[Cluster-devel] [PATCHv4 dlm/next 0/8] fs: dlm: introduce dlm re-transmission layer

[Cluster-devel] [PATCHv4 dlm/next 0/8] fs: dlm: introduce dlm re-transmission layer

[Alexander Aring]

Hi,

this is the final patch-series to make dlm reliable when re-connection
occurs. You can easily generate a couple of re-connections by running:

tcpkill -9 -i $IFACE port 21064

on your own to test these patches. At some time dlm will detect message
drops and will re-transmit messages if necessary. It introduces a new dlm
protocol behaviour and increases the dlm protocol version. I tested it
with SCTP as well and tried to be backwards compatible with dlm protocol
version 3.1. However I don't recommend at all to mix these versions
in a setup since dlm version 3.2 fixes long-term issues.

- Alex

changes since v4:
 - remove fast retransmit and the timer, introduce logic to retransmit
   all unacknowledged message when doing reconnect. The receiver side
   will deliver the next fit sequence number then.

   There might be still problems with that we don't trigger a
   reconnection again if we don't transmit anything (but still have
   something unacknowledged in midcomms). This is still an issue in
   lowcomms implementation which I want to fix in due courses.
 - Change comments/commit msg how it works regarding to the new
   behaviour.
 - Let the send_queue now have messages in order according to the seq
   this is necessary for the new behaviour that the receiver side can
   resolve drops by receiving unacknowledged messages in their order of
   delivery. If sequence .e.g. 1 3 2 is received then the receiver will
   not be able to resolve the drop because 3 will be dropped and not
   retransmitted again.
 - change the dlm fin waiting mechanism to split the wait into fin ack
   received and fin message received. Also change the timeout handling
   a little bit there.
 - add a missing flush send_queue in midcomms_close
 - update patch 04 to not be irqsafe anymore
 - fix use-after-free for dlm version 3.1 and recent nodes_srcu changes

I thought about to update patch 08 to drop all pending messages inside
the write queue, because we retransmit all unacknowledged messages
at reconnect anyway. However that makes a very bad behaviour on
reconnects with DLM version 3.1 so I only drop half-transmitted page
buffers to don't start the bytestream inside the middle of an DLM
message which is terrible as well. It might send more duplicate messages
at reconnect, but the receive should solve these duplicates.

I still have some problems with synchronization of membership with
DLM_FIN. However I think my testcase is overkill and I have zero
problems with any synchronization when not running tcpkill. It gets
a lot of more worse when I don't have any synchronization and the
"midcomms membership" and sequence numbers are out of sync with the
"cluster manager membership". I think such synchronization need to be
there but there might be more additional handling. (I hope non protocol
changes needed).

changes since v3:
 - add comment about why queues are unbound
 - move rcu usage to version receive handler

changes since v2:
 - make timer handling pending only if messages are on air, the sync
   isn't quite correct there but doesn't need to be precise
 - use before() from tcp to check if seq is before other seq with
   respect of overflows
 - change srcu handling to hold srcu in all places where nodes are
   referencing - we should not get a disadvantage of holding that
   lock. We should update also lowcomms regarding to that.
 - add some WARN_ON() to check that nothing in send/recv is going
   anymore otherwise it's likely an issue.
 - add more future work regarding to fencing of nodes if over
   cluster manager timeout/bad seq happens
 - add note about missing length size check of tail payload
   (resource name length) regarding to the receive buffer
 - remove some include which isn't necessary in recoverd.c

Thanks to Paolo Abeni and Guillaume Nault for their reviews and
recommendations.

Alexander Aring (8):
  fs: dlm: public header in out utility
  fs: dlm: add more midcomms hooks
  fs: dlm: make buffer handling per msg
  fs: dlm: add functionality to re-transmit a message
  fs: dlm: move out some hash functionality
  fs: dlm: add union in dlm header for lockspace id
  fs: dlm: add reliable connection if reconnect
  fs: dlm: don't allow half transmitted messages

 fs/dlm/config.c       |    3 +-
 fs/dlm/dlm_internal.h |   35 +-
 fs/dlm/lock.c         |   14 +-
 fs/dlm/lockspace.c    |   14 +-
 fs/dlm/lowcomms.c     |  173 ++++++-
 fs/dlm/lowcomms.h     |   23 +-
 fs/dlm/member.c       |   12 +-
 fs/dlm/midcomms.c     | 1101 +++++++++++++++++++++++++++++++++++++++--
 fs/dlm/midcomms.h     |   11 +
 fs/dlm/rcom.c         |   63 ++-
 fs/dlm/util.c         |   10 +-
 fs/dlm/util.h         |    2 +
 12 files changed, 1361 insertions(+), 100 deletions(-)

-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 1/8] fs: dlm: public header in out utility

[Alexander Aring]

This patch allows to use header_out() and header_in() outside of dlm
util functionality.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/util.c | 4 ++--
 fs/dlm/util.h | 2 ++
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/fs/dlm/util.c b/fs/dlm/util.c
index cfd0d00b19ae..74a8c5bfe9b5 100644
--- a/fs/dlm/util.c
+++ b/fs/dlm/util.c
@@ -20,7 +20,7 @@
 #define DLM_ERRNO_ETIMEDOUT	       110
 #define DLM_ERRNO_EINPROGRESS	       115
 
-static void header_out(struct dlm_header *hd)
+void header_out(struct dlm_header *hd)
 {
 	hd->h_version		= cpu_to_le32(hd->h_version);
 	hd->h_lockspace		= cpu_to_le32(hd->h_lockspace);
@@ -28,7 +28,7 @@ static void header_out(struct dlm_header *hd)
 	hd->h_length		= cpu_to_le16(hd->h_length);
 }
 
-static void header_in(struct dlm_header *hd)
+void header_in(struct dlm_header *hd)
 {
 	hd->h_version		= le32_to_cpu(hd->h_version);
 	hd->h_lockspace		= le32_to_cpu(hd->h_lockspace);
diff --git a/fs/dlm/util.h b/fs/dlm/util.h
index cc719ca9397e..d46f23c7a6a0 100644
--- a/fs/dlm/util.h
+++ b/fs/dlm/util.h
@@ -15,6 +15,8 @@ void dlm_message_out(struct dlm_message *ms);
 void dlm_message_in(struct dlm_message *ms);
 void dlm_rcom_out(struct dlm_rcom *rc);
 void dlm_rcom_in(struct dlm_rcom *rc);
+void header_out(struct dlm_header *hd);
+void header_in(struct dlm_header *hd);
 
 #endif
 
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 2/8] fs: dlm: add more midcomms hooks

[Alexander Aring]

This patch prepares hooks to redirect to the midcomms layer which will
be used by the midcomms re-transmit handling.

There exists the new concept of stateless buffers allocation and
commits. This can be used to bypass the midcomms re-transmit handling. It
is used by RCOM_STATUS and RCOM_NAMES messages, because they have their
own ping-like re-transmit handling. As well these two messages will be
used to determine the DLM version per node, because these two messages
are per observation the first messages which are exchanged.

The midcomms_remove_member() hook should be called when there is nothing
to send to the other node and the other node is still capable to
transmit dlm messages to the other node which called
midcomms_remove_member(). I experienced that the dlm protocol has a lack
of support for synchronize this event on protocol level. The result was
that there was still something to transmit but the other node was already
gone. This hook can be used to provide such synchronization. Although I
am not totally sure about the placement of this hook, I did not observed
issues yet when providing such synchronization on protocol layer.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/config.c    |  3 ++-
 fs/dlm/lock.c      |  6 ++---
 fs/dlm/lockspace.c |  7 +++---
 fs/dlm/member.c    | 12 +++++++++-
 fs/dlm/midcomms.c  | 44 +++++++++++++++++++++++++++++++++++++
 fs/dlm/midcomms.h  | 10 +++++++++
 fs/dlm/rcom.c      | 55 +++++++++++++++++++++++++++-------------------
 7 files changed, 107 insertions(+), 30 deletions(-)

diff --git a/fs/dlm/config.c b/fs/dlm/config.c
index 88d95d96e36c..01ae294743e9 100644
--- a/fs/dlm/config.c
+++ b/fs/dlm/config.c
@@ -20,6 +20,7 @@
 #include <net/sock.h>
 
 #include "config.h"
+#include "midcomms.h"
 #include "lowcomms.h"
 
 /*
@@ -532,7 +533,7 @@ static void drop_comm(struct config_group *g, struct config_item *i)
 	struct dlm_comm *cm = config_item_to_comm(i);
 	if (local_comm == cm)
 		local_comm = NULL;
-	dlm_lowcomms_close(cm->nodeid);
+	dlm_midcomms_close(cm->nodeid);
 	while (cm->addr_count--)
 		kfree(cm->addr[cm->addr_count]);
 	config_item_put(i);
diff --git a/fs/dlm/lock.c b/fs/dlm/lock.c
index b93df39d0915..b3fd823009f4 100644
--- a/fs/dlm/lock.c
+++ b/fs/dlm/lock.c
@@ -59,7 +59,7 @@
 #include "dlm_internal.h"
 #include <linux/dlm_device.h>
 #include "memory.h"
-#include "lowcomms.h"
+#include "midcomms.h"
 #include "requestqueue.h"
 #include "util.h"
 #include "dir.h"
@@ -3537,7 +3537,7 @@ static int _create_message(struct dlm_ls *ls, int mb_len,
 	   pass into lowcomms_commit and a message buffer (mb) that we
 	   write our data into */
 
-	mh = dlm_lowcomms_get_buffer(to_nodeid, mb_len, GFP_NOFS, &mb);
+	mh = dlm_midcomms_get_buffer(to_nodeid, mb_len, GFP_NOFS, &mb);
 	if (!mh)
 		return -ENOBUFS;
 
@@ -3589,7 +3589,7 @@ static int create_message(struct dlm_rsb *r, struct dlm_lkb *lkb,
 static int send_message(struct dlm_mhandle *mh, struct dlm_message *ms)
 {
 	dlm_message_out(ms);
-	dlm_lowcomms_commit_buffer(mh);
+	dlm_midcomms_commit_buffer(mh);
 	return 0;
 }
 
diff --git a/fs/dlm/lockspace.c b/fs/dlm/lockspace.c
index c14cf2b7faab..bf5c55ef9d0d 100644
--- a/fs/dlm/lockspace.c
+++ b/fs/dlm/lockspace.c
@@ -16,6 +16,7 @@
 #include "member.h"
 #include "recoverd.h"
 #include "dir.h"
+#include "midcomms.h"
 #include "lowcomms.h"
 #include "config.h"
 #include "memory.h"
@@ -390,7 +391,7 @@ static int threads_start(void)
 	}
 
 	/* Thread for sending/receiving messages for all lockspace's */
-	error = dlm_lowcomms_start();
+	error = dlm_midcomms_start();
 	if (error) {
 		log_print("cannot start dlm lowcomms %d", error);
 		goto scand_fail;
@@ -698,7 +699,7 @@ int dlm_new_lockspace(const char *name, const char *cluster,
 		error = 0;
 	if (!ls_count) {
 		dlm_scand_stop();
-		dlm_lowcomms_shutdown();
+		dlm_midcomms_shutdown();
 		dlm_lowcomms_stop();
 	}
  out:
@@ -787,7 +788,7 @@ static int release_lockspace(struct dlm_ls *ls, int force)
 
 	if (ls_count == 1) {
 		dlm_scand_stop();
-		dlm_lowcomms_shutdown();
+		dlm_midcomms_shutdown();
 	}
 
 	dlm_callback_stop(ls);
diff --git a/fs/dlm/member.c b/fs/dlm/member.c
index ceef3f2074ff..c018a995a8ab 100644
--- a/fs/dlm/member.c
+++ b/fs/dlm/member.c
@@ -15,6 +15,7 @@
 #include "recover.h"
 #include "rcom.h"
 #include "config.h"
+#include "midcomms.h"
 #include "lowcomms.h"
 
 int dlm_slots_version(struct dlm_header *h)
@@ -378,7 +379,15 @@ void dlm_clear_members(struct dlm_ls *ls)
 
 void dlm_clear_members_gone(struct dlm_ls *ls)
 {
-	clear_memb_list(&ls->ls_nodes_gone);
+	struct list_head *head = &ls->ls_nodes_gone;
+	struct dlm_member *memb;
+
+	while (!list_empty(head)) {
+		memb = list_entry(head->next, struct dlm_member, list);
+		midcomms_remove_member(memb->nodeid);
+		list_del(&memb->list);
+		kfree(memb);
+	}
 }
 
 static void make_member_array(struct dlm_ls *ls)
@@ -563,6 +572,7 @@ int dlm_recover_members(struct dlm_ls *ls, struct dlm_recover *rv, int *neg_out)
 		if (dlm_is_member(ls, node->nodeid))
 			continue;
 		dlm_add_member(ls, node);
+		midcomms_add_member(node->nodeid);
 		log_rinfo(ls, "add member %d", node->nodeid);
 	}
 
diff --git a/fs/dlm/midcomms.c b/fs/dlm/midcomms.c
index 1c6654a21ec4..bbcb242e6101 100644
--- a/fs/dlm/midcomms.c
+++ b/fs/dlm/midcomms.c
@@ -28,6 +28,50 @@
 #include "lock.h"
 #include "midcomms.h"
 
