Re: [PATCH] Reduce number of mmap calls from __libc_memalign in ld.so

["H.J. Lu" <hjl.tools@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1337 bytes --]

On Sat, Apr 2, 2016 at 10:43 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Sat, Apr 2, 2016 at 10:33 AM, Mike Frysinger <vapier@gentoo.org> wrote:
>> On 02 Apr 2016 08:34, H.J. Lu wrote:
>>> __libc_memalign in ld.so allocates one page at a time and tries to
>>> optimize consecutive __libc_memalign calls by hoping that the next
>>> mmap is after the current memory allocation.
>>>
>>> However, the kernel hands out mmap addresses in top-down order, so
>>> this optimization in practice never happens, with the result that we
>>> have more mmap calls and waste a bunch of space for each __libc_memalign.
>>>
>>> This change makes __libc_memalign to mmap one page extra.  Worst case,
>>> the kernel never puts a backing page behind it, but best case it allows
>>> __libc_memalign to operate much much better.  For elf/tst-align --direct,
>>> it reduces number of mmap calls from 12 to 9.
>>>
>>> --- a/elf/dl-minimal.c
>>> +++ b/elf/dl-minimal.c
>>> @@ -75,6 +75,7 @@ __libc_memalign (size_t align, size_t n)
>>>           return NULL;
>>>         nup = GLRO(dl_pagesize);
>>>       }
>>> +      nup += GLRO(dl_pagesize);
>>
>> should this be in the else case ?
>>
>> also the comment above this code needs updating
>> -mike
>
> You are right.  Here is the updated patch.
>

We can just always increment number of pages by one.

-- 
H.J.

[-- Attachment #2: 0001-Reduce-number-of-mmap-calls-from-__libc_memalign-in-.patch --]
[-- Type: text/x-patch, Size: 1895 bytes --]

From 4aad224c5dc8c8e8496868cc1bb00d587aa4f1ed Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <hjl.tools@gmail.com>
Date: Sat, 2 Apr 2016 08:25:31 -0700
Subject: [PATCH] Reduce number of mmap calls from __libc_memalign in ld.so

__libc_memalign in ld.so allocates one page at a time and tries to
optimize consecutive __libc_memalign calls by hoping that the next
mmap is after the current memory allocation.

However, the kernel hands out mmap addresses in top-down order, so
this optimization in practice never happens, with the result that we
have more mmap calls and waste a bunch of space for each __libc_memalign.

This change makes __libc_memalign to mmap one page extra.  Worst case,
the kernel never puts a backing page behind it, but best case it allows
__libc_memalign to operate much much better.  For elf/tst-align --direct,
it reduces number of mmap calls from 12 to 9.

	* elf/dl-minimal.c (__libc_memalign): Mmap one extra page.
---
 elf/dl-minimal.c | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/elf/dl-minimal.c b/elf/dl-minimal.c
index 762e65b..0660fcf 100644
--- a/elf/dl-minimal.c
+++ b/elf/dl-minimal.c
@@ -66,15 +66,13 @@ __libc_memalign (size_t align, size_t n)
 
   if (alloc_ptr + n >= alloc_end || n >= -(uintptr_t) alloc_ptr)
     {
-      /* Insufficient space left; allocate another page.  */
+      /* Insufficient space left; allocate another page plus one extra
+	 page to reduce number of mmap calls.  */
       caddr_t page;
       size_t nup = (n + GLRO(dl_pagesize) - 1) & ~(GLRO(dl_pagesize) - 1);
-      if (__glibc_unlikely (nup == 0))
-	{
-	  if (n)
-	    return NULL;
-	  nup = GLRO(dl_pagesize);
-	}
+      if (__glibc_unlikely (nup == 0 && n))
+	return NULL;
+      nup += GLRO(dl_pagesize);
       page = __mmap (0, nup, PROT_READ|PROT_WRITE,
 		     MAP_ANON|MAP_PRIVATE, -1, 0);
       if (page == MAP_FAILED)
-- 
2.5.5