Skip to contentSkip to conversation

procfs: Improve Scaling in proc

PATCH RESEND Superseded Nathan Zimmer 3 participants 7 messages Feb 15, 2013 – Mar 4, 2013

Subsystems (3)

v3 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

PATCH RESEND v3 1/1 procfs: Improve Scaling in proc

I am currently tracking a hotlock reported by a customer on a large system, 512 cores. I am currently running 3.8-rc7 but the issue looks like it has been this way for a very long time.
The offending lock is proc_dir_entry->pde_unload_lock.

This patch converts the replaces the lock with the rcu. However the pde_openers list still is controlled by a spin lock. I tested on a 4096 machine and the lock doesn't seem hot at least according to perf.

This is a refresh/resend of what was orignally suggested by Eric Dumazet some time ago.

Supporting numbers, lower is better, they are from the test I posted earlier.

cpuinfo baseline        Rcu
tasks   read-sec        read-sec
1       0.0141          0.0141
2       0.0140          0.0142
4       0.0140          0.0141
8       0.0145          0.0140
16      0.0553          0.0168
32      0.1688          0.0549
64      0.5017          0.1690
128     1.7005          0.5038
256     5.2513          2.0804
512     8.0529          3.0162
Cc Andrew Morton <[email protected]>
Cc "Eric W. Biederman" <[email protected]>
Cc Eric Dumazet <[email protected]>
Cc Alexander Viro <[email protected]>
Cc David Woodhouse <[email protected]>
Cc Alexey Dobriyan <[email protected]>
Cc "Paul E. McKenney" <[email protected]>
---
 fs/proc/generic.c       |  58 +++++++++---------
 fs/proc/inode.c         | 156 ++++++++++++++++++++++++++----------------------
 include/linux/proc_fs.h |   6 +-
 3 files changed, 118 insertions(+), 102 deletions(-)
fs/proc/generic.c +29 -29
index 76ddae8..6896a70 100644
@@ -191,13 +191,16 @@ proc_file_read(struct file *file, char __user *buf, size_t nbytes,
 	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
 	ssize_t rv = -EIO;
 
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	const struct file_operations *fops;
+
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
+	if (!fops) {
+		rcu_read_unlock();
 		return rv;
 	}
-	pde->pde_users++;
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	rcu_read_unlock();
 
 	rv = __proc_file_read(file, buf, nbytes, ppos);
 
@@ -213,13 +216,16 @@ proc_file_write(struct file *file, const char __user *buffer,
 	ssize_t rv = -EIO;
 
 	if (pde->write_proc) {
-		spin_lock(&pde->pde_unload_lock);
-		if (!pde->proc_fops) {
-			spin_unlock(&pde->pde_unload_lock);
+		const struct file_operations *fops;
+
+		rcu_read_lock();
+		fops = rcu_dereference(pde->proc_fops);
+		if (!fops) {
+			rcu_read_unlock();
 			return rv;
 		}
-		pde->pde_users++;
-		spin_unlock(&pde->pde_unload_lock);
+		atomic_inc(&pde->pde_users);
+		rcu_read_unlock();
 
 		/* FIXME: does this routine need ppos?  probably... */
 		rv = pde->write_proc(file, buffer, count, pde->data);
@@ -558,7 +564,7 @@ static int proc_register(struct proc_dir_entry * dir, struct proc_dir_entry * dp
 
