NeilBrown <neilb@suse.de>
[linux-2.6.git] / mm / filemap.c
index d504d6e..adbac10 100644 (file)
@@ -9,7 +9,6 @@
  * most "normal" filesystems (but you don't /have/ to use this:
  * the NFS filesystem used to do this differently, for example)
  */
-#include <linux/config.h>
 #include <linux/module.h>
 #include <linux/slab.h>
 #include <linux/compiler.h>
@@ -76,8 +75,8 @@ generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
  *  ->mmap_sem
  *    ->lock_page              (access_process_vm)
  *
- *  ->mmap_sem
- *    ->i_mutex                        (msync)
+ *  ->i_mutex                  (generic_file_buffered_write)
+ *    ->mmap_sem               (fault_in_pages_readable->do_page_fault)
  *
  *  ->i_mutex
  *    ->i_alloc_sem             (various)
@@ -120,7 +119,8 @@ void __remove_from_page_cache(struct page *page)
        radix_tree_delete(&mapping->page_tree, page->index);
        page->mapping = NULL;
        mapping->nrpages--;
-       pagecache_acct(-1);
+       __dec_zone_page_state(page, NR_FILE_PAGES);
+       BUG_ON(page_mapped(page));
 }
 
 void remove_from_page_cache(struct page *page)
@@ -328,7 +328,7 @@ EXPORT_SYMBOL(sync_page_range);
  * @pos:       beginning offset in pages to write
  * @count:     number of bytes to write
  *
- * Note: Holding i_mutex across sync_page_range_nolock is not a good idea
+ * Note: Holding i_mutex across sync_page_range_nolock() is not a good idea
  * as it forces O_SYNC writers to different parts of the same file
  * to be serialised right until io completion.
  */
@@ -449,7 +449,7 @@ int add_to_page_cache(struct page *page, struct address_space *mapping,
                        page->mapping = mapping;
                        page->index = offset;
                        mapping->nrpages++;
-                       pagecache_acct(1);
+                       __inc_zone_page_state(page, NR_FILE_PAGES);
                }
                write_unlock_irq(&mapping->tree_lock);
                radix_tree_preload_end();
@@ -468,26 +468,22 @@ int add_to_page_cache_lru(struct page *page, struct address_space *mapping,
 }
 
 #ifdef CONFIG_NUMA
-struct page *page_cache_alloc(struct address_space *x)
+struct page *__page_cache_alloc(gfp_t gfp)
 {
        if (cpuset_do_page_mem_spread()) {
                int n = cpuset_mem_spread_node();
-               return alloc_pages_node(n, mapping_gfp_mask(x), 0);
+               return alloc_pages_node(n, gfp, 0);
        }
-       return alloc_pages(mapping_gfp_mask(x), 0);
+       return alloc_pages(gfp, 0);
 }
-EXPORT_SYMBOL(page_cache_alloc);
+EXPORT_SYMBOL(__page_cache_alloc);
+#endif
 
-struct page *page_cache_alloc_cold(struct address_space *x)
+static int __sleep_on_page_lock(void *word)
 {
-       if (cpuset_do_page_mem_spread()) {
-               int n = cpuset_mem_spread_node();
-               return alloc_pages_node(n, mapping_gfp_mask(x)|__GFP_COLD, 0);
-       }
-       return alloc_pages(mapping_gfp_mask(x)|__GFP_COLD, 0);
+       io_schedule();
+       return 0;
 }
-EXPORT_SYMBOL(page_cache_alloc_cold);
-#endif
 
 /*
  * In order to wait for pages to become available there must be
@@ -578,13 +574,24 @@ void fastcall __lock_page(struct page *page)
 }
 EXPORT_SYMBOL(__lock_page);
 
+/*
+ * Variant of lock_page that does not require the caller to hold a reference
+ * on the page's mapping.
+ */
+void fastcall __lock_page_nosync(struct page *page)
+{
+       DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
+       __wait_on_bit_lock(page_waitqueue(page), &wait, __sleep_on_page_lock,
+                                                       TASK_UNINTERRUPTIBLE);
+}
+
 /**
  * find_get_page - find and get a page reference
  * @mapping: the address_space to search
  * @offset: the page index
  *
- * A rather lightweight function, finding and getting a reference to a
- * hashed page atomically.
+ * Is there a pagecache struct page at the given (mapping, offset) tuple?
+ * If yes, increment its refcount and return it; if no, return NULL.
  */
 struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 {
@@ -600,26 +607,6 @@ struct page * find_get_page(struct address_space *mapping, unsigned long offset)
 EXPORT_SYMBOL(find_get_page);
 
