hi Dmitry, thanks for your reply. I also test it use the mainline, it also takes more than 1 hours. my mechine has 64 CPUs core and the disk is SATA. when mkfs.ext4, I found the 'scsi_test_unit_ready' run more than 1000 times per second by the different kworker. and every 'scsi_test_unit_ready' takes more than 200us, and the interval less than 20us. So, I think your guess is right. but there is another question, why 4.19 branch takes more than 10 hour? I will work on it, if any information about it, I will reply you. Thanks. my script: dmzadm --format /dev/sdi echo "0 21485322240 zoned /dev/sdi" | dmsetup create dmz-sdi date; mkfs.ext4 /dev/mapper/dmz-sdi; date mainline: [root@localhost ~]# uname -a Linux localhost 5.4.0-rc5 #1 SMP Thu Oct 31 11:41:20 CST 2019 aarch64 aarch64 aarch64 GNU/Linux Thu Oct 31 13:58:55 CST 2019 mke2fs 1.43.6 (29-Aug-2017) Discarding device blocks: done Creating filesystem with 2684354560 4k blocks and 335544320 inodes Filesystem UUID: e0d8e01e-efa8-47fd-a019-b184e66f65b0 Superblock backups stored on blocks: 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208, 4096000, 7962624, 11239424, 20480000, 23887872, 71663616, 78675968, 102400000, 214990848, 512000000, 550731776, 644972544, 1934917632, 2560000000 Allocating group tables: done Writing inode tables: done Creating journal (262144 blocks): done Writing superblocks and filesystem accounting information: done Thu Oct 31 15:01:01 CST 2019 after delete the 'check_events' on mainline: [root@localhost ~]# uname -a Linux localhost 5.4.0-rc5+ #2 SMP Thu Oct 31 15:07:36 CST 2019 aarch64 aarch64 aarch64 GNU/Linux Thu Oct 31 15:19:56 CST 2019 mke2fs 1.43.6 (29-Aug-2017) Discarding device blocks: done Creating filesystem with 2684354560 4k blocks and 335544320 inodes Filesystem UUID: 735198e8-9df0-49fc-aaa8-23b0869dfa05 Superblock backups stored on blocks: 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208, 4096000, 7962624, 11239424, 20480000, 23887872, 71663616, 78675968, 102400000, 214990848, 512000000, 550731776, 644972544, 1934917632, 2560000000 Allocating group tables: done Writing inode tables: done Creating journal (262144 blocks): done Writing superblocks and filesystem accounting information: done Thu Oct 31 15:30:51 CST 2019 在 2019/10/27 10:56, Dmitry Fomichev 写道:
Zhang, I just did some testing of this scenario with a recent kernel that includes this patch. The log below is a run in QEMU with 8 CPUs and it took 18.5 minutes to create the FS on a 14TB ATA drive. Doing the same thing on bare metal with 32 CPUs takes 10.5 minutes in my environment. However, when doing the same test with a SAS drive, the run takes 43 minutes. This is not quite the degradation you are observing, but still a big performance hit. Is the disk that you are using SAS or SATA? My current guess is that sd driver may generate some TEST UNIT READY commands to check if the drive is really online as a part of check_events() processing. For ATA drives, this is nearly a NOP since all TURs are completed internally in libata. But, in SCSI case, these blocking TURs are issued to the drive and certainly may degrade performance. The check_events() call has been added to bdev_device_is_dying() because simply calling bdev_queue_dying() doesn't cover the situation when the drive gets offlined in SCSI layer. It might be possible to only call check_events() once before every reclaim run and to avoid calling it in I/O mapping path. If this works, the overhead would likely be acceptable. I am going to take a look into this. Regards, Dmitry [root@xxx dmz]# uname -a Linux xxx 5.4.0-rc1-DMZ+ #1 SMP Fri Oct 11 11:23:13 PDT 2019 x86_64 x86_64 x86_64 GNU/Linux [root@xxx dmz]# lsscsi [0:0:0:0] disk QEMU QEMU HARDDISK 2.5+ /dev/sda [1:0:0:0] zbc ATA HGST HSH721415AL T240 /dev/sdb [root@xxx dmz]# ./setup-dmz test /dev/sdb [root@xxx dmz]# cat /proc/kallsyms | grep dmz_bdev_is_dying (standard input):90782:ffffffffc070a401 t dmz_bdev_is_dying.cold [dm_zoned] (standard input):90849:ffffffffc0706e10 t dmz_bdev_is_dying [dm_zoned] [root@xxx dmz]# time mkfs.ext4 /dev/mapper/test mke2fs 1.44.6 (5-Mar-2019) Discarding device blocks: done Creating filesystem with 3660840960 4k blocks and 457605120 inodes Filesystem UUID: 4536bacd-cfb5-41b2-b0bf-c2513e6e3360 Superblock backups stored on blocks: 32768, 98304, 163840, 229376, 294912, 819200, 884736, 1605632, 2654208, 4096000, 7962624, 11239424, 20480000, 23887872, 71663616, 78675968, 102400000, 214990848, 512000000, 550731776, 644972544, 1934917632, 2560000000 Allocating group tables: done Writing inode tables: done Creating journal (262144 blocks): done Writing superblocks and filesystem accounting information: done real 18m30.867s user 0m0.172s sys 0m11.198s On Sat, 2019-10-26 at 09:56 +0800, zhangxiaoxu (A) wrote:Hi all, when I 'mkfs.ext4' on the dmz device based on 10T smr disk, it takes more than 10 hours after apply 75d66ffb48efb3 ("dm zoned: properly handle backing device failure"). After delete the 'check_events' in 'dmz_bdev_is_dying', it just take less than 12 mins. I test it based on 4.19 branch. Must we do the 'check_events' at mapping path, reclaim or metadata I/O? Thanks.
-- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
