Sphinx reports htmldocs warnings:
Documentation/gpu/xe/xe_firmware:31: ./drivers/gpu/drm/xe/xe_guc_pc.c:76: ERROR: A level 2 section cannot be used here.
Documentation/gpu/xe/xe_firmware:31: ./drivers/gpu/drm/xe/xe_guc_pc.c:87: ERROR: A level 2 section cannot be used here.
The xe_guc_pc.c documentation is included inside xe_firmware.rst.
The headers in the C file currently use '=' underlines, which conflict with
the parent document's section levels.
Fix this by demoting "Frequency management" and "Render-C States" headers
from '=' to '-' to correctly nest them as subsections.
Build environment: Python 3.13.7 Sphinx 8.2.3 docutils 0.22.3
Signed-off-by: Swaraj Gaikwad <swarajgaikwad1925@gmail.com>
Link: https://patch.msgid.link/20251209094836.18589-1-swarajgaikwad1925@gmail.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
RPe is runtime-determined by PCODE and caching it caused stale values,
leading to incorrect GuC SLPC parameter settings.
Drop the cached rpe_freq field and query fresh values from hardware
on each use to ensure GuC SLPC parameters reflect current RPe.
v2: Remove cached RPe frequency field (Rodrigo)
v3: Remove extra variable (Vinay)
Modify function name (Vinay)
v4: Maintain a separate function for PVC (Rodrigo)
v5: Avoid RPn update while fetching RPe frequency (Rodrigo)
v6: Split platform-specific RPe comments (Vinay)
Closes: https://gitlab.freedesktop.org/drm/xe/kernel/-/issues/5166
Signed-off-by: Sk Anirban <sk.anirban@intel.com>
Reviewed-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Link: https://patch.msgid.link/20251112185153.3593145-5-sk.anirban@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
When Wa_22010954014 and Wa_14022085890 were first implemented, we didn't
have a device workaround infrastructure so we hacked them into the GT
workaround list. Now that we have proper device workaround support,
move them to the proper place. Note that Wa_14022085890 specifically
applies to BMG-G21 platforms, so this requires defining a BMG
subplatform to capture the correct subset of device IDs.
Reviewed-by: Gustavo Sousa <gustavo.sousa@intel.com>
Link: https://lore.kernel.org/r/20251013200944.2499947-40-matthew.d.roper@intel.com
Signed-off-by: Matt Roper <matthew.d.roper@intel.com>
Convert wait_for_pc_state() and wait_for_act_freq_limit() to
poll_timeout_us(). This brings 2 changes in behavior: Drop the
exponential wait and fix a potential much longer sleep.
usleep_range() will wait anywhere between `wait` and `wait << 1`, so
it's not correct to assume `slept += wait`. This code is not really
accurate. Pairing this with the exponential wait increase, it could be
waiting much longer than intended.
Reviewed-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Link: https://lore.kernel.org/r/20250922-xe-iopoll-v4-2-06438311a63f@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
GuC has an interface to set a power profile for the SLPC algorithm.
Base mode is default and ensures a balanced performance, power_saving
mode has conservative up/down thresholds and is suitable for use with
apps that typically need to be power efficient. This will result in
lower GT frequencies, thus consuming lower power.
Selected power profile will be displayed in this format:
$ cat power_profile
[base] power_saving
$ echo power_saving > power_profile
$ cat power_profile
base [power_saving]
v2: Address review comments (Rodrigo)
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://lore.kernel.org/r/20250903232120.390190-1-vinay.belgaumkar@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Limit GT max frequency to 2600MHz and wait for frequency to reduce
before proceeding with a transient flush. This is really only needed for
the transient flush: if L2 flush is needed due to 16023588340 then
there's no need to do this additional wait since we are already using
the bigger hammer.
v2: Use generic names, ensure user set max frequency requests wait
for flush to complete (Rodrigo)
v3:
- User requests wait via wait_var_event_timeout (Lucas)
- Close races on flush + user requests (Lucas)
- Fix xe_guc_pc_remove_flush_freq_limit() being called on last gt
rather than root gt (Lucas)
v4:
- Only apply the freq reducing part if a TDF is needed: L2 flush trumps
the need for waiting a lower frequency
Fixes: aaa08078e7 ("drm/xe/bmg: Apply Wa_22019338487")
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Link: https://lore.kernel.org/r/20250618-wa-22019338487-v5-4-b888388477f2@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
pc_set_mert_freq_cap() currently lock()/unlock() the mutex multiple times
to stash the current frequencies. It's not a problem since
xe_guc_pc_restore_stashed_freq() is guaranteed to be called only later
in the init sequence. However, now that we have _locked() variants for
this functions, use them and avoid potential issues when called from
other places or using the same pattern.
