So here is full code of of _do_reclaim(): 1001 void reclaimer::_do_reclaim() 1002 { 1003 ssize_t target; 1004 emergency_alloc_level = 1; 1005 1006 while (true) { 1007 WITH_LOCK(free_page_ranges_lock) { 1008 _blocked.wait(free_page_ranges_lock); 1009 *target = bytes_until_normal();* 1010 } 1011 1012 // This means that we are currently ballooning, we should 1013 // try to serve the waiters from temporary memory without 1014 // going on hard mode. A big batch of more memory is likely 1015 // in its way. 1016 if (_oom_blocked.has_waiters() && throttling_needed()) { 1017 _shrinker_loop(target, [] { return false; }); 1018 WITH_LOCK(free_page_ranges_lock) { 1019 if (_oom_blocked.wake_waiters()) { 1020 continue; 1021 } 1022 } 1023 } 1024 1025 _shrinker_loop(target, [this] { return _oom_blocked.has_waiters(); }); 1026 1027 WITH_LOCK(free_page_ranges_lock) { 1028 if (target >= 0) { 1029 // Wake up all waiters that are waiting and now have a chance to succeed. 1030 // If we could not wake any, there is nothing really we can do. 1031 if (!_oom_blocked.wake_waiters()) { 1032 oom(); 1033 } 1034 } 1035 1036 if (balloon_api) { 1037 balloon_api->voluntary_return(); 1038 } 1039 } 1040 } 1041 }
We got oom() because target was '>= 0'. Now the target is calculated as the result of *bytes_until_normal()*. 495 ssize_t reclaimer::bytes_until_normal(pressure curr) 496 { 497 assert(mutex_owned(&free_page_ranges_lock)); 498 if (curr == pressure::PRESSURE) { 499 return watermark_lo - stats::free(); 500 } else { 501 return 0; 502 } 503 } which seems to indicate that when 0 is returned there no need to reclaim any memory. So here are two things that might be wrong: 1. Shouldn't if (target >= 0) be changed to if (target > 0) {? 2. Shouldn't we re-read the target in second WITH_LOCK instead of comparing the original value in the beginning of the body of the loop? The line before - _shrinker_loop(target, [this] { return _oom_blocked.has_waiters(); }); - might have just released enough memory to bring target below 0, right? In any case it would be useful to print the value of the target before oom(): if (!_oom_blocked.wake_waiters()) { printf("--> Target: %ld\n", target); oom(); } On Monday, March 9, 2020 at 10:51:02 PM UTC-4, rickp wrote: > > > We're pretty close to current on OSv, but it also happens on an older > image. We have changed some stuff in our app, but I think that may just > be provking the bug. Certainly from gdb, I can see that both mmaped and > normal memory fluctuate up and down but eveything looks sane. > > More debug in wake_waiters would be useful, but I'm losing the argument > to continue with OSv at the moment which makes testing this a bit > 'political'. > > btw - when we do run the system out of memory, it seems to hang rather > than generate an oom. Have you tried it? > > The tcp_do_segment one has been mentioned before (by someone else). The > issue is that the kassert only has effect in the debug build. I'd guess > that the socket is being closed, but still has segments that have not > been processed, or something like that. I'll try and narrow it down a > bit if I get time. > > Rick > > On Mon, 2020-03-09 at 22:32 -0400, Waldek Kozaczuk wrote: > > Does it happen with the very latest OSv code? Did it start happening > > at some point more often? > > > > I wonder if we could add some helpful printouts in wake_waiters(). > > > > Btw that assert() failure in tcp_do_segment() rings a bell. > > > > On Mon, Mar 9, 2020 at 22:25 Rick Payne <ri...@rossfell.co.uk > <javascript:>> wrote: > > > I can't add much other than I doubt its fragmentation. Sometimes > > > this > > > happens within a few minutes of the system starting. At no point do > > > I > > > think we're using more than 2GB of ram (of the 12GB) either. > > > > > > I did compile up a debug verison of OSv and built the system with > > > that, > > > but I've been unable to trigger the oom(). Worse, I hit a kassert > > > in > > > the netchannel code that seems to be ignored in the 'release' > > > build, > > > but panics in the debug build: > > > > > > [E/384 bsd-kassert]: tcp_do_segment: TCPS_LISTEN > > > Assertion failed: tp->get_state() > 1 > > > (bsd/sys/netinet/tcp_input.cc: > > > tcp_do_segment: 1076) > > > > > > [backtrace] > > > 0x0000000040221330 <abort(char const*, ...)