- 04 Aug, 2016 20 commits
-
-
Chris Wilson authored
If the user floods the GPU with so many requests that the engine stalls waiting for free space, don't automatically promote the GPU to maximum frequencies. If the GPU really is saturated with work, it will migrate to high clocks by itself, otherwise it is merely a user flooding us with busy-work. Signed-off-by:
Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by:
Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1470293567-10811-20-git-send-email-chris@chris-wilson.co.uk
-
Chris Wilson authored
Signed-off-by:
-
Chris Wilson authored
intel_overlay already tracks its last flip request, along with action to take after its completion. Refactor intel_overlay to reuse the common i915_gem_active tracker. v2: Now using i915_gem_retire_fn typedef References: https://bugs.freedesktop.org/show_bug.cgi?id=93730 References: https://bugs.freedesktop.org/show_bug.cgi?id=96851Signed-off-by:
-
Chris Wilson authored
By tracking each request occupying space inside an individual intel_ring, we can greatly simplify the logic of tracking available space and not worry about other timelines. (Each ring is an ordered timeline of committed requests.) Signed-off-by:
-
Chris Wilson authored
With the introduction of requests, we amplified the number of atomic refcounted objects we use and update every execbuffer; from none to several references, and a set of references that need to be changed. We also introduced interesting side-effects in the order of retiring requests and objects. Instead of independently tracking the last request for an object, track the active objects for each request. The object will reside in the buffer list of its most recent active request and so we reduce the kref interchange to a list_move. Now retirements are entirely driven by the request, dramatically simplifying activity tracking on the object themselves, and removing the ambiguity between retiring objects and retiring requests. Furthermore with the consolidation of managing the activity tracking centrally, we can look forward to using RCU to enable lockless lookup of the current active requests for an object. In the future, we will be able to query the status or wait upon rendering to an object without even touching the struct_mutex BKL. All told, less code, simpler and faster, and more extensible. v2: Add a typedef for the function pointer for convenience later. v3: Make the noop retirement callback explicit. Allow passing NULL to the init_request_active() which is expanded to a common noop function. Signed-off-by:
-
Chris Wilson authored
Since we track requests, and requests are always added to the GPU fully formed, we never have to flush the incomplete request and know that the given request will eventually complete without any further action on our part. Signed-off-by:
-
Chris Wilson authored
We use "list" to denote the list and "link" to denote an element on that list. Rename request->list to match this idiom. Signed-off-by:
-
Chris Wilson authored
Tidy up the for loops that handle waiting for read/write vs read-only access. Signed-off-by:
-
Chris Wilson authored
The future annotations will track the locking used for access to ensure that it is always sufficient. We make the preparations now to present the API ahead and to make sure that GCC can eliminate the unused parameter. Before: 6298417 3619610 696320 10614347 a1f64b vmlinux After: 6298417 3619610 696320 10614347 a1f64b vmlinux (with i915 builtin) Signed-off-by:
-
Chris Wilson authored
In the future, we will want to add annotations to the i915_gem_active struct. The API is thus expanded to hide direct access to the contents of i915_gem_active and mediated instead through a number of helpers. Signed-off-by:
-
Chris Wilson authored
In the next patch, request tracking is made more generic and for that we need a new expanded struct and to separate out the logic changes from the mechanical churn, we split out the structure renaming into this patch. v2: Writer's block. Add some spiel about why we track requests. v3: Now i915_gem_active. v4: Now with i915_gem_active_set() for attaching to the active request. v5: Use i915_gem_active_set() from inside the retirement handlers Signed-off-by:
-
Chris Wilson authored
The drop_pages() function is a dangerous trap in that it can release the passed in object pointer and so unless the caller is aware, it can easily trick us into using the stale object afterwards. Move it into its solitary callsite where we know it is safe. Signed-off-by:
Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1470293567-10811-9-git-send-email-chris@chris-wilson.co.uk
-
Chris Wilson authored
