generic-ipi: Optimize accesses by using DEFINE_PER_CPU_SHARED_ALIGNED for IPI data
[linux-2.6.git] / mm / filemap.c
index 8275364..96ac6b0 100644 (file)
 /*
  * FIXME: remove all knowledge of the buffer layer from the core VM
  */
-#include <linux/buffer_head.h> /* for generic_osync_inode */
+#include <linux/buffer_head.h> /* for try_to_free_buffers */
 
 #include <asm/mman.h>
 
-
 /*
  * Shared mappings implemented 30.11.1994. It's not fully working yet,
  * though.
@@ -59,7 +58,7 @@
 /*
  * Lock ordering:
  *
- *  ->i_mmap_lock              (vmtruncate)
+ *  ->i_mmap_lock              (truncate_pagecache)
  *    ->private_lock           (__free_pte->__set_page_dirty_buffers)
  *      ->swap_lock            (exclusive_swap_page, others)
  *        ->mapping->tree_lock
  *
  *  ->task->proc_lock
  *    ->dcache_lock            (proc_pid_lookup)
+ *
+ *  (code doesn't rely on that order, so you could switch it around)
+ *  ->tasklist_lock             (memory_failure, collect_procs_ao)
+ *    ->i_mmap_lock
  */
 
 /*
@@ -120,6 +123,8 @@ void __remove_from_page_cache(struct page *page)
        page->mapping = NULL;
        mapping->nrpages--;
        __dec_zone_page_state(page, NR_FILE_PAGES);
+       if (PageSwapBacked(page))
+               __dec_zone_page_state(page, NR_SHMEM);
        BUG_ON(page_mapped(page));
 
        /*
@@ -255,27 +260,27 @@ int filemap_flush(struct address_space *mapping)
 EXPORT_SYMBOL(filemap_flush);
 
 /**
- * wait_on_page_writeback_range - wait for writeback to complete
- * @mapping:   target address_space
- * @start:     beginning page index
- * @end:       ending page index
+ * filemap_fdatawait_range - wait for writeback to complete
+ * @mapping:           address space structure to wait for
+ * @start_byte:                offset in bytes where the range starts
+ * @end_byte:          offset in bytes where the range ends (inclusive)
  *
- * Wait for writeback to complete against pages indexed by start->end
- * inclusive
+ * Walk the list of under-writeback pages of the given address space
+ * in the given range and wait for all of them.
  */
-int wait_on_page_writeback_range(struct address_space *mapping,
-                               pgoff_t start, pgoff_t end)
+int filemap_fdatawait_range(struct address_space *mapping, loff_t start_byte,
+                           loff_t end_byte)
 {
+       pgoff_t index = start_byte >> PAGE_CACHE_SHIFT;
+       pgoff_t end = end_byte >> PAGE_CACHE_SHIFT;
        struct pagevec pvec;
        int nr_pages;
        int ret = 0;
-       pgoff_t index;
 
-       if (end < start)
+       if (end_byte < start_byte)
                return 0;
 
        pagevec_init(&pvec, 0);
-       index = start;
        while ((index <= end) &&
                        (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
                        PAGECACHE_TAG_WRITEBACK,
@@ -305,70 +310,7 @@ int wait_on_page_writeback_range(struct address_space *mapping,
 
        return ret;
 }
-
-/**
- * sync_page_range - write and wait on all pages in the passed range
- * @inode:     target inode
- * @mapping:   target address_space
- * @pos:       beginning offset in pages to write
- * @count:     number of bytes to write
- *
- * Write and wait upon all the pages in the passed range.  This is a "data
- * integrity" operation.  It waits upon in-flight writeout before starting and
- * waiting upon new writeout.  If there was an IO error, return it.
- *
- * We need to re-take i_mutex during the generic_osync_inode list walk because
- * it is otherwise livelockable.
- */
-int sync_page_range(struct inode *inode, struct address_space *mapping,
-                       loff_t pos, loff_t count)
-{
-       pgoff_t start = pos >> PAGE_CACHE_SHIFT;
-       pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
-       int ret;
-
-       if (!mapping_cap_writeback_dirty(mapping) || !count)
-               return 0;
-       ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
-       if (ret == 0) {
-               mutex_lock(&inode->i_mutex);
-               ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
-               mutex_unlock(&inode->i_mutex);
-       }
-       if (ret == 0)
-               ret = wait_on_page_writeback_range(mapping, start, end);
-       return ret;
-}
-EXPORT_SYMBOL(sync_page_range);
-
-/**
- * sync_page_range_nolock - write & wait on all pages in the passed range without locking
- * @inode:     target inode
- * @mapping:   target address_space
- * @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
- * as it forces O_SYNC writers to different parts of the same file
- * to be serialised right until io completion.
- */
-int sync_page_range_nolock(struct inode *inode, struct address_space *mapping,
-                          loff_t pos, loff_t count)
-{
-       pgoff_t start = pos >> PAGE_CACHE_SHIFT;
-       pgoff_t end = (pos + count - 1) >> PAGE_CACHE_SHIFT;
-       int ret;
-
-       if (!mapping_cap_writeback_dirty(mapping) || !count)
-               return 0;
-       ret = filemap_fdatawrite_range(mapping, pos, pos + count - 1);
-       if (ret == 0)
-               ret = generic_osync_inode(inode, mapping, OSYNC_METADATA);
-       if (ret == 0)
-               ret = wait_on_page_writeback_range(mapping, start, end);
-       return ret;
-}
-EXPORT_SYMBOL(sync_page_range_nolock);
+EXPORT_SYMBOL(filemap_fdatawait_range);
 
