JFFS3 design issues

JFFS3 design issues
Artem B. Bityutskiy
[email protected]
Version 0.32 (draft)
November 27, 2005
Abstract
JFFS2, the Journalling Flash File System version 2, is widely used in the embedded
systems world. It was designed for relatively small flash chips and has serious problems
when it is used on large flash devices. Unfortunately, these scalability problems are deep
inside the design of the file system, and cannot be solved without full redesign.
This document describes JFFS3 . a new flash file system which is designed to be
scalable.
Contents
1 JFFS2 overview 1
2 JFFS3 Requirements 2
3 Introduction to JFFS3 3
3.1 Indexing problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.2 Wandering trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.3 B-trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.4 Indexing in JFFS3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.5 Indexing example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.6 The Journal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.7 Garbage collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.8 The superblock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4 The tree 13
4.1 Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.2 Keys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.1 Trivial key scheme . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.2.2 Keys comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2.3 Key schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2.4 Keys compression . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Links . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5 Garbage Collection 20
6 The superblock 20
6.1 The superblock management algorithm . . . . . . . . . . . . . . . . . . . 20
6.2 The length of the chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.3 The superblock search . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
7 Issues/ideas/to be done 26
8 Definitions 27
9 Symbols 31
10 Abbreviations 32
11 Credits 32
12 References 33
1 JFFS2 overview
JFFS2, the Journalling Flash File System version 2 [1] is widely used in the embedded
systems world. JFFS2 was originally designed for small NOR flashes (less then about
32MB) and the first device with JFFS2 file system was a small bar-code scanner. Later,
when NAND flashes became widely used, NAND support was added to JFFS2. The first
NAND flashes were also small enough, but grew in size very quickly and are currently
much larger then 32MB (e.g., Samsung produces 2GB NAND flashes [4]).
JFFS2 has log-structured design, which basically means, that the whole file system
may be regarded as one large log. Any file system modification (i.e., file change, directory
creation, changing file乫s attributes, etc) is appended to the log. The log is the only data
structure on the flash media. Modifications are encapsulated into small data structures
called nodes.
So, JFFS2 is roughly a log, the log consists of nodes, each node contains a file system
modification. And this is basically all JFFS2 file system is. It is very simple from the
physical layout乫s standpoint. For more information about the design of JFFS2 and about
the log-structured design, look at [1], [2], and [3].
It is not the goal of this chapter to delve into details of JFFS2 but it is still wanted
to provide enough information to make it clear why JFFS2 has scalability problems and
why JFFS3 is needed. To keep this chapter simple, terms index or indexing information
are used.
The index is a crucial part of any file system as it is used to keep track of everything
that is stored in the file system. For example, the index may help to quickly locate the
addresses of physical blocks which correspond to the specified file at the specified offset,
or it helps to quickly find all the directory entries in a specified directory and so on.
For example, in case of ext2, the inode table, the bitmap and the set of direct, indirect,
doubly indirect and triply indirect pointers may be considered the index. In case of the
FAT file system, the File Allocation Table may be considered as the index, etc.
In traditional file systems the index is usually kept and maintained on the media, but
unfortunately, this is not the case for JFFS2. In JFFS2, the index is maintained in RAM,
not on the flash media. And this is the root of all the JFFS2 scalability problems.
Of course, as the index in kept in RAM, JFFS2 achieves extremely high throughput,
just because it does not need to update the index on flash after something has been
changed in the file system. And this works very well for relatively small flashes, for which
JFFS2 was originally designed. But as soon as one tries to use JFFS2 on large flashes
(starting from about 128MB), many problems come up.
At first, it is obvious that JFFS2 needs to build the index in RAM when it mounts
the file system. For this reason, it needs to scan the entire flash partition in order to
locate all the nodes which are present there. So, the larger is JFFS2 partition, the more
nodes it has, the longer it takes to mount it.
The second, it is evidently that the index consumes some RAM. And the larger is the
JFFS2 file system, the more nodes it has, the more memory is consumed.
To put it differently, if S is the size of the JFFS3 flash partition 1,
. JFFS2 mount time scales as O(S) (linearly);
. JFFS2 memory consumption scales as O(S) (linearly).
1Note, all the symbols used in this document are summarized in section 9
1
So, it may be stood that JFFS2 does not scale. But despite the scalability problems,
JFFS2 has many advantages, for example:
. very economical flash usage . data usually take as much flash space as it actually
need, without wasting a lot space as in case of traditional file systems for block
devices;
. admitting of 乭on-flight乭 compression which allows to fit a great deal of data to the
flash; note, there are few file systems which support compression;
. very good file system write throughput (no need to update any on-flash indexing
information as it simply does not exist there);
. unclean reboots robustness;
. good enough wear-leveling.
It is also worth noting here that there is a patch which is usually referred to as the
乭summary patch乭, that was implemented by Ferenc Havasi and was recently committed
to the JFFS2 CVS. This patch speeds up the JFFS2 mount greatly, especially in case of
NAND flashes. What the patch basically does is that it puts a small 乭summary乭 node
at the end of each flash erasable block. This node, roughly speaking, contains the copy
of headers of all the nodes in this eraseblocks. So, when JFFS2 mounts the file system,
it needs to glance to the end of each eraseblock and read the summary node. This
results in that JFFS2 only needs to read one or few NAND pages from the end of each
eraseblock. Instead, when there is no summary, JFFS2 reads almost every NAND page
in each eraseblock, because node headers are spread more or less evenly over eraseblocks.
Although the patch helps a lot, it is still a not scalable solution and it only relaxes the
coefficient of the JFFS2 mount time liner dependency. Let alone that it does not lessen
JFFS2 memory consumption.
2 JFFS3 Requirements
The following are the main user-level requirements JFFS3 has to meet.
R01 JFFS3 memory consumption must not depend on the size of JFFS3 partition, the
number of inodes in the file system, size of files, directories, and the like. Of course,
JFFS3 must be able to use the advantage of the available RAM, but only for
different kinds of caches which may be freed any time in case of memory pressure.
R02 JFFS3 have to provide very fast file system mount without the need to scan the
whole flash partition.
R03 JFFS3 have to provide good flash wear-levelling.
R04 JFFS3 must guarantee that unclean reboots cannot cause any file system corruption.
R05 JFFS3 must provide good enough performance.
R06 Unlike JFFS2, JFFS3 must implement write-behind caching for better performance.
2
R07 JFFS3 must gracefully deal with different kinds of data corruptions (flash bit-flips,
bad blocks may appear dynamically, etc).
R08 In case of serious corruptions it should be possible to reconstruct all the data which
were not damaged by means external tools like ckfs.jffs3.
R09 All the JFFS3 characteristics ought to scale not faster the logarithmic function.
JFFS2-like linear dependencies are not acceptable.
R10 JFFS3 must support extended attributes.
R11 JFFS3 must support the Access Control Lists feature (ACL).
R12 JFFS3 have to support on-flight compression.
R13 JFFS3 should provide good concurrency which means that it should be possible
to read the file system during Garbage Collection and to read/write during the
Journal Commit, read/write the file system simultaneously, etc.
3 Introduction to JFFS3
The main idea how to fix in JFFS2 to make it scalable is to move the index from RAM to
flash. Unfortunately, this requires complete JFFS2 redesign and re-implementation and
the design of JFFS3 is largely different to the design of JFFS2. This section discusses
the base JFFS3 design ideas without any detailed description.
3.1 Indexing problem
There is a large difference between block devices and flash devices in how they allow to
update the contents of a sector. Block devices admit of so-called 乭in-place updates乭, i.e.
the update may be written straight to the sector. Flash devices do not allow this unless
the whole eraseblock has been erased before.
Obviously, it is unacceptable to erase the whole eraseblock each time a sector is
updated. Instead, so-called 乭out-of-place updates乭 technique is usually used. This simply
means that no attempts to update sectors in-place are made but instead, updates are
written to some other sector and the contents of the previous sector is afterwords regarded
as garbage.
This 乭out-of-place writes乭 property of flash devices assumes that JFFS3 also has
log-structured design as in JFFS3 any update is written out-of-place. And it seems that
this is natural for any flash file system to have log-structured design.
