Provided by: erlang-manpages_25.3.2.8+dfsg-1ubuntu4.4_all bug

NAME

       dbg - The Text Based Trace Facility

DESCRIPTION

       This  module  implements  a text based interface to the trace/3 and the trace_pattern/2 BIFs. It makes it
       possible to trace functions, processes, ports and messages.

       To quickly get started on tracing function calls you can use the following code in the Erlang shell:

       1> dbg:tracer(). %% Start the default trace message receiver
       {ok,<0.36.0>}
       2> dbg:p(all, c). %% Setup call (c) tracing on all processes
       {ok,[{matched,nonode@nohost,26}]}
       3> dbg:tp(lists, seq, x). %% Setup an exception return trace (x) on lists:seq
       {ok,[{matched,nonode@nohost,2},{saved,x}]}
       4> lists:seq(1,10).
       (<0.34.0>) call lists:seq(1,10)
       (<0.34.0>) returned from lists:seq/2 -> [1,2,3,4,5,6,7,8,9,10]
       [1,2,3,4,5,6,7,8,9,10]

       For more examples of how to use dbg from the Erlang shell, see the simple example section.

       The utilities are also suitable to use in system testing on large systems, where  other  tools  have  too
       much  impact  on  the system performance. Some primitive support for sequential tracing is also included,
       see the advanced topics section.

EXPORTS

       fun2ms(LiteralFun) -> MatchSpec

              Types:

                 LiteralFun = fun() literal
                 MatchSpec = term()

              Pseudo function that by means of a parse_transform translates the literal fun() typed as parameter
              in the function call to a match specification as described in the match_spec manual of ERTS  users
              guide.  (With  literal I mean that the fun() needs to textually be written as the parameter of the
              function, it cannot be held in a variable which in turn is passed to the function).

              The parse transform is implemented in the module ms_transform and the source must include the file
              ms_transform.hrl in STDLIB for this pseudo function to work. Failing to include the  hrl  file  in
              the  source  will result in a runtime error, not a compile time ditto. The include file is easiest
              included by adding the line -include_lib("stdlib/include/ms_transform.hrl"). to the source file.

              The fun() is very restricted, it can take only a single parameter (the parameter list to match), a
              sole variable or a list. It needs to use the is_XXX  guard  tests  and  one  cannot  use  language
              constructs  that  have  no representation in a match_spec (like if, case, receive etc). The return
              value from the fun will be the return value of the resulting match_spec.

              Example:

              1> dbg:fun2ms(fun([M,N]) when N > 3 -> return_trace() end).
              [{['$1','$2'],[{'>','$2',3}],[{return_trace}]}]

              Variables from the environment can be imported, so that this works:

              2> X=3.
              3
              3> dbg:fun2ms(fun([M,N]) when N > X -> return_trace() end).
              [{['$1','$2'],[{'>','$2',{const,3}}],[{return_trace}]}]

              The imported variables will be replaced by match_spec const expressions, which is consistent  with
              the  static  scoping  for  Erlang fun()s. Local or global function calls cannot be in the guard or
              body of the fun however. Calls to builtin match_spec functions of course is allowed:

              4> dbg:fun2ms(fun([M,N]) when N > X, is_atomm(M) -> return_trace() end).
              Error: fun containing local erlang function calls ('is_atomm' called in guard)\
               cannot be translated into match_spec
              {error,transform_error}
              5> dbg:fun2ms(fun([M,N]) when N > X, is_atom(M) -> return_trace() end).
              [{['$1','$2'],[{'>','$2',{const,3}},{is_atom,'$1'}],[{return_trace}]}]

              As you can see by the example, the function can be called from the shell too. The fun()  needs  to
              be  literally  in  the call when used from the shell as well. Other means than the parse_transform
              are used in the shell case, but more or less the same  restrictions  apply  (the  exception  being
              records, as they are not handled by the shell).

          Warning:
              If  the parse_transform is not applied to a module which calls this pseudo function, the call will
              fail in runtime (with a badarg). The module dbg actually exports a function with this name, but it
              should never really  be  called  except  for  when  using  the  function  in  the  shell.  If  the
              parse_transform  is  properly applied by including the ms_transform.hrl header file, compiled code
              will never call the function, but the function call is replaced by a literal match_spec.

