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, 79 insertions, 0 deletions

diff --git a/Documentation/fb/deferred_io.rst b/Documentation/fb/deferred_io.rst
new file mode 100644
index 000000000..7fc1933b0
--- /dev/null
+++ b/Documentation/fb/deferred_io.rst
@@ -0,0 +1,79 @@
+===========
+Deferred IO
+===========
+
+Deferred IO is a way to delay and repurpose IO. It uses host memory as a
+buffer and the MMU pagefault as a pretrigger for when to perform the device
+IO. The following example may be a useful explanation of how one such setup
+works:
+
+- userspace app like Xfbdev mmaps framebuffer
+- deferred IO and driver sets up fault and page_mkwrite handlers
+- userspace app tries to write to mmapped vaddress
+- we get pagefault and reach fault handler
+- fault handler finds and returns physical page
+- we get page_mkwrite where we add this page to a list
+- schedule a workqueue task to be run after a delay
+- app continues writing to that page with no additional cost. this is
+  the key benefit.
+- the workqueue task comes in and mkcleans the pages on the list, then
+  completes the work associated with updating the framebuffer. this is
+  the real work talking to the device.
+- app tries to write to the address (that has now been mkcleaned)
+- get pagefault and the above sequence occurs again
+
+As can be seen from above, one benefit is roughly to allow bursty framebuffer
+writes to occur at minimum cost. Then after some time when hopefully things
+have gone quiet, we go and really update the framebuffer which would be
+a relatively more expensive operation.
+
+For some types of nonvolatile high latency displays, the desired image is
+the final image rather than the intermediate stages which is why it's okay
+to not update for each write that is occurring.
+
+It may be the case that this is useful in other scenarios as well. Paul Mundt
+has mentioned a case where it is beneficial to use the page count to decide
+whether to coalesce and issue SG DMA or to do memory bursts.
+
+Another one may be if one has a device framebuffer that is in an usual format,
+say diagonally shifting RGB, this may then be a mechanism for you to allow
+apps to pretend to have a normal framebuffer but reswizzle for the device
+framebuffer at vsync time based on the touched pagelist.
+
+How to use it: (for applications)
+---------------------------------
+No changes needed. mmap the framebuffer like normal and just use it.
+
+How to use it: (for fbdev drivers)
+----------------------------------
+The following example may be helpful.
+
+1. Setup your structure. Eg::
+
+	static struct fb_deferred_io hecubafb_defio = {
+		.delay		= HZ,
+		.deferred_io	= hecubafb_dpy_deferred_io,
+	};
+
+The delay is the minimum delay between when the page_mkwrite trigger occurs
+and when the deferred_io callback is called. The deferred_io callback is
+explained below.
+
+2. Setup your deferred IO callback. Eg::
+
+	static void hecubafb_dpy_deferred_io(struct fb_info *info,
+					     struct list_head *pagelist)
+
+The deferred_io callback is where you would perform all your IO to the display
+device. You receive the pagelist which is the list of pages that were written
+to during the delay. You must not modify this list. This callback is called
+from a workqueue.
+
+3. Call init::
+
+	info->fbdefio = &hecubafb_defio;
+	fb_deferred_io_init(info);
+
+4. Call cleanup::
+
+	fb_deferred_io_cleanup(info);