278 lines
15 KiB
Markdown
Raw Normal View History

# Debugging Garbage Collector Related Problems
This page contains some hints on debugging issues specific to the
Boehm-Demers-Weiser conservative garbage collector. It applies both
to debugging issues in client code that manifest themselves as collector
misbehavior, and to debugging the collector itself.
If you suspect a bug in the collector itself, it is strongly recommended that
you try the latest collector release before proceeding.
## Bus Errors and Segmentation Violations
If the fault occurred in `GC_find_limit`, or with incremental collection
enabled, this is probably normal. The collector installs handlers to take care
of these. You will not see these unless you are using a debugger. Your
debugger _should_ allow you to continue. It's often preferable to tell the
debugger to ignore SIGBUS and SIGSEGV ("handle SIGSEGV SIGBUS nostop noprint"
in gdb, "ignore SIGSEGV SIGBUS" in most versions of dbx) and set a breakpoint
in `abort`. The collector will call abort if the signal had another cause, and
there was not other handler previously installed.
We recommend debugging without incremental collection if possible. (This
applies directly to UNIX systems. Debugging with incremental collection under
win32 is worse. See README.win32.)
If the application generates an unhandled SIGSEGV or equivalent, it may often
be easiest to set the environment variable `GC_LOOP_ON_ABORT`. On many
platforms, this will cause the collector to loop in a handler when the SIGSEGV
is encountered (or when the collector aborts for some other reason), and
a debugger can then be attached to the looping process. This sidesteps common
operating system problems related to incomplete core files for multi-threaded
applications, etc.
## Other Signals
On most platforms, the multi-threaded version of the collector needs one or
two other signals for internal use by the collector in stopping threads. It is
normally wise to tell the debugger to ignore these. On Linux, the collector
currently uses SIGPWR and SIGXCPU by default.
## Warning Messages About Needing to Allocate Blacklisted Blocks
The garbage collector generates warning messages of the form:
Needed to allocate blacklisted block at 0x...
or
Repeated allocation of very large block ...
when it needs to allocate a block at a location that it knows to be referenced
by a false pointer. These false pointers can be either permanent (e.g.
a static integer variable that never changes) or temporary. In the latter
case, the warning is largely spurious, and the block will eventually
be reclaimed normally. In the former case, the program will still run
correctly, but the block will never be reclaimed. Unless the block is intended
to be permanent, the warning indicates a memory leak.
1. Ignore these warnings while you are using GC_DEBUG. Some of the routines
mentioned below don't have debugging equivalents. (Alternatively, write the
missing routines and send them to me.)
2. Replace allocator calls that request large blocks with calls to
`GC_malloc_ignore_off_page` or `GC_malloc_atomic_ignore_off_page`. You may
want to set a breakpoint in `GC_default_warn_proc` to help you identify such
calls. Make sure that a pointer to somewhere near the beginning of the
resulting block is maintained in a (preferably volatile) variable as long
as the block is needed.
3. If the large blocks are allocated with realloc, we suggest instead
allocating them with something like the following. Note that the realloc
size increment should be fairly large (e.g. a factor of 3/2) for this to
exhibit reasonable performance. But we all know we should do that anyway.
void * big_realloc(void *p, size_t new_size) {
size_t old_size = GC_size(p);
void * result;
if (new_size <= 10000) return(GC_realloc(p, new_size));
if (new_size <= old_size) return(p);
result = GC_malloc_ignore_off_page(new_size);
if (result == 0) return(0);
memcpy(result,p,old_size);
GC_free(p);
return(result);
}
4. In the unlikely case that even relatively small object (<20KB)
allocations are triggering these warnings, then your address space contains
lots of "bogus pointers", i.e. values that appear to be pointers but aren't.
Usually this can be solved by using `GC_malloc_atomic` or the routines
in `gc_typed.h` to allocate large pointer-free regions of bitmaps, etc.
Sometimes the problem can be solved with trivial changes of encoding
in certain values. It is possible, to identify the source of the bogus
pointers by building the collector with `-DPRINT_BLACK_LIST`, which will
cause it to print the "bogus pointers", along with their location.
5. If you get only a fixed number of these warnings, you are probably only
introducing a bounded leak by ignoring them. If the data structures being
allocated are intended to be permanent, then it is also safe to ignore them.