+void *dlm_midcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
+{
+	return dlm_lowcomms_get_buffer(nodeid, len, allocation, ppc);
+}
+
+void dlm_midcomms_commit_buffer(void *mh)
+{
+	dlm_lowcomms_commit_buffer(mh);
+}
+
+void *dlm_midcomms_stateless_get_buffer(int nodeid, int len, gfp_t allocation,
+					char **ppc)
+{
+	return dlm_lowcomms_get_buffer(nodeid, len, allocation, ppc);
+}
+
+void dlm_midcomms_stateless_commit_buffer(void *mh)
+{
+	dlm_lowcomms_commit_buffer(mh);
+}
+
+void midcomms_add_member(int nodeid)
+{
+}
+
+void midcomms_remove_member(int nodeid)
+{
+}
+
+int dlm_midcomms_close(int nodeid)
+{
+	return dlm_lowcomms_close(nodeid);
+}
+
+int dlm_midcomms_start(void)
+{
+	return dlm_lowcomms_start();
+}
+
+void dlm_midcomms_shutdown(void)
+{
+	dlm_lowcomms_shutdown();
+}
+
 /*
  * Called from the low-level comms layer to process a buffer of
  * commands.
diff --git a/fs/dlm/midcomms.h b/fs/dlm/midcomms.h
index 61e90a921849..6ce2f2947475 100644
--- a/fs/dlm/midcomms.h
+++ b/fs/dlm/midcomms.h
@@ -13,6 +13,16 @@
 #define __MIDCOMMS_DOT_H__
 
 int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int buflen);
+void *dlm_midcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc);
+void dlm_midcomms_commit_buffer(void *mh);
+void *dlm_midcomms_stateless_get_buffer(int nodeid, int len, gfp_t allocation,
+					char **ppc);
+void dlm_midcomms_stateless_commit_buffer(void *mh);
+void midcomms_add_member(int nodeid);
+void midcomms_remove_member(int nodeid);
+int dlm_midcomms_close(int nodeid);
+int dlm_midcomms_start(void);
+void dlm_midcomms_shutdown(void);
 
 #endif				/* __MIDCOMMS_DOT_H__ */
 
diff --git a/fs/dlm/rcom.c b/fs/dlm/rcom.c
index f5b1bd65728d..7a7d4a8e4706 100644
--- a/fs/dlm/rcom.c
+++ b/fs/dlm/rcom.c
@@ -28,14 +28,18 @@ static int rcom_response(struct dlm_ls *ls)
 }
 
 static int create_rcom(struct dlm_ls *ls, int to_nodeid, int type, int len,
-		       struct dlm_rcom **rc_ret, struct dlm_mhandle **mh_ret)
+		       struct dlm_rcom **rc_ret, struct dlm_mhandle **mh_ret,
+		       bool stateless)
 {
 	struct dlm_rcom *rc;
 	struct dlm_mhandle *mh;
 	char *mb;
 	int mb_len = sizeof(struct dlm_rcom) + len;
 
-	mh = dlm_lowcomms_get_buffer(to_nodeid, mb_len, GFP_NOFS, &mb);
+	if (stateless)
+		mh = dlm_midcomms_stateless_get_buffer(to_nodeid, mb_len, GFP_NOFS, &mb);
+	else
+		mh = dlm_midcomms_get_buffer(to_nodeid, mb_len, GFP_NOFS, &mb);
 	if (!mh) {
 		log_print("create_rcom to %d type %d len %d ENOBUFS",
 			  to_nodeid, type, len);
@@ -62,10 +66,13 @@ static int create_rcom(struct dlm_ls *ls, int to_nodeid, int type, int len,
 }
 
 static void send_rcom(struct dlm_ls *ls, struct dlm_mhandle *mh,
-		      struct dlm_rcom *rc)
+		      struct dlm_rcom *rc, bool stateless)
 {
 	dlm_rcom_out(rc);
-	dlm_lowcomms_commit_buffer(mh);
+	if (stateless)
+		dlm_midcomms_stateless_commit_buffer(mh);
+	else
+		dlm_midcomms_commit_buffer(mh);
 }
 
 static void set_rcom_status(struct dlm_ls *ls, struct rcom_status *rs,
@@ -154,7 +161,7 @@ int dlm_rcom_status(struct dlm_ls *ls, int nodeid, uint32_t status_flags)
 
 retry:
 	error = create_rcom(ls, nodeid, DLM_RCOM_STATUS,
-			    sizeof(struct rcom_status), &rc, &mh);
+			    sizeof(struct rcom_status), &rc, &mh, true);
 	if (error)
 		goto out;
 
@@ -163,7 +170,7 @@ int dlm_rcom_status(struct dlm_ls *ls, int nodeid, uint32_t status_flags)
 	allow_sync_reply(ls, &rc->rc_id);
 	memset(ls->ls_recover_buf, 0, LOWCOMMS_MAX_TX_BUFFER_LEN);
 
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, true);
 
 	error = dlm_wait_function(ls, &rcom_response);
 	disallow_sync_reply(ls);
@@ -219,7 +226,7 @@ static void receive_rcom_status(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 
  do_create:
 	error = create_rcom(ls, nodeid, DLM_RCOM_STATUS_REPLY,
-			    len, &rc, &mh);
+			    len, &rc, &mh, true);
 	if (error)
 		return;
 
@@ -246,7 +253,7 @@ static void receive_rcom_status(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 	spin_unlock(&ls->ls_recover_lock);
 
  do_send:
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, true);
 }
 
 static void receive_sync_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
@@ -277,7 +284,8 @@ int dlm_rcom_names(struct dlm_ls *ls, int nodeid, char *last_name, int last_len)
 	ls->ls_recover_nodeid = nodeid;
 
 retry:
-	error = create_rcom(ls, nodeid, DLM_RCOM_NAMES, last_len, &rc, &mh);
+	error = create_rcom(ls, nodeid, DLM_RCOM_NAMES, last_len, &rc, &mh,
+			    true);
 	if (error)
 		goto out;
 	memcpy(rc->rc_buf, last_name, last_len);
@@ -285,7 +293,7 @@ int dlm_rcom_names(struct dlm_ls *ls, int nodeid, char *last_name, int last_len)
 	allow_sync_reply(ls, &rc->rc_id);
 	memset(ls->ls_recover_buf, 0, LOWCOMMS_MAX_TX_BUFFER_LEN);
 
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, true);
 
 	error = dlm_wait_function(ls, &rcom_response);
 	disallow_sync_reply(ls);
@@ -305,7 +313,8 @@ static void receive_rcom_names(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 	inlen = rc_in->rc_header.h_length - sizeof(struct dlm_rcom);
 	outlen = LOWCOMMS_MAX_TX_BUFFER_LEN - sizeof(struct dlm_rcom);
 
-	error = create_rcom(ls, nodeid, DLM_RCOM_NAMES_REPLY, outlen, &rc, &mh);
+	error = create_rcom(ls, nodeid, DLM_RCOM_NAMES_REPLY, outlen, &rc, &mh,
+			    true);
 	if (error)
 		return;
 	rc->rc_id = rc_in->rc_id;
@@ -313,7 +322,7 @@ static void receive_rcom_names(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 
 	dlm_copy_master_names(ls, rc_in->rc_buf, inlen, rc->rc_buf, outlen,
 			      nodeid);
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, true);
 }
 
 int dlm_send_rcom_lookup(struct dlm_rsb *r, int dir_nodeid)
@@ -324,13 +333,13 @@ int dlm_send_rcom_lookup(struct dlm_rsb *r, int dir_nodeid)
 	int error;
 
 	error = create_rcom(ls, dir_nodeid, DLM_RCOM_LOOKUP, r->res_length,
-			    &rc, &mh);
+			    &rc, &mh, false);
 	if (error)
 		goto out;
 	memcpy(rc->rc_buf, r->res_name, r->res_length);
 	rc->rc_id = (unsigned long) r->res_id;
 
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, false);
  out:
 	return error;
 }
@@ -342,7 +351,8 @@ static void receive_rcom_lookup(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 	int error, ret_nodeid, nodeid = rc_in->rc_header.h_nodeid;
 	int len = rc_in->rc_header.h_length - sizeof(struct dlm_rcom);
 
-	error = create_rcom(ls, nodeid, DLM_RCOM_LOOKUP_REPLY, 0, &rc, &mh);
+	error = create_rcom(ls, nodeid, DLM_RCOM_LOOKUP_REPLY, 0, &rc, &mh,
+			    false);
 	if (error)
 		return;
 
@@ -361,7 +371,7 @@ static void receive_rcom_lookup(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 	rc->rc_id = rc_in->rc_id;
 	rc->rc_seq_reply = rc_in->rc_seq;
 
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, false);
 }
 
 static void receive_rcom_lookup_reply(struct dlm_ls *ls, struct dlm_rcom *rc_in)
@@ -410,7 +420,8 @@ int dlm_send_rcom_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
 	if (lkb->lkb_lvbptr)
 		len += ls->ls_lvblen;
 
-	error = create_rcom(ls, r->res_nodeid, DLM_RCOM_LOCK, len, &rc, &mh);
+	error = create_rcom(ls, r->res_nodeid, DLM_RCOM_LOCK, len, &rc, &mh,
+			    false);
 	if (error)
 		goto out;
 
@@ -418,7 +429,7 @@ int dlm_send_rcom_lock(struct dlm_rsb *r, struct dlm_lkb *lkb)
 	pack_rcom_lock(r, lkb, rl);
 	rc->rc_id = (unsigned long) r;
 
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, false);
  out:
 	return error;
 }
@@ -433,7 +444,7 @@ static void receive_rcom_lock(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 	dlm_recover_master_copy(ls, rc_in);
 
 	error = create_rcom(ls, nodeid, DLM_RCOM_LOCK_REPLY,
-			    sizeof(struct rcom_lock), &rc, &mh);
+			    sizeof(struct rcom_lock), &rc, &mh, false);
 	if (error)
 		return;
 
@@ -444,7 +455,7 @@ static void receive_rcom_lock(struct dlm_ls *ls, struct dlm_rcom *rc_in)
 	rc->rc_id = rc_in->rc_id;
 	rc->rc_seq_reply = rc_in->rc_seq;
 
-	send_rcom(ls, mh, rc);
+	send_rcom(ls, mh, rc, false);
 }
 
 /* If the lockspace doesn't exist then still send a status message
@@ -458,7 +469,7 @@ int dlm_send_ls_not_ready(int nodeid, struct dlm_rcom *rc_in)
 	char *mb;
 	int mb_len = sizeof(struct dlm_rcom) + sizeof(struct rcom_config);
 
-	mh = dlm_lowcomms_get_buffer(nodeid, mb_len, GFP_NOFS, &mb);
+	mh = dlm_midcomms_get_buffer(nodeid, mb_len, GFP_NOFS, &mb);
 	if (!mh)
 		return -ENOBUFS;
 
@@ -479,7 +490,7 @@ int dlm_send_ls_not_ready(int nodeid, struct dlm_rcom *rc_in)
 	rf->rf_lvblen = cpu_to_le32(~0U);
 
 	dlm_rcom_out(rc);
-	dlm_lowcomms_commit_buffer(mh);
+	dlm_midcomms_commit_buffer(mh);
 
 	return 0;
 }
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 3/8] fs: dlm: make buffer handling per msg

[Alexander Aring]

This patch makes the void pointer handle for lowcomms functionality per
message and not per page allocation entry. A refcount handling for the
handle was added to keep the message alive until the user doesn't need
it anymore.

