"Tree RCU": scalable classic RCU implementation
This patch fixes a long-standing performance bug in classic RCU that
results in massive internal-to-RCU lock contention on systems with
more than a few hundred CPUs. Although this patch creates a separate
flavor of RCU for ease of review and patch maintenance, it is intended
to replace classic RCU.
This patch still handles stress better than does mainline, so I am still
calling it ready for inclusion. This patch is against the -tip tree.
Nevertheless, experience on an actual 1000+ CPU machine would still be
most welcome.
Most of the changes noted below were found while creating an rcutiny
(which should permit ejecting the current rcuclassic) and while doing
detailed line-by-line documentation.
Updates from v9 (http://lkml.org/lkml/2008/12/2/334):
o Fixes from remainder of line-by-line code walkthrough,
including comment spelling, initialization, undesirable
narrowing due to type conversion, removing redundant memory
barriers, removing redundant local-variable initialization,
and removing redundant local variables.
I do not believe that any of these fixes address the CPU-hotplug
issues that Andi Kleen was seeing, but please do give it a whirl
in case the machine is smarter than I am.
A writeup from the walkthrough may be found at the following
URL, in case you are suffering from terminal insomnia or
masochism:
http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf
o Made rcutree tracing use seq_file, as suggested some time
ago by Lai Jiangshan.
o Added a .csv variant of the rcudata debugfs trace file, to allow
people having thousands of CPUs to drop the data into
a spreadsheet. Tested with oocalc and gnumeric. Updated
documentation to suit.
Updates from v8 (http://lkml.org/lkml/2008/11/15/139):
o Fix a theoretical race between grace-period initialization and
force_quiescent_state() that could occur if more than three
jiffies were required to carry out the grace-period
initialization. Which it might, if you had enough CPUs.
o Apply Ingo's printk-standardization patch.
o Substitute local variables for repeated accesses to global
variables.
o Fix comment misspellings and redundant (but harmless) increments
of ->n_rcu_pending (this latter after having explicitly added it).
o Apply checkpatch fixes.
Updates from v7 (http://lkml.org/lkml/2008/10/10/291):
o Fixed a number of problems noted by Gautham Shenoy, including
the cpu-stall-detection bug that he was having difficulty
convincing me was real. ;-)
o Changed cpu-stall detection to wait for ten seconds rather than
three in order to reduce false positive, as suggested by Ingo
Molnar.
o Produced a design document (http://lwn.net/Articles/305782/).
The act of writing this document uncovered a number of both
theoretical and "here and now" bugs as noted below.
o Fix dynticks_nesting accounting confusion, simplify WARN_ON()
condition, fix kerneldoc comments, and add memory barriers
in dynticks interface functions.
o Add more data to tracing.
o Remove unused "rcu_barrier" field from rcu_data structure.
o Count calls to rcu_pending() from scheduling-clock interrupt
to use as a surrogate timebase should jiffies stop counting.
o Fix a theoretical race between force_quiescent_state() and
grace-period initialization. Yes, initialization does have to
go on for some jiffies for this race to occur, but given enough
CPUs...
Updates from v6 (http://lkml.org/lkml/2008/9/23/448):
o Fix a number of checkpatch.pl complaints.
o Apply review comments from Ingo Molnar and Lai Jiangshan
on the stall-detection code.
o Fix several bugs in !CONFIG_SMP builds.
o Fix a misspelled config-parameter name so that RCU now announces
at boot time if stall detection is configured.
o Run tests on numerous combinations of configurations parameters,
which after the fixes above, now build and run correctly.
Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line):
o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a
changeset some time ago, and finally got around to retesting
this option).
o Fix some tracing bugs in rcupreempt that caused incorrect
totals to be printed.
o I now test with a more brutal random-selection online/offline
script (attached). Probably more brutal than it needs to be
on the people reading it as well, but so it goes.
o A number of optimizations and usability improvements:
o Make rcu_pending() ignore the grace-period timeout when
there is no grace period in progress.
o Make force_quiescent_state() avoid going for a global
lock in the case where there is no grace period in
progress.
o Rearrange struct fields to improve struct layout.
o Make call_rcu() initiate a grace period if RCU was
idle, rather than waiting for the next scheduling
clock interrupt.
o Invoke rcu_irq_enter() and rcu_irq_exit() only when
idle, as suggested by Andi Kleen. I still don't
completely trust this change, and might back it out.
o Make CONFIG_RCU_TRACE be the single config variable
manipulated for all forms of RCU, instead of the prior
confusion.
o Document tracing files and formats for both rcupreempt
and rcutree.
Updates from v4 for those missing v5 given its bad subject line:
o Separated dynticks interface so that NMIs and irqs call separate
functions, greatly simplifying it. In particular, this code
no longer requires a proof of correctness. ;-)
o Separated dynticks state out into its own per-CPU structure,
avoiding the duplicated accounting.
o The case where a dynticks-idle CPU runs an irq handler that
invokes call_rcu() is now correctly handled, forcing that CPU
out of dynticks-idle mode.
o Review comments have been applied (thank you all!!!).
For but one example, fixed the dynticks-ordering issue that
Manfred pointed out, saving me much debugging. ;-)
o Adjusted rcuclassic and rcupreempt to handle dynticks changes.
Attached is an updated patch to Classic RCU that applies a hierarchy,
greatly reducing the contention on the top-level lock for large machines.
This passes 10-hour concurrent rcutorture and online-offline testing on
128-CPU ppc64 without dynticks enabled, and exposes some timekeeping
bugs in presence of dynticks (exciting working on a system where
"sleep 1" hangs until interrupted...), which were fixed in the
2.6.27 kernel. It is getting more reliable than mainline by some
measures, so the next version will be against -tip for inclusion.
See also Manfred Spraul's recent patches (or his earlier work from
2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2).
We will converge onto a common patch in the fullness of time, but are
currently exploring different regions of the design space. That said,
I have already gratefully stolen quite a few of Manfred's ideas.
This patch provides CONFIG_RCU_FANOUT, which controls the bushiness
of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on
64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT,
there is no hierarchy. By default, the RCU initialization code will
adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA
architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable
this balancing, allowing the hierarchy to be exactly aligned to the
underlying hardware. Up to two levels of hierarchy are permitted
(in addition to the root node), allowing up to 16,384 CPUs on 32-bit
systems and up to 262,144 CPUs on 64-bit systems. I just know that I
am going to regret saying this, but this seems more than sufficient
for the foreseeable future. (Some architectures might wish to set
CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs.
If this becomes a real problem, additional levels can be added, but I
doubt that it will make a significant difference on real hardware.)
In the common case, a given CPU will manipulate its private rcu_data
structure and the rcu_node structure that it shares with its immediate
neighbors. This can reduce both lock and memory contention by multiple
orders of magnitude, which should eliminate the need for the strange
manipulations that are reported to be required when running Linux on
very large systems.
Some shortcomings:
o More bugs will probably surface as a result of an ongoing
line-by-line code inspection.
Patches will be provided as required.
o There are probably hangs, rcutorture failures, &c. Seems
quite stable on a 128-CPU machine, but that is kind of small
compared to 4096 CPUs. However, seems to do better than
mainline.
Patches will be provided as required.
o The memory footprint of this version is several KB larger
than rcuclassic.
A separate UP-only rcutiny patch will be provided, which will
reduce the memory footprint significantly, even compared
to the old rcuclassic. One such patch passes light testing,
and has a memory footprint smaller even than rcuclassic.
Initial reaction from various embedded guys was "it is not
worth it", so am putting it aside.
Credits:
o Manfred Spraul for ideas, review comments, and bugs spotted,
as well as some good friendly competition. ;-)
o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers,
Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton
for reviews and comments.
o Thomas Gleixner for much-needed help with some timer issues
(see patches below).
o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos,
Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton
Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines
alive despite my heavy abuse^Wtesting.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-18 21:55:32 +01:00
|
|
|
CONFIG_RCU_TRACE debugfs Files and Formats
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The rcupreempt and rcutree implementations of RCU provide debugfs trace
|
|
|
|
|
output that summarizes counters and state. This information is useful for
|
|
|
|
|
debugging RCU itself, and can sometimes also help to debug abuses of RCU.
|
|
|
|
|
Note that the rcuclassic implementation of RCU does not provide debugfs
|
|
|
|
|
trace output.
|
|
|
|
|
|
|
|
|
|
The following sections describe the debugfs files and formats for
|
|
|
|
|
preemptable RCU (rcupreempt) and hierarchical RCU (rcutree).
