From mboxrd@z Thu Jan  1 00:00:00 1970
From: gmbnomis@gmail.com (Simon Baatz)
Date: Wed, 1 May 2013 21:04:41 +0200
Subject: [PATCH V3 2/2] ARM: Handle user space mapped pages in
 flush_kernel_dcache_page
In-Reply-To: <20130501142206.GB17387@arm.com>
References: <20121008230733.GA17819@schnuecks.de>
 <20121118211005.GW22106@titan.lakedaemon.net>
 <20130418111608.GF27197@titan.lakedaemon.net>
 <20130418112201.GQ14496@n2100.arm.linux.org.uk>
 <20130418114016.GG27197@titan.lakedaemon.net>
 <20130418135104.GA18616@arm.com>
 <20130421220629.GA25571@schnuecks.de>
 <20130430112225.GD29766@arm.com>
 <20130430210403.GA18076@schnuecks.de>
 <20130501142206.GB17387@arm.com>
Message-ID: <20130501190441.GA22227@schnuecks.de>
To: linux-arm-kernel@lists.infradead.org
List-Id: linux-arm-kernel.lists.infradead.org

On Wed, May 01, 2013 at 03:22:06PM +0100, Catalin Marinas wrote:
> On Tue, Apr 30, 2013 at 10:04:03PM +0100, Simon Baatz wrote:
> > On Tue, Apr 30, 2013 at 12:22:25PM +0100, Catalin Marinas wrote:
> > > On Sun, Apr 21, 2013 at 11:06:30PM +0100, Simon Baatz wrote:
> > > > On Thu, Apr 18, 2013 at 02:51:04PM +0100, Catalin Marinas wrote:
> > > > > On Thu, Apr 18, 2013 at 12:40:16PM +0100, Jason Cooper wrote:
> > ...
> > > > 
> > > > It is not the driver itself which is using the API, it is the
> > > > generic scatterlist memory iterator. And I don't think that this is
> > > > wrong, as I have tried to explain in [1].
> > > 
> > > Trying to remember what we've discussed over the past months on this
> > > topic. It looks like sg_miter_stop() does the right thing in calling
> > > flush_kernel_dcache_page(). Commit f8b63c1 (ARM: 6382/1: Remove
> > > superfluous flush_kernel_dcache_page()) removed this function entirely.
> > > The code previously had this comment - /* highmem pages are always
> > > flushed upon kunmap already */ which I think it wasn't fully correct
> > > either. The kunmap_atomic() flushes the caches but kunmap() doesn't, so
> > > I suspect we only get the flushing if SG_MITER_ATOMIC.
> > > 
> > > So it looks to me like flush_kernel_dcache_page() should be implemented
> > > even for highmem pages (with VIVT or aliasing VIPT, at least for non
> > > kmap_atomic addresses by checking for FIXADDR_START). If highmem is
> > > disabled, I suspect we still need this function since the calling code
> > > doesn't care whether kmap/kunmap was a no-op. But can we keep it as a
> > > simple call to __cpuc_flush_dcache_area(page_address(page), PAGE_SIZE)?
> > 
> > My first version ([1]) had:
> > 
> > 	if ((cache_is_vivt() || cache_is_vipt_aliasing()) && !PageHighMem(page))
> > 		__flush_kernel_dcache_page(page);
> > 
> > If I understand this correctly, you are proposing to remove the
> > highmem exclusion.
> 
> The highmem exclusion may have been there originally because of a
> comment suggesting that kunmap() does the flushing. This is the case
> only for kunmap_atomic() AFAICT (and maybe we could remove that as well
> and rely on flush_kernel_dcache_page() being called).
> 
> > And then in __flush_kernel_dcache_page():
> > 
> > 	mapping = page_mapping(page);
> > 
> > 	if (!mapping || mapping_mapped(mapping))
> > 		__cpuc_flush_dcache_area(page_address(page), PAGE_SIZE);
> > 
> > I still prefer to have this condition here to avoid the flush when
> > there is no user mapping at all.  This is handled by lazy flushing
> > and is probably the most common case anyway (given how many people
> > seem to be affected by this problem).
> 
> Looking at the old thread, you said there is a case when this condition
> is not true (O_DIRECT case). If that's for a page cache page, then we
> can handle it lazily (for anonymous pages I don't think we can rely on
> lazy flushing since the kernel does not guarantee the clearing of the
> PG_arch_1 bit).

As Russel pointed out in a comment to a later version of the patch,
PG_arch_1 makes only sense for page cache pages. The condition above
is ok from my point of view (it is based on what flush_dcache_page()
uses):

- Page cache page without user space mapping:
mapping != NULL and mapping_mapped() == 0

-> no flush here; lazy flush based on PG_arch_1 later if needed (we
   rely on the proper initialization of the page to "dirty" here.)


- Page cache page with user space mapping:
mapping != NULL and mapping_mapped() != 0

-> kernel mapping flushed here (user mapping can be assumed to be clean)


- Anonymous page:
mapping == NULL

-> kernel mapping flushed here (user mapping can be assumed to be clean)

 
> > Additionally, although we can assume that the page is kmapped,
> > page_address(page) can still be NULL for a highmem page, right?
> 
> It looks like kmap() always sets page_address(page) but I'm not sure
> about kmap_atomic(), it doesn't seem to.

Hmm, in __flush_dcache_page() we have the following code to flush the
kernel mapping:

void __flush_dcache_page(struct address_space *mapping, struct page *page)
{
	/*
	 * Writeback any data associated with the kernel mapping of this
	 * page.  This ensures that data in the physical page is mutually
	 * coherent with the kernels mapping.
	 */
	if (!PageHighMem(page)) {
		__cpuc_flush_dcache_area(page_address(page), PAGE_SIZE);
	} else {
		void *addr = kmap_high_get(page);
		if (addr) {
			__cpuc_flush_dcache_area(addr, PAGE_SIZE);
			kunmap_high(page);
		} else if (cache_is_vipt()) {
			/* unmapped pages might still be cached */
			addr = kmap_atomic(page);
			__cpuc_flush_dcache_area(addr, PAGE_SIZE);
			kunmap_atomic(addr);
		}
	}
...


We probably should reuse it in flush_kernel_dcache_page() to flush
the kernel mapping. (The last else clause looks strange though)


- Simon