Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 1 | /* |
| 2 | * sigaltstack coroutine initialization code |
| 3 | * |
| 4 | * Copyright (C) 2006 Anthony Liguori <anthony@codemonkey.ws> |
| 5 | * Copyright (C) 2011 Kevin Wolf <kwolf@redhat.com> |
| 6 | * Copyright (C) 2012 Alex Barcelo <abarcelo@ac.upc.edu> |
| 7 | ** This file is partly based on pth_mctx.c, from the GNU Portable Threads |
| 8 | ** Copyright (c) 1999-2006 Ralf S. Engelschall <rse@engelschall.com> |
| 9 | * |
| 10 | * This library is free software; you can redistribute it and/or |
| 11 | * modify it under the terms of the GNU Lesser General Public |
| 12 | * License as published by the Free Software Foundation; either |
| 13 | * version 2.1 of the License, or (at your option) any later version. |
| 14 | * |
| 15 | * This library is distributed in the hope that it will be useful, |
| 16 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 17 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
| 18 | * Lesser General Public License for more details. |
| 19 | * |
| 20 | * You should have received a copy of the GNU Lesser General Public |
| 21 | * License along with this library; if not, see <http://www.gnu.org/licenses/>. |
| 22 | */ |
| 23 | |
| 24 | /* XXX Is there a nicer way to disable glibc's stack check for longjmp? */ |
| 25 | #ifdef _FORTIFY_SOURCE |
| 26 | #undef _FORTIFY_SOURCE |
| 27 | #endif |
| 28 | #include <stdlib.h> |
| 29 | #include <setjmp.h> |
| 30 | #include <stdint.h> |
| 31 | #include <pthread.h> |
| 32 | #include <signal.h> |
| 33 | #include "qemu-common.h" |
Paolo Bonzini | 737e150 | 2012-12-17 18:19:44 +0100 | [diff] [blame] | 34 | #include "block/coroutine_int.h" |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 35 | |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 36 | typedef struct { |
| 37 | Coroutine base; |
| 38 | void *stack; |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 39 | sigjmp_buf env; |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 40 | } CoroutineUContext; |
| 41 | |
| 42 | /** |
| 43 | * Per-thread coroutine bookkeeping |
| 44 | */ |
| 45 | typedef struct { |
| 46 | /** Currently executing coroutine */ |
| 47 | Coroutine *current; |
| 48 | |
| 49 | /** The default coroutine */ |
| 50 | CoroutineUContext leader; |
| 51 | |
| 52 | /** Information for the signal handler (trampoline) */ |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 53 | sigjmp_buf tr_reenter; |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 54 | volatile sig_atomic_t tr_called; |
| 55 | void *tr_handler; |
| 56 | } CoroutineThreadState; |
| 57 | |
| 58 | static pthread_key_t thread_state_key; |
| 59 | |
| 60 | static CoroutineThreadState *coroutine_get_thread_state(void) |
| 61 | { |
| 62 | CoroutineThreadState *s = pthread_getspecific(thread_state_key); |
| 63 | |
| 64 | if (!s) { |
| 65 | s = g_malloc0(sizeof(*s)); |
| 66 | s->current = &s->leader.base; |
| 67 | pthread_setspecific(thread_state_key, s); |
| 68 | } |
| 69 | return s; |
| 70 | } |
| 71 | |
| 72 | static void qemu_coroutine_thread_cleanup(void *opaque) |
| 73 | { |
| 74 | CoroutineThreadState *s = opaque; |
| 75 | |
| 76 | g_free(s); |
| 77 | } |
| 78 | |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 79 | static void __attribute__((constructor)) coroutine_init(void) |
| 80 | { |
| 81 | int ret; |
| 82 | |
| 83 | ret = pthread_key_create(&thread_state_key, qemu_coroutine_thread_cleanup); |
| 84 | if (ret != 0) { |
| 85 | fprintf(stderr, "unable to create leader key: %s\n", strerror(errno)); |
| 86 | abort(); |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | /* "boot" function |
| 91 | * This is what starts the coroutine, is called from the trampoline |
| 92 | * (from the signal handler when it is not signal handling, read ahead |
| 93 | * for more information). |
| 94 | */ |
| 95 | static void coroutine_bootstrap(CoroutineUContext *self, Coroutine *co) |
| 96 | { |
| 97 | /* Initialize longjmp environment and switch back the caller */ |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 98 | if (!sigsetjmp(self->env, 0)) { |