+280> > > > 0x0000000040221399 <__assert_fail+64> > > > 0x00000000402a4798 <???+1076512664> > > > 0x00000000402a97c2 <???+1076533186> > > > 0x00000000402a98a1 <???+1076533409> > > > 0x00000000402aa448 <???+1076536392> > > > 0x0000000040656a9a <std::function<void (mbuf*)>::operator()(mbuf*) > > > const+76> > > > 0x0000000040655855 <net_channel::process_queue()+61> > > > 0x000000004023b165 <???+1076080997> > > > 0x000000004023b4d7 <soclose+878> > > > 0x000000004024cd21 <socket_file::close()+51> > > > 0x00000000406a6a10 <fdrop+151> > > > 0x00000000406a64f7 <fdclose(int)+184> > > > 0x000000004067cd42 <close+41> > > > > > > So at the moment, I'm a bit stuck with getting any more info... > > > > > > Rick > > > > > > On Mon, 2020-03-09 at 08:52 -0700, Waldek Kozaczuk wrote: > > > > As I understand this stack trace the oom() was called here as > > > part of > > > > _do_reclaim(): > > > > > > > > 1025 WITH_LOCK(free_page_ranges_lock) { > > > > 1026 if (target >= 0) { > > > > 1027 // Wake up all waiters that are waiting and > > > now > > > > have a chance to succeed. > > > > 1028 // If we could not wake any, there is > > > nothing > > > > really we can do. > > > > 1029 if (!_oom_blocked.wake_waiters()) { > > > > 1030 oom(); > > > > 1031 } > > > > 1032 } > > > > 1033 > > > > 1034 if (balloon_api) { > > > > 1035 balloon_api->voluntary_return(); > > > > 1036 } > > > > 1037 } > > > > > > > > so it seems wake_waiters() returned false. I wonder if the memory > > > was > > > > heavily fragmented or there is some logical bug in there. This > > > method > > > > is called from two places and I wonder if this part of > > > wake_waiters() > > > > is correct: > > > > > > > > 921 if (!_waiters.empty()) { > > > > 922 reclaimer_thread.wake(); > > > > 923 } > > > > 924 return woken; > > > > > > > > > > > > should this if also set woken to true? > > > > > > > > Also could we also enhance the oom() logic to print out more > > > useful > > > > information if this happens once again? > > > > > > > > On Tuesday, March 3, 2020 at 2:21:40 AM UTC-5, rickp wrote: > > > > > Had a crash on a system that I don't understand. Its a VM with > > > > > 12GB > > > > > allocated, we were running without about 10.5GB free according > > > to > > > > > the > > > > > API. > > > > > > > > > > Out of the blue, we had a panic: > > > > > > > > > > Out of memory: could not reclaim any further. Current memory: > > > > > 10954988 > > > > > Kb > > > > > [backtrace] > > > > > 0x00000000403f6320 <memory::oom()+32> > > > > > 0x00000000403f71cc <memory::reclaimer::_do_reclaim()+380> > > > > > 0x00000000403f722f <???+1077899823> > > > > > 0x000000004040f29b <thread_main_c+43> > > > > > 0x00000000403ae412 <???+1077601298> > > > > > > > > > > The 'Out of memory' message seems to print stats::free() and > > > that > > > > > number suggests we have plenty of free ram. > > > > > > > > > > Have I misunderstood, or is there something I need to be > > > looking > > > > > at? > > > > > > > > > > Cheers, > > > > > Rick > > > > > > > > > > > > > -- > > > > You received this message because you are subscribed to the > > > Google > > > > Groups "OSv Development" group. > > > > To unsubscribe from this group and stop receiving emails from it, > > > > send an email to osv...@googlegroups.com <javascript:>. > > > > To view this discussion on the web visit > > > > > > > > https://groups.google.com/d/msgid/osv-dev/8f7e00a5-edfe-4487-aa5a-5072a560c6e3%40googlegroups.com > > > > > . > > > > > -- You received this message because you are subscribed to the Google Groups "OSv Development" group. To unsubscribe from this group and stop receiving emails from it, send an email to osv-dev+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/osv-dev/c6874fed-fa27-4872-bc50-bae0fe1eb4ec%40googlegroups.com.