[PATCHv6 2/2] ARM: dma-mapping: remove custom consistent dma region

[Hiroshi Doyu]

Hi Marek,

On Tue, 21 Aug 2012 17:01:08 +0200
Marek Szyprowski <m.szyprowski at samsung.com> wrote:

> > > > -__iommu_alloc_remap(struct page **pages, size_t size, gfp_t gfp, pgprot_t prot)
> > > > +__iommu_alloc_remap(struct page **pages, size_t size, gfp_t gfp, pgprot_t prot,
> > > > +                   const void *caller)
> > > >  {
> > > > -       struct arm_vmregion *c;
> > > > -       size_t align;
> > > > -       size_t count = size >> PAGE_SHIFT;
> > > > -       int bit;
> > > > +       unsigned int i, nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
> > > > +       struct vm_struct *area;
> > > > +       unsigned long p;
> > > >
> > > > -       if (!consistent_pte[0]) {
> > > > -               pr_err("%s: not initialised\n", __func__);
> > > > -               dump_stack();
> > > > +       area = get_vm_area_caller(size, VM_ARM_DMA_CONSISTENT | VM_USERMAP,
> > > > +                                 caller);
> > > > +       if (!area)
> > >
> > > This patch replaced the custom "consistent_pte" with
> > > get_vm_area_caller()", which breaks the compatibility with the
> > > existing driver. This causes the following kernel oops(*1). That
> > > driver has called dma_pool_alloc() to allocate memory from the
> > > interrupt context, and it hits BUG_ON(in_interrpt()) in
> > > "get_vm_area_caller()"(*2). Regardless of the badness of allocation
> > > from interrupt handler in the driver, I have the following question.
> > >
> > > The following "__get_vm_area_node()" can take gfp_mask, it means that
> > > this function is expected to be called from atomic context, but why
> > > it's _NOT_ allowed _ONLY_ from interrupt context?
> > >
> > > According to the following definitions, "in_interrupt()" is in "in_atomic()".
> > >
> > > #define in_interrupt()	(preempt_count() & (HARDIRQ_MASK | SOFTIRQ_MASK | NMI_MASK))
> > > #define in_atomic()	((preempt_count() & ~PREEMPT_ACTIVE) != 0)
> > >
> > > Does anyone know why BUG_ON(in_interrupt()) is set in __get_vm_area_node(*3)?
> > 
> > For arm_dma_alloc(), it allocates from the pool if GFP_ATOMIC, but for
> > arm_iommu_alloc_attrs() doesn't have pre-allocate pool at all, and it
> > always call "get_vm_area_caller()". That's why it hits BUG(). But
> > still I don't understand why it's not BUG_ON(in_atomic) as Russell
> > already pointed out(*1).
> > 
> > *1: http://article.gmane.org/gmane.linux.kernel.mm/76708
> 
> Ok, now I see the problem. I will try to find out a solution for your issue.

My explanation wasn't so good.

For a solution, I thought that, in order to allow IOMMU'able device
drivers to allocate memory from atomic context/ISR, there were the
following 2 solutions:

(1) To provide the pre-allocate area like arm_dma_alloc() does,
or
(2) __get_vm_area_node() can be called from ISR.

But (2) doesn't work because PGALLOC_GFP(GFP_KERNEL) is used to
allocate a page table. This is called from:

  arm_iommu_alloc_attrs() ->
    __iommu_alloc_remap() ->
      ioremap_page_range() ->
        .....              ->
          pte_alloc_one_kernel() ->
              pte = (pte_t *)__get_free_page(PGALLOC_GFP);

We always have to avoid changing a page table for atomic
allocation. So for me, the only remaining solution is
(1) pre-allocation. We can make use of the same atomic pool both for
DMA and IOMMU. I'll send the patch.