While at it, prefer and early return for the WA check to reduce
indentation.
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://lore.kernel.org/r/20250618-wa-22019338487-v5-2-b888388477f2@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
There are places in which the getters/setters are called one after the
other causing a multiple lock()/unlock(). These are not currently a
problem since they are all happening from the same thread, but there's a
race possibility as calls are added outside of the early init when the
max/min and stashed values need to be correlated.
Add the _locked() variants to prepare for that.
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://lore.kernel.org/r/20250618-wa-22019338487-v5-1-b888388477f2@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Define PMU events for GT frequency (actual and requested). The
instantaneous values for these frequencies will be displayed.
Following PMU events are being added:
xe_0000_00_02.0/gt-actual-frequency/ [Kernel PMU event]
xe_0000_00_02.0/gt-requested-frequency/ [Kernel PMU event]
Standard perf commands can be used to monitor GT frequency:
$ perf stat -e xe_0000_00_02.0/gt-requested-frequency,gt=0/ -I1000
1.001229762 1483 Mhz xe_0000_00_02.0/gt-requested-frequency,gt=0/
2.006175406 1483 Mhz xe_0000_00_02.0/gt-requested-frequency,gt=0/
v2: Use locks while storing/reading samples, keep track of multiple
clients (Lucas) and other general cleanup.
v3: Review comments (Lucas) and use event counts instead of mask for
active events.
v4: Add freq events to event_param_valid method (Riana)
v5: Use instantaneous values instead of aggregating (Lucas)
v6: Obtain fwake at init for freq events as well and use non fwake
variant method for reading requested freq to avoid lockdep issues (Lucas)
v7: Review comments (Rodrigo, Lucas)
Cc: Riana Tauro <riana.tauro@intel.com>
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Signed-off-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://lore.kernel.org/r/20250331204827.2535393-1-vinay.belgaumkar@intel.com
In a rare situation of thermal limit during resume, GuC can
be slow and run into delays like this:
xe 0000:00:02.0: [drm] GT1: excessive init time: 667ms! \
[status = 0x8002F034, timeouts = 0]
xe 0000:00:02.0: [drm] GT1: excessive init time: \
[freq = 100MHz (req = 800MHz), before = 100MHz, \
perf_limit_reasons = 0x1C001000]
xe 0000:00:02.0: [drm] *ERROR* GT1: GuC PC Start failed
------------[ cut here ]------------
xe 0000:00:02.0: [drm] GT1: Failed to start GuC PC: -EIO
When this happens, it will block entirely the GPU to be used.
So, let's try and with a huge timeout in the hope it comes back.
Also, let's collect some information on how long it is usually
taking on situations like this, so perhaps the time can be tuned
later.
Cc: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Cc: Jonathan Cavitt <jonathan.cavitt@intel.com>
Cc: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Jonathan Cavitt <jonathan.cavitt@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20250307160307.1093391-1-rodrigo.vivi@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Allow user to provide a low latency hint. When set, KMD sends a hint
to GuC which results in special handling for that process. SLPC will
ramp the GT frequency aggressively every time it switches to this
process.
We need to enable the use of SLPC Compute strategy during init, but
it will apply only to processes that set this bit during process
creation.
