Re: [PATCH net] tcp: fix page frag corruption on page fault

Hello,

On Fri, 2021-11-26 at 07:32 -0800, Eric Dumazet wrote:

On Fri, Nov 26, 2021 at 7:27 AM Eric Dumazet [off-list ref] wrote:

quoted

We need to find one flag that can ask  gfpflags_normal_context() to
return false for this case.

Or if too difficult, stop only using sk->sk_allocation to decide
between the per-thread frag, or the per socket one.

I presume there are few active CIFS sockets on a host. They should use
a per socket sk_frag.

A pure networking change could be to use a new MSG_  flag to force sendmsg()
to use the sk->sk_frag, but that would add yet another test in TCP fast path.

About gfpflags_normal_context being used by dlm : we can simply add our own
helper with a new name describing that we want:

Both being in process context, and not from page fault handler .

Thanks for the hints! I'm testing the following (replacing the helper
with direct usage of a suitable check). Plus some adjusting to the
sk_page_frag() description.

Paolo
---

diff --git a/include/net/sock.h b/include/net/sock.h
index b32906e1ab55..8ae6fecfa18c 100644
--- a/include/net/sock.h
+++ b/include/net/sock.h

@@ -2442,7 +2442,8 @@ static inline void sk_stream_moderate_sndbuf(struct sock *sk)
  */
 static inline struct page_frag *sk_page_frag(struct sock *sk)
 {
-       if (gfpflags_normal_context(sk->sk_allocation))
+       if ((sk->sk_allocation & (__GFP_DIRECT_RECLAIM | __GFP_MEMALLOC | __GFP_FS)) ==
+           (__GFP_DIRECT_RECLAIM | __GFP_FS))
                return &current->task_frag;
 
        return &sk->sk_frag;