| Message ID | 20240104162510.72773-2-urezki@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | Reduce synchronize_rcu() latency(v4) | expand |
On Thu, Jan 04, 2024 at 05:25:07PM +0100, Uladzislau Rezki (Sony) wrote: > A call to a synchronize_rcu() can be optimized from a latency > point of view. Workloads which depend on this can benefit of it. > > The delay of wakeme_after_rcu() callback, which unblocks a waiter, > depends on several factors: > > - how fast a process of offloading is started. Combination of: > - !CONFIG_RCU_NOCB_CPU/CONFIG_RCU_NOCB_CPU; > - !CONFIG_RCU_LAZY/CONFIG_RCU_LAZY; > - other. > - when started, invoking path is interrupted due to: > - time limit; > - need_resched(); > - if limit is reached. > - where in a nocb list it is located; > - how fast previous callbacks completed; > > Example: > > 1. On our embedded devices i can easily trigger the scenario when > it is a last in the list out of ~3600 callbacks: > > <snip> > <...>-29 [001] d..1. 21950.145313: rcu_batch_start: rcu_preempt CBs=3613 bl=28 > ... > <...>-29 [001] ..... 21950.152578: rcu_invoke_callback: rcu_preempt rhp=00000000b2d6dee8 func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152579: rcu_invoke_callback: rcu_preempt rhp=00000000a446f607 func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152580: rcu_invoke_callback: rcu_preempt rhp=00000000a5cab03b func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152581: rcu_invoke_callback: rcu_preempt rhp=0000000013b7e5ee func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152582: rcu_invoke_callback: rcu_preempt rhp=000000000a8ca6f9 func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152583: rcu_invoke_callback: rcu_preempt rhp=000000008f162ca8 func=wakeme_after_rcu.cfi_jt > <...>-29 [001] d..1. 21950.152625: rcu_batch_end: rcu_preempt CBs-invoked=3612 idle=.... > <snip> > > 2. We use cpuset/cgroup to classify tasks and assign them into > different cgroups. For example "backgrond" group which binds tasks > only to little CPUs or "foreground" which makes use of all CPUs. > Tasks can be migrated between groups by a request if an acceleration > is needed. > > See below an example how "surfaceflinger" task gets migrated. > Initially it is located in the "system-background" cgroup which > allows to run only on little cores. In order to speed it up it > can be temporary moved into "foreground" cgroup which allows > to use big/all CPUs: > > cgroup_attach_task(): > -> cgroup_migrate_execute() > -> cpuset_can_attach() > -> percpu_down_write() > -> rcu_sync_enter() > -> synchronize_rcu() > -> now move tasks to the new cgroup. > -> cgroup_migrate_finish() > > <snip> > rcuop/1-29 [000] ..... 7030.528570: rcu_invoke_callback: rcu_preempt rhp=00000000461605e0 func=wakeme_after_rcu.cfi_jt > PERFD-SERVER-1855 [000] d..1. 7030.530293: cgroup_attach_task: dst_root=3 dst_id=22 dst_level=1 dst_path=/foreground pid=1900 comm=surfaceflinger > TimerDispatch-2768 [002] d..5. 7030.537542: sched_migrate_task: comm=surfaceflinger pid=1900 prio=98 orig_cpu=0 dest_cpu=4 > <snip> > > "Boosting a task" depends on synchronize_rcu() latency: > > - first trace shows a completion of synchronize_rcu(); > - second shows attaching a task to a new group; > - last shows a final step when migration occurs. > > 3. To address this drawback, maintain a separate track that consists > of synchronize_rcu() callers only. After completion of a grace period > users are deferred to a dedicated worker to process requests. > > 4. This patch reduces the latency of synchronize_rcu() approximately > by ~30-40% on synthetic tests. The real test case, camera launch time, > shows(time is in milliseconds): > > 1-run 542 vs 489 improvement 9% > 2-run 540 vs 466 improvement 13% > 3-run 518 vs 468 improvement 9% > 4-run 531 vs 457 improvement 13% > 5-run 548 vs 475 improvement 13% > 6-run 509 vs 484 improvement 4% > > Synthetic test(no "noise" from other callbacks): > Hardware: x86_64 64 CPUs, 64GB of memory > Linux-6.6 > > - 10K tasks(simultaneous); > - each task does(1000 loops) > synchronize_rcu(); > kfree(p); > > default: CONFIG_RCU_NOCB_CPU: takes 54 seconds to complete all users; > patch: CONFIG_RCU_NOCB_CPU: takes 35 seconds to complete all users. > > Running 60K gives approximately same results on my setup. Please note > it is without any interaction with another type of callbacks, otherwise > it will impact a lot a default case. > > 5. An extra CONFIG_RCU_SR_NORMAL_DEBUG_GP kernel option is added > which enables additional debugging for detecting a grace period > incompletion for synchronize_rcu() users. If a GP is not fully > passed for any user, the warning message is emitted. > > 6. By default it is disabled. To enable this perform one of the > below sequence: > > echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp > or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1" > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Hi Uladzislau, I've tried your patches (v3) on Android with 6.1.43 kernel. The test cycles 10 apps (including camera) sequentially for 100 iterations. I've set rcu_normal to override the rcu_expedited in the boot parameters: adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu rcupdate.rcu_normal=1 rcupdate.rcu_expedited=1 rcu_nocbs=0-7 The configurations are: A - echo 0 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp B - echo 1 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp Results: = APP LAUNCH TIME = delta (B-A) ratio(%) overall_app_launch_time(ms) -11399.00 -6.65 == camera_launch_time type delta(B-A %) A_count B_count HOT -7.05 99 99 COLD -6.33 1 1 === Function Latencies === Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit nsec : count distribution nsec : count distribution 0 -> 1 : 0 | | 0 -> 1 : 0 | | 2 -> 3 : 0 | | 2 -> 3 : 0 | | 4 -> 7 : 0 | | 4 -> 7 : 0 | | 8 -> 15 : 0 | | 8 -> 15 : 0 | | 16 -> 31 : 0 | | 16 -> 31 : 0 | | 32 -> 63 : 0 | | 32 -> 63 : 0 | | 64 -> 127 : 0 | | 64 -> 127 : 0 | | 128 -> 255 : 0 | | 128 -> 255 : 0 | | 256 -> 511 : 0 | | 256 -> 511 : 0 | | 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | 4194304 -> 8388607 : 871 |** | 4194304 -> 8388607 : 1180 |**** | 8388608 -> 16777215 : 3204 |******** | 8388608 -> 16777215 : 7020 |************************* | 16777216 -> 33554431 : 15013 |****************************************| 16777216 -> 33554431 : 10952 |****************************************| Exiting trace of synchronize_rcu_expedited Exiting trace of synchronize_rcu_expedited Tracing synchronize_rcu. Hit Ctrl-C to exit Tracing synchronize_rcu. Hit Ctrl-C to exit nsec : count distribution nsec : count distribution 0 -> 1 : 0 | | 0 -> 1 : 0 | | 2 -> 3 : 0 | | 2 -> 3 : 0 | | 4 -> 7 : 0 | | 4 -> 7 : 0 | | 8 -> 15 : 0 | | 8 -> 15 : 0 | | 16 -> 31 : 0 | | 16 -> 31 : 0 | | 32 -> 63 : 0 | | 32 -> 63 : 0 | | 64 -> 127 : 0 | | 64 -> 127 : 0 | | 128 -> 255 : 0 | | 128 -> 255 : 0 | | 256 -> 511 : 0 | | 256 -> 511 : 0 | | 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | 4194304 -> 8388607 : 861 |** | 4194304 -> 8388607 : 1136 |**** | 8388608 -> 16777215 : 3111 |******** | 8388608 -> 16777215 : 6320 |************************ | 16777216 -> 33554431 : 13901 |****************************************| 16777216 -> 33554431 : 10484 |****************************************| Exiting trace of synchronize_rcu Exiting trace of synchronize_rcu Interestingly I tried the same experiment without rcu_normal=1 (leaving rcu_expedited=1): adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu rcupdate.rcu_expedited=1 rcu_nocbs=0-7 In this case I also saw the -6 to -7% decrease in the app launch times but I don't have a good explanation why that would be? (The fucntion latency histograms in this case didn't show any significant difference). Do you have any insight why this may happen? Thanks, Kalesh > --- > .../admin-guide/kernel-parameters.txt | 14 ++ > kernel/rcu/Kconfig.debug | 12 ++ > kernel/rcu/tree.c | 138 +++++++++++++++++- > kernel/rcu/tree_exp.h | 2 +- > 4 files changed, 164 insertions(+), 2 deletions(-) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index 17a454909ab4..2cca75e4f0c6 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -5047,6 +5047,20 @@ > delay, memory pressure or callback list growing too > big. > > + rcutree.rcu_normal_wake_from_gp= [KNL] > + Reduces a latency of synchronize_rcu() call. This approach > + maintains its own track of synchronize_rcu() callers, so it > + does not interact with regular callbacks because it does not > + use a call_rcu[_hurry]() path. Please note, this is for a > + normal grace period. > + > + How to enable it: > + > + echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp > + or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1" > + > + Default is 0. > + > rcuscale.gp_async= [KNL] > Measure performance of asynchronous > grace-period primitives such as call_rcu(). > diff --git a/kernel/rcu/Kconfig.debug b/kernel/rcu/Kconfig.debug > index 9b0b52e1836f..4812c6249185 100644 > --- a/kernel/rcu/Kconfig.debug > +++ b/kernel/rcu/Kconfig.debug > @@ -168,4 +168,16 @@ config RCU_STRICT_GRACE_PERIOD > when looking for certain types of RCU usage bugs, for example, > too-short RCU read-side critical sections. > > +config RCU_SR_NORMAL_DEBUG_GP > + bool "Debug synchronize_rcu() callers for a grace period completion" > + depends on DEBUG_KERNEL && RCU_EXPERT > + default n > + help > + This option enables additional debugging for detecting a grace > + period incompletion for synchronize_rcu() users. If a GP is not > + fully passed for any user, the warning message is emitted. > + > + Say Y here if you want to enable such debugging > + Say N if you are unsure. > + > endmenu # "RCU Debugging" > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c > index 499803234176..b756c40e4960 100644 > --- a/kernel/rcu/tree.c > +++ b/kernel/rcu/tree.c > @@ -1422,6 +1422,106 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap) > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > } > > +/* > + * There are three lists for handling synchronize_rcu() users. > + * A first list corresponds to new coming users, second for users > + * which wait for a grace period and third is for which a grace > + * period is passed. > + */ > +static struct sr_normal_state { > + struct llist_head srs_next; /* request a GP users. */ > + struct llist_head srs_wait; /* wait for GP users. */ > + struct llist_head srs_done; /* ready for GP users. */ > + > + /* > + * In order to add a batch of nodes to already > + * existing srs-done-list, a tail of srs-wait-list > + * is maintained. > + */ > + struct llist_node *srs_wait_tail; > +} sr; > + > +/* Disabled by default. */ > +static int rcu_normal_wake_from_gp; > +module_param(rcu_normal_wake_from_gp, int, 0644); > + > +static void rcu_sr_normal_complete(struct llist_node *node) > +{ > + struct rcu_synchronize *rs = container_of( > + (struct rcu_head *) node, struct rcu_synchronize, head); > + unsigned long oldstate = (unsigned long) rs->head.func; > + > + WARN_ONCE(IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP) && > + !poll_state_synchronize_rcu(oldstate), > + "A full grace period is not passed yet: %lu", > + rcu_seq_diff(get_state_synchronize_rcu(), oldstate)); > + > + /* Finally. */ > + complete(&rs->completion); > +} > + > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) > +{ > + struct llist_node *done, *rcu, *next; > + > + done = llist_del_all(&sr.srs_done); > + if (!done) > + return; > + > + llist_for_each_safe(rcu, next, done) > + rcu_sr_normal_complete(rcu); > +} > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); > + > +/* > + * Helper function for rcu_gp_cleanup(). > + */ > +static void rcu_sr_normal_gp_cleanup(void) > +{ > + struct llist_node *head, *tail; > + > + if (llist_empty(&sr.srs_wait)) > + return; > + > + tail = READ_ONCE(sr.srs_wait_tail); > + head = __llist_del_all(&sr.srs_wait); > + > + if (head) { > + /* Can be not empty. */ > + llist_add_batch(head, tail, &sr.srs_done); > + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); > + } > +} > + > +/* > + * Helper function for rcu_gp_init(). > + */ > +static void rcu_sr_normal_gp_init(void) > +{ > + struct llist_node *head, *tail; > + > + if (llist_empty(&sr.srs_next)) > + return; > + > + tail = llist_del_all(&sr.srs_next); > + head = llist_reverse_order(tail); > + > + /* > + * A waiting list of GP should be empty on this step, > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), > + * rolls it over. If not, it is a BUG, warn a user. > + */ > + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); > + > + WRITE_ONCE(sr.srs_wait_tail, tail); > + __llist_add_batch(head, tail, &sr.srs_wait); > +} > + > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) > +{ > + llist_add((struct llist_node *) &rs->head, &sr.srs_next); > +} > + > /* > * Initialize a new grace period. Return false if no grace period required. > */ > @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) > /* Record GP times before starting GP, hence rcu_seq_start(). */ > rcu_seq_start(&rcu_state.gp_seq); > ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); > + rcu_sr_normal_gp_init(); > trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); > rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); > raw_spin_unlock_irq_rcu_node(rnp); > @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) > } > raw_spin_unlock_irq_rcu_node(rnp); > > + // Make synchronize_rcu() users aware of the end of old grace period. > + rcu_sr_normal_gp_cleanup(); > + > // If strict, make all CPUs aware of the end of the old grace period. > if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) > on_each_cpu(rcu_strict_gp_boundary, NULL, 0); > @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) > return true; > } > > +/* > + * Helper function for the synchronize_rcu() API. > + */ > +static void synchronize_rcu_normal(void) > +{ > + struct rcu_synchronize rs; > + > + if (!READ_ONCE(rcu_normal_wake_from_gp)) { > + wait_rcu_gp(call_rcu_hurry); > + return; > + } > + > + init_rcu_head_on_stack(&rs.head); > + init_completion(&rs.completion); > + > + /* > + * This code might be preempted, therefore take a GP > + * snapshot before adding a request. > + */ > + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) > + rs.head.func = (void *) get_state_synchronize_rcu(); > + > + rcu_sr_normal_add_req(&rs); > + > + /* Kick a GP and start waiting. */ > + (void) start_poll_synchronize_rcu(); > + > + /* Now we can wait. */ > + wait_for_completion(&rs.completion); > + destroy_rcu_head_on_stack(&rs.head); > +} > + > /** > * synchronize_rcu - wait until a grace period has elapsed. > * > @@ -3612,7 +3748,7 @@ void synchronize_rcu(void) > if (rcu_gp_is_expedited()) > synchronize_rcu_expedited(); > else > - wait_rcu_gp(call_rcu_hurry); > + synchronize_rcu_normal(); > return; > } > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h > index 014ddf672165..bdc30d972d32 100644 > --- a/kernel/rcu/tree_exp.h > +++ b/kernel/rcu/tree_exp.h > @@ -985,7 +985,7 @@ void synchronize_rcu_expedited(void) > > /* If expedited grace periods are prohibited, fall back to normal. */ > if (rcu_gp_is_normal()) { > - wait_rcu_gp(call_rcu_hurry); > + synchronize_rcu_normal(); > return; > } > > -- > 2.39.2 >
Hello, Kalesh! > > Hi Uladzislau, > > I've tried your patches (v3) on Android with 6.1.43 kernel. > > The test cycles 10 apps (including camera) sequentially for 100 > iterations. > > I've set rcu_normal to override the rcu_expedited in the boot > parameters: > > adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu > > rcupdate.rcu_normal=1 > rcupdate.rcu_expedited=1 > rcu_nocbs=0-7 > > > The configurations are: > > A - echo 0 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp > B - echo 1 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp > > Results: > > = APP LAUNCH TIME = > delta (B-A) ratio(%) > overall_app_launch_time(ms) -11399.00 -6.65 > > > == camera_launch_time > type delta(B-A %) A_count B_count > HOT -7.05 99 99 > COLD -6.33 1 1 > > If i interpret it correctly you also see that this series reduces a launch time by 6/7% on your app set. Is that correct? > === Function Latencies === > > Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit > > nsec : count distribution nsec : count distribution > 0 -> 1 : 0 | | 0 -> 1 : 0 | | > 2 -> 3 : 0 | | 2 -> 3 : 0 | | > 4 -> 7 : 0 | | 4 -> 7 : 0 | | > 8 -> 15 : 0 | | 8 -> 15 : 0 | | > 16 -> 31 : 0 | | 16 -> 31 : 0 | | > 32 -> 63 : 0 | | 32 -> 63 : 0 | | > 64 -> 127 : 0 | | 64 -> 127 : 0 | | > 128 -> 255 : 0 | | 128 -> 255 : 0 | | > 256 -> 511 : 0 | | 256 -> 511 : 0 | | > 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | > 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | > 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | > 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | > 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | > 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | > 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | > 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | > 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | > 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | > 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | > 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | > 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | > 4194304 -> 8388607 : 871 |** | 4194304 -> 8388607 : 1180 |**** | > 8388608 -> 16777215 : 3204 |******** | 8388608 -> 16777215 : 7020 |************************* | > 16777216 -> 33554431 : 15013 |****************************************| 16777216 -> 33554431 : 10952 |****************************************| > Exiting trace of synchronize_rcu_expedited Exiting trace of synchronize_rcu_expedited > > > Tracing synchronize_rcu. Hit Ctrl-C to exit Tracing synchronize_rcu. Hit Ctrl-C to exit > > nsec : count distribution nsec : count distribution > 0 -> 1 : 0 | | 0 -> 1 : 0 | | > 2 -> 3 : 0 | | 2 -> 3 : 0 | | > 4 -> 7 : 0 | | 4 -> 7 : 0 | | > 8 -> 15 : 0 | | 8 -> 15 : 0 | | > 16 -> 31 : 0 | | 16 -> 31 : 0 | | > 32 -> 63 : 0 | | 32 -> 63 : 0 | | > 64 -> 127 : 0 | | 64 -> 127 : 0 | | > 128 -> 255 : 0 | | 128 -> 255 : 0 | | > 256 -> 511 : 0 | | 256 -> 511 : 0 | | > 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | > 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | > 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | > 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | > 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | > 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | > 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | > 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | > 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | > 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | > 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | > 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | > 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | > 4194304 -> 8388607 : 861 |** | 4194304 -> 8388607 : 1136 |**** | > 8388608 -> 16777215 : 3111 |******** | 8388608 -> 16777215 : 6320 |************************ | > 16777216 -> 33554431 : 13901 |****************************************| 16777216 -> 33554431 : 10484 |****************************************| > Exiting trace of synchronize_rcu Exiting trace of synchronize_rcu > Who is B and who is A? > > Interestingly I tried the same experiment without rcu_normal=1 (leaving rcu_expedited=1): > > adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu > rcupdate.rcu_expedited=1 > rcu_nocbs=0-7 > > In this case I also saw the -6 to -7% decrease in the app launch times > but I don't have a good explanation why that would be? (The fucntion > latency histograms in this case didn't show any significant difference). > Do you have any insight why this may happen? > When rcu_expedited=1 is set and rcu_normal=0 is disabled. The synchronize_rcu() call is converted into synchronize_rcu_expidited(): <snip> void synchronize_rcu(void) { unsigned long flags; struct rcu_node *rnp; RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || lock_is_held(&rcu_lock_map) || lock_is_held(&rcu_sched_lock_map), "Illegal synchronize_rcu() in RCU read-side critical section"); if (!rcu_blocking_is_gp()) { if (rcu_gp_is_expedited()) synchronize_rcu_expedited(); else synchronize_rcu_normal(); return; } ... <snip> rcu_gp_is_expidited() is true, so invoke "expedited" version. I see some concerns in preferring an expedited version as a global replacement. First of all it is related to latency sensitive workloads because in order to expedite a grace period it sends out IPIs on all online CPUs to force them to report a quiescent-state asap. I have not investigated yet how it affects such workloads. Therefore, in your case, you also see a performance boost of your app sets. Thank you for looking at it! -- Uladzislau Rezki
On Wed, Jan 10, 2024 at 1:22 AM Uladzislau Rezki <urezki@gmail.com> wrote: > > Hello, Kalesh! > > > > > Hi Uladzislau, > > > > I've tried your patches (v3) on Android with 6.1.43 kernel. > > > > The test cycles 10 apps (including camera) sequentially for 100 > > iterations. > > > > I've set rcu_normal to override the rcu_expedited in the boot > > parameters: > > > > adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu > > > > rcupdate.rcu_normal=1 > > rcupdate.rcu_expedited=1 > > rcu_nocbs=0-7 > > > > > > The configurations are: > > > > A - echo 0 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp > > B - echo 1 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp > > > > Results: > > > > = APP LAUNCH TIME = > > delta (B-A) ratio(%) > > overall_app_launch_time(ms) -11399.00 -6.65 > > > > > > == camera_launch_time > > type delta(B-A %) A_count B_count > > HOT -7.05 99 99 > > COLD -6.33 1 1 > > > > Hi Uladzislau, > If i interpret it correctly you also see that this series reduces > a launch time by 6/7% on your app set. Is that correct? Yes your understanding is correct. > > > === Function Latencies === > > > > Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit > > > > nsec : count distribution nsec : count distribution > > 0 -> 1 : 0 | | 0 -> 1 : 0 | | > > 2 -> 3 : 0 | | 2 -> 3 : 0 | | > > 4 -> 7 : 0 | | 4 -> 7 : 0 | | > > 8 -> 15 : 0 | | 8 -> 15 : 0 | | > > 16 -> 31 : 0 | | 16 -> 31 : 0 | | > > 32 -> 63 : 0 | | 32 -> 63 : 0 | | > > 64 -> 127 : 0 | | 64 -> 127 : 0 | | > > 128 -> 255 : 0 | | 128 -> 255 : 0 | | > > 256 -> 511 : 0 | | 256 -> 511 : 0 | | > > 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | > > 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | > > 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | > > 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | > > 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | > > 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | > > 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | > > 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | > > 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | > > 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | > > 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | > > 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | > > 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | > > 4194304 -> 8388607 : 871 |** | 4194304 -> 8388607 : 1180 |**** | > > 8388608 -> 16777215 : 3204 |******** | 8388608 -> 16777215 : 7020 |************************* | > > 16777216 -> 33554431 : 15013 |****************************************| 16777216 -> 33554431 : 10952 |****************************************| > > Exiting