On 29/09/2021 17:44, Andrew Donnellan wrote:

On 29/9/21 11:43 pm, Uwe Kleine-König wrote:> I'm not a huge fan either, 
I used it to keep the control flow as is and

quoted

without introducing several calls to to_pci_driver.

The whole code looks as follows:

    list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
        struct pci_driver *afu_drv;
        if (afu_dev->dev.driver &&
            (afu_drv = 
to_pci_driver(afu_dev->dev.driver))->err_handler &&
            afu_drv->err_handler->resume)
            afu_drv->err_handler->resume(afu_dev);
    }

Without assignment in the if it could look as follows:

    list_for_each_entry(afu_dev, &afu->phb->bus->devices, bus_list) {
        struct pci_driver *afu_drv;

        if (!afu_dev->dev.driver)
            continue;

        afu_drv = to_pci_driver(afu_dev->dev.driver));

        if (afu_drv->err_handler && afu_drv->err_handler->resume)
            afu_drv->err_handler->resume(afu_dev);
    }

Fine for me.

This looks fine.

As an aside while writing my email I discovered the existence of 
container_of_safe(), a version of container_of() that handles the null 
and err ptr cases... if to_pci_driver() used that, the null check in the 
caller could be moved until after the to_pci_driver() call which would 
be neater.

But then, grep tells me that container_of_safe() is used precisely zero 
times in the entire tree. Interesting.

quoted

(Sidenote: What happens if the device is unbound directly after the
check for afu_dev->dev.driver? This is a problem the old code had, too
(assuming it is a real problem, didn't check deeply).)

Looking at any of the cxl PCI error handling paths brings back 
nightmares from a few years ago... Fred: I wonder if we need to add a 
lock here?

Yes, it's indeed a potential issue, there's nothing to prevent the afu 
driver to unbind during that window. Sigh..

   Fred