NetBSD Problem Report #45441
From www@NetBSD.org Fri Oct 7 22:53:44 2011
Return-Path: <www@NetBSD.org>
Received: from mail.netbsd.org (mail.netbsd.org [204.152.190.11])
by www.NetBSD.org (Postfix) with ESMTP id A011F63D4DF
for <gnats-bugs@gnats.NetBSD.org>; Fri, 7 Oct 2011 22:53:44 +0000 (UTC)
Message-Id: <20111007225343.E288563D4AC@www.NetBSD.org>
Date: Fri, 7 Oct 2011 22:53:43 +0000 (UTC)
From: alnsn@NetBSD.org
Reply-To: alnsn@NetBSD.org
To: gnats-bugs@NetBSD.org
Subject: aio_write and aio_read connected to one pipe don't make progress
X-Send-Pr-Version: www-1.0
>Number: 45441
>Category: kern
>Synopsis: aio_write and aio_read connected to one pipe don't make progress
>Confidential: no
>Severity: non-critical
>Priority: low
>Responsible: kern-bug-people
>State: open
>Class: sw-bug
>Submitter-Id: net
>Arrival-Date: Fri Oct 07 22:55:00 +0000 2011
>Last-Modified: Tue Nov 01 07:55:01 +0000 2011
>Originator: Alexander Nasonov
>Release: NetBSD 5.99.56
>Organization:
home sweet home
>Environment:
NetBSD nebeda.localdomain 5.99.56 NetBSD 5.99.56 (GENERIC) #0: Thu Oct 6 13:58:32 UTC 2011 builds@b6.netbsd.org:/home/builds/ab/HEAD/amd64/201110061150Z-obj/home/builds/ab/HEAD/src/sys/arch/amd64/compile/GENERIC amd64
>Description:
If you initialize two aiocb structs to connect to one pipe and call aio_read before aio_write, then aio_error() calls will always be returning EINPROGRESS.
>How-To-Repeat:
#include <err.h>
#include <errno.h>
#include <limits.h>
#include <stdlib.h>
#include <string.h>
#include <aio.h>
#include <unistd.h>
static void
init_pipe_data(int fd[], struct aiocb cb[], char *buf[], size_t buf_sz)
{
int i;
if (pipe(fd) != 0)
err(EXIT_FAILURE, "pipe");
for (i = 0; i < 2; i++) {
buf[i] = malloc(buf_sz);
if (buf[i] == NULL)
err(EXIT_FAILURE, "malloc");
buf[i][0] = i;
buf[i][buf_sz-1] = CHAR_MAX - i;
memset(&cb[i], 0, sizeof(cb[i]));
cb[i].aio_buf = buf[i];
cb[i].aio_fildes = fd[i];
cb[i].aio_offset = 0;
cb[i].aio_nbytes = buf_sz;
}
}
static void
destroy_pipe_data(int fd[], char *buf[])
{
int i;
for (i = 0; i < 2; i++) {
free(buf[i]);
close(fd[i]);
}
}
int main()
{
const size_t sz = PIPE_BUF + 13;
int res[2] = { EINPROGRESS, EINPROGRESS };
int fd[2];
struct aiocb cb[2];
char *buf[2];
init_pipe_data(fd, cb, buf, sz);
if (aio_read(&cb[0]) != 0)
err(EXIT_FAILURE, "aio_read");
if (aio_write(&cb[1]) != 0)
err(EXIT_FAILURE, "aio_write");
do {
int i;
for (i = 0; i < 2; i++)
if (res[i] == EINPROGRESS)
res[i] = aio_error(&cb[i]);
} while (res[0] == EINPROGRESS || res[1] == EINPROGRESS);
destroy_pipe_data(fd, buf);
return EXIT_SUCCESS;
}
>Fix:
not known
>Audit-Trail:
From: Alexander Nasonov <alnsn@yandex.ru>
To: gnats-bugs@NetBSD.org
Cc: kern-bug-people@netbsd.org, gnats-admin@netbsd.org,
netbsd-bugs@netbsd.org
Subject: Re: kern/45441: aio_write and aio_read connected to one pipe don't
make progress
Date: Wed, 12 Oct 2011 21:26:40 +0100
alnsn@NetBSD.org wrote:
> If you initialize two aiocb structs to connect to one pipe and call aio_read before aio_write, then aio_error() calls will always be returning EINPROGRESS.
...
>Fix:
> not known
The best fix, I think, is return ESPIPE for non-seekable fildes rather
than silently ignoring aio_offset. This would indirectly solve the issue
described in this PR.
Any objections if I commit the patch below? (I'll document ESPIPE as well)
Alex
Index: sys/kern/sys_aio.c
===================================================================
RCS file: /cvsroot/src/sys/kern/sys_aio.c,v
retrieving revision 1.37
diff -p -u -u -r1.37 sys_aio.c
--- sys/kern/sys_aio.c 17 Feb 2011 19:02:50 -0000 1.37
+++ sys/kern/sys_aio.c 12 Oct 2011 20:15:34 -0000
@@ -360,6 +360,7 @@ aio_process(struct aio_job *a_job)
struct proc *p = curlwp->l_proc;
struct aiocb *aiocbp = &a_job->aiocbp;
struct file *fp;
+ struct vnode *vp;
int fd = aiocbp->aio_fildes;
int error = 0;
@@ -380,6 +381,13 @@ aio_process(struct aio_job *a_job)
goto done;
}
+ vp = (struct vnode *)fp->f_data;
+ if (fp->f_type != DTYPE_VNODE || vp->v_type == VFIFO) {
+ fd_putfile(fd);
+ error = ESPIPE;
+ goto done;
+ }
+
aiov.iov_base = (void *)(uintptr_t)aiocbp->aio_buf;
aiov.iov_len = aiocbp->aio_nbytes;
auio.uio_iov = &aiov;
@@ -427,7 +435,6 @@ aio_process(struct aio_job *a_job)
/*
* Perform a file Sync operation
*/
- struct vnode *vp;
if ((error = fd_getvnode(fd, &fp)) != 0)
goto done;
From: David Laight <david@l8s.co.uk>
To: Alexander Nasonov <alnsn@yandex.ru>
Cc: gnats-bugs@NetBSD.org
Subject: Re: kern/45441: aio_write and aio_read connected to one pipe don't make progress
Date: Thu, 13 Oct 2011 21:47:30 +0100
On Wed, Oct 12, 2011 at 09:26:40PM +0100, Alexander Nasonov wrote:
> alnsn@NetBSD.org wrote:
> > If you initialize two aiocb structs to connect to one pipe and call aio_read before aio_write, then aio_error() calls will always be returning EINPROGRESS.
