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