Module recon_trace
recon_trace
is a module that handles tracing in a safe manner for single
Erlang nodes, currently for function calls only.
Authors: Fred Hebert (mononcqc@ferd.ca
) [web site: http://ferd.ca/
].
Description
recon_trace
is a module that handles tracing in a safe manner for single
Erlang nodes, currently for function calls only. Functionality includes:
- Nicer to use interface (arguably) than
dbg
or trace BIFs. - Protection against dumb decisions (matching all calls on a node being traced, for example)
- Adding safe guards in terms of absolute trace count or rate-limitting
- Nicer formatting than default traces
Tracing Erlang Code
The Erlang Trace BIFs allow to trace any Erlang code at all. They work in two parts: pid specifications, and trace patterns.
Pid specifications let you decide which processes to target. They can be
specific pids, all
pids, existing
pids, or new
pids (those not
spawned at the time of the function call).
The trace patterns represent functions. Functions can be specified in two
parts: specifying the modules, functions, and arguments, and then with
Erlang match specifications to add constraints to arguments (see
calls/3
for details).
What defines whether you get traced or not is the intersection of both:
_,--------,_ _,--------,_ ,-' `-,,-' `-, ,-' ,-' '-, `-, | Matching -' '- Matching | | Pids | Getting | Trace | | | Traced | Patterns | | -, ,- | '-, '-, ,-' ,-' '-,_ _,-''-,_ _,-' '--------' '--------'
If either the pid specification excludes a process or a trace pattern excludes a given call, no trace will be received.
Example Session
First let's trace the queue:new
functions in any process:
1> recon_trace:calls({queue, new, '_'}, 1). 1 13:14:34.086078 <0.44.0> queue:new() Recon tracer rate limit tripped.
The limit was set to 1
trace message at most, and recon
let us
know when that limit was reached.
Let's instead look for all the queue:in/2
calls, to see what it is
we're inserting in queues:
2> recon_trace:calls({queue, in, 2}, 1). 1 13:14:55.365157 <0.44.0> queue:in(a, {[],[]}) Recon tracer rate limit tripped.
In order to see the content we want, we should change the trace patterns
to use a fun
that matches on all arguments in a list (_
) and returns
return_trace()
. This last part will generate a second trace for each
call that includes the return value:
3> recon_trace:calls({queue, in, fun(_) -> return_trace() end}, 3). 1 13:15:27.655132 <0.44.0> queue:in(a, {[],[]}) 13:15:27.655467 <0.44.0> queue:in/2 --> {[a],[]} 13:15:27.757921 <0.44.0> queue:in(a, {[],[]}) Recon tracer rate limit tripped.
Matching on argument lists can be done in a more complex manner:
4> recon_trace:calls( 4> {queue, '_', fun([A,_]) when is_list(A); is_integer(A) andalso A > 1 -> return_trace() end}, 4> {10,100} 4> ). 32 13:24:21.324309 <0.38.0> queue:in(3, {[],[]}) 13:24:21.371473 <0.38.0> queue:in/2 --> {[3],[]} 13:25:14.694865 <0.53.0> queue:split(4, {[10,9,8,7],[1,2,3,4,5,6]}) 13:25:14.695194 <0.53.0> queue:split/2 --> {{[4,3,2],[1]},{[10,9,8,7],[5,6]}} 5> recon_trace:clear(). ok
Note that in the pattern above, no specific function ('_'
) was
matched against. Instead, the fun
used restricted functions to those
having two arguments, the first of which is either a list or an integer
greater than 1
.
The limit was also set using {10,100}
instead of an integer, making the
rate-limitting at 10 messages per 100 milliseconds, instead of an absolute
value.
Any tracing can be manually interrupted by calling recon_trace:clear()
,
or killing the shell process.
Be aware that extremely broad patterns with lax rate-limitting (or very
high absolute limits) may impact your node's stability in ways
recon_trace
cannot easily help you with.
In doubt, start with the most restrictive tracing possible, with low limits, and progressively increase your scope.
See calls/3
for more details and tracing possibilities.
Structure
This library is production-safe due to taking the following structure for tracing:
[IO/Group leader] <---------------------, | | [shell] ---> [tracer process] ----> [formatter]
The tracer process receives trace messages from the node, and enforces limits in absolute terms or trace rates, before forwarding the messages to the formatter. This is done so the tracer can do as little work as possible and never block while building up a large mailbox.
The tracer process is linked to the shell, and the formatter to the tracer process. The formatter also traps exits to be able to handle all received trace messages until the tracer termination, but will then shut down as soon as possible.
