Locking was introduced in QEMU 2.7 to address the deprecation of
readdir_r(3) in glibc 2.24. It turns out that the frontend code is
the worst place to handle a critical section with a pthread mutex:
the code runs in a coroutine on behalf of the QEMU mainloop and then
yields control, waiting for the fsdev backend to process the request
in a worker thread. If the client resends another readdir request for
the same fid before the previous one finally unlocked the mutex, we're
deadlocked.
This never bit us because the linux client serializes readdir requests
for the same fid, but it is quite easy to demonstrate with a custom
client.
A good solution could be to narrow the critical section in the worker
thread code and to return a copy of the dirent to the frontend, but
this causes quite some changes in both 9p.c and codir.c. So, instead
of that, in order for people to easily backport the fix to older QEMU
versions, let's simply use a CoMutex since all the users for this
sit in coroutines.
Fixes: 7cde47d4a8 ("9p: add locking to V9fsDir")
Reviewed-by: Christian Schoenebeck <qemu_oss@crudebyte.com>
Message-Id: <158981894794.109297.3530035833368944254.stgit@bahia.lan>
Signed-off-by: Greg Kurz <groug@kaod.org>
Devices may have component devices and buses.
Device realization may fail. Realization is recursive: a device's
realize() method realizes its components, and device_set_realized()
realizes its buses (which should in turn realize the devices on that
bus, except bus_set_realized() doesn't implement that, yet).
When realization of a component or bus fails, we need to roll back:
unrealize everything we realized so far. If any of these unrealizes
failed, the device would be left in an inconsistent state. Must not
happen.
device_set_realized() lets it happen: it ignores errors in the roll
back code starting at label child_realize_fail.
Since realization is recursive, unrealization must be recursive, too.
But how could a partly failed unrealize be rolled back? We'd have to
re-realize, which can fail. This design is fundamentally broken.
device_set_realized() does not roll back at all. Instead, it keeps
unrealizing, ignoring further errors.
It can screw up even for a device with no buses: if the lone
dc->unrealize() fails, it still unregisters vmstate, and calls
listeners' unrealize() callback.
bus_set_realized() does not roll back either. Instead, it stops
unrealizing.
Fortunately, no unrealize method can fail, as we'll see below.
To fix the design error, drop parameter @errp from all the unrealize
methods.
Any unrealize method that uses @errp now needs an update. This leads
us to unrealize() methods that can fail. Merely passing it to another
unrealize method cannot cause failure, though. Here are the ones that
do other things with @errp:
* virtio_serial_device_unrealize()
Fails when qbus_set_hotplug_handler() fails, but still does all the
other work. On failure, the device would stay realized with its
resources completely gone. Oops. Can't happen, because
qbus_set_hotplug_handler() can't actually fail here. Pass
&error_abort to qbus_set_hotplug_handler() instead.
* hw/ppc/spapr_drc.c's unrealize()
Fails when object_property_del() fails, but all the other work is
already done. On failure, the device would stay realized with its
vmstate registration gone. Oops. Can't happen, because
object_property_del() can't actually fail here. Pass &error_abort
to object_property_del() instead.
* spapr_phb_unrealize()
Fails and bails out when remove_drcs() fails, but other work is
already done. On failure, the device would stay realized with some
of its resources gone. Oops. remove_drcs() fails only when
chassis_from_bus()'s object_property_get_uint() fails, and it can't
here. Pass &error_abort to remove_drcs() instead.
Therefore, no unrealize method can fail before this patch.
device_set_realized()'s recursive unrealization via bus uses
object_property_set_bool(). Can't drop @errp there, so pass
&error_abort.
We similarly unrealize with object_property_set_bool() elsewhere,
always ignoring errors. Pass &error_abort instead.
Several unrealize methods no longer handle errors from other unrealize
methods: virtio_9p_device_unrealize(),
virtio_input_device_unrealize(), scsi_qdev_unrealize(), ...
Much of the deleted error handling looks wrong anyway.
One unrealize methods no longer ignore such errors:
usb_ehci_pci_exit().
Several realize methods no longer ignore errors when rolling back:
v9fs_device_realize_common(), pci_qdev_unrealize(),
spapr_phb_realize(), usb_qdev_realize(), vfio_ccw_realize(),
virtio_device_realize().
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Reviewed-by: Paolo Bonzini <pbonzini@redhat.com>
Message-Id: <20200505152926.18877-17-armbru@redhat.com>
A client establishes a session by sending a Tversion request along with a
'msize' parameter which client uses to suggest server a maximum message
size ever to be used for communication (for both requests and replies)
between client and server during that session. If client suggests a 'msize'
smaller than 4096 then deny session by server immediately with an error
response (Rlerror for "9P2000.L" clients or Rerror for "9P2000.u" clients)
instead of replying with Rversion.
