On Thu, Nov 17, 2022 at 10:12:07AM -0500, Peter Xu wrote:
> Hi, Huacai,
>
> On Thu, Nov 17, 2022 at 12:25:32PM +0800, Huacai Chen wrote:
> > Now {pmd,pte}_mkdirty() set _PAGE_DIRTY bit unconditionally, this causes
> > random segmentation fault after commit 0ccf7f168e17bb7e ("mm/thp: carry
> > over dirty bit when thp splits on pmd").
> >
> > The reason is: when fork(), parent process use pmd_wrprotect() to clear
> > huge page's _PAGE_WRITE and _PAGE_DIRTY (for COW);
>
> Is it safe to drop dirty bit when wr-protect? It means the mm can reclaim
> the page directly assuming the page contains rubbish.
>
> Consider after fork() and memory pressure kicks the kswapd, I don't see
> anything stops the kswapd from recycling the pages and lose the data in
> both processes.
Feel free to ignore this question.. I think I got an answer from Hev (and
I then got a follow up question):
https://lore.kernel.org/all/Y3Z9Zf0jARMOkFBq@x1n/
>
> > then pte_mkdirty() set
> > _PAGE_DIRTY as well as _PAGE_MODIFIED while splitting dirty huge pages;
> > once _PAGE_DIRTY is set, there will be no tlb modify exception so the COW
> > machanism fails; and at last memory corruption occurred between parent
> > and child processes.
> >
> > So, we should set _PAGE_DIRTY only when _PAGE_WRITE is set in {pmd,pte}_
> > mkdirty().
> >
> > Cc: stable@xxxxxxxxxxxxxxx
> > Cc: Peter Xu <peterx@xxxxxxxxxx>
> > Signed-off-by: Huacai Chen <chenhuacai@xxxxxxxxxxx>
> > ---
> > Note: CC sparc maillist because they have similar issues.
>
> I also had a look on sparc64, it seems to not do the same as loongarch
> here (not removing dirty in wr-protect):
>
> static inline pmd_t pmd_wrprotect(pmd_t pmd)
> {
> pte_t pte = __pte(pmd_val(pmd));
>
> pte = pte_wrprotect(pte);
>
> return __pmd(pte_val(pte));
> }
>
> static inline pte_t pte_wrprotect(pte_t pte)
> {
> unsigned long val = pte_val(pte), tmp;
>
> __asm__ __volatile__(
> "\n661: andn %0, %3, %0\n"
> " nop\n"
> "\n662: nop\n"
> " nop\n"
> " .section .sun4v_2insn_patch, \"ax\"\n"
> " .word 661b\n"
> " sethi %%uhi(%4), %1\n"
> " sllx %1, 32, %1\n"
> " .word 662b\n"
> " or %1, %%lo(%4), %1\n"
> " andn %0, %1, %0\n"
> " .previous\n"
> : "=r" (val), "=r" (tmp)
> : "0" (val), "i" (_PAGE_WRITE_4U | _PAGE_W_4U),
> "i" (_PAGE_WRITE_4V | _PAGE_W_4V));
>
> return __pte(val);
> }
(Same here; I just overlooked what does _PAGE_W_4U meant..)
>
> >
> > arch/loongarch/include/asm/pgtable.h | 8 ++++++--
> > 1 file changed, 6 insertions(+), 2 deletions(-)
> >
> > diff --git a/arch/loongarch/include/asm/pgtable.h b/arch/loongarch/include/asm/pgtable.h
> > index 946704bee599..debbe116f105 100644
> > --- a/arch/loongarch/include/asm/pgtable.h
> > +++ b/arch/loongarch/include/asm/pgtable.h
> > @@ -349,7 +349,9 @@ static inline pte_t pte_mkclean(pte_t pte)
> >
> > static inline pte_t pte_mkdirty(pte_t pte)
> > {
> > - pte_val(pte) |= (_PAGE_DIRTY | _PAGE_MODIFIED);
> > + pte_val(pte) |= _PAGE_MODIFIED;
> > + if (pte_val(pte) & _PAGE_WRITE)
> > + pte_val(pte) |= _PAGE_DIRTY;
>
> I'm not sure whether mm has rule to always set write bit then set dirty
> bit, need to be careful here because the outcome may differ when use:
>
> pte_mkdirty(pte_mkwrite(pte))
> (expected)
>
> VS:
>
> pte_mkwrite(pte_mkdirty(pte))
> (dirty not set)
>
> I had a feeling I miss some arch-specific details here on why loongarch
> needs such implementation, but I can't quickly tell.
After a closer look I think it's fine for loongarch as pte_mkwrite will
also set the dirty bit unconditionally, so at least the two ways will still
generate the same pte (DIRTY+MODIFIED+WRITE).
But this whole thing is still confusing to me. It'll still be great if
anyone can help explain why the _DIRTY cannot be set only in pte_mkwrite()
if that's the solo place in charge of "whether the pte is writable".
The other follow up question is: how do we mark "this pte contains valid
data" (the common definition of "dirty bit"), while "this pte is not
writable" on loongarch?
It can happen when we're installing a page with non-zero data meanwhile
wr-protected. That's actually a valid case for userfaultfd wr-protect mode
where user specified UFFDIO_COPY ioctl with flag UFFDIO_COPY_MODE_WP, where
we'll install a non-zero page from user buffer but don't grant write bit.
>From code-wise, I think it can be done currently with this on loongarch:
pte_wrprotect(pte_mkwrite(pte_mkdirty(pte)))
Where pte_wrprotect(pte_mkwrite(pte)) is not a no-op but applying MODIFIED.
While on many other archs it'll be as simple as:
pte_mkdirty(pte)
But that's really error-prone and not obvious.
Copying Hev too.
Thanks,
--
Peter Xu
