IPC locks breaking down with RCU

Hi Linus,

This is the latest version of the ipc lock patch. It breaks down the three global IPC locks into one lock per IPC ID, also addresses the cache line bouncing problem introduced in the original patch. The original post could be found at:
http://marc.theaimsgroup.com/?l=linux-kernel&m=102980357802682&w=2

The original patch breaks down the global IPC locks, yet added another layer of locking to protect the IPC ID array in case of resizing. Some concern was raised that the read/write lock may cause cache line bouncing.

Since write lock is only used when the array is dynamically resized, RCU seems perfectly fit for this situation. By doing so it could reduce the possible lock contention in some applications where the IPC resources are heavily used, without introducing cache line bouncing.

Besides the RCU changes, it also remove the redundant ipc_lockall() and ipc_unlockall() as suggested by Hugh Dickins.

Patch is against 2.5.43 kernel. It requires Dipankar Sarma's read_barrier_depends RCU helper patch:
http://marc.theaimsgroup.com/?l=linux-kernel&m=103479438017486&w=2

We use the ipc lock on OracleApps and it gave us the best number. Please include.

Mingming Cao

Binary files linux-2.5.43/arch/i386/boot/compressed/vmlinux.bin.gz and linux-2.5.43-ipc/arch/i386/boot/compressed/vmlinux.bin.gz differ

diff -urN -X dontdiff linux-2.5.43/include/linux/ipc.h linux-2.5.43-ipc/include/linux/ipc.h
@@ -56,6 +56,7 @@
 /* used by in-kernel data structures */
 struct kern_ipc_perm
 {
+	spinlock_t	lock;
 	key_t		key;
 	uid_t		uid;
 	gid_t		gid;

diff -urN -X dontdiff linux-2.5.43/ipc/shm.c linux-2.5.43-ipc/ipc/shm.c
@@ -38,8 +38,6 @@
 
 #define shm_lock(id)	((struct shmid_kernel*)ipc_lock(&shm_ids,id))
 #define shm_unlock(id)	ipc_unlock(&shm_ids,id)
-#define shm_lockall()	ipc_lockall(&shm_ids)
-#define shm_unlockall()	ipc_unlockall(&shm_ids)
 #define shm_get(id)	((struct shmid_kernel*)ipc_get(&shm_ids,id))
 #define shm_buildid(id, seq) \
 	ipc_buildid(&shm_ids, id, seq)

@@ -409,14 +407,12 @@
 
 		memset(&shm_info,0,sizeof(shm_info));
 		down(&shm_ids.sem);
-		shm_lockall();
 		shm_info.used_ids = shm_ids.in_use;
 		shm_get_stat (&shm_info.shm_rss, &shm_info.shm_swp);
 		shm_info.shm_tot = shm_tot;
 		shm_info.swap_attempts = 0;
 		shm_info.swap_successes = 0;
 		err = shm_ids.max_id;
-		shm_unlockall();
 		up(&shm_ids.sem);
 		if(copy_to_user (buf, &shm_info, sizeof(shm_info)))
 			return -EFAULT;

diff -urN -X dontdiff linux-2.5.43/ipc/util.c linux-2.5.43-ipc/ipc/util.c
@@ -92,8 +92,10 @@
 {
 	int id;
 	struct kern_ipc_perm* p;
+	int max_id = ids->max_id;
 
-	for (id = 0; id <= ids->max_id; id++) {
+	read_barrier_depends();
+	for (id = 0; id <= max_id; id++) {
 		p = ids->entries[id].p;
 		if(p==NULL)
 			continue;

@@ -106,8 +108,8 @@
 static int grow_ary(struct ipc_ids* ids, int newsize)
 {
 	struct ipc_id* new;
-	struct ipc_id* old;
 	int i;
+	struct rcu_ipc_array *arg = NULL;
 
 	if(newsize > IPCMNI)
 		newsize = IPCMNI;

@@ -121,14 +123,19 @@
 	for(i=ids->size;i<newsize;i++) {
 		new[i].p = NULL;
 	}
+	arg = (struct rcu_ipc_array *) kmalloc(sizeof(*arg), GFP_KERNEL);
+	if(arg == NULL)
+		return ids->size;
+	arg->entries = ids->entries;
+	arg->size = ids->size;
+	
 	spin_lock(&ids->ary);
-
-	old = ids->entries;
 	ids->entries = new;
-	i = ids->size;
+	wmb();
 	ids->size = newsize;
 	spin_unlock(&ids->ary);
-	ipc_free(old, sizeof(struct ipc_id)*i);
+
+	call_rcu(&arg->rh, ipc_free_callback, arg);
 	return ids->size;
 }
 

@@ -166,7 +173,9 @@
 	if(ids->seq > ids->seq_max)
 		ids->seq = 0;
 
-	spin_lock(&ids->ary);
+	new->lock = SPIN_LOCK_UNLOCKED;
+	rcu_read_lock();
+	spin_lock(&new->lock);
 	ids->entries[id].p = new;
 	return id;
 }

@@ -188,6 +197,7 @@
 	int lid = id % SEQ_MULTIPLIER;
 	if(lid >= ids->size)
 		BUG();
+	rmb();
 	p = ids->entries[lid].p;
 	ids->entries[lid].p = NULL;
 	if(p==NULL)

@@ -239,7 +249,12 @@
 	else
 		kfree(ptr);
 }
-
+static void ipc_free_callback(void * arg)
+{
+	struct rcu_ipc_array *a = (struct rcu_ipc_array *)arg;
+	ipc_free(a->entries, a->size);
+	kfree(arg);
+}
 /**
  *	ipcperms	-	check IPC permissions
  *	@ipcp: IPC permission set

This looks very good to me: I'm glad you found the RCU idea works out. No need for performance numbers, this is now clearly the right way to go. And read_barrier_depends is in 2.5.44, so no problem there.

