Re: [RFC PATCH 4/7] drm/ttm: Support huge pagefaults

["Christian König" <christian.koenig@amd.com>]

Am 27.11.19 um 09:31 schrieb Thomas Hellström (VMware):
> From: Thomas Hellstrom <thellstrom@vmware.com>
>
> Support huge (PMD-size and PUD-size) page-table entries by providing a
> huge_fault() callback.
> We still support private mappings and write-notify by splitting the huge
> page-table entries on write-access.
>
> Note that for huge page-faults to occur, either the kernel needs to be
> compiled with trans-huge-pages always enabled, or the kernel needs to be
> compiled with trans-huge-pages enabled using madvise, and the user-space
> app needs to call madvise() to enable trans-huge pages on a per-mapping
> basis.
>
> Furthermore huge page-faults will not occur unless buffer objects and
> user-space addresses are aligned on huge page size boundaries.
>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Michal Hocko <mhocko@suse.com>
> Cc: "Matthew Wilcox (Oracle)" <willy@infradead.org>
> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
> Cc: Ralph Campbell <rcampbell@nvidia.com>
> Cc: "Jérôme Glisse" <jglisse@redhat.com>
> Cc: "Christian König" <christian.koenig@amd.com>
> Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
> ---
>   drivers/gpu/drm/ttm/ttm_bo_vm.c | 139 +++++++++++++++++++++++++++++++-
>   include/drm/ttm/ttm_bo_api.h    |   3 +-
>   2 files changed, 138 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c
> index 2098f8d4dfc5..8d6089880e39 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
> @@ -150,6 +150,84 @@ vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
>   }
>   EXPORT_SYMBOL(ttm_bo_vm_reserve);
>   
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +/**
> + * ttm_bo_vm_insert_huge - Insert a pfn for PUD or PMD faults
> + * @vmf: Fault data
> + * @bo: The buffer object
> + * @page_offset: Page offset from bo start
> + * @fault_page_size: The size of the fault in pages.
> + * @pgprot: The page protections.
> + * Does additional checking whether it's possible to insert a PUD or PMD
> + * pfn and performs the insertion.
> + *
> + * Return: VM_FAULT_NOPAGE on successful insertion, VM_FAULT_FALLBACK if
> + * a huge fault was not possible, and a VM_FAULT_ERROR code otherwise.
> + */
> +static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
> +					struct ttm_buffer_object *bo,
> +					pgoff_t page_offset,
> +					pgoff_t fault_page_size,
> +					pgprot_t pgprot)
> +{
> +	pgoff_t i;
> +	vm_fault_t ret;
> +	unsigned long pfn;
> +	pfn_t pfnt;
> +	struct ttm_tt *ttm = bo->ttm;
> +	bool write = vmf->flags & FAULT_FLAG_WRITE;
> +
> +
> +	/* Fault should not cross bo boundary */
> +	page_offset &= ~(fault_page_size - 1);
> +	if (page_offset + fault_page_size > bo->num_pages)
> +		goto out_fallback;
> +
> +	if (bo->mem.bus.is_iomem)
> +		pfn = ttm_bo_io_mem_pfn(bo, page_offset);
> +	else
> +		pfn = page_to_pfn(ttm->pages[page_offset]);
> +
> +	/* pfn must be fault_page_size aligned. */
> +	if ((pfn & (fault_page_size - 1)) != 0)
> +		goto out_fallback;
> +
> +	/* IO memory is OK now, TT memory must be contigous. */

That won't work correctly, IO mem might not be contiguous either.

We either need to call ttm_bo_io_mem_pfn() multiple times and check that 
the addresses are linear or return the length additional to the pfn.

Regards,
Christian.

