Messages in this thread | | | Date | Mon, 6 Apr 2009 18:27:46 -0700 | From | "Paul E. McKenney" <> | Subject | Re: [PATCH] tracing/filters: allow event filters to be set only when not tracing |
| |
On Mon, Apr 06, 2009 at 08:34:20PM -0400, Steven Rostedt wrote: > > > On Mon, 6 Apr 2009, Paul E. McKenney wrote: > > > > Hmmm... Do the start_critical_timings() and stop_critical_timings() > > functions have anything to do with ftrace()? > > Yep. > > > It does not look like I can just bury RCU-idle controls in these > > functions, given that they are also invoked around call_console_drivers(), > > but if all the idle loops are surrounded by stop_critical_timings() and > > start_critical_timings(), that would ease location of all the idle > > loops. ;-) > > They are for the irqsoff tracer. Several archs will disable interrupts and > call an assembly "halt" type instruction that can be woken up by > interrupts. But this skips the irqs off check, and we end up with irqsoff > showing the idle loop with huge interrupts off latencies. > > I'm not sure they are included in all archs. They may be.
Not a chance! :-(
Looks like one can poke into cpu_idle() on all arches, but looks like I cannot even assume that I can leverage CONFIG_NO_HZ due to the rcu_needs_cpu() case. See below for my rough (and almost certainly erroneous) notes thus far.
Thoughts?
Thanx, Paul
------------------------------------------------------------------------
o Have rcu_idle() call from spinning idle loops. This would invoke rcu_qsctr_inc() and rcu_bh_qsctr_inc().
In this case, rcu_check_callbacks() need only check for interrupt from user-level code.
o But this won't work for idle loops that yield to a hypervisor without being in nohz node. For these, do rcu_idle_start() before the yield and rcu_idle_end() after the yield. These would manipulate a per-CPU variable.
In this case, rcu_check_callbacks() needs to check for:
a. interrupt from user-level code
b. unnested interrupt while under the influence of rcu_idle_start().
o Put rcu_idle_start() after stop_critical_timings() and put rcu_idle_end() before start_critical_timings() in each arch's idle loop. But maybe pick less confusing names...
o powerpc: looks OK.
o s390: looks OK -- always dyntick.
o Super-H: Looks OK -- always dyntick.
o x86 32-bit: Looks OK. This is cpu_idle() in arch/x86/kernel/process_32.c.
o x86 64-bit: Looks OK.
o ACPI: Looks OK.
o x86 APM: invokes system default idle, so OK if that is OK.
o x86: default_idle() in arch/x86/kernel/process.c need rcu_idle_start() and rcu_idle_end(), or ensure that caller always does the right thing???
o cpuidle_idle_call() in drivers/cpuidle/cpuidle.c needs rcu_idle_start() and rcu_idle_end()???
o ARM: Looks OK -- always dyntick. See cpu_idle() in arch/arm/kernel/process.c.
o Blackfin: Looks OK -- always dyntick. see cpu_idle() in arch/blackfin/kernel/process.c.
o Cris: ??? cpu_idle() in arch/cris/kernel/process.c has strange comment:
* Mark this as an RCU critical section so that * synchronize_kernel() in the unload path waits * for our completion.
Ain't gonna happen...
But looks like we need an rcu_start_idle() and an rcu_stop_idle() around the call to idle().
o H8300: cpu_idle() in arch/h8300/kernel/process.c.
Need an rcu_start_idle() and an rcu_stop_idle() around the call to idle(). Hopefully not messing too much with the exit latency.
o IA64: cpu_idle() in arch/ia64/kernel/process.c.
Need an rcu_start_idle() and an rcu_stop_idle() around the call to idle(). Not sure how far around...
o M68k: cpu_idle() in arch/m68k/kernel/process.c.
Need an rcu_start_idle() and an rcu_stop_idle() around the call to idle(). Hopefully not messing too much with the exit latency.
o M68Knommu: Ditto, but arch/m68knommu/kernel/process.c.
o UM: Looks OK -- always nohz.
o M32r: cpu_idle() in arch/m32r/kernel/process.c.
Need an rcu_start_idle() and an rcu_stop_idle() around the call to idle(). Hopefully not messing too much with the exit latency.
o MN10300: cpu_idle() in arch/mn10300/kernel/process.c.
Need an rcu_start_idle() and an rcu_stop_idle() around the call to idle(). Hopefully not messing too much with the exit latency.
@@@ Need to recheck the "nohz OK" assumption -- might have disabled it. Maybe make rcu_idle_start() and rcu_idle_stop() no-ops in case of CONFIG_NO_HZ -- but still need rcu_needs_cpu() case to work. So follow up on cpu_idle() cases.
o Idle loops (for posterity's sake):
o Generic: default_idle() gets rcu_idle().
o Power: All invoked from cpu_idle(), which already contains stop_critical_timings() and start_critical_timings().
o 44x: ppc44x_idle() absolutely unclear.
o powermac: power4_idle() is assembly code???
Appears to need a call to rcu_idle() in the loop headed by label 1b:.
o maple: same as powermac.
o pseries: pSeries_setup_arch()
o pseries_shared_idle_sleep() needs rcu_idle_start() and rcu_idle_end() surrounding the cede_processor() call.
o pseries_dedicated_idle_sleep() needs rcu_idle_start() and rcu_idle_end() surrounding the cede_processor() call. Might also need rcu_idle() in the "while" loop...
o iseries: iSeries_setup_arch()
o iseries_shared_idle() seems to need rcu_idle_start() and rcu_idle_end() surrounding yield_shared_processor() call.
o iseries_dedicated_idle() needs no action, as it forces nohz. ;-)
o pasemi: modes[] array in platforms/pasemi/idle.c
o idle_spin() assembly function, need call to rcu_idle() in loop headed by label 1:.
| |