fidzu : kernel

Found another interesting article, this time mentioning a tool I wrote long ago and that, at least for kernel object files, has been working for a long time without much care on my part: pahole, go read a bit about it at Will Cohen's "How to avoid wasting megabytes of memory a few bytes at a time" article.

Guess I should try running a companion script that tries to process all .o files in debuginfo packages to see how bad it is for non-kernel files, with all the DWARF changes over these years…

20 Jun 2017 3:49pm GMT

A reminder that our Early Bird registration rate is ending soon. The last day at the Early Bird rate of 400$ is Sunday June 18th. We are also almost sold out of Early Bird slots (15% left of our quota). Get yours soon!
Starting June 19th registration will be at the regular rate of 550$.
Please see the Attend page for info.

15 Jun 2017 11:20pm GMT

The last month or two has seen a lot of work simplifying the Linux-kernel RCU implementation, with more than 2700 net lines of code removed. The remainder of this post lists the user-visible changes, along with alternative ways to get the corresponding job done.

1. 
   
2. The infamous CONFIG_RCU_KTHREAD_PRIO Kconfig parameter is now defunct, but the rcutree.kthread_prio kernel boot parameter gets the job done.
   
3. The CONFIG_NO_HZ_FULL_SYSIDLE Kconfig parameter has kicked the bucket. There is no replacement because no one was using it. If you need it, revert the -rcu commit tagged by sysidle.2017.05.11a.
   
4. The CONFIG_PROVE_RCU_REPEATEDLY Kconfig parameter is no more. There is no replacement because as far as I know, no one has used it for many years. It was a great help in tracking down lockdep-RCU warnings back in the day, but these warnings are now sufficiently rare that finding them one boot at a time is no longer a problem. If you need it, do the obvious hacking on Kconfig and lockdep.c.
   
5. The CONFIG_SPARSE_RCU_POINTER Kconfig parameter now rests in peace. There is no replacement because there doesn't seem to be any reason for RCU's sparse checking to be the only such checking that is optional. If you really need to disable RCU's sparse checking, hand-edit the definition as needed.
   
6. The CONFIG_CLASSIC_SRCU Kconfig parameter bought the farm. This was only present to handle massive failures of the new Tree/Tiny SRCU implementations, but these appear to be quite reliable and should be used instead of Classic SRCU.
   
7. RCU's debugfs tracing is done for. As far as I know, I was the only real user, and I haven't used it in years. If you need it, revert the -rcu commit tagged by debugfs.2017.05.15a.
   
8. The CONFIG_RCU_NOCB_CPU_NONE, CONFIG_RCU_NOCB_CPU_ZERO, and CONFIG_RCU_NOCB_CPU_ALL Kconfig parameters have departed. Use the rcu_nocbs kernel boot parameter instead, which can do quite a bit more than those Kconfig parameters ever could.
   
9. Tiny RCU's event tracing and RCU CPU stall warnings are now pushing up daisies. The point of Tiny RCU is to be tiny and educational, and these added features were not helping reach either of these two goals. The replacement is to reproduce the problem with Tree RCU.
   
10. These changes should matter only to people running rcutorture:
    1. 
       
    2. The CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT and CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY Kconfig parameters have been entombed: Use the rcutree.gp_preinit_delay kernel boot parameter instead.
       
    3. The CONFIG_RCU_TORTURE_TEST_SLOW_INIT and CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY Kconfig parameters have given up the ghost: Use the rcutree.gp_init_delay kernel boot parameter instead.
       
    4. The CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP and CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY Kconfig parameters have passed on: Use the rcutree.gp_cleanup_delay kernel boot parameter instead.
       

There will probably be a few more simplifications in the near future, but this should be at least enough for one merge window!

14 Jun 2017 9:03pm GMT