Skip to contentSkip to conversation

ceph: fix long stalls on sync/syncfs

PATCH Superseded Jeff Layton 4 participants 9 messages Apr 2, 2020 – Apr 3, 2020 Ceph client

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover ceph: fix long stalls on sync/syncfs

This is v2 of the patch I sent the other day to fix the problem of long stalls when calling sync or syncfs.

This set converts the mdsc->cap_dirty list to a per-session list, and then has the only caller that looks at cap_dirty walk the list of sessions and issue flushes for each session in turn.

With this, we can use an empty s_cap_dirty list as an indicator that the cap flush is the last one going to the session and can mark that one as one we're waiting on so the MDS can expedite it.

This also attempts to clarify some of the locking around s_cap_dirty, and adds a FIXME comment to raise the question about locking around s_cap_flushing.

Jeff Layton (2):
ceph: convert mdsc->cap_dirty to a per-session list
ceph: request expedited service on session's last cap flush

 fs/ceph/caps.c       | 72 +++++++++++++++++++++++++++++++++++++++-----
 fs/ceph/mds_client.c |  2 +-
 fs/ceph/mds_client.h |  5 +--
 fs/ceph/super.h      | 21 +++++++++++--
 4 files changed, 87 insertions(+), 13 deletions(-)
Signature · 1 line
2.25.1
#1 lore raw

PATCH v2 1/2 ceph: convert mdsc->cap_dirty to a per-session list

This is a per-sb list now, but that makes it difficult to tell when the cap is the last dirty one associated with the session. Switch this to be a per-session list, but continue using the mdsc->cap_dirty_lock to protect the lists.

This list is only ever walked in ceph_flush_dirty_caps, so change that to walk the sessions array and then flush the caps for inodes on each session's list.

If the auth cap ever changes while the inode has dirty caps, then move the inode to the appropriate session for the new auth_cap. Also, ensure that we never remove an auth cap while the inode is still on the s_cap_dirty list.

---
 fs/ceph/caps.c       | 64 ++++++++++++++++++++++++++++++++++++++++----
 fs/ceph/mds_client.c |  2 +-
 fs/ceph/mds_client.h |  5 ++--
 fs/ceph/super.h      | 21 ++++++++++++---
 4 files changed, 81 insertions(+), 11 deletions(-)
fs/ceph/caps.c +59 -5
index 61808793e0c0..95c9b25e45a6 100644
@@ -727,6 +727,19 @@ void ceph_add_cap(struct inode *inode,
 	if (flags & CEPH_CAP_FLAG_AUTH) {
 		if (!ci->i_auth_cap ||
 		    ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
+			if (ci->i_auth_cap &&
+			    ci->i_auth_cap->session != cap->session) {
+				/*
+				 * If auth cap session changed, and the cap is
+				 * dirty, move it to the correct session's list
+				 */
+				if (!list_empty(&ci->i_dirty_item)) {
+					spin_lock(&mdsc->cap_dirty_lock);
+					list_move(&ci->i_dirty_item,
+						  &cap->session->s_cap_dirty);
+					spin_unlock(&mdsc->cap_dirty_lock);
+				}
+			}
 			ci->i_auth_cap = cap;
 			cap->mds_wanted = wanted;
 		}
@@ -1123,8 +1136,10 @@ void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
 
 	/* remove from inode's cap rbtree, and clear auth cap */
 	rb_erase(&cap->ci_node, &ci->i_caps);
-	if (ci->i_auth_cap == cap)
+	if (ci->i_auth_cap == cap) {
+		WARN_ON_ONCE(!list_empty(&ci->i_dirty_item));
 		ci->i_auth_cap = NULL;
+	}
 
 	/* remove from session list */
 	spin_lock(&session->s_cap_lock);
@@ -1690,6 +1705,8 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
 	     ceph_cap_string(was | mask));
 	ci->i_dirty_caps |= mask;
 	if (was == 0) {
+		struct ceph_mds_session *session = ci->i_auth_cap->session;
+
 		WARN_ON_ONCE(ci->i_prealloc_cap_flush);
 		swap(ci->i_prealloc_cap_flush, *pcf);
 
@@ -1702,7 +1719,7 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
 		     &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
 		BUG_ON(!list_empty(&ci->i_dirty_item));
 		spin_lock(&mdsc->cap_dirty_lock);
-		list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
+		list_add(&ci->i_dirty_item, &session->s_cap_dirty);
 		spin_unlock(&mdsc->cap_dirty_lock);
 		if (ci->i_flushing_caps == 0) {
 			ihold(inode);
@@ -3753,6 +3770,13 @@ static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
 			if (cap == ci->i_auth_cap)
 				ci->i_auth_cap = tcap;
 
