| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
| |
Fixes #22
|
| | |
|
| |
|
|
| |
Fixes #35
|
| |\
| |
| |
| | |
lib, ipcpd, irmd: Register hash instead of name
|
| | |
| |
| |
| |
| |
| |
| |
| |
| | |
All information passed over the IRMd/IPCP boundary for using IPC
services (flow allocation, registration) is now hashed. This
effectively fixes the shared namespace between DIFs and the IRMDs.
This PR also fixes some API issues (adding const identifiers),
shuffles the include headers a bit and some small bugs.
|
| |/
|
|
|
|
| |
This adds the STATUS variable to the message() call in CMakeLists.txt
in places where it was missing. This ensures that the message is
printed to stdout instead of stderr.
|
| | |
|
| | |
|
| |
|
|
|
| |
Adds support for netmap (https://github.com/luigirizzo/netmap) to
increase the performance of packet transfer over the LLC shim.
|
| |
|
|
| |
This checks if the netmap headers are available on the system.
|
| | |
|
| |
|
|
|
|
| |
An assertion was done instead of NULL check, where the FRCT instance
could legitimately be NULL, resulting in the IPCP dying when it
shouldn't.
|
| |
|
|
|
| |
This avoids handling the mgmt frames under lock, since it may deadlock
if a new mgmt frame arrives in the meantime.
|
| |
|
|
|
|
| |
Setting the timeouts on flow_alloc and flow_accept will now work. This
makes some changes to the UNIX sockets used for management
communication between the APs, IRMd and IPCPs.
|
| |
|
|
|
|
| |
This fixes the overwriting of management frames by adding a list in
the shim Ethernet that keeps track of management frames instead of a
single buffer.
|
| |
|
|
|
| |
A missing else clause was missing in the fast path, causing the PCI to
be shrunk when it should not be. A double free has also been fixed.
|
| |
|
|
|
| |
This adds a lock to prevent a race condition between flow_req_arr and
flow_alloc_resp.
|
| |
|
|
|
| |
The state check should be done first to avoid blocking on the recv
call.
|
| |\
| |
| |
| | |
ipcpd: Remove unnecessary locks in local
|
| | | |
|
| |\|
| |
| |
| | |
Be irmd res
|
| | | |
|
| | |
| |
| |
| |
| | |
This removes the rwlock in the shim-eth-llc IPCP since it is already
protected by a mutex.
|
| | |
| |
| |
| |
| | |
This removes cancellation calls from the shim Ethernet. The different
threads now check if the IPCP is still operational or not.
|
| | |
| |
| |
| |
| |
| | |
This deprecates these special socket options of the raw socket since
they are less efficient for regular packet I/O. They should be more
performant for batch processing of SDUs.
|
| |/
|
|
|
|
| |
This adds a thread to handle management frames, since otherwise a
deadlock can occur by blocking SDUs until flow allocation can be
handled.
|
| | |
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
| |
Since there are no SDUs on most QoS cubes, the fmgr was always timing
out on most QoS cubes, causing considerable delays in most SDUs. The
timeout was reset to a very small value (so the bug in glibc 2.25 will
cause lockups again).
|
| | |
|
| |
|
|
|
| |
This removes some logs in the graph component and replaces prints of
uin64_t variables with an architecture agnostic variant.
|
| |
|
|
| |
This fixes the bug in handling multiple concurrent flow allocations.
|
| |\
| |
| |
| | |
lib: Revise flow allocation API
|
| | |
| |
| |
| |
| |
| | |
The flow_alloc_res and flow_alloc_resp calls have been removed. The
flow_alloc and flow_accept calls are now both blocking and take an
additional timeout argument.
|
| |/ |
|
| |
|
|
|
| |
The FRCT instance was not properly cleaned when the PFF could not find
the next hop.
|
| |
|
|
|
|
| |
This fixes a bug in enrollment where only the first enrollment would
happen correctly since the booleans indicating whether boot, members,
and DIF name were sent were not initialized back to false.
|
| |
|
|
|
| |
This adds a regression test for the graph component to test the
routing table.
|
| |
|
|
|
|
| |
This turns the directed graph into an undirected one. Only one side of
the flow creates an FSDB entry. The graph structure creates an edge
object for every vertex involved when an edge is updated or removed.
|
| |
|
|
| |
They were not consumed upon reading, causing the rdrbuff to fill up.
|
| | |
|
| | |
|
| |
|
|
|
| |
The PCI was being freed by frct, but it was stack memory which was
created in the fmgr, resulting in an illegal free.
|
| |
|
|
|
|
| |
The shm PCI was never initialized during flow manager init. This
commit will do that, and initialize the pdu length correctly as well,
since it was not being written into the RIB, nor read in shm_pci_init.
|
| |
|
|
|
| |
The index j in the function that transforms the list of predecessors
to a routing table was incremented at the wrong time.
|
| |
|
|
|
|
| |
This adds fixes the locking of the PFF which was externalized, but not
yet correctly updated within the PFF component itself and within the
flow manager.
|
| |
|
|
|
|
| |
This removes the graph_add_edge operation of the graph component. The
routing component now only listens to RO_MODIFY events, and updates
the graph accordingly.
|
| |
|
|
|
| |
This fixes several bugs in the ro sets, rib. And it fixes several bugs
in the graph and routing component of the normal IPCP.
|
