Atheros Kconfig needs to be dependent on WLAN_80211
[linux-2.6.git] / fs / splice.c
index ed2ce99..73766d2 100644 (file)
@@ -21,6 +21,7 @@
 #include <linux/file.h>
 #include <linux/pagemap.h>
 #include <linux/splice.h>
+#include <linux/memcontrol.h>
 #include <linux/mm_inline.h>
 #include <linux/swap.h>
 #include <linux/writeback.h>
@@ -28,6 +29,7 @@
 #include <linux/module.h>
 #include <linux/syscalls.h>
 #include <linux/uio.h>
+#include <linux/security.h>
 
 /*
  * Attempt to steal a page from a pipe buffer. This should perhaps go into
@@ -57,8 +59,9 @@ static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
                 */
                wait_on_page_writeback(page);
 
-               if (PagePrivate(page))
-                       try_to_release_page(page, GFP_KERNEL);
+               if (page_has_private(page) &&
+                   !try_to_release_page(page, GFP_KERNEL))
+                       goto out_unlock;
 
                /*
                 * If we succeeded in removing the mapping, set LRU flag
@@ -74,6 +77,7 @@ static int page_cache_pipe_buf_steal(struct pipe_inode_info *pipe,
         * Raced with truncate or failed to remove page from current
         * address space, unlock and return failure.
         */
+out_unlock:
        unlock_page(page);
        return 1;
 }
@@ -163,7 +167,7 @@ static const struct pipe_buf_operations user_page_pipe_buf_ops = {
  * @spd:       data to fill
  *
  * Description:
- *    @spd contains a map of pages and len/offset tupples, a long with
+ *    @spd contains a map of pages and len/offset tuples, along with
  *    the struct pipe_buf_operations associated with these pages. This
  *    function will link that data to the pipe.
  *
@@ -178,8 +182,7 @@ ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
        do_wakeup = 0;
        page_nr = 0;
 
-       if (pipe->inode)
-               mutex_lock(&pipe->inode->i_mutex);
+       pipe_lock(pipe);
 
        for (;;) {
                if (!pipe->readers) {
@@ -241,30 +244,33 @@ ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
                pipe->waiting_writers--;
        }
 
-       if (pipe->inode) {
-               mutex_unlock(&pipe->inode->i_mutex);
+       pipe_unlock(pipe);
 
-               if (do_wakeup) {
-                       smp_mb();
-                       if (waitqueue_active(&pipe->wait))
-                               wake_up_interruptible(&pipe->wait);
-                       kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
-               }
+       if (do_wakeup) {
+               smp_mb();
+               if (waitqueue_active(&pipe->wait))
+                       wake_up_interruptible(&pipe->wait);
+               kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
        }
 
        while (page_nr < spd_pages)
-               page_cache_release(spd->pages[page_nr++]);
+               spd->spd_release(spd, page_nr++);
 
        return ret;
 }
 
+static void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
+{
+       page_cache_release(spd->pages[i]);
+}
+
 static int
 __generic_file_splice_read(struct file *in, loff_t *ppos,
                           struct pipe_inode_info *pipe, size_t len,
                           unsigned int flags)
 {
        struct address_space *mapping = in->f_mapping;
-       unsigned int loff, nr_pages;
+       unsigned int loff, nr_pages, req_pages;
        struct page *pages[PIPE_BUFFERS];
        struct partial_page partial[PIPE_BUFFERS];
        struct page *page;
@@ -276,32 +282,29 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                .partial = partial,
                .flags = flags,
                .ops = &page_cache_pipe_buf_ops,
+               .spd_release = spd_release_page,
        };
 
        index = *ppos >> PAGE_CACHE_SHIFT;
        loff = *ppos & ~PAGE_CACHE_MASK;
-       nr_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
-
-       if (nr_pages > PIPE_BUFFERS)
-               nr_pages = PIPE_BUFFERS;
-
-       /*
-        * Don't try to 2nd guess the read-ahead logic, call into
-        * page_cache_readahead() like the page cache reads would do.
-        */
-       page_cache_readahead(mapping, &in->f_ra, in, index, nr_pages);
+       req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+       nr_pages = min(req_pages, (unsigned)PIPE_BUFFERS);
 
        /*
         * Lookup the (hopefully) full range of pages we need.
         */
        spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
+       index += spd.nr_pages;
 
        /*
         * If find_get_pages_contig() returned fewer pages than we needed,
-        * allocate the rest and fill in the holes.
+        * readahead/allocate the rest and fill in the holes.
         */
+       if (spd.nr_pages < nr_pages)
+               page_cache_sync_readahead(mapping, &in->f_ra, in,
+                               index, req_pages - spd.nr_pages);
+
        error = 0;
-       index += spd.nr_pages;
        while (spd.nr_pages < nr_pages) {
                /*
                 * Page could be there, find_get_pages_contig() breaks on
@@ -310,12 +313,6 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                page = find_get_page(mapping, index);
                if (!page) {
                        /*
-                        * Make sure the read-ahead engine is notified
-                        * about this failure.
-                        */
-                       handle_ra_miss(mapping, &in->f_ra, index);
-
-                       /*
                         * page didn't exist, allocate one.
                         */
                        page = page_cache_alloc_cold(mapping);
@@ -323,7 +320,7 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                                break;
 
                        error = add_to_page_cache_lru(page, mapping, index,
-                                             GFP_KERNEL);
+                                               mapping_gfp_mask(mapping));
                        if (unlikely(error)) {
                                page_cache_release(page);
                                if (error == -EEXIST)
@@ -360,6 +357,10 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
                page = pages[page_nr];
 
+               if (PageReadahead(page))
+                       page_cache_async_readahead(mapping, &in->f_ra, in,
+                                       page, index, req_pages - page_nr);
+
                /*
                 * If the page isn't uptodate, we may need to start io on it
                 */
@@ -369,19 +370,30 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                         * for an in-flight io page
                         */
                        if (flags & SPLICE_F_NONBLOCK) {
-                               if (TestSetPageLocked(page))
+                               if (!trylock_page(page)) {
+                                       error = -EAGAIN;
                                        break;
+                               }
                        } else
                                lock_page(page);
 
