DDB scripting, output capture, and textdumps
Robert Watson
rwatson at FreeBSD.org
Tue Dec 18 04:10:47 PST 2007
Dear all:
I've been hacking on-and-off for a while on a side project to improve our
kernel debugging facilities. Primarily, my concern has been to address
three problems:
- The complications of employing kernel core dumps for debugging,
including the large size of dumps making them unwieldy to distribute or
store for any extended period (even with minidumps), the requirement to
have relatively synchronized kernel source in order to use the dumps,
the need to have a kernel with debugging symbols, and the problems with
fsck causing sufficient swap use to invalidate dumps before they can be
extracted.
- The decreasing likelihood that notebooks will ship with serial ports
that can be used for interactive debugging using DDB. Making end-users
type in stack traces is cruel, photos are a pain, and X11 rules out
both.
- The fact that a great many problems are most easily diagnosed using
utility routines present in DDB, but not as easily using kgdb for
offline analysis. I find that for many bugs I analyze, simply looking
at the DDB output is sufficient to identify the source of the problem.
An idea I punted around a bit at BSDCan earlier this year (or perhaps it
was at EuroBSDCon the previous year) was an idea of a "textdump" -- that
is, a new type of kernel dump based on capturing automatically extracted
debugging information generated by DDB. The result would be an ASCII text
file that could be filed as a bug report, perhaps even automatically.
To this end, I have implemented three new facilities for use with DDB:
(1) DDB output capture. The output of DDB is stored in a memory buffer,
and can be extracted using a sysctl or textdumps (see below). This
can be turned on and off, both for use manually ("I'll want this
later, but not that") and as part of scripts (see below).
(2) DDB scripting. A limited number of named scripts can be defined to
run a series of DDB commands. No loops, etc, just simple command
lists. These can be caused to run automatically on entering DDB for
various scenarios, including WITNESS violations and kernel panics.
They can also be run by hand in order to save a bit of typing if you
use DDB in a repetitive way (as I do).
(3) Textdumps. A new dump type that stores a series of data files
containing various pieces of information, including the DDB capture
buffer, kernel message buffer, kernel configuration (if compiled into
the kernel), panic message, and kernel version string. These are
stored in the ustar format inside the dump partition (aligned to the
end) so can be easily extended, and savecore(8) requires almost no new
logic to deal with them (it just drops numbered tar files in
/var/crash). This makes it straight forward to extend the textdump format
to include new types of information and avoids the issue of how to safely
simultaneously represent information in many different formats in the same
file.
These are pretty flexible tools, and you can imagine doing the following
sorts of things:
- Setting the kdb.enter.panic script to automatically turn on output
capture, do full backtraces of all threads, show open file information,
dump UMA stats, and save it all to a textdump and then reboot.
- Setting the kdb.enter.witness script to show lock information, generate
a coredump, and reboot. Or, just to automatically do "show allocks" and
drop to the DDB prompt.
- Adding a flag to rc.conf to automatically submit textdumps via e-mail to
a specific address, perhaps including GNATS or an automated bug system.
These could be unpacked and automatically analyzed, and do to the compact
size, kept for long-term trend analysis or to identify when a problem
started occuring.
I've produced an initial snapshot of the above, which can be found here:
http://www.watson.org/~robert/freebsd/20071218-ddb.tgz
This adds three files to DDB, patches quite a few kernel files (to pass
more information into KDB about why it's being entered, in order to
trigger the right script), enhancements to savecore(8) to know how to
extract textdumps, adds a ddb(8) command line tool so that userspace can
manage DDB scripts from outside the debugger, extensions to the ddb(4) man
page, and a new textdump(4) man page.
There are a number of known limitations; I've tried to document them at
the top of the pertinent files where I am aware of them. I also regret to
say that to date I've been able to test only on i386, and not other
platforms. I'd welcome any feedback -- I'd like to get these changes into CVS
in the next week or two.
Robert N M Watson
Computer Laboratory
University of Cambridge
More information about the freebsd-arch
mailing list