On Wed, Sep 04, 2024 at 11:33:43PM +0000, Yosry Ahmed wrote: > The z3fold compressed pages allocator is rarely used, most users use > zsmalloc. The only disadvantage of zsmalloc in comparison is the > dependency on MMU, and zbud is a more common option for !MMU as it was > the default zswap allocator for a long time. > > Historically, zsmalloc had worse latency than zbud and z3fold but > offered better memory savings. This is no longer the case as shown by > a simple recent analysis [1]. That analysis showed that z3fold does not > have any advantage over zsmalloc or zbud considering both performance > and memory usage. In a kernel build test on tmpfs in a limited cgroup, > z3fold took 3% more time and used 1.8% more memory. The latency of > zswap_load() was 7% higher, and that of zswap_store() was 10% higher. > Zsmalloc is better in all metrics. > > Moreover, z3fold apparently has latent bugs, which was made noticeable > by a recent soft lockup bug report with z3fold [2]. Switching to > zsmalloc not only fixed the problem, but also reduced the swap usage > from 6~8G to 1~2G. Other users have also reported being bitten by > mistakenly enabling z3fold. > > Other than hurting users, z3fold is repeatedly causing wasted > engineering effort. Apart from investigating the above bug, it came up > in multiple development discussions (e.g. [3]) as something we need to > handle, when there aren't any legit users (at least not intentionally). > > The natural course of action is to deprecate z3fold, and remove in a few > cycles if no objections are raised from active users. Next on the list > should be zbud, as it offers marginal latency gains at the cost of huge > memory waste when compared to zsmalloc. That one will need to wait until > zsmalloc does not depend on MMU. > > Rename the user-visible config option from CONFIG_Z3FOLD to > CONFIG_Z3FOLD_DEPRECATED so that users with CONFIG_Z3FOLD=y get a new > prompt with explanation during make oldconfig. Also, remove > CONFIG_Z3FOLD=y from defconfigs. > > [1]https://lore.kernel.org/lkml/CAJD7tkbRF6od-2x_L8-A1QL3=2Ww13sCj4S3i4bNndqF+3+_Vg@xxxxxxxxxxxxxx/ > [2]https://lore.kernel.org/lkml/EF0ABD3E-A239-4111-A8AB-5C442E759CF3@xxxxxxxxx/ > [3]https://lore.kernel.org/lkml/CAJD7tkbnmeVugfunffSovJf9FAgy9rhBVt_tx=nxUveLUfqVsA@xxxxxxxxxxxxxx/ > > Acked-by: Chris Down <chris@xxxxxxxxxxxxxx> > Acked-by: Nhat Pham <nphamcs@xxxxxxxxx> > Signed-off-by: Yosry Ahmed <yosryahmed@xxxxxxxxxx> zsmalloc's CONFIG_MMU requirement was a concern in the past, but z3fold appears so bitrotted at this point that it's simply not a usable option anymore. > I think it should actually be fine to remove z3fold without deprecating > it first, but I am doing the due diligence. Yeah, you never know for sure if users exist. Deprecating it for a few cycles is the safer option. Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>
