[dm-devel] Re: [PATCH 2.6.19 5/5] fs: freeze_bdev with semaphore not mutex
Rafael J. Wysocki
rjw at sisk.pl
Wed Nov 8 23:06:11 UTC 2006
On Wednesday, 8 November 2006 16:25, Alasdair G Kergon wrote:
> On Wed, Nov 08, 2006 at 03:43:26PM +0100, Rafael J. Wysocki wrote:
> > Will it be enough to cover the interactions with dm?
>
> There are cases where you *cannot* freeze the filesystem (unless
> you wait for userspace processes to finish what they are doing) -
> and only dm can tell you that.
>
> The freeze_filesystems() code ought to do it's best in any given
> circumstances within the constraints.
>
> So:
> Get the filesystem's block device.
> Check the full tree of devices that that block device depends upon
> and for any device that belongs to device-mapper check if it is suspended.
> If it is, skip the original device.
>
> struct mapped_device *dm_get_md(dev_t dev);
> int dm_suspended(struct mapped_device *md);
> void dm_put(struct mapped_device *md);
>
> Handling the tree is the difficult bit, but sysfs could help.
> [You can get the device-mapper dependencies with:
> struct mapped_device *dm_get_md(dev_t dev);
> struct dm_table *dm_get_table(struct mapped_device *md);
> struct list_head *dm_table_get_devices(struct dm_table *t);
> void dm_table_put(struct dm_table *t);
> void dm_put(struct mapped_device *md);
> ]
>
> Consider that you could have an already-frozen filesystem containing a loop
> device containing a device-mapper device containing a not-frozen filesystem.
I think the last point is handled correctly by freezing the filesystems in the
reverse order - unless the fs below the loop has been frozen before by
someone else, but I guess that would lead to problems anyway.
> You won't be able to freeze that top filesystem because the I/O would
> queue lower down the stack. (Similar problem if the device-mapper device
> in the stack was suspended.)
The suspended dm device in the stack is not, however.
Is there any list of all dm devices in the system?
Rafael
--
You never change things by fighting the existing reality.
R. Buckminster Fuller
More information about the dm-devel
mailing list