Re: [RFC PATCH] page_ext: create page extension for all memblock memory regions

[Jaewon Kim <jaewon31.kim@gmail.com>]

64
59

2022년 5월 17일 (화) 오후 9:55, Mike Rapoport <rppt@linux.ibm.com>님이 작성:
>
> On Tue, May 17, 2022 at 08:38:18PM +0900, Jaewon Kim wrote:
> > Hello Mike Rapoport
> > Thank you for your comment.
> >
> > Oh really? Could you point out the code or the commit regarding 'all
> > struct pages in any section should be valid and
> > properly initialized' ?
>
> There were several commits that refactored the memory map initialization,
> freeing of the unused memory map and abuse of pfn_valid() as a substitute
> of "is memory valid" semantics.
>
> > Actually I am using v5.10 based source tree on an arm64 device.
>
> Then most probably your change is not relevant for the upstream kernel.
> Did you observe any issues with page_ext initialization on v5.18-rcN
> kernels?

Actually I observed only 59 sections were initialized for page_ext and
missed 5 sections.
It should be totally 64 sections * 128 MB = 8,192 MB

>
> > I tried to look up and found the following commit in v5.16-rc1, did
> > you mean this?
> > 3de360c3fdb3 arm64/mm: drop HAVE_ARCH_PFN_VALID
>
> Yes, this is one of those commits.
>
> > I guess memblock_is_memory code in pfn_valid in arch/arm64/mm/init.c, v5.10
> > might affect the page_ext_init.
>
> Yes. In 5.10 the pfn_valid() test in page_ext_init() will skip an entire
> section if the first pfn in that section is not memory that can be mapped
> in the linear map.
>
> But again, this should be fixed in the latest kernels.

Great! Thank you for your explanation.
I will check it someday later when I use the latest kernel on our devices.
The next version on our devices seems to be v5.15 though.

Thank you
Jaewon Kim

>
> > Thank you
> >
> > 2022년 5월 17일 (화) 오후 5:25, Mike Rapoport <rppt@linux.ibm.com>님이 작성:
> > >
> > > On Mon, May 16, 2022 at 05:33:21PM -0700, Andrew Morton wrote:
> > > > On Mon,  9 May 2022 16:43:30 +0900 Jaewon Kim <jaewon31.kim@samsung.com> wrote:
> > > >
> > > > > The page extension can be prepared for each section. But if the first
> > > > > page is not valid, the page extension for the section was not
> > > > > initialized though there were many other valid pages within the section.
> > >
> > > What do you mean by "first page [in a section] is not valid"?
> > > In recent kernels all struct pages in any section should be valid and
> > > properly initialized.
> > >
> > > > > To support the page extension for all sections, refer to memblock memory
> > > > > regions. If the page is valid use the nid from pfn_to_nid, otherwise use
> > > > > the previous nid.
> > > > >
> > > > > Also this pagech changed log to include total sections and a section
> > > > > size.
> > > > >
> > > > > i.e.
> > > > > allocated 100663296 bytes of page_ext for 64 sections (1 section : 0x8000000)
> > > >
> > > > Cc Joonsoo, who wrote this code.
> > > > Cc Mike, for memblock.
> > > >
> > > > Thanks.
> > > >
> > > > >
> > > > > diff --git a/mm/page_ext.c b/mm/page_ext.c
> > > > > index 2e66d934d63f..506d58b36a1d 100644
> > > > > --- a/mm/page_ext.c
