Merge branch 'bugfix/tls_threadptr_calculation' into 'master'

freertos: fix TLS run-time address calculation

See merge request espressif/esp-idf!13271

components/esp32/ld/esp32.project.ld.in

@@ -232,7 +232,6 @@ SECTIONS
   ASSERT(((_bss_end - ORIGIN(dram0_0_seg)) <= LENGTH(dram0_0_seg)),
           "DRAM segment data does not fit.")
 
-  /* When modifying the alignment, update tls_section_alignment in pxPortInitialiseStack */
   .flash.appdesc : ALIGN(0x10)
   {
     _rodata_start = ABSOLUTE(.);
@@ -248,6 +247,8 @@ SECTIONS
 
   .flash.rodata : ALIGN(0x10)
   {
+    _flash_rodata_start = ABSOLUTE(.);
+
     mapping[flash_rodata]
 
 
@@ -305,6 +306,8 @@ SECTIONS
     . = ALIGN(4);
   } >default_rodata_seg
 
+  _flash_rodata_align = ALIGNOF(.flash.rodata);
+
   .flash.text :
   {
     _stext = .;

components/esp32c3/ld/esp32c3.project.ld.in

@@ -262,7 +262,6 @@ SECTIONS
     _rodata_reserved_start = .;
   } > default_rodata_seg
 
-  /* When modifying the alignment, don't forget to update tls_section_alignment in pxPortInitialiseStack */
   .flash.appdesc : ALIGN(0x10)
   {
     _rodata_start = ABSOLUTE(.);
@@ -278,6 +277,8 @@ SECTIONS
 
   .flash.rodata : ALIGN(0x10)
   {
+    _flash_rodata_start = ABSOLUTE(.);
+
     mapping[flash_rodata]
 
     *(.irom1.text) /* catch stray ICACHE_RODATA_ATTR */

components/esp32s2/ld/esp32s2.project.ld.in

@@ -251,7 +251,6 @@ SECTIONS
     _bss_end = ABSOLUTE(.);
   } > dram0_0_seg
 
-  /* When modifying the alignment, update tls_section_alignment in pxPortInitialiseStack */
   .flash.appdesc : ALIGN(0x10)
   {
     _rodata_reserved_start = ABSOLUTE(.);
@@ -268,6 +267,8 @@ SECTIONS
 
   .flash.rodata : ALIGN(0x10)
   {
+    _flash_rodata_start = ABSOLUTE(.);
+
     mapping[flash_rodata]
 
     *(.irom1.text) /* catch stray ICACHE_RODATA_ATTR */
@@ -324,6 +325,8 @@ SECTIONS
     . = ALIGN(4);
   } >default_rodata_seg
 
+  _flash_rodata_align = ALIGNOF(.flash.rodata);
+
   .flash.text :
   {
     _stext = .;

components/esp32s3/ld/esp32s3.project.ld.in

@@ -293,7 +293,6 @@ SECTIONS
     _rodata_reserved_start = .;
   } > default_rodata_seg
 
-  /* When modifying the alignment, don't forget to update tls_section_alignment in pxPortInitialiseStack */
   .flash.appdesc : ALIGN(0x10)
   {
     _rodata_start = ABSOLUTE(.);
@@ -309,6 +308,8 @@ SECTIONS
 
   .flash.rodata : ALIGN(0x10)
   {
+    _flash_rodata_start = ABSOLUTE(.);
+
     mapping[flash_rodata]
 
     *(.irom1.text) /* catch stray ICACHE_RODATA_ATTR */
@@ -362,6 +363,8 @@ SECTIONS
     . = ALIGN(4);
   } > default_rodata_seg
 
+  _flash_rodata_align = ALIGNOF(.flash.rodata);
+
   /* Marks the end of IRAM code segment */
   .iram0.text_end (NOLOAD) :
   {

components/freertos/port/riscv/port.c

@@ -224,18 +224,58 @@ StackType_t *pxPortInitialiseStack(StackType_t *pxTopOfStack, TaskFunction_t pxC
     extern uint32_t __global_pointer$;
     uint8_t* task_thread_local_start;
     uint8_t* threadptr;
-    extern char _thread_local_start, _thread_local_end, _rodata_start;
+    extern char _thread_local_start, _thread_local_end, _flash_rodata_start;
 
     /* Byte pointer, so that subsequent calculations don't depend on sizeof(StackType_t). */
     uint8_t* sp = (uint8_t*) pxTopOfStack;
 
