percpu: fix per_cpu_ptr_to_phys() handling of non-page-aligned addresses
[linux-2.6.git] / mm / page-writeback.c
index 31f6988..0e309cd 100644 (file)
 #include <trace/events/writeback.h>
 
 /*
+ * Sleep at most 200ms at a time in balance_dirty_pages().
+ */
+#define MAX_PAUSE              max(HZ/5, 1)
+
+/*
+ * Estimate write bandwidth at 200ms intervals.
+ */
+#define BANDWIDTH_INTERVAL     max(HZ/5, 1)
+
+/*
  * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
  * will look to see if it needs to force writeback or throttling.
  */
@@ -111,6 +121,7 @@ EXPORT_SYMBOL(laptop_mode);
 
 /* End of sysctl-exported parameters */
 
+unsigned long global_dirty_limit;
 
 /*
  * Scale the writeback cache size proportional to the relative writeout speeds.
@@ -219,6 +230,7 @@ int dirty_bytes_handler(struct ctl_table *table, int write,
  */
 static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
 {
+       __inc_bdi_stat(bdi, BDI_WRITTEN);
        __prop_inc_percpu_max(&vm_completions, &bdi->completions,
                              bdi->max_prop_frac);
 }
@@ -244,13 +256,8 @@ void task_dirty_inc(struct task_struct *tsk)
 static void bdi_writeout_fraction(struct backing_dev_info *bdi,
                long *numerator, long *denominator)
 {
-       if (bdi_cap_writeback_dirty(bdi)) {
-               prop_fraction_percpu(&vm_completions, &bdi->completions,
+       prop_fraction_percpu(&vm_completions, &bdi->completions,
                                numerator, denominator);
-       } else {
-               *numerator = 0;
-               *denominator = 1;
-       }
 }
 
 static inline void task_dirties_fraction(struct task_struct *tsk,
@@ -274,12 +281,13 @@ static inline void task_dirties_fraction(struct task_struct *tsk,
  * effectively curb the growth of dirty pages. Light dirtiers with high enough
  * dirty threshold may never get throttled.
  */
+#define TASK_LIMIT_FRACTION 8
 static unsigned long task_dirty_limit(struct task_struct *tsk,
                                       unsigned long bdi_dirty)
 {
        long numerator, denominator;
        unsigned long dirty = bdi_dirty;
-       u64 inv = dirty >> 3;
+       u64 inv = dirty / TASK_LIMIT_FRACTION;
 
        task_dirties_fraction(tsk, &numerator, &denominator);
        inv *= numerator;
@@ -290,6 +298,12 @@ static unsigned long task_dirty_limit(struct task_struct *tsk,
        return max(dirty, bdi_dirty/2);
 }
 
+/* Minimum limit for any task */
+static unsigned long task_min_dirty_limit(unsigned long bdi_dirty)
+{
+       return bdi_dirty - bdi_dirty / TASK_LIMIT_FRACTION;
+}
+
 /*
  *
  */
@@ -397,6 +411,11 @@ unsigned long determine_dirtyable_memory(void)
        return x + 1;   /* Ensure that we never return 0 */
 }
 
+static unsigned long hard_dirty_limit(unsigned long thresh)
+{
+       return max(thresh, global_dirty_limit);
+}
+
 /*
  * global_dirty_limits - background-writeback and dirty-throttling thresholds
  *
@@ -435,12 +454,20 @@ void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
        }
        *pbackground = background;
        *pdirty = dirty;
+       trace_global_dirty_state(background, dirty);
 }
 
-/*
+/**
  * bdi_dirty_limit - @bdi's share of dirty throttling threshold
+ * @bdi: the backing_dev_info to query
+ * @dirty: global dirty limit in pages
  *
- * Allocate high/low dirty limits to fast/slow devices, in order to prevent
+ * Returns @bdi's dirty limit in pages. The term "dirty" in the context of
+ * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
+ * And the "limit" in the name is not seriously taken as hard limit in
+ * balance_dirty_pages().
+ *
+ * It allocates high/low dirty limits to fast/slow devices, in order to prevent
  * - starving fast devices
  * - piling up dirty pages (that will take long time to sync) on slow devices
  *
@@ -468,6 +495,153 @@ unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
        return bdi_dirty;
 }
 
