm_format_plane_cpp(params->fb.pixel_format, 0) == 2)
dpfc_ctl |= DPFC_CTL_LIMIT_2X;
else
dpfc_ctl |= DPFC_CTL_LIMIT_1X;
dpfc_ctl |= DPFC_CTL_FENCE_EN | params->fb.fence_reg;
I915_WRITE(DPFC_FENCE_YOFF, params->crtc.fence_y_offset);
/* enable it... */
I915_WRITE(DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
}
static void g4x_fbc_deactivate(struct drm_i915_private *dev_priv)
{
u32 dpfc_ctl;
/* Disable compression */
dpfc_ctl = I915_READ(DPFC_CONTROL);
if (dpfc_ctl & DPFC_CTL_EN) {
dpfc_ctl &= ~DPFC_CTL_EN;
I915_WRITE(DPFC_CONTROL, dpfc_ctl);
}
}
static bool g4x_fbc_is_active(struct drm_i915_private *dev_priv)
{
return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
}
/* This function forces a CFB recompression through the nuke operation. */
static void intel_fbc_recompress(struct drm_i915_private *dev_priv)
{
I915_WRITE(MSG_FBC_REND_STATE, FBC_REND_NUKE);
POSTING_READ(MSG_FBC_REND_STATE);
}
static void ilk_fbc_activate(struct drm_i915_private *dev_priv)
{
struct intel_fbc_reg_params *params = &dev_priv->fbc.params;
u32 dpfc_ctl;
int threshold = dev_priv->fbc.threshold;
dpfc_ctl = DPFC_CTL_PLANE(params->crtc.plane);
if (drm_format_plane_cpp(params->fb.pixel_format, 0) == 2)
threshold++;
switch (threshold) {
case 4:
case 3:
dpfc_ctl |= DPFC_CTL_LIMIT_4X;
break;
case 2:
dpfc_ctl |= DPFC_CTL_LIMIT_2X;
break;
case 1:
dpfc_ctl |= DPFC_CTL_LIMIT_1X;
break;
}
dpfc_ctl |= DPFC_CTL_FENCE_EN;
if (IS_GEN5(dev_priv))
dpfc_ctl |= params->fb.fence_reg;
I915_WRITE(ILK_DPFC_FENCE_YOFF, params->crtc.fence_y_offset);
I915_WRITE(ILK_FBC_RT_BASE, params->fb.ggtt_offset | ILK_FBC_RT_VALID);
/* enable it... */
I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
if (IS_GEN6(dev_priv)) {
I915_WRITE(SNB_DPFC_CTL_SA,
SNB_CPU_FENCE_ENABLE | params->fb.fence_reg);
I915_WRITE(DPFC_CPU_FENCE_OFFSET, params->crtc.fence_y_offset);
}
intel_fbc_recompress(dev_priv);
}
static void ilk_fbc_deactivate(struct drm_i915_private *dev_priv)
{
u32 dpfc_ctl;
/* Disable compression */
dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
if (dpfc_ctl & DPFC_CTL_EN) {
dpfc_ctl &= ~DPFC_CTL_EN;
I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
}
}
static bool ilk_fbc_is_active(struct drm_i915_private *dev_priv)
{
return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
}
static void gen7_fbc_activate(struct drm_i915_private *dev_priv)
{
struct intel_fbc_reg_params *params = &dev_priv->fbc.params;
u32 dpfc_ctl;
int threshold = dev_priv->fbc.threshold;
dpfc_ctl = 0;
if (IS_IVYBRIDGE(dev_priv))
dpfc_ctl |= IVB_DPFC_CTL_PLANE(params->crtc.plane);
if (drm_format_plane_cpp(params->fb.pixel_format, 0) == 2)
threshold++;
switch (threshold) {
case 4:
case 3:
dpfc_ctl |= DPFC_CTL_LIMIT_4X;
break;
case 2:
dpfc_ctl |= DPFC_CTL_LIMIT_2X;
break;
case 1:
dpfc_ctl |= DPFC_CTL_LIMIT_1X;
break;
}
dpfc_ctl |= IVB_DPFC_CTL_FENCE_EN;
if (dev_priv->fbc.false_color)
dpfc_ctl |= FBC_CTL_FALSE_COLOR;
if (IS_IVYBRIDGE(dev_priv)) {
/* WaFbcAsynchFlipDisableFbcQueue:ivb */
I915_WRITE(ILK_DISPLAY_CHICKEN1,
I915_READ(ILK_DISPLAY_CHICKEN1) |
ILK_FBCQ_DIS);
} else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
/* WaFbcAsynchFlipDisableFbcQueue:hsw,bdw */
I915_WRITE(CHICKEN_PIPESL_1(params->crtc.pipe),
I915_READ(CHICKEN_PIPESL_1(params->crtc.pipe)) |
HSW_FBCQ_DIS);
}
I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
I915_WRITE(SNB_DPFC_CTL_SA,
SNB_CPU_FENCE_ENABLE | params->fb.fence_reg);
I915_WRITE(DPFC_CPU_FENCE_OFFSET, params->crtc.fence_y_offset);
intel_fbc_recompress(dev_priv);
}
static bool intel_fbc_hw_is_active(struct drm_i915_private *dev_priv)
{
if (INTEL_INFO(dev_priv)->gen >= 5)
return ilk_fbc_is_active(dev_priv);
else if (IS_GM45(dev_priv))
return g4x_fbc_is_active(dev_priv);
else
return i8xx_fbc_is_active(dev_priv);
}
static void intel_fbc_hw_activate(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
fbc->active = true;
if (INTEL_INFO(dev_priv)->gen >= 7)
gen7_fbc_activate(dev_priv);
else if (INTEL_INFO(dev_priv)->gen >= 5)
ilk_fbc_activate(dev_priv);
else if (IS_GM45(dev_priv))
g4x_fbc_activate(dev_priv);
else
i8xx_fbc_activate(dev_priv);
}
static void intel_fbc_hw_deactivate(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
fbc->active = false;
if (INTEL_INFO(dev_priv)->gen >= 5)
ilk_fbc_deactivate(dev_priv);
else if (IS_GM45(dev_priv))
g4x_fbc_deactivate(dev_priv);
else
i8xx_fbc_deactivate(dev_priv);
}
/**
* intel_fbc_is_active - Is FBC active?
