Hi!
It appears that kernel time keeping may become unpredictable when HZ is
raised to large values due to negative shift operations in timer.c:
time_adj = -ltemp << (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
Suppose on x86:
HZ = 2048
and SHIFT_SCALE = 22
and SHIFT_UPDATE = (SHIFT_KG + MAXTC) = 12
then
SHIFT_HZ = 11
and (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE) = -1
This negative number is a problem for shift operations. The
following patch takes care of it (I hope).
Rolf Fokkens
[email protected]
(2nd post of this patch)
--- linux-2.5.37.orig/kernel/timer.c Mon Sep 16 20:43:47 2002
+++ linux-2.5.37/kernel/timer.c Sat Sep 21 13:08:10 2002
@@ -339,6 +339,11 @@
run_task_queue(&tq_immediate);
}
+static inline long signedshift (long val, int nshift)
+{
+ return (nshift > 0 ? val << nshift : val >> -nshift);
+}
+
/*
* this routine handles the overflow of the microsecond field
*
@@ -418,7 +423,7 @@
if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
ltemp = (MAXPHASE / MINSEC) << SHIFT_UPDATE;
time_offset += ltemp;
- time_adj = -ltemp << (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
+ time_adj = signedshift (-ltemp, SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
} else {
ltemp = time_offset;
if (!(time_status & STA_FLL))
@@ -426,7 +431,7 @@
if (ltemp > (MAXPHASE / MINSEC) << SHIFT_UPDATE)
ltemp = (MAXPHASE / MINSEC) << SHIFT_UPDATE;
time_offset -= ltemp;
- time_adj = ltemp << (SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
+ time_adj = signedshift (ltemp, SHIFT_SCALE - SHIFT_HZ - SHIFT_UPDATE);
}
/*