Re: libata, devm_*, and MSI ?

[Mark Lord <liml@rtr.ca>]

Grant Grundler wrote:
> On Tue, Jan 20, 2009 at 10:16 AM, Mark Lord <liml@rtr.ca> wrote:
> ...
>>> I don't think that's necessary if free_irq() or disable_irq() are called.
>>> However, I'm not seeing those get called either.
> ...
>> The linux/Documentation/PCI/MSI-HOWTO.txt information is rather explicit
>> about it, and the pci core code does seem to expect it.
> 
> It looks like pci_disable_msi() requires free_irq() to be called first in order
> to guarantee the vectors are no longer in use. But I didn't see any code to
> enforce that. Which code was "expecting free_irq()" to be called?
..

The devres subsystem automatically frees resources on device removal,
including calling free_irq() as necessary for any interrupt handler
that was registered with devm_request_irq(), as done in libata.
See linux/drivers/base/devres.c for the actual call to free_irq().

But it doesn't handle pci_disable_msi() in there.

Thus this code from latest sata_mv:

+static void mv_pci_remove_one(struct pci_dev *pdev)
+{
+       struct device *dev = &pdev->dev;
+       struct ata_host *host = dev_get_drvdata(dev);
+       struct mv_host_priv *hpriv = host->private_data;
+
+       if (hpriv->hp_flags & MV_HP_FLAG_MSI) {
+               devm_free_irq(host->dev, pdev->irq, host);
+               pci_disable_msi(pdev);
+       }
+       ata_pci_remove_one(pdev);
+}

I believe that other MSI users might want something similar,
or perhaps Tejun could extend devres to include a pair
of suitable functions, devm_enable_msi() and  devm_disable_msi().
Then it would be just automatic for drivers, without any fuss.

Cheers