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, 111 insertions, 0 deletions

diff --git a/Documentation/driver-api/md/raid5-cache.rst b/Documentation/driver-api/md/raid5-cache.rst
new file mode 100644
index 000000000..5f947cbc2
--- /dev/null
+++ b/Documentation/driver-api/md/raid5-cache.rst
@@ -0,0 +1,111 @@
+================
+RAID 4/5/6 cache
+================
+
+Raid 4/5/6 could include an extra disk for data cache besides normal RAID
+disks. The role of RAID disks isn't changed with the cache disk. The cache disk
+caches data to the RAID disks. The cache can be in write-through (supported
+since 4.4) or write-back mode (supported since 4.10). mdadm (supported since
+3.4) has a new option '--write-journal' to create array with cache. Please
+refer to mdadm manual for details. By default (RAID array starts), the cache is
+in write-through mode. A user can switch it to write-back mode by::
+
+	echo "write-back" > /sys/block/md0/md/journal_mode
+
+And switch it back to write-through mode by::
+
+	echo "write-through" > /sys/block/md0/md/journal_mode
+
+In both modes, all writes to the array will hit cache disk first. This means
+the cache disk must be fast and sustainable.
+
+write-through mode
+==================
+
+This mode mainly fixes the 'write hole' issue. For RAID 4/5/6 array, an unclean
+shutdown can cause data in some stripes to not be in consistent state, eg, data
+and parity don't match. The reason is that a stripe write involves several RAID
+disks and it's possible the writes don't hit all RAID disks yet before the
+unclean shutdown. We call an array degraded if it has inconsistent data. MD
+tries to resync the array to bring it back to normal state. But before the
+resync completes, any system crash will expose the chance of real data
+corruption in the RAID array. This problem is called 'write hole'.
+
+The write-through cache will cache all data on cache disk first. After the data
+is safe on the cache disk, the data will be flushed onto RAID disks. The
+two-step write will guarantee MD can recover correct data after unclean
+shutdown even the array is degraded. Thus the cache can close the 'write hole'.
+
+In write-through mode, MD reports IO completion to upper layer (usually
+filesystems) after the data is safe on RAID disks, so cache disk failure
+doesn't cause data loss. Of course cache disk failure means the array is
+exposed to 'write hole' again.
+
+In write-through mode, the cache disk isn't required to be big. Several
+hundreds megabytes are enough.
+
+write-back mode
+===============
+
+write-back mode fixes the 'write hole' issue too, since all write data is
+cached on cache disk. But the main goal of 'write-back' cache is to speed up
+write. If a write crosses all RAID disks of a stripe, we call it full-stripe
+write. For non-full-stripe writes, MD must read old data before the new parity
+can be calculated. These synchronous reads hurt write throughput. Some writes
+which are sequential but not dispatched in the same time will suffer from this
+overhead too. Write-back cache will aggregate the data and flush the data to
+RAID disks only after the data becomes a full stripe write. This will
+completely avoid the overhead, so it's very helpful for some workloads. A
+typical workload which does sequential write followed by fsync is an example.
+
+In write-back mode, MD reports IO completion to upper layer (usually
+filesystems) right after the data hits cache disk. The data is flushed to raid
+disks later after specific conditions met. So cache disk failure will cause
+data loss.
+
+In write-back mode, MD also caches data in memory. The memory cache includes
+the same data stored on cache disk, so a power loss doesn't cause data loss.
+The memory cache size has performance impact for the array. It's recommended
+the size is big. A user can configure the size by::
+
+	echo "2048" > /sys/block/md0/md/stripe_cache_size
+
+Too small cache disk will make the write aggregation less efficient in this
+mode depending on the workloads. It's recommended to use a cache disk with at
+least several gigabytes size in write-back mode.
+
+The implementation
+==================
+
+The write-through and write-back cache use the same disk format. The cache disk
+is organized as a simple write log. The log consists of 'meta data' and 'data'
+pairs. The meta data describes the data. It also includes checksum and sequence
+ID for recovery identification. Data can be IO data and parity data. Data is
+checksummed too. The checksum is stored in the meta data ahead of the data. The
+checksum is an optimization because MD can write meta and data freely without
+worry about the order. MD superblock has a field pointed to the valid meta data
+of log head.
+
+The log implementation is pretty straightforward. The difficult part is the
+order in which MD writes data to cache disk and RAID disks. Specifically, in
+write-through mode, MD calculates parity for IO data, writes both IO data and
+parity to the log, writes the data and parity to RAID disks after the data and
+parity is settled down in log and finally the IO is finished. Read just reads
+from raid disks as usual.
+
+In write-back mode, MD writes IO data to the log and reports IO completion. The
+data is also fully cached in memory at that time, which means read must query
+memory cache. If some conditions are met, MD will flush the data to RAID disks.
+MD will calculate parity for the data and write parity into the log. After this
+is finished, MD will write both data and parity into RAID disks, then MD can
+release the memory cache. The flush conditions could be stripe becomes a full
+stripe write, free cache disk space is low or free in-kernel memory cache space
+is low.
+
+After an unclean shutdown, MD does recovery. MD reads all meta data and data
+from the log. The sequence ID and checksum will help us detect corrupted meta
+data and data. If MD finds a stripe with data and valid parities (1 parity for
+raid4/5 and 2 for raid6), MD will write the data and parities to RAID disks. If
+parities are incompleted, they are discarded. If part of data is corrupted,
+they are discarded too. MD then loads valid data and writes them to RAID disks
+in normal way.