Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Date:   Thu, 29 Sep 2022 03:47:08 -0400
From:   "Michael S. Tsirkin" <mst@redhat.com>
To:     Stefano Garzarella <sgarzare@redhat.com>
Cc:     Junichi Uekawa <uekawa@chromium.org>,
        Stefan Hajnoczi <stefanha@redhat.com>,
        Jason Wang <jasowang@redhat.com>,
        Eric Dumazet <edumazet@google.com>, davem@davemloft.net,
        netdev@vger.kernel.org, Jakub Kicinski <kuba@kernel.org>,
        virtualization@lists.linux-foundation.org, kvm@vger.kernel.org,
        Paolo Abeni <pabeni@redhat.com>, linux-kernel@vger.kernel.org,
        Bobby Eshleman <bobby.eshleman@gmail.com>
Subject: Re: [PATCH] vhost/vsock: Use kvmalloc/kvfree for larger packets.
Message-ID: <20220929034552-mutt-send-email-mst@kernel.org>
References: <20220928064538.667678-1-uekawa@chromium.org>
 <20220928082823.wyxplop5wtpuurwo@sgarzare-redhat>
 <20220928052738-mutt-send-email-mst@kernel.org>
 <20220928151135.pvrlsylg6j3hzh74@sgarzare-redhat>
 <20220928160116-mutt-send-email-mst@kernel.org>
 <20220929074010.37mksjmwr3l4wlwt@sgarzare-redhat>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220929074010.37mksjmwr3l4wlwt@sgarzare-redhat>
Precedence: bulk

On Thu, Sep 29, 2022 at 09:40:10AM +0200, Stefano Garzarella wrote:
> On Wed, Sep 28, 2022 at 04:02:12PM -0400, Michael S. Tsirkin wrote:
> > On Wed, Sep 28, 2022 at 05:11:35PM +0200, Stefano Garzarella wrote:
> > > On Wed, Sep 28, 2022 at 05:31:58AM -0400, Michael S. Tsirkin wrote:
> > > > On Wed, Sep 28, 2022 at 10:28:23AM +0200, Stefano Garzarella wrote:
> > > > > On Wed, Sep 28, 2022 at 03:45:38PM +0900, Junichi Uekawa wrote:
> > > > > > When copying a large file over sftp over vsock, data size is usually 32kB,
> > > > > > and kmalloc seems to fail to try to allocate 32 32kB regions.
> > > > > >
> > > > > > Call Trace:
> > > > > >  [<ffffffffb6a0df64>] dump_stack+0x97/0xdb
> > > > > >  [<ffffffffb68d6aed>] warn_alloc_failed+0x10f/0x138
> > > > > >  [<ffffffffb68d868a>] ? __alloc_pages_direct_compact+0x38/0xc8
> > > > > >  [<ffffffffb664619f>] __alloc_pages_nodemask+0x84c/0x90d
> > > > > >  [<ffffffffb6646e56>] alloc_kmem_pages+0x17/0x19
> > > > > >  [<ffffffffb6653a26>] kmalloc_order_trace+0x2b/0xdb
> > > > > >  [<ffffffffb66682f3>] __kmalloc+0x177/0x1f7
> > > > > >  [<ffffffffb66e0d94>] ? copy_from_iter+0x8d/0x31d
> > > > > >  [<ffffffffc0689ab7>] vhost_vsock_handle_tx_kick+0x1fa/0x301 [vhost_vsock]
> > > > > >  [<ffffffffc06828d9>] vhost_worker+0xf7/0x157 [vhost]
> > > > > >  [<ffffffffb683ddce>] kthread+0xfd/0x105
> > > > > >  [<ffffffffc06827e2>] ? vhost_dev_set_owner+0x22e/0x22e [vhost]
> > > > > >  [<ffffffffb683dcd1>] ? flush_kthread_worker+0xf3/0xf3
> > > > > >  [<ffffffffb6eb332e>] ret_from_fork+0x4e/0x80
> > > > > >  [<ffffffffb683dcd1>] ? flush_kthread_worker+0xf3/0xf3
> > > > > >
> > > > > > Work around by doing kvmalloc instead.
> > > > > >
> > > > > > Signed-off-by: Junichi Uekawa <uekawa@chromium.org>
> > > >
> > > > My worry here is that this in more of a work around.
> > > > It would be better to not allocate memory so aggressively:
> > > > if we are so short on memory we should probably process
> > > > packets one at a time. Is that very hard to implement?
> > > 
> > > Currently the "virtio_vsock_pkt" is allocated in the "handle_kick" callback
> > > of TX virtqueue. Then the packet is multiplexed on the right socket queue,
> > > then the user space can de-queue it whenever they want.
> > > 
> > > So maybe we can stop processing the virtqueue if we are short on memory, but
> > > when can we restart the TX virtqueue processing?
> > 
> > Assuming you added at least one buffer, the time to restart would be
> > after that buffer has been used.
> 
> Yes, but we still might not have as many continuous pages to allocate, so I
> would use kvmalloc the same.


you would do something like
	if (is_vmalloc_addr())
		stop adding buffers.


> I agree that we should do better, I hope that moving to sk_buff will allow
> us to better manage allocation. Maybe after we merge that part we should
> spend some time to solve these problems.
> 
> Thanks,
> Stefano