2016-10-06 08:24:08 +02:00
|
|
|
// Module for network
|
|
|
|
|
|
|
|
|
|
#include "module.h"
|
|
|
|
|
#include "lauxlib.h"
|
|
|
|
|
#include "platform.h"
|
|
|
|
|
#include "lmem.h"
|
|
|
|
|
#include "ip_fmt.h"
|
|
|
|
|
#include "task/task.h"
|
|
|
|
|
|
|
|
|
|
#include <string.h>
|
|
|
|
|
#include <stdlib.h>
|
|
|
|
|
#include <string.h>
|
|
|
|
|
|
2017-07-06 23:55:32 +02:00
|
|
|
#include "lwip/api.h"
|
2016-10-06 08:24:08 +02:00
|
|
|
#include "lwip/err.h"
|
|
|
|
|
#include "lwip/ip_addr.h"
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
#include "lwip/dns.h"
|
2016-10-06 08:24:08 +02:00
|
|
|
#include "lwip/tcp.h"
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
#include "lwip/inet.h"
|
|
|
|
|
#include "lwip/igmp.h"
|
2017-07-06 23:55:32 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
#include <esp_log.h>
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
// Some LWIP macros cause complaints with ptr NULL checks, so shut them off :(
|
|
|
|
|
#pragma GCC diagnostic ignored "-Waddress"
|
|
|
|
|
|
|
|
|
|
typedef enum net_type {
|
|
|
|
|
TYPE_TCP_SERVER = 0,
|
|
|
|
|
TYPE_TCP_CLIENT,
|
|
|
|
|
TYPE_UDP_SOCKET
|
|
|
|
|
} net_type;
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static const char *NET_TABLES[] = {
|
2016-10-06 08:24:08 +02:00
|
|
|
"net.tcpserver",
|
|
|
|
|
"net.tcpsocket",
|
|
|
|
|
"net.udpsocket"
|
|
|
|
|
};
|
|
|
|
|
#define NET_TABLE_TCP_SERVER NET_TABLES[0]
|
|
|
|
|
#define NET_TABLE_TCP_CLIENT NET_TABLES[1]
|
|
|
|
|
#define NET_TABLE_UDP_SOCKET NET_TABLES[2]
|
|
|
|
|
|
|
|
|
|
#define TYPE_TCP TYPE_TCP_CLIENT
|
|
|
|
|
#define TYPE_UDP TYPE_UDP_SOCKET
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
|
|
|
|
static task_handle_t net_handler;
|
|
|
|
|
static task_handle_t dns_handler;
|
|
|
|
|
|
|
|
|
|
typedef enum {
|
|
|
|
|
EVT_SENT, EVT_RECV, EVT_ZEROREAD, EVT_CONNECTED, EVT_ERR
|
|
|
|
|
} bounce_event_t;
|
|
|
|
|
|
|
|
|
|
typedef struct netconn_bounce_event {
|
|
|
|
|
struct netconn *netconn;
|
|
|
|
|
bounce_event_t event;
|
|
|
|
|
err_t err;
|
|
|
|
|
uint16_t len;
|
|
|
|
|
} netconn_bounce_event_t;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
struct lnet_userdata;
|
|
|
|
|
|
|
|
|
|
typedef struct dns_event {
|
|
|
|
|
ip_addr_t resolved_ip;
|
|
|
|
|
enum { DNS_STATIC, DNS_SOCKET } invoker;
|
|
|
|
|
union {
|
|
|
|
|
int cb_ref;
|
|
|
|
|
struct lnet_userdata *ud;
|
|
|
|
|
};
|
|
|
|
|
} dns_event_t;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
// --- lwIP callbacks -----------------------------------------------------
|
|
|
|
|
|
|
|
|
|
// Caution - these run in a different RTOS thread!
|
|
|
|
|
|
|
|
|
|
static void lnet_netconn_callback(struct netconn *netconn, enum netconn_evt evt, u16_t len)
|
|
|
|
|
{
|
|
|
|
|
if (!netconn) return;
|
|
|
|
|
|
|
|
|
|
switch (evt)
|
|
|
|
|
{
|
|
|
|
|
case NETCONN_EVT_SENDPLUS:
|
|
|
|
|
case NETCONN_EVT_RCVPLUS:
|
|
|
|
|
case NETCONN_EVT_ERROR: break;
|
|
|
|
|
default: return; // we don't care about minus events
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
netconn_bounce_event_t *nbe = malloc(sizeof(netconn_bounce_event_t));
|
|
|
|
|
if (!nbe)
|
|
|
|
|
{
|
|
|
|
|
ESP_LOGE("net", "out of memory - lost event notification!");
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
nbe->netconn = netconn;
|
|
|
|
|
nbe->len = len;
|
|
|
|
|
nbe->err = ERR_OK;
|
|
|
|
|
task_prio_t prio = TASK_PRIORITY_MEDIUM;
|
|
|
|
|
switch (evt)
|
|
|
|
|
{
|
|
|
|
|
case NETCONN_EVT_SENDPLUS:
|
|
|
|
|
nbe->event = (len == 0) ? EVT_CONNECTED : EVT_SENT; break;
|
|
|
|
|
case NETCONN_EVT_RCVPLUS:
|
|
|
|
|
nbe->event = (len == 0) ? EVT_ZEROREAD : EVT_RECV;
|
|
|
|
|
prio = TASK_PRIORITY_HIGH;
|
|
|
|
|
break;
|
|
|
|
|
case NETCONN_EVT_ERROR:
|
|
|
|
|
nbe->event = EVT_ERR;
|
|
|
|
|
nbe->err = netconn_err(netconn); // is this safe in this thread?
