The local_rename() callback is vulnerable to symlink attacks because it
uses rename() which follows symbolic links in all path elements but the
rightmost one.
This patch simply transforms local_rename() into a wrapper around
local_renameat() which is symlink-attack safe.
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_renameat() callback is currently a wrapper around local_rename()
which is vulnerable to symlink attacks.
This patch rewrites local_renameat() to have its own implementation, based
on local_opendir_nofollow() and renameat().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_lstat() callback is vulnerable to symlink attacks because it
calls:
(1) lstat() which follows symbolic links in all path elements but the
rightmost one
(2) getxattr() which follows symbolic links in all path elements
(3) local_mapped_file_attr()->local_fopen()->openat(O_NOFOLLOW) which
follows symbolic links in all path elements but the rightmost
one
This patch converts local_lstat() to rely on opendir_nofollow() and
fstatat(AT_SYMLINK_NOFOLLOW) to fix (1), fgetxattrat_nofollow() to
fix (2).
A new local_fopenat() helper is introduced as a replacement to
local_fopen() to fix (3). No effort is made to factor out code
because local_fopen() will be dropped when all users have been
converted to call local_fopenat().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_readlink() callback is vulnerable to symlink attacks because it
calls:
(1) open(O_NOFOLLOW) which follows symbolic links for all path elements but
the rightmost one
(2) readlink() which follows symbolic links for all path elements but the
rightmost one
This patch converts local_readlink() to rely on open_nofollow() to fix (1)
and opendir_nofollow(), readlinkat() to fix (2).
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_truncate() callback is vulnerable to symlink attacks because
it calls truncate() which follows symbolic links in all path elements.
This patch converts local_truncate() to rely on open_nofollow() and
ftruncate() instead.
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_statfs() callback is vulnerable to symlink attacks because it
calls statfs() which follows symbolic links in all path elements.
This patch converts local_statfs() to rely on open_nofollow() and fstatfs()
instead.
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_utimensat() callback is vulnerable to symlink attacks because it
calls qemu_utimens()->utimensat(AT_SYMLINK_NOFOLLOW) which follows symbolic
links in all path elements but the rightmost one or qemu_utimens()->utimes()
which follows symbolic links for all path elements.
This patch converts local_utimensat() to rely on opendir_nofollow() and
utimensat(AT_SYMLINK_NOFOLLOW) directly instead of using qemu_utimens().
It is hence assumed that the OS supports utimensat(), i.e. has glibc 2.6
or higher and linux 2.6.22 or higher, which seems reasonable nowadays.
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_remove() callback is vulnerable to symlink attacks because it
calls:
(1) lstat() which follows symbolic links in all path elements but the
rightmost one
(2) remove() which follows symbolic links in all path elements but the
rightmost one
This patch converts local_remove() to rely on opendir_nofollow(),
fstatat(AT_SYMLINK_NOFOLLOW) to fix (1) and unlinkat() to fix (2).
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_unlinkat() callback is vulnerable to symlink attacks because it
calls remove() which follows symbolic links in all path elements but the
rightmost one.
This patch converts local_unlinkat() to rely on opendir_nofollow() and
unlinkat() instead.
Most of the code is moved to a separate local_unlinkat_common() helper
which will be reused in a subsequent patch to fix the same issue in
local_remove().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_lremovexattr() callback is vulnerable to symlink attacks because
it calls lremovexattr() which follows symbolic links in all path elements
but the rightmost one.
This patch introduces a helper to emulate the non-existing fremovexattrat()
function: it is implemented with /proc/self/fd which provides a trusted
path that can be safely passed to lremovexattr().
local_lremovexattr() is converted to use this helper and opendir_nofollow().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_lsetxattr() callback is vulnerable to symlink attacks because
it calls lsetxattr() which follows symbolic links in all path elements but
the rightmost one.
This patch introduces a helper to emulate the non-existing fsetxattrat()
function: it is implemented with /proc/self/fd which provides a trusted
path that can be safely passed to lsetxattr().
local_lsetxattr() is converted to use this helper and opendir_nofollow().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_llistxattr() callback is vulnerable to symlink attacks because
it calls llistxattr() which follows symbolic links in all path elements but
the rightmost one.
This patch introduces a helper to emulate the non-existing flistxattrat()
function: it is implemented with /proc/self/fd which provides a trusted
path that can be safely passed to llistxattr().
local_llistxattr() is converted to use this helper and opendir_nofollow().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_lgetxattr() callback is vulnerable to symlink attacks because
it calls lgetxattr() which follows symbolic links in all path elements but
the rightmost one.
This patch introduces a helper to emulate the non-existing fgetxattrat()
function: it is implemented with /proc/self/fd which provides a trusted
path that can be safely passed to lgetxattr().
local_lgetxattr() is converted to use this helper and opendir_nofollow().
