- 20 Dec, 2019 5 commits
-
-
Ville Syrjälä authored
We have several places where we want to allocate a pristine crtc state. Some of those currently call intel_crtc_state_reset() to properly initialize all the non-zero defaults in the state, but some places do not. Let's add intel_crtc_state_alloc() to do both the alloc and the reset, and call that everywhere we need a fresh crtc state. v2: s/kzalloc/kmalloc/ since we memset() anyway (José) Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Signed-off-by:
Ville Syrjälä <ville.syrjala@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191219111430.17527-1-ville.syrjala@linux.intel.comReviewed-by:
José Roberto de Souza <jose.souza@intel.com>
-
Chris Wilson authored
Decide whether or not we need to disable arbitration within user batches based on our intel_engine_has_preemption() flag. Signed-off-by:
Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Reviewed-by:
Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191213151331.1788371-1-chris@chris-wilson.co.uk
-
Chris Wilson authored
Instead of rummaging through the intel_context to peek at the GEM context in the middle of request submission to decide whether to use semaphores, store that information on the intel_context itself. Signed-off-by:
Andi Shyti <andi.shyti@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191220101230.256839-2-chris@chris-wilson.co.uk
-
Chris Wilson authored
Keep the intel_context as being the primary state for i915_request, with the GEM context a backpointer from the low level state for the rarer cases we need client information. Our goal is to remove such references to clients from the backend, and leave the HW submission agnostic to client interfaces and self-contained. Signed-off-by:
-
Chris Wilson authored
Since we added the context_alloc callback to intel_context_ops, we can safely install a custom hook for the deferred virtual context allocation. This means that all new contexts behave the same upon creation, simplifying later code. Signed-off-by:
-
- 19 Dec, 2019 16 commits
-
-
Chris Wilson authored
Avoid adding the retire workers to the virtual engine so that we don't end up in the unenviable situation of trying to free the virtual engine while its worker remains active. Fixes: dc93c9b6 ("drm/i915/gt: Schedule request retirement when signaler idles") Closes: https://gitlab.freedesktop.org/drm/intel/issues/867Signed-off-by:
Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191219221344.161523-1-chris@chris-wilson.co.uk
-
Ville Syrjälä authored
All the other display related tracepoints use intel_ instead if i915_ as the prefix. Do the same for the pipe update tracepoints so I don't always have to spend time looking for them. Signed-off-by:
-
Ville Syrjälä authored
I fumbled the conflict resolution a bit when applying the fbc vblank wait w/a. Because of that we now call intel_fbc_pre_update() twice. Remove the second redundant call. Reported-by:
-
Ville Syrjälä authored
icl and tgl are still affected by the modulo 4 PLANE_OFFSET.y underrun issue. Reject such configurations on all gen9+ platforms. Can be reproduced easily with the following sequence of hardware poking: while { write FBC_CTL.enable=1 wait for vblank write PLANE_OFFSET .x=0 .y=32 write PLANE_SURF wait for vblank # if PLANE_OFFSET.y is multiple of 4 the underrun won't happen write PLANE_OFFSET .x=0 .y=31 write PLANE_SURF wait for vblank # extra vblank wait is required here presumably # to get FBC into the proper state wait for vblank write FBC_CTL.enable=0 # underrun happens some time after FBC disable wait for vblank } Both 8888 and 565 pixel formats and all tilinga formats seem affected. Reproduced on KBL/GLK/ICL/TGL. BDW confirmed not affected. Closes: https://gitlab.freedesktop.org/drm/intel/issues/792Signed-off-by:Imre Deak <imre.deak@intel.com>
-
Colin Ian King authored
Currently pointers to and from are not initialized and may contain garbage values. This will cause uninitialized pointer reads in the call to intel_frontbuffer_track and later checks to see if to and from are null. Fix this by ensuring to and from are initialized to NULL. Addresses-Coverity: ("Uninitialised pointer read)" Fixes: da42104f ("drm/i915: Hold reference to intel_frontbuffer as we track activity") Signed-off-by:Colin Ian King <colin.king@canonical.com> Reviewed-by:
