Here is a link to Linus' reply to Jan's concern about making i_blkbibts byte addressable: https://marc.info/?l=linux-fsdevel&m=152882624707975&w=2
Here is a link to an lkp.org report about potential performance improvement in some workload, which could(?) be related to packing i_blkbits closer to i_bytes/i_lock: https://marc.info/?l=linux-fsdevel&m=153077048108198&w=2 Changes since v1: - Add links to relevant discussions Signed-off-by: Amir Goldstein <[email protected]> --- Al, Re-posting the patch per your request. Regarding your question: "I would like more details about the variation of timing - what's the dispersion from boot to boot, for starters?" I haven't run those performance tests, just got them from lkp robot, so perhaps Xiaolong can answer your question. I too, find the reported improvement a bit too good to be reliably true. I do see in the graphs at the bottom of the report that there are many good vs. bad samples though. Thanks, Amir. include/linux/fs.h | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/include/linux/fs.h b/include/linux/fs.h index 760d8da1b6c7..6d0489613dc1 100644 --- a/include/linux/fs.h +++ b/include/linux/fs.h @@ -274,6 +274,7 @@ struct writeback_control; /* * Write life time hint values. + * Stored in struct inode as u8. */ enum rw_hint { WRITE_LIFE_NOT_SET = 0, @@ -607,8 +608,8 @@ struct inode { struct timespec i_ctime; spinlock_t i_lock; /* i_blocks, i_bytes, maybe i_size */ unsigned short i_bytes; - unsigned int i_blkbits; - enum rw_hint i_write_hint; + u8 i_blkbits; + u8 i_write_hint; blkcnt_t i_blocks; #ifdef __NEED_I_SIZE_ORDERED -- 2.7.4
