memcg: fix vmscan count in small memcgs
[linux-2.6.git] / mm / filemap.c
index d8b34d1..867d402 100644 (file)
@@ -34,6 +34,7 @@
 #include <linux/hardirq.h> /* for BUG_ON(!in_atomic()) only */
 #include <linux/memcontrol.h>
 #include <linux/mm_inline.h> /* for page_is_file_cache() */
+#include <linux/cleancache.h>
 #include "internal.h"
 
 /*
 /*
  * Lock ordering:
  *
- *  ->i_mmap_lock              (truncate_pagecache)
+ *  ->i_mmap_mutex             (truncate_pagecache)
  *    ->private_lock           (__free_pte->__set_page_dirty_buffers)
  *      ->swap_lock            (exclusive_swap_page, others)
  *        ->mapping->tree_lock
  *
  *  ->i_mutex
- *    ->i_mmap_lock            (truncate->unmap_mapping_range)
+ *    ->i_mmap_mutex           (truncate->unmap_mapping_range)
  *
  *  ->mmap_sem
- *    ->i_mmap_lock
+ *    ->i_mmap_mutex
  *      ->page_table_lock or pte_lock  (various, mainly in memory.c)
  *        ->mapping->tree_lock (arch-dependent flush_dcache_mmap_lock)
  *
  *  ->i_mutex                  (generic_file_buffered_write)
  *    ->mmap_sem               (fault_in_pages_readable->do_page_fault)
  *
- *  ->i_mutex
- *    ->i_alloc_sem             (various)
- *
- *  inode_wb_list_lock
+ *  bdi->wb.list_lock
  *    sb_lock                  (fs/fs-writeback.c)
  *    ->mapping->tree_lock     (__sync_single_inode)
  *
- *  ->i_mmap_lock
+ *  ->i_mmap_mutex
  *    ->anon_vma.lock          (vma_adjust)
  *
  *  ->anon_vma.lock
  *    ->zone.lru_lock          (check_pte_range->isolate_lru_page)
  *    ->private_lock           (page_remove_rmap->set_page_dirty)
  *    ->tree_lock              (page_remove_rmap->set_page_dirty)
- *    inode_wb_list_lock       (page_remove_rmap->set_page_dirty)
+ *    bdi.wb->list_lock                (page_remove_rmap->set_page_dirty)
  *    ->inode->i_lock          (page_remove_rmap->set_page_dirty)
- *    inode_wb_list_lock       (zap_pte_range->set_page_dirty)
+ *    bdi.wb->list_lock                (zap_pte_range->set_page_dirty)
  *    ->inode->i_lock          (zap_pte_range->set_page_dirty)
  *    ->private_lock           (zap_pte_range->__set_page_dirty_buffers)
  *
  *  (code doesn't rely on that order, so you could switch it around)
  *  ->tasklist_lock             (memory_failure, collect_procs_ao)
- *    ->i_mmap_lock
+ *    ->i_mmap_mutex
  */
 
 /*
@@ -118,8 +116,19 @@ void __delete_from_page_cache(struct page *page)
 {
        struct address_space *mapping = page->mapping;
 
+       /*
+        * if we're uptodate, flush out into the cleancache, otherwise
+        * invalidate any existing cleancache entries.  We can't leave
+        * stale data around in the cleancache once our page is gone
+        */
+       if (PageUptodate(page) && PageMappedToDisk(page))
+               cleancache_put_page(page);
+       else
+               cleancache_flush_page(mapping, page);
+
        radix_tree_delete(&mapping->page_tree, page->index);
        page->mapping = NULL;
+       /* Leave page->index set: truncation lookup relies upon it */
        mapping->nrpages--;
        __dec_zone_page_state(page, NR_FILE_PAGES);
        if (PageSwapBacked(page))
@@ -166,45 +175,15 @@ void delete_from_page_cache(struct page *page)
 }
 EXPORT_SYMBOL(delete_from_page_cache);
 
