]> git.karo-electronics.de Git - linux-beck.git/log
linux-beck.git
11 years agodrm/i915: Fix context size calculation on SNB/IVB/VLV
Ville Syrjälä [Thu, 22 Aug 2013 16:23:13 +0000 (19:23 +0300)]
drm/i915: Fix context size calculation on SNB/IVB/VLV

All the different context sizes reported in the CXT_SIZE register
aren't meant to be simply added together.

While BSpec is somewhat unclear on the topic of the actual context
size, empirical tests have now revealed the truth. So let's add a
big fat comment to remind people how it all works.

As a result of correctly interpreting CXT_SIZE, the IVB context
size is reduced from three pages to two, while SNB context size
remains at two pages.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Use POSTING_READ in lcpll code
Daniel Vetter [Wed, 21 Aug 2013 21:38:08 +0000 (23:38 +0200)]
drm/i915: Use POSTING_READ in lcpll code

If we don't use the return value of a mmio read our coding style is to
use the POSTING_READ macro. This avoids cluttering the mmio traces.

While at it add the missing posting read in the lcpll enable function
that Paulo spotted.

v2: Drop the _NOTRACE changes, tracing such wait_for loops in the modeset
code might actually be rather useful!

Cc: Paulo Zanoni <przanoni@gmail.com>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: enable Package C8+ by default
Paulo Zanoni [Mon, 19 Aug 2013 16:18:12 +0000 (13:18 -0300)]
drm/i915: enable Package C8+ by default

This should be working, so enable it by default. Also easy to revert.

v2: Rebase, s/allow/enable/.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: add i915.pc8_timeout function
Paulo Zanoni [Mon, 19 Aug 2013 16:18:11 +0000 (13:18 -0300)]
drm/i915: add i915.pc8_timeout function

We currently only enter PC8+ after all its required conditions are
met, there's no rendering, and we stay like that for at least 5
seconds.

I chose "5 seconds" because this value is conservative and won't make
us enter/leave PC8+ thousands of times after the screen is off: some
desktop environments have applications that wake up and do rendering
every 1-3 seconds, even when the screen is off and the machine is
completely idle.

But when I was testing my PC8+ patches I set the default value to
100ms so I could use the bad-behaving desktop environments to
stress-test my patches. I also thought it would be a good idea to ask
our power management team to test different values, but I'm pretty
sure they would ask me for an easy way to change the timeout. So to
help these 2 cases I decided to create an option that would make it
easier to change the default value. I also expect people making
specific products that use our driver could try to find the perfect
timeout for them.

Anyway, fixing the bad-behaving applications will always lead to
better power savings than just changing the timeout value: you need to
stop waking the Kernel, not quickly put it back to sleep again after
you wake it for nothing. Bad sleep leads to bad mood!

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: add i915_pc8_status debugfs file
Paulo Zanoni [Mon, 19 Aug 2013 16:18:10 +0000 (13:18 -0300)]
drm/i915: add i915_pc8_status debugfs file

Make it print the value of the variables on the PC8 struct.

v2: Update to recent renames and add the new fields.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: allow package C8+ states on Haswell (disabled)
Paulo Zanoni [Mon, 19 Aug 2013 16:18:09 +0000 (13:18 -0300)]
drm/i915: allow package C8+ states on Haswell (disabled)

This patch allows PC8+ states on Haswell. These states can only be
reached when all the display outputs are disabled, and they allow some
more power savings.

The fact that the graphics device is allowing PC8+ doesn't mean that
the machine will actually enter PC8+: all the other devices also need
to allow PC8+.

For now this option is disabled by default. You need i915.allow_pc8=1
if you want it.

This patch adds a big comment inside i915_drv.h explaining how it
works and how it tracks things. Read it.

v2: (this is not really v2, many previous versions were already sent,
     but they had different names)
    - Use the new functions to enable/disable GTIMR and GEN6_PMIMR
    - Rename almost all variables and functions to names suggested by
      Chris
    - More WARNs on the IRQ handling code
    - Also disable PC8 when there's GPU work to do (thanks to Ben for
      the help on this), so apps can run caster
    - Enable PC8 on a delayed work function that is delayed for 5
      seconds. This makes sure we only enable PC8+ if we're really
      idle
    - Make sure we're not in PC8+ when suspending
v3: - WARN if IRQs are disabled on __wait_seqno
    - Replace some DRM_ERRORs with WARNs
    - Fix calls to restore GT and PM interrupts
    - Use intel_mark_busy instead of intel_ring_advance to disable PC8
v4: - Use the force_wake, Luke!
v5: - Remove the "IIR is not zero" WARNs
    - Move the force_wake chunk to its own patch
    - Only restore what's missing from RC6, not everything

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: fix SDEIMR assertion when disabling LCPLL
Paulo Zanoni [Mon, 19 Aug 2013 16:18:08 +0000 (13:18 -0300)]
drm/i915: fix SDEIMR assertion when disabling LCPLL

This was causing WARNs in one machine, so instead of trying to guess
exactly which hotplug bits should exist, just do the test on the
non-HPD bits. We don't care about the state of the hotplug bits, we
just care about the others, that need to be 1.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: grab force_wake when restoring LCPLL
Paulo Zanoni [Mon, 19 Aug 2013 16:18:07 +0000 (13:18 -0300)]
drm/i915: grab force_wake when restoring LCPLL

If LCPLL is disabled, there's a chance we might be in package C8 state
or deeper, and we'll get a hard hang when restoring LCPLL (also, a red
led lights up on my motherboard). So grab the force_wake, which will
get us out of RC6 and, as a consequence, out of PC8+ (since we need
RC6 to get into PC8+).

Note: Discussions with hw designers are still ongoing what exactly
goes boom here. But I think we can go ahead and just merge this little
hack for now until it's clear what we actually need.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
[danvet: Add small note about the current state of the discussion
around this hack.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: drop WaMbcDriverBootEnable workaround
Jesse Barnes [Wed, 21 Aug 2013 15:08:55 +0000 (08:08 -0700)]
drm/i915: drop WaMbcDriverBootEnable workaround

Turns out the BIOS will do this for us as needed, and if we try to do it
again we risk hangs or other bad behavior.

Note that this seems to break libva on ChromeOS after resumes (but
strangely _not_ after booting up).

This essentially reverts

commit b4ae3f22d238617ca11610b29fde16cf8c0bc6e0
Author: Jesse Barnes <jbarnes@virtuousgeek.org>
Date:   Thu Jun 14 11:04:48 2012 -0700

    drm/i915: load boot context at driver init time

and

commit b3bf076697a68a8577f4a5f7407de0bb2b3b56ac
Author: Paulo Zanoni <paulo.r.zanoni@intel.com>
Date:   Tue Nov 20 13:27:44 2012 -0200

    drm/i915: implement WaMbcDriverBootEnable on Haswell

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reported-and-Tested-by: Stéphane Marchesin <marcheu@chromium.org>
[danvet: Add note about impact and regression citation.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Cleaning up the relocate entry function
Rafael Barbalho [Wed, 21 Aug 2013 16:10:51 +0000 (17:10 +0100)]
drm/i915: Cleaning up the relocate entry function

As the relocate entry function was getting a bit too big I've moved
the code that used to use either the cpu or the gtt to for the
relocation into two separate functions.

Signed-off-by: Rafael Barbalho <rafael.barbalho@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: merge HSW and SNB PM irq handlers
Paulo Zanoni [Thu, 15 Aug 2013 14:51:32 +0000 (11:51 -0300)]
drm/i915: merge HSW and SNB PM irq handlers

Because hsw_pm_irq_handler does exactly what gen6_rps_irq_handler does
and also processes the 2 additional VEBOX bits. So merge those
functions and wrap the VEBOX bits on a HAS_VEBOX check. This
check isn't really necessary since the bits are reserved on
SNB/IVB/VLV, but it's a good documentation on who uses them.

v2: - Change IS_HASWELL check to HAS_VEBOX

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: fix how we mask PMIMR when adding work to the queue
Paulo Zanoni [Fri, 9 Aug 2013 20:04:36 +0000 (17:04 -0300)]
drm/i915: fix how we mask PMIMR when adding work to the queue

It seems we've been doing this ever since we started processing the
RPS events on a work queue, on commit "drm/i915: move gen6 rps
handling to workqueue", 4912d04193733a825216b926ffd290fada88ab07.

