4 * Copyright (C) 2003 Linus Torvalds
6 * Mon Mar 5, 2007: Davide Libenzi <davidel@xmailserver.org>
7 * Changed ->read() to return a siginfo strcture instead of signal number.
8 * Fixed locking in ->poll().
9 * Added sighand-detach notification.
10 * Added fd re-use in sys_signalfd() syscall.
11 * Now using anonymous inode source.
12 * Thanks to Oleg Nesterov for useful code review and suggestions.
13 * More comments and suggestions from Arnd Bergmann.
14 * Sat May 19, 2007: Davi E. M. Arnaut <davi@haxent.com.br>
15 * Retrieve multiple signals with one read() call
18 #include <linux/file.h>
19 #include <linux/poll.h>
20 #include <linux/init.h>
22 #include <linux/sched.h>
23 #include <linux/kernel.h>
24 #include <linux/signal.h>
25 #include <linux/list.h>
26 #include <linux/anon_inodes.h>
27 #include <linux/signalfd.h>
31 wait_queue_head_t wqh;
33 struct task_struct *tsk;
36 struct signalfd_lockctx {
37 struct task_struct *tsk;
42 * Tries to acquire the sighand lock. We do not increment the sighand
43 * use count, and we do not even pin the task struct, so we need to
44 * do it inside an RCU read lock, and we must be prepared for the
45 * ctx->tsk going to NULL (in signalfd_deliver()), and for the sighand
46 * being detached. We return 0 if the sighand has been detached, or
47 * 1 if we were able to pin the sighand lock.
49 static int signalfd_lock(struct signalfd_ctx *ctx, struct signalfd_lockctx *lk)
51 struct sighand_struct *sighand = NULL;
54 lk->tsk = rcu_dereference(ctx->tsk);
55 if (likely(lk->tsk != NULL))
56 sighand = lock_task_sighand(lk->tsk, &lk->flags);
59 if (sighand && !ctx->tsk) {
60 unlock_task_sighand(lk->tsk, &lk->flags);
64 return sighand != NULL;
67 static void signalfd_unlock(struct signalfd_lockctx *lk)
69 unlock_task_sighand(lk->tsk, &lk->flags);
73 * This must be called with the sighand lock held.
75 void signalfd_deliver(struct task_struct *tsk, int sig)
77 struct sighand_struct *sighand = tsk->sighand;
78 struct signalfd_ctx *ctx, *tmp;
81 list_for_each_entry_safe(ctx, tmp, &sighand->signalfd_list, lnk) {
83 * We use a negative signal value as a way to broadcast that the
84 * sighand has been orphaned, so that we can notify all the
85 * listeners about this. Remember the ctx->sigmask is inverted,
86 * so if the user is interested in a signal, that corresponding
90 if (ctx->tsk == tsk) {
92 list_del_init(&ctx->lnk);
96 if (!sigismember(&ctx->sigmask, sig))
102 static void signalfd_cleanup(struct signalfd_ctx *ctx)
104 struct signalfd_lockctx lk;
107 * This is tricky. If the sighand is gone, we do not need to remove
108 * context from the list, the list itself won't be there anymore.
110 if (signalfd_lock(ctx, &lk)) {
112 signalfd_unlock(&lk);
117 static int signalfd_release(struct inode *inode, struct file *file)
119 signalfd_cleanup(file->private_data);
123 static unsigned int signalfd_poll(struct file *file, poll_table *wait)
125 struct signalfd_ctx *ctx = file->private_data;
126 unsigned int events = 0;
127 struct signalfd_lockctx lk;
129 poll_wait(file, &ctx->wqh, wait);
132 * Let the caller get a POLLIN in this case, ala socket recv() when
133 * the peer disconnects.
135 if (signalfd_lock(ctx, &lk)) {
136 if ((lk.tsk == current &&
137 next_signal(&lk.tsk->pending, &ctx->sigmask) > 0) ||
138 next_signal(&lk.tsk->signal->shared_pending,
141 signalfd_unlock(&lk);
149 * Copied from copy_siginfo_to_user() in kernel/signal.c
151 static int signalfd_copyinfo(struct signalfd_siginfo __user *uinfo,
152 siginfo_t const *kinfo)
156 BUILD_BUG_ON(sizeof(struct signalfd_siginfo) != 128);
159 * Unused memebers should be zero ...
161 err = __clear_user(uinfo, sizeof(*uinfo));
164 * If you change siginfo_t structure, please be sure
165 * this code is fixed accordingly.
