2022-12-11 03:35:37

by Masahiro Yamada

[permalink] [raw]
Subject: [PATCH] kbuild: use .NOTINTERMEDIATE for future GNU Make versions

In Kbuild, some files are generated by chains of pattern/implicit rules.
For example, *.dtb.o files in drivers/of/unittest-data/Makefile are
generated by the chain of 3 pattern rules, like this:

%.dts -> %.dtb -> %.dtb.S -> %.dtb.o

Here, %.dts is the real source, %.dtb.o is the final target.
%.dtb and %.dtb.S are called "intermediate files".

As GNU Make manual [1] says, intermediate files are treated differently
in two ways:

(a) The first difference is what happens if the intermediate file does
not exist. If an ordinary file 'b' does not exist, and make considers
a target that depends on 'b', it invariably creates 'b' and then
updates the target from 'b'. But if 'b' is an intermediate file, then
make can leave well enough alone: it won't create 'b' unless one of
its prerequisites is out of date. This means the target depending
on 'b' won't be rebuilt either, unless there is some other reason
to update that target: for example the target doesn't exist or a
different prerequisite is newer than the target.

(b) The second difference is that if make does create 'b' in order to
update something else, it deletes 'b' later on after it is no longer
needed. Therefore, an intermediate file which did not exist before
make also does not exist after make. make reports the deletion to
you by printing a 'rm' command showing which file it is deleting.

Actually, (b) is problematic for Kbuild because most of the build rules
depend on FORCE and the if_changed* macros really determine if the
target should be updated. So, all missing files, whether they are
intermediate or not, are always rebuilt.

To see why (b) is a problem, delete ".SECONDARY:" from
scripts/Kbuild.include, and repeat this command:

$ make allmodconfig drivers/of/unittest-data/

The intermediate files will be deleted, which results in rebuilding
intermediate and final objects in the next run of make.

In the old days, people suppressed (b) in inconsistent ways.
As commit 54a702f70589 ("kbuild: mark $(targets) as .SECONDARY and
remove .PRECIOUS markers") noted, you should not use .PRECIOUS because
.PRECIOUS has the following behavior (c), which is not likely what you
want.

(c) If make is killed or interrupted during the execution of their
recipes, the target is not deleted. Also, the target is not deleted
on error even if .DELETE_ON_ERROR is specified.

.SECONDARY is a much better way to disable (b), but a small problem
is that .SECONDARY enables (a), which gives a side-effect to $?;
prerequisites marked as .SECONDARY do not appear in $?. This is a
drawback for Kbuild.

I thought it was a bug and opened a bug report. As Paul, the GNU Make
maintainer, concluded in [2], this is not a bug.

A good news is that, GNU Make 4.4 added the perfect solution,
.NOTINTERMEDIATE, which cancels both (a) and (b).

For clarificaton, my understanding of .INTERMEDIATE, .SECONDARY,
.PRECIOUS and .NOTINTERMEDIATE are as follows:

(a) (b) (c)
.INTERMEDIATE enable enable disable
.SECONDARY enable disable disable
.PRECIOUS disable disable enable
.NOTINTERMEDIATE disable disable disable

However, GNU Make 4.4 has a bug for the global .NOTINTERMEDIATE. [3]
It was fixed by commit 6164608900ad ("[SV 63417] Ensure global
.NOTINTERMEDIATE disables all intermediates"), and will be available
in the next release of GNU Make.

The following is the gain for .NOTINTERMEDIATE:

[Current Make]

$ make allnoconfig vmlinux
[ full build ]
$ rm include/linux/device.h
$ make vmlinux
CALL scripts/checksyscalls.sh

Make does not notice the removal of <linux/device.h>.

[Future Make]

$ make-latest allnoconfig vmlinux
[ full build ]
$ rm include/linux/device.h
$ make-latest vmlinux
CC arch/x86/kernel/asm-offsets.s
In file included from ./include/linux/writeback.h:13,
from ./include/linux/memcontrol.h:22,
from ./include/linux/swap.h:9,
from ./include/linux/suspend.h:5,
from arch/x86/kernel/asm-offsets.c:13:
./include/linux/blk_types.h:11:10: fatal error: linux/device.h: No such file or directory
11 | #include <linux/device.h>
| ^~~~~~~~~~~~~~~~
compilation terminated.
make-latest[1]: *** [scripts/Makefile.build:114: arch/x86/kernel/asm-offsets.s] Error 1
make-latest: *** [Makefile:1282: prepare0] Error 2

Make notices the removal of <linux/device.h>, and rebuilds objects
that depended on <linux/device.h>. There exists a source file that
includes <linux/device.h>, and it raises an error.

