Re: [PATCH] dma-debug: fix incorrect pfn calculation

[Miles Chen]

On Tue, 2017-09-26 at 15:53 +0100, Robin Murphy wrote:
> On 26/09/17 14:24, miles.chen@xxxxxxxxxxxx wrote:
> > From: Miles Chen <miles.chen@xxxxxxxxxxxx>
> > 
> > dma-debug report the following warning:
> > 
> > [name:panic&]WARNING: CPU: 3 PID: 298 at kernel-4.4/lib/dma-debug.c:604
> > debug _dma_assert_idle+0x1a8/0x230()
> > DMA-API: cpu touching an active dma mapped cacheline [cln=0x00000882300]
> > CPU: 3 PID: 298 Comm: vold Tainted: G        W  O    4.4.22+ #1
> > Hardware name: MT6739 (DT)
> > Call trace:
> > [<ffffff800808acd0>] dump_backtrace+0x0/0x1d4
> > [<ffffff800808affc>] show_stack+0x14/0x1c
> > [<ffffff800838019c>] dump_stack+0xa8/0xe0
> > [<ffffff80080a0594>] warn_slowpath_common+0xf4/0x11c
> > [<ffffff80080a061c>] warn_slowpath_fmt+0x60/0x80
> > [<ffffff80083afe24>] debug_dma_assert_idle+0x1a8/0x230
> > [<ffffff80081dca9c>] wp_page_copy.isra.96+0x118/0x520
> > [<ffffff80081de114>] do_wp_page+0x4fc/0x534
> > [<ffffff80081e0a14>] handle_mm_fault+0xd4c/0x1310
> > [<ffffff8008098798>] do_page_fault+0x1c8/0x394
> > [<ffffff800808231c>] do_mem_abort+0x50/0xec
> > 
> > I found that debug_dma_alloc_coherent() and debug_dma_free_coherent()
> > always use type "dma_debug_coherent" and assume that dma_alloc_coherent()
> > always returns a linear address.
> > 
> > However if a device returns false on is_device_dma_coherent(),
> > dma_alloc_coherent() will create another non-cacheable mapping
> > (also non linear). In this case, page_to_pfn(virt_to_page(virt)) will
> > return an incorrect pfn. If the pfn is valid and mapped as a COW page,
> > we will hit the warning when doing wp_page_copy().
> > 
> > Fix this by calculating correct pfn if is_device_dma_coherent()
> > returns false.
> 
> As the inevitable storm of kbuild robot reports will tell you soon, you
> can't do that: is_device_dma_coherent() is a private helper between
> arm{,64} arch code and xen, and should not be used anywhere else.

thanks for that. I should not use the private helper function here.

> > Signed-off-by: Miles Chen <miles.chen@xxxxxxxxxxxx>
> > ---
> >  lib/dma-debug.c | 20 ++++++++++++++------
> >  1 file changed, 14 insertions(+), 6 deletions(-)
> > 
> > diff --git a/lib/dma-debug.c b/lib/dma-debug.c
> > index ea4cc3d..b17e56e 100644
> > --- a/lib/dma-debug.c
> > +++ b/lib/dma-debug.c
> > @@ -47,6 +47,8 @@ enum {
> >  	dma_debug_sg,
> >  	dma_debug_coherent,
> >  	dma_debug_resource,
> > +	dma_debug_noncoherent,
> > +	nr_dma_debug_types,
> >  };
> >  
> >  enum map_err_types {
> > @@ -154,9 +156,9 @@ static inline bool dma_debug_disabled(void)
> >  	[MAP_ERR_CHECKED] = "dma map error checked",
> >  };
> >  
> > -static const char *type2name[5] = { "single", "page",
> > +static const char *type2name[nr_dma_debug_types] = { "single", "page",
> >  				    "scather-gather", "coherent",
> > -				    "resource" };
> > +				    "resource", "noncoherent" };
> >  
> >  static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE",
> >  				   "DMA_FROM_DEVICE", "DMA_NONE" };
> > @@ -1484,6 +1486,7 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size,
> >  			      dma_addr_t dma_addr, void *virt)
> >  {
> >  	struct dma_debug_entry *entry;
> > +	bool coherent = is_device_dma_coherent(dev);
> >  
> >  	if (unlikely(dma_debug_disabled()))
> >  		return;
> > @@ -1495,9 +1498,11 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size,
> >  	if (!entry)
> >  		return;
> >  
> > -	entry->type      = dma_debug_coherent;
> > +	entry->type      = coherent ? dma_debug_coherent :
> > +					dma_debug_noncoherent;
> >  	entry->dev       = dev;
> > -	entry->pfn	 = page_to_pfn(virt_to_page(virt));
> 
> There are more architectures where the virtual address returned by
> dma_alloc_coherent is not a linear map address - some just have a static
> offset between cacheable and non-cacheable aliases - so there may be
> other cases where this is the wrong calculation, but at least those
> probably don't trigger the problematic false-positive.
> 
> That said, the cases where coherent allocations *are* dynamically
> remapped should be easy enough to handle properly without having to
> resort to dodgy hacks:
> 
> 	if (is_vmalloc_addr(virt))
> 		pfn = vmalloc_to_pfn(virt);
> 	else
> 		pfn = page_to_pfn(virt_to_page(virt));
> 
> Simple.

Thanks for your advice. I'll try this approach.

Miles

> 
> > +	entry->pfn	 = coherent ? page_to_pfn(virt_to_page(virt)) :
> > +					dma_addr >> PAGE_SHIFT;
> 
> And in particular, this is just as likely to just give *different* false
> positives, since there's no guarantee whatsoever that dma_addr has any
> relationship to the appropriate pfn.
> 
> Robin.
> 
> >  	entry->offset	 = offset_in_page(virt);
> >  	entry->size      = size;
> >  	entry->dev_addr  = dma_addr;
> > @@ -1510,10 +1515,13 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size,
> >  void debug_dma_free_coherent(struct device *dev, size_t size,
> >  			 void *virt, dma_addr_t addr)
> >  {
> > +	bool coherent = is_device_dma_coherent(dev);
> >  	struct dma_debug_entry ref = {
> > -		.type           = dma_debug_coherent,
> > +		.type           = coherent ? dma_debug_coherent :
> > +						dma_debug_noncoherent,
> >  		.dev            = dev,
> > -		.pfn		= page_to_pfn(virt_to_page(virt)),
> > +		.pfn		= coherent ? page_to_pfn(virt_to_page(virt)) :
> > +						addr >> PAGE_SHIFT,
> >  		.offset		= offset_in_page(virt),
> >  		.dev_addr       = addr,
> >  		.size           = size,
> > 
>