trace of synchronize_rcu_expedited Exiting trace of synchronize_rcu_expedited > > > > > > Tracing synchronize_rcu. Hit Ctrl-C to exit Tracing synchronize_rcu. Hit Ctrl-C to exit > > > > nsec : count distribution nsec : count distribution > > 0 -> 1 : 0 | | 0 -> 1 : 0 | | > > 2 -> 3 : 0 | | 2 -> 3 : 0 | | > > 4 -> 7 : 0 | | 4 -> 7 : 0 | | > > 8 -> 15 : 0 | | 8 -> 15 : 0 | | > > 16 -> 31 : 0 | | 16 -> 31 : 0 | | > > 32 -> 63 : 0 | | 32 -> 63 : 0 | | > > 64 -> 127 : 0 | | 64 -> 127 : 0 | | > > 128 -> 255 : 0 | | 128 -> 255 : 0 | | > > 256 -> 511 : 0 | | 256 -> 511 : 0 | | > > 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | > > 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | > > 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | > > 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | > > 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | > > 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | > > 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | > > 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | > > 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | > > 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | > > 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | > > 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | > > 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | > > 4194304 -> 8388607 : 861 |** | 4194304 -> 8388607 : 1136 |**** | > > 8388608 -> 16777215 : 3111 |******** | 8388608 -> 16777215 : 6320 |************************ | > > 16777216 -> 33554431 : 13901 |****************************************| 16777216 -> 33554431 : 10484 |****************************************| > > Exiting trace of synchronize_rcu Exiting trace of synchronize_rcu > > > Who is B and who is A? Left is A (rcu_normal_wake_from_gp=0) and right is B (rcu_normal_wake_from_gp=1) > > > > > Interestingly I tried the same experiment without rcu_normal=1 (leaving rcu_expedited=1): > > > > adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu > > rcupdate.rcu_expedited=1 > > rcu_nocbs=0-7 > > > > In this case I also saw the -6 to -7% decrease in the app launch times > > but I don't have a good explanation why that would be? (The fucntion > > latency histograms in this case didn't show any significant difference). > > Do you have any insight why this may happen? > > > When rcu_expedited=1 is set and rcu_normal=0 is disabled. The > synchronize_rcu() call is converted into synchronize_rcu_expidited(): > > <snip> > void synchronize_rcu(void) > { > unsigned long flags; > struct rcu_node *rnp; > > RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || > lock_is_held(&rcu_lock_map) || > lock_is_held(&rcu_sched_lock_map), > "Illegal synchronize_rcu() in RCU read-side critical section"); > if (!rcu_blocking_is_gp()) { > if (rcu_gp_is_expedited()) > synchronize_rcu_expedited(); > else > synchronize_rcu_normal(); > return; > } > ... > <snip> > > rcu_gp_is_expidited() is true, so invoke "expedited" version. > > I see some concerns in preferring an expedited version as a global > replacement. First of all it is related to latency sensitive workloads > because in order to expedite a grace period it sends out IPIs on all > online CPUs to force them to report a quiescent-state asap. I have not > investigated yet how it affects such workloads. > > Therefore, in your case, you also see a performance boost of your app sets. IIUC the patch shouldn't affect the case? The only difference in A vs B is rcu_normal_wake_from_gp (both have rcu_expedited=1). Thanks, Kalesh > > Thank you for looking at it! > > -- > Uladzislau Rezki
Hello, Kalesh! > > > > > > > > Hi Uladzislau, > > > > > > I've tried your patches (v3) on Android with 6.1.43 kernel. > > > > > > The test cycles 10 apps (including camera) sequentially for 100 > > > iterations. > > > > > > I've set rcu_normal to override the rcu_expedited in the boot > > > parameters: > > > > > > adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu > > > > > > rcupdate.rcu_normal=1 > > > rcupdate.rcu_expedited=1 > > > rcu_nocbs=0-7 > > > > > > > > > The configurations are: > > > > > > A - echo 0 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp > > > B - echo 1 >/sys/module/rcutree/parameters/rcu_normal_wake_from_gp > > > > > > Results: > > > > > > = APP LAUNCH TIME = > > > delta (B-A) ratio(%) > > > overall_app_launch_time(ms) -11399.00 -6.65 > > > > > > > > > == camera_launch_time > > > type delta(B-A %) A_count B_count > > > HOT -7.05 99 99 > > > COLD -6.33 1 1 > > > > > > > > Hi Uladzislau, > > > If i interpret it correctly you also see that this series reduces > > a launch time by 6/7% on your app set. Is that correct? > > Yes your understanding is correct. > > > > > > === Function Latencies === > > > > > > Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit Tracing synchronize_rcu_expedited. Hit Ctrl-C to exit > > > > > > nsec : count distribution nsec : count distribution > > > 0 -> 1 : 0 | | 0 -> 1 : 0 | | > > > 2 -> 3 : 0 | | 2 -> 3 : 0 | | > > > 4 -> 7 : 0 | | 4 -> 7 : 0 | | > > > 8 -> 15 : 0 | | 8 -> 15 : 0 | | > > > 16 -> 31 : 0 | | 16 -> 31 : 0 | | > > > 32 -> 63 : 0 | | 32 -> 63 : 0 | | > > > 64 -> 127 : 0 | | 64 -> 127 : 0 | | > > > 128 -> 255 : 0 | | 128 -> 255 : 0 | | > > > 256 -> 511 : 0 | | 256 -> 511 : 0 | | > > > 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | > > > 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | > > > 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | > > > 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | > > > 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | > > > 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | > > > 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | > > > 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | > > > 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | > > > 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | > > > 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | > > > 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | > > > 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | > > > 4194304 -> 8388607 : 871 |** | 4194304 -> 8388607 : 1180 |**** | > > > 8388608 -> 16777215 : 3204 |******** | 8388608 -> 16777215 : 7020 |************************* | > > > 16777216 -> 33554431 : 15013 |****************************************| 16777216 -> 33554431 : 10952 |****************************************| > > > Exiting trace of synchronize_rcu_expedited Exiting trace of synchronize_rcu_expedited > > > > > > > > > Tracing synchronize_rcu. Hit Ctrl-C to exit Tracing synchronize_rcu. Hit Ctrl-C to exit > > > > > > nsec : count distribution nsec : count distribution > > > 0 -> 1 : 0 | | 0 -> 1 : 0 | | > > > 2 -> 3 : 0 | | 2 -> 3 : 0 | | > > > 4 -> 7 : 0 | | 4 -> 7 : 0 | | > > > 8 -> 15 : 0 | | 8 -> 15 : 0 | | > > > 16 -> 31 : 0 | | 16 -> 31 : 0 | | > > > 32 -> 63 : 0 | | 32 -> 63 : 0 | | > > > 64 -> 127 : 0 | | 64 -> 127 : 0 | | > > > 128 -> 255 : 0 | | 128 -> 255 : 0 | | > > > 256 -> 511 : 0 | | 256 -> 511 : 0 | | > > > 512 -> 1023 : 0 | | 512 -> 1023 : 0 | | > > > 1024 -> 2047 : 0 | | 1024 -> 2047 : 0 | | > > > 2048 -> 4095 : 0 | | 2048 -> 4095 : 0 | | > > > 4096 -> 8191 : 0 | | 4096 -> 8191 : 0 | | > > > 8192 -> 16383 : 0 | | 8192 -> 16383 : 0 | | > > > 16384 -> 32767 : 0 | | 16384 -> 32767 : 0 | | > > > 32768 -> 65535 : 0 | | 32768 -> 65535 : 0 | | > > > 65536 -> 131071 : 0 | | 65536 -> 131071 : 0 | | > > > 131072 -> 262143 : 0 | | 131072 -> 262143 : 0 | | > > > 262144 -> 524287 : 0 | | 262144 -> 524287 : 0 | | > > > 524288 -> 1048575 : 0 | | 524288 -> 1048575 : 0 | | > > > 1048576 -> 2097151 : 0 | | 1048576 -> 2097151 : 0 | | > > > 2097152 -> 4194303 : 0 | | 2097152 -> 4194303 : 0 | | > > > 4194304 -> 8388607 : 861 |** | 4194304 -> 8388607 : 1136 |**** | > > > 8388608 -> 16777215 : 3111 |******** | 8388608 -> 16777215 : 6320 |************************ | > > > 16777216 -> 33554431 : 13901 |****************************************| 16777216 -> 33554431 : 10484 |****************************************| > > > Exiting trace of synchronize_rcu Exiting trace of synchronize_rcu > > > > > Who is B and who is A? > > Left is A (rcu_normal_wake_from_gp=0) and right is B > (rcu_normal_wake_from_gp=1) > > > > > > > > Interestingly I tried the same experiment without rcu_normal=1 (leaving rcu_expedited=1): > > > > > > adb shell cat /proc/cmdline | tr ' ' '\n' | grep rcu > > > rcupdate.rcu_expedited=1 > > > rcu_nocbs=0-7 > > > > > > In this case I also saw the -6 to -7% decrease in the app launch times > > > but I don't have a good explanation why that would be? (The fucntion > > > latency histograms in this case didn't show any significant difference). > > > Do you have any insight why this may happen? > > > > > When rcu_expedited=1 is set and rcu_normal=0 is disabled. The > > synchronize_rcu() call is converted into synchronize_rcu_expidited(): > > > > <snip> > > void synchronize_rcu(void) > > { > > unsigned long flags; > > struct rcu_node *rnp; > > > > RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || > > lock_is_held(&rcu_lock_map) || > > lock_is_held(&rcu_sched_lock_map), > > "Illegal synchronize_rcu() in RCU read-side critical section"); > > if (!rcu_blocking_is_gp()) { > > if (rcu_gp_is_expedited()) > > synchronize_rcu_expedited(); > > else > > synchronize_rcu_normal(); > > return; > > } > > ... > > <snip> > > > > rcu_gp_is_expidited() is true, so invoke "expedited" version. > > > > I see some concerns in preferring an expedited version as a global > > replacement. First of all it is related to latency sensitive workloads > > because in order to expedite a grace period it sends out IPIs on all > > online CPUs to force them to report a quiescent-state asap. I have not > > investigated yet how it affects such workloads. > > > > Therefore, in your case, you also see a performance boost of your app sets. > > IIUC the patch shouldn't affect the case? The only difference in A vs > B is rcu_normal_wake_from_gp (both have rcu_expedited=1). > Right. This patch does not touch "expedited" version at all. Appreciate for test results and looking at! -- Uladzislau Rezki
