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, 301 insertions, 0 deletions

diff --git a/arch/xtensa/lib/usercopy.S b/arch/xtensa/lib/usercopy.S
new file mode 100644
index 000000000..2c665c0b4
--- /dev/null
+++ b/arch/xtensa/lib/usercopy.S
@@ -0,0 +1,301 @@
+/*
+ *  arch/xtensa/lib/usercopy.S
+ *
+ *  Copy to/from user space (derived from arch/xtensa/lib/hal/memcopy.S)
+ *
+ *  DO NOT COMBINE this function with <arch/xtensa/lib/hal/memcopy.S>.
+ *  It needs to remain separate and distinct.  The hal files are part
+ *  of the Xtensa link-time HAL, and those files may differ per
+ *  processor configuration.  Patching the kernel for another
+ *  processor configuration includes replacing the hal files, and we
+ *  could lose the special functionality for accessing user-space
+ *  memory during such a patch.  We sacrifice a little code space here
+ *  in favor to simplify code maintenance.
+ *
+ *  This file is subject to the terms and conditions of the GNU General
+ *  Public License.  See the file "COPYING" in the main directory of
+ *  this archive for more details.
+ *
+ *  Copyright (C) 2002 Tensilica Inc.
+ */
+
+
+/*
+ * size_t __xtensa_copy_user (void *dst, const void *src, size_t len);
+ *
+ * The returned value is the number of bytes not copied.  Implies zero
+ * is success.
+ *
+ * The general case algorithm is as follows:
+ *   If the destination and source are both aligned,
+ *     do 16B chunks with a loop, and then finish up with
+ *     8B, 4B, 2B, and 1B copies conditional on the length.
+ *   If destination is aligned and source unaligned,
+ *     do the same, but use SRC to align the source data.
+ *   If destination is unaligned, align it by conditionally
+ *     copying 1B and 2B and then retest.
+ *   This code tries to use fall-through braches for the common
+ *     case of aligned destinations (except for the branches to
+ *     the alignment label).
+ *
+ * Register use:
+ *	a0/ return address
+ *	a1/ stack pointer
+ *	a2/ return value
+ *	a3/ src
+ *	a4/ length
+ *	a5/ dst
+ *	a6/ tmp
+ *	a7/ tmp
+ *	a8/ tmp
+ *	a9/ tmp
+ *	a10/ tmp
+ *	a11/ original length
+ */
+
+#include <linux/linkage.h>
+#include <asm/asmmacro.h>
+#include <asm/core.h>
+
+	.text
+ENTRY(__xtensa_copy_user)
+
+#if !XCHAL_HAVE_LOOPS && defined(__XTENSA_CALL0_ABI__)
+#define STACK_SIZE 4
+#else
+#define STACK_SIZE 0
+#endif
+	abi_entry(STACK_SIZE)
+	# a2/ dst, a3/ src, a4/ len
+	mov	a5, a2		# copy dst so that a2 is return value
+	mov	a11, a4		# preserve original len for error case
+.Lcommon:
+	bbsi.l	a2, 0, .Ldst1mod2 # if dst is 1 mod 2
+	bbsi.l	a2, 1, .Ldst2mod4 # if dst is 2 mod 4
+.Ldstaligned:	# return here from .Ldstunaligned when dst is aligned
+	srli	a7, a4, 4	# number of loop iterations with 16B
