nfsd4: ignore WANT bits in open downgrade
[linux-2.6.git] / fs / splice.c
index 6349d31..fa2defa 100644 (file)
@@ -20,7 +20,8 @@
 #include <linux/fs.h>
 #include <linux/file.h>
 #include <linux/pagemap.h>
-#include <linux/pipe_fs_i.h>
+#include <linux/splice.h>
+#include <linux/memcontrol.h>
 #include <linux/mm_inline.h>
 #include <linux/swap.h>
 #include <linux/writeback.h>
 #include <linux/module.h>
 #include <linux/syscalls.h>
 #include <linux/uio.h>
-
-struct partial_page {
-       unsigned int offset;
-       unsigned int len;
-};
-
-/*
- * Passed to splice_to_pipe
- */
-struct splice_pipe_desc {
-       struct page **pages;            /* page map */
-       struct partial_page *partial;   /* pages[] may not be contig */
-       int nr_pages;                   /* number of pages in map */
-       unsigned int flags;             /* splice flags */
-       const struct pipe_buf_operations *ops;/* ops associated with output pipe */
-};
+#include <linux/security.h>
+#include <linux/gfp.h>
 
 /*
  * Attempt to steal a page from a pipe buffer. This should perhaps go into
@@ -73,8 +60,9 @@ 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
@@ -90,6 +78,7 @@ 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;
 }
@@ -101,8 +90,12 @@ static void page_cache_pipe_buf_release(struct pipe_inode_info *pipe,
        buf->flags &= ~PIPE_BUF_FLAG_LRU;
 }
 
-static int page_cache_pipe_buf_pin(struct pipe_inode_info *pipe,
-                                  struct pipe_buffer *buf)
+/*
+ * Check whether the contents of buf is OK to access. Since the content
+ * is a page cache page, IO may be in flight.
+ */
+static int page_cache_pipe_buf_confirm(struct pipe_inode_info *pipe,
+                                      struct pipe_buffer *buf)
 {
        struct page *page = buf->page;
        int err;
@@ -139,11 +132,11 @@ error:
        return err;
 }
 
-static const struct pipe_buf_operations page_cache_pipe_buf_ops = {
+const struct pipe_buf_operations page_cache_pipe_buf_ops = {
        .can_merge = 0,
        .map = generic_pipe_buf_map,
        .unmap = generic_pipe_buf_unmap,
-       .pin = page_cache_pipe_buf_pin,
+       .confirm = page_cache_pipe_buf_confirm,
        .release = page_cache_pipe_buf_release,
        .steal = page_cache_pipe_buf_steal,
        .get = generic_pipe_buf_get,
@@ -163,27 +156,42 @@ static const struct pipe_buf_operations user_page_pipe_buf_ops = {
        .can_merge = 0,
        .map = generic_pipe_buf_map,
        .unmap = generic_pipe_buf_unmap,
-       .pin = generic_pipe_buf_pin,
+       .confirm = generic_pipe_buf_confirm,
        .release = page_cache_pipe_buf_release,
        .steal = user_page_pipe_buf_steal,
        .get = generic_pipe_buf_get,
 };
 
-/*
- * Pipe output worker. This sets up our pipe format with the page cache
- * pipe buffer operations. Otherwise very similar to the regular pipe_writev().
+static void wakeup_pipe_readers(struct pipe_inode_info *pipe)
+{
+       smp_mb();
+       if (waitqueue_active(&pipe->wait))
+               wake_up_interruptible(&pipe->wait);
+       kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
+}
+
+/**
+ * splice_to_pipe - fill passed data into a pipe
+ * @pipe:      pipe to fill
+ * @spd:       data to fill
+ *
+ * Description:
+ *    @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.
+ *
  */
-static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
-                             struct splice_pipe_desc *spd)
+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;
 
        ret = 0;
        do_wakeup = 0;
        page_nr = 0;
 
-       if (pipe->inode)
-               mutex_lock(&pipe->inode->i_mutex);
+       pipe_lock(pipe);
 
        for (;;) {
                if (!pipe->readers) {
@@ -193,13 +201,14 @@ static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
                        break;
                }
 
-               if (pipe->nrbufs < PIPE_BUFFERS) {
-                       int newbuf = (pipe->curbuf + pipe->nrbufs) & (PIPE_BUFFERS - 1);
+               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;
@@ -213,7 +222,7 @@ static ssize_t splice_to_pipe(struct pipe_inode_info *pipe,
 
                        if (!--spd->nr_pages)
                                break;
-                       if (pipe->nrbufs < PIPE_BUFFERS)
+                       if (pipe->nrbufs < pipe->buffers)
                                continue;
 
                        break;
@@ -244,72 +253,96 @@ static 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)
+               wakeup_pipe_readers(pipe);
 
