Currently we blow up in trace_dma_fence_init, when calling into
get_driver_name or get_timeline_name, since both the engine and context
might be NULL(or contain some garbage address) in the case of newly
allocated slab objects via the request ctor. Note that we also use
SLAB_TYPESAFE_BY_RCU here, which allows requests to be immediately
freed, but delay freeing the underlying page by an RCU grace period.
With this scheme requests can be re-allocated, at the same time as they
are also being read by some lockless RCU lookup mechanism.
In the ctor case, which is only called for new slab objects(i.e allocate
new page and call the ctor for each object) it's safe to reset the
context/engine prior to calling into dma_fence_init, since we can be
certain that no one is doing an RCU lookup which might depend on peeking
at the engine/context, like in active_engine(), since the object can't
yet be externally visible.
In the recycled case(which might also be externally visible) the request
refcount always transitions from 0->1 after we set the context/engine
etc, which should ensure it's valid to dereference the engine for
example, when doing an RCU list-walk, so long as we can also increment
the refcount first. If the refcount is already zero, then the request is
considered complete/released. If it's non-zero, then the request might
be in the process of being re-allocated, or potentially still in flight,
however after successfully incrementing the refcount, it's possible to
carefully inspect the request state, to determine if the request is
still what we were looking for. Note that all externally visible
requests returned to the cache must have zero refcount.
One possible fix then is to move dma_fence_init out from the request
ctor. Originally this was how it was done, but it was moved in:
commit 855e39e65c
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Mon Feb 3 09:41:48 2020 +0000
drm/i915: Initialise basic fence before acquiring seqno
where it looks like intel_timeline_get_seqno() relied on some of the
rq->fence state, but that is no longer the case since:
commit 12ca695d2c
Author: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Date: Tue Mar 23 16:49:50 2021 +0100
drm/i915: Do not share hwsp across contexts any more, v8.
intel_timeline_get_seqno() could also be cleaned up slightly by dropping
the request argument.
Moving dma_fence_init back out of the ctor, should ensure we have enough
of the request initialised in case of trace_dma_fence_init.
Functionally this should be the same, and is effectively what we were
already open coding before, except now we also assign the fence->lock
and fence->ops, but since these are invariant for recycled
requests(which might be externally visible), and will therefore already
hold the same value, it shouldn't matter.
An alternative fix, since we don't yet have a fully initialised request
when in the ctor, is just setting the context/engine as NULL, but this
does require adding some extra handling in get_driver_name etc.
v2(Daniel):
- Try to make the commit message less confusing
Fixes: 855e39e65c ("drm/i915: Initialise basic fence before acquiring seqno")
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Michael Mason <michael.w.mason@intel.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210921134202.3803151-1-matthew.auld@intel.com
(cherry picked from commit be988eaee1)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Nested grids in grid-table cells are not specified as proper ReST
constructs.
Commit 572f2a5cd9 ("drm/i915/guc: Update firmware to v62.0.0")
added a couple of kerneldoc tables of the form:
+---+-------+------------------------------------------------------+
| 1 | 31:0 | +------------------------------------------------+ |
+---+-------+ | | |
|...| | | Embedded `HXG Message`_ | |
+---+-------+ | | |
| n | 31:0 | +------------------------------------------------+ |
+---+-------+------------------------------------------------------+
For "make htmldocs", they happen to work as one might expect,
but they are incompatible with "make latexdocs" and "make pdfdocs",
and cause the generated gpu.tex file to become incomplete and
unbuildable by xelatex.
Restore the compatibility by removing those nested grids in the tables.
Size comparison of generated gpu.tex:
Sphinx 2.4.4 Sphinx 4.2.0
v5.14: 3238686 3841631
v5.15-rc1: 376270 432729
with this fix: 3377846 3998095
Fixes: 572f2a5cd9 ("drm/i915/guc: Update firmware to v62.0.0")
Cc: John Harrison <John.C.Harrison@Intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Matt Roper <matthew.d.roper@intel.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Akira Yokosawa <akiyks@gmail.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/4a227569-074f-c501-58bb-d0d8f60a8ae9@gmail.com
(cherry picked from commit 017792a041)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
We currently do an explicit flush of the buffer pools within the call path
of drm_driver.release(); this removes all buffers, regardless of their age,
freeing the buffers' associated resources (objects, address space areas).
However there is other code that runs within the drm_driver.release() call
chain that expects objects and their associated address space areas have
already been flushed.