-static int sync_page(void *word)
+static int sleep_on_page(void *word)
 {
-       struct address_space *mapping;
-       struct page *page;
-
-       page = container_of((unsigned long *)word, struct page, flags);
-
-       /*
-        * page_mapping() is being called without PG_locked held.
-        * Some knowledge of the state and use of the page is used to
-        * reduce the requirements down to a memory barrier.
-        * The danger here is of a stale page_mapping() return value
-        * indicating a struct address_space different from the one it's
-        * associated with when it is associated with one.
-        * After smp_mb(), it's either the correct page_mapping() for
-        * the page, or an old page_mapping() and the page's own
-        * page_mapping() has gone NULL.
-        * The ->sync_page() address_space operation must tolerate
-        * page_mapping() going NULL. By an amazing coincidence,
-        * this comes about because none of the users of the page
-        * in the ->sync_page() methods make essential use of the
-        * page_mapping(), merely passing the page down to the backing
-        * device's unplug functions when it's non-NULL, which in turn
-        * ignore it for all cases but swap, where only page_private(page) is
-        * of interest. When page_mapping() does go NULL, the entire
-        * call stack gracefully ignores the page and returns.
-        * -- wli
-        */
-       smp_mb();
-       mapping = page_mapping(page);
-       if (mapping && mapping->a_ops && mapping->a_ops->sync_page)
-               mapping->a_ops->sync_page(page);
        io_schedule();
        return 0;
 }
 
-static int sync_page_killable(void *word)
+static int sleep_on_page_killable(void *word)
 {
-       sync_page(word);
+       sleep_on_page(word);
        return fatal_signal_pending(current) ? -EINTR : 0;
 }
 
@@ -505,6 +484,7 @@ int add_to_page_cache_locked(struct page *page, struct address_space *mapping,
                        spin_unlock_irq(&mapping->tree_lock);
                } else {
                        page->mapping = NULL;
+                       /* Leave page->index set: truncation relies upon it */
                        spin_unlock_irq(&mapping->tree_lock);
                        mem_cgroup_uncharge_cache_page(page);
                        page_cache_release(page);
@@ -560,12 +540,6 @@ struct page *__page_cache_alloc(gfp_t gfp)
 EXPORT_SYMBOL(__page_cache_alloc);
 #endif
 
-static int __sleep_on_page_lock(void *word)
-{
-       io_schedule();
-       return 0;
-}
-
 /*
  * In order to wait for pages to become available there must be
  * waitqueues associated with pages. By using a hash table of
@@ -593,11 +567,22 @@ void wait_on_page_bit(struct page *page, int bit_nr)
        DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
 
        if (test_bit(bit_nr, &page->flags))
-               __wait_on_bit(page_waitqueue(page), &wait, sync_page,
+               __wait_on_bit(page_waitqueue(page), &wait, sleep_on_page,
                                                        TASK_UNINTERRUPTIBLE);
 }
 EXPORT_SYMBOL(wait_on_page_bit);
 
+int wait_on_page_bit_killable(struct page *page, int bit_nr)
+{
+       DEFINE_WAIT_BIT(wait, &page->flags, bit_nr);
+
+       if (!test_bit(bit_nr, &page->flags))
+               return 0;
+
+       return __wait_on_bit(page_waitqueue(page), &wait,
+                            sleep_on_page_killable, TASK_KILLABLE);
+}
+
 /**
  * add_page_wait_queue - Add an arbitrary waiter to a page's wait queue
  * @page: Page defining the wait queue of interest
@@ -657,17 +642,12 @@ EXPORT_SYMBOL(end_page_writeback);
 /**
  * __lock_page - get a lock on the page, assuming we need to sleep to get it
  * @page: the page to lock
- *
- * Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary.  If some
- * random driver's requestfn sets TASK_RUNNING, we could busywait.  However
- * chances are that on the second loop, the block layer's plug list is empty,
- * so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
  */
 void __lock_page(struct page *page)
 {
        DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 
-       __wait_on_bit_lock(page_waitqueue(page), &wait, sync_page,
+       __wait_on_bit_lock(page_waitqueue(page), &wait, sleep_on_page,
                                                        TASK_UNINTERRUPTIBLE);
 }
 EXPORT_SYMBOL(__lock_page);
@@ -677,36 +657,39 @@ int __lock_page_killable(struct page *page)
        DEFINE_WAIT_BIT(wait, &page->flags, PG_locked);
 
        return __wait_on_bit_lock(page_waitqueue(page), &wait,
-                                       sync_page_killable, TASK_KILLABLE);
+                                       sleep_on_page_killable, TASK_KILLABLE);
 }
 EXPORT_SYMBOL_GPL(__lock_page_killable);
 
