- 12 May, 2014 2 commits
-
-
Brad Volkin authored
For clients that submit large batch buffers the command parser has a substantial impact on performance. On my HSW ULT system performance drops as much as ~20% on some tests. Most of the time is spent in the command lookup code. Converting that from the current naive search to a hash table lookup reduces the performance drop to ~10%. The choice of value for I915_CMD_HASH_ORDER allows all commands currently used in the parser tables to hash to their own bucket (except for one collision on the render ring). The tradeoff is that it wastes memory. Because the opcodes for the commands in the tables are not particularly well distributed, reducing the order still leaves many buckets empty. The increased collisions don't seem to have a huge impact on the performance gain, but for now anyhow, the parser trades memory for performance. NB: Ville noticed that the error paths through the ring init code will leak memory. I've not addressed that here. We can do a follow up pass to handle all of the leaks. v2: improved comment describing selection of hash key mask (Damien) replace a BUG_ON() with an error return (Tvrtko, Ville) commit message improvements Signed-off-by:
Brad Volkin <bradley.d.volkin@intel.com> Reviewed-by:
Damien Lespiau <damien.lespiau@intel.com> Reviewed-by:
Tvrtko Ursulin <tvrtko.ursulin@intel.com> Signed-off-by:
Daniel Vetter <daniel.vetter@ffwll.ch>
-
Ville Syrjälä authored
Use the same code for enabling/disabling planes on all platforms. Rename the functions to reflect that they're no longer specific to any platform. For now we leave the plane enable/disable to ccur at the same old position in the modeset sequence. Signed-off-by:
Ville Syrjälä <ville.syrjala@linux.intel.com> Reviewed-by:
Chris Wilson <chris@chris-wilson.co.uk> [danvet: Frob drm_vblank_on conflict.] Signed-off-by:
-
- 07 May, 2014 8 commits
-
-
Chris Wilson authored
During the review of commit 1f70999f Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Mon Jan 27 22:43:07 2014 +0000 drm/i915: Prevent recursion by retiring requests when the ring is full Ville raised the point that our interaction with request->tail was likely to foul up other uses elsewhere (such as hang check comparing ACTHD against requests). However, we also need to restore the implicit retire requests that certain test cases depend upon (e.g. igt/gem_exec_lut_handle), this raises the spectre that the ppgtt will randomly call i915_gpu_idle() and recurse back into intel_ring_begin(). Signed-off-by:
-
Chris Wilson authored
A few improvements to the fallback method for waiting upon ring space: 1. Fix the start/end wait tracepoints to always be paired. 2. Increase responsiveness of checking 3. Mark the process as waiting upon io 4. Check for signal interruptions Signed-off-by:
-
Ben Widawsky authored
There is a good debate to be had about how best to fit the aliasing PPGTT into the code. However, as it stands right now, getting aliasing PPGTT bindings is a hack, and done through implicit arguments. To make this absolutely clear, WARN and return an error if a driver writer tries to do something they shouldn't. I have no issue with an eventual revert of this patch. It makes sense for what we have today. Signed-off-by:
Ben Widawsky <ben@bwidawsk.net> Signed-off-by:
-
Ben Widawsky authored
It was always the intention to do the topdown allocation for context objects (Chris' idea originally). Unfortunately, I never managed to land the patch, but someone else did, so now we can use it. As a reminder, hardware contexts never need to be in the precious GTT aperture space - which is what is what happens with the normal bottom up allocation we do today. Doing a top down allocation increases the odds that the HW contexts can get out of the way, especially with per FD contexts as is done in full PPGTT Signed-off-by:
-
Imre Deak authored
Signed-off-by:
Imre Deak <imre.deak@intel.com> Signed-off-by:
-
Imre Deak authored
Add runtime PM support for VLV, but leave it disabled. The next patch enables it. The suspend/resume sequence used is based on [1] and [2]. In practice we depend on the GT RC6 mechanism to save the HW context depending on the render and media power wells. By the time we run the runtime suspend callback the display side is also off and the HW context for that is managed by the display power domain framework. Besides the above there are Gunit registers that depend on a system-wide power well. This power well goes off once the device enters any of the S0i[R123] states. To handle this scenario, save/restore these Gunit registers. Note that this is not the complete register set dictated by [2], to remove some overhead, registers that are known not to be used are ignored. Also some registers are fully setup by initialization functions called during resume, these are not saved either. The list of registers can be further reduced, see the TODO note in the code. [1] VLV_gfx_clocking_PM_reset_y12w21d3 / "Driver D3 entry/exit" [2] VLV2_S0IXRegs v2: - unchanged v3: - fix s/GEN6_PMIIR/GEN6_PMIMR/ typo when saving/restoring registers (Ville) v4: - rebased on the previous patch fixing GEN register prefixes Reviewed-by:
