* RE: SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes)
@ 2006-03-31 20:38 Stone, Joshua I
0 siblings, 0 replies; 6+ messages in thread
From: Stone, Joshua I @ 2006-03-31 20:38 UTC (permalink / raw)
To: William Cohen; +Cc: Roland McGrath, SystemTap
William Cohen wrote:
> On the earlier s390 kernels the regular timer interrupt was
> eliminated. This was done for efficiency. The s390 has had
> virtualization for years. Having the Linux kernels do the timer
> interrupt was unwanted overhead. I wonder if the xen kernel is doing
> the same, only have timer interrupt
> when something really needs to get run at a certain time.
>
> http://lwn.net/Articles/138969/
One of the comments after this article specifically mentions that
"similar patches" are in xen as well. I find it odd that the
timer.ms/jiffies probes still work fine, but timer.profile gets no
probes at all. It may just be that the dynamic-ticks patch changes the
timer path, so the softIRQ timers still happen, but the hard IRQ timers
used for timer.profile are sidestepped.
It may turn out that timer.profile will have to be emulated using PMU
sampling on clockticks...
> Do we need variations on the timer probe, wall clock time and
> virtual time?
Possibly - do you have a suggestion for how to accomplish this? The
current timer probes are using the {add,mod,del}_timer interface, which
is based purely on jiffies.
Josh
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes)
2006-03-31 2:36 Stone, Joshua I
@ 2006-03-31 14:07 ` William Cohen
0 siblings, 0 replies; 6+ messages in thread
From: William Cohen @ 2006-03-31 14:07 UTC (permalink / raw)
To: Stone, Joshua I; +Cc: Roland McGrath, SystemTap
Stone, Joshua I wrote:
> William Cohen wrote:
>
>>Sometimes the options are turned off on purpose. Do the Xen kernels
>>build with "CONFIG_KPROBES=y"?
>
>
> yes
>
>
>>If kernels are built, do the resulting kernels actually have kprobe
>>support? It use to be that xen kernels copy and pasted code from other
>>arches. What additional patches are needed to bring the kprobe support
>>in the xen kernel in line with the various architectures?
>
>
>
> In order to get the debug info working, I removed the change to
> lib/Kconfig.debug (mentioned in my other email), but left the change in
> arch/x86_64/kernel/entry-xen.S. I suspect that more will need to be
> done here before CONFIG_DEBUG_INFO really works completely, as there's
> probably some reason it was disabled, but this is good enough for
> testing SystemTap.
>
> On both xen0 and xenU, all pass-5 tests succeed, except those that use
> timer.profile. (systemtap.base/timers.stp and
> systemtap.maps/pmap_agg_overflow.stp). It seems that the timer.profile
> thinks it is registered successfully, but the callback is never
> triggered. I will investigate this further...
>
>
> Josh
Josh,
The pass-5 tests working on xen is very good news.
On the earlier s390 kernels the regular timer interrupt was eliminated.
This was done for efficiency. The s390 has had virtualization for years.
Having the Linux kernels do the timer interrupt was unwanted overhead. I
wonder if the xen kernel is doing the same, only have timer interrupt
when something really needs to get run at a certain time.
http://lwn.net/Articles/138969/
Do we need variations on the timer probe, wall clock time and virtual time?
-Will
^ permalink raw reply [flat|nested] 6+ messages in thread
* RE: SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes)
@ 2006-03-31 2:36 Stone, Joshua I
2006-03-31 14:07 ` William Cohen
0 siblings, 1 reply; 6+ messages in thread
From: Stone, Joshua I @ 2006-03-31 2:36 UTC (permalink / raw)
To: William Cohen; +Cc: Roland McGrath, SystemTap
William Cohen wrote:
> Sometimes the options are turned off on purpose. Do the Xen kernels
> build with "CONFIG_KPROBES=y"?
yes
> If kernels are built, do the resulting kernels actually have kprobe
> support? It use to be that xen kernels copy and pasted code from other
> arches. What additional patches are needed to bring the kprobe support
> in the xen kernel in line with the various architectures?
In order to get the debug info working, I removed the change to
lib/Kconfig.debug (mentioned in my other email), but left the change in
arch/x86_64/kernel/entry-xen.S. I suspect that more will need to be
done here before CONFIG_DEBUG_INFO really works completely, as there's
probably some reason it was disabled, but this is good enough for
testing SystemTap.
On both xen0 and xenU, all pass-5 tests succeed, except those that use
timer.profile. (systemtap.base/timers.stp and
systemtap.maps/pmap_agg_overflow.stp). It seems that the timer.profile
thinks it is registered successfully, but the callback is never
triggered. I will investigate this further...
Josh
^ permalink raw reply [flat|nested] 6+ messages in thread
* RE: SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes)
@ 2006-03-30 21:18 Stone, Joshua I
0 siblings, 0 replies; 6+ messages in thread
From: Stone, Joshua I @ 2006-03-30 21:18 UTC (permalink / raw)
To: William Cohen, Roland McGrath; +Cc: SystemTap
Stone, Joshua I wrote:
> I found two differences in the config files that will prevent
> SystemTap from running on the Xen kernels. In 2054_FC5 and
> 2054_FC5kdump, the config files have:
>
> CONFIG_KPROBES=y
> ...
