Message-ID: <20210507182538.GF26528@arm.com>
Date:   Fri, 7 May 2021 19:25:39 +0100
From:   Catalin Marinas <catalin.marinas@....com>
To:     Steven Price <steven.price@....com>
Cc:     Marc Zyngier <maz@...nel.org>, Will Deacon <will@...nel.org>,
        James Morse <james.morse@....com>,
        Julien Thierry <julien.thierry.kdev@...il.com>,
        Suzuki K Poulose <suzuki.poulose@....com>,
        kvmarm@...ts.cs.columbia.edu, linux-arm-kernel@...ts.infradead.org,
        linux-kernel@...r.kernel.org, Dave Martin <Dave.Martin@....com>,
        Mark Rutland <mark.rutland@....com>,
        Thomas Gleixner <tglx@...utronix.de>, qemu-devel@...gnu.org,
        Juan Quintela <quintela@...hat.com>,
        "Dr. David Alan Gilbert" <dgilbert@...hat.com>,
        Richard Henderson <richard.henderson@...aro.org>,
        Peter Maydell <peter.maydell@...aro.org>,
        Haibo Xu <Haibo.Xu@....com>, Andrew Jones <drjones@...hat.com>
Subject: Re: [PATCH v11 2/6] arm64: kvm: Introduce MTE VM feature

On Thu, May 06, 2021 at 05:15:25PM +0100, Steven Price wrote:
> On 04/05/2021 18:40, Catalin Marinas wrote:
> > On Thu, Apr 29, 2021 at 05:06:41PM +0100, Steven Price wrote:
> > > On 28/04/2021 18:07, Catalin Marinas wrote:
> > > > While the set_pte_at() race on the page flags is somewhat clearer, we
> > > > may still have a race here with the VMM's set_pte_at() if the page is
> > > > mapped as tagged. KVM has its own mmu_lock but it wouldn't be held when
> > > > handling the VMM page tables (well, not always, see below).
> > > > 
> > > > gfn_to_pfn_prot() ends up calling get_user_pages*(). At least the slow
> > > > path (hva_to_pfn_slow()) ends up with FOLL_TOUCH in gup and the VMM pte
> > > > would be set, tags cleared (if PROT_MTE) before the stage 2 pte. I'm not
> > > > sure whether get_user_page_fast_only() does the same.
> > > > 
> > > > The race with an mprotect(PROT_MTE) in the VMM is fine I think as the
> > > > KVM mmu notifier is invoked before set_pte_at() and racing with another
> > > > user_mem_abort() is serialised by the KVM mmu_lock. The subsequent
> > > > set_pte_at() would see the PG_mte_tagged set either by the current CPU
> > > > or by the one it was racing with.
> > > 
> > > Given the changes to set_pte_at() which means that tags are restored from
> > > swap even if !PROT_MTE, the only race I can see remaining is the creation of
> > > new PROT_MTE mappings. As you mention an attempt to change mappings in the
> > > VMM memory space should involve a mmu notifier call which I think serialises
> > > this. So the remaining issue is doing this in a separate address space.
> > > 
> > > So I guess the potential problem is:
> > > 
> > >   * allocate memory MAP_SHARED but !PROT_MTE
> > >   * fork()
> > >   * VM causes a fault in parent address space
> > >   * child does a mprotect(PROT_MTE)
> > > 
> > > With the last two potentially racing. Sadly I can't see a good way of
> > > handling that.
> > 
> > Ah, the mmap lock doesn't help as they are different processes
> > (mprotect() acquires it as a writer).
> > 
> > I wonder whether this is racy even in the absence of KVM. If both parent
> > and child do an mprotect(PROT_MTE), one of them may be reading stale
> > tags for a brief period.
> > 
> > Maybe we should revisit whether shared MTE pages are of any use, though
> > it's an ABI change (not bad if no-one is relying on this). However...
> 
> Shared MTE pages are certainly hard to use correctly (e.g. see the
> discussions with the VMM accessing guest memory). But I guess that boat has
> sailed.