+static void bdi_update_write_bandwidth(struct backing_dev_info *bdi,
+                                      unsigned long elapsed,
+                                      unsigned long written)
+{
+       const unsigned long period = roundup_pow_of_two(3 * HZ);
+       unsigned long avg = bdi->avg_write_bandwidth;
+       unsigned long old = bdi->write_bandwidth;
+       u64 bw;
+
+       /*
+        * bw = written * HZ / elapsed
+        *
+        *                   bw * elapsed + write_bandwidth * (period - elapsed)
+        * write_bandwidth = ---------------------------------------------------
+        *                                          period
+        */
+       bw = written - bdi->written_stamp;
+       bw *= HZ;
+       if (unlikely(elapsed > period)) {
+               do_div(bw, elapsed);
+               avg = bw;
+               goto out;
+       }
+       bw += (u64)bdi->write_bandwidth * (period - elapsed);
+       bw >>= ilog2(period);
+
+       /*
+        * one more level of smoothing, for filtering out sudden spikes
+        */
+       if (avg > old && old >= (unsigned long)bw)
+               avg -= (avg - old) >> 3;
+
+       if (avg < old && old <= (unsigned long)bw)
+               avg += (old - avg) >> 3;
+
+out:
+       bdi->write_bandwidth = bw;
+       bdi->avg_write_bandwidth = avg;
+}
+
+/*
+ * The global dirtyable memory and dirty threshold could be suddenly knocked
+ * down by a large amount (eg. on the startup of KVM in a swapless system).
+ * This may throw the system into deep dirty exceeded state and throttle
+ * heavy/light dirtiers alike. To retain good responsiveness, maintain
+ * global_dirty_limit for tracking slowly down to the knocked down dirty
+ * threshold.
+ */
+static void update_dirty_limit(unsigned long thresh, unsigned long dirty)
+{
+       unsigned long limit = global_dirty_limit;
+
+       /*
+        * Follow up in one step.
+        */
+       if (limit < thresh) {
+               limit = thresh;
+               goto update;
+       }
+
+       /*
+        * Follow down slowly. Use the higher one as the target, because thresh
+        * may drop below dirty. This is exactly the reason to introduce
+        * global_dirty_limit which is guaranteed to lie above the dirty pages.
+        */
+       thresh = max(thresh, dirty);
+       if (limit > thresh) {
+               limit -= (limit - thresh) >> 5;
+               goto update;
+       }
+       return;
+update:
+       global_dirty_limit = limit;
+}
+
+static void global_update_bandwidth(unsigned long thresh,
+                                   unsigned long dirty,
+                                   unsigned long now)
+{
+       static DEFINE_SPINLOCK(dirty_lock);
+       static unsigned long update_time;
+
+       /*
+        * check locklessly first to optimize away locking for the most time
+        */
+       if (time_before(now, update_time + BANDWIDTH_INTERVAL))
+               return;
+
+       spin_lock(&dirty_lock);
+       if (time_after_eq(now, update_time + BANDWIDTH_INTERVAL)) {
+               update_dirty_limit(thresh, dirty);
+               update_time = now;
+       }
+       spin_unlock(&dirty_lock);
+}
+
+void __bdi_update_bandwidth(struct backing_dev_info *bdi,
+                           unsigned long thresh,
+                           unsigned long dirty,
+                           unsigned long bdi_thresh,
+                           unsigned long bdi_dirty,
+                           unsigned long start_time)
+{
+       unsigned long now = jiffies;
+       unsigned long elapsed = now - bdi->bw_time_stamp;
+       unsigned long written;
+
+       /*
+        * rate-limit, only update once every 200ms.
+        */
+       if (elapsed < BANDWIDTH_INTERVAL)
+               return;
+
+       written = percpu_counter_read(&bdi->bdi_stat[BDI_WRITTEN]);
+
+       /*
+        * Skip quiet periods when disk bandwidth is under-utilized.
+        * (at least 1s idle time between two flusher runs)
+        */
+       if (elapsed > HZ && time_before(bdi->bw_time_stamp, start_time))
+               goto snapshot;
+
+       if (thresh)
+               global_update_bandwidth(thresh, dirty, now);
+
+       bdi_update_write_bandwidth(bdi, elapsed, written);
+
+snapshot:
+       bdi->written_stamp = written;
+       bdi->bw_time_stamp = now;
+}
+
+static void bdi_update_bandwidth(struct backing_dev_info *bdi,
+                                unsigned long thresh,
+                                unsigned long dirty,
+                                unsigned long bdi_thresh,
+                                unsigned long bdi_dirty,
+                                unsigned long start_time)
+{
+       if (time_is_after_eq_jiffies(bdi->bw_time_stamp + BANDWIDTH_INTERVAL))
+               return;
+       spin_lock(&bdi->wb.list_lock);
+       __bdi_update_bandwidth(bdi, thresh, dirty, bdi_thresh, bdi_dirty,
+                              start_time);
+       spin_unlock(&bdi->wb.list_lock);
+}
+
 /*
  * 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
@@ -478,27 +652,25 @@ unsigned long bdi_dirty_limit(struct backing_dev_info *bdi, unsigned long dirty)
 static void balance_dirty_pages(struct address_space *mapping,
                                unsigned long write_chunk)
 {
-       long nr_reclaimable, bdi_nr_reclaimable;
-       long nr_writeback, bdi_nr_writeback;
+       unsigned long nr_reclaimable, bdi_nr_reclaimable;
+       unsigned long nr_dirty;  /* = file_dirty + writeback + unstable_nfs */
+       unsigned long bdi_dirty;
        unsigned long background_thresh;
        unsigned long dirty_thresh;
        unsigned long bdi_thresh;
+       unsigned long task_bdi_thresh;
+       unsigned long min_task_bdi_thresh;
        unsigned long pages_written = 0;
        unsigned long pause = 1;
        bool dirty_exceeded = false;
+       bool clear_dirty_exceeded = true;
        struct backing_dev_info *bdi = mapping->backing_dev_info;
+       unsigned long start_time = jiffies;
 
