[PATCH v3 4/7] mm: replace vma->vm_flags direct modifications with modifier calls

[Mel Gorman]

On Thu, Jan 26, 2023 at 08:10:26AM -0800, Suren Baghdasaryan wrote:
> On Thu, Jan 26, 2023 at 7:10 AM Mel Gorman <mgorman at techsingularity.net> wrote:
> >
> > On Wed, Jan 25, 2023 at 03:35:51PM -0800, Suren Baghdasaryan wrote:
> > > Replace direct modifications to vma->vm_flags with calls to modifier
> > > functions to be able to track flag changes and to keep vma locking
> > > correctness.
> > >
> > > Signed-off-by: Suren Baghdasaryan <surenb at google.com>
> > > Acked-by: Michal Hocko <mhocko at suse.com>
> >
> > Acked-by: Mel Gorman <mgorman at techsingularity.net>
> >
> > Minor comments that are safe to ignore.
> >
> > I think a better name for mod_vm_flags is set_clear_vm_flags to hint that
> > the first flags are to be set and the second flags are to be cleared.
> > For this patch, it doesn't matter, but it might avoid accidental swapping
> > in the future.
> >
> > reset_vm_flags might also be better named as reinit_vma_flags (or
> > vma_flags_reinit). Maybe also encourage the use of [set|clear_mod]_vm_flags
> > where possible in the comment to track exactly what is changing and
> > why. Some cases like userfaultfd just want to clear __VM_UFFD_FLAGS but
> > altering the flow in this patch is inappropriate and error prone. Others
> > such as the infiniband changes and madvise are a lot more complex.
> 
> That's a good point, but I don't want people to use mod_vm_flags() for
> the cases when the order of set/clear really matters. In such cases
> set_vm_flags() and clear_vm_flags() should be explicitly used. Maybe
> to make that clear I should add a comment and rewrite the functions
> as:
> 
> void mod_vm_flags(vma, set, clear) {
>     vma.vm_flags = vma.vm_flags | set & clear;
> }
> 

Offhand, I'm not thinking of a case where that really matters and as they
are not necessarily ordered, it's raising a read flag so yes, it definitely
it needs a comment if the ordering matters.

> In this patchset it's not that obvious but mod_vm_flags() was really
> introduced in the original per-VMA lock patchset for efficiency to
> avoid taking extra per-VMA locks. A combo of
> set_vm_flags()+clear_vm_flags() would try to retake the same per-VMA
> lock in the second call while mod_vm_flags() takes the lock only once
> and does both operations.

Ok, that seems fair but still needs a comment on why a mod_vm_flags is
not necessarily equivalent to a set_vm_flags + clear_vm_flags in terms of
correctness if that is indeed the case.

> Not a huge overhead because we check if the
> lock is already taken and bail out early but still...
> So, would the above modification to mod_vm_flags() address your concern?
> 

My concerns are entirely with the callers, not the implementation. If
someone is modifying a call site using mod_vm_flags, they have to read
through all the preceding logic to ensure the final combination of flags
is valid.  It's a code maintenance issue, not a correctness issue.

-- 
Mel Gorman
SUSE Labs