The problem is: when we add work to the queue, instead of just masking
the bits we queued and leaving all the others on their current state,
we mask the bits we queued and unmask all the others. This basically
means we'll be unmasking a bunch of interrupts we're not going to
process. And if you look at gen6_pm_rps_work, we unmask back only
GEN6_PM_RPS_EVENTS, which means the bits we unmasked when adding work
to the queue will remain unmasked after we process the queue.

Notice that even though we unmask those unrelated interrupts, we never
enable them on IER, so they don't fire our interrupt handler, they
just stay there on IIR waiting to be cleared when something else
triggers the interrupt handler.

So this patch does what seems to make more sense: mask only the bits
we add to the queue, without unmasking anything else, and so we'll
unmask them after we process the queue.

As a side effect we also have to remove that WARN, because it is not
only making sure we don't mask useful interrupts, it is also making
sure we do unmask useless interrupts! That piece of code should not be
responsible for knowing which bits should be unmasked, so just don't
assert anything, and trust that snb_disable_pm_irq should be doing the
right thing.

With i915.enable_pc8=1 I was getting ocasional "GEN6_PMIIR is not 0"
error messages due to the fact that we unmask those unrelated
interrupts but don't enable them.

Note: if bugs start bisecting to this patch, then it probably means
someone was relying on the fact that we unmask everything by accident,
then we should fix gen5_gt_irq_postinstall or whoever needs the
accidentally unmasked interrupts. Or maybe I was just wrong and we
need to revert this patch :)

Note: This started to be a more real issue with the addition of the
VEBOX support since now we do enable more than just the minimal set of
RPS interrupts in the IER register. Which means after the first rps
interrupt has happened we will never mask the VEBOX user interrupts
again and so will blow through cpu time needlessly when running video
workloads.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Add note that this started to matter with VEBOX much more.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: don't queue PM events we won't process
Paulo Zanoni [Thu, 15 Aug 2013 14:50:01 +0000 (11:50 -0300)]
drm/i915: don't queue PM events we won't process

On SNB/IVB/VLV we only call gen6_rps_irq_handler if one of the IIR
bits set is part of GEN6_PM_RPS_EVENTS, but at gen6_rps_irq_handler we
add all the enabled IIR bits to the work queue, not only the ones that
are part of GEN6_PM_RPS_EVENTS. But then gen6_pm_rps_work only
processes GEN6_PM_RPS_EVENTS, so it's useless to add anything that's
not GEN6_PM_RPS_EVENTS to the work queue.

As a bonus, gen6_rps_irq_handler looks more similar to
hsw_pm_irq_handler, so we may be able to merge them in the future.

v2: - Add a WARN in case we queued something we're not going to
      process.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net> (v1)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: don't disable/reenable IVB error interrupts when not needed
Paulo Zanoni [Tue, 6 Aug 2013 21:57:16 +0000 (18:57 -0300)]
drm/i915: don't disable/reenable IVB error interrupts when not needed

If the error interrupts are already disabled, don't disable and
reenable them. This is going to be needed when we're in PC8+, where
all the interrupts are disabled so we won't risk re-enabling
DE_ERR_INT_IVB.

v2: Use dev_priv->irq_mask (Chris)

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: add dev_priv->pm_irq_mask
Paulo Zanoni [Tue, 6 Aug 2013 21:57:15 +0000 (18:57 -0300)]
drm/i915: add dev_priv->pm_irq_mask

Just like irq_mask and gt_irq_mask, use it to track the status of
GEN6_PMIMR so we don't need to read it again every time we call
snb_update_pm_irq.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: don't update GEN6_PMIMR when it's not needed
Paulo Zanoni [Tue, 6 Aug 2013 21:57:14 +0000 (18:57 -0300)]
drm/i915: don't update GEN6_PMIMR when it's not needed

I did some brief tests and the "new_val = pmimr" condition usually
happens a few times after exiting games.

Note: This is also prep work to track the GEN6_PMIMR register state in
dev_priv->pm_imr. This happens in the next patch.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
[danvet: Add note to explain why we want this, as per the discussion
between Chris and Paulo.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: wrap GEN6_PMIMR changes
Paulo Zanoni [Tue, 6 Aug 2013 21:57:13 +0000 (18:57 -0300)]
drm/i915: wrap GEN6_PMIMR changes

Just like we're doing with the other IMR changes.

One of the functional changes is that not every caller was doing the
POSTING_READ.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: wrap GTIMR changes
Paulo Zanoni [Tue, 6 Aug 2013 21:57:12 +0000 (18:57 -0300)]
drm/i915: wrap GTIMR changes

Just like the functions that touch DEIMR and SDEIMR, but for GTIMR.
The new functions contain a POSTING_READ(GTIMR) which was not present
at the 2 callers inside i915_irq.c.

The implementation is based on ibx_display_interrupt_update.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: add the FCLK case to intel_ddi_get_cdclk_freq
Paulo Zanoni [Tue, 6 Aug 2013 21:57:11 +0000 (18:57 -0300)]
drm/i915: add the FCLK case to intel_ddi_get_cdclk_freq

We already have code to disable LCPLL and switch to FCLK, so we need this too.
We still don't call the code to disable LCPLL, but we'll call it when we add
support for Package C8+.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Expose energy counter on SNB+ through debugfs
Jesse Barnes [Tue, 20 Aug 2013 09:29:23 +0000 (10:29 +0100)]
drm/i915: Expose energy counter on SNB+ through debugfs

On SNB and IVB, there's an MSR (also exposed through MCHBAR) we can use
to read out the amount of energy used over time.  Expose this in sysfs
to make it easy to do power comparisons with different configurations.

If the platform supports it, the file will show up under the
drm/card0/power subdirectory of the PCI device in sysfs as gt_energy_uJ.
The value in the file is a running total of energy (in microjoules)
consumed by the graphics device.

v2: move to sysfs (Ben, Daniel)
    expose a simple value (Chris)
    drop unrelated hunk (Ben)

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
v3: by Ben
Tied it into existing rc6  sysfs entries and named that a more generic
"power attrs." Fixed rebase conflicts.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
v4: Since RAPL is a real driver that already exists to serve power
monitoring, place our entry in debugfs. This gives me a fallback
location for systems that do not expose it otherwise.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove I915_READ_{NOPID, SYNC_0, SYNC_1})()
Damien Lespiau [Mon, 19 Aug 2013 18:32:03 +0000 (19:32 +0100)]
drm/i915: Remove I915_READ_{NOPID, SYNC_0, SYNC_1})()

The code directly uses the registers and ring->mmio_base.

