[linux-next] splice call weird results

[linux-next] splice call weird results

[Cyrill Gorcunov]

Hi! While been trying to run criu on linux-next (due to recent Kirill's patch related
to /proc/pid/clear_refs with THP enabled) I noticed that it fails dumping programs when
moves data from memory pages into an image file. So I wrote pretty idiotic test and
run it on current fedora 20 kernel and then on linux-next.

fedora-20 output
----------------
[cyrill@moon criu] ~/pipe 
Opened pipe-test.SQBcJa
vmspliced 8192 bytes
spliced 4096 bytes

linux-next
----------
[root@fc criu]# ~/pipe
Opened pipe-test.9nZSW7
vmspliced 8192 bytes
spliced 8192 bytes

In test I fill pipe with 8K data then splice 4K of it into a file (the test
is below). Is it intended, or I did some silly mistake?
---
#define _GNU_SOURCE

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <fcntl.h>

#include <sys/uio.h>

#define ARRAY_SIZE(x)	(sizeof(x)/sizeof(x[0]))

static char buf[8 << 20];

int main(int argc, char *argv[])
{
	char *tmpname, t[64] = "pipe-test.XXXXXX";
	int _pipe[2], fd, ret = 0;

	struct iovec iov[] = {
		[0] = {
			.iov_base	= &buf[0],
			.iov_len	= 4096,
		},

		[1] = {
			.iov_base	= &buf[4096],
			.iov_len	= 4096,
		},
	};

	if (pipe(_pipe)) {
		perror("Can't create pipe");
		exit(1);
	}

	memset(buf, 0x1, sizeof(buf));
	tmpname = mktemp(t);

	fd = open(tmpname, O_RDWR | O_CREAT | O_TRUNC);
	if (fd < 0) {
		perror("Can't open temp file");
		close(_pipe[0]);
		close(_pipe[1]);
		exit(1);
	}
	printf("Opened %s\n", tmpname);

	ret = vmsplice(_pipe[1], iov, ARRAY_SIZE(iov), SPLICE_F_GIFT);
	printf("vmspliced %li bytes\n", (long)ret);

	ret = splice(_pipe[0], NULL, fd, NULL, iov[0].iov_len, SPLICE_F_MOVE);
	printf("spliced %li bytes\n", (long)ret);

out:
	close(_pipe[0]);
	close(_pipe[1]);
	close(fd);
	unlink(tmpname);
	return ret;
}

RE: [linux-next] splice call weird results

[Kirill A. Shutemov]

Cyrill Gorcunov wrote:
> Hi! While been trying to run criu on linux-next (due to recent Kirill's patch related
> to /proc/pid/clear_refs with THP enabled) I noticed that it fails dumping programs when
> moves data from memory pages into an image file. So I wrote pretty idiotic test and
> run it on current fedora 20 kernel and then on linux-next.
> 

+Al. He reworked splice code in linux-next.

> fedora-20 output
> ----------------
> [cyrill@moon criu] ~/pipe 
> Opened pipe-test.SQBcJa
> vmspliced 8192 bytes
> spliced 4096 bytes
> 
> linux-next
> ----------
> [root@fc criu]# ~/pipe
> Opened pipe-test.9nZSW7
> vmspliced 8192 bytes
> spliced 8192 bytes
> 
> In test I fill pipe with 8K data then splice 4K of it into a file (the test
> is below). Is it intended, or I did some silly mistake?
> ---
> #define _GNU_SOURCE
> 
> #include <stdio.h>
> #include <stdlib.h>
> #include <unistd.h>
> #include <string.h>
> #include <fcntl.h>
> 
> #include <sys/uio.h>
> 
> #define ARRAY_SIZE(x)	(sizeof(x)/sizeof(x[0]))
> 
> static char buf[8 << 20];
> 
> int main(int argc, char *argv[])
> {
> 	char *tmpname, t[64] = "pipe-test.XXXXXX";
> 	int _pipe[2], fd, ret = 0;
> 
> 	struct iovec iov[] = {
> 		[0] = {
> 			.iov_base	= &buf[0],
> 			.iov_len	= 4096,
> 		},
> 
> 		[1] = {
> 			.iov_base	= &buf[4096],
> 			.iov_len	= 4096,
> 		},
> 	};
> 
> 	if (pipe(_pipe)) {
> 		perror("Can't create pipe");
> 		exit(1);
> 	}
> 
> 	memset(buf, 0x1, sizeof(buf));
> 	tmpname = mktemp(t);
> 
> 	fd = open(tmpname, O_RDWR | O_CREAT | O_TRUNC);
> 	if (fd < 0) {
> 		perror("Can't open temp file");
> 		close(_pipe[0]);
> 		close(_pipe[1]);
> 		exit(1);
> 	}
> 	printf("Opened %s\n", tmpname);
> 
> 	ret = vmsplice(_pipe[1], iov, ARRAY_SIZE(iov), SPLICE_F_GIFT);
> 	printf("vmspliced %li bytes\n", (long)ret);
> 
> 	ret = splice(_pipe[0], NULL, fd, NULL, iov[0].iov_len, SPLICE_F_MOVE);
> 	printf("spliced %li bytes\n", (long)ret);
> 
> out:
> 	close(_pipe[0]);
> 	close(_pipe[1]);
> 	close(fd);
> 	unlink(tmpname);
> 	return ret;
> }