Since buffer pools auto-flush old buffers once per second in a worker
thread, there's a small window where if we remove the driver while there
are still objects in buffers with an age of less than one second, the
assumptions of the other release code may be violated.
By moving the flush to driver remove (which executes earlier via the
pci_driver.remove() flow) we're ensuring that all buffers are flushed and
their associated objects freed before some other code in
pci_driver.remove() flushes those objects so they are released before
_any_ code in drm_driver.release() that check completness of those
flushes executes.
v2: Reword commit description as suggested by Matt.
Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matt Roper <matthew.d.roper@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210924163825.634606-1-janusz.krzysztofik@linux.intel.com
On FBC1 we can specify an arbitrary cfb stride. The hw will
simply throw away any compressed line that would exceed the
specified limit and keep using the uncompressed data instead.
Thus we can allow arbitrary compression limits.
The one thing we have to keep in mind though is that the cfb
stride is specified in units of 32B (gen2) or 64B (gen3+).
Fortunately X-tile is already 128B (gen2) or 512B (gen3+) wide
so as long as we limit outselves to the same 4x compression
limit that FBC2 has we are guaranteed to have a sufficiently
aligned cfb stride.
Reviewed-by: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210921152517.803-5-ville.syrjala@linux.intel.com
Apply the same 512 byte FBC segment alignment to glk+ as we use
on skl+. The only real difference is that we now have a dedicated
register for the FBC override stride. Not 100% sure which
platforms really need the 512B alignment, but it's easiest
to just do it on everything.
Also the hardware no longer seems to misclaculate the CFB stride
for linear, so we can omit the use of the override stride for
linear unless the stride is misaligned.
Reviewed-by: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210921152517.803-3-ville.syrjala@linux.intel.com
The code to calculate the cfb stride/size is a bit of mess.
The cfb size is getting calculated based purely on the plane
stride and plane height. That doesn't account for extra
alignment we want for the cfb stride. The gen9 override
stride OTOH is just calculated based on the plane width, and
it does try to make things more aligned but any extra alignment
added there is not considered in the cfb size calculations.
So not at all convinced this is working as intended. Additionally
the compression limit handling is split between the cfb allocation
code and g4x_dpfc_ctl_limit() (for the 16bpp case), which is just
confusing.
Let's streamline the whole thing:
- Start with the plane stride, convert that into cfb stride (cfb is
always 4 bytes per pixel). All the calculations will assume 1:1
compression limit since that will give us the max values, and we
don't yet know how much stolen memory we will be able to allocate
- Align the cfb stride to 512 bytes on modern platforms. This guarantees
the 4 line segment will be 512 byte aligned regardles of the final
compression limit we choose later. The 512 byte alignment for the
segment is required by at least some of the platforms, and just doing
it always seems like the easiest option
- Figure out if we need to use the override stride or not. For X-tiled
it's never needed since the plane stride is already 512 byte aligned,
for Y-tiled it will be needed if the plane stride is not a multiple
of 512 bytes, and for linear it's apparently always needed because the
hardware miscalculates the cfb stride as PLANE_STRIDE*512 instead of
the PLANE_STRIDE*64 that it use with linear.
- The cfb size will be calculated based on the aligned cfb stride to
guarantee we actually reserved enough stolen memory and the FBC hw
won't end up scribbling over whatever else is allocated in stolen
- The compression limit handling we just do fully in the cfb allocation
code to make things less confusing
v2: Write the min cfb segment stride calculation in a more
explicit way to make it clear what is going on
v3: Remeber to update fbc->limit when changing to 16bpp
Reviewed-by: Uma Shankar <uma.shankar@intel.com> #v2
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210923042151.19052-1-ville.syrjala@linux.intel.com
Currently we blow up in trace_dma_fence_init, when calling into
get_driver_name or get_timeline_name, since both the engine and context
might be NULL(or contain some garbage address) in the case of newly
allocated slab objects via the request ctor. Note that we also use
SLAB_TYPESAFE_BY_RCU here, which allows requests to be immediately
freed, but delay freeing the underlying page by an RCU grace period.
With this scheme requests can be re-allocated, at the same time as they
are also being read by some lockless RCU lookup mechanism.
In the ctor case, which is only called for new slab objects(i.e allocate
new page and call the ctor for each object) it's safe to reset the
context/engine prior to calling into dma_fence_init, since we can be
certain that no one is doing an RCU lookup which might depend on peeking
at the engine/context, like in active_engine(), since the object can't
yet be externally visible.