 /**
- * find_trylock_page - find and lock a page
- * @mapping: the address_space to search
- * @offset: the page index
- *
- * Same as find_get_page(), but trylock it instead of incrementing the count.
- */
-struct page *find_trylock_page(struct address_space *mapping, unsigned long offset)
-{
-       struct page *page;
-
-       read_lock_irq(&mapping->tree_lock);
-       page = radix_tree_lookup(&mapping->page_tree, offset);
-       if (page && TestSetPageLocked(page))
-               page = NULL;
-       read_unlock_irq(&mapping->tree_lock);
-       return page;
-}
-EXPORT_SYMBOL(find_trylock_page);
-
-/**
  * find_lock_page - locate, pin and lock a pagecache page
  * @mapping: the address_space to search
  * @offset: the page index
@@ -684,7 +671,8 @@ repeat:
        page = find_lock_page(mapping, index);
        if (!page) {
                if (!cached_page) {
-                       cached_page = alloc_page(gfp_mask);
+                       cached_page =
+                               __page_cache_alloc(gfp_mask);
                        if (!cached_page)
                                return NULL;
                }
@@ -764,6 +752,7 @@ unsigned find_get_pages_contig(struct address_space *mapping, pgoff_t index,
        read_unlock_irq(&mapping->tree_lock);
        return i;
 }
+EXPORT_SYMBOL(find_get_pages_contig);
 
 /**
  * find_get_pages_tag - find and return pages that match @tag
@@ -792,13 +781,14 @@ unsigned find_get_pages_tag(struct address_space *mapping, pgoff_t *index,
        read_unlock_irq(&mapping->tree_lock);
        return ret;
 }
+EXPORT_SYMBOL(find_get_pages_tag);
 
 /**
  * grab_cache_page_nowait - returns locked page at given index in given cache
  * @mapping: target address_space
  * @index: the page index
  *
- * Same as grab_cache_page, but do not wait if the page is unavailable.
+ * Same as grab_cache_page(), but do not wait if the page is unavailable.
  * This is intended for speculative data generators, where the data can
  * be regenerated if the page couldn't be grabbed.  This routine should
  * be safe to call while holding the lock for another page.
@@ -810,7 +800,6 @@ struct page *
 grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
 {
        struct page *page = find_get_page(mapping, index);
-       gfp_t gfp_mask;
 
        if (page) {
                if (!TestSetPageLocked(page))
@@ -818,9 +807,8 @@ grab_cache_page_nowait(struct address_space *mapping, unsigned long index)
                page_cache_release(page);
                return NULL;
        }
-       gfp_mask = mapping_gfp_mask(mapping) & ~__GFP_FS;
-       page = alloc_pages(gfp_mask, 0);
-       if (page && add_to_page_cache_lru(page, mapping, index, gfp_mask)) {
+       page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~__GFP_FS);
+       if (page && add_to_page_cache_lru(page, mapping, index, GFP_KERNEL)) {
                page_cache_release(page);
                page = NULL;
        }
@@ -850,8 +838,6 @@ static void shrink_readahead_size_eio(struct file *filp,
                return;
 
        ra->ra_pages /= 4;
-       printk(KERN_WARNING "Reducing readahead size to %luK\n",
-                       ra->ra_pages << (PAGE_CACHE_SHIFT - 10));
 }
 
 /**
@@ -886,6 +872,7 @@ void do_generic_mapping_read(struct address_space *mapping,
        unsigned long last_index;
        unsigned long next_index;
        unsigned long prev_index;
+       unsigned int prev_offset;
        loff_t isize;
        struct page *cached_page;
        int error;
@@ -894,7 +881,8 @@ void do_generic_mapping_read(struct address_space *mapping,
        cached_page = NULL;
        index = *ppos >> PAGE_CACHE_SHIFT;
        next_index = index;
-       prev_index = ra.prev_page;
+       prev_index = ra.prev_index;
+       prev_offset = ra.prev_offset;
        last_index = (*ppos + desc->count + PAGE_CACHE_SIZE-1) >> PAGE_CACHE_SHIFT;
        offset = *ppos & ~PAGE_CACHE_MASK;
 
@@ -942,10 +930,10 @@ page_ok:
                        flush_dcache_page(page);
 
                /*
-                * When (part of) the same page is read multiple times
-                * in succession, only mark it as accessed the first time.
+                * When a sequential read accesses a page several times,
+                * only mark it as accessed the first time.
                 */
-               if (prev_index != index)
+               if (prev_index != index || offset != prev_offset)
                        mark_page_accessed(page);
                prev_index = index;
 
@@ -963,6 +951,8 @@ page_ok:
                offset += ret;
                index += offset >> PAGE_CACHE_SHIFT;
                offset &= ~PAGE_CACHE_MASK;
+               prev_offset = offset;
+               ra.prev_offset = offset;
 
                page_cache_release(page);
                if (ret == nr && desc->count)
@@ -973,7 +963,7 @@ page_not_up_to_date:
                /* Get exclusive access to the page ... */
                lock_page(page);
 