There exists now a per message callback which will be called when
allocating a new buffer. This callback will be guaranteed to be called
according the order of the sending buffer, which can be used that the
caller increments a sequence number.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/lowcomms.c | 100 +++++++++++++++++++++++++++++++++++++++++-----
 fs/dlm/lowcomms.h |   5 ++-
 fs/dlm/midcomms.c |   8 +++-
 3 files changed, 101 insertions(+), 12 deletions(-)

diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index 166e36fcf3e4..ade1a5266e4a 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c
@@ -114,6 +114,17 @@ struct writequeue_entry {
 	int end;
 	int users;
 	struct connection *con;
+	struct list_head msgs;
+	struct kref ref;
+};
+
+struct dlm_msg {
+	struct writequeue_entry *entry;
+	void *ppc;
+	int len;
+
+	struct list_head list;
+	struct kref ref;
 };
 
 struct dlm_node_addr {
@@ -977,12 +988,36 @@ static int accept_from_sock(struct listen_connection *con)
 	return result;
 }
 
-static void free_entry(struct writequeue_entry *e)
+static void dlm_page_release(struct kref *kref)
 {
+	struct writequeue_entry *e = container_of(kref, struct writequeue_entry,
+						  ref);
+
 	__free_page(e->page);
 	kfree(e);
 }
 
+static void dlm_msg_release(struct kref *kref)
+{
+	struct dlm_msg *msg = container_of(kref, struct dlm_msg, ref);
+
+	kref_put(&msg->entry->ref, dlm_page_release);
+	kfree(msg);
+}
+
+static void free_entry(struct writequeue_entry *e)
+{
+	struct dlm_msg *msg, *tmp;
+
+	list_for_each_entry_safe(msg, tmp, &e->msgs, list) {
+		list_del(&msg->list);
+		kref_put(&msg->ref, dlm_msg_release);
+	}
+
+	list_del(&e->list);
+	kref_put(&e->ref, dlm_page_release);
+}
+
 /*
  * writequeue_entry_complete - try to delete and free write queue entry
  * @e: write queue entry to try to delete
@@ -995,10 +1030,8 @@ static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
 	e->offset += completed;
 	e->len -= completed;
 
-	if (e->len == 0 && e->users == 0) {
-		list_del(&e->list);
+	if (e->len == 0 && e->users == 0)
 		free_entry(e);
-	}
 }
 
 /*
@@ -1364,12 +1397,16 @@ static struct writequeue_entry *new_writequeue_entry(struct connection *con,
 
 	entry->con = con;
 	entry->users = 1;
+	kref_init(&entry->ref);
+	INIT_LIST_HEAD(&entry->msgs);
 
 	return entry;
 }
 
 static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
-					     gfp_t allocation, char **ppc)
+					     gfp_t allocation, char **ppc,
+					     void (*cb)(void *buf, void *priv),
+					     void *priv)
 {
 	struct writequeue_entry *e;
 
@@ -1377,7 +1414,12 @@ static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
 	if (!list_empty(&con->writequeue)) {
 		e = list_last_entry(&con->writequeue, struct writequeue_entry, list);
 		if (DLM_WQ_REMAIN_BYTES(e) >= len) {
+			kref_get(&e->ref);
+
 			*ppc = page_address(e->page) + e->end;
+			if (cb)
+				cb(*ppc, priv);
+
 			e->end += len;
 			e->users++;
 			spin_unlock(&con->writequeue_lock);
@@ -1391,19 +1433,26 @@ static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
 	if (!e)
 		return NULL;
 
+	kref_get(&e->ref);
 	*ppc = page_address(e->page);
 	e->end += len;
 
 	spin_lock(&con->writequeue_lock);
+	if (cb)
+		cb(*ppc, priv);
+
 	list_add_tail(&e->list, &con->writequeue);
 	spin_unlock(&con->writequeue_lock);
 
 	return e;
 };
 
-void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
+void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
+			      void (*cb)(void *buf, void *priv), void *priv)
 {
+	struct writequeue_entry *e;
 	struct connection *con;
+	struct dlm_msg *msg;
 
 	if (len > DEFAULT_BUFFER_SIZE ||
 	    len < sizeof(struct dlm_header)) {
@@ -1417,16 +1466,36 @@ void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
 	if (!con)
 		return NULL;
 
-	return new_wq_entry(con, len, allocation, ppc);
+	msg = kzalloc(sizeof(*msg), allocation);
+	if (!msg)
+		return NULL;
+
+	kref_init(&msg->ref);
+
+	e = new_wq_entry(con, len, allocation, ppc, cb, priv);
+	if (!e) {
+		kfree(msg);
+		return NULL;
+	}
+
+	msg->ppc = *ppc;
+	msg->len = len;
+	msg->entry = e;
+
+	return msg;
 }
 
 void dlm_lowcomms_commit_buffer(void *mh)
 {
-	struct writequeue_entry *e = (struct writequeue_entry *)mh;
+	struct dlm_msg *msg = mh;
+	struct writequeue_entry *e = msg->entry;
 	struct connection *con = e->con;
 	int users;
 
 	spin_lock(&con->writequeue_lock);
+	list_add(&msg->list, &e->msgs);
+	kref_get(&msg->ref);
+
 	users = --e->users;
 	if (users)
 		goto out;
@@ -1442,6 +1511,20 @@ void dlm_lowcomms_commit_buffer(void *mh)
 	return;
 }
 
+void dlm_lowcomms_put_buffer(void *mh)
+{
+	struct dlm_msg *msg = mh;
+
+	kref_put(&msg->ref, dlm_msg_release);
+}
+
+void dlm_lowcomms_get_buffer(void *mh)
+{
+	struct dlm_msg *msg = mh;
+
+	kref_get(&msg->ref);
+}
+
 /* Send a message */
 static void send_to_sock(struct connection *con)
 {
@@ -1520,7 +1603,6 @@ static void clean_one_writequeue(struct connection *con)
 
 	spin_lock(&con->writequeue_lock);
 	list_for_each_entry_safe(e, safe, &con->writequeue, list) {
-		list_del(&e->list);
 		free_entry(e);
 	}
 	spin_unlock(&con->writequeue_lock);
diff --git a/fs/dlm/lowcomms.h b/fs/dlm/lowcomms.h
index 48bbc4e18761..fa735497dad8 100644
--- a/fs/dlm/lowcomms.h
+++ b/fs/dlm/lowcomms.h
@@ -22,11 +22,14 @@ void dlm_lowcomms_shutdown(void);
 void dlm_lowcomms_stop(void);
 void dlm_lowcomms_exit(void);
 int dlm_lowcomms_close(int nodeid);
-void *dlm_lowcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc);
+void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
+			      void (*cb)(void *buf, void *priv), void *priv);
 void dlm_lowcomms_commit_buffer(void *mh);
 int dlm_lowcomms_connect_node(int nodeid);
 int dlm_lowcomms_nodes_set_mark(int nodeid, unsigned int mark);
 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len);
+void dlm_lowcomms_put_buffer(void *mh);
+void dlm_lowcomms_get_buffer(void *mh);
 
 #endif				/* __LOWCOMMS_DOT_H__ */
 
diff --git a/fs/dlm/midcomms.c b/fs/dlm/midcomms.c
index bbcb242e6101..2ea0449a82ab 100644
--- a/fs/dlm/midcomms.c
+++ b/fs/dlm/midcomms.c
@@ -30,23 +30,27 @@
 
 void *dlm_midcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
 {
-	return dlm_lowcomms_get_buffer(nodeid, len, allocation, ppc);
+	return dlm_lowcomms_new_buffer(nodeid, len, allocation, ppc, NULL,
+				       NULL);
 }
 
 void dlm_midcomms_commit_buffer(void *mh)
 {
 	dlm_lowcomms_commit_buffer(mh);
+	dlm_lowcomms_put_buffer(mh);
 }
 
 void *dlm_midcomms_stateless_get_buffer(int nodeid, int len, gfp_t allocation,
 					char **ppc)
 {
-	return dlm_lowcomms_get_buffer(nodeid, len, allocation, ppc);
+	return dlm_lowcomms_new_buffer(nodeid, len, allocation, ppc, NULL,
+				       NULL);
 }
 
 void dlm_midcomms_stateless_commit_buffer(void *mh)
 {
 	dlm_lowcomms_commit_buffer(mh);
+	dlm_lowcomms_put_buffer(mh);
 }
 
 void midcomms_add_member(int nodeid)
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 4/8] fs: dlm: add functionality to re-transmit a message

[Alexander Aring]

This patch introduces a retransmit functionality for a lowcomms message
handle. It's just allocates a new buffer and transmit it again, no
special handling about prioritize it because keeping bytestream in order.

To avoid another connection look some refactor was done to make a new
buffer allocation with a preexisting connection pointer.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/lowcomms.c | 55 ++++++++++++++++++++++++++++++++++-------------
 fs/dlm/lowcomms.h |  1 +
 2 files changed, 41 insertions(+), 15 deletions(-)

diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index ade1a5266e4a..fb04448c4e48 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c
@@ -1447,25 +1447,14 @@ static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
 	return e;
 };
 
-void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
-			      void (*cb)(void *buf, void *priv), void *priv)
+static void *dlm_lowcomms_new_buffer_con(struct connection *con, int len,
+					 gfp_t allocation, char **ppc,
+					 void (*cb)(void *buf, void *priv),
+					 void *priv)
 {
 	struct writequeue_entry *e;
-	struct connection *con;
 	struct dlm_msg *msg;
 
-	if (len > DEFAULT_BUFFER_SIZE ||
-	    len < sizeof(struct dlm_header)) {
-		BUILD_BUG_ON(PAGE_SIZE < DEFAULT_BUFFER_SIZE);
-		log_print("failed to allocate a buffer of size %d", len);
-		WARN_ON(1);
-		return NULL;
-	}
-
-	con = nodeid2con(nodeid, allocation);
-	if (!con)
-		return NULL;
-
 	msg = kzalloc(sizeof(*msg), allocation);
 	if (!msg)
 		return NULL;
@@ -1485,6 +1474,26 @@ void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
 	return msg;
 }
 
+void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
+			      void (*cb)(void *buf, void *priv), void *priv)
+{
+	struct connection *con;
+
+	if (len > DEFAULT_BUFFER_SIZE ||
+	    len < sizeof(struct dlm_header)) {
+		BUILD_BUG_ON(PAGE_SIZE < DEFAULT_BUFFER_SIZE);
+		log_print("failed to allocate a buffer of size %d", len);
+		WARN_ON(1);
+		return NULL;
+	}
+
+	con = nodeid2con(nodeid, allocation);
+	if (!con)
+		return NULL;
+
+	return dlm_lowcomms_new_buffer_con(con, len, GFP_ATOMIC, ppc, cb, priv);
+}
+
 void dlm_lowcomms_commit_buffer(void *mh)
 {
 	struct dlm_msg *msg = mh;
@@ -1525,6 +1534,22 @@ void dlm_lowcomms_get_buffer(void *mh)
 	kref_get(&msg->ref);
 }
 
+void dlm_lowcomms_resend_buffer(void *mh)
+{
+	struct dlm_msg *msg = mh;
+	void *mh_new;
+	char *ppc;
+
+	mh_new = dlm_lowcomms_new_buffer_con(msg->entry->con, msg->len,
+					     GFP_NOFS, &ppc, NULL, NULL);
+	if (!mh_new)
+		return;
+
+	memcpy(ppc, msg->ppc, msg->len);
+	dlm_lowcomms_commit_buffer(mh_new);
+	dlm_lowcomms_put_buffer(mh_new);
+}
+
 /* Send a message */
 static void send_to_sock(struct connection *con)
 {
diff --git a/fs/dlm/lowcomms.h b/fs/dlm/lowcomms.h
index fa735497dad8..345aed7e00cc 100644
--- a/fs/dlm/lowcomms.h
+++ b/fs/dlm/lowcomms.h
@@ -30,6 +30,7 @@ int dlm_lowcomms_nodes_set_mark(int nodeid, unsigned int mark);
 int dlm_lowcomms_addr(int nodeid, struct sockaddr_storage *addr, int len);
 void dlm_lowcomms_put_buffer(void *mh);
 void dlm_lowcomms_get_buffer(void *mh);
+void dlm_lowcomms_resend_buffer(void *mh);
 
 #endif				/* __LOWCOMMS_DOT_H__ */
 
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 5/8] fs: dlm: move out some hash functionality

[Alexander Aring]

This patch moves out some lowcomms hash functionality into lowcomms
header to provide them to other layers like midcomms as well.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/lowcomms.c |  9 ---------
 fs/dlm/lowcomms.h | 10 ++++++++++
 2 files changed, 10 insertions(+), 9 deletions(-)

diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index fb04448c4e48..0e08a59e489c 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c
@@ -59,7 +59,6 @@
 #include "config.h"
 
 #define NEEDED_RMEM (4*1024*1024)
-#define CONN_HASH_SIZE 32
 
 /* Number of messages to send before rescheduling */
 #define MAX_SEND_MSG_COUNT 25
@@ -166,14 +165,6 @@ static void sctp_connect_to_sock(struct connection *con);
 static void tcp_connect_to_sock(struct connection *con);
 static void dlm_tcp_shutdown(struct connection *con);
 
-/* This is deliberately very simple because most clusters have simple
-   sequential nodeids, so we should be able to go straight to a connection
-   struct in the array */
-static inline int nodeid_hash(int nodeid)
-{
-	return nodeid & (CONN_HASH_SIZE-1);
-}
-
 static struct connection *__find_con(int nodeid)
 {
 	int r, idx;
diff --git a/fs/dlm/lowcomms.h b/fs/dlm/lowcomms.h
index 345aed7e00cc..cacfc5620f54 100644
--- a/fs/dlm/lowcomms.h
+++ b/fs/dlm/lowcomms.h
@@ -13,6 +13,16 @@
 #define __LOWCOMMS_DOT_H__
 
