Re: [PATCH] kretprobe scalability improvement

From: Matt Wu
Date: Tue Jul 06 2021 - 23:10:21 EST

On 2021/7/7 AM12:25, Masami Hiramatsu wrote:
On Tue, 6 Jul 2021 09:21:00 +0800
Matt Wu <wuqiang.matt@xxxxxxxxxxxxx> wrote:

On 2021/7/5 PM2:59, Masami Hiramatsu wrote:
Hi,

On Sat, 3 Jul 2021 18:28:18 +0800
"wuqiang" <wuqiang.matt@xxxxxxxxxxxxx> wrote:

From: wuqiang <wuqiang.matt@xxxxxxxxxxxxx>

The original freelist is a LIFO queue based on singly linked list, which lacks
of scalability, and thus becomes bottleneck under high workloads. freelist was
introduced by Masami Hiramatsu's work of removing kretprobe hash lock:
url: https://lkml.org/lkml/2020/8/29/209.

Here an array-based MPMC lockless queue is proposed. The solution of bounded
array can nicely avoid ABA issue, while freelist or circular queue etc. have
to perform 2 CAS loops. The other advantage is that order and fairness can be
ignored, the only concern is to retrieve kretprobe instance records as fast
as possible, i.e. performance and correctness. Tests of kretprobe on 96-CORE
ARM64 show the biggest gain as 466.7x of the original freelist throughput.
The raw queue throughput can be 1,975 times of freelist. Here are the results:

Ubuntu 20.04, 5.13.0-rc6 (XEON E5-2660V3 2.4G, DDR4 2133MT/s, 10 CORES/20 THREADS):
1x 2x 4x 8x 10x 16x 20x 32x 40x
freelist: 13086080 22493637 32773854 20129772 18455899 18435561 18980332 18988603 18991334
array : 13144036 26059941 47449954 94517172 115856027 116414714 125692971 125553061 125685981

Ubuntu 21.04 - 5.12.10 QEMU 96 CORES (HUAWEI TaiShan 2280V2 KP920 96 CORES 2.6G, DDR4 2944 MT/s):
1x 2x 4x 8x 16x 24x 48x 96x 192x
freelist: 17,233,640 10,296,664 8,095,309 6,993,545 5,050,817 4,295,283 3,382,013 2,738,050 2,743,345
array: 19,360,905 37,395,225 56,417,463 10,020,136 209,876,209 328,940,014 632,754,916 1,277,862,473 1,169,076,739

Interesting result! How would you measure the overhead?
And also could you clarify the real scalability example of kretprobe usage ?
E.g. putting kretprobes at some function and profiling with perf. See following
slides for details.

https://events.static.linuxfound.org/sites/events/files/slides/Handling%20the%20Massive%20Multiple%20Kprobes%20v2_1.pdf
(BTW, these efforts totally stalls a while, needs to be reviewed again)

I did two kinds of tests: one is real kretprobe, the other is throughput
comparison of different queue implementations.

1) kretprobe upon security_file_mprotect

We found the performance bottleneck due to udp_recvmsg kretprobe in
our production environment, then re-produced the issue with a lighter
syscall: mprotect.

"perf stat" is used to count number of sys_enter_mprotect calligs:
perf stat -a -I 10000 -e 'syscalls:sys_enter_mprotect' vmstat 1 35

The user mode program is just a loop of mprotect() to trigger the
registered kretprobe callbacks. The codes are pushed to:
https://github.com/mattwuq/kretprobe/blob/main/mprotect/

I measured both kprobe and kretprobe for 4.14/5.9/5.12. The results
of kprobe is really good, but kretprobe doesn't scale well (even for
kernel 5.12 with "kprobes: Remove kretprobe hash").

Hmm, Ok if there is a real kretprobe issue (not freelist), it should
be solved. Could you also point this result from your changelog?

