On 24.10.2022 11:32, Roger Pau Monné wrote: > On Mon, Oct 24, 2022 at 08:45:07AM +0200, Jan Beulich wrote: >> On 21.10.2022 23:54, Andrew Cooper wrote: >>> On 20/10/2022 12:01, Roger Pau Monné wrote: >>>> Hello, >>>> >>>> As part of some follow up improvements to my VIRT_SPEC_CTRL series we >>>> have been discussing what the usage of SSBD should be for the >>>> hypervisor itself. There's currently a `spec-ctrl=ssbd` option [0], >>>> that has an out of date description, as now SSBD is always offered to >>>> guests on AMD hardware, either using SPEC_CTRL or VIRT_SPEC_CTRL. >>>> >>>> It has been pointed out by Andrew that toggling SSBD on AMD using >>>> VIRT_SPEC_CTRL or the non-architectural way (MSR_AMD64_LS_CFG) can >>>> have a high impact on performance, and hence switching it on every >>>> guest <-> hypervisor context switch is likely a very high >>>> performance penalty. >>>> >>>> It's been suggested that it could be more appropriate to run Xen with >>>> the guest SSBD selection on those systems, however that clashes with >>>> the current intent of the `spec-ctrl=ssbd` option. >>>> >>>> I hope I have captured the expressed opinions correctly in the text >>>> above. >>>> >>>> I see two ways to solve this: >>>> >>>> * Keep the current logic for switching SSBD on guest <-> hypervisor >>>> context switch, but only use it if `spec-ctrl=ssbd` is set on the >>>> command line. >>>> >>>> * Remove the logic for switching SSBD on guest <-> hypervisor context >>>> switch, ignore setting of `spec-ctrl=ssbd` on those systems and run >>>> hypervisor code with the guest selection of SSBD. >>>> >>>> Which has raised me the question of whether there's an use case >>>> for always running hypervisor code with SSBD enabled, or that's no >>>> longer relevant if we always offer guests a way for them to toggle the >>>> setting when required. >>>> >>>> I would like to settle on a way forward, so we can get this fixed >>>> before 4.17. >>>> >>>> Thanks, Roger. >>>> >>>> [0] >>>> https://xenbits.xen.org/docs/unstable/misc/xen-command-line.html#spec-ctrl-x86 >>> >>> There are many issues at play here. Not least that virt spec ctrl is >>> technically a leftover task that ought to force a re-issue of XSA-263. >>> >>> Accessing MSRs (even reading) is very expensive, typically >1k cycles. >>> The core CFG registers are more expensive than most, because they're >>> intended to be configured once after reset and then left alone. >>> >>> Throughout the speculation work, we've seen crippling performance hits >>> from accessing MSRs in fastpaths. The fact we're forced to use MSRs in >>> fastpaths even on new CPUs with built in (rather than retrofitted) >>> speculation support is is an area of concern still being worked on with >>> the CPU vendors. >>> >>> Case in point. We found for XSA-398 that toggling AMD's >>> MSR_SPEC_CTRL.IBRS on the PV entrypath was so bad that setting it >>> unilaterally behind the back of PV guests was the faster option. >>> (Another todo is to stop doing this on Intel eIBRS systems, and this >>> will recover us a decent chunk of performance.) >>> >>> >>> SSBD mitigations are (rightly or wrongly) off by default for performance >>> reasons. AMD are less affected than Intel, for microarchitectural >>> reasons which are discussed in relevant whitepapers, and which are >>> expected to remain true for future CPUs. >>> >>> When Xen doesn't care about the protecting itself against SSBD by >>> default, I guarantee you that it will be faster to omit the MSR accesses >>> and run in the guest kernel's choice, than to clear the SSBD >>> protection. We simply don't spend long enough in the hypervisor for the >>> hit against memory accesses to dwarf the hit for MSR accesses taken on >>> entry/exit. >>> >>> The reason we put in spec-ctrl=ssbd was as a stopgap, because at the >>> time we didn't know how bad SSB really was, and it was decided that the >>> admin should have a big hammer to use if they really needed. >>> >>> When Xen does care about protecting itself, the above reasoning bites >>> back hard. Because we spend (or should be spending!) >99% of time in >>> the guest, the hit to memory accesses is far more likely to be able >>> dwarf the hit from the MSR accesses, but now, the dominating factor for >>> performance is the vmexit rate. >>> >>> The problem is that if you've got a completely compute bound workload, >>> there are very few exits, while if you've got an IO bound workload, >>> there are plenty of exits. I honestly don't know if it will be more >>> efficient to leave SSBD active unilaterally (whether or not we hide >>> this, e.g. synthesizing SSB_NO), or to let the guest run with it kernels >>> choice. I suspect the answer is different with different workloads. >>> >>> >>> But, one other factor helps us. Given that the default is fast (rather >>> than secure), anyone opting in to spec-ctrl=ssbd is saying "I care more >>> about security than performance", at which point we can simplify what we >>> do because we don't need to cater to everyone. >>> >>> >>> As a slight tangent, there is a cost to having too many options, which >>> must not be ignored. Xen's speculation safety is far too complicated >>> already and needs to get more simple; this has a material impact on how >>> easy it is to follow, and how easy it to make changes. >>> >>> It is the way it is because we've had 6 years of drip feeding one >>> problem after another, and haven't had the time to take a step and >>> design something more sensible from having 6 years of >>> knowledge/learnings as a basis. There are definitely things which I >>> would have done differently, if 6 years ago, I'd known what I know now, >>> and part of the reason why the recent speculation security work has >>> taken so much effort is because it has involved reworking the effort >>> which came before, to a deadline which never has enough time to plan >>> properly within. >>> >>> >>> So, first question, do we care about having an "SSBD active while in >>> Xen" mode? >>> >>> Probably yes, because we a) still don't have a working solution for PV >>> guests on AMD and b) who knows if there's something far worse lurking in >>> the future. Sods law says that if we decide no here, it will be >>> critical for some future issue. >>> >>> But as it's off by default and noone's made has made any noise about >>> having it on, we ought to prioritise simplicity. >>> >>> Given that off is the default, but we know that kernels do offer it to >>> userspace, and it does get used by certain processes, we need to >>> prioritise performance. And here, this is net system performance, not >>> "ensure it's off whenever it can be". Having Xen run in the guest >>> kernel's choice of value will result in much better overall performance, >>> than trying to modify the setting in the VMentry/exit path. >> >> My takeaway from this reply of yours is: By default run with the guest's >> choice, while (I'm less certain here) you're undecided about the behavior >> with "spec-ctrl=ssbd". Please could you make explicit whether this is a >> correct understanding of mine? > > * spec-ctrl=ssbd -> SSBD always on, expose VIRT_SSBD > (VIRT_SPEC_CTRL.SSBD) but guest setting won't be propagated to > platform. As a future improvement also expose SSB_NO in that > case. > > * spec-ctrl=no-ssbd -> Run hypervisor code with guest SSBD selection > depending on hardware support. > > Default to `spec-ctrl=no-ssbd`. > > Would that be an accurate?
This matches my view, yes. Jan
