On 10.09.15 16:03, Savolainen, Petri (Nokia - FI/Espoo) wrote:
-----Original Message-----
From: EXT Ivan Khoronzhuk [mailto:[email protected]]
Sent: Thursday, September 10, 2015 3:34 PM
To: Savolainen, Petri (Nokia - FI/Espoo); [email protected];
[email protected]
Subject: Re: [lng-odp] [Patch 1/2] linux-generic: odp_schedule: fix
odp_schdule_wait_time
On 10.09.15 15:16, Savolainen, Petri (Nokia - FI/Espoo) wrote:
Change is OK. The calculation can be changed to use cpu cycle API
(with max CPU hz) later. Time measurement does not have to be accurate,
so we can continue to use CPU cycles and max hz (to spin at least 'ns'
nanosec).
We can wait more than 4s, with my cpu cycle counter it's not possible.
And don't forget about CPU freq and conversion to ns.
This is not place to write about that, lets talk about later.
With scheduler you can wait more than 4 sec, since your implementation does the
spinning and it can be sure
that counter is not wrapping (you'll check the counter every 30 cycles or so
while waiting).
Ok.
When you know the max freq (e.g. 1GHz) and calculate how many CPU cycles you
need to spin
based on that, you'll always wait at least the time requested. CPU can run
slower than max (e.g. 500MHz),
which means that you may wait longer than 'ns' nanoseconds, but it's OK
according the spec.
You did wait "at least" 'ns' nsec (e.g. 2x ns).
2x can have much bigger delay error than time API produces.
Especially when bigger delay is requested, say 50 ns, and you wait ~100ns.
In most cases I suppose rate of counter ~ 3 time less than Fmax.
Lets take 7ns time resolution, then 50ns + 7ns = 57ns.
error: 7ns << 50ns.
And for linux-generic cpu_cycles are only one source of counter.
It's reused by time API, so here cpu cycles anyway.
-Petri
The patch needs to be labeled with "v2". And preferably PATCH, in
capital letters.
It differs in numbering 1/2. I prefer to hold the same numbering within
series.
But, next time I will bear it in mind.
The log entry could be clarified a bit, see under.
-----Original Message-----
From: EXT Ivan Khoronzhuk [mailto:[email protected]]
Sent: Thursday, September 10, 2015 12:49 PM
To: [email protected]; Savolainen, Petri (Nokia - FI/Espoo);
[email protected]
Cc: Ivan Khoronzhuk
Subject: [lng-odp] [Patch 1/2] linux-generic: odp_schedule: fix
odp_schdule_wait_time
The resolution of schedule time can be more than 1ns. As result
can happen that 1ns corresponds 0 ticks of timer counter, but if
passed 1ns it's obvious that user wanted to schedule at least once.
In short: "Depending on resolution, a low nsec value could be
converted to 0 cycles, which was specified as WAIT ..."
Ok
Currently it can lead to wait forever, as 0 corresponds to
ODP_SCHED_WAIT, it can block program flow at all.
"... and would instruct a schedule call to wait forever."
Ok
Also ODP_SCHED_NO_WAIT corresponds to 1 tick, it's rarely but can
wait a little, when shced time has slower rate. It should correspond
to schedule only once.
So, change ODP_SCHED_NO_WAIT to 0 and ODP_SCHED_NO_WAIT to
UINT64_MAX.
Signed-off-by: Ivan Khoronzhuk <[email protected]>
---
platform/linux-generic/include/odp/plat/schedule_types.h | 4 ++--
platform/linux-generic/odp_schedule.c | 3 ---
2 files changed, 2 insertions(+), 5 deletions(-)
diff --git a/platform/linux-
generic/include/odp/plat/schedule_types.h
b/platform/linux-generic/include/odp/plat/schedule_types.h
index c48b652..8cdc40b 100644
--- a/platform/linux-generic/include/odp/plat/schedule_types.h
+++ b/platform/linux-generic/include/odp/plat/schedule_types.h
@@ -22,8 +22,8 @@ extern "C" {
* @{
*/
-#define ODP_SCHED_WAIT 0
-#define ODP_SCHED_NO_WAIT 1
+#define ODP_SCHED_WAIT UINT64_MAX
+#define ODP_SCHED_NO_WAIT 0
typedef int odp_schedule_prio_t;
diff --git a/platform/linux-generic/odp_schedule.c b/platform/linux-
generic/odp_schedule.c
index c6619e5..827fcf0 100644
--- a/platform/linux-generic/odp_schedule.c
+++ b/platform/linux-generic/odp_schedule.c
@@ -646,9 +646,6 @@ void odp_schedule_resume(void)
uint64_t odp_schedule_wait_time(uint64_t ns)
{
- if (ns <= ODP_SCHED_NO_WAIT)
- ns = ODP_SCHED_NO_WAIT + 1;
-
return odp_time_ns_to_cycles(ns);
}
--
1.9.1
--
Regards,
Ivan Khoronzhuk
--
Regards,
Ivan Khoronzhuk
_______________________________________________
lng-odp mailing list
[email protected]
https://lists.linaro.org/mailman/listinfo/lng-odp
