Commit 9ee32fea5f unconditionally prevents the CPU from entering idle states
until intel_dp_aux_ch completes for the first time, which never happens on my
DisplayPort-less intel gfx, causing the CPU to get rather hot.
Signed-off-by: Tomas Janousek <tomi@nomi.cz>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
More specifically, the LPT FDI RX only supports 8bpc and a maximum of
2 lanes, so anything above that won't work and should be rejected.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We were previously doing exactly what the "mode set sequence for CRT"
document mandates, but whenever we failed to train the link in the
first tentative, all the other subsequent retries always failed. In
one of my monitors that has 47 modes, I was usually getting around 3
failures when running "testdisplay -a".
After this patch, even if we fail in the first tentative, we can
succeed in the next ones. So now when running "testdisplay -a" I see
around 3 times the message "FDI link training done on step 1" and no
failures.
Notice that now the "retry" code looks a lot like the DP retry code.
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
On a machine with bit17 swizzling, we need to store the bit17 of the
physical page address in put-pages. This requires a memory allocation,
on average less than a page, which may be difficult to satisfy is the
request to put-pages is on behalf of the shrinker. We could allow that
allocation to pull from the reserved memory pools, but it seems much
safer to preallocate the array for tiled objects on affected machines.
v2: Export i915_gem_object_needs_bit17_swizzle() for reuse.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Both the dp and fdi code use the exact same computations (ignore minor
differences in conversion between bits and bytes).
This makes it even more apparent that we have a _massive_ mess between
cpu transcoder/fdi link/pch transcoder and pch link settings. And also
that we have hilarious amounts of confusion between edp and dp
(despite that they're identical at a link level).
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This has originally been added in
commit 8db9d77b1b
Author: Zhenyu Wang <zhenyuw@linux.intel.com>
Date: Wed Apr 7 16:15:54 2010 +0800
drm/i915: Support for Cougarpoint PCH display pipeline
probably to combat issues with hw state left behind by the BIOS. And
indeed, I've checked out that specific revision, and there is no DP
support yet. So the pch dp transcoder won't be correctly disabled, and
that's important since it requires a rether special disable dance:
Just writing 0 to TRANS_DP_CTL won't cut it, since we need to select
the NONE port when disabling, too.
And indeed, things seem to still work, so let's just remove this.
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This could have happened with the old crtc helper based modeset code,
but can't happen any longer with the new code.
Hence put in a WARN and adjust the comment. If no one hits this, we
can eventually remove it (like a few other such cases across our
code).
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
17 ms is eerily close to 60 Hz ^-1
Unfortunately this goes back to the original DP enabling for ilk, and
unfortunately does not come with a reason for it's existance attached.
Some closer inspection of the code and DP specs shows that we set the
idle link pattern before we disable the port. And it seems like that
the DP spec (or at least our hw) only switch to the idle pattern on
the next vblank. Hence a vblank wait at this spot makes _much_ more
sense than a really long wait.
v2: Rebase fixup.
v3: Add comment requested by Paulo Zanoni saying that we don't really
know what this wait is for.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
While reading docs I've noticed that this special workaround to select
the 1.6 GHz DP clock only applies to pre-production ilk machines.
Since the registers we're touching here are rather undocumented and
might be harmful on later chips, rip it out.
For the Bspec reference of this w/a look in "vol4g CPU Display
Registers [DevILK]", Section 4.1.7.1 "DP_A—DisplayPort A
Control Register", "DP_PLL_Frequency_Select".
v2: Keep a debug message as a hint in case something regresses.
Requested by Chris Wilson.
Reviewed-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>
Now that we enable the cpu edp pll in intel_dp->pre_enable and no
longer in crtc_mode_set, we can also move the modeset part to the
intel_dp->mode_set callback. Previously this was not possible because
the encoder ->mode_set callbacks are called after the crtc mode set
callback.
v2: Rebase on top of copy&pasted hsw crtc_mode_set.
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Especially getting rid of all things lvds is ... great!
v2: Drop the two additional pre-hsw hunks noticed by Paulo Zanoni.
v3:
- handle DP ports correctly (spoted by Paulo)
- don't leave {} behind for a single-line block (again spotted by
Paulo)
- kill another if (IBX || CPT) block
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Before queuing the flip but crucially after attaching the unpin-work to
the crtc, we continue to setup the unpin-work. However, should the
hardware fire early, we see the connected unpin-work and queue the task.
The task then promptly runs and unpins the fb before we finish taking
the required references or even pinning it... Havoc.