-
Imre Deak authored
Atm, none of the RPM callbacks can fail, but the next patch adding RPM support for VLV changes this, so prepare for it. In case one of these callbacks return error RPM will get permanently disabled until the error is explicitly cleared. In the future we could add support for re-enabling it, for example after resetting the HW, but for now - hopefully - we can live with the simpler solution. v2: - propagate the error from the resume callbacks too (Paulo) v3: - fix rebase fail typo around IS_GEN6() check in intel_runtime_suspend() Signed-off-by:
-
Imre Deak authored
Needed by the VLV S0ix context save/restore helpers. v2: - unchanged v3: - use proper GEN register prefixes (Ville) Signed-off-by:
-
- 06 May, 2014 13 commits
-
-
Ville Syrjälä authored
Due to Pipe C DPINVGTT has more bits on CHV. v2: Fix comment to say VLV/CHV (Rafael) Reviewed-by:
Rafael Barbalho <rafael.barbalho@intel.com> Signed-off-by:
-
Ville Syrjälä authored
v2: Rebase on top of Ben's GT interrupt shuffling. Signed-off-by:
-
Ville Syrjälä authored
CHV has pipe C and PSR which cause changes to DPFLIPSTAT. Signed-off-by:
-
Ville Syrjälä authored
FIXME: We probably want to sprinkle _CHV suffixes over these. Signed-off-by:
-
Ville Syrjälä authored
Enable aliasing PPGTT for CHV, but keep full PPGTT still disabled until it gets enabled for BDW. Signed-off-by:
-
Rafael Barbalho authored
Page table updates were getting stuck in the CPU cache on chv causing spurious page faults and strange behaviour. Signed-off-by:
-
Ville Syrjälä authored
Ignore the cache bits in PPAT and just set the snoop bit where appropriate. BDW WB is mapped to snooped access, while all other modes are mapped to non-snooped access. The hardware supposedly ignores everything except the snoop bit in the PPAT entries. Additionally the hardware actually enforces snooping for all page table accesses, and thus the snoop bit is ignored for PDEs. v2: Rebased on top of the bdw resume fix to reload the ppat entries. v3: Rebase on top of the i915_gem_gtt.h header extraction. Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com> (v1) Acked-by:
-
Ville Syrjälä authored
We won't be calling intel_enable_primary_plane() or intel_disable_primary_plane() with the primary plane in the wrong state. So remove the useless DISPLAY_PLANE_ENABLE checks. v2: Convert the checks to WARNs instead (Daniel,Paulo) Signed-off-by:
Paulo Zanoni <paulo.r.zanoni@intel.com> Signed-off-by:
-
Ville Syrjälä authored
On ILK when we disable a particular watermark level, we must maintain the actual watermark values for that level for some time (until the next vblank possibly). Otherwise we risk underruns. In order to achieve that result we must merge the LP1+ watermarks a bit differently since we must also merge levels that are to be disabled. We must also make sure we don't overflow the fields in the watermark registers in case the calculated watermarks come out too big to fit. As early as possbile we mark all computed watermark levels as disabled if they would exceed the register maximums. We make sure to leave the actual watermarks for such levels zeroed out. Then during merging, we take the maxium values for every level, regardless if they're disabled or not. That may seem a bit pointless since at the moment all the watermark levels we merge should have their values zeroed if the level is already disabled. However soon we will be dealing with intermediate watermarks that, in addition to the new watermark values, also contain the previous watermark values, and so levels that are disabled may no longer be zeroed out. v2: Split the patch in two (Paulo) Use if() instead of & when merging ->enable (Paulo) Signed-off-by: -
Ville Syrjälä authored
When we calculate the watermarks for a pipe make sure we leave any level fully zeroed out if it would exceed any of the maximum values that fit in the registers. This will be important later when we start to use also disabled watermark levels during LP1+ merging. Signed-off-by:
-
Ville Syrjälä authored
Add trace points for observing the atomic pipe update mechanism. v2: Rebased due to earlier changes v3: Pass intel_crtc instead of drm_crtc (Daniel) v4: Pass frame counter from the caller to evaded/end since the caller now always has that ready Reviewed-by:Jesse Barnes <jbarnes@virtuousgeek.org> Reviewed-by:
Sourab Gupta <sourabgupta@gmail.com> Reviewed-by:
Akash Goel <akash.goels@gmail.com> Signed-off-by:
-
Ville Syrjälä authored
