Re: [PATCH v3] srcu: Reduce blocking agressiveness of expedited grace periods further
From: Neeraj Upadhyay
Date: Fri Jul 01 2022 - 01:06:35 EST
On 7/1/2022 10:19 AM, Paul E. McKenney wrote:
On Fri, Jul 01, 2022 at 08:45:45AM +0530, Neeraj Upadhyay wrote:
Commit 640a7d37c3f4 ("srcu: Block less aggressively for expedited
grace periods") highlights a problem where aggressively blocking
SRCU expedited grace periods, as was introduced in commit
282d8998e997 ("srcu: Prevent expedited GPs and blocking readers
from consuming CPU"), introduces ~2 minutes delay to the overall
~3.5 minutes boot time, when starting VMs with "-bios QEMU_EFI.fd"
cmdline on qemu, which results in very high rate of memslots
add/remove, which causes > ~6000 synchronize_srcu() calls for
kvm->srcu SRCU instance.
Below table captures the experiments done by Zhangfei Gao and Shameer
to measure the boottime impact with various values of non-sleeping
per phase counts, with HZ_250 and preemption enabled:
+──────────────────────────+────────────────+
| SRCU_MAX_NODELAY_PHASE | Boot time (s) |
+──────────────────────────+────────────────+
| 100 | 30.053 |
| 150 | 25.151 |
| 200 | 20.704 |
| 250 | 15.748 |
| 500 | 11.401 |
| 1000 | 11.443 |
| 10000 | 11.258 |
| 1000000 | 11.154 |
+──────────────────────────+────────────────+
Analysis on the experiment results showed improved boot time
with non blocking delays close to one jiffy duration. This
was also seen when number of per-phase iterations were scaled
to one jiffy.
So, this change scales per-grace-period phase number of non-sleeping
polls, such that, non-sleeping polls are done for one jiffy. In addition
to this, srcu_get_delay() call in srcu_gp_end(), which is used to calculate
the delay used for scheduling callbacks, is replaced with the check for
expedited grace period. This is done, to schedule cbs for completed expedited
grace periods immediately, which results in improved boot time seen in
experiments. Testing done by Marc and Zhangfei confirms that this change recovers
most of the performance degradation in boottime; for CONFIG_HZ_250 configuration,
boottime improves from 3m50s to 41s on Marc's setup; and from 2m40s to ~9.7s
on Zhangfei's setup.
In addition to the changes to default per phase delays, this change
adds 3 new kernel parameters - srcutree.srcu_max_nodelay,
srcutree.srcu_max_nodelay_phase, srcutree.srcu_retry_check_delay.
This allows users to configure the srcu grace period scanning delays,
depending on their system configuration requirements.
Signed-off-by: Neeraj Upadhyay <quic_neeraju@xxxxxxxxxxx>
Tested-by: Marc Zyngier <maz@xxxxxxxxxx>
Tested-by: Zhangfei Gao <zhangfei.gao@xxxxxxxxxx>
OK, I am going to try pulling this one into the -rcu "dev" branch,
and thank you all!!!
Please see below for a wordsmithed/tagged version of this commit, and
please let me know if I messed something up.
Thanks! looks good to me; one minor nit below.
Thanx, Paul
------------------------------------------------------------------------
commit 22f57fad06c5fd35cf4b80c652ebbb4dc626fdaa
Author: Neeraj Upadhyay <quic_neeraju@xxxxxxxxxxx>
Date: Fri Jul 1 08:45:45 2022 +0530
srcu: Make expedited RCU grace periods block even less frequently
The purpose of commit 282d8998e997 ("srcu: Prevent expedited GPs
and blocking readers from consuming CPU") was to prevent a long
series of never-blocking expedited SRCU grace periods from blocking
kernel-live-patching (KLP) progress. Although it was successful, it also
resulted in excessive boot times on certain embedded workloads running
under qemu with the "-bios QEMU_EFI.fd" command line. Here "excessive"
means increasing the boot time up into the three-to-four minute range.
