Re: [PATCH v4 5/6] KVM: arm64: Add HDBSS fault handling and buffer flush

[Inochi Amaoto <inochiama@gmail.com>]

On Tue, Jul 21, 2026 at 03:18:09PM +0100, Leonardo Bras wrote:
> On Tue, Jul 21, 2026 at 04:53:08PM +0800, Inochi Amaoto wrote:
> > On Fri, Jul 17, 2026 at 04:44:24PM +0100, Leonardo Bras wrote:
> > > On Fri, Jul 17, 2026 at 02:51:12PM +0800, Tian Zheng wrote:
> > > > 
> > > > On 7/14/2026 10:19 PM, Leonardo Bras wrote:
> > > > > On Tue, Jul 14, 2026 at 09:27:15PM +0800, Tian Zheng wrote:
> > > > > > On 7/14/2026 6:50 PM, Leonardo Bras wrote:
> > > > > > > On Tue, Jul 14, 2026 at 03:38:39PM +0800, Tian Zheng wrote:
> > > > > > > > On 7/13/2026 10:06 PM, Leonardo Bras wrote:
> > > > > > > > > On Thu, Jul 09, 2026 at 06:40:25PM +0800, Tian Zheng wrote:
> > > > > > > > > > From: eillon <yezhenyu2@huawei.com>
> > > > > > > > > > 
> > > > > > > > > > Add HDBSS fault handling for buffer full, external abort, and general
> > > > > > > > > > protection fault (GPF) events. When the HDBSS buffer becomes full,
> > > > > > > > > > the hardware traps to EL2 with an HDBSSF event, which is handled by
> > > > > > > > > > setting a flush request.
> > > > > > > > > > 
> > > > > > > > > > Add kvm_flush_hdbss_buffer() to consume HDBSS buffer entries and
> > > > > > > > > > propagate dirty information into the userspace-visible dirty bitmap.
> > > > > > > > > > Flush is triggered on vcpu_put, check_vcpu_requests, and
> > > > > > > > > > sync_dirty_log.
> > > > > > > > > > 
> > > > > > > > > > Add esr_iss2_is_hdbssf() helper for HDBSS fault detection in guest
> > > > > > > > > > abort handling.
> > > > > > > > > > 
> > > > > > > > > > Signed-off-by: Eillon <yezhenyu2@huawei.com>
> > > > > > > > > > Signed-off-by: Tian Zheng <zhengtian10@huawei.com>
> > > > > > > > > > ---
> > > > > > > > > >     arch/arm64/include/asm/esr.h           |  5 +++
> > > > > > > > > >     arch/arm64/include/asm/kvm_dirty_bit.h | 11 +++++
> > > > > > > > > >     arch/arm64/include/asm/kvm_host.h      |  1 +
> > > > > > > > > >     arch/arm64/kvm/arm.c                   | 14 ++++++
> > > > > > > > > >     arch/arm64/kvm/dirty_bit.c             | 62 ++++++++++++++++++++++++++
> > > > > > > > > >     arch/arm64/kvm/mmu.c                   |  4 ++
> > > > > > > > > >     6 files changed, 97 insertions(+)
> > > > > > > > > > 
> > > > > > > > > > diff --git a/arch/arm64/include/asm/esr.h b/arch/arm64/include/asm/esr.h
> > > > > > > > > > index 81c17320a588..2e6b679b5908 100644
> > > > > > > > > > --- a/arch/arm64/include/asm/esr.h
> > > > > > > > > > +++ b/arch/arm64/include/asm/esr.h
> > > > > > > > > > @@ -437,6 +437,11 @@
> > > > > > > > > >     #ifndef __ASSEMBLER__
> > > > > > > > > >     #include <asm/types.h>
> > > > > > > > > > 
> > > > > > > > > > +static inline bool esr_iss2_is_hdbssf(unsigned long esr)
> > > > > > > > > > +{
> > > > > > > > > > +	return ESR_ELx_ISS2(esr) & ESR_ELx_HDBSSF;
> > > > > > > > > This will return a long, which will be casted as bool.
> > > > > > > > > In general, what I see in the kernel is something like:
> > > > > > > > > 
> > > > > > > > > 	return !!(ESR_ELx_ISS2(esr) & ESR_ELx_HDBSSF)
> > > > > > > > ok!
> > > > > > > > 
> > > > > > > > 
> > > > > > > > > > +}
> > > > > > > > > > +
> > > > > > > > > >     static inline unsigned long esr_brk_comment(unsigned long esr)