* @dev_priv: i915 device instance
*
* This function is used to verify the current state of FBC.
*
* FIXME: This should be tracked in the plane config eventually
* instead of queried at runtime for most callers.
*/
bool intel_fbc_is_active(struct drm_i915_private *dev_priv)
{
return dev_priv->fbc.active;
}
static void intel_fbc_work_fn(struct work_struct *__work)
{
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
struct drm_i915_private *dev_priv =
container_of(__work, struct drm_i915_private, fbc.work.work);
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_fbc_work *work = &fbc->work;
struct intel_crtc *crtc = fbc->crtc;
struct drm_vblank_crtc *vblank = &dev_priv->drm.vblank[crtc->pipe];
if (drm_crtc_vblank_get(&crtc->base)) {
DRM_ERROR("vblank not available for FBC on pipe %c\n",
pipe_name(crtc->pipe));
mutex_lock(&fbc->lock);
work->scheduled = false;
mutex_unlock(&fbc->lock);
return;
}
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
retry:
/* Delay the actual enabling to let pageflipping cease and the
* display to settle before starting the compression. Note that
* this delay also serves a second purpose: it allows for a
* vblank to pass after disabling the FBC before we attempt
* to modify the control registers.
*
* WaFbcWaitForVBlankBeforeEnable:ilk,snb
*
* It is also worth mentioning that since work->scheduled_vblank can be
* updated multiple times by the other threads, hitting the timeout is
* not an error condition. We'll just end up hitting the "goto retry"
* case below.
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
*/
wait_event_timeout(vblank->queue,
drm_crtc_vblank_count(&crtc->base) != work->scheduled_vblank,
msecs_to_jiffies(50));
mutex_lock(&fbc->lock);
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
/* Were we cancelled? */
if (!work->scheduled)
goto out;
/* Were we delayed again while this function was sleeping? */
if (drm_crtc_vblank_count(&crtc->base) == work->scheduled_vblank) {
mutex_unlock(&fbc->lock);
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
goto retry;
}
intel_fbc_hw_activate(dev_priv);
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
work->scheduled = false;
out:
mutex_unlock(&fbc->lock);
drm_crtc_vblank_put(&crtc->base);
}
static void intel_fbc_schedule_activation(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_fbc_work *work = &fbc->work;
WARN_ON(!mutex_is_locked(&fbc->lock));
if (drm_crtc_vblank_get(&crtc->base)) {
DRM_ERROR("vblank not available for FBC on pipe %c\n",
pipe_name(crtc->pipe));
return;
}
/* It is useless to call intel_fbc_cancel_work() or cancel_work() in
* this function since we're not releasing fbc.lock, so it won't have an
* opportunity to grab it to discover that it was cancelled. So we just
* update the expected jiffy count. */
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
work->scheduled = true;
work->scheduled_vblank = drm_crtc_vblank_count(&crtc->base);
drm_crtc_vblank_put(&crtc->base);
drm/i915: use a single intel_fbc_work struct This was already on my TODO list, and was requested both by Chris and Ville, for different reasons. The advantages are avoiding a frequent malloc/free pair, and the locality of having the work structure embedded in dev_priv. The maximum used memory is also smaller since previously we could have multiple allocated intel_fbc_work structs at the same time, and now we'll always have a single one - the one embedded on dev_priv. Of course, we're now using a little more memory on the cases where there's nothing scheduled. The biggest challenge here is to keep everything synchronized the way it was before. Currently, when we try to activate FBC, we allocate a new intel_fbc_work structure. Then later when we conclude we must delay the FBC activation a little more, we allocate a new intel_fbc_work struct, and then adjust dev_priv->fbc.fbc_work to point to the new struct. So when the old work runs - at intel_fbc_work_fn() - it will check that dev_priv->fbc.fbc_work points to something else, so it does nothing. Everything is also protected by fbc.lock. Just cancelling the old delayed work doesn't work because we might just cancel it after the work function already started to run, but while it is still waiting to grab fbc.lock. That's why we use the "dev_priv->fbc.fbc_work == work" check described in the paragraph above. So now that we have a single work struct we have to introduce a new way to synchronize everything. So we're making the work function a normal work instead of a delayed work, and it will be responsible for sleeping the appropriate amount of time itself. This way, after it wakes up it can grab the lock, ask "were we delayed or cancelled?" and then go back to sleep, enable FBC or give up. v2: - Spelling fixes. - Rebase after changing the patch order. - Fix ms/jiffies confusion. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v1) Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-27 02:27:49 +08:00