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Preemptable RCU debugfs Files and Formats
|
|
|
|
|
|
|
|
|
|
This implementation of RCU provides three debugfs files under the
|
|
|
|
|
top-level directory RCU: rcu/rcuctrs (which displays the per-CPU
|
|
|
|
|
counters used by preemptable RCU) rcu/rcugp (which displays grace-period
|
|
|
|
|
counters), and rcu/rcustats (which internal counters for debugging RCU).
|
|
|
|
|
|
|
|
|
|
The output of "cat rcu/rcuctrs" looks as follows:
|
|
|
|
|
|
|
|
|
|
CPU last cur F M
|
|
|
|
|
0 5 -5 0 0
|
|
|
|
|
1 -1 0 0 0
|
|
|
|
|
2 0 1 0 0
|
|
|
|
|
3 0 1 0 0
|
|
|
|
|
4 0 1 0 0
|
|
|
|
|
5 0 1 0 0
|
|
|
|
|
6 0 2 0 0
|
|
|
|
|
7 0 -1 0 0
|
|
|
|
|
8 0 1 0 0
|
|
|
|
|
ggp = 26226, state = waitzero
|
|
|
|
|
|
|
|
|
|
The per-CPU fields are as follows:
|
|
|
|
|
|
|
|
|
|
o "CPU" gives the CPU number. Offline CPUs are not displayed.
|
|
|
|
|
|
|
|
|
|
o "last" gives the value of the counter that is being decremented
|
|
|
|
|
for the current grace period phase. In the example above,
|
|
|
|
|
the counters sum to 4, indicating that there are still four
|
|
|
|
|
RCU read-side critical sections still running that started
|
|
|
|
|
before the last counter flip.
|
|
|
|
|
|
|
|
|
|
o "cur" gives the value of the counter that is currently being
|
|
|
|
|
both incremented (by rcu_read_lock()) and decremented (by
|
|
|
|
|
rcu_read_unlock()). In the example above, the counters sum to
|
|
|
|
|
1, indicating that there is only one RCU read-side critical section
|
|
|
|
|
still running that started after the last counter flip.
|
|
|
|
|
|
|
|
|
|
o "F" indicates whether RCU is waiting for this CPU to acknowledge
|
|
|
|
|
a counter flip. In the above example, RCU is not waiting on any,
|
|
|
|
|
which is consistent with the state being "waitzero" rather than
|
|
|
|
|
"waitack".
|
|
|
|
|
|
|
|
|
|
o "M" indicates whether RCU is waiting for this CPU to execute a
|
|
|
|
|
memory barrier. In the above example, RCU is not waiting on any,
|
|
|
|
|
which is consistent with the state being "waitzero" rather than
|
|
|
|
|
"waitmb".
|
|
|
|
|
|
|
|
|
|
o "ggp" is the global grace-period counter.
|
|
|
|
|
|
|
|
|
|
o "state" is the RCU state, which can be one of the following:
|
|
|
|
|
|
|
|
|
|
o "idle": there is no grace period in progress.
|
|
|
|
|
|
|
|
|
|
o "waitack": RCU just incremented the global grace-period
|
|
|
|
|
counter, which has the effect of reversing the roles of
|
|
|
|
|
the "last" and "cur" counters above, and is waiting for
|
|
|
|
|
all the CPUs to acknowledge the flip. Once the flip has
|
|
|
|
|
been acknowledged, CPUs will no longer be incrementing
|
|
|
|
|
what are now the "last" counters, so that their sum will
|
|
|
|
|
decrease monotonically down to zero.
|
|
|
|
|
|
|
|
|
|
o "waitzero": RCU is waiting for the sum of the "last" counters
|
|
|
|
|
to decrease to zero.
|
|
|
|
|
|
|
|
|
|
o "waitmb": RCU is waiting for each CPU to execute a memory
|
|
|
|
|
barrier, which ensures that instructions from a given CPU's
|
|
|
|
|
last RCU read-side critical section cannot be reordered
|
|
|
|
|
with instructions following the memory-barrier instruction.
|
|
|
|
|
|
|
|
|
|
The output of "cat rcu/rcugp" looks as follows:
|
|
|
|
|
|
|
|
|
|
oldggp=48870 newggp=48873
|
|
|
|
|
|
|
|
|
|
Note that reading from this file provokes a synchronize_rcu(). The
|
|
|
|
|
"oldggp" value is that of "ggp" from rcu/rcuctrs above, taken before
|
|
|
|
|
executing the synchronize_rcu(), and the "newggp" value is also the
|
|
|
|
|
"ggp" value, but taken after the synchronize_rcu() command returns.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The output of "cat rcu/rcugp" looks as follows:
|
|
|
|
|
|
|
|
|
|
na=1337955 nl=40 wa=1337915 wl=44 da=1337871 dl=0 dr=1337871 di=1337871
|
|
|
|
|
1=50989 e1=6138 i1=49722 ie1=82 g1=49640 a1=315203 ae1=265563 a2=49640
|
|
|
|
|
z1=1401244 ze1=1351605 z2=49639 m1=5661253 me1=5611614 m2=49639
|
|
|
|
|
|
|
|
|
|
These are counters tracking internal preemptable-RCU events, however,
|
|
|
|
|
some of them may be useful for debugging algorithms using RCU. In
|
|
|
|
|
particular, the "nl", "wl", and "dl" values track the number of RCU
|
|
|
|
|
callbacks in various states. The fields are as follows:
|
|
|
|
|
|
|
|
|
|
o "na" is the total number of RCU callbacks that have been enqueued
|
|
|
|
|
since boot.
|
|
|
|
|
|
|
|
|
|
o "nl" is the number of RCU callbacks waiting for the previous
|
|
|
|
|
grace period to end so that they can start waiting on the next
|
|
|
|
|
grace period.
|
|
|
|
|
|
|
|
|
|
o "wa" is the total number of RCU callbacks that have started waiting
|
|
|
|
|
for a grace period since boot. "na" should be roughly equal to
|
|
|
|
|
"nl" plus "wa".
|
|
|
|
|
|
|
|
|
|
o "wl" is the number of RCU callbacks currently waiting for their
|
|
|
|
|
grace period to end.
|
|
|
|
|
|
|
|
|
|
o "da" is the total number of RCU callbacks whose grace periods
|
|
|
|
|
have completed since boot. "wa" should be roughly equal to
|
|
|
|
|
"wl" plus "da".
|
|
|
|
|
|
|
|
|
|
o "dr" is the total number of RCU callbacks that have been removed
|
|
|
|
|
from the list of callbacks ready to invoke. "dr" should be roughly
|
|
|
|
|
equal to "da".
|
|
|
|
|
|
|
|
|
|
o "di" is the total number of RCU callbacks that have been invoked
|
|
|
|
|
since boot. "di" should be roughly equal to "da", though some
|
|
|
|
|
early versions of preemptable RCU had a bug so that only the
|
|
|
|
|
last CPU's count of invocations was displayed, rather than the
|
|
|
|
|
sum of all CPU's counts.
|
|
|
|
|
|
|
|
|
|
o "1" is the number of calls to rcu_try_flip(). This should be
|
|
|
|
|
roughly equal to the sum of "e1", "i1", "a1", "z1", and "m1"
|
|
|
|
|
described below. In other words, the number of times that
|
|
|
|
|
the state machine is visited should be equal to the sum of the
|
|
|
|
|
number of times that each state is visited plus the number of
|
|
|
|
|
times that the state-machine lock acquisition failed.
|
|
|
|
|
|
|
|
|
|
o "e1" is the number of times that rcu_try_flip() was unable to
|
|
|
|
|
acquire the fliplock.
|
|
|
|
|
|
|
|
|
|
o "i1" is the number of calls to rcu_try_flip_idle().
|
|
|
|
|
|
|
|
|
|
o "ie1" is the number of times rcu_try_flip_idle() exited early
|
|
|
|
|
due to the calling CPU having no work for RCU.
