On Fri, Aug 10, 2007 at 11:33:53AM +0200, Ingo Molnar wrote: > > * Jarek Poplawski <jarkao2@xxxxx> wrote: > > > > > > + } > > > > > #ifdef CONFIG_HARDIRQS_SW_RESEND > > > > > > we used the hw-resend method unconditionally, right? > > > > Right: unconditionally on a condition they are not edges... > > > > But, since not resending at all seems to work so good in testing, I > > thought, _SW_RESEND could be considered as an unnecessarily > > complicated alternative. > > > > Now, I'm a bit confused... > > the idea is multi-pronged: > > - Primarily, we want to fix the regression. 2.6.20 worked, 2.6.21 > didnt, that has to be fixed, no matter what - end of story. But we've > got a wide selection of patches for that purpose now, so what matters > at this point is the secondary question: > > - we want to know _why exactly_ the hang happens. We now have a pretty > good theory: hw-resend hangs the IO-APIC. (there is a delicate dance > between local APICs and IO-APICs for level-triggered irqs, and if we > interject via hw-resending via the local APIC, existing races, hw > bugs or weaknesses in our hw-resend implementation might be exposed) > > and even though we now have a wide selection of patches we really want > to get to the bottom of the problem so that we can fix the bug that got > exposed: apparently hw resend doesnt always work with level-triggered > irqs. > > Note that the hw-resend sequence can trigger _even without our original > patch that triggered the regression_, it's just much less likely to > happen, so this is a pre-existing IO-APIC/APIC code bug that could > trigger anytime, and which we want to see fixed. > > To confirm this theory - does the debug-patch below fix the hang? If it > fixes the hang then the theory is confirmed and then the right solution > is to retrigger an IRQ for level-triggered irqs with the proper > trigger-type set. > > Ingo Ingo: I think, you have to do this in x86_64, and there is probably send_IPI_mask used for this (but I can miss something...). I think, Marcin will not be able to do this and report before monday, but, Jean-Baptiste: of course current Ingo's or Thomas' patches are more urgent, so if you could break the current test and try this (maybe after Ingo acks this yet?) with eg. clean 2.6.23-rc1 or 2.6.22? ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Jarek P. > > ------------------> > Not-Signed-off-by: Ingo Molnar <mingo@xxxxxxx> > > Index: linux/arch/i386/kernel/io_apic.c > =================================================================== > --- linux.orig/arch/i386/kernel/io_apic.c > +++ linux/arch/i386/kernel/io_apic.c > @@ -735,7 +735,8 @@ void fastcall send_IPI_self(int vector) > * Wait for idle. > */ > apic_wait_icr_idle(); > - cfg = APIC_DM_FIXED | APIC_DEST_SELF | vector | APIC_DEST_LOGICAL; > + cfg = APIC_DM_FIXED | APIC_DEST_SELF | vector | APIC_DEST_LOGICAL | > + APIC_INT_LEVELTRIG; > /* > * Send the IPI. The write to APIC_ICR fires this off. > */ > - To unsubscribe from this list: send the line "unsubscribe linux-net" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