To see detailed background information, refer to commit 2d3b1b8f0da7
("kbuild: drop $(wildcard $^) check in if_changed* for faster rebuild").

[1]: https://www.gnu.org/software/make/manual/make.html#Chained-Rules
[2]: https://savannah.gnu.org/bugs/?55532
[3]: https://savannah.gnu.org/bugs/?63417

Signed-off-by: Masahiro Yamada <[email protected]>
---

scripts/Kbuild.include | 13 ++++++++++---
1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/scripts/Kbuild.include b/scripts/Kbuild.include
index abdc269a51da..555c68788d09 100644
--- a/scripts/Kbuild.include
+++ b/scripts/Kbuild.include
@@ -185,9 +185,6 @@ endif
make-cmd = $(call escsq,$(subst $(pound),$$(pound),$(subst $$,$$$$,$(cmd_$(1)))))

# Find any prerequisites that are newer than target or that do not exist.
-# (This is not true for now; $? should contain any non-existent prerequisites,
-# but it does not work as expected when .SECONDARY is present. This seems a bug
-# of GNU Make.)
# PHONY targets skipped in both cases.
newer-prereqs = $(filter-out $(PHONY),$?)

@@ -263,4 +260,14 @@ endif
.DELETE_ON_ERROR:

# do not delete intermediate files automatically
+#
+# .NOTINTERMEDIATE is more correct, but only available on newer Make versions.
+# Make 4.4 introduced .NOTINTERMEDIATE, and it appears in .FEATURES, but the
+# global .NOTINTERMEDIATE does not work. We can use it on Make > 4.4.
+# Use .SECONDARY for older Make versions, but "newer-prereq" cannot detect
+# deleted files.
+ifneq ($(and $(filter notintermediate, $(.FEATURES)),$(filter-out 4.4,$(MAKE_VERSION))),)
+.NOTINTERMEDIATE:
+else
.SECONDARY:
+endif
--
2.34.1


2022-12-23 11:56:31

by Nicolas Schier

[permalink] [raw]
Subject: Re: [PATCH] kbuild: use .NOTINTERMEDIATE for future GNU Make versions

