apparmor: global buffers spin lock may get contended

Hi,

We've notices that apparmor has switched from using per-CPU buffer pool and per-CPU spin_lock to a global spin_lock in df323337e507a0009d3db1ea.

This seems to be causing some contention on our build machines (with quite a bit of cores). Because that global spin lock is a part of the stat() sys call (and perhaps some other)

E.g.

- 9.29% 0.00% clang++ [kernel.vmlinux]
- 9.28% entry_SYSCALL_64_after_hwframe
- 8.98% do_syscall_64
- 7.43% __do_sys_newlstat
- 7.43% vfs_statx
- 7.18% security_inode_getattr
- 7.15% apparmor_inode_getattr
- aa_path_perm
- 3.53% aa_get_buffer
- 3.47% _raw_spin_lock
3.44% native_queued_spin_lock_slowpath
- 3.49% aa_put_buffer.part.0
- 3.45% _raw_spin_lock
3.43% native_queued_spin_lock_slowpath

Can we fix this contention?

Hi,

We've notices that apparmor has switched from using per-CPU buffer pool and per-CPU spin_lock to a global spin_lock in df323337e507a0009d3db1ea.

This seems to be causing some contention on our build machines (with quite a bit of cores). Because that global spin lock is a part of the stat() sys call (and perhaps some other)

E.g.

- 9.29% 0.00% clang++ [kernel.vmlinux]
- 9.28% entry_SYSCALL_64_after_hwframe
- 8.98% do_syscall_64
- 7.43% __do_sys_newlstat
- 7.43% vfs_statx
- 7.18% security_inode_getattr
- 7.15% apparmor_inode_getattr
- aa_path_perm
- 3.53% aa_get_buffer
- 3.47% _raw_spin_lock
3.44% native_queued_spin_lock_slowpath
- 3.49% aa_put_buffer.part.0
- 3.45% _raw_spin_lock
3.43% native_queued_spin_lock_slowpath

Can we fix this contention?

On 28 Oct 2022 02:34:07 -0700 John Johansen <[email protected]>

Quoted 144 lines · Sergey Senozhatsky

Quoted 26 lines

Hi,

We've notices that apparmor has switched from using per-CPU buffer pool and per-CPU spin_lock to a global spin_lock in df323337e507a0009d3db1ea.

This seems to be causing some contention on our build machines (with quite a bit of cores). Because that global spin lock is a part of the stat() sys call (and perhaps some other)

E.g.

- 9.29% 0.00% clang++ [kernel.vmlinux]
- 9.28% entry_SYSCALL_64_after_hwframe
- 8.98% do_syscall_64
- 7.43% __do_sys_newlstat
- 7.43% vfs_statx
- 7.18% security_inode_getattr
- 7.15% apparmor_inode_getattr
- aa_path_perm
- 3.53% aa_get_buffer
- 3.47% _raw_spin_lock
3.44% native_queued_spin_lock_slowpath
- 3.49% aa_put_buffer.part.0
- 3.45% _raw_spin_lock
3.43% native_queued_spin_lock_slowpath

Quoted 1 line

Can we fix this contention?

sorry for the delay on this. Below is a proposed patch that I have been testing to deal with this issue.

From d026988196fdbda7234fb87bc3e4aea22edcbaf9 Mon Sep 17 00:00:00 2001
From: John Johansen <[email protected]>
Date: Tue, 25 Oct 2022 01:18:41 -0700
Subject: [PATCH] apparmor: cache buffers on percpu list if there is lock contention

On a heavily loaded machine there can be lock contention on the global buffers lock. Add a percpu list to cache buffers on when lock contention is encountered.

When allocating buffers attempt to use cached buffers first, before taking the global buffers lock. When freeing buffers try to put them back to the global list but if contention is encountered, put the buffer on the percpu list.

The length of time a buffer is held on the percpu list is dynamically adjusted based on lock contention. The amount of hold time is rapidly increased and slow ramped down.

