Re: [PATCH] mm: mmu_gather: remove __tlb_reset_range() for force flush

From: Yang Shi
Date: Thu May 09 2019 - 15:11:33 EST




On 5/9/19 11:24 AM, Peter Zijlstra wrote:
On Thu, May 09, 2019 at 05:36:29PM +0000, Nadav Amit wrote:
On May 9, 2019, at 3:38 AM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
diff --git a/mm/mmu_gather.c b/mm/mmu_gather.c
index 99740e1dd273..fe768f8d612e 100644
--- a/mm/mmu_gather.c
+++ b/mm/mmu_gather.c
@@ -244,15 +244,20 @@ void tlb_finish_mmu(struct mmu_gather *tlb,
unsigned long start, unsigned long end)
{
/*
- * If there are parallel threads are doing PTE changes on same range
- * under non-exclusive lock(e.g., mmap_sem read-side) but defer TLB
- * flush by batching, a thread has stable TLB entry can fail to flush
- * the TLB by observing pte_none|!pte_dirty, for example so flush TLB
- * forcefully if we detect parallel PTE batching threads.
+ * Sensible comment goes here..
*/
- if (mm_tlb_flush_nested(tlb->mm)) {
- __tlb_reset_range(tlb);
- __tlb_adjust_range(tlb, start, end - start);
+ if (mm_tlb_flush_nested(tlb->mm) && !tlb->full_mm) {
+ /*
+ * Since we're can't tell what we actually should have
+ * flushed flush everything in the given range.
+ */
+ tlb->start = start;
+ tlb->end = end;
+ tlb->freed_tables = 1;
+ tlb->cleared_ptes = 1;
+ tlb->cleared_pmds = 1;
+ tlb->cleared_puds = 1;
+ tlb->cleared_p4ds = 1;
}

tlb_flush_mmu(tlb);
As a simple optimization, I think it is possible to hold multiple nesting
counters in the mm, similar to tlb_flush_pending, for freed_tables,
cleared_ptes, etc.

The first time you set tlb->freed_tables, you also atomically increase
mm->tlb_flush_freed_tables. Then, in tlb_flush_mmu(), you just use
mm->tlb_flush_freed_tables instead of tlb->freed_tables.
That sounds fraught with races and expensive; I would much prefer to not
go there for this arguably rare case.

Consider such fun cases as where CPU-0 sees and clears a PTE, CPU-1
races and doesn't see that PTE. Therefore CPU-0 sets and counts
cleared_ptes. Then if CPU-1 flushes while CPU-0 is still in mmu_gather,
it will see cleared_ptes count increased and flush that granularity,
OTOH if CPU-1 flushes after CPU-0 completes, it will not and potentiall
miss an invalidate it should have had.

This whole concurrent mmu_gather stuff is horrible.

/me ponders more....

So I think the fundamental race here is this:

CPU-0 CPU-1

tlb_gather_mmu(.start=1, tlb_gather_mmu(.start=2,
.end=3); .end=4);

ptep_get_and_clear_full(2)
tlb_remove_tlb_entry(2);
__tlb_remove_page();
if (pte_present(2)) // nope

tlb_finish_mmu();

// continue without TLBI(2)
// whoopsie

tlb_finish_mmu();
tlb_flush() -> TLBI(2)

I'm not quite sure if this is the case Jan really met. But, according to his test, once correct tlb->freed_tables and tlb->cleared_* are set, his test works well.



And we can fix that by having tlb_finish_mmu() sync up. Never let a
concurrent tlb_finish_mmu() complete until all concurrenct mmu_gathers
have completed.

Not sure if this will scale well.


This should not be too hard to make happen.