+				# per iteration
+	movi	a8, 3		  # if source is also aligned,
+	bnone	a3, a8, .Laligned # then use word copy
+	__ssa8	a3		# set shift amount from byte offset
+	bnez	a4, .Lsrcunaligned
+	movi	a2, 0		# return success for len==0
+	abi_ret(STACK_SIZE)
+
+/*
+ * Destination is unaligned
+ */
+
+.Ldst1mod2:	# dst is only byte aligned
+	bltui	a4, 7, .Lbytecopy	# do short copies byte by byte
+
+	# copy 1 byte
+EX(10f)	l8ui	a6, a3, 0
+	addi	a3, a3,  1
+EX(10f)	s8i	a6, a5,  0
+	addi	a5, a5,  1
+	addi	a4, a4, -1
+	bbci.l	a5, 1, .Ldstaligned	# if dst is now aligned, then
+					# return to main algorithm
+.Ldst2mod4:	# dst 16-bit aligned
+	# copy 2 bytes
+	bltui	a4, 6, .Lbytecopy	# do short copies byte by byte
+EX(10f)	l8ui	a6, a3, 0
+EX(10f)	l8ui	a7, a3, 1
+	addi	a3, a3,  2
+EX(10f)	s8i	a6, a5,  0
+EX(10f)	s8i	a7, a5,  1
+	addi	a5, a5,  2
+	addi	a4, a4, -2
+	j	.Ldstaligned	# dst is now aligned, return to main algorithm
+
+/*
+ * Byte by byte copy
+ */
+	.align	4
+	.byte	0		# 1 mod 4 alignment for LOOPNEZ
+				# (0 mod 4 alignment for LBEG)
+.Lbytecopy:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a4, .Lbytecopydone
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a4, .Lbytecopydone
+	add	a7, a3, a4	# a7 = end address for source
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lnextbyte:
+EX(10f)	l8ui	a6, a3, 0
+	addi	a3, a3, 1
+EX(10f)	s8i	a6, a5, 0
+	addi	a5, a5, 1
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a7, .Lnextbyte
+#endif /* !XCHAL_HAVE_LOOPS */
+.Lbytecopydone:
+	movi	a2, 0		# return success for len bytes copied
+	abi_ret(STACK_SIZE)
+
+/*
+ * Destination and source are word-aligned.
+ */
+	# copy 16 bytes per iteration for word-aligned dst and word-aligned src
+	.align	4		# 1 mod 4 alignment for LOOPNEZ
+	.byte	0		# (0 mod 4 alignment for LBEG)
+.Laligned:
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop1done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop1done
+	slli	a8, a7, 4
+	add	a8, a8, a3	# a8 = end of last 16B source chunk
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1:
+EX(10f)	l32i	a6, a3,  0
+EX(10f)	l32i	a7, a3,  4
+EX(10f)	s32i	a6, a5,  0
+EX(10f)	l32i	a6, a3,  8
+EX(10f)	s32i	a7, a5,  4
+EX(10f)	l32i	a7, a3, 12
+EX(10f)	s32i	a6, a5,  8
+	addi	a3, a3, 16
+EX(10f)	s32i	a7, a5, 12
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+	blt	a3, a8, .Loop1
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop1done:
+	bbci.l	a4, 3, .L2
+	# copy 8 bytes
+EX(10f)	l32i	a6, a3,  0
+EX(10f)	l32i	a7, a3,  4
+	addi	a3, a3,  8
+EX(10f)	s32i	a6, a5,  0
+EX(10f)	s32i	a7, a5,  4
+	addi	a5, a5,  8
+.L2:
+	bbci.l	a4, 2, .L3
+	# copy 4 bytes
+EX(10f)	l32i	a6, a3,  0
+	addi	a3, a3,  4
+EX(10f)	s32i	a6, a5,  0
+	addi	a5, a5,  4
+.L3:
+	bbci.l	a4, 1, .L4
+	# copy 2 bytes
