2019-10-29 20:13:23

by Jim Cromie

Subject: [PATCH 06/16] dyndbg: fix overcounting of ram used by dyndbg

during dyndbg init, verbose logging prints its ram overhead. But it
counted strlens of all ddebug callsite entries, which are full of
pointers into shared __dyndbg memory, and shouldnt be counted at all
(since theyre already in the __dyndbg section)

Signed-off-by: Jim Cromie <[email protected]>
---
lib/dynamic_debug.c | 8 ++------
1 file changed, 2 insertions(+), 6 deletions(-)

diff --git a/lib/dynamic_debug.c b/lib/dynamic_debug.c
index 540ca0861cf0..4ce0c53cdcfd 100644
--- a/lib/dynamic_debug.c
+++ b/lib/dynamic_debug.c
@@ -992,7 +992,6 @@ static int __init dynamic_debug_init(void)
char *cmdline;
int ret = 0;
int n = 0, entries = 0, modct = 0;
- int verbose_bytes = 0;

if (__start___dyndbg == __stop___dyndbg) {
pr_warn("_ddebug table is empty in a CONFIG_DYNAMIC_DEBUG build\n");
@@ -1003,9 +1002,6 @@ static int __init dynamic_debug_init(void)
iter_start = iter;
for (; iter < __stop___dyndbg; iter++) {
entries++;
- verbose_bytes += strlen(iter->modname) + strlen(iter->function)
- + strlen(iter->filename) + strlen(iter->format);
-
if (strcmp(modname, iter->modname)) {
modct++;
ret = ddebug_add_module(iter_start, n, modname);
@@ -1022,9 +1018,9 @@ static int __init dynamic_debug_init(void)
goto out_err;

ddebug_init_success = 1;
- vpr_info("%d modules, %d entries and %d bytes in ddebug tables, %d bytes in (readonly) verbose section\n",
+ vpr_info("%d modules, %d entries and %d bytes in ddebug tables, %d bytes in (readonly) __dyndbg section\n",
modct, entries, (int)(modct * sizeof(struct ddebug_table)),
- verbose_bytes + (int)(__stop___dyndbg - __start___dyndbg));
+ (int)(__stop___dyndbg - __start___dyndbg));

/* apply ddebug_query boot param, dont unload tables on err */
if (ddebug_setup_string[0] != '\0') {
--
2.21.0

2019-10-29 20:48:55

by Rasmus Villemoes

Subject: Re: [PATCH 06/16] dyndbg: fix overcounting of ram used by dyndbg

On 29/10/2019 21.00, Jim Cromie wrote:
> during dyndbg init, verbose logging prints its ram overhead. But it
> counted strlens of all ddebug callsite entries, which are full of
> pointers into shared __dyndbg memory, and shouldnt be counted at all
> (since theyre already in the __dyndbg section)

Hm. I agree we're probably overcounting, but the strings themselves do
not live in __dyndbg, but in .rodata. It's true that __FILE__ (and maybe
in a few cases ->format) get's deduplicated and by the nature of
__func__, ->function points at a unique-per-function string.

So I think the commit log is a bit misleading. However, I think the
patch is a good idea anyway: avoiding 4N strlen() calls during boot is a
good thing. And if anybody wants to know how much memory is used by the
pointed-to-strings, well, "wc dynamic_debug/control" should give an idea.

Rasmus