get_time_ns does a simplistic delta = cycle_now - cycle_last. It is possible that the h/w counter reached max and reset back to 0. This patch addresses this issue by checking for rollover condition.
NOTE 1: This does not guarantee that you cannot confuse get_time_ns. You could possibly wait for two reset cycles and then get a messed up value. To fix that we may need interrupt mode timer tick - something on the lines of jiffies on linux. NOTE 2: the question of cs->mask is not clear. if the mask is for the tick, then it is better done with (cycle_now & cs->mask) - (cs->cycle_last & cs->mask). Do we need min and max for register read?
Signed-off-by: Nishanth Menonx0nishan@ti.com
--- common/clock.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-)
Index: u-boot-v2.git/common/clock.c =================================================================== --- u-boot-v2.git.orig/common/clock.c 2008-05-20 17:19:44.000000000 -0500 +++ u-boot-v2.git/common/clock.c 2008-05-20 17:26:31.000000000 -0500 @@ -37,14 +37,24 @@ uint64_t get_time_ns(void) { struct clocksource *cs = current_clock; - uint64_t cycle_now, cycle_delta; + uint64_t cycle_now, cycle_delta = 0; uint64_t ns_offset;
/* read clocksource: */ cycle_now = cs->read();
/* calculate the delta since the last call: */ - cycle_delta = (cycle_now - cs->cycle_last) & cs->mask; + + /* Handle rollover case */ + if (cycle_now < cs->cycle_last) { + /* FIXME: I am not convinced about the cs->mask operation.. + * I am assuming cs->mask is max! Probably need to change + * clocksource structure to have min and max + */ + cycle_delta = cs->mask - (cs->cycle_last & cs->mask); + cs->cycle_last = 0; + } + cycle_delta += (cycle_now & cs->mask) - (cs->cycle_last & cs->mask);
/* convert to nanoseconds: */ ns_offset = cyc2ns(cs, cycle_delta);