Fix CPU spinlock lockups on secondary CPU bringup
Russell King [Wed, 22 Jun 2011 10:55:50 +0000 (11:55 +0100)]
Secondary CPU bringup typically calls calibrate_delay() during its
initialization.  However, calibrate_delay() modifies a global variable
(loops_per_jiffy) used for udelay() and __delay().

A side effect of 71c696b1 ("calibrate: extract fall-back calculation
into own helper") introduced in the 2.6.39 merge window means that we
end up with a substantial period where loops_per_jiffy is zero.  This
causes the spinlock debugging code to malfunction:

u64 loops = loops_per_jiffy * HZ;
for (;;) {
for (i = 0; i < loops; i++) {
if (arch_spin_trylock(&lock->raw_lock))
return;
__delay(1);
}
...
}

by never calling arch_spin_trylock() - resulting in the CPU locking
up in an infinite loop inside __spin_lock_debug().

Work around this by only writing to loops_per_jiffy only once we have
completed all the calibration decisions.

Tested-by: Santosh Shilimkar <santosh.shilimkar@ti.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Cc: <stable@kernel.org> (2.6.39-stable)
--
Better solutions (such as omitting the calibration for secondary CPUs,
or arranging for calibrate_delay() to return the LPJ value and leave
it to the caller to decide where to store it) are a possibility, but
would be much more invasive into each architecture.

I think this is the best solution for -rc and stable, but it should be
revisited for the next merge window.

 init/calibrate.c |   14 ++++++++------
 1 files changed, 8 insertions(+), 6 deletions(-)
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

init/calibrate.c

index 2568d22..aae2f40 100644 (file)
@@ -245,30 +245,32 @@ recalibrate:
 
 void __cpuinit calibrate_delay(void)
 {
+       unsigned long lpj;
        static bool printed;
 
        if (preset_lpj) {
-               loops_per_jiffy = preset_lpj;
+               lpj = preset_lpj;
                if (!printed)
                        pr_info("Calibrating delay loop (skipped) "
                                "preset value.. ");
        } else if ((!printed) && lpj_fine) {
-               loops_per_jiffy = lpj_fine;
+               lpj = lpj_fine;
                pr_info("Calibrating delay loop (skipped), "
                        "value calculated using timer frequency.. ");
-       } else if ((loops_per_jiffy = calibrate_delay_direct()) != 0) {
+       } else if ((lpj = calibrate_delay_direct()) != 0) {
                if (!printed)
                        pr_info("Calibrating delay using timer "
                                "specific routine.. ");
        } else {
                if (!printed)
                        pr_info("Calibrating delay loop... ");
-               loops_per_jiffy = calibrate_delay_converge();
+               lpj = calibrate_delay_converge();
        }
        if (!printed)
                pr_cont("%lu.%02lu BogoMIPS (lpj=%lu)\n",
-                       loops_per_jiffy/(500000/HZ),
-                       (loops_per_jiffy/(5000/HZ)) % 100, loops_per_jiffy);
+                       lpj/(500000/HZ),
+                       (lpj/(5000/HZ)) % 100, lpj);
 
+       loops_per_jiffy = lpj;
        printed = true;
 }