Zsh Mailing List Archive
Messages sorted by: Reverse Date, Date, Thread, Author

heap memory issues

Line numbers are based upon a clean 3.1.5.  All this is Src/mem.c.
None of it delves into the ZSH_MEM allocators -- it's all heap stuff.

The debugging array h_m is declared to be:
  static int h_m[1025];
It seems to be used for tracing the number of allocations of a given
size; however, with an H_ISIZE == sizeof(long), [assuming 4] then
scanning mem.c, it's 3/4 unused.  The first 1024/H_ISIZE entries and
1024 can be filled, the rest will never change from 0.

Also, for 64-bit platforms, might alignment constraints require H_ISIZE
to be sizeof(long long), or can all of them use 32-bit alignment?  The
halloc() and hrealloc() functions seem to rely upon H_ISIZE being a
measure of sufficient alignment for all data-types.

Take a deap breath for the logic here ...

Whilst I haven't tested this, it looks upon reading as though there are
some problems if hrealloc()ing a memory-block that is larger than
HEAP_ARENA_SIZE.  This would only manifest itself on an allocation
larger than about 8176 bytes (varying according to compiler whim).
These are probably sufficiently rare to not crop up too often.
In halloc() line 269, we have:
  n = HEAP_ARENA_SIZE > size ? HEAPSIZE : size + sizeof(*h);
Then in hrealloc() line 335:
  zfree(h, HEAPSIZE);
from a hrealloc(startofhunk, oldverybig, 0) -- the /next/ block checks
for big sizes, just too late.  Changing this to check for
(h->used > HEAP_ARENA_SIZE) first would, as it stands, work FOR THIS
case.  If a previous hrealloc() has gone from greater than
HEAP_ARENA_SIZE to less than HEAP_ARENA_SIZE then a new arena will have
been allocated (lines 338-345), so h->used can't be less than
HEAP_ARENA_SIZE for a big hunk, *BUT* there is the problem that the new
arena thus allocated will be smaller than HEAP_ARENA_SIZE and the
zfree(h, HEAPSIZE) in line 269 now frees too much memory.

Um... just re-read the code if that's confusing.

Hrm, lines 314-316 are:
    DPUTS(!h, "BUG: hrealloc() called for non-heap memory.");
    DPUTS(h->sp && arena(h) + h->sp->used > p,
          "BUG: hrealloc() wants to realloc pushed memory");
If the first DPUTS prints (ie, non-heap memory) then won't this
dereference a null-pointer?  At least dputs() fflush()s stderr, so we
see the error message before getting the core dump.  :^)

Lines 318-325 handle reallocating from the middle of an arena  --
should this do a memset() to 0xff if ZSH_MEM_DEBUG, or is that only for
memory ranges that extend to the end of an arena?  If so, how about a
memset() to 0xfe or something?

If I'm right and these are bugs, someone more comfortable with actually
altering the code might want to write the patches.  But first I want
Bart to rip holes in my logic and work out my mistakes.  :^)

--> Phil Pennock ; GAT d- s+:+ a22 C++(++++) UL++++/I+++/S+++/H+ P++@ L+++
E-@ W(+) N>++ o !K w--- O>+ M V !PS PE Y+ PGP+ t-- 5++ X+ R !tv b++>+++ DI+ D+
G+ e+ h* r y?

Messages sorted by: Reverse Date, Date, Thread, Author