It is interesting to notice that in log-structured file systems for block devices (like
the one described in [2]) not any update is 乭out-of-place乭. There are always some
fixed-position sectors present. These sectors usually refer the file system index, admit of
quick file system mount and they are updated in-place.
But flash devices have limited number of erase cycles for each eraseblock and it is
impossible to guaranty good wear-levelling if some eraseblocks are reserved for similar
purposes. So, it is important that in JFFS3 there are no in-place updates as good
wear-levelling is one of the main requirements to JFFS3 (see section 2).
3
The 乭out-of-place updates乭 property makes it difficult to maintain the index on the
flash media. Figure 1 demonstrates why.
A
B
C
D
Flash
A B C D D
1
D
1
Figure 1: JFFS3 indexing problem example.
Suppose the index is kept and maintained on flash and it consists of 4 parts A, B,
C, and D which refer each other: A refers B and C, B refers D, and C refers D. This
means, that A contains the physical flash address of B and C and so on.
Suppose D should be updated. Since it is updated out-of-place, the newer version
D1 is written to some other place. But there are B and C which still refer D, not D1,
and they ought to be updated as well. And when they are updated out-of-place, A will
still refer the old B and C, and so on. Thus, it is not that trivial to store and maintain
indexing information on the flash media.
3.2 Wandering trees
To address the above problem it is possible to use wandering trees. Figure 2 demonstrates
how do wandering trees work.
1. Suppose that the index is a tree and it is stored and maintained on the flash media.
The tree consists of nodes A, B, C, D, E, F, G, and H. Suppose node H should
be updated.
2. At first, the updated version H1 is written. Obviously, F still refers H.
3. Now the corresponding link in node F is changed and node F1 is written to flash.
F1 refers H1. But as F1 is also written out-of-place, A still refers the old node F.
4. Finally, the new root node A1 is written ant it refers F1.
5. Nodes A, F, H are now treated as garbage and the updated tree is composed by
nodes A1, B, C, D, E, F1, G, and H1.
So, wandering trees is the base idea of how the indexing information is going to be
maintained on the flash media in JFFS3. And it stands to reason that any tree may
be called 乭wandering tree乭 if any update in the tree requires updating parent nodes up
to the root. For example, it makes sense to talk about wandering Red-Black trees or
wandering B+-trees and so forth.
4
A
B E F
C D G H
(1)
A
B E F
C D G H
(2)
H
1
A
B E F
C D G H
(3)
H
1
F
1
A
B E F
C D G H
(4)
H
1
F
1
A
1
B E
C D G H
(5)
H
1
F
1
A
1
Figure 2: Wandering tree example.
3.3 B-trees
JFFS3 uses B+-trees and this subsection makes a short introduction to B+-trees. There
is a plenty of books where one may find more information. There are also many on-line
resources available, e.g. [6].
The inexact definition of B-tree may be formulated as a balanced search tree where
each node may have many children. The branching factor or the fanout defines the
maximal number of node乫s children. While B-trees may contain both useful data and
keys and links in non-leaf nodes, B+-trees are B-trees which store data only in leaf nodes,
while non-leaf nodes contain only keys and links.
. . .
... ...
. . .
Fanout (n)
Leaf nodes (level 0)
Depth
(L levels)
Root node (level L-1)
Figure 3: B+-tree example.
Figure 3 demonstrates a B+-tree with branching factor n and the number of level L.
5
Note, that in JFFS3 levels are numbered starting from leaf nodes (level 0) and ending
at the root node (level L . 1).
Leaf nodes in the B+-tree contain data which are indexed by keys. Non-leaf nodes do
not contain data, but contain only the indexing information, namely, keys and links.
...
Link 2
Link 3
Link n
Keys < Key 1
Keys > Key 1 and . Key 2
Keys . Key n-1
Link n-1
Link 1
Key 1
Key 2
Key 3
Key n-1
Keys > Key 2 and . Key 3
Keys > Key n-2 and . Key n-1
Figure 4: The structure of a non-leaf node in B+-tree.
Figure 4 depicts the structure of a non-leaf node. There are n links and n . 1 keys
in the node. Links may point to either leaf nodes or other non-leaf nodes. In the former
case, the leaf node will contain data which corresponds to the key which follows the link.
In the latter case, the pointed non-leaf node (and the whole subtree with the root in this
non-leaf node) will contain more keys in range (Key 1,Key 2].
Keys are sorted in the ascending order in non-leaf nodes, so it is not that difficult
to lookup data corresponding to any key. Furthermore, the tree is balanced, so the the
number of lookup steps does not depend on the key.
When objects are inserted or removed from the tree, re-balancing may be needed. The
tree is re-balanced by means of splitting nodes or merging them and there is a simple
enough algorithm exists. Please, refer to Donald Knuth乫s books for more information
about re-balancing B+-trees.
B+-trees are widely used when working with block devices (e.g., hard drives). Indeed,
these devices have a fixed input/output unit size (usually referred to as a sector ) and it
is natural to use B+-trees with node size multiple to the size of the sector in order to
store information on such devices.
3.4 Indexing in JFFS3
The way how JFFS3 stores and indexes the file system is similar to the approach used
by the Reiser4 file system (see [5]). All the file system objects (inodes, files, directory
entries, extended attributes, etc) are kept in one large B+-tree. Effectively, the whole
JFFS3 file system may be regarded as one large B+-tree. This tree is further referred to
just as 乭the tree乭.
Every object which is stored in the tree has a key, and objects are found in the tree
by their keys. To make it clearer what are object keys, the following is an example of
how they may look like:
6
. file data key: {inode number, offset};
. directory entry key: {parent directory inode number, direntry name hash} and the like.
The following are terms which are used in JFFS3 to refer nodes of different levels in
the tree:
. nodes of level 0 are leaf nodes;
. nodes of level 1 are twig nodes;
. nodes which are not the root, not leaf, and not twig are branch nodes;
. no-leaf nodes (i.e., the root, branch and twig) are indexing nodes.
Note, the same terminology (except indexing nodes) is used in the Reiser4 file system
[5].
Non-leaf nodes are called 乭indexing nodes乭 because they contain only indexing information,
nothing else. No file system data is kept in the indexing nodes. Indexing nodes
have fixed size which is equivalent to the flash sector size.
It is important to note that somewhat unusual terminology is used in this document.
The smallest input/output unit of the flash chip is called a sector. Since JFFS3 mainly
orients to NAND flashes, the sector is mostly the NAND page and is either 512 bytes
or 2 Kilobytes. For other flash types the sector may be different. If flash乫s minimal
input/output unit is very small (say, one bit as in case of NOR flash), there should be a
layer which emulates larger sectors (say, 512 bytes).
In opposite to indexing nodes, leaf nodes have flexible size, just like nodes in JFFS2.
So, roughly speaking, JFFS3 file system may be considered as JFFS2 file system (leaf
nodes) plus indexing information (indexing nodes) (see figure 5).
. . .
... ...
. . .
Leaf nodes have flexible size and the leaf level reminds JFFS2
Indexing nodes have fixed size
and index leaf nodes which
contain FS objects
Figure 5: The JFFS3 tree.
Similarly to JFFS2, leaf nodes consist of header and data. The header describes the
node data and contains information like the key of the node, the length, and the like.
7
Node data contains some file system data, for example a directory entry, file乫s contents,
etc.
Leaf and indexing nodes are physically separated, which means that there are eraseblocks
with only indexing nodes and with only leaf nodes. But of course, this does not
mean that the whole flash partition is divided on two parts, this only means that the
indexing and leaf nodes are not in one eraseblock. Figure 6 illustrates this.
...
Eraseblock with
indexing nodes
Eraseblock with
leaf nodes
Flash
Figure 6: An illustration of leaf and indexing nodes separation.
Eraseblocks which contain only indexing nodes are called indexing eraseblocks and
those with leaf nodes are called leaf eraseblocks.
The depth of the tree depends on how many objects are kept in the file system.
The more files, directories, etc are present in the file system, the deeper is the tree.
