Merge pull request adafruit#10027 from eightycc/issue-9837

Resolve RP2 Network Performance Issue

Author: tannewt

.gitmodules

@@ -312,7 +312,7 @@
 [submodule "ports/raspberrypi/lib/lwip"]
 	path = ports/raspberrypi/lib/lwip
 	url = https://github.com/adafruit/lwip.git
-	branch = circuitpython8
+	branch = circuitpython9
 [submodule "lib/mbedtls"]
 	path = lib/mbedtls
 	url = https://github.com/ARMmbed/mbedtls.git

ports/raspberrypi/lib/lwip

@@ -0,0 +1,13 @@
+// This file is part of the CircuitPython project: https://circuitpython.org
+//
+// SPDX-FileCopyrightText: Copyright (c) 2025 Bob Abeles
+//
+// SPDX-License-Identifier: MIT
+
+#pragma once
+
+#include <stddef.h>
+
+void *lwip_heap_malloc(size_t size);
+void lwip_heap_free(void *ptr);
+void *lwip_heap_calloc(size_t num, size_t size);

ports/raspberrypi/lwip_inc/lwip_mem.h

@@ -5,6 +5,8 @@
 // Common settings used in most of the pico_w examples
 // (see https://www.nongnu.org/lwip/2_1_x/group__lwip__opts.html for details)
 
+#include "lwip_mem.h"
+
 // allow override in some examples
 #ifndef NO_SYS
 #define NO_SYS                      1
@@ -20,14 +22,62 @@
 #define MEM_LIBC_MALLOC             0
 #endif
 #define MEM_ALIGNMENT               4
-#define MEM_SIZE                    4000
-#define MEMP_NUM_TCP_SEG            32
-#define MEMP_NUM_ARP_QUEUE          10
-#define PBUF_POOL_SIZE              24
+// MEM_USE_POOLS: mem_malloc uses pools of fixed size memory blocks. Default is 0.
+#define MEM_USE_POOLS               0
+// MEM_USE_POOLS_TRY_BIGGER_POOL: if one pool is empty, try the next bigger pool. Default is 0.
+#define MEM_USE_POOLS_TRY_BIGGER_POOL 0
+// MEMP_USE_CUSTOM_POOLS: Use custom pools defined in lwippools.h. Default is 0.
+#define MEMP_USE_CUSTOM_POOLS       0
+// MEMP_MEM_MALLOC: Use mem_malloc() for pool memory. Default is 0.
+#define MEMP_MEM_MALLOC             1
+#define MEM_CUSTOM_ALLOCATOR        1
+#define MEM_CUSTOM_FREE             lwip_heap_free
+#define MEM_CUSTOM_MALLOC           lwip_heap_malloc
+#define MEM_CUSTOM_CALLOC           lwip_heap_calloc
+
+// MEM_SIZE: The LWIP heap size. Memory for mem_malloc and mem_calloc are allocated from
+// this heap. If MEMP_MEM_MALLOC is set to 1, memory for memp_malloc is also allocated from
+// this heap; if it is 0, memory is statically pre-allocated for each pool.
+// Default is 1600.
+#define MEM_SIZE                    1600
+// MEMP_NUM_PBUF: memp pbufs used when sending from static memory. Default is 16.
+#define MEMP_NUM_PBUF               16
+// MEMP_NUM_RAW_PCB: Number of raw connection PCBs. Default is 4.
+#define MEMP_NUM_RAW_PCB            4
+// MEMP_NUM_UDP_PCB: Number of UDP PCBs. Default is 4.
+#define MEMP_NUM_UDP_PCB            4
+// MEMP_NUM_TCP_PCB: Number of simultaneously active TCP connections. Default is 5.
+#define MEMP_NUM_TCP_PCB            5
+// MEMP_NUM_TCP_PCB_LISTEN: Number of listening TCP PCBs. Default is 8.
+#define MEMP_NUM_TCP_PCB_LISTEN     8
+// MEMP_NUM_TCP_SEG: Number of simultaneously queued TCP segments. Default is 16.
+#define MEMP_NUM_TCP_SEG            16
+// MEMP_NUM_ALTCP_PCB: Number of simultaneously active altcp connections. Default is 5.
+#define MEMP_NUM_ALTCP_PCB          5
+// MEMP_NUM_REASSDATA: Number of simultaneously IP packets queued for reassembly. Default is 5.
+#define MEMP_NUM_REASSDATA          5
+// MEMP_NUM_FRAG_PBUF: Number of simultaneously IP fragments. Default is 15.
+#define MEMP_NUM_FRAG_PBUF          15
+// MEMP_NUM_ARP_QUEUE: Number of simultaneously queued ARP packets. Default is 30.
+#define MEMP_NUM_ARP_QUEUE          30
+// MEMP_NUM_IGMP_GROUP: Number of simultaneously active IGMP groups. Default is 8.
+#define MEMP_NUM_IGMP_GROUP         8
+// MEMP_NUM_SYS_TIMEOUT: Number of simultaneously active timeouts.
+// Use calculated default based on enabled modules.
+
+// PBUF_POOL_SIZE: Number of pbufs in the pbuf pool. Default is 16.
+#define PBUF_POOL_SIZE              16
+
+// LWIP's default 250 ms periodic timer interval is too long, resulting in network
+// performance issues. We reduce it to 25 ms giving a slow-timer of 50 ms and a
+// fast-timer of 25 ms.
+#define TCP_TMR_INTERVAL            25
+
 #define LWIP_ARP                    1
 #define LWIP_ETHERNET               1
 #define LWIP_ICMP                   1
 #define LWIP_RAW                    1
+
 #define TCP_WND                     (8 * TCP_MSS)
 #define TCP_MSS                     1460
 #define TCP_SND_BUF                 (8 * TCP_MSS)

