Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]
Groups > linux.kernel > #1215644 > unrolled thread
| Started by | Alexey Klimov <klimov.linux@gmail.com> |
|---|---|
| First post | 2015-08-29 03:50 +0200 |
| Last post | 2015-08-30 18:40 +0200 |
| Articles | 3 — 3 participants |
Back to article view | Back to linux.kernel
Re: [PATCH net-next 6/8] net: thunderx: Rework interrupt handler Alexey Klimov <klimov.linux@gmail.com> - 2015-08-29 03:50 +0200
Re: [PATCH net-next 6/8] net: thunderx: Rework interrupt handler Aleksey Makarov <feumilieu@gmail.com> - 2015-08-30 11:30 +0200
Re: [PATCH net-next 6/8] net: thunderx: Rework interrupt handler Sunil Kovvuri <sunil.kovvuri@gmail.com> - 2015-08-30 18:40 +0200
| From | Alexey Klimov <klimov.linux@gmail.com> |
|---|---|
| Date | 2015-08-29 03:50 +0200 |
| Subject | Re: [PATCH net-next 6/8] net: thunderx: Rework interrupt handler |
| Message-ID | <q2Drj-44B-7@gated-at.bofh.it> |
Hi Aleksey,
let me add few minor points below.
On Fri, Aug 28, 2015 at 5:59 PM, Aleksey Makarov
<aleksey.makarov@auriga.com> wrote:
> From: Sunil Goutham <sgoutham@cavium.com>
>
> Rework interrupt handler to avoid checking IRQ affinity of
> CQ interrupts. Now separate handlers are registered for each IRQ
> including RBDR. Also register interrupt handlers for only those
> which are being used.
Also add nicvf_dump_intr_status() and use it in irq handler(s).
I suggest to check and extend commit message and think about commit
name. Maybe "net: thunderx: rework interrupt handling and
registration" at least?
Please consider possibility of splitting this patch into few patches too.
>
> Signed-off-by: Sunil Goutham <sgoutham@cavium.com>
> Signed-off-by: Aleksey Makarov <aleksey.makarov@caviumnetworks.com>
> ---
> drivers/net/ethernet/cavium/thunder/nic.h | 1 +
> drivers/net/ethernet/cavium/thunder/nicvf_main.c | 172 ++++++++++++---------
> drivers/net/ethernet/cavium/thunder/nicvf_queues.h | 2 +
> 3 files changed, 103 insertions(+), 72 deletions(-)
>
> diff --git a/drivers/net/ethernet/cavium/thunder/nic.h b/drivers/net/ethernet/cavium/thunder/nic.h
> index a83f567..89b997e 100644
> --- a/drivers/net/ethernet/cavium/thunder/nic.h
> +++ b/drivers/net/ethernet/cavium/thunder/nic.h
> @@ -135,6 +135,7 @@
> #define NICVF_TX_TIMEOUT (50 * HZ)
>
> struct nicvf_cq_poll {
> + struct nicvf *nicvf;
> u8 cq_idx; /* Completion queue index */
> struct napi_struct napi;
> };
> diff --git a/drivers/net/ethernet/cavium/thunder/nicvf_main.c b/drivers/net/ethernet/cavium/thunder/nicvf_main.c
> index de51828..2198f61 100644
> --- a/drivers/net/ethernet/cavium/thunder/nicvf_main.c
> +++ b/drivers/net/ethernet/cavium/thunder/nicvf_main.c
> @@ -653,11 +653,20 @@ static void nicvf_handle_qs_err(unsigned long data)
> nicvf_enable_intr(nic, NICVF_INTR_QS_ERR, 0);
> }
>
> +static inline void nicvf_dump_intr_status(struct nicvf *nic)
> +{
> + if (netif_msg_intr(nic))
> + netdev_info(nic->netdev, "%s: interrupt status 0x%llx\n",
> + nic->netdev->name, nicvf_reg_read(nic, NIC_VF_INT));
> +}
Please check if you really need to mark this 'inline' here.
