On Fri, 7 Jun 2024, Damien Le Moal wrote:

> For zoned block devices using zone write plugging, an rcu_barrier() call
> is needed in disk_free_zone_resources() to synchronize freeing of zone
> write plugs and the destrution of the mempool used to allocate the
> plugs. The barrier call does slow down a little teardown of zoned block
> devices but should not affect teardown of regular block devices or zoned
> block devices that do not use zone write plugging (e.g. zoned DM devices
> that do not require zone append emulation).
> 
> Modify disk_free_zone_resources() to return early if we do not have a
> mempool to start with, that is, if the device does not use zone write
> plugging. This avoids the costly rcu_barrier() and speeds up disk
> teardown.
> 
> Reported-by: Mikulas Patocka <[email protected]>
> Fixes: dd291d77cc90 ("block: Introduce zone write plugging")
> Signed-off-by: Damien Le Moal <[email protected]>

This works.

Tested-by: Mikulas Patocka <[email protected]>


BTW. when we remove large number of DM devices, the process spends a lot 
of time sleeping with this stacktrace:

__schedule+0x242/0x600
schedule+0x21/0xd0
blk_mq_freeze_queue_wait+0x55/0x90
? __wake_up+0x50/0x50
del_gendisk+0x1fc/0x320
cleanup_mapped_device+0x16c/0x180 [dm_mod]
__dm_destroy+0x145/0x1f0 [dm_mod]
dm_hash_remove_all+0x5c/0x180 [dm_mod]
? dm_hash_remove_all+0x180/0x180 [dm_mod]
remove_all+0x1a/0x30 [dm_mod]
ctl_ioctl+0x1bb/0x500 [dm_mod]
dm_compat_ctl_ioctl+0x7/0x10 [dm_mod]
__x64_compat_sys_ioctl+0x133/0x160
do_syscall_64+0x17c/0x1a0
entry_SYSCALL_64_after_hwframe+0x4b/0x53

Is there some way how to remove this wait?

Mikulas

> ---
>  block/blk-zoned.c | 3 +++
>  1 file changed, 3 insertions(+)
> 
> diff --git a/block/blk-zoned.c b/block/blk-zoned.c
> index 8f89705f5e1c..137842dbb59a 100644
> --- a/block/blk-zoned.c
> +++ b/block/blk-zoned.c
> @@ -1552,6 +1552,9 @@ static void disk_destroy_zone_wplugs_hash_table(struct 
> gendisk *disk)
>  
>  void disk_free_zone_resources(struct gendisk *disk)
>  {
> +     if (!disk->zone_wplugs_pool)
> +             return;
> +
>       cancel_work_sync(&disk->zone_wplugs_work);
>  
>       if (disk->zone_wplugs_wq) {
> -- 
> 2.45.2
> 


Reply via email to