Re: [PATCH 5.15] Revert "KVM: x86: enable TDP MMU by default"

[Mathias Krause]

On 24.08.23 03:15, Sean Christopherson wrote:
> +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
>> -- 

Acked-by: Mathias Krause <minipli@xxxxxxxxxxxxxx>

I guess this means no hope for
https://lore.kernel.org/stable/234e01b6-1b5c-d682-a078-3dd91a62abf4@xxxxxxxxxxxxxx/
:/