Re: [v3 4/4] mm: hugetlb: Skip initialization of gigantic tail struct pages if freed by HVO

[Muchun Song <muchun.song@linux.dev>]

> On Aug 29, 2023, at 05:04, Mike Kravetz <mike.kravetz@oracle.com> wrote:
> 
> On 08/28/23 19:33, Muchun Song wrote:
>> 
>> 
>>> On Aug 25, 2023, at 19:18, Usama Arif <usama.arif@bytedance.com> wrote:
>>> 
>>> The new boot flow when it comes to initialization of gigantic pages
>>> is as follows:
>>> - At boot time, for a gigantic page during __alloc_bootmem_hugepage,
>>> the region after the first struct page is marked as noinit.
>>> - This results in only the first struct page to be
>>> initialized in reserve_bootmem_region. As the tail struct pages are
>>> not initialized at this point, there can be a significant saving
>>> in boot time if HVO succeeds later on.
>>> - Later on in the boot, HVO is attempted. If its successful, only the first
>>> HUGETLB_VMEMMAP_RESERVE_SIZE / sizeof(struct page) - 1 tail struct pages
>>> after the head struct page are initialized. If it is not successful,
>>> then all of the tail struct pages are initialized.
>>> 
>>> Signed-off-by: Usama Arif <usama.arif@bytedance.com>
>> 
>> This edition is simpler than before ever, thanks for your work.
>> 
>> There is premise that other subsystems do not access vmemmap pages
>> before the initialization of vmemmap pages associated withe HugeTLB
>> pages allocated from bootmem for your optimization. However, IIUC, the
>> compacting path could access arbitrary struct page when memory fails
>> to be allocated via buddy allocator. So we should make sure that
>> those struct pages are not referenced in this routine. And I know
>> if CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, it will encounter
>> the same issue, but I don't find any code to prevent this from
>> happening. I need more time to confirm this, if someone already knows,
>> please let me know, thanks. So I think HugeTLB should adopt the similar
>> way to prevent this.
> 
> In this patch, the call to hugetlb_vmemmap_optimize() is moved BEFORE
> __prep_new_hugetlb_folio or prep_new_hugetlb_folio in all code paths.
> The prep_new_hugetlb_folio routine(s) are what set the destructor (soon
> to be a flag) that identifies the set of pages as a hugetlb page.  So,
> there is now a window where a set of pages not identified as hugetlb
> will not have vmemmap pages.

Thanks for your point it out.

Seems this issue is not related to this change? hugetlb_vmemmap_optimize()
is called before the setting of destructor since the initial commit
f41f2ed43ca5. Right?

> 
> Recently, I closed the same window in the hugetlb freeing code paths with
> commit 32c877191e02 'hugetlb: do not clear hugetlb dtor until allocating'.

Yes, I saw it. 

> This patch needs to be reworked so that this window is not opened in the
> allocation paths.

So I think the fix should be a separate series.

Thanks.