Merge branch 'feature/tls_support' into 'master'

FreeRTOS TLS support

See merge request idf/esp-idf!1902

components/esp32/ld/esp32.common.ld

@@ -194,6 +194,13 @@ SECTIONS
     *(.gnu.linkonce.lit4.*)
     _lit4_end = ABSOLUTE(.);
     . = ALIGN(4);
+    _thread_local_start = ABSOLUTE(.);
+    *(.tdata)
+    *(.tdata.*)
+    *(.tbss)
+    *(.tbss.*)
+    _thread_local_end = ABSOLUTE(.);
+    . = ALIGN(4);
   } >drom0_0_seg
 
   .flash.text :

components/freertos/include/freertos/task.h

@@ -380,6 +380,9 @@ is used in assert() statements. */
  * @return pdPASS if the task was successfully created and added to a ready
  * list, otherwise an error code defined in the file projdefs.h
  *
+ * @note If program uses thread local variables (ones specified with "__thread" keyword)
+ * then storage for them will be allocated on the task's stack.
+ *
  * Example usage:
  * @code{c}
  *  // Task to be created.
@@ -530,6 +533,9 @@ is used in assert() statements. */
  * are NULL then the task will not be created and
  * errCOULD_NOT_ALLOCATE_REQUIRED_MEMORY is returned.
  *
+ * @note If program uses thread local variables (ones specified with "__thread" keyword)
+ * then storage for them will be allocated on the task's stack.
+ *
  * Example usage:
  * @code{c}
  *

components/freertos/port.c

@@ -92,6 +92,7 @@
 */
 
 #include <stdlib.h>
+#include <string.h>
 #include <xtensa/config/core.h>
 
 #include "xtensa_rtos.h"
@@ -146,9 +147,24 @@ StackType_t *pxPortInitialiseStack( StackType_t *pxTopOfStack, TaskFunction_t px
 	#if XCHAL_CP_NUM > 0
 	uint32_t *p;
 	#endif
+	uint32_t *threadptr;
+	void *task_thread_local_start;
+	extern int _thread_local_start, _thread_local_end, _rodata_start;
+	// TODO: check that TLS area fits the stack
+	uint32_t thread_local_sz = (uint8_t *)&_thread_local_end - (uint8_t *)&_thread_local_start;
 
-	/* Create interrupt stack frame aligned to 16 byte boundary */
-	sp = (StackType_t *) (((UBaseType_t)(pxTopOfStack + 1) - XT_CP_SIZE - XT_STK_FRMSZ) & ~0xf);
+	thread_local_sz = ALIGNUP(0x10, thread_local_sz);
+
+	/* Initialize task's stack so that we have the following structure at the top:
+
+		----LOW ADDRESSES ----------------------------------------HIGH ADDRESSES----------
+		 task stack | interrupt stack frame | thread local vars | co-processor save area |
+		----------------------------------------------------------------------------------
+					|																	 |
+					SP 																pxTopOfStack
+
+		All parts are aligned to 16 byte boundary. */
+	sp = (StackType_t *) (((UBaseType_t)(pxTopOfStack + 1) - XT_CP_SIZE - thread_local_sz - XT_STK_FRMSZ) & ~0xf);
 
 	/* Clear the entire frame (do not use memset() because we don't depend on C library) */
 	for (tp = sp; tp <= pxTopOfStack; ++tp)
@@ -178,6 +194,14 @@ StackType_t *pxPortInitialiseStack( StackType_t *pxTopOfStack, TaskFunction_t px
 	frame->vpri = 0xFFFFFFFF;
 	#endif
 
+	/* Init threadptr reg and TLS vars */
+	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);
+	/* shift threadptr by the offset of _thread_local_start from DROM start;
+	   need to take into account extra 16 bytes offset */
+	*threadptr = (uint32_t)task_thread_local_start - ((uint32_t)&_thread_local_start - (uint32_t)&_rodata_start) - 0x10;
+
 	#if XCHAL_CP_NUM > 0
 	/* Init the coprocessor save area (see xtensa_context.h) */
 	/* No access to TCB here, so derive indirectly. Stack growth is top to bottom.

