Damien Lespiau wondered how race the gpu reset/hang detection code is
against concurrent gpu resets/hang detections or combinations thereof.
Luckily the single work item is guranteed to never run concurrently,
so reset handling is already single-threaded.
Hence we only have to worry about concurrent hang detections, or a
hang detection firing off while we're still processing an older gpu
reset request. Due to the new mechanism of setting the reset in
progress flag and the ordering guaranteed by the schedule_work
function there's nothing to do but add a comment explaining why we're
safe.
The only thing I've noticed is that we still try to reset the gpu now,
even when it is declared terminally wedged. Add a check for that to
avoid continous warnings about failed resets, in case the hangcheck
timer ever gets stuck.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
With the previous patch the state transition handling of the reset
code itself is now (hopefully) race free and solid. But that still
leaves out everyone else - with the various lock-free wait paths
we have there's the possibility that the reset happens between the
point where we read the seqno we should wait on and the actual wait.
And if __wait_seqno then never sees the RESET_IN_PROGRESS state, we'll
happily wait for a seqno which will in all likelyhood never signal.
In practice this is not a big problem since the X server gets
constantly interrupted, and can then submit more work (hopefully) to
unblock everyone else: As soon as a new seqno write lands, all waiters
will unblock. But running the i-g-t reset testcase ZZ_hangman can
expose this race, especially on slower hw with fewer cpu cores.
Now looking forward to ARB_robustness and friends that's not the best
possible behaviour, hence this patch adds a reset_counter to be able
to detect any reset, even if a given thread never observed the
in-progress state.
The important part is to correctly order things:
- The write side needs to increment the counter after any seqno gets
reset. Hence we need to do that at the end of the reset work, and
again wake everyone up. We also need to place a barrier in between
any possible seqno changes and the counter increment, since any
unlock operations only guarantee that nothing leaks out, but not
that at later load operation gets moved ahead.
- On the read side we need to ensure that no reset can sneak in and
invalidate the seqno. In all cases we can use the one-sided barrier
that unlock operations guarantee (of the lock protecting the
respective seqno/ring pair) to ensure correct ordering. Hence it is
sufficient to place the atomic read before the mutex/spin_unlock and
no additional barriers are required.
The end-result of all this is that we need to wake up everyone twice
in a reset operation:
- First, before the reset starts, to get any lockholders of the locks,
so that the reset can proceed.
- Second, after the reset is completed, to allow waiters to properly
and reliably detect the reset condition and bail out.
I admit that this entire reset_counter thing smells a bit like
overkill, but I think it's justified since it makes it really explicit
what the bail-out condition is. And we need a reset counter anyway to
implement ARB_robustness, and imo with finer-grained locking on the
horizont this is the most resilient scheme I could think of.
v2: Drop spurious change in the wait_for_error EXIT_COND - we only
need to wait until we leave the reset-in-progress wedged state.
v3: Don't play tricks with barriers in the throttle ioctl, the
spin_unlock is barrier enough.
I've also considered using a little helper to grab the current
reset_counter, but then decided that hiding the atomic_read isn't a
great idea, since having it explicitly show up in the code is a nice
remainder to reviews to check the memory barriers.
v4: Add a comment to explain why we need to fall through in
__wait_seqno in the end variable assignments.
v5: Review from Damien:
- s/smb/smp/ in a comment
- don't increment the reset counter after we've set it to WEDGED. Now
we (again) properly wedge the gpu when the reset fails.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
More important fixes for 3.9:
- error_state improvements to help debug the new scanline wait code added
for gen6+ - bug reports started popping up :( patch from Chris Wilson.
- fix a panel power sequence confusion between the eDP and lvds detection
code resulting in black screens - regression introduce in 3.8 (Jani
Nikula)
- Chris fixed the root-cause of the ilk relocation vs. evict bug.
- Another piece of cargo-culted rc6 lore from Jani, fixes up a regression
where a system refused to go into rc6 after suspend sometimes.