schedule_work(&work->work);
}
static void intel_fbc_deactivate(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
WARN_ON(!mutex_is_locked(&fbc->lock));
/* Calling cancel_work() here won't help due to the fact that the work
* function grabs fbc->lock. Just set scheduled to false so the work
* function can know it was cancelled. */
fbc->work.scheduled = false;
if (fbc->active)
intel_fbc_hw_deactivate(dev_priv);
}
static bool multiple_pipes_ok(struct intel_crtc *crtc,
struct intel_plane_state *plane_state)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
enum pipe pipe = crtc->pipe;
/* Don't even bother tracking anything we don't need. */
if (!no_fbc_on_multiple_pipes(dev_priv))
return true;
if (plane_state->visible)
fbc->visible_pipes_mask |= (1 << pipe);
else
fbc->visible_pipes_mask &= ~(1 << pipe);
return (fbc->visible_pipes_mask & ~(1 << pipe)) != 0;
}
static int find_compression_threshold(struct drm_i915_private *dev_priv,
struct drm_mm_node *node,
int size,
int fb_cpp)
{
struct i915_ggtt *ggtt = &dev_priv->ggtt;
int compression_threshold = 1;
int ret;
u64 end;
/* The FBC hardware for BDW/SKL doesn't have access to the stolen
* reserved range size, so it always assumes the maximum (8mb) is used.
* If we enable FBC using a CFB on that memory range we'll get FIFO
* underruns, even if that range is not reserved by the BIOS. */
if (IS_BROADWELL(dev_priv) ||
IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv))
end = ggtt->stolen_size - 8 * 1024 * 1024;
else
end = ggtt->stolen_usable_size;
/* HACK: This code depends on what we will do in *_enable_fbc. If that
* code changes, this code needs to change as well.
*
* The enable_fbc code will attempt to use one of our 2 compression
* thresholds, therefore, in that case, we only have 1 resort.
*/
/* Try to over-allocate to reduce reallocations and fragmentation. */
ret = i915_gem_stolen_insert_node_in_range(dev_priv, node, size <<= 1,
4096, 0, end);
if (ret == 0)
return compression_threshold;
again:
/* HW's ability to limit the CFB is 1:4 */
if (compression_threshold > 4 ||
(fb_cpp == 2 && compression_threshold == 2))
return 0;
ret = i915_gem_stolen_insert_node_in_range(dev_priv, node, size >>= 1,
4096, 0, end);
if (ret && INTEL_INFO(dev_priv)->gen <= 4) {
return 0;
} else if (ret) {
compression_threshold <<= 1;
goto again;
} else {
return compression_threshold;
}
}
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
static int intel_fbc_alloc_cfb(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
struct drm_mm_node *uninitialized_var(compressed_llb);
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
int size, fb_cpp, ret;
WARN_ON(drm_mm_node_allocated(&fbc->compressed_fb));
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
size = intel_fbc_calculate_cfb_size(dev_priv, &fbc->state_cache);
fb_cpp = drm_format_plane_cpp(fbc->state_cache.fb.pixel_format, 0);
ret = find_compression_threshold(dev_priv, &fbc->compressed_fb,
size, fb_cpp);
if (!ret)
goto err_llb;
else if (ret > 1) {
DRM_INFO("Reducing the compressed framebuffer size. This may lead to less power savings than a non-reduced-size. Try to increase stolen memory size if available in BIOS.\n");
}
fbc->threshold = ret;
if (INTEL_INFO(dev_priv)->gen >= 5)
I915_WRITE(ILK_DPFC_CB_BASE, fbc->compressed_fb.start);
else if (IS_GM45(dev_priv)) {
I915_WRITE(DPFC_CB_BASE, fbc->compressed_fb.start);
} else {
compressed_llb = kzalloc(sizeof(*compressed_llb), GFP_KERNEL);
if (!compressed_llb)
goto err_fb;
ret = i915_gem_stolen_insert_node(dev_priv, compressed_llb,
4096, 4096);
if (ret)
goto err_fb;
fbc->compressed_llb = compressed_llb;
I915_WRITE(FBC_CFB_BASE,
dev_priv->mm.stolen_base + fbc->compressed_fb.start);
I915_WRITE(FBC_LL_BASE,
dev_priv->mm.stolen_base + compressed_llb->start);
}
DRM_DEBUG_KMS("reserved %llu bytes of contiguous stolen space for FBC, threshold: %d\n",
