Although register tuning settings are generally implemented via the
workaround infrastructure, it turns out that the DRAW_WATERMARK register
is not properly saved/restored by hardware around power events (i.e.,
RC6 entry) so updates to the value cannot be applied in the usual
manner. New workaround Wa_16014892111 informs us that any tuning
updates to this register must instead be applied via an INDIRECT_CTX
batch buffer. This will ensure that the necessary value is re-applied
when a context begins running, even if an RC6 entry had wiped the
register back to hardware defaults since the last context ran.
Fixes: 6dc85721df ("drm/i915/dg2: Add additional tuning settings")
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/6642
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Balasubramani Vivekanandan <balasubramani.vivekanandan@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220823202449.83727-1-matthew.d.roper@intel.com
pxp will not start correctly until after mei_pxp bind completes and
intel_pxp_init_hw() is called.
Wait for the bind to complete before proceeding with startup.
This fixes a race condition during bootup where we observed a small
window for pxp commands to be sent, starting pxp before mei_pxp bind
completed.
Changes since v2:
- wait for pxp_component to bind instead of returning -EAGAIN (Daniele)
Changes since v1:
- check pxp_component instead of pxp_component_added (Daniele)
- pxp_component needs tee_mutex (Daniele)
- return -EAGAIN so caller knows to retry (Daniele)
Signed-off-by: Juston Li <justonli@chromium.org>
Reviewed-by: Andrzej Hajda <andrzej.hajda@intel.com>
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220818174205.2412730-1-justonli@chromium.org
Host Turbo operates at efficient frequency when GT is not idle unless
the user or workload has forced it to a higher level. Replicate the same
behavior in SLPC by allowing the algorithm to use efficient frequency.
We had disabled it during boot due to concerns that it might break
kernel ABI for min frequency. However, this is not the case since
SLPC will still abide by the (min,max) range limits.
With this change, min freq will be at efficient frequency level at init
instead of fused min (RPn). If user chooses to reduce min freq below the
efficient freq, we will turn off usage of efficient frequency and honor
the user request. When a higher value is written, it will get toggled
back again.
The patch also corrects the register which needs to be read for obtaining
the correct efficient frequency for Gen9+.
We see much better perf numbers with benchmarks like glmark2 with
efficient frequency usage enabled as expected.
v2: Address review comments (Rodrigo)
v3: with efficient frequency being dynamic, it is possible that the req
frequency may go beyond max freq. This will cause SLPC selftests to fail.
Add a FIXME there to start the test with [RPn, RP0] instead and restore
it afterwards.
BugLink: https://gitlab.freedesktop.org/drm/intel/-/issues/5468
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220820010832.15350-1-vinay.belgaumkar@intel.com
Add a delay, configurable via debugfs (default 34ms), to disable
scheduling of a context after the pin count goes to zero. Disable
scheduling is a costly operation as it requires synchronizing with
the GuC. So the idea is that a delay allows the user to resubmit
something before doing this operation. This delay is only done if
the context isn't closed and less than a given threshold
(default is 3/4) of the guc_ids are in use.
As temporary WA disable this feature for the selftests. Selftests are
very timing sensitive and any change in timing can cause failure. A
follow up patch will fixup the selftests to understand this delay.
Alan Previn: Matt Brost first introduced this series back in Oct 2021.
However no real world workload with measured performance impact was
available to prove the intended results. Today, this series is being
republished in response to a real world workload that benefited greatly
from it along with measured performance improvement.
Workload description: 36 containers were created on a DG2 device where
each container was performing a combination of 720p 3d game rendering
and 30fps video encoding. The workload density was configured in a way
that guaranteed each container to ALWAYS be able to render and
encode no less than 30fps with a predefined maximum render + encode
latency time. That means the totality of all 36 containers and their
workloads were not saturating the engines to their max (in order to
maintain just enough headrooom to meet the min fps and max latencies
of incoming container submissions).
Problem statement: It was observed that the CPU core processing the i915
soft IRQ work was experiencing severe load. Using tracelogs and an
instrumentation patch to count specific i915 IRQ events, it was confirmed
that the majority of the CPU cycles were caused by the
gen11_other_irq_handler() -> guc_irq_handler() code path. The vast
majority of the cycles was determined to be processing a specific G2H
IRQ: i.e. INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_DONE. These IRQs are sent
by GuC in response to i915 KMD sending H2G requests:
INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_SET. Those H2G requests are sent
whenever a context goes idle so that we can unpin the context from GuC.
The high CPU utilization % symptom was limiting density scaling.
Root Cause Analysis: Because the incoming execution buffers were spread
across 36 different containers (each with multiple contexts) but the
system in totality was NOT saturated to the max, it was assumed that each
context was constantly idling between submissions. This was causing
a thrashing of unpinning contexts from GuC at one moment, followed quickly
by repinning them due to incoming workload the very next moment. These
event-pairs were being triggered across multiple contexts per container,
across all containers at the rate of > 30 times per sec per context.