|
|
|
|
|
break;
|
|
|
|
|
default: break; // impossible
|
|
|
|
|
}
|
|
|
|
|
if (!task_post(prio, net_handler, (task_param_t)nbe))
|
|
|
|
|
ESP_LOGE("net", "out of task post slots - lost data!");
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
// This one may *also* run in the LVM thread
|
|
|
|
|
static void net_dns_cb(const char *name, const ip_addr_t *ipaddr, void *callback_arg) {
|
|
|
|
|
dns_event_t *ev = (dns_event_t *)callback_arg;
|
|
|
|
|
|
|
|
|
|
ev->resolved_ip = ipaddr ? *ipaddr : ip_addr_any;
|
|
|
|
|
|
|
|
|
|
if (!task_post_medium(dns_handler, (task_param_t)ev))
|
|
|
|
|
{
|
|
|
|
|
ESP_LOGE("net", "out of task post slots - lost DNS response for %s", name);
|
|
|
|
|
// And we also leaked the ev here :(
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
// --- active list (out of reach of the lwIP thread) ---------------------
|
|
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
typedef struct lnet_userdata {
|
|
|
|
|
enum net_type type;
|
|
|
|
|
int self_ref;
|
2017-07-06 23:55:32 +02:00
|
|
|
uint16_t port;
|
|
|
|
|
struct netconn *netconn;
|
2017-07-20 21:23:38 +02:00
|
|
|
bool closing;
|
2016-10-06 08:24:08 +02:00
|
|
|
union {
|
|
|
|
|
struct {
|
|
|
|
|
int cb_accept_ref;
|
2016-12-30 05:39:07 +01:00
|
|
|
int timeout;
|
2016-10-06 08:24:08 +02:00
|
|
|
} server;
|
|
|
|
|
struct {
|
|
|
|
|
int wait_dns;
|
|
|
|
|
int cb_dns_ref;
|
|
|
|
|
int cb_receive_ref;
|
|
|
|
|
int cb_sent_ref;
|
|
|
|
|
// Only for TCP:
|
2017-07-06 23:55:32 +02:00
|
|
|
bool connecting;
|
2017-11-12 17:41:29 +01:00
|
|
|
size_t num_send;
|
2016-10-06 08:24:08 +02:00
|
|
|
int cb_connect_ref;
|
|
|
|
|
int cb_disconnect_ref;
|
|
|
|
|
int cb_reconnect_ref;
|
|
|
|
|
} client;
|
|
|
|
|
};
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
|
|
|
|
struct lnet_userdata *next;
|
2016-10-06 08:24:08 +02:00
|
|
|
} lnet_userdata;
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static lnet_userdata *active;
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static void add_userdata_to_active(lnet_userdata *ud)
|
|
|
|
|
{
|
|
|
|
|
ud->next = active;
|
|
|
|
|
active = ud;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static void remove_userdata_from_active(lnet_userdata *ud)
|
|
|
|
|
{
|
|
|
|
|
if (!ud)
|
|
|
|
|
return;
|
|
|
|
|
for (lnet_userdata **i = &active; *i; i = &(*i)->next)
|
|
|
|
|
{
|
|
|
|
|
if (*i == ud)
|
|
|
|
|
{
|
|
|
|
|
*i = ud->next;
|
|
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
static lnet_userdata *userdata_from_netconn(struct netconn *nc)
|
|
|
|
|
{
|
|
|
|
|
for (lnet_userdata *i = active; i; i = i->next)
|
|
|
|
|
if (i->netconn == nc)
|
|
|
|
|
return i;
|
|
|
|
|
return NULL;
|
|
|
|
|
}
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
// --- Forward declarations ----------------------------------------------
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static void lsent_cb (lua_State *L, lnet_userdata *ud);
|
|
|
|
|
static void lrecv_cb (lua_State *L, lnet_userdata *ud);
|
|
|
|
|
static void laccept_cb (lua_State *L, lnet_userdata *ud);
|
|
|
|
|
static void lconnected_cb (lua_State *L, lnet_userdata *ud);
|
|
|
|
|
static void lerr_cb (lua_State *L, lnet_userdata *ud, err_t err);
|
|
|
|
|
static void ldnsstatic_cb (lua_State *L, int cb_ref, ip_addr_t *addr);
|
|
|
|
|
static void ldnsfound_cb (lua_State *L, lnet_userdata *ud, ip_addr_t *addr);
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
// --- Helper functions --------------------------------------------------
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
int lwip_lua_checkerr (lua_State *L, err_t err) {
|
|
|
|
|
switch (err) {
|
|
|
|
|
case ERR_OK: return 0;
|
|
|
|
|
case ERR_MEM: return luaL_error(L, "out of memory");
|
|
|
|
|
case ERR_BUF: return luaL_error(L, "buffer error");
|
|
|
|
|
case ERR_TIMEOUT: return luaL_error(L, "timeout");
|
|
|
|
|
case ERR_RTE: return luaL_error(L, "routing problem");
|
|
|
|
|
case ERR_INPROGRESS: return luaL_error(L, "in progress");
|
|
|
|
|
case ERR_VAL: return luaL_error(L, "illegal value");
|
|
|
|
|
case ERR_WOULDBLOCK: return luaL_error(L, "would block");
|
|
|
|
|
case ERR_ABRT: return luaL_error(L, "connection aborted");
|
|
|
|
|
case ERR_RST: return luaL_error(L, "connection reset");
|
|
|
|
|
case ERR_CLSD: return luaL_error(L, "connection closed");
|
|
|
|
|
case ERR_CONN: return luaL_error(L, "not connected");
|
|
|
|
|
case ERR_ARG: return luaL_error(L, "illegal argument");
|
|
|
|
|
case ERR_USE: return luaL_error(L, "address in use");
|
|
|
|
|
case ERR_IF: return luaL_error(L, "netif error");
|
|
|
|
|
case ERR_ISCONN: return luaL_error(L, "already connected");
|
|
|
|
|
default: return luaL_error(L, "unknown error");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
lnet_userdata *net_create( lua_State *L, enum net_type type ) {
|
|
|
|
|
const char *mt = NET_TABLES[type];
|
|
|
|
|
lnet_userdata *ud = (lnet_userdata *)lua_newuserdata(L, sizeof(lnet_userdata));
|
|
|
|
|
if (!ud) return NULL;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
luaL_getmetatable(L, mt);
|
|
|
|
|
lua_setmetatable(L, -2);
|
|
|
|
|
|
|
|
|
|
ud->type = type;
|
|
|
|
|
ud->self_ref = LUA_NOREF;
|
2017-07-06 23:55:32 +02:00
|
|
|
ud->netconn = NULL;
|
2017-07-20 21:23:38 +02:00
|
|
|
ud->closing = false;
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
switch (type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
ud->client.cb_connect_ref = LUA_NOREF;
|
|
|
|
|
ud->client.cb_reconnect_ref = LUA_NOREF;
|
|
|
|
|
ud->client.cb_disconnect_ref = LUA_NOREF;
|
2017-07-06 23:55:32 +02:00
|
|
|
ud->client.connecting = false;
|
2021-07-21 04:30:51 +02:00
|
|
|
// fall-through
|
2016-10-06 08:24:08 +02:00
|
|
|
case TYPE_UDP_SOCKET:
|
|
|
|
|
ud->client.wait_dns = 0;
|
|
|
|
|
ud->client.cb_dns_ref = LUA_NOREF;
|
|
|
|
|
ud->client.cb_receive_ref = LUA_NOREF;
|
|
|
|
|
ud->client.cb_sent_ref = LUA_NOREF;
|
|
|
|
|
break;
|
|
|
|
|
case TYPE_TCP_SERVER:
|
|
|
|
|
ud->server.cb_accept_ref = LUA_NOREF;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
add_userdata_to_active(ud);
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
return ud;
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
// --- LVM thread task handlers -----------------------------------------
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
void handle_net_event(task_param_t param, task_prio_t prio)
|
2016-10-06 08:24:08 +02:00
|
|
|
{
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
(void)prio;
|
|
|
|
|
netconn_bounce_event_t *pnbe = (netconn_bounce_event_t *)param;
|
|
|
|
|
netconn_bounce_event_t nbe = *pnbe;
