Re: [dpdk-dev] [PATCH 1/2] virtio: initialize iopl when device is initialized

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2015-03-06 08:20, Stephen Hemminger:
> The issue is that virtio has no place it can do iopl() and have the IRQ thread
> work. It only shows up on real code where application is daemon, not in a toy
> demo or test application.
> 
> Right now:
>     gcc start
>        rte_virtio_pmd_init
>           iopl
>     main
>         daemon
>             fork
>                 fork 
>                   Process is now child of init not original process
> 
>                   rte_eal_init
>                      fork (pthread) for irq thread
>                                                        irq thread
>                                                         (no iopl permssion)
>                   program start
>                   rte_pmd_virtio_configure

This call tree is wrong.
Below is the right (and more complete) one:

gcc start
	driver constructor
		rte_eal_driver_register (if .a)
main
	rte_eal_init
		eal_parse_args
		rte_eal_pci_init
		rte_eal_memory_init
		rte_eal_intr_init
			pthread_create
				eal_intr_thread_main
					eal_intr_handle_interrupts
						eal_intr_process_interrupts
							virtio_interrupt_handler
		dlopen (if .so)
			driver constructor
				rte_eal_driver_register
		rte_eal_dev_init
			driver->init
				rte_virtio_pmd_init
					rte_eal_iopl_init
					rte_eth_driver_register
		pthread_create
		rte_eal_pci_probe
			driver->devinit
				rte_eth_dev_init
					rte_eth_dev_allocate
					eth_drv->eth_dev_init
						eth_virtio_dev_init
							virtio_resource_init

> So the only place where iopl() can be done in the proper context
> so that the IRQ (and other helper threads in future) have the correct
> permissions is in eal_init.

No there are 2 other possible solutions:
1) Move rte_eal_intr_init() after rte_eal_dev_init().
2) Move dlopen() before rte_eal_intr_init() and call iopl in the constructor.
With the second solution, we must also keep an iopl call in rte_virtio_pmd_init()
to return an error if iopl fails.

The second solution was proposed in the series sent by David:
	http://dpdk.org/ml/archives/dev/2015-March/014931.html
	http://dpdk.org/ml/archives/dev/2015-March/014932.html