Run FreeRTOS ISRs on a dedicated stack

Author: rhgndf

ch32v305/README.md

@@ -54,18 +54,22 @@ wchisp flash ch32v305_sdr.elf
 
   | Task | Priority | Stack |
   | --- | ---: | ---: |
-  | I2S processing | 3 | 256 words / 1024 bytes |
-  | TinyUSB device | 2 | 384 words / 1536 bytes |
-  | Application/UI | 1 | 512 words / 2048 bytes |
+  | I2S processing | 3 | 192 words / 768 bytes |
+  | TinyUSB device | 2 | 256 words / 1024 bytes |
+  | Application/UI | 1 | 384 words / 1536 bytes |
   | FreeRTOS idle | 0 | 192 words / 768 bytes |
 
 - The firmware and NMSIS DSP library are built with `-fstack-usage`. With
   release LTO, the emitted linked call paths measured 320 bytes for I2S, 576
   bytes for USB, 1040 bytes for the application, and 128 bytes for idle. The
-  allocations include the largest
-  compiler-reported nested interrupt path, the 256-byte RISC-V integer/FPU
-  switch frame, and extra margin for prebuilt newlib calls not represented in
-  this target's `.su` records.
+  allocations include the 256-byte RISC-V integer/FPU switch frame and extra
+  margin for prebuilt newlib calls not represented in this target's `.su`
+  records. Returning maskable ISRs use the 2 KiB startup stack as a dedicated
+  post-scheduler interrupt stack; their assembly wrappers preserve the
+  caller-saved FPU state before calling ordinary C bodies. The release/LTO
+  compiler call graph reports 608 bytes for the worst priority-compatible
+  three-active-interrupt path including all wrapper frames, leaving 1440 bytes
+  of static margin in that stack.
 - The kernel tick is 1 kHz, derived from a free-running 144 MHz CH32 SysTick.
   Tickless idle installs an absolute compare but intentionally busy-waits
   instead of executing `WFI`. The wait tests `CNT >= deadline` as well as

ch32v305/lib/Ld/Link.ld

@@ -168,8 +168,15 @@ SECTIONS
 	PROVIDE( _heap_end = ORIGIN(RAM) + LENGTH(RAM) - __stack_size );
 	PROVIDE( _susrstack = _heap_end );
 	PROVIDE( _eusrstack = ORIGIN(RAM) + LENGTH(RAM) );
+	PROVIDE( __freertos_irq_stack_bottom = _susrstack );
 	PROVIDE( __freertos_irq_stack_top = _eusrstack );
 
 	ASSERT( _end <= _heap_end, "RAM overflow: .data/.bss/heap overlap stack" )
+	ASSERT( (__freertos_irq_stack_bottom & 15) == 0,
+		"FreeRTOS IRQ stack bottom must be 16-byte aligned" )
+	ASSERT( (__freertos_irq_stack_top & 15) == 0,
+		"FreeRTOS IRQ stack top must be 16-byte aligned" )
+	ASSERT( (__freertos_irq_stack_top - __freertos_irq_stack_bottom) == __stack_size,
+		"FreeRTOS IRQ stack bounds must cover the startup stack" )
 
 }

ch32v305/src/ch32v30x_it.c

@@ -10,6 +10,7 @@
 * microcontroller manufactured by Nanjing Qinheng Microelectronics.
 *******************************************************************************/
 #include "ch32v30x_it.h"
+#include "freertos/port_isr.h"
 #include "tusb.h"
 
 void NMI_Handler(void) __attribute__((interrupt));
@@ -44,6 +45,7 @@ void HardFault_Handler(void)
   }
 }
 
-__attribute__((interrupt)) void USBHS_IRQHandler(void) {
+PORT_ISR_BODY(USBHS_IRQHandler)
+{
   tud_int_handler(0);
 }