-       while (page_nr < spd->nr_pages)
-               page_cache_release(spd->pages[page_nr++]);
+       while (page_nr < spd_pages)
+               spd->spd_release(spd, page_nr++);
 
        return ret;
 }
 
+void spd_release_page(struct splice_pipe_desc *spd, unsigned int i)
+{
+       page_cache_release(spd->pages[i]);
+}
+
+/*
+ * Check if we need to grow the arrays holding pages and partial page
+ * descriptions.
+ */
+int splice_grow_spd(struct pipe_inode_info *pipe, struct splice_pipe_desc *spd)
+{
+       if (pipe->buffers <= PIPE_DEF_BUFFERS)
+               return 0;
+
+       spd->pages = kmalloc(pipe->buffers * sizeof(struct page *), GFP_KERNEL);
+       spd->partial = kmalloc(pipe->buffers * sizeof(struct partial_page), GFP_KERNEL);
+
+       if (spd->pages && spd->partial)
+               return 0;
+
+       kfree(spd->pages);
+       kfree(spd->partial);
+       return -ENOMEM;
+}
+
+void splice_shrink_spd(struct pipe_inode_info *pipe,
+                      struct splice_pipe_desc *spd)
+{
+       if (pipe->buffers <= PIPE_DEF_BUFFERS)
+               return;
+
+       kfree(spd->pages);
+       kfree(spd->partial);
+}
+
 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;
-       struct page *pages[PIPE_BUFFERS];
-       struct partial_page partial[PIPE_BUFFERS];
+       unsigned int loff, nr_pages, req_pages;
+       struct page *pages[PIPE_DEF_BUFFERS];
+       struct partial_page partial[PIPE_DEF_BUFFERS];
        struct page *page;
        pgoff_t index, end_index;
        loff_t isize;
-       size_t total_len;
        int error, page_nr;
        struct splice_pipe_desc spd = {
                .pages = pages,
                .partial = partial,
                .flags = flags,
                .ops = &page_cache_pipe_buf_ops,
+               .spd_release = spd_release_page,
        };
 
+       if (splice_grow_spd(pipe, &spd))
+               return -ENOMEM;
+
        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);
-
-       /*
-        * Now fill in the holes:
-        */
-       error = 0;
-       total_len = 0;
+       req_pages = (len + loff + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+       nr_pages = min(req_pages, pipe->buffers);
 
        /*
         * Lookup the (hopefully) full range of pages we need.
         */
-       spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, pages);
+       spd.nr_pages = find_get_pages_contig(mapping, index, nr_pages, spd.pages);
+       index += spd.nr_pages;
 
        /*
         * If find_get_pages_contig() returned fewer pages than we needed,
-        * allocate the rest.
+        * readahead/allocate the rest and fill in the holes.
         */
-       index += spd.nr_pages;
+       if (spd.nr_pages < nr_pages)
+               page_cache_sync_readahead(mapping, &in->f_ra, in,
+                               index, req_pages - spd.nr_pages);
+
+       error = 0;
        while (spd.nr_pages < nr_pages) {
                /*
                 * Page could be there, find_get_pages_contig() breaks on
@@ -318,12 +351,6 @@ __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);
@@ -331,7 +358,7 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                                break;
 
                        error = add_to_page_cache_lru(page, mapping, index,
-                                             GFP_KERNEL);
+                                               GFP_KERNEL);
                        if (unlikely(error)) {
                                page_cache_release(page);
                                if (error == -EEXIST)
@@ -345,7 +372,7 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                        unlock_page(page);
                }
 
-               pages[spd.nr_pages++] = page;
+               spd.pages[spd.nr_pages++] = page;
                index++;
        }
 
@@ -366,30 +393,35 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
                 * this_len is the max we'll use from this page
                 */
                this_len = min_t(unsigned long, len, PAGE_CACHE_SIZE - loff);
-               page = pages[page_nr];
+               page = spd.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
                 */
                if (!PageUptodate(page)) {
-                       /*
-                        * If in nonblock mode then dont block on waiting
-                        * for an in-flight io page
-                        */
-                       if (flags & SPLICE_F_NONBLOCK) {
-                               if (TestSetPageLocked(page))
-                                       break;
-                       } else
-                               lock_page(page);
+                       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(spd.pages[page_nr]);
+                               spd.pages[page_nr] = page;
                        }
                        /*
                         * page was already under io and is now done, great
@@ -415,70 +447,80 @@ __generic_file_splice_read(struct file *in, loff_t *ppos,
 
                                break;
                        }
+               }
+fill_it:
+               /*
+                * i_size must be checked after PageUptodate.
+                */
+               isize = i_size_read(mapping->host);
+               end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
+               if (unlikely(!isize || index > end_index))
+                       break;
+
+               /*
+                * if this is the last page, see if we need to shrink
+                * the length and stop
+                */
+               if (end_index == index) {
+                       unsigned int plen;
 
                        /*
-                        * i_size must be checked after ->readpage().
+                        * max good bytes in this page
                         */
-                       isize = i_size_read(mapping->host);
-                       end_index = (isize - 1) >> PAGE_CACHE_SHIFT;
-                       if (unlikely(!isize || index > end_index))
+                       plen = ((isize - 1) & ~PAGE_CACHE_MASK) + 1;
+                       if (plen <= loff)
                                break;
 
