Annotation of OpenXM_contrib2/asir2000/gc/doc/debugging.html, Revision 1.1
1.1 ! noro 1: <HTML>
! 2: <HEAD>
! 3: <TITLE>Debugging Garbage Collector Related Problems</title>
! 4: </head>
! 5: <BODY>
! 6: <H1>Debugging Garbage Collector Related Problems</h1>
! 7: This page contains some hints on
! 8: debugging issues specific to
! 9: the Boehm-Demers-Weiser conservative garbage collector.
! 10: It applies both to debugging issues in client code that manifest themselves
! 11: as collector misbehavior, and to debugging the collector itself.
! 12: <P>
! 13: If you suspect a bug in the collector itself, it is strongly recommended
! 14: that you try the latest collector release, even if it is labelled as "alpha",
! 15: before proceeding.
! 16: <H2>Bus Errors and Segmentation Violations</h2>
! 17: <P>
! 18: If the fault occurred in GC_find_limit, or with incremental collection enabled,
! 19: this is probably normal. The collector installs handlers to take care of
! 20: these. You will not see these unless you are using a debugger.
! 21: Your debugger <I>should</i> allow you to continue.
! 22: It's often preferable to tell the debugger to ignore SIGBUS and SIGSEGV
! 23: ("<TT>handle SIGSEGV SIGBUS nostop noprint</tt>" in gdb,
! 24: "<TT>ignore SIGSEGV SIGBUS</tt>" in most versions of dbx)
! 25: and set a breakpoint in <TT>abort</tt>.
! 26: The collector will call abort if the signal had another cause,
! 27: and there was not other handler previously installed.
! 28: <P>
! 29: We recommend debugging without incremental collection if possible.
! 30: (This applies directly to UNIX systems.
! 31: Debugging with incremental collection under win32 is worse. See README.win32.)
! 32: <P>
! 33: If the application generates an unhandled SIGSEGV or equivalent, it may
! 34: often be easiest to set the environment variable GC_LOOP_ON_ABORT. On many
! 35: platforms, this will cause the collector to loop in a handler when the
! 36: SIGSEGV is encountered (or when the collector aborts for some other reason),
! 37: and a debugger can then be attached to the looping
! 38: process. This sidesteps common operating system problems related
! 39: to incomplete core files for multithreaded applications, etc.
! 40: <H2>Other Signals</h2>
! 41: On most platforms, the multithreaded version of the collector needs one or
! 42: two other signals for internal use by the collector in stopping threads.
! 43: It is normally wise to tell the debugger to ignore these. On Linux,
! 44: the collector currently uses SIGPWR and SIGXCPU by default.
! 45: <H2>Warning Messages About Needing to Allocate Blacklisted Blocks</h2>
! 46: The garbage collector generates warning messages of the form
! 47: <PRE>
! 48: Needed to allocate blacklisted block at 0x...
! 49: </pre>
! 50: when it needs to allocate a block at a location that it knows to be
! 51: referenced by a false pointer. These false pointers can be either permanent
! 52: (<I>e.g.</i> a static integer variable that never changes) or temporary.
! 53: In the latter case, the warning is largely spurious, and the block will
! 54: eventually be reclaimed normally.
! 55: In the former case, the program will still run correctly, but the block
! 56: will never be reclaimed. Unless the block is intended to be
! 57: permanent, the warning indicates a memory leak.
! 58: <OL>
! 59: <LI>Ignore these warnings while you are using GC_DEBUG. Some of the routines
! 60: mentioned below don't have debugging equivalents. (Alternatively, write
! 61: the missing routines and send them to me.)
! 62: <LI>Replace allocator calls that request large blocks with calls to
! 63: <TT>GC_malloc_ignore_off_page</tt> or
! 64: <TT>GC_malloc_atomic_ignore_off_page</tt>. You may want to set a
! 65: breakpoint in <TT>GC_default_warn_proc</tt> to help you identify such calls.
! 66: Make sure that a pointer to somewhere near the beginning of the resulting block
! 67: is maintained in a (preferably volatile) variable as long as
! 68: the block is needed.
! 69: <LI>
! 70: If the large blocks are allocated with realloc, we suggest instead allocating
! 71: them with something like the following. Note that the realloc size increment
! 72: should be fairly large (e.g. a factor of 3/2) for this to exhibit reasonable
! 73: performance. But we all know we should do that anyway.