-               /* Did it get unhashed before we got the lock? */
+               /* Did it get truncated before we got the lock? */
                if (!page->mapping) {
                        unlock_page(page);
                        page_cache_release(page);
@@ -1124,48 +1114,73 @@ success:
        return size;
 }
 
-/**
- * __generic_file_aio_read - generic filesystem read routine
- * @iocb:      kernel I/O control block
+/*
+ * Performs necessary checks before doing a write
  * @iov:       io vector request
  * @nr_segs:   number of segments in the iovec
- * @ppos:      current file position
+ * @count:     number of bytes to write
+ * @access_flags: type of access: %VERIFY_READ or %VERIFY_WRITE
  *
- * This is the "read()" routine for all filesystems
- * that can use the page cache directly.
+ * Adjust number of segments and amount of bytes to write (nr_segs should be
+ * properly initialized first). Returns appropriate error code that caller
+ * should return or zero in case that write should be allowed.
  */
-ssize_t
-__generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
-               unsigned long nr_segs, loff_t *ppos)
+int generic_segment_checks(const struct iovec *iov,
+                       unsigned long *nr_segs, size_t *count, int access_flags)
 {
-       struct file *filp = iocb->ki_filp;
-       ssize_t retval;
-       unsigned long seg;
-       size_t count;
-
-       count = 0;
-       for (seg = 0; seg < nr_segs; seg++) {
+       unsigned long   seg;
+       size_t cnt = 0;
+       for (seg = 0; seg < *nr_segs; seg++) {
                const struct iovec *iv = &iov[seg];
 
                /*
                 * If any segment has a negative length, or the cumulative
                 * length ever wraps negative then return -EINVAL.
                 */
-               count += iv->iov_len;
-               if (unlikely((ssize_t)(count|iv->iov_len) < 0))
+               cnt += iv->iov_len;
+               if (unlikely((ssize_t)(cnt|iv->iov_len) < 0))
                        return -EINVAL;
-               if (access_ok(VERIFY_WRITE, iv->iov_base, iv->iov_len))
+               if (access_ok(access_flags, iv->iov_base, iv->iov_len))
                        continue;
                if (seg == 0)
                        return -EFAULT;
-               nr_segs = seg;
-               count -= iv->iov_len;   /* This segment is no good */
+               *nr_segs = seg;
+               cnt -= iv->iov_len;     /* This segment is no good */
                break;
        }
+       *count = cnt;
+       return 0;
+}
+EXPORT_SYMBOL(generic_segment_checks);
+
+/**
+ * generic_file_aio_read - generic filesystem read routine
+ * @iocb:      kernel I/O control block
+ * @iov:       io vector request
+ * @nr_segs:   number of segments in the iovec
+ * @pos:       current file position
+ *
+ * This is the "read()" routine for all filesystems
+ * that can use the page cache directly.
+ */
+ssize_t
+generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
+               unsigned long nr_segs, loff_t pos)
+{
+       struct file *filp = iocb->ki_filp;
+       ssize_t retval;
+       unsigned long seg;
+       size_t count;
+       loff_t *ppos = &iocb->ki_pos;
+
+       count = 0;
+       retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
+       if (retval)
+               return retval;
 
        /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
        if (filp->f_flags & O_DIRECT) {
-               loff_t pos = *ppos, size;
+               loff_t size;
                struct address_space *mapping;
                struct inode *inode;
 
@@ -1178,13 +1193,13 @@ __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
                if (pos < size) {
                        retval = generic_file_direct_IO(READ, iocb,
                                                iov, pos, nr_segs);
-                       if (retval > 0 && !is_sync_kiocb(iocb))
-                               retval = -EIOCBQUEUED;
                        if (retval > 0)
                                *ppos = pos + retval;
                }
-               file_accessed(filp);
-               goto out;
+               if (likely(retval != 0)) {
+                       file_accessed(filp);
+                       goto out;
+               }
        }
 
        retval = 0;
@@ -1204,38 +1219,15 @@ __generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
                                retval = retval ?: desc.error;
                                break;
                        }
+                       if (desc.count > 0)
+                               break;
                }
        }
 out:
        return retval;
 }
-EXPORT_SYMBOL(__generic_file_aio_read);
-
-ssize_t
-generic_file_aio_read(struct kiocb *iocb, char __user *buf, size_t count, loff_t pos)
-{
-       struct iovec local_iov = { .iov_base = buf, .iov_len = count };
-
-       BUG_ON(iocb->ki_pos != pos);
-       return __generic_file_aio_read(iocb, &local_iov, 1, &iocb->ki_pos);
-}
 EXPORT_SYMBOL(generic_file_aio_read);
 
