v3:
- Minor comment twisting as suggested by Yosry.
- Add patches 2 and 3 to further reduce lock hold time
The purpose of this patch series is to reduce of the cpu_lock hold time
in cgroup_rstat_flush_locked() so as to reduce the latency impact when
cgroup_rstat_updated() is called as they may contend with each other
on the cpu_lock.
A parallel kernel build on a 2-socket x86-64 server is used as the
benchmarking tool for measuring the lock hold time. Below were the lock
hold time frequency distribution before and after applying different
number of patches:
Hold time Before patch Patch 1 Patches 1-2 Patches 1-3
--------- ------------ ------- ----------- -----------
0-01 us 804,139 13,738,708 14,594,545 15,484,707
01-05 us 9,772,767 1,177,194 439,926 207,382
05-10 us 4,595,028 4,984 5,960 3,174
10-15 us 303,481 3,562 3,543 3,006
15-20 us 78,971 1,314 1,397 1,066
20-25 us 24,583 18 25 15
25-30 us 6,908 12 12 10
30-40 us 8,015
40-50 us 2,192
50-60 us 316
60-70 us 43
70-80 us 7
80-90 us 2
>90 us 3
Waiman Long (3):
cgroup/rstat: Reduce cpu_lock hold time in cgroup_rstat_flush_locked()
cgroup/rstat: Optimize cgroup_rstat_updated_list()
cgroup: Avoid false cacheline sharing of read mostly rstat_cpu
include/linux/cgroup-defs.h | 14 ++++
kernel/cgroup/rstat.c | 129 +++++++++++++++++++++---------------
2 files changed, 89 insertions(+), 54 deletions(-)
--
2.39.3
The rstat_cpu and also rstat_css_list of the cgroup structure are read
mostly variables. However, they may share the same cacheline as the
subsequent rstat_flush_next and *bstat variables which can be updated
frequently. That will slow down the cgroup_rstat_cpu() call which is
called pretty frequently in the rstat code. Add a CACHELINE_PADDING()
line in between them to avoid false cacheline sharing.
A parallel kernel build on a 2-socket x86-64 server is used as the
benchmarking tool for measuring the lock hold time. Below were the lock
hold time frequency distribution before and after the patch:
Run time Before patch After patch
-------- ------------ -----------
0-01 us 14,594,545 15,484,707
01-05 us 439,926 207,382
05-10 us 5,960 3,174
10-15 us 3,543 3,006
15-20 us 1,397 1,066
20-25 us 25 15
25-30 us 12 10
It can be seen that the patch further pushes the lock hold time towards
the lower end.
Signed-off-by: Waiman Long <[email protected]>
---
include/linux/cgroup-defs.h | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/include/linux/cgroup-defs.h b/include/linux/cgroup-defs.h
index ff4b4c590f32..a4adc0580135 100644
--- a/include/linux/cgroup-defs.h
+++ b/include/linux/cgroup-defs.h
@@ -491,6 +491,13 @@ struct cgroup {
struct cgroup_rstat_cpu __percpu *rstat_cpu;
struct list_head rstat_css_list;
+ /*
+ * Add padding to separate the read mostly rstat_cpu and
+ * rstat_css_list into a different cacheline from the following
+ * rstat_flush_next and *bstat fields which can have frequent updates.
+ */
+ CACHELINE_PADDING(_pad_);
+
/*
* A singly-linked list of cgroup structures to be rstat flushed.
* This is a scratch field to be used exclusively by
--
2.39.3
When cgroup_rstat_updated() isn't being called concurrently with
cgroup_rstat_flush_locked(), its run time is pretty short. When
both are called concurrently, the cgroup_rstat_updated() run time
can spike to a pretty high value due to high cpu_lock hold time in
cgroup_rstat_flush_locked(). This can be problematic if the task calling
cgroup_rstat_updated() is a realtime task running on an isolated CPU
with a strict latency requirement. The cgroup_rstat_updated() call can
happen when there is a page fault even though the task is running in
user space most of the time.
The percpu cpu_lock is used to protect the update tree -
updated_next and updated_children. This protection is only needed when
cgroup_rstat_cpu_pop_updated() is being called. The subsequent flushing
operation which can take a much longer time does not need that protection
as it is already protected by cgroup_rstat_lock.
To reduce the cpu_lock hold time, we need to perform all the
cgroup_rstat_cpu_pop_updated() calls up front with the lock
released afterward before doing any flushing. This patch adds a new
cgroup_rstat_updated_list() function to return a singly linked list of
cgroups to be flushed.