* 'drm-intel-fixes' of git://people.freedesktop.org/~danvet/drm-intel:
drm/i915: fix FORCEWAKE posting reads
drm/i915: Invalidate the relocation presumed_offsets along the slow path
drm/i915/eDP: do not write power sequence registers for ghost eDP
drm/i915: Record DERRMR, FORCEWAKE and RING_CTL in error-state
The aim of this locking rework is that ioctls which a compositor should be
might call for every frame (set_cursor, page_flip, addfb, rmfb and
getfb/create_handle) should not be able to block on kms background
activities like output detection. And since each EDID read takes about
25ms (in the best case), that always means we'll drop at least one frame.
The solution is to add per-crtc locking for these ioctls, and restrict
background activities to only use the global lock. Change-the-world type
of events (modeset, dpms, ...) need to grab all locks.
Two tricky parts arose in the conversion:
- A lot of current code assumes that a kms fb object can't disappear while
holding the global lock, since the current code serializes fb
destruction with it. Hence proper lifetime management using the already
created refcounting for fbs need to be instantiated for all ioctls and
interfaces/users.
- The rmfb ioctl removes the to-be-deleted fb from all active users. But
unconditionally taking the global kms lock to do so introduces an
unacceptable potential stall point. And obviously changing the userspace
abi isn't on the table, either. Hence this conversion opportunistically
checks whether the rmfb ioctl holds the very last reference, which
guarantees that the fb isn't in active use on any crtc or plane (thanks
to the conversion to the new lifetime rules using proper refcounting).
Only if this is not the case will the code go through the slowpath and
grab all modeset locks. Sane compositors will never hit this path and so
avoid the stall, but userspace relying on these semantics will also not
break.
All these cases are exercised by the newly added subtests for the i-g-t
kms_flip, tested on a machine where a full detect cycle takes around 100
ms. It works, and no frames are dropped any more with these patches
applied. kms_flip also contains a special case to exercise the
above-describe rmfb slowpath.
* 'drm-kms-locking' of git://people.freedesktop.org/~danvet/drm-intel: (335 commits)
drm/fb_helper: check whether fbcon is bound
drm/doc: updates for new framebuffer lifetime rules
drm: don't hold crtc mutexes for connector ->detect callbacks
drm: only grab the crtc lock for pageflips
drm: optimize drm_framebuffer_remove
drm/vmwgfx: add proper framebuffer refcounting
drm/i915: dump refcount into framebuffer debugfs file
drm: refcounting for crtc framebuffers
drm: refcounting for sprite framebuffers
drm: fb refcounting for dirtyfb_ioctl
drm: don't take modeset locks in getfb ioctl
drm: push modeset_lock_all into ->fb_create driver callbacks
drm: nest modeset locks within fpriv->fbs_lock
drm: reference framebuffers which are on the idr
drm: revamp framebuffer cleanup interfaces
drm: create drm_framebuffer_lookup
drm: revamp locking around fb creation/destruction
drm: only take the crtc lock for ->cursor_move
drm: only take the crtc lock for ->cursor_set
drm: add per-crtc locks
...
The coup de grace of the entire journey. No more dropped frames every
10s on my testbox!
I've tried to audit all ->detect and ->get_modes callbacks, but things
became a bit fuzzy after trying to piece together the umpteenth
implemenation. Afaict most drivers just have bog-standard output
register frobbing with a notch of i2c edid reading, nothing which
could potentially race with the newly concurrent pageflip/set_cursor
code. The big exception is load-detection code which requires a
running pipe, but radeon/nouveau seem to to this without touching any
state which can be observed from page_flip (e.g. disabled crtcs
temporarily getting enabled and so a pageflip succeeding).
The only special case I could find is the i915 load detect code. That
uses the normal modeset interface to enable the load-detect crtc, and
so userspace could try to squeeze in a pageflip on the load-detect
pipe. So we need to grab the relevant crtc mutex in there, to avoid
the temporary crtc enabling to sneak out and be visible to userspace.
Note that the sysfs files already stopped grabbing the per-crtc locks,
since I didn't want to bother with doing a interruptible
modeset_lock_all. But since there's very little in-between breakage
(essentially just the ability for userspace to pageflip on load-detect
crtcs when it shouldn't on the i915 driver) I figured I don't need to
bother.