-    /* Set up TLS area */
+    /* Set up TLS area.
+     * The following diagram illustrates the layout of link-time and run-time
+     * TLS sections.
+     *
+     *          +-------------+
+     *          |Section:     |      Linker symbols:
+     *          |.flash.rodata|      ---------------
+     *       0x0+-------------+ <-- _flash_rodata_start
+     *        ^ |             |
+     *        | | Other data  |
+     *        | |     ...     |
+     *        | +-------------+ <-- _thread_local_start
+     *        | |.tbss        | ^
+     *        v |             | |
+     *    0xNNNN|int example; | | (thread_local_size)
+     *          |.tdata       | v
+     *          +-------------+ <-- _thread_local_end
+     *          | Other data  |
+     *          |     ...     |
+     *          |             |
+     *          +-------------+
+     *
+     *                                Local variables of
+     *                              pxPortInitialiseStack
+     *                             -----------------------
+     *          +-------------+ <-- pxTopOfStack
+     *          |.tdata (*)   |  ^
+     *        ^ |int example; |  |(thread_local_size
+     *        | |             |  |
+     *        | |.tbss (*)    |  v
+     *        | +-------------+ <-- task_thread_local_start
+     * 0xNNNN | |             |  ^
+     *        | |             |  |
+     *        | |             |  |_thread_local_start - _rodata_start
+     *        | |             |  |
+     *        | |             |  v
+     *        v +-------------+ <-- threadptr
+     *
+     *   (*) The stack grows downward!
+     */
+
     uint32_t thread_local_sz = (uint32_t) (&_thread_local_end - &_thread_local_start);
     thread_local_sz = ALIGNUP(0x10, thread_local_sz);
     sp -= thread_local_sz;
     task_thread_local_start = sp;
     memcpy(task_thread_local_start, &_thread_local_start, thread_local_sz);
-    threadptr = task_thread_local_start - (&_thread_local_start - &_rodata_start);
+    threadptr = task_thread_local_start - (&_thread_local_start - &_flash_rodata_start);
 
     /* Simulate the stack frame as it would be created by a context switch interrupt. */
     sp -= RV_STK_FRMSZ;

components/freertos/port/xtensa/port.c

@@ -189,7 +189,7 @@ StackType_t *pxPortInitialiseStack( StackType_t *pxTopOfStack, TaskFunction_t px
 	#endif
 	uint32_t *threadptr;
 	void *task_thread_local_start;
-	extern int _thread_local_start, _thread_local_end, _rodata_start;
+	extern int _thread_local_start, _thread_local_end, _flash_rodata_start, _flash_rodata_align;
 	// TODO: check that TLS area fits the stack
 	uint32_t thread_local_sz = (uint8_t *)&_thread_local_end - (uint8_t *)&_thread_local_start;
 
@@ -248,24 +248,25 @@ StackType_t *pxPortInitialiseStack( StackType_t *pxTopOfStack, TaskFunction_t px
 	frame->vpri = 0xFFFFFFFF;
 	#endif
 
-	/* Init threadptr reg and TLS vars */
+	/* Init threadptr register and set up TLS run-time area.
+	 * The diagram in port/riscv/port.c illustrates the calculations below.
+	 */
 	task_thread_local_start = (void *)(((uint32_t)pxTopOfStack - XT_CP_SIZE - thread_local_sz) & ~0xf);
 	memcpy(task_thread_local_start, &_thread_local_start, thread_local_sz);
 	threadptr = (uint32_t *)(sp + XT_STK_EXTRA);
-	/* Calculate THREADPTR value:
+	/* Calculate THREADPTR value.
 	 * The generated code will add THREADPTR value to a constant value determined at link time,
 	 * to get the address of the TLS variable.
 	 * The constant value is calculated by the linker as follows
 	 * (search for 'tpoff' in elf32-xtensa.c in BFD):
 	 *    offset = address - tls_section_vma + align_up(TCB_SIZE, tls_section_alignment)
-	 * where TCB_SIZE is hardcoded to 8. There doesn't seem to be a way to propagate
+	 * where TCB_SIZE is hardcoded to 8.
-	 * the section alignment value from the ld script into the code, so it is hardcoded
+	 * Note this is slightly different compared to the RISC-V port, where offset = address - tls_section_vma.
-	 * in both places.
 	 */
-	const uint32_t tls_section_alignment = 0x10;  /* has to be in sync with ALIGN value of .flash.rodata section */
+	const uint32_t tls_section_alignment = (uint32_t) &_flash_rodata_align;  /* ALIGN value of .flash.rodata section */
 	const uint32_t tcb_size = 8; /* Unrelated to FreeRTOS, this is the constant from BFD */
 	const uint32_t base = (tcb_size + tls_section_alignment - 1) & (~(tls_section_alignment - 1));
-	*threadptr = (uint32_t)task_thread_local_start - ((uint32_t)&_thread_local_start - (uint32_t)&_rodata_start) - base;
+	*threadptr = (uint32_t)task_thread_local_start - ((uint32_t)&_thread_local_start - (uint32_t)&_flash_rodata_start) - base;
 
 	#if XCHAL_CP_NUM > 0
 	/* Init the coprocessor save area (see xtensa_context.h) */