On Sun, Aug 4, 2024 at 4:03 AM Kairui Song <ryncsn@xxxxxxxxx> wrote: > > On Sun, Aug 4, 2024 at 1:09 AM Yu Zhao <yuzhao@xxxxxxxxxx> wrote: > > On Sat, Aug 3, 2024 at 2:31 AM Ge Yang <yangge1116@xxxxxxx> wrote: > > > 在 2024/8/3 4:18, Chris Li 写道: > > > > On Thu, Aug 1, 2024 at 6:56 PM Ge Yang <yangge1116@xxxxxxx> wrote: > > > >> > > > >> > > > >> > > > >>>> I can't reproduce this problem, using tmpfs to compile linux. > > > >>>> Seems you limit the memory size used to compile linux, which leads to > > > >>>> OOM. May I ask why the memory size is limited to 481280kB? Do I also > > > >>>> need to limit the memory size to 481280kB to test? > > > >>> > > > >>> Yes, you need to limit the cgroup memory size to force the swap > > > >>> action. I am using memory.max = 470M. > > > >>> > > > >>> I believe other values e.g. 800M can trigger it as well. The reason to > > > >>> limit the memory to cause the swap action. > > > >>> The goal is to intentionally overwhelm the memory load and let the > > > >>> swap system do its job. The 470M is chosen to cause a lot of swap > > > >>> action but not too high to cause OOM kills in normal kernels. > > > >>> In another word, high enough swap pressure but not too high to bust > > > >>> into OOM kill. e.g. I verify that, with your patch reverted, the > > > >>> mm-stable kernel can sustain this level of swap pressure (470M) > > > >>> without OOM kill. > > > >>> > > > >>> I borrowed the 470M magic value from Hugh and verified it works with > > > >>> my test system. Huge has a similar swab test up which is more > > > >>> complicated than mine. It is the inspiration of my swap stress test > > > >>> setup. > > > >>> > > > >>> FYI, I am using "make -j32" on a machine with 12 cores (24 > > > >>> hyperthreading). My typical swap usage is about 3-5G. I set my > > > >>> swapfile size to about 20G. > > > >>> I am using zram or ssd as the swap backend. Hope that helps you > > > >>> reproduce the problem. > > > >>> > > > >> Hi Chris, > > > >> > > > >> I try to construct the experiment according to your suggestions above. > > > > > > > > Hi Ge, > > > > > > > > Sorry to hear that you were not able to reproduce it. > > > > > > > >> High swap pressure can be triggered, but OOM can't be reproduced. The > > > >> specific steps are as follows: > > > >> root@ubuntu-server-2204:/home/yangge# cp workspace/linux/ /dev/shm/ -rf > > > > > > > > I use a slightly different way to setup the tmpfs: > > > > > > > > Here is section of my script: > > > > > > > > if ! [ -d $tmpdir ]; then > > > > sudo mkdir -p $tmpdir > > > > sudo mount -t tmpfs -o size=100% nodev $tmpdir > > > > fi > > > > > > > > sudo mkdir -p $cgroup > > > > sudo sh -c "echo $mem > $cgroup/memory.max" || echo setup > > > > memory.max error > > > > sudo sh -c "echo 1 > $cgroup/memory.oom.group" || echo setup > > > > oom.group error > > > > > > > > Per run: > > > > > > > > # $workdir is under $tmpdir > > > > sudo rm -rf $workdir > > > > mkdir -p $workdir > > > > cd $workdir > > > > echo "Extracting linux tree" > > > > XZ_OPT='-T0 -9 –memory=75%' tar xJf $linux_src || die "xz > > > > extract failed" > > > > > > > > sudo sh -c "echo $BASHPID > $cgroup/cgroup.procs" > > > > echo "Cleaning linux tree, setup defconfig" > > > > cd $workdir/linux > > > > make -j$NR_TASK clean > > > > make defconfig > /dev/null > > > > echo Kernel compile run $i > > > > /usr/bin/time -a -o $log make --silent -j$NR_TASK || die "make failed" > > > > > > > > > > > Thanks. > > > > > > >> root@ubuntu-server-2204:/home/yangge# sync > > > >> root@ubuntu-server-2204:/home/yangge# echo 3 > /proc/sys/vm/drop_caches > > > >> root@ubuntu-server-2204:/home/yangge# cd /sys/fs/cgroup/ > > > >> root@ubuntu-server-2204:/sys/fs/cgroup/# mkdir