              More information is provided by the ms_transform manual page in STDLIB.

       h() -> ok

              h stands for help. Gives a list of items for brief online help.

       h(Item) -> ok

              Types:

                 Item = atom()

              h stands for help. Gives a brief help text for functions in the dbg module.  The  available  items
              can be listed with dbg:h/0.

       p(Item) -> {ok, MatchDesc} | {error, term()}

              Equivalent to p(Item, [m]).

       p(Item, Flags) -> {ok, MatchDesc} | {error, term()}

              Types:

                 MatchDesc = [MatchNum]
                 MatchNum = {matched, node(), integer()} | {matched, node(), 0, RPCError}
                 RPCError = term()

              p  stands for process. Traces Item in accordance to the value specified by Flags. The variation of
              Item is listed below:

                pid() or port():
                  The corresponding process or port is traced. The process or port may be a  remote  process  or
                  port  (on  another  Erlang  node).  The  node must be in the list of traced nodes (see n/1 and
                  tracer/3).

                all:
                  All processes and ports in the system as well as all processes and ports created hereafter are
                  to be traced.

                processes:
                  All processes in the system as well as all processes created hereafter are to be traced.

                ports:
                  All ports in the system as well as all ports created hereafter are to be traced.

                new:
                  All processes and ports created after the call is are to be traced.

                new_processes:
                  All processes created after the call is are to be traced.

                new_ports:
                  All ports created after the call is are to be traced.

                existing:
                  All existing processes and ports are traced.

                existing_processes:
                  All existing processes are traced.

                existing_ports:
                  All existing ports are traced.

                atom():
                  The process or port with the corresponding registered name is traced. The process or port  may
                  be  a remote process (on another Erlang node). The node must be added with the n/1 or tracer/3
                  function.

                integer():
                  The process <0.Item.0> is traced.

                {X, Y, Z}:
                  The process <X.Y.Z> is traced.

                string():
                  If the Item is a string "<X.Y.Z>" as returned  from  pid_to_list/1,  the  process  <X.Y.Z>  is
                  traced.

              When enabling an Item that represents a group of processes, the Item is enabled on all nodes added
              with the n/1 or tracer/3 function.

              Flags can be a single atom, or a list of flags. The available flags are:

                s (send):
                  Traces the messages the process or port sends.

                r (receive):
                  Traces the messages the process or port receives.

                m (messages):
                  Traces the messages the process or port receives and sends.

                c (call):
                  Traces global function calls for the process according to the trace patterns set in the system
                  (see tp/2).

                p (procs):
                  Traces process related events to the process.

                ports:
                  Traces port related events to the port.

                sos (set on spawn):
                  Lets  all  processes  created  by  the  traced  process  inherit the trace flags of the traced
                  process.

                sol (set on link):
                  Lets another process, P2, inherit the trace flags of the traced process  whenever  the  traced
                  process links to P2.

                sofs (set on first spawn):
                  This is the same as sos, but only for the first process spawned by the traced process.

                sofl (set on first link):
                  This is the same as sol, but only for the first call to link/1 by the traced process.

                all:
                  Sets all flags except silent.

                clear:
                  Clears all flags.

              The list can also include any of the flags allowed in erlang:trace/3

              The  function  returns  either  an  error  tuple  or  a  tuple  {ok,  List}.  The List consists of
              specifications of how many processes and ports that matched (in the case of a pure  pid()  exactly
              1).  The  specification  of matched processes is {matched, Node, N}. If the remote processor call,
              rpc, to a remote node fails, the rpc error message is delivered  as  a  fourth  argument  and  the
              number  of  matched  processes are 0. Note that the result {ok, List} may contain a list where rpc
              calls to one, several or even all nodes failed.

       c(Mod, Fun, Args)

              Equivalent to c(Mod, Fun, Args, all).

       c(Mod, Fun, Args, Flags)

              c stands for call. Evaluates the expression apply(Mod, Fun, Args) with the trace  flags  in  Flags
              set. This is a convenient way to trace processes from the Erlang shell.

       i() -> ok

              i stands for information. Displays information about all traced processes and ports.

       tp(Module,MatchSpec)

              Same as tp({Module, '_', '_'}, MatchSpec)

       tp(Module,Function,MatchSpec)

              Same as tp({Module, Function, '_'}, MatchSpec)

       tp(Module, Function, Arity, MatchSpec)

              Same as tp({Module, Function, Arity}, MatchSpec)

       tp({Module, Function, Arity}, MatchSpec) -> {ok, MatchDesc} | {error, term()}

              Types:

                 Module = atom() | '_'
                 Function = atom() | '_'
                 Arity = integer() |'_'
                 MatchSpec = integer() | Built-inAlias | [] | match_spec()
                 Built-inAlias = x | c | cx
                 MatchDesc = [MatchInfo]
                 MatchInfo = {saved, integer()} | MatchNum
                 MatchNum = {matched, node(), integer()} | {matched, node(), 0, RPCError}

              tp  stands  for  trace  pattern.  This  function enables call trace for one or more functions. All
              exported functions matching the {Module, Function, Arity} argument  will  be  concerned,  but  the
              match_spec() may further narrow down the set of function calls generating trace messages.

              For  a  description of the match_spec() syntax, please turn to the User's guide part of the online
              documentation for the runtime system (erts). The chapter Match Specifications in  Erlang  explains
              the  general match specification "language". The most common generic match specifications used can
              be found as Built-inAlias', see ltp/0 below for details.

              The Module, Function and/or Arity parts of the tuple may be specified as the atom '_' which  is  a
              "wild-card"  matching  all modules/functions/arities. Note, if the Module is specified as '_', the
              Function and Arity parts have to be specified as  '_'  too.  The  same  holds  for  the  Functions
              relation to the Arity.