-ssize_t
-generic_file_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
-{
-       struct iovec local_iov = { .iov_base = buf, .iov_len = count };
-       struct kiocb kiocb;
-       ssize_t ret;
-
-       init_sync_kiocb(&kiocb, filp);
-       ret = __generic_file_aio_read(&kiocb, &local_iov, 1, ppos);
-       if (-EIOCBQUEUED == ret)
-               ret = wait_on_sync_kiocb(&kiocb);
-       return ret;
-}
-EXPORT_SYMBOL(generic_file_read);
-
 int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long offset, unsigned long size)
 {
        ssize_t written;
@@ -1256,26 +1248,6 @@ int file_send_actor(read_descriptor_t * desc, struct page *page, unsigned long o
        return written;
 }
 
-ssize_t generic_file_sendfile(struct file *in_file, loff_t *ppos,
-                        size_t count, read_actor_t actor, void *target)
-{
-       read_descriptor_t desc;
-
-       if (!count)
-               return 0;
-
-       desc.written = 0;
-       desc.count = count;
-       desc.arg.data = target;
-       desc.error = 0;
-
-       do_generic_file_read(in_file, ppos, &desc, actor);
-       if (desc.written)
-               return desc.written;
-       return desc.error;
-}
-EXPORT_SYMBOL(generic_file_sendfile);
-
 static ssize_t
 do_readahead(struct address_space *mapping, struct file *filp,
             unsigned long index, unsigned long nr)
@@ -1416,7 +1388,7 @@ retry_find:
                 */
                if (!did_readaround) {
                        majmin = VM_FAULT_MAJOR;
-                       inc_page_state(pgmajfault);
+                       count_vm_event(PGMAJFAULT);
                }
                did_readaround = 1;
                ra_pages = max_sane_readahead(file->f_ra.ra_pages);
@@ -1457,7 +1429,7 @@ outside_data_content:
         * accessible..
         */
        if (area->vm_mm == current->mm)
-               return NULL;
+               return NOPAGE_SIGBUS;
        /* Fall through to the non-read-ahead case */
 no_cached_page:
        /*
@@ -1465,7 +1437,6 @@ no_cached_page:
         * effect.
         */
        error = page_cache_read(file, pgoff);
-       grab_swap_token();
 
        /*
         * The page we want has now been added to the page cache.
@@ -1482,36 +1453,12 @@ no_cached_page:
         */
        if (error == -ENOMEM)
                return NOPAGE_OOM;
-       return NULL;
+       return NOPAGE_SIGBUS;
 
 page_not_uptodate:
        if (!did_readaround) {
                majmin = VM_FAULT_MAJOR;
-               inc_page_state(pgmajfault);
-       }
-       lock_page(page);
-
-       /* Did it get unhashed while we waited for it? */
-       if (!page->mapping) {
-               unlock_page(page);
-               page_cache_release(page);
-               goto retry_all;
-       }
-
-       /* Did somebody else get it up-to-date? */
-       if (PageUptodate(page)) {
-               unlock_page(page);
-               goto success;
-       }
-
-       error = mapping->a_ops->readpage(file, page);
-       if (!error) {
-               wait_on_page_locked(page);
-               if (PageUptodate(page))
-                       goto success;
-       } else if (error == AOP_TRUNCATED_PAGE) {
-               page_cache_release(page);
-               goto retry_find;
+               count_vm_event(PGMAJFAULT);
        }
 
        /*
@@ -1551,7 +1498,7 @@ page_not_uptodate:
         */
        shrink_readahead_size_eio(file, ra);
        page_cache_release(page);
-       return NULL;
+       return NOPAGE_SIGBUS;
 }
 EXPORT_SYMBOL(filemap_nopage);
 
@@ -1613,7 +1560,7 @@ no_cached_page:
 page_not_uptodate:
        lock_page(page);
 
-       /* Did it get unhashed while we waited for it? */
+       /* Did it get truncated while we waited for it? */
        if (!page->mapping) {
                unlock_page(page);
                goto err;
@@ -1769,7 +1716,7 @@ int generic_file_readonly_mmap(struct file * file, struct vm_area_struct * vma)
 EXPORT_SYMBOL(generic_file_mmap);
 EXPORT_SYMBOL(generic_file_readonly_mmap);
 
-static inline struct page *__read_cache_page(struct address_space *mapping,
+static struct page *__read_cache_page(struct address_space *mapping,
                                unsigned long index,
                                int (*filler)(void *,struct page*),
                                void *data)
@@ -1806,17 +1753,11 @@ repeat:
        return page;
 }
 
-/**
- * read_cache_page - read into page cache, fill it if needed
- * @mapping:   the page's address_space
- * @index:     the page index
- * @filler:    function to perform the read
- * @data:      destination for read data
- *
- * Read into the page cache. If a page already exists,
- * and PageUptodate() is not set, try to fill the page.
+/*
+ * Same as read_cache_page, but don't wait for page to become unlocked
+ * after submitting it to the filler.
  */
-struct page *read_cache_page(struct address_space *mapping,
+struct page *read_cache_page_async(struct address_space *mapping,
                                unsigned long index,
                                int (*filler)(void *,struct page*),
                                void *data)
@@ -1827,8 +1768,7 @@ struct page *read_cache_page(struct address_space *mapping,
 retry:
        page = __read_cache_page(mapping, index, filler, data);
        if (IS_ERR(page))
-               goto out;
-       mark_page_accessed(page);
+               return page;
        if (PageUptodate(page))
                goto out;
 
