Message-ID: <3ecfa4ff-9916-4ac1-8464-c1b6615b832a@citrix.com>
Date: Thu, 9 Jan 2025 23:00:12 +0000
From: Andrew Cooper <andrew.cooper3@...rix.com>
To: Yosry Ahmed <yosryahmed@...gle.com>
Cc: akpm@...ux-foundation.org, bp@...en8.de, dave.hansen@...ux.intel.com,
 hpa@...or.com, jackmanb@...gle.com, kernel-team@...a.com,
 linux-kernel@...r.kernel.org, linux-mm@...ck.org, luto@...nel.org,
 mingo@...hat.com, nadav.amit@...il.com, peterz@...radead.org,
 reijiw@...gle.com, riel@...riel.com, tglx@...utronix.de, x86@...nel.org,
 zhengqi.arch@...edance.com
Subject: Re: [PATCH v3 00/12] AMD broadcast TLB invalidation

On 09/01/2025 9:32 pm, Yosry Ahmed wrote:
> On Wed, Jan 8, 2025 at 6:47 PM Andrew Cooper <andrew.cooper3@...rix.com> wrote:
>>>> I suspect AMD wouldn't tell us exactly ;)
>>> Well, ideally they would just tell us the conditions under which CPUs
>>> respond to the broadcast TLB flush or the expectations around latency.
>> [Resend, complete this time]
>>
>> Disclaimer.  I'm not at AMD; I don't know how they implement it; I'm
>> just a random person on the internet.  But, here are a few things that
>> might be relevant to know.
>>
>> AMD's SEV-SNP whitepaper [1] states that RMP permissions "are cached in
>> the CPU TLB and related structures" and also "When required, hardware
>> automatically performs TLB invalidations to ensure that all processors
>> in the system see the updated RMP entry information."
>>
>> That sentence doesn't use "broadcast" or "remote", but "all processors"
>> is a pretty clear clue.  Broadcast TLB invalidations are a building
>> block of all the RMP-manipulation instructions.
>>
>> Furthermore, to be useful in this context, they need to be ordered with
>> memory.  Specifically, a new pagewalk mustn't start after an
>> invalidation, yet observe the stale RMP entry.
>>
>>
>> x86 CPUs do have reasonable forward-progress guarantees, but in order to
>> achieve forward progress, they need to e.g. guarantee that one memory
>> access doesn't displace the TLB entry backing a different memory access
>> from the same instruction, or you could livelock while trying to
>> complete a single instruction.
>>
>> A consequence is that you can't safely invalidate a TLB entry of an
>> in-progress instruction (although this means only the oldest instruction
>> in the pipeline, because everything else is speculative and potentially
>> transient).
>>
>>
>> INVLPGB invalidations are interrupt-like from the point of view of the
>> remote core, but are microarchitectural and can be taken irrespective of
>> the architectural Interrupt and Global Interrupt Flags.  As a
>> consequence, they'll need wait until an instruction boundary to be
>> processed.  While not AMD, the Intel RAR whitepaper [2] discusses the
>> handling of RARs on the remote processor, and they share a number of
>> constraints in common with INVLPGB.
>>
>>
>> Overall, I'd expect the INVLPGB instructions to be pretty quick in and
>> of themselves; interestingly, they're not identified as architecturally
>> serialising.  The broadcast is probably posted, and will be dealt with
>> by remote processors on the subsequent instruction boundary.  TLBSYNC is
>> the barrier to wait until the invalidations have been processed, and
>> this will block for an unspecified length of time, probably bounded by
>> the "longest" instruction in progress on a remote CPU.  e.g. I expect it
>> probably will suck if you have to wait for a WBINVD instruction to
>> complete on a remote CPU.
>>
>> That said, architectural IPIs have the same conditions too, except on
>> top of that you've got to run a whole interrupt handler.  So, with
>> reasonable confidence, however slow TLBSYNC might be in the worst case,
>> it's got absolutely nothing on the overhead of doing invalidations the
>> old fashioned way.
> Generally speaking, I am not arguing that TLB flush IPIs are worse
> than INLPGB/TLBSYNC, I think we should expect the latter to perform
> better in most cases.
>
> But there is a difference here because the processor executing TLBSYNC
> cannot serve interrupts or NMIs while waiting for remote CPUs, because
> they have to be served at an instruction boundary, right?

That's as per the architecture, yes.  NMIs do have to be served on
instruction boundaries.  An NMI that becomes pending while a TLBSYNC is
in progress will have to wait until the TLBSYNC completes.

(Probably.  REP string instructions and AVX scatter/gather have explicit
behaviours that them them be interrupted, and to continue from where
they left off when the interrupt handler returns.  Depending on how
TLBSYNC is implemented, it's just possible it has this property too.)

> Unless
> TLBSYNC is an exception to that rule, or its execution is considered
> completed before remote CPUs respond (i.e. the CPU executes it quickly
> then enters into a wait doing "nothing").
>
> There are also intriguing corner cases that are not documented. For
> example, you mention that it's reasonable to expect that a remote CPU
> does not serve TLBSYNC except at the instruction boundary.

INVLPGB needs to wait for an instruction boundary in order to be processed.

All TLBSYNC needs to do is wait until it's certain that all the prior
INVLPGBs issued by this CPU have been serviced.

>  What if
> that CPU is executing TLBSYNC? Do we have to wait for its execution to
> complete? Is it possible to end up in a deadlock? This goes back to my
> previous point about whether TLBSYNC is a special case or when it's
> considered to have finished executing.

Remember that the SEV-SNP instruction (PSMASH, PVALIDATE,
RMP{ADJUST,UPDATE,QUERY,READ}) have an INVLPGB/TLBSYNC pair under the
hood.  You can execute these instructions on different CPUs in parallel.

It's certainly possible AMD missed something and there's and there's a
deadlock case in there.  But Google do offer SEV-SNP VMs and have the
data and scale to know whether such a deadlock is happening in practice.

>
> I am sure people thought about that and I am probably worried over
> nothing, but there's little details here so one has to speculate.
>
> Again, sorry if I am making a fuss over nothing and it's all in my head.

It's absolutely a valid question to ask.

But x86 is full of longer delays than this.  The GIF for example can
block NMIs until the hypervisor is complete with the world switch, and
it's left as an exercise to software not to abuse this.  Taking an SMI
will be orders of magnitude more expensive than anything discussed here.

~Andrew

Powered by blists - more mailing lists