bmap introduction

This section provides general information about the block map (bmap) necessary for understanding how bmaptool works. The structure of the section is:

  • "Sparse files" - the bmap ideas are based on sparse files, so it is important to understand what sparse files are
  • "The block map" - explains what bmap is
  • "Raw images" - the main usage scenario for bmaptool is flashing raw images, which this section discusses
  • "Usage scenarios" - describes various possible bmap and bmaptool usage scenarios

Sparse files

One of the main roles of a filesystem, generally speaking, is to map blocks of file data to disk sectors. Different file-systems do this mapping differently, and filesystem performance largely depends on how well the filesystem can do the mapping. The filesystem block size is usually 4KiB, but may also be 8KiB or larger.
 
Obviously, to implement the mapping, the file-system has to maintain some kind of on-disk index. For any file on the file-system, and any offset within the file, the index allows you to find the corresponding disk sector, which stores the file's data. Whenever we write to a file, the filesystem looks up the index and writes to the corresponding disk sectors. Sometimes the filesystem has to allocate new disk sectors and update the index (such as when appending data to the file). The filesystem index is sometimes referred to as the "filesystem metadata".
 
What happens if a file area is not mapped to any disk sectors? Is this possible? The answer is yes. It is possible and these unmapped areas are often called "holes". And those files which have holes are often called "sparse files".
 
All reasonable file-systems like Linux ext[234], btrfs, XFS, or Solaris XFS, and even Windows' NTFS, support sparse files. Old and less reasonable filesystems, like FAT, do not support holes.
 
Reading holes returns zeroes. Writing to a hole causes the filesystem to allocate disk sectors for the corresponding blocks. Here is how you can create a 4GiB file with all blocks unmapped, which means that the file consists of a huge 4GiB hole:
 $ truncate -s4G image.raw
 $ stat image.raw
   File: image.raw
   Size: 4294967296   Blocks: 0     IO Block: 4096   regular file
Notice that "image.raw" is a 4GiB file, which occupies 0 blocks on the disk! So, the entire file's contents are not mapped anywhere. Reading this file would result in reading 4GiB of zeroes. If you write to the middle of the image.raw file, you'll end up with 2 holes and a mapped area in the middle.
 
Therefore:
  • Sparse files are files with holes.
  • Sparse files help save disk space, because, roughly speaking, holes do not occupy disk space.
  • A hole is an unmapped area of a file, meaning that it is not mapped anywhere on the disk.
  • Reading data from a hole returns zeroes.
  • Writing data to a hole destroys it by forcing the filesystem to map corresponding file areas to disk sectors.
  • Filesystems usually operate with blocks, so sizes and offsets of holes are aligned to the block boundary.
It is also useful to know that you should work with sparse files carefully. It is easy to accidentally expand a sparse file, that is, to map all holes to zero-filled disk areas. For example, "scp" always expands sparse files, the "tar" and "rsync" tools do the same, by default, unless you use the "--sparse" option. Compressing and then decompressing a sparse file usually expands it.
 
There are 2 ioctl's in Linux which allow you to find mapped and unmapped areas: "FIBMAP" and "FIEMAP". The former is very old and is probably supported by all Linux systems, but it is rather limited and requires root privileges. The latter is a lot more advanced and does not require root privileges, but it is relatively new (added in Linux kernel, version 2.6.28).
 
Recent versions of the Linux kernel (starting from 3.1) also support the "SEEK_HOLE" and "SEEK_DATA" values for the "whence" argument of the standard "lseek()" system call. They allow positioning to the next hole and the next mapped area of the file.
 
Advanced Linux filesystems, in modern kernels, also allow "punching holes", meaning that it is possible to unmap any aligned area and turn it into a hole. This is implemented using the "FALLOC_FL_PUNCH_HOLE" "mode" of the "fallocate()" system call.

The bmap

The bmap is an XML file, which contains a list of mapped areas, plus some additional information about
the file it was created for, for example:
  • SHA256 checksum of the bmap file itself
  • SHA256 checksum of the mapped areas
  • the original file size
  • amount of mapped data

The bmap file is designed to be both easily machine-readable and human-readable. All the machine-readable information is provided by XML tags. The human-oriented information is in XML comments, which explain the meaning of XML tags and provide useful information like amount of mapped data in percent and in MiB or GiB.

So, the best way to understand bmap is to just to read it. Here is an example of a Tizen IVI 2.0 alpha snapshot bmap file. The vast amount of block ranges have been removed, though, to keep it shorter.