                        /*
-                        * if this is the last page, see if we need to shrink
-                        * the length and stop
+                        * force quit after adding this page
                         */
-                       if (end_index == index) {
-                               loff = PAGE_CACHE_SIZE - (isize & ~PAGE_CACHE_MASK);
-                               if (total_len + loff > isize)
-                                       break;
-                               /*
-                                * force quit after adding this page
-                                */
-                               len = this_len;
-                               this_len = min(this_len, loff);
-                               loff = 0;
-                       }
+                       this_len = min(this_len, plen - loff);
+                       len = this_len;
                }
-fill_it:
-               partial[page_nr].offset = loff;
-               partial[page_nr].len = this_len;
+
+               spd.partial[page_nr].offset = loff;
+               spd.partial[page_nr].len = this_len;
                len -= this_len;
-               total_len += this_len;
                loff = 0;
                spd.nr_pages++;
                index++;
        }
 
        /*
-        * Release any pages at the end, if we quit early. 'i' is how far
+        * Release any pages at the end, if we quit early. 'page_nr' is how far
         * we got, 'nr_pages' is how many pages are in the map.
         */
        while (page_nr < nr_pages)
-               page_cache_release(pages[page_nr++]);
+               page_cache_release(spd.pages[page_nr++]);
+       in->f_ra.prev_pos = (loff_t)index << PAGE_CACHE_SHIFT;
 
        if (spd.nr_pages)
-               return splice_to_pipe(pipe, &spd);
+               error = splice_to_pipe(pipe, &spd);
 
+       splice_shrink_spd(pipe, &spd);
        return error;
 }
 
 /**
  * generic_file_splice_read - splice data from file to a pipe
  * @in:                file to splice from
+ * @ppos:      position in @in
  * @pipe:      pipe to splice to
  * @len:       number of bytes to splice
  * @flags:     splice modifier flags
  *
- * Will read pages from given file and fill them into a pipe.
+ * Description:
+ *    Will read pages from given file and fill them into a pipe. Can be
+ *    used as long as the address_space operations for the source implements
+ *    a readpage() hook.
+ *
  */
 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))
@@ -488,35 +530,153 @@ 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);
-
-               if (ret < 0)
-                       break;
-               else if (!ret) {
-                       if (spliced)
-                               break;
-                       if (flags & SPLICE_F_NONBLOCK) {
-                               ret = -EAGAIN;
-                               break;
-                       }
-               }
-
+       ret = __generic_file_splice_read(in, ppos, pipe, len, flags);
+       if (ret > 0) {
                *ppos += ret;
-               len -= ret;
-               spliced += ret;
+               file_accessed(in);
        }
 
-       if (spliced)
-               return spliced;
-
        return ret;
 }
-
 EXPORT_SYMBOL(generic_file_splice_read);
 
+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,
+};
+
+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;
+
+       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;
+}
+
+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_DEF_BUFFERS];
+       struct partial_page partial[PIPE_DEF_BUFFERS];
+       struct iovec *vec, __vec[PIPE_DEF_BUFFERS];
+       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,
+       };
+
+       if (splice_grow_spd(pipe, &spd))
+               return -ENOMEM;
+
+       res = -ENOMEM;
+       vec = __vec;
+       if (pipe->buffers > PIPE_DEF_BUFFERS) {
+               vec = kmalloc(pipe->buffers * sizeof(struct iovec), GFP_KERNEL);
+               if (!vec)
+                       goto shrink_ret;
+       }
+
+       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;
+               spd.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);
+               spd.partial[i].offset = 0;
+               spd.partial[i].len = this_len;
+               if (!this_len) {
+                       __free_page(spd.pages[i]);
+                       spd.pages[i] = NULL;
+                       nr_freed++;
+               }
+               res -= this_len;
+       }
+       spd.nr_pages -= nr_freed;
+
+       res = splice_to_pipe(pipe, &spd);
+       if (res > 0)
+               *ppos += res;
+
+shrink_ret:
+       if (vec != __vec)
+               kfree(vec);
+       splice_shrink_spd(pipe, &spd);
+       return res;
+
+err:
+       for (i = 0; i < spd.nr_pages; i++)
+               __free_page(spd.pages[i]);
+
+       res = error;
+       goto shrink_ret;
+}
+EXPORT_SYMBOL(default_file_splice_read);
+
 /*
  * Send 'sd->len' bytes to socket from 'sd->file' at position 'sd->pos'
  * using sendpage(). Return the number of bytes sent.
@@ -524,19 +684,16 @@ EXPORT_SYMBOL(generic_file_splice_read);
 static int pipe_to_sendpage(struct pipe_inode_info *pipe,
                            struct pipe_buffer *buf, struct splice_desc *sd)
 {
-       struct file *file = sd->file;
+       struct file *file = sd->u.file;
        loff_t pos = sd->pos;
-       int ret, more;
-
-       ret = buf->ops->pin(pipe, buf);
-       if (!ret) {
-               more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
+       int more;
 
-               ret = file->f_op->sendpage(file, buf->page, buf->offset,
-                                          sd->len, &pos, more);
-       }
+       if (!likely(file->f_op && file->f_op->sendpage))
+               return -EINVAL;
 