> static irqreturn_t nicvf_misc_intr_handler(int irq, void *nicvf_irq)
> {
> struct nicvf *nic = (struct nicvf *)nicvf_irq;
> u64 intr;
>
> + nicvf_dump_intr_status(nic);
> +
> intr = nicvf_reg_read(nic, NIC_VF_INT);
> /* Check for spurious interrupt */
> if (!(intr & NICVF_INTR_MBOX_MASK))
> @@ -668,59 +677,58 @@ static irqreturn_t nicvf_misc_intr_handler(int irq, void *nicvf_irq)
> return IRQ_HANDLED;
> }
>
> -static irqreturn_t nicvf_intr_handler(int irq, void *nicvf_irq)
> +static irqreturn_t nicvf_intr_handler(int irq, void *cq_irq)
> +{
> + struct nicvf_cq_poll *cq_poll = (struct nicvf_cq_poll *)cq_irq;
> + struct nicvf *nic = cq_poll->nicvf;
> + int qidx = cq_poll->cq_idx;
> +
> + nicvf_dump_intr_status(nic);
> +
> + /* Disable interrupts */
> + nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx);
> +
> + /* Schedule NAPI */
> + napi_schedule(&cq_poll->napi);
> +
> + /* Clear interrupt */
> + nicvf_clear_intr(nic, NICVF_INTR_CQ, qidx);
> +
> + return IRQ_HANDLED;
> +}
You're not considering spurious irqs in all new irq handlers here and
below and schedule napi/tasklets unconditionally. Is it correct?
For me it looks like previous implementation relied on reading of
NIC_VF_INT to understand irq type and what actions should be
performed. It generally had idea that no interrupt might occur.
> +
> +static irqreturn_t nicvf_rbdr_intr_handler(int irq, void *nicvf_irq)
> {
> - u64 qidx, intr, clear_intr = 0;
> - u64 cq_intr, rbdr_intr, qs_err_intr;
> struct nicvf *nic = (struct nicvf *)nicvf_irq;
> - struct queue_set *qs = nic->qs;
> - struct nicvf_cq_poll *cq_poll = NULL;
> + u8 qidx;
>
> - intr = nicvf_reg_read(nic, NIC_VF_INT);
> - if (netif_msg_intr(nic))
> - netdev_info(nic->netdev, "%s: interrupt status 0x%llx\n",
> - nic->netdev->name, intr);
> -
> - qs_err_intr = intr & NICVF_INTR_QS_ERR_MASK;
> - if (qs_err_intr) {
> - /* Disable Qset err interrupt and schedule softirq */
> - nicvf_disable_intr(nic, NICVF_INTR_QS_ERR, 0);
> - tasklet_hi_schedule(&nic->qs_err_task);
> - clear_intr |= qs_err_intr;
> - }
>
> - /* Disable interrupts and start polling */
> - cq_intr = (intr & NICVF_INTR_CQ_MASK) >> NICVF_INTR_CQ_SHIFT;
> - for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
> - if (!(cq_intr & (1 << qidx)))
> - continue;
> - if (!nicvf_is_intr_enabled(nic, NICVF_INTR_CQ, qidx))
> + nicvf_dump_intr_status(nic);
> +
> + /* Disable RBDR interrupt and schedule softirq */
> + for (qidx = 0; qidx < nic->qs->rbdr_cnt; qidx++) {
> + if (!nicvf_is_intr_enabled(nic, NICVF_INTR_RBDR, qidx))
> continue;
> + nicvf_disable_intr(nic, NICVF_INTR_RBDR, qidx);
> + tasklet_hi_schedule(&nic->rbdr_task);
> + /* Clear interrupt */
> + nicvf_clear_intr(nic, NICVF_INTR_RBDR, qidx);
> + }
>
> - nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx);
> - clear_intr |= ((1 << qidx) << NICVF_INTR_CQ_SHIFT);
> + return IRQ_HANDLED;
> +}
>
> - cq_poll = nic->napi[qidx];
> - /* Schedule NAPI */
> - if (cq_poll)
> - napi_schedule(&cq_poll->napi);
> - }
> +static irqreturn_t nicvf_qs_err_intr_handler(int irq, void *nicvf_irq)
> +{
> + struct nicvf *nic = (struct nicvf *)nicvf_irq;
>
> - /* Handle RBDR interrupts */
> - rbdr_intr = (intr & NICVF_INTR_RBDR_MASK) >> NICVF_INTR_RBDR_SHIFT;
> - if (rbdr_intr) {
> - /* Disable RBDR interrupt and schedule softirq */
> - for (qidx = 0; qidx < qs->rbdr_cnt; qidx++) {
> - if (!nicvf_is_intr_enabled(nic, NICVF_INTR_RBDR, qidx))
> - continue;
