ldiskfs/kernel_patches/patches/ext4-mballoc-pa_free-mismatch-rhel6.patch

   1 Index: linux-stage/fs/ext4/mballoc.c
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext4/mballoc.c  2011-03-14 16:34:39.790758415 +0800
   4 +++ linux-stage/fs/ext4/mballoc.c       2011-03-14 16:38:36.211681104 +0800
   5 @@ -3593,6 +3593,7 @@
   6         INIT_LIST_HEAD(&pa->pa_group_list);
   7         pa->pa_deleted = 0;
   8         pa->pa_type = MB_INODE_PA;
   9 +       pa->pa_error = 0;
  10
  11         mb_debug(1, "new inode pa %p: %llu/%u for %u\n", pa,
  12                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
  13 @@ -3654,6 +3655,7 @@
  14         INIT_LIST_HEAD(&pa->pa_group_list);
  15         pa->pa_deleted = 0;
  16         pa->pa_type = MB_GROUP_PA;
  17 +       pa->pa_error = 0;
  18
  19         mb_debug(1, "new group pa %p: %llu/%u for %u\n", pa,
  20                         pa->pa_pstart, pa->pa_len, pa->pa_lstart);
  21 @@ -3716,7 +3718,9 @@
  22         int err = 0;
  23         int free = 0;
  24
  25 +       assert_spin_locked(ext4_group_lock_ptr(sb, e4b->bd_group));
  26         BUG_ON(pa->pa_deleted == 0);
  27 +       BUG_ON(pa->pa_inode == NULL);
  28         ext4_get_group_no_and_offset(sb, pa->pa_pstart, &group, &bit);
  29         grp_blk_start = pa->pa_pstart - bit;
  30         BUG_ON(group != e4b->bd_group && pa->pa_len != 0);
  31 @@ -3752,19 +3756,27 @@
  32                 mb_free_blocks(pa->pa_inode, e4b, bit, next - bit);
  33                 bit = next + 1;
  34         }
  35 -       if (free != pa->pa_free) {
  36 -               printk(KERN_CRIT "pa %p: logic %lu, phys. %lu, len %lu\n",
  37 -                       pa, (unsigned long) pa->pa_lstart,
  38 -                       (unsigned long) pa->pa_pstart,
  39 -                       (unsigned long) pa->pa_len);
  40 +
  41 +       /* "free < pa->pa_free" means we maybe double alloc the same blocks,
  42 +        * otherwise maybe leave some free blocks unavailable, no need to BUG.*/
  43 +       if ((free > pa->pa_free && !pa->pa_error) || (free < pa->pa_free)) {
  44 +               ext4_error(sb, "pa free mismatch: [pa %p] "
  45 +                               "[phy %lu] [logic %lu] [len %u] [free %u] "
  46 +                               "[error %u] [inode %lu] [freed %u]", pa,
  47 +                               (unsigned long)pa->pa_pstart,
  48 +                               (unsigned long)pa->pa_lstart,
  49 +                               (unsigned)pa->pa_len, (unsigned)pa->pa_free,
  50 +                               (unsigned)pa->pa_error, pa->pa_inode->i_ino,
  51 +                               free);
  52                 ext4_grp_locked_error(sb, group,
  53 -                                       __func__, "free %u, pa_free %u",
  54 -                                       free, pa->pa_free);
  55 +                               __func__, "free %u, pa_free %u",
  56 +                               free, pa->pa_free);
  57                 /*
  58                  * pa is already deleted so we use the value obtained
  59                  * from the bitmap and continue.
