On 2023/7/14 21:42, Joel Fernandes wrote:
On Thu, Jul 13, 2023 at 11:17 PM Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx> wrote:
On 2023/7/14 10:16, Paul E. McKenney wrote:
On Thu, Jul 13, 2023 at 09:33:35AM -0700, Paul E. McKenney wrote:
On Thu, Jul 13, 2023 at 11:33:24AM -0400, Joel Fernandes wrote:
...
>From what Sandeep described, the code path is in an RCU reader. My
question is more, why doesn't it use SRCU instead since it clearly
does so if BLK_MQ_F_BLOCKING. What are the tradeoffs? IMHO, a deeper
dive needs to be made into that before concluding that the fix is to
use rcu_read_lock_any_held().
How can this be solved?
1. Always use a workqueue. Simple, but is said to have performance
issues.
2. Pass a flag in that indicates whether or not the caller is in an
RCU read-side critical section. Conceptually simple, but might
or might not be reasonable to actually implement in the code as
it exists now. (You tell me!)
3. Create a function in z_erofs that gives you a decent
approximation, maybe something like the following.
4. Other ideas here.
5. #3 plus make the corresponding Kconfig option select
PREEMPT_COUNT, assuming that any users needing compression in
non-preemptible kernels are OK with PREEMPT_COUNT being set.
(Some users of non-preemptible kernels object strenuously
to the added overhead from CONFIG_PREEMPT_COUNT=y.)
I'm not sure if it's a good idea
I think it is a fine idea.
we need to work on
CONFIG_PREEMPT_COUNT=n (why not?), we could just always trigger a
workqueue for this.
So CONFIG_PREEMPT_COUNT=n users don't deserve good performance? ;-)
I'm not sure if non-preemptible kernel users really care about
such sensitive latencies, I don't know, my 2 cents.
Thanks,
Gao Xiang
thanks,
- Joel
