Re: [PATCH] tracing: Fix memleak due to race between current_tracer and trace

[Zheng Yejian]

On 2023/8/16 09:45, Zheng Yejian wrote:
> On 2023/8/8 17:29, Zheng Yejian wrote:
> > Kmemleak report a leak in graph_trace_open():
> >
> >    unreferenced object 0xffff0040b95f4a00 (size 128):
> >      comm "cat", pid 204981, jiffies 4301155872 (age 99771.964s)
> >      hex dump (first 32 bytes):
> >        e0 05 e7 b4 ab 7d 00 00 0b 00 01 00 00 00 00 00 .....}..........
> >        f4 00 01 10 00 a0 ff ff 00 00 00 00 65 00 10 00 ............e...
> >      backtrace:
> >        [<000000005db27c8b>] kmem_cache_alloc_trace+0x348/0x5f0
> >        [<000000007df90faa>] graph_trace_open+0xb0/0x344
> >        [<00000000737524cd>] __tracing_open+0x450/0xb10
> >        [<0000000098043327>] tracing_open+0x1a0/0x2a0
> >        [<00000000291c3876>] do_dentry_open+0x3c0/0xdc0
> >        [<000000004015bcd6>] vfs_open+0x98/0xd0
> >        [<000000002b5f60c9>] do_open+0x520/0x8d0
> >        [<00000000376c7820>] path_openat+0x1c0/0x3e0
> >        [<00000000336a54b5>] do_filp_open+0x14c/0x324
> >        [<000000002802df13>] do_sys_openat2+0x2c4/0x530
> >        [<0000000094eea458>] __arm64_sys_openat+0x130/0x1c4
> >        [<00000000a71d7881>] el0_svc_common.constprop.0+0xfc/0x394
> >        [<00000000313647bf>] do_el0_svc+0xac/0xec
> >        [<000000002ef1c651>] el0_svc+0x20/0x30
> >        [<000000002fd4692a>] el0_sync_handler+0xb0/0xb4
> >        [<000000000c309c35>] el0_sync+0x160/0x180
> >
> > The root cause is descripted as follows:
> >
> >    __tracing_open() {  // 1. File 'trace' is being opened;
> >      ...
> >      *iter->trace = *tr->current_trace;  // 2. Tracer 'function_graph' is
> >                                          //    currently set;
> >      ...
> >      iter->trace->open(iter);  // 3. Call graph_trace_open() here,
> >                                //    and memory are allocated in it;
> >      ...
> >    }
> >
> >    s_start() {  // 4. The opened file is being read;
> >      ...
> >      *iter->trace = *tr->current_trace;  // 5. If tracer is switched to
> >                                          //    'nop' or others, then 
> > memory
> >                                          //    in step 3 are leaked!!!
> >      ...
> >    }
> >
> > To fix it, in s_start(), close tracer before switching then reopen the
> > new tracer after switching.
> >
> > Fixes: d7350c3f4569 ("tracing/core: make the read callbacks reentrants")
> > Signed-off-by: Zheng Yejian <zhengyejian1@xxxxxxxxxx>
> > ---
> >   kernel/trace/trace.c | 9 ++++++++-
> >   1 file changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> > index b8870078ef58..d50a0227baa3 100644
> > --- a/kernel/trace/trace.c
> > +++ b/kernel/trace/trace.c
> > @@ -4213,8 +4213,15 @@ static void *s_start(struct seq_file *m, loff_t 
> > *pos)
> >        * will point to the same string as current_trace->name.
> >        */
> >       mutex_lock(&trace_types_lock);
> > -    if (unlikely(tr->current_trace && iter->trace->name != 
> > tr->current_trace->name))
> > +    if (unlikely(tr->current_trace && iter->trace->name != 
> > tr->current_trace->name)) {
> > +        /* Close iter->trace before switching to the new current 
> > tracer */
> > +        if (iter->trace->close)
> > +            iter->trace->close(iter);
>
> Hi, kasan report an issue related to this patch in v5.10, I'll handle
> it and then send V2 if the patch needs to be changed.
>
>    BUG: KASAN: use-after-free in graph_trace_close+0x78/0x7c
>    Read of size 8 at addr ffff204054ca8f00 by task cat/193541
>    [...]
>    Call trace:
>     dump_backtrace+0x0/0x3e4
>     show_stack+0x20/0x2c
>     dump_stack+0x140/0x198
>     print_address_description.constprop.0+0x2c/0x1fc
>     __kasan_report+0xe0/0x140
>     kasan_report+0x44/0x5c
>     __asan_report_load8_noabort+0x34/0x60
>     graph_trace_close+0x78/0x7c
>     wakeup_trace_close+0x3c/0x54
>     s_start+0x4f4/0x794
>     seq_read_iter+0x210/0xd90
>     seq_read+0x288/0x410
>     vfs_read+0x13c/0x41c
>     ksys_read+0xf4/0x1e0
>     __arm64_sys_read+0x74/0xa4
>     el0_svc_common.constprop.0+0xfc/0x394
>     do_el0_svc+0xac/0xec
>     el0_svc+0x20/0x30
>     el0_sync_handler+0xb0/0xb4
>     el0_sync+0x160/0x180
>
>    Allocated by task 193541:
>     kasan_save_stack+0x28/0x60
>     __kasan_kmalloc.constprop.0+0xa4/0xd0
>     kasan_kmalloc+0x10/0x20
>     kmem_cache_alloc_trace+0x2ec/0x5f0
>     graph_trace_open+0xb0/0x344
>     __tracing_open+0x450/0xb10
>     tracing_open+0x1a0/0x2a0
>     do_dentry_open+0x3c0/0xdc0
>     vfs_open+0x98/0xd0
>     do_open+0x520/0x8d0

The root cause should be that:
  1. As __tracing_open() being called, function graph tracer is set,
     then graph_trace_open() is called;
  2. As s_start() being called, switch to wakeup tracer, then
     graph_trace_close() is called, here 'iter->private' is not cleared.
     Then wakeup_trace_open() is called, and if graph_trace_open() is
     not called by it due to is_graph(iter->tr) returns false;
  3. As next time s_start() being called, if tracer switched again then
     wakeup_trace_close() is called and it mistakenly close the
     'iter->private' in step 2 which finally causes kasan do the report.

So we may also need to set 'iter->private' to be NULL in
graph_trace_close(), I'll send v2 soon.

--- a/kernel/trace/trace_functions_graph.c
+++ b/kernel/trace/trace_functions_graph.c
@@ -1244,6 +1244,7 @@ void graph_trace_close(struct trace_iterator *iter)
        if (data) {
                free_percpu(data->cpu_data);
                kfree(data);
+               iter->private = NULL;
        }
 }

-- Zheng Yejian

> >           *iter->trace = *tr->current_trace;
> > +        /* Reopen the new current tracer */
> > +        if (iter->trace->open)
> > +            iter->trace->open(iter);
> > +    }
> >       mutex_unlock(&trace_types_lock);
> >   #ifdef CONFIG_TRACER_MAX_TRACE