I'm ignorant of RCU, and my mind goes mushy around memory barriers, but I expect you've consulted the best there; and I'll be wanting to refer to this implementation as a nice example of how to use RCU. But please make a couple of small cleanups, unless you disagree.

Now delete spinlock_t ary and all references to it: only grow_ary is using it, and it's already protected by sem, and we'd be in trouble with concurrent allocations if it were not.

And I'd be happier to see ipc_unlock without those conditionals i.e. delete the "if(lid >= ids->size) return;" and the "if (out)" - they seem to encourage calling ipc_unlock where ipc_lock did not succeed, but that would be unsafe. If you found somewhere that's being done, I think we need to fix that place, not work around it in ipc_unlock.

Linus is away this week (so I've left him off, to avoid clogging up /dev/null): perhaps Andrew could take your patch into his -mm tree when you've made those changes (or persuaded us against)?

Hugh

Sorry, MingMing, it doesn't look so good to me now.

The "if(lid >= ids->size) return;" still looks unnecessary, but I think I see why you have "if (out)" in ipc_unlock: because of ipc_rmid, which has already nulled out entries[lid].p, yes?

A minor point is, wouldn't that skipping of spin_unlock leave you with the wrong preempt count, on a CONFIG_PREEMPT y configuration? But that's easily dealt with.

A much more serious point: we could certainly bring the ipc_rmid and ipc_unlock much closer together; but however close we bring them (unlock implicit within rmid), there will still be a race with one cpu in ipc_lock spinning on out->lock, while we in ipc_rmid null entries[lid].p and unlock and free the structure containing that lock.

I think you're going to have to extend RCU to freeing the entry (though this is a much less exceptional case than growing the array, so less clear to me that RCU is appropriate here), if you're to avoid reverting to the earlier rwlock or embedded spinlock designs.

Perhaps there's a simpler solution - ask around - but I don't see it.

Hugh

Thanks a lot for your comments. Yes. That's the consideration.

Thanks for pointing this out. This is a issue that has to be addressed.

A simple solution I could think of for this problem, moving the per_id lock out of the kern_ipc_perm structure, and put it in the ipc_id structure. Actually I did this way at the first time, then I agreed with you that moving the per_id lock into there kern_ipc_perm structure will help reduce cacheline bouncing.

I think that having the per_id lock stay out of the structure it protects will easy the job of ipc_rmid(), also will avoid the wrong preempt count problem caused by the additional check "if (out)" in ipc_unlock() as you mentioned above.

Is this solution looks good to you? If so, I will update the patch for 2.5.44 soon.

Mingming

Sorry, no, doesn't look good to me. I do agree that it's an easy solution to this problem, but there's no point in having gone the RCU route to avoid cacheline bouncing, if you now reintroduce it all here.

I believe, as I said, that you'll have to go further, applying RCU to freeing the entries themselves. I had toyed with the idea of never freeing entries once allocated, which is a similarly simple solution; rejected that as in general too wasteful; and concluded that RCU is a reasonable compromise - it amounts to lazy freeing, I think.

I'm happy to be overruled by someone who understands these cacheline bounce issues better than we do, but nobody else has spoken up yet.

Hugh

Hi Andrew,

Here is the updated ipc lock patch:

- It greatly reduces the lock contention by having one lock per id. The global spinlock is removed and a spinlock is added in kern_ipc_perm structure.

- Uses ReadCopyUpdate in grow_ary() for locking-free resizing.

- In the places where ipc_rmid() is called, delay calling ipc_free() to RCU callbacks. This is to prevent ipc_lock() returning an invalid pointer after ipc_rmid(). In addition, use the workqueue to enable RCU freeing vmalloced entries.

Also some other changes:
- Remove redundant ipc_lockall/ipc_unlockall
- Now ipc_unlock() directly takes IPC ID pointer as argument, avoid extra looking up the array.

The changes are made based on the input from Huge Dickens, Manfred Spraul and Dipankar Sarma. In addition, Cliff White has run OSDL's dbt1 test on a 2 way against the earlier version of this patch. Results shows about 2-6% improvement on the average number of transactions per second. Here is the summary of his tests:

                        2.5.42-mm2      2.5.42-mm2-ipclock

----------------------------------------------------------
Average over 5 runs 85.0 BT 89.8 BT
Std Deviation 5 runs 7.4 BT 1.0 BT

Average over 4 best 88.15 BT 90.2 BT
Std Deviation 4 best 2.8 BT 0.5 BT

Full details of the tests could be found here:
http://www.osdl.org/projects/dbt1prfrns/results/mingming/index.html

patch is against 2.5.44-mm4. Please include or give any feedback.

