Related Documents
 AppArmor Technical documentation
 Format of the DFA/HFA
 PolicyDB the encoding of policy rules into the HFA
 Encoding permissions
 Putting it altogether relationship of Policy and DFA/HFA
 Steps in generating the DFA/HFA
A Brief explanation of eHFAs in AppArmor
AppArmor uses an extened Hybrid Finite Automata (eHFA) as its matching engine. An HFA is merely a finite automata (NFA) that has been specially constructed so that it is made up of finite automatas (dfas) that are joined by special transitions or border states. In its simpliest form an HFA can be reduced to a single DFA. The “extended” prefix indicates that the HFA has been extended beyond traditional nfas/dfas with additional functionality (currently limited to dynamic runtime variable matching).
AppArmor 2.1 was the first iteration of apparmor to use a dfa to perform its pattern matching as they provide a fast matching that is linear to the input stream. It has since been extended to be an eHFA.
AppArmor is capable of generating dfa graphs that can help understand what is happening in the apparmor_parser back end.
 circles represent a state
 { } within a circle are the set of expression tree nodes that make the state
 ( ) within a circle are optional and indicate an accepting state. The number is the permission set for the state
 characters along a given edge indicate valid transitions. If there is input that doesn't match a valid transition then the dfa transitions to a null nonaccepting state that can not be exited (ie it will consume all input).
Issues with AppArmor's DFA
The dfa that apparmor uses, is a simple direct representation with the most basic of table compression schemes. It also does not allow for in line variable matching or other advances that are desirable.
Specific issues with the current implementation
 it generates large tables that need to be loaded into kernel consuming memory that could be used else where
 it takes a relatively long time to compute
 the generation of the dfa without the factoring steps can be impractical even on modern hardware
 the dfa is subject to exponential state explosion.
 equivalence classes (if used) are applied post generation
 the table compression algorithm is a simple and straight forward with O(n^2) running time (slow). The compression achieved is not perfect but is fairly good given the limits of the compression format.
 the table compression format is simple and direct and does not leverage any state or transition redundancies, resulting in poor compression compared to what could be achieved
 The implementation needs to be cleaned up
Overall the dfa engine needs to be updated and replaced with a better atomata representation that keeps many of the dfas properties while being flexible, faster to compute (control state explosion), and leverages a better compression format. Better alternative exist so it is only a matter of examining the alternatives and implementing.
Rule to DFA conversion
AppArmor file rules are compiled into a dfa that is loaded into the kernel to do path matching. To do this AppArmor transforms the rules in the profiles, going through several steps to create the final dfa. The pipeline is as follows
profile rules

