| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
| |
With this patch, all the states updates are enqueued in the
tx_queue. Thus, there's a single output path. This patch adds a
simple refcounting mechanism to note when an object is sitting in
the txqueue. This patch also removes the alarm that is required by
the ftfw approach.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
| |
This patch adds a generic tx queue for all synchronization modes.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch unifies the tx_list and the tx_queue to have only one
transmission queue. Since the tx_list hold state objects and
tx_queue control messages, I have introduced a queue node type
that can be used to differenciate the kind of information that
the node stores: object or control message.
This patch also reworks the existing queue class to include a
file descriptor that can be used to know if there are new data
added to the queue (see QUEUE_F_EVFD flag).
In this change, I have also modified the current evfd to make the
file descriptor to make read operations non-blocking. Moreover,
it keeps a counter that is used to know how many messages are
inserted in the queue.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Currently, the caching system is implemented in a two layer
architecture: hashtable (inner layer) and cache (upper layer).
This patch reworks the hash-cache infrastructure to solve some
initial design problems to make it more flexible, the main strong
points of this patch are:
* Memory handling is done in the cache layer, not in the inner
hashtable layer. This removes one of the main dependencies between
the hashtable and the cache classes.
* Remove excessive encapsulation: the former cache used to hide a
lot of details of the inner hashtable implementation.
* Fix over-hashing of some operations: lookup-delete-add required
three hash calculations. Similarly, the update-or-add operation
required two hash calculations. Now, we calculate the hash once
and re-use the value how many times as we need.
This patch simplifies the caching system. As a result, we save ~130
lines of code. Small code means and less complexity means less
chance to have bugs.
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
|
| | |
|
| |
|
|
|
|
| |
the global variable "alarm" conflicts with the alarm() function from
unistd.h. resolve that conflict by giving those two global variables
a better name.
|
| |
|
|
| |
(Based on comments from Max Kellerman)
|
| |
|
|
|
|
| |
o rename `persistent' mode to `alarm'
o rename `nack' mode to `ftfw'
o Now default synchronization mode is ftfw instead of alarm
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
- add len field to nethdr
- implement buffered send/recv to batch messages
- stop using netlink format for network messages: use similar TLV-based format
- reduce synchronization messages size up to 60%
- introduce periodic alive messages for sync-nack protocol
- timeslice alarm implementation: remove alarm pthread, remove locking
- simplify debugging functions: use nfct_snprintf instead
- remove major use of libnfnetlink functions: use libnetfilter_conntrack API
- deprecate conntrackd -F, use conntrack -F instead
- major rework of the network infrastructure: much simple, less messy
|
| |
|
|
|
|
| |
- remove several debug calls
- create a child to dispatch dump requests: this will help to simplify the
current locking schema. Later.
|
| |
|
|
|
|
|
|
| |
o remove reminiscents of delay destroy message and relax transitions
o remove confusing StripNAT parameter: NAT support enabled by default
o relax event tracking: *_update callbacks use cache_update_force
o use wraparound-aware functions after/before/between
o lots of cleanups
|
| |
|
|
|
|
|
| |
- flush nack queue in the conntrackd -f path
- do not increase add_fail counter for EEXIST errors
- cleanup sync-nack code
- improve mcast_recv_netmsg: sanity check before checksumming!
|
|
|
chain
|