Signed-off-by John Johansen <[email protected]>

---
   security/apparmor/lsm.c | 74 ++++++++++++++++++++++++++++++++++++++---
   1 file changed, 69 insertions(+), 5 deletions(-)

security/apparmor/lsm.c +40 -1

index 25114735bc11..0ab70171bdb6 100644
@@ -49,12 +49,19 @@ union aa_buffer {
   	char buffer[1];
   };
 
+struct aa_local_cache {
+	unsigned int contention;
+	unsigned int hold;
+	struct list_head head;
+};
+
   #define RESERVE_COUNT 2
   static int reserve_count = RESERVE_COUNT;
   static int buffer_count;
 
   static LIST_HEAD(aa_global_buffers);
   static DEFINE_SPINLOCK(aa_buffers_lock);
+static DEFINE_PER_CPU(struct aa_local_cache, aa_local_buffers);
 
   /*
    * LSM hook functions

@@ -1622,14 +1629,44 @@ static int param_set_mode(const char *val, const struct kernel_param *kp)
   	return 0;
   }
 
+static void update_contention(struct aa_local_cache *cache)
+{
+	cache->contention += 3;
+	if (cache->contention > 9)
+		cache->contention = 9;
+	cache->hold += 1 << cache->contention;		/* 8, 64, 512 */
+}
+
   char *aa_get_buffer(bool in_atomic)
   {
   	union aa_buffer *aa_buf;
+	struct aa_local_cache *cache;
   	bool try_again = true;
   	gfp_t flags = (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
 
+	/* use per cpu cached buffers first */
+	cache = get_cpu_ptr(&aa_local_buffers);
+	if (!list_empty(&cache->head)) {
+		aa_buf = list_first_entry(&cache->head, union aa_buffer, list);
+		list_del(&aa_buf->list);
+		cache->hold--;
+		put_cpu_ptr(&aa_local_buffers);
+		return &aa_buf->buffer[0];
+	}
+	put_cpu_ptr(&aa_local_buffers);
+
+	if (!spin_trylock(&aa_buffers_lock)) {
+		cache = get_cpu_ptr(&aa_local_buffers);
+		update_contention(cache);
+		put_cpu_ptr(&aa_local_buffers);
+		spin_lock(&aa_buffers_lock);
+	} else {
+		cache = get_cpu_ptr(&aa_local_buffers);
+		if (cache->contention)
+			cache->contention--;
+		put_cpu_ptr(&aa_local_buffers);
+	}
   retry:
-	spin_lock(&aa_buffers_lock);
   	if (buffer_count > reserve_count ||
   	    (in_atomic && !list_empty(&aa_global_buffers))) {
   		aa_buf = list_first_entry(&aa_global_buffers, union aa_buffer,

@@ -1655,6 +1692,7 @@ char *aa_get_buffer(bool in_atomic)
   	if (!aa_buf) {
   		if (try_again) {
   			try_again = false;
+			spin_lock(&aa_buffers_lock);
   			goto retry;
   		}
   		pr_warn_once("AppArmor: Failed to allocate a memory buffer.\n");

@@ -1666,15 +1704,32 @@ char *aa_get_buffer(bool in_atomic)
   void aa_put_buffer(char *buf)
   {
   	union aa_buffer *aa_buf;
+	struct aa_local_cache *cache;
 
   	if (!buf)
   		return;
   	aa_buf = container_of(buf, union aa_buffer, buffer[0]);
 

security/apparmor/lsm.c char *aa_get_buffer(bool in_atomic) · line 1718

Quoted 10 lines

security/apparmor/lsm.c +6 -4

-	spin_lock(&aa_buffers_lock);
-	list_add(&aa_buf->list, &aa_global_buffers);
-	buffer_count++;
-	spin_unlock(&aa_buffers_lock);
+	cache = get_cpu_ptr(&aa_local_buffers);
+	if (!cache->hold) {
+		put_cpu_ptr(&aa_local_buffers);
+		if (spin_trylock(&aa_buffers_lock)) {
+			list_add(&aa_buf->list, &aa_global_buffers);
+			buffer_count++;

Given !hold and trylock, right time to drain the perpcu cache?

Quoted 35 lines

security/apparmor/lsm.c +23 -0

+			spin_unlock(&aa_buffers_lock);
+			cache = get_cpu_ptr(&aa_local_buffers);
+			if (cache->contention)
+				cache->contention--;
+			put_cpu_ptr(&aa_local_buffers);
+			return;
+		}
+		cache = get_cpu_ptr(&aa_local_buffers);
+		update_contention(cache);
+	}
+
+	/* cache in percpu list */
+	list_add(&aa_buf->list, &cache->head);
+	put_cpu_ptr(&aa_local_buffers);
   }
 
   /*

@@ -1716,6 +1771,15 @@ static int __init alloc_buffers(void)
   	union aa_buffer *aa_buf;
   	int i, num;
 
