Skip to content

The DejaVu runtime verification system, using BDDs for storing data occurring in observed events.

License

Notifications You must be signed in to change notification settings

havelund/dejavu

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

28 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

    _____       _    __      __    
   |  __ \     (_)   \ \    / /    
   | |  | | ___ _  __ \ \  / /   _ 
   | |  | |/ _ \ |/ _` \ \/ / | | |
   | |__| |  __/ | (_| |\  /| |_| |
   |_____/ \___| |\__,_| \/  \__,_|
              _/ |                 
             |__/                  
                        First-order past time LTL with recursive rules and time!

    Version 2.1, March 4 - 2020

Overview

DejaVu is a program written in Scala for monitoring event streams (traces) against temporal logic formulas. The formulas are written in a first-order past time linear temporal logic, with the addition of macros and recursive rules. The logic also supports reasoning about time.

An example of a property in its most basic form is the following:

prop closeOnlyOpenFiles : forall f . close(f) -> exists m . @ [open(f,m),close(f))

The property has the name closeOnlyOpenFiles and states that for any file f, if a close(f) event is observed, then there exists a mode m (e.g 'read' or 'write') such that in the previous step (@), some time in the past was observed an open(f,m) event, and since then no close(f) event has been observed.

The implementation uses BDDs (Binary Decision Diagrams) for representing assinments to quantified variables (such as f and m above).

Installing DejaVu:

The directly out contains files and directories useful for installing and running DejaVu:

  • README.pdf : this document in pdf format
  • dejavu : script to run the system
  • artifacts/dejavu_jar/dejavu.jar : the dejavu jar file
  • papers : a directory containing papers published about DejaVu
  • examples : an example directory containing properties and logs

DejavU is implemented in Scala.

  1. Install the Scala programming language if not already installed (https://www.scala-lang.org/download)

  2. Place the files dejavu and dejavu.jar mentioned above in some directory DIR (standing for the total path to this directory).

  3. cd to DIR and make the script executable:

     chmod +x dejavu
    
  4. Preferably define an alias in your shell profile to the dejavu script so it can be called from anywhere:

     alias dejavu=DIR/dejavu
    

Running DejaVu

The script is applied as follows:

dejavu <docFile> <logFile> [<bitsPerVariable> [debug]]
  • The <docFile> is the path to a file containing the specification document.
  • The <logFile> is the path to a file containing the log in CSV format to be analyzed.
  • The <bitsPerVariable> is a number indicating how many bits should be assigned to each variable in the BDD representation.
  • The debug flag will cause debugging output to be generated on standard out.

These fields will be explained in more detail in the following.

The specification document file (<docFile>) This is the temporal specification that the trace must satisfy. See explanation of the specification language below.

The log file (<logFile>) should be in comma separated value format (CSV): http://edoceo.com/utilitas/csv-file-format. For example, a file of the form:

list,chair,500
bid,chair,700
bid,chair,650
sell,chair

with no leading spaces would mean the four events:

list(chair,500)
bid(chair,700)
bid(chair,650)
sell(chair)

The log file and time

In case the log file contains time stamps, the log file name! must contain the text .timed.. E.g: log42.timed.csv. Time stamps (natural numbers) must appear as the last argument to all events. E.g. a timed version of the above trace (with time values in the range 1000 ... 1099) is:

list,chair,500,1010
bid,chair,700,1025
bid,chair,650,1067
sell,chair,1099 

Note that this last time value is not refered to explicitly in events in specifications. That is, the events with time in the above CSV format still corresponds to the following events in specification format:

list(chair,500)
bid(chair,700)
bid(chair,650)
sell(chair)

In case the log file is not timed (as described just above), time is considered as always being 0. One can still check timed properties against such a log, but of course it makes little sense.

The bits per variable (<bitsPerVariable>) indicates how many bits are assigned to each variable in the BDDs. This parameter is optional with the default value being 20. If the number is too low an error message will be issued during analysis as explained below. A too high number can have impact on the efficiency of the algorithm. Note that the number of values representable by N bits is 2^N, so one in general does not need very large numbers.

The algorithm/implementation will perform garbage collection on allocated BDDs, re-using BDDs that are no longer needed for checking the property, depending on the form of the formula.

Debugging (debug) The debugging flag can only be provided if also < bitsPerVariable > is provided. Usually one picks a low number of bits for debugging purposes (e.g. 3). The result is debugging output showing the progress of formula evaluation for each event. Amongst the output is BDD graphs visualizable with GraphViz (http://www.graphviz.org).