+			if (!list_empty(&ci->i_dirty_item)) {
+				spin_lock(&mdsc->cap_dirty_lock);
+				list_move(&ci->i_dirty_item,
+					  &tcap->session->s_cap_dirty);
+				spin_unlock(&mdsc->cap_dirty_lock);
+			}
+
 			if (!list_empty(&ci->i_cap_flush_list) &&
 			    ci->i_auth_cap == tcap) {
 				spin_lock(&mdsc->cap_dirty_lock);
@@ -4176,15 +4200,16 @@ void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
 /*
  * Flush all dirty caps to the mds
  */
-void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
+static void flush_dirty_session_caps(struct ceph_mds_session *s)
 {
+	struct ceph_mds_client *mdsc = s->s_mdsc;
 	struct ceph_inode_info *ci;
 	struct inode *inode;
 
 	dout("flush_dirty_caps\n");
 	spin_lock(&mdsc->cap_dirty_lock);
-	while (!list_empty(&mdsc->cap_dirty)) {
-		ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
+	while (!list_empty(&s->s_cap_dirty)) {
+		ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
 				      i_dirty_item);
 		inode = &ci->vfs_inode;
 		ihold(inode);
@@ -4198,6 +4223,35 @@ void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
 	dout("flush_dirty_caps done\n");
 }
 
+static void iterate_sessions(struct ceph_mds_client *mdsc,
+			     void (*cb)(struct ceph_mds_session *))
+{
+	int mds = 0;
+
+	mutex_lock(&mdsc->mutex);
+	for (mds = 0; mds < mdsc->max_sessions; ++mds) {
+		struct ceph_mds_session *s;
+
+		if (!mdsc->sessions[mds])
+			continue;
+
+		s = ceph_get_mds_session(mdsc->sessions[mds]);
+		if (!s)
+			continue;
+
+		mutex_unlock(&mdsc->mutex);
+		cb(s);
+		ceph_put_mds_session(s);
+		mutex_lock(&mdsc->mutex);
+	};
+	mutex_unlock(&mdsc->mutex);
+}
+
+void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
+{
+	iterate_sessions(mdsc, flush_dirty_session_caps);
+}
+
 void __ceph_touch_fmode(struct ceph_inode_info *ci,
 			struct ceph_mds_client *mdsc, int fmode)
 {
fs/ceph/mds_client.c +1 -1
index 9a8e7013aca1..be4ad7d28e3a 100644
@@ -755,6 +755,7 @@ static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
 	INIT_LIST_HEAD(&s->s_cap_releases);
 	INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 
+	INIT_LIST_HEAD(&s->s_cap_dirty);
 	INIT_LIST_HEAD(&s->s_cap_flushing);
 
 	mdsc->sessions[mds] = s;
@@ -4375,7 +4376,6 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
 	spin_lock_init(&mdsc->snap_flush_lock);
 	mdsc->last_cap_flush_tid = 1;
 	INIT_LIST_HEAD(&mdsc->cap_flush_list);
-	INIT_LIST_HEAD(&mdsc->cap_dirty);
 	INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
 	mdsc->num_cap_flushing = 0;
 	spin_lock_init(&mdsc->cap_dirty_lock);
fs/ceph/mds_client.h +3 -2
index 8065863321fc..bd20257fb4c2 100644
@@ -199,8 +199,10 @@ struct ceph_mds_session {
 	struct list_head  s_cap_releases; /* waiting cap_release messages */
 	struct work_struct s_cap_release_work;
 
-	/* protected by mutex */
+	/* both protected by s_mdsc->cap_dirty_lock */
+	struct list_head  s_cap_dirty;	      /* inodes w/ dirty caps */
 	struct list_head  s_cap_flushing;     /* inodes w/ flushing caps */
+
 	unsigned long     s_renew_requested; /* last time we sent a renew req */
 	u64               s_renew_seq;
 
@@ -424,7 +426,6 @@ struct ceph_mds_client {
 
 	u64               last_cap_flush_tid;
 	struct list_head  cap_flush_list;
-	struct list_head  cap_dirty;        /* inodes with dirty caps */
 	struct list_head  cap_dirty_migrating; /* ...that are migration... */
 	int               num_cap_flushing; /* # caps we are flushing */
 	spinlock_t        cap_dirty_lock;   /* protects above items */
fs/ceph/super.h +18 -3
index bb372859c0ad..3235c7e3bde7 100644
@@ -351,9 +351,24 @@ struct ceph_inode_info {
 	struct rb_root i_caps;           /* cap list */
 	struct ceph_cap *i_auth_cap;     /* authoritative cap, if any */
 	unsigned i_dirty_caps, i_flushing_caps;     /* mask of dirtied fields */
-	struct list_head i_dirty_item, i_flushing_item; /* protected by
-							 * mdsc->cap_dirty_lock
-							 */
+
+	/*
+	 * Link to the the auth cap's session's s_cap_dirty list. s_cap_dirty
+	 * is protected by the mdsc->cap_dirty_lock, but each individual item
+	 * is also protected by the inode's i_ceph_lock. Walking s_cap_dirty
+	 * requires the mdsc->cap_dirty_lock. List presence for an item can
+	 * be tested under the i_ceph_lock. Changing anything requires both.
+	 */
+	struct list_head i_dirty_item;
+
+	/* Link to session's s_cap_flushing list. Protected by
+	 * mdsc->cap_dirty_lock.
+	 *
+	 * FIXME: this list is sometimes walked without the spinlock being
+	 *	  held. What really protects it?
+	 */
+	struct list_head i_flushing_item;
+
 	/* we need to track cap writeback on a per-cap-bit basis, to allow
 	 * overlapping, pipelined cap flushes to the mds.  we can probably
 	 * reduce the tid to 8 bits if we're concerned about inode size. */
Signature · 1 line
2.25.1
#4 lore raw