Move the primary plane enable/disable to occur atomically with the sprite update that caused the primary plane visibility to change. FBC and IPS enable/disable is left to happen well before or after the primary plane change. v2: Pass intel_crtc instead of drm_crtc (Daniel) Reviewed-by:
-
Ville Syrjälä authored
Add a mechanism by which we can evade the leading edge of vblank. This guarantees that no two sprite register writes will straddle on either side of the vblank start, and that means all the writes will be latched together in one atomic operation. We do the vblank evade by checking the scanline counter, and if it's too close to the start of vblank (too close has been hardcoded to 100usec for now), we will wait for the vblank start to pass. In order to eliminate random delayes from the rest of the system, we operate with interrupts disabled, except when waiting for the vblank obviously. Note that we now go digging through pipe_to_crtc_mapping[] in the vblank interrupt handler, which is a bit dangerous since we set up interrupts before the crtcs. However in this case since it's the vblank interrupt, we don't actually unmask it until some piece of code requests it. v2: preempt_check_resched() calls after local_irq_enable() (Jesse) Hook up the vblank irq stuff on BDW as well v3: Pass intel_crtc instead of drm_crtc (Daniel) Warn if crtc.mutex isn't locked (Daniel) Add an explicit compiler barrier and document the barriers (Daniel) Note the irq vs. modeset setup madness in the commit message (Daniel) v4: Use prepare_to_wait() & co. directly and eliminate vbl_received v5: Refactor intel_pipe_handle_vblank() vs. drm_handle_vblank() (Chris) Check for min/max scanline <= 0 (Chris) Don't call intel_pipe_update_end() if start failed totally (Chris) Check that the vblank counters match on both sides of the critical section (Chris) v6: Fix atomic update for interlaced modes v7: Reorder code for better readability (Chris) v8: Drop preempt_check_resched(). It's not available to modules anymore and isn't even needed unless we ourselves cause a wakeup needing reschedule while interrupts are off Reviewed-by:
-
- 05 May, 2014 17 commits
-
-
Ben Widawsky authored
All the rest of the code to enable this is in my branch. Without my branch, hitting > 32b offsets is impossible. The code has always "supported" 64b, but it's never actually been run of tested. This change doesn't actually fix anything. [1] I am not sure why X won't work yet. I do not get hangs or obvious errors. There are 3 fixes grouped together here. First is to remove the hardcoded 0 for the upper dword of the relocation. The next fix is to use a 64b value for target_offset. The final fix is to not directly apply target_offset to reloc->delta. reloc->delta is part of ABI, and so we cannot change it. As it stands, 32b is enough to represent everything we're interested in representing anyway. The main problem is, we cannot add greater than 32b values to it directly. [1] Almost all of intel-gpu-tools is not yet ready to test 64b relocations. There are a few places that expect 32b values for offsets and these all won't work. Cc: Rafael Barbalho <rafael.barbalho@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by:
-
Ben Widawsky authored
Previously, our code only had a 32b offset value for where the batchbuffer starts. With full PPGTT, and 64b canonical GPU address space, that is an insufficient value. The code to expand is pretty straight forward, and only one platform needs to do anything with the extra bits. Signed-off-by:
-
Daniel Vetter authored
SDVO is used by both crtcs using the i9xx_ and the ironlake_ functions. For both cases there is nothing between the encoder->mode_set and the encoder->pre_enable calls that touches the hardware. The vlv_ functions are different since they enable the pll before the ->pre_enable hook. But SDVO isn't supported on vlv platforms, so this doesn't matter. We've also already clean up all the sdvo state computation logic, all relevant parts are already in the ->compute_config hook. So we can just get rid of the ->mode_set hook by converting it to a ->pre_enable hook. Reviewed-by:
-
Daniel Vetter authored
We only set a few bits in the ADPA register, which we then read back in the enable/disable hooks. So we can just move that bit of state computation code to the place where we need it since setting these bits without enabling the CRT encoder has no effects. The only exceptions are the hotplug bits since they affect the hotplug detection logic, but we already set those in the ->reset function and then never touch them. Reviewed-by:
-
Daniel Vetter authored
Currently for the i9xx crtc hooks there's nothing between the call to encoder->mode_set and encoder->pre_enable which touches the hardware. Therefore, since tv is only used on gen3/4, we can just move the hook. Yay for easy cases! The only other important thing to check is that the new ->pre_enable hook is idempotent wrt the sw state since now it can be called multiple times (due to DPMS). After a the bit of refactoring this is now easy to check: It only reads crtc->config and computes derived state but otherwise leaves it as-is, so we're good. Reviewed-by:
