Fix crash in GC_scratch_alloc under low-memory conditions

[wingo]

• headers.c (GC_scratch_alloc): Always round bytes_to_get up to a
  multiple of the page size, even if USE_MMAP is not defined. This is
  necessary because GC_unix_get_mem, which is one of the possible
  GET_MEM implementations, may use mmap even if USE_MMAP isn't defined.
  See os_dep.c for a definition of GC_unix_get_mem which tries sbrk
  first, but then falls back to mmap. Calling GC_unix_mmap_get_mem will
  abort if the allocation is not a multiple of the page size.

Here's an example stack trace of the error without this patch:

GC Warning: Out of memory - trying to allocate less
Bad GET_MEM arg

Program received signal SIGABRT, Aborted.
0x00007ffff6fd83a9 in __GI_raise (sig=sig@entry=6) at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
56 ../nptl/sysdeps/unix/sysv/linux/raise.c: No such file or directory.
(gdb) bt
#0 0x00007ffff6fd83a9 in __GI_raise (sig=sig@entry=6) at ../nptl/sysdeps/unix/sysv/linux/raise.c:56
#1 0x00007ffff6fdb4c8 in __GI_abort () at abort.c:89
#2 0x00007ffff6d50009 in GC_unix_mmap_get_mem (bytes=bytes@entry=336) at os_dep.c:2065
#3 0x00007ffff6d500e0 in GC_unix_get_mem (bytes=bytes@entry=336) at os_dep.c:2152
#4 0x00007ffff6d483ef in GC_scratch_alloc (bytes=336, bytes@entry=328) at headers.c:154
#5 0x00007ffff6d4863f in alloc_hdr () at headers.c:173
#6 GC_install_header (h=h@entry=0x7ffff2413000) at headers.c:266
#7 0x00007ffff6d41ae0 in GC_get_first_part (h=h@entry=0x7ffff2412000, hhdr=hhdr@entry=0x7ffff1dbcea0, bytes=bytes@entry=4096, index=index@entry=60)

at allchblk.c:500

#8 0x00007ffff6d41e0a in GC_allochblk_nth (sz=sz@entry=16, kind=kind@entry=1, flags=flags@entry=0, n=n@entry=60, may_split=may_split@entry=1)

at allchblk.c:783

#9 0x00007ffff6d421c6 in GC_allochblk (sz=sz@entry=16, kind=kind@entry=1, flags=flags@entry=0) at allchblk.c:624
#10 0x00007ffff6d49d85 in GC_generic_malloc_many (lb=16, k=1, result=0x68ff00) at mallocx.c:404

[ivmai]

Thanks. I discovered this bug 2 days ago and decided to develop a fix on Sunday but you did it faster.

But I think your solution should be improved - GC_unix_mmap_get_mem is defined only if MMAP_SUPPORTED (thus there is no need for alignment otherwise).

Plus, there are other uses of GET_MEM (in backgraph and misc.c) - they should be handled in the similar way, I suppose.

[ivmai]

I did the fix in other (more general) way - see commit 180b56c (it also fixes absence of rounding in backgraph.c). Please test master branch. After that, I'll cherry-pick this commit to release-7_2/7_4 branches.

[wingo]

Makes sense to me. The test case was hard for me to hit, so I can't reproduce right now. That said, I have had reports of this happening to other people; I'll be sure to report again if something goes wrong.

Thanks for the fix!