[9fans] stats.c bug

[9fans] stats.c bug

[andrey mirtchovski]

there is a small bug with stats.c's readnum() routine -- the numbers for
context switches, system calls and interrupts are unsigned long (I believe)
but are converted using strtol().

if a machine is left running cpu-heavy applications for a long time (say,
a week) those numbers will overflow and strtol() will return LONG_MAX as the
value. the net effect is that stats.c will not display context, syscalls and
interrupts properly.

here's a sample cat /dev/systat:


          0  2530252355  3182187863  2996662559     3944850           0           0           0          99           0
          0  2530252917  3182188061  2996663328     3944996           0           0           0          98           0
          0  2530253045  3182188170  2996663495     3945138           0           0           0          99           0
          0  2530253224  3182188295  2996663728     3945284           0           0           0         100           0



there is also a new entry (the 10th number) reported by /dev/sysstat, which
is undocumented in the man pages. in the kernel source (port/portdat.h) it
is documented as:

	ulong	avg_inintr;	/* avg time per clock tick in interrupt handlers */

i took the liberty of modifying stats.c to include this statistic (note:
i have not modified the manpage, nor added an argument option to invoke this
from the command line -- i figured since 'idle' isn't there yet i'd better
leave this for other people to decide).

andrey


plan9% diff stats.c /sys/src/cmd/stats.c
45d44
< 	InIntr,
65,66c64,65
< 	long		devsysstat[10];
< 	long		prevsysstat[10];
---
> 	long		devsysstat[9];
> 	long		prevsysstat[9];
108d106
< 	Minintr,
129d126
< 	"add  in intr ",
150d146
< 	inintrval(Machine*, long*, long*, int),
171d166
< 	inintrval,
434c429
< 		a[i] = strtoul(p, &p, 10);
---
> 		a[i] = strtol(p, &p, 10);
676c671
< 		present[Minintr] | present[Msyscall] | present[Mtlbmiss] | present[Mtlbpurge];
---
> 		present[Msyscall] | present[Mtlbmiss] | present[Mtlbpurge];
829,835d823
< 	*vmax = 100;
< }
<
< void
< inintrval(Machine *m, long *v, long *vmax, int)
< {
< 	*v = m->devsysstat[InIntr];
plan9%

Re: [9fans] stats.c bug

[Russ Cox]

> if a machine is left running cpu-heavy applications for a long time (say,
> a week) those numbers will overflow and strtol() will return LONG_MAX as the
> value. the net effect is that stats.c will not display context, syscalls and
> interrupts properly.

umm, what if i leave it running for two weeks?

Re: [9fans] stats.c bug

[andrey mirtchovski]

On Wed, 18 Jun 2003, Russ Cox wrote:

> umm, what if i leave it running for two weeks?
>

the numbers would wrap and there will be a single instance where stats would
report errors, afterwards it should be back to normal until the next wrap, no?

if that's not the case (i.e. if /dev/sysstat reports 'long long'-s) then we
should use strtoll()...

i don't see the catch, sorry :(