v
sort profile rules

v
merge duplicate rules

v
convert rules to expression tree

v
expression tree simplification

v
dfa creation

v
dfa minimization

v
dfa unreachable state removal

v
creation of equivalence

v
transition table compression
To illustrate the steps of the conversion pipeline the following simple profile is used through out the following discussion
/usr/bin/example {
/etc/passwd r,
/home/*/** rl,
/home/*/bin/ ix,
/home/likewise/*/*/** rwl,
/{usr,}/bin/** px,
/etc/passwd r, # duplicate
/home/*/** w, # duplicate
}
rule sorting
The first step is sorting path based rules this allows for simple merging can take place, and also aids in regular expression factoring as it put similar expressions close to each other in the tree.
/usr/bin/example {
/etc/passwd r,
/etc/passwd r, # duplicate
/home/*/** rl,
/home/*/** w, # duplicate
/home/*/bin/ ix,
/home/likewise/*/*/** rwl,
/{usr,}/bin/** px,
}
merge duplicates rules
Rules with the exact same path description are merged, merging permission as well. There are exceptions in rule merging x permission and rules that have a pair, like link rules may actually be separated. This step reduces the number of rules going through pcre conversion and also the amount of rules that expression simplifications needs to consider. The permissions merging done in this step results in smaller expression trees (tree simplification can not do permission merging as it does evaluate at the rule level), and hence smaller dfas being constructed.
/usr/bin/example {
/etc/passwd r,
/home/*/** rwl,
/home/*/bin/ ix,
/home/likewise/*/*/** rwl,
/{usr,}/bin/** px,
}
rule to expression tree conversion
AppArmor globbing rules are converted into a single pcre expression tree, that represents the rules for entire profile. This is done by first converting an individual rule into and expression tree and then using an alternation to merge it into the main expression tree (this is equivalent to combining two rules with an alternation).
Early versions of AppArmor this step is split into two parts, a text conversion to pcre syntax, and then a pcre parse tree is generated from that; e.g.:
rule1 r, rule2 w,
can be combined as(rule1 rrule2 w)
The handling of rule permissions in tree conversion ???
Note: each parse tree includes a fake node containing the permission of the rule.
This step can be examined in AppArmor 2.4 and later by doing
apparmor_parser QT D exprtree <profile>
E.G., for a single rule from a profile do:
echo "profile example { /home/*/** rl,} "  ./apparmor_parser QT D exprtree
DFA: Expression Tree
(/home/[^\0000/]([^\0000/])*/[^\0000/]([^\0000])*(((<4><16>)<65536>)<262144>)/home/[^\0000/]([^\0000/])*/[^\0000/]([^\0000])*\0000/[^/](.)*((<16><32>)<262144>))
E.G.; for the example profile above:
./apparmor_parser QT D exprtree usr.bin.example
DFA: Expression Tree
((((((/etc/passwd(<4><65536>)/home/[^\0000/]([^\0000/])*/[^\0000/]([^\0000])*(((((((<2><4>)<8>)<16>)<32768>)<65536>)<131072>)<262144>))/home/[^\0000/]([^\0000/])*/[^\0000/]([^\0000])*\0000/[^/](.)*((<16><32>)<262144>))/home/[^\0000/]([^\0000/])*/bin/(((<513><64>)<8404992>)<1048576>))/home/likewise/[^\0000/]([^\0000/])*/[^\0000/]([^\0000/])*/[^\0000/]([^\0000])*(((((((<2><4>)<8>)<16>)<32768>)<65536>)<131072>)<262144>))/home/likewise/[^\0000/]([^\0000/])*/[^\0000/]([^\0000/])*/[^\0000/]([^\0000])*\0000/[^/](.)*((<16><32>)<262144>))/(usr[])/bin/[^\0000/]([^\0000])*(<2305><37765120>))
understanding the tree output
The tree is output as as a regular expression where ( )  groups and expression precedence   alternation that separates two alternatives eg. AB *  means repeat the previous expression 0 or more times

