Date: Tue, 12 Dec 2006 19:49:41 -0500 From: Jung-uk Kim <jkim@FreeBSD.org> To: Oleg Bulyzhin <oleg@FreeBSD.org> Cc: cvs-src@FreeBSD.org, src-committers@FreeBSD.org, cvs-all@FreeBSD.org Subject: Re: cvs commit: src/sys/dev/bge if_bge.c Message-ID: <200612121949.43019.jkim@FreeBSD.org> In-Reply-To: <20061213000352.GH91560@lath.rinet.ru> References: <200612111800.kBBI0Zv0047909@repoman.freebsd.org> <200612121824.45643.jkim@FreeBSD.org> <20061213000352.GH91560@lath.rinet.ru>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tuesday 12 December 2006 07:03 pm, Oleg Bulyzhin wrote:
> On Tue, Dec 12, 2006 at 06:24:43PM -0500, Jung-uk Kim wrote:
> > On Tuesday 12 December 2006 05:25 pm, Oleg Bulyzhin wrote:
> > > On Mon, Dec 11, 2006 at 06:00:35PM +0000, Jung-uk Kim wrote:
> > > > jkim 2006-12-11 18:00:35 UTC
> > > >
> > > > FreeBSD src repository
> > > >
> > > > Modified files:
> > > > sys/dev/bge if_bge.c
> > > > Log:
> > > > - Correct collision counter for BCM5705+. This register is
> > > > read/clear. - Correct RX packet drop counter for BCM5705+.
> > > > This register is read/clear and it wraps very quickly under
> > > > heavy packet drops because only the lower ten bits are valid
> > > > according to the documentation. However, it seems few more
> > > > bits are actually valid and the rest bits are always
> > > > zeros[1]. Therefore, we don't mask them off here. To get
> > > > accurate packet drop count, we need to check the register
> > > > from bge_rxeof(). It is commented out for now, not to
> > > > penalize normal operation. Actual performance impact should
> > > > be measured later.
> > > > - Correct integer casting from u_long to uint32_t. Casting
> > > > is not really needed for all supported platforms but we
> > > > better do this correctly[2].
> > > >
> > > > Tested by: bde[1]
> > > > Suggested by: bde[2]
> > > >
> > > > Revision Changes Path
> > > > 1.158 +13 -10 src/sys/dev/bge/if_bge.c
> > >
> > > I didnt get the point of your u_long -> uint32_t changes.
> > > As i can see your change will cause more often wraps on 64bit
> > > archs: In rev. 1.153 you have converted cnt to uint32_t. Since
> > > cnt is stored in sc->bge_* counters you have shortened(on 64bit
> > > archs) them as well.
> >
> > Depending on the chip types, we use different methods to get the
> > stats. However, both methods read only 32-bit register/memory.
> > Therefore, there's no reason to use u_long at all and integer
> > math is cheap.
>
> While hardware counters are 32bit long, interface & driver's
> counters are 64bit long on 64bit chips. So you can keep numbers
> even if hardware counter wraps around.
I guess you are talking about collision counter for BCM570[0-4], right? Since we are adding four different 32-bit counters to make one counter, we have no way of knowing which one is wrapped unless we check upper-half's. To make things worse, if multiple counters wrap, the previous logic lose. If you want real solution for 64-bit machines, you have to read full 64-bit atomically.
> uint32_t & u_long are the same on i386 and u_long arithmetic is not
> slower than int one for 64 bit chips. Anyway we need few arithmetic
> operations once per second, so it's unimportant.
>
> > > P.S. Your current change is unclear to me too: since ifp
> > > counters and sc->_bge_ are both u_long i can not see any reason
> > > of converting (u_long) cast to (uint32_t) one.
> >
> > To cut the long story short, it does not really matter as I said
> > in the log but it is required if sizeof(uint32_t) < sizeof(int).
>
> Could you please explain this.
As I noted in the log, it was suggested by bde. I didn't ask his
permission but I guess it's okay to cut and paste his e-mail. ;-)
(Sorry, Bruce.)
-----------------
Of course it makes no difference on currently supported machines -- see
previous mail :-). The following uses fake ints to demonstrate the
problem.
%%%
#include <stdint.h>
#include <stdio.h>
/*
* Fake ints with more than 32 bits. Unsigned integer types strictly
* smaller than ints get promoted to signed ints for binary operations.
* Here it is not the signedness change but the loss of accidental
* wrapping at 32 bits that matters.
*/
typedef int64_t int_t;
/* Fake unsigned longs no smaller than ints. */
typedef uintmax_t ulong_t;
/* 32-bit hardware counters. */
uint32_t before_wrap = 0xffffffff;
uint32_t after_wrap = 0;
ulong_t if_ierrors = 0;
int main(void)
{
/* Should be only 1 more error, but... */
if_ierrors += (int_t)after_wrap - (int_t)before_wrap;
printf("%ju\n", (uintmax_t)if_ierrors);
}
%%%
The output is a magic number slightly less recognizable than 2^32-1:
18446744069414584321
This looks a bit like 2^64-1, but is actually 2^64-2^32+1 (1-2^32 after
overflow).
-----------------
Jung-uk Kim
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200612121949.43019.jkim>