ports/raspberrypi/lwip_inc/lwipopts.h

@@ -0,0 +1,27 @@
+// This file is part of the CircuitPython project: https://circuitpython.org
+//
+// SPDX-FileCopyrightText: Copyright (c) 2025 Bob Abeles
+//
+// SPDX-License-Identifier: MIT
+
+#include <stdint.h>
+#include <string.h>
+#include "lib/tlsf/tlsf.h"
+#include "lwip_mem.h"
+#include "supervisor/port_heap.h"
+
+void *lwip_heap_malloc(size_t size) {
+    return port_malloc(size, true);
+}
+
+void lwip_heap_free(void *ptr) {
+    port_free(ptr);
+}
+
+void *lwip_heap_calloc(size_t num, size_t size) {
+    void *ptr = lwip_heap_malloc(num * size);
+    if (ptr != NULL) {
+        memset(ptr, 0, num * size);
+    }
+    return ptr;
+}

ports/raspberrypi/lwip_src/lwip_mem.c

@@ -60,6 +60,7 @@
 
 #include "tusb.h"
 #include <cmsis_compiler.h>
+#include "lib/tlsf/tlsf.h"
 
 critical_section_t background_queue_lock;
 
@@ -87,20 +88,18 @@ extern uint32_t _ld_itcm_destination;
 extern uint32_t _ld_itcm_size;
 extern uint32_t _ld_itcm_flash_copy;
 
-#ifdef CIRCUITPY_PSRAM_CHIP_SELECT
+static tlsf_t _heap = NULL;
+static pool_t _ram_pool = NULL;
+static pool_t _psram_pool = NULL;
+static size_t _psram_size = 0;
 
-#include "lib/tlsf/tlsf.h"
+#ifdef CIRCUITPY_PSRAM_CHIP_SELECT
 
 #include "src/rp2350/hardware_regs/include/hardware/regs/qmi.h"
 #include "src/rp2350/hardware_regs/include/hardware/regs/xip.h"
 #include "src/rp2350/hardware_structs/include/hardware/structs/qmi.h"
 #include "src/rp2350/hardware_structs/include/hardware/structs/xip_ctrl.h"
 
-static tlsf_t _heap = NULL;
-static pool_t _ram_pool = NULL;
-static pool_t _psram_pool = NULL;
-static size_t _psram_size = 0;
-
 static void __no_inline_not_in_flash_func(setup_psram)(void) {
     gpio_set_function(CIRCUITPY_PSRAM_CHIP_SELECT->number, GPIO_FUNC_XIP_CS1);
     _psram_size = 0;
@@ -240,8 +239,9 @@ static void __no_inline_not_in_flash_func(setup_psram)(void) {
         return;
     }
 }
+#endif
 
-void port_heap_init(void) {
+static void _port_heap_init(void) {
     uint32_t *heap_bottom = port_heap_get_bottom();
     uint32_t *heap_top = port_heap_get_top();
     size_t size = (heap_top - heap_bottom) * sizeof(uint32_t);
@@ -252,6 +252,10 @@ void port_heap_init(void) {
     }
 }
 
+void port_heap_init(void) {
+    // We call _port_heap_init from port_init to initialize the heap early.
+}
+
 void *port_malloc(size_t size, bool dma_capable) {
     void *block = tlsf_malloc(_heap, size);
     return block;
@@ -282,7 +286,6 @@ size_t port_heap_get_largest_free_size(void) {
     // IDF does this. Not sure why.
     return tlsf_fit_size(_heap, max_size);
 }
-#endif
 
 safe_mode_t port_init(void) {
     _binary_info();
@@ -348,6 +351,9 @@ safe_mode_t port_init(void) {
     setup_psram();
     #endif
 
+    // Initialize heap early to allow for early allocation.
+    _port_heap_init();
+
     // Check brownout.
 
     #if CIRCUITPY_CYW43
@@ -356,6 +362,7 @@ safe_mode_t port_init(void) {
     // are intended to meet the power on timing requirements, but apparently
     // are inadequate. We'll back off this long delay based on future testing.
     mp_hal_delay_ms(1000);
+
     // Change this as a placeholder as to how to init with country code.
     // Default country code is CYW43_COUNTRY_WORLDWIDE)
     if (cyw43_arch_init_with_country(PICO_CYW43_ARCH_DEFAULT_COUNTRY_CODE)) {