Fix MFS builds when both MD_ROOT_SIZE and MFS_IMAGE are specified

MD_ROOT_SIZE and embed_mfs.sh were basically retired as part of
https://reviews.freebsd.org/D2903 .
However, when building a kernel with 'options MD_ROOT_SIZE' specified, this
results in a non-working MFS, as within sys/dev/md/md.c we fall within the
wrong # ifdef.

This patch implements the following:

* Allow kernels to be built without the MD_ROOT_SIZE option, which results
  in a kernel built as per D2903.
* Allow kernels to be built with the MD_ROOT_SIZE option, which results
  in a kernel built similarly to the pre-D2903 way, with the following
  differences:
  * The MFS is now put in a separate section within the kernel (oldmfs,
    so it differs from the mfs section introduced by D2903).
  * embed_mfs.sh is changed, so it looks up the oldmfs section within the
    kernel, gets its size and offset, sees if the MFS will fit within the
    allocated oldmfs section and only if all is well does a dd of the MFS
    image into the kernel.

Submitted by:	Stanislav Galabov <sgalabov@gmail.com>
Reviewed by:	brooks, imp
Differential Revision:	https://reviews.freebsd.org/D5093

sys/conf/NOTES

@@ -1078,6 +1078,14 @@ options 	UFS_GJOURNAL
 
 # Make space in the kernel for a root filesystem on a md device.
 # Define to the number of kilobytes to reserve for the filesystem.
+# This is now optional.
+# If not defined, the root filesystem passed in as the MFS_IMAGE makeoption
+# will be automatically embedded in the kernel during linking. Its exact size
+# will be consumed within the kernel.
+# If defined, the old way of embedding the filesystem in the kernel will be
+# used. That is to say MD_ROOT_SIZE KB will be allocated in the kernel and
+# later, the filesystem image passed in as the MFS_IMAGE makeoption will be
+# dd'd into the reserved space if it fits.
 options 	MD_ROOT_SIZE=10
 
 # Make the md device a potential root device, either with preloaded

sys/conf/kern.post.mk

@@ -130,6 +130,9 @@ ${FULLKERNEL}: ${SYSTEM_DEP} vers.o
 	@rm -f ${.TARGET}
 	@echo linking ${.TARGET}
 	${SYSTEM_LD}
+.if !empty(MD_ROOT_SIZE_CONFIGURED) && defined(MFS_IMAGE)
+	@sh ${S}/tools/embed_mfs.sh ${.TARGET} ${MFS_IMAGE}
+.endif
 .if ${MK_CTF} != "no"
 	@echo ${CTFMERGE} ${CTFFLAGS} -o ${.TARGET} ...
 	@${CTFMERGE} ${CTFFLAGS} -o ${.TARGET} ${SYSTEM_OBJS} vers.o
@@ -353,6 +356,7 @@ vnode_if_typedef.h:
 	${AWK} -f $S/tools/vnode_if.awk $S/kern/vnode_if.src -q
 
 .if ${MFS_IMAGE:Uno} != "no"
+.if empty(MD_ROOT_SIZE_CONFIGURED)
 # Generate an object file from the file system image to embed in the kernel
 # via linking. Make sure the contents are in the mfs section and rename the
 # start/end/size variables to __start_mfs, __stop_mfs, and mfs_size,
@@ -372,6 +376,7 @@ embedfs_${MFS_IMAGE:T:R}.o: ${MFS_IMAGE}
 		_binary_${MFS_IMAGE:C,[^[:alnum:]],_,g}_end=mfs_root_end \
 	    ${.TARGET}
 .endif
+.endif
 
 # XXX strictly, everything depends on Makefile because changes to ${PROF}
 # only appear there, but we don't handle that.

sys/conf/kern.pre.mk

@@ -195,9 +195,13 @@ SYSTEM_DEP= Makefile ${SYSTEM_OBJS}
 SYSTEM_OBJS= locore.o ${MDOBJS} ${OBJS}
 SYSTEM_OBJS+= ${SYSTEM_CFILES:.c=.o}
 SYSTEM_OBJS+= hack.So
+
+MD_ROOT_SIZE_CONFIGURED!=	grep MD_ROOT_SIZE opt_md.h || true ; echo
 .if ${MFS_IMAGE:Uno} != "no"
+.if empty(MD_ROOT_SIZE_CONFIGURED)
 SYSTEM_OBJS+= embedfs_${MFS_IMAGE:T:R}.o
 .endif
+.endif
 SYSTEM_LD= @${LD} -Bdynamic -T ${LDSCRIPT} ${_LDFLAGS} --no-warn-mismatch \
 	--warn-common --export-dynamic --dynamic-linker /red/herring \
 	-o ${.TARGET} -X ${SYSTEM_OBJS} vers.o
@@ -230,8 +234,9 @@ MKMODULESENV+=	__MPATH="${__MPATH}"
 
