Re: [PATCH V1 8/9] bnxt_en: Add TPH support in BNXT driver

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 2024-05-10 08:23, Andy Gospodarek wrote:
> On Fri, May 10, 2024 at 11:35:35AM +0100, Vadim Fedorenko wrote:
>> On 10.05.2024 04:55, Ajit Khaparde wrote:
>>> On Thu, May 9, 2024 at 2:50 PM Vadim Fedorenko
>>> <vadim.fedorenko@xxxxxxxxx> wrote:
>>>>
>>>> On 09/05/2024 17:27, Wei Huang wrote:
>>>>> From: Manoj Panicker <manoj.panicker2@xxxxxxx>
>>>>>
>>>>> As a usage example, this patch implements TPH support in Broadcom BNXT
>>>>> device driver by invoking pcie_tph_set_st() function when interrupt
>>>>> affinity is changed.
>>>>>
>>>>> Reviewed-by: Ajit Khaparde <ajit.khaparde@xxxxxxxxxxxx>
>>>>> Reviewed-by: Andy Gospodarek <andrew.gospodarek@xxxxxxxxxxxx>
>>>>> Reviewed-by: Wei Huang <wei.huang2@xxxxxxx>
>>>>> Signed-off-by: Manoj Panicker <manoj.panicker2@xxxxxxx>
>>>>> ---
>>>>>    drivers/net/ethernet/broadcom/bnxt/bnxt.c | 51 +++++++++++++++++++++++
>>>>>    drivers/net/ethernet/broadcom/bnxt/bnxt.h |  4 ++
>>>>>    2 files changed, 55 insertions(+)
>>>>>
>>>>> diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.c b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
>>>>> index 2c2ee79c4d77..be9c17566fb4 100644
>>>>> --- a/drivers/net/ethernet/broadcom/bnxt/bnxt.c
>>>>> +++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.c
>>>>> @@ -55,6 +55,7 @@
>>>>>    #include <net/page_pool/helpers.h>
>>>>>    #include <linux/align.h>
>>>>>    #include <net/netdev_queues.h>
>>>>> +#include <linux/pci-tph.h>
>>>>>
>>>>>    #include "bnxt_hsi.h"
>>>>>    #include "bnxt.h"
>>>>> @@ -10491,6 +10492,7 @@ static void bnxt_free_irq(struct bnxt *bp)
>>>>>                                free_cpumask_var(irq->cpu_mask);
>>>>>                                irq->have_cpumask = 0;
>>>>>                        }
>>>>> +                     irq_set_affinity_notifier(irq->vector, NULL);
>>>>>                        free_irq(irq->vector, bp->bnapi[i]);
>>>>>                }
>>>>>
>>>>> @@ -10498,6 +10500,45 @@ static void bnxt_free_irq(struct bnxt *bp)
>>>>>        }
>>>>>    }
>>>>>
>>>>> +static void bnxt_rtnl_lock_sp(struct bnxt *bp);
>>>>> +static void bnxt_rtnl_unlock_sp(struct bnxt *bp);
>>>>> +static void bnxt_irq_affinity_notify(struct irq_affinity_notify *notify,
>>>>> +                                  const cpumask_t *mask)
>>>>> +{
>>>>> +     struct bnxt_irq *irq;
>>>>> +
>>>>> +     irq = container_of(notify, struct bnxt_irq, affinity_notify);
>>>>> +     cpumask_copy(irq->cpu_mask, mask);
>>>>> +
>>>>> +     if (!pcie_tph_set_st(irq->bp->pdev, irq->msix_nr,
>>>>> +                          cpumask_first(irq->cpu_mask),
>>>>> +                          TPH_MEM_TYPE_VM, PCI_TPH_REQ_TPH_ONLY))
>>>>> +             pr_err("error in configuring steering tag\n");
>>>>> +
>>>>> +     if (netif_running(irq->bp->dev)) {
>>>>> +             rtnl_lock();
>>>>> +             bnxt_close_nic(irq->bp, false, false);
>>>>> +             bnxt_open_nic(irq->bp, false, false);
>>>>> +             rtnl_unlock();
>>>>> +     }
>>>>
>>>> Is it really needed? It will cause link flap and pause in the traffic
>>>> service for the device. Why the device needs full restart in this case?
>>>
>>> In that sequence only the rings are recreated for the hardware to sync
>>> up the tags.
>>>
>>> Actually its not a full restart. There is no link reinit or other
>>> heavy lifting in this sequence.
>>> The pause in traffic may be momentary. Do IRQ/CPU affinities change frequently?
>>> Probably not?
>>
>> From what I can see in bnxt_en, proper validation of link_re_init parameter is
>> not (yet?) implemented, __bnxt_open_nic will unconditionally call
>> netif_carrier_off() which will be treated as loss of carrier with counters
>> increment and proper events posted. Changes to CPU affinities were
>> non-disruptive before the patch, but now it may break user-space
>> assumptions.
> 
> From my testing the link should not flap.  I just fired up a recent net-next
> and confirmed the same by calling $ ethtool -G ens7f0np0 rx 1024 which does a
> similar bnxt_close_nic(bp, false, false)/bnxt_open_nic(bp, false, false) as
> this patch.  Link remained up -- even with a non-Broadocm link-partner.
> 
>> Does FW need full rings re-init to update target value, which is one u32 write?
>> It looks like overkill TBH.
> 
> Full rings do not, but the initialization of that particular ring associated
> with this irq does need to be done.  On my list of things we need to do in
> bnxt_en is implement the new ndo_queue_stop/start and ndo_queue_mem_alloc/free
> operations and once those are done we could make a switch as that may be less
> disruptive.