@@ -1845,7 +1785,40 @@ retry:
        err = filler(data, page);
        if (err < 0) {
                page_cache_release(page);
-               page = ERR_PTR(err);
+               return ERR_PTR(err);
+       }
+out:
+       mark_page_accessed(page);
+       return page;
+}
+EXPORT_SYMBOL(read_cache_page_async);
+
+/**
+ * read_cache_page - read into page cache, fill it if needed
+ * @mapping:   the page's address_space
+ * @index:     the page index
+ * @filler:    function to perform the read
+ * @data:      destination for read data
+ *
+ * Read into the page cache. If a page already exists, and PageUptodate() is
+ * not set, try to fill the page then wait for it to become unlocked.
+ *
+ * If the page does not get brought uptodate, return -EIO.
+ */
+struct page *read_cache_page(struct address_space *mapping,
+                               unsigned long index,
+                               int (*filler)(void *,struct page*),
+                               void *data)
+{
+       struct page *page;
+
+       page = read_cache_page_async(mapping, index, filler, data);
+       if (IS_ERR(page))
+               goto out;
+       wait_on_page_locked(page);
+       if (!PageUptodate(page)) {
+               page_cache_release(page);
+               page = ERR_PTR(-EIO);
        }
  out:
        return page;
@@ -1892,11 +1865,10 @@ repeat:
  *     if suid or (sgid and xgrp)
  *             remove privs
  */
-int remove_suid(struct dentry *dentry)
+int should_remove_suid(struct dentry *dentry)
 {
        mode_t mode = dentry->d_inode->i_mode;
        int kill = 0;
-       int result = 0;
 
        /* suid always must be killed */
        if (unlikely(mode & S_ISUID))
@@ -1909,13 +1881,29 @@ int remove_suid(struct dentry *dentry)
        if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
                kill |= ATTR_KILL_SGID;
 
-       if (unlikely(kill && !capable(CAP_FSETID))) {
-               struct iattr newattrs;
+       if (unlikely(kill && !capable(CAP_FSETID)))
+               return kill;
 
-               newattrs.ia_valid = ATTR_FORCE | kill;
-               result = notify_change(dentry, &newattrs);
-       }
-       return result;
+       return 0;
+}
+EXPORT_SYMBOL(should_remove_suid);
+
+int __remove_suid(struct dentry *dentry, int kill)
+{
+       struct iattr newattrs;
+
+       newattrs.ia_valid = ATTR_FORCE | kill;
+       return notify_change(dentry, &newattrs);
+}
+
+int remove_suid(struct dentry *dentry)
+{
+       int kill = should_remove_suid(dentry);
+
+       if (unlikely(kill))
+               return __remove_suid(dentry, kill);
+
+       return 0;
 }
 EXPORT_SYMBOL(remove_suid);
 
@@ -2006,6 +1994,7 @@ inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, i
                if (unlikely(*pos + *count > inode->i_sb->s_maxbytes))
                        *count = inode->i_sb->s_maxbytes - *pos;
        } else {
+#ifdef CONFIG_BLOCK
                loff_t isize;
                if (bdev_read_only(I_BDEV(inode)))
                        return -EPERM;
@@ -2017,6 +2006,9 @@ inline int generic_write_checks(struct file *file, loff_t *pos, size_t *count, i
 
                if (*pos + *count > isize)
                        *count = isize - *pos;
+#else
+               return -EPERM;
+#endif
        }
        return 0;
 }
@@ -2049,15 +2041,14 @@ generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
         * Sync the fs metadata but not the minor inode changes and
         * of course not the data as we did direct DMA for the IO.
         * i_mutex is held, which protects generic_osync_inode() from
-        * livelocking.
+        * livelocking.  AIO O_DIRECT ops attempt to sync metadata here.
         */
-       if (written >= 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
+       if ((written >= 0 || written == -EIOCBQUEUED) &&
+           ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
                int err = generic_osync_inode(inode, mapping, OSYNC_METADATA);
                if (err < 0)
                        written = err;
        }
-       if (written == count && !is_sync_kiocb(iocb))
-               written = -EIOCBQUEUED;
        return written;
 }
 EXPORT_SYMBOL(generic_file_direct_write);
@@ -2069,7 +2060,7 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
 {
        struct file *file = iocb->ki_filp;
        struct address_space * mapping = file->f_mapping;
-       struct address_space_operations *a_ops = mapping->a_ops;
+       const struct address_space_operations *a_ops = mapping->a_ops;
        struct inode    *inode = mapping->host;
        long            status = 0;
        struct page     *page;
@@ -2104,27 +2095,39 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
                /* Limit the size of the copy to the caller's write size */
                bytes = min(bytes, count);
 
