On Fri, Apr 07, 2023 at 09:52:52AM +0200, Otto Moerbeek wrote: > > Hi, > > > > This is work in progress. I have to think if the flags to kdump I'm > > introducing should be two or a single one. > > > > Currently, malloc.c can be compiled with MALLOC_STATS defined. If run > > with option D it dumps its state to a malloc.out file at exit. This > > state can be used to find leaks amongst other things. > > > > This is not ideal for pledged processes, as they often have no way to > > write files. > > > > This changes malloc to use utrace(2) for that. > > > > As kdump has no nice way to show those lines without all extras it > > normally shows, so add two options to it to just show the lines. > > > > To use, compile and install libc with MALLOC_STATS defined. > > > > Run : > > > > $ MALLOC_OPTIONS=D ktrace -tu your_program > > ... > > $ kdump -hu > > > > Feedback appreciated.
I can't really comment on malloc(3) stuff, but I agree that utrace(2) is a good way to get information outside a pledged process. I tend to think it is safe to use it, as the pledged process need cooperation from outside to exfiltrate informations (a process with permission to call ktrace(2) on this pid). Please note it is a somehow generic problem: at least profiled processes would also get advantage of using it. Regarding kdump options, I think that -u option should implies -h (no header). Does it would make sens to considere a process using utrace(2) with several interleaved records for different sources ? A process with MALLOC_OPTIONS=D and profiling enabled for example ? An (another) option on kdump to filter on utrace label would be useful in such case, or have -u mandate a label to filter on. $ MALLOC_OPTIONS=D ktrace -tu your_program $ kdump -u mallocdumpline and for profiling: $ kdump -u profil > gmon.out $ gprof your_program gmon.out Thanks. -- Sebastien Marie