 	if (S_ISDIR(dp->mode)) {
 		if (dp->proc_iops == NULL) {
-			dp->proc_fops = &proc_dir_operations;
+			RCU_INIT_POINTER(dp->proc_fops, &proc_dir_operations);
 			dp->proc_iops = &proc_dir_inode_operations;
 		}
 		dir->nlink++;
@@ -567,7 +573,7 @@ static int proc_register(struct proc_dir_entry * dir, struct proc_dir_entry * dp
 			dp->proc_iops = &proc_link_inode_operations;
 	} else if (S_ISREG(dp->mode)) {
 		if (dp->proc_fops == NULL)
-			dp->proc_fops = &proc_file_operations;
+			RCU_INIT_POINTER(dp->proc_fops, &proc_file_operations);
 		if (dp->proc_iops == NULL)
 			dp->proc_iops = &proc_file_inode_operations;
 	}
@@ -620,7 +626,8 @@ static struct proc_dir_entry *__proc_create(struct proc_dir_entry **parent,
 	ent->mode = mode;
 	ent->nlink = nlink;
 	atomic_set(&ent->count, 1);
-	spin_lock_init(&ent->pde_unload_lock);
+	atomic_set(&ent->pde_users, 1);
+	spin_lock_init(&ent->pde_openers_lock);
 	INIT_LIST_HEAD(&ent->pde_openers);
 out:
 	return ent;
@@ -744,7 +751,7 @@ struct proc_dir_entry *proc_create_data(const char *name, umode_t mode,
 	pde = __proc_create(&parent, name, mode, nlink);
 	if (!pde)
 		goto out;
-	pde->proc_fops = proc_fops;
+	rcu_assign_pointer(pde->proc_fops, proc_fops);
 	pde->data = data;
 	if (proc_register(parent, pde) < 0)
 		goto out_free;
@@ -802,37 +809,30 @@ void remove_proc_entry(const char *name, struct proc_dir_entry *parent)
 		return;
 	}
 
-	spin_lock(&de->pde_unload_lock);
 	/*
 	 * Stop accepting new callers into module. If you're
 	 * dynamically allocating ->proc_fops, save a pointer somewhere.
 	 */
-	de->proc_fops = NULL;
-	/* Wait until all existing callers into module are done. */
-	if (de->pde_users > 0) {
-		DECLARE_COMPLETION_ONSTACK(c);
-
-		if (!de->pde_unload_completion)
-			de->pde_unload_completion = &c;
 
-		spin_unlock(&de->pde_unload_lock);
+	rcu_assign_pointer(de->proc_fops, NULL);
+	synchronize_rcu();
+	/* Wait until all existing callers into module are done. */
 
+	DECLARE_COMPLETION_ONSTACK(c);
+	de->pde_unload_completion = &c;
+	if (!atomic_dec_and_test(&de->pde_users))
 		wait_for_completion(de->pde_unload_completion);
 
-		spin_lock(&de->pde_unload_lock);
-	}
-
+	spin_lock(&de->pde_openers_lock);
 	while (!list_empty(&de->pde_openers)) {
 		struct pde_opener *pdeo;
 
 		pdeo = list_first_entry(&de->pde_openers, struct pde_opener, lh);
 		list_del(&pdeo->lh);
-		spin_unlock(&de->pde_unload_lock);
 		pdeo->release(pdeo->inode, pdeo->file);
 		kfree(pdeo);
-		spin_lock(&de->pde_unload_lock);
 	}
-	spin_unlock(&de->pde_unload_lock);
+	spin_unlock(&de->pde_openers_lock);
 
 	if (S_ISDIR(de->mode))
 		parent->nlink--;
fs/proc/inode.c +86 -70

301 lines not on this page · raw message

include/linux/proc_fs.h +3 -3
index 32676b3..82628d9 100644
@@ -68,16 +68,16 @@ struct proc_dir_entry {
 	 * If you're allocating ->proc_fops dynamically, save a pointer
 	 * somewhere.
 	 */
-	const struct file_operations *proc_fops;
+	const struct file_operations __rcu *proc_fops;
 	struct proc_dir_entry *next, *parent, *subdir;
 	void *data;
 	read_proc_t *read_proc;
 	write_proc_t *write_proc;
 	atomic_t count;		/* use count */
-	int pde_users;	/* number of callers into module in progress */
+	atomic_t pde_users;	/* number of callers into module in progress */
 	struct completion *pde_unload_completion;
+	spinlock_t pde_openers_lock;
 	struct list_head pde_openers;	/* who did ->open, but not ->release */
-	spinlock_t pde_unload_lock; /* proc_fops checks and pde_users bumps */
 	u8 namelen;
 	char name[];
 };
Signature · 1 line
1.8.1.2
#1 lore raw
Quoted 33 lines · On Fri, 15 Feb 2013 14:47:54 -0600 Nathan Zimmer <[email protected]> wrote:

I am currently tracking a hotlock reported by a customer on a large system, 512 cores. I am currently running 3.8-rc7 but the issue looks like it has been this way for a very long time.
The offending lock is proc_dir_entry->pde_unload_lock.

