doc: patch author guide, first take

Been wanting to write this forever, have to start somewhere...
This commit is contained in:
Josh Poimboeuf 2015-11-16 14:59:41 -06:00
parent aab5240df8
commit 85b7a76acc
2 changed files with 154 additions and 0 deletions

View File

@ -322,6 +322,15 @@ Done! The kernel is now patched.
$ grep -i chunk /proc/meminfo $ grep -i chunk /proc/meminfo
VMALLOCCHUNK: 34359337092 kB VMALLOCCHUNK: 34359337092 kB
Patch Author Guide
------------------
Unfortunately, live patching isn't always as easy as the previous example, and
can have some major pitfalls if you're not careful. To learn more about how to
properly create live patches, see the [Patch Author
Guide](doc/patch-author-guide.md).
How it works How it works
------------ ------------

145
doc/patch-author-guide.md Normal file
View File

@ -0,0 +1,145 @@
kpatch Patch Author Guide
=========================
Because kpatch-build is relatively easy to use, it can be easy to assume that a
successful patch module build means that the patch is safe to apply. But in
fact that's a very dangerous assumption.
There are many pitfalls that can be encountered when creating a live patch.
This document attempts to guide the patch creation process. It's a work in
progress. If you find it useful, please contribute!
Patch Analysis
--------------
kpatch provides _some_ guarantees, but it does not guarantee that all patches
are safe to apply. Every patch must also be analyzed in-depth by a human.
The most important point here cannot be stressed enough. Here comes the bold:
**Do not blindly apply patches. There is no subsitute for human analysis and
reasoning on a per-patch basis. All patches must be thoroughly analyzed by a
human kernel expert who completely understands the patch and the affected code
and how they relate to the live patching environment.**
kpatch vs livepatch vs kGraft
-----------------------------
This document assumes that the kpatch core module is being used. Other live
patching systems (e.g., livepatch and kGraft) have different consistency
models. Each comes with its own guarantees, and there are some subtle
differences. The guidance in this document applies **only** to kpatch.
Patch upgrades
--------------
Due to potential unexpected interactions between patches, it's highly
recommended that when patching a system which has already been patched, the
second patch should be a cumulative upgrade which is a superset of the first
patch.
Data structure changes
----------------------
kpatch patches functions, not data. If the original patch involves a change to
a data structure, the patch will require some rework, as changes to data
structures are not allowed by default.
Usually you have to get creative. There are several possible ways to handle
this:
### Change the code which uses the data structure
Sometimes, instead of changing the data structure itself, you can change the
code which uses it.
For example, consider this
[patch](http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=54a20552e1eae07aa240fa370a0293e006b5faed).
which has the following hunk:
```
@@ -3270,6 +3277,7 @@ static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
[SVM_EXIT_EXCP_BASE + PF_VECTOR] = pf_interception,
[SVM_EXIT_EXCP_BASE + NM_VECTOR] = nm_interception,
[SVM_EXIT_EXCP_BASE + MC_VECTOR] = mc_interception,
+ [SVM_EXIT_EXCP_BASE + AC_VECTOR] = ac_interception,
[SVM_EXIT_INTR] = intr_interception,
[SVM_EXIT_NMI] = nmi_interception,
[SVM_EXIT_SMI] = nop_on_interception,
```
`svm_exit_handlers[]` is an array of function pointers. The patch adds a
`ac_interception` function pointer to the array at index `[SVM_EXIT_EXCP_BASE +
AC_VECTOR]`. That change is incompatible with kpatch.
Looking at the source file, we can see that this function pointer is only
accessed by a single function, `handle_exit()`:
```
if (exit_code >= ARRAY_SIZE(svm_exit_handlers)
|| !svm_exit_handlers[exit_code]) {
WARN_ONCE(1, "svm: unexpected exit reason 0x%x\n", exit_code);
kvm_queue_exception(vcpu, UD_VECTOR);
return 1;
}
return svm_exit_handlers[exit_code](svm);
```
So an easy solution here is to just change the code to manually check for the
new case before looking in the data structure:
```
@@ -3580,6 +3580,9 @@ static int handle_exit(struct kvm_vcpu *vcpu)
return 1;
}
+ if (exit_code == SVM_EXIT_EXCP_BASE + AC_VECTOR)
+ return ac_interception(svm);
+
return svm_exit_handlers[exit_code](svm);
}
```
Not only is this an easy solution, it's also safer than touching data since
kpatch creates a barrier between the calling of old functions and new
functions.
### Use a kpatch load hook
If you need to change the contents of an existing variable in-place, you can
use the KPATCH_LOAD_HOOK macro to specify a function to be called when the
patch module is loaded.
Don't forget to protect access to the data as needed.
Also be careful when upgrading. If patch A has a load hook which writes to X,
and then you load patch B which is a superset of A, in some cases you may want
to prevent patch B from writing to X, if A is already loaded.
Examples needed.
### Use a shadow variable
If you need to add a field to an existing data structure, or even many existing
data structures, you can use the `kpatch_shadow_*()` functions.
Example needed (see shadow-newpid.patch in the integration tests directory).
Data semantic changes
---------------------
Sometimes, the data itself remains the same, but how it's used is changed. A
common example is locking semantic changes.
Example needed.
Init code changes
-----------------
Any code which runs in an `__init` function or during module or device
initialization is problematic, as it may have already run before the patch was
applied. The patch may require a load hook function which detects whether such
init code has run, and which rewrites or changes the original initialization to
force it into the desired state. Some changes involving hardware init are
inherently incompatible with live patching.