2016-12-02 16:45, Jianfeng Tan: > Depends-on: http://dpdk.org/dev/patchwork/patch/17493/
The above patch is marked as rejected. Do we want to reject also this patch? > When a DPDK application grab a PCI device, device and driver work > together to Rx/Tx packets. If the DPDK app crashes or gets killed, > there's no chance for DPDK driver to stop the device, which means > rte_eth_dev_stop() will not be called. > > This could result in problems. In virtio's case, the device (the > vhost backend), will keep working. If packets come, the vhost will > copy them into the vring, which is negotiated with the previous DPDK > app. But the resources, especially hugepages, are recycled by VM > kernel. What's worse, the memory could be allocated for other usage, > and re-written. This leads to an uncertain situation. Like this post > has reported, https://bugs.launchpad.net/qemu/+bug/1558175, QEMU's > vhost finds out wrong value, and exits the whole QEMU process. > > To make it right, we need to restart (or reset) the device and make > the device go into the initial state, when uio-generic or igb_uio > recycles the PCI device. There's a chance in uio framework to disable > devices when /dev/uioX gets closed. Here we enable the pci device > in open() hook and disable it in release() hook. > > However, if device is not enabled in probe() phase any more, we can > not register irq in probe() through uio_register_device(). To address > that, we invoke request_irq() to register callback directly. > > Similar change needs to be done in uio-generic driver. And vfio-pci > seems to have done that already. > > Signed-off-by: Jianfeng Tan <jianfeng....@intel.com>
