writeback: limit write_cache_pages integrity scanning to current EOF
[safe/jmp/linux-2.6] / mm / page-writeback.c
index cbd4cba..bbd396a 100644 (file)
@@ -58,7 +58,7 @@ static inline long sync_writeback_pages(unsigned long dirtied)
 /* The following parameters are exported via /proc/sys/vm */
 
 /*
- * Start background writeback (via pdflush) at this percentage
+ * Start background writeback (via writeback threads) at this percentage
  */
 int dirty_background_ratio = 10;
 
@@ -158,37 +158,37 @@ static void update_completion_period(void)
 }
 
 int dirty_background_ratio_handler(struct ctl_table *table, int write,
-               struct file *filp, void __user *buffer, size_t *lenp,
+               void __user *buffer, size_t *lenp,
                loff_t *ppos)
 {
        int ret;
 
-       ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
+       ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
        if (ret == 0 && write)
                dirty_background_bytes = 0;
        return ret;
 }
 
 int dirty_background_bytes_handler(struct ctl_table *table, int write,
-               struct file *filp, void __user *buffer, size_t *lenp,
+               void __user *buffer, size_t *lenp,
                loff_t *ppos)
 {
        int ret;
 
-       ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
+       ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
        if (ret == 0 && write)
                dirty_background_ratio = 0;
        return ret;
 }
 
 int dirty_ratio_handler(struct ctl_table *table, int write,
-               struct file *filp, void __user *buffer, size_t *lenp,
+               void __user *buffer, size_t *lenp,
                loff_t *ppos)
 {
        int old_ratio = vm_dirty_ratio;
        int ret;
 
-       ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
+       ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
        if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
                update_completion_period();
                vm_dirty_bytes = 0;
@@ -198,13 +198,13 @@ int dirty_ratio_handler(struct ctl_table *table, int write,
 
 
 int dirty_bytes_handler(struct ctl_table *table, int write,
-               struct file *filp, void __user *buffer, size_t *lenp,
+               void __user *buffer, size_t *lenp,
                loff_t *ppos)
 {
        unsigned long old_bytes = vm_dirty_bytes;
        int ret;
 
-       ret = proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos);
+       ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
        if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
                update_completion_period();
                vm_dirty_ratio = 0;
@@ -477,8 +477,8 @@ get_dirty_limits(unsigned long *pbackground, unsigned long *pdirty,
  * balance_dirty_pages() must be called by processes which are generating dirty
  * data.  It looks at the number of dirty pages in the machine and will force
  * the caller to perform writeback if the system is over `vm_dirty_ratio'.
- * If we're over `background_thresh' then pdflush is woken to perform some
- * writeout.
+ * If we're over `background_thresh' then the writeback threads are woken to
+ * perform some writeout.
  */
 static void balance_dirty_pages(struct address_space *mapping,
                                unsigned long write_chunk)
@@ -566,7 +566,8 @@ static void balance_dirty_pages(struct address_space *mapping,
                if (pages_written >= write_chunk)
                        break;          /* We've done our duty */
 
-               schedule_timeout_interruptible(pause);
+               __set_current_state(TASK_INTERRUPTIBLE);
+               io_schedule_timeout(pause);
 
                /*
                 * Increase the delay for each loop, up to our previous
@@ -582,7 +583,7 @@ static void balance_dirty_pages(struct address_space *mapping,
                bdi->dirty_exceeded = 0;
 
        if (writeback_in_progress(bdi))
-               return;         /* pdflush is already working this queue */
+               return;
 
        /*
         * In laptop mode, we wait until hitting the higher threshold before
@@ -593,10 +594,10 @@ static void balance_dirty_pages(struct address_space *mapping,
         * background_thresh, to keep the amount of dirty memory low.
         */
        if ((laptop_mode && pages_written) ||
-           (!laptop_mode && ((nr_writeback = global_page_state(NR_FILE_DIRTY)
-                                         + global_page_state(NR_UNSTABLE_NFS))
+           (!laptop_mode && ((global_page_state(NR_FILE_DIRTY)
+                              + global_page_state(NR_UNSTABLE_NFS))
                                          > background_thresh)))
-               bdi_start_writeback(bdi, nr_writeback);
+               bdi_start_writeback(bdi, NULL, 0);
 }
 
 void set_page_dirty_balance(struct page *page, int page_mkwrite)
@@ -682,35 +683,31 @@ void throttle_vm_writeout(gfp_t gfp_mask)
         }
 }
 
-static void laptop_timer_fn(unsigned long unused);
-
-static DEFINE_TIMER(laptop_mode_wb_timer, laptop_timer_fn, 0, 0);
-
 /*
  * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
  */
 int dirty_writeback_centisecs_handler(ctl_table *table, int write,
-       struct file *file, void __user *buffer, size_t *length, loff_t *ppos)
+       void __user *buffer, size_t *length, loff_t *ppos)
 {
-       proc_dointvec(table, write, file, buffer, length, ppos);
+       proc_dointvec(table, write, buffer, length, ppos);
+       bdi_arm_supers_timer();
        return 0;
 }
 