This patch converts the replaces the lock with the rcu. However the pde_openers list still is controlled by a spin lock. I tested on a 4096 machine and the lock doesn't seem hot at least according to perf.

This is a refresh/resend of what was orignally suggested by Eric Dumazet some time ago.

Supporting numbers, lower is better, they are from the test I posted earlier.

cpuinfo baseline        Rcu
tasks   read-sec        read-sec
1       0.0141          0.0141
2       0.0140          0.0142
4       0.0140          0.0141
8       0.0145          0.0140
16      0.0553          0.0168
32      0.1688          0.0549
64      0.5017          0.1690
128     1.7005          0.5038
256     5.2513          2.0804
512     8.0529          3.0162

...

fs/proc/generic.c +0 -0
index 76ddae8..6896a70 100644

fs/proc/generic.c proc_file_read(struct file *file, char __user *buf, size_t nbytes, · line 194

Quoted 7 lines
fs/proc/generic.c +1 -3
@@ -191,13 +191,16 @@ proc_file_read(struct file *file, char __user *buf, size_t nbytes,
 	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
 	ssize_t rv = -EIO;
 
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	const struct file_operations *fops;

There's now a stray newline in the definitions section.

fs/proc/generic.c proc_file_read(struct file *file, char __user *buf, size_t nbytes, · line 203

Quoted 10 lines
fs/proc/generic.c +6 -2
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
+	if (!fops) {
+		rcu_read_unlock();
 		return rv;
 	}
-	pde->pde_users++;
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	rcu_read_unlock();

So what's up with pde_users? Seems that it's atomic_t *and* uses a form of RCU protection. We can't make it a plain old integer because it's modified under rcu_read_lock() and we can't move the atomic_inc() outside rcu_read_lock() because of the synchronization games in remove_proc_entry()?

Quoted 18 lines
fs/proc/generic.c +0 -5
 	rv = __proc_file_read(file, buf, nbytes, ppos);
 
 
...
@@ -802,37 +809,30 @@ void remove_proc_entry(const char *name, struct proc_dir_entry *parent)
 		return;
 	}
 
-	spin_lock(&de->pde_unload_lock);
 	/*
 	 * Stop accepting new callers into module. If you're
 	 * dynamically allocating ->proc_fops, save a pointer somewhere.
 	 */
-	de->proc_fops = NULL;
-	/* Wait until all existing callers into module are done. */
-	if (de->pde_users > 0) {
-		DECLARE_COMPLETION_ONSTACK(c);

fs/proc/generic.c void remove_proc_entry(const char *name, struct proc_dir_entry *parent) · line 821

Quoted 10 lines
fs/proc/generic.c +4 -4
-
-		if (!de->pde_unload_completion)
-			de->pde_unload_completion = &c;
 
-		spin_unlock(&de->pde_unload_lock);
+	rcu_assign_pointer(de->proc_fops, NULL);
+	synchronize_rcu();
+	/* Wait until all existing callers into module are done. */
 
+	DECLARE_COMPLETION_ONSTACK(c);

This should have generated a c99-style definition warning. Did your compiler version not do this?

