[PATCH] Cardbus card memory assignment on x86_64 (if base==maxbase)

From: Bernhard Kaindl (bk_at_suse.de)
Date: 08/26/04

  • Next message: David S. Miller: "Re: Linux 2.6.9-rc1" 
    Date:	Thu, 26 Aug 2004 07:48:09 +0200 (CEST)
To: linux-kernel@vger.kernel.org

    Hi,
        I've sent the patch below to Ivan Kokshaysky (bk login: ink) and got
    a mail from him back that I should forward it and telling that it looks
    correct. According to bitkeeper logs for 2.4 he initially did the pci_size
    function this patch is about.

    I've debugged a memory assignment problem which happens even with 2.6.8.1
    on Amilo Laptops from Siemens Fujitsu with a RT2500 CardBus card.

    The problem was that the PCI scan didn't report a memory range for the
    card and thus the driver finally complained about not beind able to remap
    the memory.

    After plaching a dump_stack() into bus_add_device() to know thich
    functions were really involved, I worked my way forward into the PCI
    code.

    Enabling the debug statement in the PCI code, I've found pci_setup_device().
    From there I added more printk's. Ultimatively, I've found that pci_size()
    returned 0 for the memory size at this place in drivers/pci/probe.c:

      sz = pci_size(l, sz, PCI_BASE_ADDRESS_MEM_MASK);
      if (!sz)
         continue;

    Here is the pci_size():

    /*
      * Find the extent of a PCI decode..
      */
    static u32 pci_size(u32 base, u32 maxbase, unsigned long mask)
    {
        u32 size = mask & maxbase; /* Find the significant bits */
        if (!size)
           return 0;

        /* Get the lowest of them to find the decode size, and
        from that the extent. */
        size = (size & ~(size-1)) - 1;

        /* base == maxbase can be valid only if the BAR has
        already been programmed with all 1s. */
        if (base == maxbase && ((base | size) & mask) != mask)
           return 0;

        return size;
    }

    Adding the results of my debug, it's called like this - for my card:
    pci_size(u32 base = ffffe000, maxbase = ffffe000, mask = fffffffffffffff0)

    Finally the CPU reaches this line:
    if (base == maxbase && ((base | size) & mask) != mask)
      return 0;

    With the values from my debug printks, it looks like this:

    if (ffffe000 == ffffe000
      && ((ffffe000 | 1fff) & fffffffffffffff0) != fffffffffffffff0)

    Evaluation of the if:

       1): ffffe000 == ffffe000 is true

       2): (ffffe000 | 1fff) & fffffffffffffff0) results in fffffff0

       but this means 2) will always != fffffffffffffff0.
                                         (it's a large #defined mask)

    It can never be equal, because "(base | size) & mask)" (base and size 32bit)
    - the result of ((32-bit | 32-bit) & 64-bit)) can never match a large 64-bit
    value such as fffffffffffffff0.

    So, on 64-bit, calls to

          static u32 pci_size(u32 base, u32 maxbase, unsigned long mask)

    with base == maxbase and mask = ~0x0fUL will always result in
    return 0.

    I think that's the bug here, this function is intented to work
    on 32-bit PCI values.

    PCI64 is handled in a separate way, in pci_read_bases(), just a few
    lines after pci_size() is called:

    #if BITS_PER_LONG == 64
      res->start |= ((unsigned long) l) << 32;
      res->end = res->start + sz;
      pci_write_config_dword(dev, reg+4, ~0);
      pci_read_config_dword(dev, reg+4, &sz);
      pci_write_config_dword(dev, reg+4, l);
      if (~sz)
             res->end = res->start + 0xffffffff +
                                      (((unsigned long) ~sz) << 32);

    I believe, pci_size() should work with 32-bit vaules
    only and I've tested that the diff of

    -static u32 pci_size(u32 base, u32 maxbase, unsigned long mask)
    +static u32 pci_size(u32 base, u32 maxbase, u32 mask)

    makes my card's memory appear as it does in a 32-bit kernel where mask
    is a 32-bit value too.

    It seems to be really a corner case, I don't know if many other
    Cards provide such PCI config space where base == maxbase,
    but there could be some others out there as well.

    For these cards, the above diff fixes the memory assignment problem.

    I hope this mail helps some of you, the patch is attached, I'll forward
    this to the kernel maintainers.

    Bernhard

    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

  • Next message: David S. Miller: "Re: Linux 2.6.9-rc1"

    Relevant Pages