Linus notes [1] that the introduction of new code that uses VM_BUG_ON()
is just as bad as BUG_ON(), because it will crash the kernel on
distributions that enable CONFIG_DEBUG_VM (like Fedora):
VM_BUG_ON() has the exact same semantics as BUG_ON. It is literally
no different, the only difference is "we can make the code smaller
because these are less important". [2]
This resulted in a more generic discussion about usage of BUG() and
friends. While there might be corner cases that still deserve a BUG_ON(),
most BUG_ON() cases should simply use WARN_ON_ONCE() and implement a
recovery path if reasonable:
The only possible case where BUG_ON can validly be used is "I have
some fundamental data corruption and cannot possibly return an
error". [2]
As a very good approximation is the general rule:
"absolutely no new BUG_ON() calls _ever_" [2]
... not even if something really shouldn't ever happen and is merely for
documenting that an invariant always has to hold.
There is only one good BUG_ON():
Now, that said, there is one very valid sub-form of BUG_ON():
BUILD_BUG_ON() is absolutely 100% fine. [2]
While WARN will also crash the machine with panic_on_warn set, that's
exactly to be expected:
So we have two very different cases: the "virtual machine with good
logging where a dead machine is fine" - use 'panic_on_warn'. And
the actual real hardware with real drivers, running real loads by
users. [3]
The basic idea is that warnings will similarly get reported by users
and be found during testing. However, in contrast to a BUG(), there is a
way to actually influence the expected behavior (e.g., panic_on_warn)
and to eventually keep the machine alive to extract some debug info.
Ingo notes that not all WARN_ON_ONCE cases need recovery. If we don't ever
expect this code to trigger in any case, recovery code is not really
helpful.
I'd prefer to keep all these warnings 'simple' - i.e. no attempted
recovery & control flow, unless we ever expect these to trigger.
[4]
There have been different rules floating around that were never properly
documented. Let's try to clarify.
[1] https://lkml.kernel.org/r/CAHk-=wiEAH+ojSpAgx_Ep=NKPWHU8AdO3V56BXcCsU97oYJ1EA@xxxxxxxxxxxxxx
[2] https://lore.kernel.org/r/CAHk-=wg40EAZofO16Eviaj7mfqDhZ2gVEbvfsMf6gYzspRjYvw@xxxxxxxxxxxxxx
[3] https://lore.kernel.org/r/CAHk-=wgF7K2gSSpy=m_=K3Nov4zaceUX9puQf1TjkTJLA2XC_g@xxxxxxxxxxxxxx
Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
---
Documentation/process/coding-style.rst | 27 ++++++++++++++++++++++++++
1 file changed, 27 insertions(+)
diff --git a/Documentation/process/coding-style.rst b/Documentation/process/coding-style.rst
index 03eb53fd029a..a6d81ff578fe 100644
--- a/Documentation/process/coding-style.rst
+++ b/Documentation/process/coding-style.rst
@@ -1186,6 +1186,33 @@ expression used. For instance:
#endif /* CONFIG_SOMETHING */
+22) Do not crash the kernel
+---------------------------
+
+Do not add new code that uses BUG(), BUG_ON(), VM_BUG_ON(), ... to crash
+the kernel if certain conditions are not met. Instead, use WARN_ON_ONCE()
+with recovery code (if reasonable) instead. Unavoidable data corruption /
+security issues might be a very rare exception to this rule and need good
+justification.
+
+There is no need for WARN_ON_ONCE() recovery code if we never expect it to
+possibly trigger unless something goes seriously wrong or some other code
+is changed to break invariants. Note that VM_WARN_ON_ONCE() cannot be used
+in conditionals.
+
+Usage of WARN() and friends is fine for something that is not expected to
+be triggered easily. ``panic_on_warn`` users are not an excuse to not use
+WARN(): whoever enables ``panic_on_warn`` explicitly asked the kernel to
+crash in this case.
+
+However, WARN() and friends should not be used for something that is
+expected to trigger easily, for example, by user space. pr_warn_once()
+might be a reasonable replacement to notify the user.
+
+Note that BUILD_BUG_ON() is perfectly fine because it will make compilation
+fail instead of crashing the kernel.
+
+
Appendix I) References
----------------------
--
2.37.1