-       return ret;
+       more = (sd->flags & SPLICE_F_MORE) || sd->len < sd->total_len;
+       return file->f_op->sendpage(file, buf->page, buf->offset,
+                                   sd->len, &pos, more);
 }
 
 /*
@@ -559,66 +716,26 @@ 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->file;
+       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;
 
-       /*
-        * make sure the data in this buffer is uptodate
-        */
-       ret = buf->ops->pin(pipe, 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) {
                /*
@@ -632,213 +749,234 @@ 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);
 
-/*
- * Pipe input worker. Most of this logic works like a regular pipe, the
- * key here is the 'actor' worker passed in that actually moves the data
- * to the wanted destination. See pipe_to_file/pipe_to_sendpage above.
- */
-ssize_t __splice_from_pipe(struct pipe_inode_info *pipe,
-                          struct file *out, loff_t *ppos, size_t len,
-                          unsigned int flags, splice_actor *actor)
+static void wakeup_pipe_writers(struct pipe_inode_info *pipe)
 {
-       int ret, do_wakeup, err;
-       struct splice_desc sd;
+       smp_mb();
+       if (waitqueue_active(&pipe->wait))
+               wake_up_interruptible(&pipe->wait);
+       kill_fasync(&pipe->fasync_writers, SIGIO, POLL_OUT);
+}
 
-       ret = 0;
-       do_wakeup = 0;
+/**
+ * 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 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.
+ */
+int splice_from_pipe_feed(struct pipe_inode_info *pipe, struct splice_desc *sd,
+                         splice_actor *actor)
+{
+       int ret;
 
-       sd.total_len = len;
-       sd.flags = flags;
-       sd.file = out;
-       sd.pos = *ppos;
+       while (pipe->nrbufs) {
+               struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
+               const struct pipe_buf_operations *ops = buf->ops;
 
-       for (;;) {
-               if (pipe->nrbufs) {
-                       struct pipe_buffer *buf = pipe->bufs + pipe->curbuf;
-                       const struct pipe_buf_operations *ops = buf->ops;
+               sd->len = buf->len;
+               if (sd->len > sd->total_len)
+                       sd->len = sd->total_len;
 
-                       sd.len = buf->len;
-                       if (sd.len > sd.total_len)
-                               sd.len = sd.total_len;
+               ret = buf->ops->confirm(pipe, buf);
+               if (unlikely(ret)) {
+                       if (ret == -ENODATA)
+                               ret = 0;
+                       return ret;
+               }
 
-                       err = actor(pipe, buf, &sd);
-                       if (err <= 0) {
-                               if (!ret && err != -ENODATA)
-                                       ret = err;
+               ret = actor(pipe, buf, sd);
+               if (ret <= 0)
+                       return ret;
 
-                               break;
-                       }
+               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 (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);
+EXPORT_SYMBOL(splice_from_pipe_begin);
 
-ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
-                        loff_t *ppos, size_t len, unsigned int flags,
-                        splice_actor *actor)
+/**
+ * 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)
 {
-       ssize_t ret;
-       struct inode *inode = out->f_mapping->host;
+       if (sd->need_wakeup)
+               wakeup_pipe_writers(pipe);
+}
+EXPORT_SYMBOL(splice_from_pipe_end);
 
-       /*
-        * 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);
-       ret = __splice_from_pipe(pipe, out, ppos, len, flags, actor);
-       inode_double_unlock(inode, pipe->inode);
+/**
+ * __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;
 
-       return 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);
 
 /**
- * generic_file_splice_write_nolock - generic_file_splice_write without mutexes
- * @pipe:      pipe info
- * @out:       file to write to
- * @len:       number of bytes to splice
+ * splice_from_pipe - splice data from a pipe to a file
+ * @pipe:      pipe to splice from
+ * @out:       file to splice to
+ * @ppos:      position in @out
+ * @len:       how many bytes to splice
  * @flags:     splice modifier flags
+ * @actor:     handler that splices the data
  *
- * 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.
+ * Description:
+ *    See __splice_from_pipe. This function locks the pipe inode,
+ *    otherwise it's identical to __splice_from_pipe().
  *
  */
-ssize_t
-generic_file_splice_write_nolock(struct pipe_inode_info *pipe, struct file *out,
-                                loff_t *ppos, size_t len, unsigned int flags)
+ssize_t splice_from_pipe(struct pipe_inode_info *pipe, struct file *out,
+                        loff_t *ppos, size_t len, unsigned int flags,
+                        splice_actor *actor)
 {
-       struct address_space *mapping = out->f_mapping;
-       struct inode *inode = mapping->host;
        ssize_t ret;
-       int err;
-
-       err = remove_suid(out->f_path.dentry);
-       if (unlikely(err))
-               return err;
-
-       ret = __splice_from_pipe(pipe, out, ppos, len, flags, pipe_to_file);
-       if (ret > 0) {
-               *ppos += 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))) {
-                       err = generic_osync_inode(inode, mapping,
-                                                 OSYNC_METADATA|OSYNC_DATA);
+       struct splice_desc sd = {
+               .total_len = len,
+               .flags = flags,
+               .pos = *ppos,
+               .u.file = out,
+       };
 
