2013-04-12 20:45:00

by Steve Magnani

Subject: [PATCH] sd: fix infinite kernel/udev loop on non-removable Medium Not Present

Commit eface65c336eff420d70beb0fb6787a732e05ffb (2.6.38) altered
set_media_not_present() in a way that prevents the sd driver from
remembering that a non-removable device has reported "Medium Not Present".
This condition can occur on hotplug of a (i.e.) USB Mass Storage device
whose medium is offline due to an unrecoverable controller error,
but which is otherwise capable of SCSI communication (to download new
microcode, etc.).

Under these conditions, the changed code results in an infinite loop
between the kernel and udevd. When udevd attempts to open the device
in response to a change notification, a SCSI "Medium Not Present" error
occurs which causes the kernel to signal another change. The cycle
repeats until the device is unplugged, resulting in udevd consuming ever-
increasing amounts of CPU and virtual memory.

Resolve this by remembering "media not present" whether the device has
declared itself "removable" or not.

Signed-off-by: Steven J. Magnani <[email protected]>
---
--- a/drivers/scsi/sd.c 2013-04-12 14:16:12.252531097 -0500
+++ b/drivers/scsi/sd.c 2013-04-12 14:21:55.197216521 -0500
@@ -1298,10 +1298,8 @@ out:

static void set_media_not_present(struct scsi_disk *sdkp)
{
- if (sdkp->media_present)
+ if (sdkp->media_present) {
sdkp->device->changed = 1;
-
- if (sdkp->device->removable) {
sdkp->media_present = 0;
sdkp->capacity = 0;
}

2013-04-12 20:53:30

by James Bottomley

Subject: Re: [PATCH] sd: fix infinite kernel/udev loop on non-removable Medium Not Present

Could you resend this, cc'ing the [email protected] mailing
list?

Thanks,

James


On Fri, 2013-04-12 at 15:44 -0500, Steve Magnani wrote:
> Commit eface65c336eff420d70beb0fb6787a732e05ffb (2.6.38) altered
> set_media_not_present() in a way that prevents the sd driver from
> remembering that a non-removable device has reported "Medium Not Present".
> This condition can occur on hotplug of a (i.e.) USB Mass Storage device
> whose medium is offline due to an unrecoverable controller error,
> but which is otherwise capable of SCSI communication (to download new
> microcode, etc.).
>
> Under these conditions, the changed code results in an infinite loop
> between the kernel and udevd. When udevd attempts to open the device
> in response to a change notification, a SCSI "Medium Not Present" error
> occurs which causes the kernel to signal another change. The cycle
> repeats until the device is unplugged, resulting in udevd consuming ever-
> increasing amounts of CPU and virtual memory.
>
> Resolve this by remembering "media not present" whether the device has
> declared itself "removable" or not.
>
> Signed-off-by: Steven J. Magnani <[email protected]>
> ---
> --- a/drivers/scsi/sd.c 2013-04-12 14:16:12.252531097 -0500
> +++ b/drivers/scsi/sd.c 2013-04-12 14:21:55.197216521 -0500
> @@ -1298,10 +1298,8 @@ out:
>
> static void set_media_not_present(struct scsi_disk *sdkp)
> {
> - if (sdkp->media_present)
> + if (sdkp->media_present) {
> sdkp->device->changed = 1;
> -
> - if (sdkp->device->removable) {
> sdkp->media_present = 0;
> sdkp->capacity = 0;
> }
>