|
|
|
|
|
free(pnbe); // free before we can hit any luaL_error()s
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
lnet_userdata *ud = userdata_from_netconn(nbe.netconn);
|
|
|
|
|
if (!ud)
|
|
|
|
|
return;
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
if (nbe.event == EVT_ERR)
|
2016-10-06 08:24:08 +02:00
|
|
|
{
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
lerr_cb(lua_getstate(), ud, nbe.err);
|
2017-07-06 23:55:32 +02:00
|
|
|
return;
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
if (ud->type == TYPE_TCP_CLIENT || ud->type == TYPE_UDP_SOCKET)
|
|
|
|
|
{
|
|
|
|
|
switch (nbe.event)
|
|
|
|
|
{
|
|
|
|
|
case EVT_CONNECTED:
|
|
|
|
|
if (ud->type == TYPE_TCP_CLIENT && ud->client.connecting)
|
|
|
|
|
{
|
|
|
|
|
ud->client.connecting = false;
|
|
|
|
|
lconnected_cb(lua_getstate(), ud);
|
2017-11-12 17:41:29 +01:00
|
|
|
}
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
break;
|
|
|
|
|
case EVT_SENT:
|
|
|
|
|
ud->client.num_send -= nbe.len;
|
|
|
|
|
if (ud->client.num_send == 0) // all data sent, trigger Lua callback
|
|
|
|
|
lsent_cb(lua_getstate(), ud);
|
|
|
|
|
break;
|
|
|
|
|
case EVT_RECV:
|
|
|
|
|
if (nbe.len > 0)
|
|
|
|
|
lrecv_cb(lua_getstate(), ud);
|
|
|
|
|
break;
|
|
|
|
|
case EVT_ZEROREAD: // fall-through
|
|
|
|
|
case EVT_ERR:
|
2017-07-20 21:23:38 +02:00
|
|
|
ud->closing = true;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
lerr_cb(lua_getstate(), ud, nbe.err);
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
else if (ud->type == TYPE_TCP_SERVER)
|
|
|
|
|
{
|
|
|
|
|
switch (nbe.event)
|
|
|
|
|
{
|
|
|
|
|
case EVT_ZEROREAD: // new connection
|
|
|
|
|
laccept_cb (lua_getstate(), ud);
|
|
|
|
|
break;
|
|
|
|
|
case EVT_ERR:
|
|
|
|
|
ud->closing = true;
|
|
|
|
|
lerr_cb(lua_getstate(), ud, nbe.err);
|
|
|
|
|
break;
|
|
|
|
|
default: break;
|
2017-07-06 23:55:32 +02:00
|
|
|
}
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
}
|
|
|
|
|
}
|
2016-10-06 08:24:08 +02:00
|
|
|
|
2017-07-06 23:55:32 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
void handle_dns_event(task_param_t param, task_prio_t prio)
|
|
|
|
|
{
|
|
|
|
|
(void)prio;
|
|
|
|
|
dns_event_t *pev = (dns_event_t *)param;
|
|
|
|
|
dns_event_t ev = *pev;
|
2017-07-06 23:55:32 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
lua_State *L = lua_getstate();
|
|
|
|
|
luaM_free(L, pev); // free before we can hit any luaL_error()s
|
2017-07-06 23:55:32 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
switch (ev.invoker)
|
|
|
|
|
{
|
|
|
|
|
case DNS_STATIC:
|
|
|
|
|
ldnsstatic_cb(L, ev.cb_ref, &ev.resolved_ip);
|
2017-07-06 23:55:32 +02:00
|
|
|
break;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
case DNS_SOCKET:
|
|
|
|
|
ldnsfound_cb(L, ev.ud, &ev.resolved_ip);
|
2017-07-06 23:55:32 +02:00
|
|
|
}
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
2017-07-06 23:55:32 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static err_t netconn_close_and_delete(struct netconn *conn) {
|
2017-08-14 13:54:37 +02:00
|
|
|
err_t err = netconn_close(conn);
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
if (err == ERR_OK)
|
2017-09-12 22:20:26 +02:00
|
|
|
netconn_delete(conn);
|
2017-08-14 13:54:37 +02:00
|
|
|
|
|
|
|
|
return err;
|
2017-07-06 23:55:32 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
// --- Lua API - create
|
|
|
|
|
|
|
|
|
|
// Lua: net.createUDPSocket()
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_createUDPSocket( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
net_create(L, TYPE_UDP_SOCKET);
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: net.createServer(type, timeout)
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_createServer( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
int type, timeout;
|
|
|
|
|
|
|
|
|
|
type = luaL_optlong(L, 1, TYPE_TCP);
|
|
|
|
|
timeout = luaL_optlong(L, 2, 30);
|
|
|
|
|
|
|
|
|
|
if (type == TYPE_UDP) return net_createUDPSocket( L );
|
|
|
|
|
if (type != TYPE_TCP) return luaL_error(L, "invalid type");
|
|
|
|
|
|
2016-12-30 05:39:07 +01:00
|
|
|
lnet_userdata *ud = net_create(L, TYPE_TCP_SERVER);
|
|
|
|
|
ud->server.timeout = timeout;
|
2016-10-06 08:24:08 +02:00
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: net.createConnection(type, secure)
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_createConnection( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
int type, secure;
|
|
|
|
|
|
|
|
|
|
type = luaL_optlong(L, 1, TYPE_TCP);
|
|
|
|
|
secure = luaL_optlong(L, 2, 0);
|
|
|
|
|
|
|
|
|
|
if (type == TYPE_UDP) return net_createUDPSocket( L );
|
|
|
|
|
if (type != TYPE_TCP) return luaL_error(L, "invalid type");
|
|
|
|
|
if (secure) {
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
return luaL_error(L, "secure connections not yet supported");
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
net_create(L, TYPE_TCP_CLIENT);
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// --- Get & check userdata
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static lnet_userdata *net_get_udata_s( lua_State *L, int stack ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
if (!lua_isuserdata(L, stack)) return NULL;
|
|
|
|
|
lnet_userdata *ud = (lnet_userdata *)lua_touserdata(L, stack);
|
|
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
case TYPE_TCP_SERVER:
|
|
|
|
|
case TYPE_UDP_SOCKET:
|
|
|
|
|
break;
|
|
|
|
|
default: return NULL;
|
|
|
|
|
}
|
|
|
|
|
const char *mt = NET_TABLES[ud->type];
|
|
|
|
|
ud = luaL_checkudata(L, stack, mt);
|
|
|
|
|
return ud;
|
|
|
|
|
}
|
|
|
|
|
#define net_get_udata(L) net_get_udata_s(L, 1)
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
|
|
|
|
static int lnet_socket_resolve_dns(lua_State *L, lnet_userdata *ud, const char * domain, bool close_on_err)
|
|
|
|
|
{
|
|
|
|
|
ud->client.wait_dns ++;
|
|
|
|
|
int unref = 0;
|
|
|
|
|
if (ud->self_ref == LUA_NOREF) {
|
|
|
|
|
unref = 1;
|
|
|
|
|
lua_pushvalue(L, 1);
|
|
|
|
|
ud->self_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
dns_event_t *ev = luaM_new(L, dns_event_t);
|
|
|
|
|
if (!ev)
|
|
|
|
|
return luaL_error (L, "out of memory");
|
|
|
|
|
|
|
|
|
|
ev->invoker = DNS_SOCKET;
|
|
|
|
|
ev->ud = ud;
|
|
|
|
|
|
|
|
|
|
err_t err = dns_gethostbyname(domain, &ev->resolved_ip, net_dns_cb, ev);
|
|
|
|
|
if (err == ERR_OK)
|
|
|
|
|
net_dns_cb(domain, &ev->resolved_ip, ev);
|
|
|
|
|
else if (err != ERR_INPROGRESS)
|
|
|
|
|
{
|
|
|
|
|
luaM_free(L, ev);
|
|
|
|
|
ud->client.wait_dns --;
|
|
|
|
|
if (unref) {
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