                        /*
-                        * page was truncated, stop here. if this isn't the
-                        * first page, we'll just complete what we already
-                        * added
+                        * Page was truncated, or invalidated by the
+                        * filesystem.  Redo the find/create, but this time the
+                        * page is kept locked, so there's no chance of another
+                        * race with truncate/invalidate.
                         */
                        if (!page->mapping) {
                                unlock_page(page);
-                               break;
+                               page = find_or_create_page(mapping, index,
+                                               mapping_gfp_mask(mapping));
+
+                               if (!page) {
+                                       error = -ENOMEM;
+                                       break;
+                               }
+                               page_cache_release(pages[page_nr]);
+                               pages[page_nr] = page;
                        }
                        /*
                         * page was already under io and is now done, great
@@ -452,6 +464,7 @@ fill_it:
         */
        while (page_nr < nr_pages)
                page_cache_release(pages[page_nr++]);
+       in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 
        if (spd.nr_pages)
                return splice_to_pipe(pipe, &spd);
@@ -477,9 +490,8 @@ ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
                                 struct pipe_inode_info *pipe, size_t len,
                                 unsigned int flags)
 {
-       ssize_t spliced;
-       int ret;
        loff_t isize, left;
+       int ret;
 
        isize = i_size_read(in->f_mapping->host);
        if (unlikely(*ppos >= isize))
@@ -489,34 +501,136 @@ ssize_t generic_file_splice_read(struct file *in, loff_t *ppos,
        if (unlikely(left < len))
                len = left;
 
-       ret = 0;
-       spliced = 0;
-       while (len) {
-               ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
+       ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
+       if (ret > 0)
+               *ppos += ret;
 
-               if (ret < 0)
-                       break;
-               else if (!ret) {
-                       if (spliced)
-                               break;
-                       if (flags & SPLICE_F_NONBLOCK) {
-                               ret = -EAGAIN;
-                               break;
-                       }
-               }
+       return ret;
+}
+EXPORT_SYMBOL(generic_file_splice_read);
 
-               *ppos += ret;
-               len -= ret;
-               spliced += ret;
-       }
+static const struct pipe_buf_operations default_pipe_buf_ops = {
+       .can_merge = 0,
+       .map = generic_pipe_buf_map,
+       .unmap = generic_pipe_buf_unmap,
+       .confirm = generic_pipe_buf_confirm,
+       .release = generic_pipe_buf_release,
+       .steal = generic_pipe_buf_steal,
+       .get = generic_pipe_buf_get,
+};
 
-       if (spliced)
-               return spliced;
+static ssize_t kernel_readv(struct file *file, const struct iovec *vec,
+                           unsigned long vlen, loff_t offset)
+{
+       mm_segment_t old_fs;
+       loff_t pos = offset;
+       ssize_t res;
 
-       return ret;
+       old_fs = get_fs();
+       set_fs(get_ds());
+       /* The cast to a user pointer is valid due to the set_fs() */
+       res = vfs_readv(file, (const struct iovec __user *)vec, vlen, &pos);
+       set_fs(old_fs);
+
+       return res;
 }
 
-EXPORT_SYMBOL(generic_file_splice_read);
+static ssize_t kernel_write(struct file *file, const char *buf, size_t count,
+                           loff_t pos)
+{
+       mm_segment_t old_fs;
+       ssize_t res;
+
+       old_fs = get_fs();
+       set_fs(get_ds());
+       /* The cast to a user pointer is valid due to the set_fs() */
+       res = vfs_write(file, (const char __user *)buf, count, &pos);
+       set_fs(old_fs);
+
+       return res;
+}
+
+ssize_t default_file_splice_read(struct file *in, loff_t *ppos,
+                                struct pipe_inode_info *pipe, size_t len,
+                                unsigned int flags)
+{
+       unsigned int nr_pages;
+       unsigned int nr_freed;
+       size_t offset;
+       struct page *pages[PIPE_BUFFERS];
+       struct partial_page partial[PIPE_BUFFERS];
+       struct iovec vec[PIPE_BUFFERS];
+       pgoff_t index;
+       ssize_t res;
+       size_t this_len;
+       int error;
+       int i;
+       struct splice_pipe_desc spd = {
+               .pages = pages,
+               .partial = partial,
+               .flags = flags,
+               .ops = &default_pipe_buf_ops,
+               .spd_release = spd_release_page,
+       };
+
+       index = *ppos >> PAGE_CACHE_SHIFT;
+       offset = *ppos & ~PAGE_CACHE_MASK;
+       nr_pages = (len + offset + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+
+       for (i = 0; i < nr_pages && i < PIPE_BUFFERS && len; i++) {
+               struct page *page;
+
+               page = alloc_page(GFP_USER);
+               error = -ENOMEM;
+               if (!page)
+                       goto err;
+
+               this_len = min_t(size_t, len, PAGE_CACHE_SIZE - offset);
+               vec[i].iov_base = (void __user *) page_address(page);
+               vec[i].iov_len = this_len;
+               pages[i] = page;
+               spd.nr_pages++;
+               len -= this_len;
+               offset = 0;
+       }
+
+       res = kernel_readv(in, vec, spd.nr_pages, *ppos);
+       if (res < 0) {
+               error = res;
+               goto err;
+       }
+
+       error = 0;
+       if (!res)
+               goto err;
+
+       nr_freed = 0;
+       for (i = 0; i < spd.nr_pages; i++) {
+               this_len = min_t(size_t, vec[i].iov_len, res);
+               partial[i].offset = 0;
+               partial[i].len = this_len;
+               if (!this_len) {
+                       __free_page(pages[i]);
+                       pages[i] = NULL;
+                       nr_freed++;
+               }
+               res -= this_len;
+       }
+       spd.nr_pages -= nr_freed;
+
+       res = splice_to_pipe(pipe, &spd);
+       if (res > 0)
+               *ppos += res;
+
+       return res;
+
+err:
+       for (i = 0; i < spd.nr_pages; i++)
+               __free_page(pages[i]);
+
+       return error;
+}
+EXPORT_SYMBOL(default_file_splice_read);
 