Le Thu, Jan 04, 2024 at 05:25:07PM +0100, Uladzislau Rezki (Sony) a écrit : > diff --git a/kernel/rcu/Kconfig.debug b/kernel/rcu/Kconfig.debug > index 9b0b52e1836f..4812c6249185 100644 > --- a/kernel/rcu/Kconfig.debug > +++ b/kernel/rcu/Kconfig.debug > @@ -168,4 +168,16 @@ config RCU_STRICT_GRACE_PERIOD > when looking for certain types of RCU usage bugs, for example, > too-short RCU read-side critical sections. > > +config RCU_SR_NORMAL_DEBUG_GP > + bool "Debug synchronize_rcu() callers for a grace period completion" > + depends on DEBUG_KERNEL && RCU_EXPERT > + default n > + help > + This option enables additional debugging for detecting a grace > + period incompletion for synchronize_rcu() users. If a GP is not > + fully passed for any user, the warning message is emitted. > + > + Say Y here if you want to enable such debugging > + Say N if you are unsure. How about just reuse CONFIG_PROVE_RCU instead? > + > endmenu # "RCU Debugging" > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c > index 499803234176..b756c40e4960 100644 > --- a/kernel/rcu/tree.c > +++ b/kernel/rcu/tree.c > @@ -1422,6 +1422,106 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap) > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > } > > +/* > + * There are three lists for handling synchronize_rcu() users. > + * A first list corresponds to new coming users, second for users > + * which wait for a grace period and third is for which a grace > + * period is passed. > + */ > +static struct sr_normal_state { > + struct llist_head srs_next; /* request a GP users. */ > + struct llist_head srs_wait; /* wait for GP users. */ > + struct llist_head srs_done; /* ready for GP users. */ > + > + /* > + * In order to add a batch of nodes to already > + * existing srs-done-list, a tail of srs-wait-list > + * is maintained. > + */ > + struct llist_node *srs_wait_tail; > +} sr; "sr" is good enough for a function scope variable but not for a file scope one. At least "sr_state" would be better. Or maybe you don't even need to name that struct and make instead: struct { ... ... } sr_normal_state; > + > +/* Disabled by default. */ > +static int rcu_normal_wake_from_gp; > +module_param(rcu_normal_wake_from_gp, int, 0644); > + > +static void rcu_sr_normal_complete(struct llist_node *node) > +{ > + struct rcu_synchronize *rs = container_of( > + (struct rcu_head *) node, struct rcu_synchronize, head); Should there be some union in struct rcu_synchronize between struct rcu_head and struct llist_node? Anyway it's stack allocated, they could even be separate fields. > + unsigned long oldstate = (unsigned long) rs->head.func; Luckily struct callback_head layout allows such magic but if rcu_head and llist_node were separate, reviewers would be less hurt. If stack space really matters, something like the below? struct rcu_synchronize { union { struct rcu_head head; struct { struct llist_node node; unsigned long seq; } } struct completion completion; }; > + > + WARN_ONCE(IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP) && > + !poll_state_synchronize_rcu(oldstate), > + "A full grace period is not passed yet: %lu", > + rcu_seq_diff(get_state_synchronize_rcu(), oldstate)); > + > + /* Finally. */ > + complete(&rs->completion); > +} > + [...] > + > +/* > + * Helper function for rcu_gp_cleanup(). > + */ > +static void rcu_sr_normal_gp_cleanup(void) > +{ > + struct llist_node *head, *tail; > + > + if (llist_empty(&sr.srs_wait)) > + return; > + > + tail = READ_ONCE(sr.srs_wait_tail); Is the READ_ONCE() needed? A part from those boring details: Reviewed-by: Frederic Weisbecker <frederic@kernel.org>
On Thu, Jan 04, 2024 at 05:25:07PM +0100, Uladzislau Rezki (Sony) wrote: > A call to a synchronize_rcu() can be optimized from a latency > point of view. Workloads which depend on this can benefit of it. > > The delay of wakeme_after_rcu() callback, which unblocks a waiter, > depends on several factors: > > - how fast a process of offloading is started. Combination of: > - !CONFIG_RCU_NOCB_CPU/CONFIG_RCU_NOCB_CPU; > - !CONFIG_RCU_LAZY/CONFIG_RCU_LAZY; > - other. > - when started, invoking path is interrupted due to: > - time limit; > - need_resched(); > - if limit is reached. > - where in a nocb list it is located; > - how fast previous callbacks completed; > > Example: > > 1. On our embedded devices i can easily trigger the scenario when > it is a last in the list out of ~3600 callbacks: > > <snip> > <...>-29 [001] d..1. 21950.145313: rcu_batch_start: rcu_preempt CBs=3613 bl=28 > ... > <...>-29 [001] ..... 21950.152578: rcu_invoke_callback: rcu_preempt rhp=00000000b2d6dee8 func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152579: rcu_invoke_callback: rcu_preempt rhp=00000000a446f607 func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152580: rcu_invoke_callback: rcu_preempt rhp=00000000a5cab03b func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152581: rcu_invoke_callback: rcu_preempt rhp=0000000013b7e5ee func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152582: rcu_invoke_callback: rcu_preempt rhp=000000000a8ca6f9 func=__free_vm_area_struct.cfi_jt > <...>-29 [001] ..... 21950.152583: rcu_invoke_callback: rcu_preempt rhp=000000008f162ca8 func=wakeme_after_rcu.cfi_jt > <...>-29 [001] d..1. 21950.152625: rcu_batch_end: rcu_preempt CBs-invoked=3612 idle=.... > <snip> > > 2. We use cpuset/cgroup to classify tasks and assign them into > different cgroups. For example "backgrond" group which binds tasks > only to little CPUs or "foreground" which makes use of all CPUs. > Tasks can be migrated between groups by a request if an acceleration > is needed. > > See below an example how "surfaceflinger" task gets migrated. > Initially it is located in the "system-background" cgroup which > allows to run only on little cores. In order to speed it up it > can be temporary moved into "foreground" cgroup which allows > to use big/all CPUs: > > cgroup_attach_task(): > -> cgroup_migrate_execute() > -> cpuset_can_attach() > -> percpu_down_write() > -> rcu_sync_enter() > -> synchronize_rcu() > -> now move tasks to the new cgroup. > -> cgroup_migrate_finish() > > <snip> > rcuop/1-29 [000] ..... 7030.528570: rcu_invoke_callback: rcu_preempt rhp=00000000461605e0 func=wakeme_after_rcu.cfi_jt > PERFD-SERVER-1855 [000] d..1. 7030.530293: cgroup_attach_task: dst_root=3 dst_id=22 dst_level=1 dst_path=/foreground pid=1900 comm=surfaceflinger > TimerDispatch-2768 [002] d..5. 7030.537542: sched_migrate_task: comm=surfaceflinger pid=1900 prio=98 orig_cpu=0 dest_cpu=4 > <snip> > > "Boosting a task" depends on synchronize_rcu() latency: > > - first trace shows a completion of synchronize_rcu(); > - second shows attaching a task to a new group; > - last shows a final step when migration occurs. > > 3. To address this drawback, maintain a separate track that consists > of synchronize_rcu() callers only. After completion of a grace period > users are deferred to a dedicated worker to process requests. > > 4. This patch reduces the latency of synchronize_rcu() approximately > by ~30-40% on synthetic tests. The real test case, camera launch time, > shows(time is in milliseconds): > > 1-run 542 vs 489 improvement 9% > 2-run 540 vs 466 improvement 13% > 3-run 518 vs 468 improvement 9% > 4-run 531 vs 457 improvement 13% > 5-run 548 vs 475 improvement 13% > 6-run 509 vs 484 improvement 4% > > Synthetic test(no "noise" from other callbacks): > Hardware: x86_64 64 CPUs, 64GB of memory > Linux-6.6 > > - 10K tasks(simultaneous); > - each task does(1000 loops) > synchronize_rcu(); > kfree(p); > > default: CONFIG_RCU_NOCB_CPU: takes 54 seconds to complete all users; > patch: CONFIG_RCU_NOCB_CPU: takes 35 seconds to complete all users. > > Running 60K gives approximately same results on my setup. Please note > it is without any interaction with another type of callbacks, otherwise > it will impact a lot a default case. > > 5. An extra CONFIG_RCU_SR_NORMAL_DEBUG_GP kernel option is added > which enables additional debugging for detecting a grace period > incompletion for synchronize_rcu() users. If a GP is not fully > passed for any user, the warning message is emitted. > > 6. By default it is disabled. To enable this perform one of the > below sequence: > > echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp > or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1" > > Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> Again, nice latency reductions! A few comments and questions below. Thanx, Paul > --- > .../admin-guide/kernel-parameters.txt | 14 ++ > kernel/rcu/Kconfig.debug | 12 ++ > kernel/rcu/tree.c | 138 +++++++++++++++++- > kernel/rcu/tree_exp.h | 2 +- > 4 files changed, 164 insertions(+), 2 deletions(-) > > diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt > index 17a454909ab4..2cca75e4f0c6 100644 > --- a/Documentation/admin-guide/kernel-parameters.txt > +++ b/Documentation/admin-guide/kernel-parameters.txt > @@ -5047,6 +5047,20 @@ > delay, memory pressure or callback list growing too > big. > > + rcutree.rcu_normal_wake_from_gp= [KNL] > + Reduces a latency of synchronize_rcu() call. This approach > + maintains its own track of synchronize_rcu() callers, so it > + does not interact with regular callbacks because it does not > + use a call_rcu[_hurry]() path. Please note, this is for a > + normal grace period. > + > + How to enable it: > + > + echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp > + or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1" > + > + Default is 0. > + > rcuscale.gp_async= [KNL] > Measure performance of asynchronous > grace-period primitives such as call_rcu(). > diff --git a/kernel/rcu/Kconfig.debug b/kernel/rcu/Kconfig.debug > index 9b0b52e1836f..4812c6249185 100644 > --- a/kernel/rcu/Kconfig.debug > +++ b/kernel/rcu/Kconfig.debug > @@ -168,4 +168,16 @@ config RCU_STRICT_GRACE_PERIOD > when looking for certain types of RCU usage bugs, for example, > too-short RCU read-side critical sections. > > +config RCU_SR_NORMAL_DEBUG_GP > + bool "Debug synchronize_rcu() callers for a grace period completion" > + depends on DEBUG_KERNEL && RCU_EXPERT > + default n > + help > + This option enables additional debugging for detecting a grace > + period incompletion for synchronize_rcu() users. If a GP is not > + fully passed for any user, the warning message is emitted. > + > + Say Y here if you want to enable such debugging > + Say N if you are unsure. > + > endmenu # "RCU Debugging" > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c > index 499803234176..b756c40e4960 100644 > --- a/kernel/rcu/tree.c > +++ b/kernel/rcu/tree.c > @@ -1422,6 +1422,106 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap) > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > } > > +/* > + * There are three lists for handling synchronize_rcu() users. > + * A first list corresponds to new coming users, second for users > + * which wait for a grace period and third is for which a grace > + * period is passed. > + */ > +static struct sr_normal_state { > + struct llist_head srs_next; /* request a GP users. */ > + struct llist_head srs_wait; /* wait for GP users. */ > + struct llist_head srs_done; /* ready for GP users. */ > + > + /* > + * In order to add a batch of nodes to already > + * existing srs-done-list, a tail of srs-wait-list > + * is maintained. > + */ > + struct llist_node *srs_wait_tail; > +} sr; Please put this in the rcu_state structure. Having the separate structure is fine (it does group the fields nicely, plus you can take a pointer to it in the functions using this state), but it is good to have the state in one place. Also, please add the data structures in a separate patch. This might save someone a lot of time and effort should someone breaks the kernel in a way that depends on data-structure size. It would be much easier for us if their bisection converged on the commit that adds the data structures instead of the commit that also adds a lot of code. > + > +/* Disabled by default. */ > +static int rcu_normal_wake_from_gp; > +module_param(rcu_normal_wake_from_gp, int, 0644); > + > +static void rcu_sr_normal_complete(struct llist_node *node) > +{ > + struct rcu_synchronize *rs = container_of( > + (struct rcu_head *) node, struct rcu_synchronize, head); > + unsigned long oldstate = (unsigned long) rs->head.func; > + > + WARN_ONCE(IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP) && > + !poll_state_synchronize_rcu(oldstate), > + "A full grace period is not passed yet: %lu", > + rcu_seq_diff(get_state_synchronize_rcu(), oldstate)); Good, the false-positive-prone check is now under debug. Or at least possible, even if not prone. > + /* Finally. */ > + complete(&rs->completion); > +} > + > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) > +{ > + struct llist_node *done, *rcu, *next; > + > + done = llist_del_all(&sr.srs_done); > + if (!done) > + return; > + > + llist_for_each_safe(rcu, next, done) > + rcu_sr_normal_complete(rcu); > +} > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); Why not put this into the sr_normal_state structure? You can use __WORK_INITIALIZER() to initialize it, as is done in a number of other places in the kernel. > +/* > + * Helper function for rcu_gp_cleanup(). > + */ > +static void rcu_sr_normal_gp_cleanup(void) > +{ > + struct llist_node *head, *tail; > + > + if (llist_empty(&sr.srs_wait)) > + return; > + > + tail = READ_ONCE(sr.srs_wait_tail); > + head = __llist_del_all(&sr.srs_wait); > + > + if (head) { > + /* Can be not empty. */ > + llist_add_batch(head, tail, &sr.srs_done); > + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); > + } > +} > + > +/* > + * Helper function for rcu_gp_init(). > + */ > +static void rcu_sr_normal_gp_init(void) > +{ > + struct llist_node *head, *tail; > + > + if (llist_empty(&sr.srs_next)) > + return; > + > + tail = llist_del_all(&sr.srs_next); > + head = llist_reverse_order(tail); Again, reversing the order is going to cause trouble on large systems. Let's please not do that. (I could have sworn that this was not present in the last series...) > + /* > + * A waiting list of GP should be empty on this step, > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), > + * rolls it over. If not, it is a BUG, warn a user. > + */ > + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); > + > + WRITE_ONCE(sr.srs_wait_tail, tail); > + __llist_add_batch(head, tail, &sr.srs_wait); > +} > + > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) > +{ > + llist_add((struct llist_node *) &rs->head, &sr.srs_next); > +} > + > /* > * Initialize a new grace period. Return false if no grace period required. > */ > @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) > /* Record GP times before starting GP, hence rcu_seq_start(). */ > rcu_seq_start(&rcu_state.gp_seq); > ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); > + rcu_sr_normal_gp_init(); > trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); > rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); > raw_spin_unlock_irq_rcu_node(rnp); > @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) > } > raw_spin_unlock_irq_rcu_node(rnp); > > + // Make synchronize_rcu() users aware of the end of old grace period. > + rcu_sr_normal_gp_cleanup(); > + > // If strict, make all CPUs aware of the end of the old grace period. > if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) > on_each_cpu(rcu_strict_gp_boundary, NULL, 0); > @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) > return true; > } > > +/* > + * Helper function for the synchronize_rcu() API. > + */ > +static void synchronize_rcu_normal(void) > +{ > + struct rcu_synchronize rs; > + > + if (!READ_ONCE(rcu_normal_wake_from_gp)) { > + wait_rcu_gp(call_rcu_hurry); > + return; > + } > + > + init_rcu_head_on_stack(&rs.head); > + init_completion(&rs.completion); > + > + /* > + * This code might be preempted, therefore take a GP > + * snapshot before adding a request. > + */ > + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) > + rs.head.func = (void *) get_state_synchronize_rcu(); > + > + rcu_sr_normal_add_req(&rs); > + > + /* Kick a GP and start waiting. */ > + (void) start_poll_synchronize_rcu(); It is unfortunate that the debugging requires an extra timestamp. The ways I can think of to avoid this have problems, though. If this thing was replicated per leaf rcu_node structure, the usual approach would be to protect it with that structure's ->lock. Thoughts? > + /* Now we can wait. */ > + wait_for_completion(&rs.completion); > + destroy_rcu_head_on_stack(&rs.head); > +} > + > /** > * synchronize_rcu - wait until a grace period has elapsed. > * > @@ -3612,7 +3748,7 @@ void synchronize_rcu(void) > if (rcu_gp_is_expedited()) > synchronize_rcu_expedited(); > else > - wait_rcu_gp(call_rcu_hurry); > + synchronize_rcu_normal(); > return; > } > > diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h > index 014ddf672165..bdc30d972d32 100644 > --- a/kernel/rcu/tree_exp.h > +++ b/kernel/rcu/tree_exp.h > @@ -985,7 +985,7 @@ void synchronize_rcu_expedited(void) > > /* If expedited grace periods are prohibited, fall back to normal. */ > if (rcu_gp_is_normal()) { > - wait_rcu_gp(call_rcu_hurry); > + synchronize_rcu_normal(); > return; > } > > -- > 2.39.2 >
> > +/* > > + * There are three lists for handling synchronize_rcu() users. > > + * A first list corresponds to new coming users, second for users > > + * which wait for a grace period and third is for which a grace > > + * period is passed. > > + */ > > +static struct sr_normal_state { > > + struct llist_head srs_next; /* request a GP users. */ > > + struct llist_head srs_wait; /* wait for GP users. */ > > + struct llist_head srs_done; /* ready for GP users. */ > > + > > + /* > > + * In order to add a batch of nodes to already > > + * existing srs-done-list, a tail of srs-wait-list > > + * is maintained. > > + */ > > + struct llist_node *srs_wait_tail; > > +} sr; > > Please put this in the rcu_state structure. Having the separate structure > is fine (it does group the fields nicely, plus you can take a pointer > to it in the functions using this state), but it is good to have the > state in one place. > > Also, please add the data structures in a separate patch. This might > save someone a lot of time and effort should someone breaks the kernel > in a way that depends on data-structure size. It would be much easier > for us if their bisection converged on the commit that adds the data > structures instead of the commit that also adds a lot of code. > I put the data under rcu_state in the patch-3 in this series. But i can create a separate patch for this purpose. Should i split it or not? > > + /* Finally. */ > > + complete(&rs->completion); > > +} > > + > > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) > > +{ > > + struct llist_node *done, *rcu, *next; > > + > > + done = llist_del_all(&sr.srs_done); > > + if (!done) > > + return; > > + > > + llist_for_each_safe(rcu, next, done) > > + rcu_sr_normal_complete(rcu); > > +} > > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); > > Why not put this into the sr_normal_state structure? You can use > __WORK_INITIALIZER() to initialize it, as is done in a number of other > places in the kernel. > It is not a big problem. I can move it under "rcu_state" also! > > +/* > > + * Helper function for rcu_gp_cleanup(). > > + */ > > +static void rcu_sr_normal_gp_cleanup(void) > > +{ > > + struct llist_node *head, *tail; > > + > > + if (llist_empty(&sr.srs_wait)) > > + return; > > + > > + tail = READ_ONCE(sr.srs_wait_tail); > > + head = __llist_del_all(&sr.srs_wait); > > + > > + if (head) { > > + /* Can be not empty. */ > > + llist_add_batch(head, tail, &sr.srs_done); > > + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); > > + } > > +} > > + > > +/* > > + * Helper function for rcu_gp_init(). > > + */ > > +static void rcu_sr_normal_gp_init(void) > > +{ > > + struct llist_node *head, *tail; > > + > > + if (llist_empty(&sr.srs_next)) > > + return; > > + > > + tail = llist_del_all(&sr.srs_next); > > + head = llist_reverse_order(tail); > > Again, reversing the order is going to cause trouble on large systems. > Let's please not do that. (I could have sworn that this was not present > in the last series...) > > > + /* > > + * A waiting list of GP should be empty on this step, > > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), > > + * rolls it over. If not, it is a BUG, warn a user. > > + */ > > + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); > > + > > + WRITE_ONCE(sr.srs_wait_tail, tail); > > + __llist_add_batch(head, tail, &sr.srs_wait); > > +} > > + > > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) > > +{ > > + llist_add((struct llist_node *) &rs->head, &sr.srs_next); > > +} > > + > > /* > > * Initialize a new grace period. Return false if no grace period required. > > */ > > @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) > > /* Record GP times before starting GP, hence rcu_seq_start(). */ > > rcu_seq_start(&rcu_state.gp_seq); > > ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); > > + rcu_sr_normal_gp_init(); > > trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); > > rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); > > raw_spin_unlock_irq_rcu_node(rnp); > > @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) > > } > > raw_spin_unlock_irq_rcu_node(rnp); > > > > + // Make synchronize_rcu() users aware of the end of old grace period. > > + rcu_sr_normal_gp_cleanup(); > > + > > // If strict, make all CPUs aware of the end of the old grace period. > > if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) > > on_each_cpu(rcu_strict_gp_boundary, NULL, 0); > > @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) > > return true; > > } > > > > +/* > > + * Helper function for the synchronize_rcu() API. > > + */ > > +static void synchronize_rcu_normal(void) > > +{ > > + struct rcu_synchronize rs; > > + > > + if (!READ_ONCE(rcu_normal_wake_from_gp)) { > > + wait_rcu_gp(call_rcu_hurry); > > + return; > > + } > > + > > + init_rcu_head_on_stack(&rs.head); > > + init_completion(&rs.completion); > > + > > + /* > > + * This code might be preempted, therefore take a GP > > + * snapshot before adding a request. > > + */ > > + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) > > + rs.head.func = (void *) get_state_synchronize_rcu(); > > + > > + rcu_sr_normal_add_req(&rs); > > + > > + /* Kick a GP and start waiting. */ > > + (void) start_poll_synchronize_rcu(); > > It is unfortunate that the debugging requires an extra timestamp. > The ways I can think of to avoid this have problems, though. If this > thing was replicated per leaf rcu_node structure, the usual approach > would be to protect it with that structure's ->lock. > Hmm.. a per-node approach can be deployed later. As discussed earlier :) Debugging part i do not follow, could you please elaborate a bit? Thanks! -- Uladzislau Rezki
On Sat, Jan 13, 2024 at 12:09:51AM +0100, Frederic Weisbecker wrote: > Le Thu, Jan 04, 2024 at 05:25:07PM +0100, Uladzislau Rezki (Sony) a écrit : > > diff --git a/kernel/rcu/Kconfig.debug b/kernel/rcu/Kconfig.debug > > index 9b0b52e1836f..4812c6249185 100644 > > --- a/kernel/rcu/Kconfig.debug > > +++ b/kernel/rcu/Kconfig.debug > > @@ -168,4 +168,16 @@ config RCU_STRICT_GRACE_PERIOD > > when looking for certain types of RCU usage bugs, for example, > > too-short RCU read-side critical sections. > > > > +config RCU_SR_NORMAL_DEBUG_GP > > + bool "Debug synchronize_rcu() callers for a grace period completion" > > + depends on DEBUG_KERNEL && RCU_EXPERT > > + default n > > + help > > + This option enables additional debugging for detecting a grace > > + period incompletion for synchronize_rcu() users. If a GP is not > > + fully passed for any user, the warning message is emitted. > > + > > + Say Y here if you want to enable such debugging > > + Say N if you are unsure. > > How about just reuse CONFIG_PROVE_RCU instead? > Less extra CONFIG_* configuration we have the better approach is. I do not mind, so we can reuse it. Thanks for this point :) I see in some places indeed it is used as a debugging peace. > > + > > endmenu # "RCU Debugging" > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c > > index 499803234176..b756c40e4960 100644 > > --- a/kernel/rcu/tree.c > > +++ b/kernel/rcu/tree.c > > @@ -1422,6 +1422,106 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap) > > raw_spin_unlock_irqrestore_rcu_node(rnp, flags); > > } > > > > +/* > > + * There are three lists for handling synchronize_rcu() users. > > + * A first list corresponds to new coming users, second for users > > + * which wait for a grace period and third is for which a grace > > + * period is passed. > > + */ > > +static struct sr_normal_state { > > + struct llist_head srs_next; /* request a GP users. */ > > + struct llist_head srs_wait; /* wait for GP users. */ > > + struct llist_head srs_done; /* ready for GP users. */ > > + > > + /* > > + * In order to add a batch of nodes to already > > + * existing srs-done-list, a tail of srs-wait-list > > + * is maintained. > > + */ > > + struct llist_node *srs_wait_tail; > > +} sr; > > "sr" is