| Message ID | 20200731024244.872574-1-andriin@fb.com |
|---|---|
| State | Accepted |
| Delegated to: | BPF Maintainers |
| Headers | show |
| Series | [bpf-next] tools build: propagate build failures from tools/build/Makefile.build | expand |
On Thu, Jul 30, 2020 at 7:44 PM Andrii Nakryiko <andriin@fb.com> wrote: > > The '&&' command seems to have a bad effect when $(cmd_$(1)) exits with > non-zero effect: the command failure is masked (despite `set -e`) and all but > the first command of $(dep-cmd) is executed (successfully, as they are mostly > printfs), thus overall returning 0 in the end. > > This means in practice that despite compilation errors, tools's build Makefile > will return success. We see this very reliably with libbpf's Makefile, which > doesn't get compilation error propagated properly. This in turns causes issues > with selftests build, as well as bpftool and other projects that rely on > building libbpf. > > The fix is simple: don't use &&. Given `set -e`, we don't need to chain > commands with &&. The shell will exit on first failure, giving desired > behavior and propagating error properly. > > Cc: Jiri Olsa <jolsa@kernel.org> > Cc: Arnaldo Carvalho de Melo <acme@redhat.com> > Fixes: 275e2d95591e ("tools build: Move dependency copy into function") > Signed-off-by: Andrii Nakryiko <andriin@fb.com> Jiri, Arnaldo, could you please review and ack?
On Thu, Jul 30, 2020 at 07:42:44PM -0700, Andrii Nakryiko wrote: > The '&&' command seems to have a bad effect when $(cmd_$(1)) exits with > non-zero effect: the command failure is masked (despite `set -e`) and all but > the first command of $(dep-cmd) is executed (successfully, as they are mostly > printfs), thus overall returning 0 in the end. nice, thanks for digging into this, any idea why is the failure masked? Acked-by: Jiri Olsa <jolsa@redhat.com> jirka > > This means in practice that despite compilation errors, tools's build Makefile > will return success. We see this very reliably with libbpf's Makefile, which > doesn't get compilation error propagated properly. This in turns causes issues > with selftests build, as well as bpftool and other projects that rely on > building libbpf. > > The fix is simple: don't use &&. Given `set -e`, we don't need to chain > commands with &&. The shell will exit on first failure, giving desired > behavior and propagating error properly. > > Cc: Jiri Olsa <jolsa@kernel.org> > Cc: Arnaldo Carvalho de Melo <acme@redhat.com> > Fixes: 275e2d95591e ("tools build: Move dependency copy into function") > Signed-off-by: Andrii Nakryiko <andriin@fb.com> > --- > > I'm sending this against bpf-next tree, given libbpf is affected enough for me > to debug this fun problem that no one seemed to notice (or care, at least) in > almost 5 years. If there is a better kernel tree, please let me know. > > tools/build/Build.include | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/tools/build/Build.include b/tools/build/Build.include > index 9ec01f4454f9..585486e40995 100644 > --- a/tools/build/Build.include > +++ b/tools/build/Build.include > @@ -74,7 +74,8 @@ dep-cmd = $(if $(wildcard $(fixdep)), > # dependencies in the cmd file > if_changed_dep = $(if $(strip $(any-prereq) $(arg-check)), \ > @set -e; \ > - $(echo-cmd) $(cmd_$(1)) && $(dep-cmd)) > + $(echo-cmd) $(cmd_$(1)); \ > + $(dep-cmd)) > > # if_changed - execute command if any prerequisite is newer than > # target, or command line has changed > -- > 2.24.1 >
On Sun, Aug 2, 2020 at 9:11 AM Jiri Olsa <jolsa@redhat.com> wrote: > > On Thu, Jul 30, 2020 at 07:42:44PM -0700, Andrii Nakryiko wrote: > > The '&&' command seems to have a bad effect when $(cmd_$(1)) exits with > > non-zero effect: the command failure is masked (despite `set -e`) and all but > > the first command of $(dep-cmd) is executed (successfully, as they are mostly > > printfs), thus overall returning 0 in the end. > > nice, thanks for digging into this, > any idea why is the failure masked? Two things. 1. In make, assume you have command f = a in one function and g = b; c in another. If you write f && g, you end up with (a && b); c, right? 