Mika Kuoppala [Thu, 7 May 2020 14:20:45 +0000 (17:20 +0300)]
drm/i915/gen12: Add aux table invalidate for all engines
All engines, exception being blitter as it does not
care about the form, can access compressed surfaces.
So we need to add forced aux table invalidates
for those engines.
v2: virtual instance masking (Chris)
v3: bug on if not found (Chris)
References: d248b371f747 ("drm/i915/gen12: Invalidate aux table entries forcibly")
References bspec#43904, hsdes#1809175790 Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Chuansheng Liu <chuansheng.liu@intel.com> Cc: Rafael Antognolli <rafael.antognolli@intel.com> Signed-off-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> Acked-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Link: https://patchwork.freedesktop.org/patch/msgid/20200507142045.8668-1-mika.kuoppala@linux.intel.com
Chris Wilson [Thu, 7 May 2020 15:23:38 +0000 (16:23 +0100)]
drm/i915: Remove wait priority boosting
Upon waiting a request (when asked), we gave that request a small
priority boost, not enough for it to cause preemption, but enough for it
to be scheduled next before all equals. We also used that bit to give
new clients a small priority boost, similar to FQ_CODEL, such that we
favoured short interactive tasks ahead of long running streams.
However, this is causing lots of complications with timeslicing where we
both want to honour the boost and yet ignore it. Those complications
cause unexpected user behaviour (tasks not being timesliced and run
concurrently as epxected), and the easiest way to resolve that is to
remove the boost. Hopefully, we can find a compromise again if we need
to, but in theory timeslicing itself and future more advanced schedulers
should give us the interactivity boost we seek.
Chris Wilson [Thu, 7 May 2020 15:51:09 +0000 (16:51 +0100)]
drm/i915: Mark concurrent submissions with a weak-dependency
We recorded the dependencies for WAIT_FOR_SUBMIT in order that we could
correctly perform priority inheritance from the parallel branches to the
common trunk. However, for the purpose of timeslicing and reset
handling, the dependency is weak -- as we the pair of requests are
allowed to run in parallel and not in strict succession.
The real significance though is that this allows us to rearrange
groups of WAIT_FOR_SUBMIT linked requests along the single engine, and
so can resolve user level inter-batch scheduling dependencies from user
semaphores.
Fixes: c81471f5e95c ("drm/i915: Copy across scheduler behaviour flags across submit fences")
Testcase: igt/gem_exec_fence/submit Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: <stable@vger.kernel.org> # v5.6+ Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200507155109.8892-1-chris@chris-wilson.co.uk
Mika Kuoppala [Wed, 6 May 2020 16:53:10 +0000 (19:53 +0300)]
drm/i915/gen12: Invalidate aux table entries forcibly
Aux table invalidation can fail on update. So
next access may cause memory access to be into stale entry.
Proposed workaround is to invalidate entries between
all batchbuffers.
v2: correct register address (Yang)
v3: respect the order (Chris)
References bspec#43904, hsdes#1809175790 Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Chuansheng Liu <chuansheng.liu@intel.com> Cc: Rafael Antognolli <rafael.antognolli@intel.com> Cc: Yang A Shi <yang.a.shi@intel.com> Signed-off-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> Acked-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Link: https://patchwork.freedesktop.org/patch/msgid/20200506165310.1239-1-mika.kuoppala@linux.intel.com
Chris Wilson [Mon, 4 May 2020 18:07:45 +0000 (19:07 +0100)]
drm/i915/gt: Stop holding onto the pinned_default_state
As we only restore the default context state upon banning a context, we
only need enough of the state to run the ring and nothing more. That is
we only need our bare protocontext.
Chris Wilson [Tue, 5 May 2020 08:46:29 +0000 (09:46 +0100)]
drm/i915/execlists: Record the active CCID from before reset
If we cannot trust the reset will flush out the CS event queue such that
process_csb() reports an accurate view of HW, we will need to search the
active and pending contexts to determine which was actually running at
the time we issued the reset.
We need a new PCode request commands and reply codes
to be added as a prepartion patch for QGV points
restricting for new SAGV support.
v2: - Extracted those changes into separate patch
(Ville Syrjälä)
v3: - Moved new PCode masks to another place from
PCode commands(Ville)
v4: - Moved new PCode masks to correspondent PCode
command, with identation(Ville)
- Changed naming to ICL_ instead of GEN11_
to fit more nicely into existing definition
style.
Chris Wilson [Sun, 3 May 2020 18:00:34 +0000 (19:00 +0100)]
drm/i915/display: Warn if the FBC is still writing to stolen on removal
If the FBC is still writing into stolen, it will overwrite any future
users of that stolen region. Check before release, just to ease any
concerns -- we can remove it again later if it is barking up the wrong
tree.
Sultan Alsawaf [Thu, 30 Apr 2020 21:46:54 +0000 (14:46 -0700)]
drm/i915: Don't enable WaIncreaseLatencyIPCEnabled when IPC is disabled
In commit 5a7d202b1574, a logical AND was erroneously changed to an OR,
causing WaIncreaseLatencyIPCEnabled to be enabled unconditionally for
kabylake and coffeelake, even when IPC is disabled. Fix the logic so
that WaIncreaseLatencyIPCEnabled is only used when IPC is enabled.
