[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-Id: <20180508155158.7e6a789d950bcaf957c8c3bf@linux-foundation.org>
Date: Tue, 8 May 2018 15:51:58 -0700
From: Andrew Morton <akpm@...ux-foundation.org>
To: Mark Rutland <mark.rutland@....com>
Cc: linux-kernel@...r.kernel.org, aryabinin@...tuozzo.com,
dvyukov@...gle.com, mingo@...hat.com, peterz@...radead.org
Subject: Re: [PATCH 2/3] kcov: prefault the kcov_area
On Fri, 4 May 2018 14:55:34 +0100 Mark Rutland <mark.rutland@....com> wrote:
> On many architectures the vmalloc area is lazily faulted in upon first
> access. This is problematic for KCOV, as __sanitizer_cov_trace_pc
> accesses the (vmalloc'd) kcov_area, and fault handling code may be
> instrumented. If an access to kcov_area faults, this will result in
> mutual recursion through the fault handling code and
> __sanitizer_cov_trace_pc(), eventually leading to stack corruption
> and/or overflow.
>
> We can avoid this by faulting in the kcov_area before
> __sanitizer_cov_trace_pc() is permitted to access it. Once it has been
> faulted in, it will remain present in the process page tables, and will
> not fault again.
>
> ...
>
> --- a/kernel/kcov.c
> +++ b/kernel/kcov.c
> @@ -324,6 +324,17 @@ static int kcov_close(struct inode *inode, struct file *filep)
> return 0;
> }
>
> +static void kcov_fault_in_area(struct kcov *kcov)
It would be nice to have a comment here explaining why the function
exists.
umm, this?
--- a/kernel/kcov.c~kcov-prefault-the-kcov_area-fix-fix
+++ a/kernel/kcov.c
@@ -324,6 +324,10 @@ static int kcov_close(struct inode *inod
return 0;
}
+/*
+ * fault in a lazily-faulted vmalloc area, to avoid recursion issues if the
+ * vmalloc fault handler itself is instrumented.
+ */
static void kcov_fault_in_area(struct kcov *kcov)
{
unsigned long stride = PAGE_SIZE / sizeof(unsigned long);
> +{
> + unsigned long stride = PAGE_SIZE / sizeof(unsigned long);
> + unsigned long *area = kcov->area;
> + unsigned long offset;
> +
> + for (offset = 0; offset < kcov->size; offset += stride) {
> + READ_ONCE(area[offset]);
> + }
> +}
Powered by blists - more mailing lists