+Jeremi and Mathias, my scripts for sending patches to stable don't auto-cc :-/
On Wed, Aug 23, 2023, Sean Christopherson wrote:
> Disable the TDP MMU by default in v5.15 kernels to "fix" several severe
> performance bugs that have since been found and fixed in the TDP MMU, but
> are unsuitable for backporting to v5.15.
>
> The problematic bugs are fixed by upstream commit edbdb43fc96b ("KVM:
> x86: Preserve TDP MMU roots until they are explicitly invalidated") and
> commit 01b31714bd90 ("KVM: x86: Do not unload MMU roots when only toggling
> CR0.WP with TDP enabled"). Both commits fix scenarios where KVM will
> rebuild all TDP MMU page tables in paths that are frequently hit by
> certain guest workloads. While not exactly common, the guest workloads
> are far from rare. The fallout of rebuilding TDP MMU page tables can be
> so severe in some cases that it induces soft lockups in the guest.
>
> Commit edbdb43fc96b would require _significant_ effort and churn to
> backport due it depending on a major rework that was done in v5.18.
>
> Commit 01b31714bd90 has far fewer direct conflicts, but has several subtle
> _known_ dependencies, and it's unclear whether or not there are more
> unknown dependencies that have been missed.
>
> Lastly, disabling the TDP MMU in v5.15 kernels also fixes a lurking train
> wreck started by upstream commit a955cad84cda ("KVM: x86/mmu: Retry page
> fault if root is invalidated by memslot update"). That commit was tagged
> for stable to fix a memory leak, but didn't cherry-pick cleanly and was
> never backported to v5.15. Which is extremely fortunate, as it introduced
> not one but two bugs, one of which was fixed by upstream commit
> 18c841e1f411 ("KVM: x86: Retry page fault if MMU reload is pending and
> root has no sp"), while the other was unknowingly fixed by upstream
> commit ba6e3fe25543 ("KVM: x86/mmu: Grab mmu_invalidate_seq in
> kvm_faultin_pfn()") in v6.3 (a one-off fix will be made for v6.1 kernels,
> which did receive a backport for a955cad84cda). Disabling the TDP MMU
> by default reduces the probability of breaking v5.15 kernels by
> backporting only a subset of the fixes.
>
> As far as what is lost by disabling the TDP MMU, the main selling point of
> the TDP MMU is its ability to service page fault VM-Exits in parallel,
> i.e. the main benefactors of the TDP MMU are deployments of large VMs
> (hundreds of vCPUs), and in particular delployments that live-migrate such
> VMs and thus need to fault-in huge amounts of memory on many vCPUs after
> restarting the VM after migration.
>
> Smaller VMs can see performance improvements, but nowhere enough to make
> up for the TDP MMU (in v5.15) absolutely cratering performance for some
> workloads. And practically speaking, anyone that is deploying and
> migrating VMs with hundreds of vCPUs is likely rolling their own kernel,
> not using a stock v5.15 series kernel.
>
> This reverts commit 71ba3f3189c78f756a659568fb473600fd78f207.
>
> Link: https://lore.kernel.org/all/ZDmEGM+CgYpvDLh6@xxxxxxxxxx
> Link: https://lore.kernel.org/all/f023d927-52aa-7e08-2ee5-59a2fbc65953@xxxxxxxxxxxxxxx
> Cc: Jeremi Piotrowski <jpiotrowski@xxxxxxxxxxxxxxxxxxx>
> Cc: Mathias Krause <minipli@xxxxxxxxxxxxxx>
> Signed-off-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> ---
> arch/x86/kvm/mmu/tdp_mmu.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index 6c2bb60ccd88..7a64fb238044 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -10,7 +10,7 @@
> #include <asm/cmpxchg.h>
> #include <trace/events/kvm.h>
>
> -static bool __read_mostly tdp_mmu_enabled = true;
> +static bool __read_mostly tdp_mmu_enabled = false;
> module_param_named(tdp_mmu, tdp_mmu_enabled, bool, 0644);
>
> /* Initializes the TDP MMU for the VM, if enabled. */
>
> base-commit: f6f7927ac664ba23447f8dd3c3dfe2f4ee39272f
> --
> 2.42.0.rc2.253.gd59a3bf2b4-goog
>
