On 5 November 2014 20:39, Kumar Gala [off-list ref] wrote:

On Nov 5, 2014, at 6:55 AM, Ankit Jindal [off-list ref] wrote:

quoted

Hi Kumar,

On 31 October 2014 19:09, Kumar Gala [off-list ref] wrote:

quoted

On Oct 31, 2014, at 4:30 AM, Ankit Jindal [off-list ref] wrote:

quoted

Hi Kumar,

On 21 October 2014 12:08, Kumar Gala [off-list ref] wrote:

quoted

On Oct 21, 2014, at 7:56 AM, Ankit Jindal [off-list ref] wrote:

quoted

Currently, three types of mem regions are supported: UIO_MEM_PHYS,
UIO_MEM_LOGICAL and UIO_MEM_VIRTUAL. Among these UIO_MEM_PHYS helps
UIO driver export physcial memory to user space as non-cacheable
user memory. Typcially memory-mapped registers of a device are exported
to user space as UIO_MEM_PHYS type mem region. The UIO_MEM_PHYS type
is not efficient if dma-capable devices are capable of maintaining coherency
with CPU caches.

This patch adds new type UIO_MEM_PHYS_CACHE for mem regions to enable
cacheable access to physical memory from user space.

Signed-off-by: Ankit Jindal <redacted>
Signed-off-by: Tushar Jagad <redacted>
---
drivers/uio/uio.c          |   11 ++++++++---
include/linux/uio_driver.h |    1 +
2 files changed, 9 insertions(+), 3 deletions(-)

Rather than adding a new type, why not allow the driver to set the pgprot value, this way one has full control and we don?t need to keep adding types for various different cache attributions in the future.

Do you mean to add a new field pgprot_t in the memtype structure and
uio_mmap_physical will set vma->vm_page_prot to this value provided by
driver ? If this is the case then we will need to change all the
current uio based drivers which was the reason I preferred to have a
new mem type.

Please let me know if I have misunderstood anything.

I?m suggeting in uio_mmap_physical to do something like:

if (idev->info->set_pgprot)
       idev->info->set_pgprot(vma->vm_page_prot)
else
       vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);

And add a set_prprot callback to 'struct uio_info?.

Here?s patch from several years ago:

http://patchwork.ozlabs.org/patch/119224/

The suggested solution looks okey but not sure whether there is any
available drivers using different combinations. Also, I looked at the
available pgprot routines, looks like only pgprot_noncached and
pgprot_writecombine are the available ones. So if we are not going to
use these pgprot routines then driver might have architecture
dependent switches, which we should avoid.

There are cases that are arch/driver specific that do not fall into pgprot_noncached or pgprot_writecombine.  So I don?t see why we should limit them.  For example the Freescale networking guys need cacheable-noncoherent for some of their UIO work.

We can deal with arch specific issues during review of the UIO driver themselves.

Ok. But, in our case we do not want to set any special attribute,
instead just want to avoid setting non-cacheable attribute. So if we
go by the approach as in your patch, then I need to register a dummy
routine which does nothing.

I think another approach could be to have new mem type as in this
patch and add something like this to this patch:

if (cacheable) {
        /* check if driver want to set any special cacheable attribute
combination */
        if (idev->info->set_pgprot)
                idev->info->set_pgprot(vma->vm_page_prot)
} else {
        vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
}

Please let me know your view.

- k

--
Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora Forum,
a Linux Foundation Collaborative Project

Thanks,
Ankit