-               /*
-                * Limit the size of the copy to that of the current segment,
-                * because fault_in_pages_readable() doesn't know how to walk
-                * segments.
+               /* We only need to worry about prefaulting when writes are from
+                * user-space.  NFSd uses vfs_writev with several non-aligned
+                * segments in the vector, and limiting to one segment a time is
+                * a noticeable performance for re-write
                 */
-               bytes = min(bytes, cur_iov->iov_len - iov_base);
-
-               /*
-                * Bring in the user page that we will copy from _first_.
-                * Otherwise there's a nasty deadlock on copying from the
-                * same page as we're writing to, without it being marked
-                * up-to-date.
-                */
-               fault_in_pages_readable(buf, bytes);
+               if (!segment_eq(get_fs(), KERNEL_DS)) {
+                       /*
+                        * Limit the size of the copy to that of the current
+                        * segment, because fault_in_pages_readable() doesn't
+                        * know how to walk segments.
+                        */
+                       bytes = min(bytes, cur_iov->iov_len - iov_base);
 
+                       /*
+                        * Bring in the user page that we will copy from
+                        * _first_.  Otherwise there's a nasty deadlock on
+                        * copying from the same page as we're writing to,
+                        * without it being marked up-to-date.
+                        */
+                       fault_in_pages_readable(buf, bytes);
+               }
                page = __grab_cache_page(mapping,index,&cached_page,&lru_pvec);
                if (!page) {
                        status = -ENOMEM;
                        break;
                }
 
+               if (unlikely(bytes == 0)) {
+                       status = 0;
+                       copied = 0;
+                       goto zero_length_segment;
+               }
+
                status = a_ops->prepare_write(file, page, offset, offset+bytes);
                if (unlikely(status)) {
                        loff_t isize = i_size_read(inode);
@@ -2154,7 +2157,8 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
                        page_cache_release(page);
                        continue;
                }
-               if (likely(copied > 0)) {
+zero_length_segment:
+               if (likely(copied >= 0)) {
                        if (!status)
                                status = copied;
 
@@ -2223,30 +2227,14 @@ __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
        size_t ocount;          /* original count */
        size_t count;           /* after file limit checks */
        struct inode    *inode = mapping->host;
-       unsigned long   seg;
        loff_t          pos;
        ssize_t         written;
        ssize_t         err;
 
        ocount = 0;
-       for (seg = 0; seg < nr_segs; seg++) {
-               const struct iovec *iv = &iov[seg];
-
-               /*
-                * If any segment has a negative length, or the cumulative
-                * length ever wraps negative then return -EINVAL.
-                */
-               ocount += iv->iov_len;
-               if (unlikely((ssize_t)(ocount|iv->iov_len) < 0))
-                       return -EINVAL;
-               if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
-                       continue;
-               if (seg == 0)
-                       return -EFAULT;
-               nr_segs = seg;
-               ocount -= iv->iov_len;  /* This segment is no good */
-               break;
-       }
+       err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
+       if (err)
+               return err;
 
        count = ocount;
        pos = *ppos;
@@ -2264,7 +2252,7 @@ __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
        if (count == 0)
                goto out;
 
-       err = remove_suid(file->f_dentry);
+       err = remove_suid(file->f_path.dentry);
        if (err)
                goto out;
 
@@ -2272,8 +2260,11 @@ __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
 
        /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
        if (unlikely(file->f_flags & O_DIRECT)) {
-               written = generic_file_direct_write(iocb, iov,
-                               &nr_segs, pos, ppos, count, ocount);
+               loff_t endbyte;
+               ssize_t written_buffered;
+
+               written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
+                                                       ppos, count, ocount);
                if (written < 0 || written == count)
                        goto out;
                /*
@@ -2282,30 +2273,66 @@ __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
                 */
                pos += written;
                count -= written;
-       }
+               written_buffered = generic_file_buffered_write(iocb, iov,
+                                               nr_segs, pos, ppos, count,
+                                               written);
+               /*
+                * If generic_file_buffered_write() retuned a synchronous error
+                * then we want to return the number of bytes which were
+                * direct-written, or the error code if that was zero.  Note
+                * that this differs from normal direct-io semantics, which
+                * will return -EFOO even if some bytes were written.
+                */
+               if (written_buffered < 0) {
+                       err = written_buffered;
+                       goto out;
+               }
 
