> On Aug 9, 2022, at 6:33 AM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> >
> On Mon, Aug 08, 2022 at 06:49:50PM -0700, alexlzhu@xxxxxx wrote:
>> +int thp_number_utilized_pages(struct page *page)
>> +{
>> + unsigned long page_index, page_offset, value;
>> + int thp_nr_utilized_pages = HPAGE_PMD_NR;
>> + int step_size = sizeof(unsigned long);
>> + bool is_all_zeroes;
>> + void *kaddr;
>> +
>> + if (!page || !is_anon_transparent_hugepage(page))
>> + return -1;
>> +
>> + kaddr = kmap_local_page(compound_head(page));
>> + for (page_index = 0; page_index < HPAGE_PMD_NR; page_index++) {
>> + is_all_zeroes = true;
>> + for (page_offset = 0; page_offset < PAGE_SIZE; page_offset += step_size) {
>> + value = *(unsigned long *)(kaddr + page_index * PAGE_SIZE + page_offset);
>> + if (value != 0) {
>> + is_all_zeroes = false;
>> + break;
>> + }
>> + }
>> + if (is_all_zeroes)
>> + thp_nr_utilized_pages--;
>> + }
>> + kunmap_local(kaddr);
>
> You haven't tested this on 32-bit maachines with HIGHMEM, have you?
> kmap_local_page() only maps PAGE_SIZE bytes, so you need to map it and
> unmap each page. Also, why are you calling compound_head()? Surely
> you're already assuming that page is a head page here?
>
> (this is why I introduced the struct folio type, so you know whether
> a pointer is already to a head page and you don't need to call
> compound_head() again).
>
I see. Thanks for catching this! I have reimplemented with folio in a way that will work on 32 bit
architectures with HIGHMEM. Will send out v3.