When we call i915_vma_unbind(), we will wait upon outstanding rendering. This will also trigger a retirement phase, which may update the object lists. If, we extend request tracking to the VMA itself (rather than keep it at the encompassing object), then there is a potential that the obj->vma_list be modified for other elements upon i915_vma_unbind(). As a result, if we walk over the object list and call i915_vma_unbind(), we need to be prepared for that list to change. Signed-off-by:
-
Chris Wilson authored
Since we may have VMA allocated for an object, but we interrupted their binding, there is a disparity between have elements on the obj->vma_list and being bound. i915_gem_obj_bound_any() does this check, but this is not rigorously observed - add an explicit count to make it easier. Signed-off-by:
-
Chris Wilson authored
For the global GTT (and aliasing GTT), the address space is owned by the device (it is a global resource) and so the per-file owner field is NULL. For per-process GTT (where we create an address space per context), each is owned by the opening file. We can use this ownership information to both distinguish GGTT and ppGTT address spaces, as well as occasionally inspect the owner. v2: Whitespace, tells us who owns i915_address_space Signed-off-by:
-
Chris Wilson authored
Since we have a static if-else-chain for device probing of the global GTT, we do not need to use a function pointer, let alone store it when we never use it again. So use the if-else-chain to call down into the device specific probe. Signed-off-by:
-
Chris Wilson authored
Initialising the global GTT is tricky as we wish to use the drm_mm range manager during the modesetting initialisation (to capture stolen allocations from the BIOS) before we actually enable GEM. To overcome this, we currently setup the drm_mm first and then carefully rebind them. v2: Fixup after rebasing v3: GGTT initialisation needs to be split around kicking out conflicts v4: Restore an old UMS BUG_ON(mappable > total) as a DRM_ERROR plus fixup of probe results. Signed-off-by:
-
Chris Wilson authored
Since these are internal functions they operate on drm_i915_private and not the drm_device being passed in. So pass in the drm_i915_private instead, and remove one layer of dancing. No space wins here, just conforming to the norm in function parameters. v2: Include all the probe functions Signed-off-by:
-
Chris Wilson authored
In order to handle conflicting drivers (i.e. vgacon) having a different setup of hardware, we have to remove those other drivers before we try to setup our own mappings. This requires us to split GGTT initialisation between probing for the hardware location (part of the PCI BAR) and later establishing the kernel resources for it. Signed-off-by:
-
Chris Wilson authored
As we can now have multiple VMA inside the global GTT (with partial mappings, rotations, etc), it is no longer true that there may just be a single GGTT entry and so we should walk the full vma_list to count up the actual usage. In addition to unifying the two walkers, switch from multiplying the object size for each vma to summing the bound vma sizes. Signed-off-by:
-
- 03 Aug, 2016 2 commits
-
-
Ville Syrjälä authored
Remove the CHV early bail out from intel_cleanup_gt_powersave() so that we'll clean up the extra RPM reference held due to i915.enable_rc6=0. Cc: Imre Deak <imre.deak@intel.com> Fixes: b268c699 ("drm/i915: refactor RPM disabling due to RC6 being disabled") Signed-off-by:
Ville Syrjälä <ville.syrjala@linux.intel.com> Link: http://patchwork.freedesktop.org/patch/msgid/1470136053-23276-1-git-send-email-ville.syrjala@linux.intel.comReviewed-by:
Imre Deak <imre.deak@intel.com>
-
Chris Wilson authored
Now that we initialize the state to both legacy and execlists inside intel_engine_cs, we should also clean up that state from the common functions. Signed-off-by:
-
- 02 Aug, 2016 18 commits
-
-
Chris Wilson authored
In order to be more consistent with the rest of the request construction and ring emission, use the common names for the ring and request. Rather than using signaler_req, waiter_req, and intel_ring *wait, we use plain req and ring. Signed-off-by:
-
Chris Wilson authored
Since requests can no longer be generated as a side-effect of intel_ring_begin(), we know that the seqno will be unchanged during ring-emission. This predicatablity then means we do not have to check for the seqno wrapping around whilst emitting the semaphore for engine->sync_to(). Signed-off-by:
-
Chris Wilson authored
Now that emitting requests is identical between legacy and execlists, we can use the same function to build up the ring for submitting to either engine. (With the exception of i915_switch_contexts(), but in time that will also be handled gracefully.) Signed-off-by:
-
Chris Wilson authored
GCC was inlining the init and setup functions, but was getting itself confused into thinking that variables could be used uninitialised. If we do the inline for gcc, it is happy! As a bonus we shrink the code. v2: A couple of minor tweaks from Joonas Signed-off-by:
-
Chris Wilson authored
