Messages in this thread | | | Date | Tue, 10 Oct 2023 10:31:51 +0800 | Subject | Re: [PATCH v4 2/2] mm: Init page count in reserve_bootmem_region when MEMINIT_EARLY | From | Yajun Deng <> |
| |
On 2023/10/8 16:57, Yajun Deng wrote: > > On 2023/10/2 16:30, David Hildenbrand wrote: >> On 29.09.23 10:30, Mike Rapoport wrote: >>> On Thu, Sep 28, 2023 at 04:33:02PM +0800, Yajun Deng wrote: >>>> memmap_init_range() would init page count of all pages, but the free >>>> pages count would be reset in __free_pages_core(). There are opposite >>>> operations. It's unnecessary and time-consuming when it's >>>> MEMINIT_EARLY >>>> context. >>>> >>>> Init page count in reserve_bootmem_region when in MEMINIT_EARLY >>>> context, >>>> and check the page count before reset it. >>>> >>>> At the same time, the INIT_LIST_HEAD in reserve_bootmem_region isn't >>>> need, as it already done in __init_single_page. >>>> >>>> The following data was tested on an x86 machine with 190GB of RAM. >>>> >>>> before: >>>> free_low_memory_core_early() 341ms >>>> >>>> after: >>>> free_low_memory_core_early() 285ms >>>> >>>> Signed-off-by: Yajun Deng <yajun.deng@linux.dev> >>>> --- >>>> v4: same with v2. >>>> v3: same with v2. >>>> v2: check page count instead of check context before reset it. >>>> v1: >>>> https://lore.kernel.org/all/20230922070923.355656-1-yajun.deng@linux.dev/ >>>> --- >>>> mm/mm_init.c | 18 +++++++++++++----- >>>> mm/page_alloc.c | 20 ++++++++++++-------- >>>> 2 files changed, 25 insertions(+), 13 deletions(-) >>>> >>>> diff --git a/mm/mm_init.c b/mm/mm_init.c >>>> index 9716c8a7ade9..3ab8861e1ef3 100644 >>>> --- a/mm/mm_init.c >>>> +++ b/mm/mm_init.c >>>> @@ -718,7 +718,7 @@ static void __meminit >>>> init_reserved_page(unsigned long pfn, int nid) >>>> if (zone_spans_pfn(zone, pfn)) >>>> break; >>>> } >>>> - __init_single_page(pfn_to_page(pfn), pfn, zid, nid, >>>> INIT_PAGE_COUNT); >>>> + __init_single_page(pfn_to_page(pfn), pfn, zid, nid, 0); >>>> } >>>> #else >>>> static inline void pgdat_set_deferred_range(pg_data_t *pgdat) {} >>>> @@ -756,8 +756,8 @@ void __meminit >>>> reserve_bootmem_region(phys_addr_t start, >>>> init_reserved_page(start_pfn, nid); >>>> - /* Avoid false-positive PageTail() */ >>>> - INIT_LIST_HEAD(&page->lru); >>>> + /* Init page count for reserved region */ >>> >>> Please add a comment that describes _why_ we initialize the page >>> count here. >>> >>>> + init_page_count(page); >>>> /* >>>> * no need for atomic set_bit because the struct >>>> @@ -888,9 +888,17 @@ void __meminit memmap_init_range(unsigned long >>>> size, int nid, unsigned long zone >>>> } >>>> page = pfn_to_page(pfn); >>>> - __init_single_page(page, pfn, zone, nid, INIT_PAGE_COUNT); >>>> - if (context == MEMINIT_HOTPLUG) >>>> + >>>> + /* If the context is MEMINIT_EARLY, we will init page >>>> count and >>>> + * mark page reserved in reserve_bootmem_region, the free >>>> region >>>> + * wouldn't have page count and we will check the pages count >>>> + * in __free_pages_core. >>>> + */ >>>> + __init_single_page(page, pfn, zone, nid, 0); >>>> + if (context == MEMINIT_HOTPLUG) { >>>> + init_page_count(page); >>>> __SetPageReserved(page); >>> >>> Rather than calling init_page_count() and __SetPageReserved() for >>> MEMINIT_HOTPLUG you can set flags to INIT_PAGE_COUNT | >>> INIT_PAGE_RESERVED >>> an call __init_single_page() after the check for MEMINIT_HOTPLUG. >>> >>> But more generally, I wonder if we have to differentiate HOTPLUG >>> here at all. >>> @David, can you comment please? >> >> There are a lot of details to that, and I'll share some I can briefly >> think