diff options
Diffstat (limited to 'libcr/coroutine.c')
| -rw-r--r-- | libcr/coroutine.c | 681 |
1 files changed, 456 insertions, 225 deletions
diff --git a/libcr/coroutine.c b/libcr/coroutine.c index 41d987e..182e85c 100644 --- a/libcr/coroutine.c +++ b/libcr/coroutine.c @@ -1,57 +1,70 @@ /* libcr/coroutine.c - Simple embeddable coroutine implementation * - * Copyright (C) 2024 Luke T. Shumaker <lukeshu@lukeshu.com> + * Copyright (C) 2024-2025 Luke T. Shumaker <lukeshu@lukeshu.com> * SPDX-License-Identifier: AGPL-3.0-or-later */ -#include <assert.h> -#include <setjmp.h> /* for setjmp(), longjmp(), jmp_buf */ #include <stdint.h> /* for uint8_t */ -#include <stdio.h> /* for printf(), fprintf(), stderr */ #include <stdlib.h> /* for aligned_alloc(), free() */ +#include <string.h> /* for strncpy(), memset() */ + +#include <libmisc/assert.h> +#include <libmisc/macro.h> + +#define LOG_NAME COROUTINE +#include <libmisc/log.h> #include <libcr/coroutine.h> +#undef COROUTINE /* Configuration **************************************************************/ #include "config.h" -#ifndef CONFIG_COROUTINE_DEFAULT_STACK_SIZE - #error config.h must define CONFIG_COROUTINE_DEFAULT_STACK_SIZE +#ifndef CONFIG_COROUTINE_NAME_LEN + #error config.h must define CONFIG_COROUTINE_NAME_LEN (non-negative integer) #endif #ifndef CONFIG_COROUTINE_NUM - #error config.h must define CONFIG_COROUTINE_NUM + #error config.h must define CONFIG_COROUTINE_NUM (non-negative integer) #endif #ifndef CONFIG_COROUTINE_MEASURE_STACK - #error config.h must define CONFIG_COROUTINE_MEASURE_STACK + #error config.h must define CONFIG_COROUTINE_MEASURE_STACK (bool) #endif #ifndef CONFIG_COROUTINE_PROTECT_STACK - #error config.h must define CONFIG_COROUTINE_PROTECT_STACK + #error config.h must define CONFIG_COROUTINE_PROTECT_STACK (bool) #endif #ifndef CONFIG_COROUTINE_DEBUG - #error config.h must define CONFIG_COROUTINE_DEBUG + #error config.h must define CONFIG_COROUTINE_DEBUG (bool) #endif #ifndef CONFIG_COROUTINE_VALGRIND - #error config.h must define CONFIG_COROUTINE_VALGRIND + #error config.h must define CONFIG_COROUTINE_VALGRIND (bool) +#endif +#ifndef CONFIG_COROUTINE_GDB + #error config.h must define CONFIG_COROUTINE_GDB (bool) #endif +/* Enforce that CONFIG_COROUTINE_NUM is greater than 1, to work around + * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118212 */ +static_assert(CONFIG_COROUTINE_NUM > 1); + +/* Implementation *************************************************************/ + #if CONFIG_COROUTINE_VALGRIND #include <valgrind/valgrind.h> #endif -/* Implementation *************************************************************/ - /* * Portability notes: * - * - It uses GCC `__attribute__`s, and the GNUC ({ ... }) statement - * exprs extension. + * - It uses GCC `gnu::` attributes, and the GNUC `({ ... })` + * statement exprs extension. * * - It has a small bit of platform-specific code in the "platform * support" section. Other than this, it should be portable to * other platforms CPUs. It currently contains implementations for - * __x86_64__ (assumes POSIX) and __arm__ (assumes bare-metal), and - * should be fairly easy to add implementations for other platforms. + * __unix__ and __ARM_EABI__ "operating systems" on __x86_64__ and + * __ARM_ARCH_6M__ CPUs, and should be fairly easy to add + * implementations for other platforms. * * - It uses setjmp()/longjmp() in "unsafe" ways. POSIX-2017 * longjmp(3p) says @@ -117,81 +130,151 @@ * no longer exists. */ -#define ALWAYS_INLINE inline __attribute__((always_inline)) - /* platform support ***********************************************************/ /* As part of sbc-harness, this only really needs to support ARM-32, but being * able to run it on my x86-64 GNU/Linux laptop is useful for debugging. */ #define CR_PLAT_STACK_ALIGNMENT \ - ({ __attribute__((aligned)) void fn(void) {}; __alignof__(fn); }) + ({ [[gnu::aligned]] void fn(void) {}; __alignof__(fn); }) #if 0 { /* bracket to get Emacs indentation to work how I want */ #endif /*==================================================================== - * Wrappers for setjmp()/longjmp() that do *not* save the - * interrupt mask. */ + * Interrupt management routines. */ #if __unix__ - /* On a *NIX OS, we use