On Thu, Nov 17, 2022 at 06:34AM -0800, Dave Hansen wrote:
> On 11/17/22 05:58, Marco Elver wrote:
> > [ 0.663761] WARNING: CPU: 0 PID: 0 at arch/x86/include/asm/kfence.h:46 kfence_protect+0x7b/0x120
> > [ 0.664033] WARNING: CPU: 0 PID: 0 at mm/kfence/core.c:234 kfence_protect+0x7d/0x120
> > [ 0.664465] kfence: kfence_init failed
>
> Any chance you could add some debugging and figure out what actually
> made kfence call over? Was it the pte or the level?
>
> if (WARN_ON(!pte || level != PG_LEVEL_4K))
> return false;
>
> I can see how the thing you bisected to might lead to a page table not
> being split, which could mess with the 'level' check.
Yes - it's the 'level != PG_LEVEL_4K'.
We do actually try to split the pages in arch_kfence_init_pool() (above
this function) - so with "x86/mm: Inhibit _PAGE_NX changes from
cpa_process_alias()" this somehow fails...
> Also, is there a reason this code is mucking with the page tables
> directly? It seems, uh, rather wonky. This, for instance:
>
> > if (protect)
> > set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
> > else
> > set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
> >
> > /*
> > * Flush this CPU's TLB, assuming whoever did the allocation/free is
> > * likely to continue running on this CPU.
> > */
> > preempt_disable();
> > flush_tlb_one_kernel(addr);
> > preempt_enable();
>
> Seems rather broken. I assume the preempt_disable() is there to get rid
> of some warnings. But, there is nothing I can see to *keep* the CPU
> that did the free from being different from the one where the TLB flush
> is performed until the preempt_disable(). That makes the
> flush_tlb_one_kernel() mostly useless.
>
> Is there a reason this code isn't using the existing page table
> manipulation functions and tries to code its own? What prevents it from
> using something like the attached patch?
Yes, see the comment below - it's to avoid the IPIs and TLB shoot-downs,
because KFENCE _can_ tolerate the inaccuracy even if we hit the wrong
TLB or other CPUs' TLBs aren't immediately flushed - we trade a few
false negatives for minimizing performance impact.
> diff --git a/arch/x86/include/asm/kfence.h b/arch/x86/include/asm/kfence.h
> index ff5c7134a37a..5cdb3a1f3995 100644
> --- a/arch/x86/include/asm/kfence.h
> +++ b/arch/x86/include/asm/kfence.h
> @@ -37,34 +37,13 @@ static inline bool arch_kfence_init_pool(void)
> return true;
> }
>
> -/* Protect the given page and flush TLB. */
> static inline bool kfence_protect_page(unsigned long addr, bool protect)
> {
> - unsigned int level;
> - pte_t *pte = lookup_address(addr, &level);
> -
> - if (WARN_ON(!pte || level != PG_LEVEL_4K))
> - return false;
> -
> - /*
> - * We need to avoid IPIs, as we may get KFENCE allocations or faults
> - * with interrupts disabled. Therefore, the below is best-effort, and
> - * does not flush TLBs on all CPUs. We can tolerate some inaccuracy;
> - * lazy fault handling takes care of faults after the page is PRESENT.
> - */
> -
^^ See this comment. Additionally there's a real performance concern,
and the inaccuracy is something that we deliberately accept.
> if (protect)
> - set_pte(pte, __pte(pte_val(*pte) & ~_PAGE_PRESENT));
> + set_memory_np(addr, addr + PAGE_SIZE);
> else
> - set_pte(pte, __pte(pte_val(*pte) | _PAGE_PRESENT));
> + set_memory_p(addr, addr + PAGE_SIZE);
Isn't this going to do tons of IPIs and shoot down other CPU's TLBs?
KFENCE shouldn't incur this overhead on large machines with >100 CPUs if
we can avoid it.
What does "x86/mm: Inhibit _PAGE_NX changes from cpa_process_alias()" do
that suddenly makes all this fail?
What solution do you prefer that both fixes the issue and avoids the
IPIs?
Thanks,
-- Marco
