Re: [RFC PATCH 2/5] hugetlb: enhance hugetlb fault processing to support soft dirty

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Feb 10, 2021 at 04:03:19PM -0800, Mike Kravetz wrote:
> hugetlb fault processing code would COW all write faults where the
> pte was not writable.  Soft dirty will write protect ptes as part
> of it's tracking mechanism.  The existing hugetlb_cow  code will do
> the right thing for PRIVATE mappings as it checks map_count.  However,
> for SHARED mappings it would actually allocate and install a COW page.
> Modify the code to not call hugetlb_cow for SHARED mappings and just
> update the pte.
> 
> Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
> ---
>  mm/hugetlb.c | 23 ++++++++++++++++-------
>  1 file changed, 16 insertions(+), 7 deletions(-)
> 
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index 47f3123afd1a..b561b6867ec1 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -4584,8 +4584,10 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>  	 * spinlock. For private mappings, we also lookup the pagecache
>  	 * page now as it is used to determine if a reservation has been
>  	 * consumed.
> +	 * Only non-shared mappings are sent to hugetlb_cow.
>  	 */
> -	if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry)) {
> +	if ((flags & FAULT_FLAG_WRITE) && !huge_pte_write(entry) &&
> +					!(vma->vm_flags & VM_SHARED)) {
>  		if (vma_needs_reservation(h, vma, haddr) < 0) {
>  			ret = VM_FAULT_OOM;
>  			goto out_mutex;
> @@ -4593,9 +4595,7 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>  		/* Just decrements count, does not deallocate */
>  		vma_end_reservation(h, vma, haddr);
>  
> -		if (!(vma->vm_flags & VM_MAYSHARE))
> -			pagecache_page = hugetlbfs_pagecache_page(h,
> -								vma, haddr);
> +		pagecache_page = hugetlbfs_pagecache_page(h, vma, haddr);

Pure question: I see that the check actually changed from VM_MAYSHARE into
VM_SHARE, then I noticed I'm actually unclear on the difference..  Say, when
VM_MAYSHARE is set, could VM_SHARED be cleared in any case?  Or say, is this
change intended?

I see that vma_set_page_prot() tried to remove VM_SHARED if soft dirty enabled
(which should cause vma_wants_writenotify() to return true, iiuc), however
that's temporary just to calculate vm_page_prot, and it's not applied to the
vma->vm_flags.  I failed to find a place where VM_SHARED of the vma is cleared
while VM_MAYSHARE is set..

>  	}
>  
>  	ptl = huge_pte_lock(h, mm, ptep);
> @@ -4620,9 +4620,18 @@ vm_fault_t hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
>  
>  	if (flags & FAULT_FLAG_WRITE) {
>  		if (!huge_pte_write(entry)) {
> -			ret = hugetlb_cow(mm, vma, address, ptep,
> -					  pagecache_page, ptl);
> -			goto out_put_page;
> +			if (!(vma->vm_flags & VM_SHARED)) {
> +				ret = hugetlb_cow(mm, vma, address, ptep,
> +						pagecache_page, ptl);
> +				goto out_put_page;
> +			}
> +
> +			/* write protected for soft dirty processing */
> +			if ((vma->vm_flags & VM_WRITE) &&

This VM_WRITE check seems to be redundant.  As example, do_user_addr_fault() of
x86 code will check this right after vma lookup by access_error().  So when
reach here if "flags & FAULT_FLAG_WRITE", then VM_WRITE must be set, imho.

> +					(vma->vm_flags & VM_SHARED))
> +				entry = huge_pte_mkwrite(entry);

Same question to VM_SHARED, since "(vma->vm_flags & VM_SHARED)" is just checked
above and we'll go hugetlb_cow() otherwise.

> +
> +			entry = huge_pte_mkdirty(entry);

There's another huge_pte_mkdirty() right below; likely we could merge them somehow?

Thanks,

>  		}
>  		entry = huge_pte_mkdirty(entry);
>  	}
> -- 
> 2.29.2
> 

-- 
Peter Xu





[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux OMAP]     [Linux MIPS]     [eCos]     [Asterisk Internet PBX]     [Linux API]

  Powered by Linux