1 /* 2 * fs/eventfd.c 3 * 4 * Copyright (C) 2007 Davide Libenzi <davidel@xmailserver.org> 5 * 6 */ 7 8 #include <linux/file.h> 9 #include <linux/poll.h> 10 #include <linux/init.h> 11 #include <linux/fs.h> 12 #include <linux/sched.h> 13 #include <linux/kernel.h> 14 #include <linux/list.h> 15 #include <linux/spinlock.h> 16 #include <linux/anon_inodes.h> 17 #include <linux/eventfd.h> 18 #include <linux/syscalls.h> 19 20 struct eventfd_ctx { 21 wait_queue_head_t wqh; 22 /* 23 * Every time that a write(2) is performed on an eventfd, the 24 * value of the __u64 being written is added to "count" and a 25 * wakeup is performed on "wqh". A read(2) will return the "count" 26 * value to userspace, and will reset "count" to zero. The kernel 27 * size eventfd_signal() also, adds to the "count" counter and 28 * issue a wakeup. 29 */ 30 __u64 count; 31 unsigned int flags; 32 }; 33 34 /* 35 * Adds "n" to the eventfd counter "count". Returns "n" in case of 36 * success, or a value lower then "n" in case of coutner overflow. 37 * This function is supposed to be called by the kernel in paths 38 * that do not allow sleeping. In this function we allow the counter 39 * to reach the ULLONG_MAX value, and we signal this as overflow 40 * condition by returining a POLLERR to poll(2). 41 */ 42 int eventfd_signal(struct file *file, int n) 43 { 44 struct eventfd_ctx *ctx = file->private_data; 45 unsigned long flags; 46 47 if (n < 0) 48 return -EINVAL; 49 spin_lock_irqsave(&ctx->wqh.lock, flags); 50 if (ULLONG_MAX - ctx->count < n) 51 n = (int) (ULLONG_MAX - ctx->count); 52 ctx->count += n; 53 if (waitqueue_active(&ctx->wqh)) 54 wake_up_locked_poll(&ctx->wqh, POLLIN); 55 spin_unlock_irqrestore(&ctx->wqh.lock, flags); 56 57 return n; 58 } 59 60 static int eventfd_release(struct inode *inode, struct file *file) 61 { 62 kfree(file->private_data); 63 return 0; 64 } 65 66 static unsigned int eventfd_poll(struct file *file, poll_table *wait) 67 { 68 struct eventfd_ctx *ctx = file->private_data; 69 unsigned int events = 0; 70 unsigned long flags; 71 72 poll_wait(file, &ctx->wqh, wait); 73 74 spin_lock_irqsave(&ctx->wqh.lock, flags); 75 if (ctx->count > 0) 76 events |= POLLIN; 77 if (ctx->count == ULLONG_MAX) 78 events |= POLLERR; 79 if (ULLONG_MAX - 1 > ctx->count) 80 events |= POLLOUT; 81 spin_unlock_irqrestore(&ctx->wqh.lock, flags); 82 83 return events; 84 } 85 86 static ssize_t eventfd_read(struct file *file, char __user *buf, size_t count, 87 loff_t *ppos) 88 { 89 struct eventfd_ctx *ctx = file->private_data; 90 ssize_t res; 91 __u64 ucnt = 0; 92 DECLARE_WAITQUEUE(wait, current); 93 94 if (count < sizeof(ucnt)) 95 return -EINVAL; 96 spin_lock_irq(&ctx->wqh.lock); 97 res = -EAGAIN; 98 if (ctx->count > 0) 99 res = sizeof(ucnt); 100 else if (!(file->f_flags & O_NONBLOCK)) { 101 __add_wait_queue(&ctx->wqh, &wait); 102 for (res = 0;;) { 103 set_current_state(TASK_INTERRUPTIBLE); 104 if (ctx->count > 0) { 105 res = sizeof(ucnt); 106 break; 107 } 108 if (signal_pending(current)) { 109 res = -ERESTARTSYS; 110 break; 111 } 112 spin_unlock_irq(&ctx->wqh.lock); 113 schedule(); 114 spin_lock_irq(&ctx->wqh.lock); 115 } 116 __remove_wait_queue(&ctx->wqh, &wait); 117 __set_current_state(TASK_RUNNING); 118 } 119 if (likely(res > 0)) { 120 ucnt = (ctx->flags & EFD_SEMAPHORE) ? 