fbc->compressed_fb.size, fbc->threshold);
return 0;
err_fb:
kfree(compressed_llb);
i915_gem_stolen_remove_node(dev_priv, &fbc->compressed_fb);
err_llb:
pr_info_once("drm: not enough stolen space for compressed buffer (need %d more bytes), disabling. Hint: you may be able to increase stolen memory size in the BIOS to avoid this.\n", size);
return -ENOSPC;
}
static void __intel_fbc_cleanup_cfb(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
if (drm_mm_node_allocated(&fbc->compressed_fb))
i915_gem_stolen_remove_node(dev_priv, &fbc->compressed_fb);
if (fbc->compressed_llb) {
i915_gem_stolen_remove_node(dev_priv, fbc->compressed_llb);
kfree(fbc->compressed_llb);
}
}
void intel_fbc_cleanup_cfb(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
__intel_fbc_cleanup_cfb(dev_priv);
mutex_unlock(&fbc->lock);
}
static bool stride_is_valid(struct drm_i915_private *dev_priv,
unsigned int stride)
{
/* These should have been caught earlier. */
WARN_ON(stride < 512);
WARN_ON((stride & (64 - 1)) != 0);
/* Below are the additional FBC restrictions. */
if (IS_GEN2(dev_priv) || IS_GEN3(dev_priv))
return stride == 4096 || stride == 8192;
if (IS_GEN4(dev_priv) && !IS_G4X(dev_priv) && stride < 2048)
return false;
if (stride > 16384)
return false;
return true;
}
static bool pixel_format_is_valid(struct drm_i915_private *dev_priv,
uint32_t pixel_format)
{
switch (pixel_format) {
case DRM_FORMAT_XRGB8888:
case DRM_FORMAT_XBGR8888:
return true;
case DRM_FORMAT_XRGB1555:
case DRM_FORMAT_RGB565:
/* 16bpp not supported on gen2 */
if (IS_GEN2(dev_priv))
return false;
/* WaFbcOnly1to1Ratio:ctg */
if (IS_G4X(dev_priv))
return false;
return true;
default:
return false;
}
}
/*
* For some reason, the hardware tracking starts looking at whatever we
* programmed as the display plane base address register. It does not look at
* the X and Y offset registers. That's why we look at the crtc->adjusted{x,y}
* variables instead of just looking at the pipe/plane size.
*/
static bool intel_fbc_hw_tracking_covers_screen(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
unsigned int effective_w, effective_h, max_w, max_h;
if (INTEL_INFO(dev_priv)->gen >= 8 || IS_HASWELL(dev_priv)) {
max_w = 4096;
max_h = 4096;
} else if (IS_G4X(dev_priv) || INTEL_INFO(dev_priv)->gen >= 5) {
max_w = 4096;
max_h = 2048;
} else {
max_w = 2048;
max_h = 1536;
}
intel_fbc_get_plane_source_size(&fbc->state_cache, &effective_w,
&effective_h);
effective_w += crtc->adjusted_x;
effective_h += crtc->adjusted_y;
return effective_w <= max_w && effective_h <= max_h;
}
static void intel_fbc_update_state_cache(struct intel_crtc *crtc,
struct intel_crtc_state *crtc_state,
struct intel_plane_state *plane_state)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_fbc_state_cache *cache = &fbc->state_cache;
struct drm_framebuffer *fb = plane_state->base.fb;
struct drm_i915_gem_object *obj;
cache->crtc.mode_flags = crtc_state->base.adjusted_mode.flags;
if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
cache->crtc.hsw_bdw_pixel_rate =
ilk_pipe_pixel_rate(crtc_state);
cache->plane.rotation = plane_state->base.rotation;
cache->plane.src_w = drm_rect_width(&plane_state->src) >> 16;
cache->plane.src_h = drm_rect_height(&plane_state->src) >> 16;
cache->plane.visible = plane_state->visible;
if (!cache->plane.visible)
return;
obj = intel_fb_obj(fb);
/* FIXME: We lack the proper locking here, so only run this on the
* platforms that need. */
if (IS_GEN(dev_priv, 5, 6))
cache->fb.ilk_ggtt_offset = i915_gem_obj_ggtt_offset(obj);
cache->fb.pixel_format = fb->pixel_format;
cache->fb.stride = fb->pitches[0];
cache->fb.fence_reg = obj->fence_reg;
cache->fb.tiling_mode = obj->tiling_mode;
}
static bool intel_fbc_can_activate(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_fbc_state_cache *cache = &fbc->state_cache;
if (!cache->plane.visible) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "primary plane not visible";
return false;
}
if ((cache->crtc.mode_flags & DRM_MODE_FLAG_INTERLACE) ||
(cache->crtc.mode_flags & DRM_MODE_FLAG_DBLSCAN)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "incompatible mode";
return false;
}
if (!intel_fbc_hw_tracking_covers_screen(crtc)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "mode too large for compression";
return false;
}
/* The use of a CPU fence is mandatory in order to detect writes
* by the CPU to the scanout and trigger updates to the FBC.