Reviewed-by: Rob Clark <rob@ti.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We have two classes of framebuffer
- Created by the driver (atm only for fbdev), and the driver holds
onto the last reference count until destruction.
- Created by userspace and associated with a given fd. These
framebuffers will be reaped when their assoiciated fb is closed.
Now these two cases are set up differently, the framebuffers are on
different lists and hence destruction needs to clean up different
things. Also, for userspace framebuffers we remove them from any
current usage, whereas for internal framebuffers it is assumed that
the driver has done this already.
Long story short, we need two different ways to cleanup such drivers.
Three functions are involved in total:
- drm_framebuffer_remove: Convenience function which removes the fb
from all active usage and then drops the passed-in reference.
- drm_framebuffer_unregister_private: Will remove driver-private
framebuffers from relevant lists and drop the corresponding
references. Should be called for driver-private framebuffers before
dropping the last reference (or like for a lot of the drivers where
the fbdev is embedded someplace else, before doing the cleanup
manually).
- drm_framebuffer_cleanup: Final cleanup for both classes of fbs,
should be called by the driver's ->destroy callback once the last
reference is gone.
This patch just rolls out the new interfaces and updates all drivers
(by adding calls to drm_framebuffer_unregister_private at all the
right places)- no functional changes yet. Follow-on patches will move
drm core code around and update the lifetime management for
framebuffers, so that we are no longer required to keep framebuffers
alive by locking mode_config.mutex.
I've also updated the kerneldoc already.
vmwgfx seems to again be a bit special, at least I haven't figured out
how the fbdev support in that driver works. It smells like it's
external though.
v2: The i915 driver creates another private framebuffer in the
load-detect code. Adjust its cleanup code, too.
Reviewed-by: Rob Clark <rob@ti.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Well, at least step 1. The goal here is that framebuffer objects can
survive outside of the mode_config lock, with just a reference held
as protection. The first step to get there is to introduce a special
fb_lock which protects fb lookup, creation and destruction, to make
them appear atomic.
This new fb_lock can nest within the mode_config lock. But the idea is
(once the reference counting part is completed) that we only quickly
take that fb_lock to lookup a framebuffer and grab a reference,
without any other locks involved.
vmwgfx is the only driver which does framebuffer lookups itself, also
wrap those calls to drm_mode_object_find with the new lock.
Also protect the fb_list walking in i915 and omapdrm with the new lock.
As a slight complication there's also the list of user-created fbs
attached to the file private. The problem now is that at fclose() time
we need to walk that list, eventually do a modeset call to remove the
fb from active usage (and are required to be able to take the
mode_config lock), but in the end we need to grab the new fb_lock to
remove the fb from the list. The easiest solution is to add another
mutex to protect this per-file list.
Currently that new fbs_lock nests within the modeset locks and so
appears redudant. But later patches will switch around this sequence
so that taking the modeset locks in the fb destruction path is
optional in the fastpath. Ultimately the goal is that addfb and rmfb
do not require the mode_config lock, since otherwise they have the
potential to introduce stalls in the pageflip sequence of a compositor
(if the compositor e.g. switches to a fullscreen client or if it
enables a plane). But that requires a few more steps and hoops to jump
through.
Note that framebuffer creation/destruction is now double-protected -
once by the fb_lock and in parts by the idr_lock. The later would be
unnecessariy if framebuffers would have their own idr allocator. But
that's material for another patch (series).
v2: Properly initialize the fb->filp_head list in _init, otherwise the
newly added WARN to check whether the fb isn't on a fpriv list any
more will fail for driver-private objects.
v3: Fixup two error-case unlock bugs spotted by Richard Wilbur.
Reviewed-by: Rob Clark <rob@ti.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Two exceptions:
- debugfs files only read information which is not related to crtc, so
can stay on the modeset_config lock.
- Same holds for the edp vdd work in intel_dp.c. Add a corresponding
WARN_ON and a comment next to the intel_dp struct fields for
documentation.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
With more fine-grained locking we can no longer rely on the big
mode_config lock to prevent concurrent access to mode resources
like framebuffers. Instead a framebuffer becomes accessible to
other threads as soon as it is added to the relevant lookup
structures. Hence it needs to be fully set up by the time drivers
call drm_framebuffer_init.
