[lustre-discuss] Issues of building Lustre-2.7.0 for vanilla Linux kernel

[Tung-Han Hsieh]

Dear All,

Recently we tried to build Lustre-2.7.0 server for vanilla Linux kernel
version 5, plus zfs-2.3.4, under Debian-12.13 Linux system with Mellanox
Infiniband enabled. There are some issues to report. The building options
are fixed to:

- for zfs-2.3.4:
  ./configure --prefix=/opt/lustre --with-linux=/usr/src/linux-<version>
- for lustre-2.7.0:
  export KBUILD_EXTRA_SYMBOLS=/usr/src/zfs-2.3.4/module/Module.symvers
  ./configure --prefix=/opt/lustre --with-linux=/usr/src/linux-<version> \
                    --with-zfs=/usr/src/zfs-2.3.4
--with-o2ib=yes --disable-ldiskfs \
                    --disable-tests --enable-mpitests=no

1. For vanilla Linux kernel 5.4.279:
    During compilation, we encounter the following errors:
=====================================================================
CC [M]  /usr/src/lustre-2.17.0/libcfs/libcfs/crypto/keyring.o
/usr/src/lustre-2.17.0/libcfs/libcfs/crypto/keyring.c: In function
free_master_key:
/usr/src/lustre-2.17.0/libcfs/libcfs/crypto/keyring.c:55:9: error: implicit
declaration of function kfree_sensitive; did you mean kvfree_sensitive?
[-Werror=implicit-function-declaration]
   55 |         kfree_sensitive(mk);
      |         ^~~~~~~~~~~~~~~
      |         kvfree_sensitive
cc1: all warnings being treated as errors
=====================================================================

    After debug, we have to add the following code:
=====================================================================
// In the beginning of the source file, and after all the #include files.
#ifndef HAVE_KFREE_SENSITIVE
#define kfree_sensitive(x)      kzfree(x)
#endif
=====================================================================

    into the following source files:

    - lustre-2.7.0/libcfs/libcfs/crypto/keyring.c
    - lustre-2.7.0/libcfs/libcfs/crypto/keysetup_v1.c

    to work around this problem, since kfree_sensitive() seems not
available in vanilla Linux kernel 5.4.X.

2. For vanilla Linux kernel 5.10.252:
    During compilation, we encounter the following errors:
=====================================================================
  CC [M]  /usr/src/lustre-2.17.0/lustre/llite/pcc.o
/usr/src/lustre-2.17.0/lustre/llite/pcc.c: In function
__pcc_file_reset_projid:
/usr/src/lustre-2.17.0/lustre/llite/pcc.c:3491:18: error: implicit
declaration of function get_fs; did you mean sget_fc?
[-Werror=implicit-function-declaration]
 3491 |         old_fs = get_fs();
      |                  ^~~~~~
      |                  sget_fc
/usr/src/lustre-2.17.0/lustre/llite/pcc.c:3491:18: error: incompatible
types when assigning to type mm_segment_t from type int
/usr/src/lustre-2.17.0/lustre/llite/pcc.c:3492:9: error: implicit
declaration of function set_fs; did you mean sget_fc?
[-Werror=implicit-function-declaration]
 3492 |         set_fs(KERNEL_DS);
      |         ^~~~~~
      |         sget_fc
/usr/src/lustre-2.17.0/lustre/llite/pcc.c:3492:16: error: KERNEL_DS
undeclared (first use in this function); did you mean KERNFS_NS?
 3492 |         set_fs(KERNEL_DS);
      |                ^~~~~~~~~
      |                KERNFS_NS
/usr/src/lustre-2.17.0/lustre/llite/pcc.c:3492:16: note: each undeclared
identifier is reported only once for each function it appears in
cc1: all warnings being treated as errors
=====================================================================

    When looking into this problem, the static int
__pcc_file_reset_projid() function of the file:
lustre-2.17.0/lustre/lllite/pcc.c reads:
=====================================================================
#ifdef HAVE_FILEATTR_GET
        ....
        rc = inode->i_op->fileattr_set(&nop_mnt_idmap, dentry, &fa);
#else
        ....
        old_fs = get_fs();
        set_fs(KERNEL_DS);
        ....
#endif
=====================================================================
    It is awkward that, for vanilla Linux kernel 5.10.X, the old interface
get_fs() and set_fs() were removed. However, the new
interface fileattr_set() is not available. As a result, Lustre-2.17.0
cannot be built under vanilla Linux kernel 5.10.X, unless there is a
specific patch.

3. For vanilla Linux kernel 5.15.202:
    Lustre-2.17.0 can be built successfully. However, when loading lustre
modules:
    modprobe lustre
    modprobe lnet

    Both command encountered the following error message in dmesg:

=====================================================================
lustre_symbols_init: error -38
LNetError: 85487:0:(module.c:664:libcfs_init()) cfs_arch_init: error -38
=====================================================================

    I did not debug it carefully. But I suspect that the mechanism of
loading modules in kernel 5.15.X might be changed. So Lustre cannot start
in this environment.

So, currently Lustre-2.17.0 server with zfs-2.3.4 can only work in vanilla
kernel 5.4.X (I only tested 5.4.279). I hope that this report can be
considered to add patches for vanilla kernel 5.15.X, or even 5.10.X, since
newer releases of Linux kernel is needed for new hardware as the file
servers.

Best Regards,
T.H.Hsieh
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.lustre.org/pipermail/lustre-discuss-lustre.org/attachments/20260330/d03ed4ca/attachment.htm>