|
|
|
|
|
|
|
|
|
|
o "g1" is the number of times that rcu_try_flip_idle() decided
|
|
|
|
|
to start a new grace period. "i1" should be roughly equal to
|
|
|
|
|
"ie1" plus "g1".
|
|
|
|
|
|
|
|
|
|
o "a1" is the number of calls to rcu_try_flip_waitack().
|
|
|
|
|
|
|
|
|
|
o "ae1" is the number of times that rcu_try_flip_waitack() found
|
|
|
|
|
that at least one CPU had not yet acknowledge the new grace period
|
|
|
|
|
(AKA "counter flip").
|
|
|
|
|
|
|
|
|
|
o "a2" is the number of time rcu_try_flip_waitack() found that
|
|
|
|
|
all CPUs had acknowledged. "a1" should be roughly equal to
|
|
|
|
|
"ae1" plus "a2". (This particular output was collected on
|
|
|
|
|
a 128-CPU machine, hence the smaller-than-usual fraction of
|
|
|
|
|
calls to rcu_try_flip_waitack() finding all CPUs having already
|
|
|
|
|
acknowledged.)
|
|
|
|
|
|
|
|
|
|
o "z1" is the number of calls to rcu_try_flip_waitzero().
|
|
|
|
|
|
|
|
|
|
o "ze1" is the number of times that rcu_try_flip_waitzero() found
|
|
|
|
|
that not all of the old RCU read-side critical sections had
|
|
|
|
|
completed.
|
|
|
|
|
|
|
|
|
|
o "z2" is the number of times that rcu_try_flip_waitzero() finds
|
|
|
|
|
the sum of the counters equal to zero, in other words, that
|
|
|
|
|
all of the old RCU read-side critical sections had completed.
|
|
|
|
|
The value of "z1" should be roughly equal to "ze1" plus
|
|
|
|
|
"z2".
|
|
|
|
|
|
|
|
|
|
o "m1" is the number of calls to rcu_try_flip_waitmb().
|
|
|
|
|
|
|
|
|
|
o "me1" is the number of times that rcu_try_flip_waitmb() finds
|
|
|
|
|
that at least one CPU has not yet executed a memory barrier.
|
|
|
|
|
|
|
|
|
|
o "m2" is the number of times that rcu_try_flip_waitmb() finds that
|
|
|
|
|
all CPUs have executed a memory barrier.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Hierarchical RCU debugfs Files and Formats
|
|
|
|
|
|
|
|
|
|
This implementation of RCU provides three debugfs files under the
|
|
|
|
|
top-level directory RCU: rcu/rcudata (which displays fields in struct
|
|
|
|
|
rcu_data), rcu/rcugp (which displays grace-period counters), and
|
|
|
|
|
rcu/rcuhier (which displays the struct rcu_node hierarchy).
|
|
|
|
|
|
|
|
|
|
The output of "cat rcu/rcudata" looks as follows:
|
|
|
|
|
|
|
|
|
|
rcu:
|
2009-04-14 06:31:18 +02:00
|
|
|
rcu:
|
|
|
|
|
0 c=17829 g=17829 pq=1 pqc=17829 qp=0 dt=10951/1 dn=0 df=1101 of=0 ri=36 ql=0 b=10
|
|
|
|
|
1 c=17829 g=17829 pq=1 pqc=17829 qp=0 dt=16117/1 dn=0 df=1015 of=0 ri=0 ql=0 b=10
|
|
|
|
|
2 c=17829 g=17829 pq=1 pqc=17829 qp=0 dt=1445/1 dn=0 df=1839 of=0 ri=0 ql=0 b=10
|
|
|
|
|
3 c=17829 g=17829 pq=1 pqc=17829 qp=0 dt=6681/1 dn=0 df=1545 of=0 ri=0 ql=0 b=10
|
|
|
|
|
4 c=17829 g=17829 pq=1 pqc=17829 qp=0 dt=1003/1 dn=0 df=1992 of=0 ri=0 ql=0 b=10
|
|
|
|
|
5 c=17829 g=17830 pq=1 pqc=17829 qp=1 dt=3887/1 dn=0 df=3331 of=0 ri=4 ql=2 b=10
|
|
|
|
|
6 c=17829 g=17829 pq=1 pqc=17829 qp=0 dt=859/1 dn=0 df=3224 of=0 ri=0 ql=0 b=10
|
|
|
|
|
7 c=17829 g=17830 pq=0 pqc=17829 qp=1 dt=3761/1 dn=0 df=1818 of=0 ri=0 ql=2 b=10
|
"Tree RCU": scalable classic RCU implementation
This patch fixes a long-standing performance bug in classic RCU that
results in massive internal-to-RCU lock contention on systems with
more than a few hundred CPUs. Although this patch creates a separate
flavor of RCU for ease of review and patch maintenance, it is intended
to replace classic RCU.
This patch still handles stress better than does mainline, so I am still
calling it ready for inclusion. This patch is against the -tip tree.
Nevertheless, experience on an actual 1000+ CPU machine would still be
most welcome.
Most of the changes noted below were found while creating an rcutiny
(which should permit ejecting the current rcuclassic) and while doing
detailed line-by-line documentation.
Updates from v9 (http://lkml.org/lkml/2008/12/2/334):
o Fixes from remainder of line-by-line code walkthrough,
including comment spelling, initialization, undesirable
narrowing due to type conversion, removing redundant memory
barriers, removing redundant local-variable initialization,
and removing redundant local variables.
I do not believe that any of these fixes address the CPU-hotplug
issues that Andi Kleen was seeing, but please do give it a whirl
in case the machine is smarter than I am.
A writeup from the walkthrough may be found at the following
URL, in case you are suffering from terminal insomnia or
masochism:
http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf
o Made rcutree tracing use seq_file, as suggested some time
ago by Lai Jiangshan.
o Added a .csv variant of the rcudata debugfs trace file, to allow
people having thousands of CPUs to drop the data into
a spreadsheet. Tested with oocalc and gnumeric. Updated
documentation to suit.
Updates from v8 (http://lkml.org/lkml/2008/11/15/139):
o Fix a theoretical race between grace-period initialization and
force_quiescent_state() that could occur if more than three
jiffies were required to carry out the grace-period
initialization. Which it might, if you had enough CPUs.
o Apply Ingo's printk-standardization patch.
o Substitute local variables for repeated accesses to global
variables.
o Fix comment misspellings and redundant (but harmless) increments
of ->n_rcu_pending (this latter after having explicitly added it).
o Apply checkpatch fixes.
Updates from v7 (http://lkml.org/lkml/2008/10/10/291):
o Fixed a number of problems noted by Gautham Shenoy, including
the cpu-stall-detection bug that he was having difficulty
convincing me was real. ;-)
o Changed cpu-stall detection to wait for ten seconds rather than
three in order to reduce false positive, as suggested by Ingo
Molnar.
o Produced a design document (http://lwn.net/Articles/305782/).
The act of writing this document uncovered a number of both
theoretical and "here and now" bugs as noted below.
o Fix dynticks_nesting accounting confusion, simplify WARN_ON()
condition, fix kerneldoc comments, and add memory barriers
in dynticks interface functions.
o Add more data to tracing.
o Remove unused "rcu_barrier" field from rcu_data structure.
o Count calls to rcu_pending() from scheduling-clock interrupt
to use as a surrogate timebase should jiffies stop counting.
o Fix a theoretical race between force_quiescent_state() and
grace-period initialization. Yes, initialization does have to
go on for some jiffies for this race to occur, but given enough
CPUs...
Updates from v6 (http://lkml.org/lkml/2008/9/23/448):
o Fix a number of checkpatch.pl complaints.
o Apply review comments from Ingo Molnar and Lai Jiangshan
on the stall-detection code.
o Fix several bugs in !CONFIG_SMP builds.
o Fix a misspelled config-parameter name so that RCU now announces
at boot time if stall detection is configured.
o Run tests on numerous combinations of configurations parameters,
which after the fixes above, now build and run correctly.
Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line):
o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a
changeset some time ago, and finally got around to retesting
this option).
o Fix some tracing bugs in rcupreempt that caused incorrect
totals to be printed.
o I now test with a more brutal random-selection online/offline
script (attached). Probably more brutal than it needs to be
on the people reading it as well, but so it goes.
o A number of optimizations and usability improvements:
o Make rcu_pending() ignore the grace-period timeout when
there is no grace period in progress.
o Make force_quiescent_state() avoid going for a global
lock in the case where there is no grace period in
progress.
o Rearrange struct fields to improve struct layout.
o Make call_rcu() initiate a grace period if RCU was
idle, rather than waiting for the next scheduling
clock interrupt.
o Invoke rcu_irq_enter() and rcu_irq_exit() only when
idle, as suggested by Andi Kleen. I still don't
completely trust this change, and might back it out.
o Make CONFIG_RCU_TRACE be the single config variable
manipulated for all forms of RCU, instead of the prior
confusion.
o Document tracing files and formats for both rcupreempt
and rcutree.
Updates from v4 for those missing v5 given its bad subject line:
o Separated dynticks interface so that NMIs and irqs call separate
functions, greatly simplifying it. In particular, this code
no longer requires a proof of correctness. ;-)
o Separated dynticks state out into its own per-CPU structure,
avoiding the duplicated accounting.
o The case where a dynticks-idle CPU runs an irq handler that
invokes call_rcu() is now correctly handled, forcing that CPU
out of dynticks-idle mode.
o Review comments have been applied (thank you all!!!).
For but one example, fixed the dynticks-ordering issue that
Manfred pointed out, saving me much debugging. ;-)
o Adjusted rcuclassic and rcupreempt to handle dynticks changes.
Attached is an updated patch to Classic RCU that applies a hierarchy,
greatly reducing the contention on the top-level lock for large machines.
This passes 10-hour concurrent rcutorture and online-offline testing on
128-CPU ppc64 without dynticks enabled, and exposes some timekeeping
bugs in presence of dynticks (exciting working on a system where
"sleep 1" hangs until interrupted...), which were fixed in the
2.6.27 kernel. It is getting more reliable than mainline by some
measures, so the next version will be against -tip for inclusion.
See also Manfred Spraul's recent patches (or his earlier work from
2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2).
We will converge onto a common patch in the fullness of time, but are
currently exploring different regions of the design space. That said,
I have already gratefully stolen quite a few of Manfred's ideas.
This patch provides CONFIG_RCU_FANOUT, which controls the bushiness
of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on
64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT,
there is no hierarchy. By default, the RCU initialization code will
adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA
architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable
this balancing, allowing the hierarchy to be exactly aligned to the
underlying hardware. Up to two levels of hierarchy are permitted
(in addition to the root node), allowing up to 16,384 CPUs on 32-bit
systems and up to 262,144 CPUs on 64-bit systems. I just know that I
am going to regret saying this, but this seems more than sufficient
for the foreseeable future. (Some architectures might wish to set
CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs.
If this becomes a real problem, additional levels can be added, but I
doubt that it will make a significant difference on real hardware.)
In the common case, a given CPU will manipulate its private rcu_data
structure and the rcu_node structure that it shares with its immediate
neighbors. This can reduce both lock and memory contention by multiple
orders of magnitude, which should eliminate the need for the strange
manipulations that are reported to be required when running Linux on
very large systems.
Some shortcomings:
o More bugs will probably surface as a result of an ongoing
line-by-line code inspection.
Patches will be provided as required.
o There are probably hangs, rcutorture failures, &c. Seems
quite stable on a 128-CPU machine, but that is kind of small
compared to 4096 CPUs. However, seems to do better than
mainline.
Patches will be provided as required.
o The memory footprint of this version is several KB larger
than rcuclassic.
A separate UP-only rcutiny patch will be provided, which will
reduce the memory footprint significantly, even compared
to the old rcuclassic. One such patch passes light testing,
and has a memory footprint smaller even than rcuclassic.
Initial reaction from various embedded guys was "it is not
worth it", so am putting it aside.
Credits:
o Manfred Spraul for ideas, review comments, and bugs spotted,
as well as some good friendly competition. ;-)
o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers,
Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton
for reviews and comments.
o Thomas Gleixner for much-needed help with some timer issues
(see patches below).
o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos,
Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton
Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines
alive despite my heavy abuse^Wtesting.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-18 21:55:32 +01:00
|
|
|
rcu_bh:
|
2009-04-14 06:31:18 +02:00
|
|
|
0 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=10951/1 dn=0 df=0 of=0 ri=0 ql=0 b=10
|
|
|
|
|
1 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=16117/1 dn=0 df=13 of=0 ri=0 ql=0 b=10
|
|
|
|
|
2 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=1445/1 dn=0 df=15 of=0 ri=0 ql=0 b=10
|
|
|
|
|
3 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=6681/1 dn=0 df=9 of=0 ri=0 ql=0 b=10
|
|
|
|
|
4 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=1003/1 dn=0 df=15 of=0 ri=0 ql=0 b=10
|
|
|
|
|
5 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=3887/1 dn=0 df=15 of=0 ri=0 ql=0 b=10
|
|
|
|
|
6 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=859/1 dn=0 df=15 of=0 ri=0 ql=0 b=10
|
|
|
|
|
7 c=-275 g=-275 pq=1 pqc=-275 qp=0 dt=3761/1 dn=0 df=15 of=0 ri=0 ql=0 b=10
|
"Tree RCU": scalable classic RCU implementation
This patch fixes a long-standing performance bug in classic RCU that
results in massive internal-to-RCU lock contention on systems with
more than a few hundred CPUs. Although this patch creates a separate
flavor of RCU for ease of review and patch maintenance, it is intended
to replace classic RCU.
This patch still handles stress better than does mainline, so I am still
calling it ready for inclusion. This patch is against the -tip tree.
Nevertheless, experience on an actual 1000+ CPU machine would still be
most welcome.
Most of the changes noted below were found while creating an rcutiny
(which should permit ejecting the current rcuclassic) and while doing
detailed line-by-line documentation.
Updates from v9 (http://lkml.org/lkml/2008/12/2/334):
o Fixes from remainder of line-by-line code walkthrough,
including comment spelling, initialization, undesirable
narrowing due to type conversion, removing redundant memory
barriers, removing redundant local-variable initialization,
and removing redundant local variables.
I do not believe that any of these fixes address the CPU-hotplug
issues that Andi Kleen was seeing, but please do give it a whirl
in case the machine is smarter than I am.
A writeup from the walkthrough may be found at the following
URL, in case you are suffering from terminal insomnia or
masochism:
http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf
o Made rcutree tracing use seq_file, as suggested some time
ago by Lai Jiangshan.
o Added a .csv variant of the rcudata debugfs trace file, to allow
people having thousands of CPUs to drop the data into
a spreadsheet. Tested with oocalc and gnumeric. Updated
documentation to suit.
Updates from v8 (http://lkml.org/lkml/2008/11/15/139):
o Fix a theoretical race between grace-period initialization and
force_quiescent_state() that could occur if more than three
jiffies were required to carry out the grace-period
initialization. Which it might, if you had enough CPUs.
o Apply Ingo's printk-standardization patch.
o Substitute local variables for repeated accesses to global
variables.
o Fix comment misspellings and redundant (but harmless) increments
of ->n_rcu_pending (this latter after having explicitly added it).
o Apply checkpatch fixes.
Updates from v7 (http://lkml.org/lkml/2008/10/10/291):
o Fixed a number of problems noted by Gautham Shenoy, including
the cpu-stall-detection bug that he was having difficulty
convincing me was real. ;-)
o Changed cpu-stall detection to wait for ten seconds rather than
three in order to reduce false positive, as suggested by Ingo
Molnar.
o Produced a design document (http://lwn.net/Articles/305782/).
The act of writing this document uncovered a number of both
theoretical and "here and now" bugs as noted below.
o Fix dynticks_nesting accounting confusion, simplify WARN_ON()
condition, fix kerneldoc comments, and add memory barriers
in dynticks interface functions.
o Add more data to tracing.
o Remove unused "rcu_barrier" field from rcu_data structure.
o Count calls to rcu_pending() from scheduling-clock interrupt
to use as a surrogate timebase should jiffies stop counting.
o Fix a theoretical race between force_quiescent_state() and
grace-period initialization. Yes, initialization does have to
go on for some jiffies for this race to occur, but given enough
CPUs...
