disk corruption in 2.5.66-bk5 - 2.5.69

disk corruption in 2.5.66-bk5 - 2.5.69

[Adam J. Richter]

	linux-2.5.66-bk5 changed blkdev_put in fs/block_dev.c
to only call sync_blockdev if bdev->bd_openers drops to zero.
For me, this has resulted in disk corruption when lilo runs
under kernels 2.5.66-bk5 through 2.5.69, even if I run
"sync" repeatedly after running lilo.

	I am not sure why I seem to be the first person reporting
this disastrous problem.  I suspect it may have something to do
with the fact that I boot from an initial ramdisk that changes
the root without using pivot_root (instead, process 1 does a chroot
at a time when it is the only process on the system) and the fact
that I user partitions defined by partx rather than kernel-based
partitioning.  Perhaps bdev->bd_openers never drops to zero in this
configuration as a result of some other bug that only occurs in
such a configuration.

	I notice also that that change in 2.5.66-bk5 causes
sync_blockdev to be called with the big kernel lock held, although
I do not think that is the cause of the data corruption.

	With the following kludge, the lilo data corruption 
in 2.5.69 seems to be gone.  I have deliberately not cleaned
up this patch for integration at this point, because I don't
understand the issue enough yet to know the real problem
and the right fix.

Adam J. Richter     __     ______________   575 Oroville Road
adam@yggdrasil.com     \ /                  Miplitas, California 95035
+1 408 309-6081         | g g d r a s i l   United States of America
                         "Free Software For The Rest Of Us."

--- linux-2.5.69/fs/block_dev.c.orig	2003-05-14 17:43:40.000000000 -0700
+++ linux-2.5.69/fs/block_dev.c	2003-05-14 17:44:29.000000000 -0700
@@ -635,14 +635,24 @@
 int blkdev_put(struct block_device *bdev, int kind)
 {
 	int ret = 0;
 	struct inode *bd_inode = bdev->bd_inode;
 	struct gendisk *disk = bdev->bd_disk;
 
 	down(&bdev->bd_sem);
+
+	/* AJR start */
+	switch (kind) {
+	case BDEV_FILE:
+	case BDEV_FS:
+		sync_blockdev(bd_inode->i_bdev);
+		break;
+	}
+	/* AJR end */
+
 	lock_kernel();
 	if (!--bdev->bd_openers) {
 		switch (kind) {
 		case BDEV_FILE:
 		case BDEV_FS:
 			sync_blockdev(bd_inode->i_bdev);
 			break;

Re: disk corruption in 2.5.66-bk5 - 2.5.69

[Andrew Morton]

"Adam J. Richter" <adam@yggdrasil.com> wrote:
>
> 	linux-2.5.66-bk5 changed blkdev_put in fs/block_dev.c
>  to only call sync_blockdev if bdev->bd_openers drops to zero.
>  For me, this has resulted in disk corruption when lilo runs
>  under kernels 2.5.66-bk5 through 2.5.69, even if I run
>  "sync" repeatedly after running lilo.

Works OK here with a simple test:

	sleep 10000 < /dev/hda5 &
	cat < some-file > /dev/hda5
	sync
	<press reset>
	cmp some-file /dev/hda5

LILO does a global sync itself before quitting.  Check to see if there's a
lot of writeout happening.  See if there's writeout happening when you run
sync manually.

Maybe take a copy of /dev/hda1 into a regular file, then reboot, then
compare with what's on the disk.

Check that /proc/meminfo:Dirty goes to zero after `sync'.