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);
63 unlock_task_sighand(lk->tsk, &lk->flags);
67 if (lk->tsk->tgid == current->tgid)
73 static void signalfd_unlock(struct signalfd_lockctx *lk)
75 unlock_task_sighand(lk->tsk, &lk->flags);
79 * This must be called with the sighand lock held.
81 void signalfd_deliver(struct task_struct *tsk, int sig)
83 struct sighand_struct *sighand = tsk->sighand;
84 struct signalfd_ctx *ctx, *tmp;
87 list_for_each_entry_safe(ctx, tmp, &sighand->signalfd_list, lnk) {
89 * We use a negative signal value as a way to broadcast that the
90 * sighand has been orphaned, so that we can notify all the
91 * listeners about this. Remember the ctx->sigmask is inverted,
92 * so if the user is interested in a signal, that corresponding
96 if (ctx->tsk == tsk) {
98 list_del_init(&ctx->lnk);
102 if (!sigismember(&ctx->sigmask, sig))
108 static void signalfd_cleanup(struct signalfd_ctx *ctx)
110 struct signalfd_lockctx lk;
113 * This is tricky. If the sighand is gone, we do not need to remove
114 * context from the list, the list itself won't be there anymore.
116 if (signalfd_lock(ctx, &lk)) {
118 signalfd_unlock(&lk);
123 static int signalfd_release(struct inode *inode, struct file *file)
125 signalfd_cleanup(file->private_data);
129 static unsigned int signalfd_poll(struct file *file, poll_table *wait)
131 struct signalfd_ctx *ctx = file->private_data;
132 unsigned int events = 0;
133 struct signalfd_lockctx lk;
135 poll_wait(file, &ctx->wqh, wait);
138 * Let the caller get a POLLIN in this case, ala socket recv() when
139 * the peer disconnects.
141 if (signalfd_lock(ctx, &lk)) {
142 if ((lk.tsk == current &&
143 next_signal(&lk.tsk->pending, &ctx->sigmask) > 0) ||
144 next_signal(&lk.tsk->signal->shared_pending,
147 signalfd_unlock(&lk);
155 * Copied from copy_siginfo_to_user() in kernel/signal.c
157 static int signalfd_copyinfo(struct signalfd_siginfo __user *uinfo,
158 siginfo_t const *kinfo)
162 BUILD_BUG_ON(sizeof(struct signalfd_siginfo) != 128);
165 * Unused memebers should be zero ...
167 err = __clear_user(uinfo, sizeof(*uinfo));
170 * If you change siginfo_t structure, please be sure
171 * this code is fixed accordingly.
173 err |= __put_user(kinfo->si_signo, &uinfo->signo);
174 err |= __put_user(kinfo->si_errno, &uinfo->err);
175 err |= __put_user((short)kinfo->si_code, &uinfo->code);
176 switch (kinfo->si_code & __SI_MASK) {
178 err |= __put_user(kinfo->si_pid, &uinfo->pid);
179 err |= __put_user(kinfo->si_uid, &uinfo->uid);
182 err |= __put_user(kinfo->si_tid, &uinfo->tid);
183 err |= __put_user(kinfo->si_overrun, &uinfo->overrun);
184 err |= __put_user((long)kinfo->si_ptr, &uinfo->svptr);
187 err |= __put_user(kinfo->si_band, &uinfo->band);
188 err |= __put_user(kinfo->si_fd, &uinfo->fd);
191 err |= __put_user((long)kinfo->si_addr, &uinfo->addr);
192 #ifdef __ARCH_SI_TRAPNO
193 err |= __put_user(kinfo->si_trapno, &uinfo->trapno);
197 err |= __put_user(kinfo->si_pid, &uinfo->pid);
198 err |= __put_user(kinfo->si_uid, &uinfo->uid);
199 err |= __put_user(kinfo->si_status, &uinfo->status);
200 err |= __put_user(kinfo->si_utime, &uinfo->utime);
201 err |= __put_user(kinfo->si_stime, &uinfo->stime);
203 case __SI_RT: /* This is not generated by the kernel as of now. */
204 case __SI_MESGQ: /* But this is */
205 err |= __put_user(kinfo->si_pid, &uinfo->pid);
206 err |= __put_user(kinfo->si_uid, &uinfo->uid);
207 err |= __put_user((long)kinfo->si_ptr, &uinfo->svptr);