 /**
  * filemap_fdatawait - wait for all under-writeback pages to complete
@@ -384,8 +326,7 @@ int filemap_fdatawait(struct address_space *mapping)
        if (i_size == 0)
                return 0;
 
-       return wait_on_page_writeback_range(mapping, 0,
-                               (i_size - 1) >> PAGE_CACHE_SHIFT);
+       return filemap_fdatawait_range(mapping, 0, i_size - 1);
 }
 EXPORT_SYMBOL(filemap_fdatawait);
 
@@ -432,9 +373,8 @@ int filemap_write_and_wait_range(struct address_space *mapping,
                                                 WB_SYNC_ALL);
                /* See comment of filemap_write_and_wait() */
                if (err != -EIO) {
-                       int err2 = wait_on_page_writeback_range(mapping,
-                                               lstart >> PAGE_CACHE_SHIFT,
-                                               lend >> PAGE_CACHE_SHIFT);
+                       int err2 = filemap_fdatawait_range(mapping,
+                                               lstart, lend);
                        if (!err)
                                err = err2;
                }
@@ -476,6 +416,8 @@ int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
                if (likely(!error)) {
                        mapping->nrpages++;
                        __inc_zone_page_state(page, NR_FILE_PAGES);
+                       if (PageSwapBacked(page))
+                               __inc_zone_page_state(page, NR_SHMEM);
                        spin_unlock_irq(&mapping->tree_lock);
                } else {
                        page->mapping = NULL;
@@ -521,7 +463,7 @@ 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, gfp, 0);
+               return alloc_pages_exact_node(n, gfp, 0);
        }
        return alloc_pages(gfp, 0);
 }
@@ -1004,9 +946,6 @@ EXPORT_SYMBOL(grab_cache_page_nowait);
 static void shrink_readahead_size_eio(struct file *filp,
                                        struct file_ra_state *ra)
 {
-       if (!ra->ra_pages)
-               return;
-
        ra->ra_pages /= 4;
 }
 
@@ -1472,8 +1411,10 @@ static void do_sync_mmap_readahead(struct vm_area_struct *vma,
        if (VM_RandomReadHint(vma))
                return;
 
-       if (VM_SequentialReadHint(vma)) {
-               page_cache_sync_readahead(mapping, ra, file, offset, 1);
+       if (VM_SequentialReadHint(vma) ||
+                       offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
+               page_cache_sync_readahead(mapping, ra, file, offset,
+                                         ra->ra_pages);
                return;
        }
 
@@ -1487,13 +1428,15 @@ static void do_sync_mmap_readahead(struct vm_area_struct *vma,
        if (ra->mmap_miss > MMAP_LOTSAMISS)
                return;
 
+       /*
+        * mmap read-around
+        */
        ra_pages = max_sane_readahead(ra->ra_pages);
        if (ra_pages) {
-               pgoff_t start = 0;
-
-               if (offset > ra_pages / 2)
-                       start = offset - ra_pages / 2;
-               do_page_cache_readahead(mapping, file, start, ra_pages);
+               ra->start = max_t(long, 0, offset - ra_pages/2);
+               ra->size = ra_pages;
+               ra->async_size = 0;
+               ra_submit(ra, mapping, file);
        }
 }
 
@@ -1515,7 +1458,8 @@ static void do_async_mmap_readahead(struct vm_area_struct *vma,
        if (ra->mmap_miss > 0)
                ra->mmap_miss--;
        if (PageReadahead(page))
-               page_cache_async_readahead(mapping, ra, file, page, offset, 1);
+               page_cache_async_readahead(mapping, ra, file,
+                                          page, offset, ra->ra_pages);
 }
 
 /**
@@ -1646,7 +1590,7 @@ page_not_uptodate:
 }
 EXPORT_SYMBOL(filemap_fault);
 
-struct vm_operations_struct generic_file_vm_ops = {
+const struct vm_operations_struct generic_file_vm_ops = {
        .fault          = filemap_fault,
 };
 
@@ -1879,7 +1823,7 @@ static size_t __iovec_copy_from_user_inatomic(char *vaddr,
 
 /*
  * Copy as much as we can into the page and return the number of bytes which
- * were sucessfully copied.  If a fault is encountered then return the number of
+ * were successfully copied.  If a fault is encountered then return the number of
  * bytes which were copied.
  */
 size_t iov_iter_copy_from_user_atomic(struct page *page,
@@ -2165,20 +2109,7 @@ generic_file_direct_write(struct kiocb *iocb, const struct iovec *iov,
                }
                *ppos = end;
        }
-
-       /*
-        * 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.  AIO O_DIRECT ops attempt to sync metadata here.
-        */
 out:
-       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;
-       }
        return written;
 }
 EXPORT_SYMBOL(generic_file_direct_write);
@@ -2270,6 +2201,7 @@ again:
                pagefault_enable();
                flush_dcache_page(page);
 