> > > > > > > > > >     {
> > > > > > > > > >     	return esr & ESR_ELx_BRK64_ISS_COMMENT_MASK;
> > > > > > > > > > diff --git a/arch/arm64/include/asm/kvm_dirty_bit.h b/arch/arm64/include/asm/kvm_dirty_bit.h
> > > > > > > > > > index 84b12f0a10af..4b28000e972f 100644
> > > > > > > > > > --- a/arch/arm64/include/asm/kvm_dirty_bit.h
> > > > > > > > > > +++ b/arch/arm64/include/asm/kvm_dirty_bit.h
> > > > > > > > > > @@ -10,7 +10,18 @@
> > > > > > > > > >     #include <asm/kvm_pgtable.h>
> > > > > > > > > >     #include <asm/sysreg.h>
> > > > > > > > > > 
> > > > > > > > > > +/* HDBSS entry field definitions */
> > > > > > > > > > +#define HDBSS_ENTRY_VALID	BIT(0)
> > > > > > > > > > +#define HDBSS_ENTRY_TTWL_SHIFT (1)
> > > > > > > > > > +#define HDBSS_ENTRY_TTWL_MASK (GENMASK(3, 1))
> > > > > > > > > > +#define HDBSS_ENTRY_TTWL(x) \
> > > > > > > > > > +	(((x) << HDBSS_ENTRY_TTWL_SHIFT) & HDBSS_ENTRY_TTWL_MASK)
> > > > > > > > > > +#define HDBSS_ENTRY_TTWL_RESV HDBSS_ENTRY_TTWL(-4)
> > > > > > > > > > +#define HDBSS_ENTRY_IPA	GENMASK_ULL(55, 12)
> > > > > > > > > > +
> > > > > > > > > >     int kvm_arm_vcpu_alloc_hdbss(struct kvm_vcpu *vcpu, unsigned int order);
> > > > > > > > > >     void kvm_arm_vcpu_free_hdbss(struct kvm_vcpu *vcpu);
> > > > > > > > > > +void kvm_flush_hdbss_buffer(struct kvm_vcpu *vcpu);
> > > > > > > > > > +int kvm_handle_hdbss_fault(struct kvm_vcpu *vcpu);
> > > > > > > > > > 
> > > > > > > > > >     #endif /* __ARM64_KVM_DIRTY_BIT_H__ */
> > > > > > > > > > diff --git a/arch/arm64/include/asm/kvm_host.h b/arch/arm64/include/asm/kvm_host.h
> > > > > > > > > > index c41ec6d9c45a..cecfb884a64f 100644
> > > > > > > > > > --- a/arch/arm64/include/asm/kvm_host.h
> > > > > > > > > > +++ b/arch/arm64/include/asm/kvm_host.h
> > > > > > > > > > @@ -55,6 +55,7 @@
> > > > > > > > > >     #define KVM_REQ_GUEST_HYP_IRQ_PENDING	KVM_ARCH_REQ(9)
> > > > > > > > > >     #define KVM_REQ_MAP_L1_VNCR_EL2		KVM_ARCH_REQ(10)
> > > > > > > > > >     #define KVM_REQ_VGIC_PROCESS_UPDATE	KVM_ARCH_REQ(11)
> > > > > > > > > > +#define KVM_REQ_FLUSH_HDBSS			KVM_ARCH_REQ(12)
> > > > > > > > > > 
> > > > > > > > > >     #define KVM_DIRTY_LOG_MANUAL_CAPS   (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
> > > > > > > > > >     				     KVM_DIRTY_LOG_INITIALLY_SET)
> > > > > > > > > > diff --git a/arch/arm64/kvm/arm.c b/arch/arm64/kvm/arm.c
> > > > > > > > > > index bf6688245d83..566953a4e23a 100644
> > > > > > > > > > --- a/arch/arm64/kvm/arm.c
> > > > > > > > > > +++ b/arch/arm64/kvm/arm.c
> > > > > > > > > > @@ -755,6 +755,9 @@ void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
> > > > > > > > > >     		kvm_vcpu_put_hw_mmu(vcpu);
> > > > > > > > > >     	kvm_arm_vmid_clear_active();
> > > > > > > > > > 
> > > > > > > > > > +	if (vcpu->kvm->arch.enable_hdbss)
> > > > > > > > > > +		kvm_flush_hdbss_buffer(vcpu);
> > > > > > > > > > +
> > > > > > > > > >     	vcpu_clear_on_unsupported_cpu(vcpu);
> > > > > > > > > >     	vcpu->cpu = -1;
> > > > > > > > > >     }
> > > > > > > > > > @@ -1157,6 +1160,9 @@ static int check_vcpu_requests(struct kvm_vcpu *vcpu)
> > > > > > > > > >     		if (kvm_dirty_ring_check_request(vcpu))
> > > > > > > > > >     			return 0;
> > > > > > > > > > 
> > > > > > > > > > +		if (kvm_check_request(KVM_REQ_FLUSH_HDBSS, vcpu))
