On Wed, Mar 22, 2023 at 6:50 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote: > > Yosry Ahmed <yosryahmed@xxxxxxxxxx> writes: > > > On Sun, Mar 19, 2023 at 7:56 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote: > >> > >> Yosry Ahmed <yosryahmed@xxxxxxxxxx> writes: > >> > >> > On Thu, Mar 16, 2023 at 12:51 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote: > >> >> > >> >> Yosry Ahmed <yosryahmed@xxxxxxxxxx> writes: > >> >> > >> >> > On Sun, Mar 12, 2023 at 7:13 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote: > >> >> >> > >> >> >> Yosry Ahmed <yosryahmed@xxxxxxxxxx> writes: > >> >> >> > >> >> >> <snip> > >> >> >> > > >> >> >> > My current idea is to have one xarray that stores the swap_descs > >> >> >> > (which include swap_entry, swapcache, swap_count, etc), and only for > >> >> >> > rotating disks have an additional xarray that maps swap_entry -> > >> >> >> > swap_desc for cluster readahead, assuming we can eliminate all other > >> >> >> > situations requiring a reverse mapping. > >> >> >> > > >> >> >> > I am not sure how having separate xarrays help? If we have one xarray, > >> >> >> > might as well save the other lookups on put everything in swap_desc. > >> >> >> > In fact, this should improve the locking today as swapcache / > >> >> >> > swap_count operations can be lockless or very lightly contended. > >> >> >> > >> >> >> The condition of the proposal is "reverse mapping cannot be avoided for > >> >> >> enough situation". So, if reverse mapping (or cluster readahead) can be > >> >> >> avoided for enough situations, I think your proposal is good. Otherwise, > >> >> >> I propose to use 2 xarrays. You don't need another reverse mapping > >> >> >> xarray, because you just need to read the next several swap_entry into > >> >> >> the swap cache for cluster readahead. swap_desc isn't needed for > >> >> >> cluster readahead. > >> >> > > >> >> > swap_desc would be needed for cluster readahead in my original > >> >> > proposal as the swap cache lives in swap_descs. Based on the current > >> >> > implementation, we would need a reverse mapping (swap entry -> > >> >> > swap_desc) in 3 situations: > >> >> > > >> >> > 1) __try_to_reclaim_swap(): when trying to find an empty swap slot and > >> >> > failing, we fallback to trying to find swap entries that only have a > >> >> > page in the swap cache (no references in page tables or page cache) > >> >> > and free them. This would require a reverse mapping. > >> >> > > >> >> > 2) swapoff: we need to swap in all entries in a swapfile, so we need > >> >> > to get all swap_descs associated with that swapfile. > >> >> > > >> >> > 3) swap cluster readahead. > >> >> > > >> >> > For (1), I think we can drop the dependency of a reverse mapping if we > >> >> > free swap entries once we swap a page in and add it to the swap cache, > >> >> > even if the swap count does not drop to 0. > >> >> > >> >> Now, we will not drop the swap cache even if the swap count becomes 0 if > >> >> swap space utility < 50%. Per my understanding, this avoid swap page > >> >> writing for read accesses. So I don't think we can change this directly > >> >> without necessary discussion firstly. > >> > > >> > > >> > Right. I am not sure I understand why we do this today, is it to save > >> > the overhead of allocating a new swap entry if the page is swapped out > >> > again soon? I am not sure I understand this statement "this avoid swap > >> > page > >> > writing for read accesses". > >> > > >> >> > >> >> > >> >> > For (2), instead of scanning page tables and shmem page cache to find > >> >> > swapped out pages for the swapfile, we can scan all swap_descs > >> >> > instead, we should be more efficient. This is one of the proposal's > >> >> > potential advantages. > >> >> > >> >> Good. > >> >> > >> >> > (3) is the one that would still need a reverse mapping with the > >> >> > current proposal. Today we use swap cluster readahead for anon pages > >> >> > if we have a spinning disk or vma readahead is disabled. For shmem, we > >> >> > always use cluster readahead. If we can limit cluster readahead to > >> >> > only rotating disks, then the reverse mapping can only be maintained > >> >> > for swapfiles on rotating disks. Otherwise, we will need to maintain a > >> >> > reverse mapping for all swapfiles. > >> >> > >> >> For shmem, I think that it should be good to readahead based on shmem > >> >> file offset