livepatch: Apply vmlinux-specific KLP relocations early

livepatch: Apply vmlinux-specific KLP relocations early

KLP relocations are livepatch-specific relocations which are applied to
a KLP module's text or data.  They exist for two reasons:

  1) Unexported symbols: replacement functions often need to access
     unexported symbols (e.g. static functions), which "normal"
     relocations don't allow.

  2) Late module patching: this is the ability for a KLP module to
     bypass normal module dependencies, such that the KLP module can be
     loaded *before* a to-be-patched module.  This means that
     relocations which need to access symbols in the to-be-patched
     module might need to be applied to the KLP module well after it has
     been loaded.

Non-late-patched KLP relocations are applied from the KLP module's init
function.  That usually works fine, unless the patched code wants to use
alternatives, paravirt patching, jump tables, or some other special
section which needs relocations.  Then we run into ordering issues and
crashes.

In order for those special sections to work properly, the KLP
relocations should be applied *before* the special section init code
runs, such as apply_paravirt(), apply_alternatives(), or
jump_label_apply_nops().

You might think the obvious solution would be to move the KLP relocation
initialization earlier, but it's not necessarily that simple.  The
problem is the above-mentioned late module patching, for which KLP
relocations can get applied well after the KLP module is loaded.

To "fix" this issue in the past, we created .klp.arch sections:

  .klp.arch.{module}..altinstructions
  .klp.arch.{module}..parainstructions

Those sections allow KLP late module patching code to call
apply_paravirt() and apply_alternatives() after the module-specific KLP
relocations (.klp.rela.{module}.{section}) have been applied.

But that has a lot of drawbacks, including code complexity, the need for
arch-specific code, and the (per-arch) danger that we missed some
special section -- for example the __jump_table section which is used
for jump labels.

It turns out there's a simpler and more functional approach.  There are
two kinds of KLP relocation sections:

  1) vmlinux-specific KLP relocation sections

     .klp.rela.vmlinux.{sec}

     These are relocations (applied to the KLP module) which reference
     unexported vmlinux symbols.

  2) module-specific KLP relocation sections

     .klp.rela.{module}.{sec}:

     These are relocations (applied to the KLP module) which reference
     unexported or exported module symbols.

Up until now, these have been treated the same.  However, they're
inherently different.

Because of late module patching, module-specific KLP relocations can be
applied very late, thus they can create the ordering headaches described
above.

But vmlinux-specific KLP relocations don't have that problem.  There's
nothing to prevent them from being applied earlier.  So apply them at
the same time as normal relocations, when the KLP module is being
loaded.

This means that for vmlinux-specific KLP relocations, we no longer have
any ordering issues.  vmlinux-referencing jump labels, alternatives, and
paravirt patching will work automatically, without the need for the
.klp.arch hacks.

All that said, for module-specific KLP relocations, the ordering
problems still exist and we *do* still need .klp.arch.  Or do we?  Stay
tuned.

Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Joe Lawrence <joe.lawrence@redhat.com>
Acked-by: Miroslav Benes <mbenes@suse.cz>
Acked-by: Jessica Yu <jeyu@kernel.org>
Signed-off-by: Jiri Kosina <jkosina@suse.cz>