 /*
  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
@@ -560,14 +674,14 @@ static int pipe_to_sendpage(struct pipe_inode_info *pipe,
  * SPLICE_F_MOVE isn't set, or we cannot move the page, we simply create
  * a new page in the output file page cache and fill/dirty that.
  */
-static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
-                       struct splice_desc *sd)
+int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
+                struct splice_desc *sd)
 {
        struct file *file = sd->u.file;
        struct address_space *mapping = file->f_mapping;
        unsigned int offset, this_len;
        struct page *page;
-       pgoff_t index;
+       void *fsdata;
        int ret;
 
        /*
@@ -577,49 +691,16 @@ static int pipe_to_file(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
        if (unlikely(ret))
                return ret;
 
-       index = sd->pos >> PAGE_CACHE_SHIFT;
        offset = sd->pos & ~PAGE_CACHE_MASK;
 
        this_len = sd->len;
        if (this_len + offset > PAGE_CACHE_SIZE)
                this_len = PAGE_CACHE_SIZE - offset;
 
-find_page:
-       page = find_lock_page(mapping, index);
-       if (!page) {
-               ret = -ENOMEM;
-               page = page_cache_alloc_cold(mapping);
-               if (unlikely(!page))
-                       goto out_ret;
-
-               /*
-                * This will also lock the page
-                */
-               ret = add_to_page_cache_lru(page, mapping, index,
-                                           GFP_KERNEL);
-               if (unlikely(ret))
-                       goto out;
-       }
-
-       ret = mapping->a_ops->prepare_write(file, page, offset, offset+this_len);
-       if (unlikely(ret)) {
-               loff_t isize = i_size_read(mapping->host);
-
-               if (ret != AOP_TRUNCATED_PAGE)
-                       unlock_page(page);
-               page_cache_release(page);
-               if (ret == AOP_TRUNCATED_PAGE)
-                       goto find_page;
-
-               /*
-                * prepare_write() may have instantiated a few blocks
-                * outside i_size.  Trim these off again.
-                */
-               if (sd->pos + this_len > isize)
-                       vmtruncate(mapping->host, isize);
-
-               goto out_ret;
-       }
+       ret = pagecache_write_begin(file, mapping, sd->pos, this_len,
+                               AOP_FLAG_UNINTERRUPTIBLE, &page, &fsdata);
+       if (unlikely(ret))
+               goto out;
 
        if (buf->page != page) {
                /*
@@ -633,134 +714,182 @@ find_page:
                kunmap_atomic(dst, KM_USER1);
                buf->ops->unmap(pipe, buf, src);
        }
-
-       ret = mapping->a_ops->commit_write(file, page, offset, offset+this_len);
-       if (ret) {
-               if (ret == AOP_TRUNCATED_PAGE) {
-                       page_cache_release(page);
-                       goto find_page;
-               }
-               if (ret < 0)
-                       goto out;
-               /*
-                * Partial write has happened, so 'ret' already initialized by
-                * number of bytes written, Where is nothing we have to do here.
-                */
-       } else
-               ret = this_len;
-       /*
-        * Return the number of bytes written and mark page as
-        * accessed, we are now done!
-        */
-       mark_page_accessed(page);
+       ret = pagecache_write_end(file, mapping, sd->pos, this_len, this_len,
+                               page, fsdata);
 out:
-       page_cache_release(page);
-       unlock_page(page);
-out_ret:
        return ret;
 }
+EXPORT_SYMBOL(pipe_to_file);
+
+static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
+{
+       smp_mb();
+       if (waitqueue_active(&pipe->wait))
+               wake_up_interruptible(&pipe->wait);
+       kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
+}
 
 /**
- * __splice_from_pipe - splice data from a pipe to given actor
+ * splice_from_pipe_feed - feed available data from a pipe to a file
  * @pipe:      pipe to splice from
  * @sd:                information to @actor
  * @actor:     handler that splices the data
  *
  * Description:
- *    This function does little more than loop over the pipe and call
- *    @actor to do the actual moving of a single struct pipe_buffer to
- *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
- *    pipe_to_user.
+ *    This function loops over the pipe and calls @actor to do the
+ *    actual moving of a single struct pipe_buffer to the desired
+ *    destination.  It returns when there's no more buffers left in
+ *    the pipe or if the requested number of bytes (@sd->total_len)
+ *    have been copied.  It returns a positive number (one) if the
+ *    pipe needs to be filled with more data, zero if the required
+ *    number of bytes have been copied and -errno on error.
  *
+ *    This, together with splice_from_pipe_{begin,end,next}, may be
+ *    used to implement the functionality of __splice_from_pipe() when
+ *    locking is required around copying the pipe buffers to the
+ *    destination.
  */
-ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
-                          splice_actor *actor)
+int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
+                         splice_actor *actor)
 {
-       int ret, do_wakeup, err;
-
-       ret = 0;
-       do_wakeup = 0;
-
-       for (;;) {
-               if (pipe->nrbufs) {
-                       struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
-                       const struct pipe_buf_operations *ops = buf->ops;
+       int ret;
 
-                       sd->len = buf->len;
-                       if (sd->len > sd->total_len)
-                               sd->len = sd->total_len;
+       while (pipe->nrbufs) {
+               struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
+               const struct pipe_buf_operations *ops = buf->ops;
 
-                       err = actor(pipe, buf, sd);
-                       if (err <= 0) {
-                               if (!ret && err != -ENODATA)
-                                       ret = err;
+               sd->len = buf->len;
+               if (sd->len > sd->total_len)
+                       sd->len = sd->total_len;
 
-                               break;
-                       }
+               ret = actor(pipe, buf, sd);
+               if (ret <= 0) {
+                       if (ret == -ENODATA)
+                               ret = 0;
+                       return ret;
+               }
+               buf->offset += ret;
+               buf->len -= ret;
 
