On Tue, 24 Sept 2024 at 17:17, Gao Xiang <hsiangkao@xxxxxxxxxxxxxxxxx> wrote:
>
> Just side note: I think `generic_file_vm_ops` already prepares this
> feature, so generic_file_mmap users also have fault around behaviors.
Hmm. Maybe. But it doesn't really change the fundamental issue - the
code in question seems to be just *random*.
And I mean that in a very real and very immediate sense: the
fault-around code and filemap_map_pages() only maps in pages that are
uptodate, so it literally DEPENDS ON TIMING whether some previous IO
has completed or not, and thus on whether the page fault is handled by
the fault-around in filemap_map_pages() or by the filemap_fault()
code.
In other words - I think this is all completely broken.
Put another way: explain to me why random IO timing details should
matter for the whether we do __filemap_fsnotify_fault() on a page
fault or not?
So no. I'm not taking this pull request. It makes absolutely zero
sense to me, and I don't think it has sane semantics. The argument
that it is already used by people is not an argument.
The new fsnotify hooks need to make SENSE - not be in random locations
that give some kind of random data.
Linus