! 74: <PRE>
! 75: void * big_realloc(void *p, size_t new_size)
! 76: {
! 77: size_t old_size = GC_size(p);
! 78: void * result;
! 79:
! 80: if (new_size <= 10000) return(GC_realloc(p, new_size));
! 81: if (new_size <= old_size) return(p);
! 82: result = GC_malloc_ignore_off_page(new_size);
! 83: if (result == 0) return(0);
! 84: memcpy(result,p,old_size);
! 85: GC_free(p);
! 86: return(result);
! 87: }
! 88: </pre>
! 89:
! 90: <LI> In the unlikely case that even relatively small object
! 91: (<20KB) allocations are triggering these warnings, then your address
! 92: space contains lots of "bogus pointers", i.e. values that appear to
! 93: be pointers but aren't. Usually this can be solved by using GC_malloc_atomic
! 94: 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
! 95: in certain values. It is possible, to identify the source of the bogus
! 96: pointers by building the collector with <TT>-DPRINT_BLACK_LIST</tt>,
! 97: which will cause it to print the "bogus pointers", along with their location.
! 98:
! 99: <LI> If you get only a fixed number of these warnings, you are probably only
! 100: introducing a bounded leak by ignoring them. If the data structures being
! 101: allocated are intended to be permanent, then it is also safe to ignore them.
! 102: The warnings can be turned off by calling GC_set_warn_proc with a procedure
! 103: that ignores these warnings (e.g. by doing absolutely nothing).
! 104: </ol>
! 105:
! 106: <H2>The Collector References a Bad Address in <TT>GC_malloc</tt></h2>
! 107:
! 108: This typically happens while the collector is trying to remove an entry from
! 109: its free list, and the free list pointer is bad because the free list link
! 110: in the last allocated object was bad.
! 111: <P>
! 112: With > 99% probability, you wrote past the end of an allocated object.
! 113: Try setting <TT>GC_DEBUG</tt> before including <TT>gc.h</tt> and
! 114: allocating with <TT>GC_MALLOC</tt>. This will try to detect such
! 115: overwrite errors.
! 116:
! 117: <H2>Unexpectedly Large Heap</h2>
! 118:
! 119: Unexpected heap growth can be due to one of the following:
! 120: <OL>
! 121: <LI> Data structures that are being unintentionally retained. This
! 122: is commonly caused by data structures that are no longer being used,
! 123: but were not cleared, or by caches growing without bounds.
! 124: <LI> Pointer misidentification. The garbage collector is interpreting
! 125: integers or other data as pointers and retaining the "referenced"
! 126: objects.
! 127: <LI> Heap fragmentation. This should never result in unbounded growth,
! 128: but it may account for larger heaps. This is most commonly caused
! 129: by allocation of large objects. On some platforms it can be reduced
! 130: by building with -DUSE_MUNMAP, which will cause the collector to unmap
! 131: memory corresponding to pages that have not been recently used.
! 132: <LI> Per object overhead. This is usually a relatively minor effect, but
! 133: it may be worth considering. If the collector recognizes interior
! 134: pointers, object sizes are increased, so that one-past-the-end pointers
! 135: are correctly recognized. The collector can be configured not to do this
! 136: (<TT>-DDONT_ADD_BYTE_AT_END</tt>).
! 137: <P>
! 138: The collector rounds up object sizes so the result fits well into the
! 139: chunk size (<TT>HBLKSIZE</tt>, normally 4K on 32 bit machines, 8K
! 140: on 64 bit machines) used by the collector. Thus it may be worth avoiding
! 141: objects of size 2K + 1 (or 2K if a byte is being added at the end.)
! 142: </ol>
! 143: The last two cases can often be identified by looking at the output
! 144: of a call to <TT>GC_dump()</tt>. Among other things, it will print the
! 145: list of free heap blocks, and a very brief description of all chunks in
! 146: the heap, the object sizes they correspond to, and how many live objects
! 147: were found in the chunk at the last collection.