-/**
- * __lock_page_nosync - get a lock on the page, without calling sync_page()
- * @page: the page to lock
- *
- * Variant of lock_page that does not require the caller to hold a reference
- * on the page's mapping.
- */
-void __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);
-}
-
 int __lock_page_or_retry(struct page *page, struct mm_struct *mm,
                         unsigned int flags)
 {
-       if (!(flags & FAULT_FLAG_ALLOW_RETRY)) {
-               __lock_page(page);
-               return 1;
-       } else {
-               if (!(flags & FAULT_FLAG_RETRY_NOWAIT)) {
-                       up_read(&mm->mmap_sem);
+       if (flags & FAULT_FLAG_ALLOW_RETRY) {
+               /*
+                * CAUTION! In this case, mmap_sem is not released
+                * even though return 0.
+                */
+               if (flags & FAULT_FLAG_RETRY_NOWAIT)
+                       return 0;
+
+               up_read(&mm->mmap_sem);
+               if (flags & FAULT_FLAG_KILLABLE)
+                       wait_on_page_locked_killable(page);
+               else
                        wait_on_page_locked(page);
-               }
                return 0;
+       } else {
+               if (flags & FAULT_FLAG_KILLABLE) {
+                       int ret;
+
+                       ret = __lock_page_killable(page);
+                       if (ret) {
+                               up_read(&mm->mmap_sem);
+                               return 0;
+                       }
+               } else
+                       __lock_page(page);
+               return 1;
        }
 }
 
@@ -1409,12 +1392,15 @@ generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
        unsigned long seg = 0;
        size_t count;
        loff_t *ppos = &iocb->ki_pos;
+       struct blk_plug plug;
 
        count = 0;
        retval = generic_segment_checks(iov, &nr_segs, &count, VERIFY_WRITE);
        if (retval)
                return retval;
 
+       blk_start_plug(&plug);
+
        /* coalesce the iovecs and go direct-to-BIO for O_DIRECT */
        if (filp->f_flags & O_DIRECT) {
                loff_t size;
@@ -1487,6 +1473,7 @@ generic_file_aio_read(struct kiocb *iocb, const struct iovec *iov,
                        break;
        }
 out:
+       blk_finish_plug(&plug);
        return retval;
 }
 EXPORT_SYMBOL(generic_file_aio_read);
@@ -1579,15 +1566,17 @@ static void do_sync_mmap_readahead(struct vm_area_struct *vma,
        /* If we don't want any read-ahead, don't bother */
        if (VM_RandomReadHint(vma))
                return;
+       if (!ra->ra_pages)
+               return;
 
-       if (VM_SequentialReadHint(vma) ||
-                       offset - 1 == (ra->prev_pos >> PAGE_CACHE_SHIFT)) {
+       if (VM_SequentialReadHint(vma)) {
                page_cache_sync_readahead(mapping, ra, file, offset,
                                          ra->ra_pages);
                return;
        }
 
-       if (ra->mmap_miss < INT_MAX)
+       /* Avoid banging the cache line if not needed */
+       if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
                ra->mmap_miss++;
 
        /*
@@ -1601,12 +1590,10 @@ static void do_sync_mmap_readahead(struct vm_area_struct *vma,
         * mmap read-around
         */
        ra_pages = max_sane_readahead(ra->ra_pages);
-       if (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);
-       }
+       ra->start = max_t(long, 0, offset - ra_pages / 2);
+       ra->size = ra_pages;
+       ra->async_size = ra_pages / 4;
+       ra_submit(ra, mapping, file);
 }
 
 /*
@@ -1673,6 +1660,7 @@ int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
                /* No page in the page cache at all */
                do_sync_mmap_readahead(vma, ra, file, offset);
                count_vm_event(PGMAJFAULT);
+               mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
                ret = VM_FAULT_MAJOR;
 retry_find:
                page = find_get_page(mapping, offset);
@@ -1711,7 +1699,6 @@ retry_find:
                return VM_FAULT_SIGBUS;
        }
 
-       ra->prev_pos = (loff_t)offset << PAGE_CACHE_SHIFT;
        vmf->page = page;
        return ret | VM_FAULT_LOCKED;
 
