On 1/18/23 20:57, Brooks Davis wrote:
On Wed, Jan 18, 2023 at 04:53:51PM -0700, Warner Losh wrote:
On Wed, Jan 18, 2023 at 3:07 PM Mitchell Horne <[email protected]> wrote:
On 1/17/23 12:38, Brooks Davis wrote:
The branch main has been updated by brooks:
URL:
https://cgit.FreeBSD.org/src/commit/?id=b75062f23431fbabef1e7d665cae270b144f71b1
commit b75062f23431fbabef1e7d665cae270b144f71b1
Author: Brooks Davis <[email protected]>
AuthorDate: 2023-01-17 16:36:15 +0000
Commit: Brooks Davis <[email protected]>
CommitDate: 2023-01-17 16:37:42 +0000
riscv: Fix thread0.td_kstack_pages init
Commit 0ef3ca7ae37c70e9dc83475dc2e68e98e1c2a418 initialized
thread0.td_kstack_pages to KSTACK_PAGES. Due to the lack of an
include of opt_kstack_pages.h it used the fallback value of 4 from
machine/param.h.
Does this mean that we could/should include opt_kstack_pages.h within
machine/param.h (under #ifdef _KERNEL)? This header is both a consumer
and provider of the KSTACK_PAGES definition, by virtue of the #ifndef. I
think the hidden dependency should be avoided, if possible.
No. Including opt_XXXX.h is never OK in our .h files. They are used in too
many places, some of which "cheat" and define _KERNEL becuse, well, they
need to get to the kernel bits.... That will break...
Riiiiiight, I was forgetting the _KERNEL liars always ruin the fun. You
are right, and the "never include opt_ headers in a header" rule is a
good one.
We could potentially use the __has_include extension. I don't think we
care about building the kernel with a compiler that isn't clang or gcc
and the usage pattern defined by gcc is safe for compilers that don't
define it. We could do something like:
#ifdef _KERNEL
#ifndef KSTACK_PAGES
#ifdef __has_include
#if __has_include("opt_kstack_pages.h")
#include "opt_kstack_pages.h"
#endif
#endif
#endif
#endif
<old #ifndef KSTACK_PAGES code>
Yeah this would work, but I think we can agree it's one step more
complicated than necessary for this edge-case which is unlikely to show
up again often, if at all. I say let's just add a comment to each
machine/param.h explaining the situation and that will be enough.
Mitchell
I do agree, however, that the current interface is less than ideal...
Of course, the problem at hand has been fixed and we want to keep direct
consumers of KSTACK_PAGES to a minimum, but I think the point still stands.
I think it's a good point, but the current way is likely the least-bad way
to accomplish things.
It would be much better if we could remove it from machine/param.h and
opt_XXX.h always defines it, even the default value when it's not otherwise
specified. However, we don't (currently) have a way to set default values
in config(8). We could add it, since the efforts at config++ have thus far
fallen flat....
I think this is probably the better direction to move. There aren't any
in-tree uses of KSTACK_PAGES so removing the definition should be fine.
-- Brooks
