This functions allows iterate available cpg groups
and their members. API is modelled like ckpt iteration
functions.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2399 fd59a12c-fef9-0310-b244-a6a79926bd2f
The root of the theoretical problem is that cpg_join or cpg_leave
messages are being sent via the C apis between synchronization. With
the current cpg, synchronization happens in confchg_fn, and then later
in cpg_sync_process. cpg_sync_process is called much later after
confchg_fn and introduces a small probability of a window of time for
queued in totem (but not yet ordered by totem) for those cpg_join and
cpg_leave operations to interact with the synchronization process which
should happen in one atomic operation but currently is two distinct
operations.
This patch deletes confchg_fn and make sends joinlist/downlist
in cpg_sync_process.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2365 fd59a12c-fef9-0310-b244-a6a79926bd2f
This patch allows only one connection per (node, pid, grp_name) tuple.
This means, you cannot make more connection from one process to same
group_name. This is (I hope) how cpg should behave. In case, you will
try to do that, CPG_ERR_EXISTS error is returned.
Of course, there is no problem with creating:
- more connection with same (pid, grp) if nodeid is different
- more connection with same (node, grp) if pid is different (for example
after fork, or two distinct processes)
- more connection with same (node, pid) if grp is different (connect
one process to more cpgs).
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2364 fd59a12c-fef9-0310-b244-a6a79926bd2f
This patch fixes situation, when in the middle of
sync some node will send regular message before
another node will receive confch message, and regular
message is delivered to application. From application
point of view, this node is unknown -> don't expect
any messages.
Now, no such messages are delivered to application.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2332 fd59a12c-fef9-0310-b244-a6a79926bd2f
This is needed as the objdb order will change as modules are loaded/unloaded and is
also set up to unload non-default services last (which is the opposite of what
something like Pacemaker needs).
In the worst case, the current behavior leads to cluster services (dlm, ocfs2, etc)
failing during shutdown. This patch also ensures that if, for example, cpg is unloaded
then anything that depends on it is unloaded first.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2224 fd59a12c-fef9-0310-b244-a6a79926bd2f
This could probably be more tidy to detect those OS platforms which don't do this instead of hardcoding
to a specific platform we intend to port to.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2221 fd59a12c-fef9-0310-b244-a6a79926bd2f
Make a tiny type change and watch it propagate.
* include/corosync/engine/coroapi.h
(struct corosync_lib_handler) [lib_handler_fn]: Change type
of 2nd parameter: s/void *msg/const void *msg/.
Propagate the above into cfg.c and votequorum.c:
* services/cfg.c (message_handler_req_lib_cfg_get_node_addrs):
Constification exposed a bug in this function whereby it mistakenly
modified storage through its now-const *msg parameter. Since it
did that solely to store a temporary result, we've changed it
to use a local variable instead.
* services/votequorum.c (message_handler_req_lib_votequorum_setvotes):
Likewise.
* exec/vsf_quorum.c: add const to msg param.
* services/evs.c: Likewise.
* services/pload.c: Likewise.
* services/cpg.c: Likewise.
* services/confdb.c: Likewise.
* exec/coroipcs.h: signature of coroipcs_handler_fn_lvalue must match
that of lib_handler_fn; noted via main.c.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@2047 fd59a12c-fef9-0310-b244-a6a79926bd2f
cpg_finalise is called. It can cause the process_info structure to be
removed twice from the group list.
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@1884 fd59a12c-fef9-0310-b244-a6a79926bd2f
- services/Makefile.am: add include search paths for config.h
- include/corosync/cs_config.h.in exports LCRSODIR and SOCKETDIR
- tools/Makefile.am: drop -D defines for dirs that are now in
cs_config.h or config.h
- configure.ac: sanitize prefix and exec_prefix paths. Export DIRS in
*config.h
- lib/Makefile.am: : drop -D defines for dirs that are now in
cs_config.h or config.h. Add rule to build lcr_ifact.o or building from
lib/ fails miserably
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@1877 fd59a12c-fef9-0310-b244-a6a79926bd2f
This patch causes the flow control state in the library to be set
properly when the flow control is turned off (disabled). Then it can be
read properly by the flow control apis.
This also fixes the case where the application is no longer sending
messages and it has already dispatched all its received messages
before flow control is disabled.
Also, CPG response messages with a TRY_AGAIN error did NOT contain
a valid flow control state value. This meant the library could get
stuck with flow control enabled (flow control was never enabled
for the EXEC, so no disable event occurred).
This case was hit when a new node was joining - sync_in_process()
resulted in a TRY_AGAIN for error cpg_mcast_joined).
Also, in message_handler_req_exec_cpg_mcast() the state passed
back to the library defaulted to disabled for messages received
from another node (even if flow control was still enabled)
- this meant if multiple nodes were sending CPG messages,
then the library flow control state flip-flopped between
enabled and disabled.
Author: Steven Dake <sdake@redhat.com> &
Tim Beale <tim.beale@alliedtelesis.co.nz>
git-svn-id: http://svn.fedorahosted.org/svn/corosync/trunk@1667 fd59a12c-fef9-0310-b244-a6a79926bd2f