Updates from v6 (http://lkml.org/lkml/2008/9/23/448):
o Fix a number of checkpatch.pl complaints.
o Apply review comments from Ingo Molnar and Lai Jiangshan
on the stall-detection code.
o Fix several bugs in !CONFIG_SMP builds.
o Fix a misspelled config-parameter name so that RCU now announces
at boot time if stall detection is configured.
o Run tests on numerous combinations of configurations parameters,
which after the fixes above, now build and run correctly.
Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line):
o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a
changeset some time ago, and finally got around to retesting
this option).
o Fix some tracing bugs in rcupreempt that caused incorrect
totals to be printed.
o I now test with a more brutal random-selection online/offline
script (attached). Probably more brutal than it needs to be
on the people reading it as well, but so it goes.
o A number of optimizations and usability improvements:
o Make rcu_pending() ignore the grace-period timeout when
there is no grace period in progress.
o Make force_quiescent_state() avoid going for a global
lock in the case where there is no grace period in
progress.
o Rearrange struct fields to improve struct layout.
o Make call_rcu() initiate a grace period if RCU was
idle, rather than waiting for the next scheduling
clock interrupt.
o Invoke rcu_irq_enter() and rcu_irq_exit() only when
idle, as suggested by Andi Kleen. I still don't
completely trust this change, and might back it out.
o Make CONFIG_RCU_TRACE be the single config variable
manipulated for all forms of RCU, instead of the prior
confusion.
o Document tracing files and formats for both rcupreempt
and rcutree.
Updates from v4 for those missing v5 given its bad subject line:
o Separated dynticks interface so that NMIs and irqs call separate
functions, greatly simplifying it. In particular, this code
no longer requires a proof of correctness. ;-)
o Separated dynticks state out into its own per-CPU structure,
avoiding the duplicated accounting.
o The case where a dynticks-idle CPU runs an irq handler that
invokes call_rcu() is now correctly handled, forcing that CPU
out of dynticks-idle mode.
o Review comments have been applied (thank you all!!!).
For but one example, fixed the dynticks-ordering issue that
Manfred pointed out, saving me much debugging. ;-)
o Adjusted rcuclassic and rcupreempt to handle dynticks changes.
Attached is an updated patch to Classic RCU that applies a hierarchy,
greatly reducing the contention on the top-level lock for large machines.
This passes 10-hour concurrent rcutorture and online-offline testing on
128-CPU ppc64 without dynticks enabled, and exposes some timekeeping
bugs in presence of dynticks (exciting working on a system where
"sleep 1" hangs until interrupted...), which were fixed in the
2.6.27 kernel. It is getting more reliable than mainline by some
measures, so the next version will be against -tip for inclusion.
See also Manfred Spraul's recent patches (or his earlier work from
2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2).
We will converge onto a common patch in the fullness of time, but are
currently exploring different regions of the design space. That said,
I have already gratefully stolen quite a few of Manfred's ideas.
This patch provides CONFIG_RCU_FANOUT, which controls the bushiness
of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on
64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT,
there is no hierarchy. By default, the RCU initialization code will
adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA
architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable
this balancing, allowing the hierarchy to be exactly aligned to the
underlying hardware. Up to two levels of hierarchy are permitted
(in addition to the root node), allowing up to 16,384 CPUs on 32-bit
systems and up to 262,144 CPUs on 64-bit systems. I just know that I
am going to regret saying this, but this seems more than sufficient
for the foreseeable future. (Some architectures might wish to set
CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs.
If this becomes a real problem, additional levels can be added, but I
doubt that it will make a significant difference on real hardware.)
In the common case, a given CPU will manipulate its private rcu_data
structure and the rcu_node structure that it shares with its immediate
neighbors. This can reduce both lock and memory contention by multiple
orders of magnitude, which should eliminate the need for the strange
manipulations that are reported to be required when running Linux on
very large systems.
Some shortcomings:
o More bugs will probably surface as a result of an ongoing
line-by-line code inspection.
Patches will be provided as required.
o There are probably hangs, rcutorture failures, &c. Seems
quite stable on a 128-CPU machine, but that is kind of small
compared to 4096 CPUs. However, seems to do better than
mainline.
Patches will be provided as required.
o The memory footprint of this version is several KB larger
than rcuclassic.
A separate UP-only rcutiny patch will be provided, which will
reduce the memory footprint significantly, even compared
to the old rcuclassic. One such patch passes light testing,
and has a memory footprint smaller even than rcuclassic.
Initial reaction from various embedded guys was "it is not
worth it", so am putting it aside.
Credits:
o Manfred Spraul for ideas, review comments, and bugs spotted,
as well as some good friendly competition. ;-)
o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers,
Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton
for reviews and comments.
o Thomas Gleixner for much-needed help with some timer issues
(see patches below).
o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos,
Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton
Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines
alive despite my heavy abuse^Wtesting.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-18 21:55:32 +01:00
|
|
|
|
|
|
|
|
The first section lists the rcu_data structures for rcu, the second for
|
|
|
|
|
rcu_bh. Each section has one line per CPU, or eight for this 8-CPU system.
|
|
|
|
|
The fields are as follows:
|
|
|
|
|
|
|
|
|
|
o The number at the beginning of each line is the CPU number.
|
|
|
|
|
CPUs numbers followed by an exclamation mark are offline,
|
|
|
|
|
but have been online at least once since boot. There will be
|
|
|
|
|
no output for CPUs that have never been online, which can be
|
|
|
|
|
a good thing in the surprisingly common case where NR_CPUS is
|
|
|
|
|
substantially larger than the number of actual CPUs.
|
|
|
|
|
|
|
|
|
|
o "c" is the count of grace periods that this CPU believes have
|
|
|
|
|
completed. CPUs in dynticks idle mode may lag quite a ways
|
|
|
|
|
behind, for example, CPU 4 under "rcu" above, which has slept
|
|
|
|
|
through the past 25 RCU grace periods. It is not unusual to
|
|
|
|
|
see CPUs lagging by thousands of grace periods.
|
|
|
|
|
|
|
|
|
|
o "g" is the count of grace periods that this CPU believes have
|
|
|
|
|
started. Again, CPUs in dynticks idle mode may lag behind.
|
|
|
|
|
If the "c" and "g" values are equal, this CPU has already
|
|
|
|
|
reported a quiescent state for the last RCU grace period that
|
|
|
|
|
it is aware of, otherwise, the CPU believes that it owes RCU a
|
|
|
|
|
quiescent state.
|
|
|
|
|
|
|
|
|
|
o "pq" indicates that this CPU has passed through a quiescent state
|
|
|
|
|
for the current grace period. It is possible for "pq" to be
|
|
|
|
|
"1" and "c" different than "g", which indicates that although
|
|
|
|
|
the CPU has passed through a quiescent state, either (1) this
|
|
|
|
|
CPU has not yet reported that fact, (2) some other CPU has not
|
|
|
|
|
yet reported for this grace period, or (3) both.
|
|
|
|
|
|
|
|
|
|
o "pqc" indicates which grace period the last-observed quiescent
|
|
|
|
|
state for this CPU corresponds to. This is important for handling
|
|
|
|
|
the race between CPU 0 reporting an extended dynticks-idle
|
|
|
|
|
quiescent state for CPU 1 and CPU 1 suddenly waking up and
|
|
|
|
|
reporting its own quiescent state. If CPU 1 was the last CPU
|
|
|
|
|
for the current grace period, then the CPU that loses this race
|
|
|
|
|
will attempt to incorrectly mark CPU 1 as having checked in for
|
|
|
|
|
the next grace period!
|
|
|
|
|
|
|
|
|
|
o "qp" indicates that RCU still expects a quiescent state from
|
|
|
|
|
this CPU.
|
|
|
|
|
|
|
|
|
|
o "dt" is the current value of the dyntick counter that is incremented
|
|
|
|
|
when entering or leaving dynticks idle state, either by the
|
|
|
|
|
scheduler or by irq. The number after the "/" is the interrupt
|
|
|
|
|
nesting depth when in dyntick-idle state, or one greater than
|
|
|
|
|
the interrupt-nesting depth otherwise.
