fs/eventfd.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  fs/eventfd.c
   4  *
   5  *  Copyright (C) 2007  Davide Libenzi <davidel@xmailserver.org>
   6  *
   7  */
   8
   9 #include <linux/file.h>
  10 #include <linux/poll.h>
  11 #include <linux/init.h>
  12 #include <linux/fs.h>
  13 #include <linux/sched/signal.h>
  14 #include <linux/kernel.h>
  15 #include <linux/slab.h>
  16 #include <linux/list.h>
  17 #include <linux/spinlock.h>
  18 #include <linux/anon_inodes.h>
  19 #include <linux/syscalls.h>
  20 #include <linux/export.h>
  21 #include <linux/kref.h>
  22 #include <linux/eventfd.h>
  23 #include <linux/proc_fs.h>
  24 #include <linux/seq_file.h>
  25 #include <linux/idr.h>
  26 #include <linux/uio.h>
  27
  28 static DEFINE_IDA(eventfd_ida);
  29
  30 struct eventfd_ctx {
  31         struct kref kref;
  32         wait_queue_head_t wqh;
  33         /*
  34          * Every time that a write(2) is performed on an eventfd, the
  35          * value of the __u64 being written is added to "count" and a
  36          * wakeup is performed on "wqh". A read(2) will return the "count"
  37          * value to userspace, and will reset "count" to zero. The kernel
  38          * side eventfd_signal() also, adds to the "count" counter and
  39          * issue a wakeup.
  40          */
  41         __u64 count;
  42         unsigned int flags;
  43         int id;
  44 };
  45
  46 __u64 eventfd_signal_mask(struct eventfd_ctx *ctx, __u64 n, unsigned mask)
  47 {
  48         unsigned long flags;
  49
  50         /*
  51          * Deadlock or stack overflow issues can happen if we recurse here
  52          * through waitqueue wakeup handlers. If the caller users potentially
  53          * nested waitqueues with custom wakeup handlers, then it should
  54          * check eventfd_signal_allowed() before calling this function. If
  55          * it returns false, the eventfd_signal() call should be deferred to a
  56          * safe context.
  57          */
  58         if (WARN_ON_ONCE(current->in_eventfd))
  59                 return 0;
  60
  61         spin_lock_irqsave(&ctx->wqh.lock, flags);
  62         current->in_eventfd = 1;
  63         if (ULLONG_MAX - ctx->count < n)
  64                 n = ULLONG_MAX - ctx->count;
  65         ctx->count += n;
  66         if (waitqueue_active(&ctx->wqh))
  67                 wake_up_locked_poll(&ctx->wqh, EPOLLIN | mask);
  68         current->in_eventfd = 0;
  69         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
  70
  71         return n;
  72 }
  73
  74 /**
  75  * eventfd_signal - Adds @n to the eventfd counter.
  76  * @ctx: [in] Pointer to the eventfd context.
  77  * @n: [in] Value of the counter to be added to the eventfd internal counter.
  78  *          The value cannot be negative.
  79  *
  80  * This function is supposed to be called by the kernel in paths that do not
  81  * allow sleeping. In this function we allow the counter to reach the ULLONG_MAX
  82  * value, and we signal this as overflow condition by returning a EPOLLERR
  83  * to poll(2).
  84  *
  85  * Returns the amount by which the counter was incremented.  This will be less
  86  * than @n if the counter has overflowed.
  87  */
  88 __u64 eventfd_signal(struct eventfd_ctx *ctx, __u64 n)
  89 {
  90         return eventfd_signal_mask(ctx, n, 0);
  91 }
  92 EXPORT_SYMBOL_GPL(eventfd_signal);
  93
  94 static void eventfd_free_ctx(struct eventfd_ctx *ctx)
  95 {
  96         if (ctx->id >= 0)
  97                 ida_simple_remove(&eventfd_ida, ctx->id);
  98         kfree(ctx);
  99 }
 100
 101 static void eventfd_free(struct kref *kref)
 102 {
 103         struct eventfd_ctx *ctx = container_of(kref, struct eventfd_ctx, kref);
 104
 105         eventfd_free_ctx(ctx);
 106 }
 107
 108 /**
 109  * eventfd_ctx_put - Releases a reference to the internal eventfd context.
 110  * @ctx: [in] Pointer to eventfd context.
 111  *
 112  * The eventfd context reference must have been previously acquired either
 113  * with eventfd_ctx_fdget() or eventfd_ctx_fileget().
