lkl: fix hijack and zpoline parallel builds

[ddiss]

This is a follow up change for commit 3c97822 ("lkl: add tests build barrier") tracked via #558. The hijack and zpoline libraries also share object files, so need a barrier to avoid parallel build failures.

[ddiss]

This is fixing the parallel build failures I raised in #558 (comment) with make-4.4.1-3.3 (openSUSE Tumbleweed) , but it strangely doesn't work with make-4.2.1-7.3.2 (Leap 15.6). The .WAIT install filter appears to be broken:

make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=. obj=lklfuse
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=lib obj=liblkl
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=tests obj=disk
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=tests obj=boot
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=tests obj=disk-vfio-pci
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=tests obj=net-test
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=tests obj=config
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=tests obj=test-dlmopen
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=lib/hijack obj=liblkl-hijack
make -f /home/david/kernel/lkl/tools/build/Makefile.build dir=lib/hijack obj=liblkl-zpoline
\
    install -d /home/david/install/here/usr/include ; \
    install -m 644 include/lkl.h include/lkl_host.h /home/david/kernel/lkl/tools/lkl/include/lkl_autoconf.h \
      include/lkl_config.h /home/david/install/here/usr/include ; \
    cp -r /home/david/kernel/lkl/tools/lkl/include/lkl /home/david/install/here/usr/include
\
    install -d /home/david/install/here/usr/lib ; \
    install -m 644 /home/david/kernel/lkl/tools/lkl/lib/liblkl.so /home/david/kernel/lkl/tools/lkl/lib/hijack/liblkl-hijack.so .WAIT /home/david/kernel/lkl/tools/lkl/lib/hijack/liblkl-zpoline.so /home/david/kernel/lkl/tools/lkl/liblkl.a /home/david/install/here/usr/lib
install: cannot stat '.WAIT': No such file or directory
make: *** [Makefile:180: libraries_install] Error 1

[ddiss]

This is fixing the parallel build failures I raised in #558 (comment) with make-4.4.1-3.3 (openSUSE Tumbleweed) , but it strangely doesn't work with make-4.2.1-7.3.2 (Leap 15.6). The .WAIT install filter appears to be broken:

ah, the expand-targets filter isn't supposed to remove the .WAIT target, only avoid adding $(EXESUF) / $(SOSUF) to it.

[ddiss]

This is fixing the parallel build failures I raised in #558 (comment) with make-4.4.1-3.3 (openSUSE Tumbleweed) , but it strangely doesn't work with make-4.2.1-7.3.2 (Leap 15.6). The .WAIT install filter appears to be broken:

ah, the expand-targets filter isn't supposed to remove the .WAIT target, only avoid adding $(EXESUF) / $(SOSUF) to it.

This extra change seems to work, but it's not exactly pretty:

--- a/tools/lkl/Makefile
+++ b/tools/lkl/Makefile
@@ -177,9 +177,9 @@ headers_install: $(TARGETS)
            cp -r $(OUTPUT)include/lkl $(DESTDIR)$(INCDIR)
 
 libraries_install: $(call expand-targets,$(libs-y),$(SOSUF)) $(OUTPUT)liblkl.a
-       $(call QUIET_INSTALL, libraries) \
+       $(if $(filter .%,$^),,$(call QUIET_INSTALL, libraries) \
            install -d $(DESTDIR)$(LIBDIR) ; \
-           install -m 644 $^ $(DESTDIR)$(LIBDIR)
+           install -m 644 $^ $(DESTDIR)$(LIBDIR))

Will squash this into the commit, but please feel free to suggest something cleaner.

[ddiss]

v2:

• filter .WAIT target from install invocation

[ddiss]

make install rule still isn't working for older distros. keeping this flagged as draft.

[ddiss]

I finally have a fix that I'm happy with for this. As mentioned, the .WAIT target causes issues with libraries_install on make 4.2.1 (works with 4.4.1), so I went with a simple shared -in.o prerequisite. Will do some more testing and then push here.

[ddiss]

v3.

• use a simple shared libhijack-priv-in.o prerequisite instead of .WAIT

[ddiss]

windows-2022 test failure is due to:

  TEST       not ok   nanosleep 
sleep 78556400 (ret=0), expected sleep 87654321

circle-ci qemu fails to boot:

Too long with no output (exceeded 10m0s): context deadline exceeded

Both of these are unrelated to the change here IMO.

[ddiss]

This is linking libhijack-priv-in.o into liblkl.so which is obviously wrong. I need another minor change to track the hijack libraries separate from liblkl.so...

[ddiss]

This is linking libhijack-priv-in.o into liblkl.so which is obviously wrong. I need another minor change to track the hijack libraries separate from liblkl.so...

The existing $(OUTPUT)liblkl$(SOSUF) rule doesn't match anything, due to the lack of the lib/ prefix. Fixing that up and linking there, separate to the hijack libraries, seems reasonable. v4 to follow...

[ddiss]

v4:

• the existing liblkl$(SOSUF) target is currently unmatched, so fix it and use it as a linker target which omits libhijack-priv-in.o

[ddiss]

One follow up I'd also like to propose (via a separate PR) is to use the same private common in.o approach for tests, effectively reverting the use of .WAIT in 3c97822 . Instead we'd be adding something like:

$(tests-y:%=$(OUTPUT)%$(EXESUF)): $(tests-priv-y:%=$(OUTPUT)%-in.o) 
...
tests-priv-y := tests/tests-common
...
tests-common-y += cla.o test.o