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