The warnings can be turned off by calling `GC_set_warn_proc` with
a procedure that ignores these warnings (e.g. by doing absolutely nothing).
## The Collector References a Bad Address in GC_malloc
This typically happens while the collector is trying to remove an entry from
its free list, and the free list pointer is bad because the free list link
in the last allocated object was bad.
With >99% probability, you wrote past the end of an allocated object. Try
setting `GC_DEBUG` before including `gc.h` and allocating with `GC_MALLOC`.
This will try to detect such overwrite errors.
## Unexpectedly Large Heap
Unexpected heap growth can be due to one of the following:
1. Data structures that are being unintentionally retained. This is commonly
caused by data structures that are no longer being used, but were not
cleared, or by caches growing without bounds.
2. Pointer misidentification. The garbage collector is interpreting integers
or other data as pointers and retaining the "referenced" objects. A common
symptom is that GC_dump() shows much of the heap as black-listed.
3. Heap fragmentation. This should never result in unbounded growth, but
it may account for larger heaps. This is most commonly caused by allocation
of large objects.
4. Per object overhead. This is usually a relatively minor effect, but
it may be worth considering. If the collector recognizes interior pointers,
object sizes are increased, so that one-past-the-end pointers are correctly
recognized. The collector can be configured not to do this
(`-DDONT_ADD_BYTE_AT_END`).
The collector rounds up object sizes so the result fits well into the chunk
size (`HBLKSIZE`, normally 4K on 32 bit machines, 8K on 64 bit machines) used
by the collector. Thus it may be worth avoiding objects of size 2K + 1 (or 2K
if a byte is being added at the end.) The last two cases can often
be identified by looking at the output of a call to `GC_dump`. Among other
things, it will print the list of free heap blocks, and a very brief
description of all chunks in the heap, the object sizes they correspond to,
and how many live objects were found in the chunk at the last collection.
Growing data structures can usually be identified by:
1. Building the collector with `-DKEEP_BACK_PTRS`,
2. Preferably using debugging allocation (defining `GC_DEBUG` before
including `gc.h` and allocating with `GC_MALLOC`), so that objects will
be identified by their allocation site,
3. Running the application long enough so that most of the heap is composed
of "leaked" memory, and
4. Then calling `GC_generate_random_backtrace` from gc_backptr.h a few times
to determine why some randomly sampled objects in the heap are being
retained.
The same technique can often be used to identify problems with false pointers,
by noting whether the reference chains printed
by `GC_generate_random_backtrace` involve any misidentified pointers.
An alternate technique is to build the collector with `-DPRINT_BLACK_LIST`
which will cause it to report values that are almost, but not quite, look like
heap pointers. It is very likely that actual false pointers will come from
similar sources.
In the unlikely case that false pointers are an issue, it can usually
be resolved using one or more of the following techniques:
1. Use `GC_malloc_atomic` for objects containing no pointers. This is
especially important for large arrays containing compressed data,
pseudo-random numbers, and the like. It is also likely to improve GC
performance, perhaps drastically so if the application is paging.
2. If you allocate large objects containing only one or two pointers at the
beginning, either try the typed allocation primitives is`gc_typed.h`,
or separate out the pointer-free component.
3. Consider using `GC_malloc_ignore_off_page` to allocate large objects.
(See `gc.h` and above for details. Large means >100K in most environments.)
4. If your heap size is larger than 100MB or so, build the collector with
`-DLARGE_CONFIG`. This allows the collector to keep more precise black-list
information.
5. If you are using heaps close to, or larger than, a gigabyte on a 32-bit
machine, you may want to consider moving to a platform with 64-bit pointers.
This is very likely to resolve any false pointer issues.
## Prematurely Reclaimed Objects
The usual symptom of this is a segmentation fault, or an obviously overwritten
value in a heap object. This should, of course, be impossible. In practice,
it may happen for reasons like the following:
1. The collector did not intercept the creation of threads correctly
in a multi-threaded application, e.g. because the client called
`pthread_create` without including `gc.h`, which redefines it.
2. The last pointer to an object in the garbage collected heap was stored
somewhere were the collector could not see it, e.g. in an object allocated
with system `malloc`, in certain types of `mmap`ed files, or in some data
structure visible only to the OS. (On some platforms, thread-local storage
is one of these.)