Some instrumentation code are added to measure the cpu_lock hold time
right after lock acquisition to after releasing the lock. Parallel
kernel build on a 2-socket x86-64 server is used as the benchmarking
tool for measuring the lock hold time.
The maximum cpu_lock hold time before and after the patch are 100us and
29us respectively. So the worst case time is reduced to about 30% of
the original. However, there may be some OS or hardware noises like NMI
or SMI in the test system that can worsen the worst case value. Those
noises are usually tuned out in a real production environment to get
a better result.
OTOH, the lock hold time frequency distribution should give a better
idea of the performance benefit of the patch. Below were the frequency
distribution before and after the patch:
Hold time Before patch After patch
--------- ------------ -----------
0-01 us 804,139 13,738,708
01-05 us 9,772,767 1,177,194
05-10 us 4,595,028 4,984
10-15 us 303,481 3,562
15-20 us 78,971 1,314
20-25 us 24,583 18
25-30 us 6,908 12
30-40 us 8,015
40-50 us 2,192
50-60 us 316
60-70 us 43
70-80 us 7
80-90 us 2
>90 us 3
Signed-off-by: Waiman Long <[email protected]>
Reviewed-by: Yosry Ahmed <[email protected]>
---
include/linux/cgroup-defs.h | 7 ++++++
kernel/cgroup/rstat.c | 43 ++++++++++++++++++++++++-------------
2 files changed, 35 insertions(+), 15 deletions(-)
diff --git a/include/linux/cgroup-defs.h b/include/linux/cgroup-defs.h
index 265da00a1a8b..ff4b4c590f32 100644
--- a/include/linux/cgroup-defs.h
+++ b/include/linux/cgroup-defs.h
@@ -491,6 +491,13 @@ struct cgroup {
struct cgroup_rstat_cpu __percpu *rstat_cpu;
struct list_head rstat_css_list;
+ /*
+ * A singly-linked list of cgroup structures to be rstat flushed.
+ * This is a scratch field to be used exclusively by
+ * cgroup_rstat_flush_locked() and protected by cgroup_rstat_lock.
+ */
+ struct cgroup *rstat_flush_next;
+
/* cgroup basic resource statistics */
struct cgroup_base_stat last_bstat;
struct cgroup_base_stat bstat;
diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
index d80d7a608141..1f300bf4dc40 100644
--- a/kernel/cgroup/rstat.c
+++ b/kernel/cgroup/rstat.c
@@ -145,6 +145,32 @@ static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
return pos;
}
+/* Return a list of updated cgroups to be flushed */
+static struct cgroup *cgroup_rstat_updated_list(struct cgroup *root, int cpu)
+{
+ raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
+ struct cgroup *head, *tail, *next;
+ unsigned long flags;
+
+ /*
+ * The _irqsave() is needed because cgroup_rstat_lock is
+ * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
+ * this lock with the _irq() suffix only disables interrupts on
+ * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
+ * interrupts on both configurations. The _irqsave() ensures
+ * that interrupts are always disabled and later restored.
+ */
+ raw_spin_lock_irqsave(cpu_lock, flags);
+ head = tail = cgroup_rstat_cpu_pop_updated(NULL, root, cpu);
+ while (tail) {
+ next = cgroup_rstat_cpu_pop_updated(tail, root, cpu);
+ tail->rstat_flush_next = next;
+ tail = next;
+ }
+ raw_spin_unlock_irqrestore(cpu_lock, flags);
+ return head;
+}
+
/*
* A hook for bpf stat collectors to attach to and flush their stats.
* Together with providing bpf kfuncs for cgroup_rstat_updated() and
@@ -179,21 +205,9 @@ static void cgroup_rstat_flush_locked(struct cgroup *cgrp)
lockdep_assert_held(&cgroup_rstat_lock);
for_each_possible_cpu(cpu) {
- raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock,
- cpu);
- struct cgroup *pos = NULL;
- unsigned long flags;
+ struct cgroup *pos = cgroup_rstat_updated_list(cgrp, cpu);
- /*
- * The _irqsave() is needed because cgroup_rstat_lock is
- * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
- * this lock with the _irq() suffix only disables interrupts on
- * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
- * interrupts on both configurations. The _irqsave() ensures
- * that interrupts are always disabled and later restored.
