Hello Sergey, On Tue, Aug 09, 2022 at 06:20:04PM +0900, Sergey Senozhatsky wrote: > On (22/08/09 18:11), Sergey Senozhatsky wrote: > > > > > /me needs to confirm. > > > > > > > > With that commit reverted, I see no more I/O errors, only oom-killer > > > > messages (which is OK IMO, provided I write 1G of urandom on a machine w/ > > > > 800M of RAM): > > > > > > Hmm... So handle allocation always succeeds in the slow path? (when we > > > try to allocate it second time) > > > > Yeah I can see how handle re-allocation with direct reclaim can make it more > > successful, but in exchange it oom-kills some user-space process, I suppose. > > Is oom-kill really a good alternative though? > > We likely will need to revert e7be8d1dd983 given that it has some > user visible changes. But, honestly, failing zram write vs oom-kill > a user-space is a tough choice. I think oom-kill is an inevitable escape from low memory situation if we don't solve original problem with high memory consumption in the user setup. Reclaim-based zram slow path just delays oom if memory eating root cause is not resolved. I totally agree with you that all patches which have visible user degradations should be reverted, but maybe this is more user setup problem, what do you think? If you make the decision to revert slow path removal patch, I would prefer to review the original patch with unneeded code removal again if you don't mind: https://lore.kernel.org/linux-block/20220422115959.3313-1-avromanov@xxxxxxxxxxxxxx/ -- Thank you, Dmitry
