Re: [PULL v2 31/82] vhost: Change the sequence of device start

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Nov 07, 2022 at 06:31:00PM +0100, Christian A. Ehrhardt wrote:
> On Mon, Nov 07, 2022 at 08:30:32AM -0500, Michael S. Tsirkin wrote:
> > On Sun, Nov 06, 2022 at 07:00:33PM +0100, Christian A. Ehrhardt wrote:
> > > 
> > > Hi,
> > > 
> > > On Sat, Nov 05, 2022 at 12:43:05PM -0400, Michael S. Tsirkin wrote:
> > > > On Sat, Nov 05, 2022 at 05:35:57PM +0100, Bernhard Beschow wrote:
> > > > > 
> > > > > 
> > > > > On Wed, Nov 2, 2022 at 5:24 PM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > > > 
> > > > >     From: Yajun Wu <yajunw@nvidia.com>
> > > > > 
> > > > >     This patch is part of adding vhost-user vhost_dev_start support. The
> > > > >     motivation is to improve backend configuration speed and reduce live
> > > > >     migration VM downtime.
> > > > > 
> > > > >     Moving the device start routines after finishing all the necessary device
> > > > >     and VQ configuration, further aligning to the virtio specification for
> > > > >     "device initialization sequence".
> > > > > 
> > > > >     Following patch will add vhost-user vhost_dev_start support.
> > > > > 
> > > > >     Signed-off-by: Yajun Wu <yajunw@nvidia.com>
> > > > >     Acked-by: Parav Pandit <parav@nvidia.com>
> > > > > 
> > > > >     Message-Id: <20221017064452.1226514-2-yajunw@nvidia.com>
> > > > >     Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
> > > > >     Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
> > > > >     ---
> > > > >      hw/block/vhost-user-blk.c | 18 +++++++++++-------
> > > > >      hw/net/vhost_net.c        | 12 ++++++------
> > > > >      2 files changed, 17 insertions(+), 13 deletions(-)
> > > > > 
> > > > > 
> > > > > A git bisect tells me that this is the first bad commit for failing qos-tests
> > > > > which only fail when parallel jobs are enabled, e.g. `make check-qtest -j8`:
> > > 
> > > Parallel test run is not required provided that the test machine is
> > > sufficiently busy (load > number of CPU threads). In this case a single
> > > invocation of the qos test will fail reliably with this change.
> > > 
> > > However, the change is not really the root cause of the failures.
> > > 
> > > > > Summary of Failures:
> > > > > 
> > > > >  76/541 qemu:qtest+qtest-aarch64 / qtest-aarch64/qos-test                      
> > > > >   ERROR          18.68s   killed by signal 6 SIGABRT
> > > > >  77/541 qemu:qtest+qtest-arm / qtest-arm/qos-test                              
> > > > >   ERROR          17.60s   killed by signal 6 SIGABRT
> > > > >  93/541 qemu:qtest+qtest-i386 / qtest-i386/qos-test                            
> > > > >   ERROR          18.98s   killed by signal 6 SIGABRT
> > > > > 108/541 qemu:qtest+qtest-ppc64 / qtest-ppc64/qos-test                          
> > > > >   ERROR          16.40s   killed by signal 6 SIGABRT
> > > > > 112/541 qemu:qtest+qtest-i386 / qtest-i386/bios-tables-test                    
> > > > >   ERROR          145.94s   killed by signal 6 SIGABRT
> > > > > 130/541 qemu:qtest+qtest-x86_64 / qtest-x86_64/qos-test                        
> > > > >   ERROR          17.32s   killed by signal 6 SIGABRT
> > > > > 243/541 qemu:qtest+qtest-x86_64 / qtest-x86_64/bios-tables-test                
> > > > >   ERROR          127.70s   killed by signal 6 SIGABRT
> > > > > 
> > > > > Ok:                 500
> > > > > Expected Fail:      0  
> > > > > Fail:               7  
> > > > > Unexpected Pass:    0  
> > > > > Skipped:            34  
> > > > > Timeout:            0  
> > > > > 
> > > > > Can anyone else reproduce this?
