diff options
Diffstat (limited to 'coroutine.c')
| -rw-r--r-- | coroutine.c | 131 |
1 files changed, 112 insertions, 19 deletions
diff --git a/coroutine.c b/coroutine.c index 34d84c8..9a53731 100644 --- a/coroutine.c +++ b/coroutine.c @@ -1,4 +1,4 @@ -/* coroutine.c - Simple coroutine and request/response implementation +/* coroutine.c - Simple embeddable coroutine implementation * * Copyright (C) 2024 Luke T. Shumaker <lukeshu@lukeshu.com> * SPDX-Licence-Identifier: AGPL-3.0-or-later @@ -10,15 +10,106 @@ #include <assert.h> #include <setjmp.h> -#include "cfg_limits.h" #include "coroutine.h" +/* Configuration **************************************************************/ + +#define COROUTINE_NUM 5 +#define COROUTINE_MEASURE_STACK 1 + +/* Implementation *************************************************************/ + +/* + * Portability notes: + * + * - It uses GCC `__attribute__`s, though these can likely be easily + * swapped for other compilers. + * + * - It has a small bit of CPU-specific code (assembly, and definition + * of a STACK_GROWS_DOWNWARD={0,1} macro) in + * coroutine.c:call_with_stack(). Other than this, it should be + * portable to other CPUs. It currently contains implementations + * for __x86_64__ and __arm__, and should be fairly easy to add + * implementations for other CPUs. + * + * - It uses setjmp()/longjmp() in "unsafe" ways. POSIX-2017 + * longjmp(3) says + * + * > If the most recent invocation of setjmp() with the + * > corresponding jmp_buf ... or if the function containing the + * > invocation of setjmp() has terminated execution in the interim, + * > or if the invocation of setjmp() was within the scope of an + * > identifier with variably modified type and execution has left + * > that scope in the interim, the behavior is undefined. + * + * We use longjmp() both of these scenarios, but make it OK by using + * call_with_stack() to manage the stack ourselves, assuming the + * sole reason that longjmp() behavior is undefined in such cases is + * because the stack that its saved stack-pointer points to is no + * longer around. It seems absurd that an implementation would + * choose to do something else, but I'm calling it out here because + * you never know. + * + * Note that setjmp()/longjmp() are defined in 3 places: in the libc + * (glibc/newlib), as GCC intrinsics, and the lower-level GCC + * __builtin_{setjmp,newlib} which the libc and intrinsic versions + * likely use. Our assumptions seem to be valid for + * x86_64-pc-linux-gnu/gcc-14.2.1/glibc-2.40 and + * arm-none-eabi/gcc-14.1.0/newlib-4.4.0. + * + * Why not use <ucontext.h>, the now-deprecated (was in POSIX.1-2001, + * is gone in POSIX-2008) predecesor to <setjmp.h>? It would let us + * do this without any assembly or unsafe assumptions. Simply: + * because newlib does not provide it. + */ + +/* + * Design decisions and notes: + * + * - Coroutines are launched with a zeroed-out stack. Because all + * stack variables should be initialized before they are read, this + * "shouldn't" make a difference, but: (1) Initializing it to a + * known value allows us to measure how much of the stack was + * written to, which is helpful to tune stack sizes. (2) Leaving it + * uninitialized just gives me the willies. + * + * - Because embedded programs should be adverse to using the heap, + * COROUTINE_NUM is fixed, instead of having coroutine_add() + * dynamically grow the coroutine_table as-needed. + * + * - On the flip-side, coroutine stacks are allocated on the heap + * instead of having them be statically-allocated along with + * coroutine_table. (1) This reduced the blast-area of damage for a + * stack-overflow; and indeed if the end of the stack alignes with a + * page-boundary memory-protection can even detect the overflow for + * us. (2) Having different-looking addresses for stack-area vs + * static-area is handy for making things jump out at you when + * debugging. (3) Given the above about wanting a zeroed-out stack, + * this allows us to take advantage of optimizations in calloc() + * instead of using memset, and this can likely also improve things + * with being