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
The local_open() and local_opendir() callbacks are vulnerable to symlink
attacks because they call:
(1) open(O_NOFOLLOW) which follows symbolic links in all path elements but
the rightmost one
(2) opendir() which follows symbolic links in all path elements
This patch converts both callbacks to use new helpers based on
openat_nofollow() to only open files and directories if they are
below the virtfs shared folder
This partly fixes CVE-2016-9602.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
This patch opens the shared folder and caches the file descriptor, so that
it can be used to do symlink-safe path walk.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
When using the passthrough security mode, symbolic links created by the
guest are actual symbolic links on the host file system.
Since the resolution of symbolic links during path walk is supposed to
occur on the client side. The server should hence never receive any path
pointing to an actual symbolic link. This isn't guaranteed by the protocol
though, and malicious code in the guest can trick the server to issue
various syscalls on paths whose one or more elements are symbolic links.
In the case of the "local" backend using the "passthrough" or "none"
security modes, the guest can directly create symbolic links to arbitrary
locations on the host (as per spec). The "mapped-xattr" and "mapped-file"
security modes are also affected to a lesser extent as they require some
help from an external entity to create actual symbolic links on the host,
i.e. another guest using "passthrough" mode for example.
The current code hence relies on O_NOFOLLOW and "l*()" variants of system
calls. Unfortunately, this only applies to the rightmost path component.
A guest could maliciously replace any component in a trusted path with a
symbolic link. This could allow any guest to escape a virtfs shared folder.
This patch introduces a variant of the openat() syscall that successively
opens each path element with O_NOFOLLOW. When passing a file descriptor
pointing to a trusted directory, one is guaranteed to be returned a
file descriptor pointing to a path which is beneath the trusted directory.
This will be used by subsequent patches to implement symlink-safe path walk
for any access to the backend.
Symbolic links aren't the only threats actually: a malicious guest could
change a path element to point to other types of file with undesirable
effects:
- a named pipe or any other thing that would cause openat() to block
- a terminal device which would become QEMU's controlling terminal
These issues can be addressed with O_NONBLOCK and O_NOCTTY.
Two helpers are introduced: one to open intermediate path elements and one
to open the rightmost path element.
Suggested-by: Jann Horn <jannh@google.com>
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
(renamed openat_nofollow() to relative_openat_nofollow(),
assert path is relative and doesn't contain '//',
fixed side-effect in assert, Greg Kurz)
Signed-off-by: Greg Kurz <groug@kaod.org>
If these functions fail, they should not change *fs. Let's use local
variables to fix this.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
These functions are always called indirectly. It really doesn't make sense
for them to sit in a header file.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
All that CoQueue needs in order to become thread-safe is help
from an external mutex. Add this to the API.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Reviewed-by: Fam Zheng <famz@redhat.com>
Message-id: 20170213181244.16297-6-pbonzini@redhat.com
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
cause 9p clients to hang. Other patches are minor enhancements.
-----BEGIN PGP SIGNATURE-----
iEYEABECAAYFAliIegsACgkQAvw66wEB28LjzwCeIKbBFC/hbc43UqaNX82OGd2v
soYAn0YYXJUAykyjNEMLdhhNp+rABzNk
=1PaE
-----END PGP SIGNATURE-----
Merge remote-tracking branch 'remotes/gkurz/tags/for-upstream' into staging
This pull request fixes a 2.9 regression and a long standing bug that can
cause 9p clients to hang. Other patches are minor enhancements.
# gpg: Signature made Wed 25 Jan 2017 10:12:27 GMT
# gpg: using DSA key 0x02FC3AEB0101DBC2
# gpg: Good signature from "Greg Kurz <groug@kaod.org>"
# gpg: aka "Greg Kurz <groug@free.fr>"
# gpg: aka "Greg Kurz <gkurz@fr.ibm.com>"
# gpg: aka "Greg Kurz <gkurz@linux.vnet.ibm.com>"
# gpg: aka "Gregory Kurz (Groug) <groug@free.fr>"
# gpg: aka "Gregory Kurz (Cimai Technology) <gkurz@cimai.com>"
# gpg: aka "Gregory Kurz (Meiosys Technology) <gkurz@meiosys.com>"
# gpg: WARNING: This key is not certified with a trusted signature!
# gpg: There is no indication that the signature belongs to the owner.
# Primary key fingerprint: 2BD4 3B44 535E C0A7 9894 DBA2 02FC 3AEB 0101 DBC2
* remotes/gkurz/tags/for-upstream:
9pfs: fix offset error in v9fs_xattr_read()
9pfs: local: trivial cosmetic fix in pwritev op
9pfs: fix off-by-one error in PDU free list
tests: virtio-9p: improve error reporting
9pfs: add missing coroutine_fn annotations
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
The current code tries to copy `read_count' bytes starting at offset
`offset' from a `read_count`-sized iovec. This causes v9fs_pack() to
fail with ENOBUFS.