Results from DejaVu

Wellformedness errors

Error messages will be emitted if the specification document is not wellformed. A wellformedness violation terminates the program, and can be one of the following:

  • Syntax error: the document does not follow the grammar.
  • Free variable: a used variable has not been introduced as a predicate parameter or a quantified variable.
  • Hiding: a quantified expression hides a name introduced at an outer level.
  • Unused variable: a name introduced as a predicate parameter or quantifier is not used.
  • Inconsistent: a predicate is called with inconsistent number of arguments compared to definition or other calls.
  • Duplicates: duplicate definition of a predicate or property name.
  • Undefined event: events have been defined but an event is used and not amongst the defined ones.
  • Variable duplication: a variable is introduced more than once in an event, macro, or rule parameter list.
  • Unprotected recursive rule definition: A rule is called in the body of a rule definition wthout being underneath a previous-time operator: @.

A warning does not terminate the program, and can be one of:

  • Unused macro: a predicate macro is defined but not used.
  • Unused event: an event is defined but not used.

Property violations The tool will indicate a violation of a property by printing what event number it concerns and what event. For example:

*** Property incr violated on event number 3:

#########################################################
#### bid(chair,650)
#########################################################  

indicates that event number 3 violates the property incr, and that event is a line in the CSV file having the format:

bid,chair,650   

Trace statistics A trace statistics is printed, which indicates how many events were processed in total and how they were distributed over the different event types:

Processed 1100006 events

==================
Event Counts:
------------------
logout : 20001
open   : 520001
login  : 500000
close  : 40002
access : 20002
==================

Warnings will also be here be issued if:

  • if there are events in the specification that do not occur in the trace, or dually
  • if there are events in the trace that do not occur in the specification

Both can potentially be signs of a flawed specification, but might not need to be.

Not enough bits per variable If not enough bits have been allocated for a variable to hold the number of values generated for that variable, an assertion violation like the following is printed:

*** java.lang.AssertionError: assertion failed: 
    10 bits is not enough to represent variable i.

One can/should experiment with BDD sizes.

Timing results The system will print the following timings:

  • the time spent on parsing spec and synthesizing the monitor (a Scala program)
  • the time spent on compiling the synthesized monitor
  • the time spent on verifying trace with compiled monitor

Generated files The system will generate the following files (none of which need any attention from the user, but may be informative):

  • TraceMonitor.scala : containing the synthesized monitor (a self-sufficient Scala program).
  • ast.dot : file showing the structure of the formula (used for generating BDD updating code). This can be viewed with GraphViz (http://www.graphviz.org). These two files help illustrate how the algorithm works.
  • dejavu-results : contains numbers of events that violated property if any violations occurred. Mostly used for unit testing purposes.

The DejaVu Specification Logic

Grammar

The grammar for the DejaVu temporal logic is as follows.

<doc> ::= <def> ... <def>
<def> ::= <eventdef> | <macrodef> | <propertydef>

<eventdef> ::= 'pred' <event>,...,<event>
<event>    ::= <id> [ '(' <id> ',' ... ',' <id> ')' ]

<macrodef> ::= 'pred' <id> [ '(' <id> ',' ... ',' <id> ')' ] '=' <form>

<propertydef> ::= 'prop' <id> ':' <form> ['where' <ruledef> ',' ... ',' <ruledef>]

<ruledef> ::= <id> ['(' <id> ',' ... ',' <id> ')'] ':=' <form>

<form> ::= 
     'true' 
   | 'false' 
   | <id> [ '(' <param> ',' ... ',' <param> `)' ]
   | <form> <binop> <form> 
   | '[' <form> ',' <form> ')'
   | <unop> <form>
   | <id> <oper> (<id> | <const>)
   | '(' <form> ')'
   | <quantifier> <id> '.' <form>

<param>  ::= <id> | <const>
<const>  ::= <string> | <integer>
<binop>  ::= '->' | '|' | '&' | 'S' [<time>] | Z <timeLE>
<unop>   ::= '!' |  '@' | 'P' [<time>] | 'H' [<time>] 
<oper>   ::= '<' | '<=' | '=' | '>' | '>='    
<quantifier> ::= 'exists' | 'forall' | 'Exists' | 'Forall'
<time>   ::= <timeLE> | <timeGT>
<timeLE> ::= '[<=' <number> ']'
<timeGT> ::= '[>' <number> ']'    

Event, Macro, and Property Definitions

A specification document <doc>consists of a sequence of definitions. A definition <def> can either be an event definition, a macro definition, or a property to be checked.

