| Message ID | 20220809034517.3867176-1-joel@joelfernandes.org (mailing list archive) |
|---|---|
| Headers | show |
| Series | Implement call_rcu_lazy() and miscellaneous fixes | expand |
On 8/8/2022 11:45 PM, Joel Fernandes (Google) wrote: > Just a refresh of v3 with one additional debug patch. v3's cover letter is here: > https://lore.kernel.org/all/20220713213237.1596225-1-joel@joelfernandes.org/ > > I just started working on this again while I have some time during paternity > leave ;-) So I thought I'll just send it out again. No other changes other > than that 1 debug patch I added on the top. > > Next I am going to go refine the power results as mentioned in Paul's comments > on the last cover letter. Side note: Here is another big selling point for call_rcu_lazy(). Instead of _lazy(), if you just increased jiffies_till_first_fqs, and slowed *all* call_rcu() down to achieve the same effect, that would affect percpu refcounters switching to atomic-mode, for example. They switch to atomic mode by calling __percpu_ref_switch_mode() which is called by percpu_ref_switch_to_atomic_sync(). This will slow this call down for the full lazy duration which will slow down suspend in blk_pre_runtime_suspend(). This is why, we cannot assume call_rcu() users will mostly just want to free memory. There could be cases just like this, and just blanket slow down of call_rcu() might bite at unexpected times. I am going to add this as a selling point for selective lazyfication (hey I get to invent words while I'm inventing new features), to my cover letter and slides. - Joel > > Joel Fernandes (Google) (5): > rcu: Introduce call_rcu_lazy() API implementation > rcuscale: Add laziness and kfree tests > fs: Move call_rcu() to call_rcu_lazy() in some paths > rcutorture: Add test code for call_rcu_lazy() > debug: Toggle lazy at runtime and change flush jiffies > > Vineeth Pillai (1): > rcu: shrinker for lazy rcu > > fs/dcache.c | 4 +- > fs/eventpoll.c | 2 +- > fs/file_table.c | 2 +- > fs/inode.c | 2 +- > include/linux/rcu_segcblist.h | 1 + > include/linux/rcupdate.h | 6 + > include/linux/sched/sysctl.h | 3 + > kernel/rcu/Kconfig | 8 + > kernel/rcu/rcu.h | 12 + > kernel/rcu/rcu_segcblist.c | 15 +- > kernel/rcu/rcu_segcblist.h | 20 +- > kernel/rcu/rcuscale.c | 74 +++++- > kernel/rcu/rcutorture.c | 60 ++++- > kernel/rcu/tree.c | 131 ++++++---- > kernel/rcu/tree.h | 10 +- > kernel/rcu/tree_nocb.h | 246 +++++++++++++++--- > kernel/sysctl.c | 17 ++ > .../selftests/rcutorture/configs/rcu/CFLIST | 1 + > .../selftests/rcutorture/configs/rcu/TREE11 | 18 ++ > .../rcutorture/configs/rcu/TREE11.boot | 8 + > 20 files changed, 536 insertions(+), 104 deletions(-) > create mode 100644 tools/testing/selftests/rcutorture/configs/rcu/TREE11 > create mode 100644 tools/testing/selftests/rcutorture/configs/rcu/TREE11.boot > > -- > 2.37.1.559.g78731f0fdb-goog >
On 8/10/2022 10:23 PM, Joel Fernandes wrote: > > > On 8/8/2022 11:45 PM, Joel Fernandes (Google) wrote: >> Just a refresh of v3 with one additional debug patch. v3's cover letter is here: >> https://lore.kernel.org/all/20220713213237.1596225-1-joel@joelfernandes.org/ >> >> I just started working on this again while I have some time during paternity >> leave ;-) So I thought I'll just send it out again. No other changes other >> than that 1 debug patch I added on the top. >> >> Next I am going to go refine the power results as mentioned in Paul's comments >> on the last cover letter. > > Side note: Here is another big selling point for call_rcu_lazy(). > Instead of _lazy(), if you just increased jiffies_till_first_fqs, and > slowed *all* call_rcu() down to achieve the same effect, that would > affect percpu refcounters switching to atomic-mode, for example. > > They switch to atomic mode by calling __percpu_ref_switch_mode() which > is called by percpu_ref_switch_to_atomic_sync().> > This will slow this call down for the full lazy duration which will slow > down suspend in blk_pre_runtime_suspend(). Correction while I am going on the record (got to be careful these days). It *might* slow down RCU for the full lazy duration, unless of course a fly-by rescue call_rcu() comes in. - Joel
