arm: tegra: raise cpu floor when display is on
[linux-2.6.git] / mm / readahead.c
index d7c6e14..867f9dd 100644 (file)
@@ -9,6 +9,7 @@
 
 #include <linux/kernel.h>
 #include <linux/fs.h>
+#include <linux/gfp.h>
 #include <linux/mm.h>
 #include <linux/module.h>
 #include <linux/blkdev.h>
@@ -108,9 +109,12 @@ EXPORT_SYMBOL(read_cache_pages);
 static int read_pages(struct address_space *mapping, struct file *filp,
                struct list_head *pages, unsigned nr_pages)
 {
+       struct blk_plug plug;
        unsigned page_idx;
        int ret;
 
+       blk_start_plug(&plug);
+
        if (mapping->a_ops->readpages) {
                ret = mapping->a_ops->readpages(filp, mapping, pages, nr_pages);
                /* Clean up the remaining pages */
@@ -128,20 +132,20 @@ static int read_pages(struct address_space *mapping, struct file *filp,
                page_cache_release(page);
        }
        ret = 0;
+
 out:
+       blk_finish_plug(&plug);
+
        return ret;
 }
 
 /*
- * do_page_cache_readahead actually reads a chunk of disk.  It allocates all
+ * __do_page_cache_readahead() actually reads a chunk of disk.  It allocates all
  * the pages first, then submits them all for I/O. This avoids the very bad
  * behaviour which would occur if page allocations are causing VM writeback.
  * We really don't want to intermingle reads and writes like that.
  *
  * Returns the number of pages requested, or the maximum amount of I/O allowed.
- *
- * do_page_cache_readahead() returns -1 if it encountered request queue
- * congestion.
  */
 static int
 __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
@@ -176,7 +180,7 @@ __do_page_cache_readahead(struct address_space *mapping, struct file *filp,
                if (page)
                        continue;
 
-               page = page_cache_alloc_cold(mapping);
+               page = page_cache_alloc_readahead(mapping);
                if (!page)
                        break;
                page->index = page_offset;
@@ -232,22 +236,6 @@ int force_page_cache_readahead(struct address_space *mapping, struct file *filp,
 }
 
 /*
- * This version skips the IO if the queue is read-congested, and will tell the
- * block layer to abandon the readahead if request allocation would block.
- *
- * force_page_cache_readahead() will ignore queue congestion and will block on
- * request queues.
- */
-int do_page_cache_readahead(struct address_space *mapping, struct file *filp,
-                       pgoff_t offset, unsigned long nr_to_read)
-{
-       if (bdi_read_congested(mapping->backing_dev_info))
-               return -1;
-
-       return __do_page_cache_readahead(mapping, filp, offset, nr_to_read, 0);
-}
-
-/*
  * Given a desired number of PAGE_CACHE_SIZE readahead pages, return a
  * sensible upper limit.
  */
@@ -260,7 +248,7 @@ unsigned long max_sane_readahead(unsigned long nr)
 /*
  * Submit IO for the read-ahead request in file_ra_state.
  */
-static unsigned long ra_submit(struct file_ra_state *ra,
+unsigned long ra_submit(struct file_ra_state *ra,
                       struct address_space *mapping, struct file *filp)
 {
        int actual;
@@ -349,6 +337,59 @@ static unsigned long get_next_ra_size(struct file_ra_state *ra,
  */
 