signals as interrupts. POSIX leaves - * it implementation-defined whether setjmp()/longjmp() save - * the signal mask; while glibc does not save it, let's not - * rely on that. */ - #define cr_plat_setjmp(env) sigsetjmp(env, 0) - #define cr_plat_longjmp(env, val) siglongjmp(env, val) -#elif __NEWLIB__ - /* newlib does not have sigsetjmp()/sigsetlongjmp(), but - * setjmp()/longjmp() do not save the interrupt mask, * so we - * can use them directly. */ - #define cr_plat_setjmp(env) setjmp(env) - #define cr_plat_longjmp(env, val) longjmp(env, val) -#else - #error unsupported platform (not __unix__, not __NEWLIB__) + #include <signal.h> /* for sig*, SIG* */ + + /* For a signal to be *in* the mask means that the signal is + * *blocked*. */ + + #define _CR_SIG_SENTINEL SIGURG + #if CONFIG_COROUTINE_GDB + #define _CR_SIG_GDB SIGWINCH + #endif + +#if CONFIG_COROUTINE_VALGRIND + /* Hack around a bug in Valgrind where it runs the + * sigsuspend(tmp_sigmask)-triggered handler function with the + * original mask, not the `tmp_mask`. */ + static bool _cr_plat_in_sigsuspend = false; #endif -/*==================================================================== - * Interrupt management routines. */ -#if __unix__ - #include <signal.h> /* for sig*, SIG_* */ - #include <unistd.h> /* for pause() */ + bool cr_plat_is_in_intrhandler(void) { +#if CONFIG_COROUTINE_VALGRIND + if (_cr_plat_in_sigsuspend) + return true; +#endif + sigset_t cur_mask; + sigprocmask(0, NULL, &cur_mask); + if (sigismember(&cur_mask, _CR_SIG_SENTINEL)) + /* Interrupts are disabled, so we cannot be in + * an interrupt handler. */ + return false; + for (int sig = SIGRTMIN; sig <= SIGRTMAX; sig++) + if (sigismember(&cur_mask, sig)) + return true; + return false; + } + static inline bool _cr_plat_are_interrupts_enabled(void) { + assert(!cr_plat_is_in_intrhandler()); + sigset_t cur_mask; + sigfillset(&cur_mask); + sigprocmask(0, NULL, &cur_mask); + return !sigismember(&cur_mask, _CR_SIG_SENTINEL); + } static inline void cr_plat_wait_for_interrupt(void) { - pause(); + assert(!cr_plat_is_in_intrhandler()); + assert(!_cr_plat_are_interrupts_enabled()); + sigset_t set; + sigemptyset(&set); +#if CONFIG_COROUTINE_VALGRIND + _cr_plat_in_sigsuspend = true; +#endif + sigsuspend(&set); +#if CONFIG_COROUTINE_VALGRIND + _cr_plat_in_sigsuspend = false; +#endif } - void _cr_plat_disable_interrupts(void) { - sigset_t all; + bool _cr_plat_save_and_disable_interrupts(void) { + assert(!cr_plat_is_in_intrhandler()); + sigset_t all, old; sigfillset(&all); - sigprocmask(SIG_BLOCK, &all, NULL); + sigprocmask(SIG_SETMASK, &all, &old); + return !sigismember(&old, _CR_SIG_SENTINEL); } void _cr_plat_enable_interrupts(void) { - sigset_t all; - sigfillset(&all); - sigprocmask(SIG_UNBLOCK, &all, NULL); + assert(!cr_plat_is_in_intrhandler()); + assert(!_cr_plat_are_interrupts_enabled()); + sigset_t zero; + sigemptyset(&zero); + sigprocmask(SIG_SETMASK, &zero, NULL); } -#elif __arm__ - /* Assume bare-metal if !__unix__. */ - static ALWAYS_INLINE void cr_plat_wait_for_interrupt(void) { - asm volatile ("wfi":::"memory"); + #if CONFIG_COROUTINE_GDB + static void _cr_gdb_intrhandler(int LM_UNUSED(sig)) {} + #endif + static void cr_plat_init(void) { + #if CONFIG_COROUTINE_GDB + int r; + struct sigaction action = { + .sa_handler = _cr_gdb_intrhandler, + }; + r = sigaction(_CR_SIG_GDB, &action, NULL); + assert(r == 0); + #endif + } +#elif __ARM_ARCH_6M__ && __ARM_EABI__ + bool cr_plat_is_in_intrhandler(void) { + uint32_t isr_number; + asm volatile ("mrs %0, ipsr" + : /* %0 */"=l"(isr_number) + ); + return isr_number != 0; } - void _cr_plat_disable_interrupts(void) { - asm volatile ("cpsid i":::"memory"); + LM_ALWAYS_INLINE static bool _cr_plat_are_interrupts_enabled(void) { + assert(!cr_plat_is_in_intrhandler()); + uint32_t primask; + asm volatile ("mrs %0, PRIMASK" + : /* %0 */"=l"(primask) + ); + return primask == 0; + } + + LM_ALWAYS_INLINE static void cr_plat_wait_for_interrupt(void) { + assert(!cr_plat_is_in_intrhandler()); + assert(!_cr_plat_are_interrupts_enabled()); + asm volatile ("wfi\n" + "cpsie i\n" + "isb\n" + "cpsid i" + :::"memory"); + } + bool _cr_plat_save_and_disable_interrupts(void) { + assert(!cr_plat_is_in_intrhandler()); + bool were_enabled = _cr_plat_are_interrupts_enabled(); + asm volatile ("cpsid i"); + return were_enabled; } void _cr_plat_enable_interrupts(void) { - asm volatile ("cpsie i":::"memory"); + assert(!cr_plat_is_in_intrhandler()); + assert(!