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm: Remove IS_IRONLAKE_D()
Damien Lespiau [Mon, 19 Aug 2013 18:32:02 +0000 (19:32 +0100)]
drm: Remove IS_IRONLAKE_D()

This define hasn't been used since:

  commit cfdf1fa23f4074c9f8766dc67a928bbf680b1ac9
  Author: Kristian Høgsberg <krh@bitplanet.net>
  Date:   Wed Dec 16 15:16:16 2009 -0500

      drm/i915: Implement IS_* macros using static tables

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove HAS_PIPE_CONTROL()
Damien Lespiau [Mon, 19 Aug 2013 18:32:01 +0000 (19:32 +0100)]
drm/i915: Remove HAS_PIPE_CONTROL()

The code using this was removed in:

  commit 88f23b8fa3e6357c423af24ec31c661fc12f884b
  Author: Chris Wilson <chris@chris-wilson.co.uk>
  Date:   Sun Dec 5 15:08:31 2010 +0000

      drm/i915: Avoid using PIPE_CONTROL on Ironlake

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove DSPARB_HWCONTROL()
Damien Lespiau [Mon, 19 Aug 2013 18:32:00 +0000 (19:32 +0100)]
drm/i915: Remove DSPARB_HWCONTROL()

This define hasn't been used since:

  commit 652c393a3368af84359da37c45afc35a91144960
  Author: Jesse Barnes <jbarnes@virtuousgeek.org>
  Date:   Mon Aug 17 13:31:43 2009 -0700

      drm/i915: add dynamic clock frequency control

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: make IVB FDI training match spec v3
Jesse Barnes [Mon, 19 Aug 2013 18:04:55 +0000 (11:04 -0700)]
drm/i915: make IVB FDI training match spec v3

The existing code was trying different vswing and preemphasis settings
in the wrong place, and wasn't trying them enough.  So add a loop to
walk through them, properly disabling FDI TX and RX in between if a
failure is detected.

v2: remove unneeded reg writes, add delays around bit lock checks (Jesse)
v3: fix TX and RX disable per spec (Paulo)
    fix delays per spec (Paulo)
    make RX symbol lock check match TX bit lock check (Paulo)

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=51983
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915/vma: Correct use after free in eviction
Ben Widawsky [Fri, 16 Aug 2013 20:29:33 +0000 (13:29 -0700)]
drm/i915/vma: Correct use after free in eviction

The vma will [possibly] be destroyed during unbind in eviction.
Immediately after this, we try to delete the list entry.

Chris and Ville did the debug on this before I woke up, I just get to
take credit for the fix :p

For future reference the Oops that Mika reported:

[  403.472448] BUG: unable to handle kernel paging request at 6b6b6b6b
[  403.472473] IP: [<c12c1500>] __list_del_entry+0x20/0xe0
[  403.472514] *pdpt = 000000002e89c001 *pde = 0000000000000000
[  403.472556] Oops: 0000 [#1] SMP
[  403.472582] Modules linked in: mxm_wmi snd_hda_codec_hdmi snd_hda_codec_realtek snd_hda_intel snd_hda_codec snd_hwdep snd_pcm snd_seq_midi snd_rawmidi psmouse snd_seq_midi_event snd_seq serio_raw snd_timer snd_seq_device snd soundcore snd_page_alloc wmi bnep rfcomm bluetooth mac_hid parport_pc ppdev lp parport usbhid dm_crypt firewire_ohci firewire_core crc_itu_t i915 drm_kms_helper e1000e ptp drm i2c_algo_bit pps_core xhci_hcd video
[  403.472895] CPU: 2 PID: 1940 Comm: Xorg Not tainted 3.11.0-rc2+ #827
[  403.472938] Hardware name:                  /DZ77BH-55K, BIOS BHZ7710H.86A.0070.2012.0416.2117 04/16/2012
[  403.473002] task: ec866c00 ti: ee6a2000 task.ti: ee6a2000
[  403.473039] EIP: 0060:[<c12c1500>] EFLAGS: 00013202 CPU: 2
[  403.473078] EIP is at __list_del_entry+0x20/0xe0
[  403.473109] EAX: f016d9bc EBX: f016d9bc ECX: 6b6b6b6b EDX: 6b6b6b6b
[  403.473151] ESI: 00000000 EDI: ee6a3c90 EBP: ee6a3c60 ESP: ee6a3c48
[  403.473193]  DS: 007b ES: 007b FS: 00d8 GS: 00e0 SS: 0068
[  403.473230] CR0: 80050033 CR2: 6b6b6b6b CR3: 2ec43000 CR4: 001407f0
[  403.473271] Stack:
[  403.473285]  f63b2ff0 f61f98c0 f61f8000 f016d9bc 00000000 f016d9bc ee6a3cac f8519a4a
[  403.473347]  00000000 00000000 10000000 f61f8000 0100a000 10000000 00000001 008ca000
[  403.473410]  f64ee840 f61f98c0 f016d9bc f016dcec ee6a3c98 ee6a3c98 f61f98c0 dcc58f00
[  403.473472] Call Trace:
[  403.473509]  [<f8519a4a>] i915_gem_evict_something+0x17a/0x2d0 [i915]
[  403.473567]  [<f8516ed1>] i915_gem_object_pin+0x271/0x660 [i915]
[  403.473622]  [<f851c740>] ? i915_ggtt_clear_range+0x20/0x20 [i915]
[  403.473676]  [<f8517afa>] i915_gem_object_pin_to_display_plane+0xda/0x190 [i915]
[  403.473742]  [<f852d9fa>] intel_pin_and_fence_fb_obj+0xba/0x140 [i915]
[  403.473800]  [<f852db40>] intel_gen7_queue_flip+0x30/0x1c0 [i915]
[  403.473856]  [<f85337b0>] intel_crtc_page_flip+0x1a0/0x320 [i915]
[  403.473911]  [<f847b549>] ? drm_framebuffer_reference+0x39/0x80 [drm]
[  403.473965]  [<f847f9fb>] drm_mode_page_flip_ioctl+0x28b/0x320 [drm]
[  403.474018]  [<f846fec8>] drm_ioctl+0x4b8/0x560 [drm]
[  403.474064]  [<f847f770>] ? drm_mode_gamma_get_ioctl+0xd0/0xd0 [drm]
[  403.474113]  [<c1140f8a>] ? do_sync_read+0x6a/0xa0
[  403.474154]  [<f846fa10>] ? drm_copy_field+0x80/0x80 [drm]
[  403.474193]  [<c115134c>] do_vfs_ioctl+0x7c/0x5b0
[  403.474228]  [<c1141d2f>] ? vfs_read+0xef/0x160
[  403.474263]  [<c108dcbb>] ? ktime_get_ts+0x4b/0x120
[  403.474298]  [<c1151917>] SyS_ioctl+0x97/0xa0
[  403.474330]  [<c1590bc1>] sysenter_do_call+0x12/0x22
[  403.474364] Code: 55 f4 8b 45 f8 e9 75 ff ff ff 90 55 89 e5 53 83 ec 14 8b 08 8b 50 04 81 f9 00 01 10 00 74 24 81 fa 00 02 20 00 0f 84 8e 00 00 00 <8b> 1a 39 d8 75 62 8b 59 04 39 d8 75 35 89 51 04 89 0a 83 c4 14
[  403.474566] EIP: [<c12c1500>] __list_del_entry+0x20/0xe0 SS:ESP 0068:ee6a3c48
[  403.476513] CR2: 000000006b6b6b6b

v2: Missed the drm_object_unreference use after free (Ville)
Daniel Vetter <daniel@ffwll.ch> writes:

Reported-by: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add the Oops from Mika to the commit message.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: prepare bind_to_vm for preallocated vma
Ben Widawsky [Wed, 14 Aug 2013 09:38:35 +0000 (11:38 +0200)]
drm/i915: prepare bind_to_vm for preallocated vma

In the new execbuf code we want to track buffers using the vmas even
before they're all properly mapped. Which means that bind_to_vm needs
to deal with buffers which have preallocated vmas which aren't yet
bound.

This patch implements this prep work and adjusts our WARN/BUG checks.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Split out from Ben's big execbuf patch. Also move one BUG
back to its original place to deflate the diff a notch.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Switch eviction code to use vmas
Ben Widawsky [Wed, 14 Aug 2013 09:38:34 +0000 (11:38 +0200)]
drm/i915: Switch eviction code to use vmas

The execbuf wants to do relocations usings vmas, so we need a
vma->exec_list. The eviction code also uses the old obj execbuf list
for it's own book-keeping, but would really prefer to deal in vmas
only. So switch it over to the new list.