This patch here is the drivers part of that reorg. Nothing really fancy
going on safe for three special cases.
- exynos needs to be careful to properly unref all handles.
- nouveau gets a resource leak fixed for free: one of the error
cases didn't cleanup the framebuffer, which is now moot since
the framebuffer is only registered once it is fully set up.
- vmwgfx requires a slight reordering of operations, I'm hoping I didn't
break anything (but it's refcount management only, so should be safe).
v2: Split out exynos, since it's a bit more hairy than expected.
v3: Drop bogus cirrus hunk noticed by Richard Wilbur.
v4: Split out vmwgfx since there's a small change in return values.
Reviewed-by: Rob Clark <rob@ti.com> (core + omapdrm)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Now that we seem to have brought order to the GTT barriers, the last one
to review is the terminal barrier before we unbind the buffer from the
GTT. This needs to only be performed if the buffer still resides in the
GTT domain, and so we can skip some needless barriers otherwise.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
With a fence, we only need to insert a memory barrier around the actual
fence alteration for CPU accesses through the GTT. Performing the
barrier in flush-fence was inserting unnecessary and expensive barriers
for never fenced objects.
Note removing the barriers from flush-fence, which was effectively a
barrier before every direct access through the GTT, revealed that we
where missing a barrier before the first access through the GTT. Lack of
that barrier was sufficient to cause GPU hangs.
v2: Add a couple more comments to explain the new barriers
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We have two important transitions of the wedged state in the current
code:
- 0 -> 1: This means a hang has been detected, and signals to everyone
that they please get of any locks, so that the reset work item can
do its job.
- 1 -> 0: The reset handler has completed.
Now the last transition mixes up two states: "Reset completed and
successful" and "Reset failed". To distinguish these two we do some
tricks with the reset completion, but I simply could not convince
myself that this doesn't race under odd circumstances.
Hence split this up, and add a new terminal state indicating that the
hw is gone for good.
Also add explicit #defines for both states, update comments.
v2: Split out the reset handling bugfix for the throttle ioctl.
v3: s/tmp/wedged/ sugested by Chris Wilson. Also fixup up a rebase
error which prevented this patch from actually compiling.
v4: To unify the wedged state with the reset counter, keep the
reset-in-progress state just as a flag. The terminally-wedged state is
now denoted with a big number.
v5: Add a comment to the reset_counter special values explaining that
WEDGED & RESET_IN_PROGRESS needs to be true for the code to be
correct.
v6: Fixup logic errors introduced with the wedged+reset_counter
unification. Since WEDGED implies reset-in-progress (in a way we're
terminally stuck in the dead-but-reset-not-completed state), we need
ensure that we check for this everywhere. The specific bug was in
wait_for_error, which would simply have timed out.
v7: Extract an inline i915_reset_in_progress helper to make the code
more readable. Also annote the reset-in-progress case with an
unlikely, to help the compiler optimize the fastpath. Do the same for
the terminally wedged case with i915_terminally_wedged.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
While auditing the code I've noticed one place (the throttle ioctl)
which does not yet wait for the reset handler to complete and doesn't
properly decode the wedge state into -EAGAIN/-EIO. Fix this up by
calling the right helpers. This might explain the oddball "my
compositor just died in a successfull gpu reset" reports. Or maybe not, since
current mesa doesn't use this ioctl to throttle command submission.
The throttle ioctl doesn't take the struct_mutex, so to avoid busy-looping
with -EAGAIN while a reset is in process, check for errors first and wait
for the handler to complete if a reset is pending by calling
i915_gem_wait_for_error.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
And to make Ben Widawsky happier, use the gpu_error instead of
the entire device as the argument in some functions.
Drop the outdated comment on ->wedged for now, a follow-up patch will
change the semantics and add a proper comment again.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This has been sprinkled all over the place in dev_priv. I think
it'd be good to also move all the code into a separate file like
i915_gem_error.c, but that's for another patch.
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Tha one is really big, since it contains tons of comments explaining
how things work. Which is nice ;-)
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The reasoning behind our code taking two paths depending upon whether or
not we may have been configured for IOMMU isn't clear to me. It should
always be safe to use the pci mapping functions as they are designed to
abstract the decision we were handling in i915.
