UBI: fix atomic LEB change problems
authorArtem Bityutskiy <Artem.Bityutskiy@nokia.com>
Thu, 13 Sep 2007 11:28:14 +0000 (14:28 +0300)
committerArtem Bityutskiy <Artem.Bityutskiy@nokia.com>
Sun, 14 Oct 2007 10:10:21 +0000 (13:10 +0300)
When the UBI device is nearly full, i.e. all LEBs are mapped, we have
only one spare LEB left - the one we reserved for WL purposes. Well,
I do not count the LEBs which were reserved for bad PEB handling -
suppose NOR flash for simplicity. If an "atomic LEB change operation"
is run, and the WL unit is moving a LEB, we have no spare LEBs to
finish the operation and fail, which is not good. Moreover, if there
are 2 or more simultanious "atomic LEB change" requests, only one of
them has chances to succeed, the other will fail with -ENOSPC. Not
good either.

This patch does 2 things:
1. Reserves one PEB for the "atomic LEB change" operation.
2. Serealize the operations so that only on of them may run
   at a time (by means of a mutex).

Pointed-to-by: Brijesh Singh <brijesh.s.singh@gmail.com>
Signed-off-by: Artem Bityutskiy <Artem.Bityutskiy@nokia.com>
drivers/mtd/ubi/eba.c
drivers/mtd/ubi/ubi.h

index 81bb6a33b555836a4f4bbe9d1f4cc0eb1caa66d3..7b7add6da12f31aeac6fab06299e6167e981f46a 100644 (file)
@@ -46,6 +46,9 @@
 #include <linux/err.h>
 #include "ubi.h"
 
+/* Number of physical eraseblocks reserved for atomic LEB change operation */
+#define EBA_RESERVED_PEBS 1
+
 /**
  * struct ltree_entry - an entry in the lock tree.
  * @rb: links RB-tree nodes
@@ -827,6 +830,9 @@ write_error:
  * data, which has to be aligned. This function guarantees that in case of an
  * unclean reboot the old contents is preserved. Returns zero in case of
  * success and a negative error code in case of failure.
+ *
+ * UBI reserves one LEB for the "atomic LEB change" operation, so only one
+ * LEB change may be done at a time. This is ensured by @ubi->alc_mutex.
  */
 int ubi_eba_atomic_leb_change(struct ubi_device *ubi, int vol_id, int lnum,
                              const void *buf, int len, int dtype)
@@ -843,11 +849,10 @@ int ubi_eba_atomic_leb_change(struct ubi_device *ubi, int vol_id, int lnum,
        if (!vid_hdr)
                return -ENOMEM;
 
+       mutex_lock(&ubi->alc_mutex);
        err = leb_write_lock(ubi, vol_id, lnum);
-       if (err) {
-               ubi_free_vid_hdr(ubi, vid_hdr);
-               return err;
-       }
+       if (err)
+               goto out_mutex;
 
        vid_hdr->sqnum = cpu_to_be64(next_sqnum(ubi));
        vid_hdr->vol_id = cpu_to_be32(vol_id);
@@ -864,9 +869,8 @@ int ubi_eba_atomic_leb_change(struct ubi_device *ubi, int vol_id, int lnum,
 retry:
        pnum = ubi_wl_get_peb(ubi, dtype);
        if (pnum < 0) {
-               ubi_free_vid_hdr(ubi, vid_hdr);
-               leb_write_unlock(ubi, vol_id, lnum);
-               return pnum;
+               err = pnum;
+               goto out_leb_unlock;
        }
 
        dbg_eba("change LEB %d:%d, PEB %d, write VID hdr to PEB %d",
@@ -888,17 +892,18 @@ retry:
 
        if (vol->eba_tbl[lnum] >= 0) {
                err = ubi_wl_put_peb(ubi, vol->eba_tbl[lnum], 1);
-               if (err) {
-                       ubi_free_vid_hdr(ubi, vid_hdr);
-                       leb_write_unlock(ubi, vol_id, lnum);
-                       return err;
-               }
+               if (err)
+                       goto out_leb_unlock;
        }
 
        vol->eba_tbl[lnum] = pnum;
+
+out_leb_unlock:
        leb_write_unlock(ubi, vol_id, lnum);
+out_mutex:
+       mutex_unlock(&ubi->alc_mutex);
        ubi_free_vid_hdr(ubi, vid_hdr);
-       return 0;
+       return err;
 
 write_error:
        if (err != -EIO || !ubi->bad_allowed) {
@@ -908,17 +913,13 @@ write_error:
                 * mode just in case.
                 */
                ubi_ro_mode(ubi);
-               leb_write_unlock(ubi, vol_id, lnum);
-               ubi_free_vid_hdr(ubi, vid_hdr);
-               return err;
+               goto out_leb_unlock;
        }
 
        err = ubi_wl_put_peb(ubi, pnum, 1);
        if (err || ++tries > UBI_IO_RETRIES) {
                ubi_ro_mode(ubi);
-               leb_write_unlock(ubi, vol_id, lnum);
-               ubi_free_vid_hdr(ubi, vid_hdr);
-               return err;
+               goto out_leb_unlock;
        }
 
        vid_hdr->sqnum = cpu_to_be64(next_sqnum(ubi));
@@ -1122,6 +1123,7 @@ int ubi_eba_init_scan(struct ubi_device *ubi, struct ubi_scan_info *si)
        dbg_eba("initialize EBA unit");
 
        spin_lock_init(&ubi->ltree_lock);
+       mutex_init(&ubi->alc_mutex);
        ubi->ltree = RB_ROOT;
 
        if (ubi_devices_cnt == 0) {
@@ -1183,6 +1185,14 @@ int ubi_eba_init_scan(struct ubi_device *ubi, struct ubi_scan_info *si)
                ubi->rsvd_pebs  += ubi->beb_rsvd_pebs;
        }
 
+       if (ubi->avail_pebs < EBA_RESERVED_PEBS) {
+               ubi_err("no enough physical eraseblocks (%d, need %d)",
+                       ubi->avail_pebs, EBA_RESERVED_PEBS);
+               goto out_free;
+       }
+       ubi->avail_pebs -= EBA_RESERVED_PEBS;
+       ubi->rsvd_pebs += EBA_RESERVED_PEBS;
+
        dbg_eba("EBA unit is initialized");
        return 0;
 
index cc010111264f5cb7c00d781009d9d9e8a809206a..5e941a6330303bc36271286a7861c4fdd5f63788 100644 (file)
@@ -221,14 +221,15 @@ struct ubi_wl_entry;
  * @vtbl_slots: how many slots are available in the volume table
  * @vtbl_size: size of the volume table in bytes
  * @vtbl: in-RAM volume table copy
+ * @vtbl_mutex: protects on-flash volume table
  *
  * @max_ec: current highest erase counter value
  * @mean_ec: current mean erase counter value
  *
- * global_sqnum: global sequence number
+ * @global_sqnum: global sequence number
  * @ltree_lock: protects the lock tree and @global_sqnum
  * @ltree: the lock tree
- * @vtbl_mutex: protects on-flash volume table
+ * @alc_mutex: serializes "atomic LEB change" operations
  *
  * @used: RB-tree of used physical eraseblocks
  * @free: RB-tree of free physical eraseblocks
@@ -308,6 +309,7 @@ struct ubi_device {
        unsigned long long global_sqnum;
        spinlock_t ltree_lock;
        struct rb_root ltree;
+       struct mutex alc_mutex;
 
        /* Wear-leveling unit's stuff */
        struct rb_root used;