Again this is just a prep patch for the big execbuf vma conversion.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Split out from Ben's big execbuf vma patch.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: s/obj->exec_list/obj->obj_exec_link in debugfs
Ben Widawsky [Wed, 14 Aug 2013 09:38:33 +0000 (11:38 +0200)]
drm/i915: s/obj->exec_list/obj->obj_exec_link in debugfs

To convert the execbuf code over to use vmas natively we need to
shuffle the exec_list a bit. This patch here just prepares things with
the debugfs code, which also uses the old exec_list list_head, newly
called obj_exec_link.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Split out from Ben's big patch.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agoi915: Add a Kconfig option to turn on i915.preliminary_hw_support by default
Josh Triplett [Tue, 13 Aug 2013 23:23:17 +0000 (16:23 -0700)]
i915: Add a Kconfig option to turn on i915.preliminary_hw_support by default

When building kernels for a preliminary hardware target, having to add a
kernel command-line option can prove inconvenient.  Add a Kconfig option
that changes the default of this option to 1.

Signed-off-by: Josh Triplett <josh@joshtriplett.org>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
[danvet: Pimp the Kconfig help text a bit as suggested by Damien in
his 2nd review.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: enable the power well before module unload
Paulo Zanoni [Wed, 14 Aug 2013 17:40:37 +0000 (14:40 -0300)]
drm/i915: enable the power well before module unload

Our driver initialization doesn't seem to be ready to load when the
power well is disabled: we hit a few "Unclaimed register" messages. So
do just like we already do for the suspend/resume path: enable the
power well before unloading.

At some point we'll want to be able to survive suspend/resume and
load/unload with the power well disabled, but for now let's just fix
the regression.

Regression introduced by the following commit:

commit bf51d5e2cda5d36d98e4b46ac7fca9461e512c41
Author: Paulo Zanoni <paulo.r.zanoni@intel.com>
Date:   Wed Jul 3 17:12:13 2013 -0300
    drm/i915: switch disable_power_well default value to 1

Bug can be reproduced by running the "module_reload" script from
intel-gpu-tools.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=67813
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: explicit store base gem object in dma_buf->priv
Daniel Vetter [Thu, 8 Aug 2013 07:10:38 +0000 (09:10 +0200)]
drm/i915: explicit store base gem object in dma_buf->priv

Makes it more obviously correct what tricks we play by reusing the drm
prime release helper.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: unpin backing storage in dmabuf_unmap
Daniel Vetter [Thu, 8 Aug 2013 07:10:37 +0000 (09:10 +0200)]
drm/i915: unpin backing storage in dmabuf_unmap

This fixes a WARN in i915_gem_free_object when the
obj->pages_pin_count isn't 0.

v2: Add locking to unmap, noticed by Chris Wilson. Note that even
though we call unmap with our own dev->struct_mutex held that won't
result in an immediate deadlock since we never go through the dma_buf
interfaces for our own, reimported buffers. But it's still easy to
blow up and anger lockdep, but that's already the case with our ->map
implementation. Fixing this for real will involve per dma-buf ww mutex
locking by the callers. And lots of fun. So go with the duct-tape
approach for now.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reported-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Cc: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Tested-by: Armin K. <krejzi@email.com> (v1)
Acked-by: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Only unmask required PM interrupts
Vinit Azad [Wed, 14 Aug 2013 20:34:33 +0000 (13:34 -0700)]
drm/i915: Only unmask required PM interrupts

Un-masking all PM interrupts causes hardware to generate
interrupts regardless of whether the interrupts are enabled
on the DE side. Since turbo only need up/down threshold and
rc6 timeout interrupt, mask all other interrupts bits to avoid
unnecessary overhead/wake up.

Note that our interrupt handler isn't being fired since we do set the
IER bits properly (IIR bits aren't set). The overhead isn't because
our driver is reacting to these interrupts, but because hardware keeps
generating internal messages when PMINTRMSK doesn't mask out the
up/down EI interrupts (which happen periodically).

Change-Id: I6c947df6fd5f60584d39b9e8b8c89faa51a5e827
Signed-off-by: Vinit Azad <vinit.azad@intel.com>
[danvet: Add follow-up explanation of the precise effects from Vinit
as a note to the commit message.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: clarify Haswell power well bit names
Paulo Zanoni [Fri, 2 Aug 2013 19:22:25 +0000 (16:22 -0300)]
drm/i915: clarify Haswell power well bit names

Whenever I need to work with the HSW_PWER_WELL_* register bits I have
to look at the documentation to find out which bit is to request the
power well and which one shows its current state. Rename the bits so I
won't need to look the docs every time.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: check the power well when redisabling VGA
Paulo Zanoni [Fri, 2 Aug 2013 19:22:24 +0000 (16:22 -0300)]
drm/i915: check the power well when redisabling VGA

If the power well is disabled VGA is guaranteed to be disabled.

This fixes unclaimed register messages that happen on suspend/resume.

v2: Check the actual hw power well state instead of our own tracking
to make sure VGA is _really_ off (in case the BIOS/KVMr has just its
own request bit set). Requested by Ville.

Note: Ville suggested whether it wouldn't be better to just enable the
power well over a slightly longer time in our resume code, since we
already do that. I tend to agree, but there's also the modeset force
code in the lid notifier which _also_ eventually calls redisable_vga.
We shouldn't ever need this on somewhat modern hw (everything with
opregion essentially) but the code to bail out isn't there. Hence
stick with this simple approach here for now.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=67517
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
[danvet: Summarize the discussion around the resume sequence and lid
notifier a bit.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Drop the overzealous warning from i915_gem_set_cache_level
Chris Wilson [Mon, 12 Aug 2013 10:46:17 +0000 (11:46 +0100)]
drm/i915: Drop the overzealous warning from i915_gem_set_cache_level

By our earlier reckoning, move from a snooped/llc setting to an uncached
setting, leaves the CPU cache in a consistent state irrespective of our
domain tracking - so we can forgo the warning about the lack of
invalidation. Similarly for any writes posted to the snooped CPU domain,
we know will be safely clflushed to the uncached PTEs after forcing the
domain change.

This WARN started to pop up with

commit d46f1c3f1372e3a72fab97c60480aa4a1084387f
Author:     Chris Wilson <chris@chris-wilson.co.uk>
AuthorDate: Thu Aug 8 14:41:06 2013 +0100

    drm/i915: Allow the GPU to cache stolen memory

Ville brought up a scenario where the interaction of a set_caching
ioctl call from userspace on a scanout buffer (i.e. obj->pin_display
is set) resulted in the code getting confused and not properly
flushing stale cpu cachelines. Luckily we already prevent this by
rejecting caching changes when obj->pin_count is set.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=68040
Tested-by: cancan,feng <cancan.feng@intel.com>
[danvet: Add buglink, bisect result and explain why Ville's scenario
is already taken care of.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm: WARN when removing unallocated node
Ben Widawsky [Wed, 14 Aug 2013 01:09:08 +0000 (18:09 -0700)]
drm: WARN when removing unallocated node

The conditional is usually a recoverable driver bug, and so WARNing, and
preventing the drm_mm code from doing potential damage (BUG) is
desirable.

This issue was hit and fixed twice while developing the i915 multiple
address space code. The first fix is the patch just before this, and is
hit on an not frequently occuring error path. Another was fixed during
patch iteration, so it's hard to see from the patch:

commit c6cfb325677ea6305fb19acf3a4d14ea267f923e
Author: Ben Widawsky <ben@bwidawsk.net>
Date:   Fri Jul 5 14:41:06 2013 -0700

    drm/i915: Embed drm_mm_node in i915 gem obj

From the intel-gfx mailing list, we discussed this:
References: <20130705191235.GA3057@bwidawsk.net>

Cc: Dave Airlie <airlied@redhat.com>
CC: <dri-devel@lists.freedesktop.org>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Acked-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: use vma->node directly and rewrap map&fence in bind
Daniel Vetter [Wed, 14 Aug 2013 08:21:23 +0000 (10:21 +0200)]
drm/i915: use vma->node directly and rewrap map&fence in bind

Use () to make for neater alignment of the split lines, too. With this
we ditch another jump through the obj_gtt_size/offset indirection
maze.