> - nicvf_disable_intr(nic, NICVF_INTR_RBDR, qidx);
> - tasklet_hi_schedule(&nic->rbdr_task);
> - clear_intr |= ((1 << qidx) << NICVF_INTR_RBDR_SHIFT);
> - }
> - }
> + nicvf_dump_intr_status(nic);
> +
> + /* Disable Qset err interrupt and schedule softirq */
> + nicvf_disable_intr(nic, NICVF_INTR_QS_ERR, 0);
> + tasklet_hi_schedule(&nic->qs_err_task);
> + nicvf_clear_intr(nic, NICVF_INTR_QS_ERR, 0);
>
> - /* Clear interrupts */
> - nicvf_reg_write(nic, NIC_VF_INT, clear_intr);
> return IRQ_HANDLED;
> }
>
> @@ -754,7 +762,7 @@ static void nicvf_disable_msix(struct nicvf *nic)
>
> static int nicvf_register_interrupts(struct nicvf *nic)
> {
> - int irq, free, ret = 0;
> + int irq, ret = 0;
> int vector;
>
> for_each_cq_irq(irq)
> @@ -769,44 +777,42 @@ static int nicvf_register_interrupts(struct nicvf *nic)
> sprintf(nic->irq_name[irq], "NICVF%d RBDR%d",
> nic->vf_id, irq - NICVF_INTR_ID_RBDR);
>
> - /* Register all interrupts except mailbox */
> - for (irq = 0; irq < NICVF_INTR_ID_SQ; irq++) {
> + /* Register CQ interrupts */
> + for (irq = 0; irq < nic->qs->cq_cnt; irq++) {
> vector = nic->msix_entries[irq].vector;
> ret = request_irq(vector, nicvf_intr_handler,
> - 0, nic->irq_name[irq], nic);
> + 0, nic->irq_name[irq], nic->napi[irq]);
> if (ret)
> - break;
> + goto err;
> nic->irq_allocated[irq] = true;
> }
>
> - for (irq = NICVF_INTR_ID_SQ; irq < NICVF_INTR_ID_MISC; irq++) {
> + /* Register RBDR interrupt */
> + for (irq = NICVF_INTR_ID_RBDR;
> + irq < (NICVF_INTR_ID_RBDR + nic->qs->rbdr_cnt); irq++) {
> vector = nic->msix_entries[irq].vector;
> - ret = request_irq(vector, nicvf_intr_handler,
> + ret = request_irq(vector, nicvf_rbdr_intr_handler,
> 0, nic->irq_name[irq], nic);
> if (ret)
> - break;
> + goto err;
> nic->irq_allocated[irq] = true;
> }
>
> + /* Register QS error interrupt */
> sprintf(nic->irq_name[NICVF_INTR_ID_QS_ERR],
> "NICVF%d Qset error", nic->vf_id);
> - if (!ret) {
> - vector = nic->msix_entries[NICVF_INTR_ID_QS_ERR].vector;
> - irq = NICVF_INTR_ID_QS_ERR;
> - ret = request_irq(vector, nicvf_intr_handler,
> - 0, nic->irq_name[irq], nic);
> - if (!ret)
> - nic->irq_allocated[irq] = true;
> - }
> + irq = NICVF_INTR_ID_QS_ERR;
> + ret = request_irq(nic->msix_entries[irq].vector,
> + nicvf_qs_err_intr_handler,
> + 0, nic->irq_name[irq], nic);
> + if (!ret)
> + nic->irq_allocated[irq] = true;
>
> - if (ret) {
> - netdev_err(nic->netdev, "Request irq failed\n");
> - for (free = 0; free < irq; free++)
> - free_irq(nic->msix_entries[free].vector, nic);
> - return ret;
> - }
> +err:
> + if (ret)
> + netdev_err(nic->netdev, "request_irq failed, vector %d\n", irq);
>
> - return 0;
> + return ret;
> }
>
> static void nicvf_unregister_interrupts(struct nicvf *nic)
> @@ -815,8 +821,14 @@ static void nicvf_unregister_interrupts(struct nicvf *nic)
>
> /* Free registered interrupts */
> for (irq = 0; irq < nic->num_vec; irq++) {
> - if (nic->irq_allocated[irq])
> + if (!nic->irq_allocated[irq])
> + continue;
> +
> + if (irq < NICVF_INTR_ID_SQ)
> + free_irq(nic->msix_entries[irq].vector, nic->napi[irq]);
> + else
> free_irq(nic->msix_entries[irq].vector, nic);
> +
> nic->irq_allocated[irq] = false;
> }
>
> @@ -888,6 +900,20 @@ static netdev_tx_t nicvf_xmit(struct sk_buff *skb, struct net_device *netdev)
> return NETDEV_TX_OK;
> }
>
> +static inline void nicvf_free_cq_poll(struct nicvf *nic)
> +{
> + struct nicvf_cq_poll *cq_poll = NULL;
Please check if you really need initialize it to NULL here.