An event definition introduces those events that the property will refer to. If no events are defined then the events are inferred to be those referred to in the property. It may help to reduce errors in properties to declare the events up front.

A predicate macro definition <macrodef> introduces a named shorthand for a formula, possibly parameterized with variable names (those introduced with quantifiers). For example the following macro definition defines a shorthand representing that a file is open:

pred isOpen(f) = !close(f) S open(f)

Macros can be called in properties. Macros can call other macros, but cannot be recursive. The order of the declaration is of no importance. They can furthermore be introduced after as well as before the properties referring to them.

A property definition <propdef> introduces a named property, which is a first-order past time temporal formula.

Formulas

The different formulas <form> have the following intuitive meaning:

true, false     : Boolean truth and falsehood 
id(v1,...,vn)   : event or call of predicate macro, where vi can be a constant or variable
p -> q          : p implies q
p | q           : p or q
p & q           : p and q
p S q           : p since q (q was true in the past, and since then, including that point in time, p has been true) 
p S[<=d] q      : p since q but where q occurred within d time units
p S[>d] q       : p since q but where q occurred earlier than d time units
p Z[<=d] q      : p since q but where q did not occur at the current time
[p,q)           : interval notation equivalent to: !q S p. This form may be easier to read.
! p             : not p
@ p             : in previous state p is true
P p             : in some previous state p is true
P[<=d] p        : in some previous state within d time units p is true
P[>d] p         : in some previous state earlier than d time units p is true
H p             : in all previous states p is true
H[<=d] p        : in all previous states within d time units p is true
H[>d] p         : in all previous states earlier than d time units p is true
x op k          : x is related to variable or constant k via op. E.g.: x < 10, x <= y, x = y, x >= 10, x > z   
// -- quantification over seen values in the past, see (*) below:
exists x . p(x) : there exists an x such that seen(x) and p(x) 
forall x . p(x) : for all x, if seen(x) then p(x)
// -- quantification over the infinite domain of all values:
Exists x . p(x) : there exists an x such that p(x) 
Forall x . p(x) : for all x p(x)    

(*) seen(x) holds if x has been observed in the past

Rules

A new extension of DejaVu is the notion of rules, which are part of a property definition. As an example, consider the following property about threads being spawned in an operating system. We want to ensure that when a thread y reports some data d back to another thread x, then thread y has been spawned by thread x either directly, or transitively via a sequence of spawn events. The events are spawn(x,y) (thread x spawns thread y) and report(y,x,d) (thread y reports data d back to thread x). For this we need to compute a transitive closure of spawning relationships, here expressed with the rule spawning(x,y). This property can be stated as follows using the rule spawned(x,y) (thread x spawned y, either directly or indirectly through other spawns):

prop spawning :
  Forall x . Forall y . Forall d . report(y,x,d) -> spawned(x,y) 
  where 
    spawned(x,y) := 
        @ spawned(x,y) 
      | spawn(x,y) 
      | Exists z . (@spawned(x,z) & spawn(z,y))

The property states that if there is a report(y,x,d) event (thread y reporting data d back to thread x), then spawned(x,y) must hold, defined as follows: either spawned(x,y) held in the previous state, or there is a spawn(x,y) in the current state, or, the interesting case: spawned(x,z) held in the previous state for some z, and spawn(z,y) holds in the current state. This last disjunct forms the transitive closure.

Time

Timing properties can be expressed using natural numbers as constraints. Examples formulas concerning commands being dispatched dis(m) and succeeding suc(m) are:

Forall m . suc(m) -> true S[<=3] dis(m) // succeeding command must have been dispatched within 3 time units 
Forall m . suc(m) -> true S[>3] dis(m) // dispatch earlier than 3 time units
Forall m . dis(m) -> ! (true Z[<=3] dis(m)) // if command dispatched, not dispatched before! within 3
Forall m . suc(m) -> P[<=3] dis(m) // succeeding command must have been dispatched within 3 time units 
Forall m . suc(m) -> P[>3] dis(m) // dispatch earlier than 3 time units
Forall m . suc(m) -> H[<=3] !dis(m) // for a succeeding command no dispatch within 3 time units before
Forall m . suc(m) -> H[>3] !dis(m) // no dispatch earlier than 3 time units

Further Examples Of DejaVu Properties

Auctions

We illustrate the logic with properties that an auction has to satisfy. The following observable events occur during an auction:

  • list(i,r) : item i is listed for auction with the minimal reserve sales price r.
  • bid(i,a) : the bidding of a dollars on item i.
  • sell(i) : the selling of item i to highest bidder.