 # Architecture and output format arguments for objdump to convert image to
 # object file
-.if ${MFS_IMAGE:Uno} != "no"
 
+.if ${MFS_IMAGE:Uno} != "no"
+.if empty(MD_ROOT_SIZE_CONFIGURED)
 .if !defined(EMBEDFS_FORMAT.${MACHINE_ARCH})
 EMBEDFS_FORMAT.${MACHINE_ARCH}!= awk -F'"' '/OUTPUT_FORMAT/ {print $$2}' ${LDSCRIPT}
 .if empty(EMBEDFS_FORMAT.${MACHINE_ARCH})
@@ -254,6 +259,7 @@ EMBEDFS_FORMAT.mips64?=		elf64-tradbigmips
 EMBEDFS_FORMAT.mips64el?=	elf64-tradlittlemips
 EMBEDFS_FORMAT.riscv?=		elf64-littleriscv
 .endif
+.endif
 
 # Detect kernel config options that force stack frames to be turned on.
 DDB_ENABLED!=	grep DDB opt_ddb.h || true ; echo

sys/dev/md/md.c

@@ -130,18 +130,12 @@ SYSCTL_INT(_vm, OID_AUTO, md_malloc_wait, CTLFLAG_RW, &md_malloc_wait, 0,
  */
 #if defined(MD_ROOT_SIZE)
 /*
+ * We put the mfs_root symbol into the oldmfs section of the kernel object file.
  * Applications that patch the object with the image can determine
- * the size looking at the start and end markers (strings),
+ * the size looking at the oldmfs section size within the kernel.
- * so we want them contiguous.
  */
-static struct {
+u_char mfs_root[MD_ROOT_SIZE*1024] __attribute__ ((section ("oldmfs")));
-	u_char start[MD_ROOT_SIZE*1024];
+const int mfs_root_size = sizeof(mfs_root);
-	u_char end[128];
-} mfs_root = {
-	.start = "MFS Filesystem goes here",
-	.end = "MFS Filesystem had better STOP here",
-};
-const int mfs_root_size = sizeof(mfs_root.start);
 #else
 extern volatile u_char __weak_symbol mfs_root;
 extern volatile u_char __weak_symbol mfs_root_end;

sys/tools/embed_mfs.sh

@@ -32,8 +32,20 @@
 # $2: MFS image filename
 #
 
-obs=`strings -at d $1 | grep "MFS Filesystem goes here" | awk '{print $1}'`
+mfs_size=`stat -f '%z' $2 2> /dev/null`
-dd if=$2 ibs=8192 of=$1 obs=${obs} oseek=1 conv=notrunc 2> /dev/null
+# If we can't determine MFS image size - bail.
+[ -z ${mfs_size} ] && echo "Can't determine MFS image size" && exit 1
 
-strings $1 | grep 'MFS Filesystem had better STOP here' > /dev/null || \
+sec_info=`objdump -h $1 2> /dev/null | grep " oldmfs "`
-	(rm $1 && echo "MFS image too large" && false)
+# If we can't find the mfs section within the given kernel - bail.
+[ -z "${sec_info}" ] && echo "Can't locate mfs section within kernel" && exit 1
+
+sec_size=`echo ${sec_info} | awk '{printf("%d", "0x" $3)}' 2> /dev/null`
+sec_start=`echo ${sec_info} | awk '{printf("%d", "0x" $6)}' 2> /dev/null`
+
+# If the mfs section size is smaller than the mfs image - bail.
+[ ${sec_size} -lt ${mfs_size} ] && echo "MFS image too large" && exit 1
+
+# Dump the mfs image into the mfs section
+dd if=$2 ibs=8192 of=$1 obs=${sec_start} oseek=1 conv=notrunc 2> /dev/null && \
+    echo "MFS image embedded into kernel" && exit 0