[pve-devel] [PATCH] cpg: fix issue when nodes with low id hang
Thomas Lamprecht
t.lamprecht at proxmox.com
Wed Apr 25 09:16:51 CEST 2018
The cpg config change callback where not made correctly for nodes
with a low member id (lowest id == master node) after corosync on
said node hung, due to IO, artifical suspend or other scheduling
related hangs.
This is releated to an heuristic for choosing/syncing the CPG member
downlist added in 6bbbfcb6b4af72cf35ab9fdb4412fa6c6bdacc12 (corosync
repository).
See whole issue thread:
https://lists.clusterlabs.org/pipermail/users/2018-March/014594.html
Upstream pull-request:
https://github.com/corosync/corosync/pull/347
Signed-off-by: Thomas Lamprecht <t.lamprecht at proxmox.com>
---
This is not yet applied upstream, but will be very likely the way to
go.
I propose to apply this and put it on pvetest, this should expand the
testsurface it gets and we can help upstream better to either report a
problem or 'bless' this solution.
...orm-clients-about-left-nodes-during-pause.patch | 322 +++++++++++++++++++++
patches/series | 1 +
2 files changed, 323 insertions(+)
create mode 100644 patches/0010-cpg-Inform-clients-about-left-nodes-during-pause.patch
diff --git a/patches/0010-cpg-Inform-clients-about-left-nodes-during-pause.patch b/patches/0010-cpg-Inform-clients-about-left-nodes-during-pause.patch
new file mode 100644
index 0000000..2e745c7
--- /dev/null
+++ b/patches/0010-cpg-Inform-clients-about-left-nodes-during-pause.patch
@@ -0,0 +1,322 @@
+From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
+From: Jan Friesse <jfriesse at redhat.com>
+Date: Tue, 24 Apr 2018 17:44:48 +0200
+Subject: [PATCH] cpg: Inform clients about left nodes during pause
+
+Patch tries to fix incorrect behaviour during following test-case:
+- 3 nodes
+- Node 1 is paused
+- Node 2 and 3 detects node 1 as failed and informs CPG clients
+- Node 1 is unpaused
+- Node 1 clients are informed about new membership, but not about Node 1
+ being paused, so from Node 1 point-of-view, Node 2 and 3 failure
+
+Solution is to:
+- Remove downlist master choose and always choose local node downlist.
+ For Node 1 in example above, downlist contains Node 2 and 3.
+- Keep code which informs clients about left nodes
+- Use joinlist as a authoritative source of nodes/clients which exists
+ in membership
+
+This patch doesn't break backwards compatibility.
+
+I've walked thru all the patches which changed behavior of cpg to ensure
+patch does not break CPG behavior. Most important were:
+- 058f50314cd20abe67f5e8fb3c029a63b0e10cdc - Base. Code was significantly
+ changed to handle double free by split group_info into two structures
+ cpg_pd (local node clients) and process_info (all clients). Joinlist
+ was
+- 97c28ea756cdf59316b2f609103122cc678329bd - This patch removed
+ confchg_fn and made CPG sync correct
+- feff0e8542463773207a3b2c1f6004afba1f58d5 - I've tested described
+ behavior without any issues
+- 6bbbfcb6b4af72cf35ab9fdb4412fa6c6bdacc12 - Added idea of using
+ heuristics to choose same downlist on all nodes. Sadly this idea
+ was beginning of the problems described in
+ 040fda8872a4a20340d73fa1c240b86afb2489f8,
+ ac1d79ea7c14997353427e962865781d0836d9fa,
+ 559d4083ed8355fe83f275e53b9c8f52a91694b2,
+ 02c5dffa5bb8579c223006fa1587de9ba7409a3d,
+ 64d0e5ace025cc929e42896c5d6beb3ef75b8244 and
+ b55f32fe2e1538db33a1ec584b67744c724328c6
+- 02c5dffa5bb8579c223006fa1587de9ba7409a3d - Made joinlist as
+ authoritative source of nodes/clients but left downlist_master_choose
+ as a source of information about left nodes
+
+Long story made short. This patch basically reverts
+idea of using heuristics to choose same downlist on all nodes.