*/
if (cache->fb.tiling_mode != I915_TILING_X ||
cache->fb.fence_reg == I915_FENCE_REG_NONE) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "framebuffer not tiled or fenced";
return false;
}
if (INTEL_INFO(dev_priv)->gen <= 4 && !IS_G4X(dev_priv) &&
cache->plane.rotation != BIT(DRM_ROTATE_0)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "rotation unsupported";
return false;
}
if (!stride_is_valid(dev_priv, cache->fb.stride)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "framebuffer stride not supported";
return false;
}
if (!pixel_format_is_valid(dev_priv, cache->fb.pixel_format)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "pixel format is invalid";
return false;
}
/* WaFbcExceedCdClockThreshold:hsw,bdw */
if ((IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) &&
cache->crtc.hsw_bdw_pixel_rate >= dev_priv->cdclk_freq * 95 / 100) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "pixel rate is too big";
return false;
}
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
/* It is possible for the required CFB size change without a
* crtc->disable + crtc->enable since it is possible to change the
* stride without triggering a full modeset. Since we try to
* over-allocate the CFB, there's a chance we may keep FBC enabled even
* if this happens, but if we exceed the current CFB size we'll have to
* disable FBC. Notice that it would be possible to disable FBC, wait
* for a frame, free the stolen node, then try to reenable FBC in case
* we didn't get any invalidate/deactivate calls, but this would require
* a lot of tracking just for a specific case. If we conclude it's an
* important case, we can implement it later. */
if (intel_fbc_calculate_cfb_size(dev_priv, &fbc->state_cache) >
fbc->compressed_fb.size * fbc->threshold) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "CFB requirements changed";
return false;
}
return true;
}
static bool intel_fbc_can_choose(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
struct intel_fbc *fbc = &dev_priv->fbc;
if (intel_vgpu_active(dev_priv)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "VGPU is active";
return false;
}
if (!i915.enable_fbc) {
fbc->no_fbc_reason = "disabled per module param or by default";
return false;
}
if (fbc_on_pipe_a_only(dev_priv) && crtc->pipe != PIPE_A) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "no enabled pipes can have FBC";
return false;
}
if (fbc_on_plane_a_only(dev_priv) && crtc->plane != PLANE_A) {
fbc->no_fbc_reason = "no enabled planes can have FBC";
return false;
}
return true;
}
static void intel_fbc_get_reg_params(struct intel_crtc *crtc,
struct intel_fbc_reg_params *params)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_fbc_state_cache *cache = &fbc->state_cache;
/* Since all our fields are integer types, use memset here so the
* comparison function can rely on memcmp because the padding will be
* zero. */
memset(params, 0, sizeof(*params));
params->crtc.pipe = crtc->pipe;
params->crtc.plane = crtc->plane;
params->crtc.fence_y_offset = get_crtc_fence_y_offset(crtc);
params->fb.pixel_format = cache->fb.pixel_format;
params->fb.stride = cache->fb.stride;
params->fb.fence_reg = cache->fb.fence_reg;
params->cfb_size = intel_fbc_calculate_cfb_size(dev_priv, cache);
params->fb.ggtt_offset = cache->fb.ilk_ggtt_offset;
}
static bool intel_fbc_reg_params_equal(struct intel_fbc_reg_params *params1,
struct intel_fbc_reg_params *params2)
{
/* We can use this since intel_fbc_get_reg_params() does a memset. */
return memcmp(params1, params2, sizeof(*params1)) == 0;
}
void intel_fbc_pre_update(struct intel_crtc *crtc,
struct intel_crtc_state *crtc_state,
struct intel_plane_state *plane_state)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
if (!multiple_pipes_ok(crtc, plane_state)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "more than one pipe active";
goto deactivate;
}
if (!fbc->enabled || fbc->crtc != crtc)
goto unlock;
intel_fbc_update_state_cache(crtc, crtc_state, plane_state);
deactivate:
intel_fbc_deactivate(dev_priv);
unlock:
mutex_unlock(&fbc->lock);
}
static void __intel_fbc_post_update(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_fbc_reg_params old_params;
WARN_ON(!mutex_is_locked(&fbc->lock));
if (!fbc->enabled || fbc->crtc != crtc)
return;
if (!intel_fbc_can_activate(crtc)) {
WARN_ON(fbc->active);
return;
}
old_params = fbc->params;
intel_fbc_get_reg_params(crtc, &fbc->params);
/* If the scanout has not changed, don't modify the FBC settings.
* Note that we make the fundamental assumption that the fb->obj
* cannot be unpinned (and have its GTT offset and fence revoked)
* without first being decoupled from the scanout and FBC disabled.