| 99 | siglongjmp(*(sigjmp_buf *)co->entry_arg, 1); |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 100 | } |
| 101 | |
| 102 | while (true) { |
| 103 | co->entry(co->entry_arg); |
| 104 | qemu_coroutine_switch(co, co->caller, COROUTINE_TERMINATE); |
| 105 | } |
| 106 | } |
| 107 | |
| 108 | /* |
| 109 | * This is used as the signal handler. This is called with the brand new stack |
| 110 | * (thanks to sigaltstack). We have to return, given that this is a signal |
| 111 | * handler and the sigmask and some other things are changed. |
| 112 | */ |
| 113 | static void coroutine_trampoline(int signal) |
| 114 | { |
| 115 | CoroutineUContext *self; |
| 116 | Coroutine *co; |
| 117 | CoroutineThreadState *coTS; |
| 118 | |
| 119 | /* Get the thread specific information */ |
| 120 | coTS = coroutine_get_thread_state(); |
| 121 | self = coTS->tr_handler; |
| 122 | coTS->tr_called = 1; |
| 123 | co = &self->base; |
| 124 | |
| 125 | /* |
| 126 | * Here we have to do a bit of a ping pong between the caller, given that |
| 127 | * this is a signal handler and we have to do a return "soon". Then the |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 128 | * caller can reestablish everything and do a siglongjmp here again. |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 129 | */ |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 130 | if (!sigsetjmp(coTS->tr_reenter, 0)) { |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 131 | return; |
| 132 | } |
| 133 | |
| 134 | /* |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 135 | * Ok, the caller has siglongjmp'ed back to us, so now prepare |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 136 | * us for the real machine state switching. We have to jump |
| 137 | * into another function here to get a new stack context for |
| 138 | * the auto variables (which have to be auto-variables |
| 139 | * because the start of the thread happens later). Else with |
| 140 | * PIC (i.e. Position Independent Code which is used when PTH |
| 141 | * is built as a shared library) most platforms would |
| 142 | * horrible core dump as experience showed. |
| 143 | */ |
| 144 | coroutine_bootstrap(self, co); |
| 145 | } |
| 146 | |
Paolo Bonzini | 4023978 | 2013-02-19 11:59:09 +0100 | [diff] [blame] | 147 | Coroutine *qemu_coroutine_new(void) |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 148 | { |
| 149 | const size_t stack_size = 1 << 20; |
| 150 | CoroutineUContext *co; |
| 151 | CoroutineThreadState *coTS; |
| 152 | struct sigaction sa; |
| 153 | struct sigaction osa; |
Peter Maydell | 2ad2210 | 2012-11-10 21:47:52 +0000 | [diff] [blame] | 154 | stack_t ss; |
| 155 | stack_t oss; |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 156 | sigset_t sigs; |
| 157 | sigset_t osigs; |
| 158 | jmp_buf old_env; |
| 159 | |
| 160 | /* The way to manipulate stack is with the sigaltstack function. We |
| 161 | * prepare a stack, with it delivering a signal to ourselves and then |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 162 | * put sigsetjmp/siglongjmp where needed. |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 163 | * This has been done keeping coroutine-ucontext as a model and with the |
| 164 | * pth ideas (GNU Portable Threads). See coroutine-ucontext for the basics |
| 165 | * of the coroutines and see pth_mctx.c (from the pth project) for the |
| 166 | * sigaltstack way of manipulating stacks. |
| 167 | */ |
| 168 | |
| 169 | co = g_malloc0(sizeof(*co)); |
| 170 | co->stack = g_malloc(stack_size); |
| 171 | co->base.entry_arg = &old_env; /* stash away our jmp_buf */ |
| 172 | |
| 173 | coTS = coroutine_get_thread_state(); |
| 174 | coTS->tr_handler = co; |
| 175 | |
| 176 | /* |
| 177 | * Preserve the SIGUSR2 signal state, block SIGUSR2, |
| 178 | * and establish our signal handler. The signal will |
| 179 | * later transfer control onto the signal stack. |
| 180 | */ |
| 181 | sigemptyset(&sigs); |
| 182 | sigaddset(&sigs, SIGUSR2); |
| 183 | pthread_sigmask(SIG_BLOCK, &sigs, &osigs); |
| 184 | sa.sa_handler = coroutine_trampoline; |
| 185 | sigfillset(&sa.sa_mask); |
| 186 | sa.sa_flags = SA_ONSTACK; |
| 187 | if (sigaction(SIGUSR2, &sa, &osa) != 0) { |
