[PATCH v2 5/5] ARM: XIP kernel: store .data compressed in ROM

[nicolas.pitre@linaro.org (Nicolas Pitre)]

The .data segment stored in ROM is only copied to RAM once at boot time
and never referenced afterwards. This is arguably a suboptimal usage of
ROM resources.

This patch allows for compressing the .data segment before storing it
into ROM and decompressing it to RAM rather than simply copying it,
saving on precious ROM space.

Because global data is not available yet (obviously) we must allocate
decompressor workspace memory on the stack. The .bss area is used as a
stack area for that purpose before it is cleared. The required stack
frame is 9568 bytes for __inflate_kernel_data() alone, so make sure
the .bss is large enough to cope with that plus extra room for called
functions or fail the build.

Those numbers were picked arbitrarily based on the above 9568 byte
stack frame:

10240 (2.5 * PAGE_SIZE): used to override -Wframe-larger-than whose
default value is 1024.
12288 (3 * PAGE_SIZE): minimum .bss size to contain the stack.

Signed-off-by: Nicolas Pitre <nico@linaro.org>
Reviewed-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
---
 arch/arm/Kconfig                    | 11 +++++++
 arch/arm/boot/Makefile              | 13 +++++++-
 arch/arm/boot/deflate_xip_data.sh   | 64 +++++++++++++++++++++++++++++++++++++
 arch/arm/kernel/Makefile            |  5 +++
 arch/arm/kernel/head-common.S       | 11 ++++++-
 arch/arm/kernel/head-inflate-data.c | 62 +++++++++++++++++++++++++++++++++++
 arch/arm/kernel/vmlinux-xip.lds.S   |  8 +++++
 7 files changed, 172 insertions(+), 2 deletions(-)
 create mode 100755 arch/arm/boot/deflate_xip_data.sh
 create mode 100644 arch/arm/kernel/head-inflate-data.c

diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index 61a0cb1506..bf79c461bd 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -2005,6 +2005,17 @@ config XIP_PHYS_ADDR
 	  be linked for and stored to.  This address is dependent on your
 	  own flash usage.
 
+config XIP_DEFLATED_DATA
+	bool "Store kernel .data section compressed in ROM"
+	depends on XIP_KERNEL
+	select ZLIB_INFLATE
+	help
+	  Before the kernel is actually executed, its .data section has to be
+	  copied to RAM from ROM. This option allows for storing that data
+	  in compressed form and decompressed to RAM rather than merely being
+	  copied, saving some precious ROM space. A possible drawback is a
+	  slightly longer boot delay.
+
 config KEXEC
 	bool "Kexec system call (EXPERIMENTAL)"
 	depends on (!SMP || PM_SLEEP_SMP)
diff --git a/arch/arm/boot/Makefile b/arch/arm/boot/Makefile
index 50f8d1be7f..a3af4dc08c 100644
--- a/arch/arm/boot/Makefile
+++ b/arch/arm/boot/Makefile
@@ -31,8 +31,19 @@ targets := Image zImage xipImage bootpImage uImage
 
 ifeq ($(CONFIG_XIP_KERNEL),y)
 
+cmd_deflate_xip_data = $(CONFIG_SHELL) -c \
+	'$(srctree)/$(src)/deflate_xip_data.sh $< $@ || { rm -f $@; false; }'
+
+ifeq ($(CONFIG_XIP_DEFLATED_DATA),y)
+quiet_cmd_mkxip = XIPZ    $@
+cmd_mkxip = $(cmd_objcopy) && $(cmd_deflate_xip_data)
+else
+quiet_cmd_mkxip = $(quiet_cmd_objcopy)
+cmd_mkxip = $(cmd_objcopy)
+endif
+
 $(obj)/xipImage: vmlinux FORCE
-	$(call if_changed,objcopy)
+	$(call if_changed,mkxip)
 	@$(kecho) '  Physical Address of xipImage: $(CONFIG_XIP_PHYS_ADDR)'
 
