[review v2] jit: remove bp locations when unregistering jit code

["Simon Marchi (Code Review)" <gerrit@gnutoolchain-gerrit.osci.io>]

Simon Marchi has posted comments on this change.

Change URL: https://gnutoolchain-gerrit.osci.io/r/c/binutils-gdb/+/704
......................................................................


Patch Set 2:

> Patch Set 1:
> 
> > Patch Set 1:
> > 
> > Just trying to understand the problem better.  From what I understand, when the breakpoint locations must be updated following a code change event (e.g. solib getting loaded or unloaded), we call breakpoint_re_set, which goes through all breakpoint locations and sees if they must be updated.  Have you tryied calling this from jit_unregister_code?
> 
> Have tried it now, just in case - it doesn't work. As far as I understand the problem here is that from the gdb PoV breakpoint locations with the same address are the same and `locations_are_equal (existing_locations, b->loc)` condition is true. Because of that updating breakpoint locations won't actually install breakpoint traps - gdb will still think that it is already installed. However in JIT case that instruction memory was overwritten and doesn't have a trap anymore.
> 
> Sadly I don't have enough knowledge about gdb architecture to reason if the same problem can possibly manifest with non-jit object files.

There's still something that doesn't connect in my mind.  I may have a wrong picture of what's happening, so please correct me.

From my understanding of the JIT interface (https://sourceware.org/gdb/current/onlinedocs/gdb/JIT-Interface.html), unregistering code is made by the process calling __jit_debug_register_code (on which GDB has a special breakpoint) with action == JIT_UNREGISTER.  Registering code is made by the process calling __jit_debug_register_code with action == JIT_REGISTER.

So to unregister a jit region and register a new one (that would happen to have the exact same code at the exact same address as the previous one), the process would need to call __jit_debug_register_code twice, executing between the two events.  After unregistration, when the execution resumes, shouldn't there be something that deletes the breakpoint locations related to that objfile that was removed?  And then when __jit_debug_register_code for registering the new object, we would re-create brand new breakpoint locations?

Or is it that somehow, the unregistration and re-registration is made during the same stop?


-- 
Gerrit-Project: binutils-gdb
Gerrit-Branch: master
Gerrit-Change-Id: Id9133540d67fa0c4619ac88324b0349b89e4b2b1
Gerrit-Change-Number: 704
Gerrit-PatchSet: 2
Gerrit-Owner: Mihails Strasuns <mihails.strasuns@intel.com>
Gerrit-Reviewer: Mihails Strasuns <mihails.strasuns@intel.com>
Gerrit-CC: Simon Marchi <simon.marchi@polymtl.ca>
Gerrit-Comment-Date: Tue, 26 Nov 2019 16:58:51 +0000
Gerrit-HasComments: No
Gerrit-Has-Labels: No
Gerrit-MessageType: comment