On Sun, Oct 9, 2011 at 9:55 PM, Andi Kleen <a...@firstfloor.org> wrote:
> From: Andi Kleen <a...@linux.intel.com>
>
> Use the Linux MADV_DONTNEED call to unmap free pages in the garbage
> collector.Then keep the unmapped pages in the free list. This avoid
> excessive memory fragmentation on large LTO bulds, which can lead
> to gcc bumping into the Linux vm_max_map limit per process.
>
> Based on a idea from Jakub.
Shouldn't we prefer still "mapped" pages when allocating? Thus, keep
the freepages list "sorted"?
With the new params to call release_pages less, how does this
interact with using MADV_DONTNEED? The only reason to delay
MADV_DONTNEED is to avoid splitting huge-pages? Which would
mean that we should rather be better at controlling where we allocate
from from the free-list?
Richard.
> gcc/:
>
> 2011-10-08 Andi Kleen <a...@linux.intel.com>
>
> PR other/50636
> * config.in, configure: Regenerate.
> * configure.ac (madvise): Add to AC_CHECK_FUNCS.
> * ggc-page.c (USING_MADVISE): Add.
> (page_entry): Add unmapped field.
> (alloc_page): Check for unmapped pages.
> (release_pages): Add USING_MADVISE branch.
> ---
> gcc/config.in | 6 ++++++
> gcc/configure | 2 +-
> gcc/configure.ac | 2 +-
> gcc/ggc-page.c | 48 +++++++++++++++++++++++++++++++++++++++++++++++-
> 4 files changed, 55 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/config.in b/gcc/config.in
> index f2847d8..e8148b6 100644
> --- a/gcc/config.in
> +++ b/gcc/config.in
> @@ -1276,6 +1276,12 @@
> #endif
>
>
> +/* Define to 1 if you have the `madvise' function. */
> +#ifndef USED_FOR_TARGET
> +#undef HAVE_MADVISE
> +#endif
> +
> +
> /* Define to 1 if you have the <malloc.h> header file. */
> #ifndef USED_FOR_TARGET
> #undef HAVE_MALLOC_H
> diff --git a/gcc/configure b/gcc/configure
> index cb55dda..4a54adf 100755
> --- a/gcc/configure
> +++ b/gcc/configure
> @@ -9001,7 +9001,7 @@ fi
> for ac_func in times clock kill getrlimit setrlimit atoll atoq \
> sysconf strsignal getrusage nl_langinfo \
> gettimeofday mbstowcs wcswidth mmap setlocale \
> - clearerr_unlocked feof_unlocked ferror_unlocked fflush_unlocked
> fgetc_unlocked fgets_unlocked fileno_unlocked fprintf_unlocked
> fputc_unlocked fputs_unlocked fread_unlocked fwrite_unlocked
> getchar_unlocked getc_unlocked putchar_unlocked putc_unlocked
> + clearerr_unlocked feof_unlocked ferror_unlocked fflush_unlocked
> fgetc_unlocked fgets_unlocked fileno_unlocked fprintf_unlocked
> fputc_unlocked fputs_unlocked fread_unlocked fwrite_unlocked
> getchar_unlocked getc_unlocked putchar_unlocked putc_unlocked madvise
> do :
> as_ac_var=`$as_echo "ac_cv_func_$ac_func" | $as_tr_sh`
> ac_fn_c_check_func "$LINENO" "$ac_func" "$as_ac_var"
> diff --git a/gcc/configure.ac b/gcc/configure.ac
> index a7b94e6..357902e 100644
> --- a/gcc/configure.ac
> +++ b/gcc/configure.ac
> @@ -1027,7 +1027,7 @@ define(gcc_UNLOCKED_FUNCS, clearerr_unlocked
> feof_unlocked dnl
> AC_CHECK_FUNCS(times clock kill getrlimit setrlimit atoll atoq \
> sysconf strsignal getrusage nl_langinfo \
> gettimeofday mbstowcs wcswidth mmap setlocale \
> - gcc_UNLOCKED_FUNCS)
> + gcc_UNLOCKED_FUNCS madvise)
>
> if test x$ac_cv_func_mbstowcs = xyes; then
> AC_CACHE_CHECK(whether mbstowcs works, gcc_cv_func_mbstowcs_works,
> diff --git a/gcc/ggc-page.c b/gcc/ggc-page.c
> index 624f029..b0b3b3f 100644
> --- a/gcc/ggc-page.c
> +++ b/gcc/ggc-page.c
> @@ -50,6 +50,10 @@ along with GCC; see the file COPYING3. If not see
> #define USING_MALLOC_PAGE_GROUPS
> #endif
>
> +#if defined(HAVE_MADVISE) && defined(MADV_DONTNEED)
> +# define USING_MADVISE
> +#endif
> +
> /* Strategy:
>
> This garbage-collecting allocator allocates objects on one of a set
> @@ -277,6 +281,9 @@ typedef struct page_entry
> /* The lg of size of objects allocated from this page. */
> unsigned char order;
>
> + /* Unmapped page? */
> + bool unmapped;
> +
> /* A bit vector indicating whether or not objects are in use. The
> Nth bit is one if the Nth object on this page is allocated. This
> array is dynamically sized. */
> @@ -740,6 +747,10 @@ alloc_page (unsigned order)
>
> if (p != NULL)
> {
> + if (p->unmapped)
> + G.bytes_mapped += p->bytes;
> + p->unmapped = false;
> +
> /* Recycle the allocated memory from this page ... */
> *pp = p->next;
> page = p->page;
> @@ -956,7 +967,42 @@ free_page (page_entry *entry)
> static void
> release_pages (void)
> {
> -#ifdef USING_MMAP
> +#ifdef USING_MADVISE
> + page_entry *p, *start_p;
> + char *start;
> + size_t len;
> +
> + for (p = G.free_pages; p; )
> + {
> + if (p->unmapped)
> + {
> + p = p->next;
> + continue;
> + }
> + start = p->page;
> + len = p->bytes;
> + start_p = p;
> + p = p->next;
> + while (p && p->page == start + len)
> + {
> + len += p->bytes;
> + p = p->next;
> + }
> + /* Give the page back to the kernel, but don't free the mapping.
> + This avoids fragmentation in the virtual memory map of the
> + process. Next time we can reuse it by just touching it. */
> + madvise (start, len, MADV_DONTNEED);
> + /* Don't count those pages as mapped to not touch the garbage collector
> + unnecessarily. */
> + G.bytes_mapped -= len;
> + while (start_p != p)
> + {
> + start_p->unmapped = true;
> + start_p = start_p->next;
> + }
> + }
> +#endif
> +#if defined(USING_MMAP) && !defined(USING_MADVISE)
> page_entry *p, *next;
> char *start;
> size_t len;
> --
> 1.7.5.4
>
>
