[Crash-utility] Re: [PATCH] crash: symbols: Optimize kernel module pseudo-symbol matching overhead

On 5/19/26 7:02 AM, Tao Liu wrote: Hi Tao, Thank you for the suggestion and for questioning the improvement claim. I went back and did a direct end-to-end measurement for this patch on the following workflow: 1. start ./crash /proc/kcore 2. run sym -l 3. quit crash I compared commit df23d5b7c9e0 against its parent commit on the same machine and kernel environment, using /proc/kcore as the dumpfile and driving the interaction through a pseudo-terminal to stay close to actual crash console usage. Based on the measurements I got, this patch did not show a meaningful improvement for that workflow. On my test setup, the patched version was actually slightly slower overall, so I agree that the performance benefit is not sufficiently justified here. Given that result, I think it is better for me to withdraw this patch rather than keep the claim without solid evidence. That said, I would appreciate it if you could take a closer look at the previous patch in the series: crash: symbols: Optimize symbol string matching overhead in numeric_forward() That patch targets a more focused hot path, and in my testing it did show measurable improvement, so I believe it is the one more worth considering on its own. Thanks again for the careful review and for asking for the measurement data. Best regards, Rui > On Tue, May 19, 2026 at 5:27 AM Rui Qi <qirui.001(a)bytedance.com&gt; wrote: >> >> In a modern Linux environment with numerous kernel modules loaded, the >> crash utility must process and sort tens of thousands (often 50,000+) of >> module symbols. The symbol sorting function (compare_syms) and module >> loading loops frequently check for pseudo-symbols using macros like >> MODULE_PSEUDO_SYMBOL(), MODULE_START(), and others. >> >> These macros heavily rely on the STREQ()/STRNEQ() macros, which >> internally expand to string_exists() checks followed by a full strcmp() >> or strncmp(). Because these macros are evaluated millions of times during >> module initialization and sorting (O(N log N) complexity), the function >> call overhead of strcmp() becomes a noticeable performance bottleneck. >> >> By explicitly checking the first character of the symbol name >> (e.g., (sp)->name[0] == '_') before invoking STRNEQ(), we introduce a >> lightweight "early reject" mechanism. Since most regular module symbols >> do not start with an underscore, this short-circuits the evaluation for >> the vast majority of symbols, completely avoiding the overhead of >> strncmp() macro expansion. >> >> Additionally, since the symbol name could potentially be NULL, we safely >> guard the character access with an explicit non-null check >> ((sp)->name &&) to prevent segmentation faults. >> >> This patch applies the early reject optimization to: >> 1. All MODULE_* pseudo-symbol macros. >> 2. The compare_syms() function for "__insmod" and "_MODULE_START_" checks. >> >> This cumulative optimization yields a measurable performance improvement >> in module symbol sorting and initialization speed. >> >> Signed-off-by: Rui Qi <qirui.001(a)bytedance.com&gt; >> --- >> symbols.c | 20 ++++++++++---------- >> 1 file changed, 10 insertions(+), 10 deletions(-) >> >> diff --git a/symbols.c b/symbols.c >> index 736d9d96d606..9b0de45e9a95 100644 >> --- a/symbols.c >> +++ b/symbols.c >> @@ -1314,14 +1314,14 @@ symname_hash_search(struct syment *table[], char *name) >> * Output for sym -[lL] command. >> */ >> >> -#define MODULE_PSEUDO_SYMBOL(sp) (STRNEQ((sp)->name, "_MODULE_")) >> +#define MODULE_PSEUDO_SYMBOL(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_")) >> >> -#define MODULE_START(sp) (STRNEQ((sp)->name, "_MODULE_START_")) >> -#define MODULE_END(sp) (STRNEQ((sp)->name, "_MODULE_END_")) >> -#define MODULE_INIT_START(sp) (STRNEQ((sp)->name, "_MODULE_INIT_START_")) >> -#define MODULE_INIT_END(sp) (STRNEQ((sp)->name, "_MODULE_INIT_END_")) >> -#define MODULE_SECTION_START(sp) (STRNEQ((sp)->name, "_MODULE_SECTION_START")) >> -#define MODULE_SECTION_END(sp) (STRNEQ((sp)->name, "_MODULE_SECTION_END")) >> +#define MODULE_START(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_START_")) >> +#define MODULE_END(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_END_")) >> +#define MODULE_INIT_START(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_INIT_START_")) >> +#define MODULE_INIT_END(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_INIT_END_")) >> +#define MODULE_SECTION_START(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_SECTION_START")) >> +#define MODULE_SECTION_END(sp) ((sp)->name && (sp)->name[0] == '_' && STRNEQ((sp)->name, "_MODULE_SECTION_END")) >> >> #define MODULE_MEM_START(sp,t) (STRNEQ((sp)->name, module_tag[t].start)) >> #define MODULE_MEM_END(sp,t) (STRNEQ((sp)->name, module_tag[t].end)) >> @@ -3419,12 +3419,12 @@ compare_syms(const void *v1, const void *v2) >> s2 = (struct syment *)v2; >> >> if (s1->value == s2->value) { >> - if (STRNEQ(s1->name, "__insmod")) >> + if (s1->name && s1->name[0] == '_' && STRNEQ(s1->name, "__insmod")) >> return -1; >> - if (STRNEQ(s2->name, "__insmod")) >> + if (s2->name && s2->name[0] == '_' && STRNEQ(s2->name, "__insmod")) >> return 1; >> if (MODULE_MEM_START(s2, MOD_TEXT) || >> - STRNEQ(s2->name, "_MODULE_START_")) >> + s2->name && s2->name[0] == '_' && STRNEQ(s2->name, "_MODULE_START_")) >> return 1; >> /* Get pseudo section name. */ >> if (MODULE_SECTION_START(s1)) >> -- >> 2.20.1 >> -- >> Crash-utility mailing list -- devel(a)lists.crash-utility.osci.io >> To unsubscribe send an email to devel-leave(a)lists.crash-utility.osci.io >> https://${domain_name}/admin/lists/devel.lists.crash-utility.osci.io/ >> Contribution Guidelines: https://github.com/crash-utility/crash/wiki >> >