Re: [PATCH net-next 1/2] tcp: add TCP_ZEROCOPY_RECEIVE support for zerocopy receive

[Andy Lutomirski <luto@kernel.org>]

On Wed, Apr 25, 2018 at 9:04 AM, Matthew Wilcox <willy@infradead.org> wrote:
> On Wed, Apr 25, 2018 at 06:01:02AM -0700, Eric Dumazet wrote:
>> On 04/24/2018 11:28 PM, Christoph Hellwig wrote:
>> > On Tue, Apr 24, 2018 at 10:27:21PM -0700, Eric Dumazet wrote:
>> >> When adding tcp mmap() implementation, I forgot that socket lock
>> >> had to be taken before current->mm->mmap_sem. syzbot eventually caught
>> >> the bug.
>> >>
>> >> Since we can not lock the socket in tcp mmap() handler we have to
>> >> split the operation in two phases.
>> >>
>> >> 1) mmap() on a tcp socket simply reserves VMA space, and nothing else.
>> >>   This operation does not involve any TCP locking.
>> >>
>> >> 2) setsockopt(fd, IPPROTO_TCP, TCP_ZEROCOPY_RECEIVE, ...) implements
>> >>  the transfert of pages from skbs to one VMA.
>> >>   This operation only uses down_read(&current->mm->mmap_sem) after
>> >>   holding TCP lock, thus solving the lockdep issue.
>> >>
>> >> This new implementation was suggested by Andy Lutomirski with great details.
>> >
>> > Thanks, this looks much more sensible to me.
>> >
>>
>> Thanks Christoph
>>
>> Note the high cost of zap_page_range(), needed to avoid -EBUSY being returned
>> from vm_insert_page() the second time TCP_ZEROCOPY_RECEIVE is used on one VMA.
>>
>> Ideally a vm_replace_page() would avoid this cost ?
>
> If you don't zap the page range, any of the CPUs in the system where
> any thread in this task have ever run may have a TLB entry pointing to
> this page ... if the page is being recycled into the page allocator,
> then that page might end up as a slab page or page table or page cache
> while the other CPU still have access to it.

Indeed.  This is one of the reasons that Linus has generally been
quite vocal that he doesn't like MMU-based zerocopy schemes.

>
> You could hang onto the page until you've built up a sufficiently large
> batch, then bulk-invalidate all of the TLB entries, but we start to get
> into weirdnesses on different CPU architectures.

The existing mmu_gather code should already handle this at least
moderately well.  If it's not, then it should be fixed.

On x86, there is no operation to flush a range of addresses.  You can
flush one address or you can flush all of them.  If you flush one page
at a time, then you might never recover the performance of a plain old
memcpy().  If you flush all of them, then you're hurting the
performance of everything else in the task.

In general, I suspect that the zerocopy receive mechanism will only
really be a win in single-threaded applications that consume large
amounts of receive bandwidth on a single TCP socket using lots of
memory and don't do all that much else.