ud->self_ref = LUA_NOREF;
|
|
|
|
|
}
|
|
|
|
|
if (close_on_err)
|
|
|
|
|
{
|
|
|
|
|
ud->closing = true;
|
|
|
|
|
if (netconn_close_and_delete(ud->netconn) == ERR_OK)
|
|
|
|
|
ud->netconn = NULL;
|
|
|
|
|
}
|
|
|
|
|
return lwip_lua_checkerr(L, err);
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
// --- Lua API
|
|
|
|
|
|
|
|
|
|
// Lua: server:listen(port, addr, function(c)), socket:listen(port, addr)
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_listen( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud || ud->type == TYPE_TCP_CLIENT)
|
|
|
|
|
return luaL_error(L, "invalid user data");
|
2017-07-06 23:55:32 +02:00
|
|
|
if (ud->netconn)
|
2016-10-06 08:24:08 +02:00
|
|
|
return luaL_error(L, "already listening");
|
|
|
|
|
int stack = 2;
|
|
|
|
|
uint16_t port = 0;
|
|
|
|
|
const char *domain = "0.0.0.0";
|
|
|
|
|
if (lua_isnumber(L, stack))
|
|
|
|
|
port = lua_tointeger(L, stack++);
|
|
|
|
|
if (lua_isstring(L, stack)) {
|
|
|
|
|
size_t dl = 0;
|
|
|
|
|
domain = luaL_checklstring(L, stack++, &dl);
|
|
|
|
|
}
|
|
|
|
|
ip_addr_t addr;
|
|
|
|
|
if (!ipaddr_aton(domain, &addr))
|
|
|
|
|
return luaL_error(L, "invalid IP address");
|
|
|
|
|
if (ud->type == TYPE_TCP_SERVER) {
|
2021-08-21 17:39:54 +02:00
|
|
|
if (lua_isfunction(L, stack)) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushvalue(L, stack++);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->server.cb_accept_ref);
|
|
|
|
|
ud->server.cb_accept_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
} else {
|
|
|
|
|
return luaL_error(L, "need callback");
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
err_t err = ERR_OK;
|
|
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_SERVER:
|
2017-07-06 23:55:32 +02:00
|
|
|
ud->netconn = netconn_new_with_callback(NETCONN_TCP, lnet_netconn_callback);
|
|
|
|
|
if (!ud->netconn)
|
|
|
|
|
return luaL_error(L, "cannot allocate netconn");
|
|
|
|
|
netconn_set_nonblocking(ud->netconn, 1);
|
|
|
|
|
|
|
|
|
|
err = netconn_bind(ud->netconn, &addr, port);
|
2016-10-06 08:24:08 +02:00
|
|
|
if (err == ERR_OK) {
|
2017-07-06 23:55:32 +02:00
|
|
|
err = netconn_listen(ud->netconn);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
break;
|
|
|
|
|
case TYPE_UDP_SOCKET:
|
2017-07-06 23:55:32 +02:00
|
|
|
ud->netconn = netconn_new_with_callback(NETCONN_UDP, lnet_netconn_callback);
|
|
|
|
|
if (!ud->netconn)
|
|
|
|
|
return luaL_error(L, "cannot allocate netconn");
|
|
|
|
|
netconn_set_nonblocking(ud->netconn, 1);
|
|
|
|
|
|
|
|
|
|
err = netconn_bind(ud->netconn, &addr, port);
|
2016-10-06 08:24:08 +02:00
|
|
|
break;
|
|
|
|
|
default: break;
|
|
|
|
|
}
|
|
|
|
|
if (err != ERR_OK) {
|
2017-07-20 21:23:38 +02:00
|
|
|
ud->closing = true;
|
2016-10-06 08:24:08 +02:00
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_SERVER:
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
if (netconn_close_and_delete(ud->netconn) == ERR_OK)
|
2017-08-14 13:54:37 +02:00
|
|
|
ud->netconn = NULL;
|
2016-10-06 08:24:08 +02:00
|
|
|
break;
|
|
|
|
|
case TYPE_UDP_SOCKET:
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
if (netconn_close_and_delete(ud->netconn) == ERR_OK)
|
2017-08-14 13:54:37 +02:00
|
|
|
ud->netconn = NULL;
|
2016-10-06 08:24:08 +02:00
|
|
|
break;
|
|
|
|
|
default: break;
|
|
|
|
|
}
|
|
|
|
|
return lwip_lua_checkerr(L, err);
|
|
|
|
|
}
|
|
|
|
|
if (ud->self_ref == LUA_NOREF) {
|
|
|
|
|
lua_pushvalue(L, 1);
|
|
|
|
|
ud->self_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client:connect(port, addr)
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_connect( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud || ud->type != TYPE_TCP_CLIENT)
|
|
|
|
|
return luaL_error(L, "invalid user data");
|
2017-07-06 23:55:32 +02:00
|
|
|
if (ud->netconn)
|
2016-10-06 08:24:08 +02:00
|
|
|
return luaL_error(L, "already connected");
|
|
|
|
|
uint16_t port = luaL_checkinteger(L, 2);
|
|
|
|
|
if (port == 0)
|
|
|
|
|
return luaL_error(L, "specify port");
|
|
|
|
|
const char *domain = "127.0.0.1";
|
|
|
|
|
if (lua_isstring(L, 3)) {
|
|
|
|
|
size_t dl = 0;
|
|
|
|
|
domain = luaL_checklstring(L, 3, &dl);
|
|
|
|
|
}
|
2017-07-06 23:55:32 +02:00
|
|
|
|
|
|
|
|
ud->netconn = netconn_new_with_callback(NETCONN_TCP, lnet_netconn_callback);
|
|
|
|
|
if (!ud->netconn)
|
|
|
|
|
return luaL_error(L, "cannot allocate netconn");
|
|
|
|
|
netconn_set_nonblocking(ud->netconn, 1);
|
|
|
|
|
ud->port = port;
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
return lnet_socket_resolve_dns(L, ud, domain, true);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client/socket:on(name, callback)
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_on( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud || ud->type == TYPE_TCP_SERVER)
|
|
|
|
|
return luaL_error(L, "invalid user data");
|
|
|
|
|
int *refptr = NULL;
|
|
|
|
|
const char *name = luaL_checkstring(L, 2);
|
|
|
|
|
if (!name) return luaL_error(L, "need callback name");
|
|
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
if (strcmp("connection",name)==0)
|
|
|
|
|
{ refptr = &ud->client.cb_connect_ref; break; }
|
|
|
|
|
if (strcmp("disconnection",name)==0)
|
|
|
|
|
{ refptr = &ud->client.cb_disconnect_ref; break; }
|
|
|
|
|
if (strcmp("reconnection",name)==0)
|
|
|
|
|
{ refptr = &ud->client.cb_reconnect_ref; break; }
|
2021-07-21 04:30:51 +02:00
|
|
|
// fall-through
|
2016-10-06 08:24:08 +02:00
|
|
|
case TYPE_UDP_SOCKET:
|
|
|
|
|
if (strcmp("dns",name)==0)
|
|
|
|
|
{ refptr = &ud->client.cb_dns_ref; break; }
|
|
|
|
|
if (strcmp("receive",name)==0)
|
|
|
|
|
{ refptr = &ud->client.cb_receive_ref; break; }
|
|
|
|
|
if (strcmp("sent",name)==0)
|
|
|
|
|
{ refptr = &ud->client.cb_sent_ref; break; }
|
|
|
|
|
break;
|
|
|
|
|
default: return luaL_error(L, "invalid user data");
|
|
|
|
|
}
|
|
|
|
|
if (refptr == NULL)
|
|
|
|
|
return luaL_error(L, "invalid callback name");
|
2021-08-21 17:39:54 +02:00
|
|
|
if (lua_isfunction(L, 3)) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushvalue(L, 3);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, *refptr);
|
|
|
|
|
*refptr = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
} else if (lua_isnil(L, 3)) {
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, *refptr);
|
|
|
|
|
*refptr = LUA_NOREF;
|
|
|
|
|
} else {
|
|
|
|
|
return luaL_error(L, "invalid callback function");
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client:send(data, function(c)), socket:send(port, ip, data, function(s))
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_send( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud || ud->type == TYPE_TCP_SERVER)