2. Try this shell script: set -ex false && true true It will return success. It won't execute the first true command, as expected, but won't terminate the shell as you'd expect from set -e. So basically, having a "logical operator" in a sequence of commands negates the effect of `set -e`. Intuitively I'd expect that from ||, but seems like && does that as well. if [] has similar effect -- any failing command in an if check doesn't trigger an early termination of a script. > > Acked-by: Jiri Olsa <jolsa@redhat.com> > > jirka > > > > > This means in practice that despite compilation errors, tools's build Makefile > > will return success. We see this very reliably with libbpf's Makefile, which > > doesn't get compilation error propagated properly. This in turns causes issues > > with selftests build, as well as bpftool and other projects that rely on > > building libbpf. > > > > The fix is simple: don't use &&. Given `set -e`, we don't need to chain > > commands with &&. The shell will exit on first failure, giving desired > > behavior and propagating error properly. > > > > Cc: Jiri Olsa <jolsa@kernel.org> > > Cc: Arnaldo Carvalho de Melo <acme@redhat.com> > > Fixes: 275e2d95591e ("tools build: Move dependency copy into function") > > Signed-off-by: Andrii Nakryiko <andriin@fb.com> > > --- > > > > I'm sending this against bpf-next tree, given libbpf is affected enough for me > > to debug this fun problem that no one seemed to notice (or care, at least) in > > almost 5 years. If there is a better kernel tree, please let me know. > > > > tools/build/Build.include | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/tools/build/Build.include b/tools/build/Build.include > > index 9ec01f4454f9..585486e40995 100644 > > --- a/tools/build/Build.include > > +++ b/tools/build/Build.include > > @@ -74,7 +74,8 @@ dep-cmd = $(if $(wildcard $(fixdep)), > > # dependencies in the cmd file > > if_changed_dep = $(if $(strip $(any-prereq) $(arg-check)), \ > > @set -e; \ > > - $(echo-cmd) $(cmd_$(1)) && $(dep-cmd)) > > + $(echo-cmd) $(cmd_$(1)); \ > > + $(dep-cmd)) > > > > # if_changed - execute command if any prerequisite is newer than > > # target, or command line has changed > > -- > > 2.24.1 > > >
On Sun, Aug 02, 2020 at 11:22:07AM -0700, Andrii Nakryiko wrote: > On Sun, Aug 2, 2020 at 9:11 AM Jiri Olsa <jolsa@redhat.com> wrote: > > > > On Thu, Jul 30, 2020 at 07:42:44PM -0700, Andrii Nakryiko wrote: > > > The '&&' command seems to have a bad effect when $(cmd_$(1)) exits with > > > non-zero effect: the command failure is masked (despite `set -e`) and all but > > > the first command of $(dep-cmd) is executed (successfully, as they are mostly > > > printfs), thus overall returning 0 in the end. > > > > nice, thanks for digging into this, > > any idea why is the failure masked? > > Two things. > > 1. In make, assume you have command f = a in one function and g = b; c > in another. If you write f && g, you end up with (a && b); c, right? > > 2. Try this shell script: > > set -ex > false && true > true > > It will return success. It won't execute the first true command, as > expected, but won't terminate the shell as you'd expect from set -e. > > So basically, having a "logical operator" in a sequence of commands > negates the effect of `set -e`. Intuitively I'd expect that from ||, > but seems like && does that as well. if [] has similar effect -- any > failing command in an if check doesn't trigger an early termination of > a script. nice, thanks for explanation jirka > > > > > Acked-by: Jiri Olsa <jolsa@redhat.com> > > > > jirka > > > > > > > > This means in practice that despite compilation errors, tools's build Makefile > > > will return success. We see this very reliably with libbpf's Makefile, which > > > doesn't get compilation error propagated properly. This in turns causes issues > > > with selftests build, as well as bpftool and other projects that rely on > > > building libbpf. > > > > > > The fix is simple: don't use &&. Given `set -e`, we don't need to chain > > > commands with &&. The shell will exit on first failure, giving desired > > > behavior and propagating error