Re: [PATCH] x86/mm/cpa: Flush direct map alias during cpa

From: Peter Zijlstra
Date: Fri Apr 24 2020 - 06:54:17 EST


On Thu, Apr 23, 2020 at 07:02:26PM +0000, Edgecombe, Rick P wrote:
> On Thu, 2020-04-23 at 10:41 +0200, Peter Zijlstra wrote:
> > Wouldn't something like so make more sense?
>
> Yes. Dave had commented on whether a smaller fix would be better for
> backports if needed. Since that diff is the whole fix, do you want to
> take it from here or should I put it in a patch?

I've made it look like this. Holler if you need it changed ;-)

---
Subject: x86/mm/cpa: Flush direct map alias during cpa
From: Rick Edgecombe <rick.p.edgecombe@xxxxxxxxx>
Date: Wed, 22 Apr 2020 20:13:55 -0700

From: Rick Edgecombe <rick.p.edgecombe@xxxxxxxxx>

As an optimization, cpa_flush() was changed to optionally only flush
the range in @cpa if it was small enough. However, this range does
not include any direct map aliases changed in cpa_process_alias(). So
small set_memory_() calls that touch that alias don't get the direct
map changes flushed. This situation can happen when the virtual
address taking variants are passed an address in vmalloc or modules
space.

In these cases, force a full TLB flush.

Note this issue does not extend to cases where the set_memory_() calls are
passed a direct map address, or page array, etc, as the primary target. In
those cases the direct map would be flushed.

Fixes: 935f5839827e ("x86/mm/cpa: Optimize cpa_flush_array() TLB invalidation")
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@xxxxxxxxx>
Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
arch/x86/mm/pat/set_memory.c | 12 ++++++++----
1 file changed, 8 insertions(+), 4 deletions(-)

--- a/arch/x86/mm/pat/set_memory.c
+++ b/arch/x86/mm/pat/set_memory.c
@@ -43,7 +43,8 @@ struct cpa_data {
unsigned long pfn;
unsigned int flags;
unsigned int force_split : 1,
- force_static_prot : 1;
+ force_static_prot : 1,
+ force_flush_all : 1;
struct page **pages;
};

@@ -355,10 +356,10 @@ static void cpa_flush(struct cpa_data *d
return;
}

- if (cpa->numpages <= tlb_single_page_flush_ceiling)
- on_each_cpu(__cpa_flush_tlb, cpa, 1);
- else
+ if (cpa->force_flush_all || cpa->numpages > tlb_single_page_flush_ceiling)
flush_tlb_all();
+ else
+ on_each_cpu(__cpa_flush_tlb, cpa, 1);

if (!cache)
return;
@@ -1598,6 +1599,8 @@ static int cpa_process_alias(struct cpa_
alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
alias_cpa.curpage = 0;

+ cpa->force_flush_all = 1;
+
ret = __change_page_attr_set_clr(&alias_cpa, 0);
if (ret)
return ret;
@@ -1618,6 +1621,7 @@ static int cpa_process_alias(struct cpa_
alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
alias_cpa.curpage = 0;

+ cpa->force_flush_all = 1;
/*
* The high mapping range is imprecise, so ignore the
* return value.