Re: [PATCH 3/4] [Patch-next] ACPI, APEI Fix the return value(==NULL) of acpi_pre_map always.

[Huang Ying <ying.huang@intel.com>]

On Tue, 2010-08-17 at 10:43 +0800, Jin Dongming wrote:
> (2010/08/17 10:37), Huang Ying wrote:
> > On Tue, 2010-08-17 at 08:56 +0800, Jin Dongming wrote:
> >> acpi_pre_map() is used for remapping the physical address of I/O, so
> >> it should be return NULL or remapped virtual address. The problem
> >> is whether I/O remapping is successful or not, the function returns
> >> NULL always.
> > 
> > No. NULL is returned for error path only. Please check the code again.
> 
> There three places used the local variable vaddr in acpi_pre_map() in next-tree.
>    1. 115         vaddr = __acpi_try_ioremap(paddr, size);
>    2. 122         vaddr = ioremap(pg_off, pg_sz);
>    3. 135         vaddr = __acpi_try_ioremap(paddr, size);
> 
> Let's think about the following statement.
>    Assumption: the physical address has never been remapped.
>    Step:
>        1. vaddr == NULL
>           Because the physical address is not registered in the acpi_iomaps,
>           it should be returned NULL from __acpi_try_ioremap().
> 
>        2. vaddr == the virtual address of the physical address.
>           Here if ioremap is successful, the value of vaddr should be 
>           the virtual address returned from ioremap().
> 
>        3. vaddr == NULL                <== IMPORTANT
>           Here it is because the physical address has not been registered
>           in the acpi_iomaps yet, it still return NULL from __acpi_try_ioremap().
>           So it is why vaddr == NULL, even if the physical address has never
>           been remapped.
> 
>     Result: vaddr == NULL.

return vaddr + (paddr - pg_off), is not NULL for common cases.

> And if the vaddr is not NULL, it could not be added into acpi_iomaps.
> Codes in acpi_pre_map() is like following.
> 
> 134         spin_lock_irqsave(&acpi_iomaps_lock, flags);
> 135         vaddr = __acpi_try_ioremap(paddr, size);           <== the 3rd step
> 136         if (vaddr) {
> 137                 spin_unlock_irqrestore(&acpi_iomaps_lock, flags);
> 138                 iounmap(map->vaddr);
> 139                 kfree(map);
> 140                 return vaddr;
> 141         }
> 142         list_add_tail_rcu(&map->list, &acpi_iomaps);      <== add into acpi_iomaps.
> 143         spin_unlock_irqrestore(&acpi_iomaps_lock, flags);

Oops, this is a bug, and your patch really fix it. But I think you
should change the patch description. NULL is not returned in almost all
cases. Because (paddr - pg_off) is not zero in common cases.

Best Regards,
Huang Ying