Thanks,

Mingming Cao

diff -urN 2544-mm4/include/linux/ipc.h 2544-mm4-ipc/include/linux/ipc.h
@@ -56,6 +56,8 @@
 /* used by in-kernel data structures */
 struct kern_ipc_perm
 {
+	spinlock_t	lock;
+	int		deleted;
 	key_t		key;
 	uid_t		uid;
 	gid_t		gid;

diff -urN 2544-mm4/ipc/shm.c 2544-mm4-ipc/ipc/shm.c
@@ -38,9 +38,7 @@
 static struct ipc_ids shm_ids;
 
 #define shm_lock(id)	((struct shmid_kernel*)ipc_lock(&shm_ids,id))
-#define shm_unlock(id)	ipc_unlock(&shm_ids,id)
-#define shm_lockall()	ipc_lockall(&shm_ids)
-#define shm_unlockall()	ipc_unlockall(&shm_ids)
+#define shm_unlock(shp)	ipc_unlock(&(shp)->shm_perm)
 #define shm_get(id)	((struct shmid_kernel*)ipc_get(&shm_ids,id))
 #define shm_buildid(id, seq) \
 	ipc_buildid(&shm_ids, id, seq)

@@ -93,7 +91,7 @@
 	shp->shm_atim = CURRENT_TIME;
 	shp->shm_lprid = current->pid;
 	shp->shm_nattch++;
-	shm_unlock(id);
+	shm_unlock(shp);
 }
 
 /* This is called by fork, once for every shm attach. */

@@ -114,7 +112,7 @@
 {
 	shm_tot -= (shp->shm_segsz + PAGE_SIZE - 1) >> PAGE_SHIFT;
 	shm_rmid (shp->id);
-	shm_unlock(shp->id);
+	shm_unlock(shp);
 	if (!is_file_hugepages(shp->shm_file))
 		shmem_lock(shp->shm_file, 0);
 	fput (shp->shm_file);

@@ -145,7 +143,7 @@
 	   shp->shm_flags & SHM_DEST)
 		shm_destroy (shp);
 	else
-		shm_unlock(id);
+		shm_unlock(shp);
 	up (&shm_ids.sem);
 }
 

@@ -225,7 +223,7 @@
 	else
 		file->f_op = &shm_file_operations;
 	shm_tot += numpages;
-	shm_unlock (id);
+	shm_unlock(shp);
 	return shp->id;
 
 no_id:

@@ -261,7 +259,7 @@
 			err = -EACCES;
 		else
 			err = shm_buildid(id, shp->shm_perm.seq);
-		shm_unlock(id);
+		shm_unlock(shp);
 	}
 	up(&shm_ids.sem);
 

@@ -421,14 +419,12 @@
 
 		memset(&shm_info,0,sizeof(shm_info));
 		down(&shm_ids.sem);
-		shm_lockall();
 		shm_info.used_ids = shm_ids.in_use;
 		shm_get_stat (&shm_info.shm_rss, &shm_info.shm_swp);
 		shm_info.shm_tot = shm_tot;
 		shm_info.swap_attempts = 0;
 		shm_info.swap_successes = 0;
 		err = shm_ids.max_id;
-		shm_unlockall();
 		up(&shm_ids.sem);
 		if(copy_to_user (buf, &shm_info, sizeof(shm_info))) {
 			err = -EFAULT;

@@ -470,7 +466,7 @@
 		tbuf.shm_cpid	= shp->shm_cprid;
 		tbuf.shm_lpid	= shp->shm_lprid;
 		tbuf.shm_nattch	= shp->shm_nattch;
-		shm_unlock(shmid);
+		shm_unlock(shp);
 		if(copy_shmid_to_user (buf, &tbuf, version))
 			err = -EFAULT;
 		else

@@ -505,7 +501,7 @@
 				shmem_lock(shp->shm_file, 0);
 			shp->shm_flags &= ~SHM_LOCKED;
 		}
-		shm_unlock(shmid);
+		shm_unlock(shp);
 		goto out;
 	}
 	case IPC_RMID:

@@ -538,7 +534,7 @@
 			shp->shm_flags |= SHM_DEST;
 			/* Do not find it any more */
 			shp->shm_perm.key = IPC_PRIVATE;
-			shm_unlock(shmid);
+			shm_unlock(shp);
 		} else
 			shm_destroy (shp);
 		up(&shm_ids.sem);

@@ -581,12 +577,12 @@
 
 	err = 0;
 out_unlock_up:
-	shm_unlock(shmid);
+	shm_unlock(shp);
 out_up:
 	up(&shm_ids.sem);
 	goto out;
 out_unlock:
-	shm_unlock(shmid);
+	shm_unlock(shp);
 out:
 	return err;
 }

