mm/swap: fix SWP_PFN_BITS with CONFIG_PHYS_ADDR_T_64BIT on 32bit
authorDavid Hildenbrand <david@redhat.com>
Mon, 5 Dec 2022 15:08:57 +0000 (16:08 +0100)
committerAndrew Morton <akpm@linux-foundation.org>
Sat, 10 Dec 2022 02:41:16 +0000 (18:41 -0800)
We use "unsigned long" to store a PFN in the kernel and phys_addr_t to
store a physical address.

On a 64bit system, both are 64bit wide.  However, on a 32bit system, the
latter might be 64bit wide.  This is, for example, the case on x86 with
PAE: phys_addr_t and PTEs are 64bit wide, while "unsigned long" only spans
32bit.

The current definition of SWP_PFN_BITS without MAX_PHYSMEM_BITS misses
that case, and assumes that the maximum PFN is limited by an 32bit
phys_addr_t.  This implies, that SWP_PFN_BITS will currently only be able
to cover 4 GiB - 1 on any 32bit system with 4k page size, which is wrong.

Let's rely on the number of bits in phys_addr_t instead, but make sure to
not exceed the maximum swap offset, to not make the BUILD_BUG_ON() in
is_pfn_swap_entry() unhappy.  Note that swp_entry_t is effectively an
unsigned long and the maximum swap offset shares that value with the swap
type.

For example, on an 8 GiB x86 PAE system with a kernel config based on
Debian 11.5 (-> CONFIG_FLATMEM=y, CONFIG_X86_PAE=y), we will currently
fail removing migration entries (remove_migration_ptes()), because
mm/page_vma_mapped.c:check_pte() will fail to identify a PFN match as
swp_offset_pfn() wrongly masks off PFN bits.  For example,
split_huge_page_to_list()->...->remap_page() will leave migration entries
in place and continue to unlock the page.

Later, when we stumble over these migration entries (e.g., via
/proc/self/pagemap), pfn_swap_entry_to_page() will BUG_ON() because these
migration entries shouldn't exist anymore and the page was unlocked.

[   33.067591] kernel BUG at include/linux/swapops.h:497!
[   33.067597] invalid opcode: 0000 [#1] PREEMPT SMP NOPTI
[   33.067602] CPU: 3 PID: 742 Comm: cow Tainted: G            E      6.1.0-rc8+ #16
[   33.067605] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.16.0-1.fc36 04/01/2014
[   33.067606] EIP: pagemap_pmd_range+0x644/0x650
[   33.067612] Code: 00 00 00 00 66 90 89 ce b9 00 f0 ff ff e9 ff fb ff ff 89 d8 31 db e8 48 c6 52 00 e9 23 fb ff ff e8 61 83 56 00 e9 b6 fe ff ff <0f> 0b bf 00 f0 ff ff e9 38 fa ff ff 3e 8d 74 26 00 55 89 e5 57 31
[   33.067615] EAX: ee394000 EBX: 00000002 ECX: ee394000 EDX: 00000000
[   33.067617] ESI: c1b0ded4 EDI: 00024a00 EBP: c1b0ddb4 ESP: c1b0dd68
[   33.067619] DS: 007b ES: 007b FS: 00d8 GS: 0033 SS: 0068 EFLAGS: 00010246
[   33.067624] CR0: 80050033 CR2: b7a00000 CR3: 01bbbd20 CR4: 00350ef0
[   33.067625] Call Trace:
[   33.067628]  ? madvise_free_pte_range+0x720/0x720
[   33.067632]  ? smaps_pte_range+0x4b0/0x4b0
[   33.067634]  walk_pgd_range+0x325/0x720
[   33.067637]  ? mt_find+0x1d6/0x3a0
[   33.067641]  ? mt_find+0x1d6/0x3a0
[   33.067643]  __walk_page_range+0x164/0x170
[   33.067646]  walk_page_range+0xf9/0x170
[   33.067648]  ? __kmem_cache_alloc_node+0x2a8/0x340
[   33.067653]  pagemap_read+0x124/0x280
[   33.067658]  ? default_llseek+0x101/0x160
[   33.067662]  ? smaps_account+0x1d0/0x1d0
[   33.067664]  vfs_read+0x90/0x290
[   33.067667]  ? do_madvise.part.0+0x24b/0x390
[   33.067669]  ? debug_smp_processor_id+0x12/0x20
[   33.067673]  ksys_pread64+0x58/0x90
[   33.067675]  __ia32_sys_ia32_pread64+0x1b/0x20
[   33.067680]  __do_fast_syscall_32+0x4c/0xc0
[   33.067683]  do_fast_syscall_32+0x29/0x60
[   33.067686]  do_SYSENTER_32+0x15/0x20
[   33.067689]  entry_SYSENTER_32+0x98/0xf1

Decrease the indentation level of SWP_PFN_BITS and SWP_PFN_MASK to keep it
readable and consistent.

[david@redhat.com: rely on sizeof(phys_addr_t) and min_t() instead]
Link: https://lkml.kernel.org/r/20221206105737.69478-1-david@redhat.com
[david@redhat.com: use "int" for comparison, as we're only comparing numbers < 64]
Link: https://lkml.kernel.org/r/1f157500-2676-7cef-a84e-9224ed64e540@redhat.com
Link: https://lkml.kernel.org/r/20221205150857.167583-1-david@redhat.com
Fixes: 0d206b5d2e0d ("mm/swap: add swp_offset_pfn() to fetch PFN from swap entry")
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Yang Shi <shy828301@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
include/linux/swapops.h

index 86b95cc..b07b277 100644 (file)
  * can use the extra bits to store other information besides PFN.
  */
 #ifdef MAX_PHYSMEM_BITS
-#define SWP_PFN_BITS                   (MAX_PHYSMEM_BITS - PAGE_SHIFT)
+#define SWP_PFN_BITS           (MAX_PHYSMEM_BITS - PAGE_SHIFT)
 #else  /* MAX_PHYSMEM_BITS */
-#define SWP_PFN_BITS                   (BITS_PER_LONG - PAGE_SHIFT)
+#define SWP_PFN_BITS           min_t(int, \
+                                     sizeof(phys_addr_t) * 8 - PAGE_SHIFT, \
+                                     SWP_TYPE_SHIFT)
 #endif /* MAX_PHYSMEM_BITS */
-#define SWP_PFN_MASK                   (BIT(SWP_PFN_BITS) - 1)
+#define SWP_PFN_MASK           (BIT(SWP_PFN_BITS) - 1)
 
 /**
  * Migration swap entry specific bitfield definitions.  Layout: