Re: [Xen-devel] [PATCH v3] CODING_STYLE: Document how to handle unexpected conditions

[Jan Beulich <jbeulich@xxxxxxxx>]

On 09.12.2019 12:29, George Dunlap wrote:
> --- a/CODING_STYLE
> +++ b/CODING_STYLE
> @@ -133,3 +133,97 @@ the end of files.  It should be:
>   * indent-tabs-mode: nil
>   * End:
>   */
> +
> +Handling unexpected conditions
> +------------------------------
> +
> +GUIDELINES:
> +
> +Passing errors up the stack should be used when the caller is already
> +expecting to handle errors, and the state when the error was
> +discovered isn’t broken, or isn't too hard to fix.
> +
> +domain_crash() should be used when passing errors up the stack is too
> +difficult, and/or when fixing up state of a guest is impractical, but
> +where fixing up the state of Xen will allow Xen to continue running.
> +This is particularly appropriate when the guest is exhibiting behavior
> +well-behaved guest should.

DYM "shouldn't"?

> +BUG_ON() should be used when you can’t pass errors up the stack, and
> +either continuing or crashing the guest would likely cause an
> +information leak or privilege escalation vulnerability.
> +
> +ASSERT() IS NOT AN ERROR HANDLING MECHANISM.  ASSERT is a way to move
> +detection of a bug earlier in the programming cycle; it is a
> +more-noticeable printk.  It should only be added after one of the
> +other three error-handling mechanisms has been evaluated for
> +reliability and security.
> +
> +RATIONALE:
> +
> +It's frequently the case that code is written with the assumption that
> +certain conditions can never happen.  There are several possible
> +actions programmers can take in these situations:
> +
> +* Programmers can simply not handle those cases in any way, other than
> +perhaps to write a comment documenting what the assumption is.
> +
> +* Programmers can try to handle the case gracefully -- fixing up
> +in-progress state and returning an error to the user.
> +
> +* Programmers can crash the guest.
> +
> +* Programmers can use ASSERT(), which will cause the check to be
> +executed in DEBUG builds, and cause the hypervisor to crash if it's
> +violated
> +
> +* Programmers can use BUG_ON(), which will cause the check to be
> +executed in both DEBUG and non-DEBUG builds, and cause the hypervisor
> +to crash if it's violated.
> +
> +In selecting which response to use, we want to achieve several goals:
> +
> +- To minimize risk of introducing security vulnerabilities,
> +  particularly as the code evolves over time
> +
> +- To efficiently spend programmer time
> +
> +- To detect violations of assumptions as early as possible
> +
> +- To minimize the impact of bugs on production use cases
> +
> +The guidelines above attempt to balance these:
> +
> +- When the caller is expecting to handle errors, and there is no
> +broken state at the time the unexpected condition is discovered, or
> +when fixing the state is straightforward, then fixing up the state and
> +returning an error is the most robust thing to do.  However, if the
> +caller isn't expecting to handle errors, or if the state is difficult
> +to fix, then returning an error may require extensive refactoring,
> +which is not a good use of programmer time when they're certain that
> +this condition cannot occur.
> +
> +- BUG_ON() will stop all hypervisor action immediately.  In situations
> +where continuing might allow an attacker to escalate privilege, a
> +BUG_ON() can change a privilege escalation or information leak into a
> +denial-of-service (an improvement).  But in situations where
> +continuing (say, returning an error) might be safe, then BUG_ON() can
> +change a benign failure into denial-of-service (a degradation).
> +
> +- domain_crash() is similar to BUG_ON(), but with a more limited
> +effect: it stops that domain immediately.  In situations where
> +continuing might cause guest or hypervisor corruption, but destroying
> +the guest allows the hypervisor to continue, this can change a more
> +serious bug into a guest denial-of-service.  But in situations where
> +returning an error might be safe, then domain_crash() can change a
> +benign failure into a guest denial-of-service.

Perhaps further put emphasis on the call tree still getting unwound
normally, which may imply further actions on the (now dying) domain
taken. Unfortunately it's not unusual for people to forget this; I
think the IOMMU code in particular was (hopefully isn't so much
anymore) a "good" example of this.

Jan

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxxx
https://lists.xenproject.org/mailman/listinfo/xen-devel