-                       ret += err;
-                       buf->offset += err;
-                       buf->len -= err;
+               sd->num_spliced += ret;
+               sd->len -= ret;
+               sd->pos += ret;
+               sd->total_len -= ret;
 
-                       sd->len -= err;
-                       sd->pos += err;
-                       sd->total_len -= err;
-                       if (sd->len)
-                               continue;
+               if (!buf->len) {
+                       buf->ops = NULL;
+                       ops->release(pipe, buf);
+                       pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
+                       pipe->nrbufs--;
+                       if (pipe->inode)
+                               sd->need_wakeup = true;
+               }
 
-                       if (!buf->len) {
-                               buf->ops = NULL;
-                               ops->release(pipe, buf);
-                               pipe->curbuf = (pipe->curbuf + 1) & (PIPE_BUFFERS - 1);
-                               pipe->nrbufs--;
-                               if (pipe->inode)
-                                       do_wakeup = 1;
-                       }
+               if (!sd->total_len)
+                       return 0;
+       }
 
-                       if (!sd->total_len)
-                               break;
-               }
+       return 1;
+}
+EXPORT_SYMBOL(splice_from_pipe_feed);
 
-               if (pipe->nrbufs)
-                       continue;
+/**
+ * splice_from_pipe_next - wait for some data to splice from
+ * @pipe:      pipe to splice from
+ * @sd:                information about the splice operation
+ *
+ * Description:
+ *    This function will wait for some data and return a positive
+ *    value (one) if pipe buffers are available.  It will return zero
+ *    or -errno if no more data needs to be spliced.
+ */
+int splice_from_pipe_next(struct pipe_inode_info *pipe, struct splice_desc *sd)
+{
+       while (!pipe->nrbufs) {
                if (!pipe->writers)
-                       break;
-               if (!pipe->waiting_writers) {
-                       if (ret)
-                               break;
-               }
+                       return 0;
 
-               if (sd->flags & SPLICE_F_NONBLOCK) {
-                       if (!ret)
-                               ret = -EAGAIN;
-                       break;
-               }
+               if (!pipe->waiting_writers && sd->num_spliced)
+                       return 0;
 
-               if (signal_pending(current)) {
-                       if (!ret)
-                               ret = -ERESTARTSYS;
-                       break;
-               }
+               if (sd->flags & SPLICE_F_NONBLOCK)
+                       return -EAGAIN;
 
-               if (do_wakeup) {
-                       smp_mb();
-                       if (waitqueue_active(&pipe->wait))
-                               wake_up_interruptible_sync(&pipe->wait);
-                       kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
-                       do_wakeup = 0;
+               if (signal_pending(current))
+                       return -ERESTARTSYS;
+
+               if (sd->need_wakeup) {
+                       wakeup_pipe_writers(pipe);
+                       sd->need_wakeup = false;
                }
 
                pipe_wait(pipe);
        }
 
-       if (do_wakeup) {
-               smp_mb();
-               if (waitqueue_active(&pipe->wait))
-                       wake_up_interruptible(&pipe->wait);
-               kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
-       }
+       return 1;
+}
+EXPORT_SYMBOL(splice_from_pipe_next);
 
-       return ret;
+/**
+ * splice_from_pipe_begin - start splicing from pipe
+ * @sd:                information about the splice operation
+ *
+ * Description:
+ *    This function should be called before a loop containing
+ *    splice_from_pipe_next() and splice_from_pipe_feed() to
+ *    initialize the necessary fields of @sd.
+ */
+void splice_from_pipe_begin(struct splice_desc *sd)
+{
+       sd->num_spliced = 0;
+       sd->need_wakeup = false;
+}
+EXPORT_SYMBOL(splice_from_pipe_begin);
+
+/**
+ * splice_from_pipe_end - finish splicing from pipe
+ * @pipe:      pipe to splice from
+ * @sd:                information about the splice operation
+ *
+ * Description:
+ *    This function will wake up pipe writers if necessary.  It should
+ *    be called after a loop containing splice_from_pipe_next() and
+ *    splice_from_pipe_feed().
+ */
+void splice_from_pipe_end(struct pipe_inode_info *pipe, struct splice_desc *sd)
+{
+       if (sd->need_wakeup)
+               wakeup_pipe_writers(pipe);
+}
+EXPORT_SYMBOL(splice_from_pipe_end);
+
+/**
+ * __splice_from_pipe - splice data from a pipe to given actor
+ * @pipe:      pipe to splice from
+ * @sd:                information to @actor
+ * @actor:     handler that splices the data
+ *
+ * Description:
+ *    This function does little more than loop over the pipe and call
+ *    @actor to do the actual moving of a single struct pipe_buffer to
+ *    the desired destination. See pipe_to_file, pipe_to_sendpage, or
+ *    pipe_to_user.
+ *
+ */
+ssize_t __splice_from_pipe(struct pipe_inode_info *pipe, struct splice_desc *sd,
+                          splice_actor *actor)
+{
+       int ret;
+
+       splice_from_pipe_begin(sd);
+       do {
+               ret = splice_from_pipe_next(pipe, sd);
+               if (ret > 0)
+                       ret = splice_from_pipe_feed(pipe, sd, actor);
+       } while (ret > 0);
+       splice_from_pipe_end(pipe, sd);
+
+       return sd->num_spliced ? sd->num_spliced : ret;
 }
 EXPORT_SYMBOL(__splice_from_pipe);
 
@@ -774,7 +903,7 @@ EXPORT_SYMBOL(__splice_from_pipe);
  * @actor:     handler that splices the data
  *
  * Description:
- *    See __splice_from_pipe. This function locks the input and output inodes,
+ *    See __splice_from_pipe. This function locks the pipe inode,
  *    otherwise it's identical to __splice_from_pipe().
  *
  */
@@ -783,7 +912,6 @@ ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
                         splice_actor *actor)
 {
        ssize_t ret;
-       struct inode *inode = out->f_mapping->host;
        struct splice_desc sd = {
                .total_len = len,
                .flags = flags,
@@ -791,21 +919,15 @@ ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
                .u.file = out,
        };
 
