Commit 28ee90fe6048 ("x86/mm: implement free pmd/pte page interfaces")
introduced a new location where a pmd was released, but neglected to run
the pmd page destructor. In fact, this happened previously for a
different pmd release path and was fixed by commit:
c283610e44ec ("x86, mm: do not leak page->ptl for pmd page tables").
This issue was hidden until recently because the failure mode is silent,
but commit:
b2b29d6d0119 ("mm: account PMD tables like PTE tables")
...turns the failure mode into this signature:
BUG: Bad page state in process lt-pmem-ns pfn:15943d
page:000000007262ed7b refcount:0 mapcount:-1024 mapping:0000000000000000 index:0x0 pfn:0x15943d
flags: 0xaffff800000000()
raw: 00affff800000000 dead000000000100 0000000000000000 0000000000000000
raw: 0000000000000000 ffff913a029bcc08 00000000fffffbff 0000000000000000
page dumped because: nonzero mapcount
[..]
dump_stack+0x8b/0xb0
bad_page.cold+0x63/0x94
free_pcp_prepare+0x224/0x270
free_unref_page+0x18/0xd0
pud_free_pmd_page+0x146/0x160
ioremap_pud_range+0xe3/0x350
ioremap_page_range+0x108/0x160
__ioremap_caller.constprop.0+0x174/0x2b0
? memremap+0x7a/0x110
memremap+0x7a/0x110
devm_memremap+0x53/0xa0
pmem_attach_disk+0x4ed/0x530 [nd_pmem]
? __devm_release_region+0x52/0x80
nvdimm_bus_probe+0x85/0x210 [libnvdimm]
Given this is a repeat occurrence it seemed prudent to look for other
places where this destructor might be missing and whether a better
helper is needed. try_to_free_pmd_page() looks like a candidate, but
testing with setting up and tearing down pmd mappings via the dax unit
tests is thus far not triggering the failure. As for a better helper
pmd_free() is close, but it is a messy fit due to requiring an @mm arg.
Also, ___pmd_free_tlb() wants to call paravirt_tlb_remove_table()
instead of free_page(), so open-coded pgtable_pmd_page_dtor() seems the
best way forward for now.
Fixes: 28ee90fe6048 ("x86/mm: implement free pmd/pte page interfaces")
Cc: <[email protected]>
Cc: Dave Hansen <[email protected]>
Cc: Andy Lutomirski <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: Borislav Petkov <[email protected]>
Cc: [email protected]
Cc: "H. Peter Anvin" <[email protected]>
Co-debugged-by: Matthew Wilcox <[email protected]>
Tested-by: Yi Zhang <[email protected]>
Signed-off-by: Dan Williams <[email protected]>
---
arch/x86/mm/pgtable.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/arch/x86/mm/pgtable.c b/arch/x86/mm/pgtable.c
index dfd82f51ba66..f6a9e2e36642 100644
--- a/arch/x86/mm/pgtable.c
+++ b/arch/x86/mm/pgtable.c
@@ -829,6 +829,8 @@ int pud_free_pmd_page(pud_t *pud, unsigned long addr)
}
free_page((unsigned long)pmd_sv);
+
+ pgtable_pmd_page_dtor(virt_to_page(pmd));
free_page((unsigned long)pmd);
return 1;
On Wed, Dec 02, 2020 at 10:28:12PM -0800, Dan Williams wrote:
> pmd_free() is close, but it is a messy fit due to requiring an @mm arg.
Hurpm, only parisc and s390 actually use that argument. And s390
_really_ needs it, because they're doing runtime folding per mm.
Ping, this bug is still present on v5.11-rc2, need a resend?
On Wed, Dec 2, 2020 at 10:28 PM Dan Williams <[email protected]> wrote:
>
> Commit 28ee90fe6048 ("x86/mm: implement free pmd/pte page interfaces")
> introduced a new location where a pmd was released, but neglected to run
> the pmd page destructor. In fact, this happened previously for a
> different pmd release path and was fixed by commit:
>
> c283610e44ec ("x86, mm: do not leak page->ptl for pmd page tables").