_cr_plat_are_interrupts_enabled()); + asm volatile ("cpsie i"); } + static void cr_plat_init(void) {} #else - #error unsupported platform (not __unix__, not bare-metal __arm__) + #error unsupported platform (not __unix__, not __ARM_ARCH_6M__ && __ARM_EABI__) #endif /*==================================================================== * Stack management routines. */ -#if __arm__ +#if __ARM_ARCH_6M__ #define CR_PLAT_STACK_GROWS_DOWNWARD 1 #if CONFIG_COROUTINE_MEASURE_STACK - static ALWAYS_INLINE uintptr_t cr_plat_get_sp(void) { + LM_ALWAYS_INLINE static uintptr_t cr_plat_get_sp(void) { uintptr_t sp; asm volatile ("mov %0, sp":"=r"(sp)); return sp; @@ -209,13 +292,13 @@ "mov sp, %1\n\t" /* [sp = stack] */ "mov r0, %3\n\t" /* [arg0 = args] */ "blx %2\n\t" /* [fn()] */ - "ldr r0, %0\n\t" /* [sp = staved_sp */ + "ldr r0, %0\n\t" /* [sp = saved_sp */ "mov sp, r0" /* ] */ : : /* %0 */"m"(saved_sp), - /* %1 */"r"(stack), - /* %2 */"r"(fn), - /* %3 */"r"(args) + /* %1 */"l"(stack), + /* %2 */"l"(fn), + /* %3 */"l"(args) : "r0" ); } @@ -223,7 +306,7 @@ #define CR_PLAT_STACK_GROWS_DOWNWARD 1 #if CONFIG_COROUTINE_MEASURE_STACK - static ALWAYS_INLINE uintptr_t cr_plat_get_sp(void) { + LM_ALWAYS_INLINE static uintptr_t cr_plat_get_sp(void) { uintptr_t sp; asm volatile ("movq %%rsp, %0":"=r"(sp)); return sp; @@ -247,46 +330,79 @@ ); } #else - #error unsupported platform (not __arm__, not __x86__) + #error unsupported CPU (not __ARM_ARCH_6M__, not __x86_64__) +#endif + +/*==================================================================== + * Wrappers for setjmp()/longjmp() that: + * 1. Allow us to inspect the buffer. + * 2. Do *not* save the interrupt mask. + */ + #include <setjmp.h> /* for setjmp(), longjmp(), jmp_buf */ + typedef struct { + jmp_buf raw; + #if CONFIG_COROUTINE_MEASURE_STACK + /* We aught to be able to get sp out of the raw + * `jmp_buf`, but libc authors insist on jmp_buf being + * opaque, glibc going as far as to xor it with a + * secret to obfuscate it! */ + uintptr_t sp; + #endif + } cr_plat_jmp_buf; + static void _cr_plat_setjmp_pre(cr_plat_jmp_buf *env [[maybe_unused]]) { + #if CONFIG_COROUTINE_MEASURE_STACK + env->sp = cr_plat_get_sp(); + #endif + } + #if CONFIG_COROUTINE_MEASURE_STACK + static uintptr_t cr_plat_setjmp_get_sp(cr_plat_jmp_buf *env) { return env->sp; } + #endif + /* cr_plat_setjmp *NEEDS* to be a preprocessor macro rather + * than a real function, because [[gnu::returns_twice]] + * doesn't work. + * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117469 */ +#if __unix__ + /* On __unix__, we use POSIX real-time signals as interrupts. + * POSIX leaves it implementation-defined whether + * setjmp()/longjmp() save the signal mask; while glibc does + * not save it, let's not rely on that. */ + #define cr_plat_setjmp(env) ({ _cr_plat_setjmp_pre(env); sigsetjmp((env)->raw, 0); }) + [[noreturn]] static void cr_plat_longjmp(cr_plat_jmp_buf *env, int val) { siglongjmp(env->raw, val); } +#elif __NEWLIB__ + /* newlib does not have sigsetjmp()/sigsetlongjmp(), but + * setjmp()/longjmp() do not save the interrupt mask, so we + * can use them directly. */ + #define cr_plat_setjmp(env) ({ _cr_plat_setjmp_pre(env); setjmp((env)->raw); }) + [[noreturn]] static void cr_plat_longjmp(cr_plat_jmp_buf *env, int val) { longjmp(env->raw, val); } +#else + #error unsupported platform (not __unix__, not __NEWLIB__) #endif #if 0 } #endif -/* preprocessor macros ********************************************************/ - -/** Return `n` rounded up to the nearest multiple of `d` */ -#define ROUND_UP(n, d) ( ( ((n)+(d)-1) / (d) ) * (d) ) -#define ARRAY_LEN(arr) (sizeof(arr)/sizeof((arr)[0])) -#define NEXT_POWER_OF_2(x) ((1ULL)<<((sizeof(unsigned long long)*8)-__builtin_clzll(x))) - /* types **********************************************************************/ -enum coroutine_state { - CR_NONE = 0, /* this slot in the table is empty */ - CR_INITIALIZING, /* running, before cr_begin() */ - CR_RUNNING, /* running, after cr_begin() */ - CR_PRE_RUNNABLE, /* running, after