Let's add a mode (enabled via the LIBOSMO_IO_URING_BATCH environment
variable), where we don't call io_uring_submit() after every operation
we add to the submission queue. Rather, do that once before we go into
poll.
This should massively reduce the amount of io_uring_enter() syscalls
we're seeing.
Change-Id: Ia27ada909df246f21d110d8ae0b40229aacd579d
"Assert failed 0 osmo_io_uring.c:94" is rather useless in terms of an
error message. Let's improve upon that.
Change-Id: I0ac9ae59e99c3704b3ce33919c9e7d11024476e3
If it fails, do not set the IOFD_FLAG_NOTIFY_CONNECTED flag and log an
error message.
Closes: Coverity CID#349579
Change-Id: I34e8cc9a2b9df0c624841e5f9268a15c32418da1
msghdr must be detached, because subsequent callback at
iofd_handle_send_completion() may destroy the iofd (which in turn
frees this msghdr, if still attached) and frees the msghdr, causing a
double free.
Related: OS#5751
Change-Id: Ia349f73de2145fa360b20dd40deb73a8ffc71f07
There is always a completion after cancelling a uring request.
Because uring requests use msghdr pointer as user data, we cannot just
free the msghdr after cancelling. Upon completion (received after
cancelling), the user data still points to the msghdr. To prevent a
use-after-free bug, msghdr is not freed, but detached from iofd
instance. Then upon completion, the msghdr (if it was detached from
iofd) is freed.
Additionally it is not required to keep IOFD_FLAG_IN_CALLBACK set
anymore, if there is a msghdr attached to iofd. As described above,
all msghdr get detached, if iofd is freed (uring request get cancelled)
during callback.
Related: OS#5751
Change-Id: Ic253f085dd6362db85f029f46350951472210a02
In order to receive a connect notification from SCTP socket,
poll/select event must be used instead of a write notification via
io_uring completion event.
Once the connect notification has been received, subsequent write
notifications via io_uring are used.
Change-Id: I4eca9ea72beb0d6ea4d44cce81ed620033f07270
Related: OS#5751
Add support osmo_io operations resembling sendmsg() and recvmsg() socket
operations. This is what will enable the implementation of higher-layer
functions like equivalents of sctp_recvmsg() and sctp_send() in
libosmo-netif and/or other users.
Change-Id: I89eb519b22d21011d61a7855b2364bc3c295df82
Related: OS#5751
This relocation is necessary as the backend (osmo_io_fd or
osmo_io_uring) requires a different approach in handling connect
notifications. As a result, a function call has been introduced to
struct iofd_backend_ops.
In a subsequent patch, the process for the osmo_io_uring backend will
be modified to handle SCTP connect notifications using poll/select.
If connect notification is requested using poll/select, the file
descriptior must be registered to osmo_fd, using osmo_fd_register. If
read / write notification is requested by application, the file
descriptior must be registered also. A flag is used prevent calling
osmo_fd_register / osmo_fd_unregister multiple times, which would cause
a crash.
Change-Id: I905ec85210570aff8addadfc9603335d04eb057a
Related: OS#5751
Harald Welte
Andreas Eversberg
Vadim Yanitskiy
Daniel Willmann