Re: [PATCH net-next] net: tipc: fix FB_MTU eat two pages
From: Menglong Dong
Date: Mon Jun 07 2021 - 08:52:38 EST
On Sat, Jun 05, 2021 at 10:25:53AM -0400, Jon Maloy wrote:
>
>
> On 6/4/21 9:28 PM, Menglong Dong wrote:
> > Hello Maloy,
> >
> > On Sat, Jun 5, 2021 at 3:20 AM Jon Maloy <jmaloy@xxxxxxxxxx> wrote:
> > >
> > [...]
> > > Please don't add any extra file just for this little fix. We have enough
> > > files.
> > > Keep the macros in msg.h/c where they used to be. You can still add
> > > your copyright line to those files.
> > > Regarding the macros kept inside msg.c, they are there because we design
> > > by the principle of minimal exposure, even among our module internal files.
> > > Otherwise it is ok.
> > >
> > I don't want to add a new file too, but I found it's hard to define FB_MTU. I
> > tried to define it in msg.h, and 'crypto.h' should be included, which
> > 'BUF_HEADROOM' is defined in. However, 'msg.h' is already included in
> > 'crypto.h', so it doesn't work.
> >
> > I tried to define FB_MTU in 'crypto.h', but it feels weird to define
> > it here. And
> > FB_MTU is also used in 'bcast.c', so it can't be defined in 'msg.c'.
> >
> > I will see if there is a better solution.
> I think we can leverage the fact that this by definition is a node local
> message, and those are never encrypted.
> So, if you base FB_MTU on the non-crypto versions of BUF_HEADROOM and
> BUF_TAILROOM we should be safe.
> That will even give us better utilization of the space available.
>
I think we misunderstanded something. If I base FB_MTU on the non-crypto
version of BUF_TAILROOM, it will eat 2 pages too. Because the
BUF_TAILROOM used in tipc_buf_acquire() is the crypto version, which is
larger than the non-crypto version:
struct sk_buff *tipc_buf_acquire(u32 size, gfp_t gfp)
{
struct sk_buff *skb;
#ifdef CONFIG_TIPC_CRYPTO
unsigned int buf_size = (BUF_HEADROOM + size + BUF_TAILROOM + 3) & ~3u;
#else
unsigned int buf_size = (BUF_HEADROOM + size + 3) & ~3u;
#endif
[..]
}
So if I use the non-crypto version, the size allocated will be:
PAGE_SIZE - BUF_HEADROOM_non-crypto - BUF_TAILROOM_non-crypt +
BUF_HEADROOM_crypto + BUF_TAILROOM_crypto
which is larger than PAGE_SIZE.
So, I think the simple way is to define FB_MTU in 'crypto.h'. Is this
acceptable?
Thanks!
Menglong Dong