 $(obj)/Image $(obj)/zImage: FORCE
diff --git a/arch/arm/boot/deflate_xip_data.sh b/arch/arm/boot/deflate_xip_data.sh
new file mode 100755
index 0000000000..1189598a25
--- /dev/null
+++ b/arch/arm/boot/deflate_xip_data.sh
@@ -0,0 +1,64 @@
+#!/bin/sh
+
+# XIP kernel .data segment compressor
+#
+# Created by:	Nicolas Pitre, August 2017
+# Copyright:	(C) 2017  Linaro Limited
+#
+# This program is free software; you can redistribute it and/or modify
+# it under the terms of the GNU General Public License version 2 as
+# published by the Free Software Foundation.
+
+# This script locates the start of the .data section in xipImage and
+# substitutes it with a compressed version. The needed offsets are obtained
+# from symbol addresses in vmlinux. It is expected that .data extends to
+# the end of xipImage.
+
+set -e
+
+VMLINUX="$1"
+XIPIMAGE="$2"
+
+DD="dd status=none"
+
+# Use "make V=1" to debug this script.
+case "$KBUILD_VERBOSE" in
+*1*)
+	set -x
+	;;
+esac
+
+sym_val() {
+	# extract hex value for symbol in $1
+	local val=$($NM "$VMLINUX" | sed -n "/ $1$/{s/ .*$//p;q}")
+	[ "$val" ] || { echo "can't find $1 in $VMLINUX" 1>&2; exit 1; }
+	# convert from hex to decimal
+	echo $((0x$val))
+}
+
+__data_loc=$(sym_val __data_loc)
+_edata_loc=$(sym_val _edata_loc)
+base_offset=$(sym_val _xiprom)
+
+# convert to file based offsets
+data_start=$(($__data_loc - $base_offset))
+data_end=$(($_edata_loc - $base_offset))
+
+# Make sure data occupies the last part of the file.
+file_end=$(stat -c "%s" "$XIPIMAGE")
+if [ "$file_end" != "$data_end" ]; then
+	printf "end of xipImage doesn't match with _edata_loc (%#x vs %#x)\n" \
+	       $(($file_end + $base_offset)) $_edata_loc 2>&1
+	exit 1;
+fi
+
+# be ready to clean up
+trap 'rm -f "$XIPIMAGE.tmp"' 0 1 2 3
+
+# substitute the data section by a compressed version
+$DD if="$XIPIMAGE" count=$data_start iflag=count_bytes of="$XIPIMAGE.tmp"
+$DD if="$XIPIMAGE"  skip=$data_start iflag=skip_bytes |
+gzip -9 >> "$XIPIMAGE.tmp"
+
+# replace kernel binary
+mv -f "$XIPIMAGE.tmp" "$XIPIMAGE"
diff --git a/arch/arm/kernel/Makefile b/arch/arm/kernel/Makefile
index ad325a8c7e..52f437997c 100644
--- a/arch/arm/kernel/Makefile
+++ b/arch/arm/kernel/Makefile
@@ -87,6 +87,11 @@ head-y			:= head$(MMUEXT).o
 obj-$(CONFIG_DEBUG_LL)	+= debug.o
 obj-$(CONFIG_EARLY_PRINTK)	+= early_printk.o
 
+# This is executed very early using a temporary stack when no memory allocator
+# nor global data is available. Everything has to be allocated on the stack.
+CFLAGS_head-inflate-data.o := $(call cc-option,-Wframe-larger-than=10240)
+obj-$(CONFIG_XIP_DEFLATED_DATA) += head-inflate-data.o
+
 obj-$(CONFIG_ARM_VIRT_EXT)	+= hyp-stub.o
 AFLAGS_hyp-stub.o		:=-Wa,-march=armv7-a
 ifeq ($(CONFIG_ARM_PSCI),y)
diff --git a/arch/arm/kernel/head-common.S b/arch/arm/kernel/head-common.S
index bf9c4e38ec..a25027b87a 100644
--- a/arch/arm/kernel/head-common.S
+++ b/arch/arm/kernel/head-common.S
@@ -87,7 +87,14 @@ __mmap_switched:
 	adr	r4, __mmap_switched_data
 	mov	fp, #0
 