Ville Syrjälä [Wed, 29 Apr 2020 18:54:57 +0000 (21:54 +0300)]
drm/i915: Streamline the artihmetic
All these ROUNDING_FACTORs and whatnot are making this thing hard to
read. Get rid of them. And let's massage some of the fractions to
give us less questionable intermediate results and perhaps less
divisions.
Also looks like a good helping of 64bit math stuff is needed to
avoid some of overflows present in the current code. There
might still be a few overflows, namely when calculating
link_clks_available/samples_room (would require a huge hblank
though), and potentially when calculating hblank_rise (not sure
how large link_clks_active can get).
It looks like we're still not calculating exactly what the spec says
since we truncate tu_data and tu_line early. But I'm too lazy to
figure out if we could avoid that.
Ville Syrjälä [Wed, 29 Apr 2020 18:54:56 +0000 (21:54 +0300)]
drm/i915: Rename variables to be consistent with bspec
Since the code seems insistent on using the variable names from the
bspec formulat, let's be consistent and use those names for all
the things. For some reason 'link_clk' and 'lanes' were left out
in the code until now.
Ville Syrjälä [Wed, 29 Apr 2020 18:54:55 +0000 (21:54 +0300)]
drm/i915: Nuke mode.vrefresh usage
mode.vrefresh is rounded to the nearest integer. You don't want to use
it anywhere that requires precision. Also I want to nuke it.
vtotal*vrefresh == 1000*clock/htotal, so let's use the latter.
Ville Syrjälä [Thu, 30 Apr 2020 12:58:22 +0000 (15:58 +0300)]
drm/i915: Remove cnl pre-prod workarounds
Remove all the stepping dependent cnl workarounds. Bspec lists
more steppings than this so presumably these are classed as
pre-production. And this is cnl after all so no one should
really care anyway.
Ville Syrjälä [Wed, 29 Apr 2020 10:10:23 +0000 (13:10 +0300)]
drm/i915/fbc: Require linear fb stride to be multiple of 512 bytes on gen9/glk
Display WA #1105 says that FBC requires PLANE_STRIDE to be a multiple
of 512 bytes on gen9 and glk.
This is definitely true for glk as certain tests (such as
igt/kms_big_fb/linear-16bpp-rotate-0) are now failing when the
display resolution results in a plane stride which is not a
multiple of 512 bytes.
Curiously I was not able to reproduce this on a KBL. First I
suspected that our use of the FBC override stride explain this,
but after trying to use the override stride on glk the test
still failed. I did try both the old CHICKEN_MISC_4 way and
the new FBC_STRIDE way, neither had any effect on the result.
Anyways, we need this at least on glk. But let's trust the spec
and apply the w/a for all gen9 as well, despite being unable to
reproduce the problem.
That is a preparation patch before next one where we
introduce old_bw_state and a bunch of other changes
as well.
In a review comment it was suggested to split out
at least that renaming into a separate patch, what
is done here.
We need to calculate SAGV mask also in a non-modeset
commit, however currently active_pipes are only calculated
for modesets in global atomic state, thus now we will be
tracking those also in bw_state in order to be able to
properly access global data.
v2: - Removed pre/post plane SAGV updates from modeset(Ville)
- Now tracking active pipes in intel_can_enable_sagv(Ville)
v3: - lock global state if active_pipes change as well(Ville)
drm/i915: Use bw state for per crtc SAGV evaluation
Future platforms require per-crtc SAGV evaluation
and serializing global state when those are changed
from different commits.
v2: - Add has_sagv check to intel_crtc_can_enable_sagv
so that it sets bit in reject mask.
- Use bw_state in intel_pre/post_plane_enable_sagv
instead of atomic state
v3: - Fixed rebase conflict, now using
intel_atomic_crtc_state_for_each_plane_state in
order to call it from atomic check
v4: - Use fb modifier from plane state
v5: - Make intel_has_sagv static again(Ville)
- Removed unnecessary NULL assignments(Ville)
- Removed unnecessary SAGV debug(Ville)
- Call intel_compute_sagv_mask only for modesets(Ville)
- Serialize global state only if sagv results change, but
not mask itself(Ville)
v6: - use lock global state instead of serialize(Ville)
v7: - use both global state lock and serialize depending on
if we need to change only global state or access hw
(Ville)
Chris Wilson [Mon, 4 May 2020 14:06:29 +0000 (15:06 +0100)]
drm/i915/gem: Implement legacy MI_STORE_DATA_IMM
The older arches did not convert MI_STORE_DATA_IMM to using the GTT, but
left them writing to a physical address. The notes suggest that the
primary reason would be so that the writes were cache coherent, as the
CPU cache uses physical tagging. As such we did not implement the
legacy variant of MI_STORE_DATA_IMM and so left all the relocations
synchronous -- but with a small function to convert from the vma address
into the physical address, we can implement asynchronous relocs on these
older arches, fixing up a few tests that require them.