-
Chris Wilson authored
When we park RPS, we set the GPU to run at minimum 'idle' frequency. However, as the GPU is idle, we also disable the worker and RPS interrupts - changing the RPS thresholds has no effect, it just incurs extra changes to restore them when we unpark. So on parking, leave the thresholds set to the current power level and so we expect them to be valid for our restart. References: https://gitlab.freedesktop.org/drm/intel/issues/848Signed-off-by:
-
Chris Wilson authored
Use non-forcewaked writes to queue RPS register changes that will take effect when the write buffer is flushed, rather than wake the mmio device for immediate effect. This is so that we can avoid a slow forcewake dance upon unparking, and at our irregular updates. Signed-off-by:
-
Chris Wilson authored
Knowing the round trip time of an engine is useful for tracking the health of the system as well as providing a metric for the baseline responsiveness of the engine. We can use the latter metric for automatically tuning our waits in selftests and when idling so we don't confuse a slower system with a dead one. Upon idling the engine, we send one last pulse to switch the context away from precious user state to the volatile kernel context. We know the engine is idle at this point, and the pulse is non-preemptible, so this provides us with a good measurement of the round trip time. It also provides us with faster engine parking for ringbuffer submission, which is a welcome bonus (e.g. softer-rc6). Signed-off-by:
-
Chris Wilson authored
Very similar to commit 4f88f874 ("drm/i915/gt: Schedule request retirement when timeline idles"), but this time instead of coupling into the execlists CS event interrupt, we couple into the breadcrumb interrupt and queue a timeline's retirement when the last signaler is completed. This should allow us to more rapidly park ringbuffer submission, and so help reduce power consumption on older systems. v2: Fixup intel_engine_add_retire() to handle concurrent callers References: 4f88f874 ("drm/i915/gt: Schedule request retirement when timeline idles") Signed-off-by:
-
Jani Nikula authored
Fix several issues with DSC power domains that did not take DSI transcoders into account: - On TGL+ we need to use PW2 for DSC on pipe A, not transcoder A. There is no longer an eDP transcoder, but there are two DSI transcoders which may be connected to pipe A. - On TGL+ we need to use the pipe, not transcoder, power domains for DSC on pipes other than A. Again, there are DSI transcoders. - On ICL we need to use PW2 for DSC also for DSI transcoders, not just for the eDP transcoder. Using is_pipe_dsc() also adds the warning about ICL pipe A DSC, which does not exist. Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Cc: Vandita Kulkarni <vandita.kulkarni@intel.com> Reviewed-by:
Vandita Kulkarni <vandita.kulkarni@intel.com> Signed-off-by:
Jani Nikula <jani.nikula@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191212134728.18432-1-jani.nikula@intel.com
-
Jani Nikula authored
The check for cpu_transcoder != TRANSCODER_A is more magic than necessary, and potentially misleading. Before TGL, DSC is supported on pipe A if, and only if, it's used with eDP or DSI transcoders. No functional changes. Cc: Manasi Navare <manasi.d.navare@intel.com> Cc: Vandita Kulkarni <vandita.kulkarni@intel.com> Reviewed-by:
-
Jani Nikula authored
ICL eDP and DSI transcoders have a DSC engine separate from the pipe. Abstract the register selection and fix it for ICL. Add a warning for pipe A DSC on ICL; it does not exist. Cc: Manasi Navare <manasi.d.navare@intel.com> Cc: Vandita Kulkarni <vandita.kulkarni@intel.com> Reviewed-by:
-
Andrzej Pietrasiewicz authored
Use the ddc pointer provided by the generic connector. Signed-off-by:
Andrzej Pietrasiewicz <andrzej.p@collabora.com> Acked-by:
Sam Ravnborg <sam@ravnborg.org> Reviewed-by:
Emil Velikov <emil.velikov@collabora.com> Signed-off-by:
-
Lucas De Marchi authored
Pass the correct variable as argument. Signed-off-by:
Lucas De Marchi <lucas.demarchi@intel.com> Reviewed-by:
Clinton Taylor <Clinton.A.Taylor@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191217230529.25092-3-lucas.demarchi@intel.com
-
Lucas De Marchi authored
Instead of "ungated" use the same name for the variable as the bitfield, making it clearer what's the intent of the checks. Signed-off-by:
