afs: Overhaul permit caching

Overhaul permit caching in AFS by making it per-vnode and sharing permit
lists where possible.

When most of the fileserver operations are called, they return a status
structure indicating the (revised) details of the vnode or vnodes involved
in the operation.  This includes the access mark derived from the ACL
(named CallerAccess in the protocol definition file).  This is cacheable
and if the ACL changes, the server will tell us that it is breaking the
callback promise, at which point we can discard the currently cached
permits.

With this patch, the afs_permits structure has, at the end, an array of
{ key, CallerAccess } elements, sorted by key pointer.  This is then cached
in a hash table so that it can be shared between vnodes with the same
access permits.

Permit lists can only be shared if they contain the exact same set of
key->CallerAccess mappings.

Note that that table is global rather than being per-net_ns.  If the keys
in a permit list cross net_ns boundaries, there is no problem sharing the
cached permits, since the permits are just integer masks.

Since permit lists pin keys, the permit cache also makes it easier for a
future patch to find all occurrences of a key and remove them by means of
setting the afs_permits::invalidated flag and then clearing the appropriate
key pointer.  In such an event, memory barriers will need adding.

Lastly, the permit caching is skipped if the server has sent either a
vnode-specific or an entire-server callback since the start of the
operation.

Signed-off-by: David Howells <dhowells@redhat.com>

1 files changed, 217 insertions, 150 deletions

diff --git a/fs/afs/security.c b/fs/afs/security.c
index 7cc61c8..1b5198f 100644
--- a/fs/afs/security.c
+++ b/fs/afs/security.c
@@ -1,6 +1,6 @@
 /* AFS security handling
  *
- * Copyright (C) 2007 Red Hat, Inc. All Rights Reserved.
+ * Copyright (C) 2007, 2017 Red Hat, Inc. All Rights Reserved.
  * Written by David Howells (dhowells@redhat.com)
  *
  * This program is free software; you can redistribute it and/or
@@ -14,9 +14,13 @@
 #include <linux/fs.h>
 #include <linux/ctype.h>
 #include <linux/sched.h>
+#include <linux/hashtable.h>
 #include <keys/rxrpc-type.h>
 #include "internal.h"
 
+static DEFINE_HASHTABLE(afs_permits_cache, 10);
+static DEFINE_SPINLOCK(afs_permits_lock);
+
 /*
  * get a key
  */
@@ -46,168 +50,233 @@ struct key *afs_request_key(struct afs_cell *cell)
 }
 
 /*
- * dispose of a permits list
+ * Dispose of a list of permits.
  */
-void afs_zap_permits(struct rcu_head *rcu)
+static void afs_permits_rcu(struct rcu_head *rcu)
 {
 	struct afs_permits *permits =
 		container_of(rcu, struct afs_permits, rcu);
-	int loop;
-
-	_enter("{%d}", permits->count);
+	int i;
 
-	for (loop = permits->count - 1; loop >= 0; loop--)
-		key_put(permits->permits[loop].key);
+	for (i = 0; i < permits->nr_permits; i++)
+		key_put(permits->permits[i].key);
 	kfree(permits);
 }
 
 /*
- * dispose of a permits list in which all the key pointers have been copied
+ * Discard a permission cache.
  */
-static void afs_dispose_of_permits(struct rcu_head *rcu)
+void afs_put_permits(struct afs_permits *permits)
 {
-	struct afs_permits *permits =
-		container_of(rcu, struct afs_permits, rcu);
-
-	_enter("{%d}", permits->count);
-
-	kfree(permits);
+	if (permits && refcount_dec_and_test(&permits->usage)) {
+		spin_lock(&afs_permits_lock);
+		hash_del_rcu(&permits->hash_node);
+		spin_unlock(&afs_permits_lock);
+		call_rcu(&permits->rcu, afs_permits_rcu);
+	}
 }
 
 /*
- * get the authorising vnode - this is the specified inode itself if it's a
- * directory or it's the parent directory if the specified inode is a file or
- * symlink
- * - the caller must release the ref on the inode
+ * Clear a permit cache on callback break.
  */
-static struct afs_vnode *afs_get_auth_inode(struct afs_vnode *vnode,
-					    struct key *key)
+void afs_clear_permits(struct afs_vnode *vnode)
 {
-	struct afs_vnode *auth_vnode;
-	struct inode *auth_inode;
-
-	_enter("");
+	struct afs_permits *permits;
 
