This patch replaces formerly blocking packet-stream operations by
the explicit use of 'wait_and_dispatch_io_signal' for blocking.
It also removes a misleading comment that promised a fire-and-forget
behavior whereas the implementation relied on blocking I/O anyway.
Issue #4390
Errors during IPC receive-and-wait can occur at the server side when
a client is killed. This condition is not an error from the server's
perspective. We used to print a message nevertheless, since the
condition is rather exceptional. However, when printed during the
test-sequence test, the messages interfere with the pattern matching of
the depot_autopilot, flagging the successful test as an error.
The VFS block plugin used to depend on the blocking semantics of the
packet stream's 'get_acked_packet'. This patch replaces this dependency
by the use of 'wait_and_dispatch_one_io_signal'. However, in order to
implement this change, the custom instance of a 'Signal_receiver' had to
be removed as well.
To keep this patch as little invasive as possible, it does not touch the
direct use of the block session's packet stream, which should better be
replaced by the 'Block::Connection::Job' API.
Issue #4390
The zynq_nic_drv follows a zero-copy approach and thus uses the packet
buffers as DMA memory. In order to know when the RX DMA memory can be used
for another packet, a custom ack_avail_handler is needed.
Similarly, packets received from the Uplink session are not copied to a
DMA buffer but to directly passed on as DMA memory. For this purpose,
a a custom packet_avail handler is needed.
genodelabs/genode#4384
The 'read' and 'write' utilities are from a time before the VFS API
as os/vfs.h was available. They rely on the (now removed) blocking
semantics of the packet-stream interface.
The only remaining legitimate use case of the direct interaction with
the file-system session without VFS is the back end of gcov, which needs
a way to exfiltrate the statistical data using a channel that is
independent from the libc or the VFS.
Issue #4390
This patch replaces the direct use of a file-system session via the
'file_system/util.h' helpers by the VFS using the os/vfs.h API. This
makes the component more flexible while removing the dependence from
read and write utilities of file_system/util.h, which happen to rely on
the (now removed) blocking packet-stream semantics.
Issue #4390
This patch replaces the direct interaction with the packet stream of
the block session by the use of the 'Block::Connection::Job' API,
removing the reliance on blocking packet-stream semantics.
Since I/O signals can now occur during 'Backend::submit', the patch
conditions the periodic calls of 'rump_sys_sync' by taking the backend
state into account.
Issue #4390
Since the change "block_tester: limit batching in sequential test", the
sequence test blocks infinitely when encountering a length value smaller
than the block size.
This patch takes precautions against the use of blocking packet-stream
operations like 'submit_packet'.
With the change of issue #4388, the ready-to-submit signals are no
longer implicitly handled. Hence, a call of submit_packet to a
saturated submit queue blocks infinitely.
Issue #4390
Even though the use of the C++ exception mechanism (and the implicit use
of the cxx heap) is not a problem at the server side, this patch
nevertheless replaces the exception-based return-value handling to make
the code consistent with the ipc_call path.
Issue #3612
* use 'Id_space::apply_any' instead of 'for_each' to destroy session local
buffer objects
* save session capability in session because 'cap()' is not valid in
'Session_component' destructor because it was dissolved before calling
'Root::_destruct_session', the cap is necessary to remove owned
buffers from the EP
issue #4380
Do not import a freshly allocated buffer to all contexts eagerly.
Instead check buffer list in context's 'exec_buffer' call and import
only buffers needed and not present before GPU execution. This leads to
improved performance for applications that use many OpenGL contexts
(e.g., VirtualBox 6).
issue #4380
* retrieve Genode::Env from plugin, this way no mesa applications need to
be changed.
* add 'vfs_gpu' api
* remove when all required functionality is implemented within the plugin.
issue #4380
Implement GLX and X functionality through Mesa's EGL interface. This
requires multiple OpenGL contexts and in turn GEM context support in
libdrm/iris, as implemented in libdrm and intel_gpu_drv.
Update recipes and machine.vbox6 files accordingly,
issue #4380
Retrieve multiple GPU sessions from VFS plugin, take advantage of buffer
import/export functionallity in order to implement gem context support.
Multiple contexts share all GPU buffers, but use different GPU sessions
and thus, differnt page tables and hardware contexts.
issue #4380
Add 'export_buffer'/'import_buffer' calls in order to support buffer
sharing between GPU sessions.
Reduce CAP costs for Gpu::Connection from 256 to 32.
issue #4380
Synchronize GPU completion by calling 'read' of the vfs_gpu plugin. This
enables pthreads to wait for GPU completions instead of the main EP.
issue #4380
When a <gpu> node is present in the VFS, the plugin opens a
Gpu::Connection for each open call to the 'gpu' node and installs a
completion signal handler. A read only to the fd returned only comes
back if a completion signal has been received between two read
transctions to the fd.
For now the Gpu::Connections can be retrieved by calling the
'vfs_gpu_connection(unsigned long id)' function. The id can be obtained
using 'stat' on the 'gpu' device and is located in the inode (st_ino)
field of the stat buffer.
issue #4380
* let iris handle buffer management, this implies that BOs are mapped to
the PPGTT during buffer execution and unmapped by iris later, for this
to work buffers need to be unmapped when allocating cached BOs
(vma_free) which requires a patch
* support lseek (drm_lseek for now) for determining object size
issue #4380
Account RAM and CAP resources for GPU sessions and trigger client
upgrades before allocating resources at the multiplexer. This prevents
the multiplexer from running out of resources.
issue #4380
Imagine receiving the signal for an available TX ack or an available RX packet
at the Uplink connection but a later received signal for a link-state change
(to link state "down") at the same connection is handled first and destructs
the Uplink connection before the handling of the former signals. In this case,
the methods 'Uplink_client_base::_conn_tx_handle_ack_avail' and
'Uplink_client_base::_conn_rx_handle_packet_avail' must be guarded against an
unconstructed '_conn' member, but they weren't so far.
Fixes#4384
So far, the generic Uplink connection code considered NIC drivers to transmit
connection RX packets in three different manners. Most of the drivers follow
the "normal" way of transmission with only one driver callback from generic
code. The monolithic USB NIC-driver, however, used to send in a "burst" mode
that required a dedicated path with multiple driver callbacks in the generic
code. And then there were drivers that had a fully custom function for doing
transmissions.
Also for handling connection TX acks, there was a "normal" and a
"custom handler" way.
Today, all NIC drivers in the Genode repos and the Genode-World repo follow the
"normal" way. Therefor, the unused code can be removed.
Ref #4384