Fortunately, the number of tree levels grows very slowly with the growing number of file
system objects and the tree lookup scales as O(lognS) (logarithmically).
The following are advantages of the JFFS3 indexing approach.
. Many different key assignment schemes may be used and this gives a flexibility in
how objects are sorted in the tree. Thus, one may optimize JFFS3 for specific
workloads by means of changing the format of the keys.
. Leaf nodes may be compressed, so JFFS3 admits of the on-flight compression.
. In case of corruptions of the indexing information it is possible to re-create it by
means of scanning leaf nodes乫 headers.
. There is a clear separation between data and indexing information. This implies
that the indexing information and data may be cached separately, without overlapping
in the same cache lines. This leads to better cache usage as described in the
Reiser4 paper [5].
3.5 Indexing example
This section illustrates how does JFFS3 indexing work by means of a simple example.
The example is very rough but it shows JFFS3 indexing in action. It is assumed that
keys of direntries and data objects have the same layout that is mentioned in section 3.4.
Suppose that user does the following:
1. mounts JFFS3 file system to 乭/mnt/jffs3乭 directory;
2. issues 乭ls /mnt/jffs3乭 command;
3. reads the contents of 乭/mnt/jffs3/my file乭 file.
8
The following are comments about what is going on in JFFS3 during the above steps.
1. During mount JFFS3 locates the position of the root node. This is done with help
of the JFFS3 superblock which will be described later (see section 6).
2. To get the list of directory entries in the root directory, JFFS3 looks up all objects
matching the {2, } key pattern. Indeed, direntry keys have {parent inode #, name hash} format, the root directory inode number is 2 (or another predefined constant). 乭
means wildcard. Thus, JFFS3, {2, } will match to any direntry in the root directory.
3. To read the 乭my file乭 file, JFFS3 first needs to find out its inode number. The
inode number is stored in the directory entry object. Hence, JFFS3 reads my file乫s
direntry using {1, H("my file")} key (H() is the hash function).
Then JFFS3 searches for my file乫s data objects using {I, offset} keys. Depending
on which part of file should be read, offset may take different values.
The above description is somewhat simplified e.g., JFFS3 also needs to read my file乫s
attr-data object to fetch the inode length from there (see section 4.1), etc. But the aim
of the section is just to provide an idea of how JFFS3 indexing works.
3.6 The Journal
The JFFS3 tree is both B+-tree and wandering tree. Any file system change implies
that a new node is written to the flash media, which in turn, assumes that a number of
indexing nodes must be updated. Namely, the whole path of indexing nodes up to the
root node should be updated (see section 3.2).
Evidently, it is very expensive to update several indexing nodes on each file system
change and the journal provides a mechanism to avoid this.
The journal consists of a set of eraseblocks (the journal eraseblocks) which do not
have a fixed location on flash and are not contiguous on flash. Any flash eraseblock may
be used as an journal eraseblock.
...
Journal eraseblock Leaf eraseblock Indexing eraseblock
The journal tree
(in RAM)
Describes changes which
were not yet committed
Figure 7: The JFFS3 journal.
9
When something is changed in the JFFS3 file system, the corresponding leaf node is
written to the journal, but the corresponding indexing nodes are not updated. Instead,
JFFS3 keeps track of file system changes in RAM in a data structure called the journal
tree (see figure 7).
When something is read from the file system, JFFS3 first glimpses at the in-RAM
journal tree to figure out if the needed data is in the journal. If the data are there, the
journal is read, otherwise JFFS3 performs the usual tree lookup (see figure 8).
Lookup the journal tree
and find out if the
journal contains the
requested information.
Lookup the tree
Found
Read the journal
Not found
Read
request
JFFS3
Figure 8: Read request processing in JFFS3.
The journal is committed when it is full or in some other appropriate for JFFS3 time.
This means, that the indexing nodes corresponding to the journal changes are updated
and written to the flash. The committed journal eraseblocks are then treated as leaf
eraseblocks and new journal eraseblocks are picked by JFFS3 using the common JFFS3
wear-levelling algorithm.
The journal makes it possible to postpone indexing information updates to later and
potentially more appropriate time. It also allows to merge many indexing node updates
and lessen the amount of flash write operations.
When JFFS3 file system is being mounted, the journal should be read, 乭replayed乭
and the journal tree should be built. So, the larger is the journal, the longer it may
take to mount JFFS3. From the other hand, the larger is the journal, the more writes
may be deferred and the better performance may be achieved. By the other words, there
is a trade-off between the mount time and the performance and one may vary these
characteristics by means of changing the size of the journal.
3.7 Garbage collection
Garbage collection is a vital part of any log-structured file system. Over time, JFFS3
uses up all the flash free space and it needs to reclaim flash space occupied by garbage.
And the goal of Garbage Collector is to recycle garbage and reclame flash space which
it occupies. Since the only way to reclaim it is to erase the whole eraseblock, Garbage
Collector works in terms of eraseblocks.
JFFS2 Garbage Collector is quite simple and works in several steps.
1. To reclaim dirt from an eraseblock, JFFS2 moves all valid nodes from this eraseblock
to another eraseblock.
10
2. As nodes have changed their positions, the JFFS2 in-RAM index is adjusted.
3. The first eraseblock may be erased and re-used.
Note, JFFS2 (and JFFS3) aways reserves several eraseblocks in order to guarantee
that there are always some free eraseblocks available to perform garbage collection.
JFFS3 Garbage Collector is more complex. When valid data has been moved from
an eraseblock, the corresponding indexing nodes must be updated as well. Depending on
how many space Garbage Collector has reclaimed and how many space it has spent to
update indexing nodes, it might be that Garbage Collector produces more garbage that
it reclaims. This problem is demostrated in figure 9.
A B C
D E F G
A
B C
D E F G A B C
D E F G
1 2 Reserved for GC
D E F G A B C
1 2 Reserved for GC
D E F G
1 2 Reserved for GC
1.
2.
3.
Figure 9: JFFS3 garbage collection problem illustration.
There is a subtree with root in node A depicted in the figure. At the beginning
(snapshot 1), leaf nodes D, E, F, and G are situated in the eraseblock number 1. Indexing
nodes A, B, and C are situated in the eraseblock number 2. There are also two reserved
eraseblocks. Suppose all nodes have the same size equivlent to the size of sector which is
512 bytes in this example.
At snapshot 2 JFFS3 has decided to reclaim 512 bytes of dirty space form the eraseblock
number 1. Garbage Collector moves all the valid nodes from the eraseblock number
1 to one of the reserved eraseblocks. But as indexing nodes B and C still refer old
copies of the moved nodes in the eraseblock number 1, this eraseblock cannot be erased
so far. Indexing nodes A, B, and C have to be updated first.
At snapshot 3 Garbage Collector has updated indexing nodes A, B and C, putting
them to one of the reserved eraseblocks. From now on, old copies of nodes A, B, C, D,
E, F, and G at eraseblocks 1 and 2 comprise garbage. The eraseblock number 1 was
erased and is now free.
11
But unfortunatelly, the result is that Garbage Collector made more dirt that it reclaimed
space. Indeed, GC reclaimed 512 bytes while produced three times greater
amount of garbage (see the first three sectors at eraseblock 2, snapshot 3). Compare
snapshots 1 and 2.
Hence, it is obvious that garbage collection in JFFS3 is must be more complex that
in JFFS2. Chapter 5 discusses JFFS3 Garbage Collector in details.
3.8 The superblock
The JFFS3 superblock is a data structure that describes the file system as a whole and
contains important information like the offset of the root node, the journal eraseblocks,
etc. When the file system is being mounted, it first finds and reads the JFFS3 superblock.
In case of traditional file systems the superblock usually resides at a fixed position
on the disk and may be found very quickly. Conversely, due to the 乭out-of-place write乭
flash property it is impossible to assign a fixed position for the JFFS3 superblock. Things
are getting even more complex because of the need to provide good wear-levelling . it is
incorrect to just reserve several erasable blocks for the superblock unless it is guaranteed
that these eraseblocks will not be worn out earlier then the other eraseblocks.
