On Thu, Sep 06, 2018 at 10:13:00AM +0200, Peter Zijlstra wrote:
> No, you got it the first time. There are in fact more fixmap abusers;
> see drivers/acpi/apei/ghes.c. Also, as long as set_fixmap() allows
> overwriting a _PAGE_PRESENT pte and has that dodgy
> __flush_tlb_one_kernel() in it, the broken remains (and can return).
>
> So we need to fix fixmap, to either disallow overwriting a _PAGE_PRESENT
> pte, or to issue a full TLB invalidate.
>
> Either fix will terminally break GHES, but that's OK, they've known
> about this issue since 2015 and haven't cared, so I can't be bothered
> about them.
Worse, when I boot with the below patch in, I get hits on text_poke (we
knew about this), but I also get hits from:
release_ds_buffer()
Because PEBS/BTS hardware write through the page-tables (I know right),
their buffers need to be in the PTI user map, and we do that by mapping
the pages in the CAE, and that uses fixmap.
---
diff --git a/arch/x86/mm/init_64.c b/arch/x86/mm/init_64.c
index dd519f372169..20dcd6fa690e 100644
--- a/arch/x86/mm/init_64.c
+++ b/arch/x86/mm/init_64.c
@@ -259,9 +259,16 @@ static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
static void __set_pte_vaddr(pud_t *pud, unsigned long vaddr, pte_t new_pte)
{
pmd_t *pmd = fill_pmd(pud, vaddr);
- pte_t *pte = fill_pte(pmd, vaddr);
+ pte_t *ptep = fill_pte(pmd, vaddr);
- set_pte(pte, new_pte);
+ pte_t pte = *ptep;
+
+ set_pte(ptep, new_pte);
+
+ WARN((pte_val(pte) & _PAGE_PRESENT) && (pte_val(pte) != pte_val(new_pte)),
+ "set_fixmap(%d, %lx), was: %lx",
+ (int)__virt_to_fix(vaddr),
+ pte_val(new_pte), pte_val(pte));
/*
* It's enough to flush this one mapping.