|
|
|
|
|
return luaL_error(L, "invalid user data");
|
|
|
|
|
ip_addr_t addr;
|
|
|
|
|
uint16_t port = 0;
|
|
|
|
|
const char *data;
|
|
|
|
|
size_t datalen = 0;
|
|
|
|
|
int stack = 2;
|
|
|
|
|
if (ud->type == TYPE_UDP_SOCKET) {
|
|
|
|
|
size_t dl = 0;
|
|
|
|
|
port = luaL_checkinteger(L, stack++);
|
|
|
|
|
if (port == 0) return luaL_error(L, "need port");
|
|
|
|
|
const char *domain = luaL_checklstring(L, stack++, &dl);
|
|
|
|
|
if (!domain) return luaL_error(L, "need IP address");
|
|
|
|
|
if (!ipaddr_aton(domain, &addr)) return luaL_error(L, "invalid IP address");
|
|
|
|
|
}
|
|
|
|
|
data = luaL_checklstring(L, stack++, &datalen);
|
|
|
|
|
if (!data || datalen == 0) return luaL_error(L, "no data to send");
|
2017-11-12 17:41:29 +01:00
|
|
|
ud->client.num_send = datalen;
|
2021-08-21 17:39:54 +02:00
|
|
|
if (lua_isfunction(L, stack)) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushvalue(L, stack++);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_sent_ref);
|
|
|
|
|
ud->client.cb_sent_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
}
|
2017-07-06 23:55:32 +02:00
|
|
|
|
|
|
|
|
if (ud->type == TYPE_UDP_SOCKET && !ud->netconn) {
|
|
|
|
|
ud->netconn = netconn_new_with_callback(NETCONN_UDP, lnet_netconn_callback);
|
|
|
|
|
if (!ud->netconn)
|
|
|
|
|
return luaL_error(L, "cannot allocate netconn");
|
|
|
|
|
err_t err = netconn_bind(ud->netconn, IP_ADDR_ANY, 0);
|
2016-10-06 08:24:08 +02:00
|
|
|
if (err != ERR_OK) {
|
2017-07-20 21:23:38 +02:00
|
|
|
ud->closing = true;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
if (netconn_close_and_delete(ud->netconn) == ERR_OK)
|
2017-08-14 13:54:37 +02:00
|
|
|
ud->netconn = NULL;
|
2016-10-06 08:24:08 +02:00
|
|
|
return lwip_lua_checkerr(L, err);
|
|
|
|
|
}
|
|
|
|
|
if (ud->self_ref == LUA_NOREF) {
|
|
|
|
|
lua_pushvalue(L, 1);
|
|
|
|
|
ud->self_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
}
|
|
|
|
|
}
|
2017-07-06 23:55:32 +02:00
|
|
|
|
|
|
|
|
if (!ud->netconn || ud->self_ref == LUA_NOREF)
|
2016-10-06 08:24:08 +02:00
|
|
|
return luaL_error(L, "not connected");
|
2017-07-06 23:55:32 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
err_t err;
|
|
|
|
|
if (ud->type == TYPE_UDP_SOCKET) {
|
2017-07-06 23:55:32 +02:00
|
|
|
struct netbuf *buf = netbuf_new();
|
|
|
|
|
if (!buf || !netbuf_alloc(buf, datalen))
|
2016-10-06 08:24:08 +02:00
|
|
|
return luaL_error(L, "cannot allocate message buffer");
|
2017-07-06 23:55:32 +02:00
|
|
|
netbuf_take(buf, data, datalen);
|
|
|
|
|
err = netconn_sendto(ud->netconn, buf, &addr, port);
|
|
|
|
|
netbuf_delete(buf);
|
|
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
if (ud->client.cb_sent_ref != LUA_NOREF) {
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->client.cb_sent_ref);
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
lua_call(L, 1, 0);
|
|
|
|
|
}
|
|
|
|
|
} else if (ud->type == TYPE_TCP_CLIENT) {
|
2017-07-06 23:55:32 +02:00
|
|
|
size_t bytes_written;
|
|
|
|
|
err = netconn_write_partly(ud->netconn, data, datalen, NETCONN_COPY, &bytes_written);
|
2017-07-21 22:50:36 +02:00
|
|
|
if (err == ERR_OK && (datalen != bytes_written)) {
|
|
|
|
|
// the string object is potentially gc'ed after net_send finishes and we can't ensure
|
|
|
|
|
// integrity of the data pointer to netconn -> signal error to Lua layer
|
|
|
|
|
err = ERR_BUF;
|
|
|
|
|
}
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
else {
|
|
|
|
|
err = ERR_VAL;
|
|
|
|
|
}
|
|
|
|
|
return lwip_lua_checkerr(L, err);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client/socket:dns(domain, callback(socket, addr))
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_dns( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud || ud->type == TYPE_TCP_SERVER)
|
|
|
|
|
return luaL_error(L, "invalid user data");
|
|
|
|
|
size_t dl = 0;
|
|
|
|
|
const char *domain = luaL_checklstring(L, 2, &dl);
|
|
|
|
|
if (!domain)
|
|
|
|
|
return luaL_error(L, "no domain specified");
|
2021-08-21 17:39:54 +02:00
|
|
|
if (lua_isfunction(L, 3)) {
|
2016-10-06 08:24:08 +02:00
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_dns_ref);
|
|
|
|
|
lua_pushvalue(L, 3);
|
|
|
|
|
ud->client.cb_dns_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
}
|
|
|
|
|
if (ud->client.cb_dns_ref == LUA_NOREF)
|
|
|
|
|
return luaL_error(L, "no callback specified");
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
|
|
|
|
return lnet_socket_resolve_dns(L, ud, domain, false);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client:getpeer()
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_getpeer( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud || ud->type != TYPE_TCP_CLIENT)
|
|
|
|
|
return luaL_error(L, "invalid user data");
|
2017-07-06 23:55:32 +02:00
|
|
|
if (!ud->netconn) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushnil(L);
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
return 2;
|
|
|
|
|
}
|
2017-07-06 23:55:32 +02:00
|
|
|
|
|
|
|
|
uint16_t port;
|
|
|
|
|
ip_addr_t addr;
|
|
|
|
|
netconn_peer(ud->netconn, &addr, &port);
|
|
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
if (port == 0) {
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
return 2;
|
|
|
|
|
}
|
|
|
|
|
char addr_str[IP_STR_SZ];
|
|
|
|
|
ipstr (addr_str, &addr);
|
|
|
|
|
lua_pushinteger(L, port);
|
|
|
|
|
lua_pushstring(L, addr_str);
|
|
|
|
|
return 2;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client/server/socket:getaddr()
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_getaddr( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud) return luaL_error(L, "invalid user data");
|
2017-07-06 23:55:32 +02:00
|
|
|
if (!ud->netconn) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushnil(L);
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
return 2;
|
|
|
|
|
}
|
|
|
|
|
uint16_t port = 0;
|
|
|
|
|
ip_addr_t addr;
|
|
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
case TYPE_TCP_SERVER:
|
|
|
|
|
case TYPE_UDP_SOCKET:
|
2017-07-06 23:55:32 +02:00
|
|
|
netconn_addr(ud->netconn, &addr, &port);
|
2016-10-06 08:24:08 +02:00
|
|
|
break;
|
2017-07-06 23:55:32 +02:00
|
|
|
default: break;
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
if (port == 0) {
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
return 2;
|
|
|
|
|
}
|
|
|
|
|
char addr_str[IP_STR_SZ];
|
|
|
|
|
ipstr (addr_str, &addr);
|
|
|
|
|
lua_pushinteger(L, port);
|
|
|
|
|
lua_pushstring(L, addr_str);
|
|
|
|
|
return 2;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// Lua: client/server/socket:close()
