Re: [PATCH] libnvdimm/pmem: Bypass CONFIG_HARDENED_USERCOPY overhead

[Kees Cook]

On Fri, May 17, 2019 at 10:28:48AM -0700, Dan Williams wrote:
> On Fri, May 17, 2019 at 8:57 AM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > On Fri, May 17, 2019 at 08:08:27AM -0700, Dan Williams wrote:
> > > As far as I can see it's mostly check_heap_object() that is the
> > > problem, so I'm open to finding a way to just bypass that sub-routine.
> > > However, as far as I can see none of the other block / filesystem user
> > > copy implementations submit to the hardened checks, like
> > > bio_copy_from_iter(), and iov_iter_copy_from_user_atomic() . So,
> > > either those need to grow additional checks, or the hardened copy
> > > implementation is targeting single object copy use cases, not
> > > necessarily block-I/O. Yes, Kees, please advise.
> >
> > The intention is mainly for copies that haven't had explicit bounds
> > checking already performed on them, yes. Is there something getting
> > checked out of the slab, or is it literally just the overhead of doing
> > the "is this slab?" check that you're seeing?
> 
> It's literally the overhead of "is this slab?" since it needs to go
> retrieve the struct page and read that potentially cold cacheline. In
> the case where that page is on memory media that is higher latency
> than DRAM we get the ~37% performance loss that Jeff measured.

Ah-ha! Okay, I understand now; thanks!

> The path is via the filesystem ->write_iter() file operation. In the
> DAX case the filesystem traps that path early, before submitting block
> I/O, and routes it to the dax_iomap_actor() routine. That routine
> validates that the logical file offset is within bounds of the file,
> then it does a sector-to-pfn translation which validates that the
> physical mapping is within bounds of the block device.
> 
> It seems dax_iomap_actor() is not a path where we'd be worried about
> needing hardened user copy checks.

I would agree: I think the proposed patch makes sense. :)

-- 
Kees Cook