On Thu 22-11-18 16:12:33, Theodore Y. Ts'o wrote: > On Thu, Nov 22, 2018 at 01:21:55PM +0100, Jan Kara wrote: > > > > Is there a reason for REQ_PRIO? I'm not sure all REQ_META reads are really > > a priority ones... > > Hmm, good question. With the exception of readahead, most reads will > be blocking some process. We are currently using REQ_PRIO for all > directory block reads. The ext4_sb_read() function gets used for > resizing, indirect block map to extent tree migration, and extended > attribute reads. The last is the most common, and arguably the most > justifiable to be REQ_PRIO. (Of course my understanding is that the > block layer is ignoring REQ_PRIO, so this is mostly academic...) Agreed that this is currently academic but AFAIR some blk-mq schedulers might be improved to consider REQ_PRIO again. > So I think what I'll do is this. I'll add a parameter to > ext4_sb_read() so each caller can use use REQ_PRIO. REQ_PRIO will be > used from xattr.c, but not from fs/ext4/migrate.c and > fs/ext4/resize.c. Sounds good. Thanks! Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
