Bug 232048
| Summary: | CURRENT: clang crashes during buildworld | ||
|---|---|---|---|
| Product: | Base System | Reporter: | Koop Mast <kwm> |
| Component: | bin | Assignee: | freebsd-bugs (Nobody) <bugs> |
| Status: | Closed Not A Bug | ||
| Severity: | Affects Only Me | CC: | marklmi26-fbsd |
| Priority: | --- | ||
| Version: | CURRENT | ||
| Hardware: | Any | ||
| OS: | Any | ||
|
Description
Koop Mast
2018-10-07 21:09:36 UTC
The mentioned files in the crash are too large for bugzilla so you can get them here: https://people.freebsd.org/~kwm/CGStmtOpenMP-6b8420.tar.xz SHA512 (CGStmtOpenMP-6b8420.tar.xz) = e3ed18a91d74da957b5230b3c4e527e52c7012af9ef6c2ae263fe102ea4afba9f581d2b9f616e8005d865a01314193663cd2379148bc4040863ea58d15cfe037 A common reason for: c++: error: unable to execute command: Killed is being low on free RAM for too long (too many tries to get more free RAM). The console would have reported a message (that in part may be a misnomer), something like: . . . kernel: pid 1010 (stress), uid 0, was killed: out of swap space You can likely believe the "out of swap space" if you also got messages like: . . . kernel: swap_pager_getswapspace(32): failed but otherwise you may have lots of swap space but still get "was killed" messages. In part this is because FreeBSD does not swap out processes that stay runnable, so runnable processes that keep lots of RAM in the "active" status can lead to sustained periods of low free RAM (and, so, probably paging). Changing vm.pageout_oom_seq from its default figure of 12 to something larger makes FreeBSD tolerate the low free RAM longer (more tries to free RAM). Folks have use figures like: sysctl vm.pageout_oom_seq=120 sysctl vm.pageout_oom_seq=1024 on small single board computers with 4 cores (1 hardware thread per core) to allow buildworld buildkernel to complete for -j4 with 1 GiByte or 2 GiByte of RAM. I/O/storage latency for paging and swapping can contribute to needing a larger vm.pageout_oom_seq . Another thing that might help such builds use less RAM at times is to use: LDFLAGS.lld+= -Wl,--no-threads Using a smaller N for -jN is another example (for 1<N). You were not explicit about properties such as the -jN or the amount of RAM or other such. So these notes are not informed by such contextual properties and may not apply. If they do apply, there are other bugzilla reports that this is a duplicate of. It has been an on-going issue for some time but is not obvious from what FreeBSD reports when it does the process kills. Mark Johnston provided the information tying vm.pageout_oom_seq values to controlling the behavior. This was indeed a OOM issue, another process was hogging all the memory. I should have check dmesg first to see if there was something about clang getting killed because of no more available memory. |