In the recycled case(which might also be externally visible) the request
refcount always transitions from 0->1 after we set the context/engine
etc, which should ensure it's valid to dereference the engine for
example, when doing an RCU list-walk, so long as we can also increment
the refcount first. If the refcount is already zero, then the request is
considered complete/released. If it's non-zero, then the request might
be in the process of being re-allocated, or potentially still in flight,
however after successfully incrementing the refcount, it's possible to
carefully inspect the request state, to determine if the request is
still what we were looking for. Note that all externally visible
requests returned to the cache must have zero refcount.
One possible fix then is to move dma_fence_init out from the request
ctor. Originally this was how it was done, but it was moved in:
commit 855e39e65c
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Mon Feb 3 09:41:48 2020 +0000
drm/i915: Initialise basic fence before acquiring seqno
where it looks like intel_timeline_get_seqno() relied on some of the
rq->fence state, but that is no longer the case since:
commit 12ca695d2c
Author: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Date: Tue Mar 23 16:49:50 2021 +0100
drm/i915: Do not share hwsp across contexts any more, v8.
intel_timeline_get_seqno() could also be cleaned up slightly by dropping
the request argument.
Moving dma_fence_init back out of the ctor, should ensure we have enough
of the request initialised in case of trace_dma_fence_init.
Functionally this should be the same, and is effectively what we were
already open coding before, except now we also assign the fence->lock
and fence->ops, but since these are invariant for recycled
requests(which might be externally visible), and will therefore already
hold the same value, it shouldn't matter.
An alternative fix, since we don't yet have a fully initialised request
when in the ctor, is just setting the context/engine as NULL, but this
does require adding some extra handling in get_driver_name etc.
v2(Daniel):
- Try to make the commit message less confusing
Fixes: 855e39e65c ("drm/i915: Initialise basic fence before acquiring seqno")
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Michael Mason <michael.w.mason@intel.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210921134202.3803151-1-matthew.auld@intel.com
We really only need memcpy restore for objects that affect the
operability of the migrate context. That is, primarily the page-table
objects of the migrate VM.
Add an object flag, I915_BO_ALLOC_PM_EARLY for objects that need early
restores using memcpy and a way to assign LMEM page-table object flags
to be used by the vms.
Restore objects without this flag with the gpu blitter and only objects
carrying the flag using TTM memcpy.
Initially mark the migrate, gt, gtt and vgpu vms to use this flag, and
defer for a later audit which vms actually need it. Most importantly, user-
allocated vms with pinned page-table objects can be restored using the
blitter.
Performance-wise memcpy restore is probably as fast as gpu restore if not
faster, but using gpu restore will help tackling future restrictions in
mappable LMEM size.
v4:
- Don't mark the aliasing ppgtt page table flags for early resume, but
rather the ggtt page table flags as intended. (Matthew Auld)
- The check for user buffer objects during early resume is pointless, since
they are never marked I915_BO_ALLOC_PM_EARLY. (Matthew Auld)
v5:
- Mark GuC LMEM objects with I915_BO_ALLOC_PM_EARLY to have them restored
before we fire up the migrate context.
Cc: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210922062527.865433-8-thomas.hellstrom@linux.intel.com
Pinned contexts, like the migrate contexts need reset after resume
since their context image may have been lost. Also the GuC needs to
register pinned contexts.
Add a list to struct intel_engine_cs where we add all pinned contexts on
creation, and traverse that list at resume time to reset the pinned
contexts.
This fixes the kms_pipe_crc_basic@suspend-read-crc-pipe-a selftest for now,
but proper LMEM backup / restore is needed for full suspend functionality.
However, note that even with full LMEM backup / restore it may be
desirable to keep the reset since backing up the migrate context images
must happen using memcpy() after the migrate context has become inactive,
and for performance- and other reasons we want to avoid memcpy() from
LMEM.
Also traverse the list at guc_init_lrc_mapping() calling
guc_kernel_context_pin() for the pinned contexts, like is already done
for the kernel context.
v2:
- Don't reset the contexts on each __engine_unpark() but rather at
resume time (Chris Wilson).
v3:
- Reset contexts in the engine sanitize callback. (Chris Wilson)
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Brost Matthew <matthew.brost@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210922062527.865433-6-thomas.hellstrom@linux.intel.com
Just evict unpinned objects to system. For pinned LMEM objects,
make a backup system object and blit the contents to that.
Backup is performed in three steps,
1: Opportunistically evict evictable objects using the gpu blitter.
