On Sun, 14 Oct 2007, Erez Zadok wrote: > In message <Pine.LNX.4.64.0710142049000.13119@xxxxxxxxxxxxxxxxxxxxx>, Pekka J Enberg writes: > > > > Look at mm/filemap.c:__filemap_fdatawrite_range(). You shouldn't be > > calling unionfs_writepage() _at all_ if the lower mapping has > > BDI_CAP_NO_WRITEBACK capability set. Perhaps something like the totally > > untested patch below? ... I don't disagree with your unionfs_writepages patch, Pekka, but I think it should be viewed as an optimization (don't waste time trying to write a group of pages when we know that nothing will be done) rather than as essential. Prior to unionfs's own use of AOP_WRITEPAGE_ACTIVATE, there have only been ramdisk and shmem generating it. ramdisk is careful only to return it in the wbc->for_reclaim case: I think (as in the patch I sent out before) shmem now ought to do so too for safety. Back in 2.4 days it was reasonable to assume that ->writepage would only get called from certain places, but things move faster nowadays, and the unionfs example shows others are liable to start ab/using it. I'll send Andrew that patch tomorrow (it's simple enough, but I'd like at least to try to reproduce the page_mapped bug first). > > Pekka, with a small change to your patch (to handle time-based cache > coherency), your patch worked well and passed all my tests. Thanks. > > So now I wonder if we still need the patch to prevent AOP_WRITEPAGE_ACTIVATE > from being returned to userland. I guess we still need it, b/c even with > your patch, generic_writepages() can return AOP_WRITEPAGE_ACTIVATE back to > the VFS and we need to ensure that doesn't "leak" outside the kernel. Can it now? Current git has a patch from Andrew which bears a striking resemblance to that from Pekka, stopping the leak from write_cache_pages. Hugh - 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
