include/qemu/coroutine.h

   1 /*
   2  * QEMU coroutine implementation
   3  *
   4  * Copyright IBM, Corp. 2011
   5  *
   6  * Authors:
   7  *  Stefan Hajnoczi    <stefanha@linux.vnet.ibm.com>
   8  *  Kevin Wolf         <kwolf@redhat.com>
   9  *
  10  * This work is licensed under the terms of the GNU LGPL, version 2 or later.
  11  * See the COPYING.LIB file in the top-level directory.
  12  *
  13  */
  14
  15 #ifndef QEMU_COROUTINE_H
  16 #define QEMU_COROUTINE_H
  17
  18 #include "qemu/coroutine-core.h"
  19 #include "qemu/queue.h"
  20 #include "qemu/timer.h"
  21
  22 /**
  23  * Coroutines are a mechanism for stack switching and can be used for
  24  * cooperative userspace threading.  These functions provide a simple but
  25  * useful flavor of coroutines that is suitable for writing sequential code,
  26  * rather than callbacks, for operations that need to give up control while
  27  * waiting for events to complete.
  28  *
  29  * These functions are re-entrant and may be used outside the global mutex.
  30  *
  31  * Functions that execute in coroutine context cannot be called
  32  * directly from normal functions.  Use @coroutine_fn to mark such
  33  * functions.  For example:
  34  *
  35  *   static void coroutine_fn foo(void) {
  36  *       ....
  37  *   }
  38  *
  39  * In the future it would be nice to have the compiler or a static
  40  * checker catch misuse of such functions.  This annotation might make
  41  * it possible and in the meantime it serves as documentation.
  42  */
  43
  44 /**
  45  * Provides a mutex that can be used to synchronise coroutines
  46  */
  47 struct CoWaitRecord;
  48 struct CoMutex {
  49     /* Count of pending lockers; 0 for a free mutex, 1 for an
  50      * uncontended mutex.
  51      */
  52     unsigned locked;
  53
  54     /* Context that is holding the lock.  Useful to avoid spinning
  55      * when two coroutines on the same AioContext try to get the lock. :)
  56      */
  57     AioContext *ctx;
  58
  59     /* A queue of waiters.  Elements are added atomically in front of
  60      * from_push.  to_pop is only populated, and popped from, by whoever
  61      * is in charge of the next wakeup.  This can be an unlocker or,
  62      * through the handoff protocol, a locker that is about to go to sleep.
  63      */
  64     QSLIST_HEAD(, CoWaitRecord) from_push, to_pop;
  65
  66     unsigned handoff, sequence;
  67
  68     Coroutine *holder;
  69 };
  70
  71 /**
  72  * Assert that the current coroutine holds @mutex.
  73  */
  74 static inline coroutine_fn void qemu_co_mutex_assert_locked(CoMutex *mutex)
  75 {
  76     /*
  77      * mutex->holder doesn't need any synchronisation if the assertion holds
  78      * true because the mutex protects it. If it doesn't hold true, we still
  79      * don't mind if another thread takes or releases mutex behind our back,
  80      * because the condition will be false no matter whether we read NULL or
  81      * the pointer for any other coroutine.
  82      */
  83     assert(qatomic_read(&mutex->locked) &&
  84            mutex->holder == qemu_coroutine_self());
  85 }
  86
  87 /**
  88  * CoQueues are a mechanism to queue coroutines in order to continue executing
  89  * them later.  They are similar to condition variables, but they need help
  90  * from an external mutex in order to maintain thread-safety.
  91  */
  92 typedef struct CoQueue {
  93     QSIMPLEQ_HEAD(, Coroutine) entries;
  94 } CoQueue;
  95
  96 /**
  97  * Initialise a CoQueue. This must be called before any other operation is used
  98  * on the CoQueue.
  99  */
 100 void qemu_co_queue_init(CoQueue *queue);
 101
 102 typedef enum {
 103     /*
 104      * Enqueue at front instead of back. Use this to re-queue a request when
 105      * its wait condition is not satisfied after being woken up.
 106      */
 107     CO_QUEUE_WAIT_FRONT = 0x1,
 108 } CoQueueWaitFlags;
 109
 110 /**
 111  * Adds the current coroutine to the CoQueue and transfers control to the
 112  * caller of the coroutine.  The mutex is unlocked during the wait and
 113  * locked again afterwards.