+
+Signed-off-by: Jan Friesse <jfriesse at redhat.com>
+Signed-off-by: Thomas Lamprecht <t.lamprecht at proxmox.com>
+---
+ exec/cpg.c | 164 +++++--------------------------------------------------------
+ 1 file changed, 11 insertions(+), 153 deletions(-)
+
+diff --git a/exec/cpg.c b/exec/cpg.c
+index 78ac1e9e..b851cba3 100644
+--- a/exec/cpg.c
++++ b/exec/cpg.c
+@@ -139,13 +139,6 @@ enum cpg_sync_state {
+ CPGSYNC_JOINLIST
+ };
+
+-enum cpg_downlist_state_e {
+- CPG_DOWNLIST_NONE,
+- CPG_DOWNLIST_WAITING_FOR_MESSAGES,
+- CPG_DOWNLIST_APPLYING,
+-};
+-static enum cpg_downlist_state_e downlist_state;
+-static struct list_head downlist_messages_head;
+ static struct list_head joinlist_messages_head;
+
+ struct cpg_pd {
+@@ -295,9 +288,7 @@ static int cpg_exec_send_downlist(void);
+
+ static int cpg_exec_send_joinlist(void);
+
+-static void downlist_messages_delete (void);
+-
+-static void downlist_master_choose_and_send (void);
++static void downlist_inform_clients (void);
+
+ static void joinlist_inform_clients (void);
+
+@@ -499,14 +490,6 @@ struct req_exec_cpg_downlist {
+ mar_uint32_t nodeids[PROCESSOR_COUNT_MAX] __attribute__((aligned(8)));
+ };
+
+-struct downlist_msg {
+- mar_uint32_t sender_nodeid;
+- mar_uint32_t old_members __attribute__((aligned(8)));
+- mar_uint32_t left_nodes __attribute__((aligned(8)));
+- mar_uint32_t nodeids[PROCESSOR_COUNT_MAX] __attribute__((aligned(8)));
+- struct list_head list;
+-};
+-
+ struct joinlist_msg {
+ mar_uint32_t sender_nodeid;
+ uint32_t pid;
+@@ -566,8 +549,6 @@ static void cpg_sync_init (
+ last_sync_ring_id.nodeid = ring_id->rep.nodeid;
+ last_sync_ring_id.seq = ring_id->seq;
+
+- downlist_state = CPG_DOWNLIST_WAITING_FOR_MESSAGES;
+-
+ entries = 0;
+ /*
+ * Determine list of nodeids for downlist message
+@@ -611,14 +592,10 @@ static void cpg_sync_activate (void)
+ my_member_list_entries * sizeof (unsigned int));
+ my_old_member_list_entries = my_member_list_entries;
+
+- if (downlist_state == CPG_DOWNLIST_WAITING_FOR_MESSAGES) {
+- downlist_master_choose_and_send ();
+- }
++ downlist_inform_clients ();
+
+ joinlist_inform_clients ();
+
+- downlist_messages_delete ();
+- downlist_state = CPG_DOWNLIST_NONE;
+ joinlist_messages_delete ();
+
+ notify_lib_totem_membership (NULL, my_member_list_entries, my_member_list);
+@@ -626,8 +603,7 @@ static void cpg_sync_activate (void)
+
+ static void cpg_sync_abort (void)
+ {
+- downlist_state = CPG_DOWNLIST_NONE;
+- downlist_messages_delete ();
++
+ joinlist_messages_delete ();
+ }
+
+@@ -800,76 +776,17 @@ static int notify_lib_joinlist(
+ return CS_OK;
+ }
+
+-static void downlist_log(const char *msg, struct downlist_msg* dl)
++static void downlist_log(const char *msg, struct req_exec_cpg_downlist *dl)
+ {
+ log_printf (LOG_DEBUG,
+- "%s: sender %s; members(old:%d left:%d)",
++ "%s: members(old:%d left:%d)",
+ msg,
+- api->totem_ifaces_print(dl->sender_nodeid),
+ dl->old_members,
+ dl->left_nodes);
+ }
+
+-static struct downlist_msg* downlist_master_choose (void)
++static void downlist_inform_clients (void)
+ {
+- struct downlist_msg *cmp;
+- struct downlist_msg *best = NULL;
+- struct list_head *iter;
+- uint32_t cmp_members;
+- uint32_t best_members;
+- uint32_t i;
+- int ignore_msg;
+-
+- for (iter = downlist_messages_head.next;
+- iter != &downlist_messages_head;
+- iter = iter->next) {
+-
+- cmp = list_entry(iter, struct downlist_msg, list);
+- downlist_log("comparing", cmp);
+-
+- ignore_msg = 0;
+- for (i = 0; i < cmp->left_nodes; i++) {
+- if (cmp->nodeids[i] == api->totem_nodeid_get()) {
+- log_printf (LOG_DEBUG, "Ignoring this entry because I'm in the left list\n");
+-
+- ignore_msg = 1;
+- break;
+- }
+- }
+-
+- if (ignore_msg) {
+- continue ;
+- }
+-
+- if (best == NULL) {
+- best = cmp;
+- continue;
+- }
+-
+- best_members = best->old_members - best->left_nodes;
+- cmp_members = cmp->old_members - cmp->left_nodes;
+-
+- if (cmp_members > best_members) {
+- best = cmp;
+- } else if (cmp_members == best_members) {
