Dear Nathan,


Am 21.02.22 um 16:29 schrieb Nathan Chancellor:

On Mon, Feb 21, 2022 at 12:17:40PM +0100, Paul Menzel wrote:

quoted

Am 17.02.22 um 02:16 schrieb Nathan Chancellor:

quoted

On Wed, Feb 16, 2022 at 02:19:51PM +0100, Paul Menzel wrote:

quoted

[Cc: +LLVM/clang build support folks]

[…]

quoted

quoted

To recap: On a ppc64le machine, building Linux in Ubuntu 21.10 with *llvm*
and *clang* 1:13.0-53~exp1

      $ clang --version
      Ubuntu clang version 13.0.0-2
      Target: powerpc64le-unknown-linux-gnu
      Thread model: posix
      InstalledDir: /usr/bin

results in a segmentation fault, while it works when building with GCC.

      $ gcc --version
      gcc (Ubuntu 11.2.0-7ubuntu2) 11.2.0

Thank you for keying us in. I am going to have a bit of a brain dump
here based on the information I have uncovered after a couple of hours
of debugging.

TL;DR: It seems like something is broken with __read_mostly + ld.lld
before 14.0.0.

My initial reproduction steps (boot-qemu.sh comes from
https://github.com/ClangBuiltLinux/boot-utils):

$ clang --version
clang version 13.0.1 (Fedora 13.0.1-1.fc37)
Target: x86_64-redhat-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

$ powerpc64le-linux-gnu-as --version
GNU assembler version 2.37-2.fc36
Copyright (C) 2021 Free Software Foundation, Inc.
This program is free software; you may redistribute it under the terms of
the GNU General Public License version 3 or later.
This program has absolutely no warranty.
This assembler was configured for a target of `powerpc64le-linux-gnu'.

$ curl -LSso .config https://lore.kernel.org/all/f41550c7-26c0-cf81-7de9-aa924434a565@molgen.mpg.de/3-linux-5.17-rc4-rcu-dev-config.txt (local)

$ scripts/config --set-val INITRAMFS_SOURCE '""'

$ make -skj"$(nproc)" ARCH=powerpc CROSS_COMPILE=powerpc64le-linux-gnu- LLVM=1 LLVM_IAS=0 all

$ boot-qemu.sh -a ppc64le -k . -t 45s
QEMU location: /usr/bin

QEMU version: QEMU emulator version 6.2.0 (qemu-6.2.0-5.fc37)

