Re: [PATCH net-next v3 1/5] net: phy: microchip_ptp : Add header file for Microchip ptp library

[Vadim Fedorenko <vadim.fedorenko@linux.dev>]

On 12/11/2024 23:11, Andrew Lunn wrote:
> On Tue, Nov 12, 2024 at 10:56:19PM +0000, Vadim Fedorenko wrote:
>> On 12/11/2024 22:26, Andrew Lunn wrote:
>>>> I believe, the current design of mchp_ptp_clock has some issues:
>>>>
>>>> struct mchp_ptp_clock {
>>>>           struct mii_timestamper     mii_ts;             /*     0    48 */
>>>>           struct phy_device *        phydev;             /*    48     8 */
>>>>           struct sk_buff_head        tx_queue;           /*    56    24 */
>>>>           /* --- cacheline 1 boundary (64 bytes) was 16 bytes ago --- */
>>>>           struct sk_buff_head        rx_queue;           /*    80    24 */
>>>>           struct list_head           rx_ts_list;         /*   104    16 */
>>>>           spinlock_t                 rx_ts_lock          /*   120     4 */
>>>>           int                        hwts_tx_type;       /*   124     4 */
>>>>           /* --- cacheline 2 boundary (128 bytes) --- */
>>>>           enum hwtstamp_rx_filters   rx_filter;          /*   128     4 */
>>>>           int                        layer;              /*   132     4 */
>>>>           int                        version;            /*   136     4 */
>>>>
>>>>           /* XXX 4 bytes hole, try to pack */
>>>>
>>>>           struct ptp_clock *         ptp_clock;          /*   144     8 */
>>>>           struct ptp_clock_info      caps;               /*   152   184 */
>>>>           /* --- cacheline 5 boundary (320 bytes) was 16 bytes ago --- */
>>>>           struct mutex               ptp_lock;           /*   336    32 */
>>>>           u16                        port_base_addr;     /*   368     2 */
>>>>           u16                        clk_base_addr;      /*   370     2 */
>>>>           u8                         mmd;                /*   372     1 */
>>>>
>>>>           /* size: 376, cachelines: 6, members: 16 */
>>>>           /* sum members: 369, holes: 1, sum holes: 4 */
>>>>           /* padding: 3 */
>>>>           /* last cacheline: 56 bytes */
>>>> };
>>>>
>>>> tx_queue will be splitted across 2 cache lines and will have spinlock on the
>>>> cache line next to `struct sk_buff * next`. That means 2 cachelines
>>>> will have to fetched to have an access to it - may lead to performance
>>>> issues.
>>>>
>>>> Another issue is that locks in tx_queue and rx_queue, and rx_ts_lock
>>>> share the same cache line which, again, can have performance issues on
>>>> systems which can potentially have several rx/tx queues/irqs.
>>>>
>>>> It would be great to try to reorder the struct a bit.
>>>
>>> Dumb question: How much of this is in the hot patch? If this is only
>>> used for a couple of PTP packets per second, do we care about a couple
>>> of cache misses per second? Or will every single packet the PHY
>>> processes be affected by this?
>>
>> Even with PTP packets timestamped only - imagine someone trying to run
>> PTP server part with some proper amount of clients? And it's valid to
>> configure more than 1 sync packet per second. It may become quite hot.
> 
> I'm just thinking of Donald Knuth:
> 
> “The real problem is that programmers have spent far too much time
> worrying about efficiency in the wrong places and at the wrong times;
> premature optimization is the root of all evil (or at least most of
> it) in programming.”

It's hard to object to this argument :)
I might be influenced to much by the latest findings in bnxt_en
regarding bottlenecks in PTP processing..