David Hildenbrand <david@xxxxxxxxxx> writes:
On 27.06.24 02:54, Alistair Popple wrote:
Currently DAX folio/page reference counts are managed differently to
normal pages. To allow these to be managed the same as normal pages
introduce dax_insert_pfn_pud. This will map the entire PUD-sized folio
and take references as it would for a normally mapped page.
This is distinct from the current mechanism, vmf_insert_pfn_pud,
which
simply inserts a special devmap PUD entry into the page table without
holding a reference to the page for the mapping.
Do we really have to involve mapcounts/rmap for daxfs pages at this
point? Or is this only "to make it look more like other pages" ?
The aim of the series is make FS DAX and other ZONE_DEVICE pages look
like other pages, at least with regards to the way they are refcounted.
At the moment they are not refcounted - instead their refcounts are
basically statically initialised to one and there are all these special
cases and functions requiring magic PTE bits (pXX_devmap) to do the
special DAX reference counting. This then adds some cruft to manage
pgmap references and to catch the 2->1 page refcount transition. All
this just goes away if we manage the page references the same as other
pages (and indeed we already manage DEVICE_PRIVATE and COHERENT pages
the same as normal pages).
So I think to make this work we at least need the mapcounts.
I'm asking this because:
(A) We don't support mixing PUD+PMD mappings yet. I have plans to change
that in the future, but for now you can only map using a single PUD
or by PTEs. I suspect that's good enoug for now for dax fs?
Yep, that's all we support.
(B) As long as we have subpage mapcounts, this prevents vmemmap
optimizations [1]. Is that only used for device-dax for now and are
there no plans to make use of that for fs-dax?
I don't have any plans to. This is purely focussed on refcounting pages
"like normal" so we can get rid of all the DAX special casing.
(C) We managed without so far :)
Indeed, although Christoph has asked repeatedly ([1], [2] and likely
others) that this gets fixed and I finally got sick of it coming up
everytime I need to touch something with ZONE_DEVICE pages :)
Also it removes the need for people to understand the special DAX page
recounting scheme and ends up removing a bunch of cruft as a bonus:
59 files changed, 485 insertions(+), 869 deletions(-)
And that's before I clean up all the pgmap reference handling. It also
removes the pXX_trans_huge and pXX_leaf distinction. So we managed, but
things could be better IMHO.
Having that said, with folio->_large_mapcount things like
folio_mapcount() are no longer terribly slow once we weould PTE-map a
PUD-sized folio.
Also, all ZONE_DEVICE pages should currently be marked PG_reserved,
translating to "don't touch the memmap". I think we might want to
tackle that first.
Ok. I'm keen to get this series finished and I don't quite get the
connection here, what needs to change there?