-       written = generic_file_buffered_write(iocb, iov, nr_segs,
-                       pos, ppos, count, written);
+               /*
+                * We need to ensure that the page cache pages are written to
+                * disk and invalidated to preserve the expected O_DIRECT
+                * semantics.
+                */
+               endbyte = pos + written_buffered - written - 1;
+               err = do_sync_mapping_range(file->f_mapping, pos, endbyte,
+                                           SYNC_FILE_RANGE_WAIT_BEFORE|
+                                           SYNC_FILE_RANGE_WRITE|
+                                           SYNC_FILE_RANGE_WAIT_AFTER);
+               if (err == 0) {
+                       written = written_buffered;
+                       invalidate_mapping_pages(mapping,
+                                                pos >> PAGE_CACHE_SHIFT,
+                                                endbyte >> PAGE_CACHE_SHIFT);
+               } else {
+                       /*
+                        * We don't know how much we wrote, so just return
+                        * the number of bytes which were direct-written
+                        */
+               }
+       } else {
+               written = generic_file_buffered_write(iocb, iov, nr_segs,
+                               pos, ppos, count, written);
+       }
 out:
        current->backing_dev_info = NULL;
        return written ? written : err;
 }
-EXPORT_SYMBOL(generic_file_aio_write_nolock);
 
-ssize_t
-generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
-                               unsigned long nr_segs, loff_t *ppos)
+ssize_t generic_file_aio_write_nolock(struct kiocb *iocb,
+               const struct iovec *iov, unsigned long nr_segs, loff_t pos)
 {
        struct file *file = iocb->ki_filp;
        struct address_space *mapping = file->f_mapping;
        struct inode *inode = mapping->host;
        ssize_t ret;
-       loff_t pos = *ppos;
 
-       ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs, ppos);
+       BUG_ON(iocb->ki_pos != pos);
+
+       ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
+                       &iocb->ki_pos);
 
        if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
-               int err;
+               ssize_t err;
 
                err = sync_page_range_nolock(inode, mapping, pos, ret);
                if (err < 0)
@@ -2313,51 +2340,21 @@ generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
        }
        return ret;
 }
+EXPORT_SYMBOL(generic_file_aio_write_nolock);
 
-static ssize_t
-__generic_file_write_nolock(struct file *file, const struct iovec *iov,
-                               unsigned long nr_segs, loff_t *ppos)
-{
-       struct kiocb kiocb;
-       ssize_t ret;
-
-       init_sync_kiocb(&kiocb, file);
-       ret = __generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
-       if (ret == -EIOCBQUEUED)
-               ret = wait_on_sync_kiocb(&kiocb);
-       return ret;
-}
-
-ssize_t
-generic_file_write_nolock(struct file *file, const struct iovec *iov,
-                               unsigned long nr_segs, loff_t *ppos)
-{
-       struct kiocb kiocb;
-       ssize_t ret;
-
-       init_sync_kiocb(&kiocb, file);
-       ret = generic_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
-       if (-EIOCBQUEUED == ret)
-               ret = wait_on_sync_kiocb(&kiocb);
-       return ret;
-}
-EXPORT_SYMBOL(generic_file_write_nolock);
-
-ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
-                              size_t count, loff_t pos)
+ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
+               unsigned long nr_segs, loff_t pos)
 {
        struct file *file = iocb->ki_filp;
        struct address_space *mapping = file->f_mapping;
        struct inode *inode = mapping->host;
        ssize_t ret;
-       struct iovec local_iov = { .iov_base = (void __user *)buf,
-                                       .iov_len = count };
 
        BUG_ON(iocb->ki_pos != pos);
 
        mutex_lock(&inode->i_mutex);
-       ret = __generic_file_aio_write_nolock(iocb, &local_iov, 1,
-                                               &iocb->ki_pos);
+       ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
+                       &iocb->ki_pos);
        mutex_unlock(&inode->i_mutex);
 
        if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
@@ -2371,66 +2368,6 @@ ssize_t generic_file_aio_write(struct kiocb *iocb, const char __user *buf,
 }
 EXPORT_SYMBOL(generic_file_aio_write);
 
