Messages in this thread | | | Date | Thu, 19 Mar 2009 01:05:06 +0200 | From | Avi Kivity <> | Subject | Re: Question about x86/mm/gup.c's use of disabled interrupts |
| |
Jeremy Fitzhardinge wrote: > Avi Kivity wrote: >>> Hm, awkward if flush_tlb_others doesn't IPI... >>> >> >> How can it avoid flushing the tlb on cpu [01]? It's it's >> gup_fast()ing a pte, it may as well load it into the tlb. > > xen_flush_tlb_others uses a hypercall rather than an IPI, so none of > the logic which depends on there being an IPI will work.
Right, of course, that's what we were talking about. I thought optimizations to avoid IPIs if an mm never visited a cpu.
> >>> Simplest fix is to make gup_get_pte() a pvop, but that does seem >>> like putting a red flag in front of an inner-loop hotspot, or >>> something... >>> >>> The per-cpu tlb-flush exclusion flag might really be the way to go. >> >> I don't see how it will work, without changing Xen to look at the flag? >> >> local_irq_disable() is used here to lock out a remote cpu, I don't >> see why deferring the flush helps. > > Well, no, not deferring. Making xen_flush_tlb_others() spin waiting > for "doing_gup" to clear on the target cpu. Or add an explicit notion > of a "pte update barrier" rather than implicitly relying on the tlb > IPI (which is extremely convenient when available...).
Pick up a percpu flag from all cpus and spin on each? Nasty.
You could use the irq enabled flag; it's available and what native spins on (but also means I'll need to add one if I implement this).
-- I have a truly marvellous patch that fixes the bug which this signature is too narrow to contain.
| |