@@ -1807,7 +1794,7 @@ EXPORT_SYMBOL(generic_file_readonly_mmap);
 
 static struct page *__read_cache_page(struct address_space *mapping,
                                pgoff_t index,
-                               int (*filler)(void *,struct page*),
+                               int (*filler)(void *, struct page *),
                                void *data,
                                gfp_t gfp)
 {
@@ -1838,7 +1825,7 @@ repeat:
 
 static struct page *do_read_cache_page(struct address_space *mapping,
                                pgoff_t index,
-                               int (*filler)(void *,struct page*),
+                               int (*filler)(void *, struct page *),
                                void *data,
                                gfp_t gfp)
 
@@ -1878,7 +1865,7 @@ out:
  * @mapping:   the page's address_space
  * @index:     the page index
  * @filler:    function to perform the read
- * @data:      destination for read data
+ * @data:      first arg to filler(data, page) function, often left as NULL
  *
  * Same as read_cache_page, but don't wait for page to become unlocked
  * after submitting it to the filler.
@@ -1890,7 +1877,7 @@ out:
  */
 struct page *read_cache_page_async(struct address_space *mapping,
                                pgoff_t index,
-                               int (*filler)(void *,struct page*),
+                               int (*filler)(void *, struct page *),
                                void *data)
 {
        return do_read_cache_page(mapping, index, filler, data, mapping_gfp_mask(mapping));
@@ -1938,7 +1925,7 @@ EXPORT_SYMBOL(read_cache_page_gfp);
  * @mapping:   the page's address_space
  * @index:     the page index
  * @filler:    function to perform the read
- * @data:      destination for read data
+ * @data:      first arg to filler(data, page) function, often left as NULL
  *
  * 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.
@@ -1947,7 +1934,7 @@ EXPORT_SYMBOL(read_cache_page_gfp);
  */
 struct page *read_cache_page(struct address_space *mapping,
                                pgoff_t index,
-                               int (*filler)(void *,struct page*),
+                               int (*filler)(void *, struct page *),
                                void *data)
 {
        return wait_on_page_read(read_cache_page_async(mapping, index, filler, data));
@@ -1994,16 +1981,26 @@ static int __remove_suid(struct dentry *dentry, int kill)
 int file_remove_suid(struct file *file)
 {
        struct dentry *dentry = file->f_path.dentry;
-       int killsuid = should_remove_suid(dentry);
-       int killpriv = security_inode_need_killpriv(dentry);
+       struct inode *inode = dentry->d_inode;
+       int killsuid;
+       int killpriv;
        int error = 0;
 
+       /* Fast path for nothing security related */
+       if (IS_NOSEC(inode))
+               return 0;
+
+       killsuid = should_remove_suid(dentry);
+       killpriv = security_inode_need_killpriv(dentry);
+
        if (killpriv < 0)
                return killpriv;
        if (killpriv)
                error = security_inode_killpriv(dentry);
        if (!error && killsuid)
                error = __remove_suid(dentry, killsuid);
+       if (!error && (inode->i_sb->s_flags & MS_NOSEC))
+               inode->i_flags |= S_NOSEC;
 
        return error;
 }
@@ -2339,7 +2336,7 @@ struct page *grab_cache_page_write_begin(struct address_space *mapping,
 repeat:
        page = find_lock_page(mapping, index);
        if (page)
-               return page;
+               goto found;
 
        page = __page_cache_alloc(mapping_gfp_mask(mapping) & ~gfp_notmask);
        if (!page)
@@ -2352,6 +2349,8 @@ repeat:
                        goto repeat;
                return NULL;
        }
+found:
+       wait_on_page_writeback(page);
        return page;
 }
 EXPORT_SYMBOL(grab_cache_page_write_begin);
@@ -2598,11 +2597,13 @@ ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
 {
        struct file *file = iocb->ki_filp;
        struct inode *inode = file->f_mapping->host;
+       struct blk_plug plug;
        ssize_t ret;
 
        BUG_ON(iocb->ki_pos != pos);
 
        mutex_lock(&inode->i_mutex);
+       blk_start_plug(&plug);
        ret = __generic_file_aio_write(iocb, iov, nr_segs, &iocb->ki_pos);
        mutex_unlock(&inode->i_mutex);
 
@@ -2613,6 +2614,7 @@ ssize_t generic_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
                if (err < 0 && ret > 0)
                        ret = err;
        }
+       blk_finish_plug(&plug);
        return ret;
 }
 EXPORT_SYMBOL(generic_file_aio_write);