1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
|
/* coroutine.c - Simple embeddable coroutine implementation
*
* Copyright (C) 2024 Luke T. Shumaker <lukeshu@lukeshu.com>
* SPDX-Licence-Identifier: AGPL-3.0-or-later
*/
#include <stdint.h> /* for uint8_t */
#include <stdio.h> /* for fprintf(), stderr */
#include <stdlib.h> /* for calloc(), free() */
#include <assert.h>
#include <setjmp.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, /* 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 {
enum coroutine_state state;
jmp_buf env;
size_t stack_size;
void *stack;
};
static struct coroutine coroutine_table[COROUTINE_NUM] = {0};
static cid_t coroutine_running = 0;
static jmp_buf coroutine_add_env;
static jmp_buf coroutine_main_env;
static void call_with_stack(void *stack, cr_fn_t fn, void *args) {
static void *saved_sp = NULL;
/* 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 */
"movq %1 , %%rsp\n\t" /* sp = stack */
"movq %3 , %%rdi\n\t" /* arg0 = args */
"call *%2\n\t" /* fn() */
"movq %0 , %%rsp" /* sp = saved_sp */
:
: /* %0 */"m"(saved_sp),
/* %1 */"r"(stack),
/* %2 */"r"(fn),
/* %3 */"r"(args)
: "rdi"
);
#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" /* ] */
"mov sp, %1\n\t" /* [sp = stack] */
"mov r0, %1\n\t" /* [arg0 = args] */
"blx %2\n\t" /* [fn()] */
"ldr r0, %0\n\t" /* [sp = staved_sp */
"mov sp, r0" /* ] */
:
: /* %0 */"m"(saved_sp),
/* %1 */"r"(stack),
/* %2 */"r"(fn),
/* %3 */"r"(args)
: "r0"
);
#else
#error unsupported architecture
#endif
}
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);
cid_t child;
{
size_t idx_base = last_created;
for (size_t idx_shift = 0; idx_shift < COROUTINE_NUM; idx_shift++) {
child = ((idx_base + idx_shift) % COROUTINE_NUM) + 1;
if (coroutine_table[child-1].state == CR_NONE)
goto found;
}
return 0;
found:
}
last_created = child;
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;
assert(parent == 0 || coroutine_table[parent-1].state == CR_RUNNING);
coroutine_running = child;
coroutine_table[child-1].state = CR_INITIALIZING;
if (!setjmp(coroutine_add_env)) { /* point=a */
/* run until cr_begin() */
call_with_stack(coroutine_table[child-1].stack + (STACK_GROWS_DOWNWARD ? coroutine_table[child-1].stack_size : 0), fn, args);
assert(false); /* should cr_begin() instead of returning */
}
assert(coroutine_table[child-1].state == CR_RUNNABLE);
assert(parent == 0 || coroutine_table[parent-1].state == CR_RUNNING);
coroutine_running = parent;
return child;
}
void coroutine_main(void) {
bool ran;
for (coroutine_running = 1;; coroutine_running = (coroutine_running%COROUTINE_NUM)+1) {
if (coroutine_running == 1)
ran = false;
struct coroutine *cr = &coroutine_table[coroutine_running-1];
if (cr->state == CR_RUNNABLE) {
ran = true;
cr->state = CR_RUNNING;
if (!setjmp(coroutine_main_env)) { /* point=b */
longjmp(cr->env, 1); /* jump to point=c */
assert(false); /* should cr_exit() instead of returning */
}
if (cr->state == CR_NONE) {
#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--;
printf("info: coroutine %zu exited having used %zu B stack space\n", coroutine_running, stack_used);
#endif
free(cr->stack);
coroutine_table[coroutine_running-1] = (struct coroutine){0};
}
}
if (coroutine_running == COROUTINE_NUM && !ran) {
fprintf(stderr, "error: no runnable coroutines\n");
return;
}
}
}
bool cr_begin(void) {
assert(coroutine_table[coroutine_running-1].state == CR_INITIALIZING);
coroutine_table[coroutine_running-1].state = CR_RUNNABLE;
if (!setjmp(coroutine_table[coroutine_running-1].env)) /* point=c1 */
longjmp(coroutine_add_env, 1); /* jump to point=a */
}
static inline __attribute__ ((no_split_stack)) void _cr_transition(enum coroutine_state state) {
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 */
}
void cr_yield(void) { _cr_transition(CR_RUNNABLE); }
void cr_pause_and_yield(void) { _cr_transition(CR_PAUSED); }
void cr_exit(void) {
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 */
}
void cr_unpause(cid_t cid) {
assert(coroutine_table[cid-1].state == CR_PAUSED);
coroutine_table[cid-1].state = CR_RUNNABLE;
}
cid_t cr_getcid(void) {
return coroutine_running;
}
|