-       /*
-        * The actor worker might be calling ->prepare_write and
-        * ->commit_write. Most of the time, these expect i_mutex to
-        * be held. Since this may result in an ABBA deadlock with
-        * pipe->inode, we have to order lock acquiry here.
-        */
-       inode_double_lock(inode, pipe->inode);
+       pipe_lock(pipe);
        ret = __splice_from_pipe(pipe, &sd, actor);
-       inode_double_unlock(inode, pipe->inode);
+       pipe_unlock(pipe);
 
        return ret;
 }
 
 /**
- * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
+ * generic_file_splice_write - splice data from a pipe to a file
  * @pipe:      pipe info
  * @out:       file to write to
  * @ppos:      position in @out
@@ -814,13 +936,12 @@ ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
  *
  * Description:
  *    Will either move or copy pages (determined by @flags options) from
- *    the given pipe inode to the given file. The caller is responsible
- *    for acquiring i_mutex on both inodes.
+ *    the given pipe inode to the given file.
  *
  */
 ssize_t
-generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
-                                loff_t *ppos, size_t len, unsigned int flags)
+generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
+                         loff_t *ppos, size_t len, unsigned int flags)
 {
        struct address_space *mapping = out->f_mapping;
        struct inode *inode = mapping->host;
@@ -831,13 +952,28 @@ generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
                .u.file = out,
        };
        ssize_t ret;
-       int err;
 
-       err = remove_suid(out->f_path.dentry);
-       if (unlikely(err))
-               return err;
+       pipe_lock(pipe);
+
+       splice_from_pipe_begin(&sd);
+       do {
+               ret = splice_from_pipe_next(pipe, &sd);
+               if (ret <= 0)
+                       break;
+
+               mutex_lock_nested(&inode->i_mutex, I_MUTEX_CHILD);
+               ret = file_remove_suid(out);
+               if (!ret)
+                       ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
+               mutex_unlock(&inode->i_mutex);
+       } while (ret > 0);
+       splice_from_pipe_end(pipe, &sd);
+
+       pipe_unlock(pipe);
+
+       if (sd.num_spliced)
+               ret = sd.num_spliced;
 
-       ret = __splice_from_pipe(pipe, &sd, pipe_to_file);
        if (ret > 0) {
                unsigned long nr_pages;
 
@@ -849,8 +985,12 @@ generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
                 * sync it.
                 */
                if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
+                       int err;
+
+                       mutex_lock(&inode->i_mutex);
                        err = generic_osync_inode(inode, mapping,
                                                  OSYNC_METADATA|OSYNC_DATA);
+                       mutex_unlock(&inode->i_mutex);
 
                        if (err)
                                ret = err;
@@ -861,67 +1001,38 @@ generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
        return ret;
 }
 
-EXPORT_SYMBOL(generic_file_splice_write_nolock);
+EXPORT_SYMBOL(generic_file_splice_write);
 
-/**
- * generic_file_splice_write - splice data from a pipe to a file
- * @pipe:      pipe info
- * @out:       file to write to
- * @ppos:      position in @out
- * @len:       number of bytes to splice
- * @flags:     splice modifier flags
- *
- * Description:
- *    Will either move or copy pages (determined by @flags options) from
- *    the given pipe inode to the given file.
- *
- */
-ssize_t
-generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
-                         loff_t *ppos, size_t len, unsigned int flags)
+static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
+                         struct splice_desc *sd)
 {
-       struct address_space *mapping = out->f_mapping;
-       struct inode *inode = mapping->host;
-       ssize_t ret;
-       int err;
+       int ret;
+       void *data;
 
-       err = should_remove_suid(out->f_path.dentry);
-       if (unlikely(err)) {
-               mutex_lock(&inode->i_mutex);
-               err = __remove_suid(out->f_path.dentry, err);
-               mutex_unlock(&inode->i_mutex);
-               if (err)
-                       return err;
-       }
+       ret = buf->ops->confirm(pipe, buf);
+       if (ret)
+               return ret;
 
-       ret = splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
-       if (ret > 0) {
-               unsigned long nr_pages;
+       data = buf->ops->map(pipe, buf, 0);
+       ret = kernel_write(sd->u.file, data + buf->offset, sd->len, sd->pos);
+       buf->ops->unmap(pipe, buf, data);
 
-               *ppos += ret;
-               nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+       return ret;
+}
 
-               /*
-                * If file or inode is SYNC and we actually wrote some data,
-                * sync it.
-                */
-               if (unlikely((out->f_flags & O_SYNC) || IS_SYNC(inode))) {
-                       mutex_lock(&inode->i_mutex);
-                       err = generic_osync_inode(inode, mapping,
-                                                 OSYNC_METADATA|OSYNC_DATA);
-                       mutex_unlock(&inode->i_mutex);
+static ssize_t default_file_splice_write(struct pipe_inode_info *pipe,
+                                        struct file *out, loff_t *ppos,
+                                        size_t len, unsigned int flags)
+{
+       ssize_t ret;
 
-                       if (err)
-                               ret = err;
-               }
-               balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
-       }
+       ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
+       if (ret > 0)
+               *ppos += ret;
 
        return ret;
 }
 
-EXPORT_SYMBOL(generic_file_splice_write);
-
 /**
  * generic_splice_sendpage - splice data from a pipe to a socket
  * @pipe:      pipe to splice from
@@ -949,19 +1060,25 @@ EXPORT_SYMBOL(generic_splice_sendpage);
 static long do_splice_from(struct pipe_inode_info *pipe, struct file *out,
                           loff_t *ppos, size_t len, unsigned int flags)
 {
+       ssize_t (*splice_write)(struct pipe_inode_info *, struct file *,
+                               loff_t *, size_t, unsigned int);
        int ret;
 
-       if (unlikely(!out->f_op || !out->f_op->splice_write))
-               return -EINVAL;
-
        if (unlikely(!(out->f_mode & FMODE_WRITE)))
                return -EBADF;
 