In order to be able to test the legacy paths, refactor the gpu
relocations so that we can hook them up to a selftest.
v2: Use an array of offsets not enum labels for the selftest
v3: Refactor the common igt_hexdump()
Chris Wilson [Mon, 4 May 2020 12:51:49 +0000 (13:51 +0100)]
drm/i915/gem: Specify address type for chained reloc batches
It is required that a chained batch be in the same address domain as its
parent, and also that must be specified in the command for earlier gen
as it is not inferred from the chaining until gen6.
Chris Wilson [Mon, 4 May 2020 04:48:42 +0000 (05:48 +0100)]
drm/i915: Allow some leniency in PCU reads
Extend the timeout for pcode reads to 20ms as they should not be
performed along critical paths, and succeeding after a short delay is
better than failing entirely.
Chris Wilson [Sun, 3 May 2020 17:15:13 +0000 (18:15 +0100)]
drm/i915/gem: Lazily acquire the device wakeref for freeing objects
We only need the device wakeref on freeing the objects if we have to
unbind the object from the global GTT, or otherwise update device
information. If the objects are clean, we never need the wakeref, so
avoid taking until required.
Chris Wilson [Sat, 2 May 2020 17:35:12 +0000 (18:35 +0100)]
drm/i915/gt: Sanitize RPS interrupts upon resume
Currently we clear and disable the RPS pm interrupts on module load, and
presume that they remain disabled forevermore. However, the mask is
cleared on suspend and so after resume they may start showing up again
unexepectedly.
Chris Wilson [Fri, 1 May 2020 19:29:45 +0000 (20:29 +0100)]
drm/i915/gem: Try an alternate engine for relocations
If at first we don't succeed, try try again.
Not all engines may support the MI ops we need to perform asynchronous
relocation patching, and so we end up falling back to a synchronous
operation that has a liability of blocking. However, Tvrtko pointed out
we don't need to use the same engine to perform the relocations as we
are planning to execute the execbuf on, and so if we switch over to a
working engine, we can perform the relocation asynchronously. The user
execbuf will be queued after the relocations by virtue of fencing.
This patch creates a new context per execbuf requiring asynchronous
relocations on an unusable engines. This is perhaps a bit excessive and
can be ameliorated by a small context cache, but for the moment we only
need it for working around a little used engine on Sandybridge, and only
if relocations are actually required to an active batch buffer.
Now we just need to teach the relocation code to handle physical
addressing for gen2/3, and we should then have universal support!
Chris Wilson [Fri, 1 May 2020 19:29:44 +0000 (20:29 +0100)]
drm/i915/gem: Use a single chained reloc batches for a single execbuf
As we can now keep chaining together a relocation batch to process any
number of relocations, we can keep building that relocation batch for
all of the target vma. This avoiding emitting a new request into the
ring for each target, consuming precious ring space and a potential
stall.
v2: Propagate the failure from submitting the relocation batch.
Chris Wilson [Fri, 1 May 2020 19:29:43 +0000 (20:29 +0100)]
drm/i915/gem: Use chained reloc batches
The ring is a precious resource: we anticipate to only use a few hundred
bytes for a request, and only try to reserve that before we start. If we
go beyond our guess in building the request, then instead of waiting at
the start of execbuf before we hold any locks or other resources, we
may trigger a wait inside a critical region. One example is in using gpu
relocations, where currently we emit a new MI_BB_START from the ring
every time we overflow a page of relocation entries. However, instead of
insert the command into the precious ring, we can chain the next page of
relocation entries as MI_BB_START from the end of the previous.
v2: Delay the emit_bb_start until after all the chained vma
synchronisation is complete. Since the buffer pool batches are idle, this
_should_ be a no-op, but one day we may some fancy async GPU bindings
for new vma!
v3: Use pool/batch consitently, once we start thinking in terms of the
batch vma, use batch->obj.
v4: Explain the magic number 4.
Tvrtko spotted that we lose propagation of the error for failing to
submit the relocation request; that's easier to fix up in the next
patch.
Chris Wilson [Fri, 1 May 2020 14:51:20 +0000 (15:51 +0100)]
drm/i915: Implement vm_ops->access for gdb access into mmaps
gdb uses ptrace() to peek and poke bytes of the target's address space.
The driver must implement an vm_ops->access() handler or else gdb will
be unable to inspect the pointer and report it as out-of-bounds.
Worse than useless as it causes immediate suspicion of the valid GTT
pointer, distracting the poor programmer trying to find his bug.
v2: Write-protect readonly objects (Matthew).
Testcase: igt/gem_mmap_gtt/ptrace
Testcase: igt/gem_mmap_offset/ptrace Suggested-by: Kristian H. Kristensen <hoegsberg@google.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Cc: Maciej Patelczyk <maciej.patelczyk@intel.com> Cc: Kristian H. Kristensen <hoegsberg@google.com> Reviewed-by: Matthew Auld <matthew.auld@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200501145120.18830-1-chris@chris-wilson.co.uk
Chris Wilson [Fri, 1 May 2020 12:22:49 +0000 (13:22 +0100)]
drm/i915/gt: Make timeslicing an explicit engine property
In order to allow userspace to rely on timeslicing to reorder their
batches, we must support preemption of those user batches. Declare
timeslicing as an explicit property that is a combination of having the
kernel support and HW support.
