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, 309 insertions, 0 deletions

diff --git a/drivers/rtc/rtc-mc146818-lib.c b/drivers/rtc/rtc-mc146818-lib.c
new file mode 100644
index 000000000..dbd2d5835
--- /dev/null
+++ b/drivers/rtc/rtc-mc146818-lib.c
@@ -0,0 +1,309 @@
+// SPDX-License-Identifier: GPL-2.0-only
+#include <linux/bcd.h>
+#include <linux/delay.h>
+#include <linux/export.h>
+#include <linux/mc146818rtc.h>
+
+#ifdef CONFIG_ACPI
+#include <linux/acpi.h>
+#endif
+
+#define UIP_RECHECK_DELAY		100	/* usec */
+#define UIP_RECHECK_DELAY_MS		(USEC_PER_MSEC / UIP_RECHECK_DELAY)
+#define UIP_RECHECK_LOOPS_MS(x)		(x / UIP_RECHECK_DELAY_MS)
+
+/*
+ * Execute a function while the UIP (Update-in-progress) bit of the RTC is
+ * unset. The timeout is configurable by the caller in ms.
+ *
+ * Warning: callback may be executed more then once.
+ */
+bool mc146818_avoid_UIP(void (*callback)(unsigned char seconds, void *param),
+			int timeout,
+			void *param)
+{
+	int i;
+	unsigned long flags;
+	unsigned char seconds;
+
+	for (i = 0; UIP_RECHECK_LOOPS_MS(i) < timeout; i++) {
+		spin_lock_irqsave(&rtc_lock, flags);
+
+		/*
+		 * Check whether there is an update in progress during which the
+		 * readout is unspecified. The maximum update time is ~2ms. Poll
+		 * for completion.
+		 *
+		 * Store the second value before checking UIP so a long lasting
+		 * NMI which happens to hit after the UIP check cannot make
+		 * an update cycle invisible.
+		 */
+		seconds = CMOS_READ(RTC_SECONDS);
+
+		if (CMOS_READ(RTC_FREQ_SELECT) & RTC_UIP) {
+			spin_unlock_irqrestore(&rtc_lock, flags);
+			udelay(UIP_RECHECK_DELAY);
+			continue;
+		}
+
+		/* Revalidate the above readout */
+		if (seconds != CMOS_READ(RTC_SECONDS)) {
+			spin_unlock_irqrestore(&rtc_lock, flags);
+			continue;
+		}
+
+		if (callback)
+			callback(seconds, param);
+
+		/*
+		 * Check for the UIP bit again. If it is set now then
+		 * the above values may contain garbage.
+		 */
+		if (CMOS_READ(RTC_FREQ_SELECT) & RTC_UIP) {
+			spin_unlock_irqrestore(&rtc_lock, flags);
+			udelay(UIP_RECHECK_DELAY);
+			continue;
+		}
+
+		/*
+		 * A NMI might have interrupted the above sequence so check
+		 * whether the seconds value has changed which indicates that
+		 * the NMI took longer than the UIP bit was set. Unlikely, but
+		 * possible and there is also virt...
+		 */
+		if (seconds != CMOS_READ(RTC_SECONDS)) {
+			spin_unlock_irqrestore(&rtc_lock, flags);
+			continue;
+		}
+		spin_unlock_irqrestore(&rtc_lock, flags);
+
+		if (UIP_RECHECK_LOOPS_MS(i) >= 100)
+			pr_warn("Reading current time from RTC took around %li ms\n",
+				UIP_RECHECK_LOOPS_MS(i));
+
+		return true;
+	}
+	return false;
+}
+EXPORT_SYMBOL_GPL(mc146818_avoid_UIP);
+
+/*
+ * If the UIP (Update-in-progress) bit of the RTC is set for more then
+ * 10ms, the RTC is apparently broken or not present.