        for (;;) {
-               struct writeback_control wbc = {
-                       .sync_mode      = WB_SYNC_NONE,
-                       .older_than_this = NULL,
-                       .nr_to_write    = write_chunk,
-                       .range_cyclic   = 1,
-               };
-
                nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
                                        global_page_state(NR_UNSTABLE_NFS);
-               nr_writeback = global_page_state(NR_WRITEBACK);
+               nr_dirty = nr_reclaimable + global_page_state(NR_WRITEBACK);
 
                global_dirty_limits(&background_thresh, &dirty_thresh);
 
@@ -507,12 +679,12 @@ static void balance_dirty_pages(struct address_space *mapping,
                 * catch-up. This avoids (excessively) small writeouts
                 * when the bdi limits are ramping up.
                 */
-               if (nr_reclaimable + nr_writeback <=
-                               (background_thresh + dirty_thresh) / 2)
+               if (nr_dirty <= (background_thresh + dirty_thresh) / 2)
                        break;
 
                bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
-               bdi_thresh = task_dirty_limit(current, bdi_thresh);
+               min_task_bdi_thresh = task_min_dirty_limit(bdi_thresh);
+               task_bdi_thresh = task_dirty_limit(current, bdi_thresh);
 
                /*
                 * In order to avoid the stacked BDI deadlock we need
@@ -524,12 +696,14 @@ static void balance_dirty_pages(struct address_space *mapping,
                 * actually dirty; with m+n sitting in the percpu
                 * deltas.
                 */
-               if (bdi_thresh < 2*bdi_stat_error(bdi)) {
+               if (task_bdi_thresh < 2 * bdi_stat_error(bdi)) {
                        bdi_nr_reclaimable = bdi_stat_sum(bdi, BDI_RECLAIMABLE);
-                       bdi_nr_writeback = bdi_stat_sum(bdi, BDI_WRITEBACK);
+                       bdi_dirty = bdi_nr_reclaimable +
+                                   bdi_stat_sum(bdi, BDI_WRITEBACK);
                } else {
                        bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
-                       bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
+                       bdi_dirty = bdi_nr_reclaimable +
+                                   bdi_stat(bdi, BDI_WRITEBACK);
                }
 
                /*
@@ -538,9 +712,10 @@ static void balance_dirty_pages(struct address_space *mapping,
                 * bdi or process from holding back light ones; The latter is
                 * the last resort safeguard.
                 */
-               dirty_exceeded =
-                       (bdi_nr_reclaimable + bdi_nr_writeback > bdi_thresh)
-                       || (nr_reclaimable + nr_writeback > dirty_thresh);
+               dirty_exceeded = (bdi_dirty > task_bdi_thresh) ||
+                                 (nr_dirty > dirty_thresh);
+               clear_dirty_exceeded = (bdi_dirty <= min_task_bdi_thresh) &&
+                                       (nr_dirty <= dirty_thresh);
 
                if (!dirty_exceeded)
                        break;
@@ -548,6 +723,9 @@ static void balance_dirty_pages(struct address_space *mapping,
                if (!bdi->dirty_exceeded)
                        bdi->dirty_exceeded = 1;
 