> + int qidx;
> +
> + for (qidx = 0; qidx < nic->qs->cq_cnt; qidx++) {
> + cq_poll = nic->napi[qidx];
> + if (!cq_poll)
> + continue;
> + nic->napi[qidx] = NULL;
> + kfree(cq_poll);
> + }
> +}
> +
> int nicvf_stop(struct net_device *netdev)
> {
> int irq, qidx;
> @@ -922,7 +948,6 @@ int nicvf_stop(struct net_device *netdev)
> cq_poll = nic->napi[qidx];
> if (!cq_poll)
> continue;
> - nic->napi[qidx] = NULL;
> napi_synchronize(&cq_poll->napi);
> /* CQ intr is enabled while napi_complete,
> * so disable it now
> @@ -931,7 +956,6 @@ int nicvf_stop(struct net_device *netdev)
> nicvf_clear_intr(nic, NICVF_INTR_CQ, qidx);
> napi_disable(&cq_poll->napi);
> netif_napi_del(&cq_poll->napi);
> - kfree(cq_poll);
> }
>
> netif_tx_disable(netdev);
> @@ -947,6 +971,8 @@ int nicvf_stop(struct net_device *netdev)
>
> nicvf_unregister_interrupts(nic);
>
> + nicvf_free_cq_poll(nic);
> +
> return 0;
> }
>
> @@ -973,6 +999,7 @@ int nicvf_open(struct net_device *netdev)
> goto napi_del;
> }
> cq_poll->cq_idx = qidx;
> + cq_poll->nicvf = nic;
> netif_napi_add(netdev, &cq_poll->napi, nicvf_poll,
> NAPI_POLL_WEIGHT);
> napi_enable(&cq_poll->napi);
> @@ -1040,6 +1067,8 @@ int nicvf_open(struct net_device *netdev)
> cleanup:
> nicvf_disable_intr(nic, NICVF_INTR_MBOX, 0);
> nicvf_unregister_interrupts(nic);
> + tasklet_kill(&nic->qs_err_task);
> + tasklet_kill(&nic->rbdr_task);
> napi_del:
> for (qidx = 0; qidx < qs->cq_cnt; qidx++) {
> cq_poll = nic->napi[qidx];
> @@ -1047,9 +1076,8 @@ napi_del:
> continue;
> napi_disable(&cq_poll->napi);
> netif_napi_del(&cq_poll->napi);
> - kfree(cq_poll);
> - nic->napi[qidx] = NULL;
> }
> + nicvf_free_cq_poll(nic);
> return err;
> }
>
> diff --git a/drivers/net/ethernet/cavium/thunder/nicvf_queues.h b/drivers/net/ethernet/cavium/thunder/nicvf_queues.h
> index 8b93dd6..c2ce270 100644
> --- a/drivers/net/ethernet/cavium/thunder/nicvf_queues.h
> +++ b/drivers/net/ethernet/cavium/thunder/nicvf_queues.h
> @@ -251,6 +251,8 @@ struct cmp_queue {
> void *desc;
> struct q_desc_mem dmem;
> struct cmp_queue_stats stats;
> + int irq;
> + cpumask_t affinity_mask;
> } ____cacheline_aligned_in_smp;
>
> struct snd_queue {
> --
> 2.5.0
>
>
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel
--
Best regards, Klimov Alexey
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [next] | [standalone]
| From | Aleksey Makarov <feumilieu@gmail.com> |
|---|---|
| Date | 2015-08-30 11:30 +0200 |
| Message-ID | <q3762-4RY-17@gated-at.bofh.it> |
| In reply to | #1215644 |
On 29.08.2015 04:44, Alexey Klimov wrote:
>> -static irqreturn_t nicvf_intr_handler(int irq, void *nicvf_irq)
>> +static irqreturn_t nicvf_intr_handler(int irq, void *cq_irq)
>> +{
>> + struct nicvf_cq_poll *cq_poll = (struct nicvf_cq_poll *)cq_irq;
>> + struct nicvf *nic = cq_poll->nicvf;
>> + int qidx = cq_poll->cq_idx;
>> +
>> + nicvf_dump_intr_status(nic);
>> +
>> + /* Disable interrupts */
>> + nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx);
>> +
>> + /* Schedule NAPI */
>> + napi_schedule(&cq_poll->napi);
>> +
>> + /* Clear interrupt */
>> + nicvf_clear_intr(nic, NICVF_INTR_CQ, qidx);
>> +
>> + return IRQ_HANDLED;
>> +}
>
> You're not considering spurious irqs in all new irq handlers here and
> below and schedule napi/tasklets unconditionally. Is it correct?