We have the following two requirements that ought to be met in JFFS3:
. JFFS3 must be able to quickly find the superblock;
. the superblock management techniques must not spoil the overall flash wear levelling.
In the classical file systems the superblock usually contains a lot of static data which
is rarely updated and the superblock may have any size. In JFFS3, the superblock must
be updated quite often (e.g., each time the journal is committed). This means that to
lessen the amount of I/O, the JFFS3 superblock should be as small as it is possible,
namely, one sector. And there is no reason to keep any static data in the superblock
(e.g., the size of the file system, its version, etc). For static data, JFFS3 reserves the first
eraseblock of the JFFS3 partition.
Thus, the following terms are used in this document:
. static superblock . contains only static data which are never changed by JFFS3; the
static superblock resides at the static eraseblock; the static eraseblock is the first
non-bad eraseblock of the JFFS3 partition; it is supposed that the contents of the
static eraseblock may only be changed by external user-level tools;
. superblock . contains only dynamic data, is changed quite often and requires special
methods to deal with.
JFFS3 has a rather complicated superblock management scheme which makes it possible
to quickly find the superblock without full flash scanning when the file system is
being mounted. This scheme provides good flash wear-levelling. The superblock lookup
should take few milliseconds and scale as O(log2(S)). For more detailed information
about the superblock management scheme see section 6.1.
12
4 The tree
This chapter discusses the all the aspects related to the main JFFS3 entity . the tree.
Please, refer to section 3.4 for basic information about the JFFS3 tree.
4.1 Objects
JFFS3 keeps file system objects in the leaf level of the tree (in leaf nodes) and the following
is the list of supported objects.
1. Data objects contain files乫 data and are kept in data nodes. Each data node holds
one RAM page bytes of data (i.e., PAGE SIZE which is 4K on most 32-bit architectures).
But of course, in case of small files (less then one RAM page bytes) and
files乫 tails . less data data may be put to the data node.
3.1K
4K 4K 4K 1K
2.3K 3.7K
0.8K
A 13K file (logical view)
The JFFS3 tree
Figure 10: An illustration of files乫 data representation.
Figure 10 illustrates the correspondence between files乫 contents and data objects.
Each RAM page-size piece of a 13K file corresponds to a data node in the JFFS3
tree. The 1K tail of the file also corresponds to a data node. But because of
compression actual sizes of data nodes are less then the corresponding file fragments.
The division on RAM page-sized fragments relates to the Linux Virtual Memory
Management architecture. Namely, the Linux Page Cache works in terms of RAM
pages which means, that JFFS3 is always asked to read and write files乫 in units of
RAM page size.
It is worth noting that in order to optimize flash utilization, JFFS3 may store
multiple of RAM page bytes in one data node for static files. This admits of better
compression and leads to several other benefits.
13
2. Direntry objects contain the correspondence between directory entry names and
inode numbers. Direntry objects are stored in direntry nodes. Every directory
entry in the file system has a corresponding direntry object.
3. Attr-data objects contain attributes of inodes . both standard Unix attributes like
user ID, last modification time, inode length, etc and JFFS3-specific attributes like
the type of compression, etc. Each inode has only one corresponding attr-data
object.
4. Xentry objects contain the correspondence between names of extended attributes
and xattr IDs. Every extended attribute in the file system has a corresponding
xattr entry object. This is analogous to direntry objects, but direntries contain
{direntry name)inode number} mapping, instead of {xattr name)xattr ID} mapping in xentries.
Each extended attribute in JFFS3 has its own unique number . xattr ID, just like
every inode has its own unique inode number. And in fact, JFFS3 utilizes the same
space of numbers to enumerate inodes and extended attributes.
Xentry objects are stored in xentry nodes.
5. Xattr-data objects contain the data of extended attributes. The way how xattr-data
objects are kept in the tree is equivalent to the way how data objects a kept there.
Xattr-data objects are stored in xattr-data nodes.
6. Acl objects contain Access Control Lists (ACL) of inodes (information about ACLs
may be found out at [7]). Acl objects are stored in acl nodes.
In real-world systems a vast number of files have equivalent ACL while only few
files have unique ACL. For the former group of files (or more strictly . inodes)
JFFS3 makes use of shared acl objects. This means, that there is only one acl
object instance for all of these inodes. Shared acls are referred to from attr-data
objects of these inodes. If a shared acl is written to, a new acl object is created
(copy-on-write mechanism). Conversely, for the latter group there is a distinct acl
object per each inode.
4.2 Keys
Each object has its own key and may be quickly looked up in the tree by its key. As
there are 6 object types in JFFS3, there are also 6 key types:
1. data keys . index data objects;
2. direntry keys . index direntry objects;
3. attr-data keys . index attr-data objects;
4. xentry keys . index xentry objects;
5. xattr-data keys . index xattr-data objects;
6. acl keys . index acl objects.
14
4.2.1 Trivial key scheme
Lets start discussing JFFS3 keys with an example of a simple key layout which is further
referred to as the trivial key scheme. All keys in this scheme have the same 47-bit length
(see figure 11).
. Data keys consist of the 32-bit inode number the data belongs to, the unique 3-bit
key type identifier, and the 20-bit data offset.
. Direntry keys consist of the 32-bit parent directory inode number, the unique 3-bit
key type identifier, and the 20-bit direntry name hash value.
. Attr-data keys consist of the 32-bit inode number the attributes belong to, and the
unique 3-bit key type identifier.
. Xentry keys consist of the 32-bit inode number the extended attribute belongs to,
the unique 3-bit key type identifier, and the 20-bit extended attribute name hash
value.
. Xattr-data keys consist of the 32-bit xattr ID, the unique 3-bit key type identifier,
and the 20-bit extended attribute data offset.
. Acl keys consist of the 32-bit inode number the acl object belongs to, and the unique
3-bit key type identifier.
The following is the list of key type identifiers.
1. Data keys . 0 (000 bin);
2. Direntry keys . 1 (001 bin);
3. Attr-data keys . 2 (010 bin);
4. Xentry keys . 3 (011 bin);
5. Xattr-data keys . 4 (100 bin);
6. Acl keys . 4 (101 bin);
Since data objects may only contain multiple RAM pages of data (excluding small files
and files乫 tails) offsets in keys are always RAM page size-aligned. Assuming RAM page
is 4K (12 bits), 20 bits is enough to refer up to 4GB of data. To put it differently, the
trivial key scheme limits files乫 length to 4GB providing system乫s RAM page size is 4K.
It is also worth noting that several objects of the same type may have the same
key. For example, in case of hash collision, two direntries may have equivalent keys. In
this case objects may be distinguished by means of reading the corresponding leaf node
headers.
15
inode # (32) type (3) offset (20)
parent inode # (32) type (3) name hash (20)
inode # (32) type (3) offset (20)
inode # (32) type (3) name hash (20)
xattr ID (32) type (3) offset (20)
inode # (32) type (3) unused (20)
data key:
direntry key:
attr-data key:
xentry key:
xattr-data key:
acl key:
Figure 11: The trivial key scheme.
4.2.2 Keys comparison
The important topic is how keys are compared as it defines the relative order of objects
in the tree and is crucial for searching. Note, it only makes sense to compare keys of the
same type.
JFFS3 keys are usually comprised of one or more fields, i.e., keys K1 and K2 may be
represented as
K1 = {k1
1, k2
1, ..., kp
1},K2 = {k1
2, k2
2, ..., kp
2},
where p is the number of components in keys of this type.
Keys K1 and K2 are considered to be equivalent if and only if all their fields are
equivalent, i.e. ki1
= ki2
, i = 1, 2, ..., p.
Keys are compared field-by-field starting from the first field. If on i乫th step ki1
> ki2
,
then K1 is considered to be greater then K2. Similarly, if on i乫th step ki1
< ki2
, then K1
is considered to be less then K2.
4.2.3 Key schemes
Key schemes define layout of keys for all the 6 object types. Apparently, it would be too
inflexible to hardcode JFFS3 to support only one fixed key scheme. Indeed, there may
be a great deal of reasons why users may want to use different key schemes in different
situations . some examples go bellow.
