Re: [Qemu-devel] [RESEND PATCH 2/2] hostmem-file: add an attribute 'align' to set its alignment

[Dan Williams]

On Fri, May 26, 2017 at 11:55 AM, Eduardo Habkost <address@hidden> wrote:
> On Fri, May 26, 2017 at 07:24:26AM -0700, Dan Williams wrote:
>> On Fri, May 26, 2017 at 12:16 AM, Haozhong Zhang
>> <address@hidden> wrote:
>> > On 05/26/17 07:05 +0000, Marc-André Lureau wrote:
>> >> Hi
>> >>
>> >> On Fri, May 26, 2017 at 10:51 AM Haozhong Zhang <address@hidden>
>> >> wrote:
>> >>
>> >> > On 05/26/17 06:39 +0000, Marc-André Lureau wrote:
>> >> > > Hi
>> >> > >
>> >> > > On Fri, May 26, 2017 at 6:34 AM Haozhong Zhang <address@hidden
>> >> > >
>> >> > > wrote:
>> >> > >
>> >> > > > file_ram_alloc() currently maps the backend file via mmap to a 
>> >> > > > virtual
>> >> > > > address aligned to the value returned by qemu_fd_getpagesize(). 
>> >> > > > When a
>> >> > > > DAX device (e.g. /dev/dax0.0) is used as the backend file, its 
>> >> > > > kernel
>> >> > > > mmap implementation may require an alignment larger than what
>> >> > > > qemu_fd_get_pagesize() returns (e.g. 2MB vs. 4KB), and mmap may 
>> >> > > > fail.
>> >> > > >
>> >> > >
>> >> > > How is the accepted value queried? Any chance to configure it
>> >> > > automatically?
>> >> >
>> >> > Take /dev/dax0.0 for example. The value can be read from
>> >> > /sys/class/dax/dax0.0/device/dax_region/align.
>> >> >
>> >>
>> >> Should this work be left to management layer, or could qemu figure it out
>> >> by itself (using udev?)
>> >>
>> >
>> > For DAX device only, QEMU can figure out the proper alignment by
>> > itself. However, I'm not sure whether there are other non-DAX cases
>> > requiring non-default alignment, so I think it's better to just add an
>> > interface (i.e. align attribute) in QEMU and let other management
>> > tools (e.g. libvirt?) fill a proper value.
>>
>> I can't imagine any cases where you would want to specify an
>> alignment. If it's regular file mmap any alignment is fine, and if
>> it's device-dax only the configured alignment of the device instance
>> is allowed. So, I don't think this should be a configurable option,
>> just read it from the device instance and you're done.
>
> Agreed.
>
> BTW, there's no generic interface to ask the kernel what's the
> required mmap() alignment for a file?

It's only this /dev/dax device-file case where the alignment is
strictly enforced. The required alignment is advertised in sysfs, for
example:

# ls -l /dev/dax0.0
crw------- 1 root root 253, 0 May 26 06:41 /dev/dax0.0
# cat /sys/dev/char/253\:0/device/align
2097152