@@ -646,18 +642,18 @@
 	}
 	err = shm_checkid(shp,shmid);
 	if (err) {
-		shm_unlock(shmid);
+		shm_unlock(shp);
 		goto out;
 	}
 	if (ipcperms(&shp->shm_perm, acc_mode)) {
-		shm_unlock(shmid);
+		shm_unlock(shp);
 		err = -EACCES;
 		goto out;
 	}
 	file = shp->shm_file;
 	size = file->f_dentry->d_inode->i_size;
 	shp->shm_nattch++;
-	shm_unlock(shmid);
+	shm_unlock(shp);
 
 	down_write(&current->mm->mmap_sem);
 	if (addr && !(shmflg & SHM_REMAP)) {

@@ -686,7 +682,7 @@
 	   shp->shm_flags & SHM_DEST)
 		shm_destroy (shp);
 	else
-		shm_unlock(shmid);
+		shm_unlock(shp);
 	up (&shm_ids.sem);
 
 	*raddr = (unsigned long) user_addr;

@@ -764,7 +760,7 @@
 				shp->shm_atim,
 				shp->shm_dtim,
 				shp->shm_ctim);
-			shm_unlock(i);
+			shm_unlock(shp);
 
 			pos += len;
 			if(pos < offset) {

Well I can get you a bit of testing and attention, but I'm afraid my knowledge of the IPC code is negligible.

So to be able to commend this change to Linus I'd have to rely on assurances from people who _do_ understand IPC (Hugh?) and on lots of testing.

So yes, I'll include it, and would solicit success reports from people who are actually exercising that code path, thanks.

DBT1 is really interesting, and I'm glad the OSDL team have put it together. If people would only stop sending me patches I'd be using it ;)

Could someone please help explain the results? Comparing, say,
http://www.osdl.org/projects/dbt1prfrns/results/mingming/run.2cpu.42-mm2.r5/index.html
and
http://www.osdl.org/projects/dbt1prfrns/results/mingming/run.18.r5/index.html

It would appear that 2.5 completely smoked 2.4 on response time, yet the overall bogotransactions/sec is significantly lower. What should we conclude from this?

Also I see:

14.7 minute duration
and
Time for DBT run 19:36

What is the 14.7 minutes referring to?

Also:

	2.5: Time for key creation 1:27
	2.4: Time for key creation 14:24
versus:
	2.5: Time for table creation 16:48
	2.4: Time for table creation 8:58

So it's all rather confusing. Masses of numbers usually _are_ confusing. What really adds tons of value to such an exercise is for the person who ran the test to write up some conclusions. To tell the developers what went well, what went poorly, what areas to focus on, etc. To use your own judgement to tell us what to zoom in on.

Is that something which could be added?

Manfred and I have both reviewed the patch (or the 2.5.44 version) and we both recommend it highly (well, let Manfred speak for himself).

I can't claim great expertise on IPC (never on msg, but some on shm and sem), but (unless there's an error we've missed) there's no change to IPC functionality here - it's an exercise in "self-evidently" better locking (there used to be just one spinlock covering all e.g. sems), with RCU to avoid the dirty cacheline bouncing in earlier version.

And I rarely exercise IPC paths, except when testing if I change
something: I do hope someone else can vouch for it in practice, we believe Mingming has devised a fine patch here.

Hugh

OK, thanks.

So I took a look. Wish I hadn't :( The locking rules in there are outrageously uncommented. You must be brave people.

What about this code?

void ipc_rcu_free(void* ptr, int size)

{
        struct rcu_ipc_free* arg;

        arg = (struct rcu_ipc_free *) kmalloc(sizeof(*arg), GFP_KERNEL);
        if (arg == NULL)
                return;
        arg->ptr = ptr;
        arg->size = size;
        call_rcu(&arg->rcu_head, ipc_free_callback, arg);
}

Are we sure that it's never called under locks?

And it seems that if the kmalloc fails, we decide to leak some memory, yes?

If so it would be better to use GFP_ATOMIC there. Avoids any locking problems and also increases the chance of the allocation succeeding. (With an explanatory comment, naturally :)).

Even better: is it possible to embed the rcu_ipc_free inside the object-to-be-freed? Perhaps not?

Stylistically, it is best to not typecast the return value from kmalloc, btw. You should never typecast the return value of anything which returns a void *, because it weakens your compile-time checking. Example:

	foo *bar = (foo *)zot();

The compiler will swallow that, regardless of what zot() returns. Someone could go and change zot() to return a reiserfs_inode * and you would never know about it. Whereas:

	foo *bar = zot();

Says to the compiler "zot() must return a bar * or a void *", which is much tighter checking, yes?

There is an insane amount of inlining in the ipc code. I couldn't keep my paws off it.

Before:
mnm:/usr/src/25> size ipc/*.o
text data bss dec hex filename

  28346     224     192   28762    705a ipc/built-in.o
   7390      20      64    7474    1d32 ipc/msg.o
  11236      16      64   11316    2c34 ipc/sem.o
   8136     160      64    8360    20a8 ipc/shm.o

1584 0 0 1584 630 ipc/util.o

After:
mnm:/usr/src/25> size ipc/*.o
text data bss dec hex filename

  19274     224     192   19690    4cea ipc/built-in.o
   4846      20      64    4930    1342 ipc/msg.o
   7636      16      64    7716    1e24 ipc/sem.o
   4808     160      64    5032    13a8 ipc/shm.o

1984 0 0 1984 7c0 ipc/util.o

@@ -54,63 +54,11 @@ void* ipc_alloc(int size);
 void ipc_free(void* ptr, int size);
 void ipc_rcu_free(void* arg, int size);
 