instead of swap device offset. > >> >> > >> >> It's possible that some pages in the readahead window are from HDD while > >> >> some other pages aren't. So it's a little hard to enable cluster read > >> >> for HDD only. Anyway, it's not common to use HDD for swap now. > >> >> > >> >> >> > >> >> >> > If the point is to store the swap_desc directly inside the xarray to > >> >> >> > save 8 bytes, I am concerned that having multiple xarrays for > >> >> >> > swapcache, swap_count, etc will use more than that. > >> >> >> > >> >> >> The idea is to save the memory used by reverse mapping xarray. > >> >> > > >> >> > I see. > >> >> > > >> >> >> > >> >> >> >> >> > >> >> >> >> >> > Keep in mind that the current overhead is 1 byte O(max swap pages) not > >> >> >> >> >> > O(swapped). Also, 1 byte is assuming we do not use the swap > >> >> >> >> >> > continuation pages. If we do, it may end up being more. We also > >> >> >> >> >> > allocate continuation in full 4k pages, so even if one swap_map > >> >> >> >> >> > element in a page requires continuation, we will allocate an entire > >> >> >> >> >> > page. What I am trying to say is that to get an actual comparison you > >> >> >> >> >> > need to also factor in the swap utilization and the rate of usage of > >> >> >> >> >> > swap continuation. I don't know how to come up with a formula for this > >> >> >> >> >> > tbh. > >> >> >> >> >> > > >> >> >> >> >> > Also, like Johannes said, the worst case overhead (32 bytes if you > >> >> >> >> >> > count the reverse mapping) is 0.8% of swapped memory, aka 8M for every > >> >> >> >> >> > 1G swapped. It doesn't sound *very* bad. I understand that it is pure > >> >> >> >> >> > overhead for people not using zswap, but it is not very awful. > >> >> >> >> >> > > >> >> >> >> >> >> > >> >> >> >> >> >> It seems what you really need is one bit of information to indicate > >> >> >> >> >> >> this page is backed by zswap. Then you can have a seperate pointer > >> >> >> >> >> >> for the zswap entry. > >> >> >> >> >> > > >> >> >> >> >> > If you use one bit in swp_entry_t (or one of the available swap types) > >> >> >> >> >> > to indicate whether the page is backed with a swapfile or zswap it > >> >> >> >> >> > doesn't really work. We lose the indirection layer. How do we move the > >> >> >> >> >> > page from zswap to swapfile? We need to go update the page tables and > >> >> >> >> >> > the shmem page cache, similar to swapoff. > >> >> >> >> >> > > >> >> >> >> >> > Instead, if we store a key else in swp_entry_t and use this to lookup > >> >> >> >> >> > the swp_entry_t or zswap_entry pointer then that's essentially what > >> >> >> >> >> > the swap_desc does. It just goes the extra mile of unifying the > >> >> >> >> >> > swapcache as well and storing it directly in the swap_desc instead of > >> >> >> >> >> > storing it in another lookup structure. > >> >> >> >> >> > >> >> >> >> >> If we choose to make sizeof(struct swap_desc) == 8, that is, store only > >> >> >> >> >> swap_entry in swap_desc. The added indirection appears to be another > >> >> >> >> >> level of page table with 1 entry. Then, we may use the similar method > >> >> >> >> >> as supporting system with 2 level and 3 level page tables, like the code > >> >> >> >> >> in include/asm-generic/pgtable-nopmd.h. But I haven't thought about > >> >> >> >> >> this deeply. > >> >> >> >> > > >> >> >> >> > Can you expand further on this idea? I am not sure I fully understand. > >> >> >> >> > >> >> >> >> OK. The goal is to avoid the overhead if indirection isn't enabled via > >> >> >> >> kconfig. > >> >> >> >> > >> >> >> >> If indirection isn't enabled, store swap_entry in PTE directly. > >> >> >> >> Otherwise, store index of swap_desc in PTE. Different functions (e.g., > >> >> >> >> to get/set swap_entry in PTE) are implemented based on kconfig. > >> >> >> > > >> >> >> > > >> >> >> > I thought about this, the problem is that we will have multiple > >> >> >> > implementations of multiple things. For example, swap_count without > >> >> >> > the indirection layer lives in the swap_map (with continuation logic). > >> >> >> > With the indirection layer, it lives in the swap_desc (or somewhere > >> >> >> > else). Same for the swapcache. Even if we keep the swapcache in an > >> >> >> > xarray and not inside swap_desc, it would be indexed by swap_entry if > >> >> >> > the indirection is disabled, and by swap_desc (or similar) if the > >> >> >> > indirection is enabled. I think maintaining separate implementations > >> >> >> > for when the indirection is enabled/disabled would be adding too much > >> >> >> > complexity. > >> >> >> > > >> >> >> > WDYT? > >> >> >> > >> >> >> If we go this way, swap cache and swap_count will always be indexed by > >> >> >> swap_entry. swap_desc just provides a indirection to make it possible > >> >> >> to move between swap devices. > >> >> >> > >> >> >> Why must we index swap cache and swap_count by swap_desc if indirection > >> >> >> is enabled? Yes, we can save one xarray indexing if we do so, but I > >> >> >> don't think the overhead of one xarray indexing is a showstopper. > >> >> >> > >> >> >> I think this can be one intermediate step towards your final target. > >> >> >> The changes to current implementation can be smaller. > >> >> > > >> >> > IIUC, the idea is to have two xarrays: > >> >> > (a) xarray that stores a pointer to a struct containing swap_count and > >> >> > swap cache. > >> >> > (b) xarray that stores the underlying swap entry or zswap entry. > >> >> > > >> >> > When indirection is disabled: > >> >> > page tables & page cache have swap entry directly like today, xarray > >> >> > (a) is indexed by swap entry, xarray (b) does not exist. No reverse > >> >> > mapping needed. > >> >> > > >> >> > In this case we have an extra overhead of 12-16 bytes (the struct > >> >> > containing swap_count and swap cache) vs. 24 bytes of the swap_desc. > >> >> > > >> >> > When indirection is enabled: > >> >> > page tables & page cache have a swap id (or swap_desc index), xarray > >> >> > (a) is indexed by swap id, > >> >> > >> >> xarray (a) is indexed by swap entry. > >> > > >> > > >> > How so? With the indirection enabled, the page tables & page cache > >> > have the swap id (or swap_desc index), which can point to a swap entry > >> > or a zswap entry -- which can change when the page is moved between > >> > zswap & swapfiles. How is xarray (a) indexed by the swap entry in this > >> > case? Shouldn't be indexed by the abstract swap id so that the > >> > writeback from zswap is transparent? > >> > >> In my mind, > >> > >> - swap core will define a abstract interface to swap implementations > >> (zswap, swap device/file, maybe more in the future), like VFS. > >> > >> - zswap will be a special swap implementation (compressing instead of > >> writing to disk). > >> > >> - swap core will manage the indirection layer and swap cache. > >> > >> - swap core can move swap pages between swap implementations (e.g., from > >> zswap to a swap device, or from one swap device to another swap > >> device) with the help of the indirection layer. > >> > >> In this design, the writeback from zswap becomes moving swapped pages > >> from zswap to a swap device. > > > > > > All the above matches my understanding of this proposal. swap_desc is > > the proposed indirection layer, and the swap implementations are zswap > > & swap devices. For now, we only have 2 static swap implementations > > (zswap->swapfile). In the future, we can make this more dynamic as the > > need arises. > > Great to align with you on this. Great! > > >> > >> If my understanding were correct, your suggestion is kind of moving > >> zswap logic to the swap core? And zswap will be always at a higher > >> layer on top of swap device/file? > > > > > > We do not want to move the zswap logic into the swap core, we want to > > make the swap core independent of the swap implementation, and zswap > > is just one possible implementation. > > Good! > > I found that you put zswap related data structure inside struct > swap_desc directly. I think that we should avoid that as much as > possible. That was just an initial draft to show what it would be like, I do not intend to hardcode zswap-specific items in swap_desc. > > >> > >> >> > >> >> > >> >> > xarray (b) is indexed by swap id as well > >> >> > and contain swap entry or zswap entry. Reverse mapping might be > >> >> > needed. > >> >> > >> >> Reverse mapping isn't needed. > >> > > >> > > >> > It would be needed if xarray (a) is indexed by the swap id. I am not > >> > sure I understand how it can be indexed by the swap entry if the > >> > indirection is enabled. > >> > > >> >> > >> >> > >> >> > In this case we have an extra overhead of 12-16 bytes + 8 bytes for > >> >> > xarray (b) entry + memory overhead from 2nd xarray + reverse mapping > >> >> > where