! 148: <P>
! 149: Growing data structures can usually be identified by
! 150: <OL>
! 151: <LI> Building the collector with <TT>-DKEEP_BACK_PTRS</tt>,
! 152: <LI> Preferably using debugging allocation (defining <TT>GC_DEBUG</tt>
! 153: before including <TT>gc.h</tt> and allocating with <TT>GC_MALLOC</tt>),
! 154: so that objects will be identified by their allocation site,
! 155: <LI> Running the application long enough so
! 156: that most of the heap is composed of "leaked" memory, and
! 157: <LI> Then calling <TT>GC_generate_random_backtrace()</tt> from backptr.h
! 158: a few times to determine why some randomly sampled objects in the heap are
! 159: being retained.
! 160: </ol>
! 161: <P>
! 162: The same technique can often be used to identify problems with false
! 163: pointers, by noting whether the reference chains printed by
! 164: <TT>GC_generate_random_backtrace()</tt> involve any misidentified pointers.
! 165: An alternate technique is to build the collector with
! 166: <TT>-DPRINT_BLACK_LIST</tt> which will cause it to report values that
! 167: are almost, but not quite, look like heap pointers. It is very likely that
! 168: actual false pointers will come from similar sources.
! 169: <P>
! 170: In the unlikely case that false pointers are an issue, it can usually
! 171: be resolved using one or more of the following techniques:
! 172: <OL>
! 173: <LI> Use <TT>GC_malloc_atomic</tt> for objects containing no pointers.
! 174: This is especially important for large arrays containing compressed data,
! 175: pseudo-random numbers, and the like. It is also likely to improve GC
! 176: performance, perhaps drastically so if the application is paging.
! 177: <LI> If you allocate large objects containing only
! 178: one or two pointers at the beginning, either try the typed allocation
! 179: primitives is <TT>gc_typed.h</tt>, or separate out the pointerfree component.
! 180: <LI> Consider using <TT>GC_malloc_ignore_off_page()</tt>
! 181: to allocate large objects. (See <TT>gc.h</tt> and above for details.
! 182: Large means > 100K in most environments.)
! 183: </ol>
! 184: <H2>Prematurely Reclaimed Objects</h2>
! 185: The usual symptom of this is a segmentation fault, or an obviously overwritten
! 186: value in a heap object. This should, of course, be impossible. In practice,
! 187: it may happen for reasons like the following:
! 188: <OL>
! 189: <LI> The collector did not intercept the creation of threads correctly in
! 190: a multithreaded application, <I>e.g.</i> because the client called
! 191: <TT>pthread_create</tt> without including <TT>gc.h</tt>, which redefines it.
! 192: <LI> The last pointer to an object in the garbage collected heap was stored
! 193: somewhere were the collector couldn't see it, <I>e.g.</i> in an
! 194: object allocated with system <TT>malloc</tt>, in certain types of
! 195: <TT>mmap</tt>ed files,
! 196: or in some data structure visible only to the OS. (On some platforms,
! 197: thread-local storage is one of these.)
! 198: <LI> The last pointer to an object was somehow disguised, <I>e.g.</i> by
! 199: XORing it with another pointer.
! 200: <LI> Incorrect use of <TT>GC_malloc_atomic</tt> or typed allocation.
! 201: <LI> An incorrect <TT>GC_free</tt> call.
! 202: <LI> The client program overwrote an internal garbage collector data structure.
! 203: <LI> A garbage collector bug.
! 204: <LI> (Empirically less likely than any of the above.) A compiler optimization
! 205: that disguised the last pointer.
! 206: </ol>
! 207: The following relatively simple techniques should be tried first to narrow
! 208: down the problem:
! 209: <OL>
! 210: <LI> If you are using the incremental collector try turning it off for
! 211: debugging.
! 212: <LI> If you are using shared libraries, try linking statically. If that works,
! 213: ensure that DYNAMIC_LOADING is defined on your platform.
! 214: <LI> Try to reproduce the problem with fully debuggable unoptimized code.
! 215: This will eliminate the last possibility, as well as making debugging easier.
! 216: <LI> Try replacing any suspect typed allocation and <TT>GC_malloc_atomic</tt>
! 217: calls with calls to <TT>GC_malloc</tt>.
! 218: <LI> Try removing any GC_free calls (<I>e.g.</i> with a suitable
! 219: <TT>#define</tt>).
! 220: <LI> Rebuild the collector with <TT>-DGC_ASSERTIONS</tt>.