-extern inline struct kern_ipc_perm* ipc_get(struct ipc_ids* ids, int id)
-{
-	struct kern_ipc_perm* out;
-	int lid = id % SEQ_MULTIPLIER;
-	if(lid >= ids->size)
-		return NULL;
-	rmb();
-	out = ids->entries[lid].p;
-	return out;
-}
-
-extern inline struct kern_ipc_perm* ipc_lock(struct ipc_ids* ids, int id)
-{
-	struct kern_ipc_perm* out;
-	int lid = id % SEQ_MULTIPLIER;
-
-	rcu_read_lock();
-	if(lid >= ids->size) {
-		rcu_read_unlock();
-		return NULL;
-	}
-	rmb();
-	out = ids->entries[lid].p;
-	if(out == NULL) {
-		rcu_read_unlock();
-		return NULL;
-	}
-	spin_lock(&out->lock);
-	
-	/* ipc_rmid() may have already freed the ID while ipc_lock
-	 * was spinning: here verify that the structure is still valid
-	 */
-	if (out->deleted) {
-		spin_unlock(&out->lock);
-		rcu_read_unlock();
-		return NULL;
-	}
-	return out;
-}
-
-extern inline void ipc_unlock(struct kern_ipc_perm* perm)
-{
-	spin_unlock(&perm->lock);
-	rcu_read_unlock();
-}
-
-extern inline int ipc_buildid(struct ipc_ids* ids, int id, int seq)
-{
-	return SEQ_MULTIPLIER*seq + id;
-}
-
-extern inline int ipc_checkid(struct ipc_ids* ids, struct kern_ipc_perm* ipcp, int uid)
-{
-	if(uid/SEQ_MULTIPLIER != ipcp->seq)
-		return 1;
-	return 0;
-}
+struct kern_ipc_perm* ipc_get(struct ipc_ids* ids, int id);
+struct kern_ipc_perm* ipc_lock(struct ipc_ids* ids, int id);
+void ipc_unlock(struct kern_ipc_perm* perm);
+int ipc_buildid(struct ipc_ids* ids, int id, int seq);
+int ipc_checkid(struct ipc_ids* ids, struct kern_ipc_perm* ipcp, int uid);
 
 void kernel_to_ipc64_perm(struct kern_ipc_perm *in, struct ipc64_perm *out);
 void ipc64_perm_to_ipc_perm(struct ipc64_perm *in, struct ipc_perm *out);

.

Ah, we all like to criticize the lack of comments in others' code.

Yes.

Yes, but why would it fail?
and what do you think should be the alternative?

There are no locking doubts here.
GFP_ATOMIC would _reduce_ the chance of the allocation succeeding:
GFP_KERNEL does include the __GFP_WAIT flag, GFP_ATOMIC does not.

It would certainly be possible (I did suggest it as a maybe), but it's unclear whether it's worthwhile wasting the extra memory longterm like that. Mingming chose not to embed, I see no reason to overrule.

You have too much time on your hands, Andrew :-)

I agree tempting: I thought you might like that in a subsequent patch, yes? Mingming was splitting locks, not doing a cleanup of inlines.

Hugh

slightly offtopic ...

I agree tempting: I thought you might like that in a subsequent patch, yes? Mingming was splitting locks, not doing a cleanup of inlines.

There was a time when "inline" was a very cool tool because it had been judged that the overhead of actually calling a function was just too heinous to contemplate. From comments in this and other discussions, is it safe to say that the pendulum has now swung the other way? I see a lot of people concerned about code size and apparently returning to the axiom of "if you use it more than once, make it a function." Are we as a community coming around to using inlining only on very tight, very critical functions?

Rick

I think so, at least Andrew is championing us in that direction. But I agree.

It somewhere became the notion if the function is small, it automatically should be inlined. I suspect Andrew has even stricter criteria than me (I think super small functions should be inlined) but the general "its only a couple of lines" or "it could be a macro" are not sufficient criterion for inlining.

So, my thoughts on suitable criteria would be:

- used only once and only ever in that one spot (i.e. it really could be part of the caller, but it was pulled out for cleanliness. Keep it inline to not have the cleanliness cause a performance degradation (however small)).

- small functions, where small is so small the function overhead is nearly the same size. Stuff that might not even do anything but return a permutation of an argument, etc.

- very very time critical functions in time critical places

So that removes the previous criteria of "the function is N lines or smaller" where N is some number less than 100 :)

Robert Love

Did you see any place where this is called with lock(s) hold? Maybe there is, but I could not see here. They are called from the functions which are used by IPC code only. Inside IPC there is only spin_lock per ID and sem_undo lock. Both of them are not hold when ipc_rcu_free is called.

yes.

Good point. I agree GFP_ATOMIC fits better here.

Are you saying that have a static RCU header structure in the object-to-be-freed? I think it's possible. It fits well in the rmid case, where the object to be freed is an kern_ipc_perm structure. But for the grow_ary() case, the object to be freed is a array of struct ipc_id, so it need a little bit more changes there. Maybe add a new structure ipc_entries, which include the RCU header structure and the pointer to the entries array. Then have the ipc_ids->entries point to ipc_entries. Just a little concern that this way we added a reference when looking up the IPC ID from the array.

