Rather than just assume all is well, I really would appreciate it
of you could send a positive acknowledgement that it solves
the problem without causing the universe to implode, so that
Ralf can queue up the patch for the repository.
Regards,
Kevin K.
Mikael Starvik wrote:
Yes, I thought of that variant after I sent the email yesterday.
I'll change our local implementation. If you don't hear anything
it works as expected in our case (it was pretty easy for us to
repeat).
/Mikael
-----Original Message-----
From: Kevin D. Kissell [mailto:kevink@xxxxxxxxxxxxx]
Sent: den 10 november 2009 20:46
To: Mikael Starvik
Cc: linux-mips@xxxxxxxxxxxxxx; Jesper Nilsson
Subject: Re: SMTC lookup in smtc_distribute_timer
Your failure scenario looks plausible. Mea culpa. However, I think that
a more elegant and slightly smaller (depending on just how good
the optimizer is) fix would be:
diff --git a/arch/mips/kernel/cevt-smtc.c b/arch/mips/kernel/cevt-smtc.c
index 98bd7de..b102e4f 100644
--- a/arch/mips/kernel/cevt-smtc.c
+++ b/arch/mips/kernel/cevt-smtc.c
@@ -173,11 +173,12 @@ void smtc_distribute_timer(int vpe)
unsigned int mtflags;
int cpu;
struct clock_event_device *cd;
- unsigned long nextstamp = 0L;
+ unsigned long nextstamp;
unsigned long reference;
repeat:
+ nextstamp = 0L;
for_each_online_cpu(cpu) {
/*
* Find virtual CPUs within the current VPE who have
I don't have access to SMTC-capable hardware just now, but
I guess the way to test this would be to have a test program
or kernel test stub program two events separated by the smallest
possible increment, so that the second will have passed by the
time interrupt services for the first.
Regards,
Kevin K.
Mikael Starvik wrote:
Ok, my guess is something like this:
1. At the end of smtc_distribute_timer, nextstamp is valid and has already
passed so we goto repeat.
2. Nothing updates nextstamp (only updated if the timeout is in the future
And we just decided it is in the past)
3. At the end nextstamp still has the same value so it is still valid and
in the past.
4. This repeats until read_c0_count has a value which causes nextstamp to
be in the future.
One possible patch that seams to solve it for me below. This is probably
not the correct solution so I'll need help from the SMTC experts to review
it and come up with the correct solution.
Best Regards
/Mikael
Index: cevt-smtc.c
===================================================================
RCS file: /usr/local/cvs/linux/os/linux-2.6/arch/mips/kernel/cevt-smtc.c,v
retrieving revision 1.2
diff -u -r1.2 cevt-smtc.c
--- cevt-smtc.c 2 Sep 2009 10:07:51 -0000 1.2
+++ cevt-smtc.c 10 Nov 2009 11:40:31 -0000
@@ -223,8 +223,10 @@
write_c0_compare(nextstamp);
ehb();
if ((nextstamp - (unsigned long)read_c0_count())
- > (unsigned long)LONG_MAX)
+ > (unsigned long)LONG_MAX) {
+ nextstamp = 0L;
goto repeat;
+ }
}
}