*/
if (fbc->active &&
intel_fbc_reg_params_equal(&old_params, &fbc->params))
return;
intel_fbc_deactivate(dev_priv);
intel_fbc_schedule_activation(crtc);
fbc->no_fbc_reason = "FBC enabled (active or scheduled)";
}
void intel_fbc_post_update(struct intel_crtc *crtc)
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
__intel_fbc_post_update(crtc);
mutex_unlock(&fbc->lock);
}
static unsigned int intel_fbc_get_frontbuffer_bit(struct intel_fbc *fbc)
{
if (fbc->enabled)
return to_intel_plane(fbc->crtc->base.primary)->frontbuffer_bit;
else
return fbc->possible_framebuffer_bits;
}
void intel_fbc_invalidate(struct drm_i915_private *dev_priv,
unsigned int frontbuffer_bits,
enum fb_op_origin origin)
{
struct intel_fbc *fbc = &dev_priv->fbc;
if (!fbc_supported(dev_priv))
return;
if (origin == ORIGIN_GTT || origin == ORIGIN_FLIP)
return;
mutex_lock(&fbc->lock);
fbc->busy_bits |= intel_fbc_get_frontbuffer_bit(fbc) & frontbuffer_bits;
if (fbc->enabled && fbc->busy_bits)
intel_fbc_deactivate(dev_priv);
mutex_unlock(&fbc->lock);
}
void intel_fbc_flush(struct drm_i915_private *dev_priv,
unsigned int frontbuffer_bits, enum fb_op_origin origin)
{
struct intel_fbc *fbc = &dev_priv->fbc;
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
fbc->busy_bits &= ~frontbuffer_bits;
if (origin == ORIGIN_GTT || origin == ORIGIN_FLIP)
goto out;
if (!fbc->busy_bits && fbc->enabled &&
(frontbuffer_bits & intel_fbc_get_frontbuffer_bit(fbc))) {
if (fbc->active)
intel_fbc_recompress(dev_priv);
else
__intel_fbc_post_update(fbc->crtc);
}
out:
mutex_unlock(&fbc->lock);
}
/**
* intel_fbc_choose_crtc - select a CRTC to enable FBC on
* @dev_priv: i915 device instance
* @state: the atomic state structure
*
* This function looks at the proposed state for CRTCs and planes, then chooses
* which pipe is going to have FBC by setting intel_crtc_state->enable_fbc to
* true.
*
* Later, intel_fbc_enable is going to look for state->enable_fbc and then maybe
* enable FBC for the chosen CRTC. If it does, it will set dev_priv->fbc.crtc.
*/
void intel_fbc_choose_crtc(struct drm_i915_private *dev_priv,
struct drm_atomic_state *state)
{
struct intel_fbc *fbc = &dev_priv->fbc;
struct drm_crtc *crtc;
struct drm_crtc_state *crtc_state;
struct drm_plane *plane;
struct drm_plane_state *plane_state;
bool fbc_crtc_present = false;
int i, j;
mutex_lock(&fbc->lock);
for_each_crtc_in_state(state, crtc, crtc_state, i) {
if (fbc->crtc == to_intel_crtc(crtc)) {
fbc_crtc_present = true;
break;
}
}
/* This atomic commit doesn't involve the CRTC currently tied to FBC. */
if (!fbc_crtc_present && fbc->crtc != NULL)
goto out;
/* Simply choose the first CRTC that is compatible and has a visible
* plane. We could go for fancier schemes such as checking the plane
* size, but this would just affect the few platforms that don't tie FBC
* to pipe or plane A. */
for_each_plane_in_state(state, plane, plane_state, i) {
struct intel_plane_state *intel_plane_state =
to_intel_plane_state(plane_state);
if (!intel_plane_state->visible)
continue;
for_each_crtc_in_state(state, crtc, crtc_state, j) {
struct intel_crtc_state *intel_crtc_state =
to_intel_crtc_state(crtc_state);
if (plane_state->crtc != crtc)
continue;
if (!intel_fbc_can_choose(to_intel_crtc(crtc)))
break;
intel_crtc_state->enable_fbc = true;
goto out;
}
}
out:
mutex_unlock(&fbc->lock);
}
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
/**
* intel_fbc_enable: tries to enable FBC on the CRTC
* @crtc: the CRTC
*
* This function checks if the given CRTC was chosen for FBC, then enables it if
* possible. Notice that it doesn't activate FBC. It is valid to call
* intel_fbc_enable multiple times for the same pipe without an
* intel_fbc_disable in the middle, as long as it is deactivated.
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
*/
void intel_fbc_enable(struct intel_crtc *crtc,
struct intel_crtc_state *crtc_state,
struct intel_plane_state *plane_state)
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
if (fbc->enabled) {
WARN_ON(fbc->crtc == NULL);
if (fbc->crtc == crtc) {
WARN_ON(!crtc_state->enable_fbc);
WARN_ON(fbc->active);
}
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
goto out;
}
if (!crtc_state->enable_fbc)
goto out;
WARN_ON(fbc->active);
WARN_ON(fbc->crtc != NULL);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
intel_fbc_update_state_cache(crtc, crtc_state, plane_state);
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
if (intel_fbc_alloc_cfb(crtc)) {
drm/i915/fbc: don't print no_fbc_reason to dmesg Our dmesg messages started being misleading after we converted to the enable+activate model: we always print "Disabling FBC", even when we're just deactivating it. So, for example, when I boot my machine and do "dmesg | grep -i fbc", I see: [drm:intel_fbc_enable] Enabling FBC on pipe A [drm:set_no_fbc_reason] Disabling FBC: framebuffer not tiled or fenced but then, if I read the debugfs file, I will see: $ sudo cat i915_fbc_status FBC enabled Compressing: yes so we can conclude that dmesg is misleading, since FBC is actually enabled. What happened is that we deactivated FBC due to fbcon not being tiled, but when we silently reactivated it when the display manager started. We don't print activation messages since there may be way too many of these operations per second during normal desktop usage. One possible solution would be to change set_no_fbc_reason to correctly differentiate between disable and deactivation, but we removed support from printing activation/deactivation messages in the past because they were too frequent. So instead of doing this, let's just not print anything on dmesg, and leave the debugfs file if the user needs to investigate something. We already print when we enable and disable FBC anyway on a given pipe, so this should already help triaging bugs. Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1453210558-7875-22-git-send-email-paulo.r.zanoni@intel.com
2016-01-19 21:35:54 +08:00
fbc->no_fbc_reason = "not enough stolen memory";
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
goto out;
}
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
DRM_DEBUG_KMS("Enabling FBC on pipe %c\n", pipe_name(crtc->pipe));
fbc->no_fbc_reason = "FBC enabled but not active yet\n";
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
fbc->enabled = true;
fbc->crtc = crtc;
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
out:
mutex_unlock(&fbc->lock);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
}
/**
* __intel_fbc_disable - disable FBC
* @dev_priv: i915 device instance
*
* This is the low level function that actually disables FBC. Callers should
* grab the FBC lock.