 /*
+ * Count contiguously cached pages from @offset-1 to @offset-@max,
+ * this count is a conservative estimation of
+ *     - length of the sequential read sequence, or
+ *     - thrashing threshold in memory tight systems
+ */
+static pgoff_t count_history_pages(struct address_space *mapping,
+                                  struct file_ra_state *ra,
+                                  pgoff_t offset, unsigned long max)
+{
+       pgoff_t head;
+
+       rcu_read_lock();
+       head = radix_tree_prev_hole(&mapping->page_tree, offset - 1, max);
+       rcu_read_unlock();
+
+       return offset - 1 - head;
+}
+
+/*
+ * page cache context based read-ahead
+ */
+static int try_context_readahead(struct address_space *mapping,
+                                struct file_ra_state *ra,
+                                pgoff_t offset,
+                                unsigned long req_size,
+                                unsigned long max)
+{
+       pgoff_t size;
+
+       size = count_history_pages(mapping, ra, offset, max);
+
+       /*
+        * no history pages:
+        * it could be a random read
+        */
+       if (!size)
+               return 0;
+
+       /*
+        * starts from beginning of file:
+        * it is a strong indication of long-run stream (or whole-file-read)
+        */
+       if (size >= offset)
+               size *= 2;
+
+       ra->start = offset;
+       ra->size = get_init_ra_size(size + req_size, max);
+       ra->async_size = ra->size;
+
+       return 1;
+}
+
+/*
  * A minimal readahead algorithm for trivial sequential/random reads.
  */
 static unsigned long
@@ -358,33 +399,25 @@ ondemand_readahead(struct address_space *mapping,
                   unsigned long req_size)
 {
        unsigned long max = max_sane_readahead(ra->ra_pages);
-       pgoff_t prev_offset;
-       int     sequential;
+
+       /*
+        * start of file
+        */
+       if (!offset)
+               goto initial_readahead;
 
        /*
         * It's the expected callback offset, assume sequential access.
         * Ramp up sizes, and push forward the readahead window.
         */
-       if (offset && (offset == (ra->start + ra->size - ra->async_size) ||
-                       offset == (ra->start + ra->size))) {
+       if ((offset == (ra->start + ra->size - ra->async_size) ||
+            offset == (ra->start + ra->size))) {
                ra->start += ra->size;
                ra->size = get_next_ra_size(ra, max);
                ra->async_size = ra->size;
                goto readit;
        }
 
-       prev_offset = ra->prev_pos >> PAGE_CACHE_SHIFT;
-       sequential = offset - prev_offset <= 1UL || req_size > max;
-
-       /*
-        * Standalone, small read.
-        * Read as is, and do not pollute the readahead state.
-        */
-       if (!hit_readahead_marker && !sequential) {
-               return __do_page_cache_readahead(mapping, filp,
-                                               offset, req_size, 0);
-       }
-
        /*
         * Hit a marked page without valid readahead state.
         * E.g. interleaved reads.
@@ -410,12 +443,31 @@ ondemand_readahead(struct address_space *mapping,
        }
 
        /*
-        * It may be one of
-        *      - first read on start of file
-        *      - sequential cache miss
-        *      - oversize random read
-        * Start readahead for it.
+        * oversize read
         */
+       if (req_size > max)
+               goto initial_readahead;
+
+       /*
+        * sequential cache miss
+        */
+       if (offset - (ra->prev_pos >> PAGE_CACHE_SHIFT) <= 1UL)
+               goto initial_readahead;
+
+       /*
+        * Query the page cache and look for the traces(cached history pages)
+        * that a sequential stream would leave behind.
+        */
+       if (try_context_readahead(mapping, ra, offset, req_size, max))
+               goto readit;
+
+       /*
+        * standalone, small random read
+        * Read as is, and do not pollute the readahead state.
+        */
+       return __do_page_cache_readahead(mapping, filp, offset, req_size, 0);
+
+initial_readahead:
        ra->start = offset;
        ra->size = get_init_ra_size(req_size, max);
        ra->async_size = ra->size > req_size ? ra->size - req_size : ra->size;
@@ -456,6 +508,12 @@ void page_cache_sync_readahead(struct address_space *mapping,
        if (!ra->ra_pages)
                return;
 
+       /* be dumb */
+       if (filp && (filp->f_mode & FMODE_RANDOM)) {
+               force_page_cache_readahead(mapping, filp, offset, req_size);
+               return;
+       }
+
        /* do read-ahead */
        ondemand_readahead(mapping, ra, filp, false, offset, req_size);
 }
@@ -471,7 +529,7 @@ EXPORT_SYMBOL_GPL(page_cache_sync_readahead);
  * @req_size: hint: total size of the read which the caller is performing in
  *            pagecache pages
  *
- * page_cache_async_ondemand() should be called when a page is used which
+ * page_cache_async_readahead() should be called when a page is used which
  * has the PG_readahead flag; this is a marker to suggest that the application
  * has used up enough of the readahead window that we should start pulling in
  * more pages.