<div dir="ltr">Hi all,<br><br>I thought that perhaps I'd just run into a weird edge-case being on Debian and a little bleeding edge after seeing some other folks on the mailing list compile without issue, but it looks like after some experimental work on centos trying to get Lustre 2.14 and ZFS 2.0 compiled on the same box that similar weirdness exists in the configuration steps where the autoconf logic has a hard time finding function definitions thanks to the new BSD+Linux source tree of ZFS. I'm going to revisit this and see if I can't generate some patches that make the process a bit easier.<br><br>For my own reference, are there any more recent sources-of-truth for compiling lustre from source than <a href="https://wiki.lustre.org/Installing_the_Lustre_Software">the one on the wiki?</a> There's every likelihood that I'm just missing some steps in the compilation process.<br><br>Cheers,<br>Christian</div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Mon, Feb 22, 2021 at 7:41 PM Christian Kuntz <<a href="mailto:c.kuntz@opendrives.com">c.kuntz@opendrives.com</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div>Hi Olaf,<br><br>Sure! For reference this is how I build zfs:<br><br>./autogen.sh<br>fakeroot dpkg-buildpackage -us -uc -b -j$cpus<br># Build the kernel modules, not just dkms<br>KVERS=$KERNEL_ABI KSRC=$WITHLINUX_HDR KOBJ=$WITHLINUXOBJ fakeroot debian/rules override_dh_binary-modules<br><br>Then, when I build lustre, I do like:<br>sh autogen.sh<br>./configure --with-linux=$WITHLINUX_HDR --with-linux-obj=$WITHLINUXOBJ --with-zfs=$ZFS_SRC --enable-server --enable-modules --disable-ldiskfs<br>make debs<br><br>My zfs source is given as the root zfs source directory, not the contents of explicit package dirs in ZFS_SRC/debian/ which is technically where all the build artifacts live (for instance, the development headers that would end up on the machine would be in debian/libzfslinux-dev/usr/include) after a package build. After fiddling around with autoconf a bit (and learning it's just shell!), I landed on changing this bit of the autoconf:<div>config/lustre-build-zfs.m4<br><br>AC_DEFUN([LB_ZFS_USER], [<br> dnl #<br> dnl # Detect user space zfs development headers.<br> dnl #<br> AC_MSG_CHECKING([zfs devel headers])<br> AS_IF([test -z "${zfsinc}"], [<br> AS_IF([test -d "${zfssrc}/include/os/linux"], [<br> zfsinc="-I${zfssrc}/include/os/linux/kernel -I${zfssrc}/include/os/linux/spl -I${zfssrc}/include/os/linux/zfs -I${zfssrc}/include"<br> zfslib="-L$zfssrc/lib/libzfs/.libs/ -L$zfssrc/lib/libnvpair/.libs"<br> ], [test -e "${zfssrc}/include/libzfs.h" && test -e "${zfssrc}/lib/libspl/include"], [<br> zfsinc="-I $zfssrc/lib/libspl/include -I $zfssrc/include"<br> zfslib="-L$zfssrc/lib/libzfs/.libs/ -L$zfssrc/lib/libnvpair/.libs"<br> ], [test -d /usr/include/libzfs && test -d /usr/include/libspl], [<br> zfsinc="-I/usr/include/libspl -I /usr/include/libzfs"<br> zfslib=""<br> ], [<br> zfsinc="[Not Found]"<br> zfslib=""<br> enable_zfs=no<br> ])<br> ])<br> AC_MSG_RESULT([$zfsinc])<br><br>I added a check around the new include/os style of the zfs header directories, which seems to be working OK, every config test is behaving as expected on my machine except for the multihost support one, which I'm chasing down a bit (sys/spa.h requires sys/sysmacros.h which requires... and so on). I also found that I had to modify the contents of config/lustre-build-linux.m4 to reflect the include dir change:<br><br>AC_DEFUN([LB_LINUX_COMPILE_IFELSE],<br>[m4_ifvaln([$1], [AC_LANG_CONFTEST([AC_LANG_SOURCE([$1])])])dnl<br>rm -f build/conftest.o build/conftest.mod.c build/conftest.ko<br>SUBARCH=$(echo $target_cpu | sed -e 's/powerpc.*/powerpc/' -e 's/ppc.*/powerpc/' -e 's/x86_64/x86/' -e 's/i.86/x86/' -e 's/k1om/x86/' -e 's/aarch64.*/arm64/' -e 's/armv7.