splice: lift pipe_lock out of splice_to_pipe()
authorAl Viro <viro@zeniv.linux.org.uk>
Sun, 18 Sep 2016 00:44:45 +0000 (20:44 -0400)
committerAl Viro <viro@zeniv.linux.org.uk>
Tue, 4 Oct 2016 00:40:55 +0000 (20:40 -0400)
* splice_to_pipe() stops at pipe overflow and does *not* take pipe_lock
* ->splice_read() instances do the same
* vmsplice_to_pipe() and do_splice() (ultimate callers of splice_to_pipe())
  arrange for waiting, looping, etc. themselves.

That should make pipe_lock the outermost one.

Unfortunately, existing rules for the amount passed by vmsplice_to_pipe()
and do_splice() are quite ugly _and_ userland code can be easily broken
by changing those.  It's not even "no more than the maximal capacity of
this pipe" - it's "once we'd fed pipe->nr_buffers pages into the pipe,
leave instead of waiting".

Considering how poorly these rules are documented, let's try "wait for some
space to appear, unless given SPLICE_F_NONBLOCK, then push into pipe
and if we run into overflow, we are done".

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
fs/fuse/dev.c
fs/splice.c

index a94d2ed81ab4af4bf7cc8316d5f7517a9902c3a2..eaf56c6e912389c86d975c0d2e334252df3f6ab8 100644 (file)
@@ -1364,7 +1364,6 @@ static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
                goto out;
 
        ret = 0;
-       pipe_lock(pipe);
 
        if (!pipe->readers) {
                send_sig(SIGPIPE, current, 0);
@@ -1400,7 +1399,6 @@ static ssize_t fuse_dev_splice_read(struct file *in, loff_t *ppos,
        }
 
 out_unlock:
-       pipe_unlock(pipe);
 
        if (do_wakeup) {
                smp_mb();
index 31c52e0269c2cb82d263682f809b3f172c9c4007..2ca7e6931ad560893e463df72b42a738b8c129f5 100644 (file)
@@ -183,79 +183,41 @@ ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
                       struct splice_pipe_desc *spd)
 {
        unsigned int spd_pages = spd->nr_pages;
-       int ret, do_wakeup, page_nr;
+       int ret = 0, page_nr = 0;
 
        if (!spd_pages)
                return 0;
 
-       ret = 0;
-       do_wakeup = 0;
-       page_nr = 0;
-
-       pipe_lock(pipe);
-
-       for (;;) {
-               if (!pipe->readers) {
-                       send_sig(SIGPIPE, current, 0);
-                       if (!ret)
-                               ret = -EPIPE;
-                       break;
-               }
-
-               if (pipe->nrbufs < pipe->buffers) {
-                       int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
-                       struct pipe_buffer *buf = pipe->bufs + newbuf;
-
-                       buf->page = spd->pages[page_nr];
-                       buf->offset = spd->partial[page_nr].offset;
-                       buf->len = spd->partial[page_nr].len;
-                       buf->private = spd->partial[page_nr].private;
-                       buf->ops = spd->ops;
-                       if (spd->flags & SPLICE_F_GIFT)
-                               buf->flags |= PIPE_BUF_FLAG_GIFT;
-
-                       pipe->nrbufs++;
-                       page_nr++;
-                       ret += buf->len;
-
-                       if (pipe->files)
-                               do_wakeup = 1;
+       if (unlikely(!pipe->readers)) {
+               send_sig(SIGPIPE, current, 0);
+               ret = -EPIPE;
+               goto out;
+       }
 
-                       if (!--spd->nr_pages)
-                               break;
-                       if (pipe->nrbufs < pipe->buffers)
-                               continue;
+       while (pipe->nrbufs < pipe->buffers) {
+               int newbuf = (pipe->curbuf + pipe->nrbufs) & (pipe->buffers - 1);
+               struct pipe_buffer *buf = pipe->bufs + newbuf;
 
-                       break;
-               }
+               buf->page = spd->pages[page_nr];
+               buf->offset = spd->partial[page_nr].offset;
+               buf->len = spd->partial[page_nr].len;
+               buf->private = spd->partial[page_nr].private;
+               buf->ops = spd->ops;
+               if (spd->flags & SPLICE_F_GIFT)
+                       buf->flags |= PIPE_BUF_FLAG_GIFT;
 
-               if (spd->flags & SPLICE_F_NONBLOCK) {
-                       if (!ret)
-                               ret = -EAGAIN;
-                       break;
-               }
+               pipe->nrbufs++;
+               page_nr++;
+               ret += buf->len;
 
-               if (signal_pending(current)) {
-                       if (!ret)
-                               ret = -ERESTARTSYS;
+               if (!--spd->nr_pages)
                        break;
-               }
-
-               if (do_wakeup) {
-                       wakeup_pipe_readers(pipe);
-                       do_wakeup = 0;
-               }
-
-               pipe->waiting_writers++;
-               pipe_wait(pipe);
-               pipe->waiting_writers--;
        }
 
-       pipe_unlock(pipe);
-
-       if (do_wakeup)
-               wakeup_pipe_readers(pipe);
+       if (!ret)
+               ret = -EAGAIN;
 
+out:
        while (page_nr < spd_pages)
                spd->spd_release(spd, page_nr++);
 
@@ -1339,6 +1301,20 @@ long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
 }
 EXPORT_SYMBOL(do_splice_direct);
 
+static int wait_for_space(struct pipe_inode_info *pipe, unsigned flags)
+{
+       while (pipe->nrbufs == pipe->buffers) {
+               if (flags & SPLICE_F_NONBLOCK)
+                       return -EAGAIN;
+               if (signal_pending(current))
+                       return -ERESTARTSYS;
+               pipe->waiting_writers++;
+               pipe_wait(pipe);
+               pipe->waiting_writers--;
+       }
+       return 0;
+}
+
 static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
                               struct pipe_inode_info *opipe,
                               size_t len, unsigned int flags);
@@ -1421,8 +1397,13 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                        offset = in->f_pos;
                }
 
-               ret = do_splice_to(in, &offset, opipe, len, flags);
-
+               pipe_lock(opipe);
+               ret = wait_for_space(opipe, flags);
+               if (!ret)
+                       ret = do_splice_to(in, &offset, opipe, len, flags);
+               pipe_unlock(opipe);
+               if (ret > 0)
+                       wakeup_pipe_readers(opipe);
                if (!off_in)
                        in->f_pos = offset;
                else if (copy_to_user(off_in, &offset, sizeof(loff_t)))
@@ -1546,14 +1527,20 @@ static long vmsplice_to_pipe(struct file *file, const struct iovec __user *uiov,
                return -ENOMEM;
        }
 
-       spd.nr_pages = get_iovec_page_array(&from, spd.pages,
-                                           spd.partial,
-                                           spd.nr_pages_max);
-       if (spd.nr_pages <= 0)
-               ret = spd.nr_pages;
-       else
-               ret = splice_to_pipe(pipe, &spd);
-
+       pipe_lock(pipe);
+       ret = wait_for_space(pipe, flags);
+       if (!ret) {
+               spd.nr_pages = get_iovec_page_array(&from, spd.pages,
+                                                   spd.partial,
+                                                   spd.nr_pages_max);
+               if (spd.nr_pages <= 0)
+                       ret = spd.nr_pages;
+               else
+                       ret = splice_to_pipe(pipe, &spd);
+       }
+       pipe_unlock(pipe);
+       if (ret > 0)
+               wakeup_pipe_readers(pipe);
        splice_shrink_spd(&spd);
        kfree(iov);
        return ret;