Awesome
ghc-debug
is a set of libraries which allow you to inspect the heap of
a running Haskell program from an external debugger.
For example, you could use this library to
- Implement a memory profiler, written in Haskell - GHC.Debug.Profile
- Precisely analyse other heap properties such as retainers - GHC.Debug.Retainers
- Work out any other question you want about the heap by writing your own custom analysis. The possibilities are endless!
Getting Started
There are two parts to using ghc-debug
. Firstly the application you want to
inspect has to be instrumented using the withGhcDebug
function from
GHC.Debug.Stub
. This just wraps the normal main function of your executable,
when it is executed it will create a socket by which a debugger can connect
and issue requests to. The location of the socket can be controlled by setting
the GHC_DEBUG_SOCKET
variable when the executable is run.
import GHC.Debug.Stub
main = withGhcDebug normalMain
Note: To enable source information you should also compile your application and
dependencies with -finfo-table-map
and optionally -fdistinct-constructor-tables
.
A simple debugger
The most productive way to use ghc-debug
is to write your own heap analysis
scripts. Fortunately, this is also quite simple. Here is a simple, complete, debugger
which connects to the /tmp/ghc-debug
socket, requests the GC roots and then
decodes the first one up to depth 10 before printing the result to the user.
import GHC.Debug.Client
main = withDebuggeeConnect "/tmp/ghc-debug" p1
p1 :: Debuggee -> IO ()
p1 e = do
pause e
g <- run e $ do
precacheBlocks
(r:_) <- gcRoots
buildHeapGraph (Just 10) r
putStrLn (ppHeapGraph (const "") h)
The API for writing debuggers is described in the GHC.Debug.Client
module.
There are many more examples in the test/Test.hs
file.
Snapshotting
A convenient way to use ghc-debug
is to take a snapshot of the heap and then
perform further analysis on the snapshot rather than connecting to a running
process. Snapshotting utilities are in the GHC.Debug.Snapshot
module. A
snapshot can be created using the makeSnapshot
program, it will pause
the process and then save a snapshot to the /tmp/ghc-debug-snapshot
file.
import GHC.Debug.Client
import GHC.Debug.Snapshot
main = withDebuggeeConnect "/tmp/ghc-debug" (\d -> makeSnapshot d "/tmp/ghc-debug-snapshot")
A snapshot can be then used for further analysis. For example, we can run p1
on
the snapshot by using snapshotRun
instead of withDebuggeeConnect
. The same
programs can be used with snapshots but requests such as pausing and resuming are
just ignored.
import GHC.Debug.Client
main = snapshotRun "/tmp/ghc-debug-snapshot" p1
High-Level Analysis
There are also some more high-level analysis tools already packaged with the library. Mostly as an idea about what sort of thing you could program yourself.
- Profiling - Profiling modes in the spirit of
-hT
. - Object Equivalence - Detect equivalent heap objects which could be shared.
- Count - Simple heap statistics, total number of objects, total size and maximum object size.
- Fragmentation - Functions for analysis memory fragmentation including block and mblock utilisation histograms.
- Retainers - Finding paths through the heap to work out why objects are being retained.
- Type Points From - Collapse a heap graph so that nodes are info tables and edges are references between info tables. This allows you to implement the Cork leak analysis.
These analysis modes are implemented in terms of the more low-level traversal functions.
- Sequential Traversal - Traversal with low memory overhead, accounting for cycles.
- Parallel Traversal - Experimental Parallel Traversal with low memory overhead.
Other Resources
How does it work?
We call the process we want to debug the debuggee and the process which does
the debugging the debugger.
Whilst the debuggee is
running it calls the C function start
which creates a unix domain socket (which is set from GHC_DEBUG_SOCKET
). The debugger starts and connects to the socket.
Once the debugger is connected it can send requests to the debuggee to control and inspect the RTS.
How do I use it?
The project needs to built with a development version of GHC from this branch. Then you can use normal cabal commands to build this library. The easiest way to do this is with nix. The nix shell uses the development version of GHC so you don't have to build it yourself.
$ nix-shell
$ cabal new-build all
Automated Testing
There are hspec
tests, that can be run with cabal
:
cabal new-test all
Unexpected Build Failures
If you encounter dependencies failing to build but there's a patch for
the library in head.hackage then you may need to delete ~/.cabal/packages/head.hackage.org
so that the fresh patch is visible. This is probably a bug in cabal!