. The inode number is encoded by a 32-bit integer in the trivial key scheme which
means that about 4 million inodes and extended attributes may exist on the file
system simultaneously. But in some cases this may be insufficient or, conversely,
too much and one may want to use less bits to encode inode numbers (say, only 24
bits), just for optimization.
16
. Similarly, offsets are encoded as 20-bit integers in the trivial key scheme which may
be insufficient when huge files (larger then 4G) should be supported. So one may
want to use more bits in certain cases.
. Depending on the concrete JFFS3 usage, different hash functions may be used in
direntry keys. The length of hash values may also vary depending on how many
directory entries are kept in directories. If there are huge directories with millions
of files there, long hash values should be used to avoid massive hash collisions (say,
64-bit hash values). But if it is known in advance that there will be no too large
directory entries, 2 the length of hash values may be shorter.
. It also possible that one may want to use some tricky key layouts to achieve different
kinds of optimization. For example, direntry keys may include the first 8 bytes
(64 bits) of the direntry name (see figure 12). In this case the getdents 3 Linux system
call will return direntries in 乭mostly乭 alphabetically sorted order and user-space
programs will not spend much time to sort them. In fact this technique is used in
the Reiser4 file system and it is claimed that slow sorting is a bottleneck in certain
real-life workloads. And the like.
. Different keys compression methods may be used in different key schemes (see
section 4.2.4 below).
parent inode # (32) type (3) name (64) name hash (19)
Figure 12: Direntry key layout example.
So it is obvious why JFFS3 is not attached to a fixed key scheme but instead, admits
of many different key schemes (one at a time of course) with a possibility to choose the
best suited key scheme.
4.2.4 Keys compression
The index is the most frequently re-written part of JFFS3. Indeed, every single change
at the leaf level of the tree requires re-writing L.1 indexing nodes. The number of index
updates is reduced by the write-behind cache and by the journal, but it is still changed
very often. So, it is extremely important for JFFS3 to keep the tree as shallow as it is
possible.
This means, that it makes sense to apply a sort of compression to keys in indexing
nodes. There are several ways to compress keys and the following are examples of possible
compression techniques.
Offsets coding. Offsets compression may be based on the observation that the overwhelming
majority of files in many file systems are small files. This means, that it
might makes sense to code smaller offsets by fewer bits.
2Note, when talking about directories, words 乭large乭 and 乭small乭 describe how many direntries are
kept in these directories. The more direntries a directory contains, the larger is it.
3See getdents (2) Linux manual pages
17
Table 1 contains an example of how offsets may be encoded. For offsets in range
0KB-8KB only 3 bits are enough, so the bit sequence 乭000乭 will encode offset 0,
and the bit sequence 乭001乭 will encode offset 4K 4. Offsets in range 8KB-64KB are
encoded by 6 bits and so on.
Offset range Bits in range Code prefix Code length
0KB-8KB 13 bits 00 3 bits
8KB-64KB 16 bits 01 6 bits
64KB-1MB 20 bits 10 10 bits
1MB-128MB 27 bits 110 18 bits
128MB-4GB 32 bits 111 23 bits
Table 1: An example of offset coding.
Inode number coding. If the approximate number of inodes on the file system is
known in advance, similar coding scheme may be exploited for inode numbers providing
JFFS3 may reuse deleted files乫 inode numbers.
Common prefix compression. In case of the trivial key scheme the first field of any
key is the inode number. Any other key scheme will likely also contain the inode
number in keys. If the inode number is the first component of the key, all keys
belonging to the same inode will go sequentially in indexing nodes. To put it
differently, there will be sequences of keys prefixed by the same inode number in
the tree.
inode # = 17 A
inode # = 17
Link inode # = 17 B Link inode # = 17 C Link
A Link B Link C Link
Key 1 Key 2 Key 3
Common Prefix Key 1 Key 2 Key 3
Figure 13: The common prefix compression idea illustration.
The evident compression method for these key sequences is to store the inode number
only once as the common prefix for the entire key sequence, instead of duplicating
it in every key. Figure 13 illustrates how does the prefix compression work.
Offsets sequence compression. In the trivial key scheme the last field of data keys
is the offset. The offset is multiple of RAM page size. Obviously, indexing nodes
will contain sequences of keys each of which describes data objects belonging to the
same file but with different offsets. Moreover, the keys will be ordered in increasing
key order.
4Recall, offsets in keys are RAM page-aligned and by default, the RAM page size is assumed to be
4K in this document
18
For sequences like this it is possible to only specify the starting offset, the ending
offset and the number of keys in the sequence, instead of wasting space storing the
offset in each key of the sequence.
0K 12K 4 A Link B Link C Link D Link
Key 1 Key 2 Key 3 Key 4
Start End Count Key1 Key2 Key3 Key4
A 0K Link B 4K Link C 8K Link A 12K Link
Figure 14: The Offsets sequence compression idea illustration.
Figure 14 presents an example of the offsets sequence compression method. Four
consecutive keys which describe four data objects belonging to the same inode may
be represented as a sequence of four keys without the offset field, but prefixed by
the starting offset, the ending offset and the number of keys in the sequence.
Note, the above compression methods may be combined to achieve better compression.
Because of compression, JFFS3 keys have variable size which means, that it is impossible
to directly apply the binary search algorithm to the contents of indexing nodes.
In JFFS3, indexing nodes are decompressed when read and are cached in decompressed
form. And after the indexing node has been decompressed, the binary search algorithm
is applicable.
We believe that keys compression will considerably reduce the amount of on-flash
indexing information and increase the overall performance just because the amount of
Input/Output will lessen. But only actual fixed-size keys vs. variable-size keys tests will
show if there is some real performance gain present.
4.3 Links
Links in JFFS3 have fixed length and are not compressed. The link width depends on
the size of JFFS3 partition . the larger is JFFS3 partition, the wider are links. Instead of
choosing a huge link width to suit the largest possible file systems (e.g. 64 bits), JFFS3
admits of flexible links width, depending on JFFS3 partition size.
As indexind nodes have fixed size equivalent to one sector, the width of links stored
in branch nodes and in the root nodes is
w = log2S . s.
Twig nodes refer variable-size leaf nodes so the width of links stored twih nodes is
w = log2S,
where S is the size of the JFFS3 partition and s is the size of sector.
19
5 Garbage Collection
Note! JFFS3 Garbage Collection is currently under development and this
chapter may be changed. Any suggestions and ideas are welcome.
6 The superblock
6.1 The superblock management algorithm
To implement the superblock management scheme, JFFS3 reserves the second and the
third good eraseblocks at the beginning of the flash partition (just next to the static
eraseblock). These two eraseblocks are called anchor eraseblocks, or the anchor area.
Anchor eraseblocks contain references to chain eraseblocks. Chain eraseblocks may
either refer other chain eraseblocks or the super eraseblock (see figure 15). The number
of chain eraseblocks varies depending on the size of the JFFS3 partition. If there are k
chain erase blocks, the anchor area will refer chain eraseblock 1, which will refer chain
eraseblock 2, which will refer chain eraseblock 3 and so forth. The chain eraseblock k will
refer the super eraseblock.
Anchor area chain eraseblock 1 chain eraseblock 2 super eraseblock
Figure 15: Types of eraseblocks involved to the superblock management scheme.
The super eraseblock contains the superblock which takes one sector. The chain
eraseblocks contain references to the next chain eraseblock or to the super eraseblock.
The JFFS3 superblock management mechanisms work as follows. Suppose there are k
chain eraseblocks in the current superblock management scheme. The superblock updates
are written to consecutive sectors of the super eraseblock. When the super eraseblock has
no more empty sectors, new super eraseblock is picked, the superblock update is written
to the new super eraseblock, and new reference is written to the chain eraseblock k.
Similarly, when there is no space in the chain eraseblock k, new chain eraseblock k is
picked and the corresponding reference is written to chain eraseblock k . 1, and so on.
When there are no free sectors in the chain eraseblock 1, new chain eraseblock 1 is picked
and the corresponding reference is written to the anchor area.
Figure 16 presents the example of the superblock management scheme (k = 2).