>
> This issue was hidden until recently because the failure mode is silent,
> but commit:
>
> b2b29d6d0119 ("mm: account PMD tables like PTE tables")
>
> ...turns the failure mode into this signature:
>
> BUG: Bad page state in process lt-pmem-ns pfn:15943d
> page:000000007262ed7b refcount:0 mapcount:-1024 mapping:0000000000000000 index:0x0 pfn:0x15943d
> flags: 0xaffff800000000()
> raw: 00affff800000000 dead000000000100 0000000000000000 0000000000000000
> raw: 0000000000000000 ffff913a029bcc08 00000000fffffbff 0000000000000000
> page dumped because: nonzero mapcount
> [..]
> dump_stack+0x8b/0xb0
> bad_page.cold+0x63/0x94
> free_pcp_prepare+0x224/0x270
> free_unref_page+0x18/0xd0
> pud_free_pmd_page+0x146/0x160
> ioremap_pud_range+0xe3/0x350
> ioremap_page_range+0x108/0x160
> __ioremap_caller.constprop.0+0x174/0x2b0
> ? memremap+0x7a/0x110
> memremap+0x7a/0x110
> devm_memremap+0x53/0xa0
> pmem_attach_disk+0x4ed/0x530 [nd_pmem]
> ? __devm_release_region+0x52/0x80
> nvdimm_bus_probe+0x85/0x210 [libnvdimm]
>
> Given this is a repeat occurrence it seemed prudent to look for other
> places where this destructor might be missing and whether a better
> helper is needed. try_to_free_pmd_page() looks like a candidate, but
> testing with setting up and tearing down pmd mappings via the dax unit
> tests is thus far not triggering the failure. As for a better helper
> pmd_free() is close, but it is a messy fit due to requiring an @mm arg.
> Also, ___pmd_free_tlb() wants to call paravirt_tlb_remove_table()
> instead of free_page(), so open-coded pgtable_pmd_page_dtor() seems the
> best way forward for now.
>
> Fixes: 28ee90fe6048 ("x86/mm: implement free pmd/pte page interfaces")
> Cc: <[email protected]>
> Cc: Dave Hansen <[email protected]>
> Cc: Andy Lutomirski <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Ingo Molnar <[email protected]>
> Cc: Borislav Petkov <[email protected]>
> Cc: [email protected]
> Cc: "H. Peter Anvin" <[email protected]>
> Co-debugged-by: Matthew Wilcox <[email protected]>
> Tested-by: Yi Zhang <[email protected]>
> Signed-off-by: Dan Williams <[email protected]>
> ---
> arch/x86/mm/pgtable.c | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/arch/x86/mm/pgtable.c b/arch/x86/mm/pgtable.c
> index dfd82f51ba66..f6a9e2e36642 100644
> --- a/arch/x86/mm/pgtable.c
> +++ b/arch/x86/mm/pgtable.c
> @@ -829,6 +829,8 @@ int pud_free_pmd_page(pud_t *pud, unsigned long addr)
> }
>
> free_page((unsigned long)pmd_sv);
> +
> + pgtable_pmd_page_dtor(virt_to_page(pmd));
> free_page((unsigned long)pmd);
>
> return 1;
>
On 12/2/20 10:28 PM, Dan Williams wrote:
> Commit 28ee90fe6048 ("x86/mm: implement free pmd/pte page interfaces")
> introduced a new location where a pmd was released, but neglected to run
> the pmd page destructor. In fact, this happened previously for a
> different pmd release path and was fixed by commit:
>
> c283610e44ec ("x86, mm: do not leak page->ptl for pmd page tables").
>
> This issue was hidden until recently because the failure mode is silent,
> but commit:
Looks sane. Thanks as always for the thorough changelog and the
investigation into why we're suddenly seeing this now.
I agree that ridding ourselves of open-coded free_page()'s is a good
idea, but this patch itself needs to be around for stable anyway. So,
Acked-by: Dave Hansen <[email protected]>