The helper function exec_user_callback executes within the context of an
lwIP C callback, and the user (Python) callback to be scheduled may want to
perform further TCP/IP actions, so the latter should be scheduled to run
outside the lwIP context (otherwise it's effectively a "hard IRQ" and such
callbacks have lots of restrictions).
If tcp_write returns ERR_MEM then it's not a fatal error but instead means
the caller should retry the write later on (and this is what lwIP's netconn
API does).
This fixes problems where a TCP send would raise OSError(ENOMEM) in
situations where the TCP/IP stack is under heavy load. See eg issues #1897
and #1971.
In d5f0c87bb9 this call to tcp_poll() was
added to put a timeout on closing TCP sockets. But after calling
tcp_close() the PCB may be freed and therefore invalid, so tcp_poll() can
not be used at that point. As a fix this commit calls tcp_poll() before
closing the TCP PCB. If the PCB is subsequently closed and freed by
tcp_close() or tcp_abort() then the PCB will not be on any active list and
the callback will not be executed, which is the desired behaviour (the
_lwip_tcp_close_poll() callback only needs to be called if the PCB remains
active for longer than the timeout).
Commit 2848a613ac introduced a bug where
lwip_socket_free_incoming() accessed pcb.tcp->state after the PCB was
closed. The state may have changed due to that close call, or the PCB may
be freed and therefore invalid. This commit fixes that by calling
lwip_socket_free_incoming() before the PCB is closed.
Since commit da938a83b5 the tcp_arg() that is
set for the new connection is the new connection itself, and the parent
listening socket is found in the pcb->connected entry.
Some users of this module may require the LwIP stack to run at an elevated
priority, to protect against concurrency issues with processing done by the
underlying network interface. Since LwIP doesn't provide such protection
it must be done here (the other option is to run LwIP in a separate thread,
and use thread protection mechanisms, but that is a more heavyweight
solution).
The bug polling for readability was: if alloc==0 and tcp.item==NULL then
the code would incorrectly check tcp.array[iget] which is an invalid
dereference when alloc==0. This patch refactors the code to use a helper
function lwip_socket_incoming_array() to return the correct pointer for the
incomming connection array.
Fixes issue #4511.
Python defines warnings as belonging to categories, where category is a
warning type (descending from exception type). This is useful, as e.g.
allows to disable warnings selectively and provide user-defined warning
types. So, implement this in MicroPython, except that categories are
represented just with strings. However, enough hooks are left to implement
categories differently per-port (e.g. as types), without need to patch each
and every usage.
The recent implementation of the listen backlog meant that the logic to
test for readability of such a socket changed, and this commit updates the
logic to work again.
Array to hold waiting connections is in-place if backlog=1, else is a
dynamically allocated array. Incoming connections are processed FIFO
style to maintain fairness.
Otherwise they may be called on a socket that no longer exists.
For example, if the GC calls the finaliser on the socket and then reuses
its heap memory, the "callback" entry of the old socket may contain invalid
data. If lwIP then calls the TCP callback the code may try to call the
user callback object which is now invalid. The lwIP callbacks must be
deregistered during the closing of the socket, before all the pcb pointers
are set to NULL.
This matches CPython behaviour on Linux: a socket that is new and not
listening or connected is considered "hung up".
Thanks to @rkojedzinszky for the initial patch, PR #3457.
Damien George
Paul Sokolovsky