              All  nodes  added with n/1 or tracer/3 will be affected by this call, and if Module is not '_' the
              module will be loaded on all nodes.

              The function returns either  an  error  tuple  or  a  tuple  {ok,  List}.  The  List  consists  of
              specifications  of how many functions that matched, in the same way as the processes and ports are
              presented in the return value of p/2.

              There may be a tuple {saved, N} in the return value, if  the  MatchSpec  is  other  than  [].  The
              integer  N  may then be used in subsequent calls to this function and will stand as an "alias" for
              the given expression. There are also a couple of built-in  aliases  for  common  expressions,  see
              ltp/0 below for details.

              If  an  error is returned, it can be due to errors in compilation of the match specification. Such
              errors are presented as a list  of  tuples  {error,  string()}  where  the  string  is  a  textual
              explanation of the compilation error. An example:

              (x@y)4> dbg:tp({dbg,ltp,0},[{[],[],[{message, two, arguments}, {noexist}]}]).
              {error,
               [{error,"Special form 'message' called with wrong number of
                        arguments in {message,two,arguments}."},
                {error,"Function noexist/1 does_not_exist."}]}

       tpl(Module,MatchSpec)

              Same as tpl({Module, '_', '_'}, MatchSpec)

       tpl(Module,Function,MatchSpec)

              Same as tpl({Module, Function, '_'}, MatchSpec)

       tpl(Module, Function, Arity, MatchSpec)

              Same as tpl({Module, Function, Arity}, MatchSpec)

       tpl({Module, Function, Arity}, MatchSpec) -> {ok, MatchDesc} | {error, term()}

              tpl  stands  for  trace  pattern local. This function works as tp/2, but enables tracing for local
              calls (and local functions) as well as for global calls (and functions).

       tpe(Event, MatchSpec) -> {ok, MatchDesc} | {error, term()}

              Types:

                 Event = send | 'receive'
                 MatchSpec = integer() | Built-inAlias | [] | match_spec()
                 Built-inAlias = x | c | cx
                 MatchDesc = [MatchInfo]
                 MatchInfo = {saved, integer()} | MatchNum
                 MatchNum = {matched, node(), 1} | {matched, node(), 0, RPCError}

              tpe stands for trace pattern event. This function associates  a  match  specification  with  trace
              event  send  or 'receive'. By default all executed send and 'receive' events are traced if enabled
              for a process. A match specification can be used to filter traced events based on sender, receiver
              and/or message content.

              For a description of the match_spec() syntax, please turn to the User's guide part of  the  online
              documentation  for  the runtime system (erts). The chapter Match Specifications in Erlang explains
              the general match specification "language".

              For send, the matching is done on the list [Receiver,  Msg].  Receiver  is  the  process  or  port
              identity  of  the  receiver  and  Msg  is  the message term. The pid of the sending process can be
              accessed with the guard function self/0.

              For 'receive', the matching is done on the list [Node, Sender, Msg]. Node is the node name of  the
              sender.  Sender is the process or port identity of the sender, or the atom undefined if the sender
              is not known (which may be the case for remote senders). Msg is the message term. The pid  of  the
              receiving process can be accessed with the guard function self/0.

              All nodes added with n/1 or tracer/3 will be affected by this call.

              The  return value is the same as for tp/2. The number of matched events are never larger than 1 as
              tpe/2 does not accept any form of wildcards for argument Event.

       ctp()

              Same as ctp({'_', '_', '_'})

       ctp(Module)

              Same as ctp({Module, '_', '_'})

       ctp(Module, Function)

              Same as ctp({Module, Function, '_'})

       ctp(Module, Function, Arity)

              Same as ctp({Module, Function, Arity})

       ctp({Module, Function, Arity}) -> {ok, MatchDesc} | {error, term()}

              Types:

                 Module = atom() | '_'
                 Function = atom() | '_'
                 Arity = integer() | '_'
                 MatchDesc = [MatchNum]
                 MatchNum = {matched, node(), integer()} | {matched, node(), 0, RPCError}

              ctp stands for clear  trace  pattern.  This  function  disables  call  tracing  on  the  specified
              functions.  The  semantics  of  the  parameter  is  the  same  as  for  the corresponding function
              specification in tp/2 or tpl/2. Both local and global call trace is disabled.