Quoted 46 lines
fs/proc/generic.c +9 -9
+	de->pde_unload_completion = &c;
+	if (!atomic_dec_and_test(&de->pde_users))
 		wait_for_completion(de->pde_unload_completion);
 
-		spin_lock(&de->pde_unload_lock);
-	}
-
+	spin_lock(&de->pde_openers_lock);
 	while (!list_empty(&de->pde_openers)) {
 		struct pde_opener *pdeo;
 
 		pdeo = list_first_entry(&de->pde_openers, struct pde_opener, lh);
 		list_del(&pdeo->lh);
-		spin_unlock(&de->pde_unload_lock);
 		pdeo->release(pdeo->inode, pdeo->file);
 		kfree(pdeo);
-		spin_lock(&de->pde_unload_lock);
 	}
-	spin_unlock(&de->pde_unload_lock);
+	spin_unlock(&de->pde_openers_lock);
 
 	if (S_ISDIR(de->mode))
 		parent->nlink--;
 
...
 
 static loff_t proc_reg_llseek(struct file *file, loff_t offset, int whence)
 {
+	const struct file_operations *fops;
 	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
 	loff_t rv = -EINVAL;
 	loff_t (*llseek)(struct file *, loff_t, int);
 
-	spin_lock(&pde->pde_unload_lock);
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
 	/*
 	 * remove_proc_entry() is going to delete PDE (as part of module
 	 * cleanup sequence). No new callers into module allowed.
 	 */
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	if (!fops) {
+		rcu_read_unlock();
 		return rv;
 	}

fs/proc/generic.c void remove_proc_entry(const char *name, struct proc_dir_entry *parent)

Quoted 10 lines
fs/proc/generic.c +1 -1
 	/*
 	 * Bump refcount so that remove_proc_entry will wail for ->llseek to
 	 * complete.
 	 */
-	pde->pde_users++;
+	atomic_inc(&pde->pde_users);
 	/*
 	 * Save function pointer under lock, to protect against ->proc_fops
 	 * NULL'ifying right after ->pde_unload_lock is dropped.
 	 */

This comment needs updating.

However, it doesn't appear to be true any more. With this patch we no longer set ->fops to NULL in remove_proc_entry(). (What replaced that logic?)

So are all these games with local variable `llseek' still needed? afaict the increment of pde_users will stabilize ->fops?

Quoted 18 lines
fs/proc/generic.c +5 -5
-	llseek = pde->proc_fops->llseek;
-	spin_unlock(&pde->pde_unload_lock);
+	llseek = fops->llseek;
+	rcu_read_unlock();
 
 	if (!llseek)
 		llseek = default_llseek;
@@ -182,15 +176,17 @@ static ssize_t proc_reg_read(struct file *file, char __user *buf, size_t count,
 	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
 	ssize_t rv = -EIO;
 	ssize_t (*read)(struct file *, char __user *, size_t, loff_t *);
+	const struct file_operations *fops;
 
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);

fs/proc/generic.c static ssize_t proc_reg_read(struct file *file, char __user *buf, size_t count, · line 189

Quoted 10 lines
fs/proc/generic.c +5 -3
+	if (!fops) {
+		rcu_read_unlock();
 		return rv;
 	}
-	pde->pde_users++;
-	read = pde->proc_fops->read;
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	read = fops->read;
+	rcu_read_unlock();

Many dittoes.

Quoted 26 lines
fs/proc/generic.c +8 -6
 	if (read)
 		rv = read(file, buf, count, ppos);
@@ -204,15 +200,17 @@ static ssize_t proc_reg_write(struct file *file, const char __user *buf, size_t
 	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
 	ssize_t rv = -EIO;
 	ssize_t (*write)(struct file *, const char __user *, size_t, loff_t *);
+	const struct file_operations *fops;
 
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
+	if (!fops) {
+		rcu_read_unlock();
 		return rv;
 	}
-	pde->pde_users++;
-	write = pde->proc_fops->write;
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	write = fops->write;
+	rcu_read_unlock();
 
...
#2 lore raw
Quoted 39 lines · Andrew Morton
Quoted 36 lines · On Fri, 15 Feb 2013 14:47:54 -0600 Nathan Zimmer <[email protected]> wrote:

I am currently tracking a hotlock reported by a customer on a large system, 512 cores. I am currently running 3.8-rc7 but the issue looks like it has been this way for a very long time.
The offending lock is proc_dir_entry->pde_unload_lock.

This patch converts the replaces the lock with the rcu. However the pde_openers list still is controlled by a spin lock. I tested on a 4096 machine and the lock doesn't seem hot at least according to perf.