> +	if (!bo->mem.bus.is_iomem)
> +		for (i = 1; i < fault_page_size; ++i) {
> +			if (page_to_pfn(ttm->pages[page_offset + i]) != pfn + i)
> +				goto out_fallback;
> +		}
> +
> +	pfnt = __pfn_to_pfn_t(pfn, PFN_DEV);
> +	if (fault_page_size == (HPAGE_PMD_SIZE >> PAGE_SHIFT))
> +		ret = vmf_insert_pfn_pmd_prot(vmf, pfnt, pgprot, write);
> +#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
> +	else if (fault_page_size == (HPAGE_PUD_SIZE >> PAGE_SHIFT))
> +		ret = vmf_insert_pfn_pud_prot(vmf, pfnt, pgprot, write);
> +#endif
> +	else
> +		WARN_ON_ONCE(ret = VM_FAULT_FALLBACK);
> +
> +	if (ret != VM_FAULT_NOPAGE)
> +		goto out_fallback;
> +
> +	return VM_FAULT_NOPAGE;
> +out_fallback:
> +	count_vm_event(THP_FAULT_FALLBACK);
> +	return VM_FAULT_FALLBACK;
> +}
> +#else
> +static vm_fault_t ttm_bo_vm_insert_huge(struct vm_fault *vmf,
> +					struct ttm_buffer_object *bo,
> +					pgoff_t page_offset,
> +					pgoff_t fault_page_size,
> +					pgprot_t pgprot)
> +{
> +	return VM_FAULT_NOPAGE;
> +}
> +#endif
> +
>   /**
>    * ttm_bo_vm_fault_reserved - TTM fault helper
>    * @vmf: The struct vm_fault given as argument to the fault callback
> @@ -170,7 +248,8 @@ EXPORT_SYMBOL(ttm_bo_vm_reserve);
>    */
>   vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
>   				    pgprot_t prot,
> -				    pgoff_t num_prefault)
> +				    pgoff_t num_prefault,
> +				    pgoff_t fault_page_size)
>   {
>   	struct vm_area_struct *vma = vmf->vma;
>   	struct ttm_buffer_object *bo = vma->vm_private_data;
> @@ -262,6 +341,13 @@ vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
>   		prot = pgprot_decrypted(prot);
>   	}
>   
> +	/* We don't prefault on huge faults. Yet. */
> +	if (IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) && fault_page_size != 1) {
> +		ret = ttm_bo_vm_insert_huge(vmf, bo, page_offset,
> +					    fault_page_size, prot);
> +		goto out_io_unlock;
> +	}
> +
>   	/*
>   	 * Speculatively prefault a number of pages. Only error on
>   	 * first page.
> @@ -320,7 +406,7 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
>   		return ret;
>   
>   	prot = vma->vm_page_prot;
> -	ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT);
> +	ret = ttm_bo_vm_fault_reserved(vmf, prot, TTM_BO_VM_NUM_PREFAULT, 1);
>   	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
>   		return ret;
>   
> @@ -330,6 +416,50 @@ vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf)
>   }
>   EXPORT_SYMBOL(ttm_bo_vm_fault);
>   
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +static vm_fault_t ttm_bo_vm_huge_fault(struct vm_fault *vmf,
> +				       enum page_entry_size pe_size)
> +{
> +	struct vm_area_struct *vma = vmf->vma;
> +	pgprot_t prot;
> +	struct ttm_buffer_object *bo = vma->vm_private_data;
> +	vm_fault_t ret;
> +	pgoff_t fault_page_size = 0;
> +	bool write = vmf->flags & FAULT_FLAG_WRITE;
> +
> +	switch (pe_size) {
> +	case PE_SIZE_PMD:
> +		fault_page_size = HPAGE_PMD_SIZE >> PAGE_SHIFT;
> +		break;
> +#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
> +	case PE_SIZE_PUD:
> +		fault_page_size = HPAGE_PUD_SIZE >> PAGE_SHIFT;
> +		break;
> +#endif
> +	default:
> +		WARN_ON_ONCE(1);
> +		return VM_FAULT_FALLBACK;
> +	}
> +
> +	/* Fallback on write dirty-tracking or COW */
> +	if (write && !(pgprot_val(vmf->vma->vm_page_prot) & _PAGE_RW))
> +		return VM_FAULT_FALLBACK;
> +
> +	ret = ttm_bo_vm_reserve(bo, vmf);
> +	if (ret)
> +		return ret;
> +
> +	prot = vm_get_page_prot(vma->vm_flags);
> +	ret = ttm_bo_vm_fault_reserved(vmf, prot, 1, fault_page_size);
> +	if (ret == VM_FAULT_RETRY && !(vmf->flags & FAULT_FLAG_RETRY_NOWAIT))
> +		return ret;
> +
> +	dma_resv_unlock(bo->base.resv);
> +
> +	return ret;
> +}
> +#endif
> +
>   void ttm_bo_vm_open(struct vm_area_struct *vma)
>   {
>   	struct ttm_buffer_object *bo = vma->vm_private_data;
> @@ -431,7 +561,10 @@ static const struct vm_operations_struct ttm_bo_vm_ops = {
>   	.fault = ttm_bo_vm_fault,
>   	.open = ttm_bo_vm_open,
>   	.close = ttm_bo_vm_close,
> -	.access = ttm_bo_vm_access
> +	.access = ttm_bo_vm_access,
> +#ifdef CONFIG_TRANSPARENT_HUGEPAGE
> +	.huge_fault = ttm_bo_vm_huge_fault,
> +#endif
>   };
>   
>   static struct ttm_buffer_object *ttm_bo_vm_lookup(struct ttm_bo_device *bdev,
> diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
> index 66ca49db9633..4fc90d53aa15 100644
> --- a/include/drm/ttm/ttm_bo_api.h
> +++ b/include/drm/ttm/ttm_bo_api.h
> @@ -732,7 +732,8 @@ vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
>   
>   vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
>   				    pgprot_t prot,
> -				    pgoff_t num_prefault);
> +				    pgoff_t num_prefault,
> +				    pgoff_t fault_page_size);
>   
>   vm_fault_t ttm_bo_vm_fault(struct vm_fault *vmf);
>