good enough for a function scope variable but not for a file scope one. > > At least "sr_state" would be better. Or maybe you don't even need to name that > struct and make instead: > > struct { > ... > ... > } sr_normal_state; > It is moved by the following patch in the series under the "rcu_state" struct variable. > > > + > > +/* Disabled by default. */ > > +static int rcu_normal_wake_from_gp; > > +module_param(rcu_normal_wake_from_gp, int, 0644); > > + > > +static void rcu_sr_normal_complete(struct llist_node *node) > > +{ > > + struct rcu_synchronize *rs = container_of( > > + (struct rcu_head *) node, struct rcu_synchronize, head); > > Should there be some union in struct rcu_synchronize between struct rcu_head > and struct llist_node? > > Anyway it's stack allocated, they could even be separate fields. > > > + unsigned long oldstate = (unsigned long) rs->head.func; > > Luckily struct callback_head layout allows such magic but if rcu_head > and llist_node were separate, reviewers would be less hurt. > > If stack space really matters, something like the below? > > struct rcu_synchronize { > union { > struct rcu_head head; > struct { > struct llist_node node; > unsigned long seq; > } > } > struct completion completion; > }; > > We can do that. I am not sure if should be a separate patch or as a big change. I tend to separate it. > > + > > + WARN_ONCE(IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP) && > > + !poll_state_synchronize_rcu(oldstate), > > + "A full grace period is not passed yet: %lu", > > + rcu_seq_diff(get_state_synchronize_rcu(), oldstate)); > > + > > + /* Finally. */ > > + complete(&rs->completion); > > +} > > + > [...] > > + > > +/* > > + * Helper function for rcu_gp_cleanup(). > > + */ > > +static void rcu_sr_normal_gp_cleanup(void) > > +{ > > + struct llist_node *head, *tail; > > + > > + if (llist_empty(&sr.srs_wait)) > > + return; > > + > > + tail = READ_ONCE(sr.srs_wait_tail); > > Is the READ_ONCE() needed? > > A part from those boring details: > > Reviewed-by: Frederic Weisbecker <frederic@kernel.org> > Appreciate for the review. I will fix all the comments. Thanks! -- Uladzislau Rezki
On Wed, Jan 17, 2024 at 01:26:19PM +0100, Uladzislau Rezki wrote: > > > +/* > > > + * There are three lists for handling synchronize_rcu() users. > > > + * A first list corresponds to new coming users, second for users > > > + * which wait for a grace period and third is for which a grace > > > + * period is passed. > > > + */ > > > +static struct sr_normal_state { > > > + struct llist_head srs_next; /* request a GP users. */ > > > + struct llist_head srs_wait; /* wait for GP users. */ > > > + struct llist_head srs_done; /* ready for GP users. */ > > > + > > > + /* > > > + * In order to add a batch of nodes to already > > > + * existing srs-done-list, a tail of srs-wait-list > > > + * is maintained. > > > + */ > > > + struct llist_node *srs_wait_tail; > > > +} sr; > > > > Please put this in the rcu_state structure. Having the separate structure > > is fine (it does group the fields nicely, plus you can take a pointer > > to it in the functions using this state), but it is good to have the > > state in one place. > > > > Also, please add the data structures in a separate patch. This might > > save someone a lot of time and effort should someone breaks the kernel > > in a way that depends on data-structure size. It would be much easier > > for us if their bisection converged on the commit that adds the data > > structures instead of the commit that also adds a lot of code. > > > I put the data under rcu_state in the patch-3 in this series. But i can > create a separate patch for this purpose. Should i split it or not? Bisection is best if the data-structure changes come first, keeping in mind the example where the change in data size triggers some unrelated bug. Better to have that bisection converge on a data-structure only commit than on a more complex commit. So it would be much better if the data started out in rcu_state. > > > + /* Finally. */ > > > + complete(&rs->completion); > > > +} > > > + > > > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) > > > +{ > > > + struct llist_node *done, *rcu, *next; > > > + > > > + done = llist_del_all(&sr.srs_done); > > > + if (!done) > > > + return; > > > + > > > + llist_for_each_safe(rcu, next, done) > > > + rcu_sr_normal_complete(rcu); > > > +} > > > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); > > > > Why not put this into the sr_normal_state structure? You can use > > __WORK_INITIALIZER() to initialize it, as is done in a number of other > > places in the kernel. > > > It is not a big problem. I can move it under "rcu_state" also! Very good, thank you! > > > +/* > > > + * Helper function for rcu_gp_cleanup(). > > > + */ > > > +static void rcu_sr_normal_gp_cleanup(void) > > > +{ > > > + struct llist_node *head, *tail; > > > + > > > + if (llist_empty(&sr.srs_wait)) > > > + return; > > > + > > > + tail = READ_ONCE(sr.srs_wait_tail); > > > + head = __llist_del_all(&sr.srs_wait); > > > + > > > + if (head) { > > > + /* Can be not empty. */ > > > + llist_add_batch(head, tail, &sr.srs_done); > > > + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); > > > + } > > > +} > > > + > > > +/* > > > + * Helper function for rcu_gp_init(). > > > + */ > > > +static void rcu_sr_normal_gp_init(void) > > > +{ > > > + struct llist_node *head, *tail; > > > + > > > + if (llist_empty(&sr.srs_next)) > > > + return; > > > + > > > + tail = llist_del_all(&sr.srs_next); > > > + head = llist_reverse_order(tail); > > > > Again, reversing the order is going to cause trouble on large systems. > > Let's please not do that. (I could have sworn that this was not present > > in the last series...) > > > > > + /* > > > + * A waiting list of GP should be empty on this step, > > > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), > > > + * rolls it over. If not, it is a BUG, warn a user. > > > + */ > > > + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); > > > + > > > + WRITE_ONCE(sr.srs_wait_tail, tail); > > > + __llist_add_batch(head, tail, &sr.srs_wait); > > > +} > > > + > > > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) > > > +{ > > > + llist_add((struct llist_node *) &rs->head, &sr.srs_next); > > > +} > > > + > > > /* > > > * Initialize a new grace period. Return false if no grace period required. > > > */ > > > @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) > > > /* Record GP times before starting GP, hence rcu_seq_start(). */ > > > rcu_seq_start(&rcu_state.gp_seq); > > > ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); > > > + rcu_sr_normal_gp_init(); > > > trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); > > > rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); > > > raw_spin_unlock_irq_rcu_node(rnp); > > > @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) > > > } > > > raw_spin_unlock_irq_rcu_node(rnp); > > > > > > + // Make synchronize_rcu() users aware of the end of old grace period. > > > + rcu_sr_normal_gp_cleanup(); > > > + > > > // If strict, make all CPUs aware of the end of the old grace period. > > > if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) > > > on_each_cpu(rcu_strict_gp_boundary, NULL, 0); > > > @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) > > > return true; > > > } > > > > > > +/* > > > + * Helper function for the synchronize_rcu() API. > > > + */ > > > +static void synchronize_rcu_normal(void) > > > +{ > > > + struct rcu_synchronize rs; > > > + > > > + if (!READ_ONCE(rcu_normal_wake_from_gp)) { > > > + wait_rcu_gp(call_rcu_hurry); > > > + return; > > > + } > > > + > > > + init_rcu_head_on_stack(&rs.head); > > > + init_completion(&rs.completion); > > > + > > > + /* > > > + * This code might be preempted, therefore take a GP > > > + * snapshot before adding a request. > > > + */ > > > + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) > > > + rs.head.func = (void *) get_state_synchronize_rcu(); > > > + > > > + rcu_sr_normal_add_req(&rs); > > > + > > > + /* Kick a GP and start waiting. */ > > > + (void) start_poll_synchronize_rcu(); > > > > It is unfortunate that the debugging requires an extra timestamp. > > The ways I can think of to avoid this have problems, though. If this > > thing was replicated per leaf rcu_node structure, the usual approach > > would be to protect it with that structure's ->lock. > > > Hmm.. a per-node approach can be deployed later. As discussed earlier :) Agreed! > Debugging part i do not follow, could you please elaborate a bit? Let's not worry about this unless and until we need per-rcu_node lists of tasks waiting on grace periods. At that point, we will know more and things will be more clear. Thanx, Paul
On Fri, Jan 19, 2024 at 07:24:28AM -0800, Paul E. McKenney wrote: > On Wed, Jan 17, 2024 at 01:26:19PM +0100, Uladzislau Rezki wrote: > > > > +/* > > > > + * There are three lists for handling synchronize_rcu() users. > > > > + * A first list corresponds to new coming users, second for users > > > > + * which wait for a grace period and third is for which a grace > > > > + * period is passed. > > > > + */ > > > > +static struct sr_normal_state { > > > > + struct llist_head srs_next; /* request a GP users. */ > > > > + struct llist_head srs_wait; /* wait for GP users. */ > > > > + struct llist_head srs_done; /* ready for GP users. */ > > > > + > > > > + /* > > > > + * In order to add a batch of nodes to already > > > > + * existing srs-done-list, a tail of srs-wait-list > > > > + * is maintained. > > > > + */ > > > > + struct llist_node *srs_wait_tail; > > > > +} sr; > > > > > > Please put this in the rcu_state structure. Having the separate structure > > > is fine (it does group the fields nicely, plus you can take a pointer > > > to it in the functions using this state), but it is good to have the > > > state in one place. > > > > > > Also, please add the data structures in a separate patch. This might > > > save someone a lot of time and effort should someone breaks the kernel > > > in a way that depends on data-structure size. It would be much easier > > > for us if their bisection converged on the commit that adds the data > > > structures instead of the commit that also adds a lot of code. > > > > > I put the data under rcu_state in the patch-3 in this series. But i can > > create a separate patch for this purpose. Should i split it or not? > > Bisection is best if the data-structure changes come first, keeping in > mind the example where the change in data size triggers some unrelated > bug. Better to have that bisection converge on a data-structure only > commit than on a more complex commit. > > So it would be much better if the data started out in rcu_state. > OK. Then i will also combine two patches into one: rcu: Improve handling of synchronize_rcu() users rcu: Reduce synchronize_rcu() latency to reduce the mess. > > > > + /* Finally. */ > > > > + complete(&rs->completion); > > > > +} > > > > + > > > > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) > > > > +{ > > > > + struct llist_node *done, *rcu, *next; > > > > + > > > > + done = llist_del_all(&sr.srs_done); > > > > + if (!done) > > > > + return; > > > > + > > > > + llist_for_each_safe(rcu, next, done) > > > > + rcu_sr_normal_complete(rcu); > > > > +} > > > > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); > > > > > > Why not put this into the sr_normal_state structure? You can use > > > __WORK_INITIALIZER() to initialize it, as is done in a number of other > > > places in the kernel. > > > > > It is not a big problem. I can move it under "rcu_state" also! > > Very good, thank you! > > > > > +/* > > > > + * Helper function for rcu_gp_cleanup(). > > > > + */ > > > > +static void