+ */
+bool mc146818_does_rtc_work(void)
+{
+	return mc146818_avoid_UIP(NULL, 1000, NULL);
+}
+EXPORT_SYMBOL_GPL(mc146818_does_rtc_work);
+
+struct mc146818_get_time_callback_param {
+	struct rtc_time *time;
+	unsigned char ctrl;
+#ifdef CONFIG_ACPI
+	unsigned char century;
+#endif
+#ifdef CONFIG_MACH_DECSTATION
+	unsigned int real_year;
+#endif
+};
+
+static void mc146818_get_time_callback(unsigned char seconds, void *param_in)
+{
+	struct mc146818_get_time_callback_param *p = param_in;
+
+	/*
+	 * Only the values that we read from the RTC are set. We leave
+	 * tm_wday, tm_yday and tm_isdst untouched. Even though the
+	 * RTC has RTC_DAY_OF_WEEK, we ignore it, as it is only updated
+	 * by the RTC when initially set to a non-zero value.
+	 */
+	p->time->tm_sec = seconds;
+	p->time->tm_min = CMOS_READ(RTC_MINUTES);
+	p->time->tm_hour = CMOS_READ(RTC_HOURS);
+	p->time->tm_mday = CMOS_READ(RTC_DAY_OF_MONTH);
+	p->time->tm_mon = CMOS_READ(RTC_MONTH);
+	p->time->tm_year = CMOS_READ(RTC_YEAR);
+#ifdef CONFIG_MACH_DECSTATION
+	p->real_year = CMOS_READ(RTC_DEC_YEAR);
+#endif
+#ifdef CONFIG_ACPI
+	if (acpi_gbl_FADT.header.revision >= FADT2_REVISION_ID &&
+	    acpi_gbl_FADT.century) {
+		p->century = CMOS_READ(acpi_gbl_FADT.century);
+	} else {
+		p->century = 0;
+	}
+#endif
+
+	p->ctrl = CMOS_READ(RTC_CONTROL);
+}
+
+/**
+ * mc146818_get_time - Get the current time from the RTC
+ * @time: pointer to struct rtc_time to store the current time
+ * @timeout: timeout value in ms
+ *
+ * This function reads the current time from the RTC and stores it in the
+ * provided struct rtc_time. The timeout parameter specifies the maximum
+ * time to wait for the RTC to become ready.
+ *
+ * Return: 0 on success, -ETIMEDOUT if the RTC did not become ready within
+ * the specified timeout, or another error code if an error occurred.
+ */
+int mc146818_get_time(struct rtc_time *time, int timeout)
+{
+	struct mc146818_get_time_callback_param p = {
+		.time = time
+	};
+
+	if (!mc146818_avoid_UIP(mc146818_get_time_callback, timeout, &p)) {
+		memset(time, 0, sizeof(*time));
+		return -ETIMEDOUT;
+	}
+
+	if (!(p.ctrl & RTC_DM_BINARY) || RTC_ALWAYS_BCD)
+	{
+		time->tm_sec = bcd2bin(time->tm_sec);
+		time->tm_min = bcd2bin(time->tm_min);
+		time->tm_hour = bcd2bin(time->tm_hour);
+		time->tm_mday = bcd2bin(time->tm_mday);
+		time->tm_mon = bcd2bin(time->tm_mon);
+		time->tm_year = bcd2bin(time->tm_year);
+#ifdef CONFIG_ACPI
+		p.century = bcd2bin(p.century);
+#endif
+	}
+
+#ifdef CONFIG_MACH_DECSTATION
+	time->tm_year += p.real_year - 72;
+#endif
+
+#ifdef CONFIG_ACPI
+	if (p.century > 19)
+		time->tm_year += (p.century - 19) * 100;
+#endif
+
+	/*
+	 * Account for differences between how the RTC uses the values
+	 * and how they are defined in a struct rtc_time;
+	 */
+	if (time->tm_year <= 69)
+		time->tm_year += 100;
+
+	time->tm_mon--;
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(mc146818_get_time);
+
+/* AMD systems don't allow access to AltCentury with DV1 */
+static bool apply_amd_register_a_behavior(void)
+{
+#ifdef CONFIG_X86
+	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
+	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
+		return true;
+#endif
+	return false;
+}
+
+/* Set the current date and time in the real time clock. */
+int mc146818_set_time(struct rtc_time *time)
+{
+	unsigned long flags;
+	unsigned char mon, day, hrs, min, sec;
+	unsigned char save_control, save_freq_select;
+	unsigned int yrs;
+#ifdef CONFIG_MACH_DECSTATION
+	unsigned int real_yrs;
+#endif
+	unsigned char century = 0;
+
+	yrs = time->tm_year;
+	mon = time->tm_mon + 1;   /* tm_mon starts at zero */
+	day = time->tm_mday;
+	hrs = time->tm_hour;
+	min = time->tm_min;
+	sec = time->tm_sec;
+
+	if (yrs > 255)	/* They are unsigned */
+		return -EINVAL;
+
+#ifdef CONFIG_MACH_DECSTATION
+	real_yrs = yrs;
+	yrs = 72;
+
+	/*
+	 * We want to keep the year set to 73 until March
+	 * for non-leap years, so that Feb, 29th is handled
+	 * correctly.