1 : ctx->count; 121 ctx->count -= ucnt; 122 if (waitqueue_active(&ctx->wqh)) 123 wake_up_locked_poll(&ctx->wqh, POLLOUT); 124 } 125 spin_unlock_irq(&ctx->wqh.lock); 126 if (res > 0 && put_user(ucnt, (__u64 __user *) buf)) 127 return -EFAULT; 128 129 return res; 130 } 131 132 static ssize_t eventfd_write(struct file *file, const char __user *buf, size_t count, 133 loff_t *ppos) 134 { 135 struct eventfd_ctx *ctx = file->private_data; 136 ssize_t res; 137 __u64 ucnt; 138 DECLARE_WAITQUEUE(wait, current); 139 140 if (count < sizeof(ucnt)) 141 return -EINVAL; 142 if (copy_from_user(&ucnt, buf, sizeof(ucnt))) 143 return -EFAULT; 144 if (ucnt == ULLONG_MAX) 145 return -EINVAL; 146 spin_lock_irq(&ctx->wqh.lock); 147 res = -EAGAIN; 148 if (ULLONG_MAX - ctx->count > ucnt) 149 res = sizeof(ucnt); 150 else if (!(file->f_flags & O_NONBLOCK)) { 151 __add_wait_queue(&ctx->wqh, &wait); 152 for (res = 0;;) { 153 set_current_state(TASK_INTERRUPTIBLE); 154 if (ULLONG_MAX - ctx->count > ucnt) { 155 res = sizeof(ucnt); 156 break; 157 } 158 if (signal_pending(current)) { 159 res = -ERESTARTSYS; 160 break; 161 } 162 spin_unlock_irq(&ctx->wqh.lock); 163 schedule(); 164 spin_lock_irq(&ctx->wqh.lock); 165 } 166 __remove_wait_queue(&ctx->wqh, &wait); 167 __set_current_state(TASK_RUNNING); 168 } 169 if (likely(res > 0)) { 170 ctx->count += ucnt; 171 if (waitqueue_active(&ctx->wqh)) 172 wake_up_locked_poll(&ctx->wqh, POLLIN); 173 } 174 spin_unlock_irq(&ctx->wqh.lock); 175 176 return res; 177 } 178 179 static const struct file_operations eventfd_fops = { 180 .release = eventfd_release, 181 .poll = eventfd_poll, 182 .read = eventfd_read, 183 .write = eventfd_write, 184 }; 185 186 struct file *eventfd_fget(int fd) 187 { 188 struct file *file; 189 190 file = fget(fd); 191 if (!file) 192 return ERR_PTR(-EBADF); 193 if (file->f_op != &eventfd_fops) { 194 fput(file); 195 return ERR_PTR(-EINVAL); 196 } 197 198 return file; 199 } 200 201 SYSCALL_DEFINE2(eventfd2, unsigned int, count, int, flags) 202 { 203 int fd; 204 struct eventfd_ctx *ctx; 205 206 /* Check the EFD_* constants for consistency. */ 207 BUILD_BUG_ON(EFD_CLOEXEC != O_CLOEXEC); 208 BUILD_BUG_ON(EFD_NONBLOCK != O_NONBLOCK); 209 210 if (flags & ~EFD_FLAGS_SET) 211 return -EINVAL; 212 213 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL); 214 if (!ctx) 215 return -ENOMEM; 216 217 init_waitqueue_head(&ctx->wqh); 218 ctx->count = count; 219 ctx->flags = flags; 220 221 /* 222 * When we call this, the initialization must be complete, since 223 * anon_inode_getfd() will install the fd. 224 */ 225 fd = anon_inode_getfd("[eventfd]", &eventfd_fops, ctx, 226 flags & EFD_SHARED_FCNTL_FLAGS); 227 if (fd < 0) 228 kfree(ctx); 229 return fd; 230 } 231 232 SYSCALL_DEFINE1(eventfd, unsigned int, count) 233 { 234 return sys_eventfd2(count, 0); 235 } 236 237