On 2019/3/20 下午2:37, Qu Wenruo wrote:
[snip]
> +
> +     /*
> +      * Since btrfs device add doesn't check device size at all, we could
> +      * have device item whose size is smaller than 1M which is useless, but
> +      * still valid.
> +      * So here we can only check the obviously wrong case.
> +      */
> +     if (btrfs_device_total_bytes(leaf, ditem) == 0) {
> +             dev_item_err(fs_info, leaf, slot,
> +                          "invalid total bytes: have 0");
> +             goto error;
> +     }


Hi David,

Please remove this patch from misc-next queue.

Under the following call trace, we could create device with total_bytes
== 0;

btrfs_rm_device()
|- btrfs_shrink_device()
|  |- btrfs_device_set_total_bytes(device, 0)
|  |- btrfs_update_device()
|  |- btrfs_commit_transaction() #1
|- btrfs_rm_dev_item()

This will trigger write time tree checker warning.
And further more, this can create valid btrfs with device->total_bytes
== 0 and triggering read time tree-checker if power loss happens after
above transaction #1 but before next transaction.

So this dev item check is too restrict.

And furthermore, the error output is misleading, its devid is extracted
from key->objectid, but it should be key->offset.

For the fuzzed image, I'd like to fix it by either enhancing the seed
device lookup procedure.

Thanks,
Qu

> +     if (btrfs_device_bytes_used(leaf, ditem) >
> +         btrfs_device_total_bytes(leaf, ditem)) {
> +             dev_item_err(fs_info, leaf, slot,
> +                          "invalid bytes used: have %llu expect [0, %llu]",
> +                          btrfs_device_bytes_used(leaf, ditem),
> +                          btrfs_device_total_bytes(leaf, ditem));
> +             goto error;
> +     }
> +     /*
> +      * Remaining members like io_align/type/gen/dev_group aren't really
> +      * utilized.
> +      * Skip them to make later usage of them easier.
> +      */
> +     return 0;
> +error:
> +     return -EUCLEAN;
> +}
> +
>  /*
>   * Common point to switch the item-specific validation.
>   */
> @@ -632,6 +712,9 @@ static int check_leaf_item(struct btrfs_fs_info *fs_info,
>               ret = btrfs_check_chunk_valid(fs_info, leaf, chunk,
>                                             key->offset);
>               break;
> +     case BTRFS_DEV_ITEM_KEY:
> +             ret = check_dev_item(fs_info, leaf, key, slot);
> +             break;
>       }
>       return ret;
>  }
> diff --git a/fs/btrfs/volumes.c b/fs/btrfs/volumes.c
> index 645ffc9c47b0..7510272408e8 100644
> --- a/fs/btrfs/volumes.c
> +++ b/fs/btrfs/volumes.c
> @@ -4958,15 +4958,6 @@ static void check_raid56_incompat_flag(struct 
> btrfs_fs_info *info, u64 type)
>       btrfs_set_fs_incompat(info, RAID56);
>  }
>
> -#define BTRFS_MAX_DEVS(info) ((BTRFS_MAX_ITEM_SIZE(info)     \
> -                     - sizeof(struct btrfs_chunk))           \
> -                     / sizeof(struct btrfs_stripe) + 1)
> -
> -#define BTRFS_MAX_DEVS_SYS_CHUNK ((BTRFS_SYSTEM_CHUNK_ARRAY_SIZE     \
> -                             - 2 * sizeof(struct btrfs_disk_key)     \
> -                             - 2 * sizeof(struct btrfs_chunk))       \
> -                             / sizeof(struct btrfs_stripe) + 1)
> -
>  static int __btrfs_alloc_chunk(struct btrfs_trans_handle *trans,
>                              u64 start, u64 type)
>  {
> diff --git a/fs/btrfs/volumes.h b/fs/btrfs/volumes.h
> index 3ad9d58d1b66..38ed94b77202 100644
> --- a/fs/btrfs/volumes.h
> +++ b/fs/btrfs/volumes.h
> @@ -258,6 +258,15 @@ struct btrfs_fs_devices {
>
>  #define BTRFS_BIO_INLINE_CSUM_SIZE   64
>
> +#define BTRFS_MAX_DEVS(info) ((BTRFS_MAX_ITEM_SIZE(info)     \
> +                     - sizeof(struct btrfs_chunk))           \
> +                     / sizeof(struct btrfs_stripe) + 1)
> +
> +#define BTRFS_MAX_DEVS_SYS_CHUNK ((BTRFS_SYSTEM_CHUNK_ARRAY_SIZE     \
> +                             - 2 * sizeof(struct btrfs_disk_key)     \
> +                             - 2 * sizeof(struct btrfs_chunk))       \
> +                             / sizeof(struct btrfs_stripe) + 1)
> +
>  /*
>   * we need the mirror number and stripe index to be passed around
>   * the call chain while we are processing end_io (especially errors).
>

Reply via email to