]> git.karo-electronics.de Git - karo-tx-linux.git/commitdiff
mm: consider memblock reservations for deferred memory initialization sizing
authorMichal Hocko <mhocko@suse.com>
Fri, 2 Jun 2017 21:46:49 +0000 (14:46 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 2 Jun 2017 22:07:38 +0000 (15:07 -0700)
We have seen an early OOM killer invocation on ppc64 systems with
crashkernel=4096M:

kthreadd invoked oom-killer: gfp_mask=0x16040c0(GFP_KERNEL|__GFP_COMP|__GFP_NOTRACK), nodemask=7, order=0, oom_score_adj=0
kthreadd cpuset=/ mems_allowed=7
CPU: 0 PID: 2 Comm: kthreadd Not tainted 4.4.68-1.gd7fe927-default #1
Call Trace:
  dump_stack+0xb0/0xf0 (unreliable)
  dump_header+0xb0/0x258
  out_of_memory+0x5f0/0x640
  __alloc_pages_nodemask+0xa8c/0xc80
  kmem_getpages+0x84/0x1a0
  fallback_alloc+0x2a4/0x320
  kmem_cache_alloc_node+0xc0/0x2e0
  copy_process.isra.25+0x260/0x1b30
  _do_fork+0x94/0x470
  kernel_thread+0x48/0x60
  kthreadd+0x264/0x330
  ret_from_kernel_thread+0x5c/0xa4

Mem-Info:
active_anon:0 inactive_anon:0 isolated_anon:0
 active_file:0 inactive_file:0 isolated_file:0
 unevictable:0 dirty:0 writeback:0 unstable:0
 slab_reclaimable:5 slab_unreclaimable:73
 mapped:0 shmem:0 pagetables:0 bounce:0
 free:0 free_pcp:0 free_cma:0
Node 7 DMA free:0kB min:0kB low:0kB high:0kB active_anon:0kB inactive_anon:0kB active_file:0kB inactive_file:0kB unevictable:0kB isolated(anon):0kB isolated(file):0kB present:52428800kB managed:110016kB mlocked:0kB dirty:0kB writeback:0kB mapped:0kB shmem:0kB slab_reclaimable:320kB slab_unreclaimable:4672kB kernel_stack:1152kB pagetables:0kB unstable:0kB bounce:0kB free_pcp:0kB local_pcp:0kB free_cma:0kB writeback_tmp:0kB pages_scanned:0 all_unreclaimable? yes
lowmem_reserve[]: 0 0 0 0
Node 7 DMA: 0*64kB 0*128kB 0*256kB 0*512kB 0*1024kB 0*2048kB 0*4096kB 0*8192kB 0*16384kB = 0kB
0 total pagecache pages
0 pages in swap cache
Swap cache stats: add 0, delete 0, find 0/0
Free swap  = 0kB
Total swap = 0kB
819200 pages RAM
0 pages HighMem/MovableOnly
817481 pages reserved
0 pages cma reserved
0 pages hwpoisoned

the reason is that the managed memory is too low (only 110MB) while the
rest of the the 50GB is still waiting for the deferred intialization to
be done.  update_defer_init estimates the initial memoty to initialize
to 2GB at least but it doesn't consider any memory allocated in that
range.  In this particular case we've had

Reserving 4096MB of memory at 128MB for crashkernel (System RAM: 51200MB)

so the low 2GB is mostly depleted.

Fix this by considering memblock allocations in the initial static
initialization estimation.  Move the max_initialise to
reset_deferred_meminit and implement a simple memblock_reserved_memory
helper which iterates all reserved blocks and sums the size of all that
start below the given address.  The cumulative size is than added on top
of the initial estimation.  This is still not ideal because
reset_deferred_meminit doesn't consider holes and so reservation might
be above the initial estimation whihch we ignore but let's make the
logic simpler until we really need to handle more complicated cases.

Fixes: 3a80a7fa7989 ("mm: meminit: initialise a subset of struct pages if CONFIG_DEFERRED_STRUCT_PAGE_INIT is set")
Link: http://lkml.kernel.org/r/20170531104010.GI27783@dhcp22.suse.cz
Signed-off-by: Michal Hocko <mhocko@suse.com>
Acked-by: Mel Gorman <mgorman@suse.de>
Tested-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com>
Cc: <stable@vger.kernel.org> [4.2+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/memblock.h
include/linux/mmzone.h
mm/memblock.c
mm/page_alloc.c

index 4ce24a3762627be20e805da3eaab26ba1bc47578..8098695e5d8d9dfba3815fd359823f92833a5d61 100644 (file)
@@ -425,12 +425,20 @@ static inline void early_memtest(phys_addr_t start, phys_addr_t end)
 }
 #endif
 
