[v2] mm/page_owner: fix for deferred struct page init
diff mbox series

Message ID 20181220185031.43146-1-cai@lca.pw
State New
Headers show
Series
  • [v2] mm/page_owner: fix for deferred struct page init
Related show

Commit Message

Qian Cai Dec. 20, 2018, 6:50 p.m. UTC
When booting a system with "page_owner=on",

start_kernel
  page_ext_init
    invoke_init_callbacks
      init_section_page_ext
        init_page_owner
          init_early_allocated_pages
            init_zones_in_node
              init_pages_in_zone
                lookup_page_ext
                  page_to_nid

The issue here is that page_to_nid() will not work since some page
flags have no node information until later in page_alloc_init_late() due
to DEFERRED_STRUCT_PAGE_INIT. Hence, it could trigger an out-of-bounds
access with an invalid nid.

[    8.666047] UBSAN: Undefined behaviour in ./include/linux/mm.h:1104:50
[    8.672603] index 7 is out of range for type 'zone [5]'

Also, kernel will panic since flags were poisoned earlier with,

CONFIG_DEBUG_VM_PGFLAGS=y
CONFIG_NODE_NOT_IN_PAGE_FLAGS=n

start_kernel
  setup_arch
    pagetable_init
      paging_init
        sparse_init
          sparse_init_nid
            memblock_alloc_try_nid_raw

Although later it tries to set page flags for pages in reserved bootmem
regions,

mm_init
  mem_init
    memblock_free_all
      free_low_memory_core_early
        reserve_bootmem_region

there could still have some freed pages from the page allocator but yet
to be initialized due to DEFERRED_STRUCT_PAGE_INIT. It have already been
dealt with a bit in page_ext_init().

* Take into account DEFERRED_STRUCT_PAGE_INIT.
*/
if (early_pfn_to_nid(pfn) != nid)
	continue;

However it did not handle it well in init_pages_in_zone() which end up
calling page_to_nid().

[   11.917212] page:ffffea0004200000 is uninitialized and poisoned
[   11.917220] raw: ffffffffffffffff ffffffffffffffff ffffffffffffffff
ffffffffffffffff
[   11.921745] raw: ffffffffffffffff ffffffffffffffff ffffffffffffffff
ffffffffffffffff
[   11.924523] page dumped because: VM_BUG_ON_PAGE(PagePoisoned(p))
[   11.926498] page_owner info is not active (free page?)
[   12.329560] kernel BUG at include/linux/mm.h:990!
[   12.337632] RIP: 0010:init_page_owner+0x486/0x520

Since there is no other routines depend on page_ext_init() in
start_kernel() and no real benefit to call it so early, just move it
after page_alloc_init_late() to ensure that there is no deferred pages
need to de dealt with.

Suggested-by: Michal Hocko <mhocko@kernel.org>
Signed-off-by: Qian Cai <cai@lca.pw>
---

v2: postpone init_page_ext() to after page_alloc_init_late().

 init/main.c   | 2 +-
 mm/page_ext.c | 3 +--
 2 files changed, 2 insertions(+), 3 deletions(-)

Comments

Michal Hocko Jan. 9, 2019, 7:34 a.m. UTC | #1
On Thu 20-12-18 13:50:31, Qian Cai wrote:
> When booting a system with "page_owner=on",
> 
> start_kernel
>   page_ext_init
>     invoke_init_callbacks
>       init_section_page_ext
>         init_page_owner
>           init_early_allocated_pages
>             init_zones_in_node
>               init_pages_in_zone
>                 lookup_page_ext
>                   page_to_nid
> 
> The issue here is that page_to_nid() will not work since some page
> flags have no node information until later in page_alloc_init_late() due
> to DEFERRED_STRUCT_PAGE_INIT. Hence, it could trigger an out-of-bounds
> access with an invalid nid.
> 
> [    8.666047] UBSAN: Undefined behaviour in ./include/linux/mm.h:1104:50
> [    8.672603] index 7 is out of range for type 'zone [5]'
> 
> Also, kernel will panic since flags were poisoned earlier with,
> 
> CONFIG_DEBUG_VM_PGFLAGS=y
> CONFIG_NODE_NOT_IN_PAGE_FLAGS=n
> 
> start_kernel
>   setup_arch
>     pagetable_init
>       paging_init
>         sparse_init
>           sparse_init_nid
>             memblock_alloc_try_nid_raw
> 
> Although later it tries to set page flags for pages in reserved bootmem
> regions,
> 
> mm_init
>   mem_init
>     memblock_free_all
>       free_low_memory_core_early
>         reserve_bootmem_region
> 
> there could still have some freed pages from the page allocator but yet
> to be initialized due to DEFERRED_STRUCT_PAGE_INIT. It have already been
> dealt with a bit in page_ext_init().
> 
> * Take into account DEFERRED_STRUCT_PAGE_INIT.
> */
> if (early_pfn_to_nid(pfn) != nid)
> 	continue;
> 
> However it did not handle it well in init_pages_in_zone() which end up
> calling page_to_nid().
> 
> [   11.917212] page:ffffea0004200000 is uninitialized and poisoned
> [   11.917220] raw: ffffffffffffffff ffffffffffffffff ffffffffffffffff
> ffffffffffffffff
> [   11.921745] raw: ffffffffffffffff ffffffffffffffff ffffffffffffffff
> ffffffffffffffff
> [   11.924523] page dumped because: VM_BUG_ON_PAGE(PagePoisoned(p))
> [   11.926498] page_owner info is not active (free page?)
> [   12.329560] kernel BUG at include/linux/mm.h:990!
> [   12.337632] RIP: 0010:init_page_owner+0x486/0x520
> 
> Since there is no other routines depend on page_ext_init() in
> start_kernel() and no real benefit to call it so early, just move it
> after page_alloc_init_late() to ensure that there is no deferred pages
> need to de dealt with.