> > > > > > > > > > +			kvm_flush_hdbss_buffer(vcpu);
> > > > > > > > > > +
> > > > > > > > > >     		check_nested_vcpu_requests(vcpu);
> > > > > > > > > >     	}
> > > > > > > > > > 
> > > > > > > > > > @@ -1971,7 +1977,15 @@ long kvm_arch_vcpu_unlocked_ioctl(struct file *filp, unsigned int ioctl,
> > > > > > > > > > 
> > > > > > > > > >     void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
> > > > > > > > > >     {
> > > > > > > > > > +	/*
> > > > > > > > > > +	 * Flush all CPUs' dirty log buffers to the dirty_bitmap.  Called
> > > > > > > > > > +	 * before reporting dirty_bitmap to userspace. Send a request with
> > > > > > > > > > +	 * KVM_REQUEST_WAIT to flush buffer synchronously.
> > > > > > > > > > +	 */
> > > > > > > > > > +	if (!kvm->arch.enable_hdbss)
> > > > > > > > > > +		return;
> > > > > > > > > > 
> > > > > > > > > > +	kvm_make_all_cpus_request(kvm, KVM_REQ_FLUSH_HDBSS | KVM_REQUEST_WAIT);
> > > > > > > > > >     }
> > > > > > > > > > 
> > > > > > > > > >     static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
> > > > > > > > > > diff --git a/arch/arm64/kvm/dirty_bit.c b/arch/arm64/kvm/dirty_bit.c
> > > > > > > > > > index 6c7a6ef66b5a..002366337637 100644
> > > > > > > > > > --- a/arch/arm64/kvm/dirty_bit.c
> > > > > > > > > > +++ b/arch/arm64/kvm/dirty_bit.c
> > > > > > > > > > @@ -50,3 +50,65 @@ void kvm_arm_vcpu_free_hdbss(struct kvm_vcpu *vcpu)
> > > > > > > > > > 
> > > > > > > > > >     	vcpu->arch.hdbss.hdbssbr_el2 = 0;
> > > > > > > > > >     }
> > > > > > > > > > +
> > > > > > > > > > +void kvm_flush_hdbss_buffer(struct kvm_vcpu *vcpu)
> > > > > > > > > > +{
> > > > > > > > > > +	int idx, curr_idx;
> > > > > > > > > > +	u64 *hdbss_buf;
> > > > > > > > > > +	struct kvm *kvm = vcpu->kvm;
> > > > > > > > > > +
> > > > > > > > > > +	if (!kvm->arch.enable_hdbss)
> > > > > > > > > > +		return;
> > > > > > > > > > +
> > > > > > > > > > +	curr_idx = HDBSSPROD_IDX(read_sysreg_s(SYS_HDBSSPROD_EL2));
> > > > > > > > > > +
> > > > > > > > > > +	/* Do nothing if HDBSS buffer is empty or br_el2 is NULL */
> > > > > > > > > > +	if (curr_idx == 0 || vcpu->arch.hdbss.hdbssbr_el2 == 0)
> > > > > > > > > > +		return;
> > > > > > > > > > +
> > > > > > > > > > +	hdbss_buf = page_address(phys_to_page(vcpu->arch.hdbss.base_phys));
> > > > > > > > > > +	if (!hdbss_buf)
> > > > > > > > > > +		return;
> > > > > > > > > > +
> > > > > > > > > > +	guard(write_lock_irqsave)(&vcpu->kvm->mmu_lock);
> > > > > > > > > > +	for (idx = 0; idx < curr_idx; idx++) {
> > > > > > > > > > +		u64 gpa;
> > > > > > > > > > +
> > > > > > > > > > +		gpa = hdbss_buf[idx];
> > > > > > > > > > +		if (!(gpa & HDBSS_ENTRY_VALID))
> > > > > > > > > > +			continue;
> > > > > > > > > > +
> > > > > > > > > > +		gpa &= HDBSS_ENTRY_IPA;
> > > > > > > > > > +		kvm_vcpu_mark_page_dirty(vcpu, gpa >> PAGE_SHIFT);
> > > > > > > > > You mention that it does not support dirty-ring, but above function will
> > > > > > > > > mark the page as dirty in the dirty-ring :/
> > > > > > > > > 
> > > > > > > > In kvm_arm_enable_hdbss_global(), we explicitly check and reject HDBSS
> > > > > > > > enablement if dirty-ring is active:
> > > > > > > > 
> > > > > > > > ```
> > > > > > > > if (kvm->dirty_ring_size)
> > > > > > > >       return 0;
> > > > > > > > ```
> > > > > > > > 
