kvmarm.lists.cs.columbia.edu archive mirror
 help / color / mirror / Atom feed
From: Marc Zyngier <maz@kernel.org>
To: Zenghui Yu <yuzenghui@huawei.com>
Cc: andre.przywara@arm.com, linux-kernel@vger.kernel.org,
	linux-arm-kernel@lists.infradead.org,
	kvmarm@lists.cs.columbia.edu
Subject: Re: [PATCH] KVM: arm64: vgic-v3: Clear pending bit in guest memory after synchronization
Date: Tue, 31 Mar 2020 09:07:09 +0100	[thread overview]
Message-ID: <20200331090709.17d2087d@why> (raw)
In-Reply-To: <20200331031245.1562-1-yuzenghui@huawei.com>

Hi Zenghui,

On Tue, 31 Mar 2020 11:12:45 +0800
Zenghui Yu <yuzenghui@huawei.com> wrote:

> When LPI support is enabled at redistributor level, VGIC will potentially
> load the correspond LPI penging table and sync it into the pending_latch.
> To avoid keeping the 'consumed' pending bits lying around in guest memory
> (though they're not used), let's clear them after synchronization.
> 
> The similar work had been done in vgic_v3_lpi_sync_pending_status().
> 
> Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
> ---
>  virt/kvm/arm/vgic/vgic-its.c | 23 +++++++++++++++++------
>  1 file changed, 17 insertions(+), 6 deletions(-)
> 
> diff --git a/virt/kvm/arm/vgic/vgic-its.c b/virt/kvm/arm/vgic/vgic-its.c
> index d53d34a33e35..905760bfa404 100644
> --- a/virt/kvm/arm/vgic/vgic-its.c
> +++ b/virt/kvm/arm/vgic/vgic-its.c
> @@ -435,6 +435,7 @@ static int its_sync_lpi_pending_table(struct kvm_vcpu *vcpu)
>  
>  	for (i = 0; i < nr_irqs; i++) {
>  		int byte_offset, bit_nr;
> +		bool status;
>  
>  		byte_offset = intids[i] / BITS_PER_BYTE;
>  		bit_nr = intids[i] % BITS_PER_BYTE;
> @@ -447,22 +448,32 @@ static int its_sync_lpi_pending_table(struct kvm_vcpu *vcpu)
>  			ret = kvm_read_guest_lock(vcpu->kvm,
>  						  pendbase + byte_offset,
>  						  &pendmask, 1);
> -			if (ret) {
> -				kfree(intids);
> -				return ret;
> -			}
> +			if (ret)
> +				goto out;
>  			last_byte_offset = byte_offset;
>  		}
>  
> +		status = pendmask & (1 << bit_nr);
> +
>  		irq = vgic_get_irq(vcpu->kvm, NULL, intids[i]);
>  		raw_spin_lock_irqsave(&irq->irq_lock, flags);
> -		irq->pending_latch = pendmask & (1U << bit_nr);
> +		irq->pending_latch = status;
>  		vgic_queue_irq_unlock(vcpu->kvm, irq, flags);
>  		vgic_put_irq(vcpu->kvm, irq);
> +
> +		if (status) {
> +			/* clear consumed data */
> +			pendmask &= ~(1 << bit_nr);
> +			ret = kvm_write_guest_lock(vcpu->kvm,
> +						   pendbase + byte_offset,
> +						   &pendmask, 1);
> +			if (ret)
> +				goto out;
> +		}
>  	}
>  
> +out:
>  	kfree(intids);
> -
>  	return ret;
>  }
>  

I've been thinking about this, and I wonder why we don't simply clear
the whole pending table instead of carefully wiping it one bit at a
time. My reasoning is that if a LPI isn't mapped, then it cannot be made
pending the first place.

And I think there is a similar issue in vgic_v3_lpi_sync_pending_status().
Why sync something back from the pending table when the LPI wasn't
mapped yet? This seems pretty bizarre, as the GITS_TRANSLATER spec says
that the write to this register is ignored when:

"- The EventID is mapped to an Interrupt Translation Table and the
EventID is within the range specified by MAPD on page 5-107, but the
EventID is unmapped."

(with the added bonus in the form of a type: the first instance of
"EventID" here should obviously be "DeviceID").

What do you think?

	M.
-- 
Jazz is not dead. It just smells funny...
_______________________________________________
kvmarm mailing list
kvmarm@lists.cs.columbia.edu
https://lists.cs.columbia.edu/mailman/listinfo/kvmarm

  reply	other threads:[~2020-03-31  8:07 UTC|newest]

Thread overview: 5+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-03-31  3:12 [PATCH] KVM: arm64: vgic-v3: Clear pending bit in guest memory after synchronization Zenghui Yu
2020-03-31  8:07 ` Marc Zyngier [this message]
2020-03-31  9:11   ` Zenghui Yu
2020-04-01 10:27     ` Marc Zyngier
2020-04-01 12:52       ` Zenghui Yu

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20200331090709.17d2087d@why \
    --to=maz@kernel.org \
    --cc=andre.przywara@arm.com \
    --cc=kvmarm@lists.cs.columbia.edu \
    --cc=linux-arm-kernel@lists.infradead.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=yuzenghui@huawei.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).