lkml.org 
[lkml]   [2020]   [Sep]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH RFC 09/10] kfence, Documentation: add KFENCE documentation
On Tue, Sep 08, 2020 at 08:54AM -0700, Dave Hansen wrote:
> On 9/7/20 6:40 AM, Marco Elver wrote:
> > +The most important parameter is KFENCE's sample interval, which can be set via
> > +the kernel boot parameter ``kfence.sample_interval`` in milliseconds. The
> > +sample interval determines the frequency with which heap allocations will be
> > +guarded by KFENCE. The default is configurable via the Kconfig option
> > +``CONFIG_KFENCE_SAMPLE_INTERVAL``. Setting ``kfence.sample_interval=0``
> > +disables KFENCE.
> > +
> > +With the Kconfig option ``CONFIG_KFENCE_NUM_OBJECTS`` (default 255), the number
> > +of available guarded objects can be controlled. Each object requires 2 pages,
> > +one for the object itself and the other one used as a guard page; object pages
> > +are interleaved with guard pages, and every object page is therefore surrounded
> > +by two guard pages.
>
> Is it hard to make these both tunable at runtime?

The number of objects is quite hard, because it really complicates
bookkeeping and might also have an impact on performance, which is why
we prefer the statically allocated pool (like on x86, and we're trying
to get it for arm64 as well).

The sample interval is already tunable, just write to
/sys/module/kfence/parameters/sample_interval. Although we have this
(see core.c):

module_param_named(sample_interval, kfence_sample_interval, ulong,
IS_ENABLED(CONFIG_DEBUG_KERNEL) ? 0600 : 0400);

I was wondering if it should also be tweakable on non-debug kernels, but
I fear it might be abused. Sure, you need to be root to change it, but
maybe I'm being overly conservative here? If you don't see huge problems
with it we could just make it 0600 for all builds.

> It would be nice if I hit a KFENCE error on a system to bump up the
> number of objects and turn up the frequency of guarded objects to try to
> hit it again. That would be a really nice feature for development
> environments.

Indeed, which is why we also found it might be useful to tweak
sample_interval at runtime for debug-kernels. Although I don't know how
much luck you'll have hitting it again.

My strategy at that point would be to take the stack traces, try to
construct test-cases for those code paths, and run them through KASAN
(if it isn't immediately obvious what the problem is).

> It would also be nice to have a counter somewhere (/proc/vmstat?) to
> explicitly say how many pages are currently being used.

You can check /sys/kernel/debug/kfence/stats. On a system I just booted:

[root@syzkaller][~]# cat /sys/kernel/debug/kfence/stats
enabled: 1
currently allocated: 18
total allocations: 105
total frees: 87
total bugs: 0

The "currently allocated" count is the currently used KFENCE objects (of
255 for the default config).

> I didn't mention it elsewhere, but this work looks really nice. It has
> very little impact on the core kernel and looks like a very nice tool to
> have in the toolbox. I don't see any major reasons we wouldn't want to
> merge after our typical bikeshedding. :)

Thank you!

-- Marco

\
 
 \ /
  Last update: 2020-09-08 19:34    [W:0.124 / U:0.428 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site