Details
-
Bug
-
Resolution: Fixed
-
Blocker
-
Lustre 2.3.0
-
None
-
lustre version 2.2.93
x86_64 system
ARCH=x86_64
-
3
-
4471
Description
When building lustre (version tagged 2.2.93) on x86_64 system, with ARCH environment variable set to x86_64, it fails as described below.
# echo $ARCH
x86_64
# make rpms
...
make[2]: Entering directory `/home/build/lustre-release/libcfs'
(cd libcfs && make top_distdir=../../lustre-2.2.93 distdir=../../lustre-2.2.93/libcfs/libcfs \
am__remove_distdir=: am__skip_length_check=: am__skip_mode_fix=: distdir)
make[3]: *** No rule to make target `crc32-pclmul_asm.c', needed by `distdir'. Stop.
Maybe having ARCH environment variable set when building lustre is not standard, but anyway the makefiles should not rely on its value, or it should recompute it.
in libcfs/libcfs/Makefile.in
ifeq ($(ARCH),x86)
libcfs-linux-objs += linux-crypto-crc32pclmul.o
libcfs-pclmul-obj += crc32-pclmul_asm.o
endif
ifeq ($(ARCH),i386)
libcfs-linux-objs += linux-crypto-crc32pclmul.o
libcfs-pclmul-obj += crc32-pclmul_asm.o
endif
ifeq ($(ARCH),x86_64)
libcfs-linux-objs += linux-crypto-crc32pclmul.o
libcfs-pclmul-obj += crc32-pclmul_asm.o
endif
...
libcfs-all-objs := debug.o fail.o nidstrings.o lwt.o module.o tracefile.o \
watchdog.o libcfs_string.o hash.o kernel_user_comm.o \
prng.o workitem.o upcall_cache.o libcfs_cpu.o \
libcfs_mem.o libcfs_lock.o $(libcfs-pclmul-obj)
libcfs-objs := $(libcfs-linux-objs) $(libcfs-all-objs)
in libcfs/libcfs/autoMakefile.in
...
EXTRA_DIST := $(libcfs-all-objs:%.o=%.c) Info.plist tracefile.h prng.c \
user-lock.c user-tcpip.c user-bitops.c user-prim.c workitem.c \
user-mem.c kernel_user_comm.c fail.c libcfs_cpu.c \
libcfs_mem.c libcfs_lock.c linux/linux-tracefile.h
...
DISTFILES = $(DIST_COMMON) $(DIST_SOURCES) $(TEXINFOS) $(EXTRA_DIST)
...
distdir: $(DISTFILES)
...
When the directive "$(libcfs-all-objs:%.o=%.c)" is applied to crc32-pclmul_asm.o, it adds a dependency on crc32-pclmul_asm.c file which does not exist.
The problem does not appears when ARCH environment variable is not set because in that case libcfs-pclmul-obj is null.
So I propose to update libcfs/libcfs/Makefile and move libcfs-pclmul-obj from libcfs-all-objs to libcfs-objs.
libcfs-all-objs := debug.o fail.o nidstrings.o lwt.o module.o tracefile.o \
watchdog.o libcfs_string.o hash.o kernel_user_comm.o \
prng.o workitem.o upcall_cache.o libcfs_cpu.o \
libcfs_mem.o libcfs_lock.o
libcfs-objs := $(libcfs-linux-objs) $(libcfs-all-objs) $(libcfs-pclmul-obj)
Additionally, the ARCH variable should probably be replaced in libcfs/libcfs/Makefile.in.
Attachments
Activity
Alexander,
Do you plan to replace the ARCH variable in libcfs/libcfs/Makefile.in ?
This change does not appear in the patch #3801 you submitted.
If this pans out we will need a patch for 2.3 branch as well before it closes.
Builds now. Any test besides sanity77 and running IOR with checksums enabled.
No, I don`t plan this change. The ARCH is used at Makefile to check architecture and this variable came from Kbuild. The current issue happend for user mode build only, because kernel objects was put to user mode autoMakefile.