Cc: Ben Widawsky <benjamin.widawsky@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: cleanup map&fence in bind
Ben Widawsky [Wed, 14 Aug 2013 01:09:07 +0000 (18:09 -0700)]
drm/i915: cleanup map&fence in bind

Cleanup the map and fenceable setting during bind to make more sense,
and not check i915_is_ggtt() 2 unnecessary times

v2: Move the bools into the if block (Chris) - There are ways to tidy
this function (fence calculations for instance) even further, but they
are quite invasive, so I am punting on those unless specifically asked.

v3: Add newline between variable declaration and logic (Chris)

Recommended-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove node only when allocated
Ben Widawsky [Wed, 14 Aug 2013 01:09:06 +0000 (18:09 -0700)]
drm/i915: Remove node only when allocated

VMAs can be created and not bound. One may think of it as lazy cleanup,
and safely gloss over the conditions which manufacture it. In either
case, when the object backing the i915 vma is destroyed, we must cleanup
the vma without stumbling into a bunch of pitfalls that assume the vma
is bound.

NOTE: I was pretty certain the above condition could only happen when we
introduced the use of VMAs being looked up at execbuf, and already
existing. Paulo has hit this though, so I must be missing something. As
I believe the patch is correct anyway, therefore I won't scratch my head
too hard.

v2: use goto destroy as a compromise (Chris)

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: clarify error paths in create_stolen_for_preallocated
Daniel Vetter [Wed, 14 Aug 2013 08:01:32 +0000 (10:01 +0200)]
drm/i915: clarify error paths in create_stolen_for_preallocated

Use the standard inversely ordered goto label stack for everything.
Spotted while reviewing place where we might need to to call
vma_destroy but failed to do so.

Cc: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Get VECS semaphore info on error
Ben Widawsky [Mon, 12 Aug 2013 23:53:04 +0000 (16:53 -0700)]
drm/i915: Get VECS semaphore info on error

Ideally we could use for_each_ring with the ring flags as I've done a
couple times
(http://lists.freedesktop.org/archives/intel-gfx/2013-June/029450.html).
Until Daniel merges that patch though, we can just use this.

Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Initialize seqno for VECS too
Ben Widawsky [Mon, 12 Aug 2013 23:53:03 +0000 (16:53 -0700)]
drm/i915: Initialize seqno for VECS too

We require n-1 mailboxes for proper semaphore synchronization. All
semaphore synchronization code relies on proper values in these
mailboxes. The fact that we failed to touch the vebox ring by itself
was unlikely to be an issue since the HW should be initializing the
values to 0. However the error framework for testing seqno wrap
introduced by Mika, in addition to the hangcheck via seqno, and
i915_error_first_batchbuffer() combined caused a nice explosion.

The problem is caused by seqno wrap because the wrap condition is not
properly setup. The wrap code attempts to set the sync mailboxes all
to 0, and then set the current seqno to one less than 0. In all cases,
the vebox mailbox wasn't properly being initialized. This caused a
wrap to not occur. When hangcheck kicks in with the bogus seqno
values, the rest just doesn't work. It makes me wonder if we shouldn't
consider a dumber version of hangcheck...

How we messed this up: VECS support was written before the
aforementioned other features. Upon VECS being rebased, these facts
were missed.

Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=65387
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=67198
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: tune the RC6 threshold for stability
Stéphane Marchesin [Tue, 13 Aug 2013 18:55:17 +0000 (11:55 -0700)]
drm/i915: tune the RC6 threshold for stability

It's basically the same deal as the RC6+ issues on ivy bridge
except this time with RC6 on sandy bridge. Like last time the
core of the issue is that the timings don't work 100% with our
voltage regulator. So from time to time, the kernel will print
a warning message about the GPU not getting out of RC6. In
particular, I found this fairly easy to reproduce during
suspend/resume.

Changing the threshold to 125000 instead of 50000 seems to fix
the issue. The previous patch used 150000 but as it turns out
this doesn't work everywhere. After getting such a machine, I
bisected the highest value which works, which is 125000, so here
it is.

I also measured the idle power usage before/after this patch and
didn't see a difference on a sandy bridge laptop. On haswell and
up, it makes a big difference, so we want to keep it at 50k
there. It also seems like haswell doesn't have the RC6 issues
that sandy bridge has so the 50k value is fine.

Signed-off-by: Stéphane Marchesin <marcheu@chromium.org>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: print a message when we detect an early Haswell SDV
Paulo Zanoni [Mon, 12 Aug 2013 17:34:08 +0000 (14:34 -0300)]
drm/i915: print a message when we detect an early Haswell SDV

The machines that fall in this category are the SDVs that have a PCI
ID starting with 0x0C. These are very early pre-production machines
and may not fully work. Other Haswell SDVs have PCI IDs that match the
real Haswell machines and we expect them to work better.

Even though they have problems, they still mostly work so I don't see
a reason to refuse loading our driver. But I do see a reason to reject
bug reports from these machines, so the message should help the bug
triagers.

As far as I know, we don't implement some workarounds that are
specific to these machines and suspend/resume may not work on most of
them, but besides this, they may work.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=61508
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Print the changes required for modeset
Chris Wilson [Tue, 13 Aug 2013 17:48:47 +0000 (18:48 +0100)]
drm/i915: Print the changes required for modeset

After computing the stage changes for the set_config, record those in
the debug log.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: remove set but unused variables
Paulo Zanoni [Mon, 12 Aug 2013 17:56:53 +0000 (14:56 -0300)]
drm/i915: remove set but unused variables

Caught by "make W=1 drivers/gpu/drm/i915/".

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Allow the user to set bo into the DISPLAY cache domain
Chris Wilson [Thu, 8 Aug 2013 13:41:11 +0000 (14:41 +0100)]
drm/i915: Allow the user to set bo into the DISPLAY cache domain

This is primarily for the benefit of the create2 ioctl so that the
caller can avoid the later step of rebinding the bo with new PTE bits.
After introducing WT (and possibly GFDT) cacheing for display targets,
not everything in the display is earmarked as UC, and more importantly
what is is controlled by the kernel.

Note that set_cache_level/get_cache_level for DISPLAY is not necessarily
idempotent; get_cache_level may return UC for architectures that have no
special cache domain for the display engine.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Use Write-Through cacheing for the display plane on Iris
Chris Wilson [Thu, 8 Aug 2013 13:41:10 +0000 (14:41 +0100)]
drm/i915: Use Write-Through cacheing for the display plane on Iris

Haswell GT3e has the unique feature of supporting Write-Through cacheing
of objects within the eLLC/LLC. The purpose of this is to enable the display
plane to remain coherent whilst objects lie resident in the eLLC/LLC - so
that we, in theory, get the best of both worlds, perfect display and fast
access.

However, we still need to be careful as the CPU does not see the WT when
accessing the cache. In particular, this means that we need to flush the
cache lines after writing to an object through the CPU, and on
transitioning from a cached state to WT.

v2: Actually do the clflush on transition to WT, nagging by Ville.
v3: Flush the CPU cache after writes into WT objects.
v4: Rease onto LLC updates and report WT as "uncached" for
get_cache_level_ioctl to remain symmetric with set_cache_level_ioctl.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: drop unnecessary local variable to suppress build warning
Jani Nikula [Sun, 11 Aug 2013 09:44:02 +0000 (12:44 +0300)]
drm/i915: drop unnecessary local variable to suppress build warning

Although I could not reproduce this (different compiler version,
perhaps), reportedly we get:

drivers/gpu/drm/i915/i915_irq.c:1943:27: warning: ‘score’ may be used
uninitialized in this function [-Wuninitialized]

Drop the 'score' variable altogether as it's not really needed.