 means repeat the previous expression 1 or more times .  match any character []  character class which can contain multiple alternative characters [^]  an inverted character class \xxxx  a nonprintable character
expression tree simplification (AppArmor 2.3 and later)
This stage does expression tree factoring which can remove common subexpressions that cancel out. To do this the tree is first normalized into a left or right normal form, and then a factoring pass is made finding common nodes. The normalization is then flipped and the process is repeated as necessary until the tree stabilizes and no more simplification is possible.
For example,
(abac)
would be factored intoa(bc)
This step was introduced because the dfa conversion step can cause exponential state explosion. The factoring is biased towards minimizing the cases that will cause state explosion (right normalization first). This stage is critical for large or complicated policies as it can make them possible to compute in reasonable amounts of time and memory.
Normalization
Normalization rewrites the expression tree so that alternations and concatenations always have the same consistent form, which simplifies the the Factoring stage.
left normalization rules (right normalization is opposite)
(\epsilon  a) \to a  \epsilon
(a  b)  c \to a  (b  c)
[b]  a \to a  [b]
\epsilon a \to a \epsilon
(ab)c \to a(bc)
TODO: add actual tree diagrams
Factoring Patterns
a\epsilon \to a
(a  a) \to a
a  (a  b) \to (a  b)
a  (ab) \to a (\epsilon  b) \to a (b  \epsilon)
(ab)  (ac) \to a(bc)
To see the tree after the apparmor_parser has done simplification do
./apparmor\_parser D exprsimple QT <profile_name>
DFA: Simplified Expression Tree
/((usr\[\])/bin/\[^\\0000/\](\[^\\0000\])\*(<2305><37765120>)(etc/passwd(<4><65536>)home(/\[^\\0000/\](\[^\\0000/\])\*/bin/(<513>(<8404992>(<1048576><64>)))(/likewise/\[^\\0000/\](\[^\\0000/\])\*\[\])/\[^\\0000/\](\[^\\0000/\])\*/\[^\\0000/\](\[^\\0000\])\*(<4>(<16>(<32768>(<65536>(<131072>(<262144>(<2>(<8>\\0000/\[^/\](.)\*(<16>(<32><262144>)))))))))))))
dfa creation
AppArmor does direct single regex to dfa creation as described in Compilers  Principles, Techniques, and Tools (aka the Dragon Book). This combines the traditional two step conversion of regex to NFA and then subset construction, and should general be faster and consume less memory.
At this stage of the dfa lots of information is available to do analysis on.
To see the dfa statistics and compare how previous optimization have do:
tree stats without simplification
apparmor_parser O nominimize O noremoveunreachable O noexprsimplify D dfastats QT example.txt
Created dfa: states 58 matching 69 nonmatching 56
tree with simplifications applied
apparmor_parser O nominimize O noremoveunreachable D dfastats QT example.txt
Created dfa: states 54 matching 67 nonmatching 52
The important output from dumping dfastats is the number of states created. The matching entry indicates the number of expression node comparisons that were folded into existing states while nomatching is the number of expression nodes that resulted in the creation of new states.
At this stage you can also obtain a visual graph of the generated dfa. This is done by dumping the dfa in graphviz format and using graphviz to generate a diagram. Warning it is easy to create graphs that are to big to be viewed.
To generate a graph do
apparmor_parser O nominimize O noremoveunreachable D dfagraph QT <profile_name> 2>profile.gv
dot Tpng o profile.png profile.gv
no tree simplification  tree simplification 

