perf(esp32): move WiFi statics from SRAM2 to DRAM1 (~14 KB)

espstack.go

@@ -36,6 +36,9 @@ type StackConfig struct {
 	MaxTCPPorts   int
 	MaxUDPPorts   int
 	PassivePeers  int
+	// AcceptBroadcast4 enables reception of IPv4 broadcast packets.
+	// Must be true when running a DHCP server (AP mode).
+	AcceptBroadcast4 bool
 }
 
 // DHCPConfig configures DHCP address acquisition.
@@ -57,15 +60,16 @@ func NewStack(dev *NetDev, cfg StackConfig) (*Stack, error) {
 	stack := &Stack{dev: dev}
 	const MTU = MaxFrameSize - ethernet.MaxOverheadSize + 4 // CRC not included:+4
 	xcfg := xnet.StackConfig{
-		DNSServer:         cfg.DNSServer,
-		NTPServer:         cfg.NTPServer,
-		Hostname:          cfg.Hostname,
-		MaxActiveTCPPorts: uint16(cfg.MaxTCPPorts),
-		MaxActiveUDPPorts: uint16(cfg.MaxUDPPorts),
-		RandSeed:          time.Now().UnixNano() ^ cfg.RandSeed,
-		HardwareAddress:   mac,
-		MTU:               MTU,
-		PassivePeers:      cfg.PassivePeers,
+		DNSServer:           cfg.DNSServer,
+		NTPServer:           cfg.NTPServer,
+		Hostname:            cfg.Hostname,
+		MaxActiveTCPPorts:   uint16(cfg.MaxTCPPorts),
+		MaxActiveUDPPorts:   uint16(cfg.MaxUDPPorts),
+		RandSeed:            time.Now().UnixNano() ^ cfg.RandSeed,
+		HardwareAddress:     mac,
+		MTU:                 MTU,
+		PassivePeers:        cfg.PassivePeers,
+		AcceptIPv4Broadcast: cfg.AcceptBroadcast4,
 	}
 	if cfg.StaticAddress.IsValid() && cfg.StaticAddress.Is4() {
 		xcfg.StaticAddress4 = cfg.StaticAddress.As4()

examples/ap/main-ap.go

@@ -50,9 +50,10 @@ func main() {
 
 	println("ap: creating lneto stack...")
 	stack, err := espradio.NewStack(nd, espradio.StackConfig{
-		Hostname:      ssid,
-		StaticAddress: addr,
-		MaxUDPPorts:   2,
+		Hostname:         ssid,
+		StaticAddress:    addr,
+		MaxUDPPorts:      2,
+		AcceptBroadcast4: true,
 	})
 	if err != nil {
 		failure("ap: stack err: " + err.Error())

isr.c

@@ -76,8 +76,16 @@ void espradio_user_exception(uint32_t cause, uint32_t epc, uint32_t excvaddr, ui
 
 /* ---- ISR fn/arg storage ---- */
 
-static void (*s_isr_fn[32])(void *);
-static void *s_isr_arg[32];
+/* On ESP32, place WiFi-only tables in DRAM1 (.wifibss) to free SRAM2 for
+ * the Go GC heap.  On other targets they stay in normal .bss. */
+#if CONFIG_IDF_TARGET_ESP32
+#define WIFIBSS __attribute__((section(".wifibss")))
+#else
+#define WIFIBSS
+#endif
+
+static void (*s_isr_fn[32])(void *) WIFIBSS;
+static void *s_isr_arg[32] WIFIBSS;
 
 /* Bitmask of ISR slots registered via espradio_set_intr (WiFi sources only). */
 static uint32_t s_wifi_isr_slots;
@@ -140,8 +148,8 @@ void espradio_task_yield_from_isr(void) {
 static volatile uint32_t s_isr_ring_head;
 static volatile uint32_t s_isr_ring_tail;
 static volatile uint32_t s_isr_ring_drops;
-static void             *s_isr_ring_queue[ESPRADIO_ISR_RING_SIZE];
-static uint8_t           s_isr_ring_items[ESPRADIO_ISR_RING_SIZE][ESPRADIO_ISR_ITEM_SIZE];
+static void             *s_isr_ring_queue[ESPRADIO_ISR_RING_SIZE] WIFIBSS;
+static uint8_t           s_isr_ring_items[ESPRADIO_ISR_RING_SIZE][ESPRADIO_ISR_ITEM_SIZE] WIFIBSS;
 
 int32_t espradio_queue_send_from_isr(void *queue, void *item, void *hptw) {
     if (hptw) {

netif.c

@@ -32,27 +32,18 @@ extern void (*mesh_rxcb)(void);
 