+ timeout --foreground 45s stdbuf -oL -eL qemu-system-ppc64 -initrd \
/home/nathan/cbl/github/boot-utils-ro/images/ppc64le/rootfs.cpio -device \
ipmi-bmc-sim,id=bmc0 -device isa-ipmi-bt,bmc=bmc0,irq=10 -L \
/home/nathan/cbl/github/boot-utils-ro/images/ppc64le/ -bios skiboot.lid \
-machine powernv8 -display none -kernel \
/home/nathan/cbl/src/linux/arch/powerpc/boot/zImage.epapr -m 2G \
-nodefaults -serial mon:stdio
...
[    1.478028][    T1] BUG: Kernel NULL pointer dereference on read at 0x00000000
[    1.478630][    T1] Faulting instruction address: 0xc00000000090bee0
[    1.479521][    T1] Oops: Kernel access of bad area, sig: 11 [#1]
[    1.480036][    T1] LE PAGE_SIZE=64K MMU=Hash PREEMPT SMP NR_CPUS=16 NUMA PowerNV
[    1.480853][    T1] Modules linked in:
[    1.481265][    T1] CPU: 0 PID: 1 Comm: swapper/0 Not tainted 5.17.0-rc4-00001-gfa15c7cb550f #1
[    1.481967][    T1] NIP:  c00000000090bee0 LR: c000000000d96b60 CTR: c0000000000d5b4c
[    1.482596][    T1] REGS: c000000007443330 TRAP: 0380   Not tainted  (5.17.0-rc4-00001-gfa15c7cb550f)
[    1.483305][    T1] MSR:  9000000002009033 <SF,HV,VEC,EE,ME,IR,DR,RI,LE>  CR: 22800a87  XER: 00000000
[    1.484277][    T1] CFAR: c000000000d96b5c IRQMASK: 0
[    1.484277][    T1] GPR00: c000000000d96b54 c0000000074435d0 c0000000028bc600 0000000000000000
[    1.484277][    T1] GPR04: ffffffffffffffff ffffffffff1ea558 ffffffffff1ebfe4 c00000000261ae88
[    1.484277][    T1] GPR08: 0000000000000003 0000000000000004 c00000000261ae88 0000000000000000
[    1.484277][    T1] GPR12: 0000000000800000 c000000002a60000 c000000000012518 0000000000000000
[    1.484277][    T1] GPR16: 0000000000000000 0000000000000000 0000000000000000 0000000000000000
[    1.484277][    T1] GPR20: 0000000000000000 c0000000027bff80 0000000000000cc0 0000000000000000
[    1.484277][    T1] GPR24: 0000000000000010 0000000000000000 0000000000000000 0000000000000000
[    1.484277][    T1] GPR28: c0000000028fcfd8 c000000007b83000 0000000000000000 c0000000074435d0
[    1.490325][    T1] NIP [c00000000090bee0] strlen+0x10/0x30
[    1.490788][    T1] LR [c000000000d96b60] if_nlmsg_size+0x2b0/0x390
[    1.491319][    T1] Call Trace:
[    1.491573][    T1] [c0000000074435d0] [c000000000d96b54] if_nlmsg_size+0x2a4/0x390 (unreliable)
[    1.492291][    T1] [c000000007443680] [c000000000d96790] rtmsg_ifinfo_build_skb+0x80/0x1a0
[    1.492958][    T1] [c000000007443740] [c000000000d97590] rtmsg_ifinfo+0x70/0xd0
[    1.493559][    T1] [c000000007443790] [c000000000d7d528] register_netdevice+0x5d8/0x670
[    1.494205][    T1] [c000000007443820] [c000000000d7d94c] register_netdev+0x4c/0x80
[    1.494823][    T1] [c000000007443850] [c000000000f826d8] sit_init_net+0x1b8/0x200
[    1.495426][    T1] [c0000000074438d0] [c000000000d63b5c] ops_init+0x14c/0x1c0
[    1.496014][    T1] [c000000007443930] [c000000000d6314c] register_pernet_operations+0xec/0x1e0
[    1.496716][    T1] [c000000007443990] [c000000000d633d0] register_pernet_device+0x60/0xd0
[    1.497372][    T1] [c0000000074439e0] [c000000002085194] sit_init+0x54/0x160
[    1.497950][    T1] [c000000007443a70] [c000000000011c58] do_one_initcall+0x108/0x3e0
[    1.498573][    T1] [c000000007443c70] [c000000002006190] do_initcall_level+0xe4/0x1c4
[    1.499219][    T1] [c000000007443cc0] [c00000000200604c] do_initcalls+0x84/0xe4
[    1.499799][    T1] [c000000007443d40] [c000000002005da8] kernel_init_freeable+0x160/0x1ec
[    1.500444][    T1] [c000000007443da0] [c00000000001254c] kernel_init+0x3c/0x270
[    1.501042][    T1] [c000000007443e10] [c00000000000cd64] ret_from_kernel_thread+0x5c/0x64
[    1.501721][    T1] Instruction dump:
[    1.502202][    T1] eb81ffe0 7c0803a6 4e800020 00000000 00000000 00000000 60000000 60000000
[    1.502934][    T1] 3883ffff 60000000 60000000 60000000 <8ca40001> 28050000 4082fff8 7c632050
[    1.504028][    T1] ---[ end trace 0000000000000000 ]---
...