+EX(10f)	l16ui	a6, a3,  0
+	addi	a3, a3,  2
+EX(10f)	s16i	a6, a5,  0
+	addi	a5, a5,  2
+.L4:
+	bbci.l	a4, 0, .L5
+	# copy 1 byte
+EX(10f)	l8ui	a6, a3,  0
+EX(10f)	s8i	a6, a5,  0
+.L5:
+	movi	a2, 0		# return success for len bytes copied
+	abi_ret(STACK_SIZE)
+
+/*
+ * Destination is aligned, Source is unaligned
+ */
+
+	.align	4
+	.byte	0		# 1 mod 4 alignement for LOOPNEZ
+				# (0 mod 4 alignment for LBEG)
+.Lsrcunaligned:
+	# copy 16 bytes per iteration for word-aligned dst and unaligned src
+	and	a10, a3, a8	# save unalignment offset for below
+	sub	a3, a3, a10	# align a3 (to avoid sim warnings only; not needed for hardware)
+EX(10f)	l32i	a6, a3, 0	# load first word
+#if XCHAL_HAVE_LOOPS
+	loopnez	a7, .Loop2done
+#else /* !XCHAL_HAVE_LOOPS */
+	beqz	a7, .Loop2done
+#if defined(__XTENSA_CALL0_ABI__)
+	s32i	a10, a1, 0
+	slli	a10, a7, 4
+	add	a10, a10, a3	# a10 = end of last 16B source chunk
+#else
+	slli	a12, a7, 4
+	add	a12, a12, a3	# a12 = end of last 16B source chunk
+#endif
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop2:
+EX(10f)	l32i	a7, a3,  4
+EX(10f)	l32i	a8, a3,  8
+	__src_b	a6, a6, a7
+EX(10f)	s32i	a6, a5,  0
+EX(10f)	l32i	a9, a3, 12
+	__src_b	a7, a7, a8
+EX(10f)	s32i	a7, a5,  4
+EX(10f)	l32i	a6, a3, 16
+	__src_b	a8, a8, a9
+EX(10f)	s32i	a8, a5,  8
+	addi	a3, a3, 16
+	__src_b	a9, a9, a6
+EX(10f)	s32i	a9, a5, 12
+	addi	a5, a5, 16
+#if !XCHAL_HAVE_LOOPS
+#if defined(__XTENSA_CALL0_ABI__)
+	blt	a3, a10, .Loop2
+	l32i	a10, a1, 0
+#else
+	blt	a3, a12, .Loop2
+#endif
+#endif /* !XCHAL_HAVE_LOOPS */
+.Loop2done:
+	bbci.l	a4, 3, .L12
+	# copy 8 bytes
+EX(10f)	l32i	a7, a3,  4
+EX(10f)	l32i	a8, a3,  8
+	__src_b	a6, a6, a7
+EX(10f)	s32i	a6, a5,  0
+	addi	a3, a3,  8
+	__src_b	a7, a7, a8
+EX(10f)	s32i	a7, a5,  4
+	addi	a5, a5,  8
+	mov	a6, a8
+.L12:
+	bbci.l	a4, 2, .L13
+	# copy 4 bytes
+EX(10f)	l32i	a7, a3,  4
+	addi	a3, a3,  4
+	__src_b	a6, a6, a7
+EX(10f)	s32i	a6, a5,  0
+	addi	a5, a5,  4
+	mov	a6, a7
+.L13:
+	add	a3, a3, a10	# readjust a3 with correct misalignment
+	bbci.l	a4, 1, .L14
+	# copy 2 bytes
+EX(10f)	l8ui	a6, a3,  0
+EX(10f)	l8ui	a7, a3,  1
+	addi	a3, a3,  2
+EX(10f)	s8i	a6, a5,  0
+EX(10f)	s8i	a7, a5,  1
+	addi	a5, a5,  2
+.L14:
+	bbci.l	a4, 0, .L15
+	# copy 1 byte
+EX(10f)	l8ui	a6, a3,  0
+EX(10f)	s8i	a6, a5,  0
+.L15:
+	movi	a2, 0		# return success for len bytes copied
+	abi_ret(STACK_SIZE)
+
+ENDPROC(__xtensa_copy_user)
+EXPORT_SYMBOL(__xtensa_copy_user)
+
+	.section .fixup, "ax"
+	.align	4
+
+/* a2 = original dst; a5 = current dst; a11= original len
+ * bytes_copied = a5 - a2
+ * retval = bytes_not_copied = original len - bytes_copied
+ * retval = a11 - (a5 - a2)
+ */
+
+
+10:
+	sub	a2, a5, a2	/* a2 <-- bytes copied */
+	sub	a2, a11, a2	/* a2 <-- bytes not copied */
+	abi_ret(STACK_SIZE)