Re: [PATCH v2 6/8] mm: prevent get_user_pages() from overflowing page refcount

From: Ajay Kaher
Date: Thu Oct 17 2019 - 12:29:04 EST



ïOn 09/10/19, 6:43 PM, "Vlastimil Babka" <vbabka@xxxxxxx> wrote:

>> Reported-by: Jann Horn <jannh@xxxxxxxxxx>
>> Acked-by: Matthew Wilcox <willy@xxxxxxxxxxxxx>
>> Cc: stable@xxxxxxxxxx
>> Signed-off-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
>> [ 4.4.y backport notes:
>> Ajay: Added local variable 'err' with-in follow_hugetlb_page()
>> from 2be7cfed995e, to resolve compilation error
>> Srivatsa: Replaced call to get_page_foll() with try_get_page_foll() ]
>> Signed-off-by: Srivatsa S. Bhat (VMware) <srivatsa@xxxxxxxxxxxxx>
>> Signed-off-by: Ajay Kaher <akaher@xxxxxxxxxx>
>> ---
>> mm/gup.c | 43 ++++++++++++++++++++++++++++++++-----------
>> mm/hugetlb.c | 16 +++++++++++++++-
>> 2 files changed, 47 insertions(+), 12 deletions(-)
>
> This seems to have the same issue as the 4.9 stable version [1], in not
> touching the arch-specific gup.c variants.
>
> [1]
> https://lore.kernel.org/lkml/6650323f-dbc9-f069-000b-f6b0f941a065@xxxxxxx/

Thanks Vlastimil for highlighting this here.

Yes, arch-specific gup.c variants also need to handle not only for 4.4.y,
however it should be handled till 4.19.y. I believe it's better to start
from 4.19.y and then backport those changes till 4.4.y.

Affected areas of gup.c (where page->count have been used) are:
#1: get_page() used in these files and this is safe as
it's defined in mm.h (here it's already taken care of)
#2: get_head_page_multiple() has following:
VM_BUG_ON_PAGE(page_count(page) == 0, page);
Need to change this to:
VM_BUG_ON_PAGE(page_ref_zero_or_close_to_overflow(page), page);
#3: Some of the files have used page_cache_get_speculative(),
page_cache_add_speculative() with combination of compound_head(),
this scenario needs to be handled as it was handled here:
https://lore.kernel.org/stable/1570581863-12090-7-git-send-email-akaher@xxxxxxxxxx/

Please share with me any suggestions or patches if you have already
worked on this.

Could we handle arch-specific gup.c in different patch sets and
let these patches to merge to 4.4.y?

- Ajay