 static void blob_cb_noop(void) { }
 
-/* pp_wdev_funcs relocation: the ROM ppTask dispatcher calls through a
- * heap-allocated function-pointer table (pp_wdev_funcs, 196 entries = 0x310
- * bytes).  DMA corruption can zero heap entries at runtime → pc:nil crash.
- * After esp_wifi_start() we copy the table into static .bss and redirect
- * the ROM pointer there, where DMA cannot reach. */
+/* DMA-corruption workaround tables: relocate heap-allocated function-pointer
+ * tables to static .bss where DMA cannot reach.  Only needed on ESP32-C3/S3;
+ * on the original ESP32 the blob manages these tables itself and the entry
+ * counts differ, so these are compiled out to save ~1.9 KB BSS. */
 #define PP_WDEV_FUNCS_ENTRIES  196
-static uint32_t s_pp_wdev_save[PP_WDEV_FUNCS_ENTRIES];
-
-/* net80211_funcs relocation: same DMA corruption issue as pp_wdev_funcs.
- * The blob allocates this table on the heap; we relocate to static .bss.
- * net80211_funcs_init writes ~44 entries; 128 provides safe headroom. */
 #define NET80211_FUNCS_MAX_ENTRIES  128
-static uint32_t s_net80211_funcs_save[NET80211_FUNCS_MAX_ENTRIES];
-
-/* g_phyFuns relocation: the PHY function table lives at a fixed DRAM address
- * (0x3fcef3d4, size 0x298 = 166 words) that can be corrupted at runtime —
- * entries are overwritten with arena allocation addresses, causing
- * InstructionFetchError when the PHY calibration timer reads temperature.
- * Relocate to static .bss after init and redirect g_phyFuns. */
 #define PHY_FUNCS_TABLE_WORDS  166
+#if !CONFIG_IDF_TARGET_ESP32
+static uint32_t s_pp_wdev_save[PP_WDEV_FUNCS_ENTRIES];
+static uint32_t s_net80211_funcs_save[NET80211_FUNCS_MAX_ENTRIES];
 static uint32_t s_phyFuns_save[PHY_FUNCS_TABLE_WORDS];
+#endif
 extern void *g_phyFuns;
 
 /* ppCheckTxConnTrafficIdle is called only by PM timer callbacks (pm_dream,
@@ -213,9 +204,11 @@ void espradio_restore_rom_ptrs(void) {
     if ((uint32_t)(uintptr_t)g_osi_funcs_p != s_saved_g_osi_funcs_p)
         g_osi_funcs_p = (wifi_osi_funcs_t *)(uintptr_t)s_saved_g_osi_funcs_p;
     /* esp_phy_enable resets g_phyFuns to the fixed DRAM address on every call.
-     * Redirect it back to our static copy. */
+     * Redirect it back to our static copy (C3/S3 only). */
+#if !CONFIG_IDF_TARGET_ESP32
     if (g_phyFuns != s_phyFuns_save && s_phyFuns_save[0] != 0)
         g_phyFuns = s_phyFuns_save;
+#endif
 }
 
 #define ESPRADIO_NETIF_RXRING_SIZE  6
@@ -226,7 +219,15 @@ typedef struct {
     uint16_t len;
 } espradio_rx_frame_t;
 
+/* On ESP32, place the RX ring in the dedicated DRAM1 region (SRAM1 pool 7/6)
+ * to free ~9.6 KB of SRAM2 for the Go GC heap.  The linker script's .rxring
+ * section sits before the arena in DRAM1.  On other targets the ring stays in
+ * normal .bss. */
+#if CONFIG_IDF_TARGET_ESP32
+static espradio_rx_frame_t s_rx_ring[ESPRADIO_NETIF_RXRING_SIZE] __attribute__((section(".rxring")));
+#else
 static espradio_rx_frame_t s_rx_ring[ESPRADIO_NETIF_RXRING_SIZE];
+#endif
 static volatile uint32_t   s_rx_head;
 static volatile uint32_t   s_rx_tail;
 static volatile uint32_t   s_rx_cb_count;

netlink/ap.go

@@ -102,11 +102,12 @@ func (n *Esplink) NetConnectAP(params APConnectParams) error {
 		udpPorts = 1 // reserve one slot for the DHCP server
 	}
 	espstack, err := espradio.NewStack(nd, espradio.StackConfig{
-		Hostname:      params.Hostname,
-		StaticAddress: params.StaticAddr,
-		MaxUDPPorts:   udpPorts,
-		MaxTCPPorts:   params.MaxTCPPorts,
-		PassivePeers:  params.PassivePeers,
+		Hostname:         params.Hostname,
+		StaticAddress:    params.StaticAddr,
+		MaxUDPPorts:      udpPorts,
+		MaxTCPPorts:      params.MaxTCPPorts,
+		PassivePeers:     params.PassivePeers,
+		AcceptBroadcast4: true,
 	})
 	if err != nil {
 		if debug {

osi.c

@@ -780,6 +780,13 @@ static int espradio_read_mac(uint8_t* mac, unsigned int type) {
     return rc;
 }
 
+/* On ESP32, place WiFi-only tables in DRAM1 (.wifibss) to free SRAM2. */
+#if CONFIG_IDF_TARGET_ESP32
+#define WIFIBSS __attribute__((section(".wifibss")))
+#else
+#define WIFIBSS
+#endif
+
 #define TIMER_SLOTS 32
 static struct {
     void *ptimer;
@@ -790,7 +797,7 @@ static struct {
     bool pending_setfn;
     uint64_t interval_us;
     uint64_t deadline_us;
-} timer_slots[TIMER_SLOTS];
+} timer_slots[TIMER_SLOTS] WIFIBSS;
 static unsigned timer_slots_used;
 
 void espradio_timer_fire(void *ptimer);

pcap_debug.go

@@ -19,5 +19,5 @@ var _ = pcap.Configure(machine.Serial, xnet.CapturePrinterConfig{
 })
 
 func printPacket(msg string, frame []byte) {
-	pcap.PrintPacket(msg, frame)
+	pcap.PrintEthernet(msg, frame)
 }