Active busyness of an engine is calculated using gt timestamp and the
context switch in time. While capturing the gt timestamp, it's possible
that the context switches out. This race could result in an active
busyness value that is greater than the actual context runtime value by a
small amount. This leads to a negative delta and throws off busyness
calculations for the user.
If a subsequent count is smaller than the previous one, just return the
previous one, since we expect the busyness to catch up.
Fixes: 77cdd054dd ("drm/i915/pmu: Connect engine busyness stats from GuC to pmu")
Signed-off-by: Umesh Nerlige Ramappa <umesh.nerlige.ramappa@intel.com>
Reviewed-by: John Harrison <John.C.Harrison@Intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241127174006.190128-3-umesh.nerlige.ramappa@intel.com
Commit 255fc1703e ("drm/i915/gem: Calculate object page offset for
partial memory mapping") introduced a new offset, which accounts for
userspace mapping not starting from the beginning of object's scatterlist.
This works fine for cases where first object pte is larger than the new
offset - "r->sgt.curr" counter is set to the offset to match the difference
in the number of total pages. However, if object's first pte's size is
equal to or smaller than the offset, then information about the offset
in userspace is covered up by moving "r->sgt" pointer in remap_sg():
r->sgt.curr += PAGE_SIZE;
if (r->sgt.curr >= r->sgt.max)
r->sgt = __sgt_iter(__sg_next(r->sgt.sgp), use_dma(r->iobase));
This means that two or more pages from virtual memory are counted for
only one page in object's memory, because after moving "r->sgt" pointer
"r->sgt.curr" will be 0.
We should account for this mismatch by moving "r->sgt" pointer to the
next pte. For that we may use "r.sgt.max", which already holds the max
allowed size. This change also eliminates possible confusion, when
looking at i915_scatterlist.h and remap_io_sg() code: former has
scatterlist pointer definition, which differentiates "s.max" value
based on "dma" flag (sg_dma_len() is used only when the flag is
enabled), while latter uses sg_dma_len() indiscriminately.
This patch aims to resolve issue:
https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/12031
v3:
- instead of checking if r.sgt.curr would exceed allowed max, changed
the value in the while loop to be aligned with `dma` value
v4:
- remove unnecessary parent relation
v5:
- update commit message with explanation about page counting mismatch
and link to the issue
Signed-off-by: Krzysztof Karas <krzysztof.karas@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/upbjdavlbcxku63ns4vstp5kgbn2anxwewpmnppszgb67fn66t@tfclfgkqijue
Setting event_init to NULL is mostly done to detect when the driver is
partially working: i915 probed, but pmu is not registered. However,
checking for event_init is odd as it was supposed to always be set and
kernel/events/ would just crash if it found it set to NULL.
Since there's already a "closed" boolean, use that instead and extend
it's meaning to unregistered/unregistering.
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241104213512.2314930-3-lucas.demarchi@intel.com
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Implement pmu support for gen2 so that one can use intel_gpu_top
on it once again.
Gen2 lacks MI_MODE/MODE_IDLE so we'll have to do a bit more work
to determine the state of the engine:
- to determine if the ring contains unconsumed data we can simply
compare RING_TAIL vs. RING_HEAD
- also check RING_HEAD vs. ACTHD to catch cases where the hardware
is still executing a batch buffer but the ring head has already
caught up with the tail. Not entirely sure if that's actually
possible or not, but maybe it can happen if the batch buffer is
initiated from the very end of the ring? But even if not strictly
necessary there's no real harm in checking anyway.
- MI_WAIT_FOR_EVENT can be detected via a dedicated bit in RING_HEAD
v2: Use genX_ prefix rarther than suffix (Jani)
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241008214349.23331-5-ville.syrjala@linux.intel.com
Acked-by: Jani Nikula <jani.nikula@intel.com>
Add hwmon support for fan1_input attribute, which will expose fan speed
in RPM. With this in place we can monitor fan speed using lm-sensors tool.
$ sensors
i915-pci-0300
Adapter: PCI adapter
in0: 653.00 mV
fan1: 3833 RPM
power1: N/A (max = 43.00 W)
energy1: 32.02 kJ
v2: Handle overflow, add mutex protection and ABI documentation
Aesthetic adjustments (Riana)
v3: Change rotations data type, ABI date and version
v4: Fix wakeref leak
Drop switch case and simplify hwm_fan_xx() (Andi)
v5: Rework time calculation, aesthetic adjustments (Andy)
v6: Drop redundant overflow logic (Andy)
Split fan_input_read() into dedicated helper (Badal)
v7: Fix undefined reference to __udivdi3 for i386 (Andy)
Signed-off-by: Raag Jadav <raag.jadav@intel.com>
Reviewed-by: Riana Tauro <riana.tauro@intel.com>
Reviewed-by: Andi Shyti <andi.shyti@linux.intel.com>
Reviewed-by: Badal Nilawar <badal.nilawar@intel.com>
Signed-off-by: Anshuman Gupta <anshuman.gupta@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240823034548.2670032-1-raag.jadav@intel.com
The i915 driver generates sysfs entries for each engine of the
GPU in /sys/class/drm/cardX/engines/.
