Re: [PATCH] mm: always consider THP when adjusting min_free_kbytes

[Matthew Wilcox]

On Thu, Feb 06, 2020 at 01:23:21PM -0800, Mike Kravetz wrote:
> On 2/6/20 12:39 PM, Matthew Wilcox wrote:
> > On Wed, Feb 05, 2020 at 05:36:44PM -0800, Mike Kravetz wrote:
> >> The value of min_free_kbytes is calculated in two routines:
> >> 1) init_per_zone_wmark_min based on available memory
> >> 2) set_recommended_min_free_kbytes may reserve extra space for
> >>    THP allocations
> >>
> >> In both of these routines, a user defined min_free_kbytes value will
> >> be overwritten if the value calculated in the code is larger. No message
> >> is logged if the user value is overwritten.
> >>
> >> Change code to never overwrite user defined value.  However, do log a
> >> message (once per value) showing the value calculated in code.
> > 
> > But what if the user set min_free_kbytes to, say, half of system memory,
> > and then hot-unplugs three quarters of their memory?  I think the kernel
> > should protect itself against such foolishness.
> 
> I'm not sure what we should set it to in this case.  Previously you said,
> 
> >> I'm reluctant to suggest we do a more complex adjustment of the value
> >> (eg figure out what the adjustment would have been, then apply some
> >> fraction of that adjustment to keep the ratios in proportion) because
> >> we don't really know why they adjusted it.
> 
> So, I suspect you would suggest setting it to the default computed value?
> But then, when do we start adjusting?  What if they only remove a small
> amount of memory?  And, then add the same amount back in?

I don't know about the default computed value ... we don't seem to have
any protection against the user setting min_free_kbytes to double the
amount of memory in the machine today.  Which would presumably cause
problems if I asked to maintain 32GB free at all times on my 16GB laptop?

Maybe we should have such protection?

> BTW - In the above scenario existing code would not change min_free_kbytes
> because the user defined value is greater than value computed in code.

True!