Question

I am writing a PCI driver for a simple test device.

Hardware is recognized correctly with lspci (as you can see my driver vabs has been registered):

04:02.0 Non-VGA unclassified device: Device bace:55aa
    Control: I/O+ Mem+ BusMaster+ SpecCycle- MemWINV- VGASnoop- ParErr- Stepping- SERR- FastB2B- DisINTx-
    Status: Cap- 66MHz- UDF- FastB2B- ParErr- DEVSEL=medium >TAbort- <TAbort- <MAbort- >SERR- <PERR- INTx-
    Latency: 0
    Region 0: Memory at f0000000 (32-bit, prefetchable) [size=16M]
    Kernel driver in use: vabs

Initialisation and deinitalisation of driver and PCI subystem works fine. I am getting a device number, and udev creates a device file.

When reading from the device file I get the following error message:

BUG: unable to handle kernel paging request at 00000000f0000000

I request PCI ressources in the initialisation successfully. This returns 00000000f0000000 for memstart0, which is my base address 0 for PCI.

memstart0 = pci_resource_start( pdev, 0 );
memlen = pci_resource_len( pdev, 0 );
if( request_mem_region(memstart0,memlen,pdev->dev.kobj.name)==NULL ) {
    dev_err(&pdev->dev,"Memory address conflict for device\n");
    goto cleanup_mem;
}

Trying to read from this memio address with the following code gives the mentioned error:

ssize_t driver_read(struct file *instance, char __user *buffer, size_t max_bytes_to_read, loff_t *offset) {
    u32 test;

    dev_dbg(vabs_dev,"copying from %p\n", (void *) memstart0);
    test = readl((void *) memstart0);

    return max_bytes_to_read;
}

I also tried other access functions like memcpy_fromio, ioread32, and direct pointer access with the same result.

The hardware works on a Windows machine. The only notable difference is that Windows reserves base address 0 as 00000000fd000000 while Linux reserves it as 00000000f0000000.

This is for non-profit educational purpose in a public school. Thanks for your help!

Answer 1

Read Documentation/IO-mapping.txt (search for “iomap”) and/or Chapter 15 of LDD3.

request_mem_region just ensures no other driver has already grabbed that memory region. You still need to map it into the kernel’s VM space using iomap before you can read/write it.

Note that the whole pci_resource_start etc. dance is somewhat deprecated. I believe the recommended approach these days is:

pci_request_regions(pdev, "myname");  /* to request regions for all BARs */

Then:

void __iomem *base = pci_iomap(pdev, 0, pci_resources_len(pdev,0)); /* to map BAR 0 */

Then:

ioread32(base + offset);  /* Or readl(base + offset), but this is more generic */

And finally, at the end:

pci_iounmap(pdev, base);  /* Release kernel VM mapping (undoes pci_iomap) */
pci_release_regions(pdev); /* Release all regions (undoes pci_request_regions) */

You can do the first two manually by combining pci_resource_start, pci_resource_len, request_mem_region, and iomap. But the above is (a) shorter and (b) generic between memory-mapped devices and those that use the old x86 I/O space. (Not that there are a whole lot of those around anymore.)