+               mark_page_accessed(page);
                status = a_ops->write_end(file, mapping, pos, bytes, copied,
                                                page, fsdata);
                if (unlikely(status < 0))
@@ -2308,9 +2240,6 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
                size_t count, ssize_t written)
 {
        struct file *file = iocb->ki_filp;
-       struct address_space *mapping = file->f_mapping;
-       const struct address_space_operations *a_ops = mapping->a_ops;
-       struct inode *inode = mapping->host;
        ssize_t status;
        struct iov_iter i;
 
@@ -2320,34 +2249,33 @@ generic_file_buffered_write(struct kiocb *iocb, const struct iovec *iov,
        if (likely(status >= 0)) {
                written += status;
                *ppos = pos + status;
-
-               /*
-                * For now, when the user asks for O_SYNC, we'll actually give
-                * O_DSYNC
-                */
-               if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
-                       if (!a_ops->writepage || !is_sync_kiocb(iocb))
-                               status = generic_osync_inode(inode, mapping,
-                                               OSYNC_METADATA|OSYNC_DATA);
-               }
        }
        
-       /*
-        * If we get here for O_DIRECT writes then we must have fallen through
-        * to buffered writes (block instantiation inside i_size).  So we sync
-        * the file data here, to try to honour O_DIRECT expectations.
-        */
-       if (unlikely(file->f_flags & O_DIRECT) && written)
-               status = filemap_write_and_wait_range(mapping,
-                                       pos, pos + written - 1);
-
        return written ? written : status;
 }
 EXPORT_SYMBOL(generic_file_buffered_write);
 
-static ssize_t
-__generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
-                               unsigned long nr_segs, loff_t *ppos)
+/**
+ * __generic_file_aio_write - write data to a file
+ * @iocb:      IO state structure (file, offset, etc.)
+ * @iov:       vector with data to write
+ * @nr_segs:   number of segments in the vector
+ * @ppos:      position where to write
+ *
+ * This function does all the work needed for actually writing data to a
+ * file. It does all basic checks, removes SUID from the file, updates
+ * modification times and calls proper subroutines depending on whether we
+ * do direct IO or a standard buffered write.
+ *
+ * It expects i_mutex to be grabbed unless we work on a block device or similar
+ * object which does not need locking at all.
+ *
+ * This function does *not* take care of syncing data in case of O_SYNC write.
+ * A caller has to handle it. This is mainly due to the fact that we want to
+ * avoid syncing under i_mutex.
+ */
+ssize_t __generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
+                                unsigned long nr_segs, loff_t *ppos)
 {
        struct file *file = iocb->ki_filp;
        struct address_space * mapping = file->f_mapping;
@@ -2421,10 +2349,7 @@ __generic_file_aio_write_nolock(struct kiocb *iocb, const struct iovec *iov,
                 * 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);
+               err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
                if (err == 0) {
                        written = written_buffered;
                        invalidate_mapping_pages(mapping,
@@ -2444,51 +2369,37 @@ out:
        current->backing_dev_info = NULL;
        return written ? written : err;
 }
+EXPORT_SYMBOL(__generic_file_aio_write);
 
-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;
-
-       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))) {
-               ssize_t err;
-
-               err = sync_page_range_nolock(inode, mapping, pos, ret);
-               if (err < 0)
-                       ret = err;
-       }
-       return ret;
-}
-EXPORT_SYMBOL(generic_file_aio_write_nolock);
-
+/**
+ * generic_file_aio_write - write data to a file
+ * @iocb:      IO state structure
+ * @iov:       vector with data to write
+ * @nr_segs:   number of segments in the vector
+ * @pos:       position in file where to write
+ *
+ * This is a wrapper around __generic_file_aio_write() to be used by most
+ * filesystems. It takes care of syncing the file in case of O_SYNC file
+ * and acquires i_mutex as needed.
+ */
 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;
+       struct inode *inode = file->f_mapping->host;
        ssize_t ret;
 
        BUG_ON(iocb->ki_pos != pos);
 
        mutex_lock(&inode->i_mutex);
-       ret = __generic_file_aio_write_nolock(iocb, iov, nr_segs,
-                       &iocb->ki_pos);
+       ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
        mutex_unlock(&inode->i_mutex);
 
-       if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
+       if (ret > 0 || ret == -EIOCBQUEUED) {
                ssize_t err;
 
-               err = sync_page_range(inode, mapping, pos, ret);
-               if (err < 0)
+               err = generic_write_sync(file, pos, ret);
+               if (err < 0 && ret > 0)
                        ret = err;
        }
        return ret;