+               bdi_update_bandwidth(bdi, dirty_thresh, nr_dirty,
+                                    bdi_thresh, bdi_dirty, start_time);
+
                /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
                 * Unstable writes are a feature of certain networked
                 * filesystems (i.e. NFS) in which data may have been
@@ -557,17 +735,29 @@ static void balance_dirty_pages(struct address_space *mapping,
                 * threshold otherwise wait until the disk writes catch
                 * up.
                 */
-               trace_wbc_balance_dirty_start(&wbc, bdi);
-               if (bdi_nr_reclaimable > bdi_thresh) {
-                       writeback_inodes_wb(&bdi->wb, &wbc);
-                       pages_written += write_chunk - wbc.nr_to_write;
-                       trace_wbc_balance_dirty_written(&wbc, bdi);
+               trace_balance_dirty_start(bdi);
+               if (bdi_nr_reclaimable > task_bdi_thresh) {
+                       pages_written += writeback_inodes_wb(&bdi->wb,
+                                                            write_chunk);
+                       trace_balance_dirty_written(bdi, pages_written);
                        if (pages_written >= write_chunk)
                                break;          /* We've done our duty */
                }
-               trace_wbc_balance_dirty_wait(&wbc, bdi);
                __set_current_state(TASK_UNINTERRUPTIBLE);
                io_schedule_timeout(pause);
+               trace_balance_dirty_wait(bdi);
+
+               dirty_thresh = hard_dirty_limit(dirty_thresh);
+               /*
+                * max-pause area. If dirty exceeded but still within this
+                * area, no need to sleep for more than 200ms: (a) 8 pages per
+                * 200ms is typically more than enough to curb heavy dirtiers;
+                * (b) the pause time limit makes the dirtiers more responsive.
+                */
+               if (nr_dirty < dirty_thresh &&
+                   bdi_dirty < (task_bdi_thresh + bdi_thresh) / 2 &&
+                   time_after(jiffies, start_time + MAX_PAUSE))
+                       break;
 
                /*
                 * Increase the delay for each loop, up to our previous
@@ -578,7 +768,8 @@ static void balance_dirty_pages(struct address_space *mapping,
                        pause = HZ / 10;
        }
 
-       if (!dirty_exceeded && bdi->dirty_exceeded)
+       /* Clear dirty_exceeded flag only when no task can exceed the limit */
+       if (clear_dirty_exceeded && bdi->dirty_exceeded)
                bdi->dirty_exceeded = 0;
 
        if (writeback_in_progress(bdi))
@@ -626,9 +817,13 @@ static DEFINE_PER_CPU(unsigned long, bdp_ratelimits) = 0;
 void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
                                        unsigned long nr_pages_dirtied)
 {
+       struct backing_dev_info *bdi = mapping->backing_dev_info;
        unsigned long ratelimit;
        unsigned long *p;
 
+       if (!bdi_cap_account_dirty(bdi))
+               return;
+
        ratelimit = ratelimit_pages;
        if (mapping->backing_dev_info->dirty_exceeded)
                ratelimit = 8;
@@ -892,12 +1087,12 @@ int write_cache_pages(struct address_space *mapping,
                        range_whole = 1;
                cycled = 1; /* ignore range_cyclic tests */
        }
-       if (wbc->sync_mode == WB_SYNC_ALL)
+       if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
                tag = PAGECACHE_TAG_TOWRITE;
        else
                tag = PAGECACHE_TAG_DIRTY;
 retry:
-       if (wbc->sync_mode == WB_SYNC_ALL)
+       if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
                tag_pages_for_writeback(mapping, index, end);
        done_index = index;
        while (!done && (index <= end)) {
@@ -1141,7 +1336,6 @@ EXPORT_SYMBOL(account_page_dirtied);
 void account_page_writeback(struct page *page)
 {
        inc_zone_page_state(page, NR_WRITEBACK);
-       inc_zone_page_state(page, NR_WRITTEN);
 }
 EXPORT_SYMBOL(account_page_writeback);
 
@@ -1358,8 +1552,10 @@ int test_clear_page_writeback(struct page *page)
        } else {
                ret = TestClearPageWriteback(page);
        }
-       if (ret)
+       if (ret) {
                dec_zone_page_state(page, NR_WRITEBACK);
+               inc_zone_page_state(page, NR_WRITTEN);
+       }
        return ret;
 }
 
@@ -1405,10 +1601,6 @@ EXPORT_SYMBOL(test_set_page_writeback);
  */
 int mapping_tagged(struct address_space *mapping, int tag)
 {
-       int ret;
-       rcu_read_lock();
-       ret = radix_tree_tagged(&mapping->page_tree, tag);
-       rcu_read_unlock();
-       return ret;
+       return radix_tree_tagged(&mapping->page_tree, tag);
 }
 EXPORT_SYMBOL(mapping_tagged);