+       if (unlikely(out->f_flags & O_APPEND))
+               return -EINVAL;
+
        ret = rw_verify_area(WRITE, out, ppos, len);
        if (unlikely(ret < 0))
                return ret;
 
-       return out->f_op->splice_write(pipe, out, ppos, len, flags);
+       splice_write = out->f_op->splice_write;
+       if (!splice_write)
+               splice_write = default_file_splice_write;
+
+       return splice_write(pipe, out, ppos, len, flags);
 }
 
 /*
@@ -971,11 +1088,10 @@ static long do_splice_to(struct file *in, loff_t *ppos,
                         struct pipe_inode_info *pipe, size_t len,
                         unsigned int flags)
 {
+       ssize_t (*splice_read)(struct file *, loff_t *,
+                              struct pipe_inode_info *, size_t, unsigned int);
        int ret;
 
-       if (unlikely(!in->f_op || !in->f_op->splice_read))
-               return -EINVAL;
-
        if (unlikely(!(in->f_mode & FMODE_READ)))
                return -EBADF;
 
@@ -983,7 +1099,11 @@ static long do_splice_to(struct file *in, loff_t *ppos,
        if (unlikely(ret < 0))
                return ret;
 
-       return in->f_op->splice_read(in, ppos, pipe, len, flags);
+       splice_read = in->f_op->splice_read;
+       if (!splice_read)
+               splice_read = default_file_splice_read;
+
+       return splice_read(in, ppos, pipe, len, flags);
 }
 
 /**
@@ -995,7 +1115,7 @@ static long do_splice_to(struct file *in, loff_t *ppos,
  * Description:
  *    This is a special case helper to splice directly between two
  *    points, without requiring an explicit pipe. Internally an allocated
- *    pipe is cached in the process, and reused during the life time of
+ *    pipe is cached in the process, and reused during the lifetime of
  *    that process.
  *
  */
@@ -1051,15 +1171,11 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
        sd->flags &= ~SPLICE_F_NONBLOCK;
 
        while (len) {
-               size_t read_len, max_read_len;
-
-               /*
-                * Do at most PIPE_BUFFERS pages worth of transfer:
-                */
-               max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
+               size_t read_len;
+               loff_t pos = sd->pos, prev_pos = pos;
 
-               ret = do_splice_to(in, &sd->pos, pipe, max_read_len, flags);
-               if (unlikely(ret < 0))
+               ret = do_splice_to(in, &pos, pipe, len, flags);
+               if (unlikely(ret <= 0))
                        goto out_release;
 
                read_len = ret;
@@ -1071,26 +1187,24 @@ ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
                 * could get stuck data in the internal pipe:
                 */
                ret = actor(pipe, sd);
-               if (unlikely(ret < 0))
+               if (unlikely(ret <= 0)) {
+                       sd->pos = prev_pos;
                        goto out_release;
+               }
 
                bytes += ret;
                len -= ret;
+               sd->pos = pos;
 
-               /*
-                * In nonblocking mode, if we got back a short read then
-                * that was due to either an IO error or due to the
-                * pagecache entry not being there. In the IO error case
-                * the _next_ splice attempt will produce a clean IO error
-                * return value (not a short read), so in both cases it's
-                * correct to break out of the loop here:
-                */
-               if ((flags & SPLICE_F_NONBLOCK) && (read_len < max_read_len))
-                       break;
+               if (ret < read_len) {
+                       sd->pos = prev_pos + ret;
+                       goto out_release;
+               }
        }
 
+done:
        pipe->nrbufs = pipe->curbuf = 0;
-
+       file_accessed(in);
        return bytes;
 
 out_release:
@@ -1106,16 +1220,11 @@ out_release:
                        buf->ops = NULL;
                }
        }
-       pipe->nrbufs = pipe->curbuf = 0;
 
-       /*
-        * If we transferred some data, return the number of bytes:
-        */
-       if (bytes > 0)
-               return bytes;
-
-       return ret;
+       if (!bytes)
+               bytes = ret;
 
+       goto done;
 }
 EXPORT_SYMBOL(splice_direct_to_actor);
 
@@ -1152,13 +1261,18 @@ long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
                .pos            = *ppos,
                .u.file         = out,
        };
-       size_t ret;
+       long ret;
 
        ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
-       *ppos = sd.pos;
+       if (ret > 0)
+               *ppos = sd.pos;
+
        return ret;
 }
 
+static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
+                              struct pipe_inode_info *opipe,
+                              size_t len, unsigned int flags);
 /*
  * After the inode slimming patch, i_pipe/i_bdev/i_cdev share the same
  * location, so checking ->i_pipe is not enough to verify that this is a
@@ -1179,12 +1293,32 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                      struct file *out, loff_t __user *off_out,
                      size_t len, unsigned int flags)
 {
-       struct pipe_inode_info *pipe;
+       struct pipe_inode_info *ipipe;
+       struct pipe_inode_info *opipe;
        loff_t offset, *off;
        long ret;
 
-       pipe = pipe_info(in->f_path.dentry->d_inode);
-       if (pipe) {
+       ipipe = pipe_info(in->f_path.dentry->d_inode);
+       opipe = pipe_info(out->f_path.dentry->d_inode);
+
+       if (ipipe && opipe) {
+               if (off_in || off_out)
+                       return -ESPIPE;
+
+               if (!(in->f_mode & FMODE_READ))
+                       return -EBADF;
+
+               if (!(out->f_mode & FMODE_WRITE))
+                       return -EBADF;
+
+               /* Splicing to self would be fun, but... */
+               if (ipipe == opipe)
+                       return -EINVAL;
+
+               return splice_pipe_to_pipe(ipipe, opipe, len, flags);
+       }
+
+       if (ipipe) {
                if (off_in)
                        return -ESPIPE;
                if (off_out) {
@@ -1196,7 +1330,7 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                } else
                        off = &out->f_pos;
 
-               ret = do_splice_from(pipe, out, off, len, flags);
+               ret = do_splice_from(ipipe, out, off, len, flags);
 
                if (off_out && copy_to_user(off_out, off, sizeof(loff_t)))
                        ret = -EFAULT;
@@ -1204,8 +1338,7 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                return ret;
        }
 