*/
static void __intel_fbc_disable(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
struct intel_crtc *crtc = fbc->crtc;
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
WARN_ON(!mutex_is_locked(&fbc->lock));
WARN_ON(!fbc->enabled);
WARN_ON(fbc->active);
WARN_ON(crtc->active);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
DRM_DEBUG_KMS("Disabling FBC on pipe %c\n", pipe_name(crtc->pipe));
drm/i915: alloc/free the FBC CFB during enable/disable One of the problems with the current code is that it frees the CFB and releases its drm_mm node as soon as we flip FBC's enable bit. This is bad because after we disable FBC the hardware may still use the CFB for the rest of the frame, so in theory we should only release the drm_mm node one frame after we disable FBC. Otherwise, a stolen memory allocation done right after an FBC disable may result in either corrupted memory for the new owner of that memory region or corrupted screen/underruns in case the new owner changes it while the hardware is still reading it. This case is not exactly easy to reproduce since we currently don't do a lot of stolen memory allocations, but I see patches on the mailing list trying to expose stolen memory to user space, so races will be possible. I thought about three different approaches to solve this, and they all have downsides. The first approach would be to simply use multiple drm_mm nodes and freeing the unused ones only after a frame has passed. The problem with this approach is that since stolen memory is rather small, there's a risk we just won't be able to allocate a new CFB from stolen if the previous one was not freed yet. This could happen in case we quickly disable FBC from pipe A and decide to enable it on pipe B, or just if we change pipe A's fb stride while FBC is enabled. The second approach would be similar to the first one, but maintaining a single drm_mm node and keeping track of when it can be reused. This would remove the disadvantage of not having enough space for two nodes, but would create the new problem where we may not be able to enable FBC at the point intel_fbc_update() is called, so we would have to add more code to retry updating FBC after the time has passed. And that can quickly get too complex since we can get invalidate, flush, disable and other calls in the middle of the wait. Both solutions above - and also the current code - have the problem that we unnecessarily free+realloc FBC during invalidate+flush operations even if the CFB size doesn't change. The third option would be to move the allocation/deallocation to enable/disable. This makes sure that the pipe is always disabled when we allocate/deallocate the CFB, so there's no risk that the FBC hardware may read or write to the memory right after it is freed from drm_mm. The downside is that it is possible for user space to change the buffer stride without triggering a disable/enable - only deactivate/activate -, so we'll have to handle this case somehow - see igt's kms_frontbuffer_tracking test, fbc-stridechange subtest. It could be possible to implement a way to free+alloc the CFB during said stride change, but it would involve a lot of book-keeping - exactly as mentioned above - just for on case, so for now I'll keep it simple and just deactivate FBC. Besides, we may not even need to disable FBC since we do CFB over-allocation. Note from Chris: "Starting a fullscreen client that covers a single monitor in a multi-monitor setup will trigger a change in stride on one of the CRTCs (the monitors will be flipped independently).". It shouldn't be a huge problem if we lose FBC on multi-monitor setups since these setups already have problems reaching deep PC states anyway. v2: Rebase after changing the patch order. v3: - Remove references to the stride change case being "uncommon" and paste Chris' example. - Rebase after a change in a previous patch. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-16 01:19:21 +08:00
__intel_fbc_cleanup_cfb(dev_priv);
fbc->enabled = false;
fbc->crtc = NULL;
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
}
/**
* intel_fbc_disable - disable FBC if it's associated with crtc
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
* @crtc: the CRTC
*
* This function disables FBC if it's associated with the provided CRTC.
*/
void intel_fbc_disable(struct intel_crtc *crtc)
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
{
struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
struct intel_fbc *fbc = &dev_priv->fbc;
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
if (fbc->crtc == crtc) {
WARN_ON(!fbc->enabled);
WARN_ON(fbc->active);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
__intel_fbc_disable(dev_priv);
}
mutex_unlock(&fbc->lock);
cancel_work_sync(&fbc->work.work);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
}
/**
* intel_fbc_global_disable - globally disable FBC
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
* @dev_priv: i915 device instance
*
* This function disables FBC regardless of which CRTC is associated with it.