- */
- raw_spin_lock_irqsave(cpu_lock, flags);
- while ((pos = cgroup_rstat_cpu_pop_updated(pos, cgrp, cpu))) {
+ for (; pos; pos = pos->rstat_flush_next) {
struct cgroup_subsys_state *css;
cgroup_base_stat_flush(pos, cpu);
@@ -205,7 +219,6 @@ static void cgroup_rstat_flush_locked(struct cgroup *cgrp)
css->ss->css_rstat_flush(css, cpu);
rcu_read_unlock();
}
- raw_spin_unlock_irqrestore(cpu_lock, flags);
/* play nice and yield if necessary */
if (need_resched() || spin_needbreak(&cgroup_rstat_lock)) {
--
2.39.3
The current design of cgroup_rstat_cpu_pop_updated() is to traverse
the updated tree in a way to pop out the leaf nodes first before
their parents. This can cause traversal of multiple nodes before a
leaf node can be found and popped out. IOW, a given node in the tree
can be visited multiple times before the whole operation is done. So
it is not very efficient and the code can be hard to read.
With the introduction of cgroup_rstat_updated_list() to build a list
of cgroups to be flushed first before any flushing operation is being
done, we can optimize the way the updated tree nodes are being popped
by pushing the parents first to the tail end of the list before their
children. In this way, most updated tree nodes will be visited only
once with the exception of the subtree root as we still need to go
back to its parent and popped it out of its updated_children list.
This also makes the code easier to read.
A parallel kernel build on a 2-socket x86-64 server is used as the
benchmarking tool for measuring the lock hold time. Below were the lock
hold time frequency distribution before and after the patch:
Hold time Before patch After patch
--------- ------------ -----------
0-01 us 13,738,708 14,594,545
01-05 us 1,177,194 439,926
05-10 us 4,984 5,960
10-15 us 3,562 3,543
15-20 us 1,314 1,397
20-25 us 18 25
25-30 us 12 12
It can be seen that the patch pushes the lock hold time towards the
lower end.
Signed-off-by: Waiman Long <[email protected]>
---
kernel/cgroup/rstat.c | 132 ++++++++++++++++++++++--------------------
1 file changed, 70 insertions(+), 62 deletions(-)
diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
index 1f300bf4dc40..d2b709cfeb2a 100644
--- a/kernel/cgroup/rstat.c
+++ b/kernel/cgroup/rstat.c
@@ -74,64 +74,90 @@ __bpf_kfunc void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
}
/**
- * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
- * @pos: current position
- * @root: root of the tree to traversal
+ * cgroup_rstat_push_children - push children cgroups into the given list
+ * @head: current head of the list (= parent cgroup)
+ * @prstatc: cgroup_rstat_cpu of the parent cgroup
* @cpu: target cpu
+ * Return: A new singly linked list of cgroups to be flush
*
- * Walks the updated rstat_cpu tree on @cpu from @root. %NULL @pos starts
- * the traversal and %NULL return indicates the end. During traversal,
- * each returned cgroup is unlinked from the tree. Must be called with the
- * matching cgroup_rstat_cpu_lock held.
+ * Recursively traverse down the cgroup_rstat_cpu updated tree and push
+ * parent first before its children. The parent is pushed by the caller.
+ * The recursion depth is the depth of the current updated tree.
+ */
+static struct cgroup *cgroup_rstat_push_children(struct cgroup *head,
+ struct cgroup_rstat_cpu *prstatc, int cpu)
+{
+ struct cgroup *child, *parent;
+ struct cgroup_rstat_cpu *crstatc;
+
+ parent = head;
+ child = prstatc->updated_children;
+ prstatc->updated_children = parent;
+
+ /* updated_next is parent cgroup terminated */
+ while (child != parent) {
+ child->rstat_flush_next = head;
+ head = child;
+ crstatc = cgroup_rstat_cpu(child, cpu);
+ if (crstatc->updated_children != parent)
+ head = cgroup_rstat_push_children(head, crstatc, cpu);
+ child = crstatc->updated_next;
+ crstatc->updated_next = NULL;
+ }
+ return head;
+}
+
+/**
+ * cgroup_rstat_updated_list - return a list of updated cgroups to be flushed
+ * @root: root of the cgroup subtree to traverse
+ * @cpu: target cpu
+ * Return: A singly linked list of cgroups to be flushed
+ *
+ * Walks the updated rstat_cpu tree on @cpu from @root. During traversal,
+ * each returned cgroup is unlinked from the updated tree. Must be called
+ * with the matching cgroup_rstat_cpu_lock held.