This is a refresh/resend of what was orignally suggested by Eric Dumazet some time ago.

Supporting numbers, lower is better, they are from the test I posted earlier.

cpuinfo baseline        Rcu
tasks   read-sec        read-sec
1       0.0141          0.0141
2       0.0140          0.0142
4       0.0140          0.0141
8       0.0145          0.0140
16      0.0553          0.0168
32      0.1688          0.0549
64      0.5017          0.1690
128     1.7005          0.5038
256     5.2513          2.0804
512     8.0529          3.0162

...

fs/proc/generic.c +0 -0
index 76ddae8..6896a70 100644
@@ -191,13 +191,16 @@ proc_file_read(struct file *file, char __user *buf, size_t nbytes,
  	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
  	ssize_t rv = -EIO;
 
Quoted 5 lines

fs/proc/generic.c proc_file_read(struct file *file, char __user *buf, size_t nbytes, · line 194

Quoted 4 lines
fs/proc/generic.c +1 -3
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	const struct file_operations *fops;

There's now a stray newline in the definitions section.

Noted and corrected, in a few places.

Quoted 7 lines
Quoted 7 lines
fs/proc/generic.c +4 -1
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
+	if (!fops) {
+		rcu_read_unlock();
  		return rv;
  	}
-	pde->pde_users++;
Quoted 8 lines

fs/proc/generic.c proc_file_read(struct file *file, char __user *buf, size_t nbytes, · line 203

Quoted 3 lines
fs/proc/generic.c +2 -1
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	rcu_read_unlock();

So what's up with pde_users? Seems that it's atomic_t *and* uses a form of RCU protection. We can't make it a plain old integer because it's modified under rcu_read_lock() and we can't move the atomic_inc() outside rcu_read_lock() because of the synchronization games in remove_proc_entry()?

The intent of pde_users is to let us know when it is safe to clean out the pde_openers. I probably should comment this.

Quoted 22 lines
Quoted 22 lines
fs/proc/generic.c +0 -8
  	rv = __proc_file_read(file, buf, nbytes, ppos);
 
 
...
@@ -802,37 +809,30 @@ void remove_proc_entry(const char *name, struct proc_dir_entry *parent)
  		return;
  	}
 
-	spin_lock(&de->pde_unload_lock);
  	/*
  	 * Stop accepting new callers into module. If you're
  	 * dynamically allocating ->proc_fops, save a pointer somewhere.
  	 */
-	de->proc_fops = NULL;
-	/* Wait until all existing callers into module are done. */
-	if (de->pde_users > 0) {
-		DECLARE_COMPLETION_ONSTACK(c);
-
-		if (!de->pde_unload_completion)
-			de->pde_unload_completion = &c;
 
Quoted 8 lines

fs/proc/generic.c void remove_proc_entry(const char *name, struct proc_dir_entry *parent) · line 821

Quoted 6 lines
fs/proc/generic.c +4 -1
-		spin_unlock(&de->pde_unload_lock);
+	rcu_assign_pointer(de->proc_fops, NULL);
+	synchronize_rcu();
+	/* Wait until all existing callers into module are done. */
 
+	DECLARE_COMPLETION_ONSTACK(c);

This should have generated a c99-style definition warning. Did your compiler version not do this?

A clear over site on my part.

Quoted 58 lines
Quoted 46 lines
fs/proc/generic.c +9 -9
+	de->pde_unload_completion = &c;
+	if (!atomic_dec_and_test(&de->pde_users))
  		wait_for_completion(de->pde_unload_completion);
 
-		spin_lock(&de->pde_unload_lock);
-	}
-
+	spin_lock(&de->pde_openers_lock);
  	while (!list_empty(&de->pde_openers)) {
  		struct pde_opener *pdeo;
 
  		pdeo = list_first_entry(&de->pde_openers, struct pde_opener, lh);
  		list_del(&pdeo->lh);
-		spin_unlock(&de->pde_unload_lock);
  		pdeo->release(pdeo->inode, pdeo->file);
  		kfree(pdeo);
-		spin_lock(&de->pde_unload_lock);
  	}
-	spin_unlock(&de->pde_unload_lock);
+	spin_unlock(&de->pde_openers_lock);
 
  	if (S_ISDIR(de->mode))
  		parent->nlink--;
 
