The distinction between 'ipc.h' and 'ipc_generic.h' is no more. The only
use case for platform-specific extensions of the IPC support was the
marshalling of capabilities. However, this case is accommodated by a
function interface ('_marshal_capability', '_unmarshal_capability'). By
moving the implementation of these functions from the headers into the
respective ipc libraries, we can abandon the platform-specific 'ipc.h'
headers.
This patch improves the life-time management of socket descriptors and
addresses several corner cases exposed by the 'bomb' test.
The lookup and association of file descriptors with global IDs have been
turned into an atomic operation. Otherwise, multiple threads interacting
with the singleton 'ep_sd_registry' may override each other's
associations.
Closing the socket pair used for the reply channel has been implemented
via the RAII pattern to capture all corner cases, in particular
exceptions.
If blocking operations are interrupted by signals, we throw a
'Blocking_canceled' exception.
We preserve core's socket descriptor at 'PARENT_SOCKET_HANDLE' to avoid
a corner case where the parent capability is going to dup2'ed to the
same handle.
Support for 'Thread_base::join' within core to enable leaving Genode via
Control-C.
When an IPC server is finalized two important things should happen:
First, the association of the server socket with a capability must be
invalidated. And finally, the server socket pair (server side and client
side) must be closed.
Related to #38.
When building in hybrid Linux/Genode mode, there exist two definitions
of 'size_t', one in the 'Genode' namespace and one imported from the
glibc headers.
This patch simplifies the system call bindings. The common syscall
bindings in 'src/platform/' have been reduced to the syscalls needed by
non-core programs. The additional syscalls that are needed solely by
core have been moved to 'src/core/include/core_linux_syscalls.h'.
Furthermore, the resource path is not used outside of core anymore.
Hence, we could get rid of the rpath library. The resource-path code has
been moved to 'src/core/include/resource_path.h'. The IPC-related parts
of 'src/platform' have been moved to the IPC library. So there is now a
clean separation between low-level syscall bindings (in 'src/platform')
and higher-level code.
The code for the socket-descriptor registry is now located in the
'src/base/ipc/socket_descriptor_registry.h' header. The interface is
separated from 'ipc.cc' because core needs to access the registry from
outside the ipc library.
This patch alleviates the need for any non-core process to create Unix
domain sockets locally. All sockets used for RPC communication are
created by core and subsequently passed to the other processes via RPC
or the parent interface. The immediate benefit is that no process other
than core needs to access the 'rpath' directory in order to communicate.
However, access to 'rpath' is still needed for accessing dataspaces.
Core creates one socket pair per thread on demand on the first call of
the 'Linux_cpu_session::server_sd()' or 'Linux_cpu_session::client_sd()'
functions. 'Linux_cpu_session' is a Linux-specific extension to the CPU
session interface. In addition to the socket accessors, the extension
provides a mechanism to register the PID/TID of a thread. Those
information were formerly propagated into core along with the thread
name as argument to 'create_thread()'.
Because core creates socket pairs for entrypoints, it needs to know all
threads that are potential entrypoints. For lx_hybrid programs, we
hadn't had propagated any thread information into core, yet. Hence, this
patch also contains the code for registering threads of hybrid
applications at core.
This patch eliminates the thread ID portion of the 'Native_capability'
type. The access to entrypoints is now exclusively handled by passing
socket descripts over Unix domain sockets and by inheriting the socket
descriptor of the parent entrypoint at process-creation time.
Each entrypoint creates a socket pair. The server-side socket is bound
to a unique name defined by the server. The client-side socket is then
connected to the same name. Whereas the server-side socket is meant to
be exclusively used by the server to wait for incoming requests, the
client-side socket can be delegated to other processes as payload of RPC
messages (via SCM rights). Anyone who receives a capability over RPC
receives the client-side socket of the entrypoint to which the
capability refers. Given this socket descriptor, the unique name (as
defined by the server) can be requested using 'getpeername'. Using this
name, it is possible to compare socket descriptors, which is important
to avoid duplicates from polluting the limited socket-descriptor name
space.
Wheras this patch introduces capability-based delegation of access
rights to entrypoints, it does not cover the protection of the integrity
of RPC objects. RPC objects are still referenced by a global ID passed
as normal message payload.
This patch adds prinicipal support for transmitting socket descriptors
as RPC payload. Socket descriptors are handled by the linux-specific
implementation of the capability marshalling and unmarshalling functions
in 'ipc.h'. The 'Message' type in 'src/platform/linux_socket.h' has been
extended to carry multiple descriptors in a single message.
Unfortuately, we hit a problem (and potential show stopper) here:
lx_sendmsg failed with -109 in lx_call()
The error code corresponds to ETOOMANYREFS. There is only one place in
the Linux kernel where this error code is used (net/unix/af_unix.c).
The code for 'unix_attach_fds()' suggests that there is a limit with
regard to the maximum number of references for a given Unix domain
socket. When the error occurs, core and init are running. The socket
of core's server entrypoint is present in the '/proc/pid/fd' of those
processes 8 times. The error occurs when core tries to perform an
RPC to the entrypoint to perform 'Ram_session::transfer_quota()'
(base/include/base/child.h at line 248).
By storing the reply socket descriptor inside the 'Ipc_ostream::_dst'
capability instead as part of the connection state object, we can
use the 'explicit_reply' mechanism as usual. Right now, we store
both the tid and socket handle in 'Native_capability::Dst'. In the
final version, the 'tid' member will be gone.
In the final version, the 'socket' will be the only member to remain in
the 'Dst' time. In the transition phase, we store both the old 'tid' and
the 'socket'.
This patch, which was originally created by Christian Helmuth,
represents the first step towards using SCM rights as capability
mechanism on Linux. It employs the SCM rights mechanism for transmitting
a reply capability to the server as argument of each IPC call. The
server will then send its respond to this reply file descriptor. This
way, the reply channel does not need to be globally visible anymore.
GCC warns about uninitialized local variables in cases where no
initialization is needed, in particular in the overloads of the
'Capability::call()' function. Prior this patch, we dealt with those
warnings by using an (unreliable) GCC pragma or by disabling the
particular warning altogether (which is a bad idea). This patch removes
the superfluous warnings by telling the compiler that the variable in
question is volatile.
The 'copy_to' function turned out to be not flexible enough to
accommodate the Noux fork mechanism. This patch removes the function,
adds an accessor for the capability destination and a compound type
'Native_capability::Raw' to be used wherever plain capability
information must be communicated.
Norman Feske
Christian Helmuth
Stefan Kalkowski
Genode Labs