IF YOU WOULD LIKE TO GET AN ACCOUNT, please write an
email to Administrator. User accounts are meant only to access repo
and report issues and/or generate pull requests.
This is a purpose-specific Git hosting for
BaseALT
projects. Thank you for your understanding!
Только зарегистрированные пользователи имеют доступ к сервису!
Для получения аккаунта, обратитесь к администратору.
In the future we will have multiple task id's per process. They should all be
able to benefit from the self-send local optimization.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
In the messaging_read receivers we already defer the callback: We need to
reply on potentially different tevent contexts, thus the defer_callback.
The callback case in messaging_dispatch_rec was direct before this
patch. This changes messaging_dispatch_rec to also defer the callback
in the self-send case.
Now we need only two roundtrips in local-messaging-read1 :-)
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
If we register an additional tevent context, we can now properly do
nested event contexts, listening for just one message type inside a
tevent_req_poll.
At this point this only enhances things without ctdb, but I'm working fixing
that soon.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
This delegates the decision whether to read a message to a callback
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
messaging_read_send/recv was okay for just one handler in the queue. For
multiple handlers it was pretty broken.
A handler that deletes itself as part of the callback (pretty typical use
case...) drops the message for a subsequent handler that responds to the same
message type. In messaging_dispatch_rec we walk the array, however
messaging_read_cleanup has already changed the array. tevent_req_defer_callback
does not help here: It only defers the callback, it does not defer the cleanup
function.
This also happens when a callback deletes a different handler
A handler that re-installs itself in the callback might get a message twice.
This patch changes the code such that only messaging_dispatch_rec adds records
to msg_ctx->waiters, new waiters are put into a staging area first
(msg_ctx->new_waiters). Also messaging_read_cleanup does not move anything
around in msg_ctx->waiters, it only nulls out itself. messaging_dispatch_rec is
changed to cope with this.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Callers might not be interested in the rec, just the fact that something
arrived
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
This uses a copy, will be replaced by a direct iovec call through to
sendmsg on the unix domain socket
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Messaging based on unix domain datagram sockets
This makes every process participating in messaging bind on a unix domain
datagram socket, similar to the source4 based messaging. The details are a bit
different though:
Retry after EWOULDBLOCK is done with a blocking thread, not by polling. This
was the only way I could in experiments avoid a thundering herd or high load
under Linux in extreme overload situations like many thousands of processes
sending to one blocked process. If there are better ideas to do this in a
simple way, I'm more than happy to remove the pthreadpool dependency again.
There is only one socket per process, not per task. I don't think that per-task
sockets are really necessary, we can do filtering in user space. The message
contains the destination server_id, which contains the destination task_id. I
think we can rebase the source4 based imessaging on top of this, allowing
multiple imessaging contexts on top of one messaging_context. I had planned to
do this conversion before this goes in, but Jeremy convinced me that this has
value in itself :-)
Per socket we also create a fcntl-based lockfile to allow race-free cleanup of
orphaned sockets. This lockfile contains the unique_id, which in the future
will make the server_id.tdb obsolete.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
When a smbd process dies, pending messages.tdb records for this process
might not get cleaned up. Implement a cleanup for dead records that is
triggered after a smbd dies uncleanly; the records for that PID are
deleted.
Based on a patchset from Christof Schmitt <cs@samba.org>.
Signed-off-by: Jeremy Allison <jra@samba.org>
Reviewed-by: Christof Schmitt <cs@samba.org>
All ctdb specific code is isolated in samba-cluster-support.so now.
Signed-off-by: Stefan Metzmacher <metze@samba.org>
Reviewed-by: Michael Adam <obnox@samba.org>
Autobuild-User(master): Michael Adam <obnox@samba.org>
Autobuild-Date(master): Mon Mar 24 19:08:44 CET 2014 on sn-devel-104
This is made to replace the msg_channel abstraction.
msg_channel was created to not miss any messages. For this, some
complex queueing was installed. This complexity has caused quite a
few problems in the past (see bug 10284 for example).
messaging_read_send/recv is able to achieve the same goal with a
lot less complexity. The messaging_read_send atomically installs
the reader into the messaging_context, we will not miss any messages
while this installed. messaging_send_recv will deinstall that
listener, but in the callback function you can directly call
messaging_read_send again without going through the tevent_loop_once.
