Hi Ben,
On 06/13/2018 05:25 PM, Ben Hutchings wrote:
On Mon, 2018-05-28 at 12:01 +0200, Greg Kroah-Hartman wrote:
4.4-stable review patch. If anyone has any objections, please let me know.
From: Jeremy Cline jeremy@jcline.org
[ Upstream commit 20bd1d026aacc5399464f8328f305985c493cde3 ]
If the read-only flag is true on a SCSI disk, re-reading the partition table sets the flag back to false.
To observe this bug, you can run:
- blockdev --setro /dev/sda
- blockdev --rereadpt /dev/sda
- blockdev --getro /dev/sda
This commit reads the disk's old state and combines it with the device disk-reported state rather than unconditionally marking it as RW.
It seems to me that this change is likely to cause a regression: if a SCSI device switches from read-only to read-write state then a subsequent rescan won't automatically change the block device to read- write state. The administrator will have to use the blockdev command too.
Even if this change in behaviour is acceptable, this commit does not implement it consistently. The function starts by clearing the ro flag and this commit only changes one of the three exit paths to preserve it. (The log message about Write Protect status also reports the underlying SCSI device flag and not the combined ro flag, but maybe that was intentional.)
Yes, it looks like I messed this up.
I think this commit should be reverted, both in stable and upstream. A proper fix would involve splitting the ro flag into two flags—one controlled by user-space and one read from the device—with the effective read-only status being the logical-or of those two.
This seems sensible to me, for whatever that's worth. I'm new to the kernel so I'm not certain I'll produce a reasonable patch, but I'm willing to give it a shot.
So the change would be something like:
1. Add a flag to the gendisk struct to indicate if the device is read-only. The driver sets this.
2. In hd_struct struct's policy flag is one that indicates the user-space setting and this gets set by BLKROSET ioctl/set_disk_ro.
3. The effective read-only status reported by the BLKROGET ioctl and bdev_read_only is the logical-or of the hd_struct policy flag and the gendisk device flag.
Maybe that doesn't make sense at all, though. Thoughts?
Thanks, Jeremy