This new parameter let's the application choose how often the OA
buffer should be checked on the CPU side for data availability. Longer
polling period tend to reduce CPU overhead if the application does not
care about somewhat real time data collection.
v2: Allow disabling polling completely with 0 value (Lionel)
v3: Version the new parameter (Joonas)
v4: Rebase (Umesh)
v5: Make poll delay value of 0 invalid (Umesh)
v6:
- Describe poll_oa_period (Ashutosh)
- Fix comment for new poll parameter (Lionel)
- Drop open_flags in read_properties_unlocked (Lionel)
- Rename uapi parameter (Ashutosh)
v7: Reword the comment in uapi (Ashutosh)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200324185457.14635-4-umesh.nerlige.ramappa@intel.com
We're about to introduce an options to open the perf stream, giving
the user ability to configure how often it wants the kernel to poll
the OA registers for available data.
Right now the workaround against the OA tail pointer race condition
requires at least twice the internal kernel polling timer to make any
data available.
This changes introduce checks on the OA data written into the circular
buffer to make as much data as possible available on the first
iteration of the polling timer.
v2: Use OA_TAKEN macro without the gtt_offset (Lionel)
v3: (Umesh)
- Rebase
- Change report to report32 from below review
https://patchwork.freedesktop.org/patch/330704/?series=66697&rev=1
v4: (Ashutosh, Lionel)
- Fix checkpatch errors
- Fix aging_timestamp initialization
- Check for only one valid landed report
- Fix check for unlanded report
v5: (Ashutosh)
- Fix bug in accurately determining landed report.
- Optimize the check for landed reports by going as far as the
previously determined aged tail.
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Reviewed-by: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200324185457.14635-2-umesh.nerlige.ramappa@intel.com
On Gen11 powergating half the execution units is a functional
requirement when using the VME samplers. Not fullfilling this
requirement can lead to hangs.
This unfortunately plays fairly poorly with the NOA requirements. NOA
requires a stable power configuration to maintain its configuration.
As a result using OA (and NOA feeding into it) so far has required us
to use a power configuration that can work for all contexts. The only
power configuration fullfilling this is powergating half the execution
units.
This makes performance analysis for 3D workloads somewhat pointless.
Failing to find a solution that would work for everybody, this change
introduces a new i915-perf stream open parameter that punts the
decision off to userspace. If this parameter is omitted, the existing
Gen11 behavior remains (half EU array powergating).
This change takes the initiative to move all perf related sseu
configuration into i915_perf.c
v2: Make parameter priviliged if different from default
v3: Fix context modifying its sseu config while i915-perf is enabled
v4: Always consider global sseu a privileged operation (Tvrtko)
Override req_sseu point in intel_sseu_make_rpcs() (Tvrtko)
Remove unrelated changes (Tvrtko)
v5: Some typos (Tvrtko)
Process sseu param in read_properties_unlocked() (Tvrtko)
v6: Actually commit the bits from v5...
Fixup some checkpath warnings
v7: Only compare engine uabi field (Chris)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200317132222.2638719-3-lionel.g.landwerlin@intel.com
A little bit of history :
Back when i915-perf was introduced (4.13), there was no way to
dynamically add new OA configurations to i915. Only the generated
configs baked in at build time were allowed.
It quickly became obvious that we would need to allow applications
to upload their own configurations, for instance to be able to test
new ones, and so by the next stable version (4.14) we added uAPIs
to allow uploading new configurations.
When adding that capability, we took the opportunity to remove most
HW configurations except the TestOa one which is a configuration
IGT would rely on to verify that the HW is outputting correct
values. At the time it made sense to have that confiuration in at
the same time a given HW platform added to the i915-perf driver.
Now that IGT has become the reference point for HW configurations (see
commit 53f8f541ca ("lib: Add i915_perf library"), previously this was
located in the GPUTop repository), the need for having those
configurations in i915-perf is gone.
On the Mesa side, we haven't relied on this test configuration for a
while. The MDAPI library always required 4.14 feature level and always
loaded its configuration into i915.
