linux-kernel - Re: [PATCH 4/4] perf tools: Fix struct comm

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20180719182843.GA2812@kernel.org>
Date:   Thu, 19 Jul 2018 15:28:43 -0300
From:   Arnaldo Carvalho de Melo <acme@...nel.org>
To:     Jiri Olsa <jolsa@...nel.org>
Cc:     lkml <linux-kernel@...r.kernel.org>,
        Ingo Molnar <mingo@...nel.org>,
        Namhyung Kim <namhyung@...nel.org>,
        David Ahern <dsahern@...il.com>,
        Alexander Shishkin <alexander.shishkin@...ux.intel.com>,
        Peter Zijlstra <a.p.zijlstra@...llo.nl>,
        Kan Liang <kan.liang@...el.com>,
        Andi Kleen <ak@...ux.intel.com>,
        Lukasz Odzioba <lukasz.odzioba@...el.com>,
        Wang Nan <wangnan0@...wei.com>
Subject: Re: [PATCH 4/4] perf tools: Fix struct comm_str removal crash

Em Thu, Jul 19, 2018 at 04:33:45PM +0200, Jiri Olsa escreveu:
> We occasionaly hit following assert failure in perf top,
> when processing the /proc info in multiple threads.
> 
>   perf: ...include/linux/refcount.h:109: refcount_inc:
>         Assertion `!(!refcount_inc_not_zero(r))' failed.
> 
> The gdb backtrace looks like this:
> 
>   [Switching to Thread 0x7ffff11ba700 (LWP 13749)]
>   0x00007ffff50839fb in raise () from /lib64/libc.so.6
>   (gdb)
>   #0  0x00007ffff50839fb in raise () from /lib64/libc.so.6
>   #1  0x00007ffff5085800 in abort () from /lib64/libc.so.6
>   #2  0x00007ffff507c0da in __assert_fail_base () from /lib64/libc.so.6
>   #3  0x00007ffff507c152 in __assert_fail () from /lib64/libc.so.6
>   #4  0x0000000000535373 in refcount_inc (r=0x7fffdc009be0)
>       at ...include/linux/refcount.h:109
>   #5  0x00000000005354f1 in comm_str__get (cs=0x7fffdc009bc0)
>       at util/comm.c:24
>   #6  0x00000000005356bd in __comm_str__findnew (str=0x7fffd000b260 ":2",
>       root=0xbed5c0 <comm_str_root>) at util/comm.c:72
>   #7  0x000000000053579e in comm_str__findnew (str=0x7fffd000b260 ":2",
>       root=0xbed5c0 <comm_str_root>) at util/comm.c:95
>   #8  0x000000000053582e in comm__new (str=0x7fffd000b260 ":2",
>       timestamp=0, exec=false) at util/comm.c:111
>   #9  0x00000000005363bc in thread__new (pid=2, tid=2) at util/thread.c:57
>   #10 0x0000000000523da0 in ____machine__findnew_thread (machine=0xbfde38,
>       threads=0xbfdf28, pid=2, tid=2, create=true) at util/machine.c:457
>   #11 0x0000000000523eb4 in __machine__findnew_thread (machine=0xbfde38,
>   ...
> 
> The failing assertion is this one:
> 
>   REFCOUNT_WARN(!refcount_inc_not_zero(r), ...
> 
> The problem is that we keep global comm_str_root list, which
> is accessed by multiple threads during the perf top startup
> and following 2 paths can race:
> 
>   thread 1:
>     ...
>     thread__new
>       comm__new
>         comm_str__findnew
>           down_write(&comm_str_lock);
>           __comm_str__findnew
>             comm_str__get
> 
>   thread 2:
>     ...
>     comm__override or comm__free
>       comm_str__put
>         refcount_dec_and_test
>           down_write(&comm_str_lock);
>           rb_erase(&cs->rb_node, &comm_str_root);
> 
> Because thread 2 first decrements the refcnt and only after then it
> removes the struct comm_str from the list, the thread 1 can find this
> object on the list with refcnt equls to 0 and hit the assert.
> 
> This patch fixes the thread 1 __comm_str__findnew path, by ignoring
> objects that already dropped the refcnt to 0. For the rest of the
> objects we take the refcnt before comparing its name and release
> it afterwards with comm_str__put, which can also release the object
> completely.
> 
> Link: http://lkml.kernel.org/n/tip-vrizt6sw1lu1ybsrl9l0wwln@git.kernel.org
> Signed-off-by: Jiri Olsa <jolsa@...nel.org>
> ---
>  tools/perf/util/comm.c | 15 +++++++++------
>  1 file changed, 9 insertions(+), 6 deletions(-)
> 
> diff --git a/tools/perf/util/comm.c b/tools/perf/util/comm.c
> index 7798a2cc8a86..9c4a18991e41 100644
> --- a/tools/perf/util/comm.c
> +++ b/tools/perf/util/comm.c
> @@ -18,11 +18,9 @@ struct comm_str {
>  static struct rb_root comm_str_root;
>  static struct rw_semaphore comm_str_lock = {.lock = PTHREAD_RWLOCK_INITIALIZER,};
>  
> -static struct comm_str *comm_str__get(struct comm_str *cs)
> +static bool comm_str__get(struct comm_str *cs)
>  {
> -	if (cs)
> -		refcount_inc(&cs->refcnt);
> -	return cs;
> +	return cs ? refcount_inc_not_zero(&cs->refcnt) : false;
>  }

I don't like changing the semantics of a __get() operation this way, I
think it should stay like all the others, i.e. return the object with
the desired refcount or return NULL if that is not possible.

Otherwise we'll have to switch gears when debugging refcounts in various
objects, that start having slightly different semantics for reference
counting.

We should try to find a fix that maintains the semantics of refcounting.

>  static void comm_str__put(struct comm_str *cs)
> @@ -67,9 +65,14 @@ struct comm_str *__comm_str__findnew(const char *str, struct rb_root *root)
>  		parent = *p;
>  		iter = rb_entry(parent, struct comm_str, rb_node);
>  
> +		/*
> +		 * If we race with comm_str__put, iter->refcnt is 0
> +		 * and it will be removed within comm_str__put call
> +		 * shortly, ignore it in this search.
> +		 */
>  		cmp = strcmp(str, iter->str);
> -		if (!cmp)
> -			return comm_str__get(iter);
> +		if (!cmp && comm_str__get(iter))
> +			return iter;
>  
>  		if (cmp < 0)
>  			p = &(*p)->rb_left;
> -- 
> 2.17.1