arm64 AWS host hangs during modprobe nvidia on lunar and mantic

[Francis Ginther]

Loading the nvidia driver dkms modules with "modprove nvidia" will result in the host hanging and being completely unusable. This was reproduced using both the linux generic and linux-aws kernels on lunar and mantic using an AWS g5g.xlarge instance.

To reproduce using the generic kernel:
# Deploy a arm64 host with an nvidia gpu, such as an AWS g5g.xlarge.

# Install the linux generic kernel from lunar-updates:
$ sudo DEBIAN_FRONTEND=noninteractive apt-get install -y -o DPkg::Options::=--force-confold linux-generic

# Boot to the linux-generic kernel (this can be accomplished by removing the existing kernel, in this case it was the linux-aws 6.2.0-1008-aws kernel)
$ sudo DEBIAN_FRONTEND=noninteractive apt-get purge -y -o DPkg::Options::=--force-confold linux-aws linux-aws-headers-6.2.0-1008 linux-headers-6.2.0-1008-aws linux-headers-aws linux-image-6.2.0-1008-aws linux-image-aws linux-modules-6.2.0-1008-aws linux-headers-6.2.0-1008-aws linux-image-6.2.0-1008-aws linux-modules-6.2.0-1008-aws
$ reboot

# Install the Nvidia 535-server driver DKMS package:
$ sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-driver-535-server

# Enable the driver
$ sudo modprobe nvidia

# At this point the system will hang and never return.
# A reboot instead of a modprobe will result in a system that never boots up all the way. I was able to recover the console logs from such a system and found (the full captured log is attached):

[ 1.964942] nvidia: loading out-of-tree module taints kernel.
[ 1.965475] nvidia: module license 'NVIDIA' taints kernel.
[ 1.965905] Disabling lock debugging due to kernel taint
[ 1.980905] nvidia: module verification failed: signature and/or required key missing - tainting kernel
[ 2.012067] nvidia-nvlink: Nvlink Core is being initialized, major device number 510
[ 2.012715]
[ 62.025143] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
[ 62.025807] rcu: 3-...0: (14 ticks this GP) idle=c04c/1/0x4000000000000000 softirq=653/654 fqs=3301
[ 62.026516] (detected by 0, t=15003 jiffies, g=-699, q=216 ncpus=4)
[ 62.027018] Task dump for CPU 3:
[ 62.027290] task:systemd-udevd state:R running task stack:0 pid:164 ppid:144 flags:0x0000000e
[ 62.028066] Call trace:
[ 62.028273] __switch_to+0xbc/0x100
[ 62.028567] 0x228
Timed out for waiting the udev queue being empty.
Timed out for waiting the udev queue being empty.
[ 242.045143] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
[ 242.045655] rcu: 3-...0: (14 ticks this GP) idle=c04c/1/0x4000000000000000 softirq=653/654 fqs=12303
[ 242.046373] (detected by 1, t=60008 jiffies, g=-699, q=937 ncpus=4)
[ 242.046874] Task dump for CPU 3:
[ 242.047146] task:systemd-udevd state:R running task stack:0 pid:164 ppid:144 flags:0x0000000f
[ 242.047922] Call trace:
[ 242.048128] __switch_to+0xbc/0x100
[ 242.048417] 0x228
Timed out for waiting the udev queue being empty.
Begin: Loading essential drivers ... [ 384.001142] watchdog: BUG: soft lockup - CPU#2 stuck for 22s! [modprobe:215]
[ 384.001738] Modules linked in: nvidia(POE+) crct10dif_ce video polyval_ce polyval_generic drm_kms_helper ghash_ce syscopyarea sm4 sysfillrect sha2_ce sysimgblt sha256_arm64 sha1_ce drm nvme nvme_core ena nvme_common aes_neon_bs aes_neon_blk aes_ce_blk aes_ce_cipher
[ 384.003513] CPU: 2 PID: 215 Comm: modprobe Tainted: P OE 6.2.0-26-generic #26-Ubuntu
[ 384.004210] Hardware name: Amazon EC2 g5g.xlarge/, BIOS 1.0 11/1/2018
[ 384.004715] pstate: 80400005 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 384.005259] pc : smp_call_function_many_cond+0x1b4/0x4b4
[ 384.005683] lr : smp_call_function_many_cond+0x1d0/0x4b4
[ 384.006108] sp : ffff8000089a3a70
[ 384.006381] x29: ffff8000089a3a70 x28: 0000000000000003 x27: ffff00056d1fafa0
[ 384.006954] x26: ffff00056d1d76c8 x25: ffffc87cf18bdd10 x24: 0000000000000003
[ 384.007527] x23: 0000000000000001 x22: ffff00056d1d76c8 x21: ffffc87cf18c2690
[ 384.008086] x20: ffff00056d1fafa0 x19: ffff00056d1d76c0 x18: ffff80000896d058
[ 384.008645] x17: 0000000000000000 x16: 0000000000000000 x15: 617362755f5f0073
[ 384.009209] x14: 0000000000000001 x13: 0000000000000006 x12: 4630354535323145
[ 384.009779] x11: 0101010101010101 x10: ffffb78318e9c0e0 x9 : ffffc87ceeac7da4
[ 384.010339] x8 : ffff00056d1d76f0 x7 : 0000000000000000 x6 : 0000000000000000
[ 384.010894] x5 : 0000000000000004 x4 : 0000000000000000 x3 : ffff00056d1fafa8
[ 384.011464] x2 : 0000000000000003 x1 : 0000000000000011 x0 : 0000000000000000
[ 384.012030] Call trace:
[ 384.012241] smp_call_function_many_cond+0x1b4/0x4b4
[ 384.012635] kick_all_cpus_sync+0x50/0xa0
[ 384.012961] flush_module_icache+0x64/0xd0
[ 384.013294] load_module+0x4ec/0xb54
[ 384.013588] __do_sys_finit_module+0xb0/0x150
[ 384.013944] __arm64_sys_finit_module+0x2c/0x50
[ 384.014306] invoke_syscall+0x7c/0x124
[ 384.014613] el0_svc_common.constprop.0+0x5c/0x1cc
[ 384.015000] do_el0_svc+0x38/0x60
[ 384.015280] el0_svc+0x30/0xe0
[ 384.015540] el0t_64_sync_handler+0x11c/0x150
[ 384.015896] el0t_64_sync+0x1a8/0x1ac