 114  */
 115 void eventfd_ctx_put(struct eventfd_ctx *ctx)
 116 {
 117         kref_put(&ctx->kref, eventfd_free);
 118 }
 119 EXPORT_SYMBOL_GPL(eventfd_ctx_put);
 120
 121 static int eventfd_release(struct inode *inode, struct file *file)
 122 {
 123         struct eventfd_ctx *ctx = file->private_data;
 124
 125         wake_up_poll(&ctx->wqh, EPOLLHUP);
 126         eventfd_ctx_put(ctx);
 127         return 0;
 128 }
 129
 130 static __poll_t eventfd_poll(struct file *file, poll_table *wait)
 131 {
 132         struct eventfd_ctx *ctx = file->private_data;
 133         __poll_t events = 0;
 134         u64 count;
 135
 136         poll_wait(file, &ctx->wqh, wait);
 137
 138         /*
 139          * All writes to ctx->count occur within ctx->wqh.lock.  This read
 140          * can be done outside ctx->wqh.lock because we know that poll_wait
 141          * takes that lock (through add_wait_queue) if our caller will sleep.
 142          *
 143          * The read _can_ therefore seep into add_wait_queue's critical
 144          * section, but cannot move above it!  add_wait_queue's spin_lock acts
 145          * as an acquire barrier and ensures that the read be ordered properly
 146          * against the writes.  The following CAN happen and is safe:
 147          *
 148          *     poll                               write
 149          *     -----------------                  ------------
 150          *     lock ctx->wqh.lock (in poll_wait)
 151          *     count = ctx->count
 152          *     __add_wait_queue
 153          *     unlock ctx->wqh.lock
 154          *                                        lock ctx->qwh.lock
 155          *                                        ctx->count += n
 156          *                                        if (waitqueue_active)
 157          *                                          wake_up_locked_poll
 158          *                                        unlock ctx->qwh.lock
 159          *     eventfd_poll returns 0
 160          *
 161          * but the following, which would miss a wakeup, cannot happen:
 162          *
 163          *     poll                               write
 164          *     -----------------                  ------------
 165          *     count = ctx->count (INVALID!)
 166          *                                        lock ctx->qwh.lock
 167          *                                        ctx->count += n
 168          *                                        **waitqueue_active is false**
 169          *                                        **no wake_up_locked_poll!**
 170          *                                        unlock ctx->qwh.lock
 171          *     lock ctx->wqh.lock (in poll_wait)
 172          *     __add_wait_queue
 173          *     unlock ctx->wqh.lock
 174          *     eventfd_poll returns 0
 175          */
 176         count = READ_ONCE(ctx->count);
 177
 178         if (count > 0)
 179                 events |= EPOLLIN;
 180         if (count == ULLONG_MAX)
 181                 events |= EPOLLERR;
 182         if (ULLONG_MAX - 1 > count)
 183                 events |= EPOLLOUT;
 184
 185         return events;
 186 }
 187
 188 void eventfd_ctx_do_read(struct eventfd_ctx *ctx, __u64 *cnt)
 189 {
 190         lockdep_assert_held(&ctx->wqh.lock);
 191
 192         *cnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count;
 193         ctx->count -= *cnt;
 194 }
 195 EXPORT_SYMBOL_GPL(eventfd_ctx_do_read);
 196
 197 /**
 198  * eventfd_ctx_remove_wait_queue - Read the current counter and removes wait queue.
 199  * @ctx: [in] Pointer to eventfd context.
 200  * @wait: [in] Wait queue to be removed.
 201  * @cnt: [out] Pointer to the 64-bit counter value.
 202  *
 203  * Returns %0 if successful, or the following error codes:
 204  *
 205  * -EAGAIN      : The operation would have blocked.
 206  *
 207  * This is used to atomically remove a wait queue entry from the eventfd wait
 208  * queue head, and read/reset the counter value.