*
* The only ordering guarantee is that, for a parent and a child pair
- * covered by a given traversal, if a child is visited, its parent is
- * guaranteed to be visited afterwards.
+ * covered by a given traversal, the child is before its parent in
+ * the list.
+ *
+ * Note that updated_children is self terminated while updated_next is
+ * parent cgroup terminated except the cgroup root which can be self
+ * terminated.
*/
-static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
- struct cgroup *root, int cpu)
+static struct cgroup *cgroup_rstat_updated_list(struct cgroup *root, int cpu)
{
- struct cgroup_rstat_cpu *rstatc;
- struct cgroup *parent;
-
- if (pos == root)
- return NULL;
+ raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
+ struct cgroup_rstat_cpu *rstatc = cgroup_rstat_cpu(root, cpu);
+ struct cgroup *head = NULL, *parent;
+ unsigned long flags;
/*
- * We're gonna walk down to the first leaf and visit/remove it. We
- * can pick whatever unvisited node as the starting point.
+ * The _irqsave() is needed because cgroup_rstat_lock is
+ * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
+ * this lock with the _irq() suffix only disables interrupts on
+ * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
+ * interrupts on both configurations. The _irqsave() ensures
+ * that interrupts are always disabled and later restored.
*/
- if (!pos) {
- pos = root;
- /* return NULL if this subtree is not on-list */
- if (!cgroup_rstat_cpu(pos, cpu)->updated_next)
- return NULL;
- } else {
- pos = cgroup_parent(pos);
- }
+ raw_spin_lock_irqsave(cpu_lock, flags);
- /* walk down to the first leaf */
- while (true) {
- rstatc = cgroup_rstat_cpu(pos, cpu);
- if (rstatc->updated_children == pos)
- break;
- pos = rstatc->updated_children;
- }
+ /* Return NULL if this subtree is not on-list */
+ if (!rstatc->updated_next)
+ goto unlock_ret;
/*
- * Unlink @pos from the tree. As the updated_children list is
+ * Unlink @root from its parent. As the updated_children list is
* singly linked, we have to walk it to find the removal point.
- * However, due to the way we traverse, @pos will be the first
- * child in most cases. The only exception is @root.
*/
- parent = cgroup_parent(pos);
+ parent = cgroup_parent(root);
if (parent) {
struct cgroup_rstat_cpu *prstatc;
struct cgroup **nextp;
prstatc = cgroup_rstat_cpu(parent, cpu);
nextp = &prstatc->updated_children;
- while (*nextp != pos) {
+ while (*nextp != root) {
struct cgroup_rstat_cpu *nrstatc;
nrstatc = cgroup_rstat_cpu(*nextp, cpu);
@@ -142,31 +168,13 @@ static struct cgroup *cgroup_rstat_cpu_pop_updated(struct cgroup *pos,
}
rstatc->updated_next = NULL;
- return pos;
-}
-
-/* Return a list of updated cgroups to be flushed */
-static struct cgroup *cgroup_rstat_updated_list(struct cgroup *root, int cpu)
-{
- raw_spinlock_t *cpu_lock = per_cpu_ptr(&cgroup_rstat_cpu_lock, cpu);
- struct cgroup *head, *tail, *next;
- unsigned long flags;
- /*
- * The _irqsave() is needed because cgroup_rstat_lock is
- * spinlock_t which is a sleeping lock on PREEMPT_RT. Acquiring
- * this lock with the _irq() suffix only disables interrupts on
- * a non-PREEMPT_RT kernel. The raw_spinlock_t below disables
- * interrupts on both configurations. The _irqsave() ensures
- * that interrupts are always disabled and later restored.
- */
- raw_spin_lock_irqsave(cpu_lock, flags);
- head = tail = cgroup_rstat_cpu_pop_updated(NULL, root, cpu);
- while (tail) {
- next = cgroup_rstat_cpu_pop_updated(tail, root, cpu);
- tail->rstat_flush_next = next;
- tail = next;
- }
+ /* Push @root to the list first before pushing the children */
+ head = root;
+ root->rstat_flush_next = NULL;
+ if (rstatc->updated_children != root)
+ head = cgroup_rstat_push_children(head, rstatc, cpu);
+unlock_ret:
raw_spin_unlock_irqrestore(cpu_lock, flags);
return head;
}
--
2.39.3
On Fri, Nov 3, 2023 at 8:13 PM Waiman Long <[email protected]> wrote:
>
> The current design of cgroup_rstat_cpu_pop_updated() is to traverse
> the updated tree in a way to pop out the leaf nodes first before
> their parents. This can cause traversal of multiple nodes before a
> leaf node can be found and popped out. IOW, a given node in the tree
> can be visited multiple times before the whole operation is done. So
> it is not very efficient and the code can be hard to read.
