Re: [net PATCH] octeontx2-af: Fix marking couple of structure as __packed

[Jacob Keller <jacob.e.keller@intel.com>]

On 12/19/2023 7:26 AM, David Laight wrote:
> From: Jacob Keller
>> Sent: 18 December 2023 20:44
>>
>> On 12/18/2023 12:27 AM, Suman Ghosh wrote:
>>> Couple of structures was not marked as __packed which may have some
>>> performance implication. This patch fixes the same and mark them as
>>> __packed.
>>
>> Not sure I follow why lack of __packed would have performance
>> implications? I get that __packed is important to ensure layout is
>> correct or to ensure the whole structure has the right size rather than
>> unexpected gaps. I'd guess maybe because the structures size would
>> include padding without __packed, leading to a lot of gaps when
>> combining several structures together...
>>
>> I did test on my system with pahole, and even without __packed, I don't
>> get any gaps in the npc_lt_def_cfg structure:
>>
>>
>>> struct npc_lt_def_cfg {
>>>         struct npc_lt_def          rx_ol2;               /*     0     3 */
>>>         struct npc_lt_def          rx_oip4;              /*     3     3 */
>>>         struct npc_lt_def          rx_iip4;              /*     6     3 */
>>>         struct npc_lt_def          rx_oip6;              /*     9     3 */
>>>         struct npc_lt_def          rx_iip6;              /*    12     3 */
>>>         struct npc_lt_def          rx_otcp;              /*    15     3 */
>>>         struct npc_lt_def          rx_itcp;              /*    18     3 */
>>>         struct npc_lt_def          rx_oudp;              /*    21     3 */
>>>         struct npc_lt_def          rx_iudp;              /*    24     3 */
>>>         struct npc_lt_def          rx_osctp;             /*    27     3 */
>>>         struct npc_lt_def          rx_isctp;             /*    30     3 */
>>>         struct npc_lt_def_ipsec    rx_ipsec[2];          /*    33    10 */
>>>         struct npc_lt_def          pck_ol2;              /*    43     3 */
>>>         struct npc_lt_def          pck_oip4;             /*    46     3 */
>>>         struct npc_lt_def          pck_oip6;             /*    49     3 */
>>>         struct npc_lt_def          pck_iip4;             /*    52     3 */
>>>         struct npc_lt_def_apad     rx_apad0;             /*    55     4 */
>>>         struct npc_lt_def_apad     rx_apad1;             /*    59     4 */
>>>         struct npc_lt_def_color    ovlan;                /*    63     5 */
>>>         /* --- cacheline 1 boundary (64 bytes) was 4 bytes ago --- */
>>>         struct npc_lt_def_color    ivlan;                /*    68     5 */
>>>         struct npc_lt_def_color    rx_gen0_color;        /*    73     5 */
>>>         struct npc_lt_def_color    rx_gen1_color;        /*    78     5 */
>>>         struct npc_lt_def_et       rx_et[2];             /*    83    10 */
>>>
>>>         /* size: 93, cachelines: 2, members: 23 */
>>>         /* last cacheline: 29 bytes */
>>> };
>>
>>
>> However that may not be true across all compilers etc. Also all the
>> other structures are __packed. Makes sense.
> 
> Or not - maybe all the __packed should be removed instead!
> 
> Unless these structures (or any others) appear in 'messages' which
> get transferred between systems they really shouldn't be __packed.
> And a 93 byte 'message' with all those fields seems rather odd.
> 
> The above breakdown seems to imply everything is 'unsigned char'
> so the __packed makes no difference.
> 
> Using __packed requires the compiler generate byte loads/store
> with shifts (etc) on many architectures and should really be avoided
> unless it is absolutely needed for binary compatibility.
> 
> Even then if the problem is a 64bit field that only needs to be
> 32bit aligned (as is common for some compat32 code) then the 64bit
> fields should be marked as being 32bit aligned.
> 
> 	David
> 
Right. Typically packed is only required when dealing with something
where the exact binary layout matters (i.e. copying to/from hardware or
across systems in such a way that the layout might change with different
compilers/arch).

If that isn't how this structure is used, then ya, removing __packed
seems reasonable. And at least for one system I see no difference in the
actual generated layout, making __packed redundant.

However, its not clear to me at a glance how this structure is used and
whether it really is copied between places where binary compatibility is
a requirement.

> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)