RE: [PATCH v3 4/4] KVM: arm64: Clear active_vmids on vCPU schedule out

[Shameerali Kolothum Thodi]

Hi Will,

> -----Original Message-----
> From: Will Deacon [mailto:will@xxxxxxxxxx]
> Sent: 03 August 2021 16:31
> To: Shameerali Kolothum Thodi <shameerali.kolothum.thodi@xxxxxxxxxx>
> Cc: linux-arm-kernel@xxxxxxxxxxxxxxxxxxx; kvmarm@xxxxxxxxxxxxxxxxxxxxx;
> linux-kernel@xxxxxxxxxxxxxxx; maz@xxxxxxxxxx; catalin.marinas@xxxxxxx;
> james.morse@xxxxxxx; julien.thierry.kdev@xxxxxxxxx;
> suzuki.poulose@xxxxxxx; jean-philippe@xxxxxxxxxx;
> Alexandru.Elisei@xxxxxxx; qperret@xxxxxxxxxx; Linuxarm
> <linuxarm@xxxxxxxxxx>
> Subject: Re: [PATCH v3 4/4] KVM: arm64: Clear active_vmids on vCPU
> schedule out

[...]
 
> I think we have to be really careful not to run into the "suspended
> animation" problem described in ae120d9edfe9 ("ARM: 7767/1: let the ASID
> allocator handle suspended animation") if we go down this road.
> 
> Maybe something along the lines of:
> 
> ROLLOVER
> 
>   * Take lock
>   * Inc generation
>     => This will force everybody down the slow path
>   * Record active VMIDs
>   * Broadcast TLBI
>     => Only active VMIDs can be dirty
>     => Reserve active VMIDs and mark as allocated
> 
> VCPU SCHED IN
> 
>   * Set active VMID
>   * Check generation
>   * If mismatch then:
>         * Take lock
>         * Try to match a reserved VMID
>         * If no reserved VMID, allocate new
> 
> VCPU SCHED OUT
> 
>   * Clear active VMID
> 
> but I'm not daft enough to think I got it right first time. I think it
> needs both implementing *and* modelling in TLA+ before we merge it!

I attempted to implement the above algo as below. It seems to be
working in both 16-bit vmid and 4-bit vmid test setup. Though I am
not quite sure this Is exactly what you had in mind above and covers
all corner cases.

Please take a look and let me know.
(The diff below is against this v3 series)

Thanks,
Shameer

--->8<----

--- a/arch/arm64/kvm/vmid.c
+++ b/arch/arm64/kvm/vmid.c
@@ -43,7 +43,7 @@ static void flush_context(void)
        bitmap_clear(vmid_map, 0, NUM_USER_VMIDS);

        for_each_possible_cpu(cpu) {
-               vmid = atomic64_xchg_relaxed(&per_cpu(active_vmids, cpu), 0);
+               vmid = atomic64_read(&per_cpu(active_vmids, cpu));

                /* Preserve reserved VMID */
                if (vmid == 0)
@@ -125,32 +125,17 @@ void kvm_arm_vmid_clear_active(void)
 void kvm_arm_vmid_update(struct kvm_vmid *kvm_vmid)
 {
        unsigned long flags;
-       u64 vmid, old_active_vmid;
+       u64 vmid;

        vmid = atomic64_read(&kvm_vmid->id);
-
-       /*
-        * Please refer comments in check_and_switch_context() in
-        * arch/arm64/mm/context.c.
-        */
-       old_active_vmid = atomic64_read(this_cpu_ptr(&active_vmids));
-       if (old_active_vmid && vmid_gen_match(vmid) &&
-           atomic64_cmpxchg_relaxed(this_cpu_ptr(&active_vmids),
-                                    old_active_vmid, vmid))
+       if (vmid_gen_match(vmid)) {
+               atomic64_set(this_cpu_ptr(&active_vmids), vmid);
                return;
-
-       raw_spin_lock_irqsave(&cpu_vmid_lock, flags);
-
-       /* Check that our VMID belongs to the current generation. */
-       vmid = atomic64_read(&kvm_vmid->id);
-       if (!vmid_gen_match(vmid)) {
-               vmid = new_vmid(kvm_vmid);
-               atomic64_set(&kvm_vmid->id, vmid);
        }

-
+       raw_spin_lock_irqsave(&cpu_vmid_lock, flags);
+       vmid = new_vmid(kvm_vmid);
+       atomic64_set(&kvm_vmid->id, vmid);
        atomic64_set(this_cpu_ptr(&active_vmids), vmid);
        raw_spin_unlock_irqrestore(&cpu_vmid_lock, flags);
 }
--->8<----