In case the operator is tracing from a remote shell which gets disconnected, the links between the shell and the tracer should make it so tracing is automatically turned off once you disconnect.Data Types
args()
args() = '_' | 0..255 | matchspec() | shellfun()
fn()
fn() = '_' | atom()
matchspec()
matchspec() = [{[term()], [term()], [term()]}]
max()
max() = max_traces() | max_rate()
max_rate()
max_rate() = {max_traces(), millisecs()}
max_traces()
max_traces() = non_neg_integer()
mfa()
millisecs()
millisecs() = non_neg_integer()
mod()
mod() = '_' | module()
num_matches()
num_matches() = non_neg_integer()
options()
options() = [{pid, pidspec() | [pidspec(), ...]} | {timestamp, formatter | trace} | {args, args | arity} | {scope, global | local}]
pidspec()
pidspec() = all | existing | new | recon:pid_term()
shellfun()
shellfun() = fun((term()) -> term())
Function Index
calls/2 | Equivalent to calls({Mod, Fun, Args}, Max, []) .
|
calls/3 | Allows to set trace patterns and pid specifications to trace function calls. |
clear/0 | Stops all tracing at once. |
Function Details
calls/2
calls(X1::mfa(), Max::max()) -> num_matches()
Equivalent to calls({Mod, Fun, Args}, Max, [])
.
calls/3
calls(MFAs::mfa() | [mfa(), ...], Max::max(), Opts::options()) -> num_matches()
Allows to set trace patterns and pid specifications to trace function calls.
The basic calls take the trace patterns as tuples of the form
{Module, Function, Args}
where:
Module
is any atom representing a moduleFunction
is any atom representing a function, or the wildcard'_'
Args
is either the arity of a function (0..255
), a wildcard pattern ('_'
), a match specification, or a function from a shell session that can be transformed into a match specification
There is also an argument specifying either a maximal count (a number)
of trace messages to be received, or a maximal frequency ({Num, Millisecs}
).
Here are examples of things to trace:
- All calls from the
queue
module, with 10 calls printed at most:recon_trace:calls({queue, '_', '_'}, 10)
- All calls to
lists:seq(A,B)
, with 100 calls printed at most:recon_trace:calls({lists, seq, 2}, 100)
- All calls to
lists:seq(A,B)
, with 100 calls per second at most:recon_trace:calls({lists, seq, 2}, {100, 1000})
- All calls to
lists:seq(A,B,2)
(all sequences increasing by two) with 100 calls at most:recon_trace:calls({lists, seq, fun([_,_,2]) -> ok end}, 100)
- All calls to
iolist_to_binary/1
made with a binary as an argument already (kind of useless conversion!):recon_trace:calls({erlang, iolist_to_binary, fun([X]) when is_binary(X) -> ok end}, 10)
- Calls to the queue module only in a given process
Pid
, at a rate of 50 per second at most:recon_trace:calls({queue, '_', '_'}, {50,1000}, [{pid, Pid}])
- Print the traces with the function arity instead of literal arguments:
recon_trace:calls(MFA, Max, [{args, arity}])
- Matching the
filter/2
functions of bothdict
andlists
modules, across new processes only:recon_trace:calls([{dict,filter,2},{lists,filter,2}], 10, [{pid, new]})
- Tracing the
handle_call/3
functions of a given module for all new processes, and those of an existing one registered withgproc
:recon_trace:calls({Mod,handle_call,3}, {10,100}, [{pid, [{via, gproc, Name}, new]}
- Show the result of a given function call:
recon_trace:calls({Mod,Fun,fun(_) -> return_trace() end}, Max, Opts)
orrecon_trace:calls({Mod,Fun,[{'_', [], [{return_trace}]}]}, Max, Opts)
, the important bit being thereturn_trace()
call or the{return_trace}
match spec value.
There's a few more combination possible, with multiple trace patterns per call, and more options:
{pid, PidSpec}
: which processes to trace. Valid options is any ofall
,new
,existing
, or a process descriptor ({A,B,C}
,"<A.B.C>"
, an atom representing a name,{global, Name}
,{via, Registrar, Name}
, or a pid). It's also possible to specify more than one by putting them in a list.{timestamp, formatter | trace}
: by default, the formatter process adds timestamps to messages received. If accurate timestamps are required, it's possible to force the usage of timestamps within trace messages by adding the option{timestamp, trace}
.{args, arity | args}
: whether to print arity in function calls or their (by default) literal representation.{scope, global | local}
: by default, only 'global' (fully qualified function calls) are traced, not calls made internally. To force tracing of local calls, pass in{scope, local}
. This is useful whenever you want to track the changes of code in a process that isn't called withModule:Fun(Args)
, but justFun(Args)
.
Max
values (i.e. 99999999
or
{10000,1}
) will probably negate most of the safe-guarding this library
does and be dangerous to your node. Similarly, tracing extremely large
amounts of function calls (all of them, or all of io
for example)
can be risky if more trace messages are generated than any process on
the node could ever handle, despite the precautions taken by this library.
clear/0
clear() -> ok
Stops all tracing at once.