UBIFS: do not allocate too much

Bulk-read allocates 128KiB or more using kmalloc. The allocation
starts failing often when the memory gets fragmented. UBIFS still
works fine in this case, because it falls-back to standard
(non-optimized) read method, though. This patch teaches bulk-read
to allocate exactly the amount of memory it needs, instead of
allocating 128KiB every time.

This patch is also a preparation to the further fix where we'll
have a pre-allocated bulk-read buffer as well. For example, now
the @bu object is prepared in 'ubifs_bulk_read()', so we could
path either pre-allocated or allocated information to
'ubifs_do_bulk_read()' later. Or teaching 'ubifs_do_bulk_read()'
not to allocate 'bu->buf' if it is already there.

Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>

4 files changed, 60 insertions, 33 deletions

diff --git a/fs/ubifs/file.c b/fs/ubifs/file.c
index 8be827cc7078..0c5c27d63f6e 100644
--- a/fs/ubifs/file.c
+++ b/fs/ubifs/file.c
@@ -691,32 +691,22 @@ out_err:
 /**
  * ubifs_do_bulk_read - do bulk-read.
  * @c: UBIFS file-system description object
- * @page1: first page
+ * @bu: bulk-read information
+ * @page1: first page to read
  *
  * This function returns %1 if the bulk-read is done, otherwise %0 is returned.
  */
-static int ubifs_do_bulk_read(struct ubifs_info *c, struct page *page1)
+static int ubifs_do_bulk_read(struct ubifs_info *c, struct bu_info *bu,
+                              struct page *page1)
 {
         pgoff_t offset = page1->index, end_index;
         struct address_space *mapping = page1->mapping;
         struct inode *inode = mapping->host;
         struct ubifs_inode *ui = ubifs_inode(inode);
-        struct bu_info *bu;
         int err, page_idx, page_cnt, ret = 0, n = 0;
+        int allocate = bu->buf ? 0 : 1;
         loff_t isize;
 
-        bu = kmalloc(sizeof(struct bu_info), GFP_NOFS | __GFP_NOWARN);
-        if (!bu)
-                return 0;
-
-        bu->buf_len = c->bulk_read_buf_size;
-        bu->buf = kmalloc(bu->buf_len, GFP_NOFS | __GFP_NOWARN);
-        if (!bu->buf)
-                goto out_free;
-
-        data_key_init(c, &bu->key, inode->i_ino,
-                      offset << UBIFS_BLOCKS_PER_PAGE_SHIFT);
-
         err = ubifs_tnc_get_bu_keys(c, bu);
         if (err)
                 goto out_warn;
@@ -735,12 +725,25 @@ static int ubifs_do_bulk_read(struct ubifs_info *c, struct page *page1)
                  * together. If all the pages were like this, bulk-read would
                  * reduce performance, so we turn it off for a while.
                  */
-                ui->read_in_a_row = 0;
-                ui->bulk_read = 0;
-                goto out_free;
+                goto out_bu_off;
         }
 
         if (bu->cnt) {
+                if (allocate) {
+                        /*
+                         * Allocate bulk-read buffer depending on how many data
+                         * nodes we are going to read.
+                         */
+                        bu->buf_len = bu->zbranch[bu->cnt - 1].offs +
+                                      bu->zbranch[bu->cnt - 1].len -
+                                      bu->zbranch[0].offs;
+                        ubifs_assert(bu->buf_len > 0);
+                        ubifs_assert(bu->buf_len <= c->leb_size);
+                        bu->buf = kmalloc(bu->buf_len, GFP_NOFS | __GFP_NOWARN);
+                        if (!bu->buf)
+                                goto out_bu_off;
+                }
+
                 err = ubifs_tnc_bulk_read(c, bu);
                 if (err)
                         goto out_warn;
@@ -779,13 +782,17 @@ static int ubifs_do_bulk_read(struct ubifs_info *c, struct page *page1)
         ui->last_page_read = offset + page_idx - 1;
 
 out_free:
-        kfree(bu->buf);
-        kfree(bu);
+        if (allocate)
+                kfree(bu->buf);
         return ret;
 
 out_warn:
         ubifs_warn("ignoring error %d and skipping bulk-read", err);
         goto out_free;
+
+out_bu_off:
+        ui->read_in_a_row = ui->bulk_read = 0;
+        goto out_free;
 }
 
