On 4/9/25 8:51 AM, Benjamin Marzinski wrote:
> If dm_get_live_table() returned NULL, dm_put_live_table() was never
> called. Also, it is possible that md->zone_revalidate_map will change
> while calling this function. Only read it once, so that we are always
> using the same value. Otherwise we might miss a call to
> dm_put_live_table().
>
> Finally, while md->zone_revalidate_map is set and a process is calling
> blk_revalidate_disk_zones() to set up the zone append emulation
> resources, it is possible that another process, perhaps triggered by
> blkdev_report_zones_ioctl(), will call dm_blk_report_zones(). If
> blk_revalidate_disk_zones() fails, these resources can be freed while
> the other process is still using them, causing a use-after-free error.
>
> blk_revalidate_disk_zones() will only ever be called when initially
> setting up the zone append emulation resources, such as when setting up
> a zoned dm-crypt table for the first time. Further table swaps will not
> set md->zone_revalidate_map or call blk_revalidate_disk_zones().
> However it must be called using the new table (referenced by
> md->zone_revalidate_map) and the new queue limits while the DM device is
> suspended. dm_blk_report_zones() needs some way to distinguish between a
> call from blk_revalidate_disk_zones(), which must be allowed to use
> md->zone_revalidate_map to access this not yet activated table, and all
> other calls to dm_blk_report_zones(), which should not be allowed while
> the device is suspended and cannot use md->zone_revalidate_map, since
> the zone resources might be freed by the process currently calling
> blk_revalidate_disk_zones().
>
> Solve this by tracking the process that set md->zone_revalidate_map
> dm_revalidate_zones() and only allowing that process to make use of it
> in dm_blk_report_zones().
..that sets md->zone_revalidate_map in dm_revalidate_zones() and... ?
>
> Fixes: f211268ed1f9b ("dm: Use the block layer zone append emulation")
> Reviewed-by: Damien Le Moal <dlem...@kernel.org>
I do not recall reviewing this version, but looks OK to me with the sentence
above fixed.
> Signed-off-by: Benjamin Marzinski <bmarz...@redhat.com>
--
Damien Le Moal
Western Digital Research