Since the PDU iovec is already partially filled with `offset' bytes,
let's skip them when creating `qiov_full' and have v9fs_pack() to
copy the whole of it. Moreover, this is consistent with the other
places where v9fs_init_qiov_from_pdu() is called.
This fixes commit "bcb8998fac16 9pfs: call v9fs_init_qiov_from_pdu
before v9fs_pack".
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
The server can handle MAX_REQ - 1 PDUs at a time and the virtio-9p
device has a MAX_REQ sized virtqueue. If the client manages to fill
up the virtqueue, pdu_alloc() will fail and the request won't be
processed without any notice to the client (it actually causes the
linux 9p client to hang).
This has been there since the beginning (commit 9f10751365 "virtio-9p:
Add a virtio 9p device to qemu"), but it needs an agressive workload to
run in the guest to show up.
We actually allocate MAX_REQ PDUs and I see no reason not to link them
all into the free list, so let's fix the init loop.
Reported-by: Tuomas Tynkkynen <tuomas@tuxera.com>
Suggested-by: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: Greg Kurz <groug@kaod.org>
If a migration is already in progress and somebody attempts
to add a migration blocker, this should rightly fail.
Add an errp parameter and a retcode return value to migrate_add_blocker.
Signed-off-by: John Snow <jsnow@redhat.com>
Signed-off-by: Ashijeet Acharya <ashijeetacharya@gmail.com>
Message-Id: <1484566314-3987-5-git-send-email-ashijeetacharya@gmail.com>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Acked-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Merged with recent 'Allow invtsc migration' change
The u16 and u32 types don't exist in QEMU common headers. It never broke
build because these two macros aren't use by the current code, but this
is about to change with the future addition of functional tests for 9P.
Also, these should have enclosing parenthesis to be usable in any
syntactical situation.
As suggested by Eric Blake, let's use UINT16_MAX and UINT32_MAX to address
both issues.
Signed-off-by: Greg Kurz <groug@kaod.org>
If the user passes -device virtio-9p without the corresponding -fsdev, QEMU
dereferences a NULL pointer and crashes.
This is a 2.8 regression introduced by commit 702dbcc274.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Li Qiang <liq3ea@gmail.com>
Not all 9pfs transports share memory between request and response. For
those who don't, it is necessary to know how much memory is required in
the response.
Split the existing init_iov_from_pdu function in two:
init_out_iov_from_pdu (for writes) and init_in_iov_from_pdu (for reads).
init_in_iov_from_pdu takes an additional size parameter to specify the
memory required for the response message.
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
v9fs_xattr_read should not access VirtQueueElement elems directly.
Move v9fs_init_qiov_from_pdu up in the file and call
v9fs_init_qiov_from_pdu before v9fs_pack. Use v9fs_pack on the new
iovec.
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Don't call virtio functions from 9pfs generic code, use generic function
callbacks instead.
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
pdus are initialized and used in 9pfs common code. Move the array from
V9fsVirtioState to V9fsState.
Signed-off-by: Stefano Stabellini <sstabellini@kernel.org>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
In the init operation of proxy backend dirver, it allocates a
V9fsProxy struct and some other resources. We should free these
resources when the 9pfs device is unrealized. This is what this
patch does.
Signed-off-by: Li Qiang <liq3ea@gmail.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
In the init operation of handle backend dirver, it allocates a
handle_data struct and opens a mount file. We should free these
resources when the 9pfs device is unrealized. This is what this
patch does.
Signed-off-by: Li Qiang <liq3ea@gmail.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Currently, the backend of VirtFS doesn't have a cleanup
function. This will lead resource leak issues if the backed
driver allocates resources. This patch addresses this issue.
Signed-off-by: Li Qiang <liq3ea@gmail.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Unrealize should undo things that were set during realize in
reverse order. So should do in the error path in realize.
Signed-off-by: Li Qiang <liq3ea@gmail.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
The virtfs_reset() function is called either when the virtio-9p device
gets reset, or when the client starts a new 9P session. In both cases,
if it finds fids from a previous session, the following is printed in
the monitor:
9pfs:virtfs_reset: One or more uncluncked fids found during reset
For example, if a linux guest with a mounted 9P share is reset from the
monitor with system_reset, the message will be printed. This is excessive
since these fids are now clunked and the state is clean.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
A buggy or malicious guest could pass the id of an already opened fid and
cause QEMU to abort. Let's return EINVAL to the guest instead.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
The xattrcreate operation only makes sense on a freshly cloned fid
actually, since any open state would be leaked because of the fid_type
change. This is indeed what the linux kernel client does:
fid = clone_fid(fid);
[...]
retval = p9_client_xattrcreate(fid, name, value_len, flags);
This patch also reverts commit ff55e94d23 since we are sure that a fid
with type P9_FID_NONE doesn't have a previously allocated xattr.