> > > > 
> > > > Could you pls try latest for_upstream in my tree?
> > > > That should have this fixed.
> > > 
> > > Your new pull request simply drops this change and this does fix
> > > make check-qtest. However, this looks accidental to me and the real
> > > bug is there in plain origin/master, too.
> > > 
> > > What happens is this backtrace a recursive call to vu_gpio_stop
> > > via the backtrace below. It is caused by a delayed of the TCP
> > > connection (the delayed part only triggers with heavy load on the
> > > machine).
> > > 
> > > You can get the failure back (probably in upstream) if the test is
> > > forced to us "use-started=off" which can be set on the command line.
> > > E.g. like this:
> > > 
> > > diff --git a/tests/qtest/libqos/virtio-gpio.c b/tests/qtest/libqos/virtio-gpio.c
> > > index 762aa6695b..17c6b71e8b 100644
> > > --- a/tests/qtest/libqos/virtio-gpio.c
> > > +++ b/tests/qtest/libqos/virtio-gpio.c
> > > @@ -154,14 +154,14 @@ static void virtio_gpio_register_nodes(void)
> > >      QOSGraphEdgeOptions edge_opts = { };
> > > 
> > >      /* vhost-user-gpio-device */
> > > -    edge_opts.extra_device_opts = "id=gpio0,chardev=chr-vhost-user-test";
> > > +    edge_opts.extra_device_opts = "id=gpio0,chardev=chr-vhost-user-test,use-started=off";
> > >      qos_node_create_driver("vhost-user-gpio-device",
> > >                              virtio_gpio_device_create);
> > >      qos_node_consumes("vhost-user-gpio-device", "virtio-bus", &edge_opts);
> > >      qos_node_produces("vhost-user-gpio-device", "vhost-user-gpio");
> > > 
> > >      /* virtio-gpio-pci */
> > > -    edge_opts.extra_device_opts = "id=gpio0,addr=04.0,chardev=chr-vhost-user-test";
> > > +    edge_opts.extra_device_opts = "id=gpio0,addr=04.0,chardev=chr-vhost-user-test,use-started=on";
> > >      add_qpci_address(&edge_opts, &addr);
> > >      qos_node_create_driver("vhost-user-gpio-pci", virtio_gpio_pci_create);
> > >      qos_node_consumes("vhost-user-gpio-pci", "pci-bus", &edge_opts);
> > > 
> > > 
> > > I haven't verified this but from looking at the code other types of
> > > vhost devices seem to have the same problem (e.g. vhost-user-i2c looks
> > > suspicious).
