2019-07-10 07:35:27

by Yang Bin

Subject: [PATCH] Check sk before sendpage

From: " Yang Bin "<[email protected]>

Before xmit,iscsi may disconnect just now.
So must check connection sock NULL or not,or kernel will crash for
accessing NULL pointer.

Signed-off-by: Yang Bin <[email protected]>
---
drivers/scsi/iscsi_tcp.c | 3 +++
1 file changed, 3 insertions(+)

diff --git a/drivers/scsi/iscsi_tcp.c b/drivers/scsi/iscsi_tcp.c
index 7bedbe8..a59c49f 100644
--- a/drivers/scsi/iscsi_tcp.c
+++ b/drivers/scsi/iscsi_tcp.c
@@ -264,6 +264,9 @@ static int iscsi_sw_tcp_xmit_segment(struct iscsi_tcp_conn *tcp_conn,
unsigned int copied = 0;
int r = 0;

+ if (!sk)
+ return -ENOTCONN;
+
while (!iscsi_tcp_segment_done(tcp_conn, segment, 0, r)) {
struct scatterlist *sg;
unsigned int offset, copy;
--
1.8.3.1

2019-07-10 18:38:17

by Lee Duncan

Subject: Re: [PATCH] Check sk before sendpage

On 7/10/19 12:30 AM, Yang Bin wrote:

> From: " Yang Bin "<[email protected]>
>
> Before xmit,iscsi may disconnect just now.
> So must check connection sock NULL or not,or kernel will crash for
> accessing NULL pointer.
>
> Signed-off-by: Yang Bin <[email protected]>
> ---
> drivers/scsi/iscsi_tcp.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/drivers/scsi/iscsi_tcp.c b/drivers/scsi/iscsi_tcp.c
> index 7bedbe8..a59c49f 100644
> --- a/drivers/scsi/iscsi_tcp.c
> +++ b/drivers/scsi/iscsi_tcp.c
> @@ -264,6 +264,9 @@ static int iscsi_sw_tcp_xmit_segment(struct iscsi_tcp_conn *tcp_conn,
> unsigned int copied = 0;
> int r = 0;
>
> + if (!sk)
> + return -ENOTCONN;
> +
> while (!iscsi_tcp_segment_done(tcp_conn, segment, 0, r)) {
> struct scatterlist *sg;
> unsigned int offset, copy;
>

If the socket can be closed right before iscsi_sw_tcp_xmit_segment() is
called, can it be called in the middle of sending segments? (In which
case the check would have to be in the while loop.)

--

Lee Duncan

2019-07-10 18:54:45

by James Bottomley

Subject: Re: [PATCH] Check sk before sendpage

On Wed, 2019-07-10 at 17:47 +0000, Lee Duncan wrote:
> On 7/10/19 12:30 AM, Yang Bin wrote:
>
> > From: " Yang Bin "<[email protected]>
> >
> > Before xmit,iscsi may disconnect just now.
> > So must check connection sock NULL or not,or kernel will crash for
> > accessing NULL pointer.
> >
> > Signed-off-by: Yang Bin <[email protected]>
> > ---
> > drivers/scsi/iscsi_tcp.c | 3 +++
> > 1 file changed, 3 insertions(+)
> >
> > diff --git a/drivers/scsi/iscsi_tcp.c b/drivers/scsi/iscsi_tcp.c
> > index 7bedbe8..a59c49f 100644
> > --- a/drivers/scsi/iscsi_tcp.c
> > +++ b/drivers/scsi/iscsi_tcp.c
> > @@ -264,6 +264,9 @@ static int iscsi_sw_tcp_xmit_segment(struct
> > iscsi_tcp_conn *tcp_conn,
> > unsigned int copied = 0;
> > int r = 0;
> >
> > + if (!sk)
> > + return -ENOTCONN;
> > +
> > while (!iscsi_tcp_segment_done(tcp_conn, segment, 0, r)) {
> > struct scatterlist *sg;
> > unsigned int offset, copy;
> >
>
> If the socket can be closed right before iscsi_sw_tcp_xmit_segment()
> is called, can it be called in the middle of sending segments? (In
> which case the check would have to be in the while loop.)

I think the important point is: is this an actual observed bug or just
a theoretical problem?

The reason for asking is this call is controlled directly by the
ISCSI_UEVENT_DESTROY_CONN event sent by the iscsi daemon. Obviously if
the daemon goes haywire and doesn't shut down the connection before
sending the destroy event, we may get the crash, but I would be
inclined to say fix the daemon.

James