On Sun 11 Dec 2022 12:10:59 GMT, Masahiro Yamada wrote:
> In Kbuild, some files are generated by chains of pattern/implicit
> rules.
> For example, *.dtb.o files in drivers/of/unittest-data/Makefile are
> generated by the chain of 3 pattern rules, like this:
>
> %.dts -> %.dtb -> %.dtb.S -> %.dtb.o
>
> Here, %.dts is the real source, %.dtb.o is the final target.
> %.dtb and %.dtb.S are called "intermediate files".
>
> As GNU Make manual [1] says, intermediate files are treated differently
> in two ways:
>
> (a) The first difference is what happens if the intermediate file does
> not exist. If an ordinary file 'b' does not exist, and make considers
> a target that depends on 'b', it invariably creates 'b' and then
> updates the target from 'b'. But if 'b' is an intermediate file, then
> make can leave well enough alone: it won't create 'b' unless one of
> its prerequisites is out of date. This means the target depending
> on 'b' won't be rebuilt either, unless there is some other reason
> to update that target: for example the target doesn't exist or a
> different prerequisite is newer than the target.
>
> (b) The second difference is that if make does create 'b' in order to
> update something else, it deletes 'b' later on after it is no longer
> needed. Therefore, an intermediate file which did not exist before
> make also does not exist after make. make reports the deletion to
> you by printing a 'rm' command showing which file it is deleting.
>
> Actually, (b) is problematic for Kbuild because most of the build rules
> depend on FORCE and the if_changed* macros really determine if the
> target should be updated. So, all missing files, whether they are
> intermediate or not, are always rebuilt.
>
> To see why (b) is a problem, delete ".SECONDARY:" from
> scripts/Kbuild.include, and repeat this command:
>
> $ make allmodconfig drivers/of/unittest-data/
>
> The intermediate files will be deleted, which results in rebuilding
> intermediate and final objects in the next run of make.
>
> In the old days, people suppressed (b) in inconsistent ways.
> As commit 54a702f70589 ("kbuild: mark $(targets) as .SECONDARY and
> remove .PRECIOUS markers") noted, you should not use .PRECIOUS because
> .PRECIOUS has the following behavior (c), which is not likely what you
> want.
>
> (c) If make is killed or interrupted during the execution of their
> recipes, the target is not deleted. Also, the target is not deleted
> on error even if .DELETE_ON_ERROR is specified.
>
> .SECONDARY is a much better way to disable (b), but a small problem
> is that .SECONDARY enables (a), which gives a side-effect to $?;
> prerequisites marked as .SECONDARY do not appear in $?. This is a
> drawback for Kbuild.
>
> I thought it was a bug and opened a bug report. As Paul, the GNU Make
> maintainer, concluded in [2], this is not a bug.
>
> A good news is that, GNU Make 4.4 added the perfect solution,
> .NOTINTERMEDIATE, which cancels both (a) and (b).
>
> For clarificaton, my understanding of .INTERMEDIATE, .SECONDARY,
> .PRECIOUS and .NOTINTERMEDIATE are as follows:
>
> (a) (b) (c)
> .INTERMEDIATE enable enable disable
> .SECONDARY enable disable disable
> .PRECIOUS disable disable enable
> .NOTINTERMEDIATE disable disable disable
>
> However, GNU Make 4.4 has a bug for the global .NOTINTERMEDIATE. [3]
> It was fixed by commit 6164608900ad ("[SV 63417] Ensure global
> .NOTINTERMEDIATE disables all intermediates"), and will be available
> in the next release of GNU Make.
>
> The following is the gain for .NOTINTERMEDIATE:
>
> [Current Make]
>
> $ make allnoconfig vmlinux
> [ full build ]
> $ rm include/linux/device.h
> $ make vmlinux
> CALL scripts/checksyscalls.sh
>
> Make does not notice the removal of <linux/device.h>.
>
> [Future Make]
>
> $ make-latest allnoconfig vmlinux
> [ full build ]
> $ rm include/linux/device.h
> $ make-latest vmlinux
> CC arch/x86/kernel/asm-offsets.s
> In file included from ./include/linux/writeback.h:13,
> from ./include/linux/memcontrol.h:22,
> from ./include/linux/swap.h:9,
> from ./include/linux/suspend.h:5,
> from arch/x86/kernel/asm-offsets.c:13:
> ./include/linux/blk_types.h:11:10: fatal error: linux/device.h: No such file or directory
> 11 | #include <linux/device.h>
> | ^~~~~~~~~~~~~~~~
> compilation terminated.
> make-latest[1]: *** [scripts/Makefile.build:114: arch/x86/kernel/asm-offsets.s] Error 1
> make-latest: *** [Makefile:1282: prepare0] Error 2
>
> Make notices the removal of <linux/device.h>, and rebuilds objects
> that depended on <linux/device.h>. There exists a source file that
> includes <linux/device.h>, and it raises an error.
>
> To see detailed background information, refer to commit 2d3b1b8f0da7
> ("kbuild: drop $(wildcard $^) check in if_changed* for faster rebuild").
>
> [1]: https://www.gnu.org/software/make/manual/make.html#Chained-Rules
> [2]: https://savannah.gnu.org/bugs/?55532
> [3]: https://savannah.gnu.org/bugs/?63417
>
> Signed-off-by: Masahiro Yamada <[email protected]>
> ---
>
> scripts/Kbuild.include | 13 ++++++++++---
> 1 file changed, 10 insertions(+), 3 deletions(-)
>

I really do appreciate the detailed and enlightening reasoning.

Reviewed-by: Nicolas Schier <[email protected]>


Attachments:
(No filename) (5.66 kB)
signature.asc (849.00 B)
Download all attachments