Metrics: When running this workload without this patch, we measured an
average of ~69K INTEL_GUC_ACTION_SCHED_CONTEXT_MODE_DONE events every 10
seconds or ~10 million times over ~25+ mins. With this patch, the count
reduced to ~480 every 10 seconds or about ~28K over ~10 mins. The
improvement observed is ~99% for the average counts per 10 seconds.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: Alan Previn <alan.previn.teres.alexis@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220817020511.2180747-3-alan.previn.teres.alexis@intel.com
If the GuC CTs are full and we need to stall the request submission
while waiting for space, we save the stalled request and where the stall
occurred; when the CTs have space again we pick up the request submission
from where we left off.
If a full GT reset occurs, the state of all contexts is cleared and all
non-guilty requests are unsubmitted, therefore we need to restart the
stalled request submission from scratch. To make sure that we do so,
clear the saved request after a reset.
Fixes note: the patch that introduced the bug is in 5.15, but no
officially supported platform had GuC submission enabled by default
in that kernel, so the backport to that particular version (and only
that one) can potentially be skipped.
Fixes: 925dc1cf58 ("drm/i915/guc: Implement GuC submission tasklet")
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: John Harrison <john.c.harrison@intel.com>
Cc: <stable@vger.kernel.org> # v5.15+
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220811210812.3239621-1-daniele.ceraolospurio@intel.com
There was a size check to warn if the GuC error state capture buffer
allocation would be too small to fit a reasonable amount of capture
data for the current platform. Unfortunately, the test was done too
early in the boot sequence and was actually testing 'if(-ENODEV >
size)'.
Move the check to be later. The check is only used to print a warning
message, so it doesn't really matter how early or late it is done.
Note that it is not possible to dynamically size the buffer because
the allocation needs to be done before the engine information is
available (at least, it would be in the intended two-phase GuC init
process).
Now that the check works, it is reporting size too small for newer
platforms. The check includes a 3x oversample multiplier to allow for
multiple error captures to be bufferd by GuC before i915 has a chance
to read them out. This is less important than simply being big enough
to fit the first capture.
So a) bump the default size to be large enough for one capture minimum
and b) make the warning only if one capture won't fit, instead use a
notice for the 3x size.
Note that the size estimate is a worst case scenario. Actual captures
will likely be smaller.
Lastly, use drm_warn istead of DRM_WARN as the former provides more
infmration and the latter is deprecated.
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Alan Previn <alan.previn.teres.alexis@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220728022028.2190627-3-John.C.Harrison@Intel.com
Some additional MMIO tuning settings have appeared in the bspec's
performance tuning guide section.
One of the tuning settings here is also documented as formal workaround
Wa_22012654132 for some steppings of DG2. However the tuning setting
applies to all DG2 variants and steppings, making it a superset of the
workaround.
v2:
- Move DRAW_WATERMARK to engine workaround section. It only moves into
the engine context on future platforms. (Lucas)
- CHICKEN_RASTER_2 needs to be handled as a masked register. (Lucas)
Bspec: 68331
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220816210601.2041572-2-matthew.d.roper@intel.com
The bspec performance tuning section gives recommended settings that the
driver should program for various MMIO registers. Although these
settings aren't "workarounds" we use the workaround infrastructure to do
this programming to make sure it is handled at the appropriate places
and doesn't conflict with any real workarounds.
Since more of these are starting to show up on recent platforms, it's a
good time to create a dedicated function to hold them so that there's
less ambiguity about how/where to implement new ones.
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220816210601.2041572-1-matthew.d.roper@intel.com
Crucible + recent Mesa seems to sometimes hit:
GEM_BUG_ON(num_ccs_blks > NUM_CCS_BLKS_PER_XFER)
And it looks like we can also trigger this with gem_lmem_swapping, if we
modify the test to use slightly larger object sizes.
Looking closer it looks like we have the following issues in
migrate_copy():
- We are using plain integer in various places, which we can easily
overflow with a large object.
- We pass the entire object size (when the src is lmem) into
emit_pte() and then try to copy it, which doesn't work, since we
only have a few fixed sized windows in which to map the pages and
perform the copy. With an object > 8M we therefore aren't properly
copying the pages. And then with an object > 64M we trigger the
GEM_BUG_ON(num_ccs_blks > NUM_CCS_BLKS_PER_XFER).
So it looks like our copy handling for any object > 8M (which is our
CHUNK_SZ) is currently broken on DG2.
Fixes: da0595ae91 ("drm/i915/migrate: Evict and restore the flatccs capable lmem obj")
Testcase: igt@gem_lmem_swapping
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Ramalingam C <ramalingam.c@intel.com>
Reviewed-by: Ramalingam C<ramalingam.c@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220805132240.442747-2-matthew.auld@intel.com
When the KMD sends a CLIENT_RESET request to GuC (as part of the
suspend sequence), GuC will mark the CTB buffer as 'UNUSED'. If the
KMD then checked the CTB queue, it would see a non-zero status value
and report the buffer as corrupted.
