On Fri 18-09-20 08:32:13, Vijay Balakrishna wrote: > > > On 9/17/2020 10:45 PM, Michal Hocko wrote: > > On Thu 17-09-20 11:03:56, Vijay Balakrishna wrote: > > [...] > > > > > The auto tuned value is incorrect post hotplug memory operation, in our use > > > > > case memoy hot add occurs very early during boot. > > > > Define incorrect. What are the actual values? Have you tried to increase > > > > the value manually after the hotplug? > > > > > > In our case SoC with 8GB memory, system tuned min_free_kbytes > > > - first to 22528 > > > - we perform memory hot add very early in boot > > > > What was the original and after-the-hotplug size of memory and layout? > > I suspect that all the hotplugged memory is in Movable zone, right? > > Yes, added ~1.92GB as Movable type, booting with 6GB at start. > > > > > > - now min_free_kbytes is 8703 > > > > > > Before looking at code, first I manually restored min_free_kbytes soon after > > > boot, reran stress and didn't notice symptoms I mentioned in change log. > > > > This is really surprising and I strongly suspect that an earlier reclaim > > just changed the timing enough so that workload has spread the memory > > prpessure over a longer time and that might have been enough to recycle > > some of the unreclaimable memory due to its natural life time. But this > > is a pure speculation. Much more data would be needed to analyze this. > > > > In any case your stress test is oveprovisioning your Normal zone and > > increased min_free_kbytes just papers over the sizing problem. > > > > It is a synthetic workload, likely not sized I need to check. I feel having > higher min_free_kbytes made GFP_ATOMIC allocations not to fail. Yes a higher min_free_kbytes will help GFP_ATOMIC. But only to some degree. But nobody should depend on an atomic allocation for correctness. It is just way too easy to fail under a higher memory pressure. > I have seen > NETDEV WATCHDOG timeout with stacktrace trying to allocate memory, looping > in net rx receive path. You should talk to net folks. -- Michal Hocko SUSE Labs