rcu_sr_normal_gp_cleanup(void) > > > > +{ > > > > + struct llist_node *head, *tail; > > > > + > > > > + if (llist_empty(&sr.srs_wait)) > > > > + return; > > > > + > > > > + tail = READ_ONCE(sr.srs_wait_tail); > > > > + head = __llist_del_all(&sr.srs_wait); > > > > + > > > > + if (head) { > > > > + /* Can be not empty. */ > > > > + llist_add_batch(head, tail, &sr.srs_done); > > > > + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); > > > > + } > > > > +} > > > > + > > > > +/* > > > > + * Helper function for rcu_gp_init(). > > > > + */ > > > > +static void rcu_sr_normal_gp_init(void) > > > > +{ > > > > + struct llist_node *head, *tail; > > > > + > > > > + if (llist_empty(&sr.srs_next)) > > > > + return; > > > > + > > > > + tail = llist_del_all(&sr.srs_next); > > > > + head = llist_reverse_order(tail); > > > > > > Again, reversing the order is going to cause trouble on large systems. > > > Let's please not do that. (I could have sworn that this was not present > > > in the last series...) > > > > > > > + /* > > > > + * A waiting list of GP should be empty on this step, > > > > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), > > > > + * rolls it over. If not, it is a BUG, warn a user. > > > > + */ > > > > + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); > > > > + > > > > + WRITE_ONCE(sr.srs_wait_tail, tail); > > > > + __llist_add_batch(head, tail, &sr.srs_wait); > > > > +} > > > > + > > > > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) > > > > +{ > > > > + llist_add((struct llist_node *) &rs->head, &sr.srs_next); > > > > +} > > > > + > > > > /* > > > > * Initialize a new grace period. Return false if no grace period required. > > > > */ > > > > @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) > > > > /* Record GP times before starting GP, hence rcu_seq_start(). */ > > > > rcu_seq_start(&rcu_state.gp_seq); > > > > ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); > > > > + rcu_sr_normal_gp_init(); > > > > trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); > > > > rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); > > > > raw_spin_unlock_irq_rcu_node(rnp); > > > > @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) > > > > } > > > > raw_spin_unlock_irq_rcu_node(rnp); > > > > > > > > + // Make synchronize_rcu() users aware of the end of old grace period. > > > > + rcu_sr_normal_gp_cleanup(); > > > > + > > > > // If strict, make all CPUs aware of the end of the old grace period. > > > > if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) > > > > on_each_cpu(rcu_strict_gp_boundary, NULL, 0); > > > > @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) > > > > return true; > > > > } > > > > > > > > +/* > > > > + * Helper function for the synchronize_rcu() API. > > > > + */ > > > > +static void synchronize_rcu_normal(void) > > > > +{ > > > > + struct rcu_synchronize rs; > > > > + > > > > + if (!READ_ONCE(rcu_normal_wake_from_gp)) { > > > > + wait_rcu_gp(call_rcu_hurry); > > > > + return; > > > > + } > > > > + > > > > + init_rcu_head_on_stack(&rs.head); > > > > + init_completion(&rs.completion); > > > > + > > > > + /* > > > > + * This code might be preempted, therefore take a GP > > > > + * snapshot before adding a request. > > > > + */ > > > > + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) > > > > + rs.head.func = (void *) get_state_synchronize_rcu(); > > > > + > > > > + rcu_sr_normal_add_req(&rs); > > > > + > > > > + /* Kick a GP and start waiting. */ > > > > + (void) start_poll_synchronize_rcu(); > > > > > > It is unfortunate that the debugging requires an extra timestamp. > > > The ways I can think of to avoid this have problems, though. If this > > > thing was replicated per leaf rcu_node structure, the usual approach > > > would be to protect it with that structure's ->lock. > > > > > Hmm.. a per-node approach can be deployed later. As discussed earlier :) > > Agreed! > > > Debugging part i do not follow, could you please elaborate a bit? > > Let's not worry about this unless and until we need per-rcu_node lists > of tasks waiting on grace periods. At that point, we will know more > and things will be more clear. > Good and thank you :) -- Uladzislau Rezki
On Mon, Jan 22, 2024 at 06:35:44PM +0100, Uladzislau Rezki wrote: > On Fri, Jan 19, 2024 at 07:24:28AM -0800, Paul E. McKenney wrote: > > On Wed, Jan 17, 2024 at 01:26:19PM +0100, Uladzislau Rezki wrote: > > > > > +/* > > > > > + * There are three lists for handling synchronize_rcu() users. > > > > > + * A first list corresponds to new coming users, second for users > > > > > + * which wait for a grace period and third is for which a grace > > > > > + * period is passed. > > > > > + */ > > > > > +static struct sr_normal_state { > > > > > + struct llist_head srs_next; /* request a GP users. */ > > > > > + struct llist_head srs_wait; /* wait for GP users. */ > > > > > + struct llist_head srs_done; /* ready for GP users. */ > > > > > + > > > > > + /* > > > > > + * In order to add a batch of nodes to already > > > > > + * existing srs-done-list, a tail of srs-wait-list > > > > > + * is maintained. > > > > > + */ > > > > > + struct llist_node *srs_wait_tail; > > > > > +} sr; > > > > > > > > Please put this in the rcu_state structure. Having the separate structure > > > > is fine (it does group the fields nicely, plus you can take a pointer > > > > to it in the functions using this state), but it is good to have the > > > > state in one place. > > > > > > > > Also, please add the data structures in a separate patch. This might > > > > save someone a lot of time and effort should someone breaks the kernel > > > > in a way that depends on data-structure size. It would be much easier > > > > for us if their bisection converged on the commit that adds the data > > > > structures instead of the commit that also adds a lot of code. > > > > > > > I put the data under rcu_state in the patch-3 in this series. But i can > > > create a separate patch for this purpose. Should i split it or not? > > > > Bisection is best if the data-structure changes come first, keeping in > > mind the example where the change in data size triggers some unrelated > > bug. Better to have that bisection converge on a data-structure only > > commit than on a more complex commit. > > > > So it would be much better if the data started out in rcu_state. > > > OK. Then i will also combine two patches into one: > > rcu: Improve handling of synchronize_rcu() users > rcu: Reduce synchronize_rcu() latency > > to reduce the mess. That sounds like an excellent next step, thank you! Thanx, Paul > > > > > + /* Finally. */ > > > > > + complete(&rs->completion); > > > > > +} > > > > > + > > > > > +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) > > > > > +{ > > > > > + struct llist_node *done, *rcu, *next; > > > > > + > > > > > + done = llist_del_all(&sr.srs_done); > > > > > + if (!done) > > > > > + return; > > > > > + > > > > > + llist_for_each_safe(rcu, next, done) > > > > > + rcu_sr_normal_complete(rcu); > > > > > +} > > > > > +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); > > > > > > > > Why not put this into the sr_normal_state structure? You can use > > > > __WORK_INITIALIZER() to initialize it, as is done in a number of other > > > > places in the kernel. > > > > > > > It is not a big problem. I can move it under "rcu_state" also! > > > > Very good, thank you! > > > > > > > +/* > > > > > + * Helper function for rcu_gp_cleanup(). > > > > > + */ > > > > > +static void rcu_sr_normal_gp_cleanup(void) > > > > > +{ > > > > > + struct llist_node *head, *tail; > > > > > + > > > > > + if (llist_empty(&sr.srs_wait)) > > > > > + return; > > > > > + > > > > > + tail = READ_ONCE(sr.srs_wait_tail); > > > > > + head = __llist_del_all(&sr.srs_wait); > > > > > + > > > > > + if (head) { > > > > > + /* Can be not empty. */ > > > > > + llist_add_batch(head, tail, &sr.srs_done); > > > > > + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); > > > > > + } > > > > > +} > > > > > + > > > > > +/* > > > > > + * Helper function for rcu_gp_init(). > > > > > + */ > > > > > +static void rcu_sr_normal_gp_init(void) > > > > > +{ > > > > > + struct llist_node *head, *tail; > > > > > + > > > > > + if (llist_empty(&sr.srs_next)) > > > > > + return; > > > > > + > > > > > + tail = llist_del_all(&sr.srs_next); > > > > > + head = llist_reverse_order(tail); > > > > > > > > Again, reversing the order is going to cause trouble on large systems. > > > > Let's please not do that. (I could have sworn that this was not present > > > > in the last series...) > > > > > > > > > + /* > > > > > + * A waiting list of GP should be empty on this step, > > > > > + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), > > > > > + * rolls it over. If not, it is a BUG, warn a user. > > > > > + */ > > > > > + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); > > > > > + > > > > > + WRITE_ONCE(sr.srs_wait_tail, tail); > > > > > + __llist_add_batch(head, tail, &sr.srs_wait); > > > > > +} > > > > > + > > > > > +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) > > > > > +{ > > > > > + llist_add((struct llist_node *) &rs->head, &sr.srs_next); > > > > > +} > > > > > + > > > > > /* > > > > > * Initialize a new grace period. Return false if no grace period required. > > > > > */ > > > > > @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) > > > > > /* Record GP times before starting GP, hence rcu_seq_start(). */ > > > > > rcu_seq_start(&rcu_state.gp_seq); > > > > > ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); > > > > > + rcu_sr_normal_gp_init(); > > > > > trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); > > > > > rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); > > > > > raw_spin_unlock_irq_rcu_node(rnp); > > > > > @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) > > > > > } > > > > > raw_spin_unlock_irq_rcu_node(rnp); > > > > > > > > > > + // Make synchronize_rcu() users aware of the end of old grace period. > > > > > + rcu_sr_normal_gp_cleanup(); > > > > > + > > > > > // If strict, make all CPUs aware of the end of the old grace period. > > > > > if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) > > > > > on_each_cpu(rcu_strict_gp_boundary, NULL, 0); > > > > > @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) > > > > > return true; > > > > > } > > > > > > > > > > +/* > > > > > + * Helper function for the synchronize_rcu() API. > > > > > + */ > > > > > +static void synchronize_rcu_normal(void) > > > > > +{ > > > > > + struct rcu_synchronize rs; > > > > > + > > > > > + if (!READ_ONCE(rcu_normal_wake_from_gp)) { > > > > > + wait_rcu_gp(call_rcu_hurry); > > > > > + return; > > > > > + } > > > > > + > > > > > + init_rcu_head_on_stack(&rs.head); > > > > > + init_completion(&rs.completion); > > > > > + > > > > > + /* > > > > > + * This code might be preempted, therefore take a GP > > > > > + * snapshot before adding a request. > > > > > + */ > > > > > + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) > > > > > + rs.head.func = (void *) get_state_synchronize_rcu(); > > > > > + > > > > > + rcu_sr_normal_add_req(&rs); > > > > > + > > > > > + /* Kick a GP and start waiting. */ > > > > > + (void) start_poll_synchronize_rcu(); > > > > > > > > It is unfortunate that the debugging requires an extra timestamp. > > > > The ways I can think of to avoid this have problems, though. If this > > > > thing was replicated per leaf rcu_node structure, the usual approach > > > > would be to protect it with that structure's ->lock. > > > > > > > Hmm.. a per-node approach can be deployed later. As discussed earlier :) > > > > Agreed! > > > > > Debugging part i do not follow, could you please elaborate a bit? > > > > Let's not worry about this unless and until we need per-rcu_node lists > > of tasks waiting on grace periods. At that point, we will know more > > and things will be more clear. > > > Good and thank you :) > > -- > Uladzislau Rezki
