Thanks for Steven.
I will update this patch reflecting our discussion after I completed testing.
On Fri, Oct 28, 2011 at 1:57 PM, Steven Dake <sdake@xxxxxxxxxx> wrote:
> On 10/27/2011 07:51 PM, Yunkai Zhang wrote:
>> On Fri, Oct 28, 2011 at 9:23 AM, Angus Salkeld <asalkeld@xxxxxxxxxx> wrote:
>>> Thanks for the patch! Some comments inline.
>>>
>>> -Angus
>>>
>>> On Fri, Oct 28, 2011 at 12:22:13AM +0800, Yunkai Zhang wrote:
>>>> We found that sheepdog will receive more than one confchg msg when
>>>> network partition occur. For example, suppose the cluster has 4
>>>> nodes: N1, N2, N3, N4, and they form a single-ring initially. After a
>>>> while, network partition occur, the single-ring divide into two
>>>> sub-ring: ring(N1, N2, N3) and ring(N4). The sheepdog in the ring(N4)
>>>> will receive the following confchg messages in turn:
>>>> Memb: N2,N3,N4 Left:N1 Joined:null
>>>> memb: N3,N4 Left:N2 Joined:null
>>>> memb: N4 Left:N3 Joined:null
>>>>
>>>> This patch will fixed this bug, and the client will only receive one
>>>> confchg event in this case:
>>>> memb: N4 Left:N2,N3,N4 Joined:null
>>>>
>>>> Signed-off-by: Yunkai Zhang <qiushu.zyk@xxxxxxxxxx>
>>>> ---
>>>> configure.ac | 2 +-
>>>> include/corosync/cpg.h | 1 +
>>>> services/cpg.c | 97 +++++++++++++++++++++++++++++++++++++++++-------
>>>> 3 files changed, 85 insertions(+), 15 deletions(-)
>>>>
>>>> diff --git a/configure.ac b/configure.ac
>>>> index 563f799..aaed478 100644
>>>> --- a/configure.ac
>>>> +++ b/configure.ac
>>>> @@ -88,7 +88,7 @@ AC_HEADER_SYS_WAIT
>>>> AC_CHECK_HEADERS([arpa/inet.h fcntl.h limits.h netdb.h netinet/in.h stdint.h \
>>>> stdlib.h string.h sys/ioctl.h sys/param.h sys/socket.h \
>>>> sys/time.h syslog.h unistd.h sys/types.h getopt.h malloc.h \
>>>> - sys/sockio.h utmpx.h])
>>>> + sys/sockio.h utmpx.h search.h])
>>>
>>> Why this? it's not included anywhere in corosync.
>>>
>>
>> Sorry, I forgot to clear it.
>>
>> I wrote the first version of this patch base on v1.4.1 which not
>> provide libqb and I used hash table api comes from search.h instead of
>> qbmap.h.
>>
>>>>
>>>> # Checks for typedefs, structures, and compiler characteristics.
>>>> AC_C_CONST
>>>> diff --git a/include/corosync/cpg.h b/include/corosync/cpg.h
>>>> index f7956fa..9861c5a 100644
>>>> --- a/include/corosync/cpg.h
>>>> +++ b/include/corosync/cpg.h
>>>> @@ -95,6 +95,7 @@ struct cpg_name {
>>>> };
>>>>
>>>> #define CPG_MEMBERS_MAX 128
>>>> +#define CPG_GROUPS_MAX 64
>>>
>
> Go with 8192 groups - we have some use cases where many groups are
> needed - memory consumption shouldn't be much of an issue.
>
>>> I think this is ok, though not 100% sure.
