On Fri, Sep 23, 2016 at 10:26:15AM -0700, Andi Kleen wrote: > > Afaict there's no actual need to hide the AUX buffer for this sampling > > stuff; the user knows about all this and can simply mmap() the AUX part. > > The sample could either point to locations in the AUX buffer, or (as I > > think this code does) memcpy bits out. > > This would work for perf, but not for the core dump case below. > > > Ideally we'd pass the AUX-event into the syscall, that way you avoid all > > the find_aux_event crud. I'm not sure we want to overload the group_fd > > thing more (its already very hard to create counter groups in a cgroup > > for example) .. > > > > Coredump was mentioned somewhere, but I'm not sure I've seen > > code/interfaces for that. How was that envisioned to work? > > The idea was to have a rlimit that enables PT running as a ring buffer > in the background. If something crashes the ring buffer is dumped > as part of the core dump, and then gdb can tell you how you crashed. > This extends what gdb already does explicitly today using perf > API calls.
Well, we could 'force' inject a VMA into the process's address space, we do that for a few other things as well. It also makes for less exceptions with the actual core dumping. But the worry I have is the total amount of pinned memory. If you want to inherit this on fork(), as is a reasonable expectation, then its possible to quickly exceed the total amount of pinnable memory. At which point we _should_ start failing fork(), which is a somewhat unexpected, and undesirable side-effect. Ideally we'd unpin the old buffers and repin the new buffers on context switch, but that's impossible since faulting needs scheduling, recursion, we loose. I really want to see something sensible before we go do that.