Reported-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: give more distinctive names to ring hangcheck action enums
Jani Nikula [Sun, 11 Aug 2013 09:44:01 +0000 (12:44 +0300)]
drm/i915: give more distinctive names to ring hangcheck action enums

The short lowercase names are bound to collide. The default warnings
don't even warn about shadowing.

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: remove unused leftover variable irq_received
Jani Nikula [Sun, 11 Aug 2013 09:44:26 +0000 (12:44 +0300)]
drm/i915: remove unused leftover variable irq_received

It's been there since i8xx_irq_handler() was added in
commit c2798b19bac2538393fc932bfbe59807a4734b3e
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Sun Apr 22 21:13:57 2012 +0100

    drm/i915: i8xx interrupt handler

Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agoi915: Fix SDVO potentially turning off randomly
Guillaume Clement [Sat, 10 Aug 2013 19:57:57 +0000 (21:57 +0200)]
i915: Fix SDVO potentially turning off randomly

Some Poulsbo cards seem to incorrectly report
SDVO_CMD_STATUS_TARGET_NOT_SPECIFIED instead of
SDVO_CMD_STATUS_PENDING, which causes the display to be turned off.

This could also happen to i915.

Signed-off-by: Guillaume Clement <gclement@baobob.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: WARN_ON failed map_and_fenceable
Ben Widawsky [Sat, 10 Aug 2013 05:12:12 +0000 (22:12 -0700)]
drm/i915: WARN_ON failed map_and_fenceable

I just noticed in our code we don't really check the assertion, and
given some of the code I am changing in this area, I feel a WARN is very
nice to have.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: s/&/&&/ to fix typo on the check.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: reserve I915_CACHING_DISPLAY and document cache modes
Daniel Vetter [Sat, 10 Aug 2013 12:51:11 +0000 (14:51 +0200)]
drm/i915: reserve I915_CACHING_DISPLAY and document cache modes

Resolve the catch-22 of igt needing a stable number and patches first
needing testcases by reserving the interface number up-front.

v2: Improve the spelling a bit.

v3: More spelling fail spotted by Chris.

Requested-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Only do a chipset flush after a clflush
Chris Wilson [Thu, 8 Aug 2013 13:41:09 +0000 (14:41 +0100)]
drm/i915: Only do a chipset flush after a clflush

Now that we skip clflushes more often, return a boolean indicating
whether the clflush was actually performed, and only if it was do the
chipset flush. (Though on most of the architectures where the clflush will
be skipped, the chipset flush is a no-op!)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Allow the GPU to cache stolen memory
Chris Wilson [Thu, 8 Aug 2013 13:41:06 +0000 (14:41 +0100)]
drm/i915: Allow the GPU to cache stolen memory

As a corollary to reviewing the interaction between LLC and our cache
domains, the GPU PTE bits are independent of the CPU PAT bits. As such
we can set the cache level on stolen memory based on how we wish the GPU
to cache accesses to it. So we are free to set the same default cache
levels as for normal bo, i.e. enable LLC cacheing by default where
appropriate.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Update rules for writing through the LLC with the cpu
Chris Wilson [Fri, 9 Aug 2013 11:26:45 +0000 (12:26 +0100)]
drm/i915: Update rules for writing through the LLC with the cpu

As mentioned in the previous commit, reads and writes from both the CPU
and GPU go through the LLC. This gives us coherency between the CPU and
GPU irrespective of the attribute settings either device sets. We can
use to avoid having to clflush even uncached memory.

Except for the scanout.

The scanout resides within another functional block that does not use
the LLC but reads directly from main memory. So in order to maintain
coherency with the scanout, writes to uncached memory must be flushed.
In order to optimize writes elsewhere, we start tracking whether an
framebuffer is attached to an object.

v2: Use pin_display tracking rather than fb_count (to ensure we flush
cursors as well etc) and only force the clflush along explicit writes to
the scanout paths (i.e. pin_to_display_plane and pwrite into scanout).

v3: Force the flush after hitting the slowpath in pwrite, as after
dropping the lock the object's cache domain may be invalidated. (Ville)

Based on a patch by Ville Syrjälä.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Track when an object is pinned for use by the display engine
Chris Wilson [Fri, 9 Aug 2013 11:25:09 +0000 (12:25 +0100)]
drm/i915: Track when an object is pinned for use by the display engine

The display engine has unique coherency rules such that it requires
special handling to ensure that all writes to cursors, scanouts and
sprites are clflushed. This patch introduces the infrastructure to
simply track when an object is being accessed by the display engine.

v2: Explain the is_pin_display() magic as the sources for obj->pin_count
and their individual rules is not obvious. (Ville)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Update rules for reading cache lines through the LLC
Chris Wilson [Thu, 8 Aug 2013 13:41:03 +0000 (14:41 +0100)]
drm/i915: Update rules for reading cache lines through the LLC

The LLC is a fun device. The cache is a distinct functional block within
the SA that arbitrates access from both the CPU and GPU cores. As such
all writes to memory land first in the LLC before further action is
taken. For example, an uncached write from either the CPU or GPU will
then proceed to memory and evict the cacheline from the LLC. This means that
a read from the LLC always returns the correct information even if the PTE
bit in the GPU differs from the PAT bit in the CPU. For the older
snooping architecture on non-LLC, the fundamental principle still holds
except that some coordination is required between the CPU and GPU to
explicitly perform the snooping (which is handled by our request
tracking).

The upshot of this is that we know that we can issue a read from either
LLC devices or snoopable memory and trust the contents of the cache -
i.e. we can forgo a clflush before a read in these circumstances.
Writing to memory from the CPU is a little more tricky as we have to
consider that the scanout does not read from the CPU cache at all, but
from main memory. So we have to currently treat all requests to write to
uncached memory as having to be flushed to main memory for coherency
with all consumers.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Fix FB WM for HSW
Ville Syrjälä [Fri, 9 Aug 2013 15:02:09 +0000 (18:02 +0300)]
drm/i915: Fix FB WM for HSW

Due to a misplaced memset(), we never actually enabled the FBC WM on HSW.
Move the memset() to happen a bit earlier, so that it won't clobber
results->enable_fbc_wm.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: expose HDMI connectors on port C on BYT
Jesse Barnes [Fri, 9 Aug 2013 16:34:35 +0000 (09:34 -0700)]
drm/i915: expose HDMI connectors on port C on BYT

Ryan noticed that on his board, HDMI was wired up to port C but not
exposed by the kernel, which had only expected DP on that port.  Fix
that up by enumerating both ports if possible.

Tested-by: "Matsumura, Ryan" <ryan.matsumura@intel.com>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[danvet: Fix up the whitespace fail. Tsk.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: fix a limit check in hsw_compute_wm_results()
Dan Carpenter [Fri, 9 Aug 2013 10:07:31 +0000 (13:07 +0300)]
drm/i915: fix a limit check in hsw_compute_wm_results()

The '!' here was not intended.  Since '!' has higher precedence than
compare, it means the check is never true.

This regression was introduced in

commit 71fff20ff1bb790f4defe0c880e028581ffab420
Author: Ville Syrjälä <ville.syrjala@linux.intel.com>
Date:   Tue Aug 6 22:24:03 2013 +0300

    drm/i915: Kill fbc_enable from hsw_lp_wm_results

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: unbreak i915_gem_object_ggtt_unbind()
Dan Carpenter [Fri, 9 Aug 2013 09:44:11 +0000 (12:44 +0300)]
drm/i915: unbreak i915_gem_object_ggtt_unbind()

There is an extra semi-colon here so we just leak and never unbind
anything.

This regression has been introduced in

commit 07fe0b12800d4752d729d4122c01f41f80a5ba5a
Author: Ben Widawsky <ben@bwidawsk.net>
Date:   Wed Jul 31 17:00:10 2013 -0700

    drm/i915: plumb VM into bind/unbind code

Cc: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Make intel_set_mode() static
Damien Lespiau [Thu, 8 Aug 2013 21:28:59 +0000 (22:28 +0100)]
drm/i915: Make intel_set_mode() static

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove intel_modeset_disable()
Damien Lespiau [Thu, 8 Aug 2013 21:28:58 +0000 (22:28 +0100)]
drm/i915: Remove intel_modeset_disable()

Caught by the dead code police!