167 err |= __put_user(kinfo->si_signo, &uinfo->signo);
168 err |= __put_user(kinfo->si_errno, &uinfo->err);
169 err |= __put_user((short)kinfo->si_code, &uinfo->code);
170 switch (kinfo->si_code & __SI_MASK) {
172 err |= __put_user(kinfo->si_pid, &uinfo->pid);
173 err |= __put_user(kinfo->si_uid, &uinfo->uid);
176 err |= __put_user(kinfo->si_tid, &uinfo->tid);
177 err |= __put_user(kinfo->si_overrun, &uinfo->overrun);
178 err |= __put_user((long)kinfo->si_ptr, &uinfo->svptr);
181 err |= __put_user(kinfo->si_band, &uinfo->band);
182 err |= __put_user(kinfo->si_fd, &uinfo->fd);
185 err |= __put_user((long)kinfo->si_addr, &uinfo->addr);
186 #ifdef __ARCH_SI_TRAPNO
187 err |= __put_user(kinfo->si_trapno, &uinfo->trapno);
191 err |= __put_user(kinfo->si_pid, &uinfo->pid);
192 err |= __put_user(kinfo->si_uid, &uinfo->uid);
193 err |= __put_user(kinfo->si_status, &uinfo->status);
194 err |= __put_user(kinfo->si_utime, &uinfo->utime);
195 err |= __put_user(kinfo->si_stime, &uinfo->stime);
197 case __SI_RT: /* This is not generated by the kernel as of now. */
198 case __SI_MESGQ: /* But this is */
199 err |= __put_user(kinfo->si_pid, &uinfo->pid);
200 err |= __put_user(kinfo->si_uid, &uinfo->uid);
201 err |= __put_user((long)kinfo->si_ptr, &uinfo->svptr);
203 default: /* this is just in case for now ... */
204 err |= __put_user(kinfo->si_pid, &uinfo->pid);
205 err |= __put_user(kinfo->si_uid, &uinfo->uid);
209 return err ? -EFAULT: sizeof(*uinfo);
212 static ssize_t signalfd_dequeue(struct signalfd_ctx *ctx, siginfo_t *info,
216 struct signalfd_lockctx lk;
217 DECLARE_WAITQUEUE(wait, current);
219 if (!signalfd_lock(ctx, &lk))
222 ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
229 signalfd_unlock(&lk);
233 add_wait_queue(&ctx->wqh, &wait);
235 set_current_state(TASK_INTERRUPTIBLE);
236 ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
237 signalfd_unlock(&lk);
240 if (signal_pending(current)) {
245 ret = signalfd_lock(ctx, &lk);
246 if (unlikely(!ret)) {
248 * Let the caller read zero byte, ala socket
249 * recv() when the peer disconnect. This test
250 * must be done before doing a dequeue_signal(),
251 * because if the sighand has been orphaned,
252 * the dequeue_signal() call is going to crash
253 * because ->sighand will be long gone.
259 remove_wait_queue(&ctx->wqh, &wait);
260 __set_current_state(TASK_RUNNING);
266 * Returns either the size of a "struct signalfd_siginfo", or zero if the
267 * sighand we are attached to, has been orphaned. The "count" parameter
268 * must be at least the size of a "struct signalfd_siginfo".
270 static ssize_t signalfd_read(struct file *file, char __user *buf, size_t count,
273 struct signalfd_ctx *ctx = file->private_data;
274 struct signalfd_siginfo __user *siginfo;
275 int nonblock = file->f_flags & O_NONBLOCK;
276 ssize_t ret, total = 0;
279 count /= sizeof(struct signalfd_siginfo);
283 siginfo = (struct signalfd_siginfo __user *) buf;
286 ret = signalfd_dequeue(ctx, &info, nonblock);
287 if (unlikely(ret <= 0))
289 ret = signalfd_copyinfo(siginfo, &info);
297 return total ? total : ret;
300 static const struct file_operations signalfd_fops = {
301 .release = signalfd_release,
302 .poll = signalfd_poll,
303 .read = signalfd_read,
307 * Create a file descriptor that is associated with our signal
308 * state. We can pass it around to others if we want to, but
309 * it will always be _our_ signal state.
311 asmlinkage long sys_signalfd(int ufd, sigset_t __user *user_mask, size_t sizemask)
315 struct signalfd_ctx *ctx;
316 struct sighand_struct *sighand;
319 struct signalfd_lockctx lk;
321 if (sizemask != sizeof(sigset_t) ||
322 copy_from_user(&sigmask, user_mask, sizeof(sigmask)))
323 return error = -EINVAL;
324 sigdelsetmask(&sigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
328 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
332 init_waitqueue_head(&ctx->wqh);
333 ctx->sigmask = sigmask;
336 sighand = current->sighand;
338 * Add this fd to the list of signal listeners.
340 spin_lock_irq(&sighand->siglock);
341 list_add_tail(&ctx->lnk, &sighand->signalfd_list);
342 spin_unlock_irq(&sighand->siglock);
345 * When we call this, the initialization must be complete, since
346 * anon_inode_getfd() will install the fd.
348 error = anon_inode_getfd(&ufd, &inode, &file, "[signalfd]",
349 &signalfd_fops, ctx);
356 ctx = file->private_data;
357 if (file->f_op != &signalfd_fops) {
362 * We need to be prepared of the fact that the sighand this fd
363 * is attached to, has been detched. In that case signalfd_lock()
364 * will return 0, and we'll just skip setting the new mask.
366 if (signalfd_lock(ctx, &lk)) {
367 ctx->sigmask = sigmask;
368 signalfd_unlock(&lk);
377 signalfd_cleanup(ctx);