Re: [PATCH v2 0/6] x86/alternatives: text_poke() fixes

[Andy Lutomirski <luto@amacapital.net>]

> On Sep 6, 2018, at 10:58 AM, Nadav Amit <namit@vmware.com> wrote:
> 
> at 10:17 AM, Peter Zijlstra <peterz@infradead.org> wrote:
> 
>>> On Thu, Sep 06, 2018 at 05:01:25PM +0000, Nadav Amit wrote:
>>> In addition, there might be a couple of issues with your fix:
>> 
>> It boots on my box ;-)
>> 
>>> 1. __set_pte_vaddr() is not used exclusive by set_fixmap(). This means
>>> the warning might be wrong, but also means that these code patches (Xen’s
>>> set_pte_mfn(), CPU-entry-area setup) needs to be checked. And as you said
>>> before, someone might use this function for other purposes as well.
>> 
>> CEA is fine, that actually needs it too.
>> 
>> The one thing I missed out on earlier is the below chunk, that is no
>> longer needed now that cea_set_pte() actually does the right thing.
>> 
>> diff --git a/arch/x86/events/intel/ds.c b/arch/x86/events/intel/ds.c
>> index b7b01d762d32..14ad97fa0749 100644
>> --- a/arch/x86/events/intel/ds.c
>> +++ b/arch/x86/events/intel/ds.c
>> @@ -293,12 +293,6 @@ static void ds_update_cea(void *cea, void *addr, size_t size, pgprot_t prot)
>> preempt_disable();
>> for (; msz < size; msz += PAGE_SIZE, pa += PAGE_SIZE, cea += PAGE_SIZE)
>> cea_set_pte(cea, pa, prot);
>> -
>> - /*
>> -  * This is a cross-CPU update of the cpu_entry_area, we must shoot down
>> -  * all TLB entries for it.
>> -  */
>> - flush_tlb_kernel_range(start, start + size);
>> preempt_enable();
>> }
>> 
>> @@ -310,8 +304,6 @@ static void ds_clear_cea(void *cea, size_t size)
>> preempt_disable();
>> for (; msz < size; msz += PAGE_SIZE, cea += PAGE_SIZE)
>> cea_set_pte(cea, 0, PAGE_NONE);
>> -
>> - flush_tlb_kernel_range(start, start + size);
>> preempt_enable();
>> }
>> 
>> 
>>> 2. Printing the virtual address can break KASLR.
>> 
>> Local KASLR is a myth.. but sure, we can fix the print.
>> 
>>> 3. The WARN() can introduce some overhead since checking if IRQs are
>>> disabled takes considerably long time. Perhaps VM_WARN() or something is
>>> better. I realize this code-path is not on the hot-path though...
>> 
>> Yeah, if it triggers you have bigger problems. We can make it
>> WARN_ONCE() I suppose.
>> 
>>> 4. I guess flush_tlb_kernel_range() should also have something like
>>> VM_WARN_ON(irqs_disabled()), just as an additional general sanity check.
>> 
>> It has, it's hidden in kernel/smp.c:smp_call_function_many().
> 
> Right. Thanks.
> 
>> 
>>> Let me know if you want me to make these changes and include your patch in
>>> the set.
>> 
>> The set is no longer needed. text_poke() is fine and correct with this
>> one patch.
> 
> It depends what security you want. Some may consider even the short
> time-window in which the kernel code is writable from other cores as
> insufficient for security.
> 
> In addition, the set removes the need for remote TLB shootdowns that
> text_poke() - with this fix - requires.
> 

I’m personally in favor of not needing a global broadcast flush to install kprobes.