>
> With the introduction of cgroup_rstat_updated_list() to build a list
> of cgroups to be flushed first before any flushing operation is being
> done, we can optimize the way the updated tree nodes are being popped
> by pushing the parents first to the tail end of the list before their
> children. In this way, most updated tree nodes will be visited only
> once with the exception of the subtree root as we still need to go
> back to its parent and popped it out of its updated_children list.
> This also makes the code easier to read.
>
> A parallel kernel build on a 2-socket x86-64 server is used as the
> benchmarking tool for measuring the lock hold time. Below were the lock
> hold time frequency distribution before and after the patch:
>
> Hold time Before patch After patch
> --------- ------------ -----------
> 0-01 us 13,738,708 14,594,545
> 01-05 us 1,177,194 439,926
> 05-10 us 4,984 5,960
> 10-15 us 3,562 3,543
> 15-20 us 1,314 1,397
> 20-25 us 18 25
> 25-30 us 12 12
>
> It can be seen that the patch pushes the lock hold time towards the
> lower end.
>
> Signed-off-by: Waiman Long <[email protected]>
> ---
I don't know why git decided to show this diff in the most confusing
way possible.
> kernel/cgroup/rstat.c | 132 ++++++++++++++++++++++--------------------
> 1 file changed, 70 insertions(+), 62 deletions(-)
>
> diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
> index 1f300bf4dc40..d2b709cfeb2a 100644
> --- a/kernel/cgroup/rstat.c
> +++ b/kernel/cgroup/rstat.c
> @@ -74,64 +74,90 @@ __bpf_kfunc void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
> }
>
> /**
> - * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
> - * @pos: current position
> - * @root: root of the tree to traversal
> + * cgroup_rstat_push_children - push children cgroups into the given list
> + * @head: current head of the list (= parent cgroup)
> + * @prstatc: cgroup_rstat_cpu of the parent cgroup
> * @cpu: target cpu
> + * Return: A new singly linked list of cgroups to be flush
> *
> - * Walks the updated rstat_cpu tree on @cpu from @root. %NULL @pos starts
> - * the traversal and %NULL return indicates the end. During traversal,
> - * each returned cgroup is unlinked from the tree. Must be called with the
> - * matching cgroup_rstat_cpu_lock held.
> + * Recursively traverse down the cgroup_rstat_cpu updated tree and push
> + * parent first before its children. The parent is pushed by the caller.
I think it might be useful here (and elsewhere in the patch) where
"push" is being used to elaborate that we push to the beginning in a
stack-like fashion.
> + * The recursion depth is the depth of the current updated tree.
> + */
> +static struct cgroup *cgroup_rstat_push_children(struct cgroup *head,
> + struct cgroup_rstat_cpu *prstatc, int cpu)
> +{
> + struct cgroup *child, *parent;
> + struct cgroup_rstat_cpu *crstatc;
> +
> + parent = head;
> + child = prstatc->updated_children;
> + prstatc->updated_children = parent;
> +
> + /* updated_next is parent cgroup terminated */
> + while (child != parent) {
> + child->rstat_flush_next = head;
> + head = child;
> + crstatc = cgroup_rstat_cpu(child, cpu);
> + if (crstatc->updated_children != parent)
I think cgroup->updated_children is set to the cgroup itself if it's
empty, right? Shouldn't this be crstatc->updated_children != child?
> + head = cgroup_rstat_push_children(head, crstatc, cpu);
> + child = crstatc->updated_next;
> + crstatc->updated_next = NULL;
> + }
> + return head;
> +}
> +
> +/**
> + * cgroup_rstat_updated_list - return a list of updated cgroups to be flushed
> + * @root: root of the cgroup subtree to traverse
> + * @cpu: target cpu
> + * Return: A singly linked list of cgroups to be flushed
> + *
> + * Walks the updated rstat_cpu tree on @cpu from @root. During traversal,
> + * each returned cgroup is unlinked from the updated tree. Must be called
> + * with the matching cgroup_rstat_cpu_lock held.
