[PATCH 0/2] locking, tracing: Fix incorrect use of arch_spin_lock()

From: Waiman Long
Date: Wed Sep 21 2022 - 09:23:26 EST


It is found that running the LTP read_all_proc test may cause the
following warning to show up:

[12512.905036] BUG: using smp_processor_id() in preemptible [00000000] code: read_all/133711
[12512.913499] caller is __pv_queued_spin_lock_slowpath+0x7f/0xd30
[12512.921163] CPU: 3 PID: 133711 Comm: read_all Kdump: loaded Tainted: G OE --------- --- 5.14.0-163.el9.x86_64+debug #1
[12512.936652] Hardware name: Red Hat KVM, BIOS 0.5.1 01/01/2011
[12512.944213] Call Trace:
[12512.950660] dump_stack_lvl+0x57/0x81
[12512.957400] check_preemption_disabled+0xcc/0xd0
[12512.964487] __pv_queued_spin_lock_slowpath+0x7f/0xd30
[12512.971552] ? pv_hash+0x110/0x110
[12512.978119] ? __lock_acquire+0xb72/0x1870
[12512.984683] tracing_saved_cmdlines_size_read+0x177/0x190
[12512.991655] ? saved_cmdlines_start+0x2c0/0x2c0
[12512.998355] ? inode_security+0x54/0xf0
[12513.004548] ? selinux_file_permission+0x324/0x420
[12513.011185] ? lock_downgrade+0x130/0x130
[12513.017423] ? fsnotify_perm.part.0+0x14a/0x4c0
[12513.023715] vfs_read+0x126/0x4d0
[12513.029432] ksys_read+0xf9/0x1d0
[12513.035131] ? __ia32_sys_pwrite64+0x1e0/0x1e0
[12513.041028] ? ktime_get_coarse_real_ts64+0x130/0x170
[12513.047167] do_syscall_64+0x59/0x90
[12513.052656] ? lockdep_hardirqs_on+0x79/0x100
[12513.058268] ? do_syscall_64+0x69/0x90
[12513.063593] ? lockdep_hardirqs_on+0x79/0x100
[12513.069022] ? do_syscall_64+0x69/0x90
[12513.074137] ? lockdep_hardirqs_on+0x79/0x100
[12513.079533] entry_SYSCALL_64_after_hwframe+0x63/0xcd
[12513.085015] RIP: 0033:0x7f93f09d38c2

So tracing_saved_cmdlines_size_read() does call arch_spin_lock() with
preemption enabled. Looking at other arch_spin_lock() call sites in
kernel/trace/trace.c, there are several others that may have the same
problem. Other arch_spin_lock() callers under kernel look OK as
irqs has been disabled before calling arch_spin_lock().

Add a do_arch_spin_lock() helper that disables preemption and make
kernel/trace/trace.c use it if it is not obvious that either preemption
or interrupt has been disabled.

Waiman Long (2):
locking: Provide a low overhead do_arch_spin_lock() API
tracing: Use proper do_arch_spin_lock() API

include/linux/spinlock.h | 27 +++++++++++++++++++++
kernel/trace/trace.c | 52 +++++++++++++++++++---------------------
2 files changed, 51 insertions(+), 28 deletions(-)

--
2.31.1