 #define LOWCOMMS_MAX_TX_BUFFER_LEN	4096
+#define CONN_HASH_SIZE 32
+
+/* This is deliberately very simple because most clusters have simple
+ * sequential nodeids, so we should be able to go straight to a connection
+ * struct in the array
+ */
+static inline int nodeid_hash(int nodeid)
+{
+	return nodeid & (CONN_HASH_SIZE-1);
+}
 
 /* switch to check if dlm is running */
 extern int dlm_allow_conn;
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 6/8] fs: dlm: add union in dlm header for lockspace id

[Alexander Aring]

This patch adds union inside the lockspace id to handle it also for
another use case for a different dlm command.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/dlm_internal.h | 5 ++++-
 fs/dlm/lock.c         | 8 ++++----
 fs/dlm/midcomms.c     | 1 -
 fs/dlm/rcom.c         | 4 ++--
 fs/dlm/util.c         | 6 ++++--
 5 files changed, 14 insertions(+), 10 deletions(-)

diff --git a/fs/dlm/dlm_internal.h b/fs/dlm/dlm_internal.h
index 04fe9f525ac7..917de7367a32 100644
--- a/fs/dlm/dlm_internal.h
+++ b/fs/dlm/dlm_internal.h
@@ -377,7 +377,10 @@ static inline int rsb_flag(struct dlm_rsb *r, enum rsb_flags flag)
 
 struct dlm_header {
 	uint32_t		h_version;
-	uint32_t		h_lockspace;
+	union {
+		/* for DLM_MSG and DLM_RCOM */
+		uint32_t	h_lockspace;
+	} u;
 	uint32_t		h_nodeid;	/* nodeid of sender */
 	uint16_t		h_length;
 	uint8_t			h_cmd;		/* DLM_MSG, DLM_RCOM */
diff --git a/fs/dlm/lock.c b/fs/dlm/lock.c
index b3fd823009f4..daa5747c8556 100644
--- a/fs/dlm/lock.c
+++ b/fs/dlm/lock.c
@@ -3544,7 +3544,7 @@ static int _create_message(struct dlm_ls *ls, int mb_len,
 	ms = (struct dlm_message *) mb;
 
 	ms->m_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
-	ms->m_header.h_lockspace = ls->ls_global_id;
+	ms->m_header.u.h_lockspace = ls->ls_global_id;
 	ms->m_header.h_nodeid = dlm_our_nodeid();
 	ms->m_header.h_length = mb_len;
 	ms->m_header.h_cmd = DLM_MSG;
@@ -5038,16 +5038,16 @@ void dlm_receive_buffer(union dlm_packet *p, int nodeid)
 
 	if (hd->h_nodeid != nodeid) {
 		log_print("invalid h_nodeid %d from %d lockspace %x",
-			  hd->h_nodeid, nodeid, hd->h_lockspace);
+			  hd->h_nodeid, nodeid, hd->u.h_lockspace);
 		return;
 	}
 
-	ls = dlm_find_lockspace_global(hd->h_lockspace);
+	ls = dlm_find_lockspace_global(hd->u.h_lockspace);
 	if (!ls) {
 		if (dlm_config.ci_log_debug) {
 			printk_ratelimited(KERN_DEBUG "dlm: invalid lockspace "
 				"%u from %d cmd %d type %d\n",
-				hd->h_lockspace, nodeid, hd->h_cmd, type);
+				hd->u.h_lockspace, nodeid, hd->h_cmd, type);
 		}
 
 		if (hd->h_cmd == DLM_RCOM && type == DLM_RCOM_STATUS)
diff --git a/fs/dlm/midcomms.c b/fs/dlm/midcomms.c
index 2ea0449a82ab..d86dcc95f736 100644
--- a/fs/dlm/midcomms.c
+++ b/fs/dlm/midcomms.c
@@ -149,4 +149,3 @@ int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 
 	return ret;
 }
-
diff --git a/fs/dlm/rcom.c b/fs/dlm/rcom.c
index 7a7d4a8e4706..06f7a5f1d99d 100644
--- a/fs/dlm/rcom.c
+++ b/fs/dlm/rcom.c
@@ -49,7 +49,7 @@ static int create_rcom(struct dlm_ls *ls, int to_nodeid, int type, int len,
 	rc = (struct dlm_rcom *) mb;
 
 	rc->rc_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
-	rc->rc_header.h_lockspace = ls->ls_global_id;
+	rc->rc_header.u.h_lockspace = ls->ls_global_id;
 	rc->rc_header.h_nodeid = dlm_our_nodeid();
 	rc->rc_header.h_length = mb_len;
 	rc->rc_header.h_cmd = DLM_RCOM;
@@ -476,7 +476,7 @@ int dlm_send_ls_not_ready(int nodeid, struct dlm_rcom *rc_in)
 	rc = (struct dlm_rcom *) mb;
 
 	rc->rc_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
-	rc->rc_header.h_lockspace = rc_in->rc_header.h_lockspace;
+	rc->rc_header.u.h_lockspace = rc_in->rc_header.u.h_lockspace;
 	rc->rc_header.h_nodeid = dlm_our_nodeid();
 	rc->rc_header.h_length = mb_len;
 	rc->rc_header.h_cmd = DLM_RCOM;
diff --git a/fs/dlm/util.c b/fs/dlm/util.c
index 74a8c5bfe9b5..58acbcc2081a 100644
--- a/fs/dlm/util.c
+++ b/fs/dlm/util.c
@@ -23,7 +23,8 @@
 void header_out(struct dlm_header *hd)
 {
 	hd->h_version		= cpu_to_le32(hd->h_version);
-	hd->h_lockspace		= cpu_to_le32(hd->h_lockspace);
+	/* does it for others u32 in union as well */
+	hd->u.h_lockspace	= cpu_to_le32(hd->u.h_lockspace);
 	hd->h_nodeid		= cpu_to_le32(hd->h_nodeid);
 	hd->h_length		= cpu_to_le16(hd->h_length);
 }
@@ -31,7 +32,8 @@ void header_out(struct dlm_header *hd)
 void header_in(struct dlm_header *hd)
 {
 	hd->h_version		= le32_to_cpu(hd->h_version);
-	hd->h_lockspace		= le32_to_cpu(hd->h_lockspace);
+	/* does it for others u32 in union as well */
+	hd->u.h_lockspace	= le32_to_cpu(hd->u.h_lockspace);
 	hd->h_nodeid		= le32_to_cpu(hd->h_nodeid);
 	hd->h_length		= le16_to_cpu(hd->h_length);
 }
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 7/8] fs: dlm: add reliable connection if reconnect

[Alexander Aring]

This patch introduce to make a tcp lowcomms connection reliable even if
reconnects occurs. This is done by an application layer re-transmission
handling and sequence numbers in dlm protocols. There are three new dlm
commands:

DLM_OPTS:

This will encapsulate an existing dlm message (and rcom message if they
don't have an own application side re-transmission handling). As optional
handling additional tlv's (type length fields) can be appended. This can
be for example a sequence number field. However because in DLM_OPTS the
lockspace field is unused and a sequence number is a mandatory field it
isn't made as a tlv and we put the sequence number inside the lockspace
id. The possibility to add optional options are still there for future
purposes.

DLM_ACK:

Just a dlm header to acknowledge the receive of a DLM_OPTS message to
it's sender.

DLM_FIN:

A new DLM message to synchronize pending message to the other dlm end if
the node want to disconnects. Each side waits until it receives this
message and disconnects. It's important that this message has nothing to
do with the application logic because it might run in a timeout if
acknowledge messages are dropped. The problem which we try to solve with
DLM_FIN is that the node membership is handled by the cluster manager and
not the kernel dlm protocol itself, DLM_FIN tries to synchronize the kernel
dlm protocol with cluster manager message handling.

To provide a reliable connection the node will retransmit all
unacknowledges message to it's peer on reconnect. The receiver will then
filtering out the next received message and drop all messages which are
duplicates.

As RCOM_STATUS and RCOM_NAMES messages are the first messages which are
exchanged and they have they own re-transmission handling, there exists
logic that these messages must be first. If these messages arrives we
store the dlm version field. This handling is on DLM 3.1 and after this
patch 3.2 the same. A backwards compatibility handling has been added
which seems to work on tests without tcpkill, however it's not recommended
to use DLM 3.1 and 3.2 at the same time, because DLM 3.2 tries to fix long
term bugs in the DLM protocol.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/dlm_internal.h |   30 +-
 fs/dlm/lockspace.c    |    7 +-
 fs/dlm/lowcomms.c     |    5 +-
 fs/dlm/lowcomms.h     |    7 +-
 fs/dlm/midcomms.c     | 1096 +++++++++++++++++++++++++++++++++++++++--
 fs/dlm/midcomms.h     |    1 +
 fs/dlm/rcom.c         |    4 +-
 7 files changed, 1094 insertions(+), 56 deletions(-)

diff --git a/fs/dlm/dlm_internal.h b/fs/dlm/dlm_internal.h
index 917de7367a32..da2297705713 100644
--- a/fs/dlm/dlm_internal.h
+++ b/fs/dlm/dlm_internal.h
@@ -368,18 +368,26 @@ static inline int rsb_flag(struct dlm_rsb *r, enum rsb_flags flag)
 /* dlm_header is first element of all structs sent between nodes */
 
 #define DLM_HEADER_MAJOR	0x00030000
-#define DLM_HEADER_MINOR	0x00000001
+#define DLM_HEADER_MINOR	0x00000002
+
+#define DLM_VERSION_3_1		0x00030001
+#define DLM_VERSION_3_2		0x00030002
 
 #define DLM_HEADER_SLOTS	0x00000001
 
 #define DLM_MSG			1
 #define DLM_RCOM		2
+#define DLM_OPTS		3
+#define DLM_ACK			4
+#define DLM_FIN			5
 
 struct dlm_header {
 	uint32_t		h_version;
 	union {
 		/* for DLM_MSG and DLM_RCOM */
 		uint32_t	h_lockspace;
+		/* for DLM_ACK and DLM_OPTS */
+		uint32_t	h_seq;
 	} u;
 	uint32_t		h_nodeid;	/* nodeid of sender */
 	uint16_t		h_length;
@@ -387,7 +395,6 @@ struct dlm_header {
 	uint8_t			h_pad;
 };
 
-
 #define DLM_MSG_REQUEST		1
 #define DLM_MSG_CONVERT		2
 #define DLM_MSG_UNLOCK		3
@@ -455,10 +462,29 @@ struct dlm_rcom {
 	char			rc_buf[];
 };
 
+struct dlm_opt_header {
+	uint16_t	t_type;
+	uint16_t	t_length;
+	uint32_t	o_pad;
+	/* need to be 8 byte aligned */
+	char		t_value[];
+};
+
+/* encapsulation header */
+struct dlm_opts {
+	struct dlm_header	o_header;
+	uint8_t			o_nextcmd;
+	uint8_t			o_pad;
+	uint16_t		o_optlen;
+	uint32_t		o_pad2;
+	char			o_opts[];
+};
+
 union dlm_packet {
 	struct dlm_header	header;		/* common to other two */
 	struct dlm_message	message;
 	struct dlm_rcom		rcom;
+	struct dlm_opts		opts;
 };
 
 #define DLM_RSF_NEED_SLOTS	0x00000001
diff --git a/fs/dlm/lockspace.c b/fs/dlm/lockspace.c
index bf5c55ef9d0d..2b738be8d7e4 100644
--- a/fs/dlm/lockspace.c
+++ b/fs/dlm/lockspace.c
@@ -567,7 +567,12 @@ static int new_lockspace(const char *name, const char *cluster,
 	mutex_init(&ls->ls_requestqueue_mutex);
 	mutex_init(&ls->ls_clear_proc_locks);
 
-	ls->ls_recover_buf = kmalloc(LOWCOMMS_MAX_TX_BUFFER_LEN, GFP_NOFS);
+	/* Due backwards compatibility with 3.1 we need to use maximum
+	 * possible dlm message size to be sure the message will fit and
+	 * not having out of bounds issues. However on sending side 3.2
+	 * might send less.
+	 */
+	ls->ls_recover_buf = kmalloc(DEFAULT_BUFFER_SIZE, GFP_NOFS);
 	if (!ls->ls_recover_buf)
 		goto out_lkbidr;
 
diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index 0e08a59e489c..d61a1117b1d7 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c
@@ -955,6 +955,7 @@ static int accept_from_sock(struct listen_connection *con)
 		   result in calling sk_error_report recursively. */
 		add_sock(newsock, newcon);
 		addcon = newcon;
+		dlm_midcomms_unack_msg_resend(nodeid);
 	}
 
 	set_bit(CF_CONNECTED, &addcon->flags);
@@ -1677,8 +1678,10 @@ static void process_send_sockets(struct work_struct *work)
 	struct connection *con = container_of(work, struct connection, swork);
 
 	clear_bit(CF_WRITE_PENDING, &con->flags);
-	if (con->sock == NULL) /* not mutex protected so check it inside too */
+	if (con->sock == NULL) { /* not mutex protected so check it inside too */
 		con->connect_action(con);
+		dlm_midcomms_unack_msg_resend(con->nodeid);
+	}
 	if (!list_empty(&con->writequeue))
 		send_to_sock(con);
 }
diff --git a/fs/dlm/lowcomms.h b/fs/dlm/lowcomms.h
index cacfc5620f54..1554d8c078c9 100644
--- a/fs/dlm/lowcomms.h
+++ b/fs/dlm/lowcomms.h
@@ -12,7 +12,12 @@
 #ifndef __LOWCOMMS_DOT_H__
 #define __LOWCOMMS_DOT_H__
 