components/freertos/test/test_thread_local.c

@@ -0,0 +1,105 @@
+/*
+ Test for thread local storage support.
+*/
+
+#include <string.h>
+#include <esp_types.h>
+
+#include "freertos/FreeRTOS.h"
+#include "freertos/task.h"
+#include "unity.h"
+#include "sdkconfig.h"
+
+static __thread int tl_test_var1;
+static __thread uint8_t tl_test_var2 = 55;
+static __thread uint16_t tl_test_var3 = 44;
+static __thread uint8_t tl_test_arr_var[10];
+static __thread struct test_tls_var {
+    int f32;
+    uint8_t f8;
+    uint16_t f16;
+    uint8_t farr[10];
+} tl_test_struct_var;
+
+static void task_test_tls(void *arg)
+{
+    bool *running = (bool *)arg;
+    uint32_t tp = (uint32_t)-1;
+    int test_var1_old = 0;
+    uint8_t test_var2_old = 0;
+    uint16_t test_var3_old = 0;
+    int f32_old = 0;
+    uint8_t f8_old = 0;
+    uint16_t f16_old = 0;
+
+    asm volatile ("rur.threadptr %0":"=r"(tp));
+    for (int i = 0; i < 5; i++) {
+        printf("Task[%x]: var = 0x%x 0x%x\n", tp, tl_test_var1, tl_test_var2);
+        if (i == 0) {
+            TEST_ASSERT_EQUAL(0, tl_test_var1);
+            TEST_ASSERT_EQUAL(55, tl_test_var2);
+            TEST_ASSERT_EQUAL(44, tl_test_var3);
+            for (int k = 0; k < sizeof(tl_test_arr_var); k++) {
+                TEST_ASSERT_EQUAL(0, tl_test_arr_var[k]);
+            }
+            TEST_ASSERT_EQUAL(0, tl_test_struct_var.f32);
+            TEST_ASSERT_EQUAL(0, tl_test_struct_var.f8);
+            TEST_ASSERT_EQUAL(0, tl_test_struct_var.f16);
+            for (int k = 0; k < sizeof(tl_test_struct_var.farr); k++) {
+                TEST_ASSERT_EQUAL(0, tl_test_struct_var.farr[k]);
+            }
+        } else {
+            TEST_ASSERT_EQUAL(test_var1_old+1, tl_test_var1);
+            TEST_ASSERT_EQUAL(test_var2_old+1, tl_test_var2);
+            TEST_ASSERT_EQUAL(test_var3_old+1, tl_test_var3);
+            for (int k = 0; k < sizeof(tl_test_arr_var); k++) {
+                TEST_ASSERT_EQUAL(i-1, tl_test_arr_var[k]);
+            }
+            TEST_ASSERT_EQUAL(f32_old+1, tl_test_struct_var.f32);
+            TEST_ASSERT_EQUAL(f8_old+1, tl_test_struct_var.f8);
+            TEST_ASSERT_EQUAL(f16_old+1, tl_test_struct_var.f16);
+            for (int k = 0; k < sizeof(tl_test_struct_var.farr); k++) {
+                TEST_ASSERT_EQUAL(i-1, tl_test_struct_var.farr[k]);
+            }
+        }
+        test_var1_old = tl_test_var1;
+        test_var2_old = tl_test_var2;
+        test_var3_old = tl_test_var3;
+        f32_old = tl_test_struct_var.f32;
+        f8_old = tl_test_struct_var.f8;
+        f16_old = tl_test_struct_var.f16;
+        tl_test_var1++;
+        tl_test_var2++;
+        tl_test_var3++;
+        memset(tl_test_arr_var, i, sizeof(tl_test_arr_var));
+        tl_test_struct_var.f32++;
+        tl_test_struct_var.f8++;
+        tl_test_struct_var.f16++;
+        memset(tl_test_struct_var.farr, i, sizeof(tl_test_struct_var.farr));
+        vTaskDelay(10);
+    }
+
+    if (running) {
+        *running = false;
+        vTaskDelete(NULL);
+    }
+}
+
+TEST_CASE("TLS test", "[freertos]")
+{
+    static StackType_t s_stack[2048];
+    StaticTask_t s_task;
+    bool running[2] = {true, true};
+#if CONFIG_FREERTOS_UNICORE == 0
+    int other_core = 1;
+#else
+    int other_core = 0;
+#endif
+
+    xTaskCreatePinnedToCore((TaskFunction_t)&task_test_tls, "task_test_tls", 3072, &running[0], 5, NULL, 0);
+    xTaskCreateStaticPinnedToCore((TaskFunction_t)&task_test_tls, "task_test_tls", sizeof(s_stack), &running[1],
+        5, s_stack, &s_task, other_core);
+    while (running[0] || running[1]) {
+        vTaskDelay(10);
+    }
+}