|
|
|
|
|
|
|
|
|
|
This field is displayed only for CONFIG_NO_HZ kernels.
|
|
|
|
|
|
|
|
|
|
o "dn" is the current value of the dyntick counter that is incremented
|
|
|
|
|
when entering or leaving dynticks idle state via NMI. If both
|
|
|
|
|
the "dt" and "dn" values are even, then this CPU is in dynticks
|
|
|
|
|
idle mode and may be ignored by RCU. If either of these two
|
|
|
|
|
counters is odd, then RCU must be alert to the possibility of
|
|
|
|
|
an RCU read-side critical section running on this CPU.
|
|
|
|
|
|
|
|
|
|
This field is displayed only for CONFIG_NO_HZ kernels.
|
|
|
|
|
|
|
|
|
|
o "df" is the number of times that some other CPU has forced a
|
|
|
|
|
quiescent state on behalf of this CPU due to this CPU being in
|
|
|
|
|
dynticks-idle state.
|
|
|
|
|
|
|
|
|
|
This field is displayed only for CONFIG_NO_HZ kernels.
|
|
|
|
|
|
|
|
|
|
o "of" is the number of times that some other CPU has forced a
|
|
|
|
|
quiescent state on behalf of this CPU due to this CPU being
|
|
|
|
|
offline. In a perfect world, this might neve happen, but it
|
|
|
|
|
turns out that offlining and onlining a CPU can take several grace
|
|
|
|
|
periods, and so there is likely to be an extended period of time
|
|
|
|
|
when RCU believes that the CPU is online when it really is not.
|
|
|
|
|
Please note that erring in the other direction (RCU believing a
|
|
|
|
|
CPU is offline when it is really alive and kicking) is a fatal
|
|
|
|
|
error, so it makes sense to err conservatively.
|
|
|
|
|
|
|
|
|
|
o "ri" is the number of times that RCU has seen fit to send a
|
|
|
|
|
reschedule IPI to this CPU in order to get it to report a
|
|
|
|
|
quiescent state.
|
|
|
|
|
|
|
|
|
|
o "ql" is the number of RCU callbacks currently residing on
|
|
|
|
|
this CPU. This is the total number of callbacks, regardless
|
|
|
|
|
of what state they are in (new, waiting for grace period to
|
|
|
|
|
start, waiting for grace period to end, ready to invoke).
|
|
|
|
|
|
|
|
|
|
o "b" is the batch limit for this CPU. If more than this number
|
|
|
|
|
of RCU callbacks is ready to invoke, then the remainder will
|
|
|
|
|
be deferred.
|
|
|
|
|
|
2009-04-14 06:31:18 +02:00
|
|
|
There is also an rcu/rcudata.csv file with the same information in
|
|
|
|
|
comma-separated-variable spreadsheet format.
|
|
|
|
|
|
"Tree RCU": scalable classic RCU implementation
This patch fixes a long-standing performance bug in classic RCU that
results in massive internal-to-RCU lock contention on systems with
more than a few hundred CPUs. Although this patch creates a separate
flavor of RCU for ease of review and patch maintenance, it is intended
to replace classic RCU.
This patch still handles stress better than does mainline, so I am still
calling it ready for inclusion. This patch is against the -tip tree.
Nevertheless, experience on an actual 1000+ CPU machine would still be
most welcome.
Most of the changes noted below were found while creating an rcutiny
(which should permit ejecting the current rcuclassic) and while doing
detailed line-by-line documentation.
Updates from v9 (http://lkml.org/lkml/2008/12/2/334):
o Fixes from remainder of line-by-line code walkthrough,
including comment spelling, initialization, undesirable
narrowing due to type conversion, removing redundant memory
barriers, removing redundant local-variable initialization,
and removing redundant local variables.
I do not believe that any of these fixes address the CPU-hotplug
issues that Andi Kleen was seeing, but please do give it a whirl
in case the machine is smarter than I am.
A writeup from the walkthrough may be found at the following
URL, in case you are suffering from terminal insomnia or
masochism:
http://www.kernel.org/pub/linux/kernel/people/paulmck/tmp/rcutree-walkthrough.2008.12.16a.pdf
o Made rcutree tracing use seq_file, as suggested some time
ago by Lai Jiangshan.
o Added a .csv variant of the rcudata debugfs trace file, to allow
people having thousands of CPUs to drop the data into
a spreadsheet. Tested with oocalc and gnumeric. Updated
documentation to suit.
Updates from v8 (http://lkml.org/lkml/2008/11/15/139):
o Fix a theoretical race between grace-period initialization and
force_quiescent_state() that could occur if more than three
jiffies were required to carry out the grace-period
initialization. Which it might, if you had enough CPUs.
o Apply Ingo's printk-standardization patch.
o Substitute local variables for repeated accesses to global
variables.
o Fix comment misspellings and redundant (but harmless) increments
of ->n_rcu_pending (this latter after having explicitly added it).
o Apply checkpatch fixes.
Updates from v7 (http://lkml.org/lkml/2008/10/10/291):
o Fixed a number of problems noted by Gautham Shenoy, including
the cpu-stall-detection bug that he was having difficulty
convincing me was real. ;-)
o Changed cpu-stall detection to wait for ten seconds rather than
three in order to reduce false positive, as suggested by Ingo
Molnar.
o Produced a design document (http://lwn.net/Articles/305782/).
The act of writing this document uncovered a number of both
theoretical and "here and now" bugs as noted below.
o Fix dynticks_nesting accounting confusion, simplify WARN_ON()
condition, fix kerneldoc comments, and add memory barriers
in dynticks interface functions.
o Add more data to tracing.
o Remove unused "rcu_barrier" field from rcu_data structure.
o Count calls to rcu_pending() from scheduling-clock interrupt
to use as a surrogate timebase should jiffies stop counting.
o Fix a theoretical race between force_quiescent_state() and
grace-period initialization. Yes, initialization does have to
go on for some jiffies for this race to occur, but given enough
CPUs...
Updates from v6 (http://lkml.org/lkml/2008/9/23/448):
o Fix a number of checkpatch.pl complaints.
o Apply review comments from Ingo Molnar and Lai Jiangshan
on the stall-detection code.
o Fix several bugs in !CONFIG_SMP builds.
o Fix a misspelled config-parameter name so that RCU now announces
at boot time if stall detection is configured.
o Run tests on numerous combinations of configurations parameters,
which after the fixes above, now build and run correctly.
Updates from v5 (http://lkml.org/lkml/2008/9/15/92, bad subject line):
o Fix a compiler error in the !CONFIG_FANOUT_EXACT case (blew a
changeset some time ago, and finally got around to retesting
this option).
o Fix some tracing bugs in rcupreempt that caused incorrect
totals to be printed.
o I now test with a more brutal random-selection online/offline
script (attached). Probably more brutal than it needs to be
on the people reading it as well, but so it goes.
o A number of optimizations and usability improvements:
o Make rcu_pending() ignore the grace-period timeout when
there is no grace period in progress.
o Make force_quiescent_state() avoid going for a global
lock in the case where there is no grace period in
progress.
o Rearrange struct fields to improve struct layout.
o Make call_rcu() initiate a grace period if RCU was
idle, rather than waiting for the next scheduling
clock interrupt.
o Invoke rcu_irq_enter() and rcu_irq_exit() only when
idle, as suggested by Andi Kleen. I still don't
completely trust this change, and might back it out.
o Make CONFIG_RCU_TRACE be the single config variable
manipulated for all forms of RCU, instead of the prior
confusion.
o Document tracing files and formats for both rcupreempt
and rcutree.
Updates from v4 for those missing v5 given its bad subject line:
o Separated dynticks interface so that NMIs and irqs call separate
functions, greatly simplifying it. In particular, this code
no longer requires a proof of correctness. ;-)
o Separated dynticks state out into its own per-CPU structure,
avoiding the duplicated accounting.
o The case where a dynticks-idle CPU runs an irq handler that
invokes call_rcu() is now correctly handled, forcing that CPU
out of dynticks-idle mode.
o Review comments have been applied (thank you all!!!).