 209  */
 210 int eventfd_ctx_remove_wait_queue(struct eventfd_ctx *ctx, wait_queue_entry_t *wait,
 211                                   __u64 *cnt)
 212 {
 213         unsigned long flags;
 214
 215         spin_lock_irqsave(&ctx->wqh.lock, flags);
 216         eventfd_ctx_do_read(ctx, cnt);
 217         __remove_wait_queue(&ctx->wqh, wait);
 218         if (*cnt != 0 && waitqueue_active(&ctx->wqh))
 219                 wake_up_locked_poll(&ctx->wqh, EPOLLOUT);
 220         spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 221
 222         return *cnt != 0 ? 0 : -EAGAIN;
 223 }
 224 EXPORT_SYMBOL_GPL(eventfd_ctx_remove_wait_queue);
 225
 226 static ssize_t eventfd_read(struct kiocb *iocb, struct iov_iter *to)
 227 {
 228         struct file *file = iocb->ki_filp;
 229         struct eventfd_ctx *ctx = file->private_data;
 230         __u64 ucnt = 0;
 231         DECLARE_WAITQUEUE(wait, current);
 232
 233         if (iov_iter_count(to) < sizeof(ucnt))
 234                 return -EINVAL;
 235         spin_lock_irq(&ctx->wqh.lock);
 236         if (!ctx->count) {
 237                 if ((file->f_flags & O_NONBLOCK) ||
 238                     (iocb->ki_flags & IOCB_NOWAIT)) {
 239                         spin_unlock_irq(&ctx->wqh.lock);
 240                         return -EAGAIN;
 241                 }
 242                 __add_wait_queue(&ctx->wqh, &wait);
 243                 for (;;) {
 244                         set_current_state(TASK_INTERRUPTIBLE);
 245                         if (ctx->count)
 246                                 break;
 247                         if (signal_pending(current)) {
 248                                 __remove_wait_queue(&ctx->wqh, &wait);
 249                                 __set_current_state(TASK_RUNNING);
 250                                 spin_unlock_irq(&ctx->wqh.lock);
 251                                 return -ERESTARTSYS;
 252                         }
 253                         spin_unlock_irq(&ctx->wqh.lock);
 254                         schedule();
 255                         spin_lock_irq(&ctx->wqh.lock);
 256                 }
 257                 __remove_wait_queue(&ctx->wqh, &wait);
 258                 __set_current_state(TASK_RUNNING);
 259         }
 260         eventfd_ctx_do_read(ctx, &ucnt);
 261         current->in_eventfd = 1;
 262         if (waitqueue_active(&ctx->wqh))
 263                 wake_up_locked_poll(&ctx->wqh, EPOLLOUT);
 264         current->in_eventfd = 0;
 265         spin_unlock_irq(&ctx->wqh.lock);
 266         if (unlikely(copy_to_iter(&ucnt, sizeof(ucnt), to) != sizeof(ucnt)))
 267                 return -EFAULT;
 268
 269         return sizeof(ucnt);
 270 }
 271
 272 static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count,
 273                              loff_t *ppos)
 274 {
 275         struct eventfd_ctx *ctx = file->private_data;
 276         ssize_t res;
 277         __u64 ucnt;
 278         DECLARE_WAITQUEUE(wait, current);
 279
 280         if (count < sizeof(ucnt))
 281                 return -EINVAL;
 282         if (copy_from_user(&ucnt, buf, sizeof(ucnt)))
 283                 return -EFAULT;
 284         if (ucnt == ULLONG_MAX)
 285                 return -EINVAL;
 286         spin_lock_irq(&ctx->wqh.lock);
 287         res = -EAGAIN;
 288         if (ULLONG_MAX - ctx->count > ucnt)
 289                 res = sizeof(ucnt);
 290         else if (!(file->f_flags & O_NONBLOCK)) {
 291                 __add_wait_queue(&ctx->wqh, &wait);
 292                 for (res = 0;;) {
 293                         set_current_state(TASK_INTERRUPTIBLE);
 294                         if (ULLONG_MAX - ctx->count > ucnt) {
 295                                 res = sizeof(ucnt);
 296                                 break;
 297                         }
 298                         if (signal_pending(current)) {
 299                                 res = -ERESTARTSYS;
 300                                 break;
 301                         }
 302                         spin_unlock_irq(&ctx->wqh.lock);
 303                         schedule();
 304                         spin_lock_irq(&ctx->wqh.lock);
 305                 }
 306                 __remove_wait_queue(&ctx->wqh, &wait);
 307                 __set_current_state(TASK_RUNNING);
 308         }
 309         if (likely(res > 0)) {
 310                 ctx->count += ucnt;
 311                 current->in_eventfd = 1;
 312                 if (waitqueue_active(&ctx->wqh))
 313                         wake_up_locked_poll(&ctx->wqh, EPOLLIN);
 314                 current->in_eventfd = 0;
 315         }
 316         spin_unlock_irq(&ctx->wqh.lock);
 317
 318         return res;
 319 }
 320
 321 #ifdef CONFIG_PROC_FS
 322 static void eventfd_show_fdinfo(struct seq_file *m, struct file *f)
 323 {
 324         struct eventfd_ctx *ctx = f->private_data;
 325
 326         spin_lock_irq(&ctx->wqh.lock);
 327         seq_printf(m, "eventfd-count: %16llx\n",
 328                    (unsigned long long)ctx->count);
 329         spin_unlock_irq(&ctx->wqh.lock);
 330         seq_printf(m, "eventfd-id: %d\n", ctx->id);
 331 }
 332 #endif
 333
 334 static const struct file_operations eventfd_fops = {
 335 #ifdef CONFIG_PROC_FS
 336         .show_fdinfo    = eventfd_show_fdinfo,
 337 #endif
 338         .release        = eventfd_release,
 339         .poll           = eventfd_poll,
 340         .read_iter      = eventfd_read,
 341         .write          = eventfd_write,
 342         .llseek         = noop_llseek,
 343 };
 344
 345 /**
 346  * eventfd_fget - Acquire a reference of an eventfd file descriptor.