              The return value reflects how many functions that matched, and  is  constructed  as  described  in
              tp/2. No tuple {saved, N} is however ever returned (for obvious reasons).

       ctpl()

              Same as ctpl({'_', '_', '_'})

       ctpl(Module)

              Same as ctpl({Module, '_', '_'})

       ctpl(Module, Function)

              Same as ctpl({Module, Function, '_'})

       ctpl(Module, Function, Arity)

              Same as ctpl({Module, Function, Arity})

       ctpl({Module, Function, Arity}) -> {ok, MatchDesc} | {error, term()}

              ctpl stands for clear trace pattern local. This function works as ctp/1, but only disables tracing
              set up with tpl/2 (not with tp/2).

       ctpg()

              Same as ctpg({'_', '_', '_'})

       ctpg(Module)

              Same as ctpg({Module, '_', '_'})

       ctpg(Module, Function)

              Same as ctpg({Module, Function, '_'})

       ctpg(Module, Function, Arity)

              Same as ctpg({Module, Function, Arity})

       ctpg({Module, Function, Arity}) -> {ok, MatchDesc} | {error, term()}

              ctpg  stands  for  clear  trace  pattern  global.  This function works as ctp/1, but only disables
              tracing set up with tp/2 (not with tpl/2).

       ctpe(Event) -> {ok, MatchDesc} | {error, term()}

              Types:

                 Event = send | 'receive'
                 MatchDesc = [MatchNum]
                 MatchNum = {matched, node(), 1} | {matched, node(), 0, RPCError}

              ctpe stands for clear trace pattern event. This  function  clears  match  specifications  for  the
              specified  trace event (send or 'receive'). It will revert back to the default behavior of tracing
              all triggered events.

              The return value follow the same style as for ctp/1.

       ltp() -> ok

              ltp stands for list  trace  patterns.  Use  this  function  to  recall  all  match  specifications
              previously  used  in  the session (i. e. previously saved during calls to tp/2, and built-in match
              specifications. This is very useful, as a complicated match_spec can be quite  awkward  to  write.
              Note that the match specifications are lost if stop/0 is called.

              Match  specifications used can be saved in a file (if a read-write file system is present) for use
              in later debugging sessions, see wtp/1 and rtp/1

              There are three built-in trace patterns: exception_trace, caller_trace and  caller_exception_trace
              (or  x,  c  and  cx  respectively).  Exception  trace sets a trace which will show function names,
              parameters, return values and exceptions thrown from functions.  Caller  traces  display  function
              names,  parameters  and  information  about  which function called it. An example using a built-in
              alias:

              (x@y)4> dbg:tp(lists,sort,cx).
              {ok,[{matched,nonode@nohost,2},{saved,cx}]}
              (x@y)4> lists:sort([2,1]).
              (<0.32.0>) call lists:sort([2,1]) ({erl_eval,do_apply,5})
              (<0.32.0>) returned from lists:sort/1 -> [1,2]
              [1,2]

       dtp() -> ok

              dtp stands for delete trace patterns. Use this function to "forget" all match specifications saved
              during calls to tp/2. This is useful when one wants to restore other match specifications  from  a
              file with rtp/1. Use dtp/1 to delete specific saved match specifications.

       dtp(N) -> ok

              Types:

                 N = integer()

              dtp  stands for delete trace pattern. Use this function to "forget" a specific match specification
              saved during calls to tp/2.

       wtp(Name) -> ok | {error, IOError}

              Types:

                 Name = string()
                 IOError = term()

              wtp stands for write trace patterns. This function will save all match specifications saved during
              the session (during calls to tp/2) and built-in match specifications in a text file with the  name
              designated  by Name. The format of the file is textual, why it can be edited with an ordinary text
              editor, and then restored with rtp/1.

              Each match spec in the file ends with a full  stop  (.)  and  new  (syntactically  correct)  match
              specifications can be added to the file manually.

              The  function  returns  ok  or an error tuple where the second element contains the I/O error that
              made the writing impossible.

       rtp(Name) -> ok | {error, Error}

              Types:

                 Name = string()
                 Error = term()

              rtp stands for read  trace  patterns.  This  function  reads  match  specifications  from  a  file
              (possibly)  generated  by the wtp/1 function. It checks the syntax of all match specifications and
              verifies that they are correct. The error handling principle is "all or nothing", i. e. if some of
              the match specifications are wrong, none of the specifications are added  to  the  list  of  saved
              match specifications for the running system.

              The  match specifications in the file are merged with the current match specifications, so that no
              duplicates are generated. Use ltp/0 to see what numbers were assigned to the  specifications  from
              the file.

              The function will return an error, either due to I/O problems (like a non existing or non readable
              file)  or  due  to  file  format problems. The errors from a bad format file are in a more or less
              textual format, which will give a hint to what's causing the problem.

       n(Nodename) -> {ok, Nodename} | {error, Reason}

              Types:

                 Nodename = atom()
                 Reason = term()

              n stands for node. The dbg server keeps a  list  of  nodes  where  tracing  should  be  performed.
              Whenever  a  tp/2  call or a p/2 call is made, it is executed for all nodes in this list including
              the local node (except for p/2 with a specific pid() or port() as first argument,  in  which  case
              the command is executed only on the node where the designated process or port resides).

              This  function  adds  a remote node (Nodename) to the list of nodes where tracing is performed. It
              starts a tracer process on the remote node, which will send  all  trace  messages  to  the  tracer
              process  on  the  local node (via the Erlang distribution). If no tracer process is running on the
              local node, the error reason no_local_tracer is returned. The tracer process  on  the  local  node
              must be started with the tracer/0/2 function.

              If Nodename is the local node, the error reason cant_add_local_node is returned.