properly. > > > > > > Cc: Jiri Olsa <jolsa@kernel.org> > > > Cc: Arnaldo Carvalho de Melo <acme@redhat.com> > > > Fixes: 275e2d95591e ("tools build: Move dependency copy into function") > > > Signed-off-by: Andrii Nakryiko <andriin@fb.com> > > > --- > > > > > > I'm sending this against bpf-next tree, given libbpf is affected enough for me > > > to debug this fun problem that no one seemed to notice (or care, at least) in > > > almost 5 years. If there is a better kernel tree, please let me know. > > > > > > tools/build/Build.include | 3 ++- > > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > diff --git a/tools/build/Build.include b/tools/build/Build.include > > > index 9ec01f4454f9..585486e40995 100644 > > > --- a/tools/build/Build.include > > > +++ b/tools/build/Build.include > > > @@ -74,7 +74,8 @@ dep-cmd = $(if $(wildcard $(fixdep)), > > > # dependencies in the cmd file > > > if_changed_dep = $(if $(strip $(any-prereq) $(arg-check)), \ > > > @set -e; \ > > > - $(echo-cmd) $(cmd_$(1)) && $(dep-cmd)) > > > + $(echo-cmd) $(cmd_$(1)); \ > > > + $(dep-cmd)) > > > > > > # if_changed - execute command if any prerequisite is newer than > > > # target, or command line has changed > > > -- > > > 2.24.1 > > > > > >
On 7/31/20 4:42 AM, Andrii Nakryiko wrote: > The '&&' command seems to have a bad effect when $(cmd_$(1)) exits with > non-zero effect: the command failure is masked (despite `set -e`) and all but > the first command of $(dep-cmd) is executed (successfully, as they are mostly > printfs), thus overall returning 0 in the end. > > This means in practice that despite compilation errors, tools's build Makefile > will return success. We see this very reliably with libbpf's Makefile, which > doesn't get compilation error propagated properly. This in turns causes issues > with selftests build, as well as bpftool and other projects that rely on > building libbpf. > > The fix is simple: don't use &&. Given `set -e`, we don't need to chain > commands with &&. The shell will exit on first failure, giving desired > behavior and propagating error properly. > > Cc: Jiri Olsa <jolsa@kernel.org> > Cc: Arnaldo Carvalho de Melo <acme@redhat.com> > Fixes: 275e2d95591e ("tools build: Move dependency copy into function") > Signed-off-by: Andrii Nakryiko <andriin@fb.com> Applied, thanks!
diff --git a/tools/build/Build.include b/tools/build/Build.include index 9ec01f4454f9..585486e40995 100644 --- a/tools/build/Build.include +++ b/tools/build/Build.include @@ -74,7 +74,8 @@ dep-cmd = $(if $(wildcard $(fixdep)), # dependencies in the cmd file if_changed_dep = $(if $(strip $(any-prereq) $(arg-check)), \ @set -e; \ - $(echo-cmd) $(cmd_$(1)) && $(dep-cmd)) + $(echo-cmd) $(cmd_$(1)); \ + $(dep-cmd)) # if_changed - execute command if any prerequisite is newer than # target, or command line has changed
The '&&' command seems to have a bad effect when $(cmd_$(1)) exits with non-zero effect: the command failure is masked (despite `set -e`) and all but the first command of $(dep-cmd) is executed (successfully, as they are mostly printfs), thus overall returning 0 in the end. This means in practice that despite compilation errors, tools's build Makefile will return success. We see this very reliably with libbpf's Makefile, which doesn't get compilation error propagated properly. This in turns causes issues with selftests build, as well as bpftool and other projects that rely on building libbpf. The fix is simple: don't use &&. Given `set -e`, we don't need to chain commands with &&. The shell will exit on first failure, giving desired behavior and propagating error properly. Cc: Jiri Olsa <jolsa@kernel.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Fixes: 275e2d95591e ("tools build: Move dependency copy into function") Signed-off-by: Andrii Nakryiko <andriin@fb.com> --- I'm sending this against bpf-next tree, given libbpf is affected enough for me to debug this fun problem that no one seemed to notice (or care, at least) in almost 5 years. If there is a better kernel tree, please let me know. tools/build/Build.include | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)