Chris Wilson [Thu, 30 Apr 2020 11:18:12 +0000 (12:18 +0100)]
drm/i915/gt: Move the batch buffer pool from the engine to the gt
Since the introduction of 'soft-rc6', we aim to park the device quickly
and that results in frequent idling of the whole device. Currently upon
idling we free the batch buffer pool, and so this renders the cache
ineffective for many workloads. If we want to have an effective cache of
recently allocated buffers available for reuse, we need to decouple that
cache from the engine powermanagement and make it timer based. As there
is no reason then to keep it within the engine (where it once made
retirement order easier to track), we can move it up the hierarchy to the
owner of the memory allocations.
v2: Hook up to debugfs/drop_caches to clear the cache on demand.
Extend coverage of the blitter client by exercising conversion to and
from tiled sources. In the process we perform spot checks to verify that
the tiling/detiling is being applied correctly, along with position
invariance of the tiling parameters.
Signed-off-by: Zbigniew Kempczyński <zbigniew.kempczynski@intel.com> Cc: Chris Wilson <chris@chris-wilson.co.uk> Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Link: https://patchwork.freedesktop.org/patch/msgid/20200430064957.14942-1-chris@chris-wilson.co.uk
We reduced the clocks slowly after a boost event based on the
observation that the smoothness of animations suffered. However, since
reducing the evalution intervals, we should be able to respond to the
rapidly fluctuating workload of a simple desktop animation and so
restore the more aggressive downclocking.
Chris Wilson [Wed, 29 Apr 2020 20:54:45 +0000 (21:54 +0100)]
drm/i915/gt: Apply the aggressive downclocking to parking
We treat parking as a manual RPS timeout event, and downclock the GPU
for the next unpark and batch execution. However, having restored the
aggressive downclocking and observed that we have very light workloads
whose only interaction is through the manual parking events, carry over
the aggressive downclocking to the fake RPS events.
Chris Wilson [Wed, 29 Apr 2020 20:54:44 +0000 (21:54 +0100)]
drm/i915/gt: Switch to manual evaluation of RPS
As with the realisation for soft-rc6, we respond to idling the engines
within microseconds, far faster than the response times for HW RC6 and
RPS. Furthermore, our fast parking upon idle, prevents HW RPS from
running for many desktop workloads, as the RPS evaluation intervals are
on the order of tens of milliseconds, but the typical workload is just a
couple of milliseconds, but yet we still need to determine the best
frequency for user latency versus power.
Recognising that the HW evaluation intervals are a poor fit, and that
they were deprecated [in bspec at least] from gen10, start to wean
ourselves off them and replace the EI with a timer and our accurate
busy-stats. The principle benefit of manually evaluating RPS intervals
is that we can be more responsive for better performance and powersaving
for both spiky workloads and steady-state.
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1698 Fixes: 98479ada421a ("drm/i915/gt: Treat idling as a RPS downclock event") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: Andi Shyti <andi.shyti@intel.com> Reviewed-by: Andi Shyti <andi.shyti@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200429205446.3259-4-chris@chris-wilson.co.uk
Chris Wilson [Wed, 29 Apr 2020 20:54:41 +0000 (21:54 +0100)]
drm/i915/gt: Always enable busy-stats for execlists
In the near future, we will utilize the busy-stats on each engine to
approximate the C0 cycles of each, and use that as an input to a manual
RPS mechanism. That entails having busy-stats always enabled and so we
can remove the enable/disable routines and simplify the pmu setup. As a
consequence of always having the stats enabled, we can also show the
current active time via sysfs/engine/xcs/active_time_ns.
Chris Wilson [Wed, 29 Apr 2020 17:24:29 +0000 (18:24 +0100)]
drm/i915/gt: Keep a no-frills swappable copy of the default context state
We need to keep the default context state around to instantiate new
contexts (aka golden rendercontext), and we also keep it pinned while
the engine is active so that we can quickly reset a hanging context.
However, the default contexts are large enough to merit keeping in
swappable memory as opposed to kernel memory, so we store them inside
shmemfs. Currently, we use the normal GEM objects to create the default
context image, but we can throw away all but the shmemfs file.
This greatly simplifies the tricky power management code which wants to
run underneath the normal GT locking, and we definitely do not want to
use any high level objects that may appear to recurse back into the GT.
Though perhaps the primary advantage of the complex GEM object is that
we aggressively cache the mapping, but here we are recreating the
vm_area everytime time we unpark. At the worst, we add a lightweight
cache, but first find a microbenchmark that is impacted.
Having started to create some utility functions to make working with
shmemfs objects easier, we can start putting them to wider use, where
GEM objects are overkill, such as storing persistent error state.
Dan Carpenter [Wed, 29 Apr 2020 13:24:25 +0000 (16:24 +0300)]
drm/i915/selftests: fix error handling in __live_lrc_indirect_ctx_bb()
If intel_context_create() fails then it leads to an error pointer
dereference. I shuffled things around to make error handling easier.
Fixes: 1dd47b54baea ("drm/i915: Add live selftests for indirect ctx batchbuffers") Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> Reviewed-by: Andi Shyti <andi.shyti@intel.com> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Link: https://patchwork.freedesktop.org/patch/msgid/20200429132425.GE815283@mwanda
Chris Wilson [Tue, 28 Apr 2020 09:02:55 +0000 (10:02 +0100)]
drm/i915: Avoid dereferencing a dead context
Once the intel_context is closed, the GEM context may be freed and so
the link from intel_context.gem_context is invalid.
