kpatch-build: support CONFIG_LTO_CLANG_THIN

Support CONFIG_LTO_CLANG_THIN with ld.lld --lto-obj-path option.

With CONFIG_LTO_CLANG_THIN, .o files are LLVM IR binary, so CDO doesn't
work on .o file. To solve this issue, we CDO the thinlto files generated
by the --lto-obj-path option. Clang LTO generates the thinlto files
after cross file inline, so they are good candidates for CDO. See [1] for
more discussions about this.

To achieve this, we need:

  1. kpatch-build to update kernel Makefile(s) so it generates thinlto
     files;
  2. kpatch-build and kpatch-cc to save the thinlto file properly;
  3. kpatch-build to feed these thinlto files to CDO;
  4. A small workaround in CDO that ignores changes in init.text for
     vmlinux.o.thinlto.*

[1] dynup#1320

Signed-off-by: Song Liu <[email protected]>

kpatch-build/create-diff-object.c

@@ -2825,6 +2825,21 @@ static void kpatch_mark_ignored_sections(struct kpatch_elf *kelf)
 		    !strncmp(sec->name, ".llvm_addrsig", 13) ||
 		    !strncmp(sec->name, ".llvm.", 6))
 			sec->ignore = 1;
+
+		/*
+		 * With CONFIG_LTO_CLANG, we see some weird new functions,
+		 * such as:
+		 * __initstub__kmod_syscall__728_5326_bpf_syscall_sysctl_init7
+		 * while the original function is very similar, like:
+		 * __initstub__kmod_syscall__728_5324_bpf_syscall_sysctl_init7
+		 *
+		 * We don't have very good solution for it yet. To workaround
+		 * the issue, it seems safe to ignore .init.text changes in
+		 * vmlinux.o. This will skip such new functions.
+		 */
+		if (!strncmp(childobj, "vmlinux.o", 9) &&
+		    !strncmp(sec->name, ".init.text", 10))
+			sec->ignore = 1;
 	}
 
 	sec = find_section_by_name(&kelf->sections, ".kpatch.ignore.sections");

kpatch-build/kpatch-build

@@ -179,6 +179,12 @@ cleanup() {
 	# restore original Makefile.modfinal if we updated it for the build
 	[[ -e "$TEMPDIR/Makefile.modfinal" ]] && mv -f "$TEMPDIR/Makefile.modfinal" "$KERNEL_SRCDIR/scripts"
 
+	# restore original Makefile.build if we updated it for the build
+	[[ -e "$TEMPDIR/Makefile.build" ]] && mv -f "$TEMPDIR/Makefile.build" "$KERNEL_SRCDIR/scripts"
+
+	# restore Makefile if we updated it for the build
+	[[ -e "$TEMPDIR/Makefile" ]] && mv -f "$TEMPDIR/Makefile" "$KERNEL_SRCDIR"
+
 	[[ "$DEBUG" -eq 0 ]] && rm -rf "$TEMPDIR"
 	rm -rf "$RPMTOPDIR"
 	unset KCFLAGS
@@ -1054,6 +1060,35 @@ if [[ -n "$CONFIG_DEBUG_INFO_BTF" ]]; then
 	fi
 fi
 