First thing was figuring out where the NULL pointer dereference happens,
which appears to the "strlen(ops->kind)" in rtnl_link_get_size():

515 static size_t rtnl_link_get_size(const struct net_device *dev)
516 {
517 	const struct rtnl_link_ops *ops = dev->rtnl_link_ops;
518 	size_t size;
519
520 	if (!ops)
521 		return 0;
522
523 	size = nla_total_size(sizeof(struct nlattr)) + /* IFLA_LINKINFO */
524 	       nla_total_size(strlen(ops->kind) + 1);  /* IFLA_INFO_KIND */

which I confirmed some really rudimentary printk debugging:

[    1.476862][    T1] nathan: rtnl_link_get_size(): name: sit0, ops: c0000000028fcfd8, ops->kind: (null)

diff --git a/net/core/rtnetlink.c b/net/core/rtnetlink.c
index 710da8a36729..c8d928e83aec 100644
--- a/net/core/rtnetlink.c
+++ b/net/core/rtnetlink.c

@@ -520,6 +520,9 @@ static size_t rtnl_link_get_size(const struct net_device *dev)
   	if (!ops)
   		return 0;
+	pr_err("nathan: %s(): name: %s, ops: %px, ops->kind: %s\n", __func__,
+	       dev->name, ops, ops->kind);
+
   	size = nla_total_size(sizeof(struct nlattr)) + /* IFLA_LINKINFO */
   	       nla_total_size(strlen(ops->kind) + 1);  /* IFLA_INFO_KIND */

Okay... how did sit0 end up with a NULL kind...? It is very clearly
defined as "sit":

1830 static struct rtnl_link_ops sit_link_ops __read_mostly = {
1831 	.kind		= "sit",

Adding some more debug prints to net/ipv6/sit.c:

diff --git a/net/ipv6/sit.c b/net/ipv6/sit.c
index c0b138c20992..7b9edbed2fcd 100644
--- a/net/ipv6/sit.c
+++ b/net/ipv6/sit.c

@@ -1920,6 +1920,12 @@ static int __net_init sit_init_net(struct net *net)
   	 */
   	sitn->fb_tunnel_dev->features |= NETIF_F_NETNS_LOCAL;
+	pr_err("nathan: %s(): &sit_link_ops: %px\n", __func__, &sit_link_ops);
+	pr_err("nathan: %s(): sit_link_ops.kind: %s\n", __func__, sit_link_ops.kind);
+	pr_err("nathan: %s(): sit_link_ops.maxtype: %u\n", __func__, sit_link_ops.maxtype);
+	pr_err("nathan: %s(): sitn->fb_tunnel_dev->rtnl_link_ops: %px\n", __func__, sitn->fb_tunnel_dev->rtnl_link_ops);
+	pr_err("nathan: %s(): sitn->fb_tunnel_dev->rtnl_link_ops->kind: %s\n", __func__, sitn->fb_tunnel_dev->rtnl_link_ops->kind);
+
   	err = register_netdev(sitn->fb_tunnel_dev);
   	if (err)
   		goto err_reg_dev;

reveals:

[    1.471920][    T1] sit: nathan: sit_init_net(): &sit_link_ops: c0000000028fcfd8
[    1.472534][    T1] sit: nathan: sit_init_net(): sit_link_ops.kind: (null)
[    1.473088][    T1] sit: nathan: sit_init_net(): sit_link_ops.maxtype: 20
[    1.473639][    T1] sit: nathan: sit_init_net(): sitn->fb_tunnel_dev->rtnl_link_ops: c0000000028fcfd8
[    1.474370][    T1] sit: nathan: sit_init_net(): sitn->fb_tunnel_dev->rtnl_link_ops->kind: (null)

This is super bizarre, as the maxtype member appears to have the correct
value, but how is kind's initial getting dropped on the floor?