...
 
  static loff_t proc_reg_llseek(struct file *file, loff_t offset, int whence)
  {
+	const struct file_operations *fops;
  	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
  	loff_t rv = -EINVAL;
  	loff_t (*llseek)(struct file *, loff_t, int);
 
-	spin_lock(&pde->pde_unload_lock);
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
  	/*
  	 * remove_proc_entry() is going to delete PDE (as part of module
  	 * cleanup sequence). No new callers into module allowed.
  	 */
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	if (!fops) {
+		rcu_read_unlock();
  		return rv;
  	}

fs/proc/generic.c void remove_proc_entry(const char *name, struct proc_dir_entry *parent)

Quoted 10 lines
fs/proc/generic.c +1 -1
  	/*
  	 * Bump refcount so that remove_proc_entry will wail for ->llseek to
  	 * complete.
  	 */
-	pde->pde_users++;
+	atomic_inc(&pde->pde_users);
  	/*
  	 * Save function pointer under lock, to protect against ->proc_fops
  	 * NULL'ifying right after ->pde_unload_lock is dropped.
  	 */

This comment needs updating.

Quoted 6 lines

However, it doesn't appear to be true any more. With this patch we no longer set ->fops to NULL in remove_proc_entry(). (What replaced that logic?)

So are all these games with local variable `llseek' still needed? afaict the increment of pde_users will stabilize ->fops?

We still are setting de->proc_fops to NULL to prevent new callers.
Also we still have to save fops-> since we cannot use fops outside the rcu_read_un/lock. Unless I misunderstood your question.
But yes the comment needs to be updated.

Quoted 55 lines
Quoted 19 lines
fs/proc/generic.c +5 -5
-	llseek = pde->proc_fops->llseek;
-	spin_unlock(&pde->pde_unload_lock);
+	llseek = fops->llseek;
+	rcu_read_unlock();
 
  	if (!llseek)
  		llseek = default_llseek;
@@ -182,15 +176,17 @@ static ssize_t proc_reg_read(struct file *file, char __user *buf, size_t count,
  	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
  	ssize_t rv = -EIO;
  	ssize_t (*read)(struct file *, char __user *, size_t, loff_t *);
+	const struct file_operations *fops;
 
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);

fs/proc/generic.c static ssize_t proc_reg_read(struct file *file, char __user *buf, size_t count, · line 189

Quoted 10 lines
fs/proc/generic.c +5 -3
+	if (!fops) {
+		rcu_read_unlock();
  		return rv;
  	}
-	pde->pde_users++;
-	read = pde->proc_fops->read;
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	read = fops->read;
+	rcu_read_unlock();

Many dittoes.

Quoted 24 lines
fs/proc/generic.c +8 -6
  	if (read)
  		rv = read(file, buf, count, ppos);
@@ -204,15 +200,17 @@ static ssize_t proc_reg_write(struct file *file, const char __user *buf, size_t
  	struct proc_dir_entry *pde = PDE(file->f_path.dentry->d_inode);
  	ssize_t rv = -EIO;
  	ssize_t (*write)(struct file *, const char __user *, size_t, loff_t *);
+	const struct file_operations *fops;
 
-	spin_lock(&pde->pde_unload_lock);
-	if (!pde->proc_fops) {
-		spin_unlock(&pde->pde_unload_lock);
+	rcu_read_lock();
+	fops = rcu_dereference(pde->proc_fops);
+	if (!fops) {
+		rcu_read_unlock();
  		return rv;
  	}
-	pde->pde_users++;
-	write = pde->proc_fops->write;
-	spin_unlock(&pde->pde_unload_lock);
+	atomic_inc(&pde->pde_users);
+	write = fops->write;
+	rcu_read_unlock();
 
Quoted 2 lines
Quoted 2 lines

...