> > > > > > > > So when kvm_flush_hdbss_buffer() runs (which requires enable_hdbss = true),
> > > > > > > > we know for certain that
> > > > > > > > 
> > > > > > > > kvm->dirty_ring_size == 0. Therefore, kvm_vcpu_mark_page_dirty() will always
> > > > > > > > take the dirty_bitmap path,
> > > > > > > > 
> > > > > > > > never the dirty-ring path.
> > > > > > > > 
> > > > > > > > That said, I'll add a comment in kvm_flush_hdbss_buffer() before dirty ring
> > > > > > > > mode is supported, to make this explicit:
> > > > > > > > 
> > > > > > > > ```
> > > > > > > > /*
> > > > > > > >    * HDBSS is mutually exclusive with dirty-ring mode (see
> > > > > > > >    * kvm_arm_enable_hdbss_global()), so kvm_vcpu_mark_page_dirty()
> > > > > > > >    * will update the dirty_bitmap, not the dirty-ring.
> > > > > > > >    */
> > > > > > > > ```
> > > > > > > > 
> > > > > > > Got it :)
> > > > > > > 
> > > > > > > Out of curiosity: which issues have you found on supporting dirty-ring at
> > > > > > > this point?
> > > > > > > 
> > > > > > > Thanks!
> > > > > > > Leo
> > > > > > 
> > > > > > I haven't looked deeply into dirty-ring yet — my main concern is that if
> > > > > > both the dirty
> > > > > > 
> > > > > > ring and HDBSS buffer fill up, the flush path might get blocked or
> > > > > > complicated.
> > > > > > 
> > > > > > For now, I'm planning to match the HDBSS buffer size to the dirty ring size
> > > > > > in v5 and test it.
> > > > > > 
> > > > > > Ideally, the two buffers would be the same size, and the entire dirty
> > > > > > tracking path would use
> > > > > > 
> > > > > Ah, I see the point.
> > > > > 
> > > > > IIRC, when dirty-ring gets full, the kernel returns to userspace with
> > > > > run->exit_reason == KVM_EXIT_DIRTY_RING_FULL, which will warn the VMM to
> > > > > drain the dirty-ring, and that makes space for us draining HDBSS to the
> > > > > dirty-ring again.
> > > > > 
> > > > > The best way to achieve that, as I remember, is to always drain
> > > > > HDBSS as much as possible at guest_exitting. That will make more space to
> > > > > newer HDBSS entries, and we can get userspace to drain the dirty-ring
> > > > > earlier.
> > > > > 
> > > > > I would say to even make HDBSS buffer half (entries) the dirty-ring. Then
> > > > > we can generally fully drain to the dirty-ring and even report ring full
> > > > > if the ring is above a given threshold percentage full.
> > > > > > HDBSS exclusively — no fallback to the legacy dirty bitmap path. If that
> > > > > > works, I think this approach should be fine.
> > > > > > 
> > > > > > Let me know if you have any insights on dirty ring's full-buffer behavior —
> > > > > > that would be helpful.
> > > > > > 
> > > > > > 
> > > > > Will do!
> > > > > 
> > > > > Thanks!
> > > > > Leo
> > > > 
> > > > 
> > > > Thanks for the insights — reusing PML's reservation mechanism makes sense.
> > > > 
> > > > I think we can*keep the HDBSS buffer at 512 entries* (matching
> > > 
> > > I recommend using a PAGESIZE (512 in 4k, but bigger in other sizes)
> > > 
> > > > PML_LOG_NR_ENTRIES)
> > > > 
> > > > for now, and *not expose any ioctl for userspace to configure it*. Since the
> > > > kernel
> > > > 
> > > > auto-enables HDBSS, a userspace size knob would be confusing.
> > > > 
> > > 
> > > Humm, I am in favor of letting the user change it according to it's 
