mm: meminit: initialise remaining struct pages in parallel with kswapd
authorMel Gorman <mgorman@suse.de>
Tue, 30 Jun 2015 21:57:05 +0000 (14:57 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 1 Jul 2015 02:44:56 +0000 (19:44 -0700)
Only a subset of struct pages are initialised at the moment.  When this
patch is applied kswapd initialise the remaining struct pages in parallel.

This should boot faster by spreading the work to multiple CPUs and
initialising data that is local to the CPU.  The user-visible effect on
large machines is that free memory will appear to rapidly increase early
in the lifetime of the system until kswapd reports that all memory is
initialised in the kernel log.  Once initialised there should be no other
user-visibile effects.

Signed-off-by: Mel Gorman <mgorman@suse.de>
Tested-by: Nate Zimmer <nzimmer@sgi.com>
Tested-by: Waiman Long <waiman.long@hp.com>
Tested-by: Daniel J Blueman <daniel@numascale.com>
Acked-by: Pekka Enberg <penberg@kernel.org>
Cc: Robin Holt <robinmholt@gmail.com>
Cc: Nate Zimmer <nzimmer@sgi.com>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: Waiman Long <waiman.long@hp.com>
Cc: Scott Norton <scott.norton@hp.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/internal.h
mm/mm_init.c
mm/page_alloc.c
mm/vmscan.c

index 88ac7be741ca5459f8d7d1e951a7b0177fb213d3..71d160437205724f283a8ece5aa6139060fd6c9d 100644 (file)
@@ -400,9 +400,15 @@ static inline void mminit_verify_zonelist(void)
 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
 #define __defermem_init __meminit
 #define __defer_init    __meminit
+
+void deferred_init_memmap(int nid);
 #else
 #define __defermem_init
 #define __defer_init __init
+
+static inline void deferred_init_memmap(int nid)
+{
+}
 #endif
 
 /* mminit_validate_memmodel_limits is independent of CONFIG_DEBUG_MEMORY_INIT */
index 5f420f7fafa1c6b6bd74cdf3b6e9cfb5e7874f87..28fbf87b20aaf74479e9442e67d2937bb6a74207 100644 (file)
@@ -11,6 +11,7 @@
 #include <linux/export.h>
 #include <linux/memory.h>
 #include <linux/notifier.h>
+#include <linux/sched.h>
 #include "internal.h"
 
 #ifdef CONFIG_DEBUG_MEMORY_INIT
index 7af45b2e8870647e0ba98f6a62fc802667ec8881..c30f5a0535fd55b966ee86a9520bef55b5834606 100644 (file)
@@ -252,6 +252,14 @@ static inline bool __defermem_init early_page_uninitialised(unsigned long pfn)
        return false;
 }
 
+static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
+{
+       if (pfn >= NODE_DATA(nid)->first_deferred_pfn)
+               return true;
+
+       return false;
+}
+
 /*
  * Returns false when the remaining initialisation should be deferred until
  * later in the boot cycle when it can be parallelised.
@@ -284,6 +292,11 @@ static inline bool early_page_uninitialised(unsigned long pfn)
        return false;
 }
 
+static inline bool early_page_nid_uninitialised(unsigned long pfn, int nid)
+{
+       return false;
+}
+
 static inline bool update_defer_init(pg_data_t *pgdat,
                                unsigned long pfn, unsigned long zone_end,
                                unsigned long *nr_initialised)
@@ -866,20 +879,51 @@ static void __meminit __init_single_pfn(unsigned long pfn, unsigned long zone,
        return __init_single_page(pfn_to_page(pfn), pfn, zone, nid);
 }
 
+#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
+static void init_reserved_page(unsigned long pfn)
+{
+       pg_data_t *pgdat;
+       int nid, zid;
+
+       if (!early_page_uninitialised(pfn))
+               return;
+
+       nid = early_pfn_to_nid(pfn);
+       pgdat = NODE_DATA(nid);
+
+       for (zid = 0; zid < MAX_NR_ZONES; zid++) {
+               struct zone *zone = &pgdat->node_zones[zid];
+
+               if (pfn >= zone->zone_start_pfn && pfn < zone_end_pfn(zone))
+                       break;
+       }
+       __init_single_pfn(pfn, zid, nid);
+}
+#else
+static inline void init_reserved_page(unsigned long pfn)
+{
+}
+#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
+
 /*
  * Initialised pages do not have PageReserved set. This function is
  * called for each range allocated by the bootmem allocator and
  * marks the pages PageReserved. The remaining valid pages are later
  * sent to the buddy page allocator.
  */
-void reserve_bootmem_region(unsigned long start, unsigned long end)
+void __meminit reserve_bootmem_region(unsigned long start, unsigned long end)
 {
        unsigned long start_pfn = PFN_DOWN(start);
        unsigned long end_pfn = PFN_UP(end);
 