1. Initially, there are 2 chain eraseblocks (numbers 5 and 419) and the super eraseblock
(number 501). There is a reference in the first sector of the anchor area which refers
the chain eraseblock 1. The first sector of the chain eraseblock 1 refers the chain
eraseblock 2, and the first sector of the chain eraseblock 2 refers the super eraseblock.
The first sector of the super eraseblock contains the superblock.
2. After the superblock has been updated, the second sector of the super eraseblock
contains the valid copy of the superblock and the first sector contains garbage.
20
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 5)
chain eraseblock 2
(eraseblock 419)
super eraseblock
(eraseblock 501)
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 5)
chain eraseblock 2
(eraseblock 419)
super eraseblock
(eraseblock 501)
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 5)
chain eraseblock 2
(eraseblock 419)
super eraseblock
(eraseblock 501)
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 5)
chain eraseblock 2
(eraseblock 419)
super eraseblock
(eraseblock 7)
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 5)
chain eraseblock 2
(eraseblock 419)
super eraseblock
(eraseblock 100)
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 5)
chain eraseblock 2
(eraseblock 77)
super eraseblock
(eraseblock 398)
Anchor area
(eraseblocks 0 and 1) chain eraseblock 1
(eraseblock 65)
chain eraseblock 2
(eraseblock 44)
super eraseblock
(eraseblock 120)
Anchor area
(eraseblocks 0 and 1)
chain eraseblock 1
(eraseblock 101)
chain eraseblock 2
(eraseblock 150)
super eraseblock
(eraseblock 132)
1.
2.
3.
4.
5.
6.
7.
8.
Figure 16: The superblock management example.
3. The superblock has been updated many times and the valid superblock is at the
last sector of the super eraseblock while the other sectors of the super eraseblock
contain garbage.
21
4. As there were no free sectors at the super eraseblock, new super eraseblock was
chosen (eraseblock number 7) and the superblock update was written to the first
sector of the new super eraseblock. As the super eraseblock changed its position,
the corresponding reference at the chain eraseblock 2 was updated. It was updated
out-of-place and now the first sector of the chain eraseblock 2 is dirty while the
second sector contains the valid reference to the new super eraseblock.
5. The superblock has been updated many times and the super eraseblock changed
its position many times and it is currently at the eraseblock number 100. The
reference to the super eraseblock was also updated many times and at the moment
the last sector of the chain eraseblock 2 contains the valid reference while the other
sectors are obsolete. Similarly, the last sector of the super eraseblock contains valid
superblock while the other sectors are obsolete.
6. When the next superblock update came, there were no free sectors at the super
eraseblock and new super eraseblock was picked (eraseblock number 398) and the
valid copy of the superblock is currently at the first sector of the eraseblock number
398, Also, there were no free sectors at the chain eraseblock 2 and new chain
eraseblock 2 was picked (eraseblock number 77), so the first sector of the eraseblock
77 contains the valid reference to the super eraseblock. Since the chain eraseblock
2 changed its position, the corresponding reference at the chain eraseblock 1
was updated and at the moment the second sector of the chain eraseblock 1 contains
the valid reference to the chain eraseblock 2 while the first sector is dirty.
7. And analogously, after many superblock updates, the chain eraseblock 1 was updated
many times and when it became full it changed its position. Sure, the chain
eraseblock 2 and the super eraseblock changed their positions many times as well.
So, at the moment, the chain eraseblock 1 is at the eraseblock number 65, the
chain eraseblock 2 is at the eraseblock 44 and the super eraseblock is at the eraseblock
120. When the chain eraseblock 1 changed its position, the corresponding
reference at the anchor area was updated and currently the second sector of the
anchor eraseblock 1 contains the valid reference to the chain eraseblock 1 while the
firs sector is dirty.
8. And even more superblock updates happened. The anchor area was updated many
times. When there were no free sectors at the anchor eraseblock 1, the anchor
eraseblock 2 was used. So, at the moment, the valid reference to the chain eraseblock
1 is at the first sector of the anchor eraseblock 2. From now on, the first
anchor eraseblock may be erased and may be used again when the second anchor
eraseblock is full.
The following are important notes about the JFFS3 superblock management.
. The superblock takes one sector so the super eraseblock may be updated at most
N times (N is the number of sectors in the eraseblock).
. In case of NAND flash, the sector is the real minimal physical input/output unit, so
only N updates are possible in anchor eraseblocks and in chain eraseblocks. But if
the real input/output unit is smaller then the sector (i.e., if JFFS3 works on top of
22
NOR flash) the advantage of this may be used and more references may be packed
into one anchor or chain eraseblock.
. When JFFS3 picks new chain/super eraseblock, the common JFFS3 wear-levelling
scheme is utilized.
. Anchor area has 2 eraseblocks in order to ensure the tolerance to unclean reboots .
one anchor eraseblock may be safely erased while the other is being used.
. When a new reference is written to anchor/chain eraseblocks, the previous reference
becomes dirty and on mount JFFS3 should find the valid reference. To facilitate
this, each reference has its version number. Each subsequent reference has higher
version then the previous. Hence, JFFS3 may use the binary search algorithm to
quickly find the valid reference.
. As unclean reboot may happen anytime, no anchor/chain/super eraseblocks are
erased before the whole chain has been updated. This makes it possible to recover
from unclean reboots if they happen while the chain of the superblock-related
eraseblocks is being updated.
6.2 The length of the chain
The number of required eraseblocks in the superblock management scheme depends on
the size of the JFFS3 partition. The larger the partition, the more levels are needed.
This is determined by the need to ensure that the anchor area is not worn out earlier
then the rest of the JFFS3 partition.
Denote the number of required chain eraseblocks plus one (the super eraseblock) m
and calculate m assuming the worst case scenario: any file system data update requires
the superblock update. This would correspond to synchronous JFFS3 operation mode
with zero-length journal.
Obviously, what is wanted is to be sure that the anchor area is not worn out earlier
then the data area, i.e. the following inequality should be true:
TA
TD
> 1, (1)
where TA is the period of time of the total anchor area wear and TD is the period of
time of the total data area wear. Note, the whole JFFS3 partition excluding the static
superblock and the anchor area is referred to as the data area.
If RA is the average rate of the anchor area updates (sectors per second), RD s the
average rate of the data area updates and N is the number of sectors per the eraseblock,
then the anchor area will be written to with rate RA/N eraseblocks per second and the
data area will be written to with the rate RD/N eraseblocks per second. So, JFFS3 will
need to erase RA/N eraseblocks per second in the anchor area and RD/N eraseblocks per
second in the data area. Therefore, TA and TD may be expressed as
TA =
2D 丒 N
RA
,
TD =
(M . 3) 丒 D 丒 N
RD
,
23
where D is the maximum number of flash eraseblock erase cycles, and M is the number of
non-bad eraseblock on the JFFS3 partition. We subtracted 3 from M to get the number
of eraseblocks in the data area.
TA
TD
= 2 丒
RD
(M . 3) 丒 RA
. (2)
If m = 0, i.e., there are no chain/super eraseblocks and the superblock is stored in
the anchor area, then taking into account (2) and that in this case RA = RD = R, we
have
TA
TD
=
2
(M . 2)
.
Suppose m = 1. i.e., there are no chain eraseblocks and only the super eraseblock is
used. In this case each file system data update will require (a) the superblock update in
the data area and (b) the anchor area update. Therefore, the anchor area will be written
N times less frequently then when m = 0 and the data area will be written 2 times more
frequently then when m = 0. This means, that RA = R/N and RD = 2R and from (2)
we have
TA
TD
= 2 丒
2N
M . 3
.
When m = 2, i.e. the chain eraseblock 1 and the super eraseblock are used, the
anchor area will be written N2 times less frequently, while the data area will be written
2 + 1/N times more frequently then when m = 0 (one superblock update on each file
system update and one chain eraseblock 1 update per N superblock updates). Therefore,
RA = R/N2 and RD = (2 + 1/N) 丒 R and from (2) we have
TA
TD
= 2 丒
2N2 + N
M . 3
.
