Re: [GIT PULL] PMEM driver for v4.1

[Boaz Harrosh <boaz@plexistor.com>]

On 04/13/2015 03:35 PM, Ingo Molnar wrote:
> 
<>
> 
> How does splice work with DAX files? AFAICS vmsplice() won't work, as 
> it uses get_user_pages(), which needs struct page backing. Also, how 
> will f_op->sendpage() work? That too needs page backing.
> 

I'm not sure about f_op->sendpage(). I do not see it in ext4 so I assumed
it works through f_op->splice_read/write.

The way f_op->splice_XX works is through the read/write_iter system
and this one is fully supported with DAX. It will do an extra copy.

[You can see this patch in linux-next:
	8b22559 dax: unify ext2/4_{dax,}_file_operations
 that explains a bit about this issue
]

>>>    - zero copy O_DIRECT into DAX areas.
>>
>> DAX is always O_DIRECT.
>>
>> What does not work is mmap of DAX file and use that pointer in an
>> O_DIRECT operation of another device. (unless it is a DAX device)
> 
> That is what I meant: O_DIRECT into (or from) DAX mmap()-ed areas, 
> from a different device.
> 

Yes. For me the one missing the most is RDMA to/from pmem copy-less.

Is why I'm pushing for my page-struct patches. But so far people
rejected it. I hope to resend it soon, once the dust settles and see
if people might change their mind.

> I'm not promoting the use of these APIs, some of them are quirky, just 
> wanted to list the known limitations. The pmem driver is already 
> useful as-is.
> 
> Thanks,
> 	Ingo
> 

Thanks
Boaz