page-aligned. + * + * - Coroutines must use cr_exit() instead of returning because if + * they return then they will return to call_with_stack() in + * coroutine_add() (not to after the longjmp() call in + * coroutine_main()), and besides being + * wrong-for-our-desired-flow-control, that's a stack location that + * no longer exists. + * + * Things to consider changing: + * + * - Consider having _cr_transition() go ahead and find the next + * coroutine to run and longjmp() direcly to it, instead of first + * jumping back to coroutine_main(). This could save a few cycles + * and a few bytes. + */ + enum coroutine_state { - CR_NONE = 0, - CR_INITIALIZING, - CR_RUNNING, - CR_RUNNABLE, - CR_PAUSED, + CR_NONE = 0, /* this slot in the table is empty */ + CR_INITIALIZING, /* running, before cr_begin() */ + CR_RUNNING, /* running, after cr_begin() */ + CR_RUNNABLE, /* not running, but runnable */ + CR_PAUSED, /* not running, and not runnable */ }; struct coroutine { @@ -33,13 +124,12 @@ static cid_t coroutine_running = 0; static jmp_buf coroutine_add_env; static jmp_buf coroutine_main_env; -void coroutine_init(void) {} - static void call_with_stack(void *stack, cr_fn_t fn, void *args) { static void *saved_sp = NULL; - /* This only really exists for running on ARM-32, but being - * able to run it on x86-64 is useful for debugging. */ + /* As part of sbc-harness, this only really needs to support + * ARM-32, but being able to run it on x86-64 is useful for + * debugging. */ #if __x86_64__ #define STACK_GROWS_DOWNWARD 1 asm volatile ("movq %%rsp , %0\n\t" /* saved_sp = sp */ @@ -56,11 +146,14 @@ static void call_with_stack(void *stack, cr_fn_t fn, void *args) { ); #elif __arm__ #define STACK_GROWS_DOWNWARD 1 + /* str/ldr can only work with a "lo" register, which sp is + * not, so we use r0 as an intermediate because we're going to + * clobber it with args anyway. */ asm volatile ("mov r0, sp\n\t" /* [saved_sp = sp */ - "str r0, %0\n\t" /* ] */ + "str r0, %0\n\t" /* ] */ "mov sp, %1\n\t" /* [sp = stack] */ - "mov r0, %1\n\t" /* [fn(arg0) */ - "blx %2\n\t" /* ] */ + "mov r0, %1\n\t" /* [arg0 = args] */ + "blx %2\n\t" /* [fn()] */ "ldr r0, %0\n\t" /* [sp = staved_sp */ "mov sp, r0" /* ] */ : @@ -75,7 +168,7 @@ static void call_with_stack(void *stack, cr_fn_t fn, void *args) { #endif } -cid_t coroutine_add(cr_fn_t fn, void *args) { +cid_t coroutine_add_with_stack_size(size_t stack_size, cr_fn_t fn, void *args) { static cid_t last_created = 0; assert(coroutine_running == 0 || coroutine_table[coroutine_running-1].state == CR_RUNNING); @@ -94,7 +187,7 @@ cid_t coroutine_add(cr_fn_t fn, void *args) { last_created = child; - coroutine_table[child-1].stack_size = COROUTINE_STACK_SIZE; + coroutine_table[child-1].stack_size = stack_size; coroutine_table[child-1].stack = calloc(1, coroutine_table[child-1].stack_size); cid_t parent = coroutine_running; @@ -127,7 +220,7 @@ void coroutine_main(void) { assert(false); /* should cr_exit() instead of returning */ } if (cr->state == CR_NONE) { -#if DEBUG +#if COROUTINE_MEASURE_STACK size_t stack_used = cr->stack_size; while (stack_used > 0 && ((uint8_t*)cr->stack)[STACK_GROWS_DOWNWARD ? cr->stack_size - stack_used : stack_used - 1] == 0) stack_used--; @@ -152,7 +245,7 @@ bool cr_begin(void) { } static inline __attribute__ ((no_split_stack)) void _cr_transition(enum coroutine_state state) { - assert(coroutine_table[coroutine_running-1].state == CR_RUNNING); + assert(coroutine_running && coroutine_table[coroutine_running-1].state == CR_RUNNING); coroutine_table[coroutine_running-1].state = state; if (!setjmp(coroutine_table[coroutine_running-1].env)) /* point=c2 */ longjmp(coroutine_main_env, 1); /* jump to point=b */ @@ -162,7 +255,7 @@ void cr_yield(void) { _cr_transition(CR_RUNNABLE); } void cr_pause_and_yield(void) { _cr_transition(CR_PAUSED); } void cr_exit(void) { - assert(coroutine_table[coroutine_running-1].state == CR_RUNNING); + assert(coroutine_running && coroutine_table[coroutine_running-1].state == CR_RUNNING); coroutine_table[coroutine_running-1].state = CR_NONE; longjmp(coroutine_main_env, 1); /* jump to point=b */ } |