> For me it looks like previous implementation relied on reading of
> NIC_VF_INT to understand irq type and what actions should be
> performed. It generally had idea that no interrupt might occur.
1. The previous version of the handler did not handle spurious
interrupts either. Probably that means that the author of the patch
knows for sure that they do not happen.
2. Instead of reading the status register new version registers
different handlers for different irqs. I don't see why it can be wrong.
I am going to address your other suggestions in the next version of the
patchset.
Thank you
Aleksey Makarov
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [next] | [standalone]
| From | Sunil Kovvuri <sunil.kovvuri@gmail.com> |
|---|---|
| Date | 2015-08-30 18:40 +0200 |
| Message-ID | <q3dOa-61N-29@gated-at.bofh.it> |
| In reply to | #1215893 |
On Sun, Aug 30, 2015 at 2:50 PM, Aleksey Makarov <feumilieu@gmail.com> wrote:
> On 29.08.2015 04:44, Alexey Klimov wrote:
>
>>> -static irqreturn_t nicvf_intr_handler(int irq, void *nicvf_irq)
>>> +static irqreturn_t nicvf_intr_handler(int irq, void *cq_irq)
>>> +{
>>> + struct nicvf_cq_poll *cq_poll = (struct nicvf_cq_poll *)cq_irq;
>>> + struct nicvf *nic = cq_poll->nicvf;
>>> + int qidx = cq_poll->cq_idx;
>>> +
>>> + nicvf_dump_intr_status(nic);
>>> +
>>> + /* Disable interrupts */
>>> + nicvf_disable_intr(nic, NICVF_INTR_CQ, qidx);
>>> +
>>> + /* Schedule NAPI */
>>> + napi_schedule(&cq_poll->napi);
>>> +
>>> + /* Clear interrupt */
>>> + nicvf_clear_intr(nic, NICVF_INTR_CQ, qidx);
>>> +
>>> + return IRQ_HANDLED;
>>> +}
>>
>>
>> You're not considering spurious irqs in all new irq handlers here and
>> below and schedule napi/tasklets unconditionally. Is it correct?
>> For me it looks like previous implementation relied on reading of
>> NIC_VF_INT to understand irq type and what actions should be
>> performed. It generally had idea that no interrupt might occur.
>
>
> 1. The previous version of the handler did not handle spurious interrupts
> either. Probably that means that the author of the patch knows for sure
> that they do not happen.
Yes, no spurious interrupts are expected from hardware.
Even if it does the NAPI poll routine will handle it as valid
descriptor count would be zero.
Don't think it makes sense to check for spurious interrupt upon every interrupt.
>
> 2. Instead of reading the status register new version registers different
> handlers for different irqs. I don't see why it can be wrong.
Previous implementation results on scheduling multiple NAPI poll handlers
on the same CPU even if IRQ's affinities are set to different CPUs.
Hence they are seperated now.
>
> I am going to address your other suggestions in the next version of the
> patchset.
>
> Thank you
> Aleksey Makarov
>
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[toc] | [prev] | [standalone]
Back to top | Article view | linux.kernel
csiph-web