+	/*
+	 * per cpu set of cached allocated buffers used to help reduce
+	 * lock contention
+	 */
+	for_each_possible_cpu(i) {
+		per_cpu(aa_local_buffers, i).contention = 0;
+		per_cpu(aa_local_buffers, i).hold = 0;
+		INIT_LIST_HEAD(&per_cpu(aa_local_buffers, i).head);
+	}
   	/*
   	 * A function may require two buffers at once. Usually the buffers are
   	 * used for a short period of time and are shared. On UP kernel buffers

Signature · 1 line

2.34.1

From d026988196fdbda7234fb87bc3e4aea22edcbaf9 Mon Sep 17 00:00:00 2001
From: John Johansen <[email protected]>
Date: Tue, 25 Oct 2022 01:18:41 -0700
Subject: [PATCH] apparmor: cache buffers on percpu list if there is lock
contention

On a heavily loaded machine there can be lock contention on the global buffers lock. Add a percpu list to cache buffers on when lock contention is encountered.

When allocating buffers attempt to use cached buffers first, before taking the global buffers lock. When freeing buffers try to put them back to the global list but if contention is encountered, put the buffer on the percpu list.

The length of time a buffer is held on the percpu list is dynamically adjusted based on lock contention. The amount of hold time is rapidly increased and slow ramped down.

Signed-off-by John Johansen <[email protected]>

Quoted 15 lines · John Johansen

From d026988196fdbda7234fb87bc3e4aea22edcbaf9 Mon Sep 17 00:00:00 2001
From: John Johansen <[email protected]>
Date: Tue, 25 Oct 2022 01:18:41 -0700
Subject: [PATCH] apparmor: cache buffers on percpu list if there is lock
contention

On a heavily loaded machine there can be lock contention on the global buffers lock. Add a percpu list to cache buffers on when lock contention is encountered.

When allocating buffers attempt to use cached buffers first, before taking the global buffers lock. When freeing buffers try to put them back to the global list but if contention is encountered, put the buffer on the percpu list.

Quoted 5 lines

The length of time a buffer is held on the percpu list is dynamically adjusted based on lock contention. The amount of hold time is rapidly increased and slow ramped down.

Signed-off-by John Johansen <[email protected]>

Reported-by Sergey Senozhatsky <[email protected]>

Quoted 20 lines · Sergey Senozhatsky

Quoted 19 lines · John Johansen

From d026988196fdbda7234fb87bc3e4aea22edcbaf9 Mon Sep 17 00:00:00 2001
From: John Johansen <[email protected]>
Date: Tue, 25 Oct 2022 01:18:41 -0700
Subject: [PATCH] apparmor: cache buffers on percpu list if there is lock
contention

On a heavily loaded machine there can be lock contention on the global buffers lock. Add a percpu list to cache buffers on when lock contention is encountered.

When allocating buffers attempt to use cached buffers first, before taking the global buffers lock. When freeing buffers try to put them back to the global list but if contention is encountered, put the buffer on the percpu list.

The length of time a buffer is held on the percpu list is dynamically adjusted based on lock contention. The amount of hold time is rapidly increased and slow ramped down.

Quoted 3 lines

Quoted 1 line

Signed-off-by John Johansen <[email protected]>

Reported-by Sergey Senozhatsky <[email protected]>

yep, thanks for catching that

security/apparmor/lsm.c union aa_buffer { · line 55

Quoted 10 lines · John Johansen

security/apparmor/lsm.c +5 -0

@@ -49,12 +49,19 @@ union aa_buffer {
  	char buffer[1];
  };
+struct aa_local_cache {
+	unsigned int contention;
+	unsigned int hold;
+	struct list_head head;
+};

if you stick a local_lock_t into that struct, then you could replace

	cache = get_cpu_ptr(&aa_local_buffers);
with
	local_lock(&aa_local_buffers.lock);
	cache = this_cpu_ptr(&aa_local_buffers);

You would get the preempt_disable() based locking for the per-CPU variable (as with get_cpu_ptr()) and additionally some lockdep validation which would warn if it is used outside of task context (IRQ).

I didn't parse completely the hold/contention logic but it seems to work ;)
You check "cache->count >= 2" twice but I don't see an inc/ dec of it nor is it part of aa_local_cache.

I can't parse how many items can end up on the local list if the global list is locked. My guess would be more than 2 due the ->hold parameter.

Do you have any numbers on the machine and performance it improved? It sure will be a good selling point.

