Dave Chinner [off-list ref] writes:

On Thu, Jun 27, 2024 at 10:54:15AM +1000, Alistair Popple wrote:

quoted

FS DAX pages have always maintained their own page reference counts
without following the normal rules for page reference counting. In
particular pages are considered free when the refcount hits one rather
than zero and refcounts are not added when mapping the page.

Tracking this requires special PTE bits (PTE_DEVMAP) and a secondary
mechanism for allowing GUP to hold references on the page (see
get_dev_pagemap). However there doesn't seem to be any reason why FS
DAX pages need their own reference counting scheme.

By treating the refcounts on these pages the same way as normal pages
we can remove a lot of special checks. In particular pXd_trans_huge()
becomes the same as pXd_leaf(), although I haven't made that change
here. It also frees up a valuable SW define PTE bit on architectures
that have devmap PTE bits defined.

It also almost certainly allows further clean-up of the devmap managed
functions, but I have left that as a future improvment.

This is an update to the original RFC rebased onto v6.10-rc5. Unlike
the original RFC it passes the same number of ndctl test suite
(https://github.com/pmem/ndctl) tests as my current development
environment does without these patches.

I strongly suggest running fstests on pmem devices with '-o
dax=always' mount options to get much more comprehensive fsdax test
coverage. That exercises a lot of the weird mmap corner cases that
cause problems so it would be good to actually test that nothing new
got broken in FSDAX by this patchset.

Thanks Dave, I will do that and report back. I suspect it will turn up
something, given Dan was seeing a crash with these patches.

 - Alistair

-Dave.