cr_unpause_from_intrhandler() - * but before cr_pause() */ - CR_RUNNABLE, /* not running, but runnable */ - CR_PAUSED, /* not running, and not runnable */ -}; - struct coroutine { + /* 1. state *************************************************/ volatile enum coroutine_state state; - jmp_buf env; -#if CONFIG_COROUTINE_MEASURE_STACK - /* We aught to be able to get this out of .env, but libc - * authors insist on jmp_buf being opaque, glibc going as far - * as to xor it with a secret ot obfuscate it! */ - uintptr_t sp; -#endif + + /* 2. name **************************************************/ + [[gnu::nonstring]] char name[CONFIG_COROUTINE_NAME_LEN]; + + /* 3. stack *************************************************/ + /* stack_size *includes* CR_STACK_GUARD at each end. */ size_t stack_size; + /* stack is the bottom of the CR_STACK_GUARD at the bottom of the stack. */ void *stack; #if CONFIG_COROUTINE_VALGRIND unsigned stack_id; #endif + bool stack_free; + + /* 4. env ***************************************************/ + cr_plat_jmp_buf env; }; /* constants ******************************************************************/ @@ -295,7 +411,6 @@ const char *coroutine_state_strs[] = { [CR_NONE] = "CR_NONE", [CR_INITIALIZING] = "CR_INITIALIZING", [CR_RUNNING] = "CR_RUNNING", - [CR_PRE_RUNNABLE] = "CR_PRE_RUNNABLE", [CR_RUNNABLE] = "CR_RUNNABLE", [CR_PAUSED] = "CR_PAUSED", }; @@ -309,16 +424,20 @@ static const uint8_t stack_pattern[] = { }; #endif #if CONFIG_COROUTINE_PROTECT_STACK - #define STACK_GUARD_SIZE \ - ROUND_UP(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT) + #define CR_STACK_GUARD_SIZE \ + ((size_t)LM_ROUND_UP(sizeof(stack_pattern), CR_PLAT_STACK_ALIGNMENT)) #else - #define STACK_GUARD_SIZE 0 + #define CR_STACK_GUARD_SIZE ((size_t)0) #endif /* global variables ***********************************************************/ -static jmp_buf coroutine_add_env; -static jmp_buf coroutine_main_env; +static bool coroutine_initialized = false; +static cr_plat_jmp_buf coroutine_add_env; +static cr_plat_jmp_buf coroutine_main_env; +#if CONFIG_COROUTINE_GDB +static cr_plat_jmp_buf coroutine_gdb_env; +#endif /* * Invariants (and non-invariants): @@ -339,7 +458,7 @@ static jmp_buf coroutine_main_env; * coroutine_ringbuf queue. */ -static struct coroutine coroutine_table[CONFIG_COROUTINE_NUM] = {0}; +static struct coroutine coroutine_table[CONFIG_COROUTINE_NUM] = {}; static struct { /* tail == head means empty */ /* buf[tail] is the next thing to run */ @@ -349,48 +468,47 @@ static struct { * compiler will optimize `%array_len` to &(array_len-1)`, (b) * we don't have to worry about funny wrap-around behavior * when head or tail overflow. */ - cid_t buf[NEXT_POWER_OF_2(CONFIG_COROUTINE_NUM)]; -} coroutine_ringbuf = {0}; + cid_t buf[LM_NEXT_POWER_OF_2(CONFIG_COROUTINE_NUM)]; +} coroutine_ringbuf = {}; static cid_t coroutine_running = 0; +static size_t coroutine_cnt = 0; /* utility functions **********************************************************/ -#define errorf(...) fprintf(stderr, "error: " __VA_ARGS__) -#define infof(...) printf("info: " __VA_ARGS__) -#if CONFIG_COROUTINE_DEBUG - #define debugf(...) printf("dbg: " __VA_ARGS__) -#else - #define debugf(...) -#endif - -#ifdef __GLIBC__ - #define assertf(expr, ...) ({ \ - if (!(expr)) { \ - errorf("assertion: " __VA_ARGS__); \ - __assert_fail(#expr, __FILE__, __LINE__, __func__); \ - } \ - }) -#else - #define assertf(expr, ...) assert(expr) -#endif - -static inline const char* coroutine_state_str(enum coroutine_state state) { - assert(state < ARRAY_LEN(coroutine_state_strs)); - return coroutine_state_strs[state]; -} - static inline void coroutine_ringbuf_push(cid_t val) { - coroutine_ringbuf.buf[coroutine_ringbuf.head++ % ARRAY_LEN(coroutine_ringbuf.buf)] = val; - assert((coroutine_ringbuf.head % ARRAY_LEN(coroutine_ringbuf.buf)) != - (coroutine_ringbuf.tail % ARRAY_LEN(coroutine_ringbuf.buf))); + coroutine_ringbuf.buf[coroutine_ringbuf.head++ % LM_ARRAY_LEN(coroutine_ringbuf.buf)] = val; + assert((coroutine_ringbuf.head % LM_ARRAY_LEN(coroutine_ringbuf.buf)) != + (coroutine_ringbuf.tail % LM_ARRAY_LEN(coroutine_ringbuf.buf))); } static inline cid_t coroutine_ringbuf_pop(void) { if (coroutine_ringbuf.tail == coroutine_ringbuf.head) return 0; - return coroutine_ringbuf.buf[coroutine_ringbuf.tail++ % ARRAY_LEN(coroutine_ringbuf.buf)]; + return coroutine_ringbuf.buf[coroutine_ringbuf.tail++ % LM_ARRAY_LEN(coroutine_ringbuf.buf)]; } +#if CONFIG_COROUTINE_GDB +LM_NEVER_INLINE void cr_gdb_breakpoint(void) { + /* Prevent the call from being optimized away. */ + asm (""); +} +LM_NEVER_INLINE void cr_gdb_readjmp(cr_plat_jmp_buf *env) { + if (!cr_plat_setjmp(&coroutine_gdb_env)) + cr_plat_longjmp(env, 2); +} +#define cr_setjmp(env) ({ \ + int val = cr_plat_setjmp(env); \ + if (val == 2) { \ + cr_gdb_breakpoint(); \ + cr_plat_longjmp(&coroutine_gdb_env, 1); \ + } \ + val; \ + }) +#else +#define cr_setjmp(env) cr_plat_setjmp(env) +#endif +#define cr_longjmp(env) cr_plat_longjmp(env, 1) + static inline void assert_cid(cid_t cid) { assert(cid > 0); assert(cid <= CONFIG_COROUTINE_NUM); @@ -398,7 +516,7 @@ static inline void assert_cid(cid_t cid) { assert(coroutine_table[cid-1].stack_size); uint8_t *stack = coroutine_table[cid-1].stack; assert(stack); - for (size_t i = 0; i < STACK_GUARD_SIZE; i++) { + for (size_t i = 0; i < CR_STACK_GUARD_SIZE; i++) { size_t j = coroutine_table[cid-1].stack_size - (i+1); assert(stack[i] == stack_pattern[i%sizeof(stack_pattern)]); assert(stack[j] == stack_pattern[j%sizeof(stack_pattern)]); @@ -415,96 +533,164 @@ static inline void assert_cid(cid_t cid) { /* coroutine_add() ************************************************************/ -cid_t coroutine_add_with_stack_size(size_t stack_size, cr_fn_t fn, void *args) { +LM_NEVER_INLINE +cid_t coroutine_allocate_cid(void) { static cid_t last_created = 0; + + size_t base = last_created; + for (size_t shift = 0; shift < CONFIG_COROUTINE_NUM; shift++) { + cid_t child = ((base + shift) % CONFIG_COROUTINE_NUM) + 1; + if (coroutine_table[child-1].state == CR_NONE) { + last_created = child; + return child; + } + } + return 0; +} + +static cid_t _coroutine_add(void *stack, size_t full_stack_size, + const char *name, + cr_fn_t fn, void *args) { cid_t parent = coroutine_running; if (parent) assert_cid_state(parent, state == CR_RUNNING); - assert(stack_size); + assert(full_stack_size > 2*CR_STACK_GUARD_SIZE); assert(fn); - debugf("coroutine_add_with_stack_size(%zu, %p, %p)...\n", - stack_size, fn, args); - - cid_t child; - { - size_t base = last_created; - for (size_t shift = 0; shift < CONFIG_COROUTINE_NUM; shift++) { - child = ((base + shift) % CONFIG_COROUTINE_NUM) + 1; - if (coroutine_table[child-1].state == CR_NONE) - goto found; - } - return 0; - found: + + if (!coroutine_initialized) { + cr_plat_init(); + coroutine_initialized = true; } - debugf("...child=%zu\n", child); - last_created = child; + cid_t child = coroutine_allocate_cid(); + if (!child) + return 0; + log_debugln("...child=", child); + + /* 1. state *************************************************/ + coroutine_table[child-1].state = CR_INITIALIZING; - coroutine_table[child-1].stack_size = stack_size; - coroutine_table[child-1].stack = - aligned_alloc(CR_PLAT_STACK_ALIGNMENT, stack_size); + /* 2. name **************************************************/ + if (name) + strncpy(coroutine_table[child-1].name, name, sizeof(coroutine_table[child-1].name)); + else + memset(coroutine_table[child-1].name, 0, sizeof(coroutine_table[child-1].name)); + + log_infoln("starting cid ", child, + " ", (qstrn, coroutine_table[child-1].name, sizeof(coroutine_table[child-1].name)), + ":"); + + /* 3. stack *************************************************/ + coroutine_table[child-1].stack_size = full_stack_size; + coroutine_table[child-1].stack_free = (stack == NULL); + if (!stack) { + log_infoln("allocing stack with size ", full_stack_size, "..."); + stack = aligned_alloc(CR_PLAT_STACK_ALIGNMENT, full_stack_size); + log_infoln("...done"); + } + coroutine_table[child-1].stack = stack; + log_infoln(" -> full stack is [", + (ptr, coroutine_table[child-1].stack), ",", + (ptr, coroutine_table[child-1].stack + full_stack_size), ")", + " ; size=", full_stack_size); + log_infoln(" -> usable stack is [", + (ptr, coroutine_table[child-1].stack + CR_STACK_GUARD_SIZE), ",", + (ptr, coroutine_table[child-1].stack + full_stack_size - CR_STACK_GUARD_SIZE), ")", + " ; size=", full_stack_size - 