2023-09-22 18:33:01

by Zach O'Keefe

[permalink] [raw]
Subject: Re: [EXTERNAL] Re: [PATCH v3] mm/thp: fix "mm: thp: kill __transhuge_page_enabled()"

On Fri, Sep 22, 2023 at 9:54 AM Yang Shi <[email protected]> wrote:
>
> On Tue, Sep 19, 2023 at 10:44 PM Saurabh Singh Sengar
> <[email protected]> wrote:
> >
> > On Tue, Sep 05, 2023 at 11:58:17PM -0700, Saurabh Singh Sengar wrote:
> > > On Fri, Aug 25, 2023 at 08:09:07AM -0700, Zach O'Keefe wrote:
> > > > On Fri, Aug 25, 2023 at 5:58 AM David Hildenbrand <[email protected]> wrote:
> > > > >
> > > > > On 25.08.23 14:49, Matthew Wilcox wrote:
> > > > > > On Fri, Aug 25, 2023 at 09:59:23AM +0200, David Hildenbrand wrote:
> > > > > >> Especially, we do have bigger ->huge_fault changes coming up:
> > > > > >>
> > > > > >> https://lkml.kernel.org/r/[email protected]
> > > >
> > > > FWIW, one of those patches updates the docs to read,
> > > >
> > > > "->huge_fault() is called when there is no PUD or PMD entry present. This
> > > > gives the filesystem the opportunity to install a PUD or PMD sized page.
> > > > Filesystems can also use the ->fault method to return a PMD sized page,
> > > > so implementing this function may not be necessary. In particular,
> > > > filesystems should not call filemap_fault() from ->huge_fault(). [..]"
> > > >
> > > > Which won't work (in the general case) without this patch (well, at
> > > > least the ->huge_fault() check part).
> > > >
> > > > So, if we're advertising this is the way it works, maybe that gives a
> > > > stronger argument for addressing it sooner vs when the first in-tree
> > > > user depends on it?
> > > >
> > > > > >> If the driver is not in the tree, people don't care.
> > > > > >>
> > > > > >> You really should try upstreaming that driver.
> > > > > >>
> > > > > >>
> > > > > >> So this patch here adds complexity (which I don't like) in order to keep an
> > > > > >> OOT driver working -- possibly for a short time. I'm tempted to say "please
> > > > > >> fix your driver to not use huge faults in that scenario, it is no longer
> > > > > >> supported".
> > > > > >>
> > > > > >> But I'm just about to vanish for 1.5 week into vacation :)
> > > > > >>
> > > > > >> @Willy, what are your thoughts?
> > > > > >
> > > > > > Fundamentally there was a bad assumption with the original patch --
> > > > > > it assumed that the only reason to support ->huge_fault was for DAX,
> > > > > > and that's not true. It's just that the only drivers in-tree which
> > > > > > support ->huge_fault do so in order to support DAX.
> > > > >
> > > > > Okay, and we are willing to continue supporting that then and it's
> > > > > nothing we want to stop OOT drivers from doing.
> > > > >
> > > > > Fine with me; we should probably reflect that in the patch description.
> > > >
> > > > I can change these paragraphs,
> > > >
> > > > "During the review of the above commits, it was determined that in-tree
> > > > users weren't affected by the change; most notably, since the only relevant
> > > > user (in terms of THP) of VM_MIXEDMAP or ->huge_fault is DAX, which is
> > > > explicitly approved early in approval logic. However, there is at least
> > > > one occurrence where an out-of-tree driver that used
> > > > VM_HUGEPAGE|VM_MIXEDMAP with a vm_ops->huge_fault handler, was broken.
> > > >
> > > > Remove the VM_NO_KHUGEPAGED check when not in collapse path and give
> > > > any ->huge_fault handler a chance to handle the fault. Note that we
> > > > don't validate the file mode or mapping alignment, which is consistent
> > > > with the behavior before the aforementioned commits."
> > > >
> > > > To read,
> > > >
> > > > "The above commits, however, overfit the existing in-tree use cases,
> > > > and assume that
> > > > the only reason to support ->huge_fault was for DAX (which is
> > > > explicitly approved early in the approval logic).
> > > > This is a bad assumption to make and unnecessarily prevents general
> > > > support of ->huge_fault by filesystems. Allow returning "true" if such
> > > > a handler exists, giving the fault path an opportunity to exercise it.
> > > >
> > > > Similarly, the rationale for including the VM_NO_KHUGEPAGED check
> > > > along the fault path was that it didn't alter any in-tree users, but
> > > > was likewise similarly unnecessarily restrictive (and reads odd).
> > > > Remove the check from the fault path."
> > > >
> > >
> > >
> > > Any chance this can make it to 6.6 kernel ?
> >
> > ping
>
> I think we tend to merge this patch, but anyway it is Andrew's call.
> Included Andrew in this loop.

Sorry for delay -- just back from (another) OOO,

From this back/forth with David/Matthew, seems like we're OK saying,
"this was a mistake", and that we can take the patch (need some form
of Ack or Reviewed-by from them first, to confirm)

> > Fundamentally there was a bad assumption with the original patch --
> > it assumed that the only reason to support ->huge_fault was for DAX,
> > and that's not true. It's just that the only drivers in-tree which
> > support ->huge_fault do so in order to support DAX.
>
> Okay, and we are willing to continue supporting that then and it's
> nothing we want to stop OOT drivers from doing.
>
> Fine with me; we should probably reflect that in the patch description.

But, I don't know about timing. We are in 6.6-rc2, and this hasn't
been exposed in Andrew's trees yet. 6.6 is looking like it could be a
LTS candidate, in which case this patch could flow backwards from
-stable (which would also land in 6.1-y) .. but I don't know if that
path is suitable for this.

Otherwise, perhaps you could include this fix
when you're ready to upstream your driver?

> >
> > >
> > > - Saurabh


2023-09-22 22:54:33

by Andrew Morton

[permalink] [raw]
Subject: Re: [EXTERNAL] Re: [PATCH v3] mm/thp: fix "mm: thp: kill __transhuge_page_enabled()"

On Fri, 22 Sep 2023 09:56:21 -0700 "Zach O'Keefe" <[email protected]> wrote:

> >From this back/forth with David/Matthew, seems like we're OK saying,
> "this was a mistake", and that we can take the patch (need some form
> of Ack or Reviewed-by from them first, to confirm)

Yup. And please let's update the changelog to reflect the details
which have been discussed thus far.

If the change *makes sense* for the current kernel then let's proceed,
regardless of the broken driver issue.

But adding a cc:stable would require extra argumentation, which I will
be interested to read ;)