Aside from simpler code, removing another member for the intel_gtt
struct is a nice motivation.
I ran this by Chris, and he wasn't concerned about the extra kzalloc,
and memory references vs. page_to_phys calculation in the case without
IOMMU.
v2: Update commit message
v3: Remove needs_dmar addition from Zhenyu upstream
This reverts (and then other stuff)
commit 20652097da
Author: Zhenyu Wang <zhenyuw@linux.intel.com>
Date: Thu Dec 13 23:47:47 2012 +0800
drm/i915: Fix missed needs_dmar setting
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com> (v2)
Cc: Zhenyu Wang <zhenyuw@linux.intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Squash in follow-up fix to remove the bogus hunk which
deleted the dma_mask configuration for gen6+.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
And, move it to where the rest of the logic is.
There is some slight functionality changes. There was extra paranoid
checks in AGP code making sure we never do idle maps on gen2 parts. That
was not duplicated as the simple PCI id check should do the right thing.
v2: use IS_GEN5 && IS_MOBILE check instead. For now, this is the same as
IS_IRONLAKE_M but is more future proof. The workaround docs hint that
more than one platform may be effected, but we've never seen such a
platform in the wild. (Rodrigo, Daniel)
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com> (v1)
Cc: Dave Airlie <airlied@redhat.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The AVI infoframe is able to inform the display whether the source is
sending full or limited range RGB data.
As per CEA-861 [1] we must first check whether the display reports the
quantization range as selectable, and if so we can set the approriate
bits in the AVI inforframe.
[1] CEA-861-E - 6.4 Format of Version 2 AVI InfoFrame
v2: Give the Q bits better names, add spec chapter information
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Add a new "Automatic" mode to the "Broadcast RGB" range property.
When selected the driver automagically selects between full range and
limited range output.
Based on CEA-861 [1] guidelines, limited range output is selected if the
mode is a CEA mode, except 640x480. Otherwise full range output is used.
Additionally DVI monitors should most likely default to full range
always.
As per DP1.2a [2] DisplayPort should always use full range for 18bpp, and
otherwise will follow CEA-861 rules.
NOTE: The default value for the property will now be "Automatic"
so some people may be affected in case they're relying on the
current full range default.
[1] CEA-861-E - 5.1 Default Encoding Parameters
[2] VESA DisplayPort Ver.1.2a - 5.1.1.1 Video Colorimetry
v2: Use has_hdmi_sink to check if a HDMI monitor is present
v3: Add information about relevant spec chapters
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The RGB color range select bit on the DP/SDVO/HDMI registers
disappeared when PCH was introduced, and instead a new PIPECONF bit
was added that performs the same function.
Add a new INTEL_MODE_LIMITED_COLOR_RANGE private mode flag, and set
it in the encoder mode_fixup if limited color range is requested.
Set the the PIPECONF bit 13 based on the flag.
Experimentation showed that simply toggling the bit while the pipe is
active doesn't work. We need to restart the pipe, which luckily already
happens.
The DP/SDVO/HDMI bit 8 is marked MBZ in the docs, so avoid setting it,
although it doesn't seem to do any harm in practice.
TODO:
- the PIPECONF bit too seems to have disappeared from HSW. Need a
volunteer to test if it's just a documentation issue or if it's really
gone. If the bit is gone and no easy replacement is found, then I suppose
we may need to use the pipe CSC unit to perform the range compression.
v2: Use mode private_flags instead of intel_encoder virtual functions
v3: Moved the intel_dp color_range handling after bpc check to help
later patches
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=46800
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Mappable_end, ie. size is almost always what you want as opposed to the
number of entries. Since we already have that information, we can scrap
the number of entries and only calculate it when needed.
If gtt_start is !0, this will have slightly different behavior. This
difference can only occur in DRI1, and exists when we try to kick out
the firmware fb. The new code seems like a bugfix to me.
The other case where we've changed the behavior is during init we check
the mappable region against our current known upper and lower limits
(64MB, and 512MB). This now matches the comment, and makes things more
convenient after removing gtt_mappable_entries.