of. >> >> 1) __SetPageReserved() >> >> I tried removing that a while ago, but there was a blocker (IIRC >> something about >> ZONE_DEVICE). I still have the patches at [1] and I could probably >> take a look >> if that blocker still exists (I recall that something changed at some >> point, but >> I never had the time to follow up). >> >> But once we stop setting the pages reserved, we might run into issues >> with ... >> >> >> 2) init_page_count() >> >> virtio-mem, XEN balloon and HV-balloon add memory blocks that can >> contain holes. >> set_online_page_callback() is used to intercept memory onlining and >> to expose >> only the pages that are not holes to the buddy: calling >> generic_online_page() on !hole. >> >> Holes are PageReserved but with an initialized page count. Memory >> offlining will fail on >> PageReserved pages -- has_unmovable_pages(). >> >> >> At least virtio-mem clears the PageReserved flag of holes when >> onlining memory, >> and currently relies in the page count to be reasonable (so memory >> offlining can work). >> >> static void virtio_mem_set_fake_offline(unsigned long pfn, >> unsigned long nr_pages, bool onlined) >> { >> page_offline_begin(); >> for (; nr_pages--; pfn++) { >> struct page *page = pfn_to_page(pfn); >> >> __SetPageOffline(page); >> if (!onlined) { >> SetPageDirty(page); >> /* FIXME: remove after cleanups */ >> ClearPageReserved(page); >> } >> } >> page_offline_end(); >> } >> >> >> For virtio-mem, we could initialize the page count there instead. The >> other PV drivers >> might require a bit more thought. >> >> >> [1] >> https://github.com/davidhildenbrand/linux/tree/online_reserved_cleanup >> >>> >>>> + } >>>> /* >>>> * Usually, we want to mark the pageblock MIGRATE_MOVABLE, >>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c >>>> index 06be8821d833..b868caabe8dc 100644 >>>> --- a/mm/page_alloc.c >>>> +++ b/mm/page_alloc.c >>>> @@ -1285,18 +1285,22 @@ void __free_pages_core(struct page *page, >>>> unsigned int order) >>>> unsigned int loop; >>>> /* >>>> - * When initializing the memmap, __init_single_page() sets the >>>> refcount >>>> - * of all pages to 1 ("allocated"/"not free"). We have to set the >>>> - * refcount of all involved pages to 0. >>>> + * When initializing the memmap, memmap_init_range sets the >>>> refcount >>>> + * of all pages to 1 ("reserved" and "free") in hotplug >>>> context. We >>>> + * have to set the refcount of all involved pages to 0. >>>> Otherwise, >>>> + * we don't do it, as reserve_bootmem_region only set the >>>> refcount on >>>> + * reserve region ("reserved") in early context. >>>> */ >>> >>> Again, why hotplug and early init should be different? >>> >>>> - prefetchw(p); >>>> - for (loop = 0; loop < (nr_pages - 1); loop++, p++) { >>>> - prefetchw(p + 1); >>>> + if (page_count(page)) { >>>> + prefetchw(p); >>>> + for (loop = 0; loop < (nr_pages - 1); loop++, p++) { >>>> + prefetchw(p + 1); >>>> + __ClearPageReserved(p); >>>> + set_page_count(p, 0); >>>> + } >>>> __ClearPageReserved(p); >>>> set_page_count(p, 0); >> >> That looks wrong. if the page count would by pure luck be 0 already >> for hotplugged memory, >> you wouldn't clear the reserved flag. >> >> These changes make me a bit nervous. > > > Is 'if (page_count(page) || PageReserved(page))' be safer? Or do I > need to do something else? >
How about the following if statement? But it needs to add more patch like v1 ([PATCH 2/4] mm: Introduce MEMINIT_LATE context).
It'll be safer, but more complex. Please comment...
if (context != MEMINIT_EARLY || (page_count(page) || PageReserved(page)) {
| |