<?xml version="1.0" ?>
<!-- This file contains the block map for an image file, which is basically
     a list of useful (mapped) block numbers in the image file. In other words,
     it lists only those blocks which contain data (boot sector, partition
     table, file-system metadata, files, directories, extents, etc). These
     blocks have to be copied to the target device. The other blocks do not
     contain any useful data and do not have to be copied to the target
     device.

     The block map an optimization which allows to copy or flash the image to
     the image quicker than copying of flashing the entire image. This is
     because with bmap less data is copied: <MappedBlocksCount> blocks instead
     of <BlocksCount> blocks.

     Besides the machine-readable data, this file contains useful commentaries
     which contain human-readable information like image size, percentage of
     mapped data, etc.

     The 'version' attribute is the block map file format version in the
     'major.minor' format. The version major number is increased whenever an
     incompatible block map format change is made. The minor number changes
     in case of minor backward-compatible changes. -->

<bmap version="1.4">
    <!-- Image size in bytes: 3.7 GiB -->
    <ImageSize> 3998237696 </ImageSize>

    <!-- Size of a block in bytes -->
    <BlockSize> 4096 </BlockSize>

    <!-- Count of blocks in the image file -->
    <BlocksCount> 976133 </BlocksCount>

    <!-- Count of mapped blocks: 817.1 MiB or 21.4%  -->
    <MappedBlocksCount> 209168     </MappedBlocksCount>

    <!-- Type of checksum used in this file -->
    <ChecksumType> sha256 </ChecksumType>

    <!-- The checksum of this bmap file. When it is calculated, the value of
         the checksum has be zero (all ASCII "0" symbols).  -->
    <BmapFileChecksum> 46dd0b24ff90c05407455cc43f7eb65d219f7a8a23b9d980b041a0d97550e06d </BmapFileChecksum>

    <!-- The block map which consists of elements which may either be a
         range of blocks or a single block. The 'chksum' attribute
         (if present) is the checksum of this blocks range. -->
    <BlockMap>
        <Range chksum="ea66c5762cf7a63ecd98a5b972ed57d0ca319de9196613c29d92c16e82a9a603"> 0-4 </Range>
        <Range chksum="f50042546c89460967dfb71cbaa5aed14f18b7abe618eedc6be4c250aed5fb21"> 256-1919 </Range>
        <Range chksum="bf634e74431550f26129c00cf7f55d976adf0d71bcf97bf7f4772656521308f1"> 12544-13703 </Range>
        <Range chksum="5be706250b7a79b8bd58855b19ba91094d5b8fe9001fd62dc1294630a6819bd6"> 15616-16255 </Range>
        <Range chksum="919cfe81159e97899bd4a6dbda767a43efd8b395d39d63dbb9bf3dabc83b652a"> 16384-16511 </Range>
        <Range chksum="7c98d2da7afa2bdee9765910c81cb588e14781b77538fb54df84f7e7039e4348"> 16608-16634 </Range>
        <Range chksum="c7cbaf526b3ecc5b79ac8a993722a28995b7d0239ba210a3061b358a07a16894"> 16640-16881 </Range>
        <Range chksum="effdfd76ee7384e7b506985f8bd3b5d21cd1d1895c829fc10c3938013188b669"> 16883 </Range>
        <Range chksum="effdfd76ee7384e7b506985f8bd3b5d21cd1d1895c829fc10c3938013188b669"> 16885 </Range>
        <Range chksum="33fdf647973fde05156c38ba7dedd2b5bc874fa10e1313105fbf8aa9b50601dc"> 16889 </Range>
        <Range chksum="06308f48b3dd70980231896e3da0a3250c95a0fc7203e92c3d75f179df898dcd"> 16892-16893 </Range>
        <Range chksum="52a98324357718ec7554c38ccb7a1b5cafd38e0c012e72e40ae63020df7673db"> 16908-26665 </Range>
        <Range chksum="d15508719738ac16a38b1d328bed44b4966d11c6f25adb49ebce2dae22b882d5"> 49408-49409 </Range>
.....
        <Range chksum="30cb3c6546b0fcaf5d04947b20b68a2c46f606267ed54f093eeabd8382764df0"> 901376-901377 </Range>
        <Range chksum="8d1181e3d36e97d78876bb318e8fec2f3d1d4c50225687efeeea8e375d8d5815"> 976112-976132 </Range>
    </BlockMap>
</bmap>

Raw images

Raw images are the simplest type of system images which may be flashed to the target block device, block-by-block, without any further processing. Raw images just "mirror" the target block device: they usually start with the MBR sector. There is a partition table at the beginning of the image and one or more partitions containing filesystems, like ext4. Usually, no special tools are required to flash a raw image to the target block device. The standard "dd" command can do the job:
$ dd if=tizen-ivi-image.raw of=/dev/usb_stick

At first glance, raw images do not look very appealing because they are large and it takes a lot of time to flash them. However, with bmap, raw images become a much more attractive type of image. We will demonstrate this, using Tizen IVI as an example.