-                       if (err)
-                               ret = err;
-               }
-               balance_dirty_pages_ratelimited(mapping);
-       }
+       pipe_lock(pipe);
+       ret = __splice_from_pipe(pipe, &sd, actor);
+       pipe_unlock(pipe);
 
        return ret;
 }
 
-EXPORT_SYMBOL(generic_file_splice_write_nolock);
-
 /**
  * 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
  *
- * Will either move or copy pages (determined by @flags options) from
- * the given pipe inode to the given file.
+ * Description:
+ *    Will either move or copy pages (determined by @flags options) from
+ *    the given pipe inode to the given file.
  *
  */
 ssize_t
@@ -847,36 +985,49 @@ generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 {
        struct address_space *mapping = out->f_mapping;
        struct inode *inode = mapping->host;
+       struct splice_desc sd = {
+               .total_len = len,
+               .flags = flags,
+               .pos = *ppos,
+               .u.file = out,
+       };
        ssize_t ret;
-       int err;
 
-       err = should_remove_suid(out->f_path.dentry);
-       if (unlikely(err)) {
-               mutex_lock(&inode->i_mutex);
-               err = __remove_suid(out->f_path.dentry, 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) {
+                       file_update_time(out);
+                       ret = splice_from_pipe_feed(pipe, &sd, pipe_to_file);
+               }
                mutex_unlock(&inode->i_mutex);
-               if (err)
-                       return err;
-       }
+       } 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, out, ppos, len, flags, pipe_to_file);
        if (ret > 0) {
-               *ppos += ret;
+               unsigned long nr_pages;
+               int err;
 
-               /*
-                * 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);
-
-                       if (err)
-                               ret = err;
-               }
-               balance_dirty_pages_ratelimited(mapping);
+               nr_pages = (ret + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
+
+               err = generic_write_sync(out, *ppos, ret);
+               if (err)
+                       ret = err;
+               else
+                       *ppos += ret;
+               balance_dirty_pages_ratelimited_nr(mapping, nr_pages);
        }
 
        return ret;
@@ -884,15 +1035,43 @@ generic_file_splice_write(struct pipe_inode_info *pipe, struct file *out,
 
 EXPORT_SYMBOL(generic_file_splice_write);
 
+static int write_pipe_buf(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
+                         struct splice_desc *sd)
+{
+       int ret;
+       void *data;
+
+       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);
+
+       return ret;
+}
+
+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;
+
+       ret = splice_from_pipe(pipe, out, ppos, len, flags, write_pipe_buf);
+       if (ret > 0)
+               *ppos += ret;
+
+       return ret;
+}
+
 /**
  * generic_splice_sendpage - splice data from a pipe to a socket
- * @inode:     pipe inode
+ * @pipe:      pipe to splice from
  * @out:       socket to write to
+ * @ppos:      position in @out
  * @len:       number of bytes to splice
  * @flags:     splice modifier flags
  *
- * Will send @len bytes from the pipe to a network socket. No data copying
- * is involved.
+ * Description:
+ *    Will send @len bytes from the pipe to a network socket. No data copying
+ *    is involved.
  *
  */
 ssize_t generic_splice_sendpage(struct pipe_inode_info *pipe, struct file *out,
@@ -909,19 +1088,26 @@ 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);
+       if (out->f_op && out->f_op->splice_write)
+               splice_write = out->f_op->splice_write;
+       else
+               splice_write = default_file_splice_write;
+
+       return splice_write(pipe, out, ppos, len, flags);
 }
 
 /*
@@ -931,11 +1117,10 @@ 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;
 
@@ -943,17 +1128,35 @@ 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);
+       if (in->f_op && in->f_op->splice_read)
+               splice_read = in->f_op->splice_read;
+       else
+               splice_read = default_file_splice_read;
+
+       return splice_read(in, ppos, pipe, len, flags);
 }
 
-long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
-                     size_t len, unsigned int flags)
+/**
+ * splice_direct_to_actor - splices data directly between two non-pipes
+ * @in:                file to splice from
+ * @sd:                actor information on where to splice to
+ * @actor:     handles the data splicing
+ *
+ * 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 lifetime of
+ *    that process.
+ *
+ */
+ssize_t splice_direct_to_actor(struct file *in, struct splice_desc *sd,
+                              splice_direct_actor *actor)
 {
        struct pipe_inode_info *pipe;
        long ret, bytes;
-       loff_t out_off;
        umode_t i_mode;
-       int i;
+       size_t len;
+       int i, flags;
 
        /*
         * We require the input being a regular file, as we don't want to
@@ -989,49 +1192,49 @@ long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
         */
        ret = 0;
        bytes = 0;
-       out_off = 0;
+       len = sd->total_len;
+       flags = sd->flags;
 
-       while (len) {
-               size_t read_len, max_read_len;
+       /*
+        * Don't block on output, we have to drain the direct pipe.
+        */
+       sd->flags &= ~SPLICE_F_NONBLOCK;
 
-               /*
-                * Do at most PIPE_BUFFERS pages worth of transfer:
-                */
-               max_read_len = min(len, (size_t)(PIPE_BUFFERS*PAGE_SIZE));
+       while (len) {
+               size_t read_len;
+               loff_t pos = sd->pos, prev_pos = pos;
 