This increase in boot time was due to the more than 6000 back-to-back
invocations of synchronize_rcu_expedited() within the KVM host OS, which
in turn resulted from qemu's emulation of a long series of MMIO accesses.
Commit 640a7d37c3f4 ("srcu: Block less aggressively for expedited grace
periods") did not significantly help this particular use case.
Zhangfei Gao and Shameerali Kolothum Thodi did experiments varying the
value of SRCU_MAX_NODELAY_PHASE with HZ=250 and with various values
of non-sleeping per phase counts on a system with preemption enabled,
and observed the following boot times:
+──────────────────────────+────────────────+
| SRCU_MAX_NODELAY_PHASE | Boot time (s) |
+──────────────────────────+────────────────+
| 100 | 30.053 |
| 150 | 25.151 |
| 200 | 20.704 |
| 250 | 15.748 |
| 500 | 11.401 |
| 1000 | 11.443 |
| 10000 | 11.258 |
| 1000000 | 11.154 |
+──────────────────────────+────────────────+
Analysis on the experiment results additional improvements with non
Nit: "results additional improvements" -> "results show additional
improvements" ?
Thanks
Neeraj
blocking delays approaching one jiffy in duration. This improvement was
also seen when number of per-phase iterations were scaled to one jiffy.
This commit therefore scales per-grace-period phase number of non-sleeping
polls so that non-sleeping polls extend for about one jiffy. In addition,
the delay-calculation call to srcu_get_delay() in srcu_gp_end() is
replaced with a simple check for an expedited grace period. This change
schedules callback invocation immediately after expedited grace periods
complete, which results in greatly improved boot times. Testing done
by Marc and Zhangfei confirms that this change recovers most of the
performance degradation in boottime; for CONFIG_HZ_250 configuration,
specifically, boot times improve from 3m50s to 41s on Marc's setup;
and from 2m40s to ~9.7s on Zhangfei's setup.
In addition to the changes to default per phase delays, this
change adds 3 new kernel parameters - srcutree.srcu_max_nodelay,
srcutree.srcu_max_nodelay_phase, and srcutree.srcu_retry_check_delay.
This allows users to configure the srcu grace period scanning delays in
order to more quickly react to additional use cases.
Fixes: 640a7d37c3f4 ("srcu: Block less aggressively for expedited grace periods")
Fixes: 282d8998e997 ("srcu: Prevent expedited GPs and blocking readers from consuming CPU")
Reported-by: Zhangfei Gao <zhangfei.gao@xxxxxxxxxx>
Reported-by: yueluck <yueluck@xxxxxxx>
Signed-off-by: Neeraj Upadhyay <quic_neeraju@xxxxxxxxxxx>
Tested-by: Marc Zyngier <maz@xxxxxxxxxx>
Tested-by: Zhangfei Gao <zhangfei.gao@xxxxxxxxxx>
Link: https://lore.kernel.org/all/20615615-0013-5adc-584f-2b1d5c03ebfc@xxxxxxxxxx/
Signed-off-by: Paul E. McKenney <paulmck@xxxxxxxxxx>
diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index af647714c113d..7e34086c64f5b 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -5805,6 +5805,24 @@
expediting. Set to zero to disable automatic
expediting.
+ srcutree.srcu_max_nodelay [KNL]
+ Specifies the number of no-delay instances
+ per jiffy for which the SRCU grace period
+ worker thread will be rescheduled with zero
+ delay. Beyond this limit, worker thread will
+ be rescheduled with a sleep delay of one jiffy.
+
+ srcutree.srcu_max_nodelay_phase [KNL]
+ Specifies the per-grace-period phase, number of
+ non-sleeping polls of readers. Beyond this limit,
+ grace period worker thread will be rescheduled
+ with a sleep delay of one jiffy, between each
+ rescan of the readers, for a grace period phase.
