lkml.org 
[lkml]   [2017]   [Sep]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 4/5] mm:swap: respect page_cluster for readahead
On Tue, Sep 12, 2017 at 03:29:45PM +0800, Huang, Ying wrote:
> Minchan Kim <minchan@kernel.org> writes:
>
> > On Tue, Sep 12, 2017 at 02:44:36PM +0800, Huang, Ying wrote:
> >> Minchan Kim <minchan@kernel.org> writes:
> >>
> >> > On Tue, Sep 12, 2017 at 01:23:01PM +0800, Huang, Ying wrote:
> >> >> Minchan Kim <minchan@kernel.org> writes:
> >> >>
> >> >> > page_cluster 0 means "we don't want readahead" so in the case,
> >> >> > let's skip the readahead detection logic.
> >> >> >
> >> >> > Cc: "Huang, Ying" <ying.huang@intel.com>
> >> >> > Signed-off-by: Minchan Kim <minchan@kernel.org>
> >> >> > ---
> >> >> > include/linux/swap.h | 3 ++-
> >> >> > 1 file changed, 2 insertions(+), 1 deletion(-)
> >> >> >
> >> >> > diff --git a/include/linux/swap.h b/include/linux/swap.h
> >> >> > index 0f54b491e118..739d94397c47 100644
> >> >> > --- a/include/linux/swap.h
> >> >> > +++ b/include/linux/swap.h
> >> >> > @@ -427,7 +427,8 @@ extern bool has_usable_swap(void);
> >> >> >
> >> >> > static inline bool swap_use_vma_readahead(void)
> >> >> > {
> >> >> > - return READ_ONCE(swap_vma_readahead) && !atomic_read(&nr_rotate_swap);
> >> >> > + return page_cluster > 0 && READ_ONCE(swap_vma_readahead)
> >> >> > + && !atomic_read(&nr_rotate_swap);
> >> >> > }
> >> >> >
> >> >> > /* Swap 50% full? Release swapcache more aggressively.. */
> >> >>
> >> >> Now the readahead window size of the VMA based swap readahead is
> >> >> controlled by /sys/kernel/mm/swap/vma_ra_max_order, while that of the
> >> >> original swap readahead is controlled by sysctl page_cluster. It is
> >> >> possible for anonymous memory to use VMA based swap readahead and tmpfs
> >> >> to use original swap readahead algorithm at the same time. So that, I
> >> >> think it is necessary to use different control knob to control these two
> >> >> algorithm. So if we want to disable readahead for tmpfs, but keep it
> >> >> for VMA based readahead, we can set 0 to page_cluster but non-zero to
> >> >> /sys/kernel/mm/swap/vma_ra_max_order. With your change, this will be
> >> >> impossible.
> >> >
> >> > For a long time, page-cluster have been used as controlling swap readahead.
> >> > One of example, zram users have been disabled readahead via 0 page-cluster.
> >> > However, with your change, it would be regressed if it doesn't disable
> >> > vma_ra_max_order.
> >> >
> >> > As well, all of swap users should be aware of vma_ra_max_order as well as
> >> > page-cluster to control swap readahead but I didn't see any document about
> >> > that. Acutaully, I don't like it but want to unify it with page-cluster.
> >>
> >> The document is in
> >>
> >> Documentation/ABI/testing/sysfs-kernel-mm-swap
> >>
> >> The concern of unifying it with page-cluster is as following.
> >>
> >> Original swap readahead on tmpfs may not work well because the combined
> >> workload is running, so we want to disable or constrain it. But at the
> >> same time, the VMA based swap readahead may work better. So I think it
> >> may be necessary to control them separately.
> >
> > My concern is users have been disabled swap readahead by page-cluster would
> > be regressed. Please take care of them.
>
> How about disable VMA based swap readahead if zram used as swap? Like
> we have done for hard disk?

It could be with SWP_SYNCHRONOUS_IO flag which indicates super-fast,
no seek cost swap devices if this patchset is merged so VM automatically
disables readahead. It is in my TODO but it's orthogonal work.

The problem I raised is "Why shouldn't we obey user's decision?",
not zram sepcific issue.

A user has used SSD as swap devices decided to disable swap readahead
by some reason(e.g., small memory system). Anyway, it has worked
via page-cluster for a several years but with vma-based swap devices,
it doesn't work any more.

\
 
 \ /
  Last update: 2017-09-12 09:57    [W:0.126 / U:0.404 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site