If you like running QEMU as a normal user (very common for TCG runs)
but you have to run virtiofsd as a root user you run into connection
problems. Adding support for an optional --socket-group allows the
users to keep using the command line.
Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20200925125147.26943-2-alex.bennee@linaro.org>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
dgilbert: Split long line
glib offers thread pools and it seems to support "exclusive" and "shared"
thread pools.
https://developer.gnome.org/glib/stable/glib-Thread-Pools.html#g-thread-pool-new
Currently we use "exlusive" thread pools but its performance seems to be
poor. I tried using "shared" thread pools and performance seems much
better. I posted performance results here.
https://www.redhat.com/archives/virtio-fs/2020-September/msg00080.html
So lets switch to shared thread pools. We can think of making it optional
once somebody can show in what cases exclusive thread pools offer better
results. For now, my simple performance tests across the board see
better results with shared thread pools.
Signed-off-by: Vivek Goyal <vgoyal@redhat.com>
Message-Id: <20200921213216.GE13362@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
With seccomp fix from Miklos
An assertion failure is raised during request processing if
unshare(CLONE_FS) fails. Implement a probe at startup so the problem can
be detected right away.
Unfortunately Docker/Moby does not include unshare in the seccomp.json
list unless CAP_SYS_ADMIN is given. Other seccomp.json lists always
include unshare (e.g. podman is unaffected):
https://raw.githubusercontent.com/seccomp/containers-golang/master/seccomp.json
Use "docker run --security-opt seccomp=path/to/seccomp.json ..." if the
default seccomp.json is missing unshare.
Cc: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Message-Id: <20200727190223.422280-4-stefanha@redhat.com>
Reviewed-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Current virtiofsd has problems about xattr operations and
they does not work properly for directory/symlink/special file.
The fundamental cause is that virtiofsd uses openat() + f...xattr()
systemcalls for xattr operation but we should not open symlink/special
file in the daemon. Therefore the function is restricted.
Fix this problem by:
1. during setup of each thread, call unshare(CLONE_FS)
2. in xattr operations (i.e. lo_getxattr), if inode is not a regular
file or directory, use fchdir(proc_loot_fd) + ...xattr() +
fchdir(root.fd) instead of openat() + f...xattr()
(Note: for a regular file/directory openat() + f...xattr()
is still used for performance reason)
With this patch, xfstests generic/062 passes on virtiofs.
This fix is suggested by Miklos Szeredi and Stefan Hajnoczi.
The original discussion can be found here:
https://www.redhat.com/archives/virtio-fs/2019-October/msg00046.html
Signed-off-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Message-Id: <20200227055927.24566-3-misono.tomohiro@jp.fujitsu.com>
Acked-by: Vivek Goyal <vgoyal@redhat.com>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
If we fail when bringing up the socket we can leak the listen_fd;
in practice the daemon will exit so it's not really a problem.
Fixes: Coverity CID 1413121
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
On guest graceful shutdown, virtiofsd receives VHOST_USER_GET_VRING_BASE
request from VMM and shuts down virtqueues by calling fv_set_started(),
which joins fv_queue_thread() threads. So when virtio_loop() returns,
there should be no thread is still accessing data in fuse session and/or
virtio dev.
But on abnormal exit, e.g. guest got killed for whatever reason,
vhost-user socket is closed and virtio_loop() breaks out the main loop
and returns to main(). But it's possible fv_queue_worker()s are still
working and accessing fuse session and virtio dev, which results in
crash or use-after-free.
Fix it by stopping fv_queue_thread()s before virtio_loop() returns,
to make sure there's no-one could access fuse session and virtio dev.
Reported-by: Qingming Su <qingming.su@linux.alibaba.com>
Signed-off-by: Eryu Guan <eguan@linux.alibaba.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Add an option to control the size of the thread pool. Requests are now
processed in parallel by default.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Introduce a thread pool so that fv_queue_thread() just pops
VuVirtqElements and hands them to the thread pool. For the time being
only one worker thread is allowed since passthrough_ll.c is not
thread-safe yet. Future patches will lift this restriction so that
multiple FUSE requests can be processed in parallel.
The main new concept is struct FVRequest, which contains both
VuVirtqElement and struct fuse_chan. We now have fv_VuDev for a device,
fv_QueueInfo for a virtqueue, and FVRequest for a request. Some of
fv_QueueInfo's fields are moved into FVRequest because they are
per-request. The name FVRequest conforms to QEMU coding style and I
expect the struct fv_* types will be renamed in a future refactoring.