So far any msize submitted by client with Tversion was simply accepted by
server without any check. Introduction of some minimum msize makes sense,
because e.g. a msize < 7 would not allow any subsequent 9p operation at
all, because 7 is the size of the header section common by all 9p message
types.
A substantial higher value of 4096 was chosen though to prevent potential
issues with some message types. E.g. Rreadlink may yield up to a size of
PATH_MAX which is usually 4096, and like almost all 9p message types,
Rreadlink is not allowed to be truncated by the 9p protocol. This chosen
size also prevents a similar issue with Rreaddir responses (provided client
always sends adequate 'count' parameter with Treaddir), because even though
directory entries retrieval may be split up over several T/Rreaddir
messages; a Rreaddir response must not truncate individual directory entries
though. So msize should be large enough to return at least one directory
entry with the longest possible file name supported by host. Most file
systems support a max. file name length of 255. Largest known file name
lenght limit would be currently ReiserFS with max. 4032 bytes, which is
also covered by this min. msize value because 4032 + 35 < 4096.
Furthermore 4096 is already the minimum msize of the Linux kernel's 9pfs
client.
Signed-off-by: Christian Schoenebeck <qemu_oss@crudebyte.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Message-Id: <8ceecb7fb9fdbeabbe55c04339349a36929fb8e3.1579567019.git.qemu_oss@crudebyte.com>
Signed-off-by: Greg Kurz <groug@kaod.org>
init_in_iov_from_pdu might not be able to allocate the full buffer size
requested, which comes from the client and could be larger than the
transport has available at the time of the request. Specifically, this
can happen with read operations, with the client requesting a read up to
the max allowed, which might be more than the transport has available at
the time.
Today the implementation of init_in_iov_from_pdu throws an error, both
Xen and Virtio.
Instead, change the V9fsTransport interface so that the size becomes a
pointer and can be limited by the implementation of
init_in_iov_from_pdu.
Change both the Xen and Virtio implementations to set the size to the
size of the buffer they managed to allocate, instead of throwing an
error. However, if the allocated buffer size is less than P9_IOHDRSZ
(the size of the header) still throw an error as the case is unhandable.
Signed-off-by: Stefano Stabellini <stefano.stabellini@xilinx.com>
CC: groug@kaod.org
CC: anthony.perard@citrix.com
CC: roman@zededa.com
CC: qemu_oss@crudebyte.com
[groug: fix 32-bit build]
Signed-off-by: Greg Kurz <groug@kaod.org>
Use variable length suffixes for inode remapping instead of the fixed
16 bit size prefixes before. With this change the inode numbers on guest
will typically be much smaller (e.g. around >2^1 .. >2^7 instead of >2^48
with the previous fixed size inode remapping.
Additionally this solution is more efficient, since inode numbers in
practice can take almost their entire 64 bit range on guest as well, so
there is less likely a need for generating and tracking additional suffixes,
which might also be beneficial for nested virtualization where each level of
virtualization would shift up the inode bits and increase the chance of
expensive remapping actions.
The "Exponential Golomb" algorithm is used as basis for generating the
variable length suffixes. The algorithm has a parameter k which controls the
distribution of bits on increasing indeces (minimum bits at low index vs.
maximum bits at high index). With k=0 the generated suffixes look like:
Index Dec/Bin -> Generated Suffix Bin
1 [1] -> [1] (1 bits)
2 [10] -> [010] (3 bits)
3 [11] -> [110] (3 bits)
4 [100] -> [00100] (5 bits)
5 [101] -> [10100] (5 bits)
6 [110] -> [01100] (5 bits)
7 [111] -> [11100] (5 bits)
8 [1000] -> [0001000] (7 bits)
9 [1001] -> [1001000] (7 bits)
10 [1010] -> [0101000] (7 bits)
11 [1011] -> [1101000] (7 bits)
12 [1100] -> [0011000] (7 bits)
...
65533 [1111111111111101] -> [1011111111111111000000000000000] (31 bits)
65534 [1111111111111110] -> [0111111111111111000000000000000] (31 bits)
65535 [1111111111111111] -> [1111111111111111000000000000000] (31 bits)
Hence minBits=1 maxBits=31
And with k=5 they would look like:
Index Dec/Bin -> Generated Suffix Bin
1 [1] -> [000001] (6 bits)
2 [10] -> [100001] (6 bits)
3 [11] -> [010001] (6 bits)
4 [100] -> [110001] (6 bits)
5 [101] -> [001001] (6 bits)
6 [110] -> [101001] (6 bits)
7 [111] -> [011001] (6 bits)
8 [1000] -> [111001] (6 bits)
9 [1001] -> [000101] (6 bits)
10 [1010] -> [100101] (6 bits)
11 [1011] -> [010101] (6 bits)
12 [1100] -> [110101] (6 bits)
...