For but one example, fixed the dynticks-ordering issue that
Manfred pointed out, saving me much debugging. ;-)
o Adjusted rcuclassic and rcupreempt to handle dynticks changes.
Attached is an updated patch to Classic RCU that applies a hierarchy,
greatly reducing the contention on the top-level lock for large machines.
This passes 10-hour concurrent rcutorture and online-offline testing on
128-CPU ppc64 without dynticks enabled, and exposes some timekeeping
bugs in presence of dynticks (exciting working on a system where
"sleep 1" hangs until interrupted...), which were fixed in the
2.6.27 kernel. It is getting more reliable than mainline by some
measures, so the next version will be against -tip for inclusion.
See also Manfred Spraul's recent patches (or his earlier work from
2004 at http://marc.info/?l=linux-kernel&m=108546384711797&w=2).
We will converge onto a common patch in the fullness of time, but are
currently exploring different regions of the design space. That said,
I have already gratefully stolen quite a few of Manfred's ideas.
This patch provides CONFIG_RCU_FANOUT, which controls the bushiness
of the RCU hierarchy. Defaults to 32 on 32-bit machines and 64 on
64-bit machines. If CONFIG_NR_CPUS is less than CONFIG_RCU_FANOUT,
there is no hierarchy. By default, the RCU initialization code will
adjust CONFIG_RCU_FANOUT to balance the hierarchy, so strongly NUMA
architectures may choose to set CONFIG_RCU_FANOUT_EXACT to disable
this balancing, allowing the hierarchy to be exactly aligned to the
underlying hardware. Up to two levels of hierarchy are permitted
(in addition to the root node), allowing up to 16,384 CPUs on 32-bit
systems and up to 262,144 CPUs on 64-bit systems. I just know that I
am going to regret saying this, but this seems more than sufficient
for the foreseeable future. (Some architectures might wish to set
CONFIG_RCU_FANOUT=4, which would limit such architectures to 64 CPUs.
If this becomes a real problem, additional levels can be added, but I
doubt that it will make a significant difference on real hardware.)
In the common case, a given CPU will manipulate its private rcu_data
structure and the rcu_node structure that it shares with its immediate
neighbors. This can reduce both lock and memory contention by multiple
orders of magnitude, which should eliminate the need for the strange
manipulations that are reported to be required when running Linux on
very large systems.
Some shortcomings:
o More bugs will probably surface as a result of an ongoing
line-by-line code inspection.
Patches will be provided as required.
o There are probably hangs, rcutorture failures, &c. Seems
quite stable on a 128-CPU machine, but that is kind of small
compared to 4096 CPUs. However, seems to do better than
mainline.
Patches will be provided as required.
o The memory footprint of this version is several KB larger
than rcuclassic.
A separate UP-only rcutiny patch will be provided, which will
reduce the memory footprint significantly, even compared
to the old rcuclassic. One such patch passes light testing,
and has a memory footprint smaller even than rcuclassic.
Initial reaction from various embedded guys was "it is not
worth it", so am putting it aside.
Credits:
o Manfred Spraul for ideas, review comments, and bugs spotted,
as well as some good friendly competition. ;-)
o Josh Triplett, Ingo Molnar, Peter Zijlstra, Mathieu Desnoyers,
Lai Jiangshan, Andi Kleen, Andy Whitcroft, and Andrew Morton
for reviews and comments.
o Thomas Gleixner for much-needed help with some timer issues
(see patches below).
o Jon M. Tollefson, Tim Pepper, Andrew Theurer, Jose R. Santos,
Andy Whitcroft, Darrick Wong, Nishanth Aravamudan, Anton
Blanchard, Dave Kleikamp, and Nathan Lynch for keeping machines
alive despite my heavy abuse^Wtesting.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-12-18 21:55:32 +01:00
|
|
|
|
|
|
|
|
The output of "cat rcu/rcugp" looks as follows:
|
|
|
|
|
|
|
|
|
|
rcu: completed=33062 gpnum=33063
|
|
|
|
|
rcu_bh: completed=464 gpnum=464
|
|
|
|
|
|
|
|
|
|
Again, this output is for both "rcu" and "rcu_bh". The fields are
|
|
|
|
|
taken from the rcu_state structure, and are as follows:
|
|
|
|
|
|
|
|
|
|
o "completed" is the number of grace periods that have completed.
|
|
|
|
|
It is comparable to the "c" field from rcu/rcudata in that a
|
|
|
|
|
CPU whose "c" field matches the value of "completed" is aware
|
|
|
|
|
that the corresponding RCU grace period has completed.
|
|
|
|
|
|
|
|
|
|
o "gpnum" is the number of grace periods that have started. It is
|
|
|
|
|
comparable to the "g" field from rcu/rcudata in that a CPU
|
|
|
|
|
whose "g" field matches the value of "gpnum" is aware that the
|
|
|
|
|
corresponding RCU grace period has started.
|
|
|
|
|
|
|
|
|
|
If these two fields are equal (as they are for "rcu_bh" above),
|
|
|
|
|
then there is no grace period in progress, in other words, RCU
|
|
|
|
|
is idle. On the other hand, if the two fields differ (as they
|
|
|
|
|
do for "rcu" above), then an RCU grace period is in progress.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
The output of "cat rcu/rcuhier" looks as follows, with very long lines:
|
|
|
|
|
|
|
|
|
|
c=6902 g=6903 s=2 jfq=3 j=72c7 nfqs=13142/nfqsng=0(13142) fqlh=6
|
|
|
|
|
1/1 0:127 ^0
|
|
|
|
|
3/3 0:35 ^0 0/0 36:71 ^1 0/0 72:107 ^2 0/0 108:127 ^3
|
|
|
|
|
3/3f 0:5 ^0 2/3 6:11 ^1 0/0 12:17 ^2 0/0 18:23 ^3 0/0 24:29 ^4 0/0 30:35 ^5 0/0 36:41 ^0 0/0 42:47 ^1 0/0 48:53 ^2 0/0 54:59 ^3 0/0 60:65 ^4 0/0 66:71 ^5 0/0 72:77 ^0 0/0 78:83 ^1 0/0 84:89 ^2 0/0 90:95 ^3 0/0 96:101 ^4 0/0 102:107 ^5 0/0 108:113 ^0 0/0 114:119 ^1 0/0 120:125 ^2 0/0 126:127 ^3
|
|
|
|
|
rcu_bh:
|
|
|
|
|
c=-226 g=-226 s=1 jfq=-5701 j=72c7 nfqs=88/nfqsng=0(88) fqlh=0
|
|
|
|
|
0/1 0:127 ^0
|
|
|
|
|
0/3 0:35 ^0 0/0 36:71 ^1 0/0 72:107 ^2 0/0 108:127 ^3
|
|
|
|
|
0/3f 0:5 ^0 0/3 6:11 ^1 0/0 12:17 ^2 0/0 18:23 ^3 0/0 24:29 ^4 0/0 30:35 ^5 0/0 36:41 ^0 0/0 42:47 ^1 0/0 48:53 ^2 0/0 54:59 ^3 0/0 60:65 ^4 0/0 66:71 ^5 0/0 72:77 ^0 0/0 78:83 ^1 0/0 84:89 ^2 0/0 90:95 ^3 0/0 96:101 ^4 0/0 102:107 ^5 0/0 108:113 ^0 0/0 114:119 ^1 0/0 120:125 ^2 0/0 126:127 ^3
|
|
|
|
|
|
|
|
|
|
This is once again split into "rcu" and "rcu_bh" portions. The fields are
|
|
|
|
|
as follows:
|
|
|
|
|
|
|
|
|
|
o "c" is exactly the same as "completed" under rcu/rcugp.
|
|
|
|
|
|
|
|
|
|
o "g" is exactly the same as "gpnum" under rcu/rcugp.
|
|
|
|
|
|
|
|
|
|
o "s" is the "signaled" state that drives force_quiescent_state()'s
|
|
|
|
|
state machine.
|
|
|
|
|
|
|
|
|
|
o "jfq" is the number of jiffies remaining for this grace period
|
|
|
|
|
before force_quiescent_state() is invoked to help push things
|
|
|
|
|
along. Note that CPUs in dyntick-idle mode thoughout the grace
|
|
|
|
|
period will not report on their own, but rather must be check by
|
|
|
|
|
some other CPU via force_quiescent_state().
