2007/8/6, Shaohua Li <shaoh.li@xxxxxxxxx>: > Hi, > I'm trying to swap out kvm guest pages. The idea is to free some pages > when memory pressure is high. kvm has special things to handle like > shadow page tables. Before guest page is released, we need free some > data, that is guest page has 'private' data, so we can't directly make > the guest page swapout with Linux swapout mechanism. I'd like write > guest pages to a file or swap. kvm guest pages are in its address > space and added into lru list like normal page (the address space is > very like a shmem file's, kvm has a memory based file system). When > vmscan decided to free one guest page, kvm guest pages's > aops.writepage will free the private data and then write it out to > block device. The problem is how to write guest pages. I thought we > have some choices: > > 1. swap it to swapfile. Like shmem does, using move_to_swap_cache to > move guest page from its address space to swap address space, and > finally it's written to swapfile. This method works well, but as kvm > is a module, I must export some swap relelated APIs, which Christoph > Hellwig dislike. > > 2. write it to a file. Just like the stack fs does, in kvm address > space's .writepage, let low fs's aops write the page. This involves > allocating a new page for low fs file and copy kvm page to the new > page. As this (doing swap) is done when memory is tight, allocating > new page isn't good. The copy isn't good too. > > 3.write it to a file. Using bmap to get file's block info and using > low level sumbit_bio for read/write. This is like what swapfile does, > but we do it by ourselves, so don't need use swap symbols. > > Do you have any suggestion which method is good, or if you have better > choice I could try? Can anybody share some hints? I really apprecate any comments. Thanks, Shaohua - 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