-               ret = do_splice_to(in, ppos, 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;
+               sd->total_len = read_len;
 
                /*
                 * NOTE: nonblocking mode only applies to the input. We
                 * must not do the output in nonblocking mode as then we
                 * could get stuck data in the internal pipe:
                 */
-               ret = do_splice_from(pipe, out, &out_off, read_len,
-                                    flags & ~SPLICE_F_NONBLOCK);
-               if (unlikely(ret < 0))
+               ret = actor(pipe, sd);
+               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:
@@ -1039,7 +1242,7 @@ out_release:
         * If we did an incomplete transfer we must release
         * the pipe buffers in question:
         */
-       for (i = 0; i < PIPE_BUFFERS; i++) {
+       for (i = 0; i < pipe->buffers; i++) {
                struct pipe_buffer *buf = pipe->bufs + i;
 
                if (buf->ops) {
@@ -1047,30 +1250,61 @@ 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;
+       if (!bytes)
+               bytes = ret;
 
-       return ret;
+       goto done;
 }
+EXPORT_SYMBOL(splice_direct_to_actor);
 
-/*
- * 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
- * pipe.
+static int direct_splice_actor(struct pipe_inode_info *pipe,
+                              struct splice_desc *sd)
+{
+       struct file *file = sd->u.file;
+
+       return do_splice_from(pipe, file, &file->f_pos, sd->total_len,
+                             sd->flags);
+}
+
+/**
+ * do_splice_direct - splices data directly between two files
+ * @in:                file to splice from
+ * @ppos:      input file offset
+ * @out:       file to splice to
+ * @len:       number of bytes to splice
+ * @flags:     splice modifier flags
+ *
+ * Description:
+ *    For use by do_sendfile(). splice can easily emulate sendfile, but
+ *    doing it in the application would incur an extra system call
+ *    (splice in + splice out, as compared to just sendfile()). So this helper
+ *    can splice directly through a process-private pipe.
+ *
  */
-static inline struct pipe_inode_info *pipe_info(struct inode *inode)
+long do_splice_direct(struct file *in, loff_t *ppos, struct file *out,
+                     size_t len, unsigned int flags)
 {
-       if (S_ISFIFO(inode->i_mode))
-               return inode->i_pipe;
+       struct splice_desc sd = {
+               .len            = len,
+               .total_len      = len,
+               .flags          = flags,
+               .pos            = *ppos,
+               .u.file         = out,
+       };
+       long ret;
+
+       ret = splice_direct_to_actor(in, &sd, direct_splice_actor);
+       if (ret > 0)
+               *ppos = sd.pos;
 
-       return NULL;
+       return ret;
 }
 
+static int splice_pipe_to_pipe(struct pipe_inode_info *ipipe,
+                              struct pipe_inode_info *opipe,
+                              size_t len, unsigned int flags);
+
 /*
  * Determine where to splice to/from.
  */
@@ -1078,16 +1312,36 @@ 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 = get_pipe_info(in);
+       opipe = get_pipe_info(out);
+
+       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) {
-                       if (out->f_op->llseek == no_llseek)
+                       if (!(out->f_mode & FMODE_PWRITE))
                                return -EINVAL;
                        if (copy_from_user(&offset, off_out, sizeof(loff_t)))
                                return -EFAULT;
@@ -1095,7 +1349,7 @@ 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;
@@ -1103,12 +1357,11 @@ 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) {
-                       if (in->f_op->llseek == no_llseek)
+                       if (!(in->f_mode & FMODE_PREAD))
                                return -EINVAL;
                        if (copy_from_user(&offset, off_in, sizeof(loff_t)))
                                return -EFAULT;
@@ -1116,7 +1369,7 @@ 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;
@@ -1136,39 +1389,33 @@ static long do_splice(struct file *in, loff_t __user *off_in,
  */
 static int get_iovec_page_array(const struct iovec __user *iov,
                                unsigned int nr_vecs, struct page **pages,
-                               struct partial_page *partial, int aligned)
+                               struct partial_page *partial, int aligned,
+                               unsigned int pipe_buffers)
 {
        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;
 
                /*
@@ -1186,12 +1433,11 @@ static int get_iovec_page_array(const struct iovec __user *iov,
                        break;
 
                npages = (off + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
-               if (npages > PIPE_BUFFERS - buffers)
-                       npages = PIPE_BUFFERS - buffers;
+               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;
@@ -1223,82 +1469,210 @@ static int get_iovec_page_array(const struct iovec __user *iov,
                 * or if we mapped the max number of pages that we have
                 * room for.
                 */
-               if (error < npages || buffers == PIPE_BUFFERS)
+               if (error < npages || buffers == pipe_buffers)
                        break;
 
                nr_vecs--;
                iov++;
        }
 
-       up_read(&current->mm->mmap_sem);
-
        if (buffers)
                return buffers;
 