65533 [1111111111111101] -> [0011100000000000100000000000] (28 bits)
65534 [1111111111111110] -> [1011100000000000100000000000] (28 bits)
65535 [1111111111111111] -> [0111100000000000100000000000] (28 bits)
Hence minBits=6 maxBits=28
Signed-off-by: Christian Schoenebeck <qemu_oss@crudebyte.com>
Signed-off-by: Greg Kurz <groug@kaod.org>
stat_to_qid attempts via qid_path_prefixmap to map unique files (which are
identified by 64 bit inode nr and 32 bit device id) to a 64 QID path value.
However this implementation makes some assumptions about inode number
generation on the host.
If qid_path_prefixmap fails, we still have 48 bits available in the QID
path to fall back to a less memory efficient full mapping.
Signed-off-by: Antonios Motakis <antonios.motakis@huawei.com>
[CS: - Rebased to https://github.com/gkurz/qemu/commits/9p-next
(SHA1 7fc4c49e91).
- Updated hash calls to new xxhash API.
- Removed unnecessary parantheses in qpf_lookup_func().
- Removed unnecessary g_malloc0() result checks.
- Log error message when running out of prefixes in
qid_path_fullmap().
- Log warning message about potential degraded performance in
qid_path_prefixmap().
- Wrapped qpf_table initialization to dedicated qpf_table_init()
function.
- Fixed typo in comment. ]
Signed-off-by: Christian Schoenebeck <qemu_oss@crudebyte.com>
Signed-off-by: Greg Kurz <groug@kaod.org>
'warn' (default): Only log an error message (once) on host if more than one
device is shared by same export, except of that just ignore this config
error though. This is the default behaviour for not breaking existing
installations implying that they really know what they are doing.
'forbid': Like 'warn', but except of just logging an error this
also denies access of guest to additional devices.
'remap': Allows to share more than one device per export by remapping
inodes from host to guest appropriately. To support multiple devices on the
9p share, and avoid qid path collisions we take the device id as input to
generate a unique QID path. The lowest 48 bits of the path will be set
equal to the file inode, and the top bits will be uniquely assigned based
on the top 16 bits of the inode and the device id.
Signed-off-by: Antonios Motakis <antonios.motakis@huawei.com>
[CS: - Rebased to https://github.com/gkurz/qemu/commits/9p-next
(SHA1 7fc4c49e91).
- Added virtfs option 'multidevs', original patch simply did the inode
remapping without being asked.
- Updated hash calls to new xxhash API.
- Updated docs for new option 'multidevs'.
- Fixed v9fs_do_readdir() not having remapped inodes.
- Log error message when running out of prefixes in
qid_path_prefixmap().
- Fixed definition of QPATH_INO_MASK.
- Wrapped qpp_table initialization to dedicated qpp_table_init()
function.
- Dropped unnecessary parantheses in qpp_lookup_func().
- Dropped unnecessary g_malloc0() result checks. ]
Signed-off-by: Christian Schoenebeck <qemu_oss@crudebyte.com>
[groug: - Moved "multidevs" parsing to the local backend.
- Added hint to invalid multidevs option error.
- Turn "remap" into "x-remap". ]
Signed-off-by: Greg Kurz <groug@kaod.org>
The QID path should uniquely identify a file. However, the
inode of a file is currently used as the QID path, which
on its own only uniquely identifies files within a device.
Here we track the device hosting the 9pfs share, in order
to prevent security issues with QID path collisions from
other devices.
We only print a warning for now but a subsequent patch will
allow users to have finer control over the desired behaviour.
Failing the I/O will be one the proposed behaviour, so we
also change stat_to_qid() to return an error here in order to
keep other patches simpler.
Signed-off-by: Antonios Motakis <antonios.motakis@huawei.com>
[CS: - Assign dev_id to export root's device already in
v9fs_device_realize_common(), not postponed in
stat_to_qid().
- error_report_once() if more than one device was
shared by export.
- Return -ENODEV instead of -ENOSYS in stat_to_qid().
- Fixed typo in log comment. ]
Signed-off-by: Christian Schoenebeck <qemu_oss@crudebyte.com>
[groug, changed to warning, updated message and changelog]
Signed-off-by: Greg Kurz <groug@kaod.org>
As with unlinkat, these flags come from the client and need to
be translated to their host values. The protocol values happen
to match linux, but that need not be true in general.