Thanks,
Nate

#3 lore raw
Quoted 6 lines · On Fri, 15 Feb 2013 17:39:38 -0600 Nathan Zimmer <[email protected]> wrote:
Quoted 2 lines

So are all these games with local variable `llseek' still needed? afaict the increment of pde_users will stabilize ->fops?

We still are setting de->proc_fops to NULL to prevent new callers. Also we still have to save fops-> since we cannot use fops outside the rcu_read_un/lock.
Unless I misunderstood your question.

ah, so we are.

#4 lore raw
In reply to Nathan Zimmer
Quoted 8 lines · Nathan Zimmer

I am currently tracking a hotlock reported by a customer on a large system, 512 cores. I am currently running 3.8-rc7 but the issue looks like it has been this way for a very long time.
The offending lock is proc_dir_entry->pde_unload_lock.

This patch converts the replaces the lock with the rcu. However the pde_openers list still is controlled by a spin lock. I tested on a 4096 machine and the lock doesn't seem hot at least according to perf.

Nathan,

This patch causes RCU errors on my ARM systems (which have 2 or 4 CPUs) during boot; see the log below. Reverting it fixes the problem. Let me know if you need any more information to diagnose the problem. Thanks.

Quoted 3 lines
[ 3.652115] VFS: Mounted root (ext3 filesystem) on device 179:1.
[ 3.664856] devtmpfs: mounted
[ 3.668877] Freeing init memory: 212K

...

Quoted 18 lines
[ 25.765891] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=2102 jiffies, g=4294967050, c=4294967049, q=5749)
[ 25.777786] INFO: Stall ended before state dump start
[ 88.815886] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=8407 jiffies, g=4294967050, c=4294967049, q=5818)
[ 88.827777] INFO: Stall ended before state dump start
[ 151.865886] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=14712 jiffies, g=4294967050, c=4294967049, q=5869)
[ 151.877863] INFO: Stall ended before state dump start
[ 214.915885] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=21017 jiffies, g=4294967050, c=4294967049, q=5869)
[ 214.927857] INFO: Stall ended before state dump start
[ 240.525981] INFO: task mountall:71 blocked for more than 120 seconds.
[ 240.532435] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.541452] INFO: task init:150 blocked for more than 120 seconds.
[ 240.547687] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.556352] INFO: task udevd:151 blocked for more than 120 seconds.
[ 240.562623] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.571262] INFO: task ata_id:197 blocked for more than 120 seconds.
[ 240.577662] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 277.965881] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 0, t=27322 jiffies, g=4294967050, c=4294967049, q=5869)
[ 277.977860] INFO: Stall ended before state dump start 
#5 lore raw

I found and fixed a number of them, with the help of sparse.
http://marc.info/?l=linux-fsdevel&m=136242660212004&w=2
If there are any I didn't find and fix let me know.

________________________________________
From: Stephen Warren [[email protected]]
Sent: Monday, March 04, 2013 2:16 PM
To: Nathan Zimmer
Cc: [email protected]; [email protected]; [email protected]; [email protected]; Andrew Morton; Eric W. Biederman; David Woodhouse; Alexey Dobriyan; Paul E. McKenney
Subject: Re: [PATCH v3 resend] procfs: Improve Scaling in proc

Quoted 8 lines · Nathan Zimmer

I am currently tracking a hotlock reported by a customer on a large system, 512 cores. I am currently running 3.8-rc7 but the issue looks like it has been this way for a very long time.
The offending lock is proc_dir_entry->pde_unload_lock.

This patch converts the replaces the lock with the rcu. However the pde_openers list still is controlled by a spin lock. I tested on a 4096 machine and the lock doesn't seem hot at least according to perf.

Nathan,

This patch causes RCU errors on my ARM systems (which have 2 or 4 CPUs) during boot; see the log below. Reverting it fixes the problem. Let me know if you need any more information to diagnose the problem. Thanks.

