Yeah I can do a generic alloc function.

THANK YOU!

Why will it not work in some case ?

Because from what I can see on network drivers pci_dev_get() and pci_dev_put() are reference operations, I'd assume around the device and not direct operations. I also see pdev accesses after pci_dev_put() so it cannot be freed at that point:

iwlwifi/pcie/trans.c::iwl_trans_pcie_removal_wk()
        pci_dev_put(pdev);
        pci_stop_and_remove_bus_device(pdev);

It's the only call I have there so I had ignored it so far.

I assume the above just works for you in your special case.

I further assume the normal implementation would be similar to device_get/put():

static __inline void /* XXX-BZ returns something? */
pci_dev_get(struct pci_dev *pdev)
{
    if (pdev != NULL)
        device_get(&pdev->dev);
}

static __inline void
pci_dev_put(struct pci_dev *pdev)
{
        if (pdev != NULL)
           device_put(&pdev->dev);
}

I believe to achieve what you seem to want/need you'd want to do the following:

Rename this function to

static void
linuxkpi_pci_release_direct(struct device *dev)
(
... fix/cast dev to pdev
}

and move it up before your two new linuxkpi_pci_get_* functions or a new alloc function; there for the struct pci_dev *dev you want to set:

pdev->dev.release = linuxkpi_pci_release_direct;

But that means that you probably also have to set the parent and init the kobj in your two get_* or alloc functions above as does linux_pci_attach_device(). That means embracing the struct device which embedded in pci_dev as first argument rather than working around it. And you don't have to use it much more than that I'd assume.

I haven't tried this; only put it down from the top of my head.

PS: yes, linux_pci_attach_device() also lacks a release function for devres; this is how I got to figure that out yesterday.

Yeah I had a look at what linux is doing and you're right, it's just device_put.
I'll see at how to fix that.