Re: [RFC PATCH v2 7/9] mm: Add the break COW PTE handler

[Chih-En Lin <shiyn.lin@gmail.com>]

On Tue, Sep 27, 2022 at 06:15:34PM +0000, Nadav Amit wrote:
> On Sep 27, 2022, at 9:29 AM, Chih-En Lin <shiyn.lin@gmail.com> wrote:
> 
> > To handle the COW PTE with write fault, introduce the helper function
> > handle_cow_pte(). The function provides two behaviors. One is breaking
> > COW by decreasing the refcount, pgables_bytes, and RSS. Another is
> > copying all the information in the shared PTE table by using
> > copy_pte_page() with a wrapper.
> > 
> > Also, add the wrapper functions to help us find out the COWed or
> > COW-available PTE table.
> > 
> 
> [ snip ]
> 
> > +static inline int copy_cow_pte_range(struct vm_area_struct *vma,
> > +				     pmd_t *dst_pmd, pmd_t *src_pmd,
> > +				     unsigned long start, unsigned long end)
> > +{
> > +	struct mm_struct *mm = vma->vm_mm;
> > +	struct mmu_notifier_range range;
> > +	int ret;
> > +	bool is_cow;
> > +
> > +	is_cow = is_cow_mapping(vma->vm_flags);
> > +	if (is_cow) {
> > +		mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
> > +					0, vma, mm, start, end);
> > +		mmu_notifier_invalidate_range_start(&range);
> > +		mmap_assert_write_locked(mm);
> > +		raw_write_seqcount_begin(&mm->write_protect_seq);
> > +	}
> > +
> > +	ret = copy_pte_range(vma, vma, dst_pmd, src_pmd, start, end);
> > +
> > +	if (is_cow) {
> > +		raw_write_seqcount_end(&mm->write_protect_seq);
> > +		mmu_notifier_invalidate_range_end(&range);
> 
> Usually, I would expect mmu-notifiers and TLB flushes to be initiated at the
> same point in the code. Presumably you changed protection, so you do need a
> TLB flush, right? Is it done elsewhere?

You're right.
I will add TLB flushes here.
Thanks.

> > +	}
> > +
> > +	return ret;
> > +}
> > +
> > +/*
> > + * Break COW PTE, two state here:
> > + *   - After fork :   [parent, rss=1, ref=2, write=NO , owner=parent]
> > + *                 to [parent, rss=1, ref=1, write=YES, owner=NULL  ]
> > + *                    COW PTE become [ref=1, write=NO , owner=NULL  ]
> > + *                    [child , rss=0, ref=2, write=NO , owner=parent]
> > + *                 to [child , rss=1, ref=1, write=YES, owner=NULL  ]
> > + *                    COW PTE become [ref=1, write=NO , owner=parent]
> > + *   NOTE
> > + *     - Copy the COW PTE to new PTE.
> > + *     - Clear the owner of COW PTE and set PMD entry writable when it is owner.
> > + *     - Increase RSS if it is not owner.
> > + */
> > +static int break_cow_pte(struct vm_area_struct *vma, pmd_t *pmd,
> > +			 unsigned long addr)
> > +{
> > +	struct mm_struct *mm = vma->vm_mm;
> > +	unsigned long pte_start, pte_end;
> > +	unsigned long start, end;
> > +	struct vm_area_struct *prev = vma->vm_prev;
> > +	struct vm_area_struct *next = vma->vm_next;
> > +	pmd_t cowed_entry = *pmd;
> > +
> > +	if (cow_pte_count(&cowed_entry) == 1) {
> > +		cow_pte_fallback(vma, pmd, addr);
> > +		return 1;
> > +	}
> > +
> > +	pte_start = start = addr & PMD_MASK;
> > +	pte_end = end = (addr + PMD_SIZE) & PMD_MASK;
> > +
> > +	pmd_clear(pmd);
> > +	/*
> > +	 * If the vma does not cover the entire address range of the PTE table,
> > +	 * it should check the previous and next.
> > +	 */
> > +	if (start < vma->vm_start && prev) {
> > +		/* The part of address range is covered by previous. */
> > +		if (start < prev->vm_end)
> > +			copy_cow_pte_range(prev, pmd, &cowed_entry,
> > +					   start, prev->vm_end);
> > +		start = vma->vm_start;
> > +	}
> > +	if (end > vma->vm_end && next) {
> > +		/* The part of address range is covered by next. */
> > +		if (end > next->vm_start)
> > +			copy_cow_pte_range(next, pmd, &cowed_entry,
> > +					   next->vm_start, end);
> > +		end = vma->vm_end;
> > +	}
> > +	if (copy_cow_pte_range(vma, pmd, &cowed_entry, start, end))
> > +		return -ENOMEM;
> > +
> > +	/*
> > +	 * Here, it is the owner, so clear the ownership. To keep RSS state and
> > +	 * page table bytes correct, it needs to decrease them.
> > +	 * Also, handle the address range issue here.
> > +	 */
> > +	if (cow_pte_owner_is_same(&cowed_entry, pmd)) {
> > +		set_cow_pte_owner(&cowed_entry, NULL);
> 
> Presumably there is some assumption on atomicity here. Otherwise, two
> threads can run the following code, which is wrong, no? Yet, I do not see
> anything that provides such atomicity.

I may have multiple process access here. But for the thread, I assume
that they need to hold the mmap_lock. Maybe I need to add the assert
here too.

> 
> > +		if (pte_start < vma->vm_start && prev &&
> > +		    pte_start < prev->vm_end)
> > +			cow_pte_rss(mm, vma->vm_prev, pmd,
> > +				    pte_start, prev->vm_end, false /* dec */);
> > +		if (pte_end > vma->vm_end && next &&
> > +		    pte_end > next->vm_start)
> > +			cow_pte_rss(mm, vma->vm_next, pmd,
> > +				    next->vm_start, pte_end, false /* dec */);
> > +		cow_pte_rss(mm, vma, pmd, start, end, false /* dec */);
> > +		mm_dec_nr_ptes(mm);
> > +	}
> > +
> > +	/* Already handled it, don't reuse cowed table. */
> > +	pmd_put_pte(vma, &cowed_entry, addr, false);
> > +
> > +	VM_BUG_ON(cow_pte_count(pmd) != 1);
> 
> Don’t use VM_BUG_ON().

Sure. I will change it to VM_WARN_ON().

Thanks,
Chih-En Lin