On Mon, Sep 23, 2019 at 01:38:23PM -0600, Jens Axboe wrote:
> On 9/23/19 5:19 AM, Matthew Wilcox wrote:
> >
> > Ping Jens?
> >
> > On Wed, Sep 18, 2019 at 08:49:49PM -0700, Matthew Wilcox wrote:
> >> On Thu, Sep 19, 2019 at 10:33:10AM +0800, Lin Feng wrote:
> >>> On 9/18/19 20:33, Michal Hocko wrote:
> >>>> I absolutely agree here. From you changelog it is also not clear what is
> >>>> the underlying problem. Both congestion_wait and wait_iff_congested
> >>>> should wake up early if the congestion is handled. Is this not the case?
> >>>
> >>> For now I don't know why, codes seem should work as you said, maybe I need to
> >>> trace more of the internals.
> >>> But weird thing is that once I set the people-disliked-tunable iowait
> >>> drop down instantly, this is contradictory to the code design.
> >>
> >> Yes, this is quite strange. If setting a smaller timeout makes a
> >> difference, that indicates we're not waking up soon enough. I see
> >> two possibilities; one is that a wakeup is missing somewhere -- ie the
> >> conditions under which we call clear_wb_congested() are wrong. Or we
> >> need to wake up sooner.
> >>
> >> Umm. We have clear_wb_congested() called from exactly one spot --
> >> clear_bdi_congested(). That is only called from:
> >>
> >> drivers/block/pktcdvd.c
> >> fs/ceph/addr.c
> >> fs/fuse/control.c
> >> fs/fuse/dev.c
> >> fs/nfs/write.c
> >>
> >> Jens, is something supposed to be calling clear_bdi_congested() in the
> >> block layer? blk_clear_congested() used to exist until October 29th
> >> last year. Or is something else supposed to be waking up tasks that
> >> are sleeping on congestion?
>
> Congestion isn't there anymore. It was always broken as a concept imho,
> since it was inherently racy. We used the old batching mechanism in the
> legacy stack to signal it, and it only worked for some devices.
Umm. OK. Well, something that used to work is now broken. So how
should we fix it? Take a look at shrink_node() in mm/vmscan.c. If we've
submitted a lot of writes to a device, and overloaded it, we want to
sleep until it's able to take more writes:
/*
* Stall direct reclaim for IO completions if underlying BDIs
* and node is congested. Allow kswapd to continue until it
* starts encountering unqueued dirty pages or cycling through
* the LRU too quickly.
*/
if (!sc->hibernation_mode && !current_is_kswapd() &&
current_may_throttle() && pgdat_memcg_congested(pgdat, root))
wait_iff_congested(BLK_RW_ASYNC, HZ/10);
With a standard block device, that now sleeps until the timeout (100ms)
expires, which is far too long for a modern SSD but is probably tuned
just right for some legacy piece of spinning rust (or indeed a modern
USB stick). How would the block layer like to indicate to the mm layer
"I am too busy, please let the device work for a bit"?
