Re: [PATCH] PCI device list locking

[Greg KH <greg@kroah.com>]

On Mon, Mar 04, 2002 at 03:54:11PM -0800, john stultz wrote:
> @@ -605,10 +620,12 @@
>  	int count = 0;
>  
>  	list_add_tail(&drv->node, &pci_drivers);
> +	spin_lock(&pcidevlist_lock);
>  	pci_for_each_dev(dev) {
>  		if (!pci_dev_driver(dev))
>  			count += pci_announce_device(drv, dev);
>  	}
> +	spin_unlock(&pcidevlist_lock);
>  	return count;
>  }
>  

pci_announce_device() calls the pci driver's probe() function, which can
easily sleep.  Not a good thing :(

> @@ -1935,9 +1959,11 @@
>  
>  	pcibios_init();
>  
> +	spin_lock(&pcidevlist_lock);
>  	pci_for_each_dev(dev) {
>  		pci_fixup_device(PCI_FIXUP_FINAL, dev);
>  	}
> +	spin_unlock(&pcidevlist_lock);
>  
>  #ifdef CONFIG_PM
>  	pm_register(PM_PCI_DEV, 0, pci_pm_callback);

pci_fixup_device() might also be able to sleep, but I didn't go through
all of the different fixup calls to make sure.  Did you look at these?

> diff -u linux24/drivers/pci/setup-bus.c:1.1.1.1 linux24/drivers/pci/setup-bus.c:1.1.1.1.28.3
> --- linux24/drivers/pci/setup-bus.c:1.1.1.1	Tue Dec 18 15:49:16 2001
> +++ linux24/drivers/pci/setup-bus.c	Mon Mar  4 15:33:30 2002
> @@ -222,9 +222,11 @@
>  		ranges.found_vga = 0;
>  		pbus_assign_resources(b, &ranges);
>  	}
> +	spin_lock(&pcidevlist_lock);
>  	pci_for_each_dev(dev) {
>  		pdev_enable_device(dev);
>  	}
> +	spin_unlock(&pcidevlist_lock);
>  }
>  
>  /* Check whether the bridge supports I/O forwarding.

Can all implementations of pci_write_config_* be safely called with a
spinlock held?

In short, I don't think you can use a spinlock, but rather a semaphore
is probably necessary.

thanks,

greg k-h

_______________________________________________
Linux-hotplug-devel mailing list  http://linux-hotplug.sourceforge.net
Linux-hotplug-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/linux-hotplug-devel