BUG: debug_exception_enter() disables preemption and may call sleeping functions on aarch64 with RT

["Luis Claudio R. Goncalves" <lgoncalv@redhat.com>]

Hello!

While running ssdd[1] from rt-tests on an aarch64 kernel with PREEMPT_RT and
debug features enabled, this bug was triggered on every single run:

[1] https://git.kernel.org/pub/scm/utils/rt-tests/rt-tests.git/tree/src/ssdd/ssdd.c

# ssdd

[  273.115597] BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:48
[  273.115607] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 6077, name: ssdd
[  273.115611] preempt_count: 1, expected: 0
[  273.115614] RCU nest depth: 0, expected: 0
[  273.115617] 1 lock held by ssdd/6077:
[  273.115620]  #0: ffff07ffd77893e0 (&sighand->siglock){+.+.}-{3:3}, at: force_sig_info_to_task+0x58/0x200
[  273.115642] Preemption disabled at:
[  273.115644] [<ffffad24518bd1f4>] debug_exception_enter+0x1c/0x80
[  273.115653] CPU: 47 UID: 0 PID: 6077 Comm: ssdd Not tainted 6.13.0-rt3 #1 PREEMPT_RT
[  273.115659] Hardware name: GIGABYTE R152-P31-00/MP32-AR1-00, BIOS F31n (SCP: 2.10.20220810) 09/30/2022
[  273.115662] Call trace:
[  273.115664]  show_stack+0x34/0x98 (C)
[  273.115670]  dump_stack_lvl+0xa8/0xe8
[  273.115675]  dump_stack+0x1c/0x38
[  273.115680]  __might_resched+0x254/0x330
[  273.115686]  rt_spin_lock+0xcc/0x220
[  273.115692]  force_sig_info_to_task+0x58/0x200
[  273.115697]  force_sig_fault+0xd0/0x120
[  273.115702]  arm64_force_sig_fault+0x48/0x80
[  273.115707]  send_user_sigtrap+0x88/0xe8
[  273.115712]  single_step_handler+0x100/0x160
[  273.115717]  do_debug_exception+0x94/0x160
[  273.115722]  el0_dbg+0x54/0x150
[  273.115727]  el0t_64_sync_handler+0x134/0x138
[  273.115732]  el0t_64_sync+0x1ac/0x1b0

The ptrace usage in ssdd eventually exercises the code path that starts on
el0t_64_sync_handler() and may end up calling do_debug_exception(), which
calls debug_exception_enter() that disables preemption.

Looking at the backtrace, later in the call chain force_sig_info_to_task()
tries to take a spinlock, which on PREEMPT_RT becomes a rtmutex and could
sleep in case of contention. That triggers the "BUG: sleeping function
called from invalid context" warning.

It is also possible to reproduce the problem in an aarch64 kernel with
PREEMPT_RT enabled, no extra debug features, by running ssdd in a loop.
With that we can see not only the backtrace reported above but also other
instances where the process is scheduled out while preemption is disabled:

# while :; do ssdd; done

[  754.673678] BUG: scheduling while atomic: ssdd/7340/0x00000002
[  754.673682] Modules linked in: qrtr rfkill sunrpc vfat fat acpi_ipmi ipmi_ssif arm_spe_pmu igb ipmi_devintf ipmi_msghandler arm_dmc620_pmu arm_cmn cppc_cpufreq arm_dsu_pmu loop dm_multipath nfnetlink xfs nvme ghash_ce sha2_ce sha256_arm64 nvme_core sha1_ce nvme_auth sbsa_gwdt ast i2c_algo_bit i2c_designware_platform xgene_hwmon i2c_designware_core dm_mirror dm_region_hash dm_log dm_mod fuse
[  754.673703] Preemption disabled at:
[  754.673703] [<ffffa87a17ca470c>] do_debug_exception+0x54/0x100
[  754.673710] CPU: 102 UID: 0 PID: 7340 Comm: ssdd Kdump: loaded Not tainted 6.14.0-rc1 #1
[  754.673712] Hardware name: GIGABYTE R152-P31-00/MP32-AR1-00, BIOS F31n (SCP: 2.10.20220810) 09/30/2022
[  754.673713] Call trace:
[  754.673714]  show_stack+0x34/0x98 (C)
[  754.673718]  dump_stack_lvl+0x80/0xa8
[  754.673721]  dump_stack+0x18/0x2c
[  754.673722]  __schedule_bug+0x90/0xc0
[  754.673726]  schedule_debug.isra.0+0x128/0x158
[  754.673728]  __schedule+0x68/0x690
[  754.673731]  schedule_rtlock+0x24/0x50
[  754.673733]  rtlock_slowlock_locked+0x1c0/0x350
[  754.673735]  rt_spin_lock+0xcc/0x130
[  754.673737]  obj_cgroup_charge+0x54/0x138
[  754.673740]  __memcg_slab_post_alloc_hook+0xcc/0x300
[  754.673743]  kmem_cache_alloc_noprof+0x304/0x338
[  754.673745]  __send_signal_locked+0x90/0x428
[  754.673748]  send_signal_locked+0xe4/0x140
[  754.673750]  force_sig_info_to_task+0xd0/0x160
[  754.673753]  force_sig_fault+0x6c/0xa8
[  754.673755]  arm64_force_sig_fault+0x48/0x80
[  754.673757]  send_user_sigtrap+0x54/0xd0
[  754.673759]  single_step_handler+0xc4/0xe0
[  754.673761]  do_debug_exception+0x7c/0x100
[  754.673762]  el0_dbg+0x40/0x158
[  754.673766]  el0t_64_sync_handler+0x134/0x138
[  754.673768]  el0t_64_sync+0x1ac/0x1b0

In this case one of the local_lock_* calls in (the functions called by)
obj_cgroup_charge() seems to hit contention and, as it is dealing with
rtmutexes, be effectively scheduled out to sleep.

The scary comment on top of debug_exception_enter() provides a reason for
preemption being disabled at that point, but it seems to open a can of worms
for PREEMPT_RT usage:

    /*
     * In debug exception context, we explicitly disable preemption despite
     * having interrupts disabled.
     * This serves two purposes: it makes it much less likely that we would
     * accidentally schedule in exception context and it will force a warning
     * if we somehow manage to schedule by accident.
     */

This is the data I gathered so far, using both v6.13.0-rt3 and 6.14.0-rc1
for testing. But due to my ignorance wrt the debug exception treatment in
aarch64 I can't devise a solution for the observed behavior.

Any suggestions or comments?

Best regards,
Luis