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_close( lua_State *L ) {
|
2017-08-14 13:54:37 +02:00
|
|
|
err_t err = ERR_OK;
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud) return luaL_error(L, "invalid user data");
|
2017-07-06 23:55:32 +02:00
|
|
|
if (ud->netconn) {
|
2016-10-06 08:24:08 +02:00
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
case TYPE_TCP_SERVER:
|
|
|
|
|
case TYPE_UDP_SOCKET:
|
2017-07-20 21:23:38 +02:00
|
|
|
ud->closing = true;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
err = netconn_close_and_delete(ud->netconn);
|
2017-08-14 13:54:37 +02:00
|
|
|
if (err == ERR_OK)
|
|
|
|
|
ud->netconn = NULL;
|
2016-10-06 08:24:08 +02:00
|
|
|
break;
|
2017-07-06 23:55:32 +02:00
|
|
|
default: break;
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
} else {
|
|
|
|
|
return luaL_error(L, "not connected");
|
|
|
|
|
}
|
|
|
|
|
if (ud->type == TYPE_TCP_SERVER ||
|
2017-07-06 23:55:32 +02:00
|
|
|
(ud->netconn == NULL && ud->client.wait_dns == 0)) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_gc(L, LUA_GCSTOP, 0);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
ud->self_ref = LUA_NOREF;
|
|
|
|
|
lua_gc(L, LUA_GCRESTART, 0);
|
|
|
|
|
}
|
2017-08-14 13:54:37 +02:00
|
|
|
return lwip_lua_checkerr(L, err);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
static int net_delete( lua_State *L ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lnet_userdata *ud = net_get_udata(L);
|
|
|
|
|
if (!ud) return luaL_error(L, "no user data");
|
2017-07-06 23:55:32 +02:00
|
|
|
if (ud->netconn) {
|
2016-10-06 08:24:08 +02:00
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
case TYPE_TCP_SERVER:
|
|
|
|
|
case TYPE_UDP_SOCKET:
|
2017-07-20 21:23:38 +02:00
|
|
|
ud->closing = true;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
netconn_delete(ud->netconn);
|
2017-07-06 23:55:32 +02:00
|
|
|
ud->netconn = NULL;
|
2016-10-06 08:24:08 +02:00
|
|
|
break;
|
2017-07-06 23:55:32 +02:00
|
|
|
default: break;
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
switch (ud->type) {
|
|
|
|
|
case TYPE_TCP_CLIENT:
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_connect_ref);
|
|
|
|
|
ud->client.cb_connect_ref = LUA_NOREF;
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_disconnect_ref);
|
|
|
|
|
ud->client.cb_disconnect_ref = LUA_NOREF;
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_reconnect_ref);
|
|
|
|
|
ud->client.cb_reconnect_ref = LUA_NOREF;
|
2021-07-21 04:30:51 +02:00
|
|
|
// fall-through
|
2016-10-06 08:24:08 +02:00
|
|
|
case TYPE_UDP_SOCKET:
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_dns_ref);
|
|
|
|
|
ud->client.cb_dns_ref = LUA_NOREF;
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_receive_ref);
|
|
|
|
|
ud->client.cb_receive_ref = LUA_NOREF;
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->client.cb_sent_ref);
|
|
|
|
|
ud->client.cb_sent_ref = LUA_NOREF;
|
|
|
|
|
break;
|
|
|
|
|
case TYPE_TCP_SERVER:
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->server.cb_accept_ref);
|
|
|
|
|
ud->server.cb_accept_ref = LUA_NOREF;
|
|
|
|
|
break;
|
|
|
|
|
}
|
|
|
|
|
lua_gc(L, LUA_GCSTOP, 0);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
ud->self_ref = LUA_NOREF;
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
remove_userdata_from_active(ud);
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_gc(L, LUA_GCRESTART, 0);
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
// --- Multicast
|
|
|
|
|
|
|
|
|
|
static int net_multicastJoinLeave( lua_State *L, int join) {
|
|
|
|
|
size_t il;
|
|
|
|
|
ip4_addr_t multicast_addr;
|
|
|
|
|
ip4_addr_t if_addr;
|
|
|
|
|
const char *multicast_ip;
|
|
|
|
|
const char *if_ip;
|
|
|
|
|
|
|
|
|
|
NODE_DBG("net_multicastJoin is called.\n");
|
|
|
|
|
if(! lua_isstring(L,1) ) return luaL_error( L, "wrong arg type" );
|
|
|
|
|
if_ip = luaL_checklstring( L, 1, &il );
|
|
|
|
|
if (if_ip != NULL)
|
|
|
|
|
if ( if_ip[0] == '\0' || strcasecmp(if_ip,"any") == 0)
|
|
|
|
|
{
|
|
|
|
|
if_ip = "0.0.0.0";
|
|
|
|
|
il = 7;
|
|
|
|
|
}
|
|
|
|
|
if (if_ip == NULL || il > 15 || il < 7) return luaL_error( L, "invalid if ip" );
|
|
|
|
|
ip4addr_aton (if_ip, &if_addr);
|
|
|
|
|
|
|
|
|
|
if(! lua_isstring(L,2) ) return luaL_error( L, "wrong arg type" );
|
|
|
|
|
multicast_ip = luaL_checklstring( L, 2, &il );
|
|
|
|
|
if (multicast_ip == NULL || il > 15 || il < 7) return luaL_error( L, "invalid multicast ip" );
|
|
|
|
|
ip4addr_aton (multicast_ip, &multicast_addr);
|
|
|
|
|
if (join)
|
|
|
|
|
{
|
|
|
|
|
igmp_joingroup(&if_addr, &multicast_addr);
|
|
|
|
|
}
|
|
|
|
|
else
|
|
|
|
|
{
|
|
|
|
|
igmp_leavegroup(&if_addr, &multicast_addr);
|
|
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
// Lua: net.multicastJoin(ifip, multicastip)
|
|
|
|
|
// if ifip "" or "any" all interfaces are affected
|
|
|
|
|
static int net_multicastJoin( lua_State* L ) {
|
|
|
|
|
return net_multicastJoinLeave(L,1);
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
// Lua: net.multicastLeave(ifip, multicastip)
|
|
|
|
|
// if ifip "" or "any" all interfaces are affected
|
|
|
|
|
static int net_multicastLeave( lua_State* L ) {
|
|
|
|
|
return net_multicastJoinLeave(L,0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
// --- DNS ---------------------------------------------------------------
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
// Lua: net.dns.resolve( domain, function(ip) )
|
|
|
|
|
static int net_dns_static( lua_State* L ) {
|
|
|
|
|
size_t dl;
|
|
|
|
|
const char* domain = luaL_checklstring(L, 1, &dl);
|
|
|
|
|
if (!domain && dl > 128) {
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
return luaL_error(L, "domain name too long");
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
2021-08-21 17:39:54 +02:00
|
|
|
luaL_checkfunction(L, 2);
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
lua_settop(L, 2);
|
|
|
|
|
|
|
|
|
|
dns_event_t *ev = luaM_new(L, dns_event_t);
|
2016-10-06 08:24:08 +02:00
|
|
|
if (!ev)
|
|
|
|
|
return luaL_error (L, "out of memory");
|
|
|
|
|
|
|
|
|
|
ev->cb_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
ev->invoker = DNS_STATIC;
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
err_t err = dns_gethostbyname(domain, &ev->resolved_ip, net_dns_cb, ev);
|
|
|
|
|
if (err == ERR_OK)
|
|
|
|
|
net_dns_cb(domain, &ev->resolved_ip, ev);
|
|
|
|
|
else if (err != ERR_INPROGRESS)
|
|
|
|
|
{
|
|
|
|
|
luaM_free(L, ev);
|
|
|
|
|
lwip_lua_checkerr(L, err);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
// Lua: s = net.dns.setdnsserver(ip_addr, [index])
|
|
|
|
|