+if [[ -n "$CONFIG_LTO_CLANG" ]]; then
+	[[ -n "$CONFIG_LTO_CLANG_THIN" ]] || die "Only thin lto is supported. Try enable CONFIG_LTO_CLANG_THIN."
+
+	# With CONFIG_LTO_CLANG, vmlinux has LOCAL symbols for some GLOBAL
+	# functions. For example:
+	#
+	# readelf -s vmlinux  | grep perf_event_bpf_event
+	#
+	# non-LTO:
+	#  124946: ffffffff81287070   858 FUNC    GLOBAL DEFAULT    1 perf_event_bpf_event
+	#
+	# LTO:
+	#  122028: ffffffff81298cd0   861 FUNC    LOCAL  HIDDEN     1 perf_event_bpf_event
+	#
+	# To work around this, use vmlinux.o instead.
+	[[ -e "$VMLINUX" ]] || die "For kernel with CONFIG_LTO_CLANG, please supply vmlinux.o via -v|--vmlinux option."
+
+	# This is a heuristic: use -x to check vmlinux vs. vmlinux.o
+	[[ -x "$VMLINUX" ]] && die "For kernel with CONFIG_LTO_CLANG, please supply vmlinux.o instead of vmlinux via -v|--vmlinux option."
+
+	cp -f "$KERNEL_SRCDIR/Makefile" "$TEMPDIR/Makefile" || die
+	sed -i "s/--thinlto-cache-dir=\$(extmod_prefix).thinlto-cache/--lto-obj-path=vmlinux.o.thinlto.o/g" "$KERNEL_SRCDIR"/Makefile
+
+	cp -f "$KERNEL_SRCDIR/scripts/Makefile.build" "$TEMPDIR/Makefile.build" || die
+	sed -i "s/\$(ld_flags)/\$(ld_flags) --lto-obj-path=\$@.thinlto.o/g" "$KERNEL_SRCDIR"/scripts/Makefile.build
+
+	export KPATCH_CC_HANDLE_LTO=1
+fi
+
 if [[ -n "$CONFIG_CC_IS_CLANG" ]]; then
 	echo "WARNING: Clang support is experimental"
 fi
@@ -1107,6 +1142,7 @@ declare -a MAKEVARS
 if [[ -n "$CONFIG_CC_IS_CLANG" ]]; then
 	MAKEVARS+=("CC=${KPATCH_CC_PREFIX}${CLANG}")
 	MAKEVARS+=("HOSTCC=clang")
+	MAKEVARS+=("LLVM=1")
 else
 	MAKEVARS+=("CC=${KPATCH_CC_PREFIX}${GCC}")
 fi
@@ -1174,9 +1210,14 @@ if [[ -n "$CONFIG_MODVERSIONS" ]]; then
 	trace_on
 fi
 
+if [[ -n "$CONFIG_LTO_CLANG" ]]; then
+	DIFF_OBJS="$TEMPDIR/thinlto_objs"
+else
+	DIFF_OBJS="$TEMPDIR/changed_objs"
+fi
 # Read as words, no quotes.
 # shellcheck disable=SC2013
-for i in $(cat "$TEMPDIR/changed_objs")
+for i in $(cat "$DIFF_OBJS")
 do
 	mkdir -p "$TEMPDIR/patched/$(dirname "$i")" || die
 	cp -f "$BUILDDIR/$i" "$TEMPDIR/patched/$i" || die
@@ -1205,7 +1246,8 @@ if [[ -z "$MODNAME" ]] ; then
 
 	MODNAME="$(module_name_string "$MODNAME")"
 fi
-FILES="$(cat "$TEMPDIR/changed_objs")"
+FILES="$(cat "$DIFF_OBJS")"
+
 cd "$TEMPDIR" || die
 mkdir output
 declare -a objnames
@@ -1229,7 +1271,11 @@ for i in $FILES; do
 
 	mkdir -p "output/$(dirname "$i")"
 	cd "$BUILDDIR" || die
-	find_kobj "$i"
+	if [[ -z "$CONFIG_LTO_CLANG" ]] ; then
+		find_kobj "$i"
+	else
+		KOBJFILE=${i/.o.thinlto.o*/}
+	fi
 	cd "$TEMPDIR" || die
 	if [[ -e "orig/$i" ]]; then
 		if [[ -n $OOT_MODULE ]]; then
@@ -1238,6 +1284,19 @@ for i in $FILES; do
 			KOBJFILE_PATH="$OOT_MODULE"
 			SYMTAB="${TEMPDIR}/module/${KOBJFILE_NAME}.symtab"
 			SYMVERS_FILE="$TEMPDIR/Module.symvers"
+		elif [[ -n "$CONFIG_LTO_CLANG" ]] ; then
+			if [[ $KOBJFILE = vmlinux ]] ; then
+				KOBJFILE_NAME=vmlinux
+				KOBJFILE_PATH="$VMLINUX"
+				SYMTAB="${TEMPDIR}/${KOBJFILE_NAME}.symtab"
+				SYMVERS_FILE="$BUILDDIR/Module.symvers"
+			else
+				KOBJFILE_NAME=$(basename "${KOBJFILE%.ko}")
+				KOBJFILE_NAME="${KOBJFILE_NAME//-/_}"
+				KOBJFILE_PATH="${TEMPDIR}/module/$KOBJFILE.ko"
+				SYMTAB="${KOBJFILE_PATH}.symtab"
+				SYMVERS_FILE="$BUILDDIR/Module.symvers"
+			fi
 		elif [[ "$(basename "$KOBJFILE")" = vmlinux ]]; then
 			KOBJFILE_NAME=vmlinux
 			KOBJFILE_PATH="$VMLINUX"
@@ -1251,11 +1310,21 @@ for i in $FILES; do
 			SYMVERS_FILE="$BUILDDIR/Module.symvers"
 		fi
 