<3>[ 219.782944] BUG: KASAN: use-after-free in intel_engine_coredump_alloc+0x1bc3/0x2250 [i915]
<3>[ 219.782996] Read of size 8 at addr ffff8881d7dff0b8 by task kworker/0:1/12
drm/i915/gt: Avoid uninitialized use of rpcurupei in frequency_show
When building with clang + -Wuninitialized:
drivers/gpu/drm/i915/gt/debugfs_gt_pm.c:407:7: warning: variable
'rpcurupei' is uninitialized when used here [-Wuninitialized]
rpcurupei,
^~~~~~~~~
drivers/gpu/drm/i915/gt/debugfs_gt_pm.c:304:16: note: initialize the
variable 'rpcurupei' to silence this warning
u32 rpcurupei, rpcurup, rpprevup;
^
= 0
1 warning generated.
rpupei is assigned twice; based on the second argument to
intel_uncore_read, it seems this one should have been assigned to
rpcurupei.
Chris Wilson [Tue, 28 Apr 2020 18:47:50 +0000 (19:47 +0100)]
drm/i915/execlists: Track inflight CCID
The presumption is that by using a circular counter that is twice as
large as the maximum ELSP submission, we would never reuse the same CCID
for two inflight contexts.
However, if we continually preempt an active context such that it always
remains inflight, it can be resubmitted with an arbitrary number of
paired contexts. As each of its paired contexts will use a new CCID,
eventually it will wrap and submit two ELSP with the same CCID.
Rather than use a simple circular counter, switch over to a small bitmap
of inflight ids so we can avoid reusing one that is still potentially
active.
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1796 Fixes: 2935ed5339c4 ("drm/i915: Remove logical HW ID") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: <stable@vger.kernel.org> # v5.5+ Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200428184751.11257-2-chris@chris-wilson.co.uk
Chris Wilson [Tue, 28 Apr 2020 18:47:49 +0000 (19:47 +0100)]
drm/i915/execlists: Avoid reusing the same logical CCID
The bspec is confusing on the nature of the upper 32bits of the LRC
descriptor. Once upon a time, it said that it uses the upper 32b to
decide if it should perform a lite-restore, and so we must ensure that
each unique context submitted to HW is given a unique CCID [for the
duration of it being on the HW]. Currently, this is achieved by using
a small circular tag, and assigning every context submitted to HW a
new id. However, this tag is being cleared on repinning an inflight
context such that we end up re-using the 0 tag for multiple contexts.
To avoid accidentally clearing the CCID in the upper 32bits of the LRC
descriptor, split the descriptor into two dwords so we can update the
GGTT address separately from the CCID.
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/1796 Fixes: 2935ed5339c4 ("drm/i915: Remove logical HW ID") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com> Cc: <stable@vger.kernel.org> # v5.5+ Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200428184751.11257-1-chris@chris-wilson.co.uk
Chris Wilson [Tue, 28 Apr 2020 11:43:07 +0000 (12:43 +0100)]
drm/i915/selftests: Tweak the tolerance for clock ticks to 12.5%
Give a small bump for our tolerance on comparing the expected vs
measured clock ticks/time from 10% to 12.5% to accommodate a bad result
on Sandybridge that was off by 10.3%. Hopefully, that is the worst we
will see.
Matt Roper [Fri, 24 Apr 2020 23:14:23 +0000 (16:14 -0700)]
drm/i915: Use proper fault mask in interrupt postinstall too
The IRQ postinstall handling had open-coded pipe fault mask selection
that never got updated for gen11. Switch it to use
gen8_de_pipe_fault_mask() to ensure we don't miss updates for new
platforms.
The CS TAIL pointer should have been reset by reset_csb_pointers(), so
in this case it is likely that we have read back from the CPU cache and
so we must clflush our control over that page. In doing so, push the
sanitisation to the start of the GT sequence so that our poisoning is
assuredly before we start talking to the HW.
Chris Wilson [Mon, 27 Apr 2020 09:30:38 +0000 (10:30 +0100)]
drm/i915/gt: Check cacheline is valid before acquiring
The hwsp_cacheline pointer from i915_request is very, very flimsy. The
i915_request.timeline (and the hwsp_cacheline) are lost upon retiring
(after an RCU grace). Therefore we need to confirm that once we have the
right pointer for the cacheline, it is not in the process of being
retired and disposed of before we attempt to acquire a reference to the
cacheline.
<3>[ 547.208237] BUG: KASAN: use-after-free in active_debug_hint+0x6a/0x70 [i915]
<3>[ 547.208366] Read of size 8 at addr ffff88822a0d2710 by task gem_exec_parall/2536
Chris Wilson [Sun, 26 Apr 2020 09:42:31 +0000 (10:42 +0100)]
drm/i915/execlists: Check preempt-timeout target before submit_ports
We evaluate *active, which is a pointer into execlists->inflight[]
during dequeue to decide how long a preempt-timeout we need to apply.
However, as soon as we do the submit_ports, the HW may send its ACK
interrupt causing us to promote execlists->pending[] tp
execlists->inflight[], overwriting the value of *active. We know *active
is only stable until we submit (as we only submit when there is no
pending promotion).
