Message-ID: <20190318191926.GA23138@redhat.com>
Date:   Mon, 18 Mar 2019 15:19:26 -0400
From:   Joe Lawrence <joe.lawrence@...hat.com>
To:     Joao Moreira <jmoreira@...e.de>
Cc:     live-patching@...r.kernel.org, mbenes@...e.cz, pmladek@...e.cz,
        jikos@...e.cz, nstange@...e.de, jpoimboe@...hat.com,
        khlebnikov@...dex-team.ru, jeyu@...nel.org, matz@...e.de,
        linux-kernel@...r.kernel.org, yamada.masahiro@...ionext.com,
        linux-kbuild@...r.kernel.org, michal.lkml@...kovi.net
Subject: Re: [PATCH v2 2/8] kbuild: Support for Symbols.list creation

On Fri, Mar 01, 2019 at 11:13:07AM -0300, Joao Moreira wrote:
> For automatic resolution of livepatch relocations, a file called
> Symbols.list is used. This file maps symbols within every compiled
> kernel object allowing the identification of symbols whose name is
> unique, thus relocation can be automatically inferred, or providing
> information that helps developers when code annotation is required for
> solving the matter.
> 
> Add support for creating Symbols.list in the main Makefile. First,
> ensure that built-in is compiled when CONFIG_LIVEPATCH is enabled (as
> required to achieve a complete Symbols.list file). Define the command to
> build Symbols.list (cmd_klp_map) and hook it in the modules rule.
> 
> As it is undesirable to have symbols from livepatch objects inside
> Symbols.list, make livepatches discernible by modifying
> scripts/Makefile.build to create a .livepatch file for each livepatch
> in $(MODVERDIR). This file then used by cmd_klp_map to identify and
> bypass livepatches.
> 
> For identifying livepatches during the build process, a flag variable
> LIVEPATCH_$(basetarget).o is considered in scripts/Makefile.build. This
> way, set this flag for the livepatch sample Makefile in
> samples/livepatch/Makefile.
> 
> Finally, Add a clean rule to ensure that Symbols.list is removed during
> clean.
> 
> Notes:
> 
> To achieve a correct Symbols.list file, all kernel objects must be
> considered, thus, its construction require these objects to be priorly
> built. On the other hand, invoking scripts/Makefile.modpost without
> having a complete Symbols.list in place would occasionally lead to
> in-tree livepatches being post-processed incorrectly. To prevent this
> from becoming a circular dependency, the construction of Symbols.list
> uses non-post-processed kernel objects and such does not cause harm as
> the symbols normally referenced from within livepatches are visible at
> this stage. Also due to these requirements, the spot in-between modules
> compilation and the invocation of scripts/Makefile.modpost was picked
> for hooking cmd_klp_map.
> 
> The approach based on .livepatch files was proposed as an alternative
> to using MODULE_INFO statementes. This approach was originally
                       ^^^^^^^^^^^
nit: s/statementes/statements

> proposed by Miroslav Benes as a workaround for identifying livepathces
                                                             ^^^^^^^^^^^
nit: s/livepathces/livepatches

> without depending on modinfo during the modpost stage. It was moved to
> this patch as the approach also shown to be useful while building
> Symbols.list.
> 
> Signed-off-by: Joao Moreira <jmoreira@...e.de>
> ---
>  .gitignore                 |  1 +
>  Makefile                   | 28 +++++++++++++++++++++++++---
>  samples/livepatch/Makefile |  1 +
>  scripts/Makefile.build     |  6 ++++++
>  4 files changed, 33 insertions(+), 3 deletions(-)
> 
> [ ... snip ... ]
> diff --git a/Makefile b/Makefile
> index 9ef547fc7ffe..b28eba2cad01 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -561,10 +561,13 @@ KBUILD_BUILTIN := 1
>  # If we have only "make modules", don't compile built-in objects.
>  # When we're building modules with modversions, we need to consider
>  # the built-in objects during the descend as well, in order to
> -# make sure the checksums are up to date before we record them.
> +# make sure the checksums are up to date before we record them. The
> +# same applies for building livepatches, as built-in objects may hold
> +# symbols which are referenced from livepatches and are required by
> +# klp-convert post-processing tool for resolving these cases.
>  
>  ifeq ($(MAKECMDGOALS),modules)
> -  KBUILD_BUILTIN := $(if $(CONFIG_MODVERSIONS),1)
> +  KBUILD_BUILTIN := $(if $(or $(CONFIG_MODVERSIONS), $(CONFIG_LIVEPATCH)),1)
>  endif
>  
>  # If we have "make <whatever> modules", compile modules
> @@ -1244,9 +1247,25 @@ all: modules
>  # duplicate lines in modules.order files.  Those are removed
>  # using awk while concatenating to the final file.
>  
> +quiet_cmd_klp_map = KLP	Symbols.list