-static void do_laptop_sync(struct work_struct *work)
+#ifdef CONFIG_BLOCK
+void laptop_mode_timer_fn(unsigned long data)
 {
-       wakeup_flusher_threads(0);
-       kfree(work);
-}
+       struct request_queue *q = (struct request_queue *)data;
+       int nr_pages = global_page_state(NR_FILE_DIRTY) +
+               global_page_state(NR_UNSTABLE_NFS);
 
-static void laptop_timer_fn(unsigned long unused)
-{
-       struct work_struct *work;
+       /*
+        * We want to write everything out, not just down to the dirty
+        * threshold
+        */
 
-       work = kmalloc(sizeof(*work), GFP_ATOMIC);
-       if (work) {
-               INIT_WORK(work, do_laptop_sync);
-               schedule_work(work);
-       }
+       if (bdi_has_dirty_io(&q->backing_dev_info))
+               bdi_start_writeback(&q->backing_dev_info, NULL, nr_pages);
 }
 
 /*
@@ -718,9 +715,9 @@ static void laptop_timer_fn(unsigned long unused)
  * of all dirty data a few seconds from now.  If the flush is already scheduled
  * then push it back - the user is still using the disk.
  */
-void laptop_io_completion(void)
+void laptop_io_completion(struct backing_dev_info *info)
 {
-       mod_timer(&laptop_mode_wb_timer, jiffies + laptop_mode);
+       mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
 }
 
 /*
@@ -730,8 +727,16 @@ void laptop_io_completion(void)
  */
 void laptop_sync_completion(void)
 {
-       del_timer(&laptop_mode_wb_timer);
+       struct backing_dev_info *bdi;
+
+       rcu_read_lock();
+
+       list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
+               del_timer(&bdi->laptop_mode_wb_timer);
+
+       rcu_read_unlock();
 }
+#endif
 
 /*
  * If ratelimit_pages is too high then we can get into dirty-data overload
@@ -820,7 +825,6 @@ int write_cache_pages(struct address_space *mapping,
                      struct writeback_control *wbc, writepage_t writepage,
                      void *data)
 {
-       struct backing_dev_info *bdi = mapping->backing_dev_info;
        int ret = 0;
        int done = 0;
        struct pagevec pvec;
@@ -831,12 +835,6 @@ int write_cache_pages(struct address_space *mapping,
        pgoff_t done_index;
        int cycled;
        int range_whole = 0;
-       long nr_to_write = wbc->nr_to_write;
-
-       if (wbc->nonblocking && bdi_write_congested(bdi)) {
-               wbc->encountered_congestion = 1;
-               return 0;
-       }
 
        pagevec_init(&pvec, 0);
        if (wbc->range_cyclic) {
@@ -853,7 +851,22 @@ int write_cache_pages(struct address_space *mapping,
                if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
                        range_whole = 1;
                cycled = 1; /* ignore range_cyclic tests */
+
+               /*
+                * If this is a data integrity sync, cap the writeback to the
+                * current end of file. Any extension to the file that occurs
+                * after this is a new write and we don't need to write those
+                * pages out to fulfil our data integrity requirements. If we
+                * try to write them out, we can get stuck in this scan until
+                * the concurrent writer stops adding dirty pages and extending
+                * EOF.
+                */
+               if (wbc->sync_mode == WB_SYNC_ALL &&
+                   wbc->range_end == LLONG_MAX) {
+                       end = i_size_read(mapping->host) >> PAGE_CACHE_SHIFT;
+               }
        }
+
 retry:
        done_index = index;
        while (!done && (index <= end)) {
@@ -936,11 +949,10 @@ continue_unlock:
                                        done = 1;
                                        break;
                                }
-                       }
+                       }
 
-                       if (nr_to_write > 0) {
-                               nr_to_write--;
-                               if (nr_to_write == 0 &&
+                       if (wbc->nr_to_write > 0) {
+                               if (--wbc->nr_to_write == 0 &&
                                    wbc->sync_mode == WB_SYNC_NONE) {
                                        /*
                                         * We stop writing back only if we are
@@ -956,12 +968,6 @@ continue_unlock:
                                        break;
                                }
                        }
-
-                       if (wbc->nonblocking && bdi_write_congested(bdi)) {
-                               wbc->encountered_congestion = 1;
-                               done = 1;
-                               break;
-                       }
                }
                pagevec_release(&pvec);
                cond_resched();
@@ -977,11 +983,8 @@ continue_unlock:
                end = writeback_index - 1;
                goto retry;
        }
-       if (!wbc->no_nrwrite_index_update) {
-               if (wbc->range_cyclic || (range_whole && nr_to_write > 0))
-                       mapping->writeback_index = done_index;
-               wbc->nr_to_write = nr_to_write;
-       }
+       if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
+               mapping->writeback_index = done_index;
 
        return ret;
 }
@@ -1153,6 +1156,13 @@ int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
 EXPORT_SYMBOL(redirty_page_for_writepage);
 
 /*
+ * Dirty a page.
+ *
+ * For pages with a mapping this should be done under the page lock
+ * for the benefit of asynchronous memory errors who prefer a consistent
+ * dirty state. This rule can be broken in some special cases,
+ * but should be better not to.
+ *
  * If the mapping doesn't provide a set_page_dirty a_op, then
  * just fall through and assume that it wants buffer_heads.
  */