Improvement with this approach as below:
Before,
:~$ NEOReadDebugKeys=1 EnableDirectSubmission=0 clpeak --kernel-latency
Platform: Intel(R) OpenCL Graphics
Device: Intel(R) Graphics [0xe20b]
Driver version : 24.52.0 (Linux x64)
Compute units : 160
Clock frequency : 2850 MHz
Kernel launch latency : 283.16 us
After,
:~$ NEOReadDebugKeys=1 EnableDirectSubmission=0 clpeak --kernel-latency
Platform: Intel(R) OpenCL Graphics
Device: Intel(R) Graphics [0xe20b]
Driver version : 24.52.0 (Linux x64)
Compute units : 160
Clock frequency : 2850 MHz
Kernel launch latency : 63.38 us
Compute PR: https://github.com/intel/compute-runtime/pull/794
Mesa PR: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/33214
IGT PR: https://patchwork.freedesktop.org/patch/639989/
V10(Lucas):
- Remove doc from drm-uapi.rst
v9(Vinay):
- remove extra line, align commit message
v8(Vinay):
- Add separate example for using low latency hint
v7(Jose):
- Update UMD PR
- applicable to all gpus
V6:
- init flags, remove redundant flags check (MAuld)
V5:
- Move uapi doc to documentation and GuC ABI specific change (Rodrigo)
- Modify logic to restrict exec queue flags (MAuld)
V4:
- To make it clear, dont use exec queue word (Vinay)
- Correct typo in description of flag (Jose/Vinay)
- rename set_strategy api and replace ctx with exec queue(Vinay)
- Start with 0th bit to indentify user flags (Jose)
V3:
- Conver user flag to kernel internal flag and use (Oak)
- Support query config for use to check kernel support (Jose)
- Dont need to take runtime pm (Vinay)
V2:
- DRM_XE_EXEC_QUEUE_LOW_LATENCY_HINT 1 planned for other hint(Szymon)
- Add motivation to description (Lucas)
Acked-by: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20250228070224.739295-2-tejas.upadhyay@intel.com
Signed-off-by: Tejas Upadhyay <tejas.upadhyay@intel.com>
Enable DCC (Duty Cycle Control) in Lunar Lake.
DCC is the SLPC task that tries to keep
the GT from operating inefficiently when thermally constrained.
Although the recommendation is to enable it, LNL GuC is leaving
it disabled by default on LNL.
It would minimize the GT frequency oscillation on throttled
scenarios, which could potentially reduce latencies.
v2: Move set_policies call after wait for running state, so
we ensure it is not overwritten. (Vinay)
v3: Fix English in the commit message (Jonathan)
v4: Also set disable to 0 so DCC can really get into effect.
v5: Avoid lnl_ prefix (Vinay)
v6: Finish renaming...
Reviewed-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Reviewed-by: Jonathan Cavitt <jonathan.cavitt@intel.com> #v3
Link: https://patchwork.freedesktop.org/patch/msgid/20250115145053.1142023-2-rodrigo.vivi@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
xe_force_wake_get() now returns the reference count-incremented domain
mask. If it fails for individual domains, the return value will always
be 0. However, for XE_FORCEWAKE_ALL, it may return a non-zero value even
in the event of failure. Use helper xe_force_wake_ref_has_domain to
verify all domains are initialized or not. Update the return handling of
xe_force_wake_get() to reflect this behavior, and ensure that the return
value is passed as input to xe_force_wake_put().
v3
- return xe_wakeref_t instead of int in xe_force_wake_get()
- xe_force_wake_put() error doesn't need to be checked. It internally
WARNS on domain ack failure.
v5
- return unsigned int from xe_force_wake_get()
- Remove redundant xe_gt_WARN_ON
v6
- use helper xe_force_wake_ref_has_domain()
v7
- Fix commit message
v9
- Rebase
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Signed-off-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
Reviewed-by: Nirmoy Das <nirmoy.das@intel.com>
Reviewed-by: Badal Nilawar <badal.nilawar@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241014075601.2324382-17-himal.prasad.ghimiray@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
This WA requires us to limit media GT frequency requests to a certain
cap value during driver load. Freq limits are restored after load
completes, so perf will not be affected during normal operations.
During normal driver operation, this WA requires dummy writes to media
offset 0x380D8C after every ~63 GGTT writes. This will ensure completion
of the LMEM writes originating from Gunit.
During driver unload(before FLR), the WA requires that we set requested
frequency to the cap value again.
v3: Do not use WA number in function name. Call WA wrapper from xe_device.
Rename some variables, check for locks in the correct function (Rodrigo).
Ensure reset path is also covered for this WA.
v4: Fix BAT failure
v5: Add a function pointer for ggtt_ops (Michal W)
v6: Fix name collision and use static function (Rodrigo)
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Vinay Belgaumkar <vinay.belgaumkar@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240620224928.3986377-2-vinay.belgaumkar@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Introduce a very simple 'wedged' state where any attempt
to access the GPU is entirely blocked.