Also worth noting is the setting of mappable_end is taken out of setup
because we do it earlier now in the DRI2 case and therefore need to add
that tiny hunk to support the DRI1 IOCTL.
v2: Move up mappable end to before legacy AGP init
v3: Add the dev_priv inclusion here from previous rebase error in patch
5
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com> (v2)
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: squash in fix for a printk format flag mismatch warning.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
We have enough info to not use the intel_gtt bridge stuff.
v2: Move setup of mappable_base above the legacy init stuff because we
still need that on older platforms. (Daniel)
v3: Remove the dev_priv hunk which was rebased in by accident
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com> (v2)
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The purpose of the gtt structure is to help isolate our gtt specific
properties from the rest of the code (in doing so it help us finish the
isolation from the AGP connection).
The following members are pulled out (and renamed):
gtt_start
gtt_total
gtt_mappable_end
gtt_mappable
gtt_base_addr
gsm
The gtt structure will serve as a nice place to put gen specific gtt
routines in upcoming patches. As far as what else I feel belongs in this
structure: it is meant to encapsulate the GTT's physical properties.
This is why I've not added fields which track various drm_mm properties,
or things like gtt_mtrr (which is itself a pretty transient field).
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
[Ben modified commit messages]
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
With the assertion from the previous patch in place, it should be safe
to get rid gtt_mappable_total. Keeps things saner to not have to track
the same info in two places.
In order to keep the diff as simple as possible and keep with the
existing gtt_setup semantics we opt to keep gtt_mappable_end. It's not
as consistent with the 'total' used in the previous patch, but that can
be fixed later.
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
[Ben modified commit message]
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Both DRI1 and DRI2 can never specify a mappable size which goes past the
GTT size. Don't pretend otherwise.
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Fix up some copypaste errors in the PIPESTAT register for VLV.
SPRITE0_FLIP_DONE_INT_EN_VLV is bit 22, not bit 26.
SPRITE0_FLIPDONE_INT_STATUS_VLV is bit 14, not bit 15.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Using copywinwin10 as an example that is dependent upon emitting a lot
of relocations (2 per operation), we see improvements of:
c2d/gm45: 618000.0/sec to 623000.0/sec.
i3-330m: 748000.0/sec to 789000.0/sec.
(measured relative to a baseline with neither optimisations applied).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Userspace is able to hint to the kernel that its command stream and
auxiliary state buffers already hold the correct presumed addresses and
so the relocation process may be skipped if the kernel does not need to
move any buffers in preparation for the execbuffer. Thus for the common
case where the allotment of buffers is static between batches, we can
avoid the overhead of individually checking the relocation entries.
Note that this requires userspace to supply the domain tracking and
requests for workarounds itself that would otherwise be computed based
upon the relocation entries.
Using copywinwin10 as an example that is dependent upon emitting a lot
of relocations (2 per operation), we see improvements of:
c2d/gm45: 618000.0/sec to 632000.0/sec.
i3-330m: 748000.0/sec to 830000.0/sec.
(measured relative to a baseline with neither optimisations applied).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Imre Deak <imre.deak@intel.com>
[danvet: Fixup merge conflict in userspace header due to different
baseline trees.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Instead of passing around the eb-objects hashtable and a separate object
list, we can include the object list into the eb-objects structure for
convenience.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Move the existing checking inside bind_to_gtt() to the more appropriate
layer in order to prevent recreation of the pages after they have been
explicitly truncated.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As a means to investigate some bad system behaviour related to the
purging of the active, inactive and unbound lists, it is useful to be
able to manually control when those lists should be cleared.
v2: use _safe list iterators as we kick objects from the list as we
walk.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add a small comment explaining why we don't need to check and
wait for gpu resets, acked by Chris on irc.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This variable is only used locally in the irq postinstall
functions for ivybridge and ironlake.
Signed-off-by: Egbert Eich <eich@suse.de>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Otherwise it seems like we can get stuck with concurrent waiters.
Right now this /shouldn't/ be a problem, since all pending pageflip
waiters are serialized by the one mode_config.mutex, so there's at
most on waiter. But better paranoid than sorry, since this is tricky
code.
v2: WARN_ON(waitqueue_active) before waiting, as suggested by Chris
Wilson.