+extern unsigned long memblock_reserved_memory_within(phys_addr_t start_addr,
+               phys_addr_t end_addr);
 #else
 static inline phys_addr_t memblock_alloc(phys_addr_t size, phys_addr_t align)
 {
        return 0;
 }
 
+static inline unsigned long memblock_reserved_memory_within(phys_addr_t start_addr,
+               phys_addr_t end_addr)
+{
+       return 0;
+}
+
 #endif /* CONFIG_HAVE_MEMBLOCK */
 
 #endif /* __KERNEL__ */
index ebaccd4e7d8cdc5f5ef13fed1475a8b202d93628..ef6a13b7bd3e851385bea32434e207a5cf6eec7f 100644 (file)
@@ -678,6 +678,7 @@ typedef struct pglist_data {
         * is the first PFN that needs to be initialised.
         */
        unsigned long first_deferred_pfn;
+       unsigned long static_init_size;
 #endif /* CONFIG_DEFERRED_STRUCT_PAGE_INIT */
 
 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
index b049c9b2dba8718a6f57777591c2f2753d8d40ad..7b8a5db76a2fec7331f09048f3c19ebdfddf9f16 100644 (file)
@@ -1739,6 +1739,29 @@ static void __init_memblock memblock_dump(struct memblock_type *type)
        }
 }
 
+extern unsigned long __init_memblock
+memblock_reserved_memory_within(phys_addr_t start_addr, phys_addr_t end_addr)
+{
+       struct memblock_region *rgn;
+       unsigned long size = 0;
+       int idx;
+
+       for_each_memblock_type((&memblock.reserved), rgn) {
+               phys_addr_t start, end;
+
+               if (rgn->base + rgn->size < start_addr)
+                       continue;
+               if (rgn->base > end_addr)
+                       continue;
+
+               start = rgn->base;
+               end = start + rgn->size;
+               size += end - start;
+       }
+
+       return size;
+}
+
 void __init_memblock __memblock_dump_all(void)
 {
        pr_info("MEMBLOCK configuration:\n");
index b7a6f583a3733b14d421a7ba6a92013a43050f7b..2302f250d6b1ba150e3c2e4e17cfb6c99574ab5b 100644 (file)
@@ -292,6 +292,26 @@ int page_group_by_mobility_disabled __read_mostly;
 #ifdef CONFIG_DEFERRED_STRUCT_PAGE_INIT
 static inline void reset_deferred_meminit(pg_data_t *pgdat)
 {
+       unsigned long max_initialise;
+       unsigned long reserved_lowmem;
+
+       /*
+        * Initialise at least 2G of a node but also take into account that
+        * two large system hashes that can take up 1GB for 0.25TB/node.
+        */
+       max_initialise = max(2UL << (30 - PAGE_SHIFT),
+               (pgdat->node_spanned_pages >> 8));
+
+       /*
+        * Compensate the all the memblock reservations (e.g. crash kernel)
+        * from the initial estimation to make sure we will initialize enough
+        * memory to boot.
+        */
+       reserved_lowmem = memblock_reserved_memory_within(pgdat->node_start_pfn,
+                       pgdat->node_start_pfn + max_initialise);
+       max_initialise += reserved_lowmem;
+
+       pgdat->static_init_size = min(max_initialise, pgdat->node_spanned_pages);
        pgdat->first_deferred_pfn = ULONG_MAX;
 }
 
@@ -314,20 +334,11 @@ static inline bool update_defer_init(pg_data_t *pgdat,
                                unsigned long pfn, unsigned long zone_end,
                                unsigned long *nr_initialised)
 {
-       unsigned long max_initialise;
-
        /* Always populate low zones for address-contrained allocations */
        if (zone_end < pgdat_end_pfn(pgdat))
                return true;
-       /*
-        * Initialise at least 2G of a node but also take into account that
-        * two large system hashes that can take up 1GB for 0.25TB/node.
-        */
-       max_initialise = max(2UL << (30 - PAGE_SHIFT),
-               (pgdat->node_spanned_pages >> 8));
-
        (*nr_initialised)++;
-       if ((*nr_initialised > max_initialise) &&
+       if ((*nr_initialised > pgdat->static_init_size) &&
            (pfn & (PAGES_PER_SECTION - 1)) == 0) {
                pgdat->first_deferred_pfn = pfn;
                return false;
@@ -6138,7 +6149,6 @@ void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
        /* pg_data_t should be reset to zero when it's allocated */
        WARN_ON(pgdat->nr_zones || pgdat->kswapd_classzone_idx);
 
-       reset_deferred_meminit(pgdat);
        pgdat->node_id = nid;
        pgdat->node_start_pfn = node_start_pfn;
        pgdat->per_cpu_nodestats = NULL;
@@ -6160,6 +6170,7 @@ void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
                (unsigned long)pgdat->node_mem_map);
 #endif
 
+       reset_deferred_meminit(pgdat);
        free_area_init_core(pgdat);
 }