This patch introduces new types for expressing CPU affinities. Instead
of dealing with physical CPU numbers, affinities are expressed as
rectangles in a grid of virtual CPU nodes. This clears the way to
conveniently assign sets of adjacent CPUs to subsystems, each of them
managing their respective viewport of the coordinate space.
By using 2D Cartesian coordinates, the locality of CPU nodes can be
modeled for different topologies such as SMP (simple Nx1 grid), grids of
NUMA nodes, or ring topologies.
The new core-internal 'Address_space' interface enables cores RM service
to flush mappings of a PD in which a given 'Rm_client' thread resides.
Prior this patch, each platform invented their own way to flush mappings
in the respective 'rm_session_support.cc' implementation. However, those
implementations used to deal poorly with some corner cases. In
particular, if a PD session was destroyed prior a RM session, the RM
session would try to use no longer existing PD session. The new
'Address_space' uses the just added weak-pointer mechanism to deal with
this issue.
Furthermore, the generic 'Rm_session_component::detach' function has
been improved to avoid duplicated unmap operations for platforms that
implement the 'Address_space' interface. Therefore, it is related to
issue #595. Right now, this is OKL4 only, but other platforms will follow.
If page faults are handled concurrently (as for base-nova) the traverse lookup
call in rm_session_component must be thread safe, which it isn't.
If the faulting area is backed by nested dataspaces which are managed by
various rm_sessions then a race happens under following circumstances
(triggered occasionally by the bomb test).
The traverse lookup may return a pointer to a rm_session of a nested dataspace.
If the rm_session is in parallel subject to destruction it happened that faults
got enqueued to the faulters list of the deleted rm_session and internally to
a list of the current rm_session of the Rm_client.
During destruction of the faulting Rm_client the associated rm_session will
be dissolved from the Rm_client, which leads to dereferencing the
dangling pointer of the already destructed rm_session.
On base-nova the memory of the rm_session object get unmapped eventually, so
that the de-referencing of the dangling pointer caused page faults in core.
The memory on other kernels inside core never get unmapped so that the
bug doesn't trigger visible faults.
The patch replace the keeping of a rm_session pointer by keeping a
capability instead. The rm_session object must be looked up now explicitly in
the Object_pool implementation, which implements proper reference counting on
the rm_session object.
Issue #549
First make the clients inaccessible and dissolve them from the entrypoint. If
this isn't the first step the clients may be obtained again between
the unlock and lock steps in the destructor.
Additionally the clients may be removed in between the unlock and call
sequence, which renders such client pointers dangling and causes spurious page
faults. Keep instead a lock as long as possible and when it is required to
release a lock, then the pointer to the objects must be revalidated.
Replace the dissolve function with a remove_client implementation as suggested
by #13, which avoids that the cpu_session may call dissolve with a dangling
pointer of a already removed rm_client object. Instead the pager must be
released explicitly.
Related to issue #549
Related to issue #394
Related to issue #13
Rm_client is derived from Pager_object. If the Pager_object is also
derived from Thread_base (which is the case for NOVA) then the
Rm_client object must be destructed without holding the rm_session_object
lock. The native platform specific Thread_base implementation has to take
care that all in-flight page handling requests are finished before
destruction. On NOVA it is done by doing an IPC to the pager thread.
(performed in Pager_object::dissolve() in base-nova). The
called thread than executes its operation until end which also requires
in some cases to take the rm_session_object lock.
Since _client_slab insertion/deletion also must be performed
synchronized but can't be protected by the rm_session_object lock
because of the described dead_lock situation, we have
to use a synchronized allocator object to perform insertion and
deletion of Rm_clients.
This patch extends the RAM session interface with the ability to
allocate DMA buffers. The client specifies the type of RAM dataspace to
allocate via the new 'cached' argument of the 'Ram_session::alloc()'
function. By default, 'cached' is true, which correponds to the common
case and the original behavior. When setting 'cached' to 'false', core
takes the precautions needed to register the memory as uncached in the
page table of each process that has the dataspace attached.
Currently, the support for allocating DMA buffers is implemented for
Fiasco.OC only. On x86 platforms, it is generally not needed. But on
platforms with more relaxed cache coherence (such as ARM), user-level
device drivers should always use uncacheable memory for DMA transactions.
With this patch clients of the RM service can state if they want a mapping
to be executable or not. This allows dataspaces to be mapped as
non-executable on Linux by default and as executable only if needed.
Partially fixes#176.
Norman Feske
Alexander Boettcher
Martin Stein
Christian Prochaska
Genode Labs