! 221: <LI> If the following works on your platform (i.e. if gctest still works
! 222: if you do this), try building the collector with
! 223: <TT>-DREDIRECT_MALLOC=GC_malloc_uncollectable</tt>. This will cause
! 224: the collector to scan memory allocated with malloc.
! 225: </ol>
! 226: If all else fails, you will have to attack this with a debugger.
! 227: Suggested steps:
! 228: <OL>
! 229: <LI> Call <TT>GC_dump()</tt> from the debugger around the time of the failure. Verify
! 230: that the collectors idea of the root set (i.e. static data regions which
! 231: it should scan for pointers) looks plausible. If not, i.e. if it doesn't
! 232: include some static variables, report this as
! 233: a collector bug. Be sure to describe your platform precisely, since this sort
! 234: of problem is nearly always very platform dependent.
! 235: <LI> Especially if the failure is not deterministic, try to isolate it to
! 236: a relatively small test case.
! 237: <LI> Set a break point in <TT>GC_finish_collection</tt>. This is a good
! 238: point to examine what has been marked, i.e. found reachable, by the
! 239: collector.
! 240: <LI> If the failure is deterministic, run the process
! 241: up to the last collection before the failure.
! 242: Note that the variable <TT>GC_gc_no</tt> counts collections and can be used
! 243: to set a conditional breakpoint in the right one. It is incremented just
! 244: before the call to GC_finish_collection.
! 245: If object <TT>p</tt> was prematurely recycled, it may be helpful to
! 246: look at <TT>*GC_find_header(p)</tt> at the failure point.
! 247: The <TT>hb_last_reclaimed</tt> field will identify the collection number
! 248: during which its block was last swept.
! 249: <LI> Verify that the offending object still has its correct contents at
! 250: this point.
! 251: The call <TT>GC_is_marked(p)</tt> from the debugger to verify that the
! 252: object has not been marked, and is about to be reclaimed.
! 253: <LI> Determine a path from a root, i.e. static variable, stack, or
! 254: register variable,
! 255: to the reclaimed object. Call <TT>GC_is_marked(q)</tt> for each object
! 256: <TT>q</tt> along the path, trying to locate the first unmarked object, say
! 257: <TT>r</tt>.
! 258: <LI> If <TT>r</tt> is pointed to by a static root,
! 259: verify that the location
! 260: pointing to it is part of the root set printed by <TT>GC_dump()</tt>. If it
! 261: is on the stack in the main (or only) thread, verify that
! 262: <TT>GC_stackbottom</tt> is set correctly to the base of the stack. If it is
! 263: in another thread stack, check the collector's thread data structure
! 264: (<TT>GC_thread[]</tt> on several platforms) to make sure that stack bounds
! 265: are set correctly.
! 266: <LI> If <TT>r</tt> is pointed to by heap object <TT>s</tt>, check that the
! 267: collector's layout description for <TT>s</tt> is such that the pointer field
! 268: will be scanned. Call <TT>*GC_find_header(s)</tt> to look at the descriptor
! 269: for the heap chunk. The <TT>hb_descr</tt> field specifies the layout
! 270: of objects in that chunk. See gc_mark.h for the meaning of the descriptor.
! 271: (If it's low order 2 bits are zero, then it is just the length of the
! 272: object prefix to be scanned. This form is always used for objects allocated
! 273: with <TT>GC_malloc</tt> or <TT>GC_malloc_atomic</tt>.)
! 274: <LI> If the failure is not deterministic, you may still be able to apply some
! 275: of the above technique at the point of failure. But remember that objects
! 276: allocated since the last collection will not have been marked, even if the
! 277: collector is functioning properly. On some platforms, the collector
! 278: can be configured to save call chains in objects for debugging.
! 279: Enabling this feature will also cause it to save the call stack at the
! 280: point of the last GC in GC_arrays._last_stack.
! 281: <LI> When looking at GC internal data structures remember that a number
! 282: of <TT>GC_</tt><I>xxx</i> variables are really macro defined to
! 283: <TT>GC_arrays._</tt><I>xxx</i>, so that
! 284: the collector can avoid scanning them.
! 285: </ol>
! 286: </body>
! 287: </html>
! 288:
! 289:
! 290:
! 291:
FreeBSD-CVSweb <freebsd-cvsweb@FreeBSD.org>