docs/api-guides/index.rst

@@ -10,6 +10,7 @@ API Guides
    ESP32 Core Dump <core_dump>
    Flash Encryption <../security/flash-encryption>
    FreeRTOS SMP Changes <freertos-smp>
+   Thread Local Storage <thread-local-storage>
    High Level Interrupts <hlinterrupts>
    JTAG Debugging <jtag-debugging/index>
    Partition Tables <partition-tables>

docs/api-guides/thread-local-storage.rst

@@ -0,0 +1,65 @@
+Thread Local Storage
+====================
+
+Overview
+--------
+
+Thread-local storage (TLS) is a mechanism by which variables are allocated such that there 
+is one instance of the variable per extant thread. ESP-IDF provides three ways to make use 
+of such variables:
+
+ - :ref:`freertos-native`: ESP-IDF FreeRTOS native API.
+ - :ref:`pthread-api`: ESP-IDF's pthread API.
+ - :ref:`c11-std`: C11 standard introduces special keyword to declare variables as thread local.
+
+.. _freertos-native:
+
+FreeRTOS Native API
+--------------------
+
+The ESP-IDF FreeRTOS provides the following API to manage thread local variables:
+
+ - :cpp:func:`vTaskSetThreadLocalStoragePointer`
+ - :cpp:func:`pvTaskGetThreadLocalStoragePointer`
+ - :cpp:func:`vTaskSetThreadLocalStoragePointerAndDelCallback`
+
+In this case maximum number of variables that can be allocated is limited by
+``configNUM_THREAD_LOCAL_STORAGE_POINTERS`` macro. Variables are kept in the task control block (TCB)
+and accessed by their index. Note that index 0 is reserved for ESP-IDF internal uses.
+Using that API user can allocate thread local variables of an arbitrary size and assign them to any number of tasks.
+Different tasks can have different sets of TLS variables.
+If size of the variable is more then 4 bytes then user is responsible for allocating/deallocating memory for it. 
+Variable's deallocation is initiated by FreeRTOS when task is deleted, but user must provide function (callback) 
+to do proper cleanup.
+
+.. _pthread-api:
+
+Pthread API
+----------------
+
+The ESP-IDF provides the following pthread API to manage thtread local variables:
+
+ - :cpp:func:`pthread_key_create`
+ - :cpp:func:`pthread_key_delete`
+ - :cpp:func:`pthread_getspecific`
+ - :cpp:func:`pthread_setspecific`
+
+This API has all benefits of the one above, but eliminates some its limits. The number of variables is
+limited only by size of available memory on the heap.
+Due to the dynamic nature this API introduces additional performance overhead compared to the native one.
+
+.. _c11-std:
+
+C11 Standard
+------------
+
+The ESP-IDF FreeRTOS supports thread local variables according to C11 standard (ones specified with ``__thread`` keyword).
+For details on this GCC feature please see https://gcc.gnu.org/onlinedocs/gcc-5.5.0/gcc/Thread-Local.html#Thread-Local.
+Storage for that kind of variables is allocated on the task's stack.
+Note that area for all such variables in the program will be allocated on the stack of
+every task in the system even if that task does not use such variables at all. For example
+ESP-IDF system tasks (like ``ipc``, ``timer`` tasks etc.) will also have that extra stack space allocated.
+So this feature should be used with care. There is a tradeoff: C11 thread local variables are quite handy
+to use in programming and can be accessed using just a few Xtensa instructions, but this benefit goes
+with the cost of additional stack usage for all tasks in the system.
+Due to static nature of variables allocation all tasks in the system have the same sets of C11 thread local variables.