This is a place where a mempool is appropriate. The objects have a "guaranteed to be returned if you wait for long enough" lifecycle.

But Hugh's right here. The chance of the single-page GFP_KERNEL allocation failing is tiny; the probability depending upon the VM-of-the-day. Let's leave it be.

Agreed. Here's my brief audit:

	/* We must read max_id before reading any entries */

This is unacceptable crap, sorry. You *must* allocate the resources required to free the object *at the time you allocate the object*, since freeing must not fail.

Yes, this must be done.

Rusty.

   there are those who do and those who hang on and you don't see too
   many doers quoting their contemporaries.  -- Larry McVoy

Thanks for spending some time reviewing the barriers for me. While I was thinking the reason why a rmb is needed here, I found that maybe we don't need a barrier here at all. Since ipc_findkey()(the code above) and the grow_ary() are both protected by ipc_ids.sem(there missing document for this), so both the max_id and the the entries array seen by ipc_findkey should be the latest one.

Also I think it's safe to remove the rmb() in ipc_get() for the same reason. ipc_get() is only used by shm_get_stat() through shm_get() and is called with the shm_ids.sem protected. (Maybe ipc_get should be removed totally?)

Sure. I will add such comments on all places where barriers are being used. I will do as much as I can to add more comments in the code about what lock/sem are hold before/after the funtion is called.:-)

I thought about embed rcu_ipc_free inside the ipc_ids structure before. But there could be a problem if grow_ary() is called again before the old array associated with the previous grow_ary() has not scheduled to be freed yet. I see a need to do that now, as you made very good point. I will make the changes tomorrow.

Thanks a lot for your comments.

Mingming

In message <[email protected]> you write:

You don't need to allocate it in the object, but you *do* need to fail grow_ary() if you can't allocate it.

I had the same dilemma when I tried to write a generic "kfree_rcu(void *)" last year: you simply can't do it 8(

Cheers,
Rusty.
--
Anyone who quotes me in their sig is an idiot. -- Rusty Russell.

Ahm. No idea who wrote the current locking. But the patch is very nice, it reduces the lock contention without increasing the number of spinlock calls.

--
Manfred

Actually, under memory load GFP_KERNEL will wait for the memory to become available, while GFP_ATOMIC will fail.

Using GFP_ATOMIC here will probably increase the risk of a memory leak.

regards,

Rik

Bravely reimplemented by the knights who say "NIH".
http://www.surriel.com/		http://distro.conectiva.com/
Current spamtrap:  <a href=mailto:"[email protected]">[email protected]</a>

Thanks for your quick feedback. I did LTP tests on it--it passed(well, I saw a failure on shmctl(), but the failure was there since 2.5.43 kernel). I will do more stress tests on it soon.

Mingming

Which shmctl() test is this? To my knowledge, there are no current known issues with shmctl tests. There is however one with sem02 in semctl() that last I heard has been partially fixed in the kernel and still needs to be fixed in glibc. Is that the one you are referring to, or is there really some other shmctl test in LTP that is failing?

Thanks,
Paul Larson

Here is the failure I saw on LTP test. The one failed is /ltp-20020807/testcases/kernel/syscalls/ipc/shmctl/shmctl01

<<<test_start>>>
tag=shmctl01 stime=1035475025
cmdline="shmctl01"
contacts=""
analysis=exit
initiation_status="ok"
<<<test_output>>>

shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    3  FAIL  :  # of attaches is incorrect - 0
shmctl01    4  PASS  :  new mode and change time are correct

<<<execution_status>>>
duration=1 termination_type=exited termination_id=1 corefile=no cutime=0 cstime=0
<<<test_end>>>

I guess you are running it with -i2? I just tried shmctl01 -i2 on a 2.5.44 kernel and did not get this error.
shmctl01 1 PASS : pid, size, # of attaches and mode are correct -
pass #1 shmctl01 0 INFO : shmdt() failed - 22

shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    3  PASS  :  new mode and change time are correct
shmctl01    4  PASS  :  shared memory appears to be removed
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    3  PASS  :  new mode and change time are correct
shmctl01    4  PASS  :  shared memory appears to be removed

If I can find some time, I'll try to grab your patch and see if I can reproduce the error on my machine.

-Paul Larson

No, I did not use -i2.

What I did is just run ./shmctl01

Sorry, Paul. Could you try 2.5.44-mm4? I saw the error on clean 2.5.44-mm4(without my patch). And I remember I saw this on 2.5.42-mm2 also.

Here is what I saw:

[root@elm3b83 shmctl]# ./shmctl01
shmctl01 1 PASS : pid, size, # of attaches and mode are correct -
pass #1

shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    3  FAIL  :  # of attaches is incorrect - 0
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    4  PASS  :  new mode and change time are correct

[root@elm3b83 shmctl]# ./shmctl01 -i2
shmctl01 1 PASS : pid, size, # of attaches and mode are correct -
pass #1

shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    3  FAIL  :  # of attaches is incorrect - 0
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    4  PASS  :  new mode and change time are correct
shmctl01    1  BROK  :  couldn't create the shared memory segment
shmctl01    2  BROK  :  Remaining cases broken
shmctl01    3  BROK  :  Remaining cases broken
shmctl01    4  BROK  :  Remaining cases broken

Maybe I just read it wrong.

