On 12/28, Andrey Vagin wrote:

pread(fd, buf, size, pos) with non-zero pos returns siginfo-s
without dequeuing signals.

A sequence number and a queue are encoded in pos.

pos = seq + SFD_*_OFFSET

seq is a sequence number of a signal in a queue.

SFD_PER_THREAD_QUEUE_OFFSET - read signals from a per-thread queue.
SFD_SHARED_QUEUE_OFFSET - read signals from a shared (process wide) queue.

This functionality is required for checkpointing pending signals.

v2: llseek() can't be used here, because peek_offset/f_pos/whatever
has to be shared with all processes which have this file opened.

Suppose that the task forks after sys_signalfd(). Now if parent or child
do llseek this affects them both. This is insane because signalfd is
"strange" to say at least, fork/dup/etc inherits signalfd_ctx but not
the" source" of the data. // Oleg Nesterov

I think we should cc Linus.

This patch adds the hack and it makes signalfd even more strange.

Yes, this hack was suggested by me because I can't suggest something
better. But if Linus dislikes this user-visible API it would be better
to get his nack right now.

+static ssize_t signalfd_peek(struct signalfd_ctx *ctx,
+				siginfo_t *info, loff_t *ppos)
+{
+	struct sigpending *pending;
+	struct sigqueue *q;
+	loff_t seq;
+	int ret = 0;
+
+	spin_lock_irq(&current->sighand->siglock);
+
+	if (*ppos >= SFD_SHARED_QUEUE_OFFSET) {
+		pending = &current->signal->shared_pending;
+		seq = *ppos - SFD_SHARED_QUEUE_OFFSET;
+	} else {
+		pending = &current->pending;
+		seq = *ppos - SFD_PER_THREAD_QUEUE_OFFSET;
+	}

You can do this outside of spin_lock_irq().

And I think it would be better to check SFD_PRIVATE_QUEUE_OFFSET too
although this is not strictly necessary. Otherwise this code assumes
that sys_pread() cheks pos >= 0 and SFD_PRIVATE_QUEUE_OFFSET == 1.

+	list_for_each_entry(q, &pending->list, list) {
+		if (sigismember(&ctx->sigmask, q->info.si_signo))
+			continue;
+
+		if (seq-- == 0) {
+			copy_siginfo(info, &q->info);
+			ret = info->si_signo;
+			break;
+		}
+	}
+
+	spin_unlock_irq(&current->sighand->siglock);
+
+	if (ret)
+		(*ppos)++;

We can change it unconditionally but I won't argue.

Again, this is not needed or the code was broken by the previous patch.

Given that 2/3 passes O_RDWR to anon_inode_getfile() I think FMODE_PREAD
should be already set. Note OPEN_FMODE(flags) in anon_inode_getfile().

Oleg.