Digging out some old emails (two years ago), the Chrome people may have
found a use for MTE in shared mappings (with memfd_create), though not
sure they took advantage of this yet.

> > Thinking about this, we have a similar problem with the PG_dcache_clean
> > and two processes doing mprotect(PROT_EXEC). One of them could see the
> > flag set and skip the I-cache maintenance while the other executes
> > stale instructions. change_pte_range() could acquire the page lock if
> > the page is VM_SHARED (my preferred core mm fix). It doesn't immediately
> > solve the MTE/KVM case but we could at least take the page lock via
> > user_mem_abort().
> 
> For PG_dcache_clean AFAICS the solution is actually simple: split the test
> and set parts. i.e..:
> 
>  if (!test_bit(PG_dcache_clean, &page->flags)) {
> 	sync_icache_aliases(page_address(page), page_size(page));
> 	set_bit(PG_dcache_clean, &page->flags);
>  }
> 
> There isn't a problem with repeating the sync_icache_aliases() call in the
> case of a race. Or am I missing something?

No, the fix is simple as you said.

> > Or maybe we just document this behaviour as racy both for PROT_EXEC and
> > PROT_MTE mappings and be done with this. The minor issue with PROT_MTE
> > is the potential leaking of tags (it's harder to leak information
> > through the I-cache).
> 
> This is the real issue I see - the race in PROT_MTE case is either a data
> leak (syncing after setting the bit) or data loss (syncing before setting
> the bit).

For a moment I thought an mmap(PROT_MTE, MAP_SHARED) on memfd/tmpfs file
may lead to the same situation but the mmap() itself won't directly
cause allocating the page. The subsequent fault via filemap_map_pages()
seems to take the page lock.

> But without serialising through a spinlock (in mte_sync_tags()) I haven't
> been able to come up with any way of closing the race. But with the change
> to set_pte_at() to call mte_sync_tags() even if the PTE isn't PROT_MTE that
> is likely to seriously hurt performance.

Yeah. We could add another page flag as a lock though I think it should
be the core code that prevents the race.

If we are to do it in the arch code, maybe easier with a custom
ptep_modify_prot_start/end() where we check if it's VM_SHARED and
VM_MTE, take a (big) lock.

In the core code, something like below (well, a partial hack, not tested
and it doesn't handle huge pages but just to give an idea):

diff --git a/mm/mprotect.c b/mm/mprotect.c
index 94188df1ee55..6ba96ff141a6 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -76,14 +76,13 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 		if (pte_present(oldpte)) {
 			pte_t ptent;
 			bool preserve_write = prot_numa && pte_write(oldpte);
+			struct page *page = NULL;
 
 			/*
 			 * Avoid trapping faults against the zero or KSM
 			 * pages. See similar comment in change_huge_pmd.
 			 */
 			if (prot_numa) {
-				struct page *page;
-
 				/* Avoid TLB flush if possible */
 				if (pte_protnone(oldpte))
 					continue;
@@ -114,6 +113,10 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 			}
 
 			oldpte = ptep_modify_prot_start(vma, addr, pte);
+			if (vma->vm_flags & VM_SHARED) {
+				page = vm_normal_page(vma, addr, oldpte);
+				lock_page(page);
+			}
 			ptent = pte_modify(oldpte, newprot);
 			if (preserve_write)
 				ptent = pte_mk_savedwrite(ptent);
@@ -138,6 +141,8 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd,
 				ptent = pte_mkwrite(ptent);
 			}
 			ptep_modify_prot_commit(vma, addr, pte, oldpte, ptent);
+			if (page)
+				unlock_page(page);
 			pages++;
 		} else if (is_swap_pte(oldpte)) {
 			swp_entry_t entry = pte_to_swp_entry(oldpte);

-- 
Catalin

Powered by blists - more mailing lists