Re: btrfs mount takes too long time

[David Sterba <dsterba@suse.cz>]

On Tue, Aug 31, 2021 at 01:10:38PM +0000, Johannes Thumshirn wrote:
> [ +Cc Naohiro ]
> 
> On 31/08/2021 14:11, Anand Jain wrote:
> > But let's see what part in btrfs_load_block_group_zone_info() is taking
> > most of the time. Could you please help get this output?
> 
> I'd suspect it's the btrfs_get_dev_zone() call that we need to do for 
> each block group in order to determine the current write pointer to
> set block_group->alloc_offset.
> 
> We could speed up the mount process by caching the device's REPORT ZONES
> response, as we're doing a REPORT ZONES once to get all zones and then 
> again per block group load. On a 14TB SMR drive this results in 
> (14  * 1024 * 1024) / 256 + 1 = 57345 REPORT ZONES calls. OTOH 
> struct blk_zone is 64 bytes per zone resulting in 64 * 57344 = 3584kB
> data to be cached.
> 
> So the solution should probably somewhere in between, aka caching X 
> REPORT ZONES responses before we need to do a new one.

That sounds like a good option. The number of new members in block group
struct has grown due to zoned support, this means higher memory
requirements, scaling up with device size. We'll have to do something
about that eventually.