I still have my results from testing 2.5.44-mm4, here's a cut and paste from shmctl01:

<<<test_start>>>
tag=shmctl01 stime=1035486589
cmdline="shmctl01"
contacts=""
analysis=exit
initiation_status="ok"
<<<test_output>>>

shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    0  INFO  :  shmdt() failed - 22
shmctl01    1  PASS  :  pid, size, # of attaches and mode are correct -
pass #1
shmctl01    2  PASS  :  pid, size, # of attaches and mode are correct -
pass #2
shmctl01    3  PASS  :  new mode and change time are correct
shmctl01    4  PASS  :  shared memory appears to be removed

<<<execution_status>>>
duration=1 termination_type=exited termination_id=0 corefile=no cutime=0 cstime=1
<<<test_end>>>

I havn't seen this test fail before but I'll be happy to do more testing with your patch to see if I can reproduce it. You may also want to consider updating LTP to the newest version. I'm fairly certain that shmctl01 hasn't been changed since the version you have, but just to be consistent you may want to do that.

Thanks,
Paul Larson

Ha! Sorry about the confusion. I re-install ltp test suites and the error is gone. My old tests must be dirty.

Mingming

These failures are caused by a bug in the ltp test. See the attached patch.

This one is odd. The testcase contains races, but they can only increase # of attaches. Could you strace shmctl01?
The testcase with shmat(), then fork() fails.

--
Manfred

diff -u ltp-orig/testcases/kernel/syscalls/ipc/shmctl/shmctl01.c ltp-20021008/testcases/kernel/syscalls/ipc/shmctl/shmctl01.c
@@ -252,10 +252,12 @@
 
 			if (stat_time == FIRST) {
 				test = set_shmat();
+			} else {
+				test = set_shared;
 			}
 
 			/* do an assignement for fun */
-			(int *)test = i;
+			*(int *)test = i;
 
 			/* pause until we get a signal from stat_cleanup() */
 			rval = pause();

@@ -273,7 +275,7 @@
 		}
 	}
 	/* sleep briefly to ensure correct execution order */
-	usleep(25000);
+	usleep(250000);
 }
 
 /*

Thank you very much for that :)

Whoa - we ran these 5 times for an average. The 2.5 run you picked was the 'off' run -
It has the worse results. You will notice on this run, there are a large number of errors
which didn't happen on the other runs - this lowered the BT/sec number. Use one of the
other 2.5 ones and you'll see something more sensible. ( say, 42-mm2.r3) Unfortunately,
on average, 2.4 still beats 2.5 on both response time and BT's

2.5.42-mm2 2.5.42-mm2-ipclock 2.4.18

 Average over 5 runs     85.0 BT           89.8 BT          96.92 BT
 Std Deviation 5 runs    7.4  BT           1.0 BT           2.07 BT
 Average of best 4 runs  88.15 BT          90.2 BT          97.2 BT
 Std Deviation 4 run     2.8 BT            0.5 BT            2.3 BT 

One other place to start comparing is in the system information which is at the bottom of the page.
Some points (might be minor) :
Cpu statistics:
2.4.18 - cpu %idle averages around 1.5% %system swings between 3-7% %nice
steady at ~3.6%

	2.5.42-mm2 cpu %idle 0.0 all thru run, %system steady at ~6% % nice up ~5.5
Swap (sar -r) 
	Very slight differences - we consumed ~98% of the memory in both cases, 2.4 
swapped a little
		bit (%28) more than 2.5 (%26) 

We also include profile data for both the load and run phase. (profile=2)

The 14.7 minute time comes from the workload driver log, which are parsed to get the
response numbers. The 'Time for' stamps come from the master driver script, and include some
of the workload startup and shutdown time. The workload driver waits a bit to be sure things are
stable, before the official run data is collected. The script timestamp waits until the run clients are
dead. So there's always a bit of a delta between the two.

This is a Mystery Question - we don't have an answer, we were hoping _you would see something :)
Table creation involves sequential inserts of data from a flat file to an SAPDB B-tree on a devspace.
Our devspace is a raw device, so we're doing raw io, plus some processing. This op is write-intensive
'Key creation' is establishing a foreign key column contraint on various tables. For each table, it examines every row in the table, looks up (does a B-tree index lookup) the column value in a second table to find a specific primary key that matches the
column value in the first table. So again, some I/O, a bit of processing. Key creation (foreign key) is read-intensive.
Also interesting is the delta in index creation:
2.5 Time for index creation 27:58
2.4 Time for index creation 17:21
Index creation requires a read of the table, a sort, then creation of a B-tree index. Both the index and
table creates build a B-tree for SAP-DB ( both run slower on 2.5 ) - the table creation does no sorting.
We also notice that the times for both index and key creation varies a bit more across runs with the -mm2 kernel,
as shown by the standard deviation across the runs.
mingming and 2.4.18 are a bit more consistent. ( we threw out -mm2 run 5 for this average, due to the errors)

Results are: average time[std dev]
Action 2.4.18 2.5.42-mm2 2.5.42-mm2-ipclock
table create 8:55 [0:04] 19:03 [2:40] 19.39 [0:50]

index create     17:17 [0:11]  25:19 [5:31]    28:05 [0:02]
key create       14:23 [0:16]  15:21 [6:37]    18:46 [0:17]

Also interesting is -mm2 run2 - foreign key creation took 5:26, the run completed with no errors...why so fast, only one time?
It is an ongoing mystery. We Just Don't Know Why Right Now.
We are working on better data capture of db/run errors, and we'd love to hear suggestions
on improving the instrumentation.

