For now, taint with TAINT_USER when loading a patch module so that the
user can always detect when a kpatch module has been previously loaded.
Eventually we will want a dedicated TAINT_KPATCH flag in the kernel.
When CONFIG_MODVERSIONS is enabled, loading of the patch module fails
with "no symbol version for kpatch_register". When building the patch
module, we need to point it to the core module's Module.symvers file.
This also works when CONFIG_MODVERSIONS is disabled, since
Module.symvers is created regardless.
When multiple patch modules patch the same function, there's no need to
patch all the intermediate functions. Just hook them all into the
original function and use the ftrace handler to find the newest one.
Also use a mutex in the register/unregister functions to protect changes
to kpatch_num_registered, kpatch_func_hash and calls to the ftrace
functions by other register/unregister invocations.
Use a mutex in the register/unregister functions to protect changes to
kpatch_num_registered, kpatch_func_hash and calls to the ftrace
functions by other register/unregister invocations.
There's no need to zero out the kpatch funcs array. The addr fields are
initialized by the patch module, the mod field is intialized by the core
module, and the node struct doesn't need to be initialized because its
fields are overwritten by hash_add.
My apologies for the size of this commit. I combined these two features
(updating API and using a hash table) into a single commit because their
implementations are tightly coupled and I didn't want to have to add
support for the old kpatch_funcs array with the new API just for the
sake of splitting up the commit :-)
- Update the core module API to get a more clear separation between core
module and patch module. This is cleaner and will help our case for
getting the core module merged upstream into the kernel.
- Convert the old kpatch_funcs array into a hash table. This is so much
nicer performance-wise and everything-else-wise than that ugly old
array.
- Do the incremental patching in stop machine. This ensures that the
funcs hash is up to date and we don't miss anything.
- Disable preemption in the ftrace handler when accessing the func hash.
That way we don't get conflicts with the stop_machine handler updating
the hash.
Print the loading/unloading messages after they have successfully
completed. Using the KERN_NOTICE log level which corresponds to a
"normal but significant condition."
Preemption shouldn't cause a problem with determining activeness safety.
Even if a thread is preempted, it'll be on the backtrace.
We may need to disable preemption when reading the kpatch_funcs array,
but I'm removing that comment for now because the kpatch_funcs array
will soon be replaced by a much better data structure, and we'll deal
with proper synchronization then.
Long ago, the kpatch_trampoline required being written in assembler, but
that's no longer needed now that it integrates nicely with ftrace.
Move it to a C function and rename it kpatch_ftrace_handler.
Build and install the kpatch core module with make and make install,
rather than building it every time with kpatch build.
The only downside to this approach is that the user has to make and make
install kpatch every time they get a new kernel. But this is only
temporary, until the kpatch module is delivered in an RPM.
- setup the makefiles to support "make" and "make install", which builds
the kpatch-build tools and installs everything in /usr/local.
- update kpatch-build to support new paths
- add "kpatch build" wrapper around kpatch-build
This changes the way the trampoline code works, thanks to a suggestion
by Steve Rostedt. Before, the trampoline was mucking with the stack
pointer and other registers, and jumping to the new function directly.
With this change, all it does is set regs->ip to the address of the new
function and return back to ftrace. When ftrace returns, it will return
to the beginning of the new function.
- fix real issue with 0's in the middle of a merged section (wrong
alignment)
- show patch util output in case it asks a question so it doesn't
silently fail
- fix issue with relocation of local objects (because they become global
objects)
- allow changes to .rela.initcall*. they should be instead caught by
the relocation comparison code.
- fix issue in compare symbols when the symbol section index has changed
but the sections themselves are the same
- in compare_symbols, when a new STT_SECTION symbol is added, ignore it.
it will be caught instead by the section comparison code
- fix issue in kpatch-gcc script that was causing gcc command lines
containing quotes to fail
The end-to-end patching works. From object analysis to generation to
runtime patching. It's still missing the scripting piece that will only
take a patch and kernel source dir as input.