3. The last pointer to an object was somehow disguised, e.g. by XORing
it with another pointer.
4. Incorrect use of `GC_malloc_atomic` or typed allocation.
5. An incorrect `GC_free` call.
6. The client program overwrote an internal garbage collector data
structure.
7. A garbage collector bug.
8. (Empirically less likely than any of the above.) A compiler optimization
that disguised the last pointer.
The following relatively simple techniques should be tried first to narrow
down the problem:
1. If you are using the incremental collector try turning it off for
debugging.
2. If you are using shared libraries, try linking statically. If that works,
ensure that DYNAMIC_LOADING is defined on your platform.
3. Try to reproduce the problem with fully debuggable unoptimized code. This
will eliminate the last possibility, as well as making debugging easier.
4. Try replacing any suspect typed allocation and `GC_malloc_atomic` calls
with calls to `GC_malloc`.
5. Try removing any `GC_free` calls (e.g. with a suitable `#define`).
6. Rebuild the collector with `-DGC_ASSERTIONS`.
7. If the following works on your platform (i.e. if gctest still works if
you do this), try building the collector with
`-DREDIRECT_MALLOC=GC_malloc_uncollectable`. This will cause the collector
to scan memory allocated with malloc.
If all else fails, you will have to attack this with a debugger. The suggested
steps are:
1. Call `GC_dump` from the debugger around the time of the failure. Verify
that the collectors idea of the root set (i.e. static data regions which
it should scan for pointers) looks plausible. If not, i.e. if it does not
include some static variables, report this as a collector bug. Be sure
to describe your platform precisely, since this sort of problem is nearly
always very platform dependent.
2. Especially if the failure is not deterministic, try to isolate
it to a relatively small test case.
3. Set a break point in `GC_finish_collection`. This is a good point
to examine what has been marked, i.e. found reachable, by the collector.
4. If the failure is deterministic, run the process up to the last
collection before the failure. Note that the variable `GC_gc_no` counts
collections and can be used to set a conditional breakpoint in the right
one. It is incremented just before the call to `GC_finish_collection`.
If object `p` was prematurely recycled, it may be helpful to look
at `*GC_find_header(p)` at the failure point. The `hb_last_reclaimed` field
will identify the collection number during which its block was last swept.
5. Verify that the offending object still has its correct contents at this
point. Then call `GC_is_marked(p)` from the debugger to verify that the
object has not been marked, and is about to be reclaimed. Note that
`GC_is_marked(p)` expects the real address of an object (the address of the
debug header if there is one), and thus it may be more appropriate to call
`GC_is_marked(GC_base(p))` instead.
6. Determine a path from a root, i.e. static variable, stack, or register
variable, to the reclaimed object. Call `GC_is_marked(q)` for each object
`q` along the path, trying to locate the first unmarked object, say `r`.
7. If `r` is pointed to by a static root, verify that the location pointing
to it is part of the root set printed by `GC_dump`. If it is on the stack
in the main (or only) thread, verify that `GC_stackbottom` is set correctly
to the base of the stack. If it is in another thread stack, check the
collector's thread data structure (`GC_thread[]` on several platforms)
to make sure that stack bounds are set correctly.
8. If `r` is pointed to by heap object `s`, check that the collector's
layout description for `s` is such that the pointer field will be scanned.
Call `*GC_find_header(s)` to look at the descriptor for the heap chunk.
The `hb_descr` field specifies the layout of objects in that chunk.
See `gc_mark.h` for the meaning of the descriptor. (If its low order 2 bits
are zero, then it is just the length of the object prefix to be scanned.
This form is always used for objects allocated with `GC_malloc` or
`GC_malloc_atomic`.)
9. If the failure is not deterministic, you may still be able to apply some
of the above technique at the point of failure. But remember that objects
allocated since the last collection will not have been marked, even if the
collector is functioning properly. On some platforms, the collector can
be configured to save call chains in objects for debugging. Enabling this
feature will also cause it to save the call stack at the point of the last
GC in `GC_arrays._last_stack`.
10. When looking at GC internal data structures remember that a number
of `GC_xxx` variables are really macro defined to `GC_arrays._xxx`, so that
the collector can avoid scanning them.