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Daniel writes:
- seqno wrap fixes and debug infrastructure from Mika Kuoppala and Chris
Wilson
- some leftover kill-agp on gen6+ patches from Ben
- hotplug improvements from Damien
- clear fb when allocated from stolen, avoids dirt on the fbcon (Chris)
- Stolen mem support from Chris Wilson, one of the many steps to get to
real fastboot support.
- Some DDI code cleanups from Paulo.
- Some refactorings around lvds and dp code.
- some random little bits&pieces
* tag 'drm-intel-next-2012-12-21' of git://people.freedesktop.org/~danvet/drm-intel: (93 commits)
drm/i915: Return the real error code from intel_set_mode()
drm/i915: Make GSM void
drm/i915: Move GSM mapping into dev_priv
drm/i915: Move even more gtt code to i915_gem_gtt
drm/i915: Make next_seqno debugs entry to use i915_gem_set_seqno
drm/i915: Introduce i915_gem_set_seqno()
drm/i915: Always clear semaphore mboxes on seqno wrap
drm/i915: Initialize hardware semaphore state on ring init
drm/i915: Introduce ring set_seqno
drm/i915: Missed conversion to gtt_pte_t
drm/i915: Bug on unsupported swizzled platforms
drm/i915: BUG() if fences are used on unsupported platform
drm/i915: fixup overlay stolen memory leak
drm/i915: clean up PIPECONF bpc #defines
drm/i915: add intel_dp_set_signal_levels
drm/i915: remove leftover display.update_wm assignment
drm/i915: check for the PCH when setting pch_transcoder
drm/i915: Clear the stolen fb before enabling
drm/i915: Access to snooped system memory through the GTT is incoherent
drm/i915: Remove stale comment about intel_dp_detect()
...
Conflicts:
drivers/gpu/drm/i915/intel_display.c
We stopped reading FORCEWAKE for posting reads in
commit 8dee3eea3c
Author: Ben Widawsky <ben@bwidawsk.net>
Date: Sat Sep 1 22:59:50 2012 -0700
drm/i915: Never read FORCEWAKE
and started using something from the same cacheline instead. On the
bug reporter's machine this broke entering rc6 states after a
suspend/resume cycle. It turns out reading ECOBUS as posting read
worked fine, while GTFIFODBG did not, preventing RC6 states after
suspend/resume per the bug report referenced below. It's not entirely
clear why, but clearly GTFIFODBG was nowhere near the same cacheline
or address range as FORCEWAKE.
Trying out various registers for posting reads showed that all tested
registers for which NEEDS_FORCE_WAKE() (in i915_drv.c) returns true
work. Conversely, most (but not quite all) registers for which
NEEDS_FORCE_WAKE() returns false do not work. Details in the referenced
bug.
Based on the above, add posting reads on ECOBUS where GTFIFODBG was
previously relied on.
In true cargo cult spirit, add posting reads for FORCEWAKE_VLV writes as
well, but instead of ECOBUS, use FORCEWAKE_ACK_VLV which is in the same
address range as FORCEWAKE_VLV.
v2: Add more details to the commit message. No functional changes.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=52411
Reported-and-tested-by: Alexander Bersenev <bay@hackerdom.ru>
CC: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@vger.kernel.org
[danvet: add cc: stable and make the commit message a bit clearer that
this is a regression fix and what exactly broke.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
In the slow path, we are forced to copy the relocations prior to
acquiring the struct mutex in order to handle pagefaults. We forgo
copying the new offsets back into the relocation entries in order to
prevent a recursive locking bug should we trigger a pagefault whilst
holding the mutex for the reservations of the execbuffer. Therefore, we
need to reset the presumed_offsets just in case the objects are rebound
back into their old locations after relocating for this exexbuffer - if
that were to happen we would assume the relocations were valid and leave
the actual pointers to the kernels dangling, instant hang.