> > > 
> > > Ok, here's the backtrace:
> > > 
> > > #0  vu_gpio_stop (vdev=vdev@entry=0x560e0ae449d0) at ../hw/virtio/vhost-user-gpio.c:143
> > > #1  0x0000560e0768fb1f in vu_gpio_disconnect (dev=<optimized out>) at ../hw/virtio/vhost-user-gpio.c:260
> > > #2  vu_gpio_event (opaque=<optimized out>, event=<optimized out>) at ../hw/virtio/vhost-user-gpio.c:279
> > > #3  0x0000560e078057b5 in tcp_chr_disconnect_locked (chr=0x560e0a00f800) at ../chardev/char-socket.c:470
> > > #4  0x0000560e078058d3 in tcp_chr_write (chr=0x560e0a00f800, buf=<optimized out>, len=<optimized out>) at ../chardev/char-socket.c:129
> > > #5  0x0000560e07808a63 in qemu_chr_write_buffer (s=s@entry=0x560e0a00f800, buf=buf@entry=0x7ffc76812ac0 <error: Cannot access memory at address 0x7ffc76812ac0>, len=20, offset=offset@entry=0x7ffc76812a40, write_all=write_all@entry=true) at ../chardev/char.c:121
> > > #6  0x0000560e07808e84 in qemu_chr_write (s=0x560e0a00f800, buf=buf@entry=0x7ffc76812ac0 <error: Cannot access memory at address 0x7ffc76812ac0>, len=len@entry=20, write_all=write_all@entry=true) at ../chardev/char.c:173
> > > #7  0x0000560e078010e6 in qemu_chr_fe_write_all (be=be@entry=0x560e0ae44bc0, buf=buf@entry=0x7ffc76812ac0 <error: Cannot access memory at address 0x7ffc76812ac0>, len=len@entry=20) at ../chardev/char-fe.c:53
> > > #8  0x0000560e07676d16 in vhost_user_write (msg=msg@entry=0x7ffc76812ac0, fds=fds@entry=0x0, fd_num=fd_num@entry=0, dev=<optimized out>, dev=<optimized out>) at ../hw/virtio/vhost-user.c:490
> > > #9  0x0000560e076777c0 in vhost_user_get_vring_base (dev=0x560e0ae44c08, ring=0x7ffc76812d80) at ../hw/virtio/vhost-user.c:1260
> > > #10 0x0000560e0767314e in vhost_virtqueue_stop (dev=dev@entry=0x560e0ae44c08, vdev=vdev@entry=0x560e0ae449d0, vq=0x560e0ae7a570, idx=0) at ../hw/virtio/vhost.c:1220
> > > #11 0x0000560e07675236 in vhost_dev_stop (hdev=hdev@entry=0x560e0ae44c08, vdev=vdev@entry=0x560e0ae449d0) at ../hw/virtio/vhost.c:1884
> > > #12 0x0000560e0768f2fd in vu_gpio_stop (vdev=0x560e0ae449d0) at ../hw/virtio/vhost-user-gpio.c:148
> > > #13 0x0000560e0766a6a3 in virtio_set_status (vdev=0x560e0ae449d0, val=<optimized out>) at ../hw/virtio/virtio.c:2442
> > > #14 0x0000560e0751a2e0 in vm_state_notify (running=running@entry=false, state=state@entry=RUN_STATE_SHUTDOWN) at ../softmmu/runstate.c:334
> > > #15 0x0000560e07514d70 in do_vm_stop (send_stop=false, state=RUN_STATE_SHUTDOWN) at ../softmmu/cpus.c:262
> > > #16 vm_shutdown () at ../softmmu/cpus.c:280
> > > #17 0x0000560e0751af63 in qemu_cleanup () at ../softmmu/runstate.c:827
> > > #18 0x0000560e07350f33 in qemu_default_main () at ../softmmu/main.c:38
> > > #19 0x00007f0a5655bd90 in __libc_start_call_main (main=main@entry=0x560e0734c3e0 <main>, argc=argc@entry=25, argv=argv@entry=0x7ffc768130a8) at ../sysdeps/nptl/libc_start_call_main.h:58
> > > #20 0x00007f0a5655be40 in __libc_start_main_impl (main=0x560e0734c3e0 <main>, argc=25, argv=0x7ffc768130a8, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>, stack_end=0x7ffc76813098) at ../csu/libc-start.c:392
> > > #21 0x0000560e07350e55 in _start ()
> > > 
> > > 
> > > Any Ideas how to fix this properly?
> > > 
> > >     regards    Christian
> > 
> > so fundamentally, any write into a socket can trigger stop?
> > I don't think we realized this.
> > Hmm this needs some thought there are likely more issues
> > like this.
> 
> The code in net/vhost-user.c seems to realize this and uses
> aio to schedule the close. I was wondering if chardev should
> do this for OPEN/CLOSE events by default (or by request).
> However, I couldn't figure out how to ensure that the device
> remains alive until the callback is actually called.
> 
>       regards   Christian

Cc Alex who's been looking at this.

-- 
MST