Re: [PATCH] mm: fix pfn calculation mistake in __ioremap_check_ram

[Dave Hansen <dave.hansen@intel.com>]

On 6/7/21 2:19 AM, Yaohui Wang wrote:
> According to the source code in function
> arch/x86/mm/ioremap.c:__ioremap_caller, after __ioremap_check_mem, if the
> mem range is IORES_MAP_SYSTEM_RAM, then __ioremap_caller should fail. But
> because of the pfn calculation problem, __ioremap_caller can success
> on IORES_MAP_SYSTEM_RAM region when the @size parameter is less than
> PAGE_SIZE. This may cause misuse of the ioremap function and raise the
> risk of performance issues. For example, ioremap(phys, PAGE_SIZE-1) may
> cause the direct memory mapping of @phys to be uncached, and iounmap won't
> revert this change. This patch fixes this issue.
> 
> In arch/x86/mm/ioremap.c:__ioremap_check_ram, start_pfn should wrap down
> the res->start address, and end_pfn should wrap up the res->end address.
> This makes the check more strict and should be more reasonable.

Was this found by inspection, or was there a real-world bug which this
patch addresses?