Re: [PATCH 1/1] gup: document and work around "COW can break either way" issue

[John Hubbard <jhubbard@nvidia.com>]

On 10/11/21 18:52, Suren Baghdasaryan wrote:
...
> [surenb: backport notes
>          Since gup_pgd_range does not exist, made appropriate changes on
>          the the gup_huge_pgd, gup_huge_pd and gup_pud_range calls instead.
> 	Replaced (gup_flags | FOLL_WRITE) with write=1 in gup_huge_pgd,
>          gup_huge_pd and gup_pud_range.
> 	Removed FOLL_PIN usage in should_force_cow_break since it's missing in
> 	the earlier kernels.]
> 

This backport looks accurate. At first I thought you missed the comment-only
change to i915_gem_userptr.c, and then I noticed that the older branch still
uses non-fast gup, so not applicable there after all. :)

Agree with others that this whole area is still shaky, but it does sound as if
this will help.


Reviewed-by: John Hubbard <jhubbard@nvidia.com>


thanks,
-- 
John Hubbard
NVIDIA

> Reported-by: Jann Horn <jannh@google.com>
> Tested-by: Christoph Hellwig <hch@lst.de>
> Acked-by: Oleg Nesterov <oleg@redhat.com>
> Acked-by: Kirill Shutemov <kirill@shutemov.name>
> Acked-by: Jan Kara <jack@suse.cz>
> Cc: Andrea Arcangeli <aarcange@redhat.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
> [surenb: backport to 4.4 kernel]
> Cc: stable@vger.kernel.org # 4.4.x
> Signed-off-by: Suren Baghdasaryan <surenb@google.com>
> ---
>   mm/gup.c         | 48 ++++++++++++++++++++++++++++++++++++++++--------
>   mm/huge_memory.c |  7 +++----
>   2 files changed, 43 insertions(+), 12 deletions(-)
> 
> diff --git a/mm/gup.c b/mm/gup.c
> index 4c5857889e9d..c80cdc408228 100644
> --- a/mm/gup.c
> +++ b/mm/gup.c
> @@ -59,13 +59,22 @@ static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
>   }
>   
>   /*
> - * FOLL_FORCE can write to even unwritable pte's, but only
> - * after we've gone through a COW cycle and they are dirty.
> + * FOLL_FORCE or a forced COW break can write even to unwritable pte's,
> + * but only after we've gone through a COW cycle and they are dirty.
>    */
>   static inline bool can_follow_write_pte(pte_t pte, unsigned int flags)
>   {
> -	return pte_write(pte) ||
> -		((flags & FOLL_FORCE) && (flags & FOLL_COW) && pte_dirty(pte));
> +	return pte_write(pte) || ((flags & FOLL_COW) && pte_dirty(pte));
> +}
> +
> +/*
> + * A (separate) COW fault might break the page the other way and
> + * get_user_pages() would return the page from what is now the wrong
> + * VM. So we need to force a COW break at GUP time even for reads.
> + */
> +static inline bool should_force_cow_break(struct vm_area_struct *vma, unsigned int flags)
> +{
> +	return is_cow_mapping(vma->vm_flags) && (flags & FOLL_GET);
>   }
>   
>   static struct page *follow_page_pte(struct vm_area_struct *vma,
> @@ -509,12 +518,18 @@ long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
>   			if (!vma || check_vma_flags(vma, gup_flags))
>   				return i ? : -EFAULT;
>   			if (is_vm_hugetlb_page(vma)) {
> +				if (should_force_cow_break(vma, foll_flags))
> +					foll_flags |= FOLL_WRITE;
>   				i = follow_hugetlb_page(mm, vma, pages, vmas,
>   						&start, &nr_pages, i,
> -						gup_flags);
> +						foll_flags);
>   				continue;
>   			}
>   		}
> +
> +		if (should_force_cow_break(vma, foll_flags))
> +			foll_flags |= FOLL_WRITE;
> +
>   retry:
>   		/*
>   		 * If we have a pending SIGKILL, don't keep faulting pages and
> @@ -1346,6 +1361,10 @@ static int gup_pud_range(pgd_t pgd, unsigned long addr, unsigned long end,
>   /*
>    * Like get_user_pages_fast() except it's IRQ-safe in that it won't fall back to
>    * the regular GUP. It will only return non-negative values.
> + *
> + * Careful, careful! COW breaking can go either way, so a non-write
> + * access can get ambiguous page results. If you call this function without
> + * 'write' set, you'd better be sure that you're ok with that ambiguity.
>    */
>   int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
>   			  struct page **pages)
> @@ -1375,6 +1394,12 @@ int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
>   	 *
>   	 * We do not adopt an rcu_read_lock(.) here as we also want to
>   	 * block IPIs that come from THPs splitting.
> +	 *
> +	 * NOTE! We allow read-only gup_fast() here, but you'd better be
> +	 * careful about possible COW pages. You'll get _a_ COW page, but
> +	 * not necessarily the one you intended to get depending on what
> +	 * COW event happens after this. COW may break the page copy in a
> +	 * random direction.
>   	 */
>   
>   	local_irq_save(flags);
> @@ -1385,15 +1410,22 @@ int __get_user_pages_fast(unsigned long start, int nr_pages, int write,
>   		next = pgd_addr_end(addr, end);
>   		if (pgd_none(pgd))
>   			break;
> +		/*
> +		 * The FAST_GUP case requires FOLL_WRITE even for pure reads,
> +		 * because get_user_pages() may need to cause an early COW in
> +		 * order to avoid confusing the normal COW routines. So only
> +		 * targets that are already writable are safe to do by just
> +		 * looking at the page tables.
> +		 */
>   		if (unlikely(pgd_huge(pgd))) {
> -			if (!gup_huge_pgd(pgd, pgdp, addr, next, write,
> +			if (!gup_huge_pgd(pgd, pgdp, addr, next, 1,
>   					  pages, &nr))
>   				break;
>   		} else if (unlikely(is_hugepd(__hugepd(pgd_val(pgd))))) {
>   			if (!gup_huge_pd(__hugepd(pgd_val(pgd)), addr,
> -					 PGDIR_SHIFT, next, write, pages, &nr))
> +					 PGDIR_SHIFT, next, 1, pages, &nr))
>   				break;
> -		} else if (!gup_pud_range(pgd, addr, next, write, pages, &nr))
> +		} else if (!gup_pud_range(pgd, addr, next, 1, pages, &nr))
>   			break;
>   	} while (pgdp++, addr = next, addr != end);
>   	local_irq_restore(flags);
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index 6404e4fcb4ed..fae45c56e2ee 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1268,13 +1268,12 @@ out_unlock:
>   }
>   
>   /*
> - * FOLL_FORCE can write to even unwritable pmd's, but only
> - * after we've gone through a COW cycle and they are dirty.
> + * FOLL_FORCE or a forced COW break can write even to unwritable pmd's,
> + * but only after we've gone through a COW cycle and they are dirty.
>    */
>   static inline bool can_follow_write_pmd(pmd_t pmd, unsigned int flags)
>   {
> -	return pmd_write(pmd) ||
> -	       ((flags & FOLL_FORCE) && (flags & FOLL_COW) && pmd_dirty(pmd));
> +	return pmd_write(pmd) || ((flags & FOLL_COW) && pmd_dirty(pmd));
>   }
>   
>   struct page *follow_trans_huge_pmd(struct vm_area_struct *vma,
>