*/arm/')<br>AS_IF([AC_TRY_COMMAND(cp conftest.c build && make -d [$2] LDFLAGS= ${LD:+LD="$LD"} CC="$CC" -f $PWD/build/Makefile LUSTRE_LINUX_CONFIG=$LINUX_CONFIG LINUXINCLUDE="$EXTRA_CHECK_INCLUDE -I$LINUX/arch/$SUBARCH/include -Iinclude -Iarch/$SUBARCH/include/generated -I$LINUX/include -Iinclude2 -I$LINUX/include/uapi -Iinclude/generated -I$LINUX/arch/$SUBARCH/include/uapi -Iarch/$SUBARCH/include/generated/uapi -I$LINUX/include/uapi -Iinclude/generated/uapi ${SPL_OBJ:+-include $SPL_OBJ/spl_config.h} ${ZFS_OBJ:+-include $ZFS_OBJ/zfs_config.h} ${SPL:+-I$ZFS/lib/libspl/include/ } ${ZFS:+-I$ZFS/include/os/linux/ -I$ZFS/include/os/linux/kernel -I$ZFS/include/os/linux/spl -I$ZFS/include/os/linux/zfs -I$ZFS/include } -include $CONFIG_INCLUDE" KBUILD_EXTRA_SYMBOLS="${ZFS_OBJ:+$ZFS_OBJ/Module.symvers} $KBUILD_EXTRA_SYMBOLS" -o tmp_include_depends -o scripts -o include/config/MARKER -C $LINUX_OBJ EXTRA_CFLAGS="-Werror-implicit-function-declaration $EXTRA_KCFLAGS" $MODULE_TARGET=$PWD/build) >/dev/null && AC_TRY_COMMAND([$3])], <br><br>I <i>think</i> I'll need to adjust the SPL variable in there, but not sure thus far. <br><br>With the above settings I've gotten everything to succesfully compile and package up aside from a small modification I had to make to lustre/utils/libmount_utils_zfs.c to prevent a name collision with spa_get_hostid. I believe this occurs because the configuration doesn't set HAVE_ZFS_MULTIHOST appropriately.<br><br><br>As you can see, I worked on it a little on the weekend :)<br><br>Cheers,<br>Christian<br></div></div></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Mon, Feb 22, 2021 at 11:58 AM Faaland, Olaf P. <<a href="mailto:faaland1@llnl.gov" target="_blank">faaland1@llnl.gov</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex">Hi Christian,<br>
<br>
Can you explain in more detail what you mean by "building ... from source-dirs after a package build"? And can you identify the lines in config/lustre-build-zfs.m4 that you're referring to?<br>
<br>
thank you,<br>
-Olaf<br>
<br>
________________________________________<br>
From: Christian Kuntz <<a href="mailto:c.kuntz@opendrives.com" target="_blank">c.kuntz@opendrives.com</a>><br>
Sent: Friday, February 19, 2021 7:00 PM<br>
To: Andreas Dilger<br>
Cc: Lustre-devel; Behlendorf, Brian D.; Faaland, Olaf P.<br>
Subject: Re: [lustre-devel] Testing 2.14 on Debian + ZFS<br>
<br>
Thanks for the confirmation!<br>
<br>
After going in a few circles, I think I've got it figured out and would appreciate some pointers on how to proceed:<br>
<br>
To preface, the way I'm building is from source-dirs after a package build (which is admittedly kind of janky as I write it out now). It would appear that the configure scripting will look in "zfs/include", and thus miss the new "zfs/include/os/..." directories that hold some important header files, causing configuration variables to miss-set in config.h thus creating the compile-time problems of type mismatches and redefinitions I was seeing.<br>
<br>
I was able to confirm manually setting some of the configuration defs and adding the new header dirs to the makefile fixed the problem, but I'm a little confused as to how to proceed in patching. Since the configuration scripting and the makefiles seem to expect only a single include-dir of header files for both zfs and spl respectively, I'm wondering what the best approach to patching this behavior would be.<br>
<br>
Thanks for your time and consideration,<br>
Christian<br>
<br>
On Fri, Feb 19, 2021 at 12:13 PM Andreas Dilger <<a href="mailto:adilger@whamcloud.com" target="_blank">adilger@whamcloud.com</a><mailto:<a href="mailto:adilger@whamcloud.com" target="_blank">adilger@whamcloud.com</a>>> wrote:<br>
This is exactly the right place for such discussions.<br>
<br>
Once you have patches, please submit to Gerrit for review, testing, and landing:<br>
<br>
<a href="https://wiki.lustre.org/Using_Gerrit" rel="noreferrer" target="_blank">https://wiki.lustre.org/Using_Gerrit</a><<a href="https://urldefense.us/v2/url?u=https-3A__wiki.lustre.org_Using-5FGerrit&d=DwMFaQ&c=pKoAVQro6qDbLoK0T8588B4mZJhJhC4e6QXJy0XnJec&r=2dxT0LLEUx0IbhFF8HRJVCTO_ihbEwKLn9gF7Kc8aG0&m=SU5sac04cekvaRvbUcheoU02W7EUu-F4sVkkrWHUnoE&s=gTcuGMUi5PycREbMB4La1s0dVNM8a5STb9pDVXZt-Lw&e=" rel="noreferrer" target="_blank">https://urldefense.us/v2/url?u=https-3A__wiki.lustre.org_Using-5FGerrit&d=DwMFaQ&c=pKoAVQro6qDbLoK0T8588B4mZJhJhC4e6QXJy0XnJec&r=2dxT0LLEUx0IbhFF8HRJVCTO_ihbEwKLn9gF7Kc8aG0&m=SU5sac04cekvaRvbUcheoU02W7EUu-F4sVkkrWHUnoE&s=gTcuGMUi5PycREbMB4La1s0dVNM8a5STb9pDVXZt-Lw&e=</a>><br>
<br>
Cheers, Andreas<br>
<br>
On Feb 19, 2021, at 02:11, Christian Kuntz <<a href="mailto:c.kuntz@opendrives.com" target="_blank">c.kuntz@opendrives.com</a><mailto:<a href="mailto:c.kuntz@opendrives.com" target="_blank">c.kuntz@opendrives.com</a>>> wrote:<br>
<br>
<br>