2: After gt idle, evict evictable objects using the gpu blitter. This will
be modified in an upcoming patch to backup pinned objects that are not used
by the blitter itself.
3: Backup remaining pinned objects using memcpy.
Also move uC suspend to after 2) to make sure we have a functional GuC
during 2) if using GuC submission.
v2:
- Major refactor to make sure gem_exec_suspend@hang-SX subtests work, and
suspend / resume works with a slightly modified GuC submission enabling
patch series.
v3:
- Fix a potential use-after-free (Matthew Auld)
- Use i915_gem_object_create_shmem() instead of
i915_gem_object_create_region (Matthew Auld)
- Minor simplifications (Matthew Auld)
- Fix up kerneldoc for i195_ttm_restore_region().
- Final lmem_suspend() call moved to i915_gem_backup_suspend from
i915_gem_suspend_late, since the latter gets called at driver unload
and we don't unnecessarily want to run it at that time.
v4:
- Interface change of ttm- & lmem suspend / resume functions to use
flags rather than bools. (Matthew Auld)
- Completely drop the i915_gem_backup_suspend change (Matthew Auld)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210922062527.865433-5-thomas.hellstrom@linux.intel.com
An upcoming common pattern is to traverse the region object list and
perform certain actions on all objects in a region. It's a little tricky
to get the list locking right, in particular since a gem object may
change region unless it's pinned or the object lock is held.
Define a function that does this for us and that takes an argument that
defines the action to be performed on each object.
v3:
- Improve structure documentation a bit (Matthew Auld)
Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210922062527.865433-3-thomas.hellstrom@linux.intel.com
PSR always had a requirement to only be enabled if there is active
planes but not following that never caused any issues.
But that changes in Alderlake-P, leaving PSR enabled without
active planes causes transcoder/port underruns.
Similar behavior was fixed during the pipe disable sequence by
commit 84030adb9e ("drm/i915/display: Disable audio, DRRS and PSR before planes").
intel_dp_compute_psr_vsc_sdp() had to move from
intel_psr_enable_locked() to intel_psr_compute_config() because we
need to be able to disable/enable PSR from atomic states without
connector and encoder state.
Reviewed-by: Gwan-gyeong Mun <gwan-gyeong.mun@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Gwan-gyeong Mun <gwan-gyeong.mun@intel.com>
Signed-off-by: José Roberto de Souza <jose.souza@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210922215242.66683-3-jose.souza@intel.com
Nested grids in grid-table cells are not specified as proper ReST
constructs.
Commit 572f2a5cd9 ("drm/i915/guc: Update firmware to v62.0.0")
added a couple of kerneldoc tables of the form:
+---+-------+------------------------------------------------------+
| 1 | 31:0 | +------------------------------------------------+ |
+---+-------+ | | |
|...| | | Embedded `HXG Message`_ | |
+---+-------+ | | |
| n | 31:0 | +------------------------------------------------+ |
+---+-------+------------------------------------------------------+
For "make htmldocs", they happen to work as one might expect,
but they are incompatible with "make latexdocs" and "make pdfdocs",
and cause the generated gpu.tex file to become incomplete and
unbuildable by xelatex.
Restore the compatibility by removing those nested grids in the tables.
Size comparison of generated gpu.tex:
Sphinx 2.4.4 Sphinx 4.2.0
v5.14: 3238686 3841631
v5.15-rc1: 376270 432729
with this fix: 3377846 3998095
Fixes: 572f2a5cd9 ("drm/i915/guc: Update firmware to v62.0.0")
Cc: John Harrison <John.C.Harrison@Intel.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Matt Roper <matthew.d.roper@intel.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Akira Yokosawa <akiyks@gmail.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/4a227569-074f-c501-58bb-d0d8f60a8ae9@gmail.com
Add support for remapping CCS FBs on ADL-P to remove the restriction
of the power-of-two sized stride and the 2MB surface offset alignment
for these FBs.
We can only remap the tiles on the main surface, not the tiles on the
CCS surface, so userspace has to generate the CCS surface aligning to
the POT size padded main surface stride (by programming the AUX
pagetable accordingly). For the required AUX pagetable setup, this
requires that either the main surface stride is 8 tiles or that the
stride is 16 tiles aligned (= 64 kbytes, the area mapped by one AUX
PTE).
v2:
- Init intel_remapped_info::plane_alignment only for remapped views and
do this from intel_fb_view_init().
Cc: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Juha-Pekka Heikkila <juhapekka.heikkila@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20210906182715.3915100-6-imre.deak@intel.com
