[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20170320155948.pgpp2uhgoppicdl4@deco.navytux.spb.ru>
Date: Mon, 20 Mar 2017 18:59:49 +0300
From: Kirill Smelkov <kirr@...edi.com>
To: mtk.manpages@...il.com
Cc: linux-man@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Why mmap(MAP_POPULATE | MAP_NONBLOCK) is needed (Re: [patch]
mmap.2: Add link to commit which broke MAP_POPULATE | MAP_NONBLOCK to be
noop)
On Sat, Mar 18, 2017 at 10:40:10PM +0300, Kirill Smelkov wrote:
> Signed-off-by: Kirill Smelkov <kirr@...edi.com>
> ---
> man2/mmap.2 | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/man2/mmap.2 b/man2/mmap.2
> index 96875e486..f6fd56523 100644
> --- a/man2/mmap.2
> +++ b/man2/mmap.2
> @@ -300,6 +300,7 @@ Don't perform read-ahead:
> create page tables entries only for pages
> that are already present in RAM.
> Since Linux 2.6.23, this flag causes
> +.\" commit 54cb8821de07f2ffcd28c380ce9b93d5784b40d7
> .BR MAP_POPULATE
> to do nothing.
> One day, the combination of
Please also find below benchmark which explains why
mmap(MAP_POPULATE | MAP_NONBLOCK)
is actually needed.
Thanks,
Kirill
---- 8< ---- (https://lab.nexedi.com/kirr/misc/blob/5a25f4ae/t_sysmmap_c.c)
/* This program benchmarks pagefault time.
*
* Unfortunately as of 2017-Mar-20 for data in pagecache the situation is as
* follows (i7-6600U, Linux 4.9.13):
*
* 1. minor pagefault: ~ 1200ns
* (this program)
*
* 2. read syscall + whole page copy: ~ 215ns
* (https://github.com/golang/go/issues/19563#issuecomment-287423654)
*
* 3. it is not possible to mmap(MAP_POPULATE | MAP_NONBLOCK) (i.e. prefault
* those PTE that are already in pagecache).
* ( http://www.spinics.net/lists/linux-man/msg11420.html,
* https://git.kernel.org/linus/54cb8821de07f2ffcd28c380ce9b93d5784b40d7 )
*
* 4. (Q) I'm not sure a mechanism exists in the kernel to automatically
* subscribe a VMA so that when a page becomes pagecached, associated PTE is
* adjusted so that programs won't need to pay minor pagefault time on
* access.
*
* unless 3 and 4 are solved mmap unfortunately seems to be slower choice
* compared to just pread.
*/
#define _GNU_SOURCE
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>
#include <sys/time.h>
#include <sys/user.h>
#include <sys/mman.h>
// 12345678
#define NITER 500000
// microtime returns current time as double
double microtime() {
int err;
struct timeval tv;
err = gettimeofday(&tv, NULL);
if (err == -1) {
perror("gettimeofday");
abort();
}
return tv.tv_sec + 1E-6 * tv.tv_usec;
}
int main() {
unsigned char *addr, sum = 0;
int fd, err, i;
size_t size;
double Tstart, Tend;
fd = open("/dev/shm/y.dat", O_RDWR | O_CREAT | O_TRUNC, 0666);
if (fd == -1) {
perror("open");
abort();
}
size = NITER * PAGE_SIZE;
err = ftruncate(fd, size);
if (err == -1) {
perror("ftruncate");
abort();
}
#if 1
// make sure RAM is actually allocated
Tstart = microtime();
err = fallocate(fd, /*mode*/0, 0, size);
Tend = microtime();
if (err == -1) {
perror("fallocate");
abort();
}
printf("T(fallocate):\t%.1f\t%6.1f ns / page\n", Tend - Tstart, (Tend - Tstart) * 1E9 / NITER);
#endif
Tstart = microtime();
addr = mmap(NULL, size, PROT_READ, MAP_SHARED, fd, 0);
//addr = mmap(NULL, size, PROT_READ, MAP_SHARED | MAP_POPULATE, fd, 0);
//addr = mmap(NULL, size, PROT_READ, MAP_SHARED | MAP_POPULATE | MAP_NONBLOCK, fd, 0);
if (addr == MAP_FAILED) {
perror("mmap");
abort();
}
Tend = microtime();
printf("T(mmap):\t%.1f\t%6.1f ns / page\n", Tend - Tstart, (Tend - Tstart) * 1E9 / NITER);
Tstart = microtime();
//for (int j=0; j < 100; j++)
for (i=0; i<NITER; i++) {
sum += addr[i*PAGE_SIZE];
}
Tend = microtime();
printf("T(pagefault):\t%.1f\t%6.1f ns / page\t(%i)\n", Tend - Tstart, (Tend - Tstart) * 1E9 / NITER, sum);
return 0;
}
---- 8< ----
Powered by blists - more mailing lists