needed. > >> >> > > >> >> > There is also the extra cpu overhead for an extra lookup in certain paths. > >> >> > > >> >> > Is my analysis correct? If yes, I agree that the original proposal is > >> >> > good if the reverse mapping can be avoided in enough situations, and > >> >> > that we should consider such alternatives otherwise. As I mentioned > >> >> > above, I think it comes down to whether we can completely restrict > >> >> > cluster readahead to rotating disks or not -- in which case we need to > >> >> > decide what to do for shmem and for anon when vma readahead is > >> >> > disabled. > >> >> > >> >> We can even have a minimal indirection implementation. Where, swap > >> >> cache and swap_map[] are kept as they ware before, just one xarray is > >> >> added. The xarray is indexed by swap id (or swap_desc index) to store > >> >> the corresponding swap entry. > >> >> > >> >> When indirection is disabled, no extra overhead. > >> >> > >> >> When indirection is enabled, the extra overhead is just 8 bytes per > >> >> swapped page. > >> >> > >> >> The basic migration support can be build on top of this. > >> >> > >> >> I think that this could be a baseline for indirection support. Then > >> >> further optimization can be built on top of it step by step with > >> >> supporting data. > >> > > >> > > >> > I am not sure how this works with zswap. Currently swap_map[] > >> > implementation is specific for swapfiles, it does not work for zswap > >> > unless we implement separate swap counting logic for zswap & > >> > swapfiles. Same for the swapcache, it currently supports being indexed > >> > by a swap entry, it would need to support being indexed by a swap id, > >> > or have a separate swap cache for zswap. Having separate > >> > implementation would add complexity, and we would need to perform > >> > handoffs of the swap count/cache when a page is moved from zswap to a > >> > swapfile. > >> > >> We can allocate a swap entry for each swapped page in zswap. > > > > > > This is exactly what the current implementation does and what we want > > to move away from. The current implementation uses zswap as an > > in-memory compressed cache on top of an actual swap device, and each > > swapped page in zswap has a swap entry allocated. With this > > implementation, zswap cannot be used without a swap device. > > I totally agree that we should avoid to use an actual swap device under > zswap. And, as an swap implementation, zswap can manage the swap entry > inside zswap without an underlying actual swap device. For example, > when we swap a page to zswap (actually compress), we can allocate a > (virtual) swap entry in the zswap. I understand that there's overhead > to manage the swap entry in zswap. We can consider how to reduce the > overhead. I see. So we can (for example) use one of the swap types for zswap, and then have zswap code handle this entry according to its implementation. We can then have an xarray that maps swap ID -> swap entry, and this swap entry is used to index the swap cache and such. When a swapped page is moved between backends we update the swap ID -> swap entry xarray. This is one possible implementation that I thought of (very briefly tbh), but it does have its problems: For zswap: - Managing swap entries inside zswap unnecessarily. - We need to maintain a swap entry -> zswap entry mapping in zswap -- similar to the current rbtree, which is something that we can get rid of with the initial proposal if we embed the zswap_entry pointer directly in the swap_desc (it can be encoded to avoid breaking the abstraction). For mm/swap in general: - When we allocate a swap entry today, we store it in folio->private (or page->private), which is used by the unmapping code to be placed in the page tables or shmem page cache. With this implementation, we need to store the swap ID in page->private instead, which means that every time we need to access the swap cache during reclaim/swapout we need to lookup the swap entry first. - On the fault path, we need two lookups instead of one (swap ID -> swap entry, swap entry -> swap cache), not sure how this affects fault latency. - Each swap backend will have its own separate implementation of swap counting, which is hard to maintain and very error-prone since the logic is backend-agnostic. - Handing over a page from one swap backend to another includes handing over swap cache entries and swap