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, 141 insertions, 0 deletions

diff --git a/Documentation/admin-guide/device-mapper/switch.rst b/Documentation/admin-guide/device-mapper/switch.rst
new file mode 100644
index 000000000..7dde06be1
--- /dev/null
+++ b/Documentation/admin-guide/device-mapper/switch.rst
@@ -0,0 +1,141 @@
+=========
+dm-switch
+=========
+
+The device-mapper switch target creates a device that supports an
+arbitrary mapping of fixed-size regions of I/O across a fixed set of
+paths.  The path used for any specific region can be switched
+dynamically by sending the target a message.
+
+It maps I/O to underlying block devices efficiently when there is a large
+number of fixed-sized address regions but there is no simple pattern
+that would allow for a compact representation of the mapping such as
+dm-stripe.
+
+Background
+----------
+
+Dell EqualLogic and some other iSCSI storage arrays use a distributed
+frameless architecture.  In this architecture, the storage group
+consists of a number of distinct storage arrays ("members") each having
+independent controllers, disk storage and network adapters.  When a LUN
+is created it is spread across multiple members.  The details of the
+spreading are hidden from initiators connected to this storage system.
+The storage group exposes a single target discovery portal, no matter
+how many members are being used.  When iSCSI sessions are created, each
+session is connected to an eth port on a single member.  Data to a LUN
+can be sent on any iSCSI session, and if the blocks being accessed are
+stored on another member the I/O will be forwarded as required.  This
+forwarding is invisible to the initiator.  The storage layout is also
+dynamic, and the blocks stored on disk may be moved from member to
+member as needed to balance the load.
+
+This architecture simplifies the management and configuration of both
+the storage group and initiators.  In a multipathing configuration, it
+is possible to set up multiple iSCSI sessions to use multiple network
+interfaces on both the host and target to take advantage of the
+increased network bandwidth.  An initiator could use a simple round
+robin algorithm to send I/O across all paths and let the storage array
+members forward it as necessary, but there is a performance advantage to
+sending data directly to the correct member.
+
+A device-mapper table already lets you map different regions of a
+device onto different targets.  However in this architecture the LUN is
+spread with an address region size on the order of 10s of MBs, which
+means the resulting table could have more than a million entries and
+consume far too much memory.
+
+Using this device-mapper switch target we can now build a two-layer
+device hierarchy:
+
+    Upper Tier - Determine which array member the I/O should be sent to.
+    Lower Tier - Load balance amongst paths to a particular member.
+
+The lower tier consists of a single dm multipath device for each member.
+Each of these multipath devices contains the set of paths directly to
+the array member in one priority group, and leverages existing path
+selectors to load balance amongst these paths.  We also build a
+non-preferred priority group containing paths to other array members for
+failover reasons.
+
+The upper tier consists of a single dm-switch device.  This device uses
+a bitmap to look up the location of the I/O and choose the appropriate
+lower tier device to route the I/O.  By using a bitmap we are able to
+use 4 bits for each address range in a 16 member group (which is very
+large for us).  This is a much denser representation than the dm table
+b-tree can achieve.
+
+Construction Parameters
+=======================
+
+    <num_paths> <region_size> <num_optional_args> [<optional_args>...] [<dev_path> <offset>]+
+	<num_paths>
+	    The number of paths across which to distribute the I/O.
+
+	<region_size>
+	    The number of 512-byte sectors in a region. Each region can be redirected
+	    to any of the available paths.
+
+	<num_optional_args>
+	    The number of optional arguments. Currently, no optional arguments
+	    are supported and so this must be zero.
+
+	<dev_path>
+	    The block device that represents a specific path to the device.
+
+	<offset>
+	    The offset of the start of data on the specific <dev_path> (in units
+	    of 512-byte sectors). This number is added to the sector number when
+	    forwarding the request to the specific path. Typically it is zero.
+
+Messages
+========
+
+set_region_mappings <index>:<path_nr> [<index>]:<path_nr> [<index>]:<path_nr>...
+
+Modify the region table by specifying which regions are redirected to
+which paths.
+
+<index>
+    The region number (region size was specified in constructor parameters).
+    If index is omitted, the next region (previous index + 1) is used.
+    Expressed in hexadecimal (WITHOUT any prefix like 0x).
+
+<path_nr>
+    The path number in the range 0 ... (<num_paths> - 1).
+    Expressed in hexadecimal (WITHOUT any prefix like 0x).
+
+R<n>,<m>
+    This parameter allows repetitive patterns to be loaded quickly. <n> and <m>
+    are hexadecimal numbers. The last <n> mappings are repeated in the next <m>
+    slots.
+
+Status
+======
+
+No status line is reported.
+
+Example
+=======
+
+Assume that you have volumes vg1/switch0 vg1/switch1 vg1/switch2 with
+the same size.
+
+Create a switch device with 64kB region size::
+
+    dmsetup create switch --table "0 `blockdev --getsz /dev/vg1/switch0`
+	switch 3 128 0 /dev/vg1/switch0 0 /dev/vg1/switch1 0 /dev/vg1/switch2 0"
+
+Set mappings for the first 7 entries to point to devices switch0, switch1,
+switch2, switch0, switch1, switch2, switch1::
+
+    dmsetup message switch 0 set_region_mappings 0:0 :1 :2 :0 :1 :2 :1
+
+Set repetitive mapping. This command::
+
+    dmsetup message switch 0 set_region_mappings 1000:1 :2 R2,10
+
+is equivalent to::
+
+    dmsetup message switch 0 set_region_mappings 1000:1 :2 :1 :2 :1 :2 :1 :2 \
+	:1 :2 :1 :2 :1 :2 :1 :2 :1 :2