  60                  */
  61         }
  62 +       BUG_ON(pa->pa_free != free);
  63         atomic_add(free, &sbi->s_mb_discarded);
  64
  65         return err;
  66 @@ -4450,6 +4462,24 @@
  67
  68         trace_ext4_request_blocks(ar);
  69
  70 +
  71 +       if (dev_check_rdonly(sb->s_bdev)) {
  72 +               struct block_device *bdev = sb->s_bdev;
  73 +
  74 +               printk(KERN_WARNING "Alloc from readonly device %s (%#x): "
  75 +                       "[inode %lu] [logic %llu] [goal %llu] [ll %llu] "
  76 +                       "[pl %llu] [lr %llu] [pr %llu] [len %u] [flags %u]\n",
  77 +                       bdev->bd_disk ? bdev->bd_disk->disk_name : "",
  78 +                       bdev->bd_dev, ar->inode->i_ino,
  79 +                       (unsigned long long)ar->logical,
  80 +                       (unsigned long long)ar->goal,
  81 +                       (unsigned long long)ar->lleft,
  82 +                       (unsigned long long)ar->pleft,
  83 +                       (unsigned long long)ar->lright,
  84 +                       (unsigned long long)ar->pright,
  85 +                       ar->len, ar->flags);
  86 +       }
  87 +
  88         /*
  89          * For delayed allocation, we could skip the ENOSPC and
  90          * EDQUOT check, as blocks and quotas have been already
  91 @@ -4529,6 +4559,25 @@
  92                         ac->ac_b_ex.fe_len = 0;
  93                         ar->len = 0;
  94                         ext4_mb_show_ac(ac);
  95 +                       if (ac->ac_pa) {
  96 +                               struct ext4_prealloc_space *pa = ac->ac_pa;
  97 +
  98 +                               /* We can not make sure whether the bitmap has
  99 +                                * been updated or not when fail case. So can
 100 +                                * not revert pa_free back, just mark pa_error*/
 101 +                               pa->pa_error++;
 102 +                               ext4_error(sb,
 103 +                                       "Updating bitmap error: [err %d] "
 104 +                                       "[pa %p] [phy %lu] [logic %lu] "
 105 +                                       "[len %u] [free %u] [error %u] "
 106 +                                       "[inode %lu]", *errp, pa,
 107 +                                       (unsigned long)pa->pa_pstart,
 108 +                                       (unsigned long)pa->pa_lstart,
 109 +                                       (unsigned)pa->pa_len,
 110 +                                       (unsigned)pa->pa_free,
 111 +                                       (unsigned)pa->pa_error,
 112 +                                       pa->pa_inode ? pa->pa_inode->i_ino : 0);
 113 +                       }
 114                 } else {
 115                         block = ext4_grp_offs_to_block(sb, &ac->ac_b_ex);
 116                         ar->len = ac->ac_b_ex.fe_len;
 117 @@ -4691,6 +4740,15 @@
 118                 goto error_return;
 119         }
 120
 121 +       if (dev_check_rdonly(sb->s_bdev)) {
 122 +               struct block_device *bdev = sb->s_bdev;
 123 +
 124 +               printk(KERN_WARNING "Release to readonly device %s (%#x): "
 125 +                       "[inode %lu] [block %llu] [count %lu] [is_meta %d]\n",
 126 +                       bdev->bd_disk ? bdev->bd_disk->disk_name : "",
 127 +                       bdev->bd_dev, inode->i_ino, block, count, metadata);
 128 +       }
 129 +
 130         ext4_debug("freeing block %llu\n", block);
 131         trace_ext4_free_blocks(inode, block, count, metadata);
 132
 133 Index: linux-stage/fs/ext4/mballoc.h
 134 ===================================================================
 135 --- linux-stage.orig/fs/ext4/mballoc.h  2011-03-14 16:32:50.859552482 +0800
 136 +++ linux-stage/fs/ext4/mballoc.h       2011-03-14 16:39:20.928429776 +0800
 137 @@ -20,6 +20,7 @@
 138  #include <linux/version.h>
 139  #include <linux/blkdev.h>
 140  #include <linux/mutex.h>
 141 +#include <linux/genhd.h>
 142  #include "ext4_jbd2.h"
 143  #include "ext4.h"
 144
 145 @@ -130,6 +131,7 @@
 146         ext4_grpblk_t           pa_free;        /* how many blocks are free */
 147         unsigned short          pa_type;        /* pa type. inode or group */
 148         spinlock_t              *pa_obj_lock;
 149 +       unsigned short          pa_error;
 150         struct inode            *pa_inode;      /* hack, for history only */
 151  };
 152