Re: [GIT PULL 00/42] perf/core improvements and fixes

[Ingo Molnar <mingo@kernel.org>]

* Namhyung Kim <namhyung@kernel.org> wrote:

> Hi Ingo,
> 
> On Fri, 5 Oct 2012 10:18:04 +0200, Ingo Molnar wrote:
> > I also noticed a 'perf trace' bug, after running 'perf trace' it 
> > outputs lines but then gets hung:
> >
> >   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND                               
> >  6081 mingo     20   0 18.2g  14g 3544 D 18.6 91.2   0:20.28 perf                    
> >
> > and then after half a minute it gets active again, outputting 
> > lines and then segfaulting:
> >
> >  LOST 1 events!
> >  31082 ) = 375
> >  31082 write(fd: 3, buf: 140030569454096, count: 48LOST 1 events!
> >  31082 select(n: 13, inp: 140030569376688, outp: 140030569376656, exp: 0, tvp: 031082 ) = 2
> >  Segmentation fault
> >
> > It's a 16-way box running:
> >
> >  Linux comet 3.5.4-1.fc17.x86_64 #1 SMP Mon Sep 17 15:03:59 UTC 2012 x86_64 x86_64 x86_64 GNU/Linux
> >
> > Note how much the RSS is, 14 GB of RAM with less of 1 minute 
> > running. The segfault might be related to a failed allocation 
> > not being handled correctly perhaps.
> 
> I also hit the segfault but it was due to a broken sample data:
> 
> Core was generated by `./perf trace'.
> Program terminated with signal 11, Segmentation fault.
> #0  perf_evsel__intval (evsel=evsel@entry=0xeae8c0, sample=sample@entry=0x7fff42278130, name=name@entry=0x55034d "id")
>     at util/evsel.c:1148
> 1148			value = *(u64 *)ptr;
> 
> (gdb) bt
> #0  perf_evsel__intval (evsel=evsel@entry=0xeae8c0, sample=sample@entry=0x7fff42278130, name=name@entry=0x55034d "id")
>     at util/evsel.c:1148
> #1  0x0000000000446987 in trace__syscall_info (sample=0x7fff42278130, evsel=0xeae8c0, trace=0x7fff422781b0)
>     at builtin-trace.c:147
> #2  trace__sys_exit (trace=0x7fff422781b0, evsel=0xeae8c0, sample=0x7fff42278130) at builtin-trace.c:193
> #3  0x00000000004470c1 in trace__run (argv=<optimized out>, argc=<optimized out>, trace=0x7fff422781b0) at builtin-trace.c:310
> #4  cmd_trace (argc=<optimized out>, argv=<optimized out>, prefix=<optimized out>) at builtin-trace.c:396
> #5  0x0000000000418c93 in run_builtin (p=p@entry=0x7b19d8, argc=argc@entry=1, argv=argv@entry=0x7fff4227a7a0) at perf.c:312
> #6  0x000000000041846e in handle_internal_command (argv=0x7fff4227a7a0, argc=1) at perf.c:360
> #7  run_argv (argv=0x7fff4227a590, argcp=0x7fff4227a59c) at perf.c:404
> #8  main (argc=1, argv=0x7fff4227a7a0) at perf.c:502
> (gdb) list
> 1143			break;
> 1144		case 4:
> 1145			value = *(u32 *)ptr;
> 1146			break;
> 1147		case 8:
> 1148			value = *(u64 *)ptr;
> 1149			break;
> 1150		default:
> 1151			return 0;
> 1152		}
> (gdb) p ptr
> $1 = (void *) 0x10
> (gdb) p *sample
> $2 = {ip = 0, pid = 0, tid = 0, time = 15762598695796738, addr = 0, id = 315, stream_id = 18446744073709551615, period = 1, 
>   cpu = 143, raw_size = 0, raw_data = 0x0, callchain = 0x0, branch_stack = 0x0, user_regs = {regs = 0x0}, user_stack = {
>     offset = 0, size = 0, data = 0x0}}
> 
> 
> In this case 'sample->raw_data' was NULL and other fields 
> seemed invalid as well.  I guess we need some kind of 
> protection?

Yeah, the code should assume the perf.data to be 100% untrusted, 
i.e. it can be random input and should never crash, lock up or 
misbehave.

Thanks,

	Ingo