This same procedure impacts the 525, 525-server, 535 and 535-server drivers. It does *not* hang a similarly configured host running focal or jammy.

[Daniel van Vugt (vanvugt)]

Although nvidia seems to be the trigger here, the crashing code appears to be pure generic linux: arch/arm64/kernel/syscall.c

[Dimitri John Ledkov (xnox)]

since then, we had multiple glibc srus; kernel sru's and most recently new release of 535-server.

can i request for this to be retested again?

[Simon Fels (morphis)]

I can reproduce the the same with the latest 535.154.05-0ubuntu0.22.04.1 on jammy with the 6.5 HWE kernel on an arm64 machine. The same happens with the -server driver 535.154.05-0ubuntu0.22.04.1.

Reproducing is pretty simple:

1. Boot plain Ubuntu 22.04 with either HWE already installed or manually installed to switch to it from GA
2. Install NVIDIA driver via

$ sudo apt install -y nvidia-headless-535-server

or

$ sudo apt install -y nvidia-headless-535

Doing either an nvidia-smi (triggers the modprobe of nvidia kernel modules) or a `modprobe nvidia` makes the system hang entirely.

The same works fine on the 5.15 GA kernel.

[Simon Fels (morphis)]

Verified that the issue does not exist with 535.154.05-0ubuntu0.22.04.1 of nvidia-utils-535-server on 6.2.0-1017-aws or 6.2.0-1018-aws of linux-aws.

[Simon Fels (morphis)]

Trying the same with the linux-nvidia-hwe-22.04-edge kernel from proposed linux-image-6.5.0-1011-nvidia wit the same NVIDIA driver (535.154.05-0ubuntu0.22.04.1 of nvidia-utils-535-server) and loading kernel driver and running nvidia-smi works fine without problems.

[Dimitri John Ledkov (xnox)]

I am surprised that `ubuntu-drivers list` doesn't provide any drivers to install, when it really should.

To install pre-built drivers I use

$ sudo apt install linux-modules-nvidia-535-server-aws nvidia-headless-535-server

Such that signed nvidia modules provided by Canonical are installed.

Similarly to upgrade to edge variant I did:

$ sudo apt install linux-aws-edge linux-modules-nvidia-535-server-aws-edge

[Dimitri John Ledkov (xnox)]

and everything seems to work fine.

[Dimitri John Ledkov (xnox)]

I wonder if the bug is with trying to install self-built dkms modules, instead of pre-built ones, and how come ubuntu-drivers is not offering pre-built ones...

[Launchpad Janitor (janitor)]

Status changed to 'Confirmed' because the bug affects multiple users.

[Francis Ginther (fginther)]

I can reproduce the failure on mantic with both the DKMS and LRM drivers. Specifically what I'm doing to install these are:

for DKMS:
sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-driver-535-server

for LRM:
sudo DEBIAN_FRONTEND=noninteractive apt-get install -y nvidia-headless-no-dkms-535-server linux-modules-nvidia-535-server-generic nvidia-utils-535-server

I'm intentionally not using `ubuntu-drivers` to isolate this testing to just the installation and functioning of the drivers.