-	if (S_ISDIR(vnode->vfs_inode.i_mode)) {
-		auth_inode = igrab(&vnode->vfs_inode);
-		ASSERT(auth_inode != NULL);
-	} else {
-		auth_inode = afs_iget(vnode->vfs_inode.i_sb, key,
-				      &vnode->status.parent, NULL, NULL);
-		if (IS_ERR(auth_inode))
-			return ERR_CAST(auth_inode);
-	}
+	spin_lock(&vnode->lock);
+	permits = rcu_dereference_protected(vnode->permit_cache,
+					    lockdep_is_held(&vnode->lock));
+	RCU_INIT_POINTER(vnode->permit_cache, NULL);
+	vnode->cb_break++;
+	spin_unlock(&vnode->lock);
 
-	auth_vnode = AFS_FS_I(auth_inode);
-	_leave(" = {%x}", auth_vnode->fid.vnode);
-	return auth_vnode;
+	if (permits)
+		afs_put_permits(permits);
 }
 
 /*
- * clear the permit cache on a directory vnode
+ * Hash a list of permits.  Use simple addition to make it easy to add an extra
+ * one at an as-yet indeterminate position in the list.
  */
-void afs_clear_permits(struct afs_vnode *vnode)
+static void afs_hash_permits(struct afs_permits *permits)
 {
-	struct afs_permits *permits;
+	unsigned long h = permits->nr_permits;
+	int i;
 
-	_enter("{%x:%u}", vnode->fid.vid, vnode->fid.vnode);
-
-	mutex_lock(&vnode->permits_lock);
-	permits = vnode->permits;
-	RCU_INIT_POINTER(vnode->permits, NULL);
-	vnode->cb_break++;
-	mutex_unlock(&vnode->permits_lock);
+	for (i = 0; i < permits->nr_permits; i++) {
+		h += (unsigned long)permits->permits[i].key / sizeof(void *);
+		h += permits->permits[i].access;
+	}
 
-	if (permits)
-		call_rcu(&permits->rcu, afs_zap_permits);
-	_leave("");
+	permits->h = h;
 }
 
 /*
- * add the result obtained for a vnode to its or its parent directory's cache
- * for the key used to access it
+ * Cache the CallerAccess result obtained from doing a fileserver operation
+ * that returned a vnode status for a particular key.  If a callback break
+ * occurs whilst the operation was in progress then we have to ditch the cache
+ * as the ACL *may* have changed.
  */
-void afs_cache_permit(struct afs_vnode *vnode, struct key *key, long acl_order)
+void afs_cache_permit(struct afs_vnode *vnode, struct key *key,
+		      unsigned int cb_break)
 {
-	struct afs_permits *permits, *xpermits;
-	struct afs_permit *permit;
-	struct afs_vnode *auth_vnode;
-	int count, loop;
+	struct afs_permits *permits, *xpermits, *replacement, *new = NULL;
+	afs_access_t caller_access = READ_ONCE(vnode->status.caller_access);
+	size_t size = 0;
+	bool changed = false;
+	int i, j;
+
+	_enter("{%x:%u},%x,%x",
+	       vnode->fid.vid, vnode->fid.vnode, key_serial(key), caller_access);
+
+	rcu_read_lock();
+
+	/* Check for the common case first: We got back the same access as last
+	 * time we tried and already have it recorded.
+	 */
+	permits = rcu_dereference(vnode->permit_cache);
+	if (permits) {
+		if (!permits->invalidated) {
+			for (i = 0; i < permits->nr_permits; i++) {
+				if (permits->permits[i].key < key)
+					continue;
+				if (permits->permits[i].key > key)
+					break;
+				if (permits->permits[i].access != caller_access) {
+					changed = true;
+					break;
+				}
 
-	_enter("{%x:%u},%x,%lx",
-	       vnode->fid.vid, vnode->fid.vnode, key_serial(key), acl_order);
+				if (cb_break != (vnode->cb_break +
+						 vnode->cb_interest->server->cb_s_break)) {
+					changed = true;
+					break;
+				}
 
-	auth_vnode = afs_get_auth_inode(vnode, key);
-	if (IS_ERR(auth_vnode)) {
-		_leave(" [get error %ld]", PTR_ERR(auth_vnode));
-		return;
-	}
+				/* The cache is still good. */
+				rcu_read_unlock();
+				return;
+			}
+		}
 
-	mutex_lock(&auth_vnode->permits_lock);
+		changed |= permits->invalidated;
+		size = permits->nr_permits;
 