 /**
@@ -803,18 +810,20 @@ static int ubifs_bulk_read(struct page *page)
         struct ubifs_info *c = inode->i_sb->s_fs_info;
         struct ubifs_inode *ui = ubifs_inode(inode);
         pgoff_t index = page->index, last_page_read = ui->last_page_read;
-        int ret = 0;
+        struct bu_info *bu;
+        int err = 0;
 
         ui->last_page_read = index;
-
         if (!c->bulk_read)
                 return 0;
+
         /*
          * Bulk-read is protected by ui_mutex, but it is an optimization, so
          * don't bother if we cannot lock the mutex.
          */
         if (!mutex_trylock(&ui->ui_mutex))
                 return 0;
+
         if (index != last_page_read + 1) {
                 /* Turn off bulk-read if we stop reading sequentially */
                 ui->read_in_a_row = 1;
@@ -822,6 +831,7 @@ static int ubifs_bulk_read(struct page *page)
                         ui->bulk_read = 0;
                 goto out_unlock;
         }
+
         if (!ui->bulk_read) {
                 ui->read_in_a_row += 1;
                 if (ui->read_in_a_row < 3)
@@ -829,10 +839,22 @@ static int ubifs_bulk_read(struct page *page)
                 /* Three reads in a row, so switch on bulk-read */
                 ui->bulk_read = 1;
         }
-        ret = ubifs_do_bulk_read(c, page);
+
+        bu = kmalloc(sizeof(struct bu_info), GFP_NOFS | __GFP_NOWARN);
+        if (!bu)
+                return 0;
+
+        bu->buf = NULL;
+        bu->buf_len = c->max_bu_buf_len;
+        data_key_init(c, &bu->key, inode->i_ino,
+                      page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT);
+
+        err = ubifs_do_bulk_read(c, bu, page);
+        kfree(bu);
+
 out_unlock:
         mutex_unlock(&ui->ui_mutex);
-        return ret;
+        return err;
 }
 
 static int ubifs_readpage(struct file *file, struct page *page)
diff --git a/fs/ubifs/super.c b/fs/ubifs/super.c
index ea493e6f2652..1d511569c035 100644
--- a/fs/ubifs/super.c
+++ b/fs/ubifs/super.c
@@ -569,16 +569,16 @@ static int init_constants_early(struct ubifs_info *c)
         c->leb_overhead = c->leb_size % UBIFS_MAX_DATA_NODE_SZ;
 
         /* Buffer size for bulk-reads */
-        c->bulk_read_buf_size = UBIFS_MAX_BULK_READ * UBIFS_MAX_DATA_NODE_SZ;
-        if (c->bulk_read_buf_size > c->leb_size)
-                c->bulk_read_buf_size = c->leb_size;
-        if (c->bulk_read_buf_size > UBIFS_KMALLOC_OK) {
+        c->max_bu_buf_len = UBIFS_MAX_BULK_READ * UBIFS_MAX_DATA_NODE_SZ;
+        if (c->max_bu_buf_len > c->leb_size)
+                c->max_bu_buf_len = c->leb_size;
+        if (c->max_bu_buf_len > UBIFS_KMALLOC_OK) {
                 /* Check if we can kmalloc that much */
-                void *try = kmalloc(c->bulk_read_buf_size,
+                void *try = kmalloc(c->max_bu_buf_len,
                                     GFP_KERNEL | __GFP_NOWARN);
                 kfree(try);
                 if (!try)
-                        c->bulk_read_buf_size = UBIFS_KMALLOC_OK;
+                        c->max_bu_buf_len = UBIFS_KMALLOC_OK;
         }
         return 0;
 }
diff --git a/fs/ubifs/tnc.c b/fs/ubifs/tnc.c
index 99e9a744cfd0..6eef5344a145 100644
--- a/fs/ubifs/tnc.c
+++ b/fs/ubifs/tnc.c
@@ -1501,7 +1501,12 @@ out:
  * @bu: bulk-read parameters and results
  *
  * Lookup consecutive data node keys for the same inode that reside
- * consecutively in the same LEB.
+ * consecutively in the same LEB. This function returns zero in case of success
+ * and a negative error code in case of failure.
+ *
+ * Note, if the bulk-read buffer length (@bu->buf_len) is known, this function
+ * makes sure bulk-read nodes fit the buffer. Otherwise, this function prepares
+ * maxumum possible amount of nodes for bulk-read.
  */
 int ubifs_tnc_get_bu_keys(struct ubifs_info *c, struct bu_info *bu)
 {
diff --git a/fs/ubifs/ubifs.h b/fs/ubifs/ubifs.h
index 06ba51efd65d..870b5c479e95 100644
--- a/fs/ubifs/ubifs.h
+++ b/fs/ubifs/ubifs.h
@@ -969,7 +969,7 @@ struct ubifs_mount_opts {
  * @mst_node: master node
  * @mst_offs: offset of valid master node
  * @mst_mutex: protects the master node area, @mst_node, and @mst_offs
- * @bulk_read_buf_size: buffer size for bulk-reads
+ * @max_bu_buf_len: maximum bulk-read buffer length
  *
  * @log_lebs: number of logical eraseblocks in the log
  * @log_bytes: log size in bytes
@@ -1217,7 +1217,7 @@ struct ubifs_info {
         struct ubifs_mst_node *mst_node;
         int mst_offs;
         struct mutex mst_mutex;
-        int bulk_read_buf_size;
+        int max_bu_buf_len;
 
         int log_lebs;
         long long log_bytes;