> ...
> >Fix:
> > not known
>
> The best fix, I think, is return ESPIPE for non-seekable fildes rather
> than silently ignoring aio_offset. This would indirectly solve the issue
> described in this PR.
IIRC that breaks historic practise.
The likely fubar here is that 'aio' is synchronous on a single kernel
thread.
David
--
David Laight: david@l8s.co.uk
From: Alexander Nasonov <alnsn@yandex.ru>
To: gnats-bugs@NetBSD.org
Cc: kern-bug-people@netbsd.org, gnats-admin@netbsd.org,
netbsd-bugs@netbsd.org, alnsn@NetBSD.org
Subject: Re: kern/45441: aio_write and aio_read connected to one pipe don't
make progress
Date: Thu, 13 Oct 2011 22:44:43 +0100
David Laight wrote:
> On Wed, Oct 12, 2011 at 09:26:40PM +0100, Alexander Nasonov wrote:
> > The best fix, I think, is return ESPIPE for non-seekable fildes rather
> > than silently ignoring aio_offset. This would indirectly solve the issue
> > described in this PR.
>
> IIRC that breaks historic practise.
Do you have any reference or any more details on this?
> The likely fubar here is that 'aio' is synchronous on a single kernel
> thread.
Yes, I know. I wrote the test to prove that the current implementation
has problems.
Having separate writer and reader kernel threads would solve the
problem but it's quite a lot of work.
Alex
From: Marc Balmer <mbalmer@NetBSD.org>
To: Alexander Nasonov <alnsn@yandex.ru>
Cc: gnats-bugs@NetBSD.org, kern-bug-people@NetBSD.org, gnats-admin@NetBSD.org,
netbsd-bugs@NetBSD.org
Subject: Re: kern/45441: aio_write and aio_read connected to one pipe don't
make progress
Date: Tue, 01 Nov 2011 08:53:16 +0100
Am 12.10.11 22:26, schrieb Alexander Nasonov:
> alnsn@NetBSD.org wrote:
>> If you initialize two aiocb structs to connect to one pipe and call aio_read before aio_write, then aio_error() calls will always be returning EINPROGRESS.
> ...
>> Fix:
>> not known
>
> The best fix, I think, is return ESPIPE for non-seekable fildes rather
> than silently ignoring aio_offset. This would indirectly solve the issue
> described in this PR.
IMO, this is just a hack. aio ist fundamentally broken, or rather not
working at the moment due to it's design. There is another aio PR open,
see there for details.
>
> Any objections if I commit the patch below? (I'll document ESPIPE as well)
> Alex
>
>
> Index: sys/kern/sys_aio.c
> ===================================================================
> RCS file: /cvsroot/src/sys/kern/sys_aio.c,v
> retrieving revision 1.37
> diff -p -u -u -r1.37 sys_aio.c
> --- sys/kern/sys_aio.c 17 Feb 2011 19:02:50 -0000 1.37
> +++ sys/kern/sys_aio.c 12 Oct 2011 20:15:34 -0000
> @@ -360,6 +360,7 @@ aio_process(struct aio_job *a_job)
> struct proc *p = curlwp->l_proc;
> struct aiocb *aiocbp = &a_job->aiocbp;
> struct file *fp;
> + struct vnode *vp;
> int fd = aiocbp->aio_fildes;
> int error = 0;
>
> @@ -380,6 +381,13 @@ aio_process(struct aio_job *a_job)
> goto done;
> }
>
> + vp = (struct vnode *)fp->f_data;
> + if (fp->f_type != DTYPE_VNODE || vp->v_type == VFIFO) {
> + fd_putfile(fd);
> + error = ESPIPE;
> + goto done;
> + }
> +
> aiov.iov_base = (void *)(uintptr_t)aiocbp->aio_buf;
> aiov.iov_len = aiocbp->aio_nbytes;
> auio.uio_iov = &aiov;
> @@ -427,7 +435,6 @@ aio_process(struct aio_job *a_job)
> /*
> * Perform a file Sync operation
> */
> - struct vnode *vp;
>
> if ((error = fd_getvnode(fd, &fp)) != 0)
> goto done;
--
\~~~~~. The NetBSD Foundation
\~~~~~' Marc Balmer, Developer / Marketing
NetBSD
\ mbalmer@NetBSD.org http://www.NetBSD.org/
(Contact us)
$NetBSD: query-full-pr,v 1.39 2013/11/01 18:47:49 spz Exp $
$NetBSD: gnats_config.sh,v 1.8 2006/05/07 09:23:38 tsutsui Exp $
Copyright © 1994-2007
The NetBSD Foundation, Inc. ALL RIGHTS RESERVED.
Home