> > > > > +++ b/mm/page_ext.c
> > > > > @@ -381,41 +381,43 @@ static int __meminit page_ext_callback(struct notifier_block *self,
> > > > >  void __init page_ext_init(void)
> > > > >  {
> > > > >     unsigned long pfn;
> > > > > -   int nid;
> > > > > +   int nid = 0;
> > > > > +   struct memblock_region *rgn;
> > > > > +   int nr_section = 0;
> > > > > +   unsigned long next_section_pfn = 0;
> > > > >
> > > > >     if (!invoke_need_callbacks())
> > > > >             return;
> > > > >
> > > > > -   for_each_node_state(nid, N_MEMORY) {
> > > > > +   /*
> > > > > +    * iterate each memblock memory region and do not skip a section having
> > > > > +    * !pfn_valid(pfn)
> > > > > +    */
> > > > > +   for_each_mem_region(rgn) {
> > > > >             unsigned long start_pfn, end_pfn;
> > > > >
> > > > > -           start_pfn = node_start_pfn(nid);
> > > > > -           end_pfn = node_end_pfn(nid);
> > > > > -           /*
> > > > > -            * start_pfn and end_pfn may not be aligned to SECTION and the
> > > > > -            * page->flags of out of node pages are not initialized.  So we
> > > > > -            * scan [start_pfn, the biggest section's pfn < end_pfn) here.
> > > > > -            */
> > > > > +           start_pfn = (unsigned long)(rgn->base >> PAGE_SHIFT);
> > > > > +           end_pfn = start_pfn + (unsigned long)(rgn->size >> PAGE_SHIFT);
> > > > > +
> > > > > +           if (start_pfn < next_section_pfn)
> > > > > +                   start_pfn = next_section_pfn;
> > > > > +
> > > > >             for (pfn = start_pfn; pfn < end_pfn;
> > > > >                     pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
> > > > >
> > > > > -                   if (!pfn_valid(pfn))
> > > > > -                           continue;
> > > > > -                   /*
> > > > > -                    * Nodes's pfns can be overlapping.
> > > > > -                    * We know some arch can have a nodes layout such as
> > > > > -                    * -------------pfn-------------->
> > > > > -                    * N0 | N1 | N2 | N0 | N1 | N2|....
> > > > > -                    */
> > > > > -                   if (pfn_to_nid(pfn) != nid)
> > > > > -                           continue;
> > > > > +                   if (pfn_valid(pfn))
> > > > > +                           nid = pfn_to_nid(pfn);
> > > > > +                   nr_section++;
> > > > >                     if (init_section_page_ext(pfn, nid))
> > > > >                             goto oom;
> > > > >                     cond_resched();
> > > > >             }
> > > > > +           next_section_pfn = pfn;
> > > > >     }
> > > > > +
> > > > >     hotplug_memory_notifier(page_ext_callback, 0);
> > > > > -   pr_info("allocated %ld bytes of page_ext\n", total_usage);
> > > > > +   pr_info("allocated %ld bytes of page_ext for %d sections (1 section : 0x%x)\n",
> > > > > +           total_usage, nr_section, (1 << SECTION_SIZE_BITS));
> > > > >     invoke_init_callbacks();
> > > > >     return;
> > > > >
> > > > > --
> > > > > 2.17.1
> > > > >
> > >
> > > --
> > > Sincerely yours,
> > > Mike.
>
> --
> Sincerely yours,
> Mike.

