| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
| |
Some NICs were not enabling TX IP checksum offloads. This fixes that
issue.
|
|
|
|
| |
This was old code that wouldn't compile or run anymore.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Add configuration sections to the UHD config file for NIC entries. Keys
are based on MAC addresses, and the entries beneath the section describe
which CPU and I/O thread to use for the NIC and its IPv4 address.
Make ring sizes configurable for uhd-dpdk. Ring size is now an argument
for packet buffers. Note that the maximum number of available buffers
is still determined at init!
Add ability to receive broadcasts to uhd-dpdk. This is controllable by
a boolean in the sockarg during socket creation. dpdk_zero_copy will
filter broadcast packets out.
Add dpdk_simple transport (to mirror udp_simple). This transport allows
receiving from broadcast addresses, but it only permits one outstanding
buffer at a time.
Fix IP checksum handling in UHD-DPDK.
TX checksums were not being calculated in the NIC, and in RX, the check
for IP checksums allowed values of zero (reported as none). Now packets
with bad IP checksums will be dropped.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This adds an internal wait queue API to uhd-dpdk. Socket configuration
requests had their blocking calls re-implemented on top of this API, and
it is also used to service requests to wait on RX packets (w/ timeout).
The wait API involves a multi-producer, single-consumer queue per I/O
thread (waiter_ring), with a condition variable used for sleeping. The
data structure is shared between user thread and I/O thread, and because
timeouts make resource release time non-deterministic, we use reference
counting on the shared resource.
One reference is generated by the user thread and passed to the I/O
thread to consume. A user thread that still needs the data after waking
must get() another reference, to postpone the destruction of the
resource until it is done.
Timeouts are based on CLOCK_MONOTONIC. For recv, a timeout of 0
indicates blocking indefinitely, and a negative timeout indicates no
timeout is desired.
Also drop timeout for closing sockets in uhd-dpdk.
The timeout would allow a user thread to pre-empt the I/O thread's
cleanup process. The user thread would free data structures the I/O
thread still needed to function. Since this timeout is superfluous
anyway, let's just get rid of it.
Also add some more input checking and error reporting.
|
|
|
|
|
|
|
|
|
|
|
|
| |
Set MTU of Ethernet ports:
Some NICs (like the Mellanox ones) require this to work.
Add ARP responder to uhd-dpdk.
Clean up pending ARP request list when done:
Threads waiting for an ARP request to complete would be woken up when
the request completed, but they wouldn't get removed from the list of
pending requests. This fixes the issue.
|
|
This library makes available a userspace network stack with a
socket-like interface for applications (except the sockets pass around
pointers to buffers and use the buffers directly--It's sockets + a
put/get for buffer management). Supported services are ARP and UDP.
Destinations can be unicast or broadcast. Multicast is not currently
supported.
The implementation has two driver layers. The upper layer runs within
the caller's context. The caller will make requests through lockless
ring buffers (including socket creation and packet transmission), and
the lower layer will implement the requests and provide a response.
Currently, the lower layer runs in a separate I/O thread, and the caller
will block until it receives a response.
The I/O thread's main body is in src/uhd_dpdk_driver.c. You'll find that
all I/O thread functions are prefixed by an underscore, and user thread
functions do not.
src/uhd_dpdk.c is used to initialize uhd-dpdk and bring up the network
interfaces.
src/uhd_dpdk_fops.c and src/uhd_dpdk_udp.c are for network services.
The test is a benchmark of a flow control loop using a certain made-up
protocol with credits and sequence number tracking.
|