Thanks for your time and the thoughtful response.<br>
<br>
Looking at this a bit more, it seems to be caused by the recent ZFS 2.0 changes for linux and freebsd support, namely the `include/os/linux` paths that provide `<sys/types.h>` that ZFS is trying to use. Adding them into the includes for the osd-zfs directory has fixed the references being broken, and now I'm contending some redefinitions and argument type mismatches. I'll try and sort this out, in the meantime is there a better place for me to place messages about this? I know this mailing list isn't used very frequently and I'd hate to flood it with such off-the-beaten-path stuff.<br>
<br>
Best,<br>
Christian<br>
<br>
On Thu, Feb 18, 2021 at 3:09 AM Andreas Dilger <<a href="mailto:adilger@whamcloud.com" target="_blank">adilger@whamcloud.com</a><mailto:<a href="mailto:adilger@whamcloud.com" target="_blank">adilger@whamcloud.com</a>>> wrote:<br>
On Feb 12, 2021, at 16:07, Christian Kuntz <<a href="mailto:c.kuntz@opendrives.com" target="_blank">c.kuntz@opendrives.com</a><mailto:<a href="mailto:c.kuntz@opendrives.com" target="_blank">c.kuntz@opendrives.com</a>>> wrote:<br>
<br>
Hello,<br>
<br>
I hope I'm communicating in the right place here, I'm currently working to compile Lustre 2.14.0-RC2 on Debian 10.7 with ZFS 2.0.2 for OSDs. If there's anything I can do to help with the testing effort or help Lustre's Debian support be more robust, please let me know! I hope I'm not too late in the release cycle to contribute.<br>
<br>
Thanks for the offer of testing. Unfortunately we are very late in the 2.14.0 release cycle (we are hoping to make the final release in the next week or so, barring any critical defects). However, even if any changes miss the final 2.14.0 release, having patches available to fix any issues you run into will make it easier for the next person to build on Debian.<br>
<br>
Ideally, any required fixes for Debian building should go into the Lustre tree directly rather than into a separate "downstream" package, so that "make debs" works out-of-the-box. We do build Ubuntu 18.04 and 20.04 clients regularly, and ZFS 2.0 servers on RHEL7/8, but I don't know if anyone is building the server code on Debian/Ubuntu on a regular basis.<br>
<br>
On the topic of compiling, I'm currently working to get debian packages made and running into a compilation issue that didn't seem to present on my previous endeavors with 2.13 and 0.7. I've got the kernel and zfs locally built and their packages made, and lustre's configure step successfully completes, but when it comes to compiling it appears that the osd-zfs portion has some broken include links:<br>
<br>
In file included from /zfs-linux-2.0.2/include/sys/arc.h:32,<br>
from /lustre-debian/lustre/osd-zfs/osd_internal.h:51,<br>
from /lustre-debian/lustre/osd-zfs/osd_handler.c:52:<br>
/zfs-linux-2.0.2/include/sys/zfs_context.h:45:10: fatal error: sys/types.h: No such file or directory<br>
#include <sys/types.h><br>
<br>
For reference, my configure and make looks like:<br>
./configure --with-linux=/linux-4.19.171-2/ \<br>
--with-linux-obj=linux-4.19.171-2/debian/build/build_amd64_none_amd64 \<br>
--with-zfs=/zfs-linux-2.0.2 \<br>
--enable-server --enable-modules --disable-ldiskfs<br>
make -j"$(nproc)" debs<br>
<br>
<br>
>From the error and some probing, it looks like zfs_context.h expects some header files to be in their libc standard `/usr/include/sys` (or somewhere abouts), but lustre's make process is pulling them from the linux sources in `SRC/include/linux/`. Predictably, I found this out the hard way by adding `-I/usr/include` to the Makefile for the zfs osds and getting battered with errors and warnings about duplicate definitions.<br>
<br>
It may be that you are looking at this from the wrong angle. Rather than trying to include more userspace headers, it might be worthwhile to see why <sys/arc.h> is being included by osd_internal.h and maybe there is a more appropriate header that could be used in the kernel? Failing that, it isn't clear wh zfs_context.h is including the <sys/type.h> header from userspace when building with __KERNEL__ set? Kernel code should be using <linux/types.h>.<br>
<br>
Cheers, Andreas<br>
--<br>
Andreas Dilger<br>
Principal Lustre Architect<br>
Whamcloud<br>
<br>
<br>
<br>
<br>
<br>
<br>
</blockquote></div>
</blockquote></div>