Now that we use the same vfuncs for emitting the batch buffer in both execlists and legacy, the golden render state initialisation is identical between both. v2: gcc wants so.ggtt_offset initialised (even though it is not used) Signed-off-by:
-
Chris Wilson authored
As gen6_emit_request() only differs from i9xx_emit_request() when semaphores are enabled, only use the specialised vfunc in that scenario. v2: Reorder semaphore init so as to keep engine->emit_request default vfunc selection compact. Signed-off-by:
-
Chris Wilson authored
As GEN6+ is now a simple variant on the basic breadcrumbs + tail write, reuse the common code. Signed-off-by:
-
Chris Wilson authored
Rather than pass in the num_dwords that the caller wishes to use after the signal command packet, split the breadcrumb emission into two phases and have both the signal and breadcrumb individiually acquire space on the ring. This makes the interface simpler for the reader, and will simplify for patches. Signed-off-by:
-
Chris Wilson authored
With adding engine->submit_request, we now have a bunch of functions with similar names used at different stages of the execlist submission. Try a different coat of paint, to hopefully reduce confusion between the requests, intel_engine_cs and the actual execlists submision process. Signed-off-by:
-
Chris Wilson authored
Move request submission from emit_request into its own common vfunc from i915_add_request(). v2: Convert I915_DISPATCH_flags to BIT(x) whilst passing v3: Rename a few functions to match. v4: Reenable execlists submission after disabling guc. v5: Be aware that everyone calls i915_guc_submission_disable()! Signed-off-by:
-
Chris Wilson authored
Space reservation is already safe with respect to the ring->size modulus, but hardware only expects to see values in the range 0...ring->size-1 (inclusive) and so requires the modulus to prevent us writing the value ring->size instead of 0. As this is only required for the register itself, we can defer the modulus to the register update and not perform it after every command packet. We keep the intel_ring_advance() around in the code to provide demarcation for the end-of-packet (which then can be compared against intel_ring_begin() as the number of dwords emitted must match the reserved space). v2: Assert that the ring size is a power-of-two to match assumptions in the code. Simplify the comment before writing the tail value to explain why the modulus is necessary. Signed-off-by:
-
Chris Wilson authored
If we rewrite the I915_WRITE_TAIL specialisation for the legacy ringbuffer as submitting the request onto the ringbuffer, we can unify the vfunc with both execlists and GuC in the next patch. v2: Drop the modulus from the I915_WRITE_TAIL as it is currently being applied in intel_ring_advance() after every command packet, and add a comment explaining why we need the modulus at all. Signed-off-by:
-
Chris Wilson authored
Joonas doesn't like the tiny function, especially if I go around making it more complicated and using it elsewhere. To remove that temptation, remove the function! Signed-off-by:
-
Chris Wilson authored
Both the ->dispatch_execbuffer and ->emit_bb_start callbacks do exactly the same thing, add MI_BATCHBUFFER_START to the request's ringbuffer - we need only one vfunc. Signed-off-by:
-
Chris Wilson authored
If is simpler and leads to more readable code through the callstack if the allocation returns the allocated struct through the return value. The importance of this is that it no longer looks like we accidentally allocate requests as side-effect of calling certain functions. Signed-off-by:
-
Chris Wilson authored
Rather than passing a complete set of GPU cache domains for either invalidation or for flushing, or even both, just pass a single parameter to the engine->emit_flush to determine the required operations. engine->emit_flush(GPU, 0) -> engine->emit_flush(EMIT_INVALIDATE) engine->emit_flush(0, GPU) -> engine->emit_flush(EMIT_FLUSH) engine->emit_flush(GPU, GPU) -> engine->emit_flush(EMIT_FLUSH | EMIT_INVALIDATE) This allows us to extend the behaviour easily in future, for example if we want just a command barrier without the overhead of flushing. Signed-off-by:
-
Chris Wilson authored
Space for flushing the GPU cache prior to completing the request is preallocated and so cannot fail - the GPU caches will always be flushed along with the completed request. This means we no longer have to track whether the GPU cache is dirty between batches like we had to with the outstanding_lazy_seqno. With the removal of the duplication in the per-backend entry points for emitting the obsolete lazy flush, we can then further unify the engine->emit_flush. v2: Expand a bit on the legacy of gpu_caches_dirty Signed-off-by:
-
Chris Wilson authored
For more consistent oop-naming, we would use intel_ring_verb, so pick intel_ring_pin() and intel_ring_unpin(). Signed-off-by:
-