Signed-off-by: Keno Fischer <keno@juliacomputing.com>
Signed-off-by: Greg Kurz <groug@kaod.org>
lhs/rhs doesn't tell much about how argument are handled, dst/src is
and const arguments is clearer in my mind. Use g_memdup() while at it.
Signed-off-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Signed-off-by: Greg Kurz <groug@kaod.org>
No good reasons to do this outside of v9fs_device_realize_common().
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
The 9P protocol is transport agnostic: if the guest misconfigured the
buffers, the best we can do is to set the broken flag on the device.
Signed-off-by: Greg Kurz <groug@kaod.org>
The 9p spec at http://man.cat-v.org/plan_9/5/intro reads:
"Each 9P message begins with a four-byte size field specify-
ing the length in bytes of the complete message including
the four bytes of the size field itself. The next byte is
the message type, one of the constants in the enumeration in
the include file <fcall.h>. The next two bytes are an iden-
tifying tag, described below."
ie, each message starts with a 7-byte long header.
The core 9P code already assumes this pretty much everywhere. This patch
does the following:
- makes the assumption explicit in the common 9p.h header, since it isn't
related to the transport
- open codes the header size in handle_9p_output() and hardens the sanity
check on the space needed for the reply message
Signed-off-by: Greg Kurz <groug@kaod.org>
Acked-by: Stefano Stabellini <sstabellini@kernel.org>
These bits aren't related to the transport so let's move them to the core
code.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Stefano Stabellini <sstabellini@kernel.org>
Use the new type in virtio-9p-device.
Signed-off-by: Stefano Stabellini <stefano@aporeto.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
CC: anthony.perard@citrix.com
CC: jgross@suse.com
CC: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
CC: Greg Kurz <groug@kaod.org>
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>
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>
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>
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>
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>
All these functions either call the v9fs_co_* functions which have the
coroutine_fn annotation, or pdu_complete() which calls qemu_co_queue_next().
Let's mark them to make it obvious they execute in coroutine context.
Signed-off-by: Greg Kurz <groug@kaod.org>
This helper is similar to v9fs_string_sprintf(), but it includes the
terminating NUL character in the size field.
This is to avoid doing v9fs_string_sprintf((V9fsString *) &path) and
then bumping the size.
Affected users are changed to use this new helper.
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
The 9P spec at http://man.cat-v.org/plan_9/5/intro says:
All directories must support walks to the directory .. (dot-dot) meaning
parent directory, although by convention directories contain no explicit
entry for .. or . (dot). The parent of the root directory of a server's
tree is itself.
This means that a client cannot walk further than the root directory
exported by the server. In other words, if the client wants to walk
"/.." or "/foo/../..", the server should answer like the request was
to walk "/".
This patch just does that:
- we cache the QID of the root directory at attach time
- during the walk we compare the QID of each path component with the root
QID to detect if we're in a "/.." situation
- if so, we skip the current component and go to the next one
Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Remove glib.h includes, as it is provided by osdep.h.
This commit was created with scripts/clean-includes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
Tested-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Michael Tokarev <mjt@tls.msk.ru>
If several threads concurrently call readdir() with the same directory
stream pointer, it is possible that they all get a pointer to the same
dirent structure, whose content is overwritten each time readdir() is
called.
We must thus serialize accesses to the dirent structure.
This may be achieved with a mutex like below:
lock_mutex();
readdir();
// work with the dirent
unlock_mutex();
This patch adds all the locking, to prepare the switch to readdir().
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
If we are to switch back to readdir(), we need a more complex type than
DIR * to be able to serialize concurrent accesses to the directory stream.
This patch introduces a placeholder type and fixes all users.
Reviewed-by: Eric Blake <eblake@redhat.com>
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Most of the 9p code is now virtio agnostic. This patch does a final cleanup:
- drop references to Virtio from the header comments
- fix includes
Also drop a couple of leading empty lines while here.
Signed-off-by: Greg Kurz <gkurz@linux.vnet.ibm.com>
Clean up includes so that osdep.h is included first and headers
which it implies are not included manually.
This commit was created with scripts/clean-includes.
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
V9fsState now only contains generic fields. Introduce V9fsVirtioState
for virtio transport. Change virtio-pci and virtio-ccw to use
V9fsVirtioState.
Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
Move out generic definitions from virtio-9p.h to 9p.h. Fix header
inclusions.
Signed-off-by: Wei Liu <wei.liu2@citrix.com>
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