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Make intel_encoder_dpms() static
Damien Lespiau [Thu, 8 Aug 2013 21:28:57 +0000 (22:28 +0100)]
drm/i915: Make intel_encoder_dpms() static

And also fix a small typo in the intel_encoder_dpms() comment.

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Make i915_hangcheck_elapsed() static
Damien Lespiau [Thu, 8 Aug 2013 21:28:56 +0000 (22:28 +0100)]
drm/i915: Make i915_hangcheck_elapsed() static

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Fix #endif comment
Damien Lespiau [Thu, 8 Aug 2013 21:28:55 +0000 (22:28 +0100)]
drm/i915: Fix #endif comment

Did you say OCD?

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove i915_gem_object_check_coherency()
Damien Lespiau [Thu, 8 Aug 2013 21:28:54 +0000 (22:28 +0100)]
drm/i915: Remove i915_gem_object_check_coherency()

This code was dead since:

  commit 432e58edc9de1d9c3d6a7b444b3c455b8f209a7d
  Author: Chris Wilson <chris@chris-wilson.co.uk>
  Date:   Thu Nov 25 19:32:06 2010 +0000

      drm/i915: Avoid allocation for execbuffer object list

so just put it to rest for good.

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove stale prototypes
Damien Lespiau [Thu, 8 Aug 2013 21:28:53 +0000 (22:28 +0100)]
drm/i915: Remove stale prototypes

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: List objects allocated from stolen memory in debugfs
Chris Wilson [Wed, 7 Aug 2013 17:30:54 +0000 (18:30 +0100)]
drm/i915: List objects allocated from stolen memory in debugfs

I was curious as to what objects were currently allocated from stolen
memory, and so exported it from debugfs.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Always call intel_update_sprite_watermarks() when disabling a plane
Ville Syrjälä [Tue, 6 Aug 2013 19:24:12 +0000 (22:24 +0300)]
drm/i915: Always call intel_update_sprite_watermarks() when disabling a plane

ILK and VLV codepaths didn't update sprite watermarks when disabling a
sprite. Make them do that.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Pass plane and crtc to intel_update_sprite_watermarks
Ville Syrjälä [Tue, 6 Aug 2013 19:24:11 +0000 (22:24 +0300)]
drm/i915: Pass plane and crtc to intel_update_sprite_watermarks

We're going to want to know the crtc in the watermark code to avoid
doing more work than we have to. We should also pass the plane we're
disabling so that we know where to stick our watermark parameters
without having to go look the plane up.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Don't try to disable plane if it's already disabled
Ville Syrjälä [Wed, 7 Aug 2013 10:30:23 +0000 (13:30 +0300)]
drm/i915: Don't try to disable plane if it's already disabled

Check plane->fb in intel_disable_plane() to determine if the plane
is already disabled.

If the plane has an fb, then it must also have a crtc, so we can drop
the plane->crtc check and just call intel_enable_primary() directly.

v2: WARN and bail if the plane doesn't have a crtc when it should

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Pass crtc to our update/disable_plane hooks
Ville Syrjälä [Tue, 6 Aug 2013 19:24:09 +0000 (22:24 +0300)]
drm/i915: Pass crtc to our update/disable_plane hooks

We're going to want to know which CRTC we're dealing with, so pass it
down to the update/disable_plane hooks.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Split plane watermark parameters into a separate struct
Ville Syrjälä [Wed, 7 Aug 2013 10:29:50 +0000 (13:29 +0300)]
drm/i915: Split plane watermark parameters into a separate struct

Give a name to the plane watermark related data we have currently
stored under intel_plane->wm.

We also observe that this data is more or less the same that we have
in the hsw_pipe_wm_parameters structure, so use it there as well.

v2: Make pahole happier

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Pull some watermarks state into a separate structure
Ville Syrjälä [Wed, 7 Aug 2013 10:29:12 +0000 (13:29 +0300)]
drm/i915: Pull some watermarks state into a separate structure

There is a bunch of global state that needs to be considered when
checking watermarks for validity. Move most of that to a new
structure intel_wm_config, to avoid having to pass around so
many variables.

One notable thing left out is the DDB partitioning information,
since we often anyway need to check the same watermarks against
both 1/2 and 5/6 DDB partitioning layouts.

v2: s/pipes_active/num_pipes_active

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Calculate max watermark levels for ILK+
Ville Syrjälä [Wed, 7 Aug 2013 10:28:19 +0000 (13:28 +0300)]
drm/i915: Calculate max watermark levels for ILK+

There are quite a few variables we need to take into account to
determine the maximum watermark levels, so it feels a bit cleaner
to calculate those rather than just have a bunch of what look like
magic numbers.

v2: s/pipes_active/num_pipes_active
    s/othwewise/otherwise

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Rename hsw_lp_wm_result to intel_wm_level
Ville Syrjälä [Tue, 6 Aug 2013 19:24:05 +0000 (22:24 +0300)]
drm/i915: Rename hsw_lp_wm_result to intel_wm_level

Let's call hsw_lp_wm_result intel_wm_level from now on and move it to
i915_drv.h for later use.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Pull watermark level validity check out
Ville Syrjälä [Wed, 7 Aug 2013 10:24:47 +0000 (13:24 +0300)]
drm/i915: Pull watermark level validity check out

Refactor the code a bit to split the watermark level validity check into
a separate function.

Also add hack there that allows us to use it even for LP0 watermarks.
ATM we don't pre-compute/check the LP0 watermarks, so we just have to
clamp them to the maximum and hope things work out.

v2: Add some debug prints when we exceed max WM0
    Kill pointless ret = false' assignment.
    Include the check for the already disabled 'result' which
    got shuffled around when the patchs got reorderd

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Add vma to list at creation
Ben Widawsky [Thu, 1 Aug 2013 00:00:16 +0000 (17:00 -0700)]
drm/i915: Add vma to list at creation

With the current code there shouldn't be a distinction - however with an
upcoming change we intend to allocate a vma much earlier, before it's
actually bound anywhere.

To do this we have to check node allocation as well for the _bound()
check.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: move list_del(&vma->vma_link) from vma_unbind to vma_destroy,
again fallout from the loss of "rm/i915: Cleanup more of VMA in
destroy".]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
fixup for drm/i915: Add vma to list at creation

11 years agodrm/i915: Update error capture for VMs
Ben Widawsky [Thu, 1 Aug 2013 00:00:15 +0000 (17:00 -0700)]
drm/i915: Update error capture for VMs

formerly: "drm/i915: Create VMAs (part 4) - Error capture"

Since the active/inactive lists are per VM, we need to modify the error
capture code to be aware of this, and also extend it to capture the
buffers from all the VMs. For now all the code assumes only 1 VM, but it
will become more generic over the next few patches.

NOTE: If the number of VMs in a real world system grows significantly
we'll have to focus on only capturing the guilty VM, or else it's likely
there won't be enough space for error capture.

v2: Squashed in the "part 6" which had dependencies on the mm_list
change. Since I've moved the mm_list change to an earlier point in the
series, we were able to accomplish it here and now.

v3: Rebased over new error capture

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: mm_list is per VMA
Ben Widawsky [Thu, 1 Aug 2013 00:00:14 +0000 (17:00 -0700)]
drm/i915: mm_list is per VMA

formerly: "drm/i915: Create VMAs (part 5) - move mm_list"

The mm_list is used for the active/inactive LRUs. Since those LRUs are
per address space, the link should be per VMx .

Because we'll only ever have 1 VMA before this point, it's not incorrect
to defer this change until this point in the patch series, and doing it
here makes the change much easier to understand.

