On 08/21/2012 11:06 PM, Andrea Arcangeli wrote: > On Tue, Aug 21, 2012 at 05:46:39PM +0800, Xiao Guangrong wrote: >> There has a bug in set_pte_at_notify which always set the pte to the >> new page before release the old page in secondary MMU, at this time, >> the process will access on the new page, but the secondary MMU still >> access on the old page, the memory is inconsistent between them >> >> Below scenario shows the bug more clearly: >> >> at the beginning: *p = 0, and p is write-protected by KSM or shared with >> parent process >> >> CPU 0 CPU 1 >> write 1 to p to trigger COW, >> set_pte_at_notify will be called: >> *pte = new_page + W; /* The W bit of pte is set */ >> >> *p = 1; /* pte is valid, so no #PF */ >> >> return back to secondary MMU, then >> the secondary MMU read p, but get: >> *p == 0; >> >> /* >> * !!!!!! >> * the host has already set p to 1, but the secondary >> * MMU still get the old value 0 >> */ >> >> call mmu_notifier_change_pte to release >> old page in secondary MMU > > The KSM usage of it looks safe because it will only establish readonly > ptes with it. Hmm, in KSM code, i found this code in replace_page: set_pte_at_notify(mm, addr, ptep, mk_pte(kpage, vma->vm_page_prot)); It is possible to establish a writable pte, no? > > It seems a problem only for do_wp_page. It wasn't safe to setup > writable ptes with it. I guess we first introduced it for KSM and then > we added it to do_wp_page too by mistake. > > The race window is really tiny, it's unlikely it has ever triggered, > however this one seem to be possible so it's slightly more serious > than the other race you recently found (the previous one in the exit > path I think it was impossible to trigger with KVM). Unfortunately, all these bugs are triggered by test cases. > >> We can fix it by release old page first, then set the pte to the new >> page. >> >> Note, the new page will be firstly used in secondary MMU before it is >> mapped into the page table of the process, but this is safe because it >> is protected by the page table lock, there is no race to change the pte >> >> Signed-off-by: Xiao Guangrong <xiaoguangrong@xxxxxxxxxxxxxxxxxx> >> --- >> include/linux/mmu_notifier.h | 2 +- >> 1 files changed, 1 insertions(+), 1 deletions(-) >> >> diff --git a/include/linux/mmu_notifier.h b/include/linux/mmu_notifier.h >> index 1d1b1e1..8c7435a 100644 >> --- a/include/linux/mmu_notifier.h >> +++ b/include/linux/mmu_notifier.h >> @@ -317,8 +317,8 @@ static inline void mmu_notifier_mm_destroy(struct mm_struct *mm) >> unsigned long ___address = __address; \ >> pte_t ___pte = __pte; \ >> \ >> - set_pte_at(___mm, ___address, __ptep, ___pte); \ >> mmu_notifier_change_pte(___mm, ___address, ___pte); \ >> + set_pte_at(___mm, ___address, __ptep, ___pte); \ >> }) > > If we establish the spte on the new page, what will happen is the same > race in reverse. The fundamental problem is that the first guy that > writes to the "newpage" (guest or host) won't fault again and so it > will fail to serialize against the PT lock. > > CPU0 CPU1 > oldpage[1] == 0 (both guest & host) > oldpage[0] = 1 > trigger do_wp_page > mmu_notifier_change_pte > spte = newpage + writable > guest does newpage[1] = 1 > vmexit > host read oldpage[1] == 0 > pte = newpage + writable (too late) > > I think the fix is to use ptep_clear_flush_notify whenever > set_pte_at_notify will establish a writable pte/spte. If the pte/spte > established by set_pte_at_notify/change_pte is readonly we don't need > to do the ptep_clear_flush_notify instead because when the host will > write to the page that will fault and serialize against the > PT lock (set_pte_at_notify must always run under the PT lock of course). > > How about this: > > ===== >>From 160a0b1b2be9bf96c45b30d9423f8196ecebe351 Mon Sep 17 00:00:00 2001 > From: Andrea Arcangeli <aarcange@xxxxxxxxxx> > Date: Tue, 21 Aug 2012 16:48:11 +0200 > Subject: [PATCH] mmu_notifier: fix race in set_pte_at_notify usage > > Whenever we establish a writable spte with set_pte_at_notify the > ptep_clear_flush before it must be a _notify one that clears the spte > too. > > The fundamental problem is that if the primary MMU that writes to the > "newpage" won't fault again if the pte established by > set_pte_at_notify is writable. And so it will fail to serialize > against the PT lock to wait the set_pte_at_notify to finish > updating all secondary MMUs before the write hits the newpage. > > CPU0 CPU1 > oldpage[1] == 0 (all MMUs) > oldpage[0] = 1 > trigger do_wp_page > take PT lock > ptep_clear_flush (secondary MMUs > still have read access to oldpage) > mmu_notifier_change_pte > pte = newpage + writable (primary MMU can write to > newpage) > host write newpage[1] == 1 (no fault, > failed to serialize against PT lock) > vmenter > guest read oldpage[1] == 0 Why? Why guest can read the old page? Before you set the pte to be writable, mmu_notifier_change_pte is called that all old pages have been released. -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