PATCH v2 2/2 ceph: request expedited service on session's last cap flush

In reply to Jeff Layton

When flushing a lot of caps to the MDS's at once (e.g. for syncfs), we can end up waiting a substantial amount of time for MDS replies, due to the fact that it may delay some of them so that it can batch them up together in a single journal transaction. This can lead to stalls when calling sync or syncfs.

What we'd really like to do is request expedited service on the _last_ cap we're flushing back to the server. If the CHECK_CAPS_FLUSH flag is set on the request and the current inode was the last one on the session->s_cap_dirty list, then mark the request with CEPH_CLIENT_CAPS_SYNC.

Note that this heuristic is not perfect. New inodes can race onto the list after we've started flushing, but it does seem to fix some common use cases.

---
 fs/ceph/caps.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
fs/ceph/caps.c +6 -2
index 95c9b25e45a6..3630f05993b3 100644
@@ -1987,6 +1987,7 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 	}
 
 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
+		int mflags = 0;
 		struct cap_msg_args arg;
 
 		cap = rb_entry(p, struct ceph_cap, ci_node);
@@ -2118,6 +2119,9 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 			flushing = ci->i_dirty_caps;
 			flush_tid = __mark_caps_flushing(inode, session, false,
 							 &oldest_flush_tid);
+			if (flags & CHECK_CAPS_FLUSH &&
+			    list_empty(&session->s_cap_dirty))
+				mflags |= CEPH_CLIENT_CAPS_SYNC;
 		} else {
 			flushing = 0;
 			flush_tid = 0;
@@ -2128,8 +2132,8 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 
 		mds = cap->mds;  /* remember mds, so we don't repeat */
 
-		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, 0, cap_used, want,
-			   retain, flushing, flush_tid, oldest_flush_tid);
+		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
+			   want, retain, flushing, flush_tid, oldest_flush_tid);
 		spin_unlock(&ci->i_ceph_lock);
 
 		__send_cap(mdsc, &arg, ci);
Signature · 1 line
2.25.1
#8 lore raw
In reply to Jeff Layton

Didn't look at the code too hard, but this sounds good to me.

I also realized that the userspace client isn't quite complete because it didn't send out FLAG_SYNC on a per-MDS-session basis, just for the last cap in its global dirty list. So thanks for that. :)
https://tracker.ceph.com/issues/44916

Quoted 29 lines · Jeff Layton

This is v2 of the patch I sent the other day to fix the problem of long stalls when calling sync or syncfs.

This set converts the mdsc->cap_dirty list to a per-session list, and then has the only caller that looks at cap_dirty walk the list of sessions and issue flushes for each session in turn.

With this, we can use an empty s_cap_dirty list as an indicator that the cap flush is the last one going to the session and can mark that one as one we're waiting on so the MDS can expedite it.

This also attempts to clarify some of the locking around s_cap_dirty, and adds a FIXME comment to raise the question about locking around s_cap_flushing.

Jeff Layton (2):
ceph: convert mdsc->cap_dirty to a per-session list
ceph: request expedited service on session's last cap flush

 fs/ceph/caps.c       | 72 +++++++++++++++++++++++++++++++++++++++-----
 fs/ceph/mds_client.c |  2 +-
 fs/ceph/mds_client.h |  5 +--
 fs/ceph/super.h      | 21 +++++++++++--
 4 files changed, 87 insertions(+), 13 deletions(-)

--
2.25.1

Reviewed-by on v2 1/2
#2 lore raw

Re: [PATCH v2 1/2] ceph: convert mdsc->cap_dirty to a per-session list

In reply to Jeff Layton
Quoted 224 lines · Jeff Layton

This is a per-sb list now, but that makes it difficult to tell when the cap is the last dirty one associated with the session. Switch this to be a per-session list, but continue using the mdsc->cap_dirty_lock to protect the lists.

This list is only ever walked in ceph_flush_dirty_caps, so change that to walk the sessions array and then flush the caps for inodes on each session's list.

If the auth cap ever changes while the inode has dirty caps, then move the inode to the appropriate session for the new auth_cap. Also, ensure that we never remove an auth cap while the inode is still on the s_cap_dirty list.

