Age | Commit message (Collapse) | Author | Files | Lines |
|
We have to be more careful when retiring children. We cannot stop
accepting connections as soon as the server tells us to quit because if
max_children is reached and we still have clients connected, the server will
not be able to spawn new children until one exits. And children will not exit
until all the clients closed the connections. So we keep operating past our
recall until we have 0 connections.
Also do not try to recall children that still have >= 1 clients connected, they
couldn't anyway.
Also use messaging to warn children and not SIGHUP.
Signed-off-by: Andreas Schneider <asn@samba.org>
Signed-off-by: Simo Sorce <idra@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
Signed-off-by: Simo Sorce <idra@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
Signed-off-by: Simo Sorce <idra@samba.org>
|
|
We can't have a clear idea of wether the worker is IDLE or BUSY.
The only things we can tell is if it is Alive, whether it is currently
Accepting connections or wether it is Exiting soon.
Remove PF_WORKER_IDLE, PF_WORKER_BUSY and replace their use with
PF_WORKER_ALIVE. Also properly assign PF_WORKER_ACCEPTING so that
users of the API can rely on the flag.
Signed-off-by: Andreas Schneider <asn@samba.org>
Signed-off-by: Simo Sorce <idra@samba.org>
|
|
We used a lock mimicking what apache does for preforked children.
But it doesn't work properly in our case because we do not stop once a request
has been served. Clients are allowed to perform multiple requests and keep the
connection open.
This means that if we allow multiple clients per children, then a child could
take the lock and then be asked to do a long or even locking operation by a
client it already is serving. This woulkd cause the whole server to deadlock,
as the child is now busy and also holding on the lock.
Using a race on accept() by having a tevent_fd on the listening socket wait
for read events we never deadlock. At most we cause a bit of contention among
children. But in the generic case connections are much less frequent for us as
clients tend to be long lived. So the little contention we may have is not a
big deal.
Signed-off-by: Andreas Schneider <asn@samba.org>
Signed-off-by: Simo Sorce <idra@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
Signed-off-by: Simo Sorce <idra@samba.org>
|
|
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Pair-Programmed-With: Andreas Schneider <asn@samba.org>
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
We need to properly handle preforked children so it is better to just do that
automatically.
If the parent needs/wants to intercept SIGCHLD events it can set a callback
that will be called by the prefork code once the internal cleanup function that
checks all prefork children has been executed.
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
This way the parent doesn't need to know how to handle dead children and
keeps all of that within the prefork abstraction.
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
The allowed_clients var is a parent managed variable that tell children how
many clients they are allowed to handle at the same time. This way children
can overcommit but within parent controlled limits.
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
To get a client connection it is now possible to use asynchronous functions.
Signed-off-by: Andreas Schneider <asn@samba.org>
|
|
Primarily built for forked off rpc service daemons, but not tied to rpc
services and generic enough to be used elsewhere easily.
Signed-off-by: Andreas Schneider <asn@samba.org>
|