To close the race, we use the flip-pending atomic to indicate when the
flip is finally setup and enqueued. So during the flip-done processing,
we can check more accurately whether the flip was expected.
v2: Add the appropriate mb() to ensure that the writes to the page-flip
worker are complete prior to marking it active and emitting the MI_FLIP.
On the read side, the mb should be enforced by the spinlocks.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@vger.kernel.org
[danvet: Review the barriers a bit, we need a write barrier both
before and after updating ->pending. Similarly we need a read barrier
in the interrupt handler both before and after reading ->pending. With
well-ordered irqs only one barrier in each place should be required,
but since this patch explicitly sets out to combat spurious interrupts
with is staged activation of the unpin work we need to go full-bore on
the barriers, too. Discussed with Chris Wilson on irc and changes
acked by him.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Less clutter in the traces. And in both cases we yell rather loud
into the logs if we time out. Patch suggested by Chris Wilson.
v2: Annotate another I915_READ in dp_aux to be consistent - we filter
out all register io in wait_for and similar loops. Chris also
suggested to mark all dp_aux register access as _NOTRACE, but I think
we should keep all functionally relevant access around, and filter
unneeded bits in userspace after the trace is captured.
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
At least on the platforms that have a dp aux irq and also have it
enabled - vlvhsw should have one, too. But I don't have a machine to
test this on. Judging from docs there's no dp aux interrupt for gm45.
Also, I only have an ivb cpu edp machine, so the dp aux A code for
snb/ilk is untested.
For dpcd probing when nothing is connected it slashes about 5ms of cpu
time (cpu time is now negligible), which agrees with 3 * 5 400 usec
timeouts.
A previous version of this patch increases the time required to go
through the dp_detect cycle (which includes reading the edid) from
around 33 ms to around 40 ms. Experiments indicated that this is
purely due to the irq latency - the hw doesn't allow us to queue up
dp aux transactions and hence irq latency directly affects throughput.
gmbus is much better, there we have a 8 byte buffer, and we get the
irq once another 4 bytes can be queued up.
But by using the pm_qos interface to request the lowest possible cpu
wake-up latency this slowdown completely disappeared.
Since all our output detection logic is single-threaded with the
mode_config mutex right now anyway, I've decide not ot play fancy and
to just reuse the gmbus wait queue. But this would definitely prep the
way to run dp detection on different ports in parallel
v2: Add a timeout for dp aux transfers when using interrupts - the hw
_does_ prevent this with the hw-based 400 usec timeout, but if the
irq somehow doesn't arrive we're screwed. Lesson learned while
developing this ;-)
v3: While at it also convert the busy-loop to wait_for_atomic, so that
we don't run the risk of an infinite loop any more.
v4: Ensure we have the smallest possible irq latency by using the
pm_qos interface.
v5: Add a comment to the code to explain why we frob pm_qos. Suggested
by Chris Wilson.
v6: Disable dp irq for vlv, that's easier than trying to get at docs
and hw.
v7: Squash in a fix for Haswell that Paulo Zanoni tracked down - the
dp aux registers aren't at a fixed offset any more, but can be on the
PCH while the DP port is on the cpu die.
Reviewed-by: Imre Deak <imre.deak@intel.com> (v6)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
GMBUS_ACTIVE has inverted sense and so doesn't fit into the
wait_hw_status helper, hence create a new gmbus_wait_idle functions.
Also, we only care about the idle irq event and nothing else, which
allows us to use the wait_event_timeout helper directly without
jumping through hoops to catch NAKs.
Since gen2/3 don't have gmbus interrupts, handle them separately with
the old wait_for macro.
This shaves another few ms off reading EDID from a hdmi screen on my
testbox here. EDID reading with interrupt driven gmbus is now as fast
as with busy-looping gmbus at 28 ms here (with negligible cpu
overhead).
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We need two special things to properly wire this up:
- Add another argument to gmbus_wait_hw_status to pass in the
correct interrupt bit in gmbus4.
- Since we can only get an irq for one of the two events we want,
hand-roll the wait_event_timeout code so that we wake up every
jiffie and can check for NAKs. This way we also subsume gmbus
support for platforms without interrupts (or where those are not
yet enabled).
The important bit really is to only enable one gmbus interrupt source
at the same time - with that piece of lore figured out, this seems to
work flawlessly.
Ben Widawsky rightfully complained the lack of measurements for the
claimed benefits (especially since the first version was actually
broken and fell back to bit-banging). Previously reading the 256 byte
hdmi EDID takes about 72 ms here. With this patch it's down to 33 ms.
