On Fri, Jun 04, 2010 at 08:13:14AM -0700, Dan Magenheimer wrote: > > 1) > > You mentiond PFRA in you description and I understood cleancache has > > a cold clean page which is evicted by reclaimer. > > But __remove_from_page_cache can be called by other call sites. > > > > For example, shmem_write page calls it for moving the page from page > > cache > > to swap cache. Although there isn't the page in page cache, it is in > > swap cache. > > So next read/write of shmem until swapout happens can be read/write in > > swap cache. > > > > I didn't looked into whole of callsites. But please review again them. > > I think the "if (PageUptodate(page))" eliminates all the cases > where bad things can happen. I missed it. my fisrt concern has gone. :) > > Note that there may be cases where some unnecessary puts/flushes > occur. The focus of the patch is on correctness first; it may > be possible to increase performance (marginally) in the future by > reducing unnecessary cases. I think it wouldn't be marginally. It depends on implementation of backend. I think frontend would be better to notify to backend in only exact place. As your descrption, we can call it in shrink_page_list with some check or change __remove_mapping which adding a argument to tell "this is calling of reclaim path". > > > 3) Please consider system memory pressure. > > And I hope Nitin consider this, too. > > This is definitely very important but remember that cleancache > provides a great deal of flexibility: Any page in cleancache > can be thrown away at any time as every page is clean! It > can even accept a page and throw it away immediately. Clearly > the backend needs to do this intelligently so this will > take some policy work. I admit design goal of cleancache is to give a greate deal of flexibility. But I think system memory pressure(ie, direct reclaim and even OOM) is exceptional. Whenever we implement various backend, every backend(non-virtual environemnt)have to implement policy which deal with system memory pressure emergency to prevent system hang, I think. And backend might need some hack to know the situation. It's horrible. So I hope frontend gives little information to backend, at least. If some backend don't need it, it can just ignore. But if some backend need it, it can be a big deal. :) > > Since I saw you sent a separate response to Nitin, I'll > let him answer for his in-kernel page cache compression > work. The solution to the similar problem for Xen is > described in the tmem internals document that I think > I pointed to earlier here: > http://oss.oracle.com/projects/tmem/documentation/internals/ I will read it when I have a time. Thanks for quick reply but I can't. It's time to sleep and weekend. See you soon and have a nice weekend. > > Thanks, > Dan > -- Kind regards, Minchan Kim -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
