An obscure make --shuffle bug

Most of the time make --shuffle finds somewhat trivial bugs of missing dependencies. I have collected a few examples here.

It is sometimes hard, but still feasible to trigger these bugs without a shuffle by carefully placing a few sleep 10 calls here and there to reorder rule execution.

But today I’ll talk about the bugs that could not be easily triggered by just waiting more. Should these bugs be treats as bugs in Makefiles? Or are they bugs in the make --shuffle permutation algorithm? You be the judge!

A gist of the Makefile language

Makefile language is an interesting one: it has a strong spirit of a declarative family: it should not be very important in which order you define rules and variables in the file, the rule execution order should not matter much either. That property makes the input very friendly to parallelism: you can consider a Makefile as a form of a serialized build graph.

The graph analogy holds surprisingly well for most of the build systems out there. But sometimes you get the rare ones that are clearly not designed to be executed in parallel. Those are also fine. We can add .NOTPARALLEL: to them and be done with it.

If we ignore clearly unclean build systems there are a few GNU extensions that break the graph model a bit. Things like immediate variable assignments A := B are one of them.

For example the following Makefile illustrates the subtlety of immediate and deferred assignments:

all: a b

V := 1
AI := $(V)
AD = $(V)
a:
        @echo a: V=$(V) AI=$(AI) AD=$(AD)

V := 2
BI := $(V)
BD = $(V)
b:
        @echo b: V=$(V) BI=$(BI) BD=$(BD)

Take some time time and try to predict the output here.

Ready now? Here is the correct answer:

$ make
a: V=2 AI=1 AD=2
b: V=2 BI=2 BD=2

Here both V= and AD= values are deferred until the final assignment. This is a well-known gotcha of Makefiles: immediate and deferred assignments can influence macro expansion and rule definition.

Immediate assignment arguably does not break declarative model too much.

Don’t worry just yet: make --shuffle does not change output of this example.

Target-specific variables

Another GNU make extension is Target-specific Variable Values.

The typical use looks benign:

all: foo bar
        @echo all

foo: a.o b.o
        @echo "foo ($^)"

bar: c.o d.o
        @echo "bar ($^)"

%.o:
        @echo cc $(CFLAGS) -o $@

foo: CFLAGS=-g

Here we define 2 targets: foo and bar. But for foo (and it’s prerequisites) we specify CFLAGS=-g target-specific override.

For this case the execution looks like that:

$ make
cc -g -o a.o
cc -g -o b.o
foo (a.o b.o)
cc -o c.o
cc -o d.o
bar (c.o d.o)
all

No surprises here either: all object files needed for foo use CFLAGS=-g override. None of the object files used by bar use -g:

So far so good. No ambiguities here.

make --shuffle does not change this output either.

Today’s problem

Now, what happens if we do share an object file across the targets and happen to use different overrides for it?

Here is our toy example:

all: foo bar
        @echo all

foo: a.o b.o shared.o
        @echo "foo ($^)"

bar: c.o d.o shared.o
        @echo "bar ($^)"

%.o:
        @echo cc $(CFLAGS) -o $@

foo: CFLAGS=-ffoo
bar: CFLAGS=-fbar

What output do you expect here? Does shared.o get compiled once? Or twice? Does it use -ffoo or -ffbar? Is it deterministic?

Take some time and come up with an answer.

How it all behaves

Let’s start from the simple case:

$ make
cc -ffoo -o a.o
cc -ffoo -o b.o
cc -ffoo -o shared.o
foo (a.o b.o shared.o)
cc -fbar -o c.o
cc -fbar -o d.o
bar (c.o d.o shared.o)
all

Here shared.o gets built once (it’s a single node in our build graph after all) and it uses -ffoo flag.

Now, why does it use -ffoo and not -fbar? Can you come up with a scenario when it uses -fbar?

Apparently if you use just make or make -j you will consistently get the same cc -ffoo -o shared.o command.

The way GNU make builds the graph is the topological sorting of targets and prerequisites defined by Makefile syntax.

It is always starts with a.o b.o shared.o and maybe finishes with foo c.o d.o bar all.

But there is an easy way to alter this order! Try this instead:

$ make bar foo
cc -fbar -o c.o
cc -fbar -o d.o
cc -fbar -o shared.o
bar (c.o d.o shared.o)
cc -ffoo -o a.o
cc -ffoo -o b.o
foo (a.o b.o shared.o)

Now shared.o is built with -fbar flag! And the same would happen if you did a partial make bar build. Or if you would change Makefile from all: foo bar to all: bar foo:

--- a/Makefile
+++ b/Makefile
@@ -1,4 +1,4 @@
-all: foo bar
+all: bar foo
        @echo all

 foo: a.o b.o shared.o

This is a problem. And that can be triggered by make --shuffle!

No shuffle:

$ for i in `seq 1 1000`; do make -j | grep -- '-o shared.o'; done | sort | uniq -c
   1000 cc -ffoo -o shared.o

Ideal consistency. Nothing shows the problem.

With shuffle:

$ for i in `seq 1 1000`; do make -j --shuffle | grep -- '-o shared.o'; done | sort | uniq -c
    494 cc -fbar -o shared.o
    506 cc -ffoo -o shared.o

Almost 50% perfect instability!