> > > workload. Why would that be confusing?
> > > 
> > > (Having a default size is useful just for enabling it to work without any 
> > > change in current VMs)
> > > 
> > > > 
> > > > The reservation logic would be:
> > > > 
> > > > - Implement kvm_cpu_dirty_log_size() on arm64 to return the HDBSS buffer
> > > > entry
> > > > 
> > > > count (512, or 0 if HDBSS is not enabled)
> > > 
> > > Why a get to log_size? does userspace need to know it's using HDBSS?
> > > 
> > > Just a set should do, as VMMs can just try to set a value, and if it fails 
> > > (IOCTL does not exist, or invalid value), then it can just go forward. 
> > > 
> > > > 
> > > > - Reuse kvm_dirty_ring_get_rsvd_entries() to reserve space for one full
> > > > flush:
> > > > 
> > > > KVM_DIRTY_RING_RSVD_ENTRIES + hdbss_entries
> > > > 
> > > > - So soft_limit = dirty_ring_size - (KVM_DIRTY_RING_RSVD_ENTRIES +
> > > > hdbss_entries),
> > > > 
> > > > guaranteeing a full flush always fits
> > > > 
> > > > kvm_flush_hdbss_buffer() at guest_exit pushes via mark_page_dirty_in_slot()
> > > > ->
> > > > 
> > > 
> > > I think I get the point here: since we can have sw dirtying as well as 
> > > HDBSS tracking, we may get to the point that we don't have enough space to 
> > > flush hdbss -> dirty-ring, right?
> > > 
> > > > kvm_dirty_ring_push() -> kvm_dirty_ring_soft_full, which triggers
> > > > 
> > > > KVM_EXIT_DIRTY_RING_FULL when needed.
> > > > 
> > > > If soft_limit is hit, KVM_REQ_DIRTY_RING_FULL is set and the next vcpu_run
> > > > exits
> > > > 
> > > > to userspace for QEMU to drain the ring.
> > > 
> > > So the software dirtying routine would be affected by the soft limit, but 
> > > HADBSS exit would not. It means the dirty-ring would be effectively smaller 
> > > than specified if we are having mostly sw dirtying.
> > > 
> > > Did I get that right?
> > > 
> > > > 
> > > > *One more thing: *when userspace sets the dirty ring size via
> > > > 
> > > > KVM_VM_IOCTL_ENABLE_DIRTY_LOG_RING, we already enforce that
> > > > 
> > > > size >= kvm_dirty_ring_get_rsvd_entries(kvm) * sizeof(struct kvm_dirty_gfn)
> > > > 
> > > > or size < PAGE_SIZE. With HDBSS, kvm_dirty_ring_get_rsvd_entries() will
> > > > include the
> > > > 
> > > > HDBSS entries via kvm_cpu_dirty_log_size(), so the same check will
> > > > automatically guarantee
> > > > 
> > > > the ring is large enough to accommodate the HDBSS buffer. No additional
> > > > validation is needed.
> > > > 
> > > 
> > 
> 
> Hi Inochi,
> 
> > I have a small question on this, since HDBSS supports 2MiB buffer,
> > it  will has more entries than the dirty ring.
> 
> The entries reserved for HDBSS are part of the dirty ring, but I get the 
> point: having a dirty-ring with 64k entries and HDBSS with 256k entries 
> (2MB/8) would be weird.
> 
> For the dirty-ring I think the plan is to have HDBSS buffer to be a fraction
> of the configured dirty-ring size.
> 
> > If we use 
> > kvm_cpu_dirty_log_size() to reserve entries. It will always enter
> > soft limit if the HDBSS buffer is huge. 
> > 
> 
> Yeah, but once HDBSS is active if we enable dirty_tracking, there should 
> not be a lot of entries being added to the dirty ring that don't come from 
> HDBSS, so it should be fine. If it happens, though, it will just exit guest 
> and drain as it happens nowadays.
> 

