On Mon, Dec 05, 2022 at 04:08:57PM +0100, David Hildenbrand wrote:
> 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 a 32bit
> phys_addr_t. This implies, that SWP_PFN_BITS will currently only be able to
> cover 4 GiB - 1 on any 32bit system, which is wrong. We can end up
> masking off valid PFN bits from the swap offset.
>
> Ideally, we'd use something like "sizeof(phys_addr_t) * 8 - PAGE_SHIFT",
> however, we might exceed SWP_TYPE_SHIFT and make the BUILD_BUG_ON() in
> is_pfn_swap_entry() unhappy. Note that swp_entry_t is effectively an
> unsigned long.
>
> Consequently, simply rely on SWP_TYPE_SHIFT in case we're on 32bit and
> CONFIG_PHYS_ADDR_T_64BIT is defined.
>
> For example, on an 8 GiB x86 PAE system, we currently fail removing
> migration entries (remove_migration_ptes()) that target PFNs above
> 4 GiB, because mm/page_vma_mapped.c:check_pte() will fail to identify a
> PFN match as swp_offset_pfn() wrongly masks off valid PFN bits.
>
> With THP, split_huge_page_to_list()->...->remap_page() will leave migration
> entries in place and continue to unlock the page, which is wrong. 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 ...
> [ 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
>
> Fixes: 0d206b5d2e0d ("mm/swap: add swp_offset_pfn() to fetch PFN from swap entry")
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Peter Xu <peterx@xxxxxxxxxx>
> Cc: Yang Shi <shy828301@xxxxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
> Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
Thanks for debugging this one.
> ---
>
> This makes my x86 PAE case work as expected again. Only cross compiled
> on other architectures.
IIUC it's not about PAE but !SPARSEMEM, as PAE actually has it defined when
with sparsemem:
#ifdef CONFIG_X86_32
# ifdef CONFIG_X86_PAE
# define SECTION_SIZE_BITS 29
# define MAX_PHYSMEM_BITS 36
# else
# define SECTION_SIZE_BITS 26
# define MAX_PHYSMEM_BITS 32
# endif
#else /* CONFIG_X86_32 */
# define SECTION_SIZE_BITS 27 /* matt - 128 is convenient right now */
# define MAX_PHYSMEM_BITS (pgtable_l5_enabled() ? 52 : 46)
#endif
One trivial comment below.
>
> ---
> include/linux/swapops.h | 8 +++++---
> 1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/include/linux/swapops.h b/include/linux/swapops.h
> index 86b95ccb81bb..4bb7a20f3fa5 100644
> --- a/include/linux/swapops.h
> +++ b/include/linux/swapops.h
> @@ -32,11 +32,13 @@
> * store PFN, we only need SWP_PFN_BITS bits. Each of the pfn swap entries
> * can use the extra bits to store other information besides PFN.
> */
> -#ifdef MAX_PHYSMEM_BITS
> +#if defined(MAX_PHYSMEM_BITS)
> #define SWP_PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
> -#else /* MAX_PHYSMEM_BITS */
> +#elif !defined(CONFIG_64BIT) && defined(CONFIG_PHYS_ADDR_T_64BIT)
> +#define SWP_PFN_BITS SWP_TYPE_SHIFT
Can we add a comment showing where SWP_TYPE_SHIFT comes from? It should be
a min value comes from either the limitation of phys address width, or from
definition of swp_entry_t (which is unsigned long).
Or I'd rather make this then the code explains better on itself, and the
change should be smaller too:
#ifdef MAX_PHYSMEM_BITS
#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((sizeof(phys_addr_t) * 8) - \
+ PAGE_SHIFT, SWP_TYPE_SHIFT)
#endif /* MAX_PHYSMEM_BITS */
#define SWP_PFN_MASK (BIT(SWP_PFN_BITS) - 1)
What do you think?
> +#else
> #define SWP_PFN_BITS (BITS_PER_LONG - PAGE_SHIFT)
> -#endif /* MAX_PHYSMEM_BITS */
> +#endif /* defined(MAX_PHYSMEM_BITS) */
> #define SWP_PFN_MASK (BIT(SWP_PFN_BITS) - 1)
>
> /**
>
> base-commit: 76dcd734eca23168cb008912c0f69ff408905235
> --
> 2.38.1
>
--
Peter Xu
