mm: per device dirty threshold
Peter Zijlstra [Wed, 17 Oct 2007 06:25:50 +0000 (23:25 -0700)]
Scale writeback cache per backing device, proportional to its writeout speed.

By decoupling the BDI dirty thresholds a number of problems we currently have
will go away, namely:

 - mutual interference starvation (for any number of BDIs);
 - deadlocks with stacked BDIs (loop, FUSE and local NFS mounts).

It might be that all dirty pages are for a single BDI while other BDIs are
idling. By giving each BDI a 'fair' share of the dirty limit, each one can have
dirty pages outstanding and make progress.

A global threshold also creates a deadlock for stacked BDIs; when A writes to
B, and A generates enough dirty pages to get throttled, B will never start
writeback until the dirty pages go away. Again, by giving each BDI its own
'independent' dirty limit, this problem is avoided.

So the problem is to determine how to distribute the total dirty limit across
the BDIs fairly and efficiently. A DBI that has a large dirty limit but does
not have any dirty pages outstanding is a waste.

What is done is to keep a floating proportion between the DBIs based on
writeback completions. This way faster/more active devices get a larger share
than slower/idle devices.

[akpm@linux-foundation.org: fix warnings]
[hugh@veritas.com: Fix occasional hang when a task couldn't get out of balance_dirty_pages]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

include/linux/backing-dev.h
include/linux/writeback.h
kernel/sysctl.c
mm/backing-dev.c
mm/page-writeback.c

index 4d9222c..48a62ba 100644 (file)
@@ -10,6 +10,7 @@
 
 #include <linux/percpu_counter.h>
 #include <linux/log2.h>
+#include <linux/proportions.h>
 #include <asm/atomic.h>
 
 struct page;
@@ -44,6 +45,9 @@ struct backing_dev_info {
        void *unplug_io_data;
 
        struct percpu_counter bdi_stat[NR_BDI_STAT_ITEMS];
+
+       struct prop_local_percpu completions;
+       int dirty_exceeded;
 };
 
 int bdi_init(struct backing_dev_info *bdi);
index d1321a8..52be879 100644 (file)
@@ -97,6 +97,10 @@ extern int dirty_expire_interval;
 extern int block_dump;
 extern int laptop_mode;
 
+extern int dirty_ratio_handler(struct ctl_table *table, int write,
+               struct file *filp, void __user *buffer, size_t *lenp,
+               loff_t *ppos);
+
 struct ctl_table;
 struct file;
 int dirty_writeback_centisecs_handler(struct ctl_table *, int, struct file *,
index 96efbb8..c676b5e 100644 (file)
@@ -813,7 +813,7 @@ static ctl_table vm_table[] = {
                .data           = &vm_dirty_ratio,
                .maxlen         = sizeof(vm_dirty_ratio),
                .mode           = 0644,
-               .proc_handler   = &proc_dointvec_minmax,
+               .proc_handler   = &dirty_ratio_handler,
                .strategy       = &sysctl_intvec,
                .extra1         = &zero,
                .extra2         = &one_hundred,
index a47065e..b0ceb29 100644 (file)
@@ -12,11 +12,17 @@ int bdi_init(struct backing_dev_info *bdi)
 
        for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
                err = percpu_counter_init_irq(&bdi->bdi_stat[i], 0);
-               if (err) {
-                       for (j = 0; j < i; j++)
-                               percpu_counter_destroy(&bdi->bdi_stat[i]);
-                       break;
-               }
+               if (err)
+                       goto err;
+       }
+
+       bdi->dirty_exceeded = 0;
+       err = prop_local_init_percpu(&bdi->completions);
+
+       if (err) {
+err:
+               for (j = 0; j < i; j++)
+                       percpu_counter_destroy(&bdi->bdi_stat[i]);
        }
 
        return err;
@@ -29,6 +35,8 @@ void bdi_destroy(struct backing_dev_info *bdi)
 
        for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
                percpu_counter_destroy(&bdi->bdi_stat[i]);
+
+       prop_local_destroy_percpu(&bdi->completions);
 }
 EXPORT_SYMBOL(bdi_destroy);
 
@@ -81,3 +89,4 @@ long congestion_wait(int rw, long timeout)
        return ret;
 }
 EXPORT_SYMBOL(congestion_wait);
+
index f1d201f..b036054 100644 (file)
@@ -2,6 +2,7 @@
  * mm/page-writeback.c
  *
  * Copyright (C) 2002, Linus Torvalds.
+ * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
  *
  * Contains functions related to writing back dirty pages at the
  * address_space level.
@@ -49,8 +50,6 @@
  */
 static long ratelimit_pages = 32;
 
-static int dirty_exceeded __cacheline_aligned_in_smp;  /* Dirty mem may be over limit */
-
 /*
  * When balance_dirty_pages decides that the caller needs to perform some
  * non-background writeback, this is how many pages it will attempt to write.
@@ -103,6 +102,103 @@ EXPORT_SYMBOL(laptop_mode);
 static void background_writeout(unsigned long _min_pages);
 