-- 
 Kirill A. Shutemov

Re: [linux-next] splice call weird results

[Cyrill Gorcunov]

On Tue, May 27, 2014 at 01:20:13PM +0300, Kirill A. Shutemov wrote:
> Cyrill Gorcunov wrote:
> > Hi! While been trying to run criu on linux-next (due to recent Kirill's patch related
> > to /proc/pid/clear_refs with THP enabled) I noticed that it fails dumping programs when
> > moves data from memory pages into an image file. So I wrote pretty idiotic test and
> > run it on current fedora 20 kernel and then on linux-next.
> > 
> 
> +Al. He reworked splice code in linux-next.

Thanks Kirill. Al, it looks like iter_file_splice_write no longer honor @len while
building kiocb vector but pushes as much as it can.

Re: [linux-next] splice call weird results

[Cong Wang]

On Tue, May 27, 2014 at 3:27 AM, Cyrill Gorcunov <gorcunov@gmail.com> wrote:
> On Tue, May 27, 2014 at 01:20:13PM +0300, Kirill A. Shutemov wrote:
>> Cyrill Gorcunov wrote:
>> > Hi! While been trying to run criu on linux-next (due to recent Kirill's patch related
>> > to /proc/pid/clear_refs with THP enabled) I noticed that it fails dumping programs when
>> > moves data from memory pages into an image file. So I wrote pretty idiotic test and
>> > run it on current fedora 20 kernel and then on linux-next.
>> >
>>
>> +Al. He reworked splice code in linux-next.
>
> Thanks Kirill. Al, it looks like iter_file_splice_write no longer honor @len while
> building kiocb vector but pushes as much as it can.


I guess this should fix your problem:
https://lkml.org/lkml/2014/5/27/353

Re: [linux-next] splice call weird results

[Cyrill Gorcunov]

On Tue, May 27, 2014 at 10:20:46AM -0700, Cong Wang wrote:
> On Tue, May 27, 2014 at 3:27 AM, Cyrill Gorcunov <gorcunov@gmail.com> wrote:
> > On Tue, May 27, 2014 at 01:20:13PM +0300, Kirill A. Shutemov wrote:
> >> Cyrill Gorcunov wrote:
> >> > Hi! While been trying to run criu on linux-next (due to recent Kirill's patch related
> >> > to /proc/pid/clear_refs with THP enabled) I noticed that it fails dumping programs when
> >> > moves data from memory pages into an image file. So I wrote pretty idiotic test and
> >> > run it on current fedora 20 kernel and then on linux-next.
> >> >
> >>
> >> +Al. He reworked splice code in linux-next.
> >
> > Thanks Kirill. Al, it looks like iter_file_splice_write no longer honor @len while
> > building kiocb vector but pushes as much as it can.
> 
> 
> I guess this should fix your problem:
> https://lkml.org/lkml/2014/5/27/353

Thanks, but i doubt, the problem I notice is rather in iter_file_splice_write.