static int net_setdnsserver( lua_State* L ) {
|
|
|
|
|
size_t l;
|
|
|
|
|
ip_addr_t ipaddr;
|
|
|
|
|
|
|
|
|
|
const char *server = luaL_checklstring( L, 1, &l );
|
|
|
|
|
if (l>16 || server == NULL ||
|
|
|
|
|
!ipaddr_aton (server, &ipaddr) || ip_addr_isany (&ipaddr))
|
|
|
|
|
return luaL_error( L, "invalid dns server ip" );
|
|
|
|
|
|
|
|
|
|
int numdns = luaL_optint(L, 2, 0);
|
|
|
|
|
if (numdns >= DNS_MAX_SERVERS)
|
|
|
|
|
return luaL_error( L, "server index out of range [0-%d]", DNS_MAX_SERVERS - 1);
|
|
|
|
|
|
|
|
|
|
dns_setserver(numdns,&ipaddr);
|
|
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
// Lua: s = net.dns.getdnsserver([index])
|
|
|
|
|
static int net_getdnsserver( lua_State* L ) {
|
|
|
|
|
int numdns = luaL_optint(L, 1, 0);
|
|
|
|
|
if (numdns >= DNS_MAX_SERVERS)
|
|
|
|
|
return luaL_error( L, "server index out of range [0-%d]", DNS_MAX_SERVERS - 1);
|
|
|
|
|
|
|
|
|
|
// ip_addr_t ipaddr;
|
|
|
|
|
// dns_getserver(numdns,&ipaddr);
|
|
|
|
|
// Bug fix by @md5crypt https://github.com/nodemcu/nodemcu-firmware/pull/500
|
2020-07-08 08:46:16 +02:00
|
|
|
const ip_addr_t *ipaddr = dns_getserver(numdns);
|
2016-10-06 08:24:08 +02:00
|
|
|
|
2020-07-08 08:46:16 +02:00
|
|
|
if ( ip_addr_isany(ipaddr) ) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushnil( L );
|
|
|
|
|
} else {
|
|
|
|
|
char temp[IP_STR_SZ];
|
2020-07-08 08:46:16 +02:00
|
|
|
ipstr (temp, ipaddr);
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_pushstring( L, temp );
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
return 1;
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
|
|
|
|
// --- Lua event dispatch -----------------------------------------------
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
static void ldnsfound_cb (lua_State *L, lnet_userdata *ud, ip_addr_t *addr) {
|
|
|
|
|
if (ud->self_ref != LUA_NOREF && ud->client.cb_dns_ref != LUA_NOREF) {
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->client.cb_dns_ref);
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
if (!ip_addr_isany (addr)) {
|
|
|
|
|
char iptmp[IP_STR_SZ];
|
|
|
|
|
ipstr (iptmp, addr);
|
|
|
|
|
lua_pushstring(L, iptmp);
|
|
|
|
|
} else {
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
}
|
|
|
|
|
lua_call(L, 2, 0);
|
|
|
|
|
}
|
|
|
|
|
ud->client.wait_dns --;
|
2017-07-06 23:55:32 +02:00
|
|
|
if (ud->netconn && ud->type == TYPE_TCP_CLIENT && !ud->client.connecting) {
|
|
|
|
|
ud->client.connecting = true;
|
|
|
|
|
netconn_connect(ud->netconn, addr, ud->port);
|
|
|
|
|
} else if (!ud->netconn && ud->client.wait_dns == 0) {
|
2016-10-06 08:24:08 +02:00
|
|
|
lua_gc(L, LUA_GCSTOP, 0);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
ud->self_ref = LUA_NOREF;
|
|
|
|
|
lua_gc(L, LUA_GCRESTART, 0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
static void ldnsstatic_cb (lua_State *L, int cb_ref, ip_addr_t *addr) {
|
|
|
|
|
if (cb_ref == LUA_NOREF)
|
|
|
|
|
return;
|
|
|
|
|
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, cb_ref);
|
|
|
|
|
luaL_unref (L, LUA_REGISTRYINDEX, cb_ref);
|
|
|
|
|
|
|
|
|
|
if (!ip_addr_isany (addr)) {
|
|
|
|
|
char iptmp[IP_STR_SZ];
|
|
|
|
|
ipstr (iptmp, addr);
|
|
|
|
|
lua_pushstring(L, iptmp);
|
|
|
|
|
} else {
|
|
|
|
|
lua_pushnil(L);
|
|
|
|
|
}
|
|
|
|
|
lua_call(L, 1, 0);
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
static void lconnected_cb (lua_State *L, lnet_userdata *ud) {
|
|
|
|
|
if (ud->self_ref != LUA_NOREF && ud->client.cb_connect_ref != LUA_NOREF) {
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->client.cb_connect_ref);
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
lua_call(L, 1, 0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2017-07-06 23:55:32 +02:00
|
|
|
static void lrecv_cb (lua_State *L, lnet_userdata *ud) {
|
2017-07-20 21:04:21 +02:00
|
|
|
if (!ud || !ud->netconn) return;
|
2017-07-06 23:55:32 +02:00
|
|
|
|
|
|
|
|
struct netbuf *p;
|
|
|
|
|
char *payload;
|
|
|
|
|
uint16_t len;
|
|
|
|
|
|
|
|
|
|
err_t err = netconn_recv(ud->netconn, &p);
|
2019-07-11 09:43:25 +02:00
|
|
|
if (err != ERR_OK || !p) {
|
2017-07-06 23:55:32 +02:00
|
|
|
lwip_lua_checkerr(L, err);
|
|
|
|
|
return;
|
|
|
|
|
}
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2019-07-11 09:43:25 +02:00
|
|
|
netbuf_first(p);
|
|
|
|
|
do {
|
2017-07-06 23:55:32 +02:00
|
|
|
netbuf_data(p, (void **)&payload, &len);
|
|
|
|
|
|
2019-07-11 09:43:25 +02:00
|
|
|
if (ud->client.cb_receive_ref != LUA_NOREF){
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->client.cb_receive_ref);
|
|
|
|
|
int num_args = 2;
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
lua_pushlstring(L, payload, len);
|
|
|
|
|
if (ud->type == TYPE_UDP_SOCKET) {
|
|
|
|
|
num_args += 2;
|
|
|
|
|
char iptmp[IP_STR_SZ];
|
|
|
|
|
ip_addr_t *addr = netbuf_fromaddr(p);
|
|
|
|
|
uint16_t port = netbuf_fromport(p);
|
|
|
|
|
ipstr (iptmp, addr);
|
|
|
|
|
lua_pushinteger(L, port);
|
|
|
|
|
lua_pushstring(L, iptmp);
|
|
|
|
|
}
|
|
|
|
|
lua_call(L, num_args, 0);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
2019-07-11 09:43:25 +02:00
|
|
|
} while (netbuf_next(p) != -1);
|
2017-07-06 23:55:32 +02:00
|
|
|
|
2021-07-15 08:25:25 +02:00
|
|
|
if (p)
|
2017-07-06 23:55:32 +02:00
|
|
|
netbuf_delete(p);
|
2016-10-06 08:24:08 +02:00
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
|
|
|
|
static void laccept_cb (lua_State *L, lnet_userdata *ud) {
|
|
|
|
|
if (!ud || !ud->netconn) return;
|
|
|
|
|
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->server.cb_accept_ref);
|
|
|
|
|
|
|
|
|
|
lnet_userdata *nud = net_create(L, TYPE_TCP_CLIENT);
|
|
|
|
|
lua_pushvalue(L, -1);
|
|
|
|
|
nud->self_ref = luaL_ref(L, LUA_REGISTRYINDEX);
|
|
|
|
|
|
|
|
|
|
struct netconn *newconn;
|
|
|
|
|
err_t err = netconn_accept(ud->netconn, &newconn);
|
|
|
|
|
if (err == ERR_OK) {
|
|
|
|
|
nud->netconn = newconn;
|
|
|
|
|
netconn_set_nonblocking(nud->netconn, 1);
|
|
|
|
|
nud->netconn->pcb.tcp->so_options |= SOF_KEEPALIVE;
|
|
|
|
|
nud->netconn->pcb.tcp->keep_idle = ud->server.timeout * 1000;
|
|
|
|
|
nud->netconn->pcb.tcp->keep_cnt = 1;
|
|
|
|
|
} else
|
|
|
|
|
luaL_error(L, "cannot accept new server socket connection");
|
|
|
|
|
lua_call(L, 1, 0);
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
static void lsent_cb (lua_State *L, lnet_userdata *ud) {
|
|
|
|
|
if (ud->client.cb_sent_ref != LUA_NOREF) {
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->client.cb_sent_ref);
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
lua_call(L, 1, 0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
