On Thu, Jan 28, 2021 at 5:16 PM [off-list ref] wrote:
On Thu, Jan 28, 2021 at 4:57 PM Valdis Klētnieks [off-list ref]
wrote:
quoted
On Thu, 28 Jan 2021 12:11:54 -0700, jim.cromie@gmail.com said:
quoted
In my hacking, Im finding this useful.
it adds a version of KBUILD_MODNAME without the quotes
OK, I'll bite. When and how is this useful?
Let me add more context here.
1st, the long saga.
v1 - here, back in august last year.
https://lore.kernel.org/kernelnewbies/20200805183023.586590-1-jim.cromie@gmail.com/
not really pertinent here.
gregkh pretty much said send it to lkml
v2 - RFC to LKML
https://lore.kernel.org/lkml/20201225201944.3701590-1-jim.cromie@gmail.com/
this didnt get much attention, no answers to my questions
the 0000 is still completely accurate, I cannot improve upon it,
though many questions in the commit messages Ive answered myself.
(or at least made some choices ;)
v3 - sits in github now
almost identical 1-16/19 with v2
https://github.com/jimc/linux.git dyndbg-next
I was hoping to pull off a context-free question,
focusing on just the Makefile.lib patch's suitability.
I could send the whole series here, or to LKML, or both.
I can hope a quick perusal will get you to the interesting parts
heres my use
Im looking to define a module-unique header record,
and to create storage for it in a separate section.
Then append that section to the module's __dyndbg section.
The pragma is leftover scaffolding from trying to figure out
that I needed the _sym_ version also, without the quotes.
I un-commented for demo only.
#define DEFINE_DYNAMIC_DEBUG_TABLE_(_sym_,_mod_) \
weak struct _ddebug \
__used __aligned(8) \
__section(".gnu.linkonce." _mod_ ".dyndbg") \
_sym_##_dyndbg_base = { \
.site = &_sym_##_dyndbg_site, \
.format = _mod_, \
.lineno = 0 \
}
#define DEFINE_DYNAMIC_DEBUG_TABLE() \
DEFINE_DYNAMIC_DEBUG_TABLE_(KBUILD_MODSYM, KBUILD_MODNAME);
I cannot make the _mod_ ## work at the same time as
allowing "foo" _mod_ "buzz" catenation, therefore the new _sym_
This DEFINE is "called" from the bottom of include/linux/dynamic_debug.h
the combo of weak and ".gnu.linkonce." appear to resolve all
all the redundant definitions created by each C files' inclusion
of the header.
Inspiration came from .gnu.linkonce.this_module
--- a/include/asm-generic/vmlinux.lds.h
+++ b/include/asm-generic/vmlinux.lds.h
#define DYNAMIC_DEBUG_DATA() \
. = ALIGN(8); \
__start___dyndbg_sites = .; \
- KEEP(*(__dyndbg_sites __dyndbg_sites.header)) \
+ KEEP(*(__dyndbg_sites .gnu.linkonce.*.dyndbg_site)) \
__stop___dyndbg_sites = .; \
__start___dyndbg = .; \
- KEEP(*(__dyndbg __dyndbg.header)) \
+ KEEP(*(__dyndbg .gnu.linkonce.*.dyndbg)) \
__stop___dyndbg = .;
#else
#define DYNAMIC_DEBUG_DATA()
that puts the header record into the output section, right after the
module's
set of ddebug's (pr_debug callsite descriptors)
By getting the header into a fixed position relative to a callsite,
its possible to go from callsite-descriptor to the header
(after .module_index is initialized with the offset, determined at _init),
then use a single per-module pointer to the separate vector of "decorator"
data.
Once the 2 vectors of __dyndbg and __dyndbg_site records are indexed by N,
I can drop the .site pointer from 1 to other, and shrink the footprint back
to
~parity (with + 1 record / module overhead) worst case.
with 2 vectors fully decoupled (pointer-wise),
33 more lines