Fixes regression from commit bcf50e2775
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Sun Nov 21 22:07:12 2010 +0000
drm/i915: Handle pagefaults in execbuffer user relocations
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=55984
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@fwll.ch>
Cc: stable@vger.kernel.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Some machines detect an eDP port even if it's not really there, and eDP
initialization has a fail path for this. Typically such machines have an
LVDS display instead. A regression introduced in
commit 82ed61fa1a
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date: Sat Oct 20 20:57:41 2012 +0200
drm/i915: make edp panel power sequence setup more robust
updated the power sequence registers PCH_PP_ON_DELAYS, PCH_PP_OFF_DELAYS,
and PCH_PP_DIVISOR also in the ghost eDP case, messing up the LVDS display.
Split the power sequencer initialization into two, delaying the register
updates until after we know the eDP is real.
Note: Keep the PP_CONTROL unlocking in the first part, even if it does not
update registers, per the commit message of the above mentioned commit.
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=52601
Reported-and-tested-by: Ryan Coe <ryan@rycomotorsports.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
These are useful for investigating hangs involving WAIT_FOR_EVENT.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Apply a droplet of Future-Proof in the if-ladder.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Daniel writes:
"Pretty much all just major fixes:
- 2 pieces of duct-tape for the ilk bug.
- Sprite regression fixes from Chris.
- OOPS fix for a div-by-zero from Chris, regression due to the modeset
rework in 3.7, now brought to light by a benign change in 3.8.
- Fix interrupted bo pinning, used to work around CS coherency issues on
i830/i845 (kernel also has a w/a newly in 3.8, but pinning is more efficient if
possible)."
This partially reverts
commit 6c085a728c
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date: Mon Aug 20 11:40:46 2012 +0200
drm/i915: Track unbound pages
Closer inspection of that patch revealed a bunch of unrelated changes
in the shrinker:
- The shrinker count is now in pages instead of objects.
- For counting the shrinkable objects the old code only looked at the
inactive list, the new code looks at all bounds objects (including
pinned ones). That is obviously in addition to the new unbound list.
- The shrinker cound is no longer scaled with
sysctl_vfs_cache_pressure. Note though that with the default tuning
value of vfs_cache_pressue = 100 this doesn't affect the shrinker
behaviour.
- When actually shrinking objects, the old code first dropped
purgeable objects, then normal (inactive) objects. Only then did it,
in a last-ditch effort idle the gpu and evict everything. The new
code omits the intermediate step of evicting normal inactive
objects.
Safe for the first change, which seems benign, and the shrinker count
scaling, which is a bit a different story, the endresult of all these
changes is that the shrinker is _much_ more likely to fall back to the
last-ditch resort of idling the gpu and evicting everything. The old
code could only do that if something else evicted lots of objects
meanwhile (since without any other changes the nr_to_scan will be
smaller than the object count).
Reverting the vfs_cache_pressure behaviour itself is a bit bogus: Only
dentry/inode object caches should scale their shrinker counts with
vfs_cache_pressure. Originally I've had that change reverted, too. But
Chris Wilson insisted that it's too bogus and shouldn't again see the
light of day.
Hence revert all these other changes and restore the old shrinker
behaviour, with the minor adjustment that we now first scan the
unbound list, then the inactive list for each object category
(purgeable or normal).
A similar patch has been tested by a few people affected by the gen4/5
hangs which started to appear in 3.7, which some people bisected to
the "drm/i915: Track unbound pages" commit. But just disabling the
unbound logic alone didn't change things at all.
Note that this patch doesn't fix the referenced bugs, it only hides
the underlying bug(s) well enough to restore pre-3.7 behaviour. The
key to achieve that is to massively reduce the likelyhood of going
into a full gpu stall and evicting everything.
v2: Reword commit message a bit, taking Chris Wilson's comment into
account.
v3: On Chris Wilson's insistency, do not reinstate the rather bogus
vfs_cache_pressure change.
Tested-by: Greg KH <gregkh@linuxfoundation.org>
Tested-by: Dave Kleikamp <dave.kleikamp@oracle.com>
References: https://bugs.freedesktop.org/show_bug.cgi?id=55984
References: https://bugs.freedesktop.org/show_bug.cgi?id=57122
References: https://bugs.freedesktop.org/show_bug.cgi?id=56916
References: https://bugs.freedesktop.org/show_bug.cgi?id=57136
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@vger.kernel.org
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