On some critical cases, like on gt_reset failure, we need to
block any other attempt to use the GPU. Otherwise we are at
a risk of reaching cases that would force us to reboot the machine.
So, when this cases are identified we corner and block any GPU
access. No IOCTL and not even another GT reset should be attempted.
The 'wedged' state in Xe is an end state with no way back.
Only a device "re-probe" (unbind + bind) can restore the GPU access.
v2: - s/wedged/busted (Lucas)
- use unbind+bind instead of module reload (Lucas)
- added more info on unbind operations and instruction on bug report
- only print the message once.
v3: - s/busted/wedged (Ashutosh, Tvrtko, Thomas)
- don't assume user has sudo and tee available (Lucas)
v4: - remove unnecessary cases around ct communication or migration.
Cc: Ashutosh Dixit <ashutosh.dixit@intel.com>
Cc: Tvrtko Ursulin <tursulin@ursulin.net>
Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Cc: Anshuman Gupta <anshuman.gupta@intel.com>
Reviewed-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>
Reviewed-by: Lucas De Marchi <lucas.demarchi@intel.com> #v2
Link: https://patchwork.freedesktop.org/patch/msgid/20240423221817.1285081-1-rodrigo.vivi@intel.com
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
The flags stored in the BO grew over time without following
much a naming pattern. First of all, get rid of the _BIT suffix that was
banned from everywhere else due to the guideline in
drivers/gpu/drm/i915/i915_reg.h that xe kind of follows:
Define bits using ``REG_BIT(N)``. Do **not** add ``_BIT`` suffix to the name.
Here the flags aren't for a register, but it's good practice to keep it
consistent.
Second divergence on names is the use or not of "CREATE". This is
because most of the flags are passed to xe_bo_create*() family of
functions, changing its behavior. However, since the flags are also
stored in the bo itself and checked elsewhere in the code, it seems
better to just omit the CREATE part.
With those 2 guidelines, all the flags are given the form
XE_BO_FLAG_<FLAG_NAME> with the following commands:
git grep -le "XE_BO_" -- drivers/gpu/drm/xe | xargs sed -i \
-e "s/XE_BO_\([_A-Z0-9]*\)_BIT/XE_BO_\1/g" \
-e 's/XE_BO_CREATE_/XE_BO_FLAG_/g'
git grep -le "XE_BO_" -- drivers/gpu/drm/xe | xargs sed -i -r \
-e 's/XE_BO_(DEFER_BACKING|SCANOUT|FIXED_PLACEMENT|PAGETABLE|NEEDS_CPU_ACCESS|NEEDS_UC|INTERNAL_TEST|INTERNAL_64K|GGTT_INVALIDATE)/XE_BO_FLAG_\1/g'
And then the defines in drivers/gpu/drm/xe/xe_bo.h are adjusted to
follow the coding style.
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240322142702.186529-3-lucas.demarchi@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
A force_wake_get failure means that the HW might not be awake for the
access we're doing; this can lead to an immediate error or it can be a
more subtle problem (e.g. a register read might return an incorrect
value that is still valid, leading the driver to make a wrong choice
instead of flagging an error).
We avoid an error from the force_wake function because callers might
handle or tolerate the error, but this only works if all callers
are checking the error code. The majority already do, but a few are not.
These are mainly falling into 3 categories, which are each handled
differently:
1) error capture: in this case we want to continue the capture, but we
log an info message in dmesg to notify the user that the capture
might have incorrect data.
2) ioctl: in this case we return a -EIO error to userspace
3) unabortable actions: these are scenarios where we can't simply abort
and retry and so it's better to just try it anyway because there is a
chance the HW is awake even with the failure. In this case we throw a
warning so we know there was a forcewake problem if something fails
down the line.
v2: use gt_WARN_ON where appropriate
Signed-off-by: Daniele Ceraolo Spurio <daniele.ceraolospurio@intel.com>
Cc: Tejas Upadhyay <tejas.upadhyay@intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Tejas Upadhyay <tejas.upadhyay@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240318154924.3453513-1-daniele.ceraolospurio@intel.com
