On Fri, 2018-12-07 at 18:08 +0800, Wei Yang wrote:
> After commit c9e97a1997fb ("mm: initialize pages on demand during
> boot"), the behavior of DEFERRED_STRUCT_PAGE_INIT is changed to
> initialize first section for highest zone on each node.
>
> Instead of test each pfn during iteration, we could calculate the
> first_deferred_pfn directly with necessary information.
>
> By doing so, we also get some performance benefit during bootup:
>
> +----------+-----------+-----------+--------+
> | |Base |Patched |Gain |
> +----------+-----------+-----------+--------+
> | 1 Node |0.011993 |0.011459 |-4.45% |
> +----------+-----------+-----------+--------+
> | 4 Nodes |0.006466 |0.006255 |-3.26% |
> +----------+-----------+-----------+--------+
>
> Test result is retrieved from dmesg time stamp by add printk around
> free_area_init_nodes().
>
> Signed-off-by: Wei Yang <richard.weiyang@xxxxxxxxx>
I'm pretty sure the fundamental assumption made in this patch is wrong.
It is assuming that the first deferred PFN will just be your start PFN
+ PAGES_PER_SECTION aligned to the nearest PAGES_PER_SECTION, do I have
that correct?
If I am not mistaken that can result in scenarios where you actually
start out with 0 pages allocated if your first section is in a span
belonging to another node, or is reserved memory for things like MMIO.
Ideally we don't want to do that as we have to immediately jump into
growing the zone with the code as it currently stands.
> ---
> mm/page_alloc.c | 57 +++++++++++++++++++++++++++------------------------------
> 1 file changed, 27 insertions(+), 30 deletions(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index baf473f80800..5f077bf07f3e 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -306,38 +306,33 @@ static inline bool __meminit early_page_uninitialised(unsigned long pfn)
> }
>
> /*
> - * Returns true when the remaining initialisation should be deferred until
> - * later in the boot cycle when it can be parallelised.
> + * Calculate first_deferred_pfn in case:
> + * - in MEMMAP_EARLY context
> + * - this is the last zone
> + *
> + * If the first aligned section doesn't exceed the end_pfn, set it to
> + * first_deferred_pfn and return it.
> */
> -static bool __meminit
> -defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
> +unsigned long __meminit
> +defer_pfn(int nid, unsigned long start_pfn, unsigned long end_pfn,
> + enum memmap_context context)
> {
> - static unsigned long prev_end_pfn, nr_initialised;
> + struct pglist_data *pgdat = NODE_DATA(nid);
> + unsigned long pfn;
>
> - /*
> - * prev_end_pfn static that contains the end of previous zone
> - * No need to protect because called very early in boot before smp_init.
> - */
> - if (prev_end_pfn != end_pfn) {
> - prev_end_pfn = end_pfn;
> - nr_initialised = 0;
> - }
> + if (context != MEMMAP_EARLY)
> + return end_pfn;
>
> - /* Always populate low zones for address-constrained allocations */
> - if (end_pfn < pgdat_end_pfn(NODE_DATA(nid)))
> - return false;
> + /* Always populate low zones */
> + if (end_pfn < pgdat_end_pfn(pgdat))
> + return end_pfn;
>
> - /*
> - * We start only with one section of pages, more pages are added as
> - * needed until the rest of deferred pages are initialized.
> - */
> - nr_initialised++;
> - if ((nr_initialised > PAGES_PER_SECTION) &&
> - (pfn & (PAGES_PER_SECTION - 1)) == 0) {
> - NODE_DATA(nid)->first_deferred_pfn = pfn;
> - return true;
> + pfn = roundup(start_pfn + PAGES_PER_SECTION - 1, PAGES_PER_SECTION);
> + if (end_pfn > pfn) {
> + pgdat->first_deferred_pfn = pfn;
> + end_pfn = pfn;
> }
> - return false;
> + return end_pfn;
Okay so I stand corrected. It looks like you are rounding up by
(PAGES_PER_SECTION - 1) * 2 since if I am not mistaken roundup should
do the same math you already did in side the function.
> }
> #else
> static inline bool early_page_uninitialised(unsigned long pfn)
> @@ -345,9 +340,11 @@ static inline bool early_page_uninitialised(unsigned long pfn)
> return false;
> }
>
> -static inline bool defer_init(int nid, unsigned long pfn, unsigned long end_pfn)
> +unsigned long __meminit
> +defer_pfn(int nid, unsigned long start_pfn, unsigned long end_pfn,
> + enum memmap_context context)
> {
> - return false;
> + return end_pfn;
> }
> #endif
>
> @@ -5514,6 +5511,8 @@ void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
> }
> #endif
>
> + end_pfn = defer_pfn(nid, start_pfn, end_pfn, context);
> +
A better approach for this might be to look at placing the loop within
a loop similar to how I handled this for the deferred init. You only
really need to be performing all of these checks once per section
aligned point anyway.
Basically if you added another loop and limited the loop below so that
you only fed it one section at a time then you could just pull the
defer_init check out of this section and place it in the outer loop
after you have already tried initializing at least one section worth of
pages.
You could probably also look at pulling in the logic that is currently
sitting at the end of the current function that is initializing things
until the end_pfn is aligned with PAGES_PER_SECTION.
> for (pfn = start_pfn; pfn < end_pfn; pfn++) {
> /*
> * There can be holes in boot-time mem_map[]s handed to this
> @@ -5526,8 +5525,6 @@ void __meminit memmap_init_zone(unsigned long size, int nid, unsigned long zone,
> continue;
> if (overlap_memmap_init(zone, &pfn))
> continue;
> - if (defer_init(nid, pfn, end_pfn))
> - break;
> }
So the whole reason for the "defer_init" call being placed here is
because there are checks to see if the prior PFN is valid, in our NUMA
node, or is an overlapping region. If your first section or in this
case 2 sections contain pages that fall into these categories you
aren't going to initialize any pages.
>
> page = pfn_to_page(pfn);