-	/* guard against a rename being detected whilst we waited for the
-	 * lock */
-	if (memcmp(&auth_vnode->fid, &vnode->status.parent,
-		   sizeof(struct afs_fid)) != 0) {
-		_debug("renamed");
-		goto out_unlock;
+		/* If this set of permits is now wrong, clear the permits
+		 * pointer so that no one tries to use the stale information.
+		 */
+		if (changed) {
+			spin_lock(&vnode->lock);
+			if (permits != rcu_access_pointer(vnode->permit_cache))
+				goto someone_else_changed_it_unlock;
+			RCU_INIT_POINTER(vnode->permit_cache, NULL);
+			spin_unlock(&vnode->lock);
+
+			afs_put_permits(permits);
+			permits = NULL;
+			size = 0;
+		}
 	}
 
-	/* have to be careful as the directory's callback may be broken between
-	 * us receiving the status we're trying to cache and us getting the
-	 * lock to update the cache for the status */
-	if (auth_vnode->acl_order - acl_order > 0) {
-		_debug("ACL changed?");
-		goto out_unlock;
+	if (cb_break != (vnode->cb_break + vnode->cb_interest->server->cb_s_break)) {
+		rcu_read_unlock();
+		goto someone_else_changed_it;
 	}
 
-	/* always update the anonymous mask */
-	_debug("anon access %x", vnode->status.anon_access);
-	auth_vnode->status.anon_access = vnode->status.anon_access;
-	if (key == vnode->volume->cell->anonymous_key)
-		goto out_unlock;
-
-	xpermits = auth_vnode->permits;
-	count = 0;
-	if (xpermits) {
-		/* see if the permit is already in the list
-		 * - if it is then we just amend the list
-		 */
-		count = xpermits->count;
-		permit = xpermits->permits;
-		for (loop = count; loop > 0; loop--) {
-			if (permit->key == key) {
-				permit->access_mask =
-					vnode->status.caller_access;
-				goto out_unlock;
+	/* We need a ref on any permits list we want to copy as we'll have to
+	 * drop the lock to do memory allocation.
+	 */
+	if (permits && !refcount_inc_not_zero(&permits->usage)) {
+		rcu_read_unlock();
+		goto someone_else_changed_it;
+	}
+
+	rcu_read_unlock();
+
+	/* Speculatively create a new list with the revised permission set.  We
+	 * discard this if we find an extant match already in the hash, but
+	 * it's easier to compare with memcmp this way.
+	 *
+	 * We fill in the key pointers at this time, but we don't get the refs
+	 * yet.
+	 */
+	size++;
+	new = kzalloc(sizeof(struct afs_permits) +
+		      sizeof(struct afs_permit) * size, GFP_NOFS);
+	if (!new)
+		return;
+
+	refcount_set(&new->usage, 1);
+	new->nr_permits = size;
+	i = j = 0;
+	if (permits) {
+		for (i = 0; i < permits->nr_permits; i++) {
+			if (j == i && permits->permits[i].key > key) {
+				new->permits[j].key = key;
+				new->permits[j].access = caller_access;
+				j++;
 			}
-			permit++;
+			new->permits[j].key = permits->permits[i].key;
+			new->permits[j].access = permits->permits[i].access;
+			j++;
 		}
 	}
 
-	permits = kmalloc(sizeof(*permits) + sizeof(*permit) * (count + 1),
-			  GFP_NOFS);
-	if (!permits)
-		goto out_unlock;
-
-	if (xpermits)
-		memcpy(permits->permits, xpermits->permits,
-			count * sizeof(struct afs_permit));
-
-	_debug("key %x access %x",
-	       key_serial(key), vnode->status.caller_access);
-	permits->permits[count].access_mask = vnode->status.caller_access;
-	permits->permits[count].key = key_get(key);
-	permits->count = count + 1;
-
-	rcu_assign_pointer(auth_vnode->permits, permits);
-	if (xpermits)
-		call_rcu(&xpermits->rcu, afs_dispose_of_permits);
-
-out_unlock:
-	mutex_unlock(&auth_vnode->permits_lock);
-	iput(&auth_vnode->vfs_inode);
-	_leave("");
+	if (j == i) {
+		new->permits[j].key = key;
+		new->permits[j].access = caller_access;
+	}
+
+	afs_hash_permits(new);
+
+	afs_put_permits(permits);
+
+	/* Now see if the permit list we want is actually already available */
+	spin_lock(&afs_permits_lock);
+
+	hash_for_each_possible(afs_permits_cache, xpermits, hash_node, new->h) {
+		if (xpermits->h != new->h ||
+		    xpermits->invalidated ||
+		    xpermits->nr_permits != new->nr_permits ||
+		    memcmp(xpermits->permits, new->permits,
+			   new->nr_permits * sizeof(struct afs_permit)) != 0)
+			continue;
+
+		if (refcount_inc_not_zero(&xpermits->usage)) {
+			replacement = xpermits;
+			goto found;
+		}
+
+		break;
+	}
+
+	for (i = 0; i < new->nr_permits; i++)
+		key_get(new->permits[i].key);
+	hash_add_rcu(afs_permits_cache, &new->hash_node, new->h);
+	replacement = new;
+	new = NULL;
+
+found:
+	spin_unlock(&afs_permits_lock);
+
+	kfree(new);
+
+	spin_lock(&vnode->lock);
+	if (cb_break != (vnode->cb_break + vnode->cb_interest->server->cb_s_break) ||
+	    permits != rcu_access_pointer(vnode->permit_cache))
+		goto someone_else_changed_it_unlock;
+	rcu_assign_pointer(vnode->permit_cache, replacement);
+	spin_unlock(&vnode->lock);
+	afs_put_permits(permits);
+	return;
+
+someone_else_changed_it_unlock:
+	spin_unlock(&vnode->lock);
+someone_else_changed_it:
+	/* Someone else changed the cache under us - don't recheck at this
+	 * time.
+	 */
+	return;
 }
 
