From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <vdavydov.dev@gmail.com>
Date: Tue, 16 Apr 2019 20:35:10 +0300
From: Vladimir Davydov <vdavydov.dev@gmail.com>
Subject: Re: [PATCH v2] core/coio_file: Use eio_sendfile_sync instead of a
 chunk mode
Message-ID: <20190416173510.4nxxdsqg7drrlnhe@esperanza>
References: <20190415211410.20562-1-gorcunov@gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20190415211410.20562-1-gorcunov@gmail.com>
To: Cyrill Gorcunov <gorcunov@gmail.com>
Cc: tml <tarantool-patches@freelists.org>
List-ID: <tarantool-patches.dev.tarantool.org>

On Tue, Apr 16, 2019 at 12:14:10AM +0300, Cyrill Gorcunov wrote:
> eio library provides a portable version of sendfile syscall
> which works more efficient than explicit copying fileby 4K
> chunks.
> 
> Signed-off-by: Cyrill Gorcunov <gorcunov@gmail.com>
> ---
> v2:
>  - Use sendfile in a cycle to address files more than 2G size
>  - Proper testing of large files remains opened: current CI engine
>    is hardly capable of managing it. I tested 200M files with manual
>    splitting (to make sure the offsets do really work) but for longterm
>    we still might need to invent something

We could use an error injection to reduce the size of data fed to
sendfile in one go.

>  - Another question which remains -- what to do with partially copied
>    files, neither code before the patch or after do not clean up parts
>    of copied data. Should not we clean it up on error path? If yes then
>    I'll prepare another patch on top.

Nobody complained => I don't think we need to do anything about it.

> 
>  src/lib/core/coio_file.c | 26 +++++++++++---------------
>  1 file changed, 11 insertions(+), 15 deletions(-)
> 
> diff --git a/src/lib/core/coio_file.c b/src/lib/core/coio_file.c
> index c5b2db781..541e0f05a 100644
> --- a/src/lib/core/coio_file.c
> +++ b/src/lib/core/coio_file.c
> @@ -571,7 +571,7 @@ static void
>  coio_do_copyfile(eio_req *req)
>  {
>  	struct coio_file_task *eio = (struct coio_file_task *)req->data;
> -
> +	ssize_t pos, ret, left;
>  	struct stat st;
>  	if (stat(eio->copyfile.source, &st) < 0) {
>  		goto error;
> @@ -588,22 +588,18 @@ coio_do_copyfile(eio_req *req)
>  		goto error_dest;
>  	}
>  
> -	enum { COPY_FILE_BUF_SIZE = 4096 };
> -
> -	char buf[COPY_FILE_BUF_SIZE];
> -
> -	while (true) {
> -		ssize_t nread = fio_read(source_fd, buf, sizeof(buf));
> -		if (nread < 0)
> -			goto error_copy;
> -
> -		if (nread == 0)
> -			break; /* eof */
> -
> -		ssize_t nwritten = fio_writen(dest_fd, buf, nread);
> -		if (nwritten < 0)
> +	for (left = st.st_size, pos = 0; left > 0;) {

st.st_size has time off_t while left has type ssize_t. They differ on
32-bit machines AFAIR.

> +		ret = eio_sendfile_sync(dest_fd, source_fd, pos, left);
> +		if (ret < 0) {
> +			say_syserror("sendfile, [%s -> %s]",
> +				     fio_filename(source_fd),
> +				     fio_filename(dest_fd));
>  			goto error_copy;
> +		}
> +		pos += ret;
> +		left -= ret;
>  	}
> +
>  	req->result = 0;
>  	close(source_fd);
>  	close(dest_fd);