| 188 | abort(); |
| 189 | } |
| 190 | |
| 191 | /* |
| 192 | * Set the new stack. |
| 193 | */ |
| 194 | ss.ss_sp = co->stack; |
| 195 | ss.ss_size = stack_size; |
| 196 | ss.ss_flags = 0; |
| 197 | if (sigaltstack(&ss, &oss) < 0) { |
| 198 | abort(); |
| 199 | } |
| 200 | |
| 201 | /* |
| 202 | * Now transfer control onto the signal stack and set it up. |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 203 | * It will return immediately via "return" after the sigsetjmp() |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 204 | * was performed. Be careful here with race conditions. The |
| 205 | * signal can be delivered the first time sigsuspend() is |
| 206 | * called. |
| 207 | */ |
| 208 | coTS->tr_called = 0; |
Jan Kiszka | 99b5beb | 2012-05-09 14:23:27 -0300 | [diff] [blame] | 209 | pthread_kill(pthread_self(), SIGUSR2); |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 210 | sigfillset(&sigs); |
| 211 | sigdelset(&sigs, SIGUSR2); |
| 212 | while (!coTS->tr_called) { |
| 213 | sigsuspend(&sigs); |
| 214 | } |
| 215 | |
| 216 | /* |
| 217 | * Inform the system that we are back off the signal stack by |
| 218 | * removing the alternative signal stack. Be careful here: It |
| 219 | * first has to be disabled, before it can be removed. |
| 220 | */ |
| 221 | sigaltstack(NULL, &ss); |
| 222 | ss.ss_flags = SS_DISABLE; |
| 223 | if (sigaltstack(&ss, NULL) < 0) { |
| 224 | abort(); |
| 225 | } |
| 226 | sigaltstack(NULL, &ss); |
| 227 | if (!(oss.ss_flags & SS_DISABLE)) { |
| 228 | sigaltstack(&oss, NULL); |
| 229 | } |
| 230 | |
| 231 | /* |
| 232 | * Restore the old SIGUSR2 signal handler and mask |
| 233 | */ |
| 234 | sigaction(SIGUSR2, &osa, NULL); |
| 235 | pthread_sigmask(SIG_SETMASK, &osigs, NULL); |
| 236 | |
| 237 | /* |
| 238 | * Now enter the trampoline again, but this time not as a signal |
| 239 | * handler. Instead we jump into it directly. The functionally |
Jim Meyering | a31f053 | 2012-05-09 05:12:04 +0000 | [diff] [blame] | 240 | * redundant ping-pong pointer arithmetic is necessary to avoid |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 241 | * type-conversion warnings related to the `volatile' qualifier and |
| 242 | * the fact that `jmp_buf' usually is an array type. |
| 243 | */ |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 244 | if (!sigsetjmp(old_env, 0)) { |
| 245 | siglongjmp(coTS->tr_reenter, 1); |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 246 | } |
| 247 | |
| 248 | /* |
| 249 | * Ok, we returned again, so now we're finished |
| 250 | */ |
| 251 | |
| 252 | return &co->base; |
| 253 | } |
| 254 | |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 255 | void qemu_coroutine_delete(Coroutine *co_) |
| 256 | { |
| 257 | CoroutineUContext *co = DO_UPCAST(CoroutineUContext, base, co_); |
| 258 | |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 259 | g_free(co->stack); |
| 260 | g_free(co); |
| 261 | } |
| 262 | |
| 263 | CoroutineAction qemu_coroutine_switch(Coroutine *from_, Coroutine *to_, |
| 264 | CoroutineAction action) |
| 265 | { |
| 266 | CoroutineUContext *from = DO_UPCAST(CoroutineUContext, base, from_); |
| 267 | CoroutineUContext *to = DO_UPCAST(CoroutineUContext, base, to_); |
| 268 | CoroutineThreadState *s = coroutine_get_thread_state(); |
| 269 | int ret; |
| 270 | |
| 271 | s->current = to_; |
| 272 | |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 273 | ret = sigsetjmp(from->env, 0); |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 274 | if (ret == 0) { |
Peter Maydell | 6ab7e54 | 2013-02-20 15:21:09 +0000 | [diff] [blame] | 275 | siglongjmp(to->env, action); |
Alex Barcelo | 3194c8c | 2012-02-28 12:25:49 +0100 | [diff] [blame] | 276 | } |
| 277 | return ret; |
| 278 | } |
| 279 | |
| 280 | Coroutine *qemu_coroutine_self(void) |
| 281 | { |
| 282 | CoroutineThreadState *s = coroutine_get_thread_state(); |
| 283 | |
| 284 | return s->current; |
| 285 | } |
| 286 | |
| 287 | bool qemu_in_coroutine(void) |
| 288 | { |
| 289 | CoroutineThreadState *s = pthread_getspecific(thread_state_key); |
| 290 | |
| 291 | return s && s->current->caller; |
| 292 | } |
| 293 | |