-#define LOWCOMMS_MAX_TX_BUFFER_LEN	4096
+#include "dlm_internal.h"
+
+#define DLM_MIDCOMMS_OPT_LEN		sizeof(struct dlm_opts)
+#define LOWCOMMS_MAX_TX_BUFFER_LEN	(DEFAULT_BUFFER_SIZE - \
+					 DLM_MIDCOMMS_OPT_LEN)
+
 #define CONN_HASH_SIZE 32
 
 /* This is deliberately very simple because most clusters have simple
diff --git a/fs/dlm/midcomms.c b/fs/dlm/midcomms.c
index d86dcc95f736..1d6da6ad6f58 100644
--- a/fs/dlm/midcomms.c
+++ b/fs/dlm/midcomms.c
@@ -3,7 +3,7 @@
 *******************************************************************************
 **
 **  Copyright (C) Sistina Software, Inc.  1997-2003  All rights reserved.
-**  Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
+**  Copyright (C) 2004-2021 Red Hat, Inc.  All rights reserved.
 **
 **
 *******************************************************************************
@@ -12,68 +12,665 @@
 /*
  * midcomms.c
  *
- * This is the appallingly named "mid-level" comms layer.
+ * This is the appallingly named "mid-level" comms layer. It takes care about
+ * deliver an on application layer "reliable" communication above the used
+ * lowcomms transport layer.
  *
- * Its purpose is to take packets from the "real" comms layer,
- * split them up into packets and pass them to the interested
- * part of the locking mechanism.
+ * How it works:
  *
- * It also takes messages from the locking layer, formats them
- * into packets and sends them to the comms layer.
+ * Each nodes keeps track of all send DLM messages in send_queue with a sequence
+ * number. The receive will send an DLM_ACK message back for every DLM message
+ * received at the other side. If a reconnect happens in lowcomms we will send
+ * all unacknowledged dlm messages again. The receiving side might drop any already
+ * received message by comparing sequence numbers.
+ *
+ * How version detection works:
+ *
+ * Due the fact that dlm has pre-configured node addresses on every side
+ * it is in it's nature that every side connects at starts to transmit
+ * dlm messages which ends in a race. However DLM_RCOM_NAMES, DLM_RCOM_STATUS
+ * and their replies are the first messages which are exchanges. Due backwards
+ * compatibility these messages are not covered by the midcomms re-transmission
+ * layer. These messages have their own re-transmission handling in the dlm
+ * application layer. The version field of every node will be set on these RCOM
+ * messages as soon as they arrived and the node isn't yet part of the nodes
+ * hash. There exists also logic to detect version mismatched if something weird
+ * going on or the first messages isn't an expected one.
+ *
+ * Termination:
+ *
+ * As this layer was implemented there was a problem discovered during
+ * connection termination. At midcomms_stop() call which occurs at the
+ * last dlm lockspace release the other side was still sending dlm
+ * messages after the connection was terminated. Further investigations
+ * shows that the membership organization is done by the cluster manager
+ * and not the kernel dlm protocol itself. To make sure the node membership
+ * of a lockspace is synchronized with the kernel a DLM_FIN message was
+ * introduced. It is part of a synchronization mechanism to wait for
+ * the specific cluster manager events on both endpoints so a termination
+ * of sockets can be initiated. The "DLM_QUEUE_TIMEOUT" defines some
+ * timeout handling to wait for some synchronization points at
+ * connection termination.
+ *
+ * Future improvements:
+ *
+ * There exists some known issues/improvements of the dlm handling. Some
+ * of them should be done in a next major dlm version bump which makes
+ * it incompatible with previous versions.
+ *
+ * Unaligned memory access:
+ *
+ * There exists cases when the dlm message buffer length is not aligned
+ * to 8 byte. However seems nobody detected any problem with it. This
+ * can be fixed in the next major version bump of dlm.
+ *
+ * Version detection:
+ *
+ * The version detection and how it's done is related to backwards
+ * compatibility. There exists better ways to make a better handling.
+ * However this should be changed in the next major version bump of dlm.
+ *
+ * Ack handling:
+ *
+ * Currently we send an ack message for every dlm message. However we
+ * can ack multiple dlm messages with one ack by just delaying the ack
+ * message. Will reduce some traffic but makes the drop detection slower.
+ *
+ * Termination:
+ *
+ * Maybe handle nodes membership of lockspaces in dlm protocol itself,
+ * but that requires still synchronization with the cluster manager.
+ *
+ * Tail Size checking:
+ *
+ * There exists a message tail payload in e.g. DLM_MSG however we don't
+ * check it against the message length yet regarding to the receive buffer
+ * length. That need to be validated.
+ *
+ * Fencing bad nodes:
+ *
+ * At timeout places or weird sequence number behaviours we should send
+ * a fencing request to the cluster manager.
  */
 
+/* Debug switch to enable a 5 seconds sleep waiting of a termination.
+ * This can be useful to test fencing while termination is running.
+ * This requires a setup with only gfs2 as dlm user, so that the
+ * last umount will terminate the connection.
+ *
+ * However it became useful to test, while the 5 seconds block in umount
+ * just press the reset button. In a lot of dropping the termination
+ * process can could take several seconds.
+ */
+#define DLM_DEBUG_FENCE_TERMINATION	0
+
+#include <net/tcp.h>
+
 #include "dlm_internal.h"
 #include "lowcomms.h"
 #include "config.h"
 #include "lock.h"
+#include "util.h"
 #include "midcomms.h"
 
-void *dlm_midcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
+/* init value for sequence numbers for testing purpose only e.g. overflows */
+#define DLM_SEQ_INIT		0
+/* 3 minutes wait to sync ending of dlm */
+#define DLM_QUEUE_TIMEOUT	msecs_to_jiffies(3 * 60 * 1000)
+#define DLM_FIN_TIMEOUT		msecs_to_jiffies(10 * 1000)
+#define DLM_VERSION_NOT_SET	0
+
+struct midcomms_node {
+	int nodeid;
+	uint32_t version;
+	uint32_t seq_send;
+	uint32_t seq_next;
+	/* These queues are unbound because we cannot drop any message in dlm.
+	 * We could send a fence signal for a specific node to the cluster
+	 * manager if queues hits some maximum value, however this handling
+	 * not supported yet.
+	 */
+	struct list_head send_queue;
+	spinlock_t send_queue_lock;
+	atomic_t send_queue_cnt;
+#define DLM_NODE_FLAG_FIN	1
+#define DLM_NODE_FLAG_FIN_ACK   2
+#define DLM_NODE_FLAG_CLOSE	3
+#define DLM_NODE_FLAG_STOP_TX	4
+#define DLM_NODE_FLAG_STOP_RX	5
+	unsigned long flags;
+	wait_queue_head_t fin_wait;
+	/* counts how many lockspaces are using this node
+	 * this refcount is necessary to determine if the
+	 * node wants to disconnect.
+	 */
+	int users;
+
+	struct hlist_node hlist;
+	struct rcu_head rcu;
+};
+
+struct dlm_send_msg {
+	const struct dlm_header *inner_hd;
+	struct midcomms_node *node;
+	struct dlm_opts *opts;
+	uint32_t seq;
+	void *mh;
+
+	void (*ack)(struct midcomms_node *node);
+	/* get_buffer/commit srcu idx exchange */
+	int idx;
+
+	struct list_head list;
+	struct rcu_head rcu;
+};
+
+struct dlm_recv_msg {
+	union dlm_packet *p;
+	uint32_t seq;
+
+	struct list_head list;
+};
+
+static struct hlist_head node_hash[CONN_HASH_SIZE];
+static DEFINE_SPINLOCK(nodes_lock);
+DEFINE_STATIC_SRCU(nodes_srcu);
+
+/* This mutex prevents that midcomms_close() is running while
+ * stop() or remove(). As I experienced invalid memory access
+ * behaviours when DLM_DEBUG_FENCE_TERMINATION is enabled and
+ * resetting machines. I will end in some double deletion in nodes
+ * datastructure.
+ */
+static DEFINE_MUTEX(close_lock);
+
+static struct midcomms_node *__find_node(int nodeid, int r)
 {
-	return dlm_lowcomms_new_buffer(nodeid, len, allocation, ppc, NULL,
-				       NULL);
+	struct midcomms_node *node;
+
+	hlist_for_each_entry_rcu(node, &node_hash[r], hlist) {
+		if (node->nodeid == nodeid)
+			return node;
+	}
+
+	return NULL;
 }
 
-void dlm_midcomms_commit_buffer(void *mh)
+static struct midcomms_node *nodeid2node(int nodeid, gfp_t alloc)
 {
-	dlm_lowcomms_commit_buffer(mh);
-	dlm_lowcomms_put_buffer(mh);
+	struct midcomms_node *node, *tmp;
+	int r = nodeid_hash(nodeid);
+
+	node = __find_node(nodeid, r);
+	if (node || !alloc)
+		return node;
+
+	node = kzalloc(sizeof(*node), alloc);
+	if (!node)
+		return NULL;
+
+	node->nodeid = nodeid;
+	node->seq_next = DLM_SEQ_INIT;
+	node->seq_send = DLM_SEQ_INIT;
+	spin_lock_init(&node->send_queue_lock);
+	atomic_set(&node->send_queue_cnt, 0);
+	INIT_LIST_HEAD(&node->send_queue);
+	init_waitqueue_head(&node->fin_wait);
+
+	spin_lock(&nodes_lock);
+	/* check again if there was somebody else
+	 * earlier here to add the node
+	 */
+	tmp = __find_node(nodeid, r);
+	if (tmp) {
+		spin_unlock(&nodes_lock);
+		kfree(node);
+		return tmp;
+	}
+
+	hlist_add_head_rcu(&node->hlist, &node_hash[r]);
+	spin_unlock(&nodes_lock);
+
+	return node;
 }
 
-void *dlm_midcomms_stateless_get_buffer(int nodeid, int len, gfp_t allocation,
-					char **ppc)
+static int dlm_send_ack(int nodeid, uint32_t seq)
 {
-	return dlm_lowcomms_new_buffer(nodeid, len, allocation, ppc, NULL,
-				       NULL);
+	int mb_len = sizeof(struct dlm_header);
+	struct dlm_header *m_header;
+	char *ppc;
+	void *mh;
+
+	mh = dlm_midcomms_stateless_get_buffer(nodeid, mb_len, GFP_NOFS, &ppc);
+	if (!mh)
+		return -ENOMEM;
+
+	m_header = (struct dlm_header *)ppc;
+
+	m_header->h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
+	m_header->h_nodeid = dlm_our_nodeid();
+	m_header->h_length = mb_len;
+	m_header->h_cmd = DLM_ACK;
+	m_header->u.h_seq = seq;
+
+	header_out(m_header);
+	dlm_midcomms_stateless_commit_buffer(mh);
+
+	return 0;
 }
 
-void dlm_midcomms_stateless_commit_buffer(void *mh)
+static int dlm_send_fin(int nodeid, void (*ack)(struct midcomms_node *node))
 {
-	dlm_lowcomms_commit_buffer(mh);
-	dlm_lowcomms_put_buffer(mh);
+	int mb_len = sizeof(struct dlm_header);
+	struct dlm_header *m_header;
+	struct dlm_send_msg *msg;
+	char *ppc;
+
+	msg = dlm_midcomms_get_buffer(nodeid, mb_len, GFP_NOFS, &ppc);
+	if (!msg)
+		return -ENOMEM;
+
+	msg->ack = ack;
+
+	m_header = (struct dlm_header *)ppc;
+
+	m_header->h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
+	m_header->h_nodeid = dlm_our_nodeid();
+	m_header->h_length = mb_len;
+	m_header->h_cmd = DLM_FIN;
+
+	header_out(m_header);
+	dlm_midcomms_commit_buffer(msg);
+
+	return 0;
 }
 
-void midcomms_add_member(int nodeid)
+static void dlm_send_msg_release(struct rcu_head *rcu)
 {
+	struct dlm_send_msg *msg = container_of(rcu, struct dlm_send_msg, rcu);
+
+	dlm_lowcomms_put_buffer(msg->mh);
+	kfree(msg);
 }
 