Signed-off-by: Greg Kurz <groug@kaod.org>
We shouldn't allow guests to create extended attribute with arbitrary sizes.
On linux hosts, the limit is XATTR_SIZE_MAX. Let's use it.
Signed-off-by: Greg Kurz <groug@kaod.org>
The v9fs_xattr_read() and v9fs_xattr_write() are passed a guest
originated offset: they must ensure this offset does not go beyond
the size of the extended attribute that was set in v9fs_xattrcreate().
Unfortunately, the current code implement these checks with unsafe
calculations on 32 and 64 bit values, which may allow a malicious
guest to cause OOB access anyway.
Fix this by comparing the offset and the xattr size, which are
both uint64_t, before trying to compute the effective number of bytes
to read or write.
Suggested-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-By: Guido Günther <agx@sigxcpu.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
The 'len' in V9fsXattr comes from the 'size' argument in setxattr()
function in guest. The setxattr() function's declaration is this:
int setxattr(const char *path, const char *name,
const void *value, size_t size, int flags);
and 'size' is treated as u64 in linux kernel client code:
int p9_client_xattrcreate(struct p9_fid *fid, const char *name,
u64 attr_size, int flags)
So the 'len' should have an type of 'uint64_t'.
The 'copied_len' in V9fsXattr is used to account for copied bytes, it
should also have an type of 'uint64_t'.
Suggested-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Currently, 9pfs sets the 'copied_len' field in V9fsXattr
to -1 to tag xattr walk fid. As the 'copied_len' is also
used to account for copied bytes, this may make confusion. This patch
add a bool 'xattrwalk_fid' to tag the xattr walk fid.
Suggested-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
If an error occurs when marshalling the transfer length to the guest, the
v9fs_write() function doesn't free an IO vector, thus leading to a memory
leak. This patch fixes the issue.
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
[groug, rephrased the changelog]
Signed-off-by: Greg Kurz <groug@kaod.org>
The v9fs_link() function keeps a reference on the source fid object. This
causes a memory leak since the reference never goes down to 0. This patch
fixes the issue.
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
[groug, rephrased the changelog]
Signed-off-by: Greg Kurz <groug@kaod.org>
The 'fs.xattr.value' field in V9fsFidState object doesn't consider the
situation that this field has been allocated previously. Every time, it
will be allocated directly. This leads to a host memory leak issue if
the client sends another Txattrcreate message with the same fid number
before the fid from the previous time got clunked.
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
[groug, updated the changelog to indicate how the leak can occur]
Signed-off-by: Greg Kurz <groug@kaod.org>
9pfs uses g_malloc() to allocate the xattr memory space, if the guest
reads this memory before writing to it, this will leak host heap memory
to the guest. This patch avoid this.
Signed-off-by: Li Qiang <liqiang6-s@360.cn>
Reviewed-by: Greg Kurz <groug@kaod.org>
Signed-off-by: Greg Kurz <groug@kaod.org>
Virtio devices should implement the VirtIODevice->reset() function to
perform necessary cleanup actions and to bring the device to a quiescent
state.
In the case of the virtio-9p device, this means:
- emptying the list of active PDUs (i.e. draining all in-flight I/O)
- freeing all fids (i.e. close open file descriptors and free memory)
That's what this patch does.
The reset handler first waits for all active PDUs to complete. Since
completion happens in the QEMU global aio context, we just have to
loop around aio_poll() until the active list is empty.
The freeing part involves some actions to be performed on the backend,
like closing file descriptors or flushing extended attributes to the
underlying filesystem. The virtfs_reset() function already does the
job: it calls free_fid() for all open fids not involved in an ongoing
I/O operation. We are sure this is the case since we have drained
the PDU active list.
The current code implements all backend accesses with coroutines, but we
want to stay synchronous on the reset path. We can either change the
current code to be able to run when not in coroutine context, or create
a coroutine context and wait for virtfs_reset() to complete. This patch
goes for the latter because it results in simpler code.
Note that we also need to create a dummy PDU because it is also an API
to pass the FsContext pointer to all backend callbacks.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
If a PDU has a flush request pending, the current code calls pdu_free()
twice:
1) pdu_complete()->pdu_free() with pdu->cancelled set, which does nothing
2) v9fs_flush()->pdu_free() with pdu->cancelled cleared, which moves the
PDU back to the free list.
This works but it complexifies the logic of pdu_free().
With this patch, pdu_complete() only calls pdu_free() if no flush request
is pending, i.e. qemu_co_queue_next() returns false.
Since pdu_free() is now supposed to be called with pdu->cancelled cleared,
the check in pdu_free() is dropped and replaced by an assertion.
Signed-off-by: Greg Kurz <groug@kaod.org>