This function takes care of holding the lock actually. I think that
sentence should be applied to cgroup_rstat_push_children() above?
> *
> * The only ordering guarantee is that, for a parent and a child pair
> - * covered by a given traversal, if a child is visited, its parent is
> - * guaranteed to be visited afterwards.
> + * covered by a given traversal, the child is before its parent in
> + * the list.
> + *
> + * Note that updated_children is self terminated while updated_next is
> + * parent cgroup terminated except the cgroup root which can be self
> + * terminated.
IIUC updated_children and updated_next is the same list.
updated_children is the head, and updated_next is how the list items
are linked. This comment makes it seem like they are two different
lists.
I am actually wondering if it's worth using the singly linked list
here. We are saving 8 bytes percpu, but the semantics are fairly
confusing. Wouldn't this be easier to reason about if you just use
list_head?
updated_children would be replaced with LIST_HEAD (or similar), and
the list would be NULL terminated instead of terminated by self/parent
cgroup. IIUC the reason it's not NULL-terminated now is because we use
cgroup->updated_next to check quickly if a cgroup is on the list or
not. If we use list_heads, we can just use list_emtpy() IIUC.
We can also simplify the semantics of unlinking @root from the updated
tree below, it would just be list_del() IIUC, which is actually more
performant as well. It seems like overall we would simplify a lot of
things. When forming the updated_list, we can just walk the tree and
splice the lists in the correct order.
It seems to me that saving 8 bytes percpu is not worth the complexity
of the custom list semantics here. Am I missing something here?
On 11/6/23 15:07, Yosry Ahmed wrote:
> On Fri, Nov 3, 2023 at 8:13 PM Waiman Long <[email protected]> wrote:
>> The current design of cgroup_rstat_cpu_pop_updated() is to traverse
>> the updated tree in a way to pop out the leaf nodes first before
>> their parents. This can cause traversal of multiple nodes before a
>> leaf node can be found and popped out. IOW, a given node in the tree
>> can be visited multiple times before the whole operation is done. So
>> it is not very efficient and the code can be hard to read.
>>
>> With the introduction of cgroup_rstat_updated_list() to build a list
>> of cgroups to be flushed first before any flushing operation is being
>> done, we can optimize the way the updated tree nodes are being popped
>> by pushing the parents first to the tail end of the list before their
>> children. In this way, most updated tree nodes will be visited only
>> once with the exception of the subtree root as we still need to go
>> back to its parent and popped it out of its updated_children list.
>> This also makes the code easier to read.
>>
>> A parallel kernel build on a 2-socket x86-64 server is used as the
>> benchmarking tool for measuring the lock hold time. Below were the lock
>> hold time frequency distribution before and after the patch:
>>
>> Hold time Before patch After patch
>> --------- ------------ -----------
>> 0-01 us 13,738,708 14,594,545
>> 01-05 us 1,177,194 439,926
>> 05-10 us 4,984 5,960
>> 10-15 us 3,562 3,543
>> 15-20 us 1,314 1,397
>> 20-25 us 18 25
>> 25-30 us 12 12
>>
>> It can be seen that the patch pushes the lock hold time towards the
>> lower end.
>>
>> Signed-off-by: Waiman Long <[email protected]>
>> ---
> I don't know why git decided to show this diff in the most confusing
> way possible.
I agree. The diff is really hard to look at. It will be easier to apply
the patch & looks at the actual rstat.c file.
>
>> kernel/cgroup/rstat.c | 132 ++++++++++++++++++++++--------------------
>> 1 file changed, 70 insertions(+), 62 deletions(-)
>>
>> diff --git a/kernel/cgroup/rstat.c b/kernel/cgroup/rstat.c
>> index 1f300bf4dc40..d2b709cfeb2a 100644
>> --- a/kernel/cgroup/rstat.c
>> +++ b/kernel/cgroup/rstat.c
>> @@ -74,64 +74,90 @@ __bpf_kfunc void cgroup_rstat_updated(struct cgroup *cgrp, int cpu)
>> }
>>
>> /**
>> - * cgroup_rstat_cpu_pop_updated - iterate and dismantle rstat_cpu updated tree
>> - * @pos: current position
>> - * @root: root of the tree to traversal
>> + * cgroup_rstat_push_children - push children cgroups into the given list
>> + * @head: current head of the list (= parent cgroup)
>> + * @prstatc: cgroup_rstat_cpu of the parent cgroup
>> * @cpu: target cpu
>> + * Return: A new singly linked list of cgroups to be flush
>> *
>> - * Walks the updated rstat_cpu tree on @cpu from @root. %NULL @pos starts
>> - * the traversal and %NULL return indicates the end. During traversal,
>> - * each returned cgroup is unlinked from the tree. Must be called with the
>> - * matching cgroup_rstat_cpu_lock held.
