sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

alb_break_llc() decides whether to break LLC preference during active
load balance. It does so by testing that every runnable fair task on the
source rq prefers its LLC:

	env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable

But the two counters cover different sets. nr_pref_llc_running is updated
in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued,
so it follows queued tasks. h_nr_runnable is updated in set_delayed()/
clear_delayed() and drops delay-dequeued tasks.

So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted
in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks,
alb_break_llc() returns false, and active balance is free to pull a task
off its preferred LLC. Active balance only moves runnable tasks, and this
is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB
skips the per-task test in can_migrate_task(). The runnable set is the one
we want.

Fix it on the counter side. A task should be counted in
nr_pref_llc_running exactly while it is both queued on its preferred LLC
(pref_llc_queued) and runnable (!sched_delayed). Define that membership
once in task_pref_llc_runnable(), and adjust the counter only through
pref_llc_running_inc()/pref_llc_running_dec() from the four sites that
change either input: account_llc_enqueue(), account_llc_dequeue(),
set_delayed() and clear_delayed(). Gating every update on the same
predicate keeps the delay, wake and dequeue paths from double-counting
or underflowing; see the comments at those sites for the ordering.

nr_llc_running and sd->llc_counts are not touched and stay on queued
semantics.

Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC")
Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/
Reported-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 210 insertions, 0 deletions

