| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
The writev system call can return -1 and errno ENOMEM, as a
retriable condition.
Reviewed-by: Jeremy Allison <jra@samba.org>
Autobuild-User(master): Jeremy Allison <jra@samba.org>
Autobuild-Date(master): Wed Mar 13 23:50:05 CET 2013 on sn-devel-104
|
|
This matches what was done for lib/socket/socket_unix.c in
c692bb02b039ae8fef6ba968fd13b36ad7d62a72.
(and is based on that patch by Landon Fuller <landonf@bikemonkey.org>)
Andrew Bartlett
Reviewed-by: Stefan Metzmacher <metze@samba.org>
Autobuild-User(master): Stefan Metzmacher <metze@samba.org>
Autobuild-Date(master): Mon Mar 4 11:15:35 CET 2013 on sn-devel-104
|
|
We only do the optimization on recvfrom/readv if the caller asked for it.
This is needed because in most cases we preferr to flush send
buffers before receiving incoming requests.
Signed-off-by: Stefan Metzmacher <metze@samba.org>
|
|
If the socket is not readable yet, we need to retry
if tsocket_bsd_pending() returns 0.
See also
https://lists.samba.org/archive/samba-technical/2012-October/087164.html
metze
Autobuild-User(master): Andrew Bartlett <abartlet@samba.org>
Autobuild-Date(master): Tue Oct 23 14:44:21 CEST 2012 on sn-devel-104
|
|
It's possible for a client to send 0 bytes in a UDP packet,
we need still need to call recvfrom() and skip the invalid
packet at a higher level. Otherwise the kernel receive queue
is blocked.
metze
|
|
metze
|
|
local address
tstream_inet_tcp_connect_send() usually only gets no local port number
and it may use the wildcard address '0.0.0.0' or '::'.
tstream_inet_tcp_connect_recv() provides the used local address and port
which are used on the wire.
metze
|
|
metze
|
|
We got random double free errors, when getting events from
epoll_wait() and try to dereference the private talloc pointer
attached to it.
Before doing the close() in the tstream_disconnect_send() function
we need to delete the fd event.
commit 38f505530ba06323a56c7d3914630efffcd12629 only fixed it for
tdgram sockets.
metze
|
|
We got random double free errors, when getting events from
epoll_wait() and try to dereference the private talloc pointer
attached to it.
Before doing the close() in the tstream_disconnect_send() function
we need to delete the fd event.
metze
Autobuild-User: Stefan Metzmacher <metze@samba.org>
Autobuild-Date: Tue Sep 28 01:02:55 UTC 2010 on sn-devel-104
|
|
the code used this pattent:
if (fd < 0) {
...various cleanups...
return fd;
}
it is much clearer to do this:
if (fd < 0) {
...various cleanups...
return -1;
}
as otherwise when reading the code you think this function may return
a fd.
Pair-Programmed-With: Andrew Bartlett <abartlet@samba.org>
|
|
This reverts commit 3d4fb698660381e650d7caeb5b7cff12847c0fb8.
This was wrong... The problem was in the caller, sa_len is a macro on irix
metze
|
|
metze
|
|
|
|
metze
|
|
metze
|
|
This reverts commit 63fc3978089df403d7e63462593784ff3d05e6e7.
We now don't support IPv6 when IPV6_V6ONLY isn't available,
as we rely on the behavior.
metze
|
|
The previous code assumed the OS would happily accept sizeof(struct
sockaddr_storage). It seems some versions of Solaris do not like
this.
Andrew Bartlett
|
|
|
|
metze
|
|
Which was:
tsocket/bsd: fix bug #7115 FreeBSD includes the UDP header in FIONREAD
Metze, this has to have been wrong - you are throwing away the talloc_realloc
pointer returned. Also no error checking. Please review.
Thank goodness for gcc warnings :-).
Jeremy.
|
|
metze
|
|
Some system already have this as default. It's easier
to behave the same way on all systems and handle ipv6
and ipv4 sockets separate.
metze
|
|
address if the local address is any
metze
|
|
tstream_bsd_connect_send()
This is needed because, we can't use sizeof(sockaddr_storage) for AF_UNIX
sockets. Also some platforms require exact values for AF_INET and AF_INET6.
metze
|
|
We want the explicit bind() when we don't use the any address.
metze
|
|
This makes sure we fill the unspefic bytes with 0
and have zero termination for sockaddr_un.
metze
|
|
This is needed as getpeername() truncates the length to 2 if the
peer socket has no name.
metze
|
|
|
|
tsocket_address_bsd_from_sockaddr()
metze
|
|
long
metze
|
|
platforms
metze
|
|
socket
Otherwise we would not notice a broken connection.
metze
|
|
I some cases the pointer value of tevent_context is the same again,
if we do something like:
ev1 = tevent_context_init();
...
fde = tevent_add_fd(ev1, fd, TEVENT_FD_READ...);
...
talloc_free(ev1);
...
ev2 = tevent_context_init();
if (ev1 == ev2) {
/* this can happen! */
}
if (tevent_fd_get_flags(fde) == 0) {
/* this is always true */
}
But the "talloc_free(ev1)" will set fde->event_ctx to NULL
and tevent_fd_get_flags() will always return 0.
metze
|
|
metze
|
|
tevent_add_fd does not properly set that. At least in epoll and select this is
the only error condition.
Metze, please check!
|
|
Remove two indentation levels by returning early on error.
Metze, please check!
|
|
metze
|
|
metze
|
|
all backends
We need to make sure that we free the sendto and recvfrom requests before
the tdgram_context
metze
|
|
sizeof(struct sockaddr_storage) generates EINVAL.
metze
|
|
It will be replaced by tdgram_context and tstream_context.
metze
|
|
metze
|
|
metze
|
|
metze
|
|
metze
|
|
The desire is to do as less syscalls during the
tdgram_sendto_send/recv() and tdgram_recvfrom_send/recv()
operations.
1. we first try the sendto()/recvfrom() syscall and
only use a fd event if we got EAGAIN.
2. we cache the fd event and only change it's flags
if really needed.
For the highload case we do almost no epoll_ctl() and epoll_wait()/select()
syscalls anymore. This speeds up the LDAP-BENCH-CLDAP test
by more than 20%. (With a modified version of this test
which let the server skip any ldb calls and just return success
I'm getting about 8000 requests per second, while I'm getting
just about 6000 requests per second without optimization)
metze
|
|
The idea is to have a tdgram and a tstream abstraction
which only provide tevent_req based io functions.
metze
|
|
metze
|
|
metze
|