400px  400px 
dfa minimization (AppArmor 2.5 and later)
AppArmor uses Hopcroft's dfa minimization algorithm to reduce the number of states in the dfa to the smallest dfa possible. This step was added to help reduce kernel memory consumption by the dfa and compilation times.
TODO: insert dfa state example pic
hashing partition setup
Traditionally minimization starts with 2 partitions (1 accepting, 1 nonaccepting) and the partitions are iteratively split until all states in a partition are indistinguishable. A single representative state is then chosen from each partition, removing the other states and a minimized dfa is created. Instead of splitting the dfa states into 2 partitions, and repartitioning, a linear pass is made through the states setting up an initial set of partitions using hashing to separate states that can never be equivalent.
transition based hashing
By default transition based hashing is used.
The hashing algorithm takes advantage of the fact that states are stored in a compressed form with a default transition and then a set of N nondefault transitions.
The rules for the hash are
 states can not be used as part of the hash as different states may be reduced to the same Partition. Only transitions and known unique partitions can be considered
 states will only be equivalent if they have the same set of nondefault transitions.
 number of none default transitions are the same
 the set of characters used for nondefault transitions are the same
permission based hashing
The permissions of the accept state can be used as part of the initial partition setup (it is defaulted on in AppArmor 2.4  2.6). This can lead to even more initial partitions (speeding up minimization) but it can result in a nonminimum dfa as some states that would have been merged end up in separate partitions, which will increase the final size and can slow down total creation time as state minimization has more states to compress.
partition merging
The current algorithm for state comparison within a partition is not the most efficient form, but since the partition setup does a very good job, it is adequate for the number of states currently present in partitions.
dfa unreachable state removal (AppArmor 2.5 and later)
The creation and or minimization of a dfa can result in unreachable states. This phase walks the dfa from the start state finding all reachable states and deletes any that can not be reached.
TODO: insert dfa state example pic
Default state differential encoding (compression)
To reduce the number of transitions a state has to encode states can be made to encode their transitions as a differential to a “default” state. If a transition is not represented in the current state the default state is entered and the process is repeated until the transition is found or the state is not marked to be encoded relative to it default.
This type of state encoding can be quite effective in reducing the number of states but it can increase both encoding and matching time (as multiple states must be traversed for a single match). However if the states to encode against are chosen carefully, then both the encoding time and matching time can be bounded, and differential encoding can even result in a faster HFA as it can reduce the memory bandwidth required by the HFA.
The requirements AppArmor uses to choose the states to encode against are:
 The state must have been previously matched while walking the dfa (it will be hot in the cache then)
 or the state must be at the same level in a DAG decomposition of the dfa, sharing a common ancestor (more on this below)
The first requirement was primarily for performance concerns but in practice works out well for compression too, as states that are close to each other often have similar transitions. The second allows expanding the reach of the compression to a few more likely options while keeping a potentially common hot path, and without breaking other properties of only referencing previously matched states.
In practice requirement 1 can not be met as each match string takes a different path through the dfa. It can however be approximated by converting the dfa into a directed acyclic graph (DAG) with the start states as the root. The DAG provides a good approximation for requirement 1 and at the same time limits how many states have to be considered for compression (only backwards in the DAG). It also provides guarentees on how many states will be walked at run time (at most 2n).
Converting the HFA into DAG for compression does have a limitation in that it removes many of a states immediate neighbours from consideration. In a DAG a states neighbours can be broken into five classes, immediate predecessor, predecessor on another branch, sibling on another branch, immediate successor, successor on another branch. The immediate predecessor and immediate successor cases are covered by the predecessor differential compression scheme described above (successor as the current state is the successor state predecessor, and thus will be considered when the successor is differentially encoded). However the successor and predecessor on another branch and sibling cases are not covered, and they maybe the more optimal path for encoding, and may be the hot path the match came through.
To help account for this AppArmor also compares to the immediate successors of the state being consider if there are transitions between the states. Sibling states are also considered if there are transitions between the state and the sibling is differentially encoded against a predecessor (not another sibling), or not differentially encoded. This broadens the set of states considered but limits it to states that were potentially matched against and thus in the cache. It also has the property of looking backwards in the DAG thus keeping the maximum number of states that are required to be transitioned to in a match to a linear constant. If only branch predecessor where used then the limit could be kept at 2n but because immediate siblings can be used iff they transition to a predecessor the limit is bounded to a slightly higher value of 5/2n.
When considering which state to differentially encode against AppArmor computes a weighted value and chooses the best one. The value is computed as follows.

For each defined transition in the state

+0 to candidate state weight
if the transition is undefined in the candidate state (the transition must be represented in the current state)

+1 to candidate state weight
if the candidate state has the same transition to the same state (the transition can be eliminated from the current state)

1 to candidate state weight
if the transition is defined in the candidate state and it is not the same transition (current state must add a transition entry to override candidate transition)


For each undefined transition in the state

+0 to candidate state weight
if the transition is undefined in the candidate state

1 to candidate state weight
if the transition is defined in candidate state (current state must add an entry to override the candidate transition)