diff --git a/Documentation/gpu/amdgpu/debugfs.rst b/Documentation/gpu/amdgpu/debugfs.rst
new file mode 100644
index 000000000..151d8bfc7
--- /dev/null
+++ b/Documentation/gpu/amdgpu/debugfs.rst
@@ -0,0 +1,210 @@
+==============
+AMDGPU DebugFS
+==============
+
+The amdgpu driver provides a number of debugfs files to aid in debugging
+issues in the driver.  These are usually found in
+/sys/kernel/debug/dri/<num>.
+
+DebugFS Files
+=============
+
+amdgpu_benchmark
+----------------
+
+Run benchmarks using the DMA engine the driver uses for GPU memory paging.
+Write a number to the file to run the test.  The results are written to the
+kernel log.  VRAM is on device memory (dGPUs) or carve out (APUs) and GTT
+(Graphics Translation Tables) is system memory that is accessible by the GPU.
+The following tests are available:
+
+- 1: simple test, VRAM to GTT and GTT to VRAM
+- 2: simple test, VRAM to VRAM
+- 3: GTT to VRAM, buffer size sweep, powers of 2
+- 4: VRAM to GTT, buffer size sweep, powers of 2
+- 5: VRAM to VRAM, buffer size sweep, powers of 2
+- 6: GTT to VRAM, buffer size sweep, common display sizes
+- 7: VRAM to GTT, buffer size sweep, common display sizes
+- 8: VRAM to VRAM, buffer size sweep, common display sizes
+
+amdgpu_test_ib
+--------------
+
+Read this file to run simple IB (Indirect Buffer) tests on all kernel managed
+rings.  IBs are command buffers usually generated by userspace applications
+which are submitted to the kernel for execution on an particular GPU engine.
+This just runs the simple IB tests included in the kernel.  These tests
+are engine specific and verify that IB submission works.
+
+amdgpu_discovery
+----------------
+
+Provides raw access to the IP discovery binary provided by the GPU.  Read this
+file to access the raw binary.  This is useful for verifying the contents of
+the IP discovery table.  It is chip specific.
+
+amdgpu_vbios
+------------
+
+Provides raw access to the ROM binary image from the GPU.  Read this file to
+access the raw binary.  This is useful for verifying the contents of the
+video BIOS ROM.  It is board specific.
+
+amdgpu_evict_gtt
+----------------
+
+Evict all buffers from the GTT memory pool.  Read this file to evict all
+buffers from this pool.
+
+amdgpu_evict_vram
+-----------------
+
+Evict all buffers from the VRAM memory pool.  Read this file to evict all
+buffers from this pool.
+
+amdgpu_gpu_recover
+------------------
+
+Trigger a GPU reset.  Read this file to trigger reset the entire GPU.
+All work currently running  on the GPU will be lost.
+
+amdgpu_ring_<name>
+------------------
+
+Provides read access to the kernel managed ring buffers for each ring <name>.
+These are useful for debugging problems on a particular ring.  The ring buffer
+is how the CPU sends commands to the GPU.  The CPU writes commands into the
+buffer and then asks the GPU engine to process it.  This is the raw binary
+contents of the ring buffer.  Use a tool like UMR to decode the rings into human
+readable form.
+
+amdgpu_mqd_<name>
+-----------------
+
+Provides read access to the kernel managed MQD (Memory Queue Descriptor) for
+ring <name> managed by the kernel driver.  MQDs define the features of the ring
+and are used to store the ring's state when it is not connected to hardware.
+The driver writes the requested ring features and metadata (GPU addresses of
+the ring itself and associated buffers) to the MQD and the firmware uses the MQD
+to populate the hardware when the ring is mapped to a hardware slot.  Only
+available on engines which use MQDs.  This provides access to the raw MQD
+binary.
+
+amdgpu_error_<name>
+-------------------
+
+Provides an interface to set an error code on the dma fences associated with
+ring <name>.  The error code specified is propagated to all fences associated
+with the ring.  Use this to inject a fence error into a ring.
+
+amdgpu_pm_info
+--------------
+
+Provides human readable information about the power management features
+and state of the GPU.  This includes current GFX clock, Memory clock,
+voltages, average SoC power, temperature, GFX load, Memory load, SMU
+feature mask, VCN power state, clock and power gating features.
+
+amdgpu_firmware_info
+--------------------
+
+Lists the firmware versions for all firmwares used by the GPU.  Only
+entries with a non-0 version are valid.  If the version is 0, the firmware
+is not valid for the GPU.
+
+amdgpu_fence_info
+-----------------
+
+Shows the last signalled and emitted fence sequence numbers for each
+kernel driver managed ring.  Fences are associated with submissions
+to the engine.  Emitted fences have been submitted to the ring
+and signalled fences have been signalled by the GPU.  Rings with a
+larger emitted fence value have outstanding work that is still being
+processed by the engine that owns that ring.  When the emitted and
+signalled fence values are equal, the ring is idle.
+
+amdgpu_gem_info
+---------------
+
+Lists all of the PIDs using the GPU and the GPU buffers that they have
+allocated.  This lists the buffer size, pool (VRAM, GTT, etc.), and buffer
+attributes (CPU access required, CPU cache attributes, etc.).
+
+amdgpu_vm_info
+--------------
+
+Lists all of the PIDs using the GPU and the GPU buffers that they have
+allocated as well as the status of those buffers relative to that process'
+GPU virtual address space (e.g., evicted, idle, invalidated, etc.).
+
+amdgpu_sa_info
+--------------
+
+Prints out all of the suballocations (sa) by the suballocation manager in the
+kernel driver.  Prints the GPU address, size, and fence info associated
+with each suballocation.  The suballocations are used internally within
+the kernel driver for various things.
+
+amdgpu_<pool>_mm
+----------------
+
+Prints TTM information about the memory pool <pool>.
+
+amdgpu_vram
+-----------
+
+Provides direct access to VRAM.  Used by tools like UMR to inspect
+objects in VRAM.
+
+amdgpu_iomem
+------------
+
+Provides direct access to GTT memory.  Used by tools like UMR to inspect
+GTT memory.
+
+amdgpu_regs_*
+-------------
+
+Provides direct access to various register apertures on the GPU.  Used
+by tools like UMR to access GPU registers.
+
+amdgpu_regs2
+------------
+
+Provides an IOCTL interface used by UMR for interacting with GPU registers.
+
+
+amdgpu_sensors
+--------------
+
+Provides an interface to query GPU power metrics (temperature, average
+power, etc.).  Used by tools like UMR to query GPU power metrics.
+
+
+amdgpu_gca_config
+-----------------
+
+Provides an interface to query GPU details (Graphics/Compute Array config,
+PCI config, GPU family, etc.).  Used by tools like UMR to query GPU details.
+
+amdgpu_wave
+-----------
+
+Used to query GFX/compute wave information from the hardware.  Used by tools
+like UMR to query GFX/compute wave information.
+
+amdgpu_gpr
+----------
+
+Used to query GFX/compute GPR (General Purpose Register) information from the
+hardware.  Used by tools like UMR to query GPRs when debugging shaders.
+
+amdgpu_gprwave
+--------------
+
+Provides an IOCTL interface used by UMR for interacting with shader waves.
+
+amdgpu_fw_attestation
+---------------------
+
+Provides an interface for reading back firmware attestation records.