Here are the resuts:

Linux 5.8.0-45-AMD64 Ubuntu T490 (i7-10510U 1.80G & DDR4 2667)
threads baseline kprobe kretprobe
1p 72,816,571 59,411,825 34,578,853
2p 111,336,194 91,219,319 40,303,484
3p 144,082,415 112,813,784 41,762,717
4p 142,233,213 118,947,750 33,103,895

Linux 5.12.0-AMD64 Ubuntu T490 (i7-10510U 1.80G & DDR4 2667)
threads baseline kprobe kretprobe
1p 72,705,816 59,523,413 39,391,428
2p 108,577,114 90,913,449 48,940,956
3p 143,493,477 118,791,390 41,067,841
4p 170,406,366 139,667,883 32,398,033

The chart picture is available at:
https://github.com/mattwuq/kretprobe/tree/main/doc/kprobe_krp_perf.png

For 5.8 the kretprobe performance is limited by kretprobe hash locking. Then I tried 5.12 with your patch of "Remove kretprobe hash" landed. But
kretprobe still don't scale, then I digged further and found freelist is
the culprit.

2) raw queue throughput benchmarks

I wrote a module with dedicated kernel threads performing insertions
and deletions of several freelist implementations for 10ms.

The codes and test scripts are available at:
https://github.com/mattwuq/kretprobe/blob/main/scalable/

1) fl.h: original freelist, LIFO queue based on singly linked list
2) ra.h: read from random position, write to last read pos
3) sa.h: array-based queue, per-cpu slot to be equally distributed
4) saca.h: the proposed version, allocating array with L1 cache line
aligned for each core
5) saea.h: make every elelment cache_line aligned
6) zz.h: a.k.a zigzag, remap numerical order to L1 cache distance,
for 64bit pointers, 0 to 0, 1 to 8, 2 to 16
7) cq.h: native circular queue, not used, can not handle reentrance

Two types of tests are performanced:
1) throughput bench: with no delay between deletion and insertion
2) emulation bench of real kretprobe: 1us delay before inserting back

All the results and charts are available at:
https://github.com/mattwuq/kretprobe/tree/main/doc/


OK, this test report is also great :)

Thanks. I will give the bpf-percpu-freelist a try this weekend.

So linear scalability is still not available, limited by the following two
considerations:

1. keep it simple: best solution could be an implementation of per-cpu queues,
but it's not applicable for this case due to complexity. After all for
most cases the number of pre-allocated kretprobe instances (maxactive) is
only a small value. If not specified by user during registering, maxactive
is set as CPU cores or 2x when preemption is enabled
2. keep it compact: cache-line-alignment can solve false-sharing and minimize
cache thrashing, but it introduces memory wasting, considering the small
body of structure kretprobe_instance. Secondly the performance improvement
of cache-line-aligned is not significant as expected

If you really need the linear scalability, drop useless entry-handler and per
instance data (or just leave the data pointer) and allocate the instance pool
for each task struct. This is perfectly scalable, because kretprobe instance
is only for making a shadow stack for the task, not CPU.

Yes, per-task list of kretprobe instances would deliver best throughput.
But the penality could be high in memory efficency and implementations.

How much penalty it would make? If we allocate a 4kb pool for each task,
it would be enough small compared with other resources (and we may be
able to select the pool on-line or compile option)

Servers here (typically 96 cores) can have 2000 tasks, but the hosts
providing docker services ccould have > 5000 tasks. One task can have
several threads, likely < 2 on average. So estimated penalty could be
5000 * 2 * 4K, 39M bytes, i.e. 0.4M bytes per core.

kretprobe is not a certain thing. It's might not used at all, or a task
might only trigger once in lifetime. The proposed solutions could provide
promising results with less than 0.4M bytes memory usage per core.

Inspired by your idea, I'm thinking of allocating from stack:

1) from stack top: need modify stack top limit, might “violate” the
purpose of guard page
2）reserve from current stack: need modify trampolines of fltrace and
kprobe, but there are many challenges.

No, I don't like this change because it will disturb the stack unwinder
and consuming the stack itself.

got it.

With a pre-built kernel, further performance tuning can be done by increasing
maxactive when registering kretprobe. Tests show 4x cores number is a fair
choice for both performance and memory efficiency.

