On 11/7/23, Kees Cook <keescook@xxxxxxxxxxxx> wrote: > On Tue, Nov 07, 2023 at 10:23:16PM +0100, Mateusz Guzik wrote: >> If the patch which dodges second lookup still somehow appears slower a >> flamegraph or other profile would be nice. I can volunteer to take a >> look at what's going on provided above measurements will be done and >> show funkyness. > > When I looked at this last, it seemed like all the work done in > do_filp_open() (my patch, which moved the lookup earlier) was heavier > than the duplicate filename_lookup(). > > What I didn't test was moving the sched_exec() before the mm creation, > which Peter confirmed shouldn't be a problem, but I think that might be > only a tiny benefit, if at all. > > If you can do some comparisons, that would be great; it always takes me > a fair bit of time to get set up for flame graph generation, etc. :) > So I spawned *one* process executing one statocally linked binary in a loop, test case from http://apollo.backplane.com/DFlyMisc/doexec.c . The profile is definitely not what I expected: 5.85% [kernel] [k] asm_exc_page_fault 5.84% [kernel] [k] __pv_queued_spin_lock_slowpath [snip] I'm going to have to recompile with lock profiling, meanwhile according to bpftrace (bpftrace -e 'kprobe:__pv_queued_spin_lock_slowpath { @[kstack()] = count(); }') top hits would be: @[ __pv_queued_spin_lock_slowpath+1 _raw_spin_lock+37 __schedule+192 schedule_idle+38 do_idle+366 cpu_startup_entry+38 start_secondary+282 secondary_startup_64_no_verify+381 ]: 181 @[ __pv_queued_spin_lock_slowpath+1 _raw_spin_lock_irq+43 wait_for_completion+141 stop_one_cpu+127 sched_exec+165 bprm_execve+328 do_execveat_common.isra.0+429 __x64_sys_execve+50 do_syscall_64+46 entry_SYSCALL_64_after_hwframe+110 ]: 206 I did not see this coming for sure. I'll poke around maybe this weekend. -- Mateusz Guzik <mjguzik gmail.com>