-#ifdef CONFIG_XIP_KERNEL
+#if defined(CONFIG_XIP_DEFLATED_DATA)
+   ARM(	ldr	sp, [r4], #4 )
+ THUMB(	ldr	sp, [r4] )
+ THUMB(	add	r4, #4 )
+	bl	__inflate_kernel_data		@ decompress .data to RAM
+	teq	r0, #0
+	bne	__error
+#elif defined(CONFIG_XIP_KERNEL)
    ARM(	ldmia	r4!, {r0, r1, r2, sp} )
  THUMB(	ldmia	r4!, {r0, r1, r2, r3} )
  THUMB(	mov	sp, r3 )
@@ -114,9 +121,11 @@ ENDPROC(__mmap_switched)
 	.type	__mmap_switched_data, %object
 __mmap_switched_data:
 #ifdef CONFIG_XIP_KERNEL
+#ifndef CONFIG_XIP_DEFLATED_DATA
 	.long	_sdata				@ r0
 	.long	__data_loc			@ r1
 	.long	_edata_loc			@ r2
+#endif
 	.long	__bss_stop			@ sp (temporary stack in .bss)
 #endif
 
diff --git a/arch/arm/kernel/head-inflate-data.c b/arch/arm/kernel/head-inflate-data.c
new file mode 100644
index 0000000000..4598b959c9
--- /dev/null
+++ b/arch/arm/kernel/head-inflate-data.c
@@ -0,0 +1,62 @@
+/*
+ * XIP kernel .data segment decompressor
+ *
+ * Created by:	Nicolas Pitre, August 2017
+ * Copyright:	(C) 2017  Linaro Limited
+ *
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
+ */
+
+#include <linux/init.h>
+#include <linux/zutil.h>
+
+/* for struct inflate_state */
+#include "../../../lib/zlib_inflate/inftrees.h"
+#include "../../../lib/zlib_inflate/inflate.h"
+#include "../../../lib/zlib_inflate/infutil.h"
+
+extern char __data_loc[];
+extern char _edata_loc[];
+extern char _sdata[];
+
+/*
+ * This code is called very early during the boot process to decompress
+ * the .data segment stored compressed in ROM. Therefore none of the global
+ * variables are valid yet, hence no kernel services such as memory
+ * allocation is available. Everything must be allocated on the stack and
+ * we must avoid any global data access. We use a temporary stack located
+ * in the .bss area. The linker script makes sure the .bss is big enough
+ * to hold our stack frame plus some room for called functions.
+ *
+ * We mimic the code in lib/decompress_inflate.c to use the smallest work
+ * area possible. And because everything is statically allocated on the
+ * stack then there is no need to clean up before returning.
+ */
+
+int __init __inflate_kernel_data(void)
+{
+	struct z_stream_s stream, *strm = &stream;
+	struct inflate_state state;
+	char *in = __data_loc;
+	int rc;
+
+	/* Check and skip gzip header (assume no filename) */
+	if (in[0] != 0x1f || in[1] != 0x8b || in[2] != 0x08 || in[3] & ~3)
+		return -1;
+	in += 10;
+
+	strm->workspace = &state;
+	strm->next_in = in;
+	strm->avail_in = _edata_loc - __data_loc;  /* upper bound */
+	strm->next_out = _sdata;
+	strm->avail_out = _edata_loc - __data_loc;
+	zlib_inflateInit2(strm, -MAX_WBITS);
+	WS(strm)->inflate_state.wsize = 0;
+	WS(strm)->inflate_state.window = NULL;
+	rc = zlib_inflate(strm, Z_FINISH);
+	if (rc == Z_STREAM_END)
+		rc = strm->avail_out;  /* should be 0 */
+	return rc;
+}
diff --git a/arch/arm/kernel/vmlinux-xip.lds.S b/arch/arm/kernel/vmlinux-xip.lds.S
index 39b1fb470a..7a84431008 100644
--- a/arch/arm/kernel/vmlinux-xip.lds.S
+++ b/arch/arm/kernel/vmlinux-xip.lds.S
@@ -306,3 +306,11 @@ ASSERT((__arch_info_end - __arch_info_begin), "no machine record defined")
  */
 ASSERT(__hyp_idmap_text_end - (__hyp_idmap_text_start & PAGE_MASK) <= PAGE_SIZE,
 	"HYP init code too big or misaligned")
+
+#ifdef CONFIG_XIP_DEFLATED_DATA
+/*
+ * The .bss is used as a stack area for __inflate_kernel_data() whose stack
+ * frame is 9568 bytes. Make sure it has extra room left.
+ */
+ASSERT((_end - __bss_start) >= 12288, ".bss too small for CONFIG_XIP_DEFLATED_DATA")
+#endif
-- 
2.9.5