Which test should I check? If it is also good for the current freelist,
I would like to expand default maxactive. (actually, current maxactive
is chosen by the minimum availability)

I tested with difference maxactive values. For current freelist, bigger
maxactive values have less effects upon performance.

So bigger 'maxactive' will scale better?

Yes, I guess it can reduce cache conflicts. Later I could do a measure
of cache misses.

XEON / X86_64 (preempt=0 / cycleus=0)
1x 10x 20x
zigzag:max=10 142649937 102381284 87993433
freelist:max=10 90050953 14533279 12234181
array:max=10 170718610 101061189 84507025
strided:max=10 170885073 1645070467 471586589
zigzag:max=20 142833611 251344437 124256740
freelist:max=20 83193711 13796546 12035244
array:max=20 157751314 208385189 139943284
strided:max=20 157810810 1818188777 2188112898
zigzag:max=40 154501823 682233175 242334634
freelist:max=40 83284714 13852153 11654861
array:max=40 157817022 361685213 251139824
strided:max=40 158885047 1791159293 1973298443

The chart url:
https://github.com/mattwuq/kretprobe/tree/main/doc/kretprobe_maxactive.png

"missed cases" was also tracked. Based on testings, so long as maxactive
is more than cores number, there won't be "missed cases".

That depends on where you put the probe. kretprobe can be nested and
sleepable. If you put a kretprobe on the function which doesn't yield,
you don't need bigger maxactive. But kretprobe on the function which
can sleep or yield, you may need more than that.

Sure, it's depends on the environment (loads & apps). So that should be
the caller's duty to specify when registering kreprobe.


More info is available at: https://github.com/mattwuq/kretprobe

Signed-off-by: wuqiang <wuqiang.matt@xxxxxxxxxxxxx>
---
include/linux/freelist.h | 187 +++++++++++++++++++--------------------
kernel/kprobes.c | 29 +++---
2 files changed, 107 insertions(+), 109 deletions(-)

diff --git a/include/linux/freelist.h b/include/linux/freelist.h
index fc1842b96469..3d4a0bc425b2 100644
--- a/include/linux/freelist.h
+++ b/include/linux/freelist.h
@@ -1,129 +1,122 @@
-/* SPDX-License-Identifier: GPL-2.0-only OR BSD-2-Clause */
+/* SPDX-License-Identifier: GPL-2.0-or-later */

Please do NOT change the license without the agreement of all copyright holders.
Or, add a new file and remove the current freelist.h.

What about this?

Ok, it's fine to me. Actually it's a totally rewrite of freelist.h. I'll
change it back in next version.

Thanks.


