Re: [PATCH RFC] usercopy: optimize stack check flow when the page-spanning test is disabled

From: Matthew Wilcox
Date: Tue Aug 14 2018 - 09:28:41 EST

On Tue, Aug 14, 2018 at 01:02:55PM +0000, Yuanxiaofeng (XiAn) wrote:
> 1, When the THREAD_SIZE is less than PAGE_SIZE, the stack will allocate memory by kmem_cache_alloc_node(), it's slab memory and will execute __check_heap_object().
> 2, When CONFIG_HARDENED_USERCOPY_PAGESPAN is enabled, the multiple-pages stacks will do some check in check_page_span().

I understand the checks will still do something useful, but I do not see the
scenario in which an object would satisfy the stack checks but fail the heap

> So, I set some restrictions to make sure the useful check will not be skipped.
> -----Original Message-----
> From: Matthew Wilcox [mailto:willy@xxxxxxxxxxxxx]
> Sent: Tuesday, August 14, 2018 8:35 PM
> To: Yuanxiaofeng (XiAn)
> Cc: keescook@xxxxxxxxxxxx; linux-mm@xxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH RFC] usercopy: optimize stack check flow when the
> On Tue, Aug 14, 2018 at 08:17:31PM +0800, Xiaofeng Yuan wrote:
> > The check_heap_object() checks the spanning multiple pages and slab.
> > When the page-spanning test is disabled, the check_heap_object() is
> > redundant for spanning multiple pages. However, the kernel stacks are
> > multiple pages under certain conditions: CONFIG_ARCH_THREAD_STACK_ALLOCATOR
> > is not defined and (THREAD_SIZE >= PAGE_SIZE). At this point, We can skip
> > the check_heap_object() for kernel stacks to improve performance.
> > Similarly, the virtually-mapped stack can skip check_heap_object() also,
> > beacause virt_addr_valid() will return.
> Why not just check_stack_object() first, then check_heap_object() second?