counts, which I imagine will involve considerable synchronization. Do you have any thoughts on this? > > Best Regards, > Huang, Ying > > >> > >> > >> >> > >> >> > >> >> >> > >> >> >> >> >> >> > >> >> >> >> >> >> Depending on how much you are going to reuse the swap cache, you might > >> >> >> >> >> >> need to have something like a swap_info_struct to keep the locks happy. > >> >> >> >> >> > > >> >> >> >> >> > My current intention is to reimplement the swapcache completely as a > >> >> >> >> >> > pointer in struct swap_desc. This would eliminate this need and a lot > >> >> >> >> >> > of the locking we do today if I get things right. > >> >> >> >> >> > > >> >> >> >> >> >> > >> >> >> >> >> >> > Another potential concern is readahead. With this design, we have no > >> >> >> >> >> >> > >> >> >> >> >> >> Readahead is for spinning disk :-) Even a normal swap file with an SSD can > >> >> >> >> >> >> use some modernization. > >> >> >> >> >> > > >> >> >> >> >> > Yeah, I initially thought we would only need the swp_entry_t -> > >> >> >> >> >> > swap_desc reverse mapping for readahead, and that we can only store > >> >> >> >> >> > that for spinning disks, but I was wrong. We need for other things as > >> >> >> >> >> > well today: swapoff, when trying to find an empty swap slot and we > >> >> >> >> >> > start trying to free swap slots used only by the swapcache. However, I > >> >> >> >> >> > think both of these cases can be fixed (I can share more details if > >> >> >> >> >> > you want). If everything goes well we should only need to maintain the > >> >> >> >> >> > reverse mapping (extra overhead above 24 bytes) for swap files on > >> >> >> >> >> > spinning disks for readahead. > >> >> >> >> >> > > >> >> >> >> >> >> > >> >> >> >> >> >> Looking forward to your discussion. > >> >> >> >> > >> >> >> >> Per my understanding, the indirection is to make it easy to move > >> >> >> >> (swapped) pages among swap devices based on hot/cold. This is similar > >> >> >> >> as the target of memory tiering. It appears that we can extend the > >> >> >> >> memory tiering (mm/memory-tiers.c) framework to cover swap devices too? > >> >> >> >> Is it possible for zswap to be faster than some slow memory media? > >> >> >> > > >> >> >> > > >> >> >> > Agree with Chris that this may require a much larger overhaul. A slow > >> >> >> > memory tier is still addressable memory, swap/zswap requires a page > >> >> >> > fault to read the pages. I think (at least for now) there is a > >> >> >> > fundamental difference. We want reclaim to eventually treat slow > >> >> >> > memory & swap as just different tiers to place cold memory in with > >> >> >> > different characteristics, but otherwise I think the swapping > >> >> >> > implementation itself is very different. Am I missing something? > >> >> >> > >> >> >> Is it possible that zswap is faster than a really slow memory > >> >> >> addressable device backed by NAND? TBH, I don't have the answer. > >> >> > > >> >> > I am not sure either. > >> >> > > >> >> >> > >> >> >> Anyway, do you need a way to describe the tiers of the swap devices? > >> >> >> So, you can move the cold pages among the swap devices based on that? > >> >> > > >> >> > For now I think the "tiers" in this proposal are just zswap and normal > >> >> > swapfiles. We can later extend it to support more explicit tiering. > >> >> > >> >> IIUC, in original zswap implementation, there's 1:1 relationship between > >> >> zswap and normal swapfile. But now, you make demoting among swap > >> >> devices more general. Then we need some general way to specify which > >> >> swap devices are fast and which are slow, and the demoting relationship > >> >> among them. It can be memory tiers or something else, but we need one. > >> > > >> > > >> > I think for this proposal, there are only 2 hardcoded tiers. Zswap is > >> > fast, swapfile is slow. In the future, we can support more dynamic > >> > tiering if the need arises. > >> > >> We can start from a simple implementation. And I think that it's better > >> to consider the general design too. Try not to make it impossible now. > > > > > > Right. I am proposing we come up with an abstract generic interface > > for swap implementations, and have 2 implementations statically > > defined (swapfiles and zswap). If the need arises, we can make swap > > implementations more dynamic in the future. > > > >> > >> > >> Best Regards, > >> Huang, Ying >
