On 13 Aug 2025, at 17:12, Wei Yang wrote:
> On Tue, Aug 12, 2025 at 11:55:10AM -0400, Zi Yan wrote:
> [...]
>> +/*
>> + * gather_folio_orders - scan through [vaddr_start, len) and record folio
>> orders
>> + * @vaddr_start: start vaddr
>> + * @len: range length
>> + * @pagemap_fd: file descriptor to /proc/<pid>/pagemap
>> + * @kpageflags_fd: file descriptor to /proc/kpageflags
>> + * @orders: output folio order array
>> + * @nr_orders: folio order array size
>> + *
>> + * gather_folio_orders() scan through [vaddr_start, len) and check all
>> folios
>> + * within the range and record their orders. All order-0 pages will be
>> recorded.
>
> I feel a little confused about the description here. Especially on the
> behavior when the range is not aligned on folio boundary.
I was too ambitious on this function. It is intended to just check after
split folio orders. I will move the function to split_huge_page_test.c
and rename it to gather_after_split_folio_orders() and
check_after_split_folio_orders().
>
> See following code at 1) and 2).
>
>> + * Non-present vaddr is skipped.
>> + *
>> + *
>> + * Return: 0 - no error, -1 - unhandled cases
>> + */
>> +static int gather_folio_orders(char *vaddr_start, size_t len,
>> + int pagemap_fd, int kpageflags_fd,
>> + int orders[], int nr_orders)
>> +{
>> + uint64_t page_flags = 0;
>> + int cur_order = -1;
>> + char *vaddr;
>> +
>> + if (!pagemap_fd || !kpageflags_fd)
>> + return -1;
>
> If my understanding is correct, we use open() to get a file descriptor.
>
> On error it returns -1. And 0 is a possible valid value, but usually used by
> stdin. The code may work in most cases, but seems not right.
Will fix it to
if (pagemap_fd == -1 || kpageflags_fd == -1)
>
>> + if (nr_orders <= 0)
>> + return -1;
>> +
>
> Maybe we want to check orders[] here too?
>
>> + for (vaddr = vaddr_start; vaddr < vaddr_start + len;) {
>> + char *next_folio_vaddr;
>> + int status;
>> +
>> + status = get_page_flags(vaddr, pagemap_fd, kpageflags_fd,
>> + &page_flags);
>> + if (status < 0)
>> + return -1;
>> +
>> + /* skip non present vaddr */
>> + if (status == 1) {
>> + vaddr += psize();
>> + continue;
>> + }
>> +
>> + /* all order-0 pages with possible false postive (non folio) */
>
> Do we still false positive case? Non-present page returns 1, which is handled
> above.
Any order-0 non folio will be counted, like GFP_KERNEL pages.
>
>> + if (!(page_flags & (KPF_COMPOUND_HEAD | KPF_COMPOUND_TAIL))) {
>> + orders[0]++;
>> + vaddr += psize();
>> + continue;
>> + }
>> +
>> + /* skip non thp compound pages */
>> + if (!(page_flags & KPF_THP)) {
>> + vaddr += psize();
>> + continue;
>> + }
>> +
>> + /* vpn points to part of a THP at this point */
>> + if (page_flags & KPF_COMPOUND_HEAD)
>> + cur_order = 1;
>> + else {
>> + /* not a head nor a tail in a THP? */
>> + if (!(page_flags & KPF_COMPOUND_TAIL))
>> + return -1;
>
> When reaches here, we know (page_flags & (KPF_COMPOUND_HEAD |
> KPF_COMPOUND_TAIL)).
> So we have at least one of it set.
>
> Looks not possible to hit it?
Will remove it.
>
>> +
>> + vaddr += psize();
>> + continue;
>
> 1)
>
> In case vaddr points to the middle of a large folio, this will skip this folio
> and count from next one.
>
>> + }
>> +
>> + next_folio_vaddr = vaddr + (1UL << (cur_order + pshift()));
>> +
>> + if (next_folio_vaddr >= vaddr_start + len)
>> + break;
>> +
>> + while ((status = get_page_flags(next_folio_vaddr, pagemap_fd,
>> + kpageflags_fd,
>> + &page_flags)) >= 0) {
>> + /*
>> + * non present vaddr, next compound head page, or
>> + * order-0 page
>> + */
>> + if (status == 1 ||
>> + (page_flags & KPF_COMPOUND_HEAD) ||
>> + !(page_flags & (KPF_COMPOUND_HEAD |
>> KPF_COMPOUND_TAIL))) {
>> + if (cur_order < nr_orders) {
>> + orders[cur_order]++;
>> + cur_order = -1;
>> + vaddr = next_folio_vaddr;
>> + }
>> + break;
>> + }
>> +
>> + /* not a head nor a tail in a THP? */
>> + if (!(page_flags & KPF_COMPOUND_TAIL))
>> + return -1;
>> +
>> + cur_order++;
>> + next_folio_vaddr = vaddr + (1UL << (cur_order +
>> pshift()));
>
> 2)
>
> If (vaddr_start + len) points to the middle of a large folio and folio is more
> than order 1 size, we may continue the loop and still count this last folio.
> Because we don't check next_folio_vaddr and (vaddr_start + len).
>
> A simple chart of these case.
>
> vaddr_start + len
> | |
> v v
> +---------------------+ +-----------------+
> |folio 1 | |folio 2 |
> +---------------------+ +-----------------+
>
> folio 1 is not counted, but folio 2 is counted.
>
> So at 1) and 2) handles the boundary differently. Not sure this is designed
> behavior. If so I think it would be better to record in document, otherwise
> the behavior is not obvious to user.
Will document it.
>
>> + }
>> +
>> + if (status < 0)
>> + return status;
>> + }
>> + if (cur_order > 0 && cur_order < nr_orders)
>> + orders[cur_order]++;
>
> Another boundary case here.
>
> If we come here because (next_folio_vaddr >= vaddr_start + len) in the for
> loop instead of the while loop. This means we found the folio head at vaddr,
> but the left range (vaddr_start + len - vaddr) is less than or equal to order
> 1 page size.
>
> But we haven't detected the real end of this folio. If this folio is more than
> order 1 size, we still count it an order 1 folio.
Yes. Will document it.
Thanks for the review.
>
>> + return 0;
>> +}
>> +
>> +int check_folio_orders(char *vaddr_start, size_t len, int pagemap_fd,
>> + int kpageflags_fd, int orders[], int nr_orders)
>> +{
>> + int *vaddr_orders;
>> + int status;
>> + int i;
>> +
>> + vaddr_orders = (int *)malloc(sizeof(int) * nr_orders);
>> +
>> + if (!vaddr_orders)
>> + ksft_exit_fail_msg("Cannot allocate memory for vaddr_orders");
>> +
>> + memset(vaddr_orders, 0, sizeof(int) * nr_orders);
>> + status = gather_folio_orders(vaddr_start, len, pagemap_fd,
>> + kpageflags_fd, vaddr_orders, nr_orders);
>> + if (status)
>> + goto out;
>> +
>> + status = 0;
>> + for (i = 0; i < nr_orders; i++)
>> + if (vaddr_orders[i] != orders[i]) {
>> + ksft_print_msg("order %d: expected: %d got %d\n", i,
>> + orders[i], vaddr_orders[i]);
>> + status = -1;
>> + }
>> +
>> +out:
>> + free(vaddr_orders);
>> + return status;
>> +}
>
> --
> Wei Yang
> Help you, Help me
Best Regards,
Yan, Zi