 114  */
 115 #define qemu_co_queue_wait(queue, lock) \
 116     qemu_co_queue_wait_impl(queue, QEMU_MAKE_LOCKABLE(lock), 0)
 117 #define qemu_co_queue_wait_flags(queue, lock, flags) \
 118     qemu_co_queue_wait_impl(queue, QEMU_MAKE_LOCKABLE(lock), (flags))
 119 void coroutine_fn qemu_co_queue_wait_impl(CoQueue *queue, QemuLockable *lock,
 120                                           CoQueueWaitFlags flags);
 121
 122 /**
 123  * Removes the next coroutine from the CoQueue, and queue it to run after
 124  * the currently-running coroutine yields.
 125  * Returns true if a coroutine was removed, false if the queue is empty.
 126  * Used from coroutine context, use qemu_co_enter_next outside.
 127  */
 128 bool coroutine_fn qemu_co_queue_next(CoQueue *queue);
 129
 130 /**
 131  * Empties the CoQueue and queues the coroutine to run after
 132  * the currently-running coroutine yields.
 133  * Used from coroutine context, use qemu_co_enter_all outside.
 134  */
 135 void coroutine_fn qemu_co_queue_restart_all(CoQueue *queue);
 136
 137 /**
 138  * Removes the next coroutine from the CoQueue, and wake it up.  Unlike
 139  * qemu_co_queue_next, this function releases the lock during aio_co_wake
 140  * because it is meant to be used outside coroutine context; in that case, the
 141  * coroutine is entered immediately, before qemu_co_enter_next returns.
 142  *
 143  * If used in coroutine context, qemu_co_enter_next is equivalent to
 144  * qemu_co_queue_next.
 145  */
 146 #define qemu_co_enter_next(queue, lock) \
 147     qemu_co_enter_next_impl(queue, QEMU_MAKE_LOCKABLE(lock))
 148 bool qemu_co_enter_next_impl(CoQueue *queue, QemuLockable *lock);
 149
 150 /**
 151  * Empties the CoQueue, waking the waiting coroutine one at a time.  Unlike
 152  * qemu_co_queue_all, this function releases the lock during aio_co_wake
 153  * because it is meant to be used outside coroutine context; in that case, the
 154  * coroutine is entered immediately, before qemu_co_enter_all returns.
 155  *
 156  * If used in coroutine context, qemu_co_enter_all is equivalent to
 157  * qemu_co_queue_all.
 158  */
 159 #define qemu_co_enter_all(queue, lock) \
 160     qemu_co_enter_all_impl(queue, QEMU_MAKE_LOCKABLE(lock))
 161 void qemu_co_enter_all_impl(CoQueue *queue, QemuLockable *lock);
 162
 163 /**
 164  * Checks if the CoQueue is empty.
 165  */
 166 bool qemu_co_queue_empty(CoQueue *queue);
 167
 168
 169 typedef struct CoRwTicket CoRwTicket;
 170 typedef struct CoRwlock {
 171     CoMutex mutex;
 172
 173     /* Number of readers, or -1 if owned for writing.  */
 174     int owners;
 175
 176     /* Waiting coroutines.  */
 177     QSIMPLEQ_HEAD(, CoRwTicket) tickets;
 178 } CoRwlock;
 179
 180 /**
 181  * Initialises a CoRwlock. This must be called before any other operation
 182  * is used on the CoRwlock
 183  */
 184 void qemu_co_rwlock_init(CoRwlock *lock);
 185
 186 /**
 187  * Read locks the CoRwlock. If the lock cannot be taken immediately because
 188  * of a parallel writer, control is transferred to the caller of the current
 189  * coroutine.
 190  */
 191 void coroutine_fn qemu_co_rwlock_rdlock(CoRwlock *lock);
 192
 193 /**
 194  * Write Locks the CoRwlock from a reader.  This is a bit more efficient than
 195  * @qemu_co_rwlock_unlock followed by a separate @qemu_co_rwlock_wrlock.
 196  * Note that if the lock cannot be upgraded immediately, control is transferred
 197  * to the caller of the current coroutine; another writer might run while
 198  * @qemu_co_rwlock_upgrade blocks.
 199  */
 200 void coroutine_fn qemu_co_rwlock_upgrade(CoRwlock *lock);
 201
 202 /**
 203  * Downgrades a write-side critical section to a reader.  Downgrading with
 204  * @qemu_co_rwlock_downgrade never blocks, unlike @qemu_co_rwlock_unlock
 205  * followed by @qemu_co_rwlock_rdlock.  This makes it more efficient, but
 206  * may also sometimes be necessary for correctness.
 207  */
 208 void coroutine_fn qemu_co_rwlock_downgrade(CoRwlock *lock);
 209
 210 /**
 211  * Write Locks the mutex. If the lock cannot be taken immediately because
 212  * of a parallel reader, control is transferred to the caller of the current
 213  * coroutine.
