[dpdk-dev] [PATCH v5 2/7] ethdev: allocate max space for internal queue array

[Andrew Rybchenko]

On 10/12/21 2:06 AM, Ananyev, Konstantin wrote:
>>>>> At queue configure stage always allocate space for maximum possible
>>>>> number (RTE_MAX_QUEUES_PER_PORT) of queue pointers.
>>>>> That will allow 'fast' inline functions (eth_rx_burst, etc.) to refer
>>>>> pointer to internal queue data without extra checking of current number
>>>>> of configured queues.
>>>>> That would help in future to hide rte_eth_dev and related structures.
>>>>> It means that from now on, each ethdev port will always consume:
>>>>> ((2*sizeof(uintptr_t))* RTE_MAX_QUEUES_PER_PORT)
>>>>> bytes of memory for its queue pointers.
>>>>> With RTE_MAX_QUEUES_PER_PORT==1024 (default value) it is 16KB per port.
>>>>>
>>>>> Signed-off-by: Konstantin Ananyev <konstantin.ananyev at intel.com>
>>>>> ---
>>>>>   lib/ethdev/rte_ethdev.c | 36 +++++++++---------------------------
>>>>>   1 file changed, 9 insertions(+), 27 deletions(-)
>>>>>
>>>>> diff --git a/lib/ethdev/rte_ethdev.c b/lib/ethdev/rte_ethdev.c
>>>>> index ed37f8871b..c8abda6dd7 100644
>>>>> --- a/lib/ethdev/rte_ethdev.c
>>>>> +++ b/lib/ethdev/rte_ethdev.c
>>>>> @@ -897,7 +897,8 @@ eth_dev_rx_queue_config(struct rte_eth_dev *dev, uint16_t nb_queues)
>>>>>
>>>>>   	if (dev->data->rx_queues == NULL && nb_queues != 0) { /* first time configuration */
>>>>>   		dev->data->rx_queues = rte_zmalloc("ethdev->rx_queues",
>>>>> -				sizeof(dev->data->rx_queues[0]) * nb_queues,
>>>>> +				sizeof(dev->data->rx_queues[0]) *
>>>>> +				RTE_MAX_QUEUES_PER_PORT,
>>>>>   				RTE_CACHE_LINE_SIZE);
>>>>
>>>> Looking at it I have few questions:
>>>> 1. Why is nb_queues == 0 case kept as an exception? Yes,
>>>>     strictly speaking it is not the problem of the patch,
>>>>     DPDK will still segfault (non-debug build) if I
>>>>     allocate Tx queues only but call rte_eth_rx_burst().
>>>
>>> eth_dev_rx_queue_config(.., nb_queues=0) is used in few places to clean-up things.
>>
>> No, as far as I know. For Tx only application (e.g. traffic generator)
>> it is 100% legal to configure with tx_queues=X, rx_queues=0.
>> The same is for Rx only application (e.g. packet capture).
> 
> Yes, that is valid config for sure.
> I just pointed that simply ignoring 'nb_queues' value and
> always allocating space for max possible queues, i.e:
> 
> eth_dev_rx_queue_config(struct rte_eth_dev *dev, uint16_t nb_queues) 
> {
> ....
> - if (dev->data->rx_queues == NULL && nb_queues != 0) { /* first time configuration */
> + if (dev->data->rx_queues == NULL) {
> wouldn't work, as right now nb_queues == 0 has extra special meaning -
> do final cleanup and free dev->data->rx_queues.
> But re-reading the text below, it seems that I misunderstood you
> and it probably wasn't your intention anyway.
> 
>>
>>>
>>>>     After reading the patch description I thought that
>>>>     we're trying to address it.
>>>
>>> We do, though I can't see how we can address it in this patch.
>>> Though it is a good idea - I think I can add extra check in eth_dev_fp_ops_setup()
>>> or around and setup RX function pointers only when dev->data->rx_queues != NULL.
>>> Same for TX.
>>
>> You don't need to care about these pointers, if these arrays are
>> always allocated. See (3) below.
>>
>>>
>>>> 2. Why do we need to allocate memory dynamically?
>>>>     Can we just make rx_queues an array of appropriate size?
>>>
>>> Pavan already asked same question.
>>> My answer to him:
>>> Yep we can, and yes it will simplify this peace of code.
>>> The main reason I decided no to do this change now -
>>> it will change layout of the_eth_dev_data structure.
>>> In this series I tried to mininize(/avoid) changes in rte_eth_dev and rte_eth_dev_data,
>>> as much as possible to avoid any unforeseen performance and functional impacts.
>>> If we'll manage to make rte_eth_dev and rte_eth_dev_data private we can in future
>>> consider that one and other changes in rte_eth_dev and rte_eth_dev_data layouts
>>> without worrying about ABI breakage
>>
>> Thanks a lot. Makes sense.
>>
>>>>     May be wasting 512K unconditionally is too much.
>>>> 3. If wasting 512K is too much, I'd consider to move
>>>>     allocation to eth_dev_get(). If
>>>
>>> Don't understand where 512KB came from.
>>
>> 32 port * 1024 queues * 2 types * 8 pointer size
>> if we allocate as in (2) above.
>>
>>> each ethdev port will always consume:
>>> ((2*sizeof(uintptr_t))* RTE_MAX_QUEUES_PER_PORT)
>>> bytes of memory for its queue pointers.
>>> With RTE_MAX_QUEUES_PER_PORT==1024 (default value) it is 16KB per port.
>>
>> IMHO it will be a bit nicer if queue pointers arrays are allocated
>> on device get if size is fixed. It is just a suggestion. If you
>> disagree, feel free to drop it.
> 
> You mean - allocate these arrays somewhere at rte_eth_dev_allocate() path?

Yes, eth_dev_get() mentioned above is called from
rte_eth_dev_allocate().

> That sounds like an interesting idea, but seems too drastic to me at that stage.

Yes, of course, we can address it later.

> 
>>
>>>>>   		if (dev->data->rx_queues == NULL) {
>>>>>   			dev->data->nb_rx_queues = 0;
>>>>> @@ -908,21 +909,11 @@ eth_dev_rx_queue_config(struct rte_eth_dev *dev, uint16_t nb_queues)
>>>>>
>>>>>   		rxq = dev->data->rx_queues;
>>>>>
>>>>> -		for (i = nb_queues; i < old_nb_queues; i++)
>>>>> +		for (i = nb_queues; i < old_nb_queues; i++) {
>>>>>   			(*dev->dev_ops->rx_queue_release)(rxq[i]);
>>>>> -		rxq = rte_realloc(rxq, sizeof(rxq[0]) * nb_queues,
>>>>> -				RTE_CACHE_LINE_SIZE);
>>>>> -		if (rxq == NULL)
>>>>> -			return -(ENOMEM);
>>>>> -		if (nb_queues > old_nb_queues) {
>>>>> -			uint16_t new_qs = nb_queues - old_nb_queues;
>>>>> -
>>>>> -			memset(rxq + old_nb_queues, 0,
>>>>> -				sizeof(rxq[0]) * new_qs);
>>>>> +			rxq[i] = NULL;
>>>>
>>>> It looks like the patch should be rebased on top of
>>>> next-net main because of queue release patches.
>>>>
>>>> [snip]
>