+
+ srcutree.srcu_retry_check_delay [KNL]
+ Specifies number of microseconds of non-sleeping
+ delay between each non-sleeping poll of readers.
+
srcutree.small_contention_lim [KNL]
Specifies the number of update-side contention
events per jiffy will be tolerated before
diff --git a/kernel/rcu/srcutree.c b/kernel/rcu/srcutree.c
index 0db7873f4e95b..1c304fec89c02 100644
--- a/kernel/rcu/srcutree.c
+++ b/kernel/rcu/srcutree.c
@@ -511,10 +511,52 @@ static bool srcu_readers_active(struct srcu_struct *ssp)
return sum;
}
-#define SRCU_INTERVAL 1 // Base delay if no expedited GPs pending.
-#define SRCU_MAX_INTERVAL 10 // Maximum incremental delay from slow readers.
-#define SRCU_MAX_NODELAY_PHASE 3 // Maximum per-GP-phase consecutive no-delay instances.
-#define SRCU_MAX_NODELAY 100 // Maximum consecutive no-delay instances.
+/*
+ * We use an adaptive strategy for synchronize_srcu() and especially for
+ * synchronize_srcu_expedited(). We spin for a fixed time period
+ * (defined below, boot time configurable) to allow SRCU readers to exit
+ * their read-side critical sections. If there are still some readers
+ * after one jiffy, we repeatedly block for one jiffy time periods.
+ * The blocking time is increased as the grace-period age increases,
+ * with max blocking time capped at 10 jiffies.
+ */
+#define SRCU_DEFAULT_RETRY_CHECK_DELAY 5
+
+static ulong srcu_retry_check_delay = SRCU_DEFAULT_RETRY_CHECK_DELAY;
+module_param(srcu_retry_check_delay, ulong, 0444);
+
+#define SRCU_INTERVAL 1 // Base delay if no expedited GPs pending.
+#define SRCU_MAX_INTERVAL 10 // Maximum incremental delay from slow readers.
+
+#define SRCU_DEFAULT_MAX_NODELAY_PHASE_LO 3UL // Lowmark on default per-GP-phase
+ // no-delay instances.
+#define SRCU_DEFAULT_MAX_NODELAY_PHASE_HI 1000UL // Highmark on default per-GP-phase
+ // no-delay instances.
+
+#define SRCU_UL_CLAMP_LO(val, low) ((val) > (low) ? (val) : (low))
+#define SRCU_UL_CLAMP_HI(val, high) ((val) < (high) ? (val) : (high))
+#define SRCU_UL_CLAMP(val, low, high) SRCU_UL_CLAMP_HI(SRCU_UL_CLAMP_LO((val), (low)), (high))
+// per-GP-phase no-delay instances adjusted to allow non-sleeping poll upto
+// one jiffies time duration. Mult by 2 is done to factor in the srcu_get_delay()
+// called from process_srcu().
+#define SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED \
+ (2UL * USEC_PER_SEC / HZ / SRCU_DEFAULT_RETRY_CHECK_DELAY)
+
+// Maximum per-GP-phase consecutive no-delay instances.
+#define SRCU_DEFAULT_MAX_NODELAY_PHASE \
+ SRCU_UL_CLAMP(SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED, \
+ SRCU_DEFAULT_MAX_NODELAY_PHASE_LO, \
+ SRCU_DEFAULT_MAX_NODELAY_PHASE_HI)
+
+static ulong srcu_max_nodelay_phase = SRCU_DEFAULT_MAX_NODELAY_PHASE;
+module_param(srcu_max_nodelay_phase, ulong, 0444);
+
+// Maximum consecutive no-delay instances.
