Skip to contentSkip to conversation

Where do we stand with the Xen patches?

DISCUSSION [email protected] 3 participants 3 messages May 17, 2009 – May 19, 2009

lore
Share & tools
lore
#0 lore raw
Quoted 20 lines
Quoted 13 lines
Quoted 13 lines

As in the past, my main worry is performance overhead of paravirt in general.

The patches that dont affect any native kernel fast path are probably OK (but still pending final review).

Regarding patches that do change the fastpath i'll do a round of measurements of CONFIG_PARAVIRT against !CONFIG_PARAVIRT kernels, and make up my mind based on that.

You could accelerate this by sending some "perf stat" hard numbers to give us an idea about where we stand today.

Quoted 6 lines
Quoted 1 line

Ingo

maybe this is iust a stupid comment (please forgive, i?m no advanced kernel hacker), but can?t the code inserted by the patches and which changes the fastpath just #IFDEFed at the critical offsets ? (as building a dom0 kernel is just another build target, isnt it ?)

Quoted 6 lines

no, if dom0 is going to be widely deployed, it will be because the distros turn on dom0 support by default. as a result any penalties due to xen support will be felt by all users of those distros (even if they don't use xen)

David Lang

so what?

print a huge warning on boot that running dom0 for xen may affect performance and that you should better run a normal kernel instead if you dont use xen , and youre done.

or is maintaining two different kernel packages a problem?

if so, instead of using IFDEFs, cant the critical path`s being generously circumvented by default, (if, else...), needing some dom0 kernel bootparam to be activated (i.e. use the kernel as dom0 kernel) ?

or is this too short-sighted view of the things ?

regards
roland

______________________________________________________
GRATIS für alle WEB.DE-Nutzer: Die maxdome Movie-FLAT! Jetzt freischalten unter http://movieflat.web.de

#1 lore raw
Quoted 2 lines · [email protected] wrote:

or is maintaining two different kernel packages a problem?

Yes, distros hate the proliferation of kernel packages with different config options, partly because of the combinatorial explosion (32 vs 64, UP vs SMP, PAE vs non-PAE...). An explicit design intent of all the Xen work is that it can be compile-time enabled without any (significant) effect on native performance, so that the decision to enable Xen doesn't have any downsides (either in terms of raw performance or maintenance of the kernel package).

Quoted 3 lines

if so, instead of using IFDEFs, cant the critical path`s being generously circumvented by default, (if, else...), needing some dom0 kernel bootparam to be activated (i.e. use the kernel as dom0 kernel) ?

Well, broadly speaking, yes. We try to avoid putting if/thens in critical paths, and where there are changes to hot patches, we use dynamic code patching to make it as efficient as possible.

J

#2 lore raw
Quoted 11 lines · Jeremy Fitzhardinge
Quoted 2 lines · [email protected] wrote:

or is maintaining two different kernel packages a problem?

Yes, distros hate the proliferation of kernel packages with different config options, partly because of the combinatorial explosion (32 vs 64, UP vs SMP, PAE vs non-PAE...). An explicit design intent of all the Xen work is that it can be compile-time enabled without any (significant) effect on native performance, so that the decision to enable Xen doesn't have any downsides (either in terms of raw performance or maintenance of the kernel package).

There is a long list of CONFIG_* that had performance impacts when enabled later had that impact tuned away. Especially now that the source of the performance problem is understood, it makes sense to me to merge and then focus energy on fixing it in tree instead of spending time maintaining the out of tree patches.

-chris