> Am 19.07.2019 um 20:17 schrieb Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>: > > On Fri, 19 Jul 2019 15:12:24 +0200, Ilya Leoshkevich wrote: >>> Am 18.07.2019 um 20:51 schrieb Jakub Kicinski <jakub.kicinski@xxxxxxxxxxxxx>: >>> >>> We should probably make a script with all the ways of calling make >>> should work. Otherwise we can lose track too easily. >> >> Thanks for the script! >> >> I’m trying to make it all pass now, and hitting a weird issue in the >> Kbuild case. The build prints "No rule to make target >> 'scripts/Makefile.ubsan.o'" and proceeds with an empty BPFTOOL_VERSION, >> which causes problems later on. > > Does it only break with UBSAN enabled? No, all the time. I think this is a coincidence - make happens to scan scripts/Makefile.ubsan first. > >> I've found that this is caused by sub_make_done=1 environment variable, >> and unsetting it indeed fixes the problem, since the root Makefile no >> longer uses the implicit %.o rule. >> >> However, I wonder if that would be acceptable in the final version of >> the patch, and whether there is a cleaner way to achieve the same >> effect? > > I'm not sure to be honest. Did you check how perf deals with that? perf obtains the version using "git describe". However, if we are building it from a tarball, it falls back to "make kernelversion" and fails in a similar way: linux-5.3-rc1$ make defconfig linux-5.3-rc1$ make tools/perf <snip> make[6]: Circular scripts/Makefile.ubsan.mod <- scripts/Makefile.ubsan.o dependency dropped. make[6]: m2c: Command not found make[6]: *** [<builtin>: scripts/Makefile.ubsan.o] Error 127 make[5]: *** [Makefile:1765: scripts/Makefile.ubsan.o] Error 2 <snip> The same trick helps: --- tools/perf/util/PERF-VERSION-GEN.orig 2019-07-23 17:12:07.621123187 +0200 +++ tools/perf/util/PERF-VERSION-GEN 2019-07-23 17:12:33.441133619 +0200 @@ -26,7 +26,7 @@ fi if test -z "$TAG" then - TAG=$(MAKEFLAGS= make -sC ../.. kernelversion) + TAG=$(MAKEFLAGS= sub_make_done= make -sC ../.. kernelversion) fi VN="$TAG$CID" if test -n "$CID"
