On 6/6/24 21:49, Mikulas Patocka wrote: > Hi > > The patch dd291d77cc90eb6a86e9860ba8e6e38eebd57d12 (block: Introduce zone > write plugging) causes a performance regression when removing block > devices. > > The kernel 6.9 removes a DM block device in 73ms. The kernel 6.10-rc in > 103ms. > > That patch adds a rcu_barrier() call to the disk-removal code path and it > causes the slowdown. We get this stacktrace when we attempt to remove > large amount of DM devices. Note that the removed devices aren't zoned at > all. OK. Let me send a fix patch to avoid the barrier for non-zoned devices. For zoned devices, it can only be avoided if the device does not need zone write plugging, which is the case for most zoned DM setup. > > [<0>] rcu_barrier+0x208/0x320 > [<0>] disk_free_zone_resources+0x102/0x160 > [<0>] disk_release+0x72/0xe0 > [<0>] device_release+0x34/0x90 > [<0>] kobject_put+0x8b/0x1d0 > [<0>] cleanup_mapped_device+0xd8/0x160 > [<0>] __dm_destroy+0x12a/0x1d0 > [<0>] dm_hash_remove_all+0x77/0x1a0 > [<0>] remove_all+0x23/0x40 > [<0>] ctl_ioctl+0x1dc/0x530 > [<0>] dm_ctl_ioctl+0xe/0x20 > [<0>] __x64_sys_ioctl+0x94/0xd0 > [<0>] do_syscall_64+0x82/0x160 > [<0>] entry_SYSCALL_64_after_hwframe+0x76/0x7e > > Mikulas > -- Damien Le Moal Western Digital Research
