Re: [PATCH] Add mmap_assert_locked() annotations to find_vma*()

From: Luigi Rizzo
Date: Sun Aug 01 2021 - 20:16:45 EST


(repost in plain text)

On Sun, Aug 1, 2021 at 9:33 PM Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> On Sat, 31 Jul 2021 10:53:41 -0700 Luigi Rizzo <lrizzo@xxxxxxxxxx> wrote:
>
> > find_vma() and variants need protection when used.
> > This patch adds mmap_assert_lock() calls in the functions.
> >
> > To make sure the invariant is satisfied, we also need to add a
> > mmap_read_loc() around the get_user_pages_remote() call in
> > get_arg_page(). The lock is not strictly necessary because the mm
> > has been newly created, but the extra cost is limited because
> > the same mutex was also acquired shortly before in __bprm_mm_init(),
> > so it is hot and uncontended.
> >
>
> Well, it isn't cost-free. find_vma() is called a lot and a surprising
> number of systems apparently run with CONFIG_DEBUG_VM. Why do you
> think this cost is justified?

I assume you are concerned with the cost of mmap_assert_locked() ?

I'd say the justification is the same as for all asserts:
at some point some code change may miss the required lock, and the
asserts are there to catch elusive race conditions,

There are in fact already instances of mmap_locked_assert()
right before find_vma() in walk_page_range(), and a couple before
calls to __get_user_pages().

As for the cost, I'd think that if CONFIG_DEBUG_VM is set,
one does it on purpose to catch errors and is prepared to pay
the cost (in this case the atomic_read(counter) in rwsem_is_locked(),
the counter should be hot).

FWIW I have instrumented find_vma() on a fast machine using kstats

https://github.com/luigirizzo/lr-cstats

(load the module then enable the trace with
echo "trace pcpu:find_vma bits 3" > /sys/kernel/debug/kstats/_control
and monitor the time with
watch "grep CPUS /sys/kernel/debug/kstats/find_vma"

I didn't run anything especially intensive except some network
benchmarks, but I have collected ~2M samples with the following
distribution of find_vma() time in nanoseconds in 3 configs:

CONFIGURATION p10 p50 p90 p95 p98

no-debug 89 109 214 332 605
debug 331 369 603 862 1338
debug+this patch 337 369 603 863 1339

As you can see, just compiling a debug kernel, even without this patch,
makes the function 3x more expensive. The effect of this patch is
not measurable (the differences are below measurement error).

cheers
luigi