Re: [PATCH 1/6] pmem: add force casts to avoid __iomem annotation

[Dan Williams <dan.j.williams@intel.com>]

On Fri, May 29, 2015 at 5:53 AM, Dan Williams <dan.j.williams@intel.com> wrote:
> On Fri, May 29, 2015 at 4:39 AM, Ross Zwisler <zwisler@gmail.com> wrote:
>> On Thu, 2015-05-28 at 15:47 -0700, Dan Williams wrote:
>>> On Thu, May 28, 2015 at 3:35 PM, Ross Zwisler
>>> <ross.zwisler@linux.intel.com> wrote:
>>> > Even though we use ioremap_nocache() to map our persistent memory
>>> > in the
>>> > pmem driver, the memory we are mapping behaves like normal memory
>>> > and
>>> > not I/O memory in that it can be accessed using regular memcpy()
>>> > operations and doesn't need to go through memcpy_toio() and
>>> > memcpy_fromio().  Force casting the pointers received from
>>> > ioremap_nocache() and given to iounmap() gives us the correct
>>> > behavior
>>> > and makes sparse happy.
>>> >
>>> > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
>>> > Cc: Dan Williams <dan.j.williams@intel.com>
>>> > Cc: linux-nvdimm@lists.01.org
>>> > ---
>>> >  drivers/block/nd/pmem.c | 7 ++++---
>>> >  1 file changed, 4 insertions(+), 3 deletions(-)
>>> >
>>> > diff --git a/drivers/block/nd/pmem.c b/drivers/block/nd/pmem.c
>>> > index 5e8c9c629f22..a8712e41e7f5 100644
>>> > --- a/drivers/block/nd/pmem.c
>>> > +++ b/drivers/block/nd/pmem.c
>>> > @@ -163,7 +163,8 @@ static struct pmem_device *pmem_alloc(struct
>>> > device *dev, struct resource *res,
>>> >          * of the CPU caches in case of a crash.
>>> >          */
>>> >         err = -ENOMEM;
>>> > -       pmem->virt_addr = ioremap_nocache(pmem->phys_addr, pmem
>>> > ->size);
>>> > +       pmem->virt_addr = (__force void *)ioremap_nocache(pmem
>>> > ->phys_addr,
>>> > +                       pmem->size);
>>>
>>> I think I'd rather see casting when ->virt_addr is used (the
>>> __io_virt() helper can be used to make this a tad cleaner), or
>>> provide
>>> ioremap apis that don't attach __iomem to their return value.
>>>  Because
>>> in this and other cases ioremap() is being on non "i/o" memory.
>>
>> The reason that I thought this was cleaner was that now when you look
>> at the pmem->virt_addr definition it is just a clean void* with no
>> annotations.  This correctly describes the memory to the user (it's
>> usable as regular memory, it's in the kernel address space, etc.).
>>
>> Having the pointer itself annotated with __iomem feels weird to me
>> because a random well meaning user could incorrectly try to use it as
>> I/O memory.
>
> pmem->virt_addr does not leak outside the driver to random well
> meaning users.  I think we have two options, provide physical address
> remap helpers from the outset (memremap()?) that don't attach __iomem,
> or put the casts on the non-iomem usages.

In other words, I don't like the fact that we apply a coarse / private
hack to fix a general kernel mis-annotation problem.  Either fix it
for everybody with something like memremap() or use __io_virt() to
document the non-iomem usages of pmem->virt_addr like the other
iomem-to-undecorated pointer helper routines in the kernel.