The process is straightforward: we loop over the UABI engines and
for each one, we:
- Create the object.
- Create basic files.
- If the engine supports timeslicing, create timeslice duration files.
- If the engine supports preemption, create preemption-related files.
- Create default value files.
Currently, if any of these steps fail, the process stops, and no
further sysfs files are created.
However, it's not necessary to stop the process on failure.
Instead, we can continue creating the remaining sysfs files for
the other engines. Even if some files fail to be created, the
list of engines can still be retrieved by querying i915.
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240819113140.325235-1-andi.shyti@linux.intel.com
Kernel test robot reports i915 can hit a warn in kvmalloc_node which has
a purpose of dissalowing crazy size kernel allocations. This was added in
7661809d49 ("mm: don't allow oversized kvmalloc() calls"):
/* Don't even allow crazy sizes */
if (WARN_ON_ONCE(size > INT_MAX))
return NULL;
This would be kind of okay since i915 at one point dropped the need for
making a shadow copy of the relocation list, but then it got re-added in
fd1500fcd4 ("Revert "drm/i915/gem: Drop relocation slowpath".") a year
after Linus added the above warning.
It is plausible that the issue was not seen until now because to trigger
gem_exec_reloc test requires a combination of an relatively older
generation hardware but with at least 8GiB of RAM installed. Probably even
more depending on runtime checks.
Lets cap what we allow userspace to pass in using the matching limit.
There should be no issue for real userspace since we are talking about
"crazy" number of relocations which have no practical purpose.
*) Well IGT tests might get upset but they can be easily adjusted.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/oe-lkp/202405151008.6ddd1aaf-oliver.sang@intel.com
Cc: Kees Cook <keescook@chromium.org>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Tvrtko Ursulin <tursulin@ursulin.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20240521101201.18978-1-tursulin@igalia.com
In commit a78a8da51b ("drm/ttm: replace busy placement with flags v6"),
__i915_ttm_get_pages was updated to use flags instead of the separate
'busy' placement list. However, the behaviour was subtly changed.
Originally, the function would attempt to use the preferred placement
without eviction, and give an opportunity to restart the operation
before falling back to allowing eviction.
This was unintentionally changed, as the preferred placement was not
given the TTM_PL_FLAG_DESIRED flag, and so eviction could be triggered
in that first pass. This caused thrashing, and a significant performance
regression on DG2 systems with small BAR. For example, Minecraft and
Team Fortress 2 would drop to single-digit framerates.
Restore the original behaviour by marking the initial placement as
desired on that first attempt. Also, rework this to use a separate
struct ttm_palcement, as the individual placements are marked 'const',
so hot-patching the flags is even more dodgy than before.
Thanks to Justin Brewer for bisecting this.
Fixes: a78a8da51b ("drm/ttm: replace busy placement with flags v6")
Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/11255
Signed-off-by: David Gow <david@davidgow.net>
Reviewed-by: Jonathan Cavitt <jonathan.cavitt@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/20240804091851.122186-3-david@davidgow.net
In commit a78a8da51b ("drm/ttm: replace busy placement with flags v6"),
the old system of having a separate placement list (for placements
which should be used without eviction) and a 'busy' placement list (for
placements which should be attempted if eviction is required) was
replaced with a new one where placements could be marked 'FALLBACK' (to
be attempted if eviction is required) or 'DESIRED' (to be attempted
first, but not if eviction is required).
i915 had always included the requested placement in the list of
'busy' placements: i.e., the placement could be used either if eviction
is required or not. But when the new system was put in place, the
requested (first) placement was marked 'DESIRED', so would never be used
if eviction became necessary. While a bug in the original commit
prevented this flag from working, when this was fixed in
4a0e7b3c ("drm/i915: fix applying placement flag"), it caused long hangs
on DG2 systems with small BAR.
Don't mark the requested placement DESIRED (or FALLBACK), allowing it to
be used in both situations. This matches the old behaviour, and resolves
the hangs.
Thanks to Justin Brewer for bisecting the issue.
Fixes: a78a8da51b ("drm/ttm: replace busy placement with flags v6")
Fixes: 4a0e7b3c37 ("drm/i915: fix applying placement flag")
Link: https://gitlab.freedesktop.org/drm/i915/kernel/-/issues/11255
Signed-off-by: David Gow <david@davidgow.net>
Reviewed-by: Jonathan Cavitt <jonathan.cavitt@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/20240804091851.122186-2-david@davidgow.net
Commit 05da7d9f71 ("drm/i915/gem: Downgrade stolen lmem setup
warning") adds a debug message where the "lmem_size" and
"dsm_base" variables are printed using the %lli identifier.
