On Fri, Nov 19, 2021 at 02:32:05PM +0300, Andrey Ryabinin wrote:

On 11/16/21 8:00 AM, Jason Wang wrote:

quoted

On Mon, Nov 15, 2021 at 11:32 PM Andrey Ryabinin [off-list ref] wrote:

quoted

Currently vhost_net_release() uses synchronize_rcu() to synchronize
freeing with vhost_zerocopy_callback(). However synchronize_rcu()
is quite costly operation. It take more than 10 seconds
to shutdown qemu launched with couple net devices like this:
        -netdev tap,id=tap0,..,vhost=on,queues=80
because we end up calling synchronize_rcu() netdev_count*queues times.

Free vhost net structures in rcu callback instead of using
synchronize_rcu() to fix the problem.

I admit the release code is somehow hard to understand. But I wonder
if the following case can still happen with this:

CPU 0 (vhost_dev_cleanup)   CPU1
(vhost_net_zerocopy_callback()->vhost_work_queue())
                                                if (!dev->worker)
dev->worker = NULL

wake_up_process(dev->worker)

If this is true. It seems the fix is to move RCU synchronization stuff
in vhost_net_ubuf_put_and_wait()?

It all depends whether vhost_zerocopy_callback() can be called outside of vhost
thread context or not. If it can run after vhost thread stopped, than the race you
describe seems possible and the fix in commit b0c057ca7e83 ("vhost: fix a theoretical race in device cleanup")
wasn't complete. I would fix it by calling synchronize_rcu() after vhost_net_flush()
and before vhost_dev_cleanup().

As for the performance problem, it can be solved by replacing synchronize_rcu() with synchronize_rcu_expedited().

expedited causes a stop of IPIs though, so it's problematic to
do it upon a userspace syscall.

But now I'm not sure that this race is actually exists and that synchronize_rcu() needed at all.
I did a bit of testing and I only see callback being called from vhost thread:

vhost-3724  3733 [002]  2701.768731: probe:vhost_zerocopy_callback: (ffffffff81af8c10)
        ffffffff81af8c11 vhost_zerocopy_callback+0x1 ([kernel.kallsyms])
        ffffffff81bb34f6 skb_copy_ubufs+0x256 ([kernel.kallsyms])
        ffffffff81bce621 __netif_receive_skb_core.constprop.0+0xac1 ([kernel.kallsyms])
        ffffffff81bd062d __netif_receive_skb_one_core+0x3d ([kernel.kallsyms])
        ffffffff81bd0748 netif_receive_skb+0x38 ([kernel.kallsyms])
        ffffffff819a2a1e tun_get_user+0xdce ([kernel.kallsyms])
        ffffffff819a2cf4 tun_sendmsg+0xa4 ([kernel.kallsyms])
        ffffffff81af9229 handle_tx_zerocopy+0x149 ([kernel.kallsyms])
        ffffffff81afaf05 handle_tx+0xc5 ([kernel.kallsyms])
        ffffffff81afce86 vhost_worker+0x76 ([kernel.kallsyms])
        ffffffff811581e9 kthread+0x169 ([kernel.kallsyms])
        ffffffff810018cf ret_from_fork+0x1f ([kernel.kallsyms])
                       0 [unknown] ([unknown])

This means that the callback can't run after kthread_stop() in vhost_dev_cleanup() and no synchronize_rcu() needed.

I'm not confident that my quite limited testing cover all possible vhost_zerocopy_callback() callstacks.