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, 189 insertions, 0 deletions

diff --git a/include/drm/drm_panic.h b/include/drm/drm_panic.h
new file mode 100644
index 000000000..ac0e46b73
--- /dev/null
+++ b/include/drm/drm_panic.h
@@ -0,0 +1,189 @@
+/* SPDX-License-Identifier: GPL-2.0 or MIT */
+
+/*
+ * Copyright (c) 2024 Intel
+ * Copyright (c) 2024 Red Hat
+ */
+
+#ifndef __DRM_PANIC_H__
+#define __DRM_PANIC_H__
+
+#include <linux/module.h>
+#include <linux/types.h>
+#include <linux/iosys-map.h>
+
+#include <drm/drm_device.h>
+#include <drm/drm_fourcc.h>
+
+/**
+ * struct drm_scanout_buffer - DRM scanout buffer
+ *
+ * This structure holds the information necessary for drm_panic to draw the
+ * panic screen, and display it.
+ */
+struct drm_scanout_buffer {
+	/**
+	 * @format:
+	 *
+	 * drm format of the scanout buffer.
+	 */
+	const struct drm_format_info *format;
+
+	/**
+	 * @map:
+	 *
+	 * Virtual address of the scanout buffer, either in memory or iomem.
+	 * The scanout buffer should be in linear format, and can be directly
+	 * sent to the display hardware. Tearing is not an issue for the panic
+	 * screen.
+	 */
+	struct iosys_map map[DRM_FORMAT_MAX_PLANES];
+
+	/**
+	 * @pages: Optional, if the scanout buffer is not mapped, set this field
+	 * to the array of pages of the scanout buffer. The panic code will use
+	 * kmap_local_page_try_from_panic() to map one page at a time to write
+	 * all the pixels. This array shouldn't be allocated from the
+	 * get_scanoutbuffer() callback.
+	 * The scanout buffer should be in linear format.
+	 */
+	struct page **pages;
+
+	/**
+	 * @width: Width of the scanout buffer, in pixels.
+	 */
+	unsigned int width;
+
+	/**
+	 * @height: Height of the scanout buffer, in pixels.
+	 */
+	unsigned int height;
+
+	/**
+	 * @pitch: Length in bytes between the start of two consecutive lines.
+	 */
+	unsigned int pitch[DRM_FORMAT_MAX_PLANES];
+
+	/**
+	 * @set_pixel: Optional function, to set a pixel color on the
+	 * framebuffer. It allows to handle special tiling format inside the
+	 * driver. It takes precedence over the @map and @pages fields.
+	 */
+	void (*set_pixel)(struct drm_scanout_buffer *sb, unsigned int x,
+			  unsigned int y, u32 color);
+
+	/**
+	 * @private: private pointer that you can use in the callbacks
+	 * set_pixel()
+	 */
+	void *private;
+
+};
+
+#ifdef CONFIG_DRM_PANIC
+
+/**
+ * drm_panic_trylock - try to enter the panic printing critical section
+ * @dev: struct drm_device
+ * @flags: unsigned long irq flags you need to pass to the unlock() counterpart
+ *
+ * This function must be called by any panic printing code. The panic printing
+ * attempt must be aborted if the trylock fails.
+ *
+ * Panic printing code can make the following assumptions while holding the
+ * panic lock:
+ *
+ * - Anything protected by drm_panic_lock() and drm_panic_unlock() pairs is safe
+ *   to access.
+ *
+ * - Furthermore the panic printing code only registers in drm_dev_unregister()
+ *   and gets removed in drm_dev_unregister(). This allows the panic code to
+ *   safely access any state which is invariant in between these two function
+ *   calls, like the list of planes &drm_mode_config.plane_list or most of the
+ *   struct drm_plane structure.
+ *
+ * Specifically thanks to the protection around plane updates in
+ * drm_atomic_helper_swap_state() the following additional guarantees hold:
+ *
+ * - It is safe to deference the drm_plane.state pointer.
+ *
+ * - Anything in struct drm_plane_state or the driver's subclass thereof which
+ *   stays invariant after the atomic check code has finished is safe to access.
+ *   Specifically this includes the reference counted pointers to framebuffer
+ *   and buffer objects.
+ *
+ * - Anything set up by &drm_plane_helper_funcs.fb_prepare and cleaned up
+ *   &drm_plane_helper_funcs.fb_cleanup is safe to access, as long as it stays
+ *   invariant between these two calls. This also means that for drivers using
+ *   dynamic buffer management the framebuffer is pinned, and therefer all
+ *   relevant datastructures can be accessed without taking any further locks
+ *   (which would be impossible in panic context anyway).
+ *
+ * - Importantly, software and hardware state set up by
+ *   &drm_plane_helper_funcs.begin_fb_access and
+ *   &drm_plane_helper_funcs.end_fb_access is not safe to access.
+ *
+ * Drivers must not make any assumptions about the actual state of the hardware,
+ * unless they explicitly protected these hardware access with drm_panic_lock()
+ * and drm_panic_unlock().
+ *
+ * Return:
+ * %0 when failing to acquire the raw spinlock, nonzero on success.
+ */
+#define drm_panic_trylock(dev, flags) \
+	raw_spin_trylock_irqsave(&(dev)->mode_config.panic_lock, flags)
+
+/**
+ * drm_panic_lock - protect panic printing relevant state
+ * @dev: struct drm_device
+ * @flags: unsigned long irq flags you need to pass to the unlock() counterpart
+ *
+ * This function must be called to protect software and hardware state that the
+ * panic printing code must be able to rely on. The protected sections must be
+ * as small as possible. It uses the irqsave/irqrestore variant, and can be
+ * called from irq handler. Examples include:
+ *
+ * - Access to peek/poke or other similar registers, if that is the way the
+ *   driver prints the pixels into the scanout buffer at panic time.
+ *
+ * - Updates to pointers like &drm_plane.state, allowing the panic handler to
+ *   safely deference these. This is done in drm_atomic_helper_swap_state().
+ *
+ * - An state that isn't invariant and that the driver must be able to access
+ *   during panic printing.
+ */
+
+#define drm_panic_lock(dev, flags) \
+	raw_spin_lock_irqsave(&(dev)->mode_config.panic_lock, flags)
+
+/**
+ * drm_panic_unlock - end of the panic printing critical section
+ * @dev: struct drm_device
+ * @flags: irq flags that were returned when acquiring the lock
+ *
+ * Unlocks the raw spinlock acquired by either drm_panic_lock() or
+ * drm_panic_trylock().
+ */
+#define drm_panic_unlock(dev, flags) \
+	raw_spin_unlock_irqrestore(&(dev)->mode_config.panic_lock, flags)
+
+#else
+
+static inline bool drm_panic_trylock(struct drm_device *dev, unsigned long flags)
+{
+	return true;
+}
+
+static inline void drm_panic_lock(struct drm_device *dev, unsigned long flags) {}
+static inline void drm_panic_unlock(struct drm_device *dev, unsigned long flags) {}
+
+#endif
+
+#if defined(CONFIG_DRM_PANIC_SCREEN_QR_CODE)
+size_t drm_panic_qr_max_data_size(u8 version, size_t url_len);
+
+u8 drm_panic_qr_generate(const char *url, u8 *data, size_t data_len, size_t data_size,
+			 u8 *tmp, size_t tmp_size);
+#endif
+
+#endif /* __DRM_PANIC_H__ */