Re: [PATCH] x86/mm: Add warning to ioremap() for conflicting cache type

[Luis R. Rodriguez]

On Wed, Apr 27, 2016 at 09:25:57AM -0600, Toshi Kani wrote:
> On x86, ioremap() and remap_pfn_range() fail on conflicting cache
> type of an alias mapping request only if the case is not allowed
> by the rule set in is_new_memtype_allowed(). While this exemption
> is necessary for remap_pfn_range() called for /dev/mem mappings,
> it is not necessary for ioremap().

The wording here is a bit unclear. How about:

Currently we allow for conflicting cache requests to account for
reading out /dev/mem, however /dev/mem uses remap_pfn_range(). An
explicit exception is not currently known to exist for conflicting
cache type requests for ioremap(), however we are not certain if
such conflicting use type currently does exist.

> Drivers should never call
> ioremap() with a conflicting cache type. This exemption handling
> may hide possible bugs in drivers.

And this to:

Upon review, we've determined that at least device drivers have
no business in calling ioremap() with a conflicting cache type.

> Add a warning message to ioremap() when a conflicting cache type
> is allowed by is_new_memtype_allowed(). This 

will help us

> helps us identify
> such usages in drivers.

such valid uses in device drivers

> 
> There is no change in the behavior.

There are no functional changes introduced by this commit other than
the addition of the new warning.

> Link: https://lkml.org/lkml/2016/4/16/54

lkml.org references should be avoided, we have no control over
that domain. Please instead use:

http://lkml.kernel.org/r/20160416092018.GA8453@xxxxxxxxx

> Signed-off-by: Toshi Kani <toshi.kani@xxxxxxx>
> Cc: Ingo Molnar <mingo@xxxxxxxxxx>
> Cc: Thomas Gleixner <tglx@xxxxxxxxxxxxx>
> Cc: "H. Peter Anvin" <hpa@xxxxxxxxx>
> Cc: Borislav Petkov <bp@xxxxxxx>
> Cc: Luis R. Rodriguez <mcgrof@xxxxxxxx>
> Cc: Andy Lutomirski <luto@xxxxxxxxxx>
> Cc: Dan Williams <dan.j.williams@xxxxxxxxx>
> ---
>  arch/x86/mm/ioremap.c |   10 ++++++----
>  1 file changed, 6 insertions(+), 4 deletions(-)
> 
> diff --git a/arch/x86/mm/ioremap.c b/arch/x86/mm/ioremap.c
> index 0d8d53d..16c5887 100644
> --- a/arch/x86/mm/ioremap.c
> +++ b/arch/x86/mm/ioremap.c
> @@ -138,14 +138,16 @@ static void __iomem *__ioremap_caller(resource_size_t phys_addr,
>  	}
>  
>  	if (pcm != new_pcm) {
> -		if (!is_new_memtype_allowed(phys_addr, size, pcm, new_pcm)) {
> -			printk(KERN_ERR
> -		"ioremap error for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
> +		retval = is_new_memtype_allowed(phys_addr, size, pcm, new_pcm);
> +		pr_err(
> +		    "ioremap %s for 0x%llx-0x%llx, requested 0x%x, got 0x%x\n",
> +				retval ? "warning" : "error",
>  				(unsigned long long)phys_addr,
>  				(unsigned long long)(phys_addr + size),
>  				pcm, new_pcm);

This doesn't really make it humanly consumable, a pcm_to_str may help,
such use already exists but not generalized see pat_get_cache_mode().

> +		if (!retval)
>  			goto err_free_memtype;
> -		}
> +
>  		pcm = new_pcm;
>  	}

In the future we may be able to proactively hunt for these invalid uses
semantically but to do so we need first a demo driver with all known possible
bugs we can think of.

  Luis