-		"$READELF" -s --wide "$KOBJFILE_PATH" > "$SYMTAB"
+		# With CONFIG_LTO_CLANG, multiple .thinlto files share a
+		# symtab file. Only generate the symtab file once.
+		[[ -e "$SYMTAB" ]] || "$READELF" -s --wide "$KOBJFILE_PATH" > "$SYMTAB"
 		if [[ "$ARCH" = "ppc64le" ]]; then
 			sed -ri 's/\s+\[<localentry>: 8\]//' "$SYMTAB"
 		fi
 
+		if [[ -n "$CONFIG_LTO_CLANG" ]] ; then
+			# skip .thinlto file that didn't change at all
+			diff "orig/$i" "patched/$i" 2> /dev/null && continue
+			# skip .thinlto file without any functions
+			num_func=$("$READELF" --symbols "orig/$i" | grep -c FUNC)
+			[[ $num_func -eq 0 ]] && continue
+		fi
+
 		# create-diff-object orig.o patched.o parent-name parent-symtab
 		#		     Module.symvers patch-mod-name output.o
 		"$TOOLSDIR"/create-diff-object $CDO_FLAGS "orig/$i" "patched/$i" "$KOBJFILE_NAME" \
@@ -1308,7 +1377,7 @@ fi
 cd "$TEMPDIR/output" || die
 # $KPATCH_LDFLAGS and result of find used as list, no quotes.
 # shellcheck disable=SC2086,SC2046
-"$LD" -r $KPATCH_LDFLAGS -o ../patch/tmp_output.o $(find . -name "*.o") 2>&1 | logger || die
+"$LD" -r $KPATCH_LDFLAGS -o ../patch/tmp_output.o $(find . -name "*.o*") 2>&1 | logger || die
 
 if [[ "$USE_KLP" -eq 1 ]]; then
 	cp -f "$TEMPDIR"/patch/tmp_output.o "$TEMPDIR"/patch/output.o || die

kpatch-build/kpatch-cc

@@ -79,6 +79,24 @@ elif [[ "$TOOLCHAINCMD" =~ ^(.*-)?ld || "$TOOLCHAINCMD" =~ ^(.*-)?ld.lld ]] ; th
 					args+=(--warn-unresolved-symbols)
 					break
 					;;
+				*/.tmp_*.o)
+					# .tmp_*.o is used for single file modules.
+					# See "cmd_ld_single_m" in scripts/Makefile.build.
+					if [[ $KPATCH_CC_HANDLE_LTO -ne 0 ]] ; then
+						mkdir -p "$KPATCH_GCC_TEMPDIR/orig/$(dirname "$relobj")"
+						cp "${obj/.tmp_/}".thinlto.o* "$KPATCH_GCC_TEMPDIR/orig/$(dirname "$relobj")"
+						echo "${obj/.tmp_/}".thinlto.o* >> "$KPATCH_GCC_TEMPDIR/thinlto_objs"
+					fi
+					break
+					;;
+				*.o)
+					if [[ $KPATCH_CC_HANDLE_LTO -ne 0 ]] ; then
+						mkdir -p "$KPATCH_GCC_TEMPDIR/orig/$(dirname "$relobj")"
+						cp "$obj".thinlto* "$KPATCH_GCC_TEMPDIR/orig/$(dirname "$relobj")"
+						echo "$obj".thinlto.o* >> "$KPATCH_GCC_TEMPDIR/thinlto_objs"
+					fi
+					break
+					;;
 				*)
 					break
 					;;