 214  */
 215 void coroutine_fn qemu_co_rwlock_wrlock(CoRwlock *lock);
 216
 217 /**
 218  * Unlocks the read/write lock and schedules the next coroutine that was
 219  * waiting for this lock to be run.
 220  */
 221 void coroutine_fn qemu_co_rwlock_unlock(CoRwlock *lock);
 222
 223 typedef struct QemuCoSleep {
 224     Coroutine *to_wake;
 225 } QemuCoSleep;
 226
 227 /**
 228  * Yield the coroutine for a given duration. Initializes @w so that,
 229  * during this yield, it can be passed to qemu_co_sleep_wake() to
 230  * terminate the sleep.
 231  */
 232 void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w,
 233                                             QEMUClockType type, int64_t ns);
 234
 235 /**
 236  * Yield the coroutine until the next call to qemu_co_sleep_wake.
 237  */
 238 void coroutine_fn qemu_co_sleep(QemuCoSleep *w);
 239
 240 static inline void coroutine_fn qemu_co_sleep_ns(QEMUClockType type, int64_t ns)
 241 {
 242     QemuCoSleep w = { 0 };
 243     qemu_co_sleep_ns_wakeable(&w, type, ns);
 244 }
 245
 246 typedef void CleanupFunc(void *opaque);
 247 /**
 248  * Run entry in a coroutine and start timer. Wait for entry to finish or for
 249  * timer to elapse, what happen first. If entry finished, return 0, if timer
 250  * elapsed earlier, return -ETIMEDOUT.
 251  *
 252  * Be careful, entry execution is not canceled, user should handle it somehow.
 253  * If @clean is provided, it's called after coroutine finish if timeout
 254  * happened.
 255  */
 256 int coroutine_fn qemu_co_timeout(CoroutineEntry *entry, void *opaque,
 257                                  uint64_t timeout_ns, CleanupFunc clean);
 258
 259 /**
 260  * Wake a coroutine if it is sleeping in qemu_co_sleep_ns. The timer will be
 261  * deleted. @sleep_state must be the variable whose address was given to
 262  * qemu_co_sleep_ns() and should be checked to be non-NULL before calling
 263  * qemu_co_sleep_wake().
 264  */
 265 void qemu_co_sleep_wake(QemuCoSleep *w);
 266
 267 /**
 268  * Yield until a file descriptor becomes readable
 269  *
 270  * Note that this function clobbers the handlers for the file descriptor.
 271  */
 272 void coroutine_fn yield_until_fd_readable(int fd);
 273
 274 /**
 275  * Increase coroutine pool size
 276  */
 277 void qemu_coroutine_inc_pool_size(unsigned int additional_pool_size);
 278
 279 /**
 280  * Decrease coroutine pool size
 281  */
 282 void qemu_coroutine_dec_pool_size(unsigned int additional_pool_size);
 283
 284 #include "qemu/lockable.h"
 285
 286 /**
 287  * Sends a (part of) iovec down a socket, yielding when the socket is full, or
 288  * Receives data into a (part of) iovec from a socket,
 289  * yielding when there is no data in the socket.
 290  * The same interface as qemu_sendv_recvv(), with added yielding.
 291  * XXX should mark these as coroutine_fn
 292  */
 293 ssize_t coroutine_fn qemu_co_sendv_recvv(int sockfd, struct iovec *iov,
 294                                          unsigned iov_cnt, size_t offset,
 295                                          size_t bytes, bool do_send);
 296 #define qemu_co_recvv(sockfd, iov, iov_cnt, offset, bytes) \
 297   qemu_co_sendv_recvv(sockfd, iov, iov_cnt, offset, bytes, false)
 298 #define qemu_co_sendv(sockfd, iov, iov_cnt, offset, bytes) \
 299   qemu_co_sendv_recvv(sockfd, iov, iov_cnt, offset, bytes, true)
 300
 301 /**
 302  * The same as above, but with just a single buffer
 303  */
 304 ssize_t coroutine_fn qemu_co_send_recv(int sockfd, void *buf, size_t bytes,
 305                                        bool do_send);
 306 #define qemu_co_recv(sockfd, buf, bytes) \
 307   qemu_co_send_recv(sockfd, buf, bytes, false)
 308 #define qemu_co_send(sockfd, buf, bytes) \
 309   qemu_co_send_recv(sockfd, buf, bytes, true)
 310
 311 #endif /* QEMU_COROUTINE_H */