-
Lucas De Marchi authored
This allows us to isolate reading and writing to the ICL_DPCLKA_CFGCR0 during the sanitize phase. Signed-off-by:
-
- 18 Dec, 2019 17 commits
-
-
Chris Wilson authored
When doing our global park, we like to be a good citizen and shrink our slab caches (of which we have quite a few now), but each kmem_cache_shrink() incurs a stop_machine() and so ends up being quite expensive, causing machine-wide stalls. While ideally we would like to throw away unused pages in our slab caches whenever it appears that we are idling, doing so will require a much cheaper mechanism. In the meantime use a delayed worked to impose a rate-limit that means we have to have been idle for more than 2 seconds before we start shrinking. References: https://gitlab.freedesktop.org/drm/intel/issues/848Signed-off-by:
-
Chris Wilson authored
Only signal the breadcrumbs from inside the irq_work, simplifying our interface and calling conventions. The micro-optimisation here is that by always using the irq_work interface, we know we are always inside an irq-off critical section for the breadcrumb signaling and can ellide save/restore of the irq flags. Signed-off-by:
-
Tvrtko Ursulin authored
Avoid rc6 counter going backward in close to 0% RC6 scenarios like: 15.005477996 114,246,613 ns i915/rc6-residency/ 16.005876662 667,657 ns i915/rc6-residency/ 17.006131417 7,286 ns i915/rc6-residency/ 18.006615031 18,446,744,073,708,914,688 ns i915/rc6-residency/ 19.007158361 18,446,744,073,709,447,168 ns i915/rc6-residency/ 20.007806498 0 ns i915/rc6-residency/ 21.008227495 1,440,403 ns i915/rc6-residency/ There are two aspects to this fix. First is not assuming rc6 value zero means GT is asleep since that can also mean GPU is fully busy and we do not want to enter the estimation path in that case. Second is ensuring monotonicity on the estimation path itself. I suspect what is happening is with extremely rapid park/unpark cycles we get no updates on the real rc6 and therefore have to careful not to unconditionally trust use last known real rc6 when creating a new estimation. v2: * Simplify logic by not tracking the estimate but last reported value. Signed-off-by: -
Ville Syrjälä authored
Move all of haswell_crtc_disable() into the encoder .post_disable() hooks. Now we're left with just calling the .disable() and .post_disable() hooks back to back. I chose to move the code into the .post_disable() hook instead of the .disable() hook as most of the sequence is currently implemented in the .post_disable() hook. We should collapse it all down to just one hook and then the encoders can drive the modeset sequence fully. But that may need some further refactoring as we currently call the ddi .post_disable() hook from mst code and we can't just replace that with a call to the ddi .disable() hook. Should also follow up with similar treatment for the enable sequence but let's start here where it's easier. Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Signed-off-by:
-
Ville Syrjälä authored
To make life easier in the future let's pass the old crtc state to intel_crtc_vblank_off() just like we already do for its counterpart intel_crtc_vblank_on(). Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Signed-off-by:
-
Ville Syrjälä authored
To make life easier in the future let's pass the old crtc state to skylake_scaler_disable() just like we already do for for its ancestor ironlake_pfit_disable(). Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Signed-off-by:
-
Ville Syrjälä authored
HSW+ platforms call encoder .post_disable() and .post_pll_disable() back to back. And since we don't even disable the PLL in between let's just move everything into .post_disable(). intel_dp_mst does forward the .post_disable() call to intel_ddi at the very end of its own .post_disable() hook, so this time MST I shouldn't even break MST by accident. Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Signed-off-by:
-
Ville Syrjälä authored
Remove the pointless vfunc detour for hsw_fdi_link_train() and just call it directly. Also pass the encoder in so we can nuke the silly encoder loop within. Cc: José Roberto de Souza <jose.souza@intel.com> Cc: Manasi Navare <manasi.d.navare@intel.com> Signed-off-by:
-
Ville Syrjälä authored