2*CR_STACK_GUARD_SIZE); #if CONFIG_COROUTINE_MEASURE_STACK || CONFIG_COROUTINE_PROTECT_STACK - for (size_t i = 0; i < stack_size; i++) + for (size_t i = 0; i < full_stack_size; i++) ((uint8_t *)coroutine_table[child-1].stack)[i] = stack_pattern[i%sizeof(stack_pattern)]; #endif #if CONFIG_COROUTINE_VALGRIND coroutine_table[child-1].stack_id = VALGRIND_STACK_REGISTER( - coroutine_table[child-1].stack + STACK_GUARD_SIZE, - coroutine_table[child-1].stack + stack_size - STACK_GUARD_SIZE); + coroutine_table[child-1].stack + CR_STACK_GUARD_SIZE, + coroutine_table[child-1].stack + full_stack_size - CR_STACK_GUARD_SIZE); #endif + /* 4. env ***************************************************/ coroutine_running = child; - coroutine_table[child-1].state = CR_INITIALIZING; - if (!cr_plat_setjmp(coroutine_add_env)) { /* point=a */ + coroutine_cnt++; + if (!cr_setjmp(&coroutine_add_env)) { /* point=a */ void *stack_base = coroutine_table[child-1].stack #if CR_PLAT_STACK_GROWS_DOWNWARD - + stack_size - - STACK_GUARD_SIZE + + full_stack_size - CR_STACK_GUARD_SIZE #else - + STACK_GUARD_SIZE + + CR_STACK_GUARD_SIZE #endif ; - debugf("...stack =%p\n", coroutine_table[child-1].stack); - debugf("...stack_base=%p\n", stack_base); + log_debugln("...stack_base=", (ptr, stack_base)); /* run until cr_begin() */ cr_plat_call_with_stack(stack_base, fn, args); - __builtin_unreachable(); /* should cr_begin() instead of returning */ + assert_notreached("should cr_begin() instead of returning"); } assert_cid_state(child, state == CR_RUNNABLE); if (parent) assert_cid_state(parent, state == CR_RUNNING); + /* Restore interrupts because cr_begin() disables interrupts + * before the context switch. XXX: This assumes that + * interrupts were enabled when _add() was called, which we + * didn't actually check. */ + cr_restore_interrupts(true); coroutine_running = parent; + log_infoln(" -> done"); return child; } -cid_t coroutine_add(cr_fn_t fn, void *args) { - return coroutine_add_with_stack_size( - CONFIG_COROUTINE_DEFAULT_STACK_SIZE, fn, args); +cid_t coroutine_add_with_stack(void *stack, size_t full_stack_size, + const char *name, + cr_fn_t fn, void *args) { + if (name) + log_debugln("coroutine_add_with_stack(", (ptr, stack), full_stack_size, ", ", (qstr, name), ", ", (ptr, fn), ", ", (ptr, args), ")..."); + else + log_debugln("coroutine_add_with_stack(", (ptr, stack), full_stack_size, ", ", (ptr, name), ", ", (ptr, fn), ", ", (ptr, args), ")..."); + + return _coroutine_add(stack, full_stack_size, name, fn, args); +} + +cid_t coroutine_add_with_stack_size(size_t usable_stack_size, + const char *name, + cr_fn_t fn, void *args) { + if (name) + log_debugln("coroutine_add_with_stack_size(", usable_stack_size, ", ", (qstr, name), ", ", (ptr, fn), ", ", (ptr, args), ")..."); + else + log_debugln("coroutine_add_with_stack_size(", usable_stack_size, ", ", (ptr, name), ", ", (ptr, fn), ", ", (ptr, args), ")..."); + + return _coroutine_add(NULL, usable_stack_size + 2*CR_STACK_GUARD_SIZE, name, fn, args); } /* coroutine_main() ***********************************************************/ void coroutine_main(void) { - debugf("coroutine_main()\n"); - cr_disable_interrupts(); + log_debugln("coroutine_main()"); + if (!coroutine_initialized) { + cr_plat_init(); + coroutine_initialized = true; + } + bool saved = cr_save_and_disable_interrupts(); + assert(saved); + assert(!cr_plat_is_in_intrhandler()); coroutine_running = 0; - for (;;) { +#if CONFIG_COROUTINE_GDB + /* Some pointless call to prevent cr_gdb_readjmp() from + * getting pruned out by `ld --gc-sections`. */ + if (coroutine_table[0].state != CR_NONE) + cr_gdb_readjmp(&coroutine_table[0].env); +#endif + while (coroutine_cnt) { cid_t next; while ( !