Yes, agreed. We don't yet know enough to map from test results to an exact kernel area.
We just added a database expert to staff (Mary Edie Meredith) so we intend to get better.
We'll probably be nagging you a bit, and again we very much appreciate all suggestions.

To

It is something we are working on adding.
cliffw

Andrew,

I tested Mingming's RCU ipc lock patch using a *new* microbenchmark - semopbench. semopbench was written to test the performance of Mingming's patch. I also ran a 3 hour stress and it completed successfully.

Explanation of the microbenchmark is below the results. Here is a link to the microbenchmark source.

http://www-124.ibm.com/developerworks/opensource/linuxperf/semopbench/semopbench.c

SUT : 8-way 700 Mhz PIII

I tested 2.5.44-mm2 and 2.5.44-mm2 + RCU ipc patch

The metric is seconds / per repetition. Lower is better.

kernel              run 1     run 2
                    seconds   seconds
==================  =======   =======

2.5.44-mm2 515.1 515.4
2.5.44-mm2+rcu-ipc 46.7 46.7

With Mingming's patch, the test completes 10X faster.

-----

2.4.44-mm2 readprofile shows 70 % of 8 CPUs spinning on .text.lock.sem :

http://www-124.ibm.com/developerworks/opensource/linuxperf/semopbench/sem.profile.1.out

2.5.44-mm2 + Mingming's patch shows that the spin on .text.lock.sem is gone :

http://www-124.ibm.com/developerworks/opensource/linuxperf/semopbench/sem.rcu.profile.1.out

Here is the semopbench results for 2.5.44-mm2 :

http://www-124.ibm.com/developerworks/opensource/linuxperf/semopbench/sem.results.1.out

Here is the semopbench results for 2.5.44-mm2 + Mingming's patch :

http://www-124.ibm.com/developerworks/opensource/linuxperf/semopbench/sem.rcu.results.1.out

-----

Here is some info on how the microbenchmark works :

-g 64 creates 64 sema4 groups

group0
group1
...
group63

-s 16 creates 16 sema4s in each group

group0 - sem0, sem1, ... sem15
group1 - sem0, sem1, ... sem15
...
group63 - sem0, sem1, ... sem15

For each of the 1024 (64*16) sema4s, a process is forked and sleeps on it's own sema4. When the test starts, the master process will post the sema4 for the 1st process in each group.

When the 1st process in each group wakes up it will :

(a) resets it's own sema4
(b) post the sema4 for the next process in the group
(c) waits on his own sema4

-n 16384 runs through each sema4 group in the above manner 16384 times.

semopbench reports :

(1) average microseconds that it takes each process to complete repetitions.
(2) CPU utilization

-d turns on debug printfs
-v turns on per process times.
-r does a readprofile -r , reset of the profile buffer before test starts

Bill Hartner

IBM Linux Technology Center Performance Team
http://www-124.ibm.com/developerworks/oss/linux
[email protected]

I took a quick look at the original ipc code and I don't understand something - it seems to me the ipc_ids structs are protected by the semaphore inside for all operations, so why do we need the spinlock in the first place ? Am I missing something here ?

Thanks

Dipankar Sarma  <[email protected]> http://lse.sourceforge.net
Linux Technology Center, IBM Software Lab, Bangalore, India.

I made that mistake too at first, Mingming set me straight. Many of the entry points down() the ipc_ids.sem semaphore, but the most significant ones do not. ipc/sem.c is probably the best example (if confusing, since it involves quite different meanings of semaphore):
sys_semop() is the frequent, fast entry point, uses sem_lock without down.

Hugh

The semaphore is used to protect the fields in ipc_ids structure, while the spinlock is used to protect IPC ids. For the current implementation, there is one spinlock for all IPC ids of the same type(i.e. for all messages queues). The patch is intend to breaks down the global spinlock and have a lock per IPC id.

Well, if the semaphore is grabbed then the critical section is serialized including accessing of IPC ids, so there would be no need to have a separate spinlock for the IPC ids. Hugh pointed out the right reason, semaphore IPCs use the spinlock for serialization in many paths, not the semaphore in ipc_ids.

Hugh's point is right - you have two data structures to protect - ipc_id arrays and the kern_ipc_perms attached to it. I would use a single spinlock in ipc_ids to serialize all the updates and use RCU for both to implement safe lockfree lookups. For the second, you would probably want to add a rcu_head to struct kern_ipc_perms and an RCU callback to free it.

Thanks

Dipankar Sarma  <[email protected]> http://lse.sourceforge.net
Linux Technology Center, IBM Software Lab, Bangalore, India.

Yes the RCU patch author Dipankar has already reviewed the memory barriers in ipclock patch.

Oh, right. grow_ary does not need spinlock_t ary anymore.:-)