kernel-build > > > >> root@ubuntu-server-2204:/sys/fs/cgroup/# cd kernel-build > > > >> root@ubuntu-server-2204:/sys/fs/cgroup/kernel-build# echo 470M > memory.max > > > >> root@ubuntu-server-2204:/sys/fs/cgroup/kernel-build# echo $$ > cgroup.procs > > > >> root@ubuntu-server-2204:/sys/fs/cgroup/kernel-build# cd /dev/shm/linux/ > > > >> root@ubuntu-server-2204:/dev/shm/linux# make clean && make -j24 > > > > > > > > I am using make -j 32. > > > > > > > > Your step should work. > > > > > > > > Did you enable MGLRU in your .config file? Mine did. I attached my > > > > config file here. > > > > > > > > > > The above test didn't enable MGLRU. > > > > > > When MGLRU is enabled, I can reproduce OOM very soon. The cause of > > > triggering OOM is being analyzed. > > Hi Ge, > > Just in case, maybe you can try to revert your patch and run the test > again? I'm also seeing OOM with MGLRU with this test, Active/Inactive > LRU is fine. But after reverting your patch, the OOM issue still > exists. > > > I think this is one of the potential side effects -- Huge mentioned > > earlier about isolate_lru_folios(): > > https://lore.kernel.org/linux-mm/503f0df7-91e8-07c1-c4a6-124cad9e65e7@xxxxxxxxxx/ > > > > Try this: > > diff --git a/mm/vmscan.c b/mm/vmscan.c > > index cfa839284b92..778bf5b7ef97 100644 > > --- a/mm/vmscan.c > > +++ b/mm/vmscan.c > > @@ -4320,7 +4320,7 @@ static bool sort_folio(struct lruvec *lruvec, > > struct folio *folio, struct scan_c > > } > > > > /* ineligible */ > > - if (zone > sc->reclaim_idx || skip_cma(folio, sc)) { > > + if (!folio_test_lru(folio) || zone > sc->reclaim_idx || > > skip_cma(folio, sc)) { > > gen = folio_inc_gen(lruvec, folio, false); > > list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]); > > return true; > > Hi Yu, I tested your patch, on my system, the OOM still exists (96 > core and 256G RAM), test memcg is limited to 512M and 32 thread (). > > And I found the OOM seems irrelevant to either your patch or Ge's > patch. (it may changed the OOM chance slight though) > > After the very quick OOM (it failed to untar the linux source code), > checking lru_gen_full: > memcg 47 /build-kernel-tmpfs > node 0 > 442 1691 29405 0 > 0 0r 0e 0p 57r > 617e 0p > 1 0r 0e 0p 0r > 4e 0p > 2 0r 0e 0p 0r > 0e 0p > 3 0r 0e 0p 0r > 0e 0p > 0 0 0 0 > 0 0 > 443 1683 57748 832 > 0 0 0 0 0 > 0 0 > 1 0 0 0 0 > 0 0 > 2 0 0 0 0 > 0 0 > 3 0 0 0 0 > 0 0 > 0 0 0 0 > 0 0 > 444 1670 30207 133 > 0 0 0 0 0 > 0 0 > 1 0 0 0 0 > 0 0 > 2 0 0 0 0 > 0 0 > 3 0 0 0 0 > 0 0 > 0 0 0 0 > 0 0 > 445 1662 0 0 > 0 0R 34T 0 57R > 238T 0 > 1 0R 0T 0 0R > 0T 0 > 2 0R 0T 0 0R > 0T 0 > 3 0R 0T 0 0R > 81T 0 > 13807L 324O 867Y 2538N > 63F 18A > > If I repeat the test many times, it may succeed by chance, but the > untar process is very slow and generates about 7000 generations. > > But if I change the untar cmdline to: > python -c "import sys; sys.stdout.buffer.write(open('$linux_src', > mode='rb').read())" | tar zx > > Then the problem is gone, it can untar the file successfully and very fast. > > This might be a different issue reported by Chris, I'm not sure. After more testing, I think these are two problems (note I changed the memcg limit to 600m later so the compile test can run smoothly). 1. OOM during the untar progress (can be workarounded by the untar cmdline I mentioned above). 2. OOM during the compile progress (this should be the one Chris encountered). Both 1 and 2 only exist for MGLRU. 1 can be workarounded using the cmdline I mentioned above. 2 is caused by Ge's patch, and 1 is not. I can confirm Yu's patch fixed 2 on my system, but the 1 seems still a problem, it's not related to this patch, maybe can be discussed elsewhere.