>> + * Recursively traverse down the cgroup_rstat_cpu updated tree and push
>> + * parent first before its children. The parent is pushed by the caller.
> I think it might be useful here (and elsewhere in the patch) where
> "push" is being used to elaborate that we push to the beginning in a
> stack-like fashion.
Right, I am thinking about a stack when I use the word "push". I will
clarify that in the comment.
>
>> + * The recursion depth is the depth of the current updated tree.
>> + */
>> +static struct cgroup *cgroup_rstat_push_children(struct cgroup *head,
>> + struct cgroup_rstat_cpu *prstatc, int cpu)
>> +{
>> + struct cgroup *child, *parent;
>> + struct cgroup_rstat_cpu *crstatc;
>> +
>> + parent = head;
>> + child = prstatc->updated_children;
>> + prstatc->updated_children = parent;
>> +
>> + /* updated_next is parent cgroup terminated */
>> + while (child != parent) {
>> + child->rstat_flush_next = head;
>> + head = child;
>> + crstatc = cgroup_rstat_cpu(child, cpu);
>> + if (crstatc->updated_children != parent)
> I think cgroup->updated_children is set to the cgroup itself if it's
> empty, right? Shouldn't this be crstatc->updated_children != child?
My mistake. Will fix it in the next version.
>
>> + head = cgroup_rstat_push_children(head, crstatc, cpu);
>> + child = crstatc->updated_next;
>> + crstatc->updated_next = NULL;
>> + }
>> + return head;
>> +}
>> +
>> +/**
>> + * cgroup_rstat_updated_list - return a list of updated cgroups to be flushed
>> + * @root: root of the cgroup subtree to traverse
>> + * @cpu: target cpu
>> + * Return: A singly linked list of cgroups to be flushed
>> + *
>> + * Walks the updated rstat_cpu tree on @cpu from @root. During traversal,
>> + * each returned cgroup is unlinked from the updated tree. Must be called
>> + * with the matching cgroup_rstat_cpu_lock held.
> This function takes care of holding the lock actually. I think that
> sentence should be applied to cgroup_rstat_push_children() above?
It is left over from before this patch. Will remove that.
>
>> *
>> * The only ordering guarantee is that, for a parent and a child pair
>> - * covered by a given traversal, if a child is visited, its parent is
>> - * guaranteed to be visited afterwards.
>> + * covered by a given traversal, the child is before its parent in
>> + * the list.
>> + *
>> + * Note that updated_children is self terminated while updated_next is
>> + * parent cgroup terminated except the cgroup root which can be self
>> + * terminated.
> IIUC updated_children and updated_next is the same list.
> updated_children is the head, and updated_next is how the list items
> are linked. This comment makes it seem like they are two different
> lists.
Thanks for the comment. I will rework the comment to clarify that a bit
more.
>
> I am actually wondering if it's worth using the singly linked list
> here. We are saving 8 bytes percpu, but the semantics are fairly
> confusing. Wouldn't this be easier to reason about if you just use
> list_head?
>
> updated_children would be replaced with LIST_HEAD (or similar), and
> the list would be NULL terminated instead of terminated by self/parent
> cgroup. IIUC the reason it's not NULL-terminated now is because we use
> cgroup->updated_next to check quickly if a cgroup is on the list or
> not. If we use list_heads, we can just use list_emtpy() IIUC.
>
> We can also simplify the semantics of unlinking @root from the updated
> tree below, it would just be list_del() IIUC, which is actually more
> performant as well. It seems like overall we would simplify a lot of
> things. When forming the updated_list, we can just walk the tree and
> splice the lists in the correct order.
>
> It seems to me that saving 8 bytes percpu is not worth the complexity
> of the custom list semantics here. Am I missing something here?
It will cost an additional 16 bytes of percpu memory if converted to
list_heads. Like other lists, there will be sibling and children
list_heads. There are also 2 pointers to update instead of one. Anyway,
I don't have an objection to convert them to list_heads if agreed by Tejun.
