Mel Gorman <mgorman@xxxxxxx> writes:
> On Wed, Mar 04, 2020 at 07:15:20PM +0800, Huang, Ying wrote:
>> In which situation the cost to reconstruct MADV_FREE pages can be higher
>> than the cost to allocate file cache page and read from disk? Heavy
>> contention on mmap_sem?
>>
>
> MADV_FREE should be anonymous only
>
> if (behavior == MADV_FREE)
> return madvise_free_single_vma(vma, start, end);
>
> .....
>
> static int madvise_free_single_vma(struct vm_area_struct *vma,
> unsigned long start_addr, unsigned long end_addr)
> {
> struct mm_struct *mm = vma->vm_mm;
> struct mmu_notifier_range range;
> struct mmu_gather tlb;
>
> /* MADV_FREE works for only anon vma at the moment */
> if (!vma_is_anonymous(vma))
> return -EINVAL
>
> So the question is not applicable. For anonymous memory, the cost of
> updating a PTE is lower than allocating a page, zeroing it and updating
> the PTE.
Sorry for confusing. The original question is NOT about comparing the
reconstruction cost between MADV_FREE anon pages and MADV_FREE file
pages, BUT about comparing the reconstruction cost between MADV_FREE
anon pages and ordinary clean file cache pages. You can find more
details in conversation between Michal and me.
> It has been repeatedly stated now for almost a week that a semantic
> change to MADV_FREE should be based on a problem encountered by a real
> application that can benefit from the new semantics. I think the only
> concrete outcome has been that userspace potentially benefits if the total
> number of MADV_FREE pages is reported globally. Even that is marginal as
> smaps has the information to tell the difference between high RSS due to
> a memory leak and high RSS usage due to MADV_FREE. The /proc/vmstats for
> MADV_FREE are of marginal benefit given that they do not tell us much
> about the current number of MADV_FREE pages in the system.
Got it! Thanks a lot for your patience and sharing. That's very
helpful.
Best Regards,
Huang, Ying