-       pipe = pipe_info(out->f_path.dentry->d_inode);
-       if (pipe) {
+       if (opipe) {
                if (off_out)
                        return -ESPIPE;
                if (off_in) {
@@ -1217,7 +1350,7 @@ static long do_splice(struct file *in, loff_t __user *off_in,
                } else
                        off = &in->f_pos;
 
-               ret = do_splice_to(in, off, pipe, len, flags);
+               ret = do_splice_to(in, off, opipe, len, flags);
 
                if (off_in && copy_to_user(off_in, off, sizeof(loff_t)))
                        ret = -EFAULT;
@@ -1241,35 +1374,28 @@ static int get_iovec_page_array(const struct iovec __user *iov,
 {
        int buffers = 0, error = 0;
 
-       /*
-        * It's ok to take the mmap_sem for reading, even
-        * across a "get_user()".
-        */
-       down_read(&current->mm->mmap_sem);
-
        while (nr_vecs) {
                unsigned long off, npages;
+               struct iovec entry;
                void __user *base;
                size_t len;
                int i;
 
-               /*
-                * Get user address base and length for this iovec.
-                */
-               error = get_user(base, &iov->iov_base);
-               if (unlikely(error))
-                       break;
-               error = get_user(len, &iov->iov_len);
-               if (unlikely(error))
+               error = -EFAULT;
+               if (copy_from_user(&entry, iov, sizeof(entry)))
                        break;
 
+               base = entry.iov_base;
+               len = entry.iov_len;
+
                /*
                 * Sanity check this iovec. 0 read succeeds.
                 */
+               error = 0;
                if (unlikely(!len))
                        break;
                error = -EFAULT;
-               if (unlikely(!base))
+               if (!access_ok(VERIFY_READ, base, len))
                        break;
 
                /*
@@ -1290,9 +1416,8 @@ static int get_iovec_page_array(const struct iovec __user *iov,
                if (npages > PIPE_BUFFERS - buffers)
                        npages = PIPE_BUFFERS - buffers;
 
-               error = get_user_pages(current, current->mm,
-                                      (unsigned long) base, npages, 0, 0,
-                                      &pages[buffers], NULL);
+               error = get_user_pages_fast((unsigned long)base, npages,
+                                       0, &pages[buffers]);
 
                if (unlikely(error <= 0))
                        break;
@@ -1331,8 +1456,6 @@ static int get_iovec_page_array(const struct iovec __user *iov,
                iov++;
        }
 
-       up_read(&current->mm->mmap_sem);
-
        if (buffers)
                return buffers;
 
@@ -1373,10 +1496,10 @@ static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
        if (copy_to_user(sd->u.userptr, src + buf->offset, sd->len))
                ret = -EFAULT;
 
+       buf->ops->unmap(pipe, buf, src);
 out:
        if (ret > 0)
                sd->u.userptr += ret;
-       buf->ops->unmap(pipe, buf, src);
        return ret;
 }
 
@@ -1397,8 +1520,7 @@ static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
        if (!pipe)
                return -EBADF;
 
-       if (pipe->inode)
-               mutex_lock(&pipe->inode->i_mutex);
+       pipe_lock(pipe);
 
        error = ret = 0;
        while (nr_segs) {
@@ -1425,6 +1547,11 @@ static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
                        break;
                }
 
+               if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
+                       error = -EFAULT;
+                       break;
+               }
+
                sd.len = 0;
                sd.total_len = len;
                sd.flags = flags;
@@ -1448,8 +1575,7 @@ static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
                iov++;
        }
 
-       if (pipe->inode)
-               mutex_unlock(&pipe->inode->i_mutex);
+       pipe_unlock(pipe);
 
        if (!ret)
                ret = error;
@@ -1473,6 +1599,7 @@ static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
                .partial = partial,
                .flags = flags,
                .ops = &user_page_pipe_buf_ops,
+               .spd_release = spd_release_page,
        };
 
        pipe = pipe_info(file->f_path.dentry->d_inode);
@@ -1503,8 +1630,8 @@ static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
  * Currently we punt and implement it as a normal copy, see pipe_to_user().
  *
  */
-asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
-                            unsigned long nr_segs, unsigned int flags)
+SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
+               unsigned long, nr_segs, unsigned int, flags)
 {
        struct file *file;
        long error;
@@ -1529,9 +1656,9 @@ asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
        return error;
 }
 
-asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
-                          int fd_out, loff_t __user *off_out,
-                          size_t len, unsigned int flags)
+SYSCALL_DEFINE6(splice, int, fd_in, loff_t __user *, off_in,
+               int, fd_out, loff_t __user *, off_out,
+               size_t, len, unsigned int, flags)
 {
        long error;
        struct file *in, *out;
@@ -1564,7 +1691,7 @@ asmlinkage long sys_splice(int fd_in, loff_t __user *off_in,
  * Make sure there's data to read. Wait for input if we can, otherwise
  * return an appropriate error.
  */
-static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
+static int ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
 {
        int ret;
 
@@ -1576,7 +1703,7 @@ static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
                return 0;
 
        ret = 0;
-       mutex_lock(&pipe->inode->i_mutex);
+       pipe_lock(pipe);
 
        while (!pipe->nrbufs) {
                if (signal_pending(current)) {
@@ -1594,7 +1721,7 @@ static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
                pipe_wait(pipe);
        }
 
-       mutex_unlock(&pipe->inode->i_mutex);
+       pipe_unlock(pipe);
        return ret;
 }
 
@@ -1602,7 +1729,7 @@ static int link_ipipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
  * Make sure there's writeable room. Wait for room if we can, otherwise
  * return an appropriate error.
  */
-static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
+static int opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
 {
        int ret;
 