---
 fs/ceph/caps.c       | 64 ++++++++++++++++++++++++++++++++++++++++----
 fs/ceph/mds_client.c |  2 +-
 fs/ceph/mds_client.h |  5 ++--
 fs/ceph/super.h      | 21 ++++++++++++---
 4 files changed, 81 insertions(+), 11 deletions(-)
fs/ceph/caps.c +59 -5
index 61808793e0c0..95c9b25e45a6 100644
@@ -727,6 +727,19 @@ void ceph_add_cap(struct inode *inode,
        if (flags & CEPH_CAP_FLAG_AUTH) {
                if (!ci->i_auth_cap ||
                    ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
+                       if (ci->i_auth_cap &&
+                           ci->i_auth_cap->session != cap->session) {
+                               /*
+                                * If auth cap session changed, and the cap is
+                                * dirty, move it to the correct session's list
+                                */
+                               if (!list_empty(&ci->i_dirty_item)) {
+                                       spin_lock(&mdsc->cap_dirty_lock);
+                                       list_move(&ci->i_dirty_item,
+                                                 &cap->session->s_cap_dirty);
+                                       spin_unlock(&mdsc->cap_dirty_lock);
+                               }
+                       }
                        ci->i_auth_cap = cap;
                        cap->mds_wanted = wanted;
                }
@@ -1123,8 +1136,10 @@ void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
 
        /* remove from inode's cap rbtree, and clear auth cap */
        rb_erase(&cap->ci_node, &ci->i_caps);
-       if (ci->i_auth_cap == cap)
+       if (ci->i_auth_cap == cap) {
+               WARN_ON_ONCE(!list_empty(&ci->i_dirty_item));
                ci->i_auth_cap = NULL;
+       }
 
        /* remove from session list */
        spin_lock(&session->s_cap_lock);
@@ -1690,6 +1705,8 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
             ceph_cap_string(was | mask));
        ci->i_dirty_caps |= mask;
        if (was == 0) {
+               struct ceph_mds_session *session = ci->i_auth_cap->session;
+
                WARN_ON_ONCE(ci->i_prealloc_cap_flush);
                swap(ci->i_prealloc_cap_flush, *pcf);
 
@@ -1702,7 +1719,7 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
                     &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
                BUG_ON(!list_empty(&ci->i_dirty_item));
                spin_lock(&mdsc->cap_dirty_lock);
-               list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
+               list_add(&ci->i_dirty_item, &session->s_cap_dirty);
                spin_unlock(&mdsc->cap_dirty_lock);
                if (ci->i_flushing_caps == 0) {
                        ihold(inode);
@@ -3753,6 +3770,13 @@ static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
                        if (cap == ci->i_auth_cap)
                                ci->i_auth_cap = tcap;
 
+                       if (!list_empty(&ci->i_dirty_item)) {
+                               spin_lock(&mdsc->cap_dirty_lock);
+                               list_move(&ci->i_dirty_item,
+                                         &tcap->session->s_cap_dirty);
+                               spin_unlock(&mdsc->cap_dirty_lock);
+                       }
+
                        if (!list_empty(&ci->i_cap_flush_list) &&
                            ci->i_auth_cap == tcap) {
                                spin_lock(&mdsc->cap_dirty_lock);
@@ -4176,15 +4200,16 @@ void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
 /*
  * Flush all dirty caps to the mds
  */
-void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
+static void flush_dirty_session_caps(struct ceph_mds_session *s)
 {
+       struct ceph_mds_client *mdsc = s->s_mdsc;
        struct ceph_inode_info *ci;
        struct inode *inode;
 
        dout("flush_dirty_caps\n");
        spin_lock(&mdsc->cap_dirty_lock);
-       while (!list_empty(&mdsc->cap_dirty)) {
-               ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
+       while (!list_empty(&s->s_cap_dirty)) {
+               ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
                                      i_dirty_item);
                inode = &ci->vfs_inode;
                ihold(inode);
@@ -4198,6 +4223,35 @@ void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
        dout("flush_dirty_caps done\n");
 }
 
+static void iterate_sessions(struct ceph_mds_client *mdsc,
+                            void (*cb)(struct ceph_mds_session *))
+{
+       int mds = 0;
+
+       mutex_lock(&mdsc->mutex);
+       for (mds = 0; mds < mdsc->max_sessions; ++mds) {
+               struct ceph_mds_session *s;
+
+               if (!mdsc->sessions[mds])
+                       continue;
+
+               s = ceph_get_mds_session(mdsc->sessions[mds]);
+               if (!s)
+                       continue;
+
+               mutex_unlock(&mdsc->mutex);
+               cb(s);
+               ceph_put_mds_session(s);
+               mutex_lock(&mdsc->mutex);
+       };
+       mutex_unlock(&mdsc->mutex);
+}
+
+void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
+{
+       iterate_sessions(mdsc, flush_dirty_session_caps);
+}
+
 void __ceph_touch_fmode(struct ceph_inode_info *ci,
                        struct ceph_mds_client *mdsc, int fmode)
 {
fs/ceph/mds_client.c +1 -1
index 9a8e7013aca1..be4ad7d28e3a 100644
@@ -755,6 +755,7 @@ static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
        INIT_LIST_HEAD(&s->s_cap_releases);
        INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 
+       INIT_LIST_HEAD(&s->s_cap_dirty);
        INIT_LIST_HEAD(&s->s_cap_flushing);
 
        mdsc->sessions[mds] = s;
@@ -4375,7 +4376,6 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
        spin_lock_init(&mdsc->snap_flush_lock);
        mdsc->last_cap_flush_tid = 1;
        INIT_LIST_HEAD(&mdsc->cap_flush_list);
-       INIT_LIST_HEAD(&mdsc->cap_dirty);
        INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
        mdsc->num_cap_flushing = 0;
        spin_lock_init(&mdsc->cap_dirty_lock);
fs/ceph/mds_client.h +3 -2
index 8065863321fc..bd20257fb4c2 100644
@@ -199,8 +199,10 @@ struct ceph_mds_session {
        struct list_head  s_cap_releases; /* waiting cap_release messages */
        struct work_struct s_cap_release_work;
 