The Tizen IVI project uses raw images which take 3.7GiB in Tizen IVI 2.0 alpha. The images are created by the MIC tool. Here is a brief description of how MIC creates them:

  1. create a 3.7GiB sparse file, which will become the Tizen IVI image in the end
  2. partition the file using the "parted" tool
  3. format the partitions using the "mkfs.ext4" tool
  4. loop-back mount all the partitions
  5. install all the required packages to the partitions: copy all the needed files and do all the tweaks
  6. unmount all loop-back-mounted image partitions, the image is ready
  7. generate the block map file for the image
  8. compress the image using "bzip2", turning them into a small file, around 300MiB

The Tizen IVI raw images are initially sparse files. All the mapped blocks represent useful data and all the holes represent unused regions, which "contain" zeroes and do not have to be copied when flashing the image. Although information about holes is lost once the image gets compressed, the bmap file still has it and it can be used to reconstruct the uncompressed image or to flash the image quickly, by copying only the mapped regions.

Raw images compress extremely well because the holes are essentially zeroes, which compress perfectly. This is why 3.7GiB Tizen IVI raw images, which contain about 1.1GiB of mapped blocks, take only 300MiB in a compressed form. And the important point is that you  need to decompress them only while flashing. The bmaptool does this "on-the-fly".

Therefore:

  1. raw images are distributed in a compressed form, and they are almost as small as a tarball (that includes all the data the image would take)
  2. the bmap file and the bmaptool make it possible to quickly flash the compressed raw image to the target block device
  3. optionally, the bmaptool can reconstruct the original uncompressed sparse raw image file

And, what is even more important, is that flashing raw images is extremely fast because you write directly to the block device, and write sequentially.

Another great thing about raw images is that they may be 100% ready-to-go and all you need to do is to put the image on your device "as-is". You do not have to know the image format, which partitions and filesystems it contains, etc. This is simple and robust.

Usage scenarios

Flashing images

Flashing or copying large images is the main bmaptool use case. The idea is that if you have a raw image file and its bmap, you can flash it to a device by writing only the mapped blocks and skipping the unmapped blocks.
 
What this basically means is that with bmap it is not necessary to try to minimize the raw image size by making the partitions small, which would require resizing them. The image can contain huge multi-gigabyte partitions, just like the target device requires. The image will then be a huge sparse file, with little mapped data. And because unmapped areas "contain" zeroes, the huge image will compress extremely well, so the huge image will be very small in compressed form. It can then be distributed in compressed form, and flashed very quickly with bmaptool and the bmap file, because bmaptool will decompress the image on-the-fly and write only mapped areas.
 
The additional benefit of using bmap for flashing is the checksum verification. Indeed, the "bmaptool create" command generates SHA256 checksums for all mapped block ranges, and the "bmaptool copy" command verifies the checksums while writing. Integrity of the bmap file itself is also protected by a SHA256 checksum and bmaptool verifies it before starting flashing.
 
On top of this, the bmap file can be signed using OpenPGP (gpg) and bmaptool automatically verifies the signature if it is present. This allows for verifying the bmap file integrity and authoring. And since the bmap file contains SHA256 checksums for all the mapped image data, the bmap file signature verification should be enough to guarantee integrity and authoring of the image file.

Reconstructing sparse files

Generally speaking, if you had a sparse file but then expanded it, there is no way to reconstruct it. In some cases, something like
  $ cp --sparse=always expanded.file reconstructed.file

would be enough. However, a file reconstructed this way will not necessarily be the same as the original sparse file. The original sparse file could have contained mapped blocks filled with all zeroes (not holes), and, in the reconstructed file, these blocks will become holes. In some cases, this does not matter. For example, if you just want to save disk space. However, for raw images, flashing it does matter, because it is essential to write zero-filled blocks and not skip them. Indeed, if you do not write the zero-filled block to corresponding disk sectors which, presumably, contain garbage, you end up with garbage in those blocks. In other words, when we are talking about flashing raw images, the difference between zero-filled blocks and holes in the original image is essential because zero-filled blocks are the required blocks which are expected to contain zeroes, while holes are just unneeded blocks with no expectations regarding the contents.

Bmaptool may be helpful for reconstructing sparse files properly. Before the sparse file is expanded, you should generate its bmap (for example, by using the "bmaptool create" command). Then you may compress your file or, otherwise, expand it. Later on, you may reconstruct it using the "bmaptool copy" command.