An auction system has to satisfy the four properties shown in below expressed over these three kinds of events, using a predicate macro inAuction(x) to express when an item x is in active auction (when it has been listed and not yet sold):

pred inAuction(x) = exists r . @ [list(x,r),sell(x))

prop incr : 
  Forall i . Forall a1 . Forall a2 . @ P bid(i,a1) & bid(i,a2) -> a1 < a2

prop sell : 
  Forall i . Forall r . P list(i,r) & sell(i) ->  exists a . P bid(i,a) & a >= r

prop open : 
  Forall i . Forall a . (bid(i,a) | sell(i)) -> inAuction(i)

prop once : 
  Forall i . Forall r . list(i,r) -> ! exists s . @ P list(i,s)

Property incr states that bidding must be increasing.

Property sell states that when an item is sold, there must exist a bidding on that item which is bigger than or equal to the reserve price.

Property open states that bidding on and selling of an item are only allowed if the item has been listed and not yet sold.

Finally, Property once states that an item can only be listed once.

Locks in a multithreaded system

We observe the following events:

  • acq(t,l) : thread t acquires lock l.
  • rel(t,l) : thread t releases lock l.
  • read(t,x) : thread t reads variable x.
  • write(t,x) : thread t writes variable x.
  • sleep(t) : thread t goes to sleep.

Basic properties

The first set of properties are stated as a conjucntion of three sub-properties:

  • A thread going to sleep should not hold any locks.
  • At most one thread can acquire a lock at a time.
  • A thread can only release a lock it has acquired.

This is formalized as follows:

prop locksBasic :
  Forall t . Forall l .
    (
      (sleep(t) -> ![acq(t,l),rel(t,l))) &
      (acq(t,l) -> ! exists s . @ [acq(s,l),rel(s,l))) &
      (rel(t,l) -> @ [acq(t,l),rel(t,l)))
    )

No deadlocks

Locks should not be acquired in a cyclic manner amongst threads (dining philosopher problem). That is, if a thread t1 takes a lock l1 and then a lock l2 (without having released l1), then at no time should another thread t2 take the locks in reverse order. Obeying this principle will prevent cyclic deadlocks.

This is formalized as follows:

prop locksDeadlocks :
  Forall t1 . Forall t2 . Forall l1 . Forall l2 .
    (@ [acq(t1,l1),rel(t1,l1)) & acq(t1,l2))
    ->
    (! @ P (@ [acq(t2,l2),rel(t2,l2)) & acq(t2,l1)))

No dataraces

If two threads access (read or write) the same shared variable, and one of the threads write to the variable, there must exist a lock, which both threads hold whenever they access the variable.

This is formalized as follows:

prop locksDataraces :
  Forall t1 . Forall t2 . Forall x .
    (
      (P (read(t1,x) | write(t1,x)))
      &
      (P write(t2,x))
    )
    ->
    Exists l .
      (
        H ((read(t1,x) | write(t1,x)) -> [acq(t1,l),rel(t1,l)))
        &
        H ((read(t2,x) | write(t2,x)) -> [acq(t2,l),rel(t2,l)))
      )

Calling Earth

This property concerns a radio on board a spacecraft, which communicates over different channels (quantified over in the formula), which each can be turned on and off with a toggle(x) - they are all initially off. Telemetry can only be sent to ground over a channel x, with the telem(x) event, when radio channel x is toggled on. The property cannot be expressed in pure past time LTL, that is, without the use of rules.

The property is formalized as follows:

prop telemetry1: 
  Forall x . closed(x) -> !telem(x) 
  where closed(x) := toggle(x) <-> @!closed(x) 

The same property can altermatively be expressed using two rules, more closely reflecting how we would model this using a state machine with two states for each channel x: closed(x) and open(x):

prop telemetry2: 
  Forall x . closed(x) -> !telem(x) 
    where
    closed(x) := 
        (!@true & !toggle(x)) 
      | (@closed(x) & !toggle(x)) 
      | (@open(x) & toggle(x)),
    open(x) := 
        (@open(x) & !toggle(x)) 
      | (@closed(x) & toggle(x))

The rule closed(x) is defined as a disjunction between three alternatives. The first alternative states that this predicate is true if we are in the initial state (the only state where @true is false), and there is no toggle(x) event. The next alternative states that closed(x) was true in the previous state and there is no toggle(x) event now. The third alternative states that we in the previous state were in the open(x) state and we observe a toggle(x) event. Similarly for the open(x) rule.

Experiments for publications

Contributors

About

The DejaVu runtime verification system, using BDDs for storing data occurring in observed events.

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages