Hi,
On 11/09/2016 05:23 PM, Thomas Gleixner wrote:
> On Tue, 1 Nov 2016, Lu Baolu wrote:
>> +static int __init xdbc_init(void)
>> +{
> ...
>> + base = ioremap_nocache(xdbc.xhci_start, xdbc.xhci_length);
>> + if (!base) {
>> + xdbc_trace("failed to remap the io address\n");
>> + ret = -ENOMEM;
>> + goto free_and_quit;
>> + }
>> +
>> + early_iounmap(xdbc.xhci_base, xdbc.xhci_length);
>> + xdbc_trace("early mapped IO address released\n");
>> +
>> + xdbc.xhci_base = base;
>> + offset = xhci_find_next_ext_cap(xdbc.xhci_base, 0, XHCI_EXT_CAPS_DEBUG);
>> + xdbc.xdbc_reg = (struct xdbc_regs __iomem *)(xdbc.xhci_base + offset);
> This is broken. What prevents that
>
> - a printk is in progress on another cpu?
>
> - a printk happens between the unmap and storing the new base ?
>
> Nothing AFAICT. So this needs to be done in a safe way. And just making it
>
> oldbase = xdbc.xhci_base;
> base = ioremap(....);
> xdbc.xhci_base = base;
> early_iounmap(oldbase);
>
> does not work either because the compiler can rightfully cache
> xdbc.xhci_base in the write related functions. The same issue with
> xdbc.xdbc_reg.
If there isn't a good solution, I will remove the ioremap code and let
it use the early mapped one.
Best regards,
Lu Baolu