2022년 5월 17일 (화) 오후 9:55, Mike Rapoport <rppt@linux.ibm.com>님이 작성:
>
> On Tue, May 17, 2022 at 08:38:18PM +0900, Jaewon Kim wrote:
> > Hello Mike Rapoport
> > Thank you for your comment.
> >
> > Oh really? Could you point out the code or the commit regarding 'all
> > struct pages in any section should be valid and
> > properly initialized' ?
>
> There were several commits that refactored the memory map initialization,
> freeing of the unused memory map and abuse of pfn_valid() as a substitute
> of "is memory valid" semantics.
>
> > Actually I am using v5.10 based source tree on an arm64 device.
>
> Then most probably your change is not relevant for the upstream kernel.
> Did you observe any issues with page_ext initialization on v5.18-rcN
> kernels?
>
> > I tried to look up and found the following commit in v5.16-rc1, did
> > you mean this?
> > 3de360c3fdb3 arm64/mm: drop HAVE_ARCH_PFN_VALID
>
> Yes, this is one of those commits.
>
> > I guess memblock_is_memory code in pfn_valid in arch/arm64/mm/init.c, v5.10
> > might affect the page_ext_init.
>
> Yes. In 5.10 the pfn_valid() test in page_ext_init() will skip an entire
> section if the first pfn in that section is not memory that can be mapped
> in the linear map.
>
> But again, this should be fixed in the latest kernels.
>
> > Thank you
> >
> > 2022년 5월 17일 (화) 오후 5:25, Mike Rapoport <rppt@linux.ibm.com>님이 작성:
> > >
> > > On Mon, May 16, 2022 at 05:33:21PM -0700, Andrew Morton wrote:
> > > > On Mon,  9 May 2022 16:43:30 +0900 Jaewon Kim <jaewon31.kim@samsung.com> wrote:
> > > >
> > > > > The page extension can be prepared for each section. But if the first
> > > > > page is not valid, the page extension for the section was not
> > > > > initialized though there were many other valid pages within the section.
> > >
> > > What do you mean by "first page [in a section] is not valid"?
> > > In recent kernels all struct pages in any section should be valid and
> > > properly initialized.
> > >
> > > > > To support the page extension for all sections, refer to memblock memory
> > > > > regions. If the page is valid use the nid from pfn_to_nid, otherwise use
> > > > > the previous nid.
> > > > >
> > > > > Also this pagech changed log to include total sections and a section
> > > > > size.
> > > > >
> > > > > i.e.
> > > > > allocated 100663296 bytes of page_ext for 64 sections (1 section : 0x8000000)
> > > >
> > > > Cc Joonsoo, who wrote this code.
> > > > Cc Mike, for memblock.
> > > >
> > > > Thanks.
> > > >
> > > > >
> > > > > diff --git a/mm/page_ext.c b/mm/page_ext.c
> > > > > index 2e66d934d63f..506d58b36a1d 100644
> > > > > --- a/mm/page_ext.c
> > > > > +++ b/mm/page_ext.c
> > > > > @@ -381,41 +381,43 @@ static int __meminit page_ext_callback(struct notifier_block *self,
> > > > >  void __init page_ext_init(void)
> > > > >  {
> > > > >     unsigned long pfn;
> > > > > -   int nid;
> > > > > +   int nid = 0;
> > > > > +   struct memblock_region *rgn;
> > > > > +   int nr_section = 0;
> > > > > +   unsigned long next_section_pfn = 0;
> > > > >
> > > > >     if (!invoke_need_callbacks())
> > > > >             return;
> > > > >
> > > > > -   for_each_node_state(nid, N_MEMORY) {
> > > > > +   /*
> > > > > +    * iterate each memblock memory region and do not skip a section having
> > > > > +    * !pfn_valid(pfn)
> > > > > +    */
> > > > > +   for_each_mem_region(rgn) {
> > > > >             unsigned long start_pfn, end_pfn;
> > > > >
> > > > > -           start_pfn = node_start_pfn(nid);
> > > > > -           end_pfn = node_end_pfn(nid);
> > > > > -           /*
> > > > > -            * start_pfn and end_pfn may not be aligned to SECTION and the
> > > > > -            * page->flags of out of node pages are not initialized.  So we
> > > > > -            * scan [start_pfn, the biggest section's pfn < end_pfn) here.
> > > > > -            */
> > > > > +           start_pfn = (unsigned long)(rgn->base >> PAGE_SHIFT);
> > > > > +           end_pfn = start_pfn + (unsigned long)(rgn->size >> PAGE_SHIFT);
> > > > > +
> > > > > +           if (start_pfn < next_section_pfn)
> > > > > +                   start_pfn = next_section_pfn;
> > > > > +
> > > > >             for (pfn = start_pfn; pfn < end_pfn;
> > > > >                     pfn = ALIGN(pfn + 1, PAGES_PER_SECTION)) {
> > > > >
> > > > > -                   if (!pfn_valid(pfn))
> > > > > -                           continue;
> > > > > -                   /*
> > > > > -                    * Nodes's pfns can be overlapping.
> > > > > -                    * We know some arch can have a nodes layout such as
> > > > > -                    * -------------pfn-------------->
> > > > > -                    * N0 | N1 | N2 | N0 | N1 | N2|....
> > > > > -                    */
> > > > > -                   if (pfn_to_nid(pfn) != nid)
> > > > > -                           continue;
> > > > > +                   if (pfn_valid(pfn))
> > > > > +                           nid = pfn_to_nid(pfn);
> > > > > +                   nr_section++;
> > > > >                     if (init_section_page_ext(pfn, nid))
> > > > >                             goto oom;
> > > > >                     cond_resched();
> > > > >             }
> > > > > +           next_section_pfn = pfn;
> > > > >     }
> > > > > +
> > > > >     hotplug_memory_notifier(page_ext_callback, 0);
> > > > > -   pr_info("allocated %ld bytes of page_ext\n", total_usage);
> > > > > +   pr_info("allocated %ld bytes of page_ext for %d sections (1 section : 0x%x)\n",
> > > > > +           total_usage, nr_section, (1 << SECTION_SIZE_BITS));
> > > > >     invoke_init_callbacks();
> > > > >     return;
> > > > >
> > > > > --
> > > > > 2.17.1
> > > > >
> > >
> > > --
> > > Sincerely yours,
> > > Mike.
>
> --
> Sincerely yours,
> Mike.