You are true, I found the IOMMU reports the dirty log in a different
way, so when enabling HDBSS, the dirty log in this scene should mainly
come from HDBSS, which means the limit check mainly replies on the 
the state of HDBSS. 

> 
> > Since I think users set large buffer expect less context switch.
> > I think this may require some change on the dirty ring framework.
> > At least I think the soft limit check should be adjusted.
> > 
> 
> That may not be the case: the memory usage of the dirty-ring is 
> (16 * dirty_ring_size * vcpus) in bytes. 
> 
> On a VM with 1024 vcpus, only the HDBSS buffer alone would be 2GB, plus 4GB 
> of the reserved dirty-ring for that buffer.
> 

Yeah, it is the case, the memory usage is huge in this case, and it
also hits the case what I said, the user uses huge buffer to reduce
the number of KVM exit.

Although it is kind of off topic, I think if a machine can have a
VM with 1024 vCPU in production. It should has so much memory (in TB)
that it may be acceptable for the a big dirty log buffer in a short
time.

> Which brings the discussion: maybe we should have a buffer management 
> happen instead of using the reserving strategy: If the buffer does not fit 
> in the dirty-ring, just move the remaining entries to the start, adjust the 
> index register, and continue. 
> 
> (On a 2MB size it would be bad, as there could be many remaining entries)
> 

I guess this could be something bad in most cases, even a 256KB buffer
(32K entry) means we have to move half of the buffer to the start...

> Alternatively, we can just keep guest_exitting until all HDBSS entries fit 
> the dirty_ring, which introduce overhead, but should work as well.
> 
> All options have some tradeoff, though.
> 

Right, everything method has some tradeoff, I guess for the first
version, just using the alternative method could be a simple idea.
This is because the HDBSS does provide an improvement on the dirty
log tracking even the implement is not very perfect.

Regards,
Inochi