On 2020/3/24 6:52, Jason Gunthorpe wrote: > On Mon, Mar 23, 2020 at 01:35:07PM -0700, Mike Kravetz wrote: >> On 3/23/20 11:07 AM, Jason Gunthorpe wrote: >>> On Mon, Mar 23, 2020 at 10:27:48AM -0700, Mike Kravetz wrote: >>> >>>>> pgd = pgd_offset(mm, addr); >>>>> - if (!pgd_present(*pgd)) >>>>> + if (!pgd_present(READ_ONCE(*pgd))) >>>>> return NULL; >>>>> p4d = p4d_offset(pgd, addr); >>>>> - if (!p4d_present(*p4d)) >>>>> + if (!p4d_present(READ_ONCE(*p4d))) >>>>> return NULL; >>>>> >>>>> pud = pud_offset(p4d, addr); >>>> >>>> One would argue that pgd and p4d can not change from present to !present >>>> during the execution of this code. To me, that seems like the issue which >>>> would cause an issue. Of course, I could be missing something. >>> >>> This I am not sure of, I think it must be true under the read side of >>> the mmap_sem, but probably not guarenteed under RCU.. >>> >>> In any case, it doesn't matter, the fact that *p4d can change at all >>> is problematic. Unwinding the above inlines we get: >>> >>> p4d = p4d_offset(pgd, addr) >>> if (!p4d_present(*p4d)) >>> return NULL; >>> pud = (pud_t *)p4d_page_vaddr(*p4d) + pud_index(address); >>> >>> According to our memory model the compiler/CPU is free to execute this >>> as: >>> >>> p4d = p4d_offset(pgd, addr) >>> p4d_for_vaddr = *p4d; >>> if (!p4d_present(*p4d)) >>> return NULL; >>> pud = (pud_t *)p4d_page_vaddr(p4d_for_vaddr) + pud_index(address); >>> >> >> Wow! How do you know this? You don't need to answer :) > > It says explicitly in Documentation/memory-barriers.txt - see > section COMPILER BARRIER: > > (*) The compiler is within its rights to reorder loads and stores > to the same variable, and in some cases, the CPU is within its > rights to reorder loads to the same variable. This means that > the following code: > > a[0] = x; > a[1] = x; > > Might result in an older value of x stored in a[1] than in a[0]. > > It also says READ_ONCE puts things in program order, but we don't use > READ_ONCE inside pud_offset(), so it doesn't help us. > > Best answer is to code things so there is exactly one dereference of > the pointer protected by READ_ONCE. Very clear to read, very safe. > > Maybe Longpeng can rework the patch around these principles? > Thanks Jason and Mike, I learn a lot from your analysis. So... the patch should like this ? @@ -4909,29 +4909,33 @@ pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr, unsigned long sz) { pgd_t *pgd; - p4d_t *p4d; - pud_t *pud; - pmd_t *pmd; + p4d_t *p4g, p4d_entry; + pud_t *pud, pud_entry; + pmd_t *pmd, pmd_entry; pgd = pgd_offset(mm, addr); if (!pgd_present(*pgd)) return NULL; - p4d = p4d_offset(pgd, addr); - if (!p4d_present(*p4d)) + + p4g = p4d_offset(pgd, addr); + p4d_entry = READ_ONCE(*p4g); + if (!p4d_present(p4d_entry)) return NULL; - pud = pud_offset(p4d, addr); - if (sz != PUD_SIZE && pud_none(*pud)) + pud = pud_offset(&p4d_entry, addr); + pud_entry = READ_ONCE(*pud); + if (sz != PUD_SIZE && pud_none(pud_entry)) return NULL; /* hugepage or swap? */ - if (pud_huge(*pud) || !pud_present(*pud)) + if (pud_huge(pud_entry) || !pud_present(pud_entry)) return (pte_t *)pud; - pmd = pmd_offset(pud, addr); - if (sz != PMD_SIZE && pmd_none(*pmd)) + pmd = pmd_offset(&pud_entry, addr); + pmd_entry = READ_ONCE(*pmd); + if (sz != PMD_SIZE && pmd_none(pmd_entry)) return NULL; /* hugepage or swap? */ - if (pmd_huge(*pmd) || !pmd_present(*pmd)) + if (pmd_huge(pmd_entry) || !pmd_present(pmd_entry)) return (pte_t *)pmd; > Also I wonder if the READ_ONCE(*pmdp) is OK. gup_pmd_range() uses it, > but I can't explain why it shouldn't be pmd_read_atomic(). > >>> In the case where p4 goes from !present -> present (ie >>> handle_mm_fault()): >>> >>> p4d_for_vaddr == p4d_none, and p4d_present(*p4d) == true, meaning the >>> p4d_page_vaddr() will crash. >>> >>> Basically the problem here is not just missing READ_ONCE, but that the >>> p4d is read multiple times at all. It should be written like gup_fast >>> does, to guarantee a single CPU read of the unstable data: >>> >>> p4d = READ_ONCE(*p4d_offset(pgdp, addr)); >>> if (!p4d_present(p4)) >>> return NULL; >>> pud = pud_offset(&p4d, addr); >>> >>> At least this is what I've been able to figure out :\ >> >> In that case, I believe there are a bunch of similar routines with this issue. > > Yes, my look around page walk related users makes me come to a similar > worry. > > Fortunately, I think this is largely theoretical as most likely the > compiler will generate a single store for these coding patterns. > > That said, there have been bugs in the past, see commit 26c191788f18 > ("mm: pmd_read_atomic: fix 32bit PAE pmd walk vs pmd_populate SMP race > condition") which is significantly related to the compiler lifting a > load inside pte_offset to before the required 'if (pmd_*)' checks. > >> For this patch, I was primarily interested in seeing the obvious >> multiple dereferences in C fixed up. This is above and beyond that! >> :) > > Well, I think it is worth solving the underlying problem > properly. Otherwise we get weird solutions to data races like > pmd_trans_unstable()... > > Jason > . > --- Regards, Longpeng(Mike)