+	 */
+	if (!is_leap_year(real_yrs + 1900) && mon < 3) {
+		real_yrs--;
+		yrs = 73;
+	}
+#endif
+
+#ifdef CONFIG_ACPI
+	if (acpi_gbl_FADT.header.revision >= FADT2_REVISION_ID &&
+	    acpi_gbl_FADT.century) {
+		century = (yrs + 1900) / 100;
+		yrs %= 100;
+	}
+#endif
+
+	/* These limits and adjustments are independent of
+	 * whether the chip is in binary mode or not.
+	 */
+	if (yrs > 169)
+		return -EINVAL;
+
+	if (yrs >= 100)
+		yrs -= 100;
+
+	spin_lock_irqsave(&rtc_lock, flags);
+	save_control = CMOS_READ(RTC_CONTROL);
+	spin_unlock_irqrestore(&rtc_lock, flags);
+	if (!(save_control & RTC_DM_BINARY) || RTC_ALWAYS_BCD) {
+		sec = bin2bcd(sec);
+		min = bin2bcd(min);
+		hrs = bin2bcd(hrs);
+		day = bin2bcd(day);
+		mon = bin2bcd(mon);
+		yrs = bin2bcd(yrs);
+		century = bin2bcd(century);
+	}
+
+	spin_lock_irqsave(&rtc_lock, flags);
+	save_control = CMOS_READ(RTC_CONTROL);
+	CMOS_WRITE((save_control|RTC_SET), RTC_CONTROL);
+	save_freq_select = CMOS_READ(RTC_FREQ_SELECT);
+	if (apply_amd_register_a_behavior())
+		CMOS_WRITE((save_freq_select & ~RTC_AMD_BANK_SELECT), RTC_FREQ_SELECT);
+	else
+		CMOS_WRITE((save_freq_select|RTC_DIV_RESET2), RTC_FREQ_SELECT);
+
+#ifdef CONFIG_MACH_DECSTATION
+	CMOS_WRITE(real_yrs, RTC_DEC_YEAR);
+#endif
+	CMOS_WRITE(yrs, RTC_YEAR);
+	CMOS_WRITE(mon, RTC_MONTH);
+	CMOS_WRITE(day, RTC_DAY_OF_MONTH);
+	CMOS_WRITE(hrs, RTC_HOURS);
+	CMOS_WRITE(min, RTC_MINUTES);
+	CMOS_WRITE(sec, RTC_SECONDS);
+#ifdef CONFIG_ACPI
+	if (acpi_gbl_FADT.header.revision >= FADT2_REVISION_ID &&
+	    acpi_gbl_FADT.century)
+		CMOS_WRITE(century, acpi_gbl_FADT.century);
+#endif
+
+	CMOS_WRITE(save_control, RTC_CONTROL);
+	CMOS_WRITE(save_freq_select, RTC_FREQ_SELECT);
+
+	spin_unlock_irqrestore(&rtc_lock, flags);
+
+	return 0;
+}
+EXPORT_SYMBOL_GPL(mc146818_set_time);