For m = 3, analogously,
TA
TD
= 2 丒
2N3 + N2 + N
M . 3
,
and for m = 0, 1, 2, . . .
TA
TD
= 2 丒
2Nm + Nm.1 + . . . + N
M . 3
.
Consequently, from (1) we have the following inequality:
2 丒
2Nm + Nm.1 + . . . + N
M . 3 > 1,
or neglecting the minor components,
4Nm
M . 3 > 1,
or
24
m > logN
M . 3
4
. (3)
Thus, form (3) it is obvious that the JFFS3 superblock management scheme scales
logarithmically.
Table 2 shows the value of m for different types of existing NAND flashes (see [10],
[11], [12], and [13]).
Type Size Sect. size M N m
Toshiba TC58DVM92A1FT 64MB 16KB 4096 32 2
Toshiba TH58NVG1S3AFT05 512MB 128KB 4096 64 2
ST Micro NAND08G-B 1GB 128KB 8192 64 2
Samsung K9K1G08X0B 2GB 128KB 16384 64 2
Table 2: The length of the JFFS3 superblock management chain for different types of
existing NAND flashes.
Note, providing that N = 64, m = 3 is enough to guarantee acceptable anchor area
wear leveling for up to 128GB flash, m = 4 . for up to 8TB flash (the inequality 3).
6.3 The superblock search
To find the superblock during mount, JFFS3 finds the valid reference in the anchor
eraseblocks, then finds the valid reference in chain erase blocks 1, 2, . . ., m . 1, and
finally finds the valid superblock in the super eraseblock. Since JFFS3 assigns versions
to records in anchor/chain/super eraseblocks and the versions are increased by one on
every update, the binary search algorithm may be used to quickly find the valid sector.
The valid reference in the anchor area may be found after log2(2N) + 2 steps (one
step involves one sector read operation), the reference in chain/super eraseblocks . after
log2(N) + 2 steps. Thus, to find the superblock, JFFS3 must read
S = 2m + log2(2N) + (m . 1) 丒 log2(N)
sectors.
Table 3 contains the approximate superblock search time for different existing NAND
flashes 5.
Type Size N m Sect. read S SB find
Toshiba TC58DVM92A1FT 64MB 32 2 50兪s 22 1.1ms
ST Micro NAND08G-B 1GB 64 2 130兪s 25 3.3ms
Samsung K9K1G08X0B 2GB 64 2 70兪s 25 1.6ms
Table 3: The superblock search time for different existing NAND flashes.
5the calculated superblock search time does not contain the ECC/CRC checking overhead as well as
any other CPU overhead.
25
For larger flash chips which would utilize the superblock management scheme with
m = 3 (no such flashes exist at the moment), the superblock search time would be about
4.3ms, providing the flash characteristics are the same as ST Micro乫s (see table 3).
7 Issues/ideas/to be done
This section contains a temporary list of issues which should be solved, ideas which
should be thought and analyzed deeper or things which were thought about but are not
yet described in this document.
The following is the list of things which should be thought about more.
1. Quota support. Will quota be supported? How will it look like . lust generic linux
quota or something better?
2. Transactions:
transaction open()/do many fs modifications()/transaction close() semantics?
Reiser4 pretends to support this via special sys reiser4() syscall. Would
be nice.
3. How can one select the compression mode on the per-inode basis? Xattrs with some
reserved name?
4. Orphaned files.
5. Holes.
6. Direct I/O.
7. How to chose/add a key scheme?
8. Extents.
The following is the list of topics which should be highlighted in this document as
well.
1. Garbage collection.
2. Tree balancing.
3. Tree locking.
4. Caching, write-behind cache.
5. An assumed flash model and the model of interactions between JFFS3 and the flash
I/O subsystem.
6. How the track of eraseblocks will be kept? Space accounting, good/bad, erase
count?
7. The wear-levelling algorithms.
8. The format of keys.
26
9. Branch nodes乫 links are sector numbers, twig nodes乫 links are absolute flash offsets.
So, the length of twig and branch keys are different and branches have greater
fanout.
10. Different optimizations may be achieved by means of changing the format of keys.
So, JFFS3 should be flexible in this respect and have a mechanism to change/select
the formats of keys.
11. The minimal amount of file乫s data in a node is PAGE SIZE. No way to create smaller
nodes as it it possible in JFFS2.
12. Portability (e.g., move FS between machines with different RAM page size, etc).
13. Errors handling.
14. Bad blocks handling.
The following is the list of ideas which were thought about but are not yet in the
document.
1. If the compression is disabled for an inode, then its nodes are (PAGE SIZE + header
size) in size, i.e., they do not fit into integer number of flash sectors. For these
nodes we may keep the header in the OOB area. In this case we should not mix
compressed nodes and uncompressed nodes in one eraseblock.
2. For large files which are mostly read-only, we may fit more then one page of data
in one node. This will mace compression better. When the file is read, all the
uncompressed pages are propagated to the page cache, like in the zisofs file system.
3. If there are few data in the superblock, we may keep this data in the root node. In
this case the root will have smaller fanout then branches.
The 乭to do乭 list.
1. Re-calculate digits for SB search time and m.
2. For now only the idea of keys compression methods is provides. Would be nice to
describe algorithms more strictly.
8 Definitions
1. Access Control Lists, ACL . a modern mechanism to control accesses to files
which provides much more flexibility that the standard Unix mechanism of owner/group/others
permissions, see [7] for more details.
2. Acl . an object in the tree containing inode乫s ACL. Refer to section 4.1 for more
information.
3. Acl key . acl object乫s key.
4. Acl node . a leaf node containing an acl object.
27
5. Anchor eraseblock, anchor area . the second and the third good eraseblocks
of the JFFS3 partition which are reserved for the superblock management. See
section 6.1 for more details.
6. Attr-data . an object in the tree where inode乫s attributes are stored (standard
Unix attributes like creation time, owner ID, etc and other attributes like the type
of compression, etc). Refer to section 4.1 for more information.
7. Attr-data key . attr-data object乫s key.
8. Attr-data node . a leaf node containing an attr-data object.
9. B-tree . a balanced search tree where each node has many children. See section 3.3.
10. B+-tree . a B-tree where no data is stored in non-leaf nodes but instead, is stored
only in leaf nodes.
11. Branch node . any node that is not leaf, not twig and not root.
12. Branching factor . the branching factor of the B-tree is the number of children
of a node.
13. Chain eraseblock . an eraseblock containing references to other chain eraseblocks
or to the super eraseblock. Chain eraseblocks facilitate quick SB searching and
are the part of the JFFS3 superblock management scheme (see section 6.1). The
main reason why chain eraseblocks are needed is the need to provide good flash
wear-levelling.
14. Clean eraseblock . an eraseblock which contains no garbage, only valid information.
15. Data area . the whole JFFS3 partition excluding the static superblock and anchor
eraseblocks.
16. Data key . data object乫s key.
17. Data node . a leaf node with file乫s data.
18. Directory entry, direntry . basically an association between the name and the
inode number.
19. Direntry key . direntry object乫s key.
20. Direntry node . a leaf node containing a direntry object.
21. Dirt, dirty space . information on flash which is not valid due to out-of-place
updates or objects deletion. It is the aim if the Garbage Collector to reclaim the
space occupied by dirt.
22. Dirty eraseblock . an eraseblock which contains some dirt along with valid nodes.
23. Dirty sector . a sector which contains dirt.
28
24. Erasable block, eraseblock . the minimal erasable unit of the flash chip from
the JFFS3乫s viewpoint.
25. Extended attributes, xattr . an association between names and data for files
and directories. See attr(5) Linux manual pages for more information.
26. Fanout . the same as branching factor.
27. Free eraseblock . an erased eraseblock (contains only 0xFF words).
28. Garbage . the same as dirt.
29. Garbage Collector . a part of any Flash File System which is responsible for
recycling dirty space and producing free eraseblocks.
30. Indexing information, index . data structures which do not contain any file
system data (files, directories, extended attributes, etc) but instead, keep track of
this data. For example, indexing information allows to quickly find all the directory
entries for any specified directory. In case of the FAT file system, the File Allocation
Table is may be treated as the index, in case of the ext2 file system the inode table,
the bitmap and the set of direct, indirect, doubly indirect and triply indirect pointers
may be considered as the index. In JFFS3, the index is comprised by the indexing
nodes. See section 3.4 for more information.
