On 3/11/25 08:42, Benjamin Marzinski wrote: > On Tue, Mar 11, 2025 at 08:19:29AM +0900, Damien Le Moal wrote: >> On 3/11/25 02:43, Benjamin Marzinski wrote: >>> On Mon, Mar 10, 2025 at 08:59:26AM +0900, Damien Le Moal wrote: >>>> On 3/10/25 07:29, Benjamin Marzinski wrote: >>>>> dm_revalidate_zones() only allowed devices that had no zone resources >>>>> set up to call blk_revalidate_disk_zones(). If the device already had >>>>> zone resources, disk->nr_zones would always equal md->nr_zones so >>>>> dm_revalidate_zones() returned without doing any work. Instead, always >>>>> call blk_revalidate_disk_zones() if you are loading a new zoned table. >>>>> >>>>> However, if the device emulates zone append operations and already has >>>>> zone append emulation resources, the table size cannot change when >>>>> loading a new table. Otherwise, all those resources will be garbage. >>>>> >>>>> If emulated zone append operations are needed and the zone write pointer >>>>> offsets of the new table do not match those of the old table, writes to >>>>> the device will still fail. This patch allows users to safely grow and >>>>> shrink zone devices. But swapping arbitrary zoned tables will still not >>>>> work. >>>> >>>> I do not think that this patch correctly address the shrinking of dm zoned >>>> device: blk_revalidate_disk_zones() will look at a smaller set of zones, which >>>> will leave already hashed zone write plugs outside of that new zone range in the >>>> disk zwplug hash table. disk_revalidate_zone_resources() does not cleanup and >>>> reallocate the hash table if the number of zones shrinks. >>> >>> This is necessary for DM. There could be plugged bios that are on on >>> these no longer in-range zones. They will obviously fail when they get >>> reissued, but we need to keep the plugs around so that they *do* get >>> reissued. A cleaner alternative would be to add code to immediately >>> error out all the plugged bios on shrinks, but I was trying to avoid >>> adding a bunch of code to deal with these cases (of course simply >>> disallowing them adds even less code). >> >> I am confused now :) >> Under the assumption that we do not allow switching to a new table that changes >> the zone configuration (in particualr, there is no grow/shrink of the device), >> then I do not think we have to do anything special for DM. > > If we don't allow switching between zoned tables, then we obviously > don't need to make DM call blk_revalidate_disk_zones() on a zoned table > switch. I was just saying that I know that this patch would leave > out-of-range zone plugs behind on a shrink, but that is necessary to > allow shrinking while there could still be outstanding plugged bios > attached to those plugs. > > So, if we wanted to allow shrinking, then I think this patch is correct > (although erroring out all the out-of-range bios would be a cleaner > solution). But assuming we don't allow shrinking, then you are correct. > We don't need to do anything special for DM. OK. Got it. As mentioned, I think we really still need to allow switching (or inserting ?) dm-error. In that case, calling blk_revalidate_disk_zones() should still be fine as long as there is no change to the zone config, which is true for dm-error. Do you agree ? -- Damien Le Moal Western Digital Research
