On Tue, 19 Nov 2019 at 13:22, Ingo Molnar mingo@kernel.org wrote:
- Ard Biesheuvel ard.biesheuvel@linaro.org wrote:
On Tue, 19 Nov 2019 at 12:10, Ingo Molnar mingo@kernel.org wrote:
- Sean Christopherson sean.j.christopherson@intel.com wrote:
x86's asm/realmode.h, which defines low level structures, variables and helpers used to bring up APs during SMP boot, ends up getting included in practically every nook and cranny of the kernel because the address used by ACPI for resuming from S3 also happens to be stored in the real mode header, and ACPI bleeds the dependency into its widely included headers.
As a result, modifying realmode.h for even the most trivial change to the boot code triggers a full kernel rebuild, which is frustrating to say the least as it some of the most difficult code to get exactly right *and* is also some of the most functionally isolated code in the kernel.
To break the kernel's widespread dependency on realmode.h, add a wrapper in the aforementioned ACPI S3 code to access the real mode header instead of derefencing the header directly in asm/acpi.h and thereby exposing it to the world via linux/acpi.h.
Build tested on x86 with allyesconfig and allmodconfig, so hopefully there aren't more build issues lurking, but at this point it wouldn't surprise me in the least if this somehow manages to break the build.
Based on tip/master, commit ceceaf1f12ba ("Merge branch 'WIP.x86/cleanups'").
Patch Synopsis:
- Patches 01-09 fix a variety of build errors that arise when patch 12 drops realmode.h from asm/acpi.h. Most of the errors are quite absurb as they have no relation whatsoever to x86's RM boot code, but occur because realmode.h happens to include asm/io.h.
Yeah, these kind of parasitic header dependencies are the main driving force behind kernel header spaghetti hell: it's super easy to add a new header, but very hard to remove them...
Hence they practically only accumulate.
As a result header removal patches get priority, from me at least. :-)
Patch 10 removes a spurious include of realmode.h from an ACPI header.
Patches 11 and 12 implement the wrapper and move it out of acpi.h.
So if the ACPI maintainers are fine with -tip carrying patches #11 and #12 then I'd be glad to route these patches upstream.
I've applied them to tip:WIP.core/headers as a work-in-progress tree, and I'm testing them on randconfigs to make sure there's no broken dependencies. I'll wait for the ACPI acks.
I edited the title of patch 12 slightly, to:
c8bceb321209: x86/ACPI/sleep: Move acpi_wakeup_address() definition into sleep.c, remove <asm/realmode.h> from <asm/acpi.h>
to make sure the big header dependency change is obvious at first sight.
I'm fine with the patches but can we drop the fixes headers please? This doesn't actually fix anything, and touching early boot stuff for no good reason should be avoided imo.
Agreed and done.
Thanks Ingo