Hi Andy, I have an implementation of the new ndo_queue_stop/start() API
[1] and would appreciate comments. I've been trying to test it but
without avail due to FW issues.

[1]: https://lore.kernel.org/netdev/20240502045410.3524155-1-dw@xxxxxxxxxxx/

> 
>> And yes, affinities can be change on fly according to the changes of the
>> workload on the host.
>>
>>>>
>>>>
>>>>> +}
>>>>> +
>>>>> +static void bnxt_irq_affinity_release(struct kref __always_unused *ref)
>>>>> +{
>>>>> +}
>>>>> +
>>>>> +static inline void __bnxt_register_notify_irqchanges(struct bnxt_irq *irq)
>>>>
>>>> No inlines in .c files, please. Let compiler decide what to inline.
>>>>
>>>>> +{
>>>>> +     struct irq_affinity_notify *notify;
>>>>> +
>>>>> +     notify = &irq->affinity_notify;
>>>>> +     notify->irq = irq->vector;
>>>>> +     notify->notify = bnxt_irq_affinity_notify;
>>>>> +     notify->release = bnxt_irq_affinity_release;
>>>>> +
>>>>> +     irq_set_affinity_notifier(irq->vector, notify);
>>>>> +}
>>>>> +
>>>>>    static int bnxt_request_irq(struct bnxt *bp)
>>>>>    {
>>>>>        int i, j, rc = 0;
>>>>> @@ -10543,6 +10584,7 @@ static int bnxt_request_irq(struct bnxt *bp)
>>>>>                        int numa_node = dev_to_node(&bp->pdev->dev);
>>>>>
>>>>>                        irq->have_cpumask = 1;
>>>>> +                     irq->msix_nr = map_idx;
>>>>>                        cpumask_set_cpu(cpumask_local_spread(i, numa_node),
>>>>>                                        irq->cpu_mask);
>>>>>                        rc = irq_set_affinity_hint(irq->vector, irq->cpu_mask);
>>>>> @@ -10552,6 +10594,15 @@ static int bnxt_request_irq(struct bnxt *bp)
>>>>>                                            irq->vector);
>>>>>                                break;
>>>>>                        }
>>>>> +
>>>>> +                     if (!pcie_tph_set_st(bp->pdev, i,
>>>>> +                                          cpumask_first(irq->cpu_mask),
>>>>> +                                          TPH_MEM_TYPE_VM, PCI_TPH_REQ_TPH_ONLY)) {
>>>>> +                             netdev_err(bp->dev, "error in setting steering tag\n");
>>>>> +                     } else {
>>>>> +                             irq->bp = bp;
>>>>> +                             __bnxt_register_notify_irqchanges(irq);
>>>>> +                     }
>>>>>                }
>>>>>        }
>>>>>        return rc;
>>>>> diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt.h b/drivers/net/ethernet/broadcom/bnxt/bnxt.h
>>>>> index dd849e715c9b..0d3442590bb4 100644
>>>>> --- a/drivers/net/ethernet/broadcom/bnxt/bnxt.h
>>>>> +++ b/drivers/net/ethernet/broadcom/bnxt/bnxt.h
>>>>> @@ -1195,6 +1195,10 @@ struct bnxt_irq {
>>>>>        u8              have_cpumask:1;
>>>>>        char            name[IFNAMSIZ + 2];
>>>>>        cpumask_var_t   cpu_mask;
>>>>> +
>>>>> +     int             msix_nr;
>>>>> +     struct bnxt     *bp;
>>>>> +     struct irq_affinity_notify affinity_notify;
>>>>>    };
>>>>>
>>>>>    #define HWRM_RING_ALLOC_TX  0x1
>>>>
>>




[Index of Archives]     [DMA Engine]     [Linux Coverity]     [Linux USB]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Greybus]

  Powered by Linux