I'm sure nobody will miss this generated stuff in i915 :)
v2: Fix selftests by creating an empty config
v3: Fix unlocking on allocation error (Dan Carpenter)
v4: Fixup checkpatch warnings
v5: Fix incorrect unlock in error path (Umesh)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200317132222.2638719-1-lionel.g.landwerlin@intel.com
Inside the general i915_oa_init_reg_state() we avoid using the
perf->mutex. However, we rely on perf->exclusive_stream being valid to
access at that point, and for that we have to control the race with
disabling perf. This relies on the disabling being a heavy barrier that
inspects all active contexts, after marking the perf->exclusive_stream
as not available. This should ensure that there are no more concurrent
accesses to the perf->exclusive_stream as we destroy it.
Mark up the races around the perf->exclusive_stream so that they stand
out much more. (And hopefully we will be running kcsan to start
validating that the only races we have are carefully controlled.)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200227085723.1961649-2-chris@chris-wilson.co.uk
Converts various instances of the printk drm logging macros to the
struct drm_device based logging macros in the drm/i915 folder using the
following coccinelle script that transforms based on the existence of
the struct drm_i915_private device pointer:
@@
identifier fn, T;
@@
fn(...) {
...
struct drm_i915_private *T = ...;
<+...
(
-DRM_INFO(
+drm_info(&T->drm,
...)
|
-DRM_ERROR(
+drm_err(&T->drm,
...)
|
-DRM_WARN(
+drm_warn(&T->drm,
...)
|
-DRM_DEBUG_KMS(
+drm_dbg_kms(&T->drm,
...)
|
-DRM_DEBUG_DRIVER(
+drm_dbg(&T->drm,
...)
|
-DRM_DEBUG_ATOMIC(
+drm_dbg_atomic(&T->drm,
...)
)
...+>
}
@@
identifier fn, T;
@@
fn(...,struct drm_i915_private *T,...) {
<+...
(
-DRM_INFO(
+drm_info(&T->drm,
...)
|
-DRM_ERROR(
+drm_err(&T->drm,
...)
|
-DRM_WARN(
+drm_warn(&T->drm,
...)
|
-DRM_DEBUG_DRIVER(
+drm_dbg(&T->drm,
...)
|
-DRM_DEBUG_KMS(
+drm_dbg_kms(&T->drm,
...)
|
-DRM_DEBUG_ATOMIC(
+drm_dbg_atomic(&T->drm,
...)
)
...+>
}
Checkpatch warnings were fixed manually.
Instances of the DRM_DEBUG macro were not converted due to lack of a
consensus of an analogous struct drm_device based macro.
References: https://lists.freedesktop.org/archives/dri-devel/2020-January/253381.html
Signed-off-by: Wambui Karuga <wambui.karugax@gmail.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200131093416.28431-2-wambui.karugax@gmail.com
Allocate only an internal intel_context for the kernel_context, forgoing
a global GEM context for internal use as we only require a separate
address space (for our own protection).
Now having weaned GT from requiring ce->gem_context, we can stop
referencing it entirely. This also means we no longer have to create random
and unnecessary GEM contexts for internal use.
GEM contexts are now entirely for tracking GEM clients, and intel_context
the execution environment on the GPU.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Andi Shyti <andi.shyti@intel.com>
Acked-by: Andi Shyti <andi.shyti@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191221160324.1073045-1-chris@chris-wilson.co.uk
As the engine->kernel_context is used within the engine-pm barrier, we
have to be careful when emitting requests outside of the barrier, as the
strict timeline locking rules do not apply. Instead, we must ensure the
engine_park() cannot be entered as we build the request, which is
simplest by taking an explicit engine-pm wakeref around the request
construction.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191125105858.1718307-1-chris@chris-wilson.co.uk
The design of the OA unit has been split into several units. We now
have a global unit (OAG) and a render specific unit (OAR). This leads
to some changes on how we program things. Some details :
OAR:
- has its own set of counter registers, they are per-context
saved/restored
- counters are not written to the circular OA buffer
- a snapshot of the counters can be acquired with
MI_RECORD_PERF_COUNT, or a single counter can be read with
MI_STORE_REGISTER_MEM.
OAG:
- has global counters that increment across context switches
- counters are written into the circular OA buffer (if requested)
v2: Fix checkpatch warnings on code style (Lucas)
v3: (Umesh)
- Update register from which tail, status and head are read
- Update logic to sample context reports
- Update whitelist mux and b counter regs
v4: Fix a bug when updating context image for new contexts (Umesh)
v5: Squash patch enabling save/restore of counters into context image
We want this so we can preempt performance queries and keep the
system responsive even when long running queries are ongoing. We
avoid doing it for all contexts.