The current state will be differentially encoded against the candidate state with the largest weight > 0. If there is no weighting > 0 then no differential encoding for the state will be done as their is no benefit to doing so.
Note: differential encoding can in one special case reduce a state to 0 stored transitions. This can happen when two states have the exact same transitions but belong in different partitions when minimized. This would happen for example when one state was an accept state and the other a none accepting state. Other wise if states have the same transitions they are redundant and removed during state minimization.
creation of equivalence classes
AppArmor can use (they are optional) equivalence classes to further compress a states transitions. AppArmor uses a single HFA table to encode equivalence classes at the byte level (ie 256 entries). The equivalence class can be used in two ways to improve compression.
 map multiple characters to a single transition
 remap characters to reduce the distance between characters allowing the better comb compression
The mapping of multiple characters to a single character transition is the traditional use of equivalence classes. Each equivalence class in the HFA consists of the set of characters that have the exact same transitions for every state in the HFA. That is 'a' and 'A' are in the same equivalence class if and only if for every state that 'a' has a defined transition, 'A' also has a defined transition and 'a' and 'A' transition to the same state. This allows 'a' and 'A' to represented by a single transition, and for an equivalence class table to be used to remap 'a' and 'A' to the equivalence class.
Even if there are no equivalence classes to be found within the HFA the equivalence class table can be used to remap characters, so that they can be better compressed by the following comb compression step. In this case the distance between common transitions is reduced (ideally to zero). For example assume that 'a' and 'A' appear as transition with in states frequently but are not equivalence classes, and that no characters between 'a' and 'A' are used in transitions. When the transitions for 'a' and 'A' are encoded in the transition table there is a gap between the two transitions as the transition table is index by character. Comb compression is a technique to try and fill the gap between those characters so that the transition table is fully packed, but this is a hard problem (np complete) and so comb compression often does a less than optimal job. The equivalence class can be used to map these two characters to be adjacent so that there is no gap to fill in the transition table. This can result in higher and faster compression as the comb compressor then has less combinations to consider.
Implementation wise the equivalence class is a 256 byte table that is index by the input character to find the equivalence class that is then used as the index for the state transition. The equivalence class transformation is one way as multiple characters can map to a single equivalence class (which is just a character). While equivalence classes require and extra lookup they can actually increase performances as they reduce the memory bandwidth required to traverse the HFA and the equivalence class table is small and cache hot.
The equivalence class computation is performed after state differential transition encoding as that reduces the number of transitions that must be considered. This in turn can improve the chance of finding equivalence classes, and even if it doesn't can result character transitions being mapped tighter together for comb compression.
dfa table compression
Currently there are is a single compression scheme with some options that can affect compression times, match times and size. The compression format is based on the split tables and comb (interleaved column) compression from the Dragon book. The current compression schemes are designed to be used at dfa runtime such that the the data is kept in an immediately accessible form using naturally accessible word entries with no bit compression. A consequence is that the tables could be further compressed using a gzip style compression for on disk storage.
The simplest form of the compression is straight comb compression with nondifferential entries. Other compression variants modify the base compression by making entries differential to other states as discussed above.
Compression Efficiency
The base compression format of split tables + comb compression works well as long as the average number of transitions per state is low. The use of differential encoding does not change the base compression format but serves to further reduce the average number of transitions that need to represented.
The effectiveness is determined by comparing the compressed format to an uncompressed dfa representation which is a table indexing the rows by state, the columns by input character, and the cells containing the next state and accept information. Ignoring any extended HFA information, this would be
size = (sizeof(next state)) * (# of input characters) * (# of states) + sizeof(accept index) * (# of states)
= (2) * 256 * s + 2 * x
= 512s + 2s
= 514s
Therefore AppArmor's uncompressed encoding (2 bytes for next state, 2 bytes for accept index) takes up 514bytes of memory for each state. The compressed table size is represented by:
size = ((sizeof(base index) + sizeof(default) + sizeof(accept index)) + ((sizeof(next) + sizeof(check)) * (Average transitions per state) * (Packing factor))) * (# of states)
= ((4 + 2 + 2) + ((2 + 2) * Ave * Pf)) * s
= 8s + 4(Ave * Pf)s
The packing factor (Pf) expresses the efficiency of the packing done by comb compression. A perfect packing of 1000 transitions would be a table with 1000 entries. When the packing isn't perfect the comb compression leaves unused entries (gaps) in the table so a Pf of 1.1x for a 1000 transitions would result in a table with a size of 1100 entries.
The theoretical maximum compression of the split table + comb compression scheme is 64.25x (514s/8s) and only occurs when there is only a default transition for every state in the DFA. Since this makes the average number of transition per state 0, the second term of the table size equation can be eliminated.
In practice, 64.25x compression is not achievable but good compression is still achieved. For a given profile, the average number of transitions per state ranges from 3 up to approximately 16 (see apparmor_parser D stats) and the packing factor ranging from 1.1 to 1.4.
Compression efficiency for different transition averages and packing factors:
Ave  1.0x  1.1x  1.2x  1.3x  1.4x  1.5x 

