On Mon, Oct 28, 2024 at 4:03 PM Barry Song <21cnbao@xxxxxxxxx> wrote: > > On Tue, Oct 29, 2024 at 6:54 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > > > On Mon, Oct 28, 2024 at 3:52 PM Barry Song <21cnbao@xxxxxxxxx> wrote: > > > > > > On Tue, Oct 29, 2024 at 6:33 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote: > > > > > > > > [..] > > > > > > > By the way, I recently had an idea: if we can conduct the zeromap check > > > > > > > earlier - for example - before allocating swap slots and pageout(), could > > > > > > > we completely eliminate swap slot occupation and allocation/release > > > > > > > for zeromap data? For example, we could use a special swap > > > > > > > entry value in the PTE to indicate zero content and directly fill it with > > > > > > > zeros when swapping back. We've observed that swap slot allocation and > > > > > > > freeing can consume a lot of CPU and slow down functions like > > > > > > > zap_pte_range and swap-in. If we can entirely skip these steps, it > > > > > > > could improve performance. However, I'm uncertain about the benefits we > > > > > > > would gain if we only have 1-2% zeromap data. > > > > > > > > > > > > If I remember correctly this was one of the ideas floated around in the > > > > > > initial version of the zeromap series, but it was evaluated as a lot more > > > > > > complicated to do than what the current zeromap code looks like. But I > > > > > > think its definitely worth looking into! > > > > > > > > Yup, I did suggest this on the first version: > > > > https://lore.kernel.org/linux-mm/CAJD7tkYcTV_GOZV3qR6uxgFEvYXw1rP-h7WQjDnsdwM=g9cpAw@xxxxxxxxxxxxxx/ > > > > > > > > , and Usama took a stab at implementing it in the second version: > > > > https://lore.kernel.org/linux-mm/20240604105950.1134192-1-usamaarif642@xxxxxxxxx/ > > > > > > > > David and Shakeel pointed out a few problems. I think they are > > > > fixable, but the complexity/benefit tradeoff was getting unclear at > > > > that point. > > > > > > > > If we can make it work without too much complexity, that would be > > > > great of course. > > > > > > > > > > > > > > Sorry for the noise. I didn't review the initial discussion. But my feeling > > > > > is that it might be valuable considering the report from Zhiguo: > > > > > > > > > > https://lore.kernel.org/linux-mm/20240805153639.1057-1-justinjiang@xxxxxxxx/ > > > > > > > > > > In fact, our recent benchmark also indicates that swap free could account > > > > > for a significant portion in do_swap_page(). > > > > > > > > As Shakeel mentioned in a reply to Usama's patch mentioned above, we > > > > would need to check the contents of the page after it's unmapped. So > > > > likely we need to allocate a swap slot, walk the rmap and unmap, check > > > > contents, walk the rmap again and update the PTEs, free the swap slot. > > > > > > > > > > So the issue is that we can't check the content before allocating slots and > > > unmapping during reclamation? If we find the content is zero, can we skip > > > all slot operations and go directly to rmap/unmap by using a special PTE? > > > > We need to unmap first before checking the content, otherwise the > > content can change right after we check it. > > Well, do we have a way to terminate the unmap if we find pte_dirty and ensure > the folio is still mapped after try_to_unmap_one()? Then we could > activate it again > after try_to_unmap. > > It might just be noise. Let me take some more time to think about it. :-) FWIW, the swap abstraction layer Yosry proposed last year (and I'm working on right now) will allow you to store these zeromapped swap entries without requiring any swap slots allocated on the swapfile. It's basically the same thing as swap/zswap decoupling. Not stopping you guys from optimizing it, since all I have right now is a (most certainly buggy) prototype + there might be benefits if we can get around the swap subsystem altogether for these zero mapped entries. Just letting you know there's a backup plan :)