static void lerr_cb (lua_State *L, lnet_userdata *ud, err_t err)
|
|
|
|
|
{
|
2017-07-20 21:04:21 +02:00
|
|
|
if (!ud || !ud->netconn) return;
|
2017-07-06 23:55:32 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
int ref;
|
|
|
|
|
if (err != ERR_OK && ud->client.cb_reconnect_ref != LUA_NOREF)
|
|
|
|
|
ref = ud->client.cb_reconnect_ref;
|
|
|
|
|
else ref = ud->client.cb_disconnect_ref;
|
|
|
|
|
if (ref != LUA_NOREF) {
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ref);
|
|
|
|
|
lua_rawgeti(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
lua_pushinteger(L, err);
|
|
|
|
|
lua_call(L, 2, 0);
|
|
|
|
|
}
|
|
|
|
|
if (ud->client.wait_dns == 0) {
|
|
|
|
|
lua_gc(L, LUA_GCSTOP, 0);
|
|
|
|
|
luaL_unref(L, LUA_REGISTRYINDEX, ud->self_ref);
|
|
|
|
|
ud->self_ref = LUA_NOREF;
|
|
|
|
|
lua_gc(L, LUA_GCRESTART, 0);
|
|
|
|
|
}
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
// --- Tables
|
|
|
|
|
|
|
|
|
|
// Module function map
|
2021-08-21 17:39:54 +02:00
|
|
|
LROT_BEGIN(net_tcpserver, NULL, 0)
|
2019-07-22 11:13:43 +02:00
|
|
|
LROT_FUNCENTRY( listen, net_listen )
|
|
|
|
|
LROT_FUNCENTRY( getaddr, net_getaddr )
|
|
|
|
|
LROT_FUNCENTRY( close, net_close )
|
|
|
|
|
LROT_FUNCENTRY( __gc, net_delete )
|
|
|
|
|
LROT_TABENTRY ( __index, net_tcpserver )
|
|
|
|
|
LROT_END(net_tcpserver, NULL, 0)
|
|
|
|
|
|
2021-08-21 17:39:54 +02:00
|
|
|
LROT_BEGIN(net_tcpsocket, NULL, 0)
|
2019-07-22 11:13:43 +02:00
|
|
|
LROT_FUNCENTRY( connect, net_connect )
|
|
|
|
|
LROT_FUNCENTRY( close, net_close )
|
|
|
|
|
LROT_FUNCENTRY( on, net_on )
|
|
|
|
|
LROT_FUNCENTRY( send, net_send )
|
|
|
|
|
LROT_FUNCENTRY( dns, net_dns )
|
|
|
|
|
LROT_FUNCENTRY( getpeer, net_getpeer )
|
|
|
|
|
LROT_FUNCENTRY( getaddr, net_getaddr )
|
|
|
|
|
LROT_FUNCENTRY( __gc, net_delete )
|
|
|
|
|
LROT_TABENTRY ( __index, net_tcpsocket )
|
|
|
|
|
LROT_END(net_tcpsocket, NULL, 0)
|
|
|
|
|
|
2021-08-21 17:39:54 +02:00
|
|
|
LROT_BEGIN(net_udpsocket, NULL, 0)
|
2019-07-22 11:13:43 +02:00
|
|
|
LROT_FUNCENTRY( listen, net_listen )
|
|
|
|
|
LROT_FUNCENTRY( close, net_close )
|
|
|
|
|
LROT_FUNCENTRY( on, net_on )
|
|
|
|
|
LROT_FUNCENTRY( send, net_send )
|
|
|
|
|
LROT_FUNCENTRY( dns, net_dns )
|
|
|
|
|
LROT_FUNCENTRY( getaddr, net_getaddr )
|
|
|
|
|
LROT_FUNCENTRY( __gc, net_delete )
|
|
|
|
|
LROT_TABENTRY ( __index, net_udpsocket )
|
|
|
|
|
LROT_END(net_udpsocket, NULL, 0)
|
|
|
|
|
|
2021-08-21 17:39:54 +02:00
|
|
|
LROT_BEGIN(net_dns, NULL, 0)
|
2019-07-22 11:13:43 +02:00
|
|
|
LROT_FUNCENTRY( setdnsserver, net_setdnsserver )
|
|
|
|
|
LROT_FUNCENTRY( getdnsserver, net_getdnsserver )
|
|
|
|
|
LROT_FUNCENTRY( resolve, net_dns_static )
|
|
|
|
|
LROT_END(net_dns, NULL, 0)
|
|
|
|
|
|
2021-08-21 17:39:54 +02:00
|
|
|
LROT_BEGIN(net, NULL, 0)
|
2019-07-22 11:13:43 +02:00
|
|
|
LROT_FUNCENTRY( createServer, net_createServer )
|
|
|
|
|
LROT_FUNCENTRY( createConnection, net_createConnection )
|
|
|
|
|
LROT_FUNCENTRY( createUDPSocket, net_createUDPSocket )
|
|
|
|
|
LROT_FUNCENTRY( multicastJoin, net_multicastJoin )
|
|
|
|
|
LROT_FUNCENTRY( multicastLeave, net_multicastLeave )
|
|
|
|
|
LROT_TABENTRY ( dns, net_dns )
|
|
|
|
|
LROT_NUMENTRY ( TCP, TYPE_TCP )
|
|
|
|
|
LROT_NUMENTRY ( UDP, TYPE_UDP )
|
|
|
|
|
LROT_TABENTRY ( __metatable, net )
|
|
|
|
|
LROT_END(net, NULL, 0)
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
|
2016-10-06 08:24:08 +02:00
|
|
|
int luaopen_net( lua_State *L ) {
|
|
|
|
|
igmp_init();
|
|
|
|
|
|
2021-08-21 17:39:54 +02:00
|
|
|
luaL_rometatable(L, NET_TABLE_TCP_SERVER, LROT_TABLEREF(net_tcpserver));
|
|
|
|
|
luaL_rometatable(L, NET_TABLE_TCP_CLIENT, LROT_TABLEREF(net_tcpsocket));
|
|
|
|
|
luaL_rometatable(L, NET_TABLE_UDP_SOCKET, LROT_TABLEREF(net_udpsocket));
|
2016-10-06 08:24:08 +02:00
|
|
|
|
Fix net module data loss & RTOS task unsafety (#2829)
To avoid races between the lwIP callbacks (lwIP RTOS task) and the Lua
handlers (LVM RTOS task), the data flow and ownership has been simplified
and cleaned up.
lwIP callbacks now have no visibility of the userdata struct. They are
limited to creating small event objects and task_post()ing them over
to the LVM "thread", passing ownership in doing so. The shared identifier
then becomes the struct netconn*.
On the LVM side, we keep a linked list of active userdata objects. This
allows us to retrieve the correct userdata when we get an event with
a netconn pointer. Because this list is only ever used within the LVM
task, no locking is necessary.
The old approach of stashing a userdata pointer into the 'socket' field
on the netconn has been removed entirely, as this was both not
thread/RTOS-task safe, and also interfered with the IDFs internal use
of the socket field (even when using only the netconn layer). As an
added benefit, this removed the need for all the SYS_ARCH_PROTECT()
locking stuff.
The need to track receive events before the corresponding userdata object
has been established has been removed by virtue of not reordering the
"accept" and the "recv" events any more (previously accepts were posted
with medium priority, while the receives where high priority, leading
to the observed reordering and associated headaches).
The workaround for IDF issue 784 has been removed as it is now not needed
and is in fact directly harmful as it results in a double-free. Yay for
getting rid of old workarounds!
DNS resolution code paths were merged for the two instances of "socket"
initiated resolves (connect/dns functions).
Also fixed an instance of using a stack variable for receiving the resolved
IP address, with said variable going out of scope before the DNS resolution
necessarily completed (hello, memory corruption!).
Where possible, moved to use the Lua allocator rather than plain malloc.
Finally, the NodeMCU task posting mechanism got a polish and an adjustment.
Given all the Bad(tm) that tends to happen if something fails task posting,
I went through a couple of iterations on how to avoid that. Alas, the
preferred solution of blocking non-LVM RTOS tasks until a slot is free
turned out to not be viable, as this easily resulted in deadlocks with the
lwIP stack. After much deliberation I settled on increasing the number of
available queue slots for the task_post() mechanism, but in the interest
of user control also now made it user configurable via Kconfig.
2019-07-14 23:20:20 +02:00
|
|
|
net_handler = task_get_id(handle_net_event);
|
|
|
|
|
dns_handler = task_get_id(handle_dns_event);
|
2016-10-06 08:24:08 +02:00
|
|
|
|
|
|
|
|
return 0;
|
|
|
|
|
}
|
|
|
|
|
|
2019-07-22 11:13:43 +02:00
|
|
|
NODEMCU_MODULE(NET, "net", net, luaopen_net);
|
