Re: [tipc-discussion] [PATCH v2 net-next 0/2] net: tipc: fix FB_MTU eat two pages and do some code cleanup

[Jon M. <jm...@re...>]

> -----Original Message-----
> From: Jon Maloy <jm...@re...>
> Sent: Thursday, June 10, 2021 4:12 PM
> To: Menglong Dong <men...@gm...>
> Cc: yin...@wi...; tip...@li...; Xin Long <lx...@re...>; tipc-dek <tip...@de...>
> Subject: Re: [PATCH v2 net-next 0/2] net: tipc: fix FB_MTU eat two pages and do some code cleanup
>
>
> On 6/9/21 8:56 AM, Menglong Dong wrote:
>> On Wed, Jun 9, 2021 at 6:47 PM Jon Maloy <jm...@re...> wrote:
>>>
>>> On 6/9/21 6:32 AM, men...@gm... wrote:
>>>> From: Menglong Dong <don...@zt...
[...]
>> No, no, I didn't miss your mail. I think it can make us clear about what and how
>> to do by sending the V2 patches.
>>
>> So we can define two versions 'FB_MTU' for bcast.c and msg.c, such as CRYPTO_MTU
>> and NON_CRYPTO_MTU. And within tipc_buf_acquire(), we decide which version
>> BUF_HEADROOM to use by the data size? Such as:
>>
>> int buf_size;
>> if (IS_ENABLED(CONFIG_TIPC_CRYPTO) && size == NON_CRYPTO_MTU) {
>>       buf_size = size + BUF_HEADROOM_non-crypto + BUF_TAILROOM_non-crypto;
>> } else {
>>       buf_size = size + BUF_HEADROOM_crypto + BUF_TAILROOM_crypto;
>> }
>>
>> Is this feeling?
>> (It's a little weird to check whether the data should be crypto by data size).
>>
>> Thanks!
>> Menglong Dong
> (Removed netdev, David Miller etc from thread)
>
> I think our main mistake is that we are trying to "compensate" for a
> behavior in tipc_buf_acquire() that doens't fit our purposes.
>
> What if we do the following:
> 1) We define BUF_HEADROOM and BUF_TAILROOM as we do now, plus a FB_MTU
> as you suggest, all inside msg.c.
> 2) We create a new inline function tipc_alloc_skb(int headroom, int
> tailroom, int size) in msg.c
>        This function does the job that tipc_buf_acqire() does now.
> 3) We let tipc_buf_acquire() call this function as
> tipc_alloc_skb(BUF_HEADROOM, BUF_TAILROOM, size);
>       Alternatively, we skip the BUF_HEADROOM/BUF_TAILROOM macros
> altogether andmake the code in tipc_buf_acquire() conditional as you
> did, but without testing for size. Actually, I think I like this better.
Maybe even better: we keep BUF_HEADROOM, but only define it as 
LL_MAX_HEADER + 48, i.e. the non-crypto value.
We still keep the conditional call in tipc_buf_acquire(), and add the 
crypto-header extras explictly when needed.
That way, we only need to change in one place if we decide to expand the 
headroom.
///jon

> 4) We let the fallback function in msg.c call it as
> tipc_alloc_skb(LL_MAX_HEADER + 48, 0, ONEPAGE_SKB);
> 5) We let the user in bcast.c get acces to the calculated value of
> FB_MTU as a function or global value, like you did in v2.
>
> I suggest you send v3 to tipc-discussion (if you are a member) and the
> recipients of this mail first, so we don't spam the netdev list with
> intermediate versions of this series.
>
> ///jon
>
>