If rte_eth_conf.intr_conf.rxq != 0 and rte_eth_conf.intr_conf.lsc = 0 when
rte_eth_dev_configure() is called, rx queue interrupts can be enabled/disabled
with rte_eth_dev_rx_intr_{enable|disable} and DPDK applications can wait for rx
queue interrupts with rte_epoll_wait(). This is the case for both igb_uio and
vfio-pci drivers.
vfio-pci already uses non-threaded ISR, but currently we can only use igb_uio
when running DPDK applications in virtual machines.
David
-----Original Message-----
From: Stephen Hemminger [mailto:step...@networkplumber.org]
Sent: Friday, January 20, 2017 3:51 PM
To: Su, David W <david.w...@intel.com>
Cc: dev@dpdk.org
Subject: Re: [dpdk-dev] [PATCH] igb_uio: use non-threaded ISR
On Fri, 20 Jan 2017 15:08:19 -0800
David Su <david.w...@intel.com> wrote:
> This eliminates the overhead of a task switch when an interrupt arrives.
>
> Signed-off-by: David Su <david.w...@intel.com>
> ---
> lib/librte_eal/linuxapp/igb_uio/igb_uio.c | 3 ++-
> 1 files changed, 2 insertions(+), 1 deletions(-)
>
> diff --git a/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> b/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> index df41e45..9338e14 100644
> --- a/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> +++ b/lib/librte_eal/linuxapp/igb_uio/igb_uio.c
> @@ -382,6 +382,7 @@ igbuio_pci_probe(struct pci_dev *dev, const struct
> pci_device_id *id)
> msix_entry.entry = 0;
> if (pci_enable_msix(dev, &msix_entry, 1) == 0) {
> dev_dbg(&dev->dev, "using MSI-X");
> + udev->info.irq_flags = IRQF_NO_THREAD;
> udev->info.irq = msix_entry.vector;
> udev->mode = RTE_INTR_MODE_MSIX;
> break;
> @@ -390,7 +391,7 @@ igbuio_pci_probe(struct pci_dev *dev, const struct
> pci_device_id *id)
> case RTE_INTR_MODE_LEGACY:
> if (pci_intx_mask_supported(dev)) {
> dev_dbg(&dev->dev, "using INTX");
> - udev->info.irq_flags = IRQF_SHARED;
> + udev->info.irq_flags = IRQF_SHARED | IRQF_NO_THREAD;
> udev->info.irq = dev->irq;
> udev->mode = RTE_INTR_MODE_LEGACY;
> break;
Since interrupts are only used for link state transistions with igb_uio,
I can't see how the overhead of task switch would matter.
