Re: kernel dtrace and current

On Sun, Sep 13, 2015 at 04:23:23PM +0300, Alexander V. Chernikov wrote:
> Hello all,
> 
> I keep running in
> "dtrace: failed to compile script: "/usr/lib/dtrace/psinfo.d", line 39: failed to copy type of 'pr_uid': Type information is in parent and unavailable"
> message more and more often while trying to trace different -current kernels.
> 
> Typically the reason besides that is the number of types embedded in kernel CTF:
> ctfdump -S /boot/kernel/kernel | awk '$0~/of types/{print$6}'
> 37160
> 
> We are bound to 32k of types by CTF format (and numbers above 32k (e.g.w/ highest bit set) are considered "child" types with the information stored in "parent").
> ctfmerge ignores this fact and instead of yelling emits type indices above 32k. On the other hand, libctf sees such indices while parsing sections and since there is no
> "parent" for kernel, it emits the error above and stops.
> 
> Thankfully, r287234 really improved the situation for ctfmerge, but there are still several thousands of identical structures and the total number is close to 32k.

r281797 and r287234 should have fixed most instances of duplicate type
definitions. At the moment, amd64 GENERIC and GENERIC-NODEBUG have
roughly 25K types in their respective CTF containers; there is a small
handful of duplicates, but at least some of them are legitimate (some
pairs of drivers redefine the same types, e.g. aac(4)/aacraid(4) or
mps(4)/mpr(4)).

Could you post a config that results in the large number of duplicates
you mention above?

> 
> Personally I solved this by removing unneeded devices from GENERIC-inherited configs.
> I wonder, however if this can be handled better.

FWIW, removing old drivers from GENERIC would be straightforward if we
could auto-load KLDs based on device IDs.

> 
> E.g. either show better error in dtrace(1) or make ctfmerge fail causing kernel build to stop (since we asked for dtrace but in reality it wouldn't work), or remove some stale devices from GENERIC, or .something totally different?

One more radical option is to extend the width of CTF type IDs. I've
been holding off on doing this for a few reasons:
- Doing so would change the binary format, making us incompatible with
  the reference CTF code in illumos.
- Type IDs are embedded in quite a few places in the various CTF
  structures, so enlarging them from 16 bits to 32 bits will bloat
  CTF containers somewhat.
- I was under the impression that r287234 addressed the problem
  sufficiently for now.

If type ID space is still a problem post-r287234, I think it's time to
just go ahead and change the format. But first I'd like to understand
the cause of the duplication you're seeing.