-       for (; start_pfn < end_pfn; start_pfn++)
-               if (pfn_valid(start_pfn))
-                       SetPageReserved(pfn_to_page(start_pfn));
+       for (; start_pfn < end_pfn; start_pfn++) {
+               if (pfn_valid(start_pfn)) {
+                       struct page *page = pfn_to_page(start_pfn);
+
+                       init_reserved_page(start_pfn);
+                       SetPageReserved(page);
+               }
+       }
 }
 
 static bool free_pages_prepare(struct page *page, unsigned int order)
@@ -1017,6 +1061,74 @@ void __defer_init __free_pages_bootmem(struct page *page, unsigned long pfn,
        return __free_pages_boot_core(page, pfn, order);
 }
 
+#ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
+/* Initialise remaining memory on a node */
+void __defermem_init deferred_init_memmap(int nid)
+{
+       struct mminit_pfnnid_cache nid_init_state = { };
+       unsigned long start = jiffies;
+       unsigned long nr_pages = 0;
+       unsigned long walk_start, walk_end;
+       int i, zid;
+       struct zone *zone;
+       pg_data_t *pgdat = NODE_DATA(nid);
+       unsigned long first_init_pfn = pgdat->first_deferred_pfn;
+
+       if (first_init_pfn == ULONG_MAX)
+               return;
+
+       /* Sanity check boundaries */
+       BUG_ON(pgdat->first_deferred_pfn < pgdat->node_start_pfn);
+       BUG_ON(pgdat->first_deferred_pfn > pgdat_end_pfn(pgdat));
+       pgdat->first_deferred_pfn = ULONG_MAX;
+
+       /* Only the highest zone is deferred so find it */
+       for (zid = 0; zid < MAX_NR_ZONES; zid++) {
+               zone = pgdat->node_zones + zid;
+               if (first_init_pfn < zone_end_pfn(zone))
+                       break;
+       }
+
+       for_each_mem_pfn_range(i, nid, &walk_start, &walk_end, NULL) {
+               unsigned long pfn, end_pfn;
+
+               end_pfn = min(walk_end, zone_end_pfn(zone));
+               pfn = first_init_pfn;
+               if (pfn < walk_start)
+                       pfn = walk_start;
+               if (pfn < zone->zone_start_pfn)
+                       pfn = zone->zone_start_pfn;
+
+               for (; pfn < end_pfn; pfn++) {
+                       struct page *page;
+
+                       if (!pfn_valid(pfn))
+                               continue;
+
+                       if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state))
+                               continue;
+
+                       if (page->flags) {
+                               VM_BUG_ON(page_zone(page) != zone);
+                               continue;
+                       }
+
+                       __init_single_page(page, pfn, zid, nid);
+                       __free_pages_boot_core(page, pfn, 0);
+                       nr_pages++;
+                       cond_resched();
+               }
+               first_init_pfn = max(end_pfn, first_init_pfn);
+       }
+
+       /* Sanity check that the next zone really is unpopulated */
+       WARN_ON(++zid < MAX_NR_ZONES && populated_zone(++zone));
+
+       pr_info("kswapd %d initialised %lu pages in %ums\n", nid, nr_pages,
+                                       jiffies_to_msecs(jiffies - start));
+}
+#endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
+
 #ifdef CONFIG_CMA
 /* Free whole pageblock and set its migration type to MIGRATE_CMA. */
 void __init init_cma_reserved_pageblock(struct page *page)
@@ -4329,6 +4441,9 @@ static void setup_zone_migrate_reserve(struct zone *zone)
        zone->nr_migrate_reserve_block = reserve;
 
        for (pfn = start_pfn; pfn < end_pfn; pfn += pageblock_nr_pages) {
+               if (!early_page_nid_uninitialised(pfn, zone_to_nid(zone)))
+                       return;
+
                if (!pfn_valid(pfn))
                        continue;
                page = pfn_to_page(pfn);
index e61445dce04e3cc83e9704e84f3d5bf9074b31db..f4a4871107645ddacd22d6ab9c3db0c492e628ee 100644 (file)
@@ -3386,7 +3386,7 @@ static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
  * If there are applications that are active memory-allocators
  * (most normal use), this basically shouldn't matter.
  */
-static int kswapd(void *p)
+static int __defermem_init kswapd(void *p)
 {
        unsigned long order, new_order;
        unsigned balanced_order;
@@ -3421,6 +3421,8 @@ static int kswapd(void *p)
        tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
        set_freezable();
 
+       deferred_init_memmap(pgdat->node_id);
+
        order = new_order = 0;
        balanced_order = 0;
        classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
@@ -3576,7 +3578,7 @@ static int cpu_callback(struct notifier_block *nfb, unsigned long action,
  * This kswapd start function will be called by init and node-hot-add.
  * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
  */
-int kswapd_run(int nid)
+int __defermem_init kswapd_run(int nid)
 {
        pg_data_t *pgdat = NODE_DATA(nid);
        int ret = 0;