Nit: the tab between "KLP" and "Symbols.list" doesn't match the same
alignment as other build commands:

  [squash] kbuild: align KLP prefix 
  https://github.com/torvalds/linux/commit/6419a05560731eb1306e441aadcaa2dfd9f9f935

> [ ... snip ... ]
> @@ -1341,7 +1360,10 @@ vmlinuxclean:
>  	$(Q)$(CONFIG_SHELL) $(srctree)/scripts/link-vmlinux.sh clean
>  	$(Q)$(if $(ARCH_POSTLINK), $(MAKE) -f $(ARCH_POSTLINK) clean)
>  
> -clean: archclean vmlinuxclean
> +klpclean:
> +	$(Q) rm -f $(objtree)/Symbols.list
> +
> +clean: archclean vmlinuxclean klpclean

Another nit, but the klpclean target doesn't create a klpclean file, so
we should add it to PHONY:

  [squash] kbuild: add klpclean to PHONY 
  https://github.com/torvalds/linux/commit/7a8b0bea2177c8150ead4848c4498a6081b4a5ae

> [ ... snip ... ]
> diff --git a/samples/livepatch/Makefile b/samples/livepatch/Makefile
> index 2472ce39a18d..9705df7f9a86 100644
> --- a/samples/livepatch/Makefile
> +++ b/samples/livepatch/Makefile
> @@ -1,3 +1,4 @@
> +LIVEPATCH_livepatch-sample.o := y
>  obj-$(CONFIG_SAMPLE_LIVEPATCH) += livepatch-sample.o
>  obj-$(CONFIG_SAMPLE_LIVEPATCH) += livepatch-shadow-mod.o
>  obj-$(CONFIG_SAMPLE_LIVEPATCH) += livepatch-shadow-fix1.o

Would this change be better located in [PATCH 6/8] modpost: Add modinfo
flag to livepatch modules?  Or was it needed to test-drive the rest of
this patch?

> diff --git a/scripts/Makefile.build b/scripts/Makefile.build
> index fd03d60f6c5a..1e28ad21314c 100644
> --- a/scripts/Makefile.build
> +++ b/scripts/Makefile.build
> @@ -247,6 +247,11 @@ cmd_gen_ksymdeps = \
>  	$(CONFIG_SHELL) $(srctree)/scripts/gen_ksymdeps.sh $@ >> $(dot-target).cmd
>  endif
>  
> +ifdef CONFIG_LIVEPATCH
> +cmd_livepatch = $(if $(LIVEPATCH_$(basetarget).o),			\
> +	$(shell touch $(MODVERDIR)/$(basetarget).livepatch))
> +endif
> +
>  define rule_cc_o_c
>  	$(call cmd,checksrc)
>  	$(call cmd_and_fixdep,cc_o_c)
> @@ -283,6 +288,7 @@ $(single-used-m): $(obj)/%.o: $(src)/%.c $(recordmcount_source) $(objtool_dep) F
>  	$(call if_changed_rule,cc_o_c)
>  	@{ echo $(@:.o=.ko); echo $@; \
>  	   $(cmd_undef_syms); } > $(MODVERDIR)/$(@F:.o=.mod)
> +	$(call cmd_livepatch)
>  
>  quiet_cmd_cc_lst_c = MKLST   $@
>        cmd_cc_lst_c = $(CC) $(c_flags) -g -c -o $*.o $< && \

Since cmd_livepatch is only called for single-used-m, does this mean
that we can only klp-convert single object file livepatch modules?

I stumbled upon this when trying to create a self-test module that
incorporated two object files.  I tried adding a $(call cmd_livepatch)
in the recipe for $(obj)/%.o, but that didn't help.  My kbuild foo
wasn't good enough to figure this one out.

-- Joe

Powered by blists - more mailing lists