For the sake of symmetry with the crtc stuff let's add a helper to reset the plane state to sane default values. For the moment this only gets caller from the plane init. Signed-off-by:
-
Ville Syrjälä authored
We have a few places where we want to reset a crtc state to its default values. Let's add a helper for that. We'll need the new __drm_atomic_helper_crtc_state_reset() helper for this to allow us to just reset the state itself without clobbering the crtc->state pointer. And while at it let's zero out the whole thing, except a few choice member which we'll mark as "invalid". And thanks to this we can now nuke intel_crtc_init_scalers(). Signed-off-by:
-
Ville Syrjälä authored
We already have alloc/free helpers for planes, add the same for crtcs. The main benefit is we get to move all the annoying state initialization out of the main crtc_init() flow. Signed-off-by:
-
Ville Syrjälä authored
Let's get rid of the redundant intel_ prefix on our variables. Signed-off-by:
-
Ville Syrjälä authored
Annoyingly __drm_atomic_helper_crtc_reset() does two totally separate things: a) reset the state to defaults values b) assign the crtc->state pointer I just want a) without the b) so let's split out part a) into __drm_atomic_helper_crtc_state_reset(). And of course we'll do the same thing for planes and connectors. v2: Fix conn__state vs. conn_state typo (Lucas) Make code and kerneldoc match for __drm_atomic_helper_plane_state_reset() Signed-off-by:Daniel Vetter <daniel.vetter@ffwll.ch> Acked-by:
-
Chris Wilson authored
For very light workloads that frequently park, acquiring the display power well (required to prevent the dmc from trashing the system) takes longer than the execution. A good example is the igt_coherency selftest, which is slowed down by an order of magnitude in the worst case with powerwell cycling. To prevent frequent cycling, while keeping our fast soft-rc6, use a timer to delay release of the display powerwell. Fixes: 31177017 ("drm/i915/gt: Schedule request retirement when timeline idles") References: https://gitlab.freedesktop.org/drm/intel/issues/848Signed-off-by:
-
Chris Wilson authored
Since obj->frontbuffer is no longer protected by the struct_mutex, as we are processing the execbuf, it may be removed. Mark the intel_frontbuffer as rcu protected, and so acquire a reference to the struct as we track activity upon it. Closes: https://gitlab.freedesktop.org/drm/intel/issues/827 Fixes: 8e7cb179 ("drm/i915: Extract intel_frontbuffer active tracking") Signed-off-by:
-
Chris Wilson authored
If the whole GT is asleep, we know that each engine must also be asleep and so we can quickly return without checking them all. Signed-off-by:
-
Chris Wilson authored
If we inherit an error along the fence chain, we skip the main work callback and go straight to the error. In the case of the vma bind worker, we only dropped the pinned pages from the worker. In the process, make sure we call the release earlier rather than wait until the final reference to the fence is dropped (as a reference is kept while being listened upon). Signed-off-by:
-
- 17 Dec, 2019 2 commits
-
-
Daniele Ceraolo Spurio authored
The Gen11+ and the legacy function differ in the register and value written to interrupt the GuC. However, while on older gen the value matches a bit on the register, on Gen11+ the value is a SW defined payload that is sent to the FW. Since the FW behaves the same no matter what value we pass to it, we can just write the same thing on all gens and get rid of the function pointer by saving the register offset. Signed-off-by:
Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com> Cc: Michal Wajdeczko <michal.wajdeczko@intel.com> Cc: John Harrison <John.C.Harrison@Intel.com> Cc: Matthew Brost <matthew.brost@intel.com> Reviewed-by:
Michal Wajdeczko <michal.wajdeczko@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20191217012316.13271-6-daniele.ceraolospurio@intel.com
-
Daniele Ceraolo Spurio authored
Since we started using CT buffers on all gens, the function pointers can only be set to either the _nop() or the _ct() functions. Since the _nop() case applies to when the CT are disabled, we can just handle that case in the _ct() functions and call them directly. v2: keep intel_guc_send() and make the CT send/receive functions work on intel_guc_ct. (Michal) Signed-off-by:
-