What is the correct fix here? Is it a make --shuffle bug to introduce non-determinism? I would say: no, Makefile should be more robust in face of target sequence passed to it.

How would one fix it properly? This default recursive propagation is non-intuitive, GNU make provides a few keywords including private. private makes the variable override local to specified targets and does not recurse.

I suggest the following fix:

all: foo bar
        @echo all

foo: a.o b.o shared.o
        @echo "foo ($^)"

bar: c.o d.o shared.o
        @echo "bar ($^)"

%.o:
        @echo cc $(CFLAGS) -o $@

a.o b.o: private CFLAGS=-ffoo
c.o d.o: private CFLAGS=-fbar
shared.o: private CFLAGS=-fshared

Note that here we use the new -fshared value to make it distinct from the other two. Running:

$ for i in `seq 1 1000`; do make -j --shuffle | grep -- '-o shared.o'; done | sort | uniq -c
   1000 cc -fshared -o shared.o

No instability!

The need for private

Is private really needed? In this toy example it would work the same with and without it.

The troubles would start if any of .o files would have in prerequisites another .o file, say a generator that builds header files or some other input:

all: foo bar
        @echo all

foo: a.o b.o shared.o
        @echo "foo ($^)"

bar: c.o d.o shared.o
        @echo "bar ($^)"

%.o:
        @echo cc $(CFLAGS) -o $@

a.o b.o: CFLAGS=-ffoo
c.o d.o: CFLAGS=-fbar
shared.o: CFLAGS=-fshared

# simulate that also shared.o depends on `generated.h`
# generated by `gen` tool.
all: gen

gen: gen.o
        @echo gen
gen.h: gen
        @echo gen.h

shared.o: gen.h

This time we will check what flags are passed to gen.o.

Without shuffle:

$ for i in `seq 1 1000`; do make -j | grep -- '-o gen.o'; done | sort | uniq -c
   1000 cc -fshared -o gen.o

Whoops: -fshared got leaked into gen.o build, but at least it’s consistent.

With --shuffle:

$ for i in `seq 1 1000`; do make -j --shuffle | grep -- '-o gen.o'; done | sort | uniq -c
    697 cc -fshared -o gen.o
    303 cc -o gen.o

33% instability! Why it’s just 33% (and not, say, 50%) is an exercise for the reader.

Adding private to all of CFLAGS= fixes the issue:

$ for i in `seq 1 1000`; do make -j --shuffle | grep -- '-o gen.o'; done | sort | uniq -c
   1000 cc -o gen.o

It’s not just stable: it also removes unexpected -fshared flag as one would probably expect.

The real world example

The above hypothetical example is almost 1-to-1 example of what happened to efivar in PR #245.

There make --shuffle build occasionally failed as:

$ gcc -Og  -g3 -Wall -Wextra  -Werror  -std=gnu11 -funsigned-char -fvisibility=hidden -specs=/build/source/src/include/gcc.specs -fno-merge-constants  -L.   -Wl,--build-id -Wl,--no-allow-shlib-undefined -Wl,--no-undefined-version -Wl,-z,now -Wl,-z,muldefs -Wl,-z,relro -Wl,--fatal-warnings     -DLIBEFIVAR_VERSION=38 -D_GNU_SOURCE -D_FILE_OFFSET_BITS=64 -I/build/source/src/include/  -shared -Wl,-soname,libefisec.so.1 -Wl,--version-script=libefisec.map  -o libefisec.so sec.o secdb.o esl-iter.o util.o -lefivar -lefisec -ldl
ld: cannot find -lefivar: No such file or directory
collect2: error: ld returned 1 exit status
make[1]: *** [/build/source/src/include/rules.mk:38: libefisec.so] Error 1 shuffle=721268944

Initially I thought it was just a matter of missing libefivar.so dependency on libefisec. But Artem Klimov pointed out that those two are not expected to depend on one another.

And indeed Makefile uses LIBS variable to specify dependencies:

all: libefivar.so libefisec.so efisecdb # and more ...
# ...
libefivar.so : LIBS=dl
# ...
efisecdb : $(EFISECDB_OBJECTS) | libefisec.so
efisecdb : LIBS=efivar efisec dl
# ...
libefisec.so : # nothing

The intent of the Makefile is to say that libefisec.so has no external LIBS= dependencies. But LIBS= override gets leaked via efisecdb : LIBS=efivar efisec dl and into libefisec.so via order-only dependency.

The fix is to add private keyword to prevent the LIBS=... leakage from efisecdb into libraries in it’s prerequisite list:

--- a/src/Makefile
+++ b/src/Makefile
@@ -42,3 +42,3 @@
 # ...
 efisecdb : $(EFISECDB_OBJECTS) | libefisec.so
-efisecdb : LIBS=efivar efisec dl
+efisecdb : private LIBS=efivar efisec dl

Parting words

Target-specific Variable Values is a tool with very sharp edges. While it allows writing very concise build rules their recursive nature quickly get out of hand if you don’t see the whole Makefile in front of you.

One of the mitigations is to use private keyword at an assignment site.

Another mitigation is to use different global variables for different purposes. Individual global variables are also easier to override from the outside.

make --shuffle might be able to expose instability problems in Makefiles that use target-specific variables not reproducible by using make -j alone.

Have fun!