|
|
|
|
|
|
|
|
|
|
o "j" is the low-order four hex digits of the jiffies counter.
|
|
|
|
|
Yes, Paul did run into a number of problems that turned out to
|
|
|
|
|
be due to the jiffies counter no longer counting. Why do you ask?
|
|
|
|
|
|
|
|
|
|
o "nfqs" is the number of calls to force_quiescent_state() since
|
|
|
|
|
boot.
|
|
|
|
|
|
|
|
|
|
o "nfqsng" is the number of useless calls to force_quiescent_state(),
|
|
|
|
|
where there wasn't actually a grace period active. This can
|
|
|
|
|
happen due to races. The number in parentheses is the difference
|
|
|
|
|
between "nfqs" and "nfqsng", or the number of times that
|
|
|
|
|
force_quiescent_state() actually did some real work.
|
|
|
|
|
|
|
|
|
|
o "fqlh" is the number of calls to force_quiescent_state() that
|
|
|
|
|
exited immediately (without even being counted in nfqs above)
|
|
|
|
|
due to contention on ->fqslock.
|
|
|
|
|
|
|
|
|
|
o Each element of the form "1/1 0:127 ^0" represents one struct
|
|
|
|
|
rcu_node. Each line represents one level of the hierarchy, from
|
|
|
|
|
root to leaves. It is best to think of the rcu_data structures
|
|
|
|
|
as forming yet another level after the leaves. Note that there
|
|
|
|
|
might be either one, two, or three levels of rcu_node structures,
|
|
|
|
|
depending on the relationship between CONFIG_RCU_FANOUT and
|
|
|
|
|
CONFIG_NR_CPUS.
|
|
|
|
|
|
|
|
|
|
o The numbers separated by the "/" are the qsmask followed
|
|
|
|
|
by the qsmaskinit. The qsmask will have one bit
|
|
|
|
|
set for each entity in the next lower level that
|
|
|
|
|
has not yet checked in for the current grace period.
|
|
|
|
|
The qsmaskinit will have one bit for each entity that is
|
|
|
|
|
currently expected to check in during each grace period.
|
|
|
|
|
The value of qsmaskinit is assigned to that of qsmask
|
|
|
|
|
at the beginning of each grace period.
|
|
|
|
|
|
|
|
|
|
For example, for "rcu", the qsmask of the first entry
|
|
|
|
|
of the lowest level is 0x14, meaning that we are still
|
|
|
|
|
waiting for CPUs 2 and 4 to check in for the current
|
|
|
|
|
grace period.
|
|
|
|
|
|
|
|
|
|
o The numbers separated by the ":" are the range of CPUs
|
|
|
|
|
served by this struct rcu_node. This can be helpful
|
|
|
|
|
in working out how the hierarchy is wired together.
|
|
|
|
|
|
|
|
|
|
For example, the first entry at the lowest level shows
|
|
|
|
|
"0:5", indicating that it covers CPUs 0 through 5.
|
|
|
|
|
|
|
|
|
|
o The number after the "^" indicates the bit in the
|
|
|
|
|
next higher level rcu_node structure that this
|
|
|
|
|
rcu_node structure corresponds to.
|
|
|
|
|
|
|
|
|
|
For example, the first entry at the lowest level shows
|
|
|
|
|
"^0", indicating that it corresponds to bit zero in
|
|
|
|
|
the first entry at the middle level.
|
2009-04-14 06:31:18 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
The output of "cat rcu/rcu_pending" looks as follows:
|
|
|
|
|
|
|
|
|
|
rcu:
|
|
|
|
|
0 np=255892 qsp=53936 cbr=0 cng=14417 gpc=10033 gps=24320 nf=6445 nn=146741
|
|
|
|
|
1 np=261224 qsp=54638 cbr=0 cng=25723 gpc=16310 gps=2849 nf=5912 nn=155792
|
|
|
|
|
2 np=237496 qsp=49664 cbr=0 cng=2762 gpc=45478 gps=1762 nf=1201 nn=136629
|
|
|
|
|
3 np=236249 qsp=48766 cbr=0 cng=286 gpc=48049 gps=1218 nf=207 nn=137723
|
|
|
|
|
4 np=221310 qsp=46850 cbr=0 cng=26 gpc=43161 gps=4634 nf=3529 nn=123110
|
|
|
|
|
5 np=237332 qsp=48449 cbr=0 cng=54 gpc=47920 gps=3252 nf=201 nn=137456
|
|
|
|
|
6 np=219995 qsp=46718 cbr=0 cng=50 gpc=42098 gps=6093 nf=4202 nn=120834
|
|
|
|
|
7 np=249893 qsp=49390 cbr=0 cng=72 gpc=38400 gps=17102 nf=41 nn=144888
|
|
|
|
|
rcu_bh:
|
|
|
|
|
0 np=146741 qsp=1419 cbr=0 cng=6 gpc=0 gps=0 nf=2 nn=145314
|
|
|
|
|
1 np=155792 qsp=12597 cbr=0 cng=0 gpc=4 gps=8 nf=3 nn=143180
|
|
|
|
|
2 np=136629 qsp=18680 cbr=0 cng=0 gpc=7 gps=6 nf=0 nn=117936
|
|
|
|
|
3 np=137723 qsp=2843 cbr=0 cng=0 gpc=10 gps=7 nf=0 nn=134863
|
|
|
|
|
4 np=123110 qsp=12433 cbr=0 cng=0 gpc=4 gps=2 nf=0 nn=110671
|
|
|
|
|
5 np=137456 qsp=4210 cbr=0 cng=0 gpc=6 gps=5 nf=0 nn=133235
|
|
|
|
|
6 np=120834 qsp=9902 cbr=0 cng=0 gpc=6 gps=3 nf=2 nn=110921
|
|
|
|
|
7 np=144888 qsp=26336 cbr=0 cng=0 gpc=8 gps=2 nf=0 nn=118542
|
|
|
|
|
|
|
|
|
|
As always, this is once again split into "rcu" and "rcu_bh" portions.
|
|
|
|
|
The fields are as follows:
|
|
|
|
|
|
|
|
|
|
o "np" is the number of times that __rcu_pending() has been invoked
|
|
|
|
|
for the corresponding flavor of RCU.
|
|
|
|
|
|
|
|
|
|
o "qsp" is the number of times that the RCU was waiting for a
|
|
|
|
|
quiescent state from this CPU.
|
|
|
|
|
|
|
|
|
|
o "cbr" is the number of times that this CPU had RCU callbacks
|
|
|
|
|
that had passed through a grace period, and were thus ready
|
|
|
|
|
to be invoked.
|
|
|
|
|
|
|
|
|
|
o "cng" is the number of times that this CPU needed another
|
|
|
|
|
grace period while RCU was idle.
|
|
|
|
|
|
|
|
|
|
o "gpc" is the number of times that an old grace period had
|
|
|
|
|
completed, but this CPU was not yet aware of it.
|
|
|
|
|
|
|
|
|
|
o "gps" is the number of times that a new grace period had started,
|
|
|
|
|
but this CPU was not yet aware of it.
|
|
|
|
|
|
|
|
|
|
o "nf" is the number of times that this CPU suspected that the
|
|
|
|
|
current grace period had run for too long, and thus needed to
|
|
|
|
|
be forced.
|
|
|
|
|
|
|
|
|
|
Please note that "forcing" consists of sending resched IPIs
|
|
|
|
|
to holdout CPUs. If that CPU really still is in an old RCU
|
|
|
|
|
read-side critical section, then we really do have to wait for it.
|
|
|
|
|
The assumption behing "forcing" is that the CPU is not still in
|
|
|
|
|
an old RCU read-side critical section, but has not yet responded
|
|
|
|
|
for some other reason.
|
|
|
|
|
|
|
|
|
|
o "nn" is the number of times that this CPU needed nothing. Alert
|
|
|
|
|
readers will note that the rcu "nn" number for a given CPU very
|
|
|
|
|
closely matches the rcu_bh "np" number for that same CPU. This
|
|
|
|
|
is due to short-circuit evaluation in rcu_pending().
|