              If  a  trace  port  (see trace_port/2) is running on the local node, remote nodes cannot be traced
              with a tracer process. The error reason cant_trace_remote_pid_to_local_port is returned.  A  trace
              port can however be started on the remote node with the tracer/3 function.

              The function will also return an error if the node Nodename is not reachable.

       cn(Nodename) -> ok

              Types:

                 Nodename = atom()

              cn  stands  for  clear node. Clears a node from the list of traced nodes. Subsequent calls to tp/2
              and p/2 will not consider that node, but tracing already activated on the node will continue to be
              in effect.

              Returns ok, cannot fail.

       ln() -> ok

              ln stands for list nodes. Shows the list of traced nodes on the console.

       tracer() -> {ok, pid()} | {error, already_started}

              This function starts a server on the local node that will be the recipient of all trace  messages.
              All subsequent calls to p/2 will result in messages sent to the newly started trace server.

              A  trace  server  started in this way will simply display the trace messages in a formatted way in
              the Erlang shell (i. e. use io:format). See tracer/2 for a description of how  the  trace  message
              handler can be customized.

              To start a similar tracer on a remote node, use n/1.

       tracer(Type, Data) -> {ok, pid()} | {error, Error}

              Types:

                 Type = port | process | module
                 Data = PortGenerator | HandlerSpec | ModuleSpec
                 PortGenerator = fun() (no arguments)
                 Error = term()
                 HandlerSpec = {HandlerFun, InitialData}
                 HandlerFun = fun() (two arguments)
                 ModuleSpec = fun() (no arguments) | {TracerModule, TracerState}
                 TracerModule = atom()
                 InitialData = TracerState = term()

              This  function  starts  a  tracer  server  with additional parameters on the local node. The first
              parameter, the Type, indicates if  trace  messages  should  be  handled  by  a  receiving  process
              (process),  by a tracer port (port) or by a tracer module (module). For a description about tracer
              ports see trace_port/2 and for a tracer modules see erl_tracer.

              If Type is process, a message  handler  function  can  be  specified  (HandlerSpec).  The  handler
              function,  which should be a fun taking two arguments, will be called for each trace message, with
              the first argument containing the message as it is and the second argument containing  the  return
              value  from the last invocation of the fun. The initial value of the second parameter is specified
              in the InitialData part of the HandlerSpec. The HandlerFun may choose any  appropriate  action  to
              take when invoked, and can save a state for the next invocation by returning it.

              If  Type is port, then the second parameter should be a fun which takes no arguments and returns a
              newly opened trace port when called. Such a fun is preferably generated by calling trace_port/2.

              if Type is module, then the second parameter should be either a tuple  describing  the  erl_tracer
              module  to  be used for tracing and the state to be used for that tracer module or a fun returning
              the same tuple.

              If  an  error  is  returned,  it  can  either  be  due  to  a  tracer   server   already   running
              ({error,already_started}) or due to the HandlerFun throwing an exception.

              To start a similar tracer on a remote node, use tracer/3.

       tracer(Nodename, Type, Data) -> {ok, Nodename} | {error, Reason}

              Types:

                 Nodename = atom()

              This  function  is  equivalent to tracer/2, but acts on the given node. A tracer is started on the
              node (Nodename) and the node is added to the list of traced nodes.

          Note:
              This function is not equivalent to n/1. While n/1 starts a  process  tracer  which  redirects  all
              trace  information  to  a process tracer on the local node (i.e. the trace control node), tracer/3
              starts a tracer of any type which is independent of the tracer on the trace control node.

              For details, see tracer/2.

       trace_port(Type, Parameters) -> fun()

              Types:

                 Type = ip | file
                 Parameters = Filename | WrapFilesSpec | IPPortSpec
                 Filename = string() | [string()] | atom()
                 WrapFilesSpec = {Filename, wrap, Suffix} | {Filename,  wrap,  Suffix,  WrapSize}  |  {Filename,
                 wrap, Suffix, WrapSize, WrapCnt}
                 Suffix = string()
                 WrapSize = integer() >= 0 | {time, WrapTime}
                 WrapTime = integer() >= 1
                 WrapCnt = integer() >= 1
                 IpPortSpec = PortNumber | {PortNumber, QueSize}
                 PortNumber = integer()
                 QueSize = integer()

              This  function creates a trace port generating fun. The fun takes no arguments and returns a newly
              opened trace port. The return value from this function  is  suitable  as  a  second  parameter  to
              tracer/2, i.e. dbg:tracer(port, dbg:trace_port(ip, 4711)).

              A  trace  port  is  an  Erlang  port to a dynamically linked in driver that handles trace messages
              directly, without the overhead of sending them as messages in the Erlang virtual machine.

              Two trace drivers are currently implemented, the file and the ip trace drivers.  The  file  driver
              sends  all  trace  messages into one or several binary files, from where they later can be fetched
              and processed with the trace_client/2 function. The ip driver opens a TCP/IP port where it listens
              for connections. When a client (preferably started by calling  trace_client/2  on  another  Erlang
              node)  connects,  all  trace  messages  are sent over the IP network for further processing by the
              remote client.