Nick Desaulniers [Sun, 26 Apr 2020 21:42:15 +0000 (14:42 -0700)]
drm/i915: re-disable -Wframe-address
The top level Makefile disables this warning. When building an
i386_defconfig with Clang, this warning is triggered a whole bunch via
includes of headers from perf.
Mika Kuoppala [Fri, 24 Apr 2020 21:48:40 +0000 (00:48 +0300)]
drm/i915: Add live selftests for indirect ctx batchbuffers
Indirect ctx batchbuffers are a hw feature of which
batch can be run, by hardware, during context restoration stage.
Driver can setup a batchbuffer and also an offset into the
context image. When context image is marshalled from
memory to registers, and when the offset from the start of
context register state is equal of what driver pre-determined,
batch will run. So one can manipulate context restoration
process at cacheline granularity, given some limitations,
as you need to have rudimentaries in place before you can
run a batch.
Add selftest which will write the ring start register
to a canary spot. This will test that hardware will run a
batchbuffer for the context in question.
v2: request wait fix, naming (Chris)
v3: test order (Chris)
v4: rebase
Mika Kuoppala [Fri, 24 Apr 2020 23:05:46 +0000 (02:05 +0300)]
drm/i915: Add per ctx batchbuffer wa for timestamp
Restoration of a previous timestamp can collide
with updating the timestamp, causing a value corruption.
Combat this issue by using indirect ctx bb to
modify the context image during restoring process.
We can preload value into scratch register. From which
we then do the actual write with LRR. LRR is faster and
thus less error prone as probability of race drops.
v2: tidying (Chris)
v3: lrr for all engines
v4: grp
v5: reg bit
v6: wa_bb_offset, virtual engines (Chris)
Chris Wilson [Fri, 24 Apr 2020 19:14:10 +0000 (20:14 +0100)]
drm/i915: Drop rq->ring->vma peeking from error capture
We only hold the active spinlock while dumping the error state, and this
does not prevent another thread from retiring the request -- as it is
quite possible that despite us capturing the current state, the GPU has
completed the request. As such, it is dangerous to dereference state
below the request as it may already be freed, and the simplest way to
avoid the danger is not include it in the error state.
Chris Wilson [Fri, 24 Apr 2020 16:28:05 +0000 (17:28 +0100)]
drm/i915/gt: Use the RPM config register to determine clk frequencies
For many configuration details within RC6 and RPS we are programming
intervals for the internal clocks. From gen11, these clocks are
configuration via the RPM_CONFIG and so for convenience, we would like
to convert to/from more natural units (ns).
Chris Wilson [Fri, 24 Apr 2020 16:28:04 +0000 (17:28 +0100)]
drm/i915/gt: Trace RPS events
Add tracek to the RPS events (interrupts, worker, enabling, threshold
selection, frequency setting), so that if we have to debug reticent HW
we have some traces to start from.
Chris Wilson [Wed, 22 Apr 2020 00:17:01 +0000 (01:17 +0100)]
drm/i915/gt: Prefer soft-rc6 over RPS DOWN_TIMEOUT
The RPS DOWN_TIMEOUT interrupt is signaled after a period of rc6, and
upon receipt of that interrupt we reprogram the GPU clocks down to the
next idle notch [to help convserve power during rc6]. However, on
execlists, we benefit from soft-rc6 immediately parking the GPU and
setting idle frequencies upon idling [within a jiffie], and here the
interrupt prevents us from restarting from our last frequency.
In the process, we can simply opt for a static pm_events mask and rely
on the enable/disable interrupts to flush the worker on parking.
This will reduce the amount of oscillation observed during steady
workloads with microsleeps, as each time the rc6 timeout occurs we
immediately follow with a waitboost for a dropped frame.
Ville Syrjälä [Wed, 22 Apr 2020 16:19:17 +0000 (19:19 +0300)]
drm/i915: Have pfit calculations return an error code
Change intel_{gmch,pch}_panel_fitting() to return a normal
error vs. success int. We'll need this later to validate that
the margin properties aren't misconfigured.
Ville Syrjälä [Wed, 22 Apr 2020 16:19:16 +0000 (19:19 +0300)]
drm/i915: Pass connector state to pfit calculations
Pass the entire connector state to intel_{gmch,pch}_panel_fitting().
For now we just need to get at .scaling_mode but in the future we'll
want access to the margin properties as well.
Ville Syrjälä [Wed, 22 Apr 2020 16:19:14 +0000 (19:19 +0300)]
drm/i915: Use drm_rect to store the pfit window pos/size
Make things a bit more abstract by replacing the pch_pfit.pos/size
raw register values with a drm_rect. Makes it slighly more convenient
to eg. compute the scaling factors.
Ville Syrjälä [Wed, 22 Apr 2020 16:19:12 +0000 (19:19 +0300)]
drm/i915: Fix skl+ non-scaled pfit modes
Fix skl_update_scaler_crtc() to deal with different scaling
modes correctly. The current implementation assumes
DRM_MODE_SCALE_FULLSCREEN. Fortunately we don't expose any
border properties currently so the code does actually end
up doing the right thing (assigning a scaler for pfit).
