Received-SPF: pass (google.com: domain of linux-kernel+bounces-212965-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
Sender: Ingo Molnar <mingo.kernel.org@gmail.com>
Date: Thu, 13 Jun 2024 11:36:38 +0200
From: Ingo Molnar <mingo@kernel.org>
To: Oleg Nesterov <oleg@redhat.com>
Cc: Nathan Chancellor <nathan@kernel.org>, linux-kernel@vger.kernel.org,
	Andy Lutomirski <luto@amacapital.net>,
	Andrew Morton <akpm@linux-foundation.org>,
	Dave Hansen <dave@sr71.net>, Peter Zijlstra <peterz@infradead.org>,
	Borislav Petkov <bp@alien8.de>, Brian Gerst <brgerst@gmail.com>,
	"H . Peter Anvin" <hpa@zytor.com>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Uros Bizjak <ubizjak@gmail.com>
Subject: [PATCH 10/9] x86/fpu: Fix 'struct fpu' misalignment on 32-bit kernels
Message-ID: <Zmq9ppuIZJ9IMZDr@gmail.com>
References: <20240608073134.264210-1-mingo@kernel.org>
 <20240608073134.264210-4-mingo@kernel.org>
 <20240610211350.GA1613053@thelio-3990X>
 <20240611124145.GA26798@redhat.com>
 <ZmlZiHVF8w09mExw@gmail.com>
 <20240612184148.GB23973@redhat.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240612184148.GB23973@redhat.com>


* Oleg Nesterov <oleg@redhat.com> wrote:

> The patch below seems to fix the problem.
> 
> Again, the changes in fpu__init_system_early_generic() are not
> strictly needed to fix it, but I believe make sense anyway.
> 
> Oleg.
> 
> 
> diff --git a/arch/x86/kernel/fpu/init.c b/arch/x86/kernel/fpu/init.c
> index 4e8d37b5a90b..848ea79886ba 100644
> --- a/arch/x86/kernel/fpu/init.c
> +++ b/arch/x86/kernel/fpu/init.c
> @@ -71,16 +71,14 @@ static bool __init fpu__probe_without_cpuid(void)
>  	return fsw == 0 && (fcw & 0x103f) == 0x003f;
>  }
>  
> -static struct fpu x86_init_fpu __read_mostly;
> +static struct fpu x86_init_fpu __attribute__ ((aligned (64))) __read_mostly;
>  
>  static void __init fpu__init_system_early_generic(void)
>  {
> -	int this_cpu = smp_processor_id();
> -
>  	fpstate_reset(&x86_init_fpu);
>  	current->thread.fpu = &x86_init_fpu;
> -	per_cpu(fpu_fpregs_owner_ctx, this_cpu) = &x86_init_fpu;
> -	x86_init_fpu.last_cpu = this_cpu;
> +	set_thread_flag(TIF_NEED_FPU_LOAD);
> +	x86_init_fpu.last_cpu = -1;
>  
>  	if (!boot_cpu_has(X86_FEATURE_CPUID) &&
>  	    !test_bit(X86_FEATURE_FPU, (unsigned long *)cpu_caps_cleared)) {
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 215a7380e41c..ec22b9bf27f5 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1562,7 +1562,7 @@ struct task_struct {
>  	 * they are included in the randomized portion of task_struct.
>  	 */
>  	randomized_struct_fields_end
> -};
> +} __attribute__ ((aligned (64)));

Oh ... indeed, FPU context save area must be 64 bytes aligned!

On 64-bit kernels this was a given, accidentally, but on 32-bit kernels 
init_task was only 32-byte aligned:

  c22f04e0 D init_task

... which misaligned the struct fpu as well, I think. With your fix:

  c22f0500 D init_task

What happened is that due to my series 'struct task_struct' lost its 
64-byte alignment attribute, which broke the fpu struct allocation code on 
32-bit kernels and made the 64-bit one probably unrobust as well.

To add insult to injury, I was aware of the alignment requirement, and 
tried to cover it with an assert, but doubly mis-coded it:

+       BUILD_BUG_ON(sizeof(*dst) % SMP_CACHE_BYTES != 0);

Which is buggy:

 - As on 32-bit kernels CONFIG_X86_L1_CACHE_SHIFT=5, ie. 32 bytes ...

 - Nor does it really check the alignment of the FPU context save area 
   within struct fpu as it's allocated after task_struct ...

The interim patch below against the full WIP.x86/fpu series is what fixes 
Nathan's 32-bit testcase.

Further improvements:

 - The extra alignment attribute in <linux/sched.h> will affect other 
   architecture as well, although in practice the alignment of init_task is 
   not critical, and is very likely at least 32 bytes, probably more. 
   Still, it's a bit ugly in its current form.

 - Also, because this was pretty hard to debug, we should probably add an 
   alignment check to fpu__init_task_struct_size() where we allocate the 
   fpu context structure, and fix the buggy size-assert.

Thanks a lot for your help Oleg! I've added this tag of yours:

  Fixed-by: Oleg Nesterov <oleg@redhat.com>

... and would appreciate your Acked-by or Reviewed-by for the eventual 
final version of the series, but I don't insist. ;-)

	Ingo

=================>
diff --git a/arch/x86/kernel/fpu/init.c b/arch/x86/kernel/fpu/init.c
index 53580e59e5db..16b6611634c3 100644
--- a/arch/x86/kernel/fpu/init.c
+++ b/arch/x86/kernel/fpu/init.c
@@ -71,15 +71,13 @@ static bool __init fpu__probe_without_cpuid(void)
 	return fsw == 0 && (fcw & 0x103f) == 0x003f;
 }
 
-static struct fpu x86_init_fpu __read_mostly;
+static struct fpu x86_init_fpu __attribute__ ((aligned (64))) __read_mostly;
 
 static void __init fpu__init_system_early_generic(void)
 {
-	int this_cpu = smp_processor_id();
-
 	fpstate_reset(&x86_init_fpu);
-	per_cpu(fpu_fpregs_owner_ctx, this_cpu) = &x86_init_fpu;
-	x86_init_fpu.last_cpu = this_cpu;
+	set_thread_flag(TIF_NEED_FPU_LOAD);
+	x86_init_fpu.last_cpu = -1;
 
 	if (!boot_cpu_has(X86_FEATURE_CPUID) &&
 	    !test_bit(X86_FEATURE_FPU, (unsigned long *)cpu_caps_cleared)) {
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 215a7380e41c..ec22b9bf27f5 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -1562,7 +1562,7 @@ struct task_struct {
 	 * they are included in the randomized portion of task_struct.
 	 */
 	randomized_struct_fields_end
-};
+} __attribute__ ((aligned (64)));
 
 #define TASK_REPORT_IDLE	(TASK_REPORT + 1)
 #define TASK_REPORT_MAX		(TASK_REPORT_IDLE << 1)