209 default: /* this is just in case for now ... */
210 err |= __put_user(kinfo->si_pid, &uinfo->pid);
211 err |= __put_user(kinfo->si_uid, &uinfo->uid);
215 return err ? -EFAULT: sizeof(*uinfo);
218 static ssize_t signalfd_dequeue(struct signalfd_ctx *ctx, siginfo_t *info,
222 struct signalfd_lockctx lk;
223 DECLARE_WAITQUEUE(wait, current);
225 if (!signalfd_lock(ctx, &lk))
228 ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
235 signalfd_unlock(&lk);
239 add_wait_queue(&ctx->wqh, &wait);
241 set_current_state(TASK_INTERRUPTIBLE);
242 ret = dequeue_signal(lk.tsk, &ctx->sigmask, info);
243 signalfd_unlock(&lk);
246 if (signal_pending(current)) {
251 ret = signalfd_lock(ctx, &lk);
252 if (unlikely(!ret)) {
254 * Let the caller read zero byte, ala socket
255 * recv() when the peer disconnect. This test
256 * must be done before doing a dequeue_signal(),
257 * because if the sighand has been orphaned,
258 * the dequeue_signal() call is going to crash
259 * because ->sighand will be long gone.
265 remove_wait_queue(&ctx->wqh, &wait);
266 __set_current_state(TASK_RUNNING);
272 * Returns either the size of a "struct signalfd_siginfo", or zero if the
273 * sighand we are attached to, has been orphaned. The "count" parameter
274 * must be at least the size of a "struct signalfd_siginfo".
276 static ssize_t signalfd_read(struct file *file, char __user *buf, size_t count,
279 struct signalfd_ctx *ctx = file->private_data;
280 struct signalfd_siginfo __user *siginfo;
281 int nonblock = file->f_flags & O_NONBLOCK;
282 ssize_t ret, total = 0;
285 count /= sizeof(struct signalfd_siginfo);
289 siginfo = (struct signalfd_siginfo __user *) buf;
292 ret = signalfd_dequeue(ctx, &info, nonblock);
293 if (unlikely(ret <= 0))
295 ret = signalfd_copyinfo(siginfo, &info);
303 return total ? total : ret;
306 static const struct file_operations signalfd_fops = {
307 .release = signalfd_release,
308 .poll = signalfd_poll,
309 .read = signalfd_read,
313 * Create a file descriptor that is associated with our signal
314 * state. We can pass it around to others if we want to, but
315 * it will always be _our_ signal state.
317 asmlinkage long sys_signalfd(int ufd, sigset_t __user *user_mask, size_t sizemask)
321 struct signalfd_ctx *ctx;
322 struct sighand_struct *sighand;
325 struct signalfd_lockctx lk;
327 if (sizemask != sizeof(sigset_t) ||
328 copy_from_user(&sigmask, user_mask, sizeof(sigmask)))
330 sigdelsetmask(&sigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
334 ctx = kmalloc(sizeof(*ctx), GFP_KERNEL);
338 init_waitqueue_head(&ctx->wqh);
339 ctx->sigmask = sigmask;
340 ctx->tsk = current->group_leader;
342 sighand = current->sighand;
344 * Add this fd to the list of signal listeners.
346 spin_lock_irq(&sighand->siglock);
347 list_add_tail(&ctx->lnk, &sighand->signalfd_list);
348 spin_unlock_irq(&sighand->siglock);
351 * When we call this, the initialization must be complete, since
352 * anon_inode_getfd() will install the fd.
354 error = anon_inode_getfd(&ufd, &inode, &file, "[signalfd]",
355 &signalfd_fops, ctx);
362 ctx = file->private_data;
363 if (file->f_op != &signalfd_fops) {
368 * We need to be prepared of the fact that the sighand this fd
369 * is attached to, has been detched. In that case signalfd_lock()
370 * will return 0, and we'll just skip setting the new mask.
372 if (signalfd_lock(ctx, &lk)) {
373 ctx->sigmask = sigmask;
374 signalfd_unlock(&lk);
383 signalfd_cleanup(ctx);