 /*
@@ -219,55 +288,45 @@ static int afs_check_permit(struct afs_vnode *vnode, struct key *key,
 			    afs_access_t *_access)
 {
 	struct afs_permits *permits;
-	struct afs_permit *permit;
-	struct afs_vnode *auth_vnode;
-	bool valid;
-	int loop, ret;
+	bool valid = false;
+	int i, ret;
 
 	_enter("{%x:%u},%x",
 	       vnode->fid.vid, vnode->fid.vnode, key_serial(key));
 
-	auth_vnode = afs_get_auth_inode(vnode, key);
-	if (IS_ERR(auth_vnode)) {
-		*_access = 0;
-		_leave(" = %ld", PTR_ERR(auth_vnode));
-		return PTR_ERR(auth_vnode);
-	}
-
-	ASSERT(S_ISDIR(auth_vnode->vfs_inode.i_mode));
+	permits = vnode->permit_cache;
 
 	/* check the permits to see if we've got one yet */
-	if (key == auth_vnode->volume->cell->anonymous_key) {
+	if (key == vnode->volume->cell->anonymous_key) {
 		_debug("anon");
-		*_access = auth_vnode->status.anon_access;
+		*_access = vnode->status.anon_access;
 		valid = true;
 	} else {
-		valid = false;
 		rcu_read_lock();
-		permits = rcu_dereference(auth_vnode->permits);
+		permits = rcu_dereference(vnode->permit_cache);
 		if (permits) {
-			permit = permits->permits;
-			for (loop = permits->count; loop > 0; loop--) {
-				if (permit->key == key) {
-					_debug("found in cache");
-					*_access = permit->access_mask;
-					valid = true;
+			for (i = 0; i < permits->nr_permits; i++) {
+				if (permits->permits[i].key < key)
+					continue;
+				if (permits->permits[i].key > key)
 					break;
-				}
-				permit++;
+
+				*_access = permits->permits[i].access;
+				valid = !permits->invalidated;
+				break;
 			}
 		}
 		rcu_read_unlock();
 	}
 
 	if (!valid) {
-		/* check the status on the file we're actually interested in
-		 * (the post-processing will cache the result on auth_vnode) */
+		/* Check the status on the file we're actually interested in
+		 * (the post-processing will cache the result).
+		 */
 		_debug("no valid permit");
 
-		ret = afs_vnode_fetch_status(vnode, auth_vnode, key, true);
+		ret = afs_vnode_fetch_status(vnode, key, true);
 		if (ret < 0) {
-			iput(&auth_vnode->vfs_inode);
 			*_access = 0;
 			_leave(" = %d", ret);
 			return ret;
@@ -275,7 +334,6 @@ static int afs_check_permit(struct afs_vnode *vnode, struct key *key,
 		*_access = vnode->status.caller_access;
 	}
 
-	iput(&auth_vnode->vfs_inode);
 	_leave(" = 0 [access %x]", *_access);
 	return 0;
 }
@@ -360,3 +418,12 @@ error:
 	_leave(" = %d", ret);
 	return ret;
 }
+
+void __exit afs_clean_up_permit_cache(void)
+{
+	int i;
+
+	for (i = 0; i < HASH_SIZE(afs_permits_cache); i++)
+		WARN_ON_ONCE(!hlist_empty(&afs_permits_cache[i]));
+
+}