[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] Why frontswap and cleancache make copies in tmem?

On 02/17/2014 07:40 PM, George Dunlap wrote:
> On Mon, Feb 17, 2014 at 11:38 AM, George Dunlap
> <George.Dunlap@xxxxxxxxxxxxx> wrote:
>> On Sun, Feb 16, 2014 at 10:14 PM, Jinchun Kim <cienlux@xxxxxxxxx> wrote:
>>> Hi, All.
>>> While I was digging tmem and its source code, I found suspicious things
>>> about frontswap and cleancache.
>>> When the guest OS wants to evict either a dirty or clean page, frontswap and
>>> cleancache will store it in tmem.
>>> The linux kernel document tells that
>>> [Documentation/vm/frontswap.txt]
>>> A "store" will copy the page to transcendent memory ....
>>> A "load" will copy the page, if found, from transcendent memory into kernel
>>> memory ...
>>> [Documentation/vm/cleancache.txt]
>>> A "put_page" will copy a page (presumably about-to-be-evicted) page into
>>> cleancache ...
>>> A "get_page" will copy the page, if found, from cleancache into kernel
>>> memory ...

Jinchun, I'm glad you are interested with tmem.

>>> My colleagues and I think copying the page is not necessary (especially for
>>> cleancache) because both kernel memory and tmem have same data. Why don't we
>>> just change the pointer to the page and let it belong to tmem? We were

No, kernel memory won't have the data any more.
For cleancache, the data exist only in tmem and the disk storage.

It's impossible to 'change the pointer to the page and let it belong to
tmem', because the page need to be reclaimed by guest os.
One possible way is to allocate a new page and return it to guest os, so
that we can let the original page belong to tmem instead of copy data.
But by this way, there is also some map/unmap cost and will make tmem
more complicated.

Welcome any ideas about making tmem better.

>>> wondering if there is any specific reason not to copy the pages to tmem.
>> Konrad / Boris?  (Sorry, I can't remember the other person who's been
>> sending tmem patches recently.)
> Oh, it was Bob Liu. :-)

George, thanks for your cc.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.