+- if (cmp->old_members > best->old_members) {
+- best = cmp;
+- } else if (cmp->old_members == best->old_members) {
+- if (cmp->sender_nodeid < best->sender_nodeid) {
+- best = cmp;
+- }
+- }
+- }
+- }
+-
+- assert (best != NULL);
+-
+- return best;
+-}
+-
+-static void downlist_master_choose_and_send (void)
+-{
+- struct downlist_msg *stored_msg;
+ struct list_head *iter;
+ struct process_info *left_pi;
+ qb_map_t *group_map;
+@@ -884,14 +801,7 @@ static void downlist_master_choose_and_send (void)
+ qb_map_iter_t *miter;
+ int i, size;
+
+- downlist_state = CPG_DOWNLIST_APPLYING;
+-
+- stored_msg = downlist_master_choose ();
+- if (!stored_msg) {
+- log_printf (LOGSYS_LEVEL_DEBUG, "NO chosen downlist");
+- return;
+- }
+- downlist_log("chosen downlist", stored_msg);
++ downlist_log("my downlist", &g_req_exec_cpg_downlist);
+
+ group_map = qb_skiplist_create();
+
+@@ -905,9 +815,9 @@ static void downlist_master_choose_and_send (void)
+ iter = iter->next;
+
+ left_pi = NULL;
+- for (i = 0; i < stored_msg->left_nodes; i++) {
++ for (i = 0; i < g_req_exec_cpg_downlist.left_nodes; i++) {
+
+- if (pi->nodeid == stored_msg->nodeids[i]) {
++ if (pi->nodeid == g_req_exec_cpg_downlist.nodeids[i]) {
+ left_pi = pi;
+ break;
+ }
+@@ -1039,23 +949,6 @@ static void joinlist_inform_clients (void)
+ joinlist_remove_zombie_pi_entries ();
+ }
+
+-static void downlist_messages_delete (void)
+-{
+- struct downlist_msg *stored_msg;
+- struct list_head *iter, *iter_next;
+-
+- for (iter = downlist_messages_head.next;
+- iter != &downlist_messages_head;
+- iter = iter_next) {
+-
+- iter_next = iter->next;
+-
+- stored_msg = list_entry(iter, struct downlist_msg, list);
+- list_del (&stored_msg->list);
+- free (stored_msg);
+- }
+-}
+-
+ static void joinlist_messages_delete (void)
+ {
+ struct joinlist_msg *stored_msg;
+@@ -1076,7 +969,6 @@ static void joinlist_messages_delete (void)
+
+ static char *cpg_exec_init_fn (struct corosync_api_v1 *corosync_api)
+ {
+- list_init (&downlist_messages_head);
+ list_init (&joinlist_messages_head);
+ api = corosync_api;
+ return (NULL);
+@@ -1338,43 +1230,9 @@ static void message_handler_req_exec_cpg_downlist(
+ unsigned int nodeid)
+ {
+ const struct req_exec_cpg_downlist *req_exec_cpg_downlist = message;
+- int i;
+- struct list_head *iter;
+- struct downlist_msg *stored_msg;
+- int found;
+
+- if (downlist_state != CPG_DOWNLIST_WAITING_FOR_MESSAGES) {
+- log_printf (LOGSYS_LEVEL_WARNING, "downlist left_list: %d received in state %d",
+- req_exec_cpg_downlist->left_nodes, downlist_state);
+- return;
+- }
+-
+- stored_msg = malloc (sizeof (struct downlist_msg));
+- stored_msg->sender_nodeid = nodeid;
+- stored_msg->old_members = req_exec_cpg_downlist->old_members;
+- stored_msg->left_nodes = req_exec_cpg_downlist->left_nodes;
+- memcpy (stored_msg->nodeids, req_exec_cpg_downlist->nodeids,
+- req_exec_cpg_downlist->left_nodes * sizeof (mar_uint32_t));
+- list_init (&stored_msg->list);
+- list_add (&stored_msg->list, &downlist_messages_head);
+-
+- for (i = 0; i < my_member_list_entries; i++) {
+- found = 0;
+- for (iter = downlist_messages_head.next;
+- iter != &downlist_messages_head;
+- iter = iter->next) {
+-
+- stored_msg = list_entry(iter, struct downlist_msg, list);
+- if (my_member_list[i] == stored_msg->sender_nodeid) {
+- found = 1;
+- }
+- }
+- if (!found) {
+- return;
+- }
+- }
+-
+- downlist_master_choose_and_send ();
++ log_printf (LOGSYS_LEVEL_WARNING, "downlist left_list: %d received",
++ req_exec_cpg_downlist->left_nodes);
+ }
+
+
+--
+2.14.2
+
diff --git a/patches/series b/patches/series
index 501b4f4..7931482 100644
--- a/patches/series
+++ b/patches/series
@@ -7,3 +7,4 @@
0007-only-start-corosync.service-if-conf-exists.patch
0008-remove-unecessary-and-problematic-corosync-qdevice.i.patch
0009-totemcrypto-Check-length-of-the-packet.patch
+0010-cpg-Inform-clients-about-left-nodes-during-pause.patch
--
2.14.2
More information about the pve-devel
mailing list