Re: [RFC v2 14/14] kvm/arm: Align the VMID allocation with the arm64 ASID one

From: Julien Grall
Date: Mon Jul 15 2019 - 13:06:34 EST


On 03/07/2019 18:36, James Morse wrote:
Hi Julien,

Hi James,

On 20/06/2019 14:06, Julien Grall wrote:
At the moment, the VMID algorithm will send an SGI to all the CPUs to
force an exit and then broadcast a full TLB flush and I-Cache
invalidation.

This patch re-use the new ASID allocator. The
benefits are:
- CPUs are not forced to exit at roll-over. Instead the VMID will be
marked reserved and the context will be flushed at next exit. This
will reduce the IPIs traffic.
- Context invalidation is now per-CPU rather than broadcasted.

+ Catalin has a model of the asid-allocator.

That's a good point :).



With the new algo, the code is now adapted:
- The function __kvm_flush_vm_context() has been renamed to
__kvm_flush_cpu_vmid_context and now only flushing the current CPU context.
- The call to update_vttbr() will be done with preemption disabled
as the new algo requires to store information per-CPU.
- The TLBs associated to EL1 will be flushed when booting a CPU to
deal with stale information. This was previously done on the
allocation of the first VMID of a new generation.

The measurement was made on a Seattle based SoC (8 CPUs), with the
number of VMID limited to 4-bit. The test involves running concurrently 40
guests with 2 vCPUs. Each guest will then execute hackbench 5 times
before exiting.

diff --git a/arch/arm64/include/asm/kvm_asid.h b/arch/arm64/include/asm/kvm_asid.h
new file mode 100644
index 000000000000..8b586e43c094
--- /dev/null
+++ b/arch/arm64/include/asm/kvm_asid.h
@@ -0,0 +1,8 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __ARM64_KVM_ASID_H__
+#define __ARM64_KVM_ASID_H__
+
+#include <asm/asid.h>
+
+#endif /* __ARM64_KVM_ASID_H__ */
+
diff --git a/arch/arm64/include/asm/kvm_asm.h b/arch/arm64/include/asm/kvm_asm.h
index ff73f5462aca..06821f548c0f 100644
--- a/arch/arm64/include/asm/kvm_asm.h
+++ b/arch/arm64/include/asm/kvm_asm.h
@@ -62,7 +62,7 @@ extern char __kvm_hyp_init_end[];
extern char __kvm_hyp_vector[];
-extern void __kvm_flush_vm_context(void);
+extern void __kvm_flush_cpu_vmid_context(void);
extern void __kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa);

As we've got a __kvm_tlb_flush_local_vmid(), would __kvm_tlb_flush_local_all() fit in
better? (This mirrors local_flush_tlb_all() too)

I am happy with the renaming here.



extern void __kvm_tlb_flush_vmid(struct kvm *kvm);
extern void __kvm_tlb_flush_local_vmid(struct kvm_vcpu *vcpu);
diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
index 4bcd9c1291d5..7ef45b7da4eb 100644
--- a/arch/arm64/include/asm/kvm_host.h
+++ b/arch/arm64/include/asm/kvm_host.h
@@ -68,8 +68,8 @@ int kvm_arch_vm_ioctl_check_extension(struct kvm *kvm, long ext);
void __extended_idmap_trampoline(phys_addr_t boot_pgd, phys_addr_t idmap_start);
struct kvm_vmid {
- /* The VMID generation used for the virt. memory system */
- u64 vmid_gen;
+ /* The ASID used for the ASID allocator */
+ atomic64_t asid;

Can we call this 'id' as happens in mm_context_t? (calling it asid is confusing)

I am fine with this suggestion.


u32 vmid;

Can we filter out the generation bits in kvm_get_vttbr() in the same way the arch code
does in cpu_do_switch_mm().

I think this saves writing back a cached pre-filtered version every time, or needing
special hooks to know when the value changed. (so we can remove this variable)

[...]

+static void vmid_update_ctxt(void *ctxt)
{
+ struct kvm_vmid *vmid = ctxt;
+ u64 asid = atomic64_read(&vmid->asid);

+ vmid->vmid = asid & ((1ULL << kvm_get_vmid_bits()) - 1);

I don't like having to poke this through the asid-allocator as a kvm-specific hack. Can we
do it in kvm_get_vttbr()?

I will have a look.



}

@@ -487,48 +467,11 @@ static bool need_new_vmid_gen(struct kvm_vmid *vmid)

(git made a mess of the diff here... squashed to just the new code:)

static void update_vmid(struct kvm_vmid *vmid)
{

+ int cpu = get_cpu();
+ asid_check_context(&vmid_info, &vmid->asid, cpu, vmid);
+ put_cpu();

If we're calling update_vmid() in a pre-emptible context, aren't we already doomed?

Yes we are. This made me realize that Linux-RT replaced the preempt_disable() in the caller by migrate_disable(). The latter will prevent the task to move to another CPU but allow preemption.

This patch will likely makes things awfully broken for Linux-RT. I will have a look to see if we can call this from preempt notifier.


Could we use smp_processor_id() instead.


}


@@ -1322,6 +1271,8 @@ static void cpu_init_hyp_mode(void *dummy)
__cpu_init_hyp_mode(pgd_ptr, hyp_stack_ptr, vector_ptr);
__cpu_init_stage2();


+ kvm_call_hyp(__kvm_flush_cpu_vmid_context);

I think we only need to do this for VHE systems too. cpu_hyp_reinit() only does the call
to cpu_init_hyp_mode() if !is_kernel_in_hyp_mode().

I guess you mean we need to do this for VHE system. If so, I agree that cpu_init_hyp_mode() is not the best place. I will move it to cpu_hyp_reinit().



}
static void cpu_hyp_reset(void)
@@ -1429,6 +1380,17 @@ static inline void hyp_cpu_pm_exit(void)
static int init_common_resources(void)
{
+ /*
+ * Initialize the ASID allocator telling it to allocate a single
+ * VMID per VM.
+ */
+ if (asid_allocator_init(&vmid_info, kvm_get_vmid_bits(), 1,
+ vmid_flush_cpu_ctxt, vmid_update_ctxt))
+ panic("Failed to initialize VMID allocator\n");

Couldn't we return an error instead? The asid allocator is needed for user-space, its
pointless to keep running if it fails. The same isn't true here. (and it would make it
easier to debug what went wrong!)

Fair point. I will update the next version.

Cheers,

--
Julien Grall