- use LRI to modify context control (Chris)
- use MASKED_FIELD to program just the masked bits (Chris)
- disable save/restore of counters on cleanup (Chris)
v6: Do not use implicit parameters (Chris)
BSpec: 28727, 30021
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Acked-by: Chris Wilson <chris.p.wilson@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191025193746.47155-2-umesh.nerlige.ramappa@intel.com
The actual conditions are that we know the GPU is not accessing the
context, and we hold a pin on the context image to allow CPU access. We
used a fake lock on ce->pin_mutex so that we could try and use lockdep
to assert that access is serialised, but the various different
hardirq/softirq contexts where we need to *fake* holding the pin_mutex
are causing more trouble.
Still it would be nice if we did have a way to reassure ourselves that
the direct update to the context image is serialised with GPU execution.
In the meantime, stop lockdep complaining about false irq inversions.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111923
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191022122845.25038-1-chris@chris-wilson.co.uk
We would like to make use of perf in Vulkan. The Vulkan API is much
lower level than OpenGL, with applications directly exposed to the
concept of command buffers (pretty much equivalent to our batch
buffers). In Vulkan, queries are always limited in scope to a command
buffer. In OpenGL, the lack of command buffer concept meant that
queries' duration could span multiple command buffers.
With that restriction gone in Vulkan, we would like to simplify
measuring performance just by measuring the deltas between the counter
snapshots written by 2 MI_RECORD_PERF_COUNT commands, rather than the
more complex scheme we currently have in the GL driver, using 2
MI_RECORD_PERF_COUNT commands and doing some post processing on the
stream of OA reports, coming from the global OA buffer, to remove any
unrelated deltas in between the 2 MI_RECORD_PERF_COUNT.
Disabling preemption only apply to a single context with which want to
query performance counters for and is considered a privileged
operation, by default protected by CAP_SYS_ADMIN. It is possible to
enable it for a normal user by disabling the paranoid stream setting.
v2: Store preemption setting in intel_context (Chris)
v3: Use priorities to avoid preemption rather than the HW mechanism
v4: Just modify the port priority reporting function
v5: Add nopreempt flag on gem context and always flag requests
appropriately, regarless of OA reconfiguration.
Link: https://gitlab.freedesktop.org/mesa/mesa/merge_requests/932
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@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/20191014201404.22468-4-chris@chris-wilson.co.uk
Listing configurations at the moment is supported only through sysfs.
This might cause issues for applications wanting to list
configurations from a container where sysfs isn't available.
This change adds a way to query the number of configurations and their
content through the i915 query uAPI.
v2: Fix sparse warnings (Lionel)
Add support to query configuration using uuid (Lionel)
v3: Fix some inconsistency in uapi header (Lionel)
Fix unlocking when not locked issue (Lionel)
Add debug messages (Lionel)
v4: Fix missing unlock (Dan)
v5: Drop lock when copying config content to userspace (Chris)
v6: Drop lock when copying config list to userspace (Chris)
Fix deadlock when calling i915_perf_get_oa_config() under
perf.metrics_lock (Lionel)
Add i915_oa_config_get() (Chris)
Link: https://gitlab.freedesktop.org/mesa/mesa/merge_requests/932
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@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/20191014201404.22468-2-chris@chris-wilson.co.uk
We haven't run into issues with programming the global OA/NOA
registers configuration from CPU so far, but HW engineers actually
recommend doing this from the command streamer. On TGL in particular
one of the clock domain in which some of that programming goes might
not be powered when we poke things from the CPU.
Since we have a command buffer prepared for the execbuffer side of
things, we can reuse that approach here too.