ch32v305/src/freertos/port_ch32v30x.S

@@ -8,6 +8,111 @@
 
 .extern vTaskSwitchContext
 
+/*
+ * A normal C interrupt body is called through one of these wrappers.  QingKe
+ * HWSTK preserves the caller-saved integer registers and interrupt return
+ * state.  The C ABI preserves the callee-saved integer/FPU registers.  Save
+ * the remaining caller-saved FPU state here.
+ *
+ * The 96-byte frame is 16-byte aligned and contains:
+ *   0..76  f0-f7, f10-f17, f28-f31
+ *      80  fcsr
+ *      84  interrupted SP when switching from a task stack, otherwise zero
+ *   88..95 alignment padding
+ *
+ * Test the actual SP against the IRQ-stack bounds instead of relying on PFIC
+ * nesting state.  This also handles a higher-priority interrupt arriving at
+ * any instruction in another wrapper's stack-switch or restore sequence.
+ * No SP-relative store is performed until the stack choice is complete.
+ */
+.macro portIRQ_WRAPPER vector, body
+.extern \body
+.global \vector
+.type \vector, @function
+.section .text.\vector, "ax", @progbits
+.p2align 2
+\vector:
+    la t0, __freertos_irq_stack_bottom
+    la t1, __freertos_irq_stack_top
+    mv t2, sp
+    bltu sp, t0, .Lirq_use_top_\@
+    bgeu sp, t1, .Lirq_use_top_\@
+    li t2, 0
+    j .Lirq_frame_\@
+
+.Lirq_use_top_\@:
+    mv sp, t1
+
+.Lirq_frame_\@:
+    addi sp, sp, -96
+    sw t2, 84(sp)
+
+    fsw f0,   0(sp)
+    fsw f1,   4(sp)
+    fsw f2,   8(sp)
+    fsw f3,  12(sp)
+    fsw f4,  16(sp)
+    fsw f5,  20(sp)
+    fsw f6,  24(sp)
+    fsw f7,  28(sp)
+    fsw f10, 32(sp)
+    fsw f11, 36(sp)
+    fsw f12, 40(sp)
+    fsw f13, 44(sp)
+    fsw f14, 48(sp)
+    fsw f15, 52(sp)
+    fsw f16, 56(sp)
+    fsw f17, 60(sp)
+    fsw f28, 64(sp)
+    fsw f29, 68(sp)
+    fsw f30, 72(sp)
+    fsw f31, 76(sp)
+    csrr t0, fcsr
+    sw t0, 80(sp)
+
+    call \body
+
+    flw f0,   0(sp)
+    flw f1,   4(sp)
+    flw f2,   8(sp)
+    flw f3,  12(sp)
+    flw f4,  16(sp)
+    flw f5,  20(sp)
+    flw f6,  24(sp)
+    flw f7,  28(sp)
+    flw f10, 32(sp)
+    flw f11, 36(sp)
+    flw f12, 40(sp)
+    flw f13, 44(sp)
+    flw f14, 48(sp)
+    flw f15, 52(sp)
+    flw f16, 56(sp)
+    flw f17, 60(sp)
+    flw f28, 64(sp)
+    flw f29, 68(sp)
+    flw f30, 72(sp)
+    flw f31, 76(sp)
+    lw t0, 80(sp)
+    csrw fcsr, t0
+
+    lw t2, 84(sp)
+    addi sp, sp, 96
+    beqz t2, .Lirq_return_\@
+    mv sp, t2
+
+.Lirq_return_\@:
+    mret
+.size \vector, .-\vector
+.endm
+
+portIRQ_WRAPPER SysTick_Handler, SysTick_Handler_Body
+portIRQ_WRAPPER USBHS_IRQHandler, USBHS_IRQHandler_Body
+portIRQ_WRAPPER DMA1_Channel4_IRQHandler, DMA1_Channel4_IRQHandler_Body
+portIRQ_WRAPPER DMA2_Channel2_IRQHandler, DMA2_Channel2_IRQHandler_Body
+portIRQ_WRAPPER EXTI15_10_IRQHandler, EXTI15_10_IRQHandler_Body
+portIRQ_WRAPPER TIM10_CC_IRQHandler, TIM10_CC_IRQHandler_Body
+portIRQ_WRAPPER TIM6_IRQHandler, TIM6_IRQHandler_Body
+
 .global SW_Handler
 .type SW_Handler, @function
 .section .text.SW_Handler, "ax", @progbits

ch32v305/src/freertos/port_ch32v30x.c

@@ -2,6 +2,7 @@
 #include "task.h"
 
 #include "debug.h"
+#include "port_isr.h"
 
 #include <stdint.h>
 
@@ -59,6 +60,8 @@ static BaseType_t pfic_has_enabled_pending_irq(void)
 void vPortYield(void)
 {
     NVIC_SetPendingIRQ(Software_IRQn);
+    /* Publish the PFIC MMIO write before execution can continue past a yield. */
+    __asm volatile("fence iorw, iorw" ::: "memory");
 }
 
 void vPortSetupTimerInterrupt(void)
@@ -85,8 +88,7 @@ void vPortSetupTimerInterrupt(void)
                     CH32_SYSTICK_CTLR_STCLK;
 }
 
