Message-ID: <zlhixzduyindq24osaedkt2xnukmatwhugfkqmaugvor6wlcol@56jsodxn4rhi>
Date: Thu, 11 Dec 2025 14:56:53 +0100
From: Stefano Garzarella <sgarzare@...hat.com>
To: "Michael S. Tsirkin" <mst@...hat.com>
Cc: Melbin K Mathew <mlbnkm1@...il.com>, stefanha@...hat.com, 
	kvm@...r.kernel.org, netdev@...r.kernel.org, virtualization@...ts.linux.dev, 
	linux-kernel@...r.kernel.org, jasowang@...hat.com, xuanzhuo@...ux.alibaba.com, 
	eperezma@...hat.com, davem@...emloft.net, edumazet@...gle.com, kuba@...nel.org, 
	pabeni@...hat.com, horms@...nel.org
Subject: Re: [PATCH net v3] vsock/virtio: cap TX credit to local buffer size

On Thu, Dec 11, 2025 at 08:05:11AM -0500, Michael S. Tsirkin wrote:
>On Thu, Dec 11, 2025 at 01:51:04PM +0100, Melbin K Mathew wrote:
>> The virtio vsock transport currently derives its TX credit directly from
>> peer_buf_alloc, which is populated from the remote endpoint's
>> SO_VM_SOCKETS_BUFFER_SIZE value.
>>
>> On the host side, this means the amount of data we are willing to queue
>> for a given connection is scaled purely by a peer-chosen value, rather
>> than by the host's own vsock buffer configuration. A guest that
>> advertises a very large buffer and reads slowly can cause the host to
>> allocate a correspondingly large amount of sk_buff memory for that
>> connection.
>>
>> In practice, a malicious guest can:
>>
>>   - set a large AF_VSOCK buffer size (e.g. 2 GiB) with
>>     SO_VM_SOCKETS_BUFFER_MAX_SIZE / SO_VM_SOCKETS_BUFFER_SIZE, and
>>
>>   - open multiple connections to a host vsock service that sends data
>>     while the guest drains slowly.
>>
>> On an unconstrained host this can drive Slab/SUnreclaim into the tens of
>> GiB range, causing allocation failures and OOM kills in unrelated host
>> processes while the offending VM remains running.
>>
>> On non-virtio transports and compatibility:
>>
>>   - VMCI uses the AF_VSOCK buffer knobs to size its queue pairs per
>>     socket based on the local vsk->buffer_* values; the remote side
>>     can’t enlarge those queues beyond what the local endpoint
>>     configured.
>>
>>   - Hyper-V’s vsock transport uses fixed-size VMBus ring buffers and
>>     an MTU bound; there is no peer-controlled credit field comparable
>>     to peer_buf_alloc, and the remote endpoint can’t drive in-flight
>>     kernel memory above those ring sizes.
>>
>>   - The loopback path reuses virtio_transport_common.c, so it
>>     naturally follows the same semantics as the virtio transport.
>>
>> Make virtio-vsock consistent with that model by intersecting the peer’s
>> advertised receive window with the local vsock buffer size when
>> computing TX credit. We introduce a small helper and use it in
>> virtio_transport_get_credit(), virtio_transport_has_space() and
>> virtio_transport_seqpacket_enqueue(), so that:
>>
>>     effective_tx_window = min(peer_buf_alloc, buf_alloc)
>>
>> This prevents a remote endpoint from forcing us to queue more data than
>> our own configuration allows, while preserving the existing credit
>> semantics and keeping virtio-vsock compatible with the other transports.
>>
>> On an unpatched Ubuntu 22.04 host (~64 GiB RAM), running a PoC with
>> 32 guest vsock connections advertising 2 GiB each and reading slowly
>> drove Slab/SUnreclaim from ~0.5 GiB to ~57 GiB and the system only
>> recovered after killing the QEMU process.
>>
>> With this patch applied, rerunning the same PoC yields:
>>
>>   Before:
>>     MemFree:        ~61.6 GiB