Given that transfering the 256 bytes over i2c at wire speed takes
20.5ms alone, the reduction in additional overhead is rather nice.
v2: Chris Wilson wondered whether GMBUS4 might contain some set bits
when booting up an hence result in some spurious interrupts. Since we
clear GMBUS4 after every wait and we do gmbus transfer really early in
the setup sequence to detect displays the window is small, but still
be paranoid and clear it properly.
v3: Clarify the comment that gmbus irq generation can only support one
kind of event, why it bothers us and how we work around that limit.
Cc: Daniel Kurtz <djkurtz@chromium.org>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Only enables the interrupt and puts a irq handler into place, doesn't
do anything yet.
Unfortunately there's no gmbus interrupt support for gen2/3 (safe for
pnv, but there the irq is marked as "Test mode").
v2: Wire up the irq handler for vlv and gen4 properly.
v3: i915_enable_pipestat expects the mask bit, not the status bits ... and
for added hilarity those are rather inconsistently named.
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The gmbus interrupt generation is rather fiddly: We can only ever
enable one interrupt source (but we always want to check for NAK
in addition to the real bit). And the bits in the gmbus status
register don't map at all to the bis in the irq register.
To prepare for this mess, start by extracting the hw status wait
loop into it's own function, consolidate the NAK error handling a
bit. To keep things flexible, pass in the status bit we care about
(in addition to any NAK signalling).
v2: I've failed to notice that the sense of GMBUS_ACTIVE is inverted,
Chris Wilson gladly pointed that out for me. To keep things simple,
ignore that case for now (we only need to idle the gmbus controller
at the end of an entire i2c transaction, not after every message).
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Otherwise the new&shiny irq-driven gmbus and dp aux code won't work that
well. Noticed since the dp aux code doesn't have an automatic fallback
with a timeout (since the hw provides for that already).
v2: Simple move drm_irq_install before intel_modeset_gem_init, as
suggested by Ben Widawsky.
v3: Now that interrupts are enabled before all connectors are fully
set up, we might fall over serving a HPD interrupt while things are
still being set up. Instead of jumping through massive hoops and
complicating the code with a separate hpd irq enable step, simply
block out the hotplug work item from doing anything until things are
in place.
v4: Actually, we can enable hotplug processing only after the fbdev is
fully set up, since we call down into the fbdev from the hotplug work
functions. So stick the hpd enabling right next to the poll helper
initialization.
v5: We need to enable irqs before intel_modeset_init, since that
function sets up the outputs.
v6: Fixup cleanup sequence, too.
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
... together with all the other irq related resources in
intel_irq_init. I've managed to oops in the notify_ring function on my
ilk, presumably because of the powerctx setup call to i915_gpu_idle.
Note that this is only a problem with the reorder irq setup sequence
for irq-driver gmbus/dp aux.
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This is for legacy legacy stuff, and checking with the leftover
pipe from the previous loop is propably not what we want. Since
pipe == 2 after the loop ... Then we only assing a variable and do
nothing with it.
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
If there are pre-wrap values in semaphore-mbox registers after wrap,
syncing against some after-wrap request will complete immediately.
Fix this by emitting ring commands to set mbox registers to zero
when the wrap happens.
v2: Use __intel_ring_begin to emit ring commands, from
Chris Wilson.
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add a small comment to handle_seqno_wrap.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
In preparation for handling ring seqno wrapping, split
intel_ring_begin into helper part which doesn't allocate
seqno.
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This is needed for testing seqno wrapping. Be careful to not bump next_seqno
more than 0x7FFFFFFF at a time (between some handled requests) as
i915_seqno_passed() can't handle bigger difference in between.
v2: Address review comments from Chris Wilson.
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Squash in fixup to properly remove the debugfs file on driver
unload again.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
In a couple of places we attempt to adjust the existing watermark
registers to update them for the new cursor watermarks. This goes
horribly wrong as instead of clearing the cursor bits prior to or'ing in
the new values, we clear the rest of the register with the result that
the watermark registers contain bogus values.
References: https://bugs.freedesktop.org/show_bug.cgi?id=47034
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>
- __iomem where there is none (I love how we mix these things up).
- Use gfp_t instead of an other plain type.
- Unconfuse one place about enum pipe vs enum transcoder - for the pch
transcoder we actually use the pipe enum. Fixup the other cases
where we assign the pipe to the cpu transcoder with explicit casts.