#ifndef FREELIST_H
#define FREELIST_H
+#include <linux/slab.h>
#include <linux/atomic.h>
/*
- * Copyright: cameron@xxxxxxxxxxxxxx
+ * lockless queue for kretprobe instances
*
- * A simple CAS-based lock-free free list. Not the fastest thing in the world
- * under heavy contention, but simple and correct (assuming nodes are never
- * freed until after the free list is destroyed), and fairly speedy under low
- * contention.
- *
- * Adapted from: https://moodycamel.com/blog/2014/solving-the-aba-problem-for-lock-free-free-lists
+ * It's an array-based MPMC lockless queue, solely for better scalability
+ * and contention mitigation. It's simple in implementation and compact in
+ * memory efficiency. The only concern is to retrieve kretprobe instance
+ * records as fast as possible, with both order and fairness ignored.
*/
struct freelist_node {
- atomic_t refs;
- struct freelist_node *next;
+ struct freelist_node *next;
};
-
struct freelist_head {
- struct freelist_node *head;
+ uint32_t fh_size; /* rounded to power of 2 */
+ uint32_t fh_mask; /* (fh_size - 1) */
+ uint16_t fh_bits; /* log2(fh_size) */
+ uint16_t fh_step; /* per-core shift stride */
+ uint32_t fh_used; /* num of elements in list */
+ struct freelist_node **fh_ents; /* array for krp instances */
};
-#define REFS_ON_FREELIST 0x80000000
-#define REFS_MASK 0x7FFFFFFF
+static inline int freelist_init(struct freelist_head *list, int max)
+{
+ uint32_t size, cores = num_possible_cpus();
+
+ list->fh_used = 0;
+ list->fh_step = ilog2(L1_CACHE_BYTES / sizeof(void *));
+ if (max < (cores << list->fh_step))
+ list->fh_size = roundup_pow_of_two(cores) << list->fh_step;
+ else
+ list->fh_size = roundup_pow_of_two(max);
+ list->fh_mask = list->fh_size - 1;
+ list->fh_bits = (uint16_t)ilog2(list->fh_size);
+ size = list->fh_size * sizeof(struct freelist_node *);
+ list->fh_ents = kzalloc(size, GFP_KERNEL);
+ if (!list->fh_ents)
+ return -ENOMEM;
+
+ return 0;
+}
-static inline void __freelist_add(struct freelist_node *node, struct freelist_head *list)
+static inline int freelist_try_add(struct freelist_node *node, struct freelist_head *list)
{
- /*
- * Since the refcount is zero, and nobody can increase it once it's
- * zero (except us, and we run only one copy of this method per node at
- * a time, i.e. the single thread case), then we know we can safely
- * change the next pointer of the node; however, once the refcount is
- * back above zero, then other threads could increase it (happens under
- * heavy contention, when the refcount goes to zero in between a load
- * and a refcount increment of a node in try_get, then back up to
- * something non-zero, then the refcount increment is done by the other
- * thread) -- so if the CAS to add the node to the actual list fails,
- * decrese the refcount and leave the add operation to the next thread
- * who puts the refcount back to zero (which could be us, hence the
- * loop).
- */
- struct freelist_node *head = READ_ONCE(list->head);
-
- for (;;) {
- WRITE_ONCE(node->next, head);
- atomic_set_release(&node->refs, 1);
-
- if (!try_cmpxchg_release(&list->head, &head, node)) {
- /*
- * Hmm, the add failed, but we can only try again when
- * the refcount goes back to zero.
- */
- if (atomic_fetch_add_release(REFS_ON_FREELIST - 1, &node->refs) == 1)
- continue;
+ uint32_t i, hint = list->fh_used << list->fh_step;
+
+ for (i = 0; i < list->fh_size; i++) {
+ struct freelist_node *item = NULL;
+ uint32_t slot = (i + hint) & list->fh_mask;
+ if (try_cmpxchg_release(&list->fh_ents[slot], &item, node)) {
+ list->fh_used++;
+ break;
}
- return;
}
+
+ return (i >= list->fh_size);
}
-static inline void freelist_add(struct freelist_node *node, struct freelist_head *list)
+static inline int freelist_add(struct freelist_node *node, struct freelist_head *list)
{
- /*
- * We know that the should-be-on-freelist bit is 0 at this point, so
- * it's safe to set it using a fetch_add.
- */
- if (!atomic_fetch_add_release(REFS_ON_FREELIST, &node->refs)) {
- /*
- * Oh look! We were the last ones referencing this node, and we