-void midcomms_remove_member(int nodeid)
+/* send_queue_lock must be held */
+static void dlm_send_msg_del(struct midcomms_node *node,
+			     struct dlm_send_msg *msg)
 {
+	list_del_rcu(&msg->list);
+
+	if (msg->ack)
+		msg->ack(node);
+
+	if (atomic_dec_and_test(&node->send_queue_cnt))
+		wake_up(&node->fin_wait);
+
+	call_rcu(&msg->rcu, dlm_send_msg_release);
 }
 
-int dlm_midcomms_close(int nodeid)
+static void dlm_send_queue_flush(struct midcomms_node *node)
 {
-	return dlm_lowcomms_close(nodeid);
+	struct dlm_send_msg *msg;
+
+	rcu_read_lock();
+	spin_lock(&node->send_queue_lock);
+	list_for_each_entry_rcu(msg, &node->send_queue, list) {
+		dlm_send_msg_del(node, msg);
+	}
+	spin_unlock(&node->send_queue_lock);
+	rcu_read_unlock();
 }
 
-int dlm_midcomms_start(void)
+static void dlm_midcomms_seq_receive_buffer(struct midcomms_node *node,
+					    union dlm_packet *p)
 {
-	return dlm_lowcomms_start();
+	dlm_receive_buffer(p, node->nodeid);
+	node->seq_next++;
 }
 
-void dlm_midcomms_shutdown(void)
+static void dlm_receive_ack(struct midcomms_node *node, uint32_t seq)
 {
-	dlm_lowcomms_shutdown();
+	struct dlm_send_msg *msg;
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(msg, &node->send_queue, list) {
+		if (before(msg->seq, seq)) {
+			spin_lock(&node->send_queue_lock);
+			dlm_send_msg_del(node, msg);
+			spin_unlock(&node->send_queue_lock);
+		}
+	}
+	rcu_read_unlock();
+}
+
+static void dlm_midcomms_receive_buffer(union dlm_packet *p,
+					struct midcomms_node *node,
+					uint32_t seq)
+{
+	if (seq == node->seq_next) {
+		dlm_midcomms_seq_receive_buffer(node, p);
+	} else {
+		log_print("ignore dlm msg because seq mismatch, seq: %u, expected: %u, nodeid: %d",
+			  seq, node->seq_next, node->nodeid);
+	}
+
+	dlm_send_ack(node->nodeid, node->seq_next);
+}
+
+static struct midcomms_node *
+dlm_midcomms_recv_node_lookup(int nodeid, const union dlm_packet *p,
+			      uint16_t msglen, int (*cb)(struct midcomms_node *node))
+{
+	struct midcomms_node *node;
+	gfp_t allocation = 0;
+	int ret;
+
+	switch (p->header.h_cmd) {
+	case DLM_RCOM:
+		if (msglen < sizeof(struct dlm_rcom)) {
+			log_print("rcom msg too small: %u, will skip this message from node %d",
+				  msglen, nodeid);
+			return NULL;
+		}
+
+		switch (le32_to_cpu(p->rcom.rc_type)) {
+		case DLM_RCOM_NAMES:
+			fallthrough;
+		case DLM_RCOM_NAMES_REPLY:
+			fallthrough;
+		case DLM_RCOM_STATUS:
+			fallthrough;
+		case DLM_RCOM_STATUS_REPLY:
+			allocation = GFP_NOFS;
+			break;
+		default:
+			break;
+		}
+
+		break;
+	default:
+		break;
+	}
+
+	node = nodeid2node(nodeid, allocation);
+	if (!node) {
+		log_print("received dlm message cmd %d from node %d in an invalid sequence",
+			  p->opts.o_nextcmd, nodeid);
+		return NULL;
+	}
+
+	if (test_bit(DLM_NODE_FLAG_STOP_RX, &node->flags)) {
+		WARN_ON(1);
+		return NULL;
+	}
+
+	ret = cb(node);
+	if (ret < 0)
+		return NULL;
+
+	return node;
+}
+
+static int dlm_midcomms_version_check_3_2(struct midcomms_node *node)
+{
+	switch (node->version) {
+	case DLM_VERSION_NOT_SET:
+		node->version = DLM_VERSION_3_2;
+		log_print("version 0x%08x for node %d detected", DLM_VERSION_3_2,
+			  node->nodeid);
+		break;
+	case DLM_VERSION_3_2:
+		break;
+	default:
+		log_print("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
+			  DLM_VERSION_3_2, node->nodeid, node->version);
+		return -1;
+	}
+
+	return 0;
+}
+
+#if 0
+static int dlm_parse_opts(void *opts, uint16_t optlen)
+{
+	struct dlm_opt_header *hd;
+	void *ptr = opts;
+	uint16_t length;
+
+	while (optlen > 0) {
+		hd = ptr;
+		switch (le16_to_cpu(hd->t_type)) {
+		/* add options here */
+		default:
+			break;
+		}
+
+		length = le16_to_cpu(hd->t_length);
+		ptr += length;
+		optlen -= length;
+	}
+
+	return 0;
+}
+#endif
+
+static int dlm_opts_check_msglen(union dlm_packet *p, uint16_t msglen, int nodeid)
+{
+	int len = msglen;
+
+	/* we only trust outer header msglen because
+	 * it's checked against receive buffer length.
+	 */
+	if (len < sizeof(struct dlm_opts))
+		return -1;
+	len -= sizeof(struct dlm_opts);
+
+	if (len < le16_to_cpu(p->opts.o_optlen))
+		return -1;
+	len -= le16_to_cpu(p->opts.o_optlen);
+
+	switch (p->opts.o_nextcmd) {
+	case DLM_FIN:
+		if (len < sizeof(struct dlm_header)) {
+			log_print("fin too small: %d, will skip this message from node %d",
+				  len, nodeid);
+			return -1;
+		}
+
+		break;
+	case DLM_MSG:
+		if (len < sizeof(struct dlm_message)) {
+			log_print("msg too small: %d, will skip this message from node %d",
+				  msglen, nodeid);
+			return -1;
+		}
+
+		break;
+	case DLM_RCOM:
+		if (len < sizeof(struct dlm_rcom)) {
+			log_print("rcom msg too small: %d, will skip this message from node %d",
+				  len, nodeid);
+			return -1;
+		}
+
+		break;
+	default:
+		log_print("unsupported o_nextcmd received: %u, will skip this message from node %d",
+			  p->opts.o_nextcmd, nodeid);
+		return -1;
+	}
+
+	return 0;
+}
+
+static void dlm_midcomms_receive_buffer_3_2(union dlm_packet *p, int nodeid)
+{
+	uint16_t msglen = le16_to_cpu(p->header.h_length);
+	struct midcomms_node *node;
+	uint32_t seq;
+	int ret, idx;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
+					     dlm_midcomms_version_check_3_2);
+	if (!node)
+		goto out;
+
+	switch (p->header.h_cmd) {
+	case DLM_RCOM:
+		/* these rcom message we use to determine version.
+		 * they have their own retransmission handling and
+		 * are the first messages of dlm.
+		 *
+		 * length already checked.
+		 */
+		switch (le32_to_cpu(p->rcom.rc_type)) {
+		case DLM_RCOM_NAMES:
+			fallthrough;
+		case DLM_RCOM_NAMES_REPLY:
+			fallthrough;
+		case DLM_RCOM_STATUS:
+			fallthrough;
+		case DLM_RCOM_STATUS_REPLY:
+			break;
+		default:
+			log_print("unsupported rcom type received: %u, will skip this message from node %d",
+				  le32_to_cpu(p->rcom.rc_type), nodeid);
+			goto out;
+		}
+
+		dlm_receive_buffer(p, nodeid);
+		break;
+	case DLM_OPTS:
+		seq = le32_to_cpu(p->header.u.h_seq);
+
+		ret = dlm_opts_check_msglen(p, msglen, nodeid);
+		if (ret < 0) {
+			log_print("opts msg too small: %u, will skip this message from node %d",
+				  msglen, nodeid);
+			goto out;
+		}
+#if 0
+		ret = dlm_parse_opts(p->opts.o_opts, p->opts.o_optlen);
+		if (ret < 0)
+			goto out;
+#endif
+
+		p = (union dlm_packet *)((unsigned char *)p->opts.o_opts +
+					 le16_to_cpu(p->opts.o_optlen));
+
+		/* recheck inner msglen just if it's not garbage */
+		msglen = le16_to_cpu(p->header.h_length);
+		switch (p->header.h_cmd) {
+		case DLM_RCOM:
+			if (msglen < sizeof(struct dlm_rcom)) {
+				log_print("inner rcom msg too small: %u, will skip this message from node %d",
+					  msglen, nodeid);
+				goto out;
+			}
+
+			break;
+		case DLM_MSG:
+			if (msglen < sizeof(struct dlm_message)) {
+				log_print("inner msg too small: %u, will skip this message from node %d",
+					  msglen, nodeid);
+				goto out;
+			}
+
+			break;
+		case DLM_FIN:
+			if (msglen < sizeof(struct dlm_header)) {
+				log_print("inner fin too small: %u, will skip this message from node %d",
+					  msglen, nodeid);
+				goto out;
+			}
+
+			if (seq != node->seq_next) {
+				dlm_send_ack(nodeid, node->seq_next);
+				goto out;
+			}
+
+			dlm_send_ack(nodeid, seq + 1);
+
+			set_bit(DLM_NODE_FLAG_FIN, &node->flags);
+			wake_up(&node->fin_wait);
+			/* midcomms only */
+			goto out;
+		default:
+			log_print("unsupported inner h_cmd received: %u, will skip this message from node %d",
+				  msglen, nodeid);
+			goto out;
+		}
+
+		dlm_midcomms_receive_buffer(p, node, seq);
+		break;
+	case DLM_ACK:
+		seq = le32_to_cpu(p->header.u.h_seq);
+		dlm_receive_ack(node, seq);
+		break;
+	default:
+		log_print("unsupported h_cmd received: %u, will skip this message from node %d",
+			  p->header.h_cmd, nodeid);
+		break;
+	}
+
+out:
+	srcu_read_unlock(&nodes_srcu, idx);
+}
+
+static int dlm_midcomms_version_check_3_1(struct midcomms_node *node)
+{
+	switch (node->version) {
+	case DLM_VERSION_NOT_SET:
+		node->version = DLM_VERSION_3_1;
+		log_print("version 0x%08x for node %d detected", DLM_VERSION_3_1,
+			  node->nodeid);
+		break;
+	case DLM_VERSION_3_1:
+		break;
+	default:
+		log_print("version mismatch detected, assumed 0x%08x but node %d has 0x%08x",
+			  DLM_VERSION_3_1, node->nodeid, node->version);
+		return -1;
+	}
+
+	return 0;
+}
+
+static void dlm_midcomms_receive_buffer_3_1(union dlm_packet *p, int nodeid)
+{
+	uint16_t msglen = le16_to_cpu(p->header.h_length);
+	struct midcomms_node *node;
+	int idx;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	node = dlm_midcomms_recv_node_lookup(nodeid, p, msglen,
+					     dlm_midcomms_version_check_3_1);
+	if (!node) {
+		srcu_read_unlock(&nodes_srcu, idx);
+		return;
+	}
+	srcu_read_unlock(&nodes_srcu, idx);
+
+	switch (p->header.h_cmd) {
+	case DLM_RCOM:
+		/* length already checked */
+		break;
+	case DLM_MSG:
+		if (msglen < sizeof(struct dlm_message)) {
+			log_print("msg too small: %u, will skip this message from node %d",
+				  msglen, nodeid);
+			return;
+		}
+
+		break;
+	default:
+		log_print("unsupported h_cmd received: %u, will skip this message from node %d",
+			  p->header.h_cmd, nodeid);
+		return;
+	}
+
+	dlm_receive_buffer(p, nodeid);
 }
 
 /*
@@ -116,32 +713,19 @@ int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 		if (msglen > len)
 			break;
 
-		switch (hd->h_cmd) {
-		case DLM_MSG:
-			if (msglen < sizeof(struct dlm_message)) {
-				log_print("dlm msg too small: %u, will skip this message",
-					  msglen);
-				goto skip;
-			}
-
+		switch (le32_to_cpu(hd->h_version)) {
+		case DLM_VERSION_3_1:
+			dlm_midcomms_receive_buffer_3_1((union dlm_packet *)ptr, nodeid);
 			break;
-		case DLM_RCOM:
-			if (msglen < sizeof(struct dlm_rcom)) {
-				log_print("dlm rcom msg too small: %u, will skip this message",
-					  msglen);
-				goto skip;
-			}
-
+		case DLM_VERSION_3_2:
+			dlm_midcomms_receive_buffer_3_2((union dlm_packet *)ptr, nodeid);
 			break;
 		default:
-			log_print("unsupported h_cmd received: %u, will skip this message",
-				  hd->h_cmd);
-			goto skip;
+			log_print("received invalid version header: %u from node %d, will skip this message",
+				  le32_to_cpu(hd->h_version), nodeid);
+			break;
 		}
 