-void SysTick_Handler(void) __attribute__((interrupt));
-void SysTick_Handler(void)
+PORT_ISR_BODY(SysTick_Handler)
 {
     UBaseType_t saved_mie = portSET_INTERRUPT_MASK_FROM_ISR();
     BaseType_t switch_required = pdFALSE;

ch32v305/src/freertos/port_isr.h

@@ -0,0 +1,18 @@
+#ifndef HFSDR_FREERTOS_PORT_ISR_H
+#define HFSDR_FREERTOS_PORT_ISR_H
+
+/*
+ * Returning maskable interrupt vectors are assembly wrappers in
+ * port_ch32v30x.S.  Their ordinary C bodies use the platform ABI and return
+ * with ret; the wrapper owns the dedicated IRQ stack and the final mret.
+ *
+ * used and externally_visible keep the body available to the assembly-only
+ * reference when link-time optimisation is enabled.  noinline keeps the
+ * compiler-emitted stack-usage record attributable to the ISR body.
+ */
+#define PORT_ISR_BODY(vector_name)                                      \
+    void vector_name##_Body(void)                                       \
+        __attribute__((used, noinline, externally_visible));            \
+    void vector_name##_Body(void)
+
+#endif

ch32v305/src/freertos/task_stacks.h

@@ -5,12 +5,15 @@
  * Sizes are in 32-bit StackType_t words.  The release/LTO -fstack-usage
  * records give worst linked call paths of 1040 bytes for Application_Task,
  * 576 bytes for usb_task, and 320 bytes for i2s_task.  The allocations below
- * also cover the maximum compiler-reported nested interrupt path, the 256-byte
- * integer/FPU context-switch frame, and margin for prebuilt newlib routines
- * whose stack frames are not present in this target's .su files.
+ * cover the 256-byte integer/FPU context-switch frame and retain conservative
+ * margin for prebuilt newlib routines whose stack frames are not present in
+ * this target's .su files.  Returning maskable ISRs use the dedicated 2 KiB
+ * IRQ stack instead of application task stacks.  The remaining measured
+ * margins are 240 bytes for Application_Task and 192 bytes each for usb_task
+ * and i2s_task.
  */
-#define APP_TASK_STACK_WORDS    512U
-#define USB_TASK_STACK_WORDS    384U
-#define I2S_TASK_STACK_WORDS    256U
+#define APP_TASK_STACK_WORDS    384U
+#define USB_TASK_STACK_WORDS    256U
+#define I2S_TASK_STACK_WORDS    192U
 
 #endif

ch32v305/src/hw/encoder.c

@@ -1,5 +1,6 @@
 #include "encoder.h"
 
+#include "freertos/port_isr.h"
 #include "hw/pinout.h"
 #include "ui/ui.h"
 
@@ -61,7 +62,7 @@ static int16_t encoder_sync_delta(void)
     return delta;
 }
 
-__attribute__((interrupt)) void EXTI15_10_IRQHandler(void)
+PORT_ISR_BODY(EXTI15_10_IRQHandler)
 {
     if(EXTI_GetITStatus(ENCODER_BUTTON_EXTI_LINE) == RESET)
     {
@@ -106,7 +107,7 @@ __attribute__((interrupt)) void EXTI15_10_IRQHandler(void)
     }
 }
 
-__attribute__((interrupt)) void TIM10_CC_IRQHandler(void)
+PORT_ISR_BODY(TIM10_CC_IRQHandler)
 {
     uint8_t have_cc1 = (uint8_t)(TIM_GetITStatus(ENCODER_TIMER, TIM_IT_CC1) != RESET);
     uint8_t have_cc2 = (uint8_t)(TIM_GetITStatus(ENCODER_TIMER, TIM_IT_CC2) != RESET);

ch32v305/src/hw/i2s.c

@@ -8,6 +8,7 @@
 
 #include "debug.h"
 #include "demod/demod.h"
+#include "freertos/port_isr.h"
 #include "main.h"
 #include "pinout.h"
 #include "usb.h"
@@ -56,7 +57,6 @@ typedef enum
 int32_t i2s_fft_sample_arr[I2S_HW_COMPLEX_SAMPLE_COUNT * 2];
 static volatile uint32_t s_fft_sample_cnt = 0U;
 
-void DMA1_Channel4_IRQHandler(void) __attribute__((interrupt));
 extern void audio_usb_mic_write(volatile uint16_t const *src_words, size_t word_count);
 
 static_assert(sizeof(uintptr_t) <= sizeof(uint32_t),
@@ -455,7 +455,7 @@ void i2s_hw_enable(FunctionalState state)
     i2s_dma_rx_start();
 }
 
-void DMA1_Channel4_IRQHandler(void)
+PORT_ISR_BODY(DMA1_Channel4_IRQHandler)
 {
     BaseType_t higher_priority_task_woken = pdFALSE;
 

ch32v305/src/hw/si5351.c

@@ -1,5 +1,6 @@
 #include "si5351.h"
 
+#include "freertos/port_isr.h"
 #include "hw/i2c.h"
 
 #include "debug.h"
@@ -610,7 +611,7 @@ static void si5351_tim6_handle_finalize_wait(void)
     si5351_tim6_finalize_clk1();
 }
 
-__attribute__((interrupt)) void TIM6_IRQHandler(void)
+PORT_ISR_BODY(TIM6_IRQHandler)
 {
     if(TIM_GetITStatus(TIM6, TIM_IT_Update) == RESET)
     {