[Devel] [PATCH RH9] x86/cpu: init_cpu_flags -- use raw spinlock

[Kirill Tkhai]

On 14.10.2021 11:20, Cyrill Gorcunov wrote:
> The @cpu_flags_lock spinlock guards manipulations with
> per-cpu @cpu_flags which is used to hide some features
> in cpuinfo output inside VE. Still the init_cpu_flags
> is called from irq context leading to the following
> 
>  | [   13.827635] =============================
>  | [   13.827636] [ BUG: Invalid wait context ]
>  | [   13.827637] 5.14.0.ovz9.10.1+ #41 Tainted: G         C     X --------- --- 
>  | [   13.827638] -----------------------------
>  | [   13.827638] systemd/1 is trying to lock:
>  | [   13.827639] ffffffffa4c9d258 (cpu_flags_lock){....}-{3:3}, at: init_cpu_flags+0xc8/0x220
>  | [   13.827649] other info that might help us debug this:
>  | [   13.827651] context-{2:2}
>  | [   13.827651] 3 locks held by systemd/1:
>  | [   13.827652]  #0: ffffffffa56e8c60 (dup_mmap_sem){.+.+}-{0:0}, at: dup_mm+0x83/0x5f0
>  | [   13.827660]  #1: ffff97574a37d138 (&mm->mmap_lock#2){++++}-{4:4}, at: dup_mm+0x9c/0x5f0
>  | [   13.827664]  #2: ffff97574489c138 (&mm->mmap_lock/1){+.+.}-{4:4}, at: dup_mm+0xd5/0x5f0
>  | [   13.827667] stack backtrace:
>  | [   13.827668] CPU: 0 PID: 1 Comm: systemd ve: / Tainted: G         C     X --------- ---
>  | [   13.827670] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.14.0-4.fc34 04/01/2014
>  | [   13.827672] Call Trace:
>  | [   13.827673]  <IRQ>
>  | [   13.827675]  dump_stack_lvl+0x57/0x7d
>  | [   13.827686]  __lock_acquire.cold+0x28b/0x2cd
>  | [   13.827694]  lock_acquire+0xca/0x300
>  | [   13.827700]  ? init_cpu_flags+0xc8/0x220
>  | [   13.827703]  _raw_spin_lock+0x34/0x80
>  | [   13.827708]  ? init_cpu_flags+0xc8/0x220
>  | [   13.827710]  init_cpu_flags+0xc8/0x220
>  | [   13.827713]  flush_smp_call_function_queue+0x13f/0x1e0
>  | [   13.827717]  __sysvec_call_function_single+0x43/0x1c0
>  | [   13.827722]  sysvec_call_function_single+0x9d/0xd0
>  | [   13.827724]  </IRQ>
>  | [   13.827724]  asm_sysvec_call_function_single+0x12/0x20
>  | [   13.827728] RIP: 0010:lock_release+0x178/0x460
>  | ...
>  | [   13.827741]  up_write+0x2f/0x1c0
>  | [   13.827743]  anon_vma_clone+0x158/0x1f0
>  | [   13.827749]  anon_vma_fork+0x33/0x180
>  | [   13.827751]  dup_mm+0x45b/0x5f0
>  | [   13.827755]  copy_process+0x1e5a/0x2050
>  | [   13.827758]  kernel_clone+0x9b/0x3f0
>  | [   13.827760]  ? vfs_statx+0x74/0x130
>  | [   13.827766]  __do_sys_clone+0x60/0x80
>  | [   13.827769]  do_syscall_64+0x3b/0x90
>  | [   13.827771]  entry_SYSCALL_64_after_hwframe+0x44/0xae
> 
> The problem is rather coming from rt camp where splinlocks
> become sleepable thus can't be used in irq context (and for our kernel
> it requires the CONFIG_PROVE_RAW_LOCK_NESTING to be set), thus since
> we know that we're operating in irq context lets use raw spinlocks
> instead.
> 
> https://jira.sw.ru/browse/PSBM-134761
> 
> CC: Kirill Tkhai <ktkhai at virtuozzo.com>
> Signed-off-by: Cyrill Gorcunov <gorcunov at gmail.com>

Reviewed-by: Kirill Tkhai <ktkhai at virtuozzo.com>

> ---
>  arch/x86/kernel/cpu/proc.c |   10 +++++-----
>  1 file changed, 5 insertions(+), 5 deletions(-)
> 
> --- vzkernel.orig/arch/x86/kernel/cpu/proc.c
> +++ vzkernel/arch/x86/kernel/cpu/proc.c
> @@ -69,7 +69,7 @@ struct cpu_flags {
>  };
>  
>  static DEFINE_PER_CPU(struct cpu_flags, cpu_flags);
> -static DEFINE_SPINLOCK(cpu_flags_lock);
> +static DEFINE_RAW_SPINLOCK(cpu_flags_lock);
>  
>  static void init_cpu_flags(void *dummy)
>  {
> @@ -107,9 +107,9 @@ static void init_cpu_flags(void *dummy)
>  		flags.val[10] &= eax;
>  	}
>  
> -	spin_lock(&cpu_flags_lock);
> +	raw_spin_lock(&cpu_flags_lock);
>  	memcpy(&per_cpu(cpu_flags, cpu), &flags, sizeof(flags));
> -	spin_unlock(&cpu_flags_lock);
> +	raw_spin_unlock(&cpu_flags_lock);
>  }
>  
>  static int show_cpuinfo(struct seq_file *m, void *v)
> @@ -158,9 +158,9 @@ static int show_cpuinfo(struct seq_file
>  	show_cpuinfo_misc(m, c);
>  
>  	if (!is_super) {
> -		spin_lock_irq(&cpu_flags_lock);
> +		raw_spin_lock_irq(&cpu_flags_lock);
>  		memcpy(&ve_flags, &per_cpu(cpu_flags, cpu), sizeof(ve_flags));
> -		spin_unlock_irq(&cpu_flags_lock);
> +		raw_spin_unlock_irq(&cpu_flags_lock);
>  	}
>  
>  
>