On Thu, 2020-02-06 at 14:33 +0100, David Hildenbrand wrote:
> On 06.02.20 14:17, Qian Cai wrote:
> > page->flags could be accessed concurrently as noticied by KCSAN,
> >
> > BUG: KCSAN: data-race in page_cpupid_xchg_last / put_page
> >
> > write (marked) to 0xfffffc0d48ec1a00 of 8 bytes by task 91442 on cpu 3:
> > page_cpupid_xchg_last+0x51/0x80
> > page_cpupid_xchg_last at mm/mmzone.c:109 (discriminator 11)
> > wp_page_reuse+0x3e/0xc0
> > wp_page_reuse at mm/memory.c:2453
> > do_wp_page+0x472/0x7b0
> > do_wp_page at mm/memory.c:2798
> > __handle_mm_fault+0xcb0/0xd00
> > handle_pte_fault at mm/memory.c:4049
> > (inlined by) __handle_mm_fault at mm/memory.c:4163
> > handle_mm_fault+0xfc/0x2f0
> > handle_mm_fault at mm/memory.c:4200
> > do_page_fault+0x263/0x6f9
> > do_user_addr_fault at arch/x86/mm/fault.c:1465
> > (inlined by) do_page_fault at arch/x86/mm/fault.c:1539
> > page_fault+0x34/0x40
> >
> > read to 0xfffffc0d48ec1a00 of 8 bytes by task 94817 on cpu 69:
> > put_page+0x15a/0x1f0
> > page_zonenum at include/linux/mm.h:923
> > (inlined by) is_zone_device_page at include/linux/mm.h:929
> > (inlined by) page_is_devmap_managed at include/linux/mm.h:948
> > (inlined by) put_page at include/linux/mm.h:1023
> > wp_page_copy+0x571/0x930
> > wp_page_copy at mm/memory.c:2615
> > do_wp_page+0x107/0x7b0
> > __handle_mm_fault+0xcb0/0xd00
> > handle_mm_fault+0xfc/0x2f0
> > do_page_fault+0x263/0x6f9
> > page_fault+0x34/0x40
> >
> > Reported by Kernel Concurrency Sanitizer on:
> > CPU: 69 PID: 94817 Comm: systemd-udevd Tainted: G W O L 5.5.0-next-20200204+ #6
> > Hardware name: HPE ProLiant DL385 Gen10/ProLiant DL385 Gen10, BIOS A40 07/10/2019
> >
> > Both the read and write are done only with the non-exclusive mmap_sem
> > held. Since the read will check for specific bits (up to three bits for
> > now) in the flag, load tearing could in theory trigger a logic bug.
> >
> > To fix it, it could introduce put_page_lockless() in those places but
> > that could be an overkill, and difficult to use. Thus, just add
> > READ_ONCE() for the read in page_zonenum() for now where it should not
> > affect the performance and correctness with a small trade-off that
> > compilers might generate less efficient optimization in some places.
> >
> > Signed-off-by: Qian Cai <[email protected]>
> > ---
> > include/linux/mm.h | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/include/linux/mm.h b/include/linux/mm.h
> > index 52269e56c514..f8529aa971c0 100644
> > --- a/include/linux/mm.h
> > +++ b/include/linux/mm.h
> > @@ -920,7 +920,7 @@ vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
> >
> > static inline enum zone_type page_zonenum(const struct page *page)
> > {
> > - return (page->flags >> ZONES_PGSHIFT) & ZONES_MASK;
> > + return (READ_ONCE(page->flags) >> ZONES_PGSHIFT) & ZONES_MASK;
>
> I can understand why other bits/flags might change, but not the zone
> number? Nobody should be changing that without heavy locking (out of
> memory hot(un)plug code). Or am I missing something? Can load tearing
> actually produce an issue if these 3 bits will never change?
>
I think you are right in this case that the writer page_cpupid_xchg_last() will
not change those 3 bits. I don't have an answer if other writers might race
here. Until someone could chime in probably from ZONE_DEVICE side to prove that
it otherwise, I'll just update the previous patch to mark it as a harmless data
race.