On Tue, Dec 17, 2019 at 01:34:16PM +0100, Peter Zijlstra wrote:
> Perhaps if we replace !tlb->end with something like:
> 
>   !tlb->freed_tables && !tlb->cleared_p*
> 
> (which GCC should be able to do with a single load and mask)
> 
> I've not really thought too hard about it yet, I need to run some
> errands, but I'll look at it more closely when I get back.

AFAICT this should work.

---
Subject: asm-generic/tlb: Avoid potential double flush

Aneesh reported that:

        tlb_flush_mmu()
          tlb_flush_mmu_tlbonly()
            tlb_flush()                 <-- #1
          tlb_flush_mmu_free()
            tlb_table_flush()
              tlb_table_invalidate()
                tlb_flush_mmu_tlbonly()
                  tlb_flush()           <-- #2

does two TLBIs when tlb->fullmm, because __tlb_reset_range() will not
clear tlb->end in that case.

Observe that any caller to __tlb_adjust_range() also sets at least one
of the tlb->freed_tables || tlb->cleared_p* bits, and those are
unconditionally cleared by __tlb_reset_range().

Change the condition for actually issuing TLBI to having one of those
bits set, as opposed to having tlb->end != 0.

Reported-by: "Aneesh Kumar K.V" <aneesh.ku...@linux.ibm.com>
Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org>
---
 include/asm-generic/tlb.h | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
index fe0ea6ff3636..c9a25c5a83e8 100644
--- a/include/asm-generic/tlb.h
+++ b/include/asm-generic/tlb.h
@@ -402,7 +402,12 @@ tlb_update_vma_flags(struct mmu_gather *tlb, struct 
vm_area_struct *vma) { }
 
 static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
 {
-       if (!tlb->end)
+       /*
+        * Anything calling __tlb_adjust_range() also sets at least one of
+        * these bits.
+        */
+       if (!(tlb->freed_tables || tlb->cleared_ptes || tlb->cleared_pmds ||
+             tlb->cleared_puds || tlb->cleared_p4ds))
                return;
 
        tlb_flush(tlb);

Reply via email to