-       /* protected by mutex */
+       /* both protected by s_mdsc->cap_dirty_lock */
+       struct list_head  s_cap_dirty;        /* inodes w/ dirty caps */
        struct list_head  s_cap_flushing;     /* inodes w/ flushing caps */
+
        unsigned long     s_renew_requested; /* last time we sent a renew req */
        u64               s_renew_seq;
 
@@ -424,7 +426,6 @@ struct ceph_mds_client {
 
        u64               last_cap_flush_tid;
        struct list_head  cap_flush_list;
-       struct list_head  cap_dirty;        /* inodes with dirty caps */
        struct list_head  cap_dirty_migrating; /* ...that are migration... */
        int               num_cap_flushing; /* # caps we are flushing */
        spinlock_t        cap_dirty_lock;   /* protects above items */
fs/ceph/super.h +18 -4
index bb372859c0ad..3235c7e3bde7 100644
@@ -351,9 +351,24 @@ struct ceph_inode_info {
        struct rb_root i_caps;           /* cap list */
        struct ceph_cap *i_auth_cap;     /* authoritative cap, if any */
        unsigned i_dirty_caps, i_flushing_caps;     /* mask of dirtied fields */
-       struct list_head i_dirty_item, i_flushing_item; /* protected by
-                                                        * mdsc->cap_dirty_lock
-                                                        */
+
+       /*
+        * Link to the the auth cap's session's s_cap_dirty list. s_cap_dirty
+        * is protected by the mdsc->cap_dirty_lock, but each individual item
+        * is also protected by the inode's i_ceph_lock. Walking s_cap_dirty
+        * requires the mdsc->cap_dirty_lock. List presence for an item can
+        * be tested under the i_ceph_lock. Changing anything requires both.
+        */
+       struct list_head i_dirty_item;
+
+       /* Link to session's s_cap_flushing list. Protected by
+        * mdsc->cap_dirty_lock.
+        *
+        * FIXME: this list is sometimes walked without the spinlock being
+        *        held. What really protects it?
+        */
+       struct list_head i_flushing_item;
+
        /* we need to track cap writeback on a per-cap-bit basis, to allow
         * overlapping, pipelined cap flushes to the mds.  we can probably
         * reduce the tid to 8 bits if we're concerned about inode size. */
--
Quoted 2 lines

2.25.1

I think it's better to unify i_dirty_item and i_flushing_item handling. how about add a new patch that puts code that move i_flushing to new auth session together with the code that move i_dirty_item.

Reviewed-by on v2 2/2
#5 lore raw

Re: [PATCH v2 2/2] ceph: request expedited service on session's last cap flush

In reply to Jeff Layton
Quoted 57 lines · Jeff Layton

When flushing a lot of caps to the MDS's at once (e.g. for syncfs), we can end up waiting a substantial amount of time for MDS replies, due to the fact that it may delay some of them so that it can batch them up together in a single journal transaction. This can lead to stalls when calling sync or syncfs.

What we'd really like to do is request expedited service on the _last_ cap we're flushing back to the server. If the CHECK_CAPS_FLUSH flag is set on the request and the current inode was the last one on the session->s_cap_dirty list, then mark the request with CEPH_CLIENT_CAPS_SYNC.

Note that this heuristic is not perfect. New inodes can race onto the list after we've started flushing, but it does seem to fix some common use cases.

---
 fs/ceph/caps.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
fs/ceph/caps.c +6 -3
index 95c9b25e45a6..3630f05993b3 100644
@@ -1987,6 +1987,7 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
        }
 
        for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
+               int mflags = 0;
                struct cap_msg_args arg;
 
                cap = rb_entry(p, struct ceph_cap, ci_node);
@@ -2118,6 +2119,9 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
                        flushing = ci->i_dirty_caps;
                        flush_tid = __mark_caps_flushing(inode, session, false,
                                                         &oldest_flush_tid);
+                       if (flags & CHECK_CAPS_FLUSH &&
+                           list_empty(&session->s_cap_dirty))
+                               mflags |= CEPH_CLIENT_CAPS_SYNC;
                } else {
                        flushing = 0;
                        flush_tid = 0;
@@ -2128,8 +2132,8 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 
                mds = cap->mds;  /* remember mds, so we don't repeat */
 
-               __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, 0, cap_used, want,
-                          retain, flushing, flush_tid, oldest_flush_tid);
+               __prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
+                          want, retain, flushing, flush_tid, oldest_flush_tid);
                spin_unlock(&ci->i_ceph_lock);
 
                __send_cap(mdsc, &arg, ci);
--
Quoted 2 lines

2.25.1

#6 lore raw

Re: [PATCH v2 2/2] ceph: request expedited service on session's last cap flush

In reply to Jeff Layton
Quoted 58 lines · Jeff Layton

When flushing a lot of caps to the MDS's at once (e.g. for syncfs), we can end up waiting a substantial amount of time for MDS replies, due to the fact that it may delay some of them so that it can batch them up together in a single journal transaction. This can lead to stalls when calling sync or syncfs.