[Simon Fels (morphis)]

Verified that with linux-aws-edge 6.5.0.1012.12~22.04.1 the DKMS installation via

$ sudo apt install -y nvidia-driver-535-server

on an AWS g5g.xlarge goes through the driver comes up fine.

Trying the same with linux-generic-hwe-22.04-edge 6.5.0-17-generic #17~22.04.1 on an Ampere Altra with 2x NVIDIA L4 still runs into the same hang with nvidia-headless-535-server (535.154.05-0ubuntu0.22.04.1).

[Simon Fels (morphis)]

I gave this another spin today with 6.5.0-17-generic #17~22.04.1 and the LRM modules of the 535 driver (6.5.0-17.17~22.04.1+1 of linux-modules-nvidia-535-server-generic-hwe-22.04) on our Altra system with 2x L4 GPUs and the same problem exists as with the DKMS modules:

[ 39.437849] watchdog: BUG: soft lockup - CPU#62 stuck for 26s! [systemd-udevd:850]
[ 39.445411] Modules linked in: nvidia(POE+) crct10dif_ce polyval_ce polyval_generic ghash_ce ast mlx5_core video drm_shmem_helper sm4 mlxfw sha2_ce drm_kms_helper nvme psample sha256_arm64 sha1_ce nvme_core igb drm tls xhci_pci nvme_common pci_hyperv_intf xhci_pci_renesas i2c_algo_bit aes_neon_bs aes_neon_blk aes_ce_blk aes_ce_cipher
[ 39.474949] CPU: 62 PID: 850 Comm: systemd-udevd Tainted: P OE 6.5.0-17-generic #17~22.04.1-Ubuntu
[ 39.485196] Hardware name: GIGABYTE G242-P30-JG/MP32-AR0-JG, BIOS F07 03/22/2021
[ 39.492578] pstate: 80400009 (Nzcv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
[ 39.499526] pc : smp_call_function_many_cond+0x19c/0x720
[ 39.504830] lr : smp_call_function_many_cond+0x1b8/0x720
[ 39.510130] sp : ffff80008934b920
[ 39.513431] x29: ffff80008934b920 x28: ffffaef99146dd10 x27: 0000000000000000
[ 39.520554] x26: 000000000000004f x25: ffff085dcfffbb80 x24: 0000000000000026
[ 39.527677] x23: 0000000000000001 x22: ffff085dcfdd6708 x21: ffffaef9914726e0
[ 39.534799] x20: ffff085dcfadbb80 x19: ffff085dcfdd6700 x18: ffff800089341060
[ 39.541921] x17: 0000000000000000 x16: 0000000000000000 x15: 43535f5f00656c75
[ 39.549044] x14: 0c030b111b111303 x13: 0000000000000006 x12: 3931413337353339
[ 39.556166] x11: 0101010101010101 x10: 000000000000004f x9 : ffffaef98ee015b8
[ 39.563289] x8 : 0000000000000000 x7 : 0000000000000000 x6 : 000000000000003e
[ 39.570411] x5 : ffffaef99146d000 x4 : 0000000000000000 x3 : ffff085dcfadbb88
[ 39.577533] x2 : 0000000000000026 x1 : 0000000000000011 x0 : 0000000000000000
[ 39.584656] Call trace:
[ 39.587090] smp_call_function_many_cond+0x19c/0x720
[ 39.592043] kick_all_cpus_sync+0x50/0xa8
[ 39.596040] flush_module_icache+0x94/0xf8
[ 39.600125] load_module+0x448/0x8e0
[ 39.603688] init_module_from_file+0x94/0x110
[ 39.608033] idempotent_init_module+0x194/0x2b0
[ 39.612551] __arm64_sys_finit_module+0x74/0x100
[ 39.617155] invoke_syscall+0x7c/0x130
[ 39.620892] el0_svc_common.constprop.0+0x5c/0x170
[ 39.625670] do_el0_svc+0x38/0x68
[ 39.628972] el0_svc+0x30/0xe0
[ 39.632016] el0t_64_sync_handler+0x128/0x158
[ 39.636360] el0t_64_sync+0x1b0/0x1b8

[Mitchell Augustin (mitchellaugustin)]

I identified a similar bug today when installing nvidia-fabricmanager-535 on a noble dev build for arm64 that may be related:

https://bugs.launchpad.net/ubuntu/+source/fabric-manager-535/+bug/2052663

[Abhishek Chauhan (abchauhan)]

Hi all,
This should be fixed on the latest driver 550.67 - https://www.nvidia.com/Download/driverResults.aspx/223429/en-us/.
Please help verify if this is resolved on your systems. Thanks!

[Abhishek Chauhan (abchauhan)]

The fix is also available on 535.171.04 available here - https://www.nvidia.com/Download/driverResults.aspx/223761/en-us/