Cheers,
Longman
On Mon, Nov 6, 2023 at 12:37 PM Waiman Long <[email protected]> wrote:
>
> On 11/6/23 15:07, Yosry Ahmed wrote:
> > On Fri, Nov 3, 2023 at 8:13 PM Waiman Long <[email protected]> wrote:
> >> The current design of cgroup_rstat_cpu_pop_updated() is to traverse
> >> the updated tree in a way to pop out the leaf nodes first before
> >> their parents. This can cause traversal of multiple nodes before a
> >> leaf node can be found and popped out. IOW, a given node in the tree
> >> can be visited multiple times before the whole operation is done. So
> >> it is not very efficient and the code can be hard to read.
> >>
> >> With the introduction of cgroup_rstat_updated_list() to build a list
> >> of cgroups to be flushed first before any flushing operation is being
> >> done, we can optimize the way the updated tree nodes are being popped
> >> by pushing the parents first to the tail end of the list before their
> >> children. In this way, most updated tree nodes will be visited only
> >> once with the exception of the subtree root as we still need to go
> >> back to its parent and popped it out of its updated_children list.
> >> This also makes the code easier to read.
> >>
> >> A parallel kernel build on a 2-socket x86-64 server is used as the
> >> benchmarking tool for measuring the lock hold time. Below were the lock
> >> hold time frequency distribution before and after the patch:
> >>
> >> Hold time Before patch After patch
> >> --------- ------------ -----------
> >> 0-01 us 13,738,708 14,594,545
> >> 01-05 us 1,177,194 439,926
> >> 05-10 us 4,984 5,960
> >> 10-15 us 3,562 3,543
> >> 15-20 us 1,314 1,397
> >> 20-25 us 18 25
> >> 25-30 us 12 12
> >>
> >> It can be seen that the patch pushes the lock hold time towards the
> >> lower end.
> >>
> >> Signed-off-by: Waiman Long <[email protected]>
> >> ---
> > I don't know why git decided to show this diff in the most confusing
> > way possible.
> I agree. The diff is really hard to look at. It will be easier to apply
> the patch & looks at the actual rstat.c file.
Would the diff be simpler if patches 1 & 2 were squashed?
[..]
> >
> >> *
> >> * The only ordering guarantee is that, for a parent and a child pair
> >> - * covered by a given traversal, if a child is visited, its parent is
> >> - * guaranteed to be visited afterwards.
> >> + * covered by a given traversal, the child is before its parent in
> >> + * the list.
> >> + *
> >> + * Note that updated_children is self terminated while updated_next is
> >> + * parent cgroup terminated except the cgroup root which can be self
> >> + * terminated.
> > IIUC updated_children and updated_next is the same list.
> > updated_children is the head, and updated_next is how the list items
> > are linked. This comment makes it seem like they are two different
> > lists.
> Thanks for the comment. I will rework the comment to clarify that a bit
> more.
> >
> > I am actually wondering if it's worth using the singly linked list
> > here. We are saving 8 bytes percpu, but the semantics are fairly
> > confusing. Wouldn't this be easier to reason about if you just use
> > list_head?
> >
> > updated_children would be replaced with LIST_HEAD (or similar), and
> > the list would be NULL terminated instead of terminated by self/parent
> > cgroup. IIUC the reason it's not NULL-terminated now is because we use
> > cgroup->updated_next to check quickly if a cgroup is on the list or
> > not. If we use list_heads, we can just use list_emtpy() IIUC.
> >
> > We can also simplify the semantics of unlinking @root from the updated
> > tree below, it would just be list_del() IIUC, which is actually more
> > performant as well. It seems like overall we would simplify a lot of
> > things. When forming the updated_list, we can just walk the tree and
> > splice the lists in the correct order.
> >
> > It seems to me that saving 8 bytes percpu is not worth the complexity
> > of the custom list semantics here. Am I missing something here?
>
> It will cost an additional 16 bytes of percpu memory if converted to
> list_heads. Like other lists, there will be sibling and children
> list_heads. There are also 2 pointers to update instead of one. Anyway,
> I don't have an objection to convert them to list_heads if agreed by Tejun.
Yes you are right. It's definitely not free, but it's also not super
costly. It's just that every time I look at the rstat code I need to
remind myself of how updated_next and updated_children work. I will
let Tejun decide.