              Using a trace port significantly lowers the overhead imposed by using tracing.

              The file trace driver expects a filename or a wrap files specification as  parameter.  A  file  is
              written  with a high degree of buffering, why all trace messages are not guaranteed to be saved in
              the file in case of a system crash. That is the price to pay for low tracing overhead.

              A wrap files specification is used to limit the disk space consumed by the  trace.  The  trace  is
              written  to  a limited number of files each with a limited size. The actual filenames are Filename
              ++ SeqCnt ++ Suffix, where SeqCnt counts as a decimal string from 0 to  WrapCnt  and  then  around
              again  from  0.  When a trace term written to the current file makes it longer than WrapSize, that
              file is closed, if the number of files in this wrap trace is as many as WrapCnt the oldest file is
              deleted then a new file is opened to become the current. Thus, when a wrap trace has been stopped,
              there are at most WrapCnt trace files saved with a  size  of  at  least  WrapSize  (but  not  much
              bigger),  except  for  the  last  file that might even be empty. The default values are WrapSize =
              128*1024 and WrapCnt = 8.

              The SeqCnt values in the filenames are all in the range 0  through  WrapCnt  with  a  gap  in  the
              circular sequence. The gap is needed to find the end of the trace.

              If the WrapSize is specified as {time, WrapTime}, the current file is closed when it has been open
              more than WrapTime milliseconds, regardless of it being empty or not.

              The  ip trace driver has a queue of QueSize messages waiting to be delivered. If the driver cannot
              deliver messages as fast as they are produced by the runtime system, a special  message  is  sent,
              which  indicates  how  many  messages  that  are  dropped. That message will arrive at the handler
              function specified in trace_client/3 as the tuple {drop, N} where N is the number  of  consecutive
              messages  dropped.  In case of heavy tracing, drop's are likely to occur, and they surely occur if
              no client is reading the trace messages. The default value of QueSize is 200.

       flush_trace_port()

              Equivalent to flush_trace_port(node()).

       flush_trace_port(Nodename) -> ok | {error, Reason}

              Equivalent to trace_port_control(Nodename,flush).

       trace_port_control(Operation)

              Equivalent to trace_port_control(node(),Operation).

       trace_port_control(Nodename,Operation) -> ok | {ok, Result} | {error, Reason}

              Types:

                 Nodename = atom()

              This function is used to do a control operation on the active trace port driver on the given  node
              (Nodename).  Which  operations  are  allowed  as well as their return values depend on which trace
              driver is used.

              Returns either ok or {ok, Result} if the operation was  successful,  or  {error,  Reason}  if  the
              current tracer is a process or if it is a port not supporting the operation.

              The allowed values for Operation are:

                flush:
                  This  function  is  used  to flush the internal buffers held by a trace port driver. Currently
                  only the file trace driver supports this operation. Returns ok.

                get_listen_port:
                  Returns {ok, IpPort} where IpPort is the IP port number used by the driver listen socket. Only
                  the ip trace driver supports this operation.

       trace_client(Type, Parameters) -> pid()

              Types:

                 Type = ip | file | follow_file
                 Parameters = Filename | WrapFilesSpec | IPClientPortSpec
                 Filename = string() | [string()] | atom()
                 WrapFilesSpec = see trace_port/2
                 Suffix = string()
                 IpClientPortSpec = PortNumber | {Hostname, PortNumber}
                 PortNumber = integer()
                 Hostname = string()

              This function starts a trace client that reads the output created  by  a  trace  port  driver  and
              handles it in mostly the same way as a tracer process created by the tracer/0 function.

              If  Type  is  file,  the  client  reads  all  trace  messages stored in the file named Filename or
              specified by WrapFilesSpec (must be the same as used when creating the  trace,  see  trace_port/2)
              and  let's  the  default  handler  function format the messages on the console. This is one way to
              interpret the data stored in a file by the file trace port driver.

              If Type is follow_file, the client behaves as in the file case, but  keeps  trying  to  read  (and
              process)  more  data  from  the  file  until  stopped by stop_trace_client/1. WrapFilesSpec is not
              allowed as second argument for this Type.

              If Type is ip, the client connects to the TCP/IP port PortNumber on the host Hostname, from  where
              it  reads  trace  messages until the TCP/IP connection is closed. If no Hostname is specified, the
              local host is assumed.