+#define SRCU_DEFAULT_MAX_NODELAY (SRCU_DEFAULT_MAX_NODELAY_PHASE > 100 ? \
+ SRCU_DEFAULT_MAX_NODELAY_PHASE : 100)
+
+static ulong srcu_max_nodelay = SRCU_DEFAULT_MAX_NODELAY;
+module_param(srcu_max_nodelay, ulong, 0444);
/*
* Return grace-period delay, zero if there are expedited grace
@@ -535,7 +577,7 @@ static unsigned long srcu_get_delay(struct srcu_struct *ssp)
jbase += j - gpstart;
if (!jbase) {
WRITE_ONCE(ssp->srcu_n_exp_nodelay, READ_ONCE(ssp->srcu_n_exp_nodelay) + 1);
- if (READ_ONCE(ssp->srcu_n_exp_nodelay) > SRCU_MAX_NODELAY_PHASE)
+ if (READ_ONCE(ssp->srcu_n_exp_nodelay) > srcu_max_nodelay_phase)
jbase = 1;
}
}
@@ -612,15 +654,6 @@ void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
}
EXPORT_SYMBOL_GPL(__srcu_read_unlock);
-/*
- * We use an adaptive strategy for synchronize_srcu() and especially for
- * synchronize_srcu_expedited(). We spin for a fixed time period
- * (defined below) to allow SRCU readers to exit their read-side critical
- * sections. If there are still some readers after a few microseconds,
- * we repeatedly block for 1-millisecond time periods.
- */
-#define SRCU_RETRY_CHECK_DELAY 5
-
/*
* Start an SRCU grace period.
*/
@@ -706,7 +739,7 @@ static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp
*/
static void srcu_gp_end(struct srcu_struct *ssp)
{
- unsigned long cbdelay;
+ unsigned long cbdelay = 1;
bool cbs;
bool last_lvl;
int cpu;
@@ -726,7 +759,9 @@ static void srcu_gp_end(struct srcu_struct *ssp)
spin_lock_irq_rcu_node(ssp);
idx = rcu_seq_state(ssp->srcu_gp_seq);
WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
- cbdelay = !!srcu_get_delay(ssp);
+ if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq), READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
+ cbdelay = 0;
+
WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
rcu_seq_end(&ssp->srcu_gp_seq);
gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
@@ -927,12 +962,16 @@ static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
*/
static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
{
+ unsigned long curdelay;
+
+ curdelay = !srcu_get_delay(ssp);
+
for (;;) {
if (srcu_readers_active_idx_check(ssp, idx))
return true;
- if (--trycount + !srcu_get_delay(ssp) <= 0)
+ if ((--trycount + curdelay) <= 0)
return false;
- udelay(SRCU_RETRY_CHECK_DELAY);
+ udelay(srcu_retry_check_delay);
}
}
@@ -1588,7 +1627,7 @@ static void process_srcu(struct work_struct *work)
j = jiffies;
if (READ_ONCE(ssp->reschedule_jiffies) == j) {
WRITE_ONCE(ssp->reschedule_count, READ_ONCE(ssp->reschedule_count) + 1);
- if (READ_ONCE(ssp->reschedule_count) > SRCU_MAX_NODELAY)
+ if (READ_ONCE(ssp->reschedule_count) > srcu_max_nodelay)
curdelay = 1;
} else {
WRITE_ONCE(ssp->reschedule_count, 1);
@@ -1680,6 +1719,11 @@ static int __init srcu_bootup_announce(void)
pr_info("Hierarchical SRCU implementation.\n");
if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
+ if (srcu_retry_check_delay != SRCU_DEFAULT_RETRY_CHECK_DELAY)
+ pr_info("\tNon-default retry check delay of %lu us.\n", srcu_retry_check_delay);
+ if (srcu_max_nodelay != SRCU_DEFAULT_MAX_NODELAY)
+ pr_info("\tNon-default max no-delay of %lu.\n", srcu_max_nodelay);
+ pr_info("\tMax phase no-delay instances is %lu.\n", srcu_max_nodelay_phase);
return 0;
}
early_initcall(srcu_bootup_announce);