Web lists-archives.com

Re: [RFC PATCH v2 11/12] x86/mm/tlb: Use async and inline messages for flushing




> On May 31, 2019, at 2:14 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> 
> On Thu, May 30, 2019 at 11:37 PM Nadav Amit <namit@xxxxxxxxxx> wrote:
>> When we flush userspace mappings, we can defer the TLB flushes, as long
>> the following conditions are met:
>> 
>> 1. No tables are freed, since otherwise speculative page walks might
>>   cause machine-checks.
>> 
>> 2. No one would access userspace before flush takes place. Specifically,
>>   NMI handlers and kprobes would avoid accessing userspace.
> 
> I think I need to ask the big picture question.  When someone calls
> flush_tlb_mm_range() (or the other entry points), if no page tables
> were freed, they want the guarantee that future accesses (initiated
> observably after the flush returns) will not use paging entries that
> were replaced by stores ordered before flush_tlb_mm_range().  We also
> need the guarantee that any effects from any memory access using the
> old paging entries will become globally visible before
> flush_tlb_mm_range().
> 
> I'm wondering if receipt of an IPI is enough to guarantee any of this.
> If CPU 1 sets a dirty bit and CPU 2 writes to the APIC to send an IPI
> to CPU 1, at what point is CPU 2 guaranteed to be able to observe the
> dirty bit?  An interrupt entry today is fully serializing by the time
> it finishes, but interrupt entries are epicly slow, and I don't know
> if the APIC waits long enough.  Heck, what if IRQs are off on the
> remote CPU?  There are a handful of places where we touch user memory
> with IRQs off, and it's (sadly) possible for user code to turn off
> IRQs with iopl().
> 
> I *think* that Intel has stated recently that SMT siblings are
> guaranteed to stop speculating when you write to the APIC ICR to poke
> them, but SMT is very special.
> 
> My general conclusion is that I think the code needs to document what
> is guaranteed and why.

I think I might have managed to confuse you with a bug I made (last minute
bug when I was doing some cleanup). This bug does not affect the performance
much, but it might led you to think that I use the APIC sending as
synchronization.

The idea is not for us to rely on write to ICR as something serializing. The
flow should be as follows:


	CPU0					CPU1

flush_tlb_mm_range()
 __smp_call_function_many()
  [ prepare call_single_data (csd) ]
  [ lock csd ] 
  [ send IPI ]
	(*)
  [ wait for csd to be unlocked ]
					[ interrupt ]
					[ copy csd info to stack ]
					[ csd unlock ]
  [ find csd is unlocked ]
  [ continue (**) ]
					[ flush TLB ]


At (**) the pages might be recycled, written-back to disk, etc. Note that
during (*), CPU0 might do some local TLB flushes, making it very likely that
CSD will be unlocked by the time it gets there.

As you can see, I don’t rely on any special micro-architectural behavior.
The synchronization is done purely in software.

Does it make more sense now?