Quoted 3 lines
[ 3.652115] VFS: Mounted root (ext3 filesystem) on device 179:1.
[ 3.664856] devtmpfs: mounted
[ 3.668877] Freeing init memory: 212K

...

Quoted 18 lines
[ 25.765891] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=2102 jiffies, g=4294967050, c=4294967049, q=5749)
[ 25.777786] INFO: Stall ended before state dump start
[ 88.815886] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=8407 jiffies, g=4294967050, c=4294967049, q=5818)
[ 88.827777] INFO: Stall ended before state dump start
[ 151.865886] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=14712 jiffies, g=4294967050, c=4294967049, q=5869)
[ 151.877863] INFO: Stall ended before state dump start
[ 214.915885] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=21017 jiffies, g=4294967050, c=4294967049, q=5869)
[ 214.927857] INFO: Stall ended before state dump start
[ 240.525981] INFO: task mountall:71 blocked for more than 120 seconds.
[ 240.532435] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.541452] INFO: task init:150 blocked for more than 120 seconds.
[ 240.547687] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.556352] INFO: task udevd:151 blocked for more than 120 seconds.
[ 240.562623] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.571262] INFO: task ata_id:197 blocked for more than 120 seconds.
[ 240.577662] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 277.965881] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 0, t=27322 jiffies, g=4294967050, c=4294967049, q=5869)
[ 277.977860] INFO: Stall ended before state dump start
#6 lore raw
Quoted 13 lines · Nathan Zimmer

Stephen Warren wrote at Monday, March 04, 2013 2:16 PM:

Quoted 12 lines · Nathan Zimmer
Quoted 8 lines

I am currently tracking a hotlock reported by a customer on a large system, 512 cores. I am currently running 3.8-rc7 but the issue looks like it has been this way for a very long time.
The offending lock is proc_dir_entry->pde_unload_lock.

This patch converts the replaces the lock with the rcu. However the pde_openers list still is controlled by a spin lock. I tested on a 4096 machine and the lock doesn't seem hot at least according to perf.

Nathan,

Quoted 7 lines
Quoted 3 lines

This patch causes RCU errors on my ARM systems (which have 2 or 4 CPUs) during boot; see the log below. Reverting it fixes the problem. Let me know if you need any more information to diagnose the problem. Thanks.

I found and fixed a number of them, with the help of sparse.
http://marc.info/?l=linux-fsdevel&m=136242660212004&w=2
If there are any I didn't find and fix let me know.

Yes, if I revert the version of this patch that's in next-20130304, and replace it with V5 that you linked to above, then the problem is no longer present. Thanks.

Quoted 22 lines
Quoted 3 lines
[ 3.652115] VFS: Mounted root (ext3 filesystem) on device 179:1.
[ 3.664856] devtmpfs: mounted
[ 3.668877] Freeing init memory: 212K

...

Quoted 18 lines
[ 25.765891] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=2102 jiffies, g=4294967050, c=4294967049, q=5749)
[ 25.777786] INFO: Stall ended before state dump start
[ 88.815886] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=8407 jiffies, g=4294967050, c=4294967049, q=5818)
[ 88.827777] INFO: Stall ended before state dump start
[ 151.865886] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=14712 jiffies, g=4294967050, c=4294967049, q=5869)
[ 151.877863] INFO: Stall ended before state dump start
[ 214.915885] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 1, t=21017 jiffies, g=4294967050, c=4294967049, q=5869)
[ 214.927857] INFO: Stall ended before state dump start
[ 240.525981] INFO: task mountall:71 blocked for more than 120 seconds.
[ 240.532435] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.541452] INFO: task init:150 blocked for more than 120 seconds.
[ 240.547687] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.556352] INFO: task udevd:151 blocked for more than 120 seconds.
[ 240.562623] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 240.571262] INFO: task ata_id:197 blocked for more than 120 seconds.
[ 240.577662] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
[ 277.965881] INFO: rcu_preempt detected stalls on CPUs/tasks: {} (detected by 0, t=27322 jiffies, g=4294967050, c=4294967049, q=5869)
[ 277.977860] INFO: Stall ended before state dump start--