- * know we want to add it to the free list, so let's do it!
- */
- __freelist_add(node, list);
- }
+ uint32_t hint = raw_smp_processor_id() << list->fh_step;
+ uint32_t slot = ((uint32_t) hint) & list->fh_mask;
+
+ do {
+ struct freelist_node *item = NULL;
+ if (try_cmpxchg_release(&list->fh_ents[slot], &item, node))
+ return 0;
+ slot = (slot + 1) & list->fh_mask;
+ } while (1);
+
+ return -1;
}
static inline struct freelist_node *freelist_try_get(struct freelist_head *list)
{
- struct freelist_node *prev, *next, *head = smp_load_acquire(&list->head);
- unsigned int refs;
-
- while (head) {
- prev = head;
- refs = atomic_read(&head->refs);
- if ((refs & REFS_MASK) == 0 ||
- !atomic_try_cmpxchg_acquire(&head->refs, &refs, refs+1)) {
- head = smp_load_acquire(&list->head);
- continue;
+ struct freelist_node *node = NULL;
+ uint32_t i, hint = raw_smp_processor_id() << list->fh_step;
+
+ for (i = 0; i < list->fh_size; i++) {
+ uint32_t slot = (hint + i) & list->fh_mask;
+ struct freelist_node *item = smp_load_acquire(&list->fh_ents[slot]);
+ if (item && try_cmpxchg_release(&list->fh_ents[slot], &item, NULL)) {
+ node = item;
+ break;
}
+ }
- /*
- * Good, reference count has been incremented (it wasn't at
- * zero), which means we can read the next and not worry about
- * it changing between now and the time we do the CAS.
- */
- next = READ_ONCE(head->next);
- if (try_cmpxchg_acquire(&list->head, &head, next)) {
- /*
- * Yay, got the node. This means it was on the list,
- * which means should-be-on-freelist must be false no
- * matter the refcount (because nobody else knows it's
- * been taken off yet, it can't have been put back on).
- */
- WARN_ON_ONCE(atomic_read(&head->refs) & REFS_ON_FREELIST);
-
- /*
- * Decrease refcount twice, once for our ref, and once
- * for the list's ref.
- */
- atomic_fetch_add(-2, &head->refs);
-
- return head;
- }
+ return node;
+}
- /*
- * OK, the head must have changed on us, but we still need to decrement
- * the refcount we increased.
- */
- refs = atomic_fetch_add(-1, &prev->refs);
- if (refs == REFS_ON_FREELIST + 1)
- __freelist_add(prev, list);
+static inline void freelist_destroy(struct freelist_head *list, void *context,
+ int (*release)(void *, void *))
+{
+ uint32_t i;
+
+ if (!list->fh_ents)
+ return;
+
+ for (i = 0; i < list->fh_size; i++) {
+ uint32_t slot = i & list->fh_mask;
+ struct freelist_node *item = smp_load_acquire(&list->fh_ents[slot]);
+ while (item) {
+ if (try_cmpxchg_release(&list->fh_ents[slot], &item, NULL)) {
+ if (release)
+ release(context, item);
+ break;
+ }
+ }
}
- return NULL;
+ if (list->fh_ents) {
+ kfree(list->fh_ents);
+ list->fh_ents = NULL;
+ }
}
-
#endif /* FREELIST_H */
diff --git a/kernel/kprobes.c b/kernel/kprobes.c
index 471b1d18a92f..5c41bee25983 100644
--- a/kernel/kprobes.c
+++ b/kernel/kprobes.c
@@ -1277,20 +1277,21 @@ void kprobe_flush_task(struct task_struct *tk)
}
NOKPROBE_SYMBOL(kprobe_flush_task);
-static inline void free_rp_inst(struct kretprobe *rp)
+static int release_ri(void *context, void *node)
{
struct kretprobe_instance *ri;
- struct freelist_node *node;
- int count = 0;
+ ri = container_of(node, struct kretprobe_instance, freelist);
+ kfree(ri);
+ if (context)
+ (*((int *)context))++;
+ return 0;
+}
- node = rp->freelist.head;
- while (node) {
- ri = container_of(node, struct kretprobe_instance, freelist);
- node = node->next;
+static inline void free_rp_inst(struct kretprobe *rp)
+{
+ int count = 0;
- kfree(ri);
- count++;
- }
+ freelist_destroy(&rp->freelist, &count, release_ri);
if (refcount_sub_and_test(count, &rp->rph->ref)) {
kfree(rp->rph);
@@ -2015,10 +2016,14 @@ int register_kretprobe(struct kretprobe *rp)
rp->maxactive = num_possible_cpus();
#endif
}
- rp->freelist.head = NULL;
+ if (freelist_init(&rp->freelist, rp->maxactive))
+ return -ENOMEM;
+
rp->rph = kzalloc(sizeof(struct kretprobe_holder), GFP_KERNEL);
- if (!rp->rph)
+ if (!rp->rph) {
+ freelist_destroy(&rp->freelist, NULL, NULL);
return -ENOMEM;
+ }
rp->rph->rp = rp;
for (i = 0; i < rp->maxactive; i++) {
--
2.25.1