*/
void intel_fbc_global_disable(struct drm_i915_private *dev_priv)
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
{
struct intel_fbc *fbc = &dev_priv->fbc;
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
if (!fbc_supported(dev_priv))
return;
mutex_lock(&fbc->lock);
if (fbc->enabled)
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
__intel_fbc_disable(dev_priv);
mutex_unlock(&fbc->lock);
cancel_work_sync(&fbc->work.work);
drm/i915: introduce intel_fbc_{enable,disable} The goal is to call FBC enable/disable only once per modeset, while activate/deactivate/update will be called multiple times. The enable() function will be responsible for deciding if a CRTC will have FBC on it and then it will "lock" FBC on this CRTC: it won't be possible to change FBC's CRTC until disable(). With this, all checks and resource acquisition that only need to be done once per modeset can be moved from update() to enable(). And then the update(), activate() and deactivate() code will also get simpler since they won't need to worry about the CRTC being changed. The disable() function will do the reverse operation of enable(). One of its features is that it should only be called while the pipe is already off. This guarantees that FBC is stopped and nothing is using the CFB. With this, the activate() and deactivate() functions just start and temporarily stop FBC. They are the ones touching the hardware enable bit, so HW state reflects dev_priv->crtc.active. The last function remaining is update(). A lot of times I thought about renaming update() to activate() or try_to_activate() since it's called when we want to activate FBC. The thing is that update() may not only decide to activate FBC, but also deactivate or keep it on the same state, so I'll leave this name for now. Moving code to enable() and disable() will also help in case we decide to move FBC to pipe_config or something else later. The current patch only puts the very basic code on enable() and disable(). The next commits will take care of moving more stuff from update() to the new functions. v2: - Rebase. - Improve commit message (Chris). v3: Rebase after changing the patch order. v4: Rebase again after upstream changes. Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/
2015-10-15 21:44:46 +08:00
}
/**
* intel_fbc_init_pipe_state - initialize FBC's CRTC visibility tracking
* @dev_priv: i915 device instance
*
* The FBC code needs to track CRTC visibility since the older platforms can't
* have FBC enabled while multiple pipes are used. This function does the
* initial setup at driver load to make sure FBC is matching the real hardware.
*/
void intel_fbc_init_pipe_state(struct drm_i915_private *dev_priv)
{
struct intel_crtc *crtc;
/* Don't even bother tracking anything if we don't need. */
if (!no_fbc_on_multiple_pipes(dev_priv))
return;
for_each_intel_crtc(&dev_priv->drm, crtc)
if (intel_crtc_active(&crtc->base) &&
to_intel_plane_state(crtc->base.primary->state)->visible)
dev_priv->fbc.visible_pipes_mask |= (1 << crtc->pipe);
}
/*
* The DDX driver changes its behavior depending on the value it reads from
* i915.enable_fbc, so sanitize it by translating the default value into either
* 0 or 1 in order to allow it to know what's going on.
*
* Notice that this is done at driver initialization and we still allow user
* space to change the value during runtime without sanitizing it again. IGT
* relies on being able to change i915.enable_fbc at runtime.
*/
static int intel_sanitize_fbc_option(struct drm_i915_private *dev_priv)
{
if (i915.enable_fbc >= 0)
return !!i915.enable_fbc;
if (IS_BROADWELL(dev_priv))
return 1;
return 0;
}
/**
* intel_fbc_init - Initialize FBC
* @dev_priv: the i915 device
*
* This function might be called during PM init process.
*/
void intel_fbc_init(struct drm_i915_private *dev_priv)
{
struct intel_fbc *fbc = &dev_priv->fbc;
enum pipe pipe;
INIT_WORK(&fbc->work.work, intel_fbc_work_fn);
mutex_init(&fbc->lock);
fbc->enabled = false;
fbc->active = false;
fbc->work.scheduled = false;
i915.enable_fbc = intel_sanitize_fbc_option(dev_priv);
DRM_DEBUG_KMS("Sanitized enable_fbc value: %d\n", i915.enable_fbc);
if (!HAS_FBC(dev_priv)) {
fbc->no_fbc_reason = "unsupported by this chipset";
return;
}
for_each_pipe(dev_priv, pipe) {
fbc->possible_framebuffer_bits |=
INTEL_FRONTBUFFER_PRIMARY(pipe);
if (fbc_on_pipe_a_only(dev_priv))
break;
}
/* This value was pulled out of someone's hat */
if (INTEL_INFO(dev_priv)->gen <= 4 && !IS_GM45(dev_priv))
I915_WRITE(FBC_CONTROL, 500 << FBC_CTL_INTERVAL_SHIFT);
/* We still don't have any sort of hardware state readout for FBC, so
* deactivate it in case the BIOS activated it to make sure software
* matches the hardware state. */
if (intel_fbc_hw_is_active(dev_priv))
intel_fbc_hw_deactivate(dev_priv);
}