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, 114 insertions, 0 deletions

diff --git a/Documentation/admin-guide/device-mapper/writecache.rst b/Documentation/admin-guide/device-mapper/writecache.rst
new file mode 100644
index 000000000..60c16b7fd
--- /dev/null
+++ b/Documentation/admin-guide/device-mapper/writecache.rst
@@ -0,0 +1,114 @@
+=================
+Writecache target
+=================
+
+The writecache target caches writes on persistent memory or on SSD. It
+doesn't cache reads because reads are supposed to be cached in page cache
+in normal RAM.
+
+When the device is constructed, the first sector should be zeroed or the
+first sector should contain valid superblock from previous invocation.
+
+Constructor parameters:
+
+1. type of the cache device - "p" or "s"
+	- p - persistent memory
+	- s - SSD
+2. the underlying device that will be cached
+3. the cache device
+4. block size (4096 is recommended; the maximum block size is the page
+   size)
+5. the number of optional parameters (the parameters with an argument
+   count as two)
+
+	start_sector n		(default: 0)
+		offset from the start of cache device in 512-byte sectors
+	high_watermark n	(default: 50)
+		start writeback when the number of used blocks reach this
+		watermark
+	low_watermark x		(default: 45)
+		stop writeback when the number of used blocks drops below
+		this watermark
+	writeback_jobs n	(default: unlimited)
+		limit the number of blocks that are in flight during
+		writeback. Setting this value reduces writeback
+		throughput, but it may improve latency of read requests
+	autocommit_blocks n	(default: 64 for pmem, 65536 for ssd)
+		when the application writes this amount of blocks without
+		issuing the FLUSH request, the blocks are automatically
+		committed
+	autocommit_time ms	(default: 1000)
+		autocommit time in milliseconds. The data is automatically
+		committed if this time passes and no FLUSH request is
+		received
+	fua			(by default on)
+		applicable only to persistent memory - use the FUA flag
+		when writing data from persistent memory back to the
+		underlying device
+	nofua
+		applicable only to persistent memory - don't use the FUA
+		flag when writing back data and send the FLUSH request
+		afterwards
+
+		- some underlying devices perform better with fua, some
+		  with nofua. The user should test it
+	cleaner
+		when this option is activated (either in the constructor
+		arguments or by a message), the cache will not promote
+		new writes (however, writes to already cached blocks are
+		promoted, to avoid data corruption due to misordered
+		writes) and it will gradually writeback any cached
+		data. The userspace can then monitor the cleaning
+		process with "dmsetup status". When the number of cached
+		blocks drops to zero, userspace can unload the
+		dm-writecache target and replace it with dm-linear or
+		other targets.
+	max_age n
+		specifies the maximum age of a block in milliseconds. If
+		a block is stored in the cache for too long, it will be
+		written to the underlying device and cleaned up.
+	metadata_only
+		only metadata is promoted to the cache. This option
+		improves performance for heavier REQ_META workloads.
+	pause_writeback n	(default: 3000)
+		pause writeback if there was some write I/O redirected to
+		the origin volume in the last n milliseconds
+
+Status:
+
+1. error indicator - 0 if there was no error, otherwise error number
+2. the number of blocks
+3. the number of free blocks
+4. the number of blocks under writeback
+5. the number of read blocks
+6. the number of read blocks that hit the cache
+7. the number of write blocks
+8. the number of write blocks that hit uncommitted block
+9. the number of write blocks that hit committed block
+10. the number of write blocks that bypass the cache
+11. the number of write blocks that are allocated in the cache
+12. the number of write requests that are blocked on the freelist
+13. the number of flush requests
+14. the number of discarded blocks
+
+Messages:
+	flush
+		Flush the cache device. The message returns successfully
+		if the cache device was flushed without an error
+	flush_on_suspend
+		Flush the cache device on next suspend. Use this message
+		when you are going to remove the cache device. The proper
+		sequence for removing the cache device is:
+
+		1. send the "flush_on_suspend" message
+		2. load an inactive table with a linear target that maps
+		   to the underlying device
+		3. suspend the device
+		4. ask for status and verify that there are no errors
+		5. resume the device, so that it will use the linear
+		   target
+		6. the cache device is now inactive and it can be deleted
+	cleaner
+		See above "cleaner" constructor documentation.
+	clear_stats
+		Clear the statistics that are reported on the status line