|
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH v1 00/28] x86: initial SVM AVIC support
On Fri, Nov 20, 2026 at 03:47:15PM +0000, Andrew Cooper wrote: > On 05/11/2026 1:33 am, dmukhin@xxxxxxxx wrote: > > Patch Series > > ============ > > > > Basically the entire work is divided into three big series: > > (A) Current series: initial SVM AVIC [5] > > (B) XTF interrupt latency test [6] > > (C) XTF infra in public CI [5] > > > > For this AVIC series: > > > > Patches 1, 5, 6, 9, 11-16 introduce various cleanups and fixups. > > > > Patch 20 changes XEN_DOMCTL_debug_op implementation to allow > > more debug operations added (like enabling/disabling AVIC at > > run-time - used by the XTF test). > > > > Patches 2, 3, 4, 7, 8, 10, 17, 18-25 represent AVIC implementation > > including performance counters and run-time controls via debug > > hypercall. > > > > Patches 26-28 cover various documentation updates. > > Sorry this reply has taken longer to get to than I had intended. I have > committed a couple of trivial patches, but I think the rest of the work > is going to take a bit more complicated rearranging. > > First of all, I think there are two areas where we want to take other > pending work as prerequisite. > > 1) Moving the SVM/SEV leaves into CPU Policy. (Series of mine needing > dusting off and re-posting.) > > In patch 4, you introduce a new command line option. I know Intel has > one, but it is part of a bigger anti-pattern we're trying to remove. > Wit the SVM leaf in policy, `cpuid=no-avic` on the command line will > work to hide the feature from Xen, and without much effort we can make > `cpuid=avic` work as an opt-in until it's enabled by default. I have parameterized MSR doorbell path vs vcpu_kick() via svm_avic=nodoorbell in the current series. It was pretty useful to run experiments. So AVIC may still need its own command line switch. What do you think? > > Also my series removes the SVM_FEATURE_* constants (they'll be > auto-generated now), which simplifies your earlier patches. > > 2) Removing the vcpu_id <-> apic_id relation. (Series of Alejandro's, > or at least the first part of it, needing dusting off and re-posting.) > > In patch 7, you encode the guest's APIC_ID as it's vCPU ID * 2. Yes > this is currently true, but we're desperately trying to break this bogus > relation. It prohibits providing a sane topology to guests, and is one > of the hard limits on HVM_MAX_VCPUs being 128. > > Alejandro did a series making topology selection work. The first half > of it at least gets into a position where the toolstack chooses the > APIC_IDs and we've got a proper variable in Xen with which to derive > everything else from. Getting at least this sorted will give you a > proper base to build the logical/physical table upon. > > 3) (Bonus). We need to take a decision on whether we're going to block > writes to APIC_ID in xAPIC mode or not. I sent an RFC patch. If we do > decide to block writes to APIC_ID in xAPIC mode, then it means the > logical/physical table is static for the lifetime of the domain. > > > Now for some feedback in no specific order > > a) Please have one single patch adding all the architectural enumeration > in one go (CPUID, VMCB, etc). i.e. merge parts of patch 2, 3 and 7. (I > have your patch 7 rebased over the VMCB cleanup which was merged after > posting). I find this helpful when reviewing because it separates "what > is needed for the hardware" from "what is just internal book-keeping". > > b) In patch 4 you have a single static avic_apic_access_page[]. That > was XSA-412 (3b5beaf49033cd). > > It's actually worse on AMD. On Intel, any writes targeting a > non-register part of the backing page are discarded. On AMD, they're > deposited in the backing page. So, this page *must* be allocated per > domain, and not global. > > c) AMD has extensions to the APIC spec. AVIC even accelerates some of > these, and you feed the accesses into the current vlapic > infrastructure. For current CPUs, at a minimum that means reverting > changes to eLVTs, but I think more generally we should probably have a > CPU generation check during init and refuse to turn on AVIC until a > human has checked which new hardware features leak through. > > d) One very fun bug with AVIC is that a misaligned xmm store can write > to two different registers at once (ymm can hit 3, and zmm can hit 5). > The whole write gets deposited (contrary to Intel, which explicitly > discards outside of the first register hit), but only one VMExit > occurs. This allows the guest to write reserved bits. We need to > confirm that nothing in Xen is going to malfunction (e.g. check for > BUG_ON()'s/etc. which ought to be downgraded to domain_crash()/etc.) > because of finding non-architectural state. > > e) You've got a number of fixes for interrupt-related things. I think > we need to get those fixed first and with robust testing in place before > trying to add AVIC in on top. I'll see about dusting off an XTF test I > have on the matter, but we really need to fix CR8 handling before > building further. > > f) I'm not overly convinced with making use of debugop like that. For > now, maybe defer it to the end of the series. I have had several > occasions to need extra hypercalls to make testing happen nicely, so I > think there is room to make something, but I suspect it will want to be > a little different in nature. > > I'm sure there's stuff I've missed, but I think this a long enough email > for now. > > There's an outstanding question on patch 5. If you can answer it, I can > commit. > > Thanks, > > ~Andrew >
|
![]() |
Lists.xenproject.org is hosted with RackSpace, monitoring our |