Removing the __read_mostly annotation "fixes" it:

[    1.481708][    T1] sit: nathan: sit_init_net(): &sit_link_ops: c0000000027d3f60
[    1.482319][    T1] sit: nathan: sit_init_net(): sit_link_ops.kind: sit
[    1.482878][    T1] sit: nathan: sit_init_net(): sit_link_ops.maxtype: 20
[    1.483429][    T1] sit: nathan: sit_init_net(): sitn->fb_tunnel_dev->rtnl_link_ops: c0000000027d3f60
[    1.484174][    T1] sit: nathan: sit_init_net(): sitn->fb_tunnel_dev->rtnl_link_ops->kind: sit
...
Linux version 5.17.0-rc4-00001-g956f02ad5c31-dirty (nathan@dev-fedora.archlinux-ax161) (clang version 13.0.1 (Fedora 13.0.1-1.fc37), LLD 13.0.1) #2 SMP PREEMPT Wed Feb 16 13:29:49 MST 2022
...

diff --git a/net/ipv6/sit.c b/net/ipv6/sit.c
index 7b9edbed2fcd..f109c7a0233b 100644
--- a/net/ipv6/sit.c
+++ b/net/ipv6/sit.c

@@ -70,7 +70,7 @@ static void ipip6_tunnel_setup(struct net_device *dev);
   static void ipip6_dev_free(struct net_device *dev);
   static bool check_6rd(struct ip_tunnel *tunnel, const struct in6_addr *v6dst,
   		      __be32 *v4dst);
-static struct rtnl_link_ops sit_link_ops __read_mostly;
+static struct rtnl_link_ops sit_link_ops;
   static unsigned int sit_net_id __read_mostly;
   struct sit_net {

@@ -1827,7 +1827,7 @@ static void ipip6_dellink(struct net_device *dev, struct list_head *head)
   		unregister_netdevice_queue(dev, head);
   }
-static struct rtnl_link_ops sit_link_ops __read_mostly = {
+static struct rtnl_link_ops sit_link_ops = {
   	.kind		= "sit",
   	.maxtype	= IFLA_IPTUN_MAX,
   	.policy		= ipip6_policy,

Switching to ld.bfd also resolves it:

[    1.470405][    T1] sit: nathan: sit_init_net(): &sit_link_ops: c0000000028acfd8
[    1.471016][    T1] sit: nathan: sit_init_net(): sit_link_ops.kind: sit
[    1.471534][    T1] sit: nathan: sit_init_net(): sit_link_ops.maxtype: 20
[    1.472062][    T1] sit: nathan: sit_init_net(): sitn->fb_tunnel_dev->rtnl_link_ops: c0000000028acfd8
[    1.472790][    T1] sit: nathan: sit_init_net(): sitn->fb_tunnel_dev->rtnl_link_ops->kind: sit
...
Linux version 5.17.0-rc4-00001-g956f02ad5c31 (nathan@dev-fedora.archlinux-ax161) (clang version 13.0.1 (Fedora 13.0.1-1.fc37), GNU ld version 2.37-2.fc36) #3 SMP PREEMPT Wed Feb 16 13:33:42 MST 2022
...

I tested with ToT LLVM (or at least, close to it, since there is an
unrelated ld.lld regression there) and I could not reproduce it there,
so I did a reverse bisect to see what commit fixes this issue in LLVM 14
and I landed on:

commit 55c14d6dbfd8e7b86c15d2613fea3490078e2ae4
Author: Fangrui Song [off-list ref]
Date:   Thu Nov 25 14:12:34 2021 -0800

      [ELF] Simplify DynamicSection content computation. NFC

      The new code computes the content twice, but avoides the tricky
      std::function<uint64_t()>. Removed 13KiB code in a Release build.

   lld/ELF/SyntheticSections.cpp | 117 ++++++++++++++++--------------------------
   lld/ELF/SyntheticSections.h   |  12 +----
   2 files changed, 44 insertions(+), 85 deletions(-)

That's... interesting, given that commit title says No Functional
Change, even though there clearly is one. That commit has a couple
mentions of PowerPC synthetic sections, so it is possible that the
new content calculation lines up with ld.bfd?

