Messages in this thread | | | Date | Thu, 27 Nov 2008 15:03:46 +0200 | From | Török Edwin <> | Subject | Re: [PATCH 2/2] tracing: identify which executable object the userspace address belongs to |
| |
On 2008-11-27 14:48, Frank Ch. Eigler wrote: > Hi - > > On Thu, Nov 27, 2008 at 11:41:45AM +0100, Peter Zijlstra wrote: > >>>> Impact: modify+improve the userstacktrace tracing visualization feature >>>> [...] >>>> You'll see stack entries like: >>>> /lib/libpthread-2.7.so[+0xd370] >>>> [...] >>>> >>> Can you suggest an actual distribution & architecture where this >>> facility may be tested/used? It appears to require frame-pointer >>> stuff that AFAIK is not generally turned on for user-space. >>> >> gentoo, just rebuild world with frame pointers ;-) >> > > Well, that only goes so far. If this feature turns out unable to work > without distributors recompiling all their stuff on, for example, x86-64, > then expectations need to be reset.
My assumption is that this feature will be used to trace individual applications, and not the system as a whole. Then you only need libc to be recompiled with frame pointers on, and your own application/your own application's libraries.
That is what I want to use it for, and there isn't another solution that allows me to do this. Sure I can trace userspace alone using ptrace (which has its own overhead), and the kernel alone by using ftrace, but I can't combine those traces in a meaningful manner. If/when the kernel will support dwarf unwinding, it will only need to provide an alternate implementation for save_stack_trace_user.
Even without frame pointers you can at least get the return address to userspace, which may be inside your application for page faults.
If I need to do system-wide tracing, I can use my 32-bit chroot [*], or boot my laptop which is 32-bit.
I don't think that this feature should get rejected just because it is not easily usable from x86_64.
[*] I haven't tested yet if tracing 32-bit applications from a 64-bit kernel works. It probably won't, and I'll need to use a different struct stack_frame with 32-bit addresses.
Another approach I've though of would be to deliver a signal to userspace on demand, and have the signal handler do the backtrace, but that would unnecesary overhead.
Best regards, --Edwin
| |