The code does need to be fixed before any borders are
exposed.
Also we have redundant calls to skl_update_scaler_crtc() in
dp/hdmi .compute_config() which can be nuked. They were anyway
called before we had even computed the pfit state so were
basically nonsense. The real call we need to keep is in
intel_crtc_atomic_check().
v2: Deal witrh skl_update_scaler_crtc() in intel_dp_ycbcr420_config()
Chris Wilson [Wed, 22 Apr 2020 19:05:58 +0000 (20:05 +0100)]
drm/i915: Only close vma we open
The history of i915_vma_close() is confusing, as is its use. As the
lifetime of the i915_vma is currently bounded by the object it is
attached to, we needed a means of identify when a vma was no longer in
use by userspace (via the user's fd). This is further complicated by
that only ppgtt vma should be closed at the user's behest, as the ggtt
were always shared.
Now that we attach the vma to a lut on the user's context, the open
count does indicate how many unique and open context/vm are referencing
this vma from the user. As such, we can and should just use the
open_count to track when the vma is still in use by userspace.
It's a poor man's replacement for reference counting.
Chris Wilson [Wed, 22 Apr 2020 07:42:03 +0000 (08:42 +0100)]
drm/i915/selftests: Add request throughput measurement to perf
Under ideal circumstances, the driver should be able to keep the GPU
fully saturated with work. Measure how close to ideal we get under the
harshest of conditions with no user payload.
v2: Also measure throughput using only one thread.
Chris Wilson [Thu, 23 Apr 2020 08:59:39 +0000 (09:59 +0100)]
drm/i915/gt: Check carefully for an idle engine in wait-for-idle
intel_gt_wait_for_idle() tries to wait until all the outstanding requests
are retired and the GPU is idle. As a side effect of retiring requests,
we may submit more work to flush any pm barriers, and so the
wait-for-idle tries to flush the background pm work and catch the new
requests. However, if the work completed in the background before we
were able to flush, it would queue the extra barrier request without us
noticing -- and so we would return from wait-for-idle with one request
remaining. (This breaks e.g. record_default_state where we need to wait
until that barrier is retired, and it may slow suspend down by causing
us to wait on the background retirement worker as opposed to immediately
retiring the barrier.)
However, since we track if there has been a submission since the engine
pm barrier, we can very quickly detect if the idle barrier is still
outstanding.
Chris Wilson [Thu, 23 Apr 2020 11:53:15 +0000 (12:53 +0100)]
drm/i915/gt: Carefully order virtual_submission_tasklet
During the virtual engine's submission tasklet, we take the request and
insert into the submission queue on each of our siblings. This seems
quite simply, and so no problems with ordering. However, the sibling
execlists' submission tasklets may run concurrently with the virtual
engine's tasklet, submitting the request to HW before the virtual
finishes its task of telling all the siblings. If this happens, the
sibling tasklet may *reorder* the ve->sibling[] array that the virtual
engine tasklet is processing. This can *only* reorder within the
elements already processed by the virtual engine, nevertheless the
race is detected by KCSAN:
[ 185.580014] BUG: KCSAN: data-race in execlists_dequeue [i915] / virtual_submission_tasklet [i915]
[ 185.580054]
[ 185.580076] write to 0xffff8881f1919860 of 8 bytes by interrupt on cpu 2:
[ 185.580553] execlists_dequeue+0x6ad/0x1600 [i915]
[ 185.581044] __execlists_submission_tasklet+0x48/0x60 [i915]
[ 185.581517] execlists_submission_tasklet+0xd3/0x170 [i915]
[ 185.581554] tasklet_action_common.isra.0+0x42/0x90
[ 185.581585] __do_softirq+0xc8/0x206
[ 185.581613] run_ksoftirqd+0x15/0x20
[ 185.581641] smpboot_thread_fn+0x15a/0x270
[ 185.581669] kthread+0x19a/0x1e0
[ 185.581695] ret_from_fork+0x1f/0x30
[ 185.581717]
[ 185.581736] read to 0xffff8881f1919860 of 8 bytes by interrupt on cpu 0:
[ 185.582231] virtual_submission_tasklet+0x10e/0x5c0 [i915]
[ 185.582265] tasklet_action_common.isra.0+0x42/0x90
[ 185.582291] __do_softirq+0xc8/0x206
[ 185.582315] run_ksoftirqd+0x15/0x20
[ 185.582340] smpboot_thread_fn+0x15a/0x270
[ 185.582368] kthread+0x19a/0x1e0
[ 185.582395] ret_from_fork+0x1f/0x30
[ 185.582417]
We can prevent this race by checking for the ve->request after looking
up the sibling array.
Imre Deak [Wed, 22 Apr 2020 12:34:40 +0000 (15:34 +0300)]
drm/i915/icl: Fix timeout handling during TypeC AUX power well enabling
Fix the check for when an AUX power well enabling timeout is expected on
a legacy TypeC port.
