Date: Mon, 8 Jan 2007 13:36:44 -0500 From: John Baldwin <jhb@freebsd.org> To: Peter Jeremy <peterjeremy@optushome.com.au> Cc: cvs-src@freebsd.org, src-committers@freebsd.org, cvs-all@freebsd.org Subject: Re: cvs commit: src/usr.bin/kdump kdump.c Message-ID: <200701081336.45214.jhb@freebsd.org> In-Reply-To: <20070106043515.GD839@turion.vk2pj.dyndns.org> References: <200701052104.l05L4cO7037092@repoman.freebsd.org> <200701051607.59334.jhb@freebsd.org> <20070106043515.GD839@turion.vk2pj.dyndns.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Friday 05 January 2007 23:35, Peter Jeremy wrote: > On Fri, 2007-Jan-05 16:07:58 -0500, John Baldwin wrote: > >On Friday 05 January 2007 16:04, John Baldwin wrote: > >> jhb 2007-01-05 21:04:37 UTC > >> > >> FreeBSD src repository > >> > >> Modified files: > >> usr.bin/kdump kdump.c > >> Log: > >> Add code to parse the utrace(2) entries generated by malloc(3) in a more > >> human-readable format. Note that we report 'realloc(p, 0)' as 'free(p)' > >> since both cases are encoded the same way and 'free()' is more common > >> than a realloc() to 0. > >> > >> MFC after: 1 week > > This is much nicer than having to run kdump output thru my perl script > to do this. The only downside I see is that the code in kdump assumes > that any utrace records that are sizeof(struct utrace_malloc) are > generated by malloc. This isn't necessarily true - whilst nothing in > the base system apart from malloc currently uses utrace, it's possible > that people are using utrace in their own code. I'd prefer to see > this decoding controlled by a command line option. (Ideally, kdump > would grow a configuration file so that a user could define their own > decoding rules - but that is a lot of work). We could turn it off if we want. I agree that having to depend on size to determine the malloc case sucks as a signature, but that format has been fairly well established at this point. > >I also have patches I use at work that allow kdump to recognize a 32-bit > >malloc utrace on an amd64 machine (for when you run an i386 binary) if folks > >are interested. I'm not sure how many i386 on amd64 hacks we want in the > >official CVS tree. :) > > Personally, I'd like FreeBSD to behave similarly to Solaris: You choose > whether to compile 32-bit or 64-bit executables with a compiler switch > and everything else is transparent. FreeBSD 3.x had smarts so that nm, > ld, gdb etc could transparently handle either a.out or ELF executables. > It would be nice if FreeBSD/amd64 could do the same (though I realise > that we don't want the overheads on other platforms, which would make it > more difficult to implement). We have several things at work that we don't currently feed back because it makes things ugly. For example, at work we let 32-bit tcpdump work on amd64 via compat ioctl's, etc. (That one is ugly as the kernel actually has to send a different header on each packet due to struct timeval being different sizes.) > >I also have another set of patches to add various utrace(2) events to the > >runtime linker as well as logic in kdump to parse them that I hope to commit > >in the near future. > > Sounds good. This goes back to my first point above - I don't think it's > safe to rely on the size of a utrace record to determine its type. For RTLD I am using a 4 byte signature 'RTLD' followed by a byte specifying the type of record that follows and do not depend on the size. I think it would be best if any new utrace's used a similar 4-char signature (you could even dynamically add handlers based on the signature to kdump that way), but for malloc() I was stuck with what we had. See www.freebsd.org/~jhb/patches/rtld_utrace.patch for the approach I used. -- John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200701081336.45214.jhb>