However, these variables are defined as resource_size_t, which
are unsigned long for 32-bit machines and unsigned long long for
64-bit machines.
The documentation (core-api/printk-formats.rst) recommends using
the %pa specifier for printing addresses and sizes of resources.
Replace %lli with %pa.
This patch also mutes the following sparse warning when compiling
with:
make W=1 ARCH=i386 drivers/gpu/drm/i915
>> drivers/gpu/drm/i915/gem/i915_gem_stolen.c:941:5: error: format '%lli'
expects argument of type 'long long int', but argument 5 has type
'resource_size_t' {aka 'unsigned int'} [-Werror=format=]
Signed-off-by: Andi Shyti <andi.shyti@linux.intel.com>
Cc: Jonathan Cavitt <jonathan.cavitt@intel.com>
Reviewed-by: Jonathan Cavitt <jonathan.cavitt@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240617184243.330231-3-andi.shyti@linux.intel.com
CI has been sporadically reporting the following issue triggered by
igt@i915_selftest@live@hangcheck on ADL-P and similar machines:
<6> [414.049203] i915: Running intel_hangcheck_live_selftests/igt_reset_evict_fence
...
<6> [414.068804] i915 0000:00:02.0: [drm] GT0: GUC: submission enabled
<6> [414.068812] i915 0000:00:02.0: [drm] GT0: GUC: SLPC enabled
<3> [414.070354] Unable to pin Y-tiled fence; err:-4
<3> [414.071282] i915_vma_revoke_fence:301 GEM_BUG_ON(!i915_active_is_idle(&fence->active))
...
<4>[ 609.603992] ------------[ cut here ]------------
<2>[ 609.603995] kernel BUG at drivers/gpu/drm/i915/gt/intel_ggtt_fencing.c:301!
<4>[ 609.604003] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
<4>[ 609.604006] CPU: 0 PID: 268 Comm: kworker/u64:3 Tainted: G U W 6.9.0-CI_DRM_14785-g1ba62f8cea9c+ #1
<4>[ 609.604008] Hardware name: Intel Corporation Alder Lake Client Platform/AlderLake-P DDR4 RVP, BIOS RPLPFWI1.R00.4035.A00.2301200723 01/20/2023
<4>[ 609.604010] Workqueue: i915 __i915_gem_free_work [i915]
<4>[ 609.604149] RIP: 0010:i915_vma_revoke_fence+0x187/0x1f0 [i915]
...
<4>[ 609.604271] Call Trace:
<4>[ 609.604273] <TASK>
...
<4>[ 609.604716] __i915_vma_evict+0x2e9/0x550 [i915]
<4>[ 609.604852] __i915_vma_unbind+0x7c/0x160 [i915]
<4>[ 609.604977] force_unbind+0x24/0xa0 [i915]
<4>[ 609.605098] i915_vma_destroy+0x2f/0xa0 [i915]
<4>[ 609.605210] __i915_gem_object_pages_fini+0x51/0x2f0 [i915]
<4>[ 609.605330] __i915_gem_free_objects.isra.0+0x6a/0xc0 [i915]
<4>[ 609.605440] process_scheduled_works+0x351/0x690
...
In the past, there were similar failures reported by CI from other IGT
tests, observed on other platforms.
Before commit 63baf4f3d5 ("drm/i915/gt: Only wait for GPU activity
before unbinding a GGTT fence"), i915_vma_revoke_fence() was waiting for
idleness of vma->active via fence_update(). That commit introduced
vma->fence->active in order for the fence_update() to be able to wait
selectively on that one instead of vma->active since only idleness of
fence registers was needed. But then, another commit 0d86ee3509
("drm/i915/gt: Make fence revocation unequivocal") replaced the call to
fence_update() in i915_vma_revoke_fence() with only fence_write(), and
also added that GEM_BUG_ON(!i915_active_is_idle(&fence->active)) in front.
No justification was provided on why we might then expect idleness of
vma->fence->active without first waiting on it.
The issue can be potentially caused by a race among revocation of fence
registers on one side and sequential execution of signal callbacks invoked
on completion of a request that was using them on the other, still
processed in parallel to revocation of those fence registers. Fix it by
waiting for idleness of vma->fence->active in i915_vma_revoke_fence().
Fixes: 0d86ee3509 ("drm/i915/gt: Make fence revocation unequivocal")
Closes: https://gitlab.freedesktop.org/drm/intel/issues/10021
Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@linux.intel.com>
Cc: stable@vger.kernel.org # v5.8+
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/20240603195446.297690-2-janusz.krzysztofik@linux.intel.com