What we'd really like to do is request expedited service on the _last_ cap we're flushing back to the server. If the CHECK_CAPS_FLUSH flag is set on the request and the current inode was the last one on the session->s_cap_dirty list, then mark the request with CEPH_CLIENT_CAPS_SYNC.

Note that this heuristic is not perfect. New inodes can race onto the list after we've started flushing, but it does seem to fix some common use cases.

---
 fs/ceph/caps.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
fs/ceph/caps.c +6 -2
index 95c9b25e45a6..3630f05993b3 100644
@@ -1987,6 +1987,7 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 	}
 
 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
+		int mflags = 0;
 		struct cap_msg_args arg;
 
 		cap = rb_entry(p, struct ceph_cap, ci_node);
@@ -2118,6 +2119,9 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 			flushing = ci->i_dirty_caps;
 			flush_tid = __mark_caps_flushing(inode, session, false,
 							 &oldest_flush_tid);
+			if (flags & CHECK_CAPS_FLUSH &&
+			    list_empty(&session->s_cap_dirty))
+				mflags |= CEPH_CLIENT_CAPS_SYNC;
 		} else {
 			flushing = 0;
 			flush_tid = 0;
@@ -2128,8 +2132,8 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 
 		mds = cap->mds;  /* remember mds, so we don't repeat */
 
-		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, 0, cap_used, want,
-			   retain, flushing, flush_tid, oldest_flush_tid);
+		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
+			   want, retain, flushing, flush_tid, oldest_flush_tid);
 		spin_unlock(&ci->i_ceph_lock);
 
 		__send_cap(mdsc, &arg, ci);
Signature · 1 line
2.25.1

FWIW, it looks good to me. I've also tested it and I confirm it improves (a lot!) the testcase reported by Jan. Maybe it's worth adding the URL to the tracker.

Cheers,
--
Luis

#7 lore raw

Re: [PATCH v2 2/2] ceph: request expedited service on session's last cap flush

Quoted 55 lines · Luis Henriques
Quoted 54 lines · Jeff Layton

When flushing a lot of caps to the MDS's at once (e.g. for syncfs), we can end up waiting a substantial amount of time for MDS replies, due to the fact that it may delay some of them so that it can batch them up together in a single journal transaction. This can lead to stalls when calling sync or syncfs.

What we'd really like to do is request expedited service on the _last_ cap we're flushing back to the server. If the CHECK_CAPS_FLUSH flag is set on the request and the current inode was the last one on the session->s_cap_dirty list, then mark the request with CEPH_CLIENT_CAPS_SYNC.

Note that this heuristic is not perfect. New inodes can race onto the list after we've started flushing, but it does seem to fix some common use cases.

---
 fs/ceph/caps.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
fs/ceph/caps.c +6 -2
index 95c9b25e45a6..3630f05993b3 100644
@@ -1987,6 +1987,7 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 	}
 
 	for (p = rb_first(&ci->i_caps); p; p = rb_next(p)) {
+		int mflags = 0;
 		struct cap_msg_args arg;
 
 		cap = rb_entry(p, struct ceph_cap, ci_node);
@@ -2118,6 +2119,9 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 			flushing = ci->i_dirty_caps;
 			flush_tid = __mark_caps_flushing(inode, session, false,
 							 &oldest_flush_tid);
+			if (flags & CHECK_CAPS_FLUSH &&
+			    list_empty(&session->s_cap_dirty))
+				mflags |= CEPH_CLIENT_CAPS_SYNC;
 		} else {
 			flushing = 0;
 			flush_tid = 0;
@@ -2128,8 +2132,8 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 
 		mds = cap->mds;  /* remember mds, so we don't repeat */
 
-		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, 0, cap_used, want,
-			   retain, flushing, flush_tid, oldest_flush_tid);
+		__prep_cap(&arg, cap, CEPH_CAP_OP_UPDATE, mflags, cap_used,
+			   want, retain, flushing, flush_tid, oldest_flush_tid);
 		spin_unlock(&ci->i_ceph_lock);
 
Quoted 9 lines
Quoted 4 lines
fs/ceph/caps.c +0 -0
 		__send_cap(mdsc, &arg, ci);
Signature · 1 line
2.25.1

FWIW, it looks good to me. I've also tested it and I confirm it improves (a lot!) the testcase reported by Jan. Maybe it's worth adding the URL to the tracker.

Will do. Thanks for the reminder!