@@ -1614,7 +1741,7 @@ static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
                return 0;
 
        ret = 0;
-       mutex_lock(&pipe->inode->i_mutex);
+       pipe_lock(pipe);
 
        while (pipe->nrbufs >= PIPE_BUFFERS) {
                if (!pipe->readers) {
@@ -1635,7 +1762,125 @@ static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
                pipe->waiting_writers--;
        }
 
-       mutex_unlock(&pipe->inode->i_mutex);
+       pipe_unlock(pipe);
+       return ret;
+}
+
+/*
+ * Splice contents of ipipe to opipe.
+ */
+static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
+                              struct pipe_inode_info *opipe,
+                              size_t len, unsigned int flags)
+{
+       struct pipe_buffer *ibuf, *obuf;
+       int ret = 0, nbuf;
+       bool input_wakeup = false;
+
+
+retry:
+       ret = ipipe_prep(ipipe, flags);
+       if (ret)
+               return ret;
+
+       ret = opipe_prep(opipe, flags);
+       if (ret)
+               return ret;
+
+       /*
+        * Potential ABBA deadlock, work around it by ordering lock
+        * grabbing by pipe info address. Otherwise two different processes
+        * could deadlock (one doing tee from A -> B, the other from B -> A).
+        */
+       pipe_double_lock(ipipe, opipe);
+
+       do {
+               if (!opipe->readers) {
+                       send_sig(SIGPIPE, current, 0);
+                       if (!ret)
+                               ret = -EPIPE;
+                       break;
+               }
+
+               if (!ipipe->nrbufs && !ipipe->writers)
+                       break;
+
+               /*
+                * Cannot make any progress, because either the input
+                * pipe is empty or the output pipe is full.
+                */
+               if (!ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS) {
+                       /* Already processed some buffers, break */
+                       if (ret)
+                               break;
+
+                       if (flags & SPLICE_F_NONBLOCK) {
+                               ret = -EAGAIN;
+                               break;
+                       }
+
+                       /*
+                        * We raced with another reader/writer and haven't
+                        * managed to process any buffers.  A zero return
+                        * value means EOF, so retry instead.
+                        */
+                       pipe_unlock(ipipe);
+                       pipe_unlock(opipe);
+                       goto retry;
+               }
+
+               ibuf = ipipe->bufs + ipipe->curbuf;
+               nbuf = (opipe->curbuf + opipe->nrbufs) % PIPE_BUFFERS;
+               obuf = opipe->bufs + nbuf;
+
+               if (len >= ibuf->len) {
+                       /*
+                        * Simply move the whole buffer from ipipe to opipe
+                        */
+                       *obuf = *ibuf;
+                       ibuf->ops = NULL;
+                       opipe->nrbufs++;
+                       ipipe->curbuf = (ipipe->curbuf + 1) % PIPE_BUFFERS;
+                       ipipe->nrbufs--;
+                       input_wakeup = true;
+               } else {
+                       /*
+                        * Get a reference to this pipe buffer,
+                        * so we can copy the contents over.
+                        */
+                       ibuf->ops->get(ipipe, ibuf);
+                       *obuf = *ibuf;
+
+                       /*
+                        * Don't inherit the gift flag, we need to
+                        * prevent multiple steals of this page.
+                        */
+                       obuf->flags &= ~PIPE_BUF_FLAG_GIFT;
+
+                       obuf->len = len;
+                       opipe->nrbufs++;
+                       ibuf->offset += obuf->len;
+                       ibuf->len -= obuf->len;
+               }
+               ret += obuf->len;
+               len -= obuf->len;
+       } while (len);
+
+       pipe_unlock(ipipe);
+       pipe_unlock(opipe);
+
+       /*
+        * If we put data in the output pipe, wakeup any potential readers.
+        */
+       if (ret > 0) {
+               smp_mb();
+               if (waitqueue_active(&opipe->wait))
+                       wake_up_interruptible(&opipe->wait);
+               kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
+       }
+       if (input_wakeup)
+               wakeup_pipe_writers(ipipe);
+
        return ret;
 }
 
@@ -1651,10 +1896,10 @@ static int link_pipe(struct pipe_inode_info *ipipe,
 
        /*
         * Potential ABBA deadlock, work around it by ordering lock
-        * grabbing by inode address. Otherwise two different processes
+        * grabbing by pipe info address. Otherwise two different processes
         * could deadlock (one doing tee from A -> B, the other from B -> A).
         */
-       inode_double_lock(ipipe->inode, opipe->inode);
+       pipe_double_lock(ipipe, opipe);
 
        do {
                if (!opipe->readers) {
@@ -1698,7 +1943,15 @@ static int link_pipe(struct pipe_inode_info *ipipe,
                i++;
        } while (len);
 
-       inode_double_unlock(ipipe->inode, opipe->inode);
+       /*
+        * return EAGAIN if we have the potential of some data in the
+        * future, otherwise just return 0
+        */
+       if (!ret && ipipe->waiting_writers && (flags & SPLICE_F_NONBLOCK))
+               ret = -EAGAIN;
+
+       pipe_unlock(ipipe);
+       pipe_unlock(opipe);
 
        /*
         * If we put data in the output pipe, wakeup any potential readers.
@@ -1735,21 +1988,18 @@ static long do_tee(struct file *in, struct file *out, size_t len,
                 * Keep going, unless we encounter an error. The ipipe/opipe
                 * ordering doesn't really matter.
                 */
-               ret = link_ipipe_prep(ipipe, flags);
+               ret = ipipe_prep(ipipe, flags);
                if (!ret) {
-                       ret = link_opipe_prep(opipe, flags);
-                       if (!ret) {
+                       ret = opipe_prep(opipe, flags);
+                       if (!ret)
                                ret = link_pipe(ipipe, opipe, len, flags);
-                               if (!ret && (flags & SPLICE_F_NONBLOCK))
-                                       ret = -EAGAIN;
-                       }
                }
        }
 
        return ret;
 }
 
-asmlinkage long sys_tee(int fdin, int fdout, size_t len, unsigned int flags)
+SYSCALL_DEFINE4(tee, int, fdin, int, fdout, size_t, len, unsigned int, flags)
 {
        struct file *in;
        int error, fput_in;