> CONFIG_DEBUG_INFO=y
>
> Whereas in 2054_FC5xen0 and 2054_xenU, the config files have:
>
> # CONFIG_KPROBES is not set
>
> ... with no mention of CONFIG_DEBUG_INFO. This might explain the
> missing dwarf info in the kernel-debuginfo package. And of course for
> SystemTap we need CONFIG_KPROBES=y.
I believe I've found the root-cause for the missing debuginfo -
linux-2.6-xen.patch has this hunk:
diff -Nru -p --exclude='.*' ref-linux-2.6.16-rc5/lib/Kconfig.debug
linux-2.6.16-rc5-xen0/lib/Kconfig.debug
--- ref-linux-2.6.16-rc5/lib/Kconfig.debug 2006-03-01
11:12:31.000000000 -0500
+++ linux-2.6.16-rc5-xen0/lib/Kconfig.debug 2006-03-01
11:09:38.000000000 -0500
@@ -145,7 +145,7 @@ config DEBUG_BUGVERBOSE
config DEBUG_INFO
bool "Compile the kernel with debug info"
- depends on DEBUG_KERNEL
+ depends on DEBUG_KERNEL && !X86_64_XEN
help
If you say Y here the resulting kernel image will include
debugging info resulting in a larger kernel image.
Also, the same patch has this chunk at the top of
arch/x86_64/kernel/entry-xen.S:
+#ifdef CONFIG_DEBUG_INFO
+#undef CONFIG_DEBUG_INFO
+#endif
I'll leave open the question of *why* DEBUG_INFO is explicitly
disabled...
Josh
^ permalink raw reply [flat|nested] 6+ messages in thread
* Re: SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes)
2006-03-30 1:05 Stone, Joshua I
@ 2006-03-30 14:44 ` William Cohen
0 siblings, 0 replies; 6+ messages in thread
From: William Cohen @ 2006-03-30 14:44 UTC (permalink / raw)
To: Stone, Joshua I; +Cc: Roland McGrath, SystemTap
Stone, Joshua I wrote:
> I found two differences in the config files that will prevent SystemTap
> from running on the Xen kernels. In 2054_FC5 and 2054_FC5kdump, the
> config files have:
>
> CONFIG_KPROBES=y
> ...
> CONFIG_DEBUG_INFO=y
>
> Whereas in 2054_FC5xen0 and 2054_xenU, the config files have:
>
> # CONFIG_KPROBES is not set
>
> ... with no mention of CONFIG_DEBUG_INFO. This might explain the
> missing dwarf info in the kernel-debuginfo package. And of course for
> SystemTap we need CONFIG_KPROBES=y.
>
>
> Josh
Sometimes the options are turned off on purpose. Do the Xen kernels
build with "CONFIG_KPROBES=y"?
If kernels are built, do the resulting kernels actually have kprobe
support? It use to be that xen kernels copy and pasted code from other
arches. What additional patches are needed to bring the kprobe support
in the xen kernel in line with the various architectures?
-Will
^ permalink raw reply [flat|nested] 6+ messages in thread
* SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes)
@ 2006-03-30 1:05 Stone, Joshua I
2006-03-30 14:44 ` William Cohen
0 siblings, 1 reply; 6+ messages in thread
From: Stone, Joshua I @ 2006-03-30 1:05 UTC (permalink / raw)
To: William Cohen, Roland McGrath; +Cc: SystemTap
William Cohen wrote:
> Roland McGrath wrote:
>>> I tried this on x86_64 - and immediately ran into a problem with the
>>> debuginfo.
>>
>>
>> This was already noticed and is some problem in the kernel rpm specs.
>> There may be an fc5 bug report for it already on bugzilla.redhat.com,
>> I'm not sure.
>
> Xen x86_64 kernels in kernel debuginfo package do not contain dwarf
> symbols
>
> https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=185662
I found two differences in the config files that will prevent SystemTap
from running on the Xen kernels. In 2054_FC5 and 2054_FC5kdump, the
config files have:
CONFIG_KPROBES=y
...
CONFIG_DEBUG_INFO=y
Whereas in 2054_FC5xen0 and 2054_xenU, the config files have:
# CONFIG_KPROBES is not set
... with no mention of CONFIG_DEBUG_INFO. This might explain the
missing dwarf info in the kernel-debuginfo package. And of course for
SystemTap we need CONFIG_KPROBES=y.
Josh
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2006-03-31 20:38 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2006-03-31 20:38 SystemTap vs. FC5 Xen kernels (was: 03-23-2006 Meeting minutes) Stone, Joshua I
-- strict thread matches above, loose matches on Subject: below --
2006-03-31 2:36 Stone, Joshua I
2006-03-31 14:07 ` William Cohen
2006-03-30 21:18 Stone, Joshua I
2006-03-30 1:05 Stone, Joshua I
2006-03-30 14:44 ` William Cohen
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).