 347  * @fd: [in] Eventfd file descriptor.
 348  *
 349  * Returns a pointer to the eventfd file structure in case of success, or the
 350  * following error pointer:
 351  *
 352  * -EBADF    : Invalid @fd file descriptor.
 353  * -EINVAL   : The @fd file descriptor is not an eventfd file.
 354  */
 355 struct file *eventfd_fget(int fd)
 356 {
 357         struct file *file;
 358
 359         file = fget(fd);
 360         if (!file)
 361                 return ERR_PTR(-EBADF);
 362         if (file->f_op != &eventfd_fops) {
 363                 fput(file);
 364                 return ERR_PTR(-EINVAL);
 365         }
 366
 367         return file;
 368 }
 369 EXPORT_SYMBOL_GPL(eventfd_fget);
 370
 371 /**
 372  * eventfd_ctx_fdget - Acquires a reference to the internal eventfd context.
 373  * @fd: [in] Eventfd file descriptor.
 374  *
 375  * Returns a pointer to the internal eventfd context, otherwise the error
 376  * pointers returned by the following functions:
 377  *
 378  * eventfd_fget
 379  */
 380 struct eventfd_ctx *eventfd_ctx_fdget(int fd)
 381 {
 382         struct eventfd_ctx *ctx;
 383         struct fd f = fdget(fd);
 384         if (!f.file)
 385                 return ERR_PTR(-EBADF);
 386         ctx = eventfd_ctx_fileget(f.file);
 387         fdput(f);
 388         return ctx;
 389 }
 390 EXPORT_SYMBOL_GPL(eventfd_ctx_fdget);
 391
 392 /**
 393  * eventfd_ctx_fileget - Acquires a reference to the internal eventfd context.
 394  * @file: [in] Eventfd file pointer.
 395  *
 396  * Returns a pointer to the internal eventfd context, otherwise the error
 397  * pointer:
 398  *
 399  * -EINVAL   : The @fd file descriptor is not an eventfd file.
 400  */
 401 struct eventfd_ctx *eventfd_ctx_fileget(struct file *file)
 402 {
 403         struct eventfd_ctx *ctx;
 404
 405         if (file->f_op != &eventfd_fops)
 406                 return ERR_PTR(-EINVAL);
 407
 408         ctx = file->private_data;
 409         kref_get(&ctx->kref);
 410         return ctx;
 411 }
 412 EXPORT_SYMBOL_GPL(eventfd_ctx_fileget);
 413
 414 static int do_eventfd(unsigned int count, int flags)
 415 {
 416         struct eventfd_ctx *ctx;
 417         struct file *file;
 418         int fd;
 419
 420         /* Check the EFD_* constants for consistency.  */
 421         BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC);
 422         BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK);
 423
 424         if (flags & ~EFD_FLAGS_SET)
 425                 return -EINVAL;
 426
 427         ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
 428         if (!ctx)
 429                 return -ENOMEM;
 430
 431         kref_init(&ctx->kref);
 432         init_waitqueue_head(&ctx->wqh);
 433         ctx->count = count;
 434         ctx->flags = flags;
 435         ctx->id = ida_simple_get(&eventfd_ida, 0, 0, GFP_KERNEL);
 436
 437         flags &= EFD_SHARED_FCNTL_FLAGS;
 438         flags |= O_RDWR;
 439         fd = get_unused_fd_flags(flags);
 440         if (fd < 0)
 441                 goto err;
 442
 443         file = anon_inode_getfile("[eventfd]", &eventfd_fops, ctx, flags);
 444         if (IS_ERR(file)) {
 445                 put_unused_fd(fd);
 446                 fd = PTR_ERR(file);
 447                 goto err;
 448         }
 449
 450         file->f_mode |= FMODE_NOWAIT;
 451         fd_install(fd, file);
 452         return fd;
 453 err:
 454         eventfd_free_ctx(ctx);
 455         return fd;
 456 }
 457
 458 SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags)
 459 {
 460         return do_eventfd(count, flags);
 461 }
 462
 463 SYSCALL_DEFINE1(eventfd, unsigned int, count)
 464 {
 465         return do_eventfd(count, 0);
 466 }
 467