Quoted 13 lines · Sebastian Andrzej Siewior

security/apparmor/lsm.c union aa_buffer { · line 55

Quoted 10 lines · John Johansen

security/apparmor/lsm.c +5 -0

@@ -49,12 +49,19 @@ union aa_buffer {
     char buffer[1];
  };
+struct aa_local_cache {
+    unsigned int contention;
+    unsigned int hold;
+    struct list_head head;
+};

if you stick a local_lock_t into that struct, then you could replace

Quoted 9 lines

      cache = get_cpu_ptr(&aa_local_buffers);
with
      local_lock(&aa_local_buffers.lock);
      cache = this_cpu_ptr(&aa_local_buffers);

You would get the preempt_disable() based locking for the per-CPU variable (as with get_cpu_ptr()) and additionally some lockdep validation which would warn if it is used outside of task context (IRQ).

I did look at local_locks and there was a reason I didn't use them. I can't recall as the original iteration of this is over a year old now. I will have to dig into it again.

Quoted 5 lines

I didn't parse completely the hold/contention logic but it seems to work ;)
You check "cache->count >= 2" twice but I don't see an inc/ dec of it nor is it part of aa_local_cache.

sadly I messed up the reordering of this and the debug patch. This will be fixed in v4.

Quoted 3 lines

I can't parse how many items can end up on the local list if the global list is locked. My guess would be more than 2 due the ->hold parameter.

So this iteration, forces pushing back to global list if there are already two on the local list. The hold parameter just affects how long the buffers remain on the local list, before trying to place them back on the global list.

Originally before the count was added more than 2 buffers could end up on the local list, and having too many local buffers is a waste of memory. The count got added to address this. The value of 2 (which should be switched to a define) was chosen because no mediation routine currently uses more than 2 buffers.

Note that this doesn't mean that more than two buffers can be allocated to a tasks on a cpu. Its possible in some cases to have a task have allocated buffers and to still have buffers on the local cache list.

Quoted 3 lines

Do you have any numbers on the machine and performance it improved? It sure will be a good selling point.

I can include some supporting info, for a 16 core machine. But it will take some time to for me to get access to a bigger machine, where this is much more important. Hence the call for some of the other people on this thread to test.

thanks for the feedback

I can test the patch with 5.10 and 5.15 kernels in different machines. Just let me know which machine types you would like me to test.

Quoted 65 lines · John Johansen

Quoted 13 lines · Sebastian Andrzej Siewior

security/apparmor/lsm.c union aa_buffer { · line 55

Quoted 10 lines · John Johansen

security/apparmor/lsm.c +5 -0

@@ -49,12 +49,19 @@ union aa_buffer {
     char buffer[1];
  };
+struct aa_local_cache {
+    unsigned int contention;
+    unsigned int hold;
+    struct list_head head;
+};

if you stick a local_lock_t into that struct, then you could replace

Quoted 9 lines

      cache = get_cpu_ptr(&aa_local_buffers);
with
      local_lock(&aa_local_buffers.lock);
      cache = this_cpu_ptr(&aa_local_buffers);

You would get the preempt_disable() based locking for the per-CPU variable (as with get_cpu_ptr()) and additionally some lockdep validation which would warn if it is used outside of task context (IRQ).

I did look at local_locks and there was a reason I didn't use them. I can't recall as the original iteration of this is over a year old now. I will have to dig into it again.

Quoted 5 lines

I didn't parse completely the hold/contention logic but it seems to work ;)
You check "cache->count >= 2" twice but I don't see an inc/ dec of it nor is it part of aa_local_cache.

sadly I messed up the reordering of this and the debug patch. This will be fixed in v4.

Quoted 3 lines

I can't parse how many items can end up on the local list if the global list is locked. My guess would be more than 2 due the ->hold parameter.

So this iteration, forces pushing back to global list if there are already two on the local list. The hold parameter just affects how long the buffers remain on the local list, before trying to place them back on the global list.

Originally before the count was added more than 2 buffers could end up on the local list, and having too many local buffers is a waste of memory. The count got added to address this. The value of 2 (which should be switched to a define) was chosen because no mediation routine currently uses more than 2 buffers.

Note that this doesn't mean that more than two buffers can be allocated to a tasks on a cpu. Its possible in some cases to have a task have allocated buffers and to still have buffers on the local cache list.

Quoted 3 lines

Do you have any numbers on the machine and performance it improved? It sure will be a good selling point.