-		dlm_receive_buffer((union dlm_packet *)ptr, nodeid);
-
-skip:
 		ret += msglen;
 		len -= msglen;
 		ptr += msglen;
@@ -149,3 +733,417 @@ int dlm_process_incoming_buffer(int nodeid, unsigned char *buf, int len)
 
 	return ret;
 }
+
+void dlm_midcomms_unack_msg_resend(int nodeid)
+{
+	struct midcomms_node *node;
+	struct dlm_send_msg *msg;
+	int idx;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	node = nodeid2node(nodeid, 0);
+	if (!node) {
+		srcu_read_unlock(&nodes_srcu, idx);
+		return;
+	}
+
+	/* old protocol, we don't support to retransmit on failure */
+	switch (node->version) {
+	case DLM_VERSION_3_2:
+		break;
+	default:
+		srcu_read_unlock(&nodes_srcu, idx);
+		return;
+	}
+
+	rcu_read_lock();
+	list_for_each_entry_rcu(msg, &node->send_queue, list) {
+		log_print("retransmit dlm msg, seq %u, nodeid %d",
+			  msg->seq, node->nodeid);
+		dlm_lowcomms_resend_buffer(msg->mh);
+	}
+	rcu_read_unlock();
+	srcu_read_unlock(&nodes_srcu, idx);
+}
+
+void *dlm_midcomms_stateless_get_buffer(int nodeid, int len, gfp_t allocation,
+					char **ppc)
+{
+	return dlm_lowcomms_new_buffer(nodeid, len, allocation,
+				       ppc, NULL, NULL);
+}
+
+void dlm_midcomms_stateless_commit_buffer(void *mh)
+{
+	dlm_lowcomms_commit_buffer(mh);
+	dlm_lowcomms_put_buffer(mh);
+}
+
+static void dlm_fill_opts_header(struct dlm_opts *opts, uint16_t inner_len,
+				 uint32_t seq)
+{
+	opts->o_header.h_cmd = DLM_OPTS;
+	opts->o_header.h_version = (DLM_HEADER_MAJOR | DLM_HEADER_MINOR);
+	opts->o_header.h_nodeid = dlm_our_nodeid();
+	opts->o_header.h_length = DLM_MIDCOMMS_OPT_LEN + inner_len;
+	opts->o_header.u.h_seq = seq;
+	header_out(&opts->o_header);
+}
+
+static void midcomms_get_buffer_cb(void *buf, void *priv)
+{
+	struct dlm_send_msg *msg = priv;
+
+	atomic_inc(&msg->node->send_queue_cnt);
+
+	spin_lock(&msg->node->send_queue_lock);
+	list_add_tail_rcu(&msg->list, &msg->node->send_queue);
+	spin_unlock(&msg->node->send_queue_lock);
+
+	msg->seq = msg->node->seq_send++;
+}
+
+static void *dlm_midcomms_get_buffer_3_2(struct dlm_send_msg *msg, int nodeid,
+					 int len, gfp_t allocation, char **ppc)
+{
+	struct dlm_opts *opts;
+	void *mh;
+
+	mh = dlm_lowcomms_new_buffer(nodeid, len + DLM_MIDCOMMS_OPT_LEN,
+				     allocation, ppc, midcomms_get_buffer_cb, msg);
+	if (!mh)
+		return NULL;
+
+	opts = (struct dlm_opts *)*ppc;
+	msg->opts = opts;
+
+	/* add possible options here */
+	dlm_fill_opts_header(opts, len, msg->seq);
+
+	*ppc += sizeof(*opts);
+	msg->inner_hd = (const struct dlm_header *)*ppc;
+	return mh;
+}
+
+void *dlm_midcomms_get_buffer(int nodeid, int len, gfp_t allocation, char **ppc)
+{
+	struct midcomms_node *node;
+	struct dlm_send_msg *msg;
+	void *mh;
+	int idx;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	node = nodeid2node(nodeid, 0);
+	if (!node) {
+		WARN_ON_ONCE(1);
+		goto err;
+	}
+
+	if (test_bit(DLM_NODE_FLAG_STOP_TX, &node->flags)) {
+		WARN_ON(1);
+		goto err;
+	}
+
+	msg = kzalloc(sizeof(*msg), GFP_NOFS);
+	if (!msg)
+		goto err;
+
+	msg->idx = idx;
+	msg->node = node;
+
+	switch (node->version) {
+	case DLM_VERSION_3_1:
+		mh = dlm_midcomms_stateless_get_buffer(nodeid, len, allocation,
+						       ppc);
+		if (!mh) {
+			kfree(msg);
+			goto err;
+		}
+
+		break;
+	case DLM_VERSION_3_2:
+		mh = dlm_midcomms_get_buffer_3_2(msg, nodeid, len, allocation,
+						 ppc);
+		if (!mh) {
+			kfree(msg);
+			goto err;
+		}
+
+		break;
+	default:
+		kfree(msg);
+		WARN_ON(1);
+		goto err;
+	}
+
+	msg->mh = mh;
+
+	/* keep in mind that is a must to call
+	 * dlm_midcomms_commit_buffer() which releases
+	 * nodes_srcu using msg->idx which is assumed
+	 * here that the application will call it.
+	 */
+	return msg;
+
+err:
+	srcu_read_unlock(&nodes_srcu, idx);
+	return NULL;
+}
+
+static void dlm_midcomms_commit_buffer_3_2(struct dlm_send_msg *msg,
+					   struct midcomms_node *node)
+{
+	/* nexthdr chain for fast lookup */
+	msg->opts->o_nextcmd = msg->inner_hd->h_cmd;
+	dlm_lowcomms_commit_buffer(msg->mh);
+}
+
+static void dlm_midcomms_commit_buffer_3_1(void *mh)
+{
+	dlm_midcomms_stateless_commit_buffer(mh);
+}
+
+void dlm_midcomms_commit_buffer(void *mh)
+{
+	struct dlm_send_msg *msg = mh;
+
+	switch (msg->node->version) {
+	case DLM_VERSION_3_1:
+		srcu_read_unlock(&nodes_srcu, msg->idx);
+
+		dlm_midcomms_commit_buffer_3_1(msg->mh);
+		/* msg is not part of rcu list in this case */
+		kfree(msg);
+		break;
+	case DLM_VERSION_3_2:
+		dlm_midcomms_commit_buffer_3_2(msg, msg->node);
+		srcu_read_unlock(&nodes_srcu, msg->idx);
+		break;
+	default:
+		srcu_read_unlock(&nodes_srcu, msg->idx);
+		WARN_ON(1);
+		break;
+	}
+}
+
+void midcomms_add_member(int nodeid)
+{
+	struct midcomms_node *node;
+	int idx;
+
+	if (nodeid == dlm_our_nodeid())
+		return;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	node = nodeid2node(nodeid, GFP_NOFS);
+	if (!node) {
+		srcu_read_unlock(&nodes_srcu, idx);
+		return;
+	}
+
+	node->users++;
+	srcu_read_unlock(&nodes_srcu, idx);
+}
+
+static void dlm_fin_ack(struct midcomms_node *node)
+{
+	set_bit(DLM_NODE_FLAG_FIN_ACK, &node->flags);
+	wake_up(&node->fin_wait);
+}
+
+static void midcomms_drain_and_fin(struct midcomms_node *node, bool active)
+{
+	int ret;
+
+	/* old protocol, we don't wait for pending operations */
+	switch (node->version) {
+	case DLM_VERSION_3_2:
+		break;
+	default:
+		return;
+	}
+
+	if (DLM_DEBUG_FENCE_TERMINATION)
+		msleep(5000);
+
+	ret = wait_event_timeout(node->fin_wait,
+				 !atomic_read(&node->send_queue_cnt) ||
+				 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
+				 DLM_QUEUE_TIMEOUT);
+	if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
+		goto timed_out;
+
+	if (active) {
+		dlm_send_fin(node->nodeid, dlm_fin_ack);
+		set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
+
+		/* wait for other side dlm + fin */
+		ret = wait_event_timeout(node->fin_wait,
+					 (!atomic_read(&node->send_queue_cnt) &&
+					  test_bit(DLM_NODE_FLAG_FIN_ACK, &node->flags) &&
+					  test_bit(DLM_NODE_FLAG_FIN, &node->flags)) ||
+					 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
+					 DLM_FIN_TIMEOUT);
+		if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
+			goto timed_out;
+	} else {
+		/* wait other side is finished by it's fin */
+		ret = wait_event_timeout(node->fin_wait,
+					 test_bit(DLM_NODE_FLAG_FIN, &node->flags) ||
+					 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
+					 DLM_FIN_TIMEOUT);
+		if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
+			goto timed_out;
+
+		dlm_send_fin(node->nodeid, dlm_fin_ack);
+		set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
+
+		/* wait fin ack is back */
+		ret = wait_event_timeout(node->fin_wait,
+					 (!atomic_read(&node->send_queue_cnt) &&
+					  test_bit(DLM_NODE_FLAG_FIN_ACK, &node->flags)) ||
+					 test_bit(DLM_NODE_FLAG_CLOSE, &node->flags),
+					 DLM_FIN_TIMEOUT);
+		if (!ret || test_bit(DLM_NODE_FLAG_CLOSE, &node->flags))
+			goto timed_out;
+	}
+
+	set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
+
+	return;
+
+timed_out:
+	if (!ret)
+		log_print("timed out from node %d", node->nodeid);
+
+	dlm_send_queue_flush(node);
+	synchronize_rcu();
+}
+
+static void midcomms_node_release(struct rcu_head *rcu)
+{
+	struct midcomms_node *node = container_of(rcu, struct midcomms_node, rcu);
+
+	/* warn if we have anything in queues which indicates a bug */
+	WARN_ON(atomic_read(&node->send_queue_cnt));
+
+	kfree(node);
+}
+
+void midcomms_remove_member(int nodeid)
+{
+	struct midcomms_node *node;
+	int idx;
+
+	if (nodeid == dlm_our_nodeid())
+		return;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	mutex_lock(&close_lock);
+	node = nodeid2node(nodeid, 0);
+	if (!node) {
+		mutex_unlock(&close_lock);
+		srcu_read_unlock(&nodes_srcu, idx);
+		return;
+	}
+
+	node->users--;
+
+	/* hitting users count to zero means the
+	 * other side is running dlm_midcomms_stop()
+	 * we meet us to have a clean disconnect.
+	 */
+	if (!node->users) {
+		midcomms_drain_and_fin(node, false);
+
+		spin_lock(&nodes_lock);
+		hlist_del_rcu(&node->hlist);
+		spin_unlock(&nodes_lock);
+
+		call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
+	}
+	mutex_unlock(&close_lock);
+	srcu_read_unlock(&nodes_srcu, idx);
+}
+
+int dlm_midcomms_start(void)
+{
+	int i;
+
+	for (i = 0; i < CONN_HASH_SIZE; i++)
+		INIT_HLIST_HEAD(&node_hash[i]);
+
+	return dlm_lowcomms_start();
+}
+
+void dlm_midcomms_shutdown(void)
+{
+	struct midcomms_node *node;
+	int i, idx;
+
+	mutex_lock(&close_lock);
+	idx = srcu_read_lock(&nodes_srcu);
+	for (i = 0; i < CONN_HASH_SIZE; i++) {
+		hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
+			midcomms_drain_and_fin(node, true);
+		}
+	}
+
+	dlm_lowcomms_shutdown();
+
+	for (i = 0; i < CONN_HASH_SIZE; i++) {
+		hlist_for_each_entry_rcu(node, &node_hash[i], hlist) {
+			spin_lock(&nodes_lock);
+			hlist_del_rcu(&node->hlist);
+			spin_unlock(&nodes_lock);
+
+			call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
+		}
+	}
+	srcu_read_unlock(&nodes_srcu, idx);
+	mutex_unlock(&close_lock);
+}
+
+int dlm_midcomms_close(int nodeid)
+{
+	struct midcomms_node *node;
+	int ret, idx;
+
+	if (nodeid == dlm_our_nodeid())
+		return 0;
+
+	idx = srcu_read_lock(&nodes_srcu);
+	/* Abort pending close/remove operation */
+	node = nodeid2node(nodeid, 0);
+	if (node) {
+		set_bit(DLM_NODE_FLAG_STOP_TX, &node->flags);
+		set_bit(DLM_NODE_FLAG_CLOSE, &node->flags);
+		wake_up(&node->fin_wait);
+	}
+
+	mutex_lock(&close_lock);
+	node = nodeid2node(nodeid, 0);
+	if (!node) {
+		mutex_unlock(&close_lock);
+		srcu_read_unlock(&nodes_srcu, idx);
+		return dlm_lowcomms_close(nodeid);
+	}
+
+	spin_lock(&nodes_lock);
+	hlist_del_rcu(&node->hlist);
+	spin_unlock(&nodes_lock);
+
+	dlm_send_queue_flush(node);
+	synchronize_rcu();
+	srcu_read_unlock(&nodes_srcu, idx);
+
+	/* wait until all readers left srcu */
+	synchronize_srcu(&nodes_srcu);
+	/* wait until there are all readers left e.g. timer */
+	mutex_unlock(&close_lock);
+
+	ret = dlm_lowcomms_close(nodeid);
+	set_bit(DLM_NODE_FLAG_STOP_RX, &node->flags);
+	call_srcu(&nodes_srcu, &node->rcu, midcomms_node_release);
+	return ret;
+}
diff --git a/fs/dlm/midcomms.h b/fs/dlm/midcomms.h
index 6ce2f2947475..a81bfb6abb4b 100644
--- a/fs/dlm/midcomms.h
+++ b/fs/dlm/midcomms.h
@@ -23,6 +23,7 @@ void midcomms_remove_member(int nodeid);
 int dlm_midcomms_close(int nodeid);
 int dlm_midcomms_start(void);
 void dlm_midcomms_shutdown(void);
+void dlm_midcomms_unack_msg_resend(int nodeid);
 