-
Daniel Vetter authored
The pipe and plane _are_ disabled when we call this. So replace it all with the corresponding assert (as self-documenting code) and rip out all the lore. Checking for a disabled plane would require us to export those macros from intel_display.c, but if the pipe is off the plane isn't working either. So this single check is good enough. Reviewed-by:
-
Daniel Vetter authored
We only support TV-out on gen3/4 mobile platforms, and i915gm is the only one that matches. Reviewed-by:
-
Daniel Vetter authored
intel_tv_mode_set is still too bug. Reviewed-by:
-
Daniel Vetter authored
intel_tv_mode_set is just too big. Reviewed-by:
-
Daniel Vetter authored
Currently for the i9xx crtc hooks there's nothing between the call to encoder->mode_set and encoder->pre_enable which touches the hardware. Therefore, since dvo is only used on gen2, we can just move the hook. Yay for easy cases! The only other important thing to check is that the new ->pre_enable hook is idempotent wrt the sw state since now it can be called multiple times (due to DPMS). It only reads crtc->config but otherwise leaves it as-is, so we're good. Reviewed-by:
-
Daniel Vetter authored
For a bunch of reasons we want to move away from the ->mode_set callbacks: All hw state setup needs to move into ->enable hooks (so that DOMS can do runtime pm) and all the configuration setup needs to move into the compute_config functions. To start with this make the enocer->mode_set callback optional. Reviewed-by:
-
Ville Syrjälä authored
The BIOS can enable a pipe but leave the primary plane disabled. This coflicts with out current idea of primary_enabled. Read the actual hardware plane state and set primary_enabled appropriately. We currently assume that primary_enabled is always true when we're about to disable a crtc. That needs to change now as the plane may not be enabled. So replace the relevant WARNs with early returns in intel_{enable,disable}_primary_hw_plane(). Fixes the following warning [ 3.831602] WARNING: CPU: 0 PID: 1112 at linux/drivers/gpu/drm/i915/intel_display.c:1918 intel_disable_primary_hw_plane+0xe4/0xf0 [i915]() which got introduced here by me: commit e9e39655c0c30cddc3f8c09a757678a24dd36737 Author: Ville Syrjälä <ville.syrjala@linux.intel.com> Date: Mon Apr 28 15:53:25 2014 +0300 drm/i915: Remove useless checks from primary enable/disable Signed-off-by: -
Ben Widawsky authored
Add_request has always contained both the semaphore mailbox updates as well as the breadcrumb writes. Since the semaphore signal is the one which actually knows about the number of dwords it needs to emit to the ring, we move the ring_begin to that function. This allows us to remove the hideously shared #define On a related not, gen8 will use a different number of dwords for semaphores, but not for add request. v2: Make number of dwords an explicit part of signalling (via function argument). (Chris) v3: very slight comment change Reviewed-by:
-
Ben Widawsky authored
This abstraction again is in preparation for gen8. Gen8 will bring new semantics for doing this operation. While here, make the writes of MI_NOOPs explicit for non-existent rings. This should have been implicit before. NOTE: This is going to be removed in a few patches. Reviewed-by:
-
Ben Widawsky authored
This will be helpful in abstracting some of the code in preparation for gen8 semaphores. v2: Move mbox stuff to a separate struct v3: Rebased over VCS2 work Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com> (v1) Signed-off-by:
-
Imre Deak authored
During the initial power well enabling on the driver init/resume path we can avoid initialzing part of the HW/SW state that will be initialized anyway by the subsequent init/resume code. For some steps like HPD initialization this redundancy is not only an overhead but an actual problem, since they can't be run this early in the overall init sequence. Add a flag marking the init phase and skip reinitialzing state that is not strictly necessary based on that. This is also needed by the upcoming HPD init restructuring by Thierry and Daniel. Signed-off-by:
-
Chris Wilson authored
In commit 691e6415 Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Wed Apr 9 09:07:36 2014 +0100 drm/i915: Always use kref tracking for all contexts. we populated fake contexts on all platforms. These were identical to the full hardware context tracking structs, except for the ctx->obj used to store the hardware state. However, there remained one place where we assumed that if a context existed, it would have an object associated with it. Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=77717 Testcase: igt/drv_suspend/debugfs-reader Signed-off-by:
-