Signature · 1 line
Jeff Layton <[email protected]>
#3 lore raw

Re: [PATCH v2 1/2] ceph: convert mdsc->cap_dirty to a per-session list

In reply to Yan, Zheng
Quoted 229 lines · Zheng
Quoted 223 lines · Jeff Layton

This is a per-sb list now, but that makes it difficult to tell when the cap is the last dirty one associated with the session. Switch this to be a per-session list, but continue using the mdsc->cap_dirty_lock to protect the lists.

This list is only ever walked in ceph_flush_dirty_caps, so change that to walk the sessions array and then flush the caps for inodes on each session's list.

If the auth cap ever changes while the inode has dirty caps, then move the inode to the appropriate session for the new auth_cap. Also, ensure that we never remove an auth cap while the inode is still on the s_cap_dirty list.

---
 fs/ceph/caps.c       | 64 ++++++++++++++++++++++++++++++++++++++++----
 fs/ceph/mds_client.c |  2 +-
 fs/ceph/mds_client.h |  5 ++--
 fs/ceph/super.h      | 21 ++++++++++++---
 4 files changed, 81 insertions(+), 11 deletions(-)
fs/ceph/caps.c +59 -5
index 61808793e0c0..95c9b25e45a6 100644
@@ -727,6 +727,19 @@ void ceph_add_cap(struct inode *inode,
        if (flags & CEPH_CAP_FLAG_AUTH) {
                if (!ci->i_auth_cap ||
                    ceph_seq_cmp(ci->i_auth_cap->mseq, mseq) < 0) {
+                       if (ci->i_auth_cap &&
+                           ci->i_auth_cap->session != cap->session) {
+                               /*
+                                * If auth cap session changed, and the cap is
+                                * dirty, move it to the correct session's list
+                                */
+                               if (!list_empty(&ci->i_dirty_item)) {
+                                       spin_lock(&mdsc->cap_dirty_lock);
+                                       list_move(&ci->i_dirty_item,
+                                                 &cap->session->s_cap_dirty);
+                                       spin_unlock(&mdsc->cap_dirty_lock);
+                               }
+                       }
                        ci->i_auth_cap = cap;
                        cap->mds_wanted = wanted;
                }
@@ -1123,8 +1136,10 @@ void __ceph_remove_cap(struct ceph_cap *cap, bool queue_release)
 
        /* remove from inode's cap rbtree, and clear auth cap */
        rb_erase(&cap->ci_node, &ci->i_caps);
-       if (ci->i_auth_cap == cap)
+       if (ci->i_auth_cap == cap) {
+               WARN_ON_ONCE(!list_empty(&ci->i_dirty_item));
                ci->i_auth_cap = NULL;
+       }
 
        /* remove from session list */
        spin_lock(&session->s_cap_lock);
@@ -1690,6 +1705,8 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
             ceph_cap_string(was | mask));
        ci->i_dirty_caps |= mask;
        if (was == 0) {
+               struct ceph_mds_session *session = ci->i_auth_cap->session;
+
                WARN_ON_ONCE(ci->i_prealloc_cap_flush);
                swap(ci->i_prealloc_cap_flush, *pcf);
 
@@ -1702,7 +1719,7 @@ int __ceph_mark_dirty_caps(struct ceph_inode_info *ci, int mask,
                     &ci->vfs_inode, ci->i_head_snapc, ci->i_auth_cap);
                BUG_ON(!list_empty(&ci->i_dirty_item));
                spin_lock(&mdsc->cap_dirty_lock);
-               list_add(&ci->i_dirty_item, &mdsc->cap_dirty);
+               list_add(&ci->i_dirty_item, &session->s_cap_dirty);
                spin_unlock(&mdsc->cap_dirty_lock);
                if (ci->i_flushing_caps == 0) {
                        ihold(inode);
@@ -3753,6 +3770,13 @@ static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
                        if (cap == ci->i_auth_cap)
                                ci->i_auth_cap = tcap;
 
+                       if (!list_empty(&ci->i_dirty_item)) {
+                               spin_lock(&mdsc->cap_dirty_lock);
+                               list_move(&ci->i_dirty_item,
+                                         &tcap->session->s_cap_dirty);
+                               spin_unlock(&mdsc->cap_dirty_lock);
+                       }
+
                        if (!list_empty(&ci->i_cap_flush_list) &&
                            ci->i_auth_cap == tcap) {
                                spin_lock(&mdsc->cap_dirty_lock);
@@ -4176,15 +4200,16 @@ void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
 /*
  * Flush all dirty caps to the mds
  */
-void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
+static void flush_dirty_session_caps(struct ceph_mds_session *s)
 {
+       struct ceph_mds_client *mdsc = s->s_mdsc;
        struct ceph_inode_info *ci;
        struct inode *inode;
 
        dout("flush_dirty_caps\n");
        spin_lock(&mdsc->cap_dirty_lock);
-       while (!list_empty(&mdsc->cap_dirty)) {
-               ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
+       while (!list_empty(&s->s_cap_dirty)) {
+               ci = list_first_entry(&s->s_cap_dirty, struct ceph_inode_info,
                                      i_dirty_item);
                inode = &ci->vfs_inode;
                ihold(inode);
@@ -4198,6 +4223,35 @@ void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
        dout("flush_dirty_caps done\n");
 }
 