((next = coroutine_ringbuf_pop())) ) { /* No coroutines are runnable, wait for an interrupt * to change that. */ - cr_enable_interrupts(); cr_plat_wait_for_interrupt(); - cr_disable_interrupts(); } - if (!cr_plat_setjmp(coroutine_main_env)) { /* point=b */ + if (!cr_setjmp(&coroutine_main_env)) { /* point=b */ coroutine_running = next; coroutine_table[coroutine_running-1].state = CR_RUNNING; - cr_plat_longjmp(coroutine_table[coroutine_running-1].env, 1); /* jump to point=c */ + cr_longjmp(&coroutine_table[coroutine_running-1].env); /* jump to point=c */ } /* This is where we jump to from cr_exit(), and from * nowhere else. */ @@ -512,33 +698,35 @@ void coroutine_main(void) { #if CONFIG_COROUTINE_VALGRIND VALGRIND_STACK_DEREGISTER(coroutine_table[coroutine_running-1].stack_id); #endif - free(coroutine_table[coroutine_running-1].stack); - coroutine_table[coroutine_running-1] = (struct coroutine){0}; + if (coroutine_table[coroutine_running-1].stack_free) + free(coroutine_table[coroutine_running-1].stack); + coroutine_table[coroutine_running-1] = (struct coroutine){}; + coroutine_cnt--; } + coroutine_running = 0; + cr_restore_interrupts(saved); } /* cr_*() *********************************************************************/ void cr_begin(void) { - debugf("cid=%zu: cr_begin()\n", coroutine_running); + log_debugln("cid=", coroutine_running, ": cr_begin()"); assert_cid_state(coroutine_running, state == CR_INITIALIZING); + bool saved = cr_save_and_disable_interrupts(); coroutine_table[coroutine_running-1].state = CR_RUNNABLE; coroutine_ringbuf_push(coroutine_running); - coroutine_table[coroutine_running-1].sp = cr_plat_get_sp(); - if (!cr_plat_setjmp(coroutine_table[coroutine_running-1].env)) /* point=c1 */ - cr_plat_longjmp(coroutine_add_env, 1); /* jump to point=a */ - cr_enable_interrupts(); + if (!cr_setjmp(&coroutine_table[coroutine_running-1].env)) /* point=c1 */ + cr_longjmp(&coroutine_add_env); /* jump to point=a */ + cr_restore_interrupts(saved); } -static inline void _cr_yield() { +static inline void _cr_yield(void) { cid_t next; while ( !((next = coroutine_ringbuf_pop())) ) { /* No coroutines are runnable, wait for an interrupt * to change that. */ - cr_enable_interrupts(); cr_plat_wait_for_interrupt(); - cr_disable_interrupts(); } if (next == coroutine_running) { @@ -546,99 +734,135 @@ static inline void _cr_yield() { return; } - coroutine_table[coroutine_running-1].sp = cr_plat_get_sp(); - if (!cr_plat_setjmp(coroutine_table[coroutine_running-1].env)) { /* point=c2 */ + if (!cr_setjmp(&coroutine_table[coroutine_running-1].env)) { /* point=c2 */ coroutine_running = next; coroutine_table[coroutine_running-1].state = CR_RUNNING; - cr_plat_longjmp(coroutine_table[coroutine_running-1].env, 1); /* jump to point=c */ + cr_longjmp(&coroutine_table[coroutine_running-1].env); /* jump to point=c */ } } void cr_yield(void) { - debugf("cid=%zu: cr_yield()\n", coroutine_running); + log_debugln("cid=", coroutine_running ,": cr_yield()"); + assert(!cr_plat_is_in_intrhandler()); assert_cid_state(coroutine_running, state == CR_RUNNING); - cr_disable_interrupts(); + bool saved = cr_save_and_disable_interrupts(); coroutine_table[coroutine_running-1].state = CR_RUNNABLE; coroutine_ringbuf_push(coroutine_running); _cr_yield(); - cr_enable_interrupts(); + cr_restore_interrupts(saved); } void cr_pause_and_yield(void) { - debugf("cid=%zu: cr_pause_and_yield()\n", coroutine_running); - assert_cid_state(coroutine_running, state == CR_RUNNING || state == CR_PRE_RUNNABLE); - - cr_disable_interrupts(); - if (coroutine_table[coroutine_running-1].state == CR_PRE_RUNNABLE) { - coroutine_table[coroutine_running-1].state = CR_RUNNABLE; - coroutine_ringbuf_push(coroutine_running); - } else - coroutine_table[coroutine_running-1].state = CR_PAUSED; + log_debugln("cid=", coroutine_running, ": cr_pause_and_yield()"); + assert(!cr_plat_is_in_intrhandler()); + assert_cid_state(coroutine_running, state == CR_RUNNING); + + bool saved = cr_save_and_disable_interrupts(); + coroutine_table[coroutine_running-1].state = CR_PAUSED; _cr_yield(); - cr_enable_interrupts(); + cr_restore_interrupts(saved); } -void cr_exit(void) { - debugf("cid=%zu: cr_exit()\n", coroutine_running); +[[noreturn]] void cr_exit(void) { + log_debugln("cid=", coroutine_running, ": cr_exit()"); + assert(!cr_plat_is_in_intrhandler()); assert_cid_state(coroutine_running, state == CR_RUNNING); - cr_disable_interrupts(); + (void)cr_save_and_disable_interrupts(); coroutine_table[coroutine_running-1].state = CR_NONE; - cr_plat_longjmp(coroutine_main_env, 1); /* jump to point=b */ + cr_longjmp(&coroutine_main_env); /* jump to point=b */ } -void cr_unpause(cid_t cid) { - debugf("cr_unpause(%zu)\n", cid); +static void _cr_unpause(cid_t cid) { assert_cid_state(cid, state == CR_PAUSED); coroutine_table[cid-1].state = CR_RUNNABLE; coroutine_ringbuf_push(cid); } +void cr_unpause(cid_t cid) { + log_debugln("cr_unpause(", cid, ")"); + assert(!cr_plat_is_in_intrhandler()); + assert_cid_state(coroutine_running, state == CR_RUNNING); + + bool saved = cr_save_and_disable_interrupts(); + _cr_unpause(cid); + cr_restore_interrupts(saved); +} + void cr_unpause_from_intrhandler(cid_t cid) { - debugf("cr_unpause_from_intrhandler(%zu)\n", cid); - assert_cid_state(cid, state == CR_RUNNING || state == CR_PAUSED); - - if (coroutine_table[cid-1].state == CR_RUNNING) { - assert(cid == coroutine_running); - debugf("... raced, deferring unpause\n"); - coroutine_table[cid-1].state = CR_PRE_RUNNABLE; - } else { - debugf("... actual unpause\n"); - coroutine_table[cid-1].state = CR_RUNNABLE; - coroutine_ringbuf_push(cid); - } + log_debugln("cr_unpause_from_intrhandler(", cid, ")"); + assert(cr_plat_is_in_intrhandler()); + + _cr_unpause(cid); } cid_t cr_getcid(void) { + assert(!cr_plat_is_in_intrhandler()); + assert_cid_state(coroutine_running, state == CR_RUNNING); return coroutine_running; } -/* cr_cid_info() **************************************************************/ +#ifndef NDEBUG +void cr_assert_in_coroutine(void) { + assert(!cr_plat_is_in_intrhandler()); + assert_cid_state(coroutine_running, state == CR_RUNNING); +} -#if CONFIG_COROUTINE_MEASURE_STACK +void cr_assert_in_intrhandler(void) { + assert(cr_plat_is_in_intrhandler()); +} +#endif + +/* answering questions about coroutines ***************************************/ void cr_cid_info(cid_t cid, struct cr_cid_info *ret) { - assert_cid(cid); + assert(cid > 0); + assert(cid <= CONFIG_COROUTINE_NUM); assert(ret); + memset(ret, 0, sizeof(*ret)); + if (coroutine_table[cid-1].state == CR_NONE) + return; + assert_cid(cid); + + /* 1. state *************************************************/ + ret->state = coroutine_table[cid-1].state; + + /* 2. name **************************************************/ + memcpy(ret->name, coroutine_table[cid-1].name, CONFIG_COROUTINE_NAME_LEN); + + /* 3. stack *************************************************/ +#if CONFIG_COROUTINE_MEASURE_STACK + uint8_t *stack = (uint8_t *)coroutine_table[cid-1].stack; + uint8_t *stack_lo = stack + CR_STACK_GUARD_SIZE; + uint8_t *stack_hi = stack + coroutine_table[cid-1].stack_size - CR_STACK_GUARD_SIZE; + /* stack_cap */ - ret->stack_cap = coroutine_table[cid-1].stack_size - 2*STACK_GUARD_SIZE; + ret->stack_cap = stack_hi - stack_lo; /* stack_max */ ret->stack_max = ret->stack_cap; - uint8_t *stack = (uint8_t *)coroutine_table[cid-1].stack; for (;;) { size_t i = #if CR_PLAT_STACK_GROWS_DOWNWARD - STACK_GUARD_SIZE + ret->stack_cap - ret->stack_max + CR_STACK_GUARD_SIZE + ret->stack_cap - ret->stack_max #else - ret->stack_max - 1 - STACK_GUARD_SIZE + ret->stack_max - 1 - CR_STACK_GUARD_SIZE #endif ; - if (ret->stack_max == 0 || - stack[i] != stack_pattern[i%sizeof(stack_pattern)]) + if (ret->stack_max == 0) + break; + assert(stack_lo <= &stack[i] && &stack[i] < stack_hi); +#if CONFIG_COROUTINE_VALGRIND + VALGRIND_DISABLE_ERROR_REPORTING; +#endif + uint8_t v = stack[i]; +#if CONFIG_COROUTINE_VALGRIND + VALGRIND_ENABLE_ERROR_REPORTING; +#endif + if (v != stack_pattern[i%sizeof(stack_pattern)]) break; ret->stack_max--; } @@ -647,15 +871,22 @@ void cr_cid_info(cid_t cid, struct cr_cid_info *ret) { uintptr_t sp; if (cid == coroutine_running) sp = cr_plat_get_sp(); + else if (coroutine_table[cid-1].state == CR_RUNNING) + sp = cr_plat_setjmp_get_sp(&coroutine_add_env); else - sp = coroutine_table[cid-1].sp; + sp = cr_plat_setjmp_get_sp(&coroutine_table[cid-1].env); assert(sp); - uintptr_t sb = (uintptr_t)coroutine_table[cid-1].stack; #if CR_PLAT_STACK_GROWS_DOWNWARD - ret->stack_cur = (sb - STACK_GUARD_SIZE) - sp; + uintptr_t sb = (uintptr_t)stack_hi; + ret->stack_cur = sb - sp; #else - ret->stack_cur = sp - (sb + STACK_GUARD_SIZE); + uintptr_t sb = (uintptr_t)stack_lo; + ret->stack_cur = sp - sb; #endif +#endif /* CONFIG_COROUTINE_MEASURE_STACK */ } -#endif /* CONFIG_COROUTINE_MEASURE_STACK */ +const char *coroutine_state_str(enum coroutine_state state) { + assert(state < LM_ARRAY_LEN(coroutine_state_strs)); + return coroutine_state_strs[state]; +} |