On 10/12/23 01:49, Kanchan Joshi wrote:
Function does OR bio->bi_ioprio with whatever is the return of
get_current_ioprio().
No, that's not what ioprio_set_class_and_level() does. It clears the
hint bits before it performs a logical OR.
So if lifetime bits were set in get_current_ioprio(), you will end up
setting that in bio->bi_ioprio too.
I'm not sure there are any use cases where it is useful to set the data
lifetime for an entire process.
Anyway, how about replacing this patch with the patch below? This will
allow to set hint information for an entire process.
diff --git a/block/blk-mq.c b/block/blk-mq.c
index e2d11183f62e..3419ca4c1bf4 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2924,9 +2924,14 @@ static inline struct request
*blk_mq_get_cached_request(struct request_queue *q,
static void bio_set_ioprio(struct bio *bio)
{
+ u16 cur_ioprio = get_current_ioprio();
+
/* Nobody set ioprio so far? Initialize it based on task's nice value */
if (IOPRIO_PRIO_CLASS(bio->bi_ioprio) == IOPRIO_CLASS_NONE)
- bio->bi_ioprio = get_current_ioprio();
+ bio->bi_ioprio |= cur_ioprio & IOPRIO_CLASS_LEVEL_MASK;
+ if (IOPRIO_PRIO_HINT(bio->bi_ioprio) == 0)
+ bio->bi_ioprio |= cur_ioprio &
+ (IOPRIO_HINT_MASK << IOPRIO_HINT_SHIFT);
blkcg_set_ioprio(bio);
}
diff --git a/include/linux/ioprio.h b/include/linux/ioprio.h
index 7578d4f6a969..5697832f35a3 100644
--- a/include/linux/ioprio.h
+++ b/include/linux/ioprio.h
@@ -71,4 +71,7 @@ static inline int ioprio_check_cap(int ioprio)
}
#endif /* CONFIG_BLOCK */
+#define IOPRIO_CLASS_LEVEL_MASK ((IOPRIO_CLASS_MASK <<
IOPRIO_CLASS_SHIFT) | \
+ (IOPRIO_LEVEL_MASK << 0))
+
#endif
ioprio_set_class_and_level() preserves the hint bits set by F2FS.
And what is the user interface you have in mind. Is it ioprio based, or
write-hint based or mix of both?
Since the data lifetime is encoded in the hint bits, the hint bits need
to be set by user space to set a data lifetime.
I asked because more than one way seems to emerge here. Parts of this
series (Patch 4) are taking inode->i_write_hint (and not ioprio value)
and putting that into bio.
I wonder what to expect if application get to send one lifetime with
fcntl (write-hints) and different one with ioprio. Is that not racy?
There is no race condition. F_SET_RW_HINT can be used to set
inode->i_write_hint. The filesystem may use the inode->i_write_hint
information. I think F2FS uses this information in its block allocator.
+ --prio=$((i<<6))
This will not work as prio can only take values between 0-7.
Perhaps you want to use "priohint" to send lifetime.
Thanks for having mentioned the priohint option. The above works on my
test setup since I modified fio locally to accept larger prio values. I
will switch to the priohint option.
Bart.