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt index 17a454909ab4..2cca75e4f0c6 100644 --- a/Documentation/admin-guide/kernel-parameters.txt +++ b/Documentation/admin-guide/kernel-parameters.txt @@ -5047,6 +5047,20 @@ delay, memory pressure or callback list growing too big. + rcutree.rcu_normal_wake_from_gp= [KNL] + Reduces a latency of synchronize_rcu() call. This approach + maintains its own track of synchronize_rcu() callers, so it + does not interact with regular callbacks because it does not + use a call_rcu[_hurry]() path. Please note, this is for a + normal grace period. + + How to enable it: + + echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp + or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1" + + Default is 0. + rcuscale.gp_async= [KNL] Measure performance of asynchronous grace-period primitives such as call_rcu(). diff --git a/kernel/rcu/Kconfig.debug b/kernel/rcu/Kconfig.debug index 9b0b52e1836f..4812c6249185 100644 --- a/kernel/rcu/Kconfig.debug +++ b/kernel/rcu/Kconfig.debug @@ -168,4 +168,16 @@ config RCU_STRICT_GRACE_PERIOD when looking for certain types of RCU usage bugs, for example, too-short RCU read-side critical sections. +config RCU_SR_NORMAL_DEBUG_GP + bool "Debug synchronize_rcu() callers for a grace period completion" + depends on DEBUG_KERNEL && RCU_EXPERT + default n + help + This option enables additional debugging for detecting a grace + period incompletion for synchronize_rcu() users. If a GP is not + fully passed for any user, the warning message is emitted. + + Say Y here if you want to enable such debugging + Say N if you are unsure. + endmenu # "RCU Debugging" diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c index 499803234176..b756c40e4960 100644 --- a/kernel/rcu/tree.c +++ b/kernel/rcu/tree.c @@ -1422,6 +1422,106 @@ static void rcu_poll_gp_seq_end_unlocked(unsigned long *snap) raw_spin_unlock_irqrestore_rcu_node(rnp, flags); } +/* + * There are three lists for handling synchronize_rcu() users. + * A first list corresponds to new coming users, second for users + * which wait for a grace period and third is for which a grace + * period is passed. + */ +static struct sr_normal_state { + struct llist_head srs_next; /* request a GP users. */ + struct llist_head srs_wait; /* wait for GP users. */ + struct llist_head srs_done; /* ready for GP users. */ + + /* + * In order to add a batch of nodes to already + * existing srs-done-list, a tail of srs-wait-list + * is maintained. + */ + struct llist_node *srs_wait_tail; +} sr; + +/* Disabled by default. */ +static int rcu_normal_wake_from_gp; +module_param(rcu_normal_wake_from_gp, int, 0644); + +static void rcu_sr_normal_complete(struct llist_node *node) +{ + struct rcu_synchronize *rs = container_of( + (struct rcu_head *) node, struct rcu_synchronize, head); + unsigned long oldstate = (unsigned long) rs->head.func; + + WARN_ONCE(IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP) && + !poll_state_synchronize_rcu(oldstate), + "A full grace period is not passed yet: %lu", + rcu_seq_diff(get_state_synchronize_rcu(), oldstate)); + + /* Finally. */ + complete(&rs->completion); +} + +static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work) +{ + struct llist_node *done, *rcu, *next; + + done = llist_del_all(&sr.srs_done); + if (!done) + return; + + llist_for_each_safe(rcu, next, done) + rcu_sr_normal_complete(rcu); +} +static DECLARE_WORK(sr_normal_gp_cleanup, rcu_sr_normal_gp_cleanup_work); + +/* + * Helper function for rcu_gp_cleanup(). + */ +static void rcu_sr_normal_gp_cleanup(void) +{ + struct llist_node *head, *tail; + + if (llist_empty(&sr.srs_wait)) + return; + + tail = READ_ONCE(sr.srs_wait_tail); + head = __llist_del_all(&sr.srs_wait); + + if (head) { + /* Can be not empty. */ + llist_add_batch(head, tail, &sr.srs_done); + queue_work(system_highpri_wq, &sr_normal_gp_cleanup); + } +} + +/* + * Helper function for rcu_gp_init(). + */ +static void rcu_sr_normal_gp_init(void) +{ + struct llist_node *head, *tail; + + if (llist_empty(&sr.srs_next)) + return; + + tail = llist_del_all(&sr.srs_next); + head = llist_reverse_order(tail); + + /* + * A waiting list of GP should be empty on this step, + * since a GP-kthread, rcu_gp_init() -> gp_cleanup(), + * rolls it over. If not, it is a BUG, warn a user. + */ + WARN_ON_ONCE(!llist_empty(&sr.srs_wait)); + + WRITE_ONCE(sr.srs_wait_tail, tail); + __llist_add_batch(head, tail, &sr.srs_wait); +} + +static void rcu_sr_normal_add_req(struct rcu_synchronize *rs) +{ + llist_add((struct llist_node *) &rs->head, &sr.srs_next); +} + /* * Initialize a new grace period. Return false if no grace period required. */ @@ -1456,6 +1556,7 @@ static noinline_for_stack bool rcu_gp_init(void) /* Record GP times before starting GP, hence rcu_seq_start(). */ rcu_seq_start(&rcu_state.gp_seq); ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq); + rcu_sr_normal_gp_init(); trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start")); rcu_poll_gp_seq_start(&rcu_state.gp_seq_polled_snap); raw_spin_unlock_irq_rcu_node(rnp); @@ -1825,6 +1926,9 @@ static noinline void rcu_gp_cleanup(void) } raw_spin_unlock_irq_rcu_node(rnp); + // Make synchronize_rcu() users aware of the end of old grace period. + rcu_sr_normal_gp_cleanup(); + // If strict, make all CPUs aware of the end of the old grace period. if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) on_each_cpu(rcu_strict_gp_boundary, NULL, 0); @@ -3561,6 +3665,38 @@ static int rcu_blocking_is_gp(void) return true; } +/* + * Helper function for the synchronize_rcu() API. + */ +static void synchronize_rcu_normal(void) +{ + struct rcu_synchronize rs; + + if (!READ_ONCE(rcu_normal_wake_from_gp)) { + wait_rcu_gp(call_rcu_hurry); + return; + } + + init_rcu_head_on_stack(&rs.head); + init_completion(&rs.completion); + + /* + * This code might be preempted, therefore take a GP + * snapshot before adding a request. + */ + if (IS_ENABLED(CONFIG_RCU_SR_NORMAL_DEBUG_GP)) + rs.head.func = (void *) get_state_synchronize_rcu(); + + rcu_sr_normal_add_req(&rs); + + /* Kick a GP and start waiting. */ + (void) start_poll_synchronize_rcu(); + + /* Now we can wait. */ + wait_for_completion(&rs.completion); + destroy_rcu_head_on_stack(&rs.head); +} + /** * synchronize_rcu - wait until a grace period has elapsed. * @@ -3612,7 +3748,7 @@ void synchronize_rcu(void) if (rcu_gp_is_expedited()) synchronize_rcu_expedited(); else - wait_rcu_gp(call_rcu_hurry); + synchronize_rcu_normal(); return; } diff --git a/kernel/rcu/tree_exp.h b/kernel/rcu/tree_exp.h index 014ddf672165..bdc30d972d32 100644 --- a/kernel/rcu/tree_exp.h +++ b/kernel/rcu/tree_exp.h @@ -985,7 +985,7 @@ void synchronize_rcu_expedited(void) /* If expedited grace periods are prohibited, fall back to normal. */ if (rcu_gp_is_normal()) { - wait_rcu_gp(call_rcu_hurry); + synchronize_rcu_normal(); return; }
A call to a synchronize_rcu() can be optimized from a latency point of view. Workloads which depend on this can benefit of it. The delay of wakeme_after_rcu() callback, which unblocks a waiter, depends on several factors: - how fast a process of offloading is started. Combination of: - !CONFIG_RCU_NOCB_CPU/CONFIG_RCU_NOCB_CPU; - !CONFIG_RCU_LAZY/CONFIG_RCU_LAZY; - other. - when started, invoking path is interrupted due to: - time limit; - need_resched(); - if limit is reached. - where in a nocb list it is located; - how fast previous callbacks completed; Example: 1. On our embedded devices i can easily trigger the scenario when it is a last in the list out of ~3600 callbacks: <snip> <...>-29 [001] d..1. 21950.145313: rcu_batch_start: rcu_preempt CBs=3613 bl=28 ... <...>-29 [001] ..... 21950.152578: rcu_invoke_callback: rcu_preempt rhp=00000000b2d6dee8 func=__free_vm_area_struct.cfi_jt <...>-29 [001] ..... 21950.152579: rcu_invoke_callback: rcu_preempt rhp=00000000a446f607 func=__free_vm_area_struct.cfi_jt <...>-29 [001] ..... 21950.152580: rcu_invoke_callback: rcu_preempt rhp=00000000a5cab03b func=__free_vm_area_struct.cfi_jt <...>-29 [001] ..... 21950.152581: rcu_invoke_callback: rcu_preempt rhp=0000000013b7e5ee func=__free_vm_area_struct.cfi_jt <...>-29 [001] ..... 21950.152582: rcu_invoke_callback: rcu_preempt rhp=000000000a8ca6f9 func=__free_vm_area_struct.cfi_jt <...>-29 [001] ..... 21950.152583: rcu_invoke_callback: rcu_preempt rhp=000000008f162ca8 func=wakeme_after_rcu.cfi_jt <...>-29 [001] d..1. 21950.152625: rcu_batch_end: rcu_preempt CBs-invoked=3612 idle=.... <snip> 2. We use cpuset/cgroup to classify tasks and assign them into different cgroups. For example "backgrond" group which binds tasks only to little CPUs or "foreground" which makes use of all CPUs. Tasks can be migrated between groups by a request if an acceleration is needed. See below an example how "surfaceflinger" task gets migrated. Initially it is located in the "system-background" cgroup which allows to run only on little cores. In order to speed it up it can be temporary moved into "foreground" cgroup which allows to use big/all CPUs: cgroup_attach_task(): -> cgroup_migrate_execute() -> cpuset_can_attach() -> percpu_down_write() -> rcu_sync_enter() -> synchronize_rcu() -> now move tasks to the new cgroup. -> cgroup_migrate_finish() <snip> rcuop/1-29 [000] ..... 7030.528570: rcu_invoke_callback: rcu_preempt rhp=00000000461605e0 func=wakeme_after_rcu.cfi_jt PERFD-SERVER-1855 [000] d..1. 7030.530293: cgroup_attach_task: dst_root=3 dst_id=22 dst_level=1 dst_path=/foreground pid=1900 comm=surfaceflinger TimerDispatch-2768 [002] d..5. 7030.537542: sched_migrate_task: comm=surfaceflinger pid=1900 prio=98 orig_cpu=0 dest_cpu=4 <snip> "Boosting a task" depends on synchronize_rcu() latency: - first trace shows a completion of synchronize_rcu(); - second shows attaching a task to a new group; - last shows a final step when migration occurs. 3. To address this drawback, maintain a separate track that consists of synchronize_rcu() callers only. After completion of a grace period users are deferred to a dedicated worker to process requests. 4. This patch reduces the latency of synchronize_rcu() approximately by ~30-40% on synthetic tests. The real test case, camera launch time, shows(time is in milliseconds): 1-run 542 vs 489 improvement 9% 2-run 540 vs 466 improvement 13% 3-run 518 vs 468 improvement 9% 4-run 531 vs 457 improvement 13% 5-run 548 vs 475 improvement 13% 6-run 509 vs 484 improvement 4% Synthetic test(no "noise" from other callbacks): Hardware: x86_64 64 CPUs, 64GB of memory Linux-6.6 - 10K tasks(simultaneous); - each task does(1000 loops) synchronize_rcu(); kfree(p); default: CONFIG_RCU_NOCB_CPU: takes 54 seconds to complete all users; patch: CONFIG_RCU_NOCB_CPU: takes 35 seconds to complete all users. Running 60K gives approximately same results on my setup. Please note it is without any interaction with another type of callbacks, otherwise it will impact a lot a default case. 5. An extra CONFIG_RCU_SR_NORMAL_DEBUG_GP kernel option is added which enables additional debugging for detecting a grace period incompletion for synchronize_rcu() users. If a GP is not fully passed for any user, the warning message is emitted. 6. By default it is disabled. To enable this perform one of the below sequence: echo 1 > /sys/module/rcutree/parameters/rcu_normal_wake_from_gp or pass a boot parameter "rcutree.rcu_normal_wake_from_gp=1" Signed-off-by: Uladzislau Rezki (Sony) <urezki@gmail.com> --- .../admin-guide/kernel-parameters.txt | 14 ++ kernel/rcu/Kconfig.debug | 12 ++ kernel/rcu/tree.c | 138 +++++++++++++++++- kernel/rcu/tree_exp.h | 2 +- 4 files changed, 164 insertions(+), 2 deletions(-)