Re: [PATCH 00/18] KVM: Consolidate and optimize MMU notifiers

[Paolo Bonzini <pbonzini@redhat.com>]

On 26/03/21 03:19, Sean Christopherson wrote:
> The end goal of this series is to optimize the MMU notifiers to take
> mmu_lock if and only if the notification is relevant to KVM, i.e. the hva
> range overlaps a memslot.   Large VMs (hundreds of vCPUs) are very
> sensitive to mmu_lock being taken for write at inopportune times, and
> such VMs also tend to be "static", e.g. backed by HugeTLB with minimal
> page shenanigans.  The vast majority of notifications for these VMs will
> be spurious (for KVM), and eliding mmu_lock for spurious notifications
> avoids an otherwise unacceptable disruption to the guest.
> 
> To get there without potentially degrading performance, e.g. due to
> multiple memslot lookups, especially on non-x86 where the use cases are
> largely unknown (from my perspective), first consolidate the MMU notifier
> logic by moving the hva->gfn lookups into common KVM.
> 
> Applies on my TDP MMU TLB flushing bug fixes[*], which conflict horribly
> with the TDP MMU changes in this series.  That code applies on kvm/queue
> (commit 4a98623d5d90, "KVM: x86/mmu: Mark the PAE roots as decrypted for
> shadow paging").
> 
> Speaking of conflicts, Ben will soon be posting a series to convert a
> bunch of TDP MMU flows to take mmu_lock only for read.  Presumably there
> will be an absurd number of conflicts; Ben and I will sort out the
> conflicts in whichever series loses the race.
> 
> Well tested on Intel and AMD.  Compile tested for arm64, MIPS, PPC,
> PPC e500, and s390.  Absolutely needs to be tested for real on non-x86,
> I give it even odds that I introduced an off-by-one bug somewhere.
> 
> [*] https://lkml.kernel.org/r/20210325200119.1359384-1-seanjc@google.com
> 
> 
> Patches 1-7 are x86 specific prep patches to play nice with moving
> the hva->gfn memslot lookups into common code.  There ended up being waaay
> more of these than I expected/wanted, but I had a hell of a time getting
> the flushing logic right when shuffling the memslot and address space
> loops.  In the end, I was more confident I got things correct by batching
> the flushes.
> 
> Patch 8 moves the existing API prototypes into common code.  It could
> technically be dropped since the old APIs are gone in the end, but I
> thought the switch to the new APIs would suck a bit less this way.
> 
> Patch 9 moves arm64's MMU notifier tracepoints into common code so that
> they are not lost when arm64 is converted to the new APIs, and so that all
> architectures can benefit.
> 
> Patch 10 moves x86's memslot walkers into common KVM.  I chose x86 purely
> because I could actually test it.  All architectures use nearly identical
> code, so I don't think it actually matters in the end.
> 
> Patches 11-13 move arm64, MIPS, and PPC to the new APIs.
> 
> Patch 14 yanks out the old APIs.
> 
> Patch 15 adds the mmu_lock elision, but only for unpaired notifications.
> 
> Patch 16 adds mmu_lock elision for paired .invalidate_range_{start,end}().
> This is quite nasty and no small part of me thinks the patch should be
> burned with fire (I won't spoil it any further), but it's also the most
> problematic scenario for our particular use case.  :-/
> 
> Patches 17-18 are additional x86 cleanups.

Queued and 1-9 and 18, thanks.  There's a small issue in patch 10 that 
prevented me from committing 10-15, but they mostly look good.

Paolo