Re: [PATCH 1/3] perf/core: Make sure the ring-buffer is mapped in all page-tables

[Thomas Gleixner <tglx@linutronix.de>]

On Fri, 20 Jul 2018, Thomas Gleixner wrote:
> On Fri, 20 Jul 2018, Andy Lutomirski wrote:
> > On Fri, Jul 20, 2018 at 12:27 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
> > > On Fri, 20 Jul 2018, Andy Lutomirski wrote:
> > >> > On Jul 20, 2018, at 6:22 AM, Joerg Roedel <joro@8bytes.org> wrote:
> > >> >
> > >> > From: Joerg Roedel <jroedel@suse.de>
> > >> >
> > >> > The ring-buffer is accessed in the NMI handler, so we better
> > >> > avoid faulting on it. Sync the vmalloc range with all
> > >> > page-tables in system to make sure everyone has it mapped.
> > >> >
> > >> > This fixes a WARN_ON_ONCE() that can be triggered with PTI
> > >> > enabled on x86-32:
> > >> >
> > >> >    WARNING: CPU: 4 PID: 0 at arch/x86/mm/fault.c:320 vmalloc_fault+0x220/0x230
> > >> >
> > >> > This triggers because with PTI enabled on an PAE kernel the
> > >> > PMDs are no longer shared between the page-tables, so the
> > >> > vmalloc changes do not propagate automatically.
> > >>
> > >> It seems like it would be much more robust to fix the vmalloc_fault()
> > >> code instead.
> > >
> > > Right, but now the obvious fix for the issue at hand is this. We surely
> > > should revisit this.
> > 
> > If you commit this under this reasoning, then please at least make it say:
> > 
> > /* XXX: The vmalloc_fault() code is buggy on PTI+PAE systems, and this
> > is a workaround. */
> > 
> > Let's not have code in the kernel that pretends to make sense but is
> > actually voodoo magic that works around bugs elsewhere.  It's no fun
> > to maintain down the road.
> 
> Fair enough. Lemme amend it. Joerg is looking into it, but I surely want to
> get that stuff some exposure in next ASAP.

Delta patch below.

Thanks.

	tglx

8<-------------
--- a/kernel/events/ring_buffer.c
+++ b/kernel/events/ring_buffer.c
@@ -815,8 +815,12 @@ static void rb_free_work(struct work_str
 	vfree(base);
 	kfree(rb);
 
-	/* Make sure buffer is unmapped in all page-tables */
-	vmalloc_sync_all();
+	/*
+	 * FIXME: PAE workaround for vmalloc_fault(): Make sure buffer is
+	 * unmapped in all page-tables.
+	 */
+	if (IS_ENABLED(CONFIG_X86_PAE))
+		vmalloc_sync_all();
 }
 
 void rb_free(struct ring_buffer *rb)
@@ -844,11 +848,13 @@ struct ring_buffer *rb_alloc(int nr_page
 		goto fail_all_buf;
 
 	/*
-	 * The buffer is accessed in NMI handlers, make sure it is
-	 * mapped in all page-tables in the system so that we don't
-	 * fault on the range in an NMI handler.
+	 * FIXME: PAE workaround for vmalloc_fault(): The buffer is
+	 * accessed in NMI handlers, make sure it is mapped in all
+	 * page-tables in the system so that we don't fault on the range in
+	 * an NMI handler.
 	 */
-	vmalloc_sync_all();
+	if (IS_ENABLED(CONFIG_X86_PAE))
+		vmalloc_sync_all();
 
 	rb->user_page = all_buf;
 	rb->data_pages[0] = all_buf + PAGE_SIZE;