              As an example, one can let trace messages  be  sent  over  the  network  to  another  Erlang  node
              (preferably not distributed), where the formatting occurs:

              On the node stack there's an Erlang node ant@stack, in the shell, type the following:

              ant@stack> dbg:tracer(port, dbg:trace_port(ip,4711)).
              <0.17.0>
              ant@stack> dbg:p(self(), send).
              {ok,1}

              All trace messages are now sent to the trace port driver, which in turn listens for connections on
              the  TCP/IP port 4711. If we want to see the messages on another node, preferably on another host,
              we do like this:

              -> dbg:trace_client(ip, {"stack", 4711}).
              <0.42.0>

              If we now send a message from the shell on the node ant@stack, where all sends from the shell  are
              traced:

              ant@stack> self() ! hello.
              hello

              The following will appear at the console on the node that started the trace client:

              (<0.23.0>) <0.23.0> ! hello
              (<0.23.0>) <0.22.0> ! {shell_rep,<0.23.0>,{value,hello,[],[]}}

              The  last  line is generated due to internal message passing in the Erlang shell. The process id's
              will vary.

       trace_client(Type, Parameters, HandlerSpec) -> pid()

              Types:

                 Type = ip | file | follow_file
                 Parameters = Filename | WrapFilesSpec | IPClientPortSpec
                 Filename = string() | [string()] | atom()
                 WrapFilesSpec = see trace_port/2
                 Suffix = string()
                 IpClientPortSpec = PortNumber | {Hostname, PortNumber}
                 PortNumber = integer()
                 Hostname = string()
                 HandlerSpec = {HandlerFun, InitialData}
                 HandlerFun = fun() (two arguments)
                 InitialData = term()

              This function works exactly as trace_client/2, but allows you to write your own handler  function.
              The  handler  function  works  mostly  as  the one described in tracer/2, but will also have to be
              prepared to handle trace messages of the form  {drop,  N},  where  N  is  the  number  of  dropped
              messages. This pseudo trace message will only occur if the ip trace driver is used.

              For  trace  type  file, the pseudo trace message end_of_trace will appear at the end of the trace.
              The return value from the handler function is in this case ignored.

       stop_trace_client(Pid) -> ok

              Types:

                 Pid = pid()

              This function shuts down a previously started trace client. The Pid argument  is  the  process  id
              returned from the trace_client/2 or trace_client/3 call.

       get_tracer()

              Equivalent to get_tracer(node()).

       get_tracer(Nodename) -> {ok, Tracer}

              Types:

                 Nodename = atom()
                 Tracer = port() | pid() | {module(), term()}

              Returns the process, port or tracer module to which all trace messages are sent.

       stop() -> ok

              Stops  the dbg server, clears all trace flags for all processes, clears all trace patterns for all
              functions, clears trace patterns for send/receive, shuts down all trace clients,  and  closes  all
              trace ports.

SIMPLE EXAMPLES - TRACING FROM THE SHELL

       The simplest way of tracing from the Erlang shell is to use dbg:c/3 or dbg:c/4, e.g. tracing the function
       dbg:get_tracer/0:

       (tiger@durin)84> dbg:c(dbg,get_tracer,[]).
       (<0.154.0>) <0.152.0> ! {<0.154.0>,{get_tracer,tiger@durin}}
       (<0.154.0>) out {dbg,req,1}
       (<0.154.0>) << {dbg,{ok,<0.153.0>}}
       (<0.154.0>) in {dbg,req,1}
       (<0.154.0>) << timeout
       {ok,<0.153.0>}
       (tiger@durin)85>

       Another  way  of  tracing  from the shell is to explicitly start a tracer and then set the trace flags of
       your choice on the processes you want to trace, e.g. trace messages and process events:

       (tiger@durin)66> Pid = spawn(fun() -> receive {From,Msg} -> From ! Msg end end).
       <0.126.0>
       (tiger@durin)67> dbg:tracer().
       {ok,<0.128.0>}
       (tiger@durin)68> dbg:p(Pid,[m,procs]).
       {ok,[{matched,tiger@durin,1}]}
       (tiger@durin)69> Pid ! {self(),hello}.
       (<0.126.0>) << {<0.116.0>,hello}
       {<0.116.0>,hello}
       (<0.126.0>) << timeout
       (<0.126.0>) <0.116.0> ! hello
       (<0.126.0>) exit normal
       (tiger@durin)70> flush().
       Shell got hello
       ok
       (tiger@durin)71>

       If you set the call trace flag, you also have to set a trace pattern for the functions you want to trace:

       (tiger@durin)77> dbg:tracer().
       {ok,<0.142.0>}
       (tiger@durin)78> dbg:p(all,call).
       {ok,[{matched,tiger@durin,3}]}
       (tiger@durin)79> dbg:tp(dbg,get_tracer,0,[]).
       {ok,[{matched,tiger@durin,1}]}
       (tiger@durin)80> dbg:get_tracer().
       (<0.116.0>) call dbg:get_tracer()
       {ok,<0.143.0>}
       (tiger@durin)81> dbg:tp(dbg,get_tracer,0,[{'_',[],[{return_trace}]}]).
       {ok,[{matched,tiger@durin,1},{saved,1}]}
       (tiger@durin)82> dbg:get_tracer().
       (<0.116.0>) call dbg:get_tracer()
       (<0.116.0>) returned from dbg:get_tracer/0 -> {ok,<0.143.0>}
       {ok,<0.143.0>}
       (tiger@durin)83>

ADVANCED TOPICS - COMBINING WITH SEQ_TRACE

       The dbg module is primarily targeted towards tracing through the erlang:trace/3 function. It is sometimes
       desired to trace messages in a more delicate way, which can be  done  with  the  help  of  the  seq_trace
       module.

       seq_trace  implements  sequential  tracing  (known  in  the  AXE10  world,  and sometimes called "forlopp
       tracing"). dbg can interpret messages generated from seq_trace and the  same  tracer  function  for  both
       types  of  tracing  can  be  used.  The  seq_trace  messages can even be sent to a trace port for further
       analysis.