        return error;
 }
 
+static int pipe_to_user(struct pipe_inode_info *pipe, struct pipe_buffer *buf,
+                       struct splice_desc *sd)
+{
+       char *src;
+       int ret;
+
+       /*
+        * See if we can use the atomic maps, by prefaulting in the
+        * pages and doing an atomic copy
+        */
+       if (!fault_in_pages_writeable(sd->u.userptr, sd->len)) {
+               src = buf->ops->map(pipe, buf, 1);
+               ret = __copy_to_user_inatomic(sd->u.userptr, src + buf->offset,
+                                                       sd->len);
+               buf->ops->unmap(pipe, buf, src);
+               if (!ret) {
+                       ret = sd->len;
+                       goto out;
+               }
+       }
+
+       /*
+        * No dice, use slow non-atomic map and copy
+        */
+       src = buf->ops->map(pipe, buf, 0);
+
+       ret = sd->len;
+       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;
+       return ret;
+}
+
+/*
+ * For lack of a better implementation, implement vmsplice() to userspace
+ * as a simple copy of the pipes pages to the user iov.
+ */
+static long vmsplice_to_user(struct file *file, const struct iovec __user *iov,
+                            unsigned long nr_segs, unsigned int flags)
+{
+       struct pipe_inode_info *pipe;
+       struct splice_desc sd;
+       ssize_t size;
+       int error;
+       long ret;
+
+       pipe = get_pipe_info(file);
+       if (!pipe)
+               return -EBADF;
+
+       pipe_lock(pipe);
+
+       error = ret = 0;
+       while (nr_segs) {
+               void __user *base;
+               size_t len;
+
+               /*
+                * 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))
+                       break;
+
+               /*
+                * Sanity check this iovec. 0 read succeeds.
+                */
+               if (unlikely(!len))
+                       break;
+               if (unlikely(!base)) {
+                       error = -EFAULT;
+                       break;
+               }
+
+               if (unlikely(!access_ok(VERIFY_WRITE, base, len))) {
+                       error = -EFAULT;
+                       break;
+               }
+
+               sd.len = 0;
+               sd.total_len = len;
+               sd.flags = flags;
+               sd.u.userptr = base;
+               sd.pos = 0;
+
+               size = __splice_from_pipe(pipe, &sd, pipe_to_user);
+               if (size < 0) {
+                       if (!ret)
+                               ret = size;
+
+                       break;
+               }
+
+               ret += size;
+
+               if (size < len)
+                       break;
+
+               nr_segs--;
+               iov++;
+       }
+
+       pipe_unlock(pipe);
+
+       if (!ret)
+               ret = error;
+
+       return ret;
+}
+
 /*
  * vmsplice splices a user address range into a pipe. It can be thought of
  * as splice-from-memory, where the regular splice is splice-from-file (or
  * to file). In both cases the output is a pipe, naturally.
- *
- * Note that vmsplice only supports splicing _from_ user memory to a pipe,
- * not the other way around. Splicing from user memory is a simple operation
- * that can be supported without any funky alignment restrictions or nasty
- * vm tricks. We simply map in the user memory and fill them into a pipe.
- * The reverse isn't quite as easy, though. There are two possible solutions
- * for that:
- *
- *     - memcpy() the data internally, at which point we might as well just
- *       do a regular read() on the buffer anyway.
- *     - Lots of nasty vm tricks, that are neither fast nor flexible (it
- *       has restriction limitations on both ends of the pipe).
- *
- * Alas, it isn't here.
- *
  */
-static long do_vmsplice(struct file *file, const struct iovec __user *iov,
-                       unsigned long nr_segs, unsigned int flags)
+static long vmsplice_to_pipe(struct file *file, const struct iovec __user *iov,
+                            unsigned long nr_segs, unsigned int flags)
 {
        struct pipe_inode_info *pipe;
-       struct page *pages[PIPE_BUFFERS];
-       struct partial_page partial[PIPE_BUFFERS];
+       struct page *pages[PIPE_DEF_BUFFERS];
+       struct partial_page partial[PIPE_DEF_BUFFERS];
        struct splice_pipe_desc spd = {
                .pages = pages,
                .partial = partial,
                .flags = flags,
                .ops = &user_page_pipe_buf_ops,
+               .spd_release = spd_release_page,
        };
+       long ret;
 
-       pipe = pipe_info(file->f_path.dentry->d_inode);
+       pipe = get_pipe_info(file);
        if (!pipe)
                return -EBADF;
-       if (unlikely(nr_segs > UIO_MAXIOV))
-               return -EINVAL;
-       else if (unlikely(!nr_segs))
-               return 0;
 
-       spd.nr_pages = get_iovec_page_array(iov, nr_segs, pages, partial,
-                                           flags & SPLICE_F_GIFT);
+       if (splice_grow_spd(pipe, &spd))
+               return -ENOMEM;
+
+       spd.nr_pages = get_iovec_page_array(iov, nr_segs, spd.pages,
+                                           spd.partial, flags & SPLICE_F_GIFT,
+                                           pipe->buffers);
        if (spd.nr_pages <= 0)
-               return spd.nr_pages;
+               ret = spd.nr_pages;
+       else
+               ret = splice_to_pipe(pipe, &spd);
 
-       return splice_to_pipe(pipe, &spd);
+       splice_shrink_spd(pipe, &spd);
+       return ret;
 }
 