 /*
+ * Scale the writeback cache size proportional to the relative writeout speeds.
+ *
+ * We do this by keeping a floating proportion between BDIs, based on page
+ * writeback completions [end_page_writeback()]. Those devices that write out
+ * pages fastest will get the larger share, while the slower will get a smaller
+ * share.
+ *
+ * We use page writeout completions because we are interested in getting rid of
+ * dirty pages. Having them written out is the primary goal.
+ *
+ * We introduce a concept of time, a period over which we measure these events,
+ * because demand can/will vary over time. The length of this period itself is
+ * measured in page writeback completions.
+ *
+ */
+static struct prop_descriptor vm_completions;
+
+static unsigned long determine_dirtyable_memory(void);
+
+/*
+ * couple the period to the dirty_ratio:
+ *
+ *   period/2 ~ roundup_pow_of_two(dirty limit)
+ */
+static int calc_period_shift(void)
+{
+       unsigned long dirty_total;
+
+       dirty_total = (vm_dirty_ratio * determine_dirtyable_memory()) / 100;
+       return 2 + ilog2(dirty_total - 1);
+}
+
+/*
+ * update the period when the dirty ratio changes.
+ */
+int dirty_ratio_handler(struct ctl_table *table, int write,
+               struct file *filp, void __user *buffer, size_t *lenp,
+               loff_t *ppos)
+{
+       int old_ratio = vm_dirty_ratio;
+       int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
+       if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
+               int shift = calc_period_shift();
+               prop_change_shift(&vm_completions, shift);
+       }
+       return ret;
+}
+
+/*
+ * Increment the BDI's writeout completion count and the global writeout
+ * completion count. Called from test_clear_page_writeback().
+ */
+static inline void __bdi_writeout_inc(struct backing_dev_info *bdi)
+{
+       __prop_inc_percpu(&vm_completions, &bdi->completions);
+}
+
+/*
+ * Obtain an accurate fraction of the BDI's portion.
+ */
+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,
+                               numerator, denominator);
+       } else {
+               *numerator = 0;
+               *denominator = 1;
+       }
+}
+
+/*
+ * Clip the earned share of dirty pages to that which is actually available.
+ * This avoids exceeding the total dirty_limit when the floating averages
+ * fluctuate too quickly.
+ */
+static void
+clip_bdi_dirty_limit(struct backing_dev_info *bdi, long dirty, long *pbdi_dirty)
+{
+       long avail_dirty;
+
+       avail_dirty = dirty -
+               (global_page_state(NR_FILE_DIRTY) +
+                global_page_state(NR_WRITEBACK) +
+                global_page_state(NR_UNSTABLE_NFS));
+
+       if (avail_dirty < 0)
+               avail_dirty = 0;
+
+       avail_dirty += bdi_stat(bdi, BDI_RECLAIMABLE) +
+               bdi_stat(bdi, BDI_WRITEBACK);
+
+       *pbdi_dirty = min(*pbdi_dirty, avail_dirty);
+}
+
+/*
  * Work out the current dirty-memory clamping and background writeout
  * thresholds.
  *
@@ -158,8 +254,8 @@ static unsigned long determine_dirtyable_memory(void)
 }
 
 static void
-get_dirty_limits(long *pbackground, long *pdirty,
-                                       struct address_space *mapping)
+get_dirty_limits(long *pbackground, long *pdirty, long *pbdi_dirty,
+                struct backing_dev_info *bdi)
 {
        int background_ratio;           /* Percentages */
        int dirty_ratio;
@@ -193,6 +289,22 @@ get_dirty_limits(long *pbackground, long *pdirty,
        }
        *pbackground = background;
        *pdirty = dirty;
+
+       if (bdi) {
+               u64 bdi_dirty = dirty;
+               long numerator, denominator;
+
+               /*
+                * Calculate this BDI's share of the dirty ratio.
+                */
+               bdi_writeout_fraction(bdi, &numerator, &denominator);
+
+               bdi_dirty *= numerator;
+               do_div(bdi_dirty, denominator);
+
+               *pbdi_dirty = bdi_dirty;
+               clip_bdi_dirty_limit(bdi, dirty, pbdi_dirty);
+       }
 }
 
 /*
@@ -204,9 +316,11 @@ get_dirty_limits(long *pbackground, long *pdirty,
  */
 static void balance_dirty_pages(struct address_space *mapping)
 {
-       long nr_reclaimable;
+       long bdi_nr_reclaimable;
+       long bdi_nr_writeback;
        long background_thresh;
        long dirty_thresh;
+       long bdi_thresh;
        unsigned long pages_written = 0;
        unsigned long write_chunk = sync_writeback_pages();
 
@@ -221,15 +335,15 @@ static void balance_dirty_pages(struct address_space *mapping)
                        .range_cyclic   = 1,
                };
 
-               get_dirty_limits(&background_thresh, &dirty_thresh, mapping);
-               nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
-                                       global_page_state(NR_UNSTABLE_NFS);
-               if (nr_reclaimable + global_page_state(NR_WRITEBACK) <=
-                       dirty_thresh)
-                               break;
+               get_dirty_limits(&background_thresh, &dirty_thresh,
+                               &bdi_thresh, bdi);
+               bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
+               bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
+               if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
+                       break;
 