This also allows us to significantly reduce the amount of time we hold
the main lock.
v2: Drop the global lock as much as possible
v3: Take global lock to pin global
v4: Create i915 request in emit_oa_config() to avoid deadlocks (Lionel)
v5: Move locking to the stream (Lionel)
v6: Move active reconfiguration request into i915_perf_stream (Lionel)
v7: Pin VMA outside request creation (Chris)
Lock VMA before move to active (Chris)
v8: Fix double free on stream->initial_oa_config_bo (Lionel)
Don't allow interruption when waiting on active config request
(Lionel)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@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/20191012072308.30312-3-chris@chris-wilson.co.uk
NOA configuration take some amount of time to apply. That amount of
time depends on the size of the GT. There is no documented time for
this. For example, past experimentations with powergating
configuration changes seem to indicate a 60~70us delay. We go with
500us as default for now which should be over the required amount of
time (according to HW architects).
v2: Don't forget to save/restore registers used for the wait (Chris)
v3: Name used CS_GPR registers (Chris)
Fix compile issue due to rebase (Lionel)
v4: Fix save/restore helpers (Umesh)
v5: Move noa_wait from drm_i915_private to i915_perf_stream (Lionel)
v6: Add missing struct declarations in i915_perf.h
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@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/20191012072308.30312-2-chris@chris-wilson.co.uk
Here we introduce a mechanism by which the execbuf part of the i915
driver will be able to request that a batch buffer containing the
programming for a particular OA config be created.
We'll execute these OA configuration buffers right before executing a
set of userspace commands so that a particular user batchbuffer be
executed with a given OA configuration.
This mechanism essentially allows the userspace driver to go through
several OA configuration without having to open/close the i915/perf
stream.
v2: No need for locking on object OA config object creation (Chris)
Flush cpu mapping of OA config (Chris)
v3: Properly deal with the perf_metric lock (Chris/Lionel)
v4: Fix oa config unref/put when not found (Lionel)
v5: Allocate BOs for configurations on the stream instead of globally
(Lionel)
v6: Fix 64bit division (Chris)
v7: Store allocated config BOs into the stream (Lionel)
Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@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/20191012072308.30312-1-chris@chris-wilson.co.uk
With the introduction of ctx->engines[] we allow multiple logical
contexts to be used on the same engine (e.g. with virtual engines).
According to bspec, aach logical context requires a unique tag in order
for context-switching to occur correctly between them. [Simple
experiments show that it is not so easy to trick the HW into performing
a lite-restore with matching logical IDs, though my memory from early
Broadwell experiments do suggest that it should be generating
lite-restores.]
We only need to keep a unique tag for the active lifetime of the
context, and for as long as we need to identify that context. The HW
uses the tag to determine if it should use a lite-restore (why not the
LRCA?) and passes the tag back for various status identifies. The only
status we need to track is for OA, so when using perf, we assign the
specific context a unique tag.
v2: Calculate required number of tags to fill ELSP.
Fixes: 976b55f0e1 ("drm/i915: Allow a context to define its set of engines")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=111895
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191004134015.13204-14-chris@chris-wilson.co.uk
Replace the struct_mutex requirement for pinning the i915_vma with the
local vm->mutex instead. Note that the vm->mutex is tainted by the
shrinker (we require unbinding from inside fs-reclaim) and so we cannot
allocate while holding that mutex. Instead we have to preallocate
workers to do allocate and apply the PTE updates after we have we
reserved their slot in the drm_mm (using fences to order the PTE writes
with the GPU work and with later unbind).
In adding the asynchronous vma binding, one subtle requirement is to
avoid coupling the binding fence into the backing object->resv. That is
the asynchronous binding only applies to the vma timeline itself and not
to the pages as that is a more global timeline (the binding of one vma
does not need to be ordered with another vma, nor does the implicit GEM
fencing depend on a vma, only on writes to the backing store). Keeping
the vma binding distinct from the backing store timelines is verified by
a number of async gem_exec_fence and gem_exec_schedule tests. The way we
do this is quite simple, we keep the fence for the vma binding separate
and only wait on it as required, and never add it to the obj->resv
itself.
Another consequence in reducing the locking around the vma is the
destruction of the vma is no longer globally serialised by struct_mutex.
A natural solution would be to add a kref to i915_vma, but that requires
decoupling the reference cycles, possibly by introducing a new
i915_mm_pages object that is own by both obj->mm and vma->pages.
However, we have not taken that route due to the overshadowing lmem/ttm
discussions, and instead play a series of complicated games with
trylocks to (hopefully) ensure that only one destruction path is called!
v2: Add some commentary, and some helpers to reduce patch churn.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191004134015.13204-4-chris@chris-wilson.co.uk