I can include some supporting info, for a 16 core machine. But it will take some time to for me to get access to a bigger machine, where this is much more important. Hence the call for some of the other people on this thread to test.

thanks for the feedback

Hi John,

I was wondering if you get a chance to work on patch v4. Please let me know if you need help with testing.

Best,
Anil

On Tue, Feb 21, 2023 at 1:27 PM Anil Altinay <[email protected] <mailto:[email protected]>> wrote:

I can test the patch with 5.10 and 5.15 kernels in different machines. Just let me know which machine types you would like me to test.

On Mon, Feb 20, 2023 at 12:42 AM John Johansen <[email protected] <mailto:[email protected]>> wrote:
>
> On 2/17/23 02:44, Sebastian Andrzej Siewior wrote:
> > On 2023-02-16 16:08:10 [-0800], John Johansen wrote:
> >> --- a/security/apparmor/lsm.c
> >> +++ b/security/apparmor/lsm.c
> >> @@ -49,12 +49,19 @@ union aa_buffer {
> >>      char buffer[1];
> >>   };
> >> +struct aa_local_cache {
> >> +    unsigned int contention;
> >> +    unsigned int hold;
> >> +    struct list_head head;
> >> +};
> >
> > if you stick a local_lock_t into that struct, then you could replace > >       cache = get_cpu_ptr(&aa_local_buffers);
> > with
> >       local_lock(&aa_local_buffers.lock);
> >       cache = this_cpu_ptr(&aa_local_buffers);
> >
> > You would get the preempt_disable() based locking for the per-CPU > > variable (as with get_cpu_ptr()) and additionally some lockdep > > validation which would warn if it is used outside of task context (IRQ). > >
> I did look at local_locks and there was a reason I didn't use them. I > can't recall as the original iteration of this is over a year old now. > I will have to dig into it again.
>
> > I didn't parse completely the hold/contention logic but it seems to work > > ;)
> > You check "cache->count >=  2" twice but I don't see an inc/ dec of it > > nor is it part of aa_local_cache.
> >
> sadly I messed up the reordering of this and the debug patch. This will be > fixed in v4.
>
> > I can't parse how many items can end up on the local list if the global > > list is locked. My guess would be more than 2 due the ->hold parameter. > >
> So this iteration, forces pushing back to global list if there are already > two on the local list. The hold parameter just affects how long the > buffers remain on the local list, before trying to place them back on > the global list.
>
> Originally before the count was added more than 2 buffers could end up > on the local list, and having too many local buffers is a waste of > memory. The count got added to address this. The value of 2 (which should > be switched to a define) was chosen because no mediation routine currently > uses more than 2 buffers.
>
> Note that this doesn't mean that more than two buffers can be allocated > to a tasks on a cpu. Its possible in some cases to have a task have > allocated buffers and to still have buffers on the local cache list. >
> > Do you have any numbers on the machine and performance it improved? It > > sure will be a good selling point.
> >
>
> I can include some supporting info, for a 16 core machine. But it will > take some time to for me to get access to a bigger machine, where this > is much more important. Hence the call for some of the other people > on this thread to test.
>
> thanks for the feedback
>

Quoted 4 lines · Anil Altinay

Hi John,

I was wondering if you get a chance to work on patch v4. Please let me know if you need help with testing.

yeah, testing help is always much appreciated. I have a v4, and I am working on 3 alternate version to compare against, to help give a better sense if we can get away with simplifying or tweak the scaling.

I should be able to post them out some time tonight.

Quoted 10 lines · John Johansen

Quoted 4 lines · Anil Altinay

Hi John,

I was wondering if you get a chance to work on patch v4. Please let me know if you need help with testing.

yeah, testing help is always much appreciated. I have a v4, and I am working on 3 alternate version to compare against, to help give a better sense if we can get away with simplifying or tweak the scaling.

Quoted 1 line

I should be able to post them out some time tonight.

Hi John,

Did you get a chance to post v4? I may be able to give it some testing on our real-life case.

Quoted 8 lines

Quoted 5 lines

yeah, testing help is always much appreciated. I have a v4, and I am working on 3 alternate version to compare against, to help give a better sense if we can get away with simplifying or tweak the scaling.

I should be able to post them out some time tonight.

Hi John,

Quoted 2 lines

Did you get a chance to post v4? I may be able to give it some testing on our real-life case.

sorry yes, how about a v5. That is simplified with 3 follow on patches that aren't strictly necessary, but some combination of them might be better than just the base patch, but splitting them out makes the individual changes easier to review.