On 2020/3/27 03:36, Pavel Tatashin wrote:
I agree with Daniel, we should look into approach where
pgdat_resize_lock is taken only for the duration of updating tracking
values such as pgdat->first_deferred_pfn (perhaps we would need to add
another tracker that would show chunks that are currently being worked
on).
The vast duration of struct page initialization process should happen
outside of this lock, and only be taken when we update globally seen
data structures: lists, tracking variables. This way we can solve
several problems: 1. allow interrupt threads to grow zones if
required. 2. keep jiffies happy. 3. allow future scaling when we will
add inner node threads to initialize struct pages (i.e. ktasks from
Daniel).
It make sense, looking forward to the inner node parallel init.
@Daniel
Is there schedule about ktasks?
Thanks!
Pasha
On Thu, Mar 26, 2020 at 2:58 PM Daniel Jordan
<daniel.m.jordan@xxxxxxxxxx> wrote:
On Thu, Mar 19, 2020 at 03:05:12PM -0400, Daniel Jordan wrote:
Regardless,
Reviewed-by: Daniel Jordan <daniel.m.jordan@xxxxxxxxxx>
Darn, I spoke too soon.
On a two-socket Xeon, smaller values of TICK_PAGE_COUNT caused the deferred
init timestamp to grow by over 25%. This was with pgdatinit0 bound to the
timer interrupt CPU to make sure the issue always reproduces.
TICK_PAGE_COUNT node 0 deferred
init time (ms)
--------------- ---------------
4096 610
8192 587
16384 487
32768 480 // used in the patch
Instead of trying to find a constant that lets the timer interrupt run often
enough, I think a better way forward is to reconsider how we handle the resize
lock. I plan to prototype something and reply back with what I get.
Yes, the time spend of pages init depends on the CPU frequency,
and the jiffies update controlled by HZ, so it's hard to find a general
constant.
It seems we need a bigger refactors about the lock to get a better solution.