-               if (!dirty_exceeded)
-                       dirty_exceeded = 1;
+               if (!bdi->dirty_exceeded)
+                       bdi->dirty_exceeded = 1;
 
                /* Note: nr_reclaimable denotes nr_dirty + nr_unstable.
                 * Unstable writes are a feature of certain networked
@@ -237,26 +351,42 @@ static void balance_dirty_pages(struct address_space *mapping)
                 * written to the server's write cache, but has not yet
                 * been flushed to permanent storage.
                 */
-               if (nr_reclaimable) {
+               if (bdi_nr_reclaimable) {
                        writeback_inodes(&wbc);
-                       get_dirty_limits(&background_thresh,
-                                               &dirty_thresh, mapping);
-                       nr_reclaimable = global_page_state(NR_FILE_DIRTY) +
-                                       global_page_state(NR_UNSTABLE_NFS);
-                       if (nr_reclaimable +
-                               global_page_state(NR_WRITEBACK)
-                                       <= dirty_thresh)
-                                               break;
                        pages_written += write_chunk - wbc.nr_to_write;
-                       if (pages_written >= write_chunk)
-                               break;          /* We've done our duty */
+                       get_dirty_limits(&background_thresh, &dirty_thresh,
+                                      &bdi_thresh, bdi);
+               }
+
+               /*
+                * In order to avoid the stacked BDI deadlock we need
+                * to ensure we accurately count the 'dirty' pages when
+                * the threshold is low.
+                *
+                * Otherwise it would be possible to get thresh+n pages
+                * reported dirty, even though there are thresh-m pages
+                * actually dirty; with m+n sitting in the percpu
+                * deltas.
+                */
+               if (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);
+               } else if (bdi_nr_reclaimable) {
+                       bdi_nr_reclaimable = bdi_stat(bdi, BDI_RECLAIMABLE);
+                       bdi_nr_writeback = bdi_stat(bdi, BDI_WRITEBACK);
                }
+
+               if (bdi_nr_reclaimable + bdi_nr_writeback <= bdi_thresh)
+                       break;
+               if (pages_written >= write_chunk)
+                       break;          /* We've done our duty */
+
                congestion_wait(WRITE, HZ/10);
        }
 
-       if (nr_reclaimable + global_page_state(NR_WRITEBACK)
-               <= dirty_thresh && dirty_exceeded)
-                       dirty_exceeded = 0;
+       if (bdi_nr_reclaimable + bdi_nr_writeback < bdi_thresh &&
+                       bdi->dirty_exceeded)
+               bdi->dirty_exceeded = 0;
 
        if (writeback_in_progress(bdi))
                return;         /* pdflush is already working this queue */
@@ -270,7 +400,9 @@ 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_reclaimable > background_thresh)))
+                       (!laptop_mode && (global_page_state(NR_FILE_DIRTY)
+                                         + global_page_state(NR_UNSTABLE_NFS)
+                                         > background_thresh)))
                pdflush_operation(background_writeout, 0);
 }
 
@@ -306,7 +438,7 @@ void balance_dirty_pages_ratelimited_nr(struct address_space *mapping,
        unsigned long *p;
 
        ratelimit = ratelimit_pages;
-       if (dirty_exceeded)
+       if (mapping->backing_dev_info->dirty_exceeded)
                ratelimit = 8;
 
        /*
@@ -342,7 +474,7 @@ void throttle_vm_writeout(gfp_t gfp_mask)
        }
 
         for ( ; ; ) {
-               get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
+               get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
 
                 /*
                  * Boost the allowable dirty threshold a bit for page
@@ -377,7 +509,7 @@ static void background_writeout(unsigned long _min_pages)
                long background_thresh;
                long dirty_thresh;
 
-               get_dirty_limits(&background_thresh, &dirty_thresh, NULL);
+               get_dirty_limits(&background_thresh, &dirty_thresh, NULL, NULL);
                if (global_page_state(NR_FILE_DIRTY) +
                        global_page_state(NR_UNSTABLE_NFS) < background_thresh
                                && min_pages <= 0)
@@ -580,9 +712,14 @@ static struct notifier_block __cpuinitdata ratelimit_nb = {
  */
 void __init page_writeback_init(void)
 {
+       int shift;
+
        mod_timer(&wb_timer, jiffies + dirty_writeback_interval);
        writeback_set_ratelimit();
        register_cpu_notifier(&ratelimit_nb);
+
+       shift = calc_period_shift();
+       prop_descriptor_init(&vm_completions, shift);
 }
 
 /**
@@ -988,8 +1125,10 @@ int test_clear_page_writeback(struct page *page)
                        radix_tree_tag_clear(&mapping->page_tree,
                                                page_index(page),
                                                PAGECACHE_TAG_WRITEBACK);
-                       if (bdi_cap_writeback_dirty(bdi))
+                       if (bdi_cap_writeback_dirty(bdi)) {
                                __dec_bdi_stat(bdi, BDI_WRITEBACK);
+                               __bdi_writeout_inc(bdi);
+                       }
                }
                write_unlock_irqrestore(&mapping->tree_lock, flags);
        } else {