 #endif				/* __MIDCOMMS_DOT_H__ */
 
diff --git a/fs/dlm/rcom.c b/fs/dlm/rcom.c
index 06f7a5f1d99d..d731d9e65c3b 100644
--- a/fs/dlm/rcom.c
+++ b/fs/dlm/rcom.c
@@ -168,7 +168,7 @@ int dlm_rcom_status(struct dlm_ls *ls, int nodeid, uint32_t status_flags)
 	set_rcom_status(ls, (struct rcom_status *)rc->rc_buf, status_flags);
 
 	allow_sync_reply(ls, &rc->rc_id);
-	memset(ls->ls_recover_buf, 0, LOWCOMMS_MAX_TX_BUFFER_LEN);
+	memset(ls->ls_recover_buf, 0, DEFAULT_BUFFER_SIZE);
 
 	send_rcom(ls, mh, rc, true);
 
@@ -291,7 +291,7 @@ int dlm_rcom_names(struct dlm_ls *ls, int nodeid, char *last_name, int last_len)
 	memcpy(rc->rc_buf, last_name, last_len);
 
 	allow_sync_reply(ls, &rc->rc_id);
-	memset(ls->ls_recover_buf, 0, LOWCOMMS_MAX_TX_BUFFER_LEN);
+	memset(ls->ls_recover_buf, 0, DEFAULT_BUFFER_SIZE);
 
 	send_rcom(ls, mh, rc, true);
 
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 8/8] fs: dlm: don't allow half transmitted messages

[Alexander Aring]

This patch will clean a dirty page buffer if a reconnect occurs. If a page
buffer was half transmitted we cannot start inside the middle of a dlm
message if a node connects again. I observed invalid length receptions
errors and was guessing that this behaviour occurs, after this patch I
never saw an invalid message length again. This patch might drops more
messages for dlm version 3.1 but 3.1 can't deal with half messages as
well, for 3.2 it might trigger more re-transmissions but will not leave dlm
in a broken state.

Signed-off-by: Alexander Aring <aahringo@redhat.com>
---
 fs/dlm/lowcomms.c | 84 ++++++++++++++++++++++++++++++-----------------
 1 file changed, 54 insertions(+), 30 deletions(-)

diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
index d61a1117b1d7..7d9874ed03ef 100644
--- a/fs/dlm/lowcomms.c
+++ b/fs/dlm/lowcomms.c
@@ -112,6 +112,7 @@ struct writequeue_entry {
 	int len;
 	int end;
 	int users;
+	bool dirty;
 	struct connection *con;
 	struct list_head msgs;
 	struct kref ref;
@@ -671,6 +672,36 @@ static void make_sockaddr(struct sockaddr_storage *saddr, uint16_t port,
 	memset((char *)saddr + *addr_len, 0, sizeof(struct sockaddr_storage) - *addr_len);
 }
 
+static void dlm_page_release(struct kref *kref)
+{
+	struct writequeue_entry *e = container_of(kref, struct writequeue_entry,
+						  ref);
+
+	__free_page(e->page);
+	kfree(e);
+}
+
+static void dlm_msg_release(struct kref *kref)
+{
+	struct dlm_msg *msg = container_of(kref, struct dlm_msg, ref);
+
+	kref_put(&msg->entry->ref, dlm_page_release);
+	kfree(msg);
+}
+
+static void free_entry(struct writequeue_entry *e)
+{
+	struct dlm_msg *msg, *tmp;
+
+	list_for_each_entry_safe(msg, tmp, &e->msgs, list) {
+		list_del(&msg->list);
+		kref_put(&msg->ref, dlm_msg_release);
+	}
+
+	list_del(&e->list);
+	kref_put(&e->ref, dlm_page_release);
+}
+
 static void dlm_close_sock(struct socket **sock)
 {
 	if (*sock) {
@@ -685,6 +716,7 @@ static void close_connection(struct connection *con, bool and_other,
 			     bool tx, bool rx)
 {
 	bool closing = test_and_set_bit(CF_CLOSING, &con->flags);
+	struct writequeue_entry *e;
 
 	if (tx && !closing && cancel_work_sync(&con->swork)) {
 		log_print("canceled swork for node %d", con->nodeid);
@@ -703,6 +735,26 @@ static void close_connection(struct connection *con, bool and_other,
 		close_connection(con->othercon, false, true, true);
 	}
 
+	/* if we send a writequeue entry only a half way, we drop the
+	 * whole entry because reconnection and that we not start of the
+	 * middle of a msg which will confuse the other end.
+	 *
+	 * we can always drop messages because retransmits, but what we
+	 * cannot allow is to transmit half messages which may be processed
+	 * at the other side.
+	 *
+	 * our policy is to start on a clean state when disconnects, we don't
+	 * know what's send/received on transport layer in this case.
+	 */
+	spin_lock_bh(&con->writequeue_lock);
+	if (!list_empty(&con->writequeue)) {
+		e = list_first_entry(&con->writequeue, struct writequeue_entry,
+				     list);
+		if (e->dirty)
+			free_entry(e);
+	}
+	spin_unlock_bh(&con->writequeue_lock);
+
 	con->rx_leftover = 0;
 	con->retries = 0;
 	clear_bit(CF_CONNECTED, &con->flags);
@@ -980,36 +1032,6 @@ static int accept_from_sock(struct listen_connection *con)
 	return result;
 }
 
-static void dlm_page_release(struct kref *kref)
-{
-	struct writequeue_entry *e = container_of(kref, struct writequeue_entry,
-						  ref);
-
-	__free_page(e->page);
-	kfree(e);
-}
-
-static void dlm_msg_release(struct kref *kref)
-{
-	struct dlm_msg *msg = container_of(kref, struct dlm_msg, ref);
-
-	kref_put(&msg->entry->ref, dlm_page_release);
-	kfree(msg);
-}
-
-static void free_entry(struct writequeue_entry *e)
-{
-	struct dlm_msg *msg, *tmp;
-
-	list_for_each_entry_safe(msg, tmp, &e->msgs, list) {
-		list_del(&msg->list);
-		kref_put(&msg->ref, dlm_msg_release);
-	}
-
-	list_del(&e->list);
-	kref_put(&e->ref, dlm_page_release);
-}
-
 /*
  * writequeue_entry_complete - try to delete and free write queue entry
  * @e: write queue entry to try to delete
@@ -1021,6 +1043,8 @@ static void writequeue_entry_complete(struct writequeue_entry *e, int completed)
 {
 	e->offset += completed;
 	e->len -= completed;
+	/* signal that page was half way transmitted */
+	e->dirty = true;
 
 	if (e->len == 0 && e->users == 0)
 		free_entry(e);
-- 
2.26.3

[Cluster-devel] [PATCHv4 dlm/next 4/8] fs: dlm: add functionality to re-transmit a message

[Alexander Ahring Oder Aring]

Hi,

On Fri, Apr 9, 2021 at 10:49 AM Alexander Aring <aahringo@redhat.com> wrote:
...
>  }
>
> +void dlm_lowcomms_resend_buffer(void *mh)
> +{
> +       struct dlm_msg *msg = mh;
> +       void *mh_new;
> +       char *ppc;
> +
> +       mh_new = dlm_lowcomms_new_buffer_con(msg->entry->con, msg->len,
> +                                            GFP_NOFS, &ppc, NULL, NULL);

This should be GFP_ATOMIC, because we later call this function while
rcu lock is held.

I am worried about the con lookup here, but I think it's fine so far
we call this function inside some workqueue which should synchronize
the free of "con" here. However I will fix that srcu usage later to be
more correct...

- Alex

[Cluster-devel] [PATCHv4 dlm/next 4/8] fs: dlm: add functionality to re-transmit a message

[Alexander Ahring Oder Aring]

Hi,

On Fri, Apr 9, 2021 at 10:49 AM Alexander Aring <aahringo@redhat.com> wrote:
>
> This patch introduces a retransmit functionality for a lowcomms message
> handle. It's just allocates a new buffer and transmit it again, no
> special handling about prioritize it because keeping bytestream in order.
>
> To avoid another connection look some refactor was done to make a new
> buffer allocation with a preexisting connection pointer.
>
> Signed-off-by: Alexander Aring <aahringo@redhat.com>
> ---
>  fs/dlm/lowcomms.c | 55 ++++++++++++++++++++++++++++++++++-------------
>  fs/dlm/lowcomms.h |  1 +
>  2 files changed, 41 insertions(+), 15 deletions(-)
>
> diff --git a/fs/dlm/lowcomms.c b/fs/dlm/lowcomms.c
> index ade1a5266e4a..fb04448c4e48 100644
> --- a/fs/dlm/lowcomms.c
> +++ b/fs/dlm/lowcomms.c
> @@ -1447,25 +1447,14 @@ static struct writequeue_entry *new_wq_entry(struct connection *con, int len,
>         return e;
>  };
>
> -void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
> -                             void (*cb)(void *buf, void *priv), void *priv)
> +static void *dlm_lowcomms_new_buffer_con(struct connection *con, int len,
> +                                        gfp_t allocation, char **ppc,
> +                                        void (*cb)(void *buf, void *priv),
> +                                        void *priv)
>  {
>         struct writequeue_entry *e;
> -       struct connection *con;
>         struct dlm_msg *msg;
>
> -       if (len > DEFAULT_BUFFER_SIZE ||
> -           len < sizeof(struct dlm_header)) {
> -               BUILD_BUG_ON(PAGE_SIZE < DEFAULT_BUFFER_SIZE);
> -               log_print("failed to allocate a buffer of size %d", len);
> -               WARN_ON(1);
> -               return NULL;
> -       }
> -
> -       con = nodeid2con(nodeid, allocation);
> -       if (!con)
> -               return NULL;
> -
>         msg = kzalloc(sizeof(*msg), allocation);
>         if (!msg)
>                 return NULL;
> @@ -1485,6 +1474,26 @@ void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
>         return msg;
>  }
>
> +void *dlm_lowcomms_new_buffer(int nodeid, int len, gfp_t allocation, char **ppc,
> +                             void (*cb)(void *buf, void *priv), void *priv)
> +{
> +       struct connection *con;
> +
> +       if (len > DEFAULT_BUFFER_SIZE ||
> +           len < sizeof(struct dlm_header)) {
> +               BUILD_BUG_ON(PAGE_SIZE < DEFAULT_BUFFER_SIZE);
> +               log_print("failed to allocate a buffer of size %d", len);
> +               WARN_ON(1);
> +               return NULL;
> +       }
> +
> +       con = nodeid2con(nodeid, allocation);
> +       if (!con)
> +               return NULL;
> +
> +       return dlm_lowcomms_new_buffer_con(con, len, GFP_ATOMIC, ppc, cb, priv);
> +}
> +
>  void dlm_lowcomms_commit_buffer(void *mh)
>  {
>         struct dlm_msg *msg = mh;
> @@ -1525,6 +1534,22 @@ void dlm_lowcomms_get_buffer(void *mh)
>         kref_get(&msg->ref);
>  }
>
> +void dlm_lowcomms_resend_buffer(void *mh)
> +{
> +       struct dlm_msg *msg = mh;
> +       void *mh_new;
> +       char *ppc;
> +
> +       mh_new = dlm_lowcomms_new_buffer_con(msg->entry->con, msg->len,
> +                                            GFP_NOFS, &ppc, NULL, NULL);
> +       if (!mh_new)
> +               return;
> +
> +       memcpy(ppc, msg->ppc, msg->len);
> +       dlm_lowcomms_commit_buffer(mh_new);
> +       dlm_lowcomms_put_buffer(mh_new);
> +}

I will change this functionality so that it checks if the actual
message is already in "retransmit" state. If yes it will not queue a
message again which sits somewhere inside the "dlm lowcomms send
queue", the retransmit state will be dropped when the dlm message was
sent out. It might be triggered again for retransmission then.

I had some terrible experience with a lot of reconnects due tcpkill
test and the "lowcomms queue send" was getting bigger and bigger with
a bunch of retransmits and the other peer was doing "a lot" of work to
filter these messages out because already received.

- Alex