-ssize_t generic_file_write(struct file *file, const char __user *buf,
-                          size_t count, loff_t *ppos)
-{
-       struct address_space *mapping = file->f_mapping;
-       struct inode *inode = mapping->host;
-       ssize_t ret;
-       struct iovec local_iov = { .iov_base = (void __user *)buf,
-                                       .iov_len = count };
-
-       mutex_lock(&inode->i_mutex);
-       ret = __generic_file_write_nolock(file, &local_iov, 1, ppos);
-       mutex_unlock(&inode->i_mutex);
-
-       if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
-               ssize_t err;
-
-               err = sync_page_range(inode, mapping, *ppos - ret, ret);
-               if (err < 0)
-                       ret = err;
-       }
-       return ret;
-}
-EXPORT_SYMBOL(generic_file_write);
-
-ssize_t generic_file_readv(struct file *filp, const struct iovec *iov,
-                       unsigned long nr_segs, loff_t *ppos)
-{
-       struct kiocb kiocb;
-       ssize_t ret;
-
-       init_sync_kiocb(&kiocb, filp);
-       ret = __generic_file_aio_read(&kiocb, iov, nr_segs, ppos);
-       if (-EIOCBQUEUED == ret)
-               ret = wait_on_sync_kiocb(&kiocb);
-       return ret;
-}
-EXPORT_SYMBOL(generic_file_readv);
-
-ssize_t generic_file_writev(struct file *file, const struct iovec *iov,
-                       unsigned long nr_segs, loff_t *ppos)
-{
-       struct address_space *mapping = file->f_mapping;
-       struct inode *inode = mapping->host;
-       ssize_t ret;
-
-       mutex_lock(&inode->i_mutex);
-       ret = __generic_file_write_nolock(file, iov, nr_segs, ppos);
-       mutex_unlock(&inode->i_mutex);
-
-       if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
-               int err;
-
-               err = sync_page_range(inode, mapping, *ppos - ret, ret);
-               if (err < 0)
-                       ret = err;
-       }
-       return ret;
-}
-EXPORT_SYMBOL(generic_file_writev);
-
 /*
  * Called under i_mutex for writes to S_ISREG files.   Returns -EIO if something
  * went wrong during pagecache shootdown.
@@ -2442,7 +2379,8 @@ generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
        struct file *file = iocb->ki_filp;
        struct address_space *mapping = file->f_mapping;
        ssize_t retval;
-       size_t write_len = 0;
+       size_t write_len;
+       pgoff_t end = 0; /* silence gcc */
 
        /*
         * If it's a write, unmap all mmappings of the file up-front.  This
@@ -2451,22 +2389,75 @@ generic_file_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
         */
        if (rw == WRITE) {
                write_len = iov_length(iov, nr_segs);
+               end = (offset + write_len - 1) >> PAGE_CACHE_SHIFT;
                if (mapping_mapped(mapping))
                        unmap_mapping_range(mapping, offset, write_len, 0);
        }
 
        retval = filemap_write_and_wait(mapping);
-       if (retval == 0) {
-               retval = mapping->a_ops->direct_IO(rw, iocb, iov,
-                                               offset, nr_segs);
-               if (rw == WRITE && mapping->nrpages) {
-                       pgoff_t end = (offset + write_len - 1)
-                                               >> PAGE_CACHE_SHIFT;
-                       int err = invalidate_inode_pages2_range(mapping,
+       if (retval)
+               goto out;
+
+       /*
+        * After a write we want buffered reads to be sure to go to disk to get
+        * the new data.  We invalidate clean cached page from the region we're
+        * about to write.  We do this *before* the write so that we can return
+        * -EIO without clobbering -EIOCBQUEUED from ->direct_IO().
+        */
+       if (rw == WRITE && mapping->nrpages) {
+               retval = invalidate_inode_pages2_range(mapping,
                                        offset >> PAGE_CACHE_SHIFT, end);
-                       if (err)
-                               retval = err;
-               }
+               if (retval)
+                       goto out;
+       }
+
+       retval = mapping->a_ops->direct_IO(rw, iocb, iov, offset, nr_segs);
+       if (retval)
+               goto out;
+
+       /*
+        * Finally, try again to invalidate clean pages which might have been
+        * faulted in by get_user_pages() if the source of the write was an
+        * mmap()ed region of the file we're writing.  That's a pretty crazy
+        * thing to do, so we don't support it 100%.  If this invalidation
+        * fails and we have -EIOCBQUEUED we ignore the failure.
+        */
+       if (rw == WRITE && mapping->nrpages) {
+               int err = invalidate_inode_pages2_range(mapping,
+                                             offset >> PAGE_CACHE_SHIFT, end);
+               if (err && retval >= 0)
+                       retval = err;
        }
+out:
        return retval;
 }
+
+/**
+ * try_to_release_page() - release old fs-specific metadata on a page
+ *
+ * @page: the page which the kernel is trying to free
+ * @gfp_mask: memory allocation flags (and I/O mode)
+ *
+ * The address_space is to try to release any data against the page
+ * (presumably at page->private).  If the release was successful, return `1'.
+ * Otherwise return zero.
+ *
+ * The @gfp_mask argument specifies whether I/O may be performed to release
+ * this page (__GFP_IO), and whether the call may block (__GFP_WAIT).
+ *
+ * NOTE: @gfp_mask may go away, and this function may become non-blocking.
+ */
+int try_to_release_page(struct page *page, gfp_t gfp_mask)
+{
+       struct address_space * const mapping = page->mapping;
+
+       BUG_ON(!PageLocked(page));
+       if (PageWriteback(page))
+               return 0;
+
+       if (mapping && mapping->a_ops->releasepage)
+               return mapping->a_ops->releasepage(page, gfp_mask);
+       return try_to_free_buffers(page);
+}
+
+EXPORT_SYMBOL(try_to_release_page);