The last paragraph should be updated. I would propose something like
this.

"
This means that assumptions behind fe53ca54270a ("mm: use
early_pfn_to_nid in page_ext_init") are incomplete. Therefore revert the
commit for now. A proper way to move the page_owner initialization to
sooner is to hook into memmap initialization.
"
> 

Fixes: fe53ca54270a ("mm: use early_pfn_to_nid in page_ext_init")

> Suggested-by: Michal Hocko <mhocko@kernel.org>
> Signed-off-by: Qian Cai <cai@lca.pw>

Acked-by: Michal Hocko <mhocko@suse.com>

> ---
> 
> v2: postpone init_page_ext() to after page_alloc_init_late().
> 
>  init/main.c   | 2 +-
>  mm/page_ext.c | 3 +--
>  2 files changed, 2 insertions(+), 3 deletions(-)
> 
> diff --git a/init/main.c b/init/main.c
> index 2b7b7fe173c9..1aeb062b2cb7 100644
> --- a/init/main.c
> +++ b/init/main.c
> @@ -696,7 +696,6 @@ asmlinkage __visible void __init start_kernel(void)
>  		initrd_start = 0;
>  	}
>  #endif
> -	page_ext_init();
>  	kmemleak_init();
>  	setup_per_cpu_pageset();
>  	numa_policy_init();
> @@ -1147,6 +1146,7 @@ static noinline void __init kernel_init_freeable(void)
>  	sched_init_smp();
>  
>  	page_alloc_init_late();
> +	page_ext_init();
>  
>  	do_basic_setup();
>  
> diff --git a/mm/page_ext.c b/mm/page_ext.c
> index ae44f7adbe07..d76fd51e312a 100644
> --- a/mm/page_ext.c
> +++ b/mm/page_ext.c
> @@ -399,9 +399,8 @@ void __init page_ext_init(void)
>  			 * -------------pfn-------------->
>  			 * N0 | N1 | N2 | N0 | N1 | N2|....
>  			 *
> -			 * Take into account DEFERRED_STRUCT_PAGE_INIT.
>  			 */
> -			if (early_pfn_to_nid(pfn) != nid)
> +			if (pfn_to_nid(pfn) != nid)
>  				continue;
>  			if (init_section_page_ext(pfn, nid))
>  				goto oom;
> -- 
> 2.17.2 (Apple Git-113)

Patch
diff mbox series

diff --git a/init/main.c b/init/main.c
index 2b7b7fe173c9..1aeb062b2cb7 100644
--- a/init/main.c
+++ b/init/main.c
@@ -696,7 +696,6 @@  asmlinkage __visible void __init start_kernel(void)
 		initrd_start = 0;
 	}
 #endif
-	page_ext_init();
 	kmemleak_init();
 	setup_per_cpu_pageset();
 	numa_policy_init();
@@ -1147,6 +1146,7 @@  static noinline void __init kernel_init_freeable(void)
 	sched_init_smp();
 
 	page_alloc_init_late();
+	page_ext_init();
 
 	do_basic_setup();
 
diff --git a/mm/page_ext.c b/mm/page_ext.c
index ae44f7adbe07..d76fd51e312a 100644
--- a/mm/page_ext.c
+++ b/mm/page_ext.c
@@ -399,9 +399,8 @@  void __init page_ext_init(void)
 			 * -------------pfn-------------->
 			 * N0 | N1 | N2 | N0 | N1 | N2|....
 			 *
-			 * Take into account DEFERRED_STRUCT_PAGE_INIT.
 			 */
-			if (early_pfn_to_nid(pfn) != nid)
+			if (pfn_to_nid(pfn) != nid)
 				continue;
 			if (init_section_page_ext(pfn, nid))
 				goto oom;