       As a match specification can turn on sequential tracing, the combination of  dbg  and  seq_trace  can  be
       quite powerful. This brief example shows a session where sequential tracing is used:

       1> dbg:tracer().
       {ok,<0.30.0>}
       2> {ok, Tracer} = dbg:get_tracer().
       {ok,<0.31.0>}
       3> seq_trace:set_system_tracer(Tracer).
       false
       4> dbg:tp(dbg, get_tracer, 0, [{[],[],[{set_seq_token, send, true}]}]).
       {ok,[{matched,nonode@nohost,1},{saved,1}]}
       5> dbg:p(all,call).
       {ok,[{matched,nonode@nohost,22}]}
       6> dbg:get_tracer(), seq_trace:set_token([]).
       (<0.25.0>) call dbg:get_tracer()
       SeqTrace [0]: (<0.25.0>) <0.30.0> ! {<0.25.0>,get_tracer} [Serial: {2,4}]
       SeqTrace [0]: (<0.30.0>) <0.25.0> ! {dbg,{ok,<0.31.0>}} [Serial: {4,5}]
       {1,0,5,<0.30.0>,4}

       This  session  sets the system_tracer to the same process as the ordinary tracer process (i. e. <0.31.0>)
       and sets the trace pattern for the function dbg:get_tracer to one  that  has  the  action  of  setting  a
       sequential  token.  When  the  function  is  called by a traced process (all processes are traced in this
       case), the process gets "contaminated" by the token and seq_trace messages are sent both for  the  server
       request  and  the response. The seq_trace:set_token([]) after the call clears the seq_trace token, why no
       messages are sent when the answer propagates via  the  shell  to  the  console  port.  The  output  would
       otherwise have been more noisy.

NOTE OF CAUTION

       When  tracing  function  calls  on  a  group  leader  process (an IO process), there is risk of causing a
       deadlock. This will happen if a group leader process generates a trace message and the tracer process, by
       calling the trace handler function, sends an IO request to the same group leader. The  problem  can  only
       occur  if  the  trace  handler  prints  to  tty  using  an  io  function such as format/2. Note that when
       dbg:p(all,call) is called, IO processes are also traced. Here's an example:

       %% Using a default line editing shell
       1> dbg:tracer(process, {fun(Msg,_) -> io:format("~p~n", [Msg]), 0 end, 0}).
       {ok,<0.37.0>}
       2> dbg:p(all, [call]).
       {ok,[{matched,nonode@nohost,25}]}
       3> dbg:tp(mymod,[{'_',[],[]}]).
       {ok,[{matched,nonode@nohost,0},{saved,1}]}
       4> mymod: % TAB pressed here
       %% -- Deadlock --

       Here's another example:

       %% Using a shell without line editing (oldshell)
       1> dbg:tracer(process).
       {ok,<0.31.0>}
       2> dbg:p(all, [call]).
       {ok,[{matched,nonode@nohost,25}]}
       3> dbg:tp(lists,[{'_',[],[]}]).
       {ok,[{matched,nonode@nohost,0},{saved,1}]}
       % -- Deadlock --

       The reason we get a deadlock in the first example is because when TAB is pressed to expand  the  function
       name,  the group leader (which handles character input) calls mymod:module_info(). This generates a trace
       message which, in turn, causes the tracer process to send an IO request to the group leader  (by  calling
       io:format/2). We end up in a deadlock.

       In the second example we use the default trace handler function. This handler prints to tty by sending IO
       requests  to  the user process. When Erlang is started in oldshell mode, the shell process will have user
       as its group leader and so will the tracer process in this example. Since user calls functions  in  lists
       we end up in a deadlock as soon as the first IO request is sent.

       Here are a few suggestions for how to avoid deadlock:

         * Don't  trace  the  group  leader  of  the  tracer  process.  If  tracing has been switched on for all
           processes,  call  dbg:p(TracerGLPid,clear)  to  stop  tracing   the   group   leader   (TracerGLPid).
           process_info(TracerPid,group_leader)  tells  you  which  process  this is (TracerPid is returned from
           dbg:get_tracer/0).

         * Don't trace the user process if using the default trace handler function.

         * In your own trace handler function, call erlang:display/1 instead of an io function or,  if  user  is
           not   used   as  group  leader,  print  to  user  instead  of  the  default  group  leader.  Example:
           io:format(user,Str,Args).

Ericsson AB                                    runtime_tools 1.19                                      dbg(3erl)