mirror of
https://github.com/esphome/esphome.git
synced 2026-09-15 17:18:40 +00:00
[rp2040] Fix W5500 Ethernet pbuf corruption by mirroring LWIPMutex semantics
The previous LwIPLock for RP2040+Ethernet only called ethernet_arch_lwip_begin/end, taking the async_context lock but never bumping arduino-pico's __inLWIP counter. The W5500 GPIO IRQ path (LwipIntfDev::_irq) checks __inLWIP to decide whether to defer packet processing; with __inLWIP == 0 it would proceed to run _lwipCallback immediately from IRQ context even while the main thread held the async_context lock, re-entering lwIP and corrupting pbuf chains. The corruption surfaced later as the pbuf_cat assertion "p->tot_len == p->len (of last pbuf in chain)" panicking the device on wiznet-w5500-evb-pico boards configured with an interrupt pin. Mirror arduino-pico's LWIPMutex (cores/rp2040/lwip_wrap.h) exactly: bump __inLWIP, take the lock, and on release re-unmask any GPIO IRQs that were deferred while we held it (__needsIRQEN). We can't write `using LwIPLock = LWIPMutex;` in core/helpers.h because pulling lwip_wrap.h there transitively poisons many translation units with lwIP types in an order that breaks err_t resolution, so the body is inlined in the rp2040 helpers.cpp where the include is scoped.
This commit is contained in:
@@ -9,7 +9,7 @@
|
||||
#include <WiFi.h>
|
||||
#include <pico/cyw43_arch.h> // For cyw43_arch_lwip_begin/end (LwIPLock)
|
||||
#elif defined(USE_ETHERNET)
|
||||
#include <LwipEthernet.h> // For ethernet_arch_lwip_begin/end (LwIPLock)
|
||||
#include <lwip_wrap.h> // For LWIPMutex — LwIPLock delegates to it (see below)
|
||||
#include "esphome/components/ethernet/ethernet_component.h"
|
||||
#endif
|
||||
#include <hardware/structs/rosc.h>
|
||||
@@ -43,9 +43,18 @@ IRAM_ATTR InterruptLock::~InterruptLock() { restore_interrupts(state_); }
|
||||
// main loop, corrupting the shared rx_buf_ pbuf chain (use-after-free, pbuf_cat
|
||||
// assertion failures). See esphome#10681.
|
||||
//
|
||||
// WiFi uses cyw43_arch_lwip_begin/end; Ethernet uses ethernet_arch_lwip_begin/end.
|
||||
// Both acquire the async_context recursive mutex to prevent IRQ callbacks from
|
||||
// firing during critical sections.
|
||||
// WiFi uses cyw43_arch_lwip_begin/end.
|
||||
//
|
||||
// For wired Ethernet, taking only the async_context lock is NOT enough. The
|
||||
// W5500 GPIO IRQ path (LwipIntfDev::_irq) checks arduino-pico's `__inLWIP`
|
||||
// counter to decide whether to defer packet processing. If we hold the
|
||||
// async_context lock without bumping `__inLWIP`, an interrupt-driven packet
|
||||
// arrival re-enters lwIP from IRQ context and corrupts pbufs (the `pbuf_cat`
|
||||
// assertion crash on wiznet-w5500-evb-pico). We mirror arduino-pico's
|
||||
// LWIPMutex (cores/rp2040/lwip_wrap.h) exactly: bump `__inLWIP`, take the
|
||||
// lock, and on release re-unmask any GPIO IRQs that were deferred while we
|
||||
// held it. We can't `using LwIPLock = LWIPMutex;` in helpers.h because
|
||||
// pulling lwip_wrap.h there poisons many TUs with lwIP types.
|
||||
//
|
||||
// When neither WiFi nor Ethernet is configured, this is a no-op since
|
||||
// there's no network stack and no lwip callbacks to race with.
|
||||
@@ -53,8 +62,18 @@ IRAM_ATTR InterruptLock::~InterruptLock() { restore_interrupts(state_); }
|
||||
LwIPLock::LwIPLock() { cyw43_arch_lwip_begin(); }
|
||||
LwIPLock::~LwIPLock() { cyw43_arch_lwip_end(); }
|
||||
#elif defined(USE_ETHERNET)
|
||||
LwIPLock::LwIPLock() { ethernet_arch_lwip_begin(); }
|
||||
LwIPLock::~LwIPLock() { ethernet_arch_lwip_end(); }
|
||||
LwIPLock::LwIPLock() {
|
||||
__inLWIP++;
|
||||
ethernet_arch_lwip_begin();
|
||||
}
|
||||
LwIPLock::~LwIPLock() {
|
||||
ethernet_arch_lwip_end();
|
||||
__inLWIP--;
|
||||
if (__needsIRQEN && !__inLWIP) {
|
||||
__needsIRQEN = false;
|
||||
ethernet_arch_lwip_gpio_unmask();
|
||||
}
|
||||
}
|
||||
#else
|
||||
LwIPLock::LwIPLock() {}
|
||||
LwIPLock::~LwIPLock() {}
|
||||
|
||||
Reference in New Issue
Block a user