Fixes: 89e01caac641 ("drm/i915: Use single set of AUX powerwell ops for gen11+") Cc: Matt Roper <matthew.d.roper@intel.com> Cc: José Roberto de Souza <jose.souza@intel.com> Signed-off-by: Imre Deak <imre.deak@intel.com> Reviewed-by: José Roberto de Souza <jose.souza@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200422123440.19522-1-imre.deak@intel.com
Chris Wilson [Wed, 22 Apr 2020 14:17:49 +0000 (15:17 +0100)]
drm/i915/execlists: Drop request-before-CS assertion
When we migrated to execlists, one of the conditions we wanted to test
for was whether the breadcrumb seqno was being written before the
breadcumb interrupt was delivered. This was following on from issues
observed on previous generations which were not so strongly ordered. With
the removal of the missed interrupt detection, we have not reliable
means of detecting the out-of-order seqno/interrupt but instead tried to
assert that the relationship between the CS event interrupt and the
breadwrite should be strongly ordered. However, Icelake proves it is
possible for the HW implementation to forget about minor little details
such as write ordering and so the order between *processing* the CS
event and the breadcrumb is unreliable.
Remove the unreliable assertion, but leave a debug telltale in case we
have reason to suspect.
Chris Wilson [Wed, 22 Apr 2020 07:28:05 +0000 (08:28 +0100)]
drm/i915/gem: Hold obj->vma.lock over for_each_ggtt_vma()
While the ggtt vma are protected by their object lifetime, the list
continues until it hits a non-ggtt vma, and that vma is not protected
and may be freed as we inspect it. Hence, we require the obj->vma.lock
to protect the list as we iterate.
An example of forgetting to hold the obj->vma.lock is
Now to take the spinlock during the list iteration, we need to break it
down into two phases. In the first phase under the lock, we cannot sleep
and so must defer the actual work to a second list, protected by the
ggtt->mutex.
We also need to hold the spinlock during creation of a new vma to
serialise with updates of the tiling on the object.
Reported-by: Dave Airlie <airlied@redhat.com> Fixes: 2850748ef876 ("drm/i915: Pull i915_vma_pin under the vm->mutex") Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Cc: Dave Airlie <airlied@redhat.com> Cc: <stable@vger.kernel.org> # v5.5+ Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Link: https://patchwork.freedesktop.org/patch/msgid/20200422072805.17340-1-chris@chris-wilson.co.uk
Chris Wilson [Wed, 22 Apr 2020 10:09:03 +0000 (11:09 +0100)]
drm/i915/selftests: Try to detect rollback during batchbuffer preemption
Since batch buffers dominant execution time, most preemption requests
should naturally occur during execution of a batch buffer. We wish to
verify that should a preemption occur within a batch buffer, when we
come to restart that batch buffer, it occurs at the interrupted
instruction and most importantly does not rollback to an earlier point.
v2: Do not clear the GPR at the start of the batch, but rely on them
being clear for new contexts.
Chris Wilson [Wed, 22 Apr 2020 08:38:55 +0000 (09:38 +0100)]
drm/i915/selftests: Disable heartbeat around RPS interrupt testing
For verifying reciving the EI interrupts, we need to hold the GPU in
very precise conditions (in terms of C0 cycles during the EI). If we
preempt the busy load to handle the heartbeat, this may perturb the busy
load causing us to miss the interrupt.
The other tests, while not as time sensitive, may also be slightly
perturbed, so apply the heartbeat protection across all the
measurements.
Chris Wilson [Tue, 21 Apr 2020 17:13:51 +0000 (18:13 +0100)]
drm/i915/selftests: Unroll the CS frequency loop
Having noticed that MI_BB_START is incurring a memory stall (see the
correlation with uncore frequency), we have to unroll the loop in order
to diminish the impact of the MI_BB_START on the instruction throughput.
Chris Wilson [Tue, 21 Apr 2020 09:25:04 +0000 (10:25 +0100)]
drm/i915/gt: Poison residual state [HWSP] across resume.
Since we may lose the content of any buffer when we relinquish control
of the system (e.g. suspend/resume), we have to be careful not to rely
on regaining control. A good method to detect when we might be using
garbage is by always injecting that garbage prior to first use on
load/resume/etc.
v2: Drop sanitize callback on cleanup
v3: Move seqno reset to timeline enter, so we reset all timelines.
However, this is done on every activation during runtime and not reset.
The similar level of paranoia we apply to correcting context state after
a period of inactivity.
Chris Wilson [Tue, 21 Apr 2020 14:22:36 +0000 (15:22 +0100)]
drm/i915/selftests: Disable C-states when measuring RPS frequency response
Let's isolate the impact of cpu frequency selection on determing the GPU
throughput in response to selection of RPS frequencies.
For real systems, we do have to be concerned with the impact of
integrating c-states, p-states and rp-states, but for the sake of
proving whether or not RPS works, one baby step at a time.
For the record, as one would hope, it does not seem to impact on the
measured performance, but we do it anyway to reduce the number of
variables. Later, we can extend the testing to encourage the the
cpu/pkg to try and sleep while the GPU is busy.
Chris Wilson [Tue, 21 Apr 2020 12:46:36 +0000 (13:46 +0100)]
drm/i915/selftests: Show the full scaling curve on failure
If we detect that the RPS end points do not scale perfectly, take the
time to measure all the in between values as well. We are aborting the
test, so we might as well spend the available time gathering critical
debug information instead.
projects / kernel.git / log