Re: [PATCH 1/4] mm: Trial do_wp_page() simplification

From: Gal Pressman <galpress@amazon.com>
To: Peter Xu <peterx@redhat.com>, Jason Gunthorpe <jgg@ziepe.ca>
Cc: Linus Torvalds <torvalds@linux-foundation.org>,
	John Hubbard <jhubbard@nvidia.com>,
	Leon Romanovsky <leonro@nvidia.com>,
	Linux-MM <linux-mm@kvack.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	"Maya B . Gokhale" <gokhale2@llnl.gov>,
	Yang Shi <yang.shi@linux.alibaba.com>,
	Marty Mcfadden <mcfadden8@llnl.gov>,
	Kirill Shutemov <kirill@shutemov.name>,
	Oleg Nesterov <oleg@redhat.com>, Jann Horn <jannh@google.com>,
	Jan Kara <jack@suse.cz>, Kirill Tkhai <ktkhai@virtuozzo.com>,
	Andrea Arcangeli <aarcange@redhat.com>,
	Christoph Hellwig <hch@lst.de>,
	Andrew Morton <akpm@linux-foundation.org>,
	"Zhang, Wei" <wzam@amazon.com>
Subject: Re: [PATCH 1/4] mm: Trial do_wp_page() simplification
Date: Tue, 2 Feb 2021 16:40:33 +0200	[thread overview]
Message-ID: <27564187-4a08-f187-5a84-3df50009f6ca@amazon.com> (raw)

> On Thu, Sep 17, 2020 at 03:03:32PM -0400, Peter Xu wrote:
> 
>> Another side effect I can think of is that we'll bring some uncertainty to
>> fork() starting from when page_maybe_dma_pinned() is used, since it's sometimes
>> bogus (hpage_pincount_available()==false) so some COWs might be triggered
>> during fork() even when not necessary if we've got some normal pages with too
>> many refcounts (over GUP_PIN_COUNTING_BIAS).  But assuming that's not a big
>> deal since it should be extremely rare, or is it?..
> 
> Looking at this a bit more.. A complete implementation will have to
> touch all four places doing write protect during fork:
> 
> copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> 		pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
> 		unsigned long addr, int *rss)
> {
> [..]
> 	if (is_cow_mapping(vm_flags) && pte_write(pte)) {
> 		ptep_set_wrprotect(src_mm, addr, src_pte);
> 		pte = pte_wrprotect(pte);
> 
> int copy_huge_pmd(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> 		  pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
> 		  struct vm_area_struct *vma)
> {
> [..]
> 	pmdp_set_wrprotect(src_mm, addr, src_pmd);
> 	pmd = pmd_mkold(pmd_wrprotect(pmd));
> 
> int copy_huge_pud(struct mm_struct *dst_mm, struct mm_struct *src_mm,
> 		  pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
> 		  struct vm_area_struct *vma)
> {
> [..]
> 	pudp_set_wrprotect(src_mm, addr, src_pud);
> 	pud = pud_mkold(pud_wrprotect(pud));
> 
> int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
>                             struct vm_area_struct *vma)
> {
> [..]
>                         if (cow) {
>                                 huge_ptep_set_wrprotect(src, addr, src_pte);
> 
> As a regression I'm pretty sure we will hit only the PTE and PMD
> cases.
> 
> Most likely the other two could be done outside the rc cycle

Hi Peter & Jason,

It seems the hugetlb part was overlooked?
We're testing if the RDMA fork MADV_DONTFORK stuff can be removed on appropriate
kernels, but our tests still fail due to lacking explicit huge pages support [1].

Peter, was it left unchanged on purpose?
Are you planning to submit the hugetlb changes as well?

[1] https://github.com/linux-rdma/rdma-core/pull/883#issuecomment-770398171