-asmlinkage long sys_vmsplice(int fd, const struct iovec __user *iov,
-                            unsigned long nr_segs, unsigned int flags)
+/*
+ * Note that vmsplice only really supports true splicing _from_ user memory
+ * to a pipe, not the other way around. Splicing from user memory is a simple
+ * operation that can be supported without any funky alignment restrictions
+ * or nasty vm tricks. We simply map in the user memory and fill them into
+ * a pipe. The reverse isn't quite as easy, though. There are two possible
+ * solutions for that:
+ *
+ *     - memcpy() the data internally, at which point we might as well just
+ *       do a regular read() on the buffer anyway.
+ *     - Lots of nasty vm tricks, that are neither fast nor flexible (it
+ *       has restriction limitations on both ends of the pipe).
+ *
+ * Currently we punt and implement it as a normal copy, see pipe_to_user().
+ *
+ */
+SYSCALL_DEFINE4(vmsplice, int, fd, const struct iovec __user *, iov,
+               unsigned long, nr_segs, unsigned int, flags)
 {
        struct file *file;
        long error;
        int fput;
 
+       if (unlikely(nr_segs > UIO_MAXIOV))
+               return -EINVAL;
+       else if (unlikely(!nr_segs))
+               return 0;
+
        error = -EBADF;
        file = fget_light(fd, &fput);
        if (file) {
                if (file->f_mode & FMODE_WRITE)
-                       error = do_vmsplice(file, iov, nr_segs, flags);
+                       error = vmsplice_to_pipe(file, iov, nr_segs, flags);
+               else if (file->f_mode & FMODE_READ)
+                       error = vmsplice_to_user(file, iov, nr_segs, flags);
 
                fput_light(file, fput);
        }
@@ -1306,9 +1680,9 @@ 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;
@@ -1341,7 +1715,7 @@ 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;
 
@@ -1353,7 +1727,7 @@ 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)) {
@@ -1371,7 +1745,7 @@ 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;
 }
 
@@ -1379,7 +1753,7 @@ 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;
 
@@ -1387,13 +1761,13 @@ static int link_opipe_prep(struct pipe_inode_info *pipe, unsigned int flags)
         * Check ->nrbufs without the inode lock first. This function
         * is speculative anyways, so missing one is ok.
         */
-       if (pipe->nrbufs < PIPE_BUFFERS)
+       if (pipe->nrbufs < pipe->buffers)
                return 0;
 
        ret = 0;
-       mutex_lock(&pipe->inode->i_mutex);
+       pipe_lock(pipe);
 
-       while (pipe->nrbufs >= PIPE_BUFFERS) {
+       while (pipe->nrbufs >= pipe->buffers) {
                if (!pipe->readers) {
                        send_sig(SIGPIPE, current, 0);
                        ret = -EPIPE;
@@ -1412,7 +1786,122 @@ 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 >= opipe->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) & (opipe->buffers - 1);
+               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) & (ipipe->buffers - 1);
+                       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)
+               wakeup_pipe_readers(opipe);
+
+       if (input_wakeup)
+               wakeup_pipe_writers(ipipe);
+
        return ret;
 }
 
@@ -1428,10 +1917,10 @@ 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) {
@@ -1445,11 +1934,11 @@ static int link_pipe(struct pipe_inode_info *ipipe,
                 * If we have iterated all input buffers or ran out of
                 * output room, break.
                 */
-               if (i >= ipipe->nrbufs || opipe->nrbufs >= PIPE_BUFFERS)
+               if (i >= ipipe->nrbufs || opipe->nrbufs >= opipe->buffers)
                        break;
 
-               ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (PIPE_BUFFERS - 1));
-               nbuf = (opipe->curbuf + opipe->nrbufs) & (PIPE_BUFFERS - 1);
+               ibuf = ipipe->bufs + ((ipipe->curbuf + i) & (ipipe->buffers-1));
+               nbuf = (opipe->curbuf + opipe->nrbufs) & (opipe->buffers - 1);
 
                /*
                 * Get a reference to this pipe buffer,
@@ -1475,17 +1964,21 @@ 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.
         */
-       if (ret > 0) {
-               smp_mb();
-               if (waitqueue_active(&opipe->wait))
-                       wake_up_interruptible(&opipe->wait);
-               kill_fasync(&opipe->fasync_readers, SIGIO, POLL_IN);
-       }
+       if (ret > 0)
+               wakeup_pipe_readers(opipe);
 
        return ret;
 }
@@ -1499,8 +1992,8 @@ static int link_pipe(struct pipe_inode_info *ipipe,
 static long do_tee(struct file *in, struct file *out, size_t len,
                   unsigned int flags)
 {
-       struct pipe_inode_info *ipipe = pipe_info(in->f_path.dentry->d_inode);
-       struct pipe_inode_info *opipe = pipe_info(out->f_path.dentry->d_inode);
+       struct pipe_inode_info *ipipe = get_pipe_info(in);
+       struct pipe_inode_info *opipe = get_pipe_info(out);
        int ret = -EINVAL;
 
        /*
@@ -1512,21 +2005,18 @@ 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;