+static void iterate_sessions(struct ceph_mds_client *mdsc,
+                            void (*cb)(struct ceph_mds_session *))
+{
+       int mds = 0;
+
+       mutex_lock(&mdsc->mutex);
+       for (mds = 0; mds < mdsc->max_sessions; ++mds) {
+               struct ceph_mds_session *s;
+
+               if (!mdsc->sessions[mds])
+                       continue;
+
+               s = ceph_get_mds_session(mdsc->sessions[mds]);
+               if (!s)
+                       continue;
+
+               mutex_unlock(&mdsc->mutex);
+               cb(s);
+               ceph_put_mds_session(s);
+               mutex_lock(&mdsc->mutex);
+       };
+       mutex_unlock(&mdsc->mutex);
+}
+
+void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
+{
+       iterate_sessions(mdsc, flush_dirty_session_caps);
+}
+
 void __ceph_touch_fmode(struct ceph_inode_info *ci,
                        struct ceph_mds_client *mdsc, int fmode)
 {
fs/ceph/mds_client.c +1 -1
index 9a8e7013aca1..be4ad7d28e3a 100644
@@ -755,6 +755,7 @@ static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
        INIT_LIST_HEAD(&s->s_cap_releases);
        INIT_WORK(&s->s_cap_release_work, ceph_cap_release_work);
 
+       INIT_LIST_HEAD(&s->s_cap_dirty);
        INIT_LIST_HEAD(&s->s_cap_flushing);
 
        mdsc->sessions[mds] = s;
@@ -4375,7 +4376,6 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
        spin_lock_init(&mdsc->snap_flush_lock);
        mdsc->last_cap_flush_tid = 1;
        INIT_LIST_HEAD(&mdsc->cap_flush_list);
-       INIT_LIST_HEAD(&mdsc->cap_dirty);
        INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
        mdsc->num_cap_flushing = 0;
        spin_lock_init(&mdsc->cap_dirty_lock);
fs/ceph/mds_client.h +3 -2
index 8065863321fc..bd20257fb4c2 100644
@@ -199,8 +199,10 @@ struct ceph_mds_session {
        struct list_head  s_cap_releases; /* waiting cap_release messages */
        struct work_struct s_cap_release_work;
 
-       /* protected by mutex */
+       /* both protected by s_mdsc->cap_dirty_lock */
+       struct list_head  s_cap_dirty;        /* inodes w/ dirty caps */
        struct list_head  s_cap_flushing;     /* inodes w/ flushing caps */
+
        unsigned long     s_renew_requested; /* last time we sent a renew req */
        u64               s_renew_seq;
 
@@ -424,7 +426,6 @@ struct ceph_mds_client {
 
        u64               last_cap_flush_tid;
        struct list_head  cap_flush_list;
-       struct list_head  cap_dirty;        /* inodes with dirty caps */
        struct list_head  cap_dirty_migrating; /* ...that are migration... */
        int               num_cap_flushing; /* # caps we are flushing */
        spinlock_t        cap_dirty_lock;   /* protects above items */
fs/ceph/super.h +18 -4
index bb372859c0ad..3235c7e3bde7 100644
@@ -351,9 +351,24 @@ struct ceph_inode_info {
        struct rb_root i_caps;           /* cap list */
        struct ceph_cap *i_auth_cap;     /* authoritative cap, if any */
        unsigned i_dirty_caps, i_flushing_caps;     /* mask of dirtied fields */
-       struct list_head i_dirty_item, i_flushing_item; /* protected by
-                                                        * mdsc->cap_dirty_lock
-                                                        */
+
+       /*
+        * Link to the the auth cap's session's s_cap_dirty list. s_cap_dirty
+        * is protected by the mdsc->cap_dirty_lock, but each individual item
+        * is also protected by the inode's i_ceph_lock. Walking s_cap_dirty
+        * requires the mdsc->cap_dirty_lock. List presence for an item can
+        * be tested under the i_ceph_lock. Changing anything requires both.
+        */
+       struct list_head i_dirty_item;
+
+       /* Link to session's s_cap_flushing list. Protected by
+        * mdsc->cap_dirty_lock.
+        *
+        * FIXME: this list is sometimes walked without the spinlock being
+        *        held. What really protects it?
+        */
+       struct list_head i_flushing_item;
+
        /* we need to track cap writeback on a per-cap-bit basis, to allow
         * overlapping, pipelined cap flushes to the mds.  we can probably
         * reduce the tid to 8 bits if we're concerned about inode size. */
--
Quoted 2 lines

2.25.1

Quoted 4 lines

I think it's better to unify i_dirty_item and i_flushing_item handling. how about add a new patch that puts code that move i_flushing to new auth session together with the code that move i_dirty_item.

Yeah, that does sound like a good idea. I'll see what I can come up with.

Signature · 1 line
Jeff Layton <[email protected]>