It took way too long, but I finally added support for tier 1 and 2 architectures. Besides that, the branching part of the interface changed to use likely/unlikely variants instead of true/false.

The mechanism is now refactored to use disable_wp() instead of the separate pmap scheme.

In D46379#1058143, @markj wrote:

db_write_bytes doesn’t bother with any of this, instead disabling WP on amd64 and using the DMAP on arm64…

While it's true that one can use the DMAP on arm64, that's hardly a desirable state of affairs. Assuming that the portion of the DMAP covered by the kernel text will be readonly, on arm64 and other platforms we'd instead need to temporarily relax protections on the 4KB kernel map page containing the target address, modify the text, and re-apply RO protections. If the first step involves demoting 2MB mappings, the last step should be careful to recreate them; I think this last piece is missing (see pmap_change_prot() on amd64 and arm64), but the use of large mappings for kernel text is important for performance. Perhaps that might be a more fruitful direction to consider.

In D46379#1057083, @markj wrote:

From the perspective of a reviewer, this patch is doing at least two things: adding a new mechanism to modify kernel text, and adding a new interface to define patch points. They should really be considered separately.