Technically, no G2H messages should be received once the CLIENT_RESET
has been sent. However, if a context was outstanding on an engine then
it would get reset and a reset notification would be sent. So, don't
actually treat UNUSED as a catastrophic error. Just flag it up as
unexpected and keep going.
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220728024225.2363663-7-John.C.Harrison@Intel.com
Having semaphores results in different behavior when a dependent request
is cancelled. In the case of semaphores the request could be on the HW
and complete successfully while without the request is held in the
driver and the error from the dependent request is propagated. Fix
live_preempt_cancel to take this behavior into account.
Also update live_preempt_cancel to use new function intel_context_ban
rather than intel_context_set_banned.
Signed-off-by: Matthew Brost <matthew.brost@intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220728024225.2363663-3-John.C.Harrison@Intel.com
TLB cache invalidation can happen on two different situations:
1. synchronously, at __vma_put_pages();
2. asynchronously.
On the first case, TLB cache invalidation happens inside
__vma_put_pages(). So, no need to do it later on.
However, on the second case, the pages will keep in memory
until __i915_vma_evict() is called.
So, we need to store the TLB data at struct i915_vma_resource,
in order to do a TLB cache invalidation before allowing
userspace to re-use the same memory.
So, i915_vma_resource_unbind() has gained a new parameter
in order to store the TLB data at the second case.
Document it.
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Signed-off-by: Mauro Carvalho Chehab <mchehab@kernel.org>
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/aa55eef7e63b8f3d0f69b525db2dd2eb87e9db6b.1658924372.git.mchehab@kernel.org
Invalidate TLB in batches, in order to reduce performance regressions.
Currently, every caller performs a full barrier around a TLB
invalidation, ignoring all other invalidations that may have already
removed their PTEs from the cache. As this is a synchronous operation
and can be quite slow, we cause multiple threads to contend on the TLB
invalidate mutex blocking userspace.
We only need to invalidate the TLB once after replacing our PTE to
ensure that there is no possible continued access to the physical
address before releasing our pages. By tracking a seqno for each full
TLB invalidate we can quickly determine if one has been performed since
rewriting the PTE, and only if necessary trigger one for ourselves.
That helps to reduce the performance regression introduced by TLB
invalidate logic.
[mchehab: rebased to not require moving the code to a separate file]
Cc: stable@vger.kernel.org
Fixes: 7938d61591 ("drm/i915: Flush TLBs before releasing backing store")
Suggested-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Signed-off-by: Chris Wilson <chris.p.wilson@intel.com>
Cc: Fei Yang <fei.yang@intel.com>
Signed-off-by: Mauro Carvalho Chehab <mchehab@kernel.org>
Acked-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/4e97ef5deb6739cadaaf40aa45620547e9c4ec06.1658924372.git.mchehab@kernel.org
PCI bar resize only works with 64 bit BAR so disable
this on 32-bit machine and resolve below compilation error:
drivers/gpu/drm/i915/gt/intel_region_lmem.c:94:23: error: result of
comparison of constant 4294967296 with expression of type
'resource_size_t' (aka 'unsigned int') is always false
[-Werror,-Wtautological-constant-out-of-range-compare]
root_res->start > 0x100000000ull)
Fixes: a91d1a17cd ("drm/i915: Add support for LMEM PCIe resizable bar")
Reported-by: Linux Kernel Functional Testing <lkft@linaro.org>
Tested-by: Linux Kernel Functional Testing <lkft@linaro.org>
Acked-by: Matthew Auld <matthew.auld@intel.com>
Signed-off-by: Nirmoy Das <nirmoy.das@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Signed-off-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220727173306.16247-1-nirmoy.das@intel.com
This patch re-introduces support for GuC v69 in parallel to v70. As this
is a quick fix, v69 has been re-introduced as the single "fallback" guc
version in case v70 is not available on disk and only for platforms that
are out of force_probe and require the GuC by default. All v69 specific
code has been labeled as such for easy identification, and the same was
done for all v70 functions for which there is a separate v69 version,
to avoid accidentally calling the wrong version via the unlabeled name.
When the fallback mode kicks in, a drm_notice message is printed in
dmesg to inform the user of the required update. The existing
logging of the fetch function has also been updated so that we no
longer complain immediately if we can't find a fw and we only throw an
error if the fetch of both the base and fallback blobs fails.
The plan is to follow this up with a more complex rework to allow for
multiple different GuC versions to be supported at the same time.
v2: reduce the fallback to platform that require it, switch to
firmware_request_nowarn(), improve logs.
Fixes: 2584b3549f ("drm/i915/guc: Update to GuC version 70.1.1")
Link: https://lists.freedesktop.org/archives/intel-gfx/2022-July/301640.html
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: John Harrison <John.C.Harrison@Intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Matt Roper <matthew.d.roper@intel.com>
Cc: Dave Airlie <airlied@gmail.com>
Cc: Michal Wajdeczko <michal.wajdeczko@intel.com>
Acked-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220718230732.1409641-1-daniele.ceraolospurio@intel.com
