| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
| |
| |
| |
| |
| | |
Now correctly deallocates the local fd if the remote already
deallocated the flow.
|
| |
| |
| |
| |
| | |
A condition variable is now used to wait for an auto-executed AP to
accept the requested flow.
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| | |
The registration function has been moved to the irm tool, applications
now need to be registered by an administrator. Currently only supports
one instance per registered name, and an AP can be registered under
only one name.
The irmd can now start a registered server application on demand.
For the full functionality of the tool, execute "irm register".
AP name removed from flow allocation. Flow allocation does not send
the source ap name as it is quite useless. The accept() call now only
returns the AE name.
|
| |
| |
| |
| |
| | |
This addresses some comments Dimitri had on the shim Ethernet with
LLC.
|
|/
|
|
|
| |
This adds a shim over LLC over Ethernet. It uses the raw socket API to
send messages directly over an interface.
|
|
|
|
|
| |
This adds wildcarding of DIF names so that application developers can
for instance specify home.* to specify all home DIFs.
|
| |
|
| |
|
|
|
|
|
| |
When the ring buffer is empty, a read call will sleep. A write call on
an empty ring buffer will wake up sleeping readers.
|
|
|
|
|
|
|
| |
This stops IPC processes from hanging due to bad locks. It first
informs all threads that the IPCP is shutting down, then cancels all
threads. Cancellation is still required because threads may be in
accept() or receivefrom().
|
| |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This introduces an IPC process that relays data between two local AP's
over the shm_du_map. Only configuration it has is a DIF name.
It required small modification elsewhere:
lib: added support for the IPCP_LOCAL type
irm: added support for the IPCP_LOCAL type
dif_config: added the IPCP_LOCAL type
tools: added support for the IPCP_LOCAL type
|
|
|
|
|
| |
The shim UDP did a non-blocking write call, causing loss of
SDU's. This has been changed to a blocking write call.
|
|
|
|
|
|
|
| |
The cbr client will now use busy waiting by default to control the
sending rate. A --sleep option has been added to allow low CPU usage
when sending at low data rates. A --flood option has been added that
writes SDU's as fast as possible.
|
|\
| |
| |
| | |
ipcpd: fixed the shim-udp grinding the cpu
|
| |
| |
| |
| |
| |
| |
| |
| | |
In linux, the select call changes the timeout variable to the time not
spent in sleep, causing the timeout to decrease and that loop to grind
the cpu. The timeval is now reset to the correct value every loop.
Also removed a LOG_ERR message that was hampering performance.
|
|/ |
|
|
|
|
|
| |
non-blocking write would return when the buffer was full
functions in dev now check validity of input file descripters
|
|\
| |
| |
| | |
Be cbr
|
| |
| |
| |
| |
| | |
This adds a cleanup handler for the main message loop, which closes
the socket upon pthread_cancel so that the ipcp can exit cleanly.
|
| |
| |
| |
| |
| | |
This adds a threadpool to cbr, so that it is not overflooded with
client requests.
|
|\ \
| |/
|/|
| | |
lib: allow running shm_du_map in single block mode
|
| |
| |
| |
| |
| | |
this allows disabling support for SDU's that are larger than a block
in the DU MAP by defining the variable SHM_MAP_SINGLE_BLOCK
|
| |
| |
| |
| |
| | |
the CBR tool would stop if the buffer is full. Now it will continue
sending SDU's until the duration expires.
|
| | |
|
| |
| |
| |
| |
| |
| |
| |
| | |
The tool sends constant bandwidth traffic between a client and a
server. cbr --help for more info.
Adds time_utils.h to the library containing useful functions for
arithmetic with timespec and timeval structures.
|
|/
|
|
| |
Fixes some memleaks in dev.c in the case of error conditions.
|
|
|
|
|
| |
This makes sure only one shm_du_map is running in the system.
Starting a second irmd would invalidate the shared memory map.
|
|\
| |
| |
| | |
irmd: Change read lock to write lock
|
| |
| |
| |
| |
| | |
A read lock was taken instead of a write lock, which resulted in
double frees.
|
|/
|
|
| |
Conf was not checked on bootstrap.
|
|
|
|
|
|
| |
The fast path has been rewritten to have certainty to read the correct
flow. Deallocation will not release port_id's or fd's until they are
explicitly released locally.
|
|
|
|
|
| |
This removes an install command from the build, since the protobuf
header file was being installed in the source directory.
|
| |
|
| |
|
|
|
|
|
|
| |
The shim UDP now supports deallocating a flow end-to-end. Contains
some stability fixes for flow allocation and some missing close()
calls in lib/sockets.
|
| |
|
|
|
|
|
|
| |
dev.c: read now only reads an SDU if is is for the correct port_id
shm_ap_rbuff: added a function peek() that returns the port_id of the
tail.
|
|
|
|
|
| |
There could be a theoretical race condition that the server sends a
message before the FD is set on the client side, which is resolved.
|
|
|
|
|
|
|
|
|
|
|
| |
There were two race conditions. The first was where client received
the response message before the client UDP port was connected to the
server UDP port and the first message of the client was sent to the
LISTEN port instead of the server application port. This was solved by
reordering the operations so the response is sent after the port is
correctly connected. The second race condition is that the FD is not
added to the set in time at the server side before the first SDU
arrives. This was solved by adding a variable proteceted by a mutex.
|
|
|
|
|
| |
some debug code for locking and a sleep check are removed from the
shim-udp.
|
|
|
|
|
|
|
|
|
| |
It uses UDP port 0x0D1F on all hosts to send and receive flow
allocation messages. It supports communication between server and
client AP over a single shim IPCP.
Implementation of full flow deallocation is pending. Both the client
and the server still have to call flow_dealloc();
|
|
|
|
|
|
|
| |
There was a check missing upon garbage collection. If the whole DU map
has been used once, all DUs will have their garbage bit set to
one. This meant that the garbage collector would keep moving the tail
indefinitely.
|
|
|
|
|
| |
When moving the tail pointer in the DU map, blocks was never reset to
zero. This caused it to sometimes clean up too many DUs.
|
| |
|
|
|
|
|
|
| |
Adds a missing modulo calculation that would result in index -1 being
returned instead of zero when the ringbuffer skipped back to the
beginning.
|
|\
| |
| |
| | |
lib: dev: small locking improvements
|
| |
| |
| |
| |
| |
| | |
This avoids holding the long locks during accept() and alloc_res()
calls. Small reduction of time of other locks held. No changes in
locking logic.
|
| |\ |
|