- Declare the mch_lock properly in a header.
There is still a decent mess in intel_bios.c about __iomem, but heck,
this is x86 and we're allowed to do that.
Makes-sparse-happy: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Use a space after the cast consistently and fix up the
newly-added cast in i915_irq.c to properly use __iomem.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Simply use the last write-domain set for the object in the batch,
trusting userspace to have correctly flushed the caches between usage as
a write target. This check dates back from the golden age of having only
a single operation per batch with the kernel repeating it for each
cliprect, and conflicts both with userspace trying to efficiently batch
multiple operations and with reducing the kernel overhead of relocation
processing.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
From BSpec:
"If the Ring Buffer Head Pointer and the Tail Pointer are on the same
cacheline, the Head Pointer must not be greater than the Tail
Pointer."
The easiest way to enforce this is to reduce the reported ring space.
References:
Gen2 BSpec "1. Programming Environment" / 1.4.4.6 "Ring Buffer Use"
Gen3 BSpec "vol1c Memory Interface Functions" / 2.3.4.5 "Ring Buffer Use"
Gen4+ BSpec "vol1c Memory Interface and Command Stream" / 5.3.4.5 "Ring Buffer Use"
v2: Include the exact BSpec references in the description
v3: s/64/I915_RING_FREE_SPACE, and add the BSpec information to the code
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>
As we may actually allocate in order to save the physical swizzling bits
during the free, we have to be careful not to trigger the shrinker on
the same object.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Added a small comment in the code to really drive the
scariness of this patch home.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The primary purpose of this was to debug some use-after-free memory
corruption that was causing an OOPS inside drm/i915. As it turned out
the corruption was being caused elsewhere and i915.ko as a major user of
many objects was being hit hardest.
Indeed as we do frequent the generic kmalloc caches, dedicating one to
ourselves (or at least naming one for us depending upon the core) aids
debugging our own slab usage.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Allow for the creation of GEM objects backed by stolen memory. As these
are not backed by ordinary pages, we create a fake dma mapping and store
the address in the scatterlist rather than obj->pages.
v2: Mark _i915_gem_object_create_stolen() as static, as noticed by Jesse
Barnes.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
In order to accommodate objects that are not backed by struct pages, but
instead point into a contiguous region of stolen space, we need to make
various changes to avoid dereferencing obj->pages or obj->base.filp.
First introduce a marker for the stolen object, that specifies its
offset into the stolen region and implies that it has no backing pages.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As FBC is commonly disabled due to limitations of the chipset upon
output configurations, on many systems FBC is never enabled. For those
systems, it is advantageous to make use of the stolen memory for other
objects and so we defer allocation of the FBC chunk until we actually
require it. This increases the likelihood of that allocation failing,
but that in turns means that we are already taking advantage of the
stolen memory!
As well as delaying the allocation from driver initialisation until the
first use of FBC, we also return the stolen block after we finish using
it - allowing greater flexibility in our usage of stolen space. A side
effect of this is that we can then attempt to allocate only the required
amount of space (with a little slack to reduce reallocation rate and
avoid fragmentation).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As yet we do not do any preallocation (chicken-and-egg problem), but we
may like to preserve anything already allocated by the BIOS or grub and
reuse for own purposes after initialising the driver.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The routine to query the base of stolen memory was using the wrong
registers and the wrong encodings on virtually every platform.
It was not until the G33 refresh, that a PCI config register was
introduced that explicitly said where the stolen memory was. Prior to
865G there was not even a register that said where the end of usable
low memory was and where the stolen memory began (or ended depending
upon chipset). Before then, one has to look at the BIOS memory maps to
find the Top of Memory. Alas that is not exported by arch/x86 and so we
have to resort to disabling stolen memory on gen2 for the time being.
Then SandyBridge enlarged the PCI register to a full 32-bits and change
the encoding of the address, so even though we happened to be querying
the right register, we read the wrong bits and ended up using address 0
for our stolen data, i.e. notably FBC.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We only need to read/write the south interrupt register if the
corresponding bit is set in the north master interrupt register.
Noticed while reading our interrupt handling code.
Same optimization has already been applied on ivb in
commit 0e43406bcc
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Wed May 9 21:45:44 2012 +0100
drm/i915: Simplify interrupt processing for IvyBridge
We can take advantage that the PCH_IIR is a subordinate register to
reduce one of the required IIR reads, and that we only need to clear
interrupts handled to reduce the writes. And by simply tidying the code
we can reduce the line count and hopefully make it more readable.
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