Shamelessly manipulated out of Daniel:
"active/inactive stuff is used by eviction when we run out of address
space, so needs to be per-vma and per-address space. Bound/unbound otoh
is used by the shrinker which only cares about the amount of memory used
and not one bit about in which address space this memory is all used in.
Of course to actual kick out an object we need to unbind it from every
address space, but for that we have the per-object list of vmas."

v2: only bump GGTT LRU in i915_gem_object_set_to_gtt_domain (Chris)

v3: Moved earlier in the series

v4: Add dropped message from v3

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Frob patch to apply and use vma->node.size directly as
discused with Ben. Also drop a needles BUG_ON before move_to_inactive,
the function itself has the same check.]
[danvet 2nd: Rebase on top of the lost "drm/i915: Cleanup more of VMA
in destroy", specifically unlink the vma from the mm_list in
vma_unbind (to keep it symmetric with bind_to_vm) instead of
vma_destroy.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Fix up map and fenceable for VMA
Ben Widawsky [Thu, 1 Aug 2013 00:00:13 +0000 (17:00 -0700)]
drm/i915: Fix up map and fenceable for VMA

formerly: "drm/i915: Create VMAs (part 3.5) - map and fenceable
tracking"

The map_and_fenceable tracking is per object. GTT mapping, and fences
only apply to global GTT. As such,  object operations which are not
performed on the global GTT should not effect mappable or fenceable
characteristics.

Functionally, this commit could very well be squashed in to a previous
patch which updated object operations to take a VM argument.  This
commit is split out because it's a bit tricky (or at least it was for
me).

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Drop the bogus hunk in i915_vma_unbind as discussed with
Ben.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Silence a sparse warning
Ville Syrjälä [Wed, 7 Aug 2013 12:11:52 +0000 (15:11 +0300)]
drm/i915: Silence a sparse warning

drivers/gpu/drm/i915/i915_debugfs.c:2136:3: warning: symbol
'i915_debugfs_files' was not declared. Should it be static?

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Rename hsw_data_buf_partitioning to intel_ddb_partitioning
Ville Syrjälä [Tue, 6 Aug 2013 19:24:04 +0000 (22:24 +0300)]
drm/i915: Rename hsw_data_buf_partitioning to intel_ddb_partitioning

We're going to use the 1/2 vs. 5/6 split option already on IVB so the
HSW name is not proper. Just give it an intel_ prefix and move it to
i915_drv.h so that we can use it there later.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Kill fbc_enable from hsw_lp_wm_results
Ville Syrjälä [Tue, 6 Aug 2013 19:24:03 +0000 (22:24 +0300)]
drm/i915: Kill fbc_enable from hsw_lp_wm_results

We don't need to store the FBC WM enabled status in each watermark
level. We anyway have to reduce it down to a single boolean, so just
delay checking the FBC WM limit until we're computing the final
value.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Split watermark level computation from the code
Ville Syrjälä [Tue, 6 Aug 2013 19:24:02 +0000 (22:24 +0300)]
drm/i915: Split watermark level computation from the code

Refactor the watermarks computation for one level to a separate
function. This function will now set the ->enable flag to true,
even if the watermark level wasn't actually checked yet. In the
future we will delay the checking so we must consider all unchecked
watermarks as possibly valid.

v2: Preserve comment about latency units

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Use 'enabled' instead of 'enable' consistently in sprite WM code
Ville Syrjälä [Tue, 6 Aug 2013 19:24:00 +0000 (22:24 +0300)]
drm/i915: Use 'enabled' instead of 'enable' consistently in sprite WM code

Let's be consistent and always call our variables 'enabled' insted of
the occasional 'enable'.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
[danvet: Spelling fix in the commit message, spotted by Chris.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915/hmdi: Rename set_infoframe() to write_infoframe()
Damien Lespiau [Tue, 6 Aug 2013 19:32:24 +0000 (20:32 +0100)]
drm/i915/hmdi: Rename set_infoframe() to write_infoframe()

set_frame() wraps the write_frame() vfunc. Be consistent and name the
wrapping function like the vfunc being called.

It's doubly confusing as we also have a set_infoframes() vfunc and
set_infoframe() doesn't wrap it.

Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm: Set aspect ratio fields in the AVI infoframe even for non CEA modes
Damien Lespiau [Tue, 6 Aug 2013 19:32:23 +0000 (20:32 +0100)]
drm: Set aspect ratio fields in the AVI infoframe even for non CEA modes

I cannot find any evidence what we shouldn't try to set those fields
when setting a non-CEA mode on an HDMI sink. So just kill that return.

Suggested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Acked-by: Dave Airlie <airlied@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm: Handle the DBLCLK flag in the common infoframe helper
Damien Lespiau [Tue, 6 Aug 2013 19:32:22 +0000 (20:32 +0100)]
drm: Handle the DBLCLK flag in the common infoframe helper

Suggested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Acked-by: Dave Airlie <airlied@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915: Remove the now obsolete infoframe definitions
Damien Lespiau [Tue, 6 Aug 2013 19:32:21 +0000 (20:32 +0100)]
drm/i915: Remove the now obsolete infoframe definitions

All the HDMI infoframe code has been ported to use video/hdmi.c, so it's
time to say bye bye to this code.

Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915/sdvo: Port the infoframe code to the shared infrastructure
Damien Lespiau [Tue, 6 Aug 2013 19:32:20 +0000 (20:32 +0100)]
drm/i915/sdvo: Port the infoframe code to the shared infrastructure

Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
Signed-off-by: Thierry Reding <thierry.reding at avionic-design.de>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915/hdmi: Port the infoframe code to the common hdmi helpers
Damien Lespiau [Tue, 6 Aug 2013 19:32:19 +0000 (20:32 +0100)]
drm/i915/hdmi: Port the infoframe code to the common hdmi helpers

Let's use the drivers/video/hmdi.c and drm infoframe helpers to build
our infoframes.

v2: Simplify the logic to compute the buffer size. We can just take the
maximum infoframe size rounded to 32, which happens to be what the
hardware let us write anyway.

v3: Remove unnecessary memset() (Ville Syrjälä)

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm/i915/hdmi: Change the write_infoframe vfunc to take a buffer and a type
Damien Lespiau [Tue, 6 Aug 2013 19:32:18 +0000 (20:32 +0100)]
drm/i915/hdmi: Change the write_infoframe vfunc to take a buffer and a type

First step in the move to the shared infoframe infrastructure, let's
move the different infoframe helpers and the write_infoframe() vfunc to
a type (enum hdmi_infoframe_type) and a buffer + len instead of using
our struct dip_infoframe.

v2: constify the infoframe pointer and don't mix signs (Ville Syrjälä)

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni at intel.com>
Signed-off-by: Thierry Reding <thierry.reding at avionic-design.de>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agodrm: Don't generate invalid AVI infoframes for CEA modes
Damien Lespiau [Tue, 6 Aug 2013 19:32:17 +0000 (20:32 +0100)]
drm: Don't generate invalid AVI infoframes for CEA modes

From CEA-861:

  Data Byte 1, bit A0 indicates whether Active Format Data is present in
  Data Byte 2 bits R3 through R0. A source device shall set A0=1 when
  any of the AFD bits are set.

ie. if we want to set active_aspect, we need to set the
active_info_valid bit to 1 as well.

Cc: Thierry Reding <thierry.reding@avionic-design.de>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Acked-by: Dave Airlie <airlied@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
11 years agovideo/hmdi: Clear the whole incoming buffer, not just the infoframe size
Damien Lespiau [Tue, 6 Aug 2013 19:32:16 +0000 (20:32 +0100)]
video/hmdi: Clear the whole incoming buffer, not just the infoframe size

If the user if this API is providing a bigger buffer than the infoframe
size, it could be for a could reason. For instance it could be because
it gives the buffer that will be written to the hardware, up to the
maximum of an infoframe size.

Instead of just zeroing up to the infoframe size, let's zero the whole
incoming buffer as those extra bytes are also used to compute the
ECC and need to be 0.

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Acked-by: Dave Airlie <airlied@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>