31. Indexing eraseblock . an eraseblock which contains indexing nodes.
32. Indexing node . a non-leaf node. Indexing nodes have fixed size (one sector) and
contain only keys and links.
33. In-place updates, in-place writes . a method of updating on-media data when
the update is written to the physical position where the data resides (in opposite
to out-of-place updates).
34. Journal . contains recent JFFS3 changes and all the file system updates first go
to the journal. The purpose of the Journal is to accumulate a bunch of JFFS3
file system changes and to postpone updating the index. See section 3.6 for more
information.
35. Journal commit . the process of re-building the indexing information for the
data which is in the journal. After the journal has been committed the journal
eraseblocks become just leaf eraseblocks.
36. Journal eraseblock . an eraseblock containing the journal data.
37. Journal tree . an in-memory tree referring Journal nodes which were not committed
so far. When JFFS3 reads, it first looks up the journal tree to find out whether
the searched information is there. See section 3.6 for more details.
38. Key . an identifier of objects in the tree.
39. Key type . an unique identifier of the key type. There are 6 key types in JFFS3
(see section 4.3).
29
40. Leaf eraseblock . an eraseblock containing leaf nodes.
41. Leaf node . any node from the leaf level of the tree (level 0). Leaf nodes contain
only data and do not further refer other nodes. For more information see section 3.4.
42. NAND page . a basic input output unit of NAND flash chips. ECC is calculated
on the per-NAND page basis. See any NAND flash manual for more details, e.g., 10.
43. Node . a pile of the tree (the tree consists of nodes) as well as the container for file
system data. There are different types of nodes in JFFS3. For more information
see section 3.4.
44. Obsolete nodes/data/sectors . the same as dirty nodes, data or sectors.
45. Out-of-place updates, out-of-place writes . a sort of data updates when the
update is not written to the same physical position, but instead, is written to some
other place and the previous contents is treated as garbage afterwords. Opposite
to in-place updates.
46. RAM page . an unit of memory management in Virtual Memory Management
subsystem of most modern operating systems (including Linux). See [9] for more
information.
47. Sector . the smallest writable unit of the flash chip from JFFS3乫s viewpoint. May
be equivalent to the minimal physical input/output unit (like in case of NAND
flashes) or larger (like in case of NOR flashes).
48. Shared acl . an acl object which is shared by many inodes for optimization purposes.
Refer to section 4.1 for more information.
49. Static eraseblock . the fist good erasable block of the JFFS3 partition where the
file system static data is stored. JFFS3 may only read it and it is created/changed
by external formatting tools.
50. Superblock . a data structure which describes the whole JFFS3 file system. Only
dynamic data is stored in the superblock, all the static data is kept in the static
superblock. There is a comprehensive superblock management scheme in JFFS3,
see section 6.1.
51. Super eraseblock . an eraseblock where the superblock is kept. See section 6.1
details.
52. Target eraseblock . the eraseblock which is currently being processed by the
Garbage Collector, i.e., nodes are moved from the target eraseblock and it is erased
afterwords.
53. Quota . a mechanism which allows to assign different limits on the file system
(e.g., restrict users in the number of files they may create or in the amount of space
they may consume, etc). See [8] for more details about quota support in Linux.
54. Tree . the main entity JFFS3 design revolves about. The JFFS3 tree is a wandering
B+-tree where all the file system stuff (files, directories, extended attributes, etc)
is stored.
30
55. Twig nodes . nodes which reside one level upper then leaf nodes (level 1).
56. Valid nodes . nodes which contain actual information, non-obsolete nodes.
57. Wandering tree . a method of updating trees when there is no possibility to
perform in-place updates. The JFFS3 tree is a wandering B+-tree. See section 3.2
for more information.
58. Xattr . a widely used contracted form for extended attributes.
59. Xattr-data . an object in the tree containing the data of an extended attribute.
60. xattr-data key . xattr-data object乫s key.
61. Xattr-data node . a leaf node containing an attr-data object.
62. Xattr ID . an unique identifier of an extended attribute. Refer to section 4.1 for
more information.
63. Xentry . an object in the tree which stores the association between the name of
an extended attribute and its xattr ID. Refer to section 4.1 for more information.
64. Xentry key . xentry object乫s key.
65. Xentry node . a leaf node containing a xentry object.
9 Symbols
The following is the list of symbols which are used to denote different things thought this
document.
. D . the number of guaranteed erases of flash eraseblocks (typically  105 for NAND
flashes);
. H() . the hash function JFFS3 uses to calculate names乫 hash for keys.
. I . inode number.
. K, Kx . tree乫s keys.
. k, kx . keys乫 fields.
. L . the number of levels in the tree.
. m . the number of eraseblocks used in the superblock management scheme without
the anchor eraseblocks, i.e. the number of chain eraseblocks plus one (the super
eraseblock).
. M . the total number of non-bad eraseblocks on the JFFS3 partition.
. n . the branching factor (fanout) of the tree.
. N . the number of sectors per eraseblock.
31
. S . the size of the JFFS3 flash partition (assuming there are no bad block).
. s . the size of sector.
. w . the bit-width of links.
10 Abbreviations
1. ACL . Access Control List
2. ECC . Error Correction Code
3. CRC . Cyclic Redundancy Check
4. JFFS2 . Journalling Flash File System version 2
5. JFFS3 . Journalling Flash File System version 3
6. MTD . Memory Technology Devices
7. RAM . Random Access Memory
8. VFS . Virtual File System
11 Credits
The following are the people I am very grateful for help (alphabetical order):
. David Woodhouse < [email protected]> . the author of JFFS2, answered a
great deal of my questions about MTD and JFFS2 and suggested some interesting
ideas for JFFS3.
. Joern Engel < [email protected]> . discussed some aspects of a new
scalable flash file system with me. Joern is developing his own flash file system
LogFS.
. Nikita Danilov < [email protected]> . used to work in Namesys and implemented
ReiserFS and Reiser4 file systems. Nikita answered my questions about
Reiser4 FS internals.
. Thomas Gleixner < [email protected]> . helped me with MTD-related things,
especially concerning flash hardware and low-level flash software.
. Victor V. Vengerov < [email protected]> . my colleague from OKTET Labs
who discussed some JFFS3 design approaches with me and suggested several interesting
ideas.
32
12 References
1. JFFS : The Journalling Flash File System,
http://sources.redhat.com/jffs2/jffs2-html/
2. The Design and Implementation of a Log-Structured File System,
http://www.cs.berkeley.edu/~brewer/cs262/LFS.pdf
3. Who wants another filesystem?,
http://cgi.cse.unsw.edu.au/~neilb/conf/lca2003/paper.pdf
4. Samsung Flash memory products,
http://www.samsung.com/Products/Semiconductor/Flash/index.htm
5. Reiser4 File System, http://www.namesys.com/
6. B-Trees, http://www.bluerwhite.org/btree/
7. POSIX Access Control Lists on Linux,
http://www.suse.de/~agruen/acl/linux-acls/
8. Quota mini-HOWTO, http://www.tldp.org/HOWTO/Quota.html
9. Wikipedia: Virtual Memory, http://en.wikipedia.org/wiki/Virtual_memory
10. Toshiba TC58DVM92A1FT NAND flash chip, http:
//www.toshiba.com/taec/components/Datasheet/TC58DVM92A1FT_030110.pdf
11. Toshiba TH58NVG1S3AFT05 NAND flash chip, http://www.toshiba.com/
taec/components/Datasheet/TH58NVG1S3AFT05_030519A.pdf
12. ST-micro NAND08G-B NAND flash chip,
http://www.st.com/stonline/products/literature/ds/11241.pdf
13. Samsung K9K1G08X0B NAND flash chip,
http://www.samsung.com/Products/Semiconductor/NANDFlash/SLC_
SmallBlock/1Gbit/K9K1G08U0B/ds_k9k1g08x0b_rev02.pdf
33