| Commit message (Collapse) | Author | Age | Files | Lines |
|
|
|
|
|
|
|
| |
The fast path will now use an incoming ring buffer per flow per
process. This necessitated the development of a new method for the
asynchronous io call, which is now based on an event queue system for
scalability (fqueue). The ipcpd's and tools have been updated to this
API.
|
| |
|
|
|
|
|
|
| |
This will allow to finalize deallocating flows until all SDUs have
been processed. Read and write calls will now block when a flow was
deallocated. Replaces NULL checks in the fast path with asserts.
|
|
|
|
|
| |
The SDU's were not correctly removed upon shutdown, peek should be
pop.
|
|
|
|
|
|
|
|
|
|
|
|
| |
IPCPs can now use ap_init() to initialize the memory. All flows are
accessed using flow descriptors, this greatly simplifies IPCP
development. Reverts the fast path to a single ap_rbuff per process.
Splits lib/ipcp into irmd/ipcp and lib/ipcp-dev. Adds a lib/shim-dev
holding tailored functions for shims. Moves the buffer_t to utils.h.
Fixes the shim-eth-llc length field. Removes the flow from shared.h.
Fixes #4
Fixes #5
|
|
|
|
|
|
|
| |
The flow_select call now takes as a parameter a flow_set_t, which
specifies a set of flow descriptors that will unblock the select call
when an SDU for one of them arrives. The select call has been moved to
its own header.
|
|
|
|
| |
Fast path is split in north and southbound paths.
|
|
|
|
|
|
|
|
|
|
| |
The shm_du_map is renamed to shm_rdrbuff to reflect the Random
Deletion Ringbuffer used in the implementation. The close_on_exit call
is removed and SDUs are cleaned up by the application in the ap_fini()
call. This required a non-blocking peek() operation in the shm_ap_rbuff.
Some initial implementation for future support of qos cubes has been
added to the shm_rdrbuff.
|
|
|
|
|
| |
Fix missing set of api upon flow_alloc.
Various locking fixes.
|
|
|
|
|
|
|
|
|
| |
Blocking I/O now uses condition variables in the shared memory instead
of busy waiting. Timeouts can be specified. This requires the size of
the rbuffs and du_map to be the same, to guarantee that when the
shm_du_map is not full, the ap_rbuffs can't be full either.
Added the timeout option to the flow for future use.
|
|
|
|
|
|
|
| |
This adds a flow_select() call that will sleep until an SDU can be
read on a flow. It returns the file descriptor for which an SDU is
ready. It takes as optional argument a timespec struct to specify a
timeout.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This PR enhances the shared memory providing recovery if a process
crashes. It adds a SHM_DU_TIMEOUT_MICROS variable, setting an
expiration time for SDU's when shared memory is full. If an
application doesn't read a blocking SDU within this time, the shared
memory will be cleansed of all SDU's for this application and the
application's rbuff will be cleared.
Some refactoring of the API's. Fixed wrong pthread checks in IRMd.
Fixes #13
Fixes #14
|
|
|
|
|
|
|
| |
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.
|
|
|
|
|
|
| |
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.
|
|
|
|
|
|
| |
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.
|
|
|
|
|
|
| |
Adds a missing modulo calculation that would result in index -1 being
returned instead of zero when the ringbuffer skipped back to the
beginning.
|
|
|
|
| |
This better reflects the use as the pid is appended to that name.
|
|
This commit has a first implementation of flow allocation (the "slow
path") and read/write (the "fast path") for ouroboros. It provides
basic but unstable communications over the shared memory.
It required a lot of changes all over the stack, and fixes a number of
previously undetected issues.
This PR still need heavy revision regarding data model, locking and
cleanup.
lib/dev: modifications to the API. It now uses an ap_init() call to
set the AP name and sets the Instance ID to the pid of the process. It
also binds the AP to the shared memory and creates tables for mappings
in the fast path. A call to ap_fini() releases the resources.
lib/shm_ap_rbuff: added ring buffer for data exchange between
processes in the fast path. It passes an index in the shm_du_map.
lib/shm_du_map: rewrote API to work with calls from dev.c. Garbage
collector added. Tests updated to new API.
ipcpd/ipcp-data: removed everything related to flows, as these are
universal for all ap's and kept in ap_data (dev.c), or similar structs
for shim ipcps.
shim-udp: added flow allocator and read/write functions and shm
elements.
irmd: revised data model and structures necessary for flow allocation.
tools: echo updated to new dev.h API.
messaging system was updated to comply with new flow allocation
messages. All exchanges use pid and port_id to bootstrap the fast
path.
|