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, 99 insertions, 0 deletions

diff --git a/Documentation/admin-guide/syscall-user-dispatch.rst b/Documentation/admin-guide/syscall-user-dispatch.rst
new file mode 100644
index 000000000..c1768d9e8
--- /dev/null
+++ b/Documentation/admin-guide/syscall-user-dispatch.rst
@@ -0,0 +1,99 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+=====================
+Syscall User Dispatch
+=====================
+
+Background
+----------
+
+Compatibility layers like Wine need a way to efficiently emulate system
+calls of only a part of their process - the part that has the
+incompatible code - while being able to execute native syscalls without
+a high performance penalty on the native part of the process.  Seccomp
+falls short on this task, since it has limited support to efficiently
+filter syscalls based on memory regions, and it doesn't support removing
+filters.  Therefore a new mechanism is necessary.
+
+Syscall User Dispatch brings the filtering of the syscall dispatcher
+address back to userspace.  The application is in control of a flip
+switch, indicating the current personality of the process.  A
+multiple-personality application can then flip the switch without
+invoking the kernel, when crossing the compatibility layer API
+boundaries, to enable/disable the syscall redirection and execute
+syscalls directly (disabled) or send them to be emulated in userspace
+through a SIGSYS.
+
+The goal of this design is to provide very quick compatibility layer
+boundary crosses, which is achieved by not executing a syscall to change
+personality every time the compatibility layer executes.  Instead, a
+userspace memory region exposed to the kernel indicates the current
+personality, and the application simply modifies that variable to
+configure the mechanism.
+
+There is a relatively high cost associated with handling signals on most
+architectures, like x86, but at least for Wine, syscalls issued by
+native Windows code are currently not known to be a performance problem,
+since they are quite rare, at least for modern gaming applications.
+
+Since this mechanism is designed to capture syscalls issued by
+non-native applications, it must function on syscalls whose invocation
+ABI is completely unexpected to Linux.  Syscall User Dispatch, therefore
+doesn't rely on any of the syscall ABI to make the filtering.  It uses
+only the syscall dispatcher address and the userspace key.
+
+As the ABI of these intercepted syscalls is unknown to Linux, these
+syscalls are not instrumentable via ptrace or the syscall tracepoints.
+
+Interface
+---------
+
+A thread can setup this mechanism on supported kernels by executing the
+following prctl:
+
+  prctl(PR_SET_SYSCALL_USER_DISPATCH, <op>, <offset>, <length>, [selector])
+
+<op> is either PR_SYS_DISPATCH_EXCLUSIVE_ON/PR_SYS_DISPATCH_INCLUSIVE_ON
+or PR_SYS_DISPATCH_OFF, to enable and disable the mechanism globally for
+that thread.  When PR_SYS_DISPATCH_OFF is used, the other fields must be zero.
+
+For PR_SYS_DISPATCH_EXCLUSIVE_ON [<offset>, <offset>+<length>) delimit
+a memory region interval from which syscalls are always executed directly,
+regardless of the userspace selector.  This provides a fast path for the
+C library, which includes the most common syscall dispatchers in the native
+code applications, and also provides a way for the signal handler to return
+without triggering a nested SIGSYS on (rt\_)sigreturn.  Users of this
+interface should make sure that at least the signal trampoline code is
+included in this region. In addition, for syscalls that implement the
+trampoline code on the vDSO, that trampoline is never intercepted.
+
+For PR_SYS_DISPATCH_INCLUSIVE_ON [<offset>, <offset>+<length>) delimit
+a memory region interval from which syscalls are dispatched based on
+the userspace selector. Syscalls from outside of the range are always
+executed directly.
+
+[selector] is a pointer to a char-sized region in the process memory
+region, that provides a quick way to enable disable syscall redirection
+thread-wide, without the need to invoke the kernel directly.  selector
+can be set to SYSCALL_DISPATCH_FILTER_ALLOW or SYSCALL_DISPATCH_FILTER_BLOCK.
+Any other value should terminate the program with a SIGSYS.
+
+Additionally, a tasks syscall user dispatch configuration can be peeked
+and poked via the PTRACE_(GET|SET)_SYSCALL_USER_DISPATCH_CONFIG ptrace
+requests. This is useful for checkpoint/restart software.
+
+Security Notes
+--------------
+
+Syscall User Dispatch provides functionality for compatibility layers to
+quickly capture system calls issued by a non-native part of the
+application, while not impacting the Linux native regions of the
+process.  It is not a mechanism for sandboxing system calls, and it
+should not be seen as a security mechanism, since it is trivial for a
+malicious application to subvert the mechanism by jumping to an allowed
+dispatcher region prior to executing the syscall, or to discover the
+address and modify the selector value.  If the use case requires any
+kind of security sandboxing, Seccomp should be used instead.
+
+Any fork or exec of the existing process resets the mechanism to
+PR_SYS_DISPATCH_OFF.