On 19.11.24 05:49, John Hubbard wrote:
The recent addition of "pofs" (pages or folios) handling to gup has a
flaw: it assumes that unpin_user_pages() handles NULL pages in the
pages** array. That's not the case, as I discovered when I ran on a new
configuration on my test machine.
Fix this by skipping NULL pages in unpin_user_pages(), just like
unpin_folios() already does.
Details: when booting on x86 with "numa=fake=2 movablecore=4G" on Linux
6.12, and running this:
tools/testing/selftests/mm/gup_longterm
...I get the following crash:
BUG: kernel NULL pointer dereference, address: 0000000000000008
RIP: 0010:sanity_check_pinned_pages+0x3a/0x2d0
...
Call Trace:
<TASK>
? __die_body+0x66/0xb0
? page_fault_oops+0x30c/0x3b0
? do_user_addr_fault+0x6c3/0x720
? irqentry_enter+0x34/0x60
? exc_page_fault+0x68/0x100
? asm_exc_page_fault+0x22/0x30
? sanity_check_pinned_pages+0x3a/0x2d0
unpin_user_pages+0x24/0xe0
check_and_migrate_movable_pages_or_folios+0x455/0x4b0
__gup_longterm_locked+0x3bf/0x820
? mmap_read_lock_killable+0x12/0x50
? __pfx_mmap_read_lock_killable+0x10/0x10
pin_user_pages+0x66/0xa0
gup_test_ioctl+0x358/0xb20
__se_sys_ioctl+0x6b/0xc0
do_syscall_64+0x7b/0x150
entry_SYSCALL_64_after_hwframe+0x76/0x7e
Fixes: 94efde1d1539 ("mm/gup: avoid an unnecessary allocation call for FOLL_LONGTERM cases")
Cc: David Hildenbrand <david@xxxxxxxxxx>
Cc: Oscar Salvador <osalvador@xxxxxxx>
Cc: Vivek Kasireddy <vivek.kasireddy@xxxxxxxxx>
Cc: Dave Airlie <airlied@xxxxxxxxxx>
Cc: Gerd Hoffmann <kraxel@xxxxxxxxxx>
Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
Cc: Christoph Hellwig <hch@xxxxxxxxxxxxx>
Cc: Jason Gunthorpe <jgg@xxxxxxxxxx>
Cc: Peter Xu <peterx@xxxxxxxxxx>
Cc: Arnd Bergmann <arnd@xxxxxxxx>
Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
Cc: Dongwon Kim <dongwon.kim@xxxxxxxxx>
Cc: Hugh Dickins <hughd@xxxxxxxxxx>
Cc: Junxiao Chang <junxiao.chang@xxxxxxxxx>
Cc: <stable@xxxxxxxxxxxxxxx>
Signed-off-by: John Hubbard <jhubbard@xxxxxxxxxx>
---
Hi,
I got a nasty shock when I tried out a new test machine setup last
night--I wish I'd noticed the problem earlier! But anyway, this should
make it all better...
I've asked Greg K-H to hold off on including commit 94efde1d1539
("mm/gup: avoid an unnecessary allocation call for FOLL_LONGTERM cases")
in linux-stable (6.11.y), but if this fix-to-the-fix looks good, then
maybe both fixes can ultimately end up in stable.
Ouch!
thanks,
John Hubbard
mm/gup.c | 17 +++++++++++++++--
1 file changed, 15 insertions(+), 2 deletions(-)
diff --git a/mm/gup.c b/mm/gup.c
index ad0c8922dac3..6e417502728a 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -52,7 +52,12 @@ static inline void sanity_check_pinned_pages(struct page **pages,
*/
for (; npages; npages--, pages++) {
struct page *page = *pages;
- struct folio *folio = page_folio(page);
+ struct folio *folio;
+
+ if (!page)
+ continue;
+
+ folio = page_folio(page);
if (is_zero_page(page) ||
!folio_test_anon(folio))
@@ -248,9 +253,14 @@ static inline struct folio *gup_folio_range_next(struct page *start,
static inline struct folio *gup_folio_next(struct page **list,
unsigned long npages, unsigned long i, unsigned int *ntails)
{
- struct folio *folio = page_folio(list[i]);
+ struct folio *folio;
unsigned int nr;
+ if (!list[i])
+ return NULL;
+
I don't particularly enjoy returning NULL here, if we don't teach the
other users of that function about that possibility. There are two other
users.
Also: we are not setting "ntails" to 1. I think the callers uses that as
"nr" to advance npages. So the caller has to make sure to set "nr = 1"
in case it sees "NULL".
Alternatively ...
+ folio = page_folio(list[i]);
+
for (nr = i + 1; nr < npages; nr++) {
if (page_folio(list[nr]) != folio)
break;
@@ -410,6 +420,9 @@ void unpin_user_pages(struct page **pages, unsigned long npages)
sanity_check_pinned_pages(pages, npages);
for (i = 0; i < npages; i += nr) {
... handle it here
if (!pages[i]) {
nr = 1;
continue;
}
No strong opinion. But I think we should either update all callers to
deal with returning NULL from this function, and set "nr = 1".
--
Cheers,
David / dhildenb