I am not really sure where to go from here, as I don't fully understand
what the problem was before that LLD change. I'll see if I can do some
more investigation tomorrow (unless someone wants to beat me to it ;)

Thank you for looking into this, and sharing your analysis.

I built LLVM/clang from the master branch, rebuilt, but can still reproduce
this.

     $ git clone --depth=1 https://github.com/llvm/llvm-project.git
     $ cd llvm-project/
     $ git log --oneline
     41cb504b7 [mlir][linalg][bufferize][NFC] Move interface impl to Linalg Transforms
     $ mkdir build
     $ cd build
     $ cmake -DLLVM_ENABLE_PROJECTS=clang -G "Unix Makefiles"

Since this is something related to ld.lld, not clang, this should be:

... -DLLVM_ENABLE_PROJECTS="clang;lld" ...

quoted

-DCMAKE_BUILD_TYPE=Release -DLLVM_INSTALL_UTILS=ON
-DCMAKE_INSTALL_PREFIX=/scratch/local2/llvm ../llvm
     $ make -j20
     $ make -j20 clang-check

You can also do 'check-lld' if you want.

quoted

     $ make install
     $ /scratch/local2/llvm/bin/clang --version
     clang version 15.0.0 (https://github.com/llvm/llvm-project.git 41cb504b7c4b18ac15830107431a0c1eec73a6b2)
     Target: powerpc64le-unknown-linux-gnu
     Thread model: posix
     InstalledDir: /scratch/local2/llvm/bin

Then build Linux after `make clean` with `/scratch/local2/llvm/bin` in the
path.

     $ LLVM=1 LLVM_IAS=0 eatmydata make -j20

     $ qemu-system-ppc64 -enable-kvm -nographic -smp cores=1,threads=1 -net none -enable-kvm -M pseries -nodefaults -device spapr-vscsi -serial stdio -m 512 -kernel /dev/shm/linux/vmlinux -append "debug_boot_weak_hash panic=-1 console=ttyS0 rcupdate.rcu_cpu_stall_suppress_at_boot=1 torture.disable_onoff_at_boot rcupdate.rcu_task_stall_timeout=30000 rcupdate.rcu_self_test=1 rcutorture.onoff_interval=1000 rcutorture.onoff_holdoff=30 rcutorture.n_barrier_cbs=4 rcutorture.stat_interval=15 rcutorture.shutdown_secs=420 rcutorture.test_no_idle_hz=1 rcutorture.verbose=1"
     […]
     Preparing to boot Linux version 5.17.0-rc5-00178-ga4b9a8fb20e7 (pmenzel@flughafenberlinbrandenburgwillybrandt.molgen.mpg.de) (clang version 15.0.0 (https://github.com/llvm/llvm-project.git 41cb504b7c4b18ac15830107431a0c1eec73a6b2), LLD 13.0.0) #29 SMP PREEMPT Mon

                                              ^ still using ld.lld 13.0.0.

If you want to test the master branch, I would checkout LLVM at
460830a9c664e8cce959c660648faa7747ad8bdc, as the next commit introduces
a boot regression unrelated to this issue:

https://github.com/ClangBuiltLinux/linux/issues/1581

That should at least confirm this is resolved in a newer release.

Sorry for missing to update ld.lld. Indeed with the commit you 
mentioned, the segmentation fault is gone.

     $ /scratch/local2/llvm/bin/ld.lld --version
     LLD 14.0.0 (compatible with GNU linkers)

quoted

Feb 21 10:58:54 CET 2022
     […]
     [    0.465889][    T1] BUG: Kernel NULL pointer dereference on read at 0x00000000
     [    0.466749][    T1] Faulting instruction address: 0xc0000000008fc300
     [    0.467507][    T1] Oops: Kernel access of bad area, sig: 11 [#1]
     […]

I do intend to do further analysis at some point over the next few days
to see if I can figure out exactly why that commit that I mentioned
above fixes the issue then we can look into what we should do about it
in the kernel sources.

Awesome. Thank you for working on that.


Kind regards,

Paul