>>     MemAvailable:   ~62.3 GiB
>>     Slab:           ~142 MiB
>>     SUnreclaim:     ~117 MiB
>>
>>   After 32 high-credit connections:
>>     MemFree:        ~61.5 GiB
>>     MemAvailable:   ~62.3 GiB
>>     Slab:           ~178 MiB
>>     SUnreclaim:     ~152 MiB
>>
>> i.e. only ~35 MiB increase in Slab/SUnreclaim, no host OOM, and the
>> guest remains responsive.
>>
>> Fixes: 06a8fc78367d ("VSOCK: Introduce virtio_vsock_common.ko")
>> Suggested-by: Stefano Garzarella <sgarzare@...hat.com>
>> Signed-off-by: Melbin K Mathew <mlbnkm1@...il.com>
>> ---
>>  net/vmw_vsock/virtio_transport_common.c | 27 ++++++++++++++++++++++---
>>  1 file changed, 24 insertions(+), 3 deletions(-)
>>
>> diff --git a/net/vmw_vsock/virtio_transport_common.c b/net/vmw_vsock/virtio_transport_common.c
>> index dcc8a1d58..02eeb96dd 100644
>> --- a/net/vmw_vsock/virtio_transport_common.c
>> +++ b/net/vmw_vsock/virtio_transport_common.c
>> @@ -491,6 +491,25 @@ void virtio_transport_consume_skb_sent(struct sk_buff *skb, bool consume)
>>  }
>>  EXPORT_SYMBOL_GPL(virtio_transport_consume_skb_sent);
>>
>> +/* Return the effective peer buffer size for TX credit computation.
>> + *
>> + * The peer advertises its receive buffer via peer_buf_alloc, but we
>> + * cap that to our local buf_alloc (derived from
>> + * SO_VM_SOCKETS_BUFFER_SIZE and already clamped to buffer_max_size)
>> + * so that a remote endpoint cannot force us to queue more data than
>> + * our own configuration allows.
>> + */
>> +static u32 virtio_transport_tx_buf_alloc(struct virtio_vsock_sock *vvs)
>> +{
>> +	return min(vvs->peer_buf_alloc, vvs->buf_alloc);
>> +}
>> +
>>  u32 virtio_transport_get_credit(struct virtio_vsock_sock *vvs, u32 credit)
>>  {
>>  	u32 ret;
>> @@ -499,7 +518,8 @@ u32 virtio_transport_get_credit(struct virtio_vsock_sock *vvs, u32 credit)
>>  		return 0;
>>
>>  	spin_lock_bh(&vvs->tx_lock);
>> -	ret = vvs->peer_buf_alloc - (vvs->tx_cnt - vvs->peer_fwd_cnt);
>> +	ret = virtio_transport_tx_buf_alloc(vvs) -
>> +		(vvs->tx_cnt - vvs->peer_fwd_cnt);
>>  	if (ret > credit)
>>  		ret = credit;
>>  	vvs->tx_cnt += ret;
>> @@ -831,7 +851,7 @@ virtio_transport_seqpacket_enqueue(struct vsock_sock *vsk,
>>
>>  	spin_lock_bh(&vvs->tx_lock);
>>
>> -	if (len > vvs->peer_buf_alloc) {
>> +	if (len > virtio_transport_tx_buf_alloc(vvs)) {
>>  		spin_unlock_bh(&vvs->tx_lock);
>>  		return -EMSGSIZE;
>>  	}
>> @@ -882,7 +902,8 @@ static s64 virtio_transport_has_space(struct vsock_sock *vsk)
>>  	struct virtio_vsock_sock *vvs = vsk->trans;
>>  	s64 bytes;
>>
>> -	bytes = (s64)vvs->peer_buf_alloc - (vvs->tx_cnt - vvs->peer_fwd_cnt);
>> +	bytes = (s64)virtio_transport_tx_buf_alloc(vvs) -
>> +		(vvs->tx_cnt - vvs->peer_fwd_cnt);
>>  	if (bytes < 0)
>>  		bytes = 0;
>>
>
>Acked-by: Michael S. Tsirkin <mst@...hat.com>
>
>
>Looking at this, why is one place casting to s64 the other is not?

Yeah, I pointed out that too in previous interactions. IMO we should fix 
virtio_transport_get_credit() since the peer can reduce `peer_buf_alloc` 
so it will overflow. Fortunately, we are limited by the credit requested 
by the caller, but we are still sending stuff when we shouldn't be.

@Melbin let me know if you will fix it, otherwise I can do that, but I'd 
like to do in a single series (multiple patches), since they depends on 
each other.

So if you prefer, I can pickup this patch and post a series with this + 
the other fix + the fix on the test I posted on the v2.

Stefano

Powered by blists - more mailing lists