On Wed, Aug 10, 2022 at 10:31:56PM -0400, Joel Fernandes wrote: > > > On 8/10/2022 10:23 PM, Joel Fernandes wrote: > > > > > > On 8/8/2022 11:45 PM, Joel Fernandes (Google) wrote: > >> Just a refresh of v3 with one additional debug patch. v3's cover letter is here: > >> https://lore.kernel.org/all/20220713213237.1596225-1-joel@joelfernandes.org/ > >> > >> I just started working on this again while I have some time during paternity > >> leave ;-) So I thought I'll just send it out again. No other changes other > >> than that 1 debug patch I added on the top. > >> > >> Next I am going to go refine the power results as mentioned in Paul's comments > >> on the last cover letter. > > > > Side note: Here is another big selling point for call_rcu_lazy(). > > Instead of _lazy(), if you just increased jiffies_till_first_fqs, and > > slowed *all* call_rcu() down to achieve the same effect, that would > > affect percpu refcounters switching to atomic-mode, for example. > > > > They switch to atomic mode by calling __percpu_ref_switch_mode() which > > is called by percpu_ref_switch_to_atomic_sync().> > > This will slow this call down for the full lazy duration which will slow > > down suspend in blk_pre_runtime_suspend(). > > Correction while I am going on the record (got to be careful these > days). It *might* slow down RCU for the full lazy duration, unless of > course a fly-by rescue call_rcu() comes in. Just unload a module, which if I remember correctly invokes rcu_barrier(). Lots of rescue callbacks. ;-) Thanx, Paul
On 8/10/2022 10:51 PM, Paul E. McKenney wrote: > On Wed, Aug 10, 2022 at 10:31:56PM -0400, Joel Fernandes wrote: >> >> >> On 8/10/2022 10:23 PM, Joel Fernandes wrote: >>> >>> >>> On 8/8/2022 11:45 PM, Joel Fernandes (Google) wrote: >>>> Just a refresh of v3 with one additional debug patch. v3's cover letter is here: >>>> https://lore.kernel.org/all/20220713213237.1596225-1-joel@joelfernandes.org/ >>>> >>>> I just started working on this again while I have some time during paternity >>>> leave ;-) So I thought I'll just send it out again. No other changes other >>>> than that 1 debug patch I added on the top. >>>> >>>> Next I am going to go refine the power results as mentioned in Paul's comments >>>> on the last cover letter. >>> >>> Side note: Here is another big selling point for call_rcu_lazy(). >>> Instead of _lazy(), if you just increased jiffies_till_first_fqs, and >>> slowed *all* call_rcu() down to achieve the same effect, that would >>> affect percpu refcounters switching to atomic-mode, for example. >>> >>> They switch to atomic mode by calling __percpu_ref_switch_mode() which >>> is called by percpu_ref_switch_to_atomic_sync().> >>> This will slow this call down for the full lazy duration which will slow >>> down suspend in blk_pre_runtime_suspend(). >> >> Correction while I am going on the record (got to be careful these >> days). It *might* slow down RCU for the full lazy duration, unless of >> course a fly-by rescue call_rcu() comes in. > > Just unload a module, which if I remember correctly invokes rcu_barrier(). > Lots of rescue callbacks. ;-) Haha. Yes I suppose the per-cpu atomic switch paths can also invoke rcu_barrier() but I suspect somebody might complain about IPIs :-P Thanks, - Joel
On Wed, Aug 10, 2022 at 11:22:13PM -0400, Joel Fernandes wrote: > > > On 8/10/2022 10:51 PM, Paul E. McKenney wrote: > > On Wed, Aug 10, 2022 at 10:31:56PM -0400, Joel Fernandes wrote: > >> > >> > >> On 8/10/2022 10:23 PM, Joel Fernandes wrote: > >>> > >>> > >>> On 8/8/2022 11:45 PM, Joel Fernandes (Google) wrote: > >>>> Just a refresh of v3 with one additional debug patch. v3's cover letter is here: > >>>> https://lore.kernel.org/all/20220713213237.1596225-1-joel@joelfernandes.org/ > >>>> > >>>> I just started working on this again while I have some time during paternity > >>>> leave ;-) So I thought I'll just send it out again. No other changes other > >>>> than that 1 debug patch I added on the top. > >>>> > >>>> Next I am going to go refine the power results as mentioned in Paul's comments > >>>> on the last cover letter. > >>> > >>> Side note: Here is another big selling point for call_rcu_lazy(). > >>> Instead of _lazy(), if you just increased jiffies_till_first_fqs, and > >>> slowed *all* call_rcu() down to achieve the same effect, that would > >>> affect percpu refcounters switching to atomic-mode, for example. > >>> > >>> They switch to atomic mode by calling __percpu_ref_switch_mode() which > >>> is called by percpu_ref_switch_to_atomic_sync().> > >>> This will slow this call down for the full lazy duration which will slow > >>> down suspend in blk_pre_runtime_suspend(). > >> > >> Correction while I am going on the record (got to be careful these > >> days). It *might* slow down RCU for the full lazy duration, unless of > >> course a fly-by rescue call_rcu() comes in. > > > > Just unload a module, which if I remember correctly invokes rcu_barrier(). > > Lots of rescue callbacks. ;-) > > Haha. Yes I suppose the per-cpu atomic switch paths can also invoke > rcu_barrier() but I suspect somebody might complain about IPIs :-P There is always a critic! ;-) Thanx, Paul