1  42.83  41.45  40.16  38.94  37.79  36.71 
2  32.13  30.60  29.20  27.93  26.77  25.70 
3  25.70  24.25  22.95  21.78  20.73  19.77 
4  21.42  20.08  18.90  17.85  16.91  16.06 
6  16.06  14.94  13.97  13.11  12.36  11.68 
8  12.85  11.90  11.08  10.36  9.73  9.18 
12  9.18  8.45  7.84  7.30  6.84  6.43 
16  7.14  6.56  6.06  5.64  5.27  4.94 
32  3.78  3.45  3.18  2.95  2.75  2.57 
64  1.95  1.77  1.63  1.51  1.40  1.31 
128  0.99  0.90  0.83  0.76  0.71  0.66 
240  0.53  0.48  0.44  0.41  0.38  0.35 
250  0.51  0.46  0.43  0.39  0.37  0.34 
256  0.50  0.45  0.42  0.38  0.36  0.33 
When differential state compression is used, the average number of transitions stays closer to the low end (2 and 3) and the packing factor approaches 1. This is because as the number of transitions approach the upper and lower bounds (1 and 256), the vector are denser and tend to pack better.
Effect various options have on the size of the dfa
add no treeexpr, no min
old style accept
An example of the effectiveness of the compression scheme is the evince profile, which compiles into 3 profiles with DFAs.
Profile 1  options  DFA # of states  no compression  2 table (no comb)  2 table + comb  2 table + comb + differential  

1  no minimize, no sharing  11977  5.87 MB  563KB ave=7.54, Pf=1.33  
2  no minimize, no sharing  8172  4.01 MB  306KB ave=6.22, Pf=1.22  
3  no minimize, no sharing  5954  2.92 MB  230KB ave=6.40, Pf=1.23  
1  minimize, no sharing  8852  4.34 MB  425KB ave=7.74, Pf=1.33  
2  minimize, no sharing  6016  2.99 MB  237KB ave=6.61, Pf=1.22  
3  minimize, no sharing  4086  2.95 MB  168KB ave=6.93, Pf=1.23  
1,2,3  minimize + sharing  ???  Ave=, Pf= 
The Basic format
The HFA is split into multiple tables, for a basic DFA (subset of the EHFA) there are 5 table types used, default, base, next, check, and accept. The default, base, and accept tables store the basic information about a state and are indexed by the state number. While the next and check tables store state transitions are index by the base value (obtained from base table) + input character.
The accept table provides the information to determine the permission associated with that state, whether directly stored in the table or an index into another structure.
The default table stores the default transition that should be taken for the state, if the check entry does not match.
The base table stores the start position value into the next check tables for the state.
The next check, tables store the nondefault transitions for the dfa. Each state begins in the table with a location determined by its base value and then it is index from that position by the character (byte) being checked. The next table stores the state to transition to for the character being checked, and the check table stores whether the entry being checked belongs to the current state.
Each state with in the table is treated as a sparse matrix column only storing non default entries. The states are then interleaved so that one states entries are mapped into the holes of another state.
Comb compression affects the base value which determines the position of the states transitions in the next, check tables. State differential encoding affects the default state and number of transition that need to be stored for the state.
performance optimization, grouping tables together so they are the same cache line.
Basic default entry compression
State  \000  \001  \002  ..  \254  \255 

0 (dead)  0  0  0  ..  0  0 
1 (start)  ..  
2  ..  
. . .  
53  ..  
54  .. 
State  Accept  Default  Base 

0 (dead)  0  0  0 
1 (start)  
2  
3  
4 
Base  Next  Check 

0  
1  
2  
3  
4 
Recursive default entry compression
Computing State Permissions from Rules
accept nodes
 unique per permission
deny
x permission
 intersection test  dominance