As long as this is always made sure, no messages will be lost.
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Immediate tevents don't free themselves as timed events do :-)
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
Autobuild-User(master): Jeremy Allison <jra@samba.org>
Autobuild-Date(master): Fri Jan 10 01:20:04 CET 2014 on sn-devel-104
... not as a speed improvement, it saves the second NULL check
Signed-off-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Jeremy Allison <jra@samba.org>
We've always had the corresponding deregister message
Autobuild-User: Volker Lendecke <vlendec@samba.org>
Autobuild-Date: Tue Jan 24 15:27:51 CET 2012 on sn-devel-104
This has been broken since ff0ac5b0 (May 2007).
Basically all messages were belonging to the General class except for CTDB
messages.
This fixed the message_send_all() function to correctly compute the class, and
fixes registrations to include all they need to cope with the fact not all
messages are of calss general (registrations rotted a bit because as long as
FLAG_MSG_GENERAL was defined the process woould receive all messages).
Signed-off-by: Andreas Schneider <asn@samba.org>
The FLAG_MSG_PRINT_NOTIFY class is actually obsolete and never used, as the
only message belonging to it is not used either.
Signed-off-by: Andreas Schneider <asn@samba.org>
the daemons themselves. Allows client utilities to silently
fail to create a messaging context due to access denied on the
messaging tdb (which I need for the following patch).
Jeremy.
If we call messaging_ctdbd_connection() we end up with the wrong vnn in our
messaging context.
This is a bit of a hack, get_my_vnn() needs to go eventually along with
procid_self()
When a samba server process dies hard, it has no chance to clean up its entries
in locking.tdb, brlock.tdb, connections.tdb and sessionid.tdb.
For locking.tdb and brlock.tdb Samba is robust by checking every time we read
an entry from the database if the corresponding process still exists. If it
does not exist anymore, the entry is deleted. This is not 100% failsafe though:
On systems with a limited PID space there is a non-zero chance that between the
smbd's death and the fresh access, the PID is recycled by another long-running
process. This renders all files that had been locked by the killed smbd
potentially unusable until the new process also dies.
This patch is supposed to fix the problem the following way: Every process ID
in every database is augmented by a random 64-bit number that is stored in a
serverid.tdb. Whenever we need to check if a process still exists we know its
PID and the 64-bit number. We look up the PID in serverid.tdb and compare the
64-bit number. If it's the same, the process still is a valid smbd holding the
lock. If it is different, a new smbd has taken over.
I believe this is safe against an smbd that has died hard and the PID has been
taken over by a non-samba process. This process would not have registered
itself with a fresh 64-bit number in serverid.tdb, so the old one still exists
in serverid.tdb. We protect against this case by the parent smbd taking care of
deregistering PIDs from serverid.tdb and the fact that serverid.tdb is
CLEAR_IF_FIRST.
CLEAR_IF_FIRST does not work in a cluster, so the automatic cleanup does not
work when all smbds are restarted. For this, "net serverid wipe" has to be run
before smbd starts up. As a convenience, "net serverid wipedbs" also cleans up
sessionid.tdb and connections.tdb.
While there, this also cleans up overloading connections.tdb with all the
process entries just for messaging_send_all().
Volker
The change notify code registered a separate message handler for each
tree connect. This registration uses the global messaging context.
The messaging code would consider a 2nd registration for the same
messaging type as being an 'update' of the handler, rather than a new
handler. It also would only call the first handler in the linked list
for a given message type when dispatching messages.
This patch changes the messaging code to allow for multiple
registrations of the same message type, and allow for multiple calls
to different messaging handler for one incoming message.
This fixes the problem with the test_notify_tcon() test that I
recently committed to the S4 smbtorture