This patch series is not optimal. fbuf reuse is dropped so each request
does malloc(se->bufsize), but there is no clean and cheap way to keep
this with a thread pool. The vq_lock mutex is held for longer than
necessary, especially during the eventfd_write() syscall. Performance
can be improved in the future.
prctl(2) had to be added to the seccomp whitelist because glib invokes
it.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
We call into libvhost-user from the virtqueue handler thread and the
vhost-user message processing thread without a lock. There is nothing
protecting the virtqueue handler thread if the vhost-user message
processing thread changes the virtqueue or memory table while it is
running.
This patch introduces a read-write lock. Virtqueue handler threads are
readers. The vhost-user message processing thread is a writer. This
will allow concurrency for multiqueue in the future while protecting
against fv_queue_thread() vs virtio_loop() races.
Note that the critical sections could be made smaller but it would be
more invasive and require libvhost-user changes. Let's start simple and
improve performance later, if necessary. Another option would be an
RCU-style approach with lighter-weight primitives.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
For fuse's queueinfo, both queueinfo array and queueinfos are allocated in
fv_queue_set_started() but not cleaned up when the daemon process quits.
This fixes the leak in proper places.
Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
Signed-off-by: Eric Ren <renzhen@linux.alibaba.com>
Reviewed-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
virtiofsd can run multiply even if the vhost_user_socket is same path.
]# ./virtiofsd -o vhost_user_socket=/tmp/vhostqemu -o source=/tmp/share &
[1] 244965
virtio_session_mount: Waiting for vhost-user socket connection...
]# ./virtiofsd -o vhost_user_socket=/tmp/vhostqemu -o source=/tmp/share &
[2] 244966
virtio_session_mount: Waiting for vhost-user socket connection...
]#
The user will get confused about the situation and maybe the cause of the
unexpected problem. So it's better to prevent the multiple running.
Create a regular file under localstatedir directory to exclude the
vhost_user_socket. To create and lock the file, use qemu_write_pidfile()
because the API has some sanity checks and file lock.
Signed-off-by: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Applied fixes from Stefan's review and moved osdep include
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
This cleans up unfreed resources in se on quiting, including
se->virtio_dev, se->vu_socket_path, se->vu_socketfd.
Signed-off-by: Liu Bo <bo.liu@linux.alibaba.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Kill the threads we've started when the queues get stopped.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
With improvements by:
Signed-off-by: Eryu Guan <eguan@linux.alibaba.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Pass the write iov pointing to guest RAM all the way through rather
than copying the data.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Xiao Yang <yangx.jy@cn.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Let fuse_session_process_buf_int take a fuse_bufvec * instead of a
fuse_buf; and then through to do_write_buf - where in the best
case it can pass that straight through to op.write_buf without copying
(other than skipping a header).
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Although --socket-path=PATH is useful for manual invocations, management
tools typically create the UNIX domain socket themselves and pass it to
the vhost-user device backend. This way QEMU can be launched
immediately with a valid socket. No waiting for the vhost-user device
backend is required when fd passing is used.
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Readv the data straight into the guests buffer.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
With fix by:
Signed-off-by: Eryu Guan <eguan@linux.alibaba.com>
Reviewed-by: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Keep track of whether we sent a reply to a request; this is a bit
paranoid but it means:
a) We should always recycle an element even if there was an error
in the request
b) Never try and send two replies on one queue element
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Route fuse out messages back through the same queue elements
that had the command that triggered the request.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Pop queue elements off queues, copy the data from them and
pass that to fuse.
Note: 'out' in a VuVirtqElement is from QEMU
'in' in libfuse is into the daemon
So we read from the out iov's to get a fuse_in_header
When we get a kick we've got to read all the elements until the queue
is empty.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
In the queue thread poll the kick_fd we're passed.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Start a thread for each queue when we get notified it's been started.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
fix by:
Signed-off-by: Jun Piao <piaojun@huawei.com>
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Add the get/set features callbacks.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Processes incoming requests on the vhost-user fd.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Listen on our unix socket for the connection from QEMU, when we get it
initialise vhost-user and dive into our own loop variant (currently
dummy).
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
When run with vhost-user options we conect to the QEMU instead
via a socket. Start this off by creating the socket.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Misono Tomohiro <misono.tomohiro@jp.fujitsu.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Alex Bennée
Vivek Goyal
Stefan Hajnoczi
Misono Tomohiro
Dr. David Alan Gilbert
Eryu Guan
Liu Bo
Masayoshi Mizuma