[DRBD-user] drbd 9.0.26 and newer fails to build on 5.4 kernels

[Natanael Copa]

On Fri, 19 Feb 2021 19:50:13 +0100
Natanael Copa <ncopa at alpinelinux.org> wrote:

> On Fri, 19 Feb 2021 19:38:53 +0100
> Natanael Copa <ncopa at alpinelinux.org> wrote:
> 
> > Hi,
> > 
> > I tried to update the kernel for alpine 3.12-stable branch from 5.4.84
> > to 5.4.99. The 3rd part kernel module drbd 9.0.22-2 failed to build so
> > I updated it to 9.0.27-1. This passed on the x86_64 machine I tested
> > built it on so I pushed it.
> > 
> > But it failed on 32 bit arm builders:
> > 
> > ...
> > In file included from ./include/linux/module.h:27,
> >                  from /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_req.h:16,
> >                  from /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_state.c:22:
> > ./arch/arm/include/asm/module.h:59: warning: "MODULE_ARCH_VERMAGIC" redefined
> >    59 | #define MODULE_ARCH_VERMAGIC \
> >       | 
> > In file included from /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_state.c:19:
> > ./include/linux/vermagic.h:28: note: this is the location of the previous definition
> >    28 | #define MODULE_ARCH_VERMAGIC ""
> >       | 
> >   GEN     /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_buildtag.c 
> >   CC [M]  /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_buildtag.o
> >   LD [M]  /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd.o
> >   Building modules, stage 2.
> >   MODPOST 2 modules
> > ERROR: "__aeabi_ldivmod" [/home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd.ko] undefined!
> > ERROR: "__aeabi_uldivmod" [/home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd.ko] undefined!
> > make[3]: *** [scripts/Makefile.modpost:94: __modpost] Error 1
> > make[2]: *** [Makefile:1639: modules] Error 2
> > make[1]: *** [Makefile:132: kbuild] Error 2
> > make[1]: Leaving directory '/home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd'
> > make: *** [Makefile:131: module] Error 2  
> > >>> ERROR: drbd-lts: build failed  
> > 
> > And on 32 bit x86:
> > 
> >   GEN     /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_buildtag.c 
> >   CC [M]  /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd_buildtag.o
> >   LD [M]  /home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd.o
> >   Building modules, stage 2.
> >   MODPOST 2 modules
> > ERROR: "__udivdi3" [/home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd.ko] undefined!
> > ERROR: "__divdi3" [/home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd/drbd.ko] undefined!
> > make[3]: *** [scripts/Makefile.modpost:94: __modpost] Error 1
> > make[2]: *** [Makefile:1639: modules] Error 2
> > make[1]: *** [Makefile:132: kbuild] Error 2
> > make[1]: Leaving directory '/home/buildozer/aports/main/drbd-lts/src/drbd-9.0.27-1/drbd'
> > make: *** [Makefile:131: module] Error 2  
> > >>> ERROR: drbd-lts: build failed  
> > 
> > I read that this can happen if do_div() is not used and libgcc is not
> > linked in (which I assume we shouldnt on kernel). I tried to look at
> > the git log if I could find anything evident but couldn't find anything.
> > 
> > So I tried to find older version. Turns out that the lastest version
> > that compiles with 5.4 kernel is 9.0.25-2, but it only compiles with
> > 5.4.84 kernel and not the 5.4.99.
> > 
> > Does anyone have a clue how to solve this? If not, I guess I have to
> > try git bisect it.  


I found the commit that introduces the issue using objdump and git log:

commit 8dc8ede32de4410e99148b39d0f960e975eeddea
Author: Joel Colledge <joel.colledge at linbit.com>
Date:   Wed Nov 11 18:11:26 2020 +0100

    drbd: fix slow sync when sync requests are answered quickly


It is the DIV_ROUND_UP in that commit that introduces this.

This change fixes it:
diff --git a/drbd/drbd_sender.c b/drbd/drbd_sender.c
index 30e301fb..34030870 100644
--- a/drbd/drbd_sender.c
+++ b/drbd/drbd_sender.c
@@ -687,7 +687,7 @@ static int drbd_resync_delay(struct drbd_peer_device *peer_device)
                         * that the rate limiting prevents any new requests
                         * from being made. Wait just long enough so that we
                         * can request some data next time. */
-                       delay = DIV_ROUND_UP(HZ * BM_SECT_PER_BIT, pdc->c_max_rate * 2);
+                       delay = DIV_ROUND_UP((unsigned long)(HZ * BM_SECT_PER_BIT / 2), pdc->c_max_rate);
                }
        } else {
                /* Fixed resync rate. Use the standard delay. */


As I understand, it is the compiler that converts the HZ *
BM_SECT_PER_BIT to 64 bit type which triggers the error.

As I understand the source BM_SECT_PER_BIT = 1 << (12-9) = 8, so it
should be safe to move the `* 2` on the right side to a `/ 2` on the
left, which results in HZ * 4 / pdc->c_max_rate.

HZ is to my understand seldomly (never?) set to anything above 1000?
And 1000 * 4 fits just fine in an unsigned long, so I think the above diff should be safe.

-nc