On 9/27/19 7:55 AM, Thomas Hellström (VMware) wrote:
On 9/27/19 12:20 AM, Linus Torvalds wrote:
On Thu, Sep 26, 2019 at 1:55 PM Thomas Hellström (VMware)
<thomas_os@xxxxxxxxxxxx> wrote:
Well, we're working on supporting huge puds and pmds in the graphics
VMAs, although in the write-notify cases we're looking at here, we
would
probably want to split them down to PTE level.
Well, that's what the existing walker code does if you don't have that
"pud_entry()" callback.
That said, I assume you would *not* want to do that if the huge
pud/pmd is already clean and read-only, but just continue.
So you may want to have a special pud_entry() that handles that case.
Eventually. Maybe. Although honestly, if you're doing dirty tracking,
I doubt it makes much sense to use largepages.
The approach we're looking at in this case is to keep huge entries
write-protected and split them in the wp_huge_xxx() code's fallback
path with the mmap_sem held. This means that there will actually be
huge entries in the page-walking code soon, but as you say, only
entries that we want to ignore and not split. So we'd also need a way
to avoid the pagewalk splitting for the situation when someone faults
a huge entry in just before the call to split_huge_xxx.
Looking at zap_pud_range() which when called from unmap_mapping_pages()
uses identical locking (no mmap_sem), it seems we should be able to get
away with i_mmap_lock(), making sure the whole page table doesn't
disappear under us. So it's not clear to me why the mmap_sem is
strictly
needed here. Better to sort those restrictions out now rather than when
huge entries start appearing.
zap_pud_range()actually does have that
VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
exactly for the case where it might have to split the pud entry.
Yes. My take on this is that locking the PUD ptl can be done either
with the mmap_sem or the i_mmap_lock if present and that we should
update the asserts in xxx_trans_huge_lock to reflect that. But when
actually splitting transhuge pages you don't want to race with
khugepaged, so you need the mmap_sem. For the graphics VMAs
(MIXEDMAP), khugepaged never touches them. Yet.
It's why they've never gotten translated to use the generic walker code.
OK. Yes there are a number of various specialized pagewalks all over
the mm code.
But another thing that worries me is that the page-table modifications
that happen in the callback use functionality that is not guaranteed
to be exported, and that mm people don't want them to be exported
because you don't want the drivers to go hacking around in page
tables, which means that the two callbacks used here would need to be
a set of core helpers anyway.
So I figure what I would end up with would actually be an extern
__walk_page_range anyway, and slightly modified asserts. Do you think
that could be acceptible?
Actually, I'll give your original suggestion a try and see what I come
up with.
Thanks,
Thomas