>>>
>>>>
>>>> struct cpg_iteration_description_t {
>>>> struct cpg_name group;
>>>> diff --git a/services/cpg.c b/services/cpg.c
>>>> index caec097..dc72c84 100644
>>>> --- a/services/cpg.c
>>>> +++ b/services/cpg.c
>>>> @@ -55,6 +55,7 @@
>>>> #include <netinet/in.h>
>>>> #include <arpa/inet.h>
>>>> #include <sys/mman.h>
>>>> +#include <qb/qbmap.h>
>>>>
>>>> #include <corosync/corotypes.h>
>>>> #include <qb/qbipc_common.h>
>>>> @@ -705,7 +706,6 @@ static int notify_lib_joinlist(
>>>> for (iter = cpg_pd_list_head.next; iter != &cpg_pd_list_head; iter = iter->next) {
>>>> struct cpg_pd *cpd = list_entry (iter, struct cpg_pd, list);
>>>> if (mar_name_compare (&cpd->group_name, group_name) == 0) {
>>>> - assert (left_list_entries <= 1);
>>>> assert (joined_list_entries <= 1);
>>>> if (joined_list_entries) {
>>>> if (joined_list[0].pid == cpd->pid &&
>>>> @@ -798,8 +798,18 @@ static void downlist_master_choose_and_send (void)
>>>> {
>>>> struct downlist_msg *stored_msg;
>>>> struct list_head *iter;
>>>> - mar_cpg_address_t left_list;
>>>> - int i;
>>>> + struct process_info *left_pi;
>>>> + struct qb_map_t *group_htab;
>>>> + struct cpg_name cpg_group;
>>>> + mar_cpg_name_t group;
>>>> + struct confchg_data{
>>>> + struct cpg_name cpg_group;
>>>> + mar_cpg_address_t left_list[CPG_MEMBERS_MAX];
>>>> + int left_list_entries;
>>>> + struct list_head list;
>>>> + } *pcd;
>>>> + DECLARE_LIST_INIT(confchg_data_list_head);
>>>> + int i, size;
>>>
>>> Why the list? You could use qb_skiplist_create() and use the
>>> iterator.
>>>
>>> miter = qb_map_iter_create(group_map);
>>> while ((p = qb_map_iter_next(miter, (void**)&pcd))) {
>>> ...
>>> }
>>>
>>> Though using libqb here is going to make it impossible to cherry-pick
>>> the patch into the flatiron branch (Steven/Honzaf would this be needed
>>> there?).
>>>
>>
>
> agree use skiplist
>
>> I worry about the same thing for my project using v1.4.1 now.
>>
>> Does Steven/Honza give some advice?
>>
>
> For an invasive patch like this, what would be ideal is a master version
> and a separate patch for flatiron. Typically the branch maintainer
> (Honza) would cherry-pick simple bug fixes, but this is major rework and
> warrants a separate patch. It must go into master prior to flatiron to
> avoid behavioural regressions when moving forward in versions.
>
>>>>
>>>> downlist_state = CPG_DOWNLIST_APPLYING;
>>>>
>>>> @@ -810,27 +820,86 @@ static void downlist_master_choose_and_send (void)
>>>> }
>>>> downlist_log("chosen downlist", stored_msg);
>>>>
>>>> - /* send events */
>>>> + group_htab = qb_hashtable_create(CPG_GROUPS_MAX);
>>>> +
>>>> + /*
>>>> + * only the cpg groups included in left nodes should receive
>>>> + * confchg event, so we will collect these cpg groups and
>>>> + * relative left_lists here
>>>> + */
>>>> for (iter = process_info_list_head.next; iter != &process_info_list_head; ) {
>>>> struct process_info *pi = list_entry(iter, struct process_info, list);
>>>> iter = iter->next;
>>>>
>>>> + left_pi = NULL;
>>>> for (i = 0; i < stored_msg->left_nodes; i++) {
>>>> +
>>>> if (pi->nodeid == stored_msg->nodeids[i]) {
>>>> - left_list.nodeid = pi->nodeid;
>>>> - left_list.pid = pi->pid;
>>>> - left_list.reason = CONFCHG_CPG_REASON_NODEDOWN;
>>>> -
>>>> - notify_lib_joinlist(&pi->group, NULL,
>>>> - 0, NULL,
>>>> - 1, &left_list,
>>>> - MESSAGE_RES_CPG_CONFCHG_CALLBACK);
>>>> - list_del (&pi->list);
>>>> - free (pi);
>>>> + left_pi = pi;
>>>> break;
>>>> }
>>>> }
>>>> +
>>>> + if (left_pi) {
>>>> + marshall_from_mar_cpg_name_t(&cpg_group, &left_pi->group);
>>>> + log_printf (LOG_DEBUG, "cpg group name:%s", cpg_group.value);
>>>> + cpg_group.value[cpg_group.length] = 0;
>>>> +
>>>> + pcd = (struct confchg_data *)qb_map_get(group_htab,
>>>> + cpg_group.value);
>>>> + if (pcd != NULL) {
>>>> + size = pcd->left_list_entries;
>>>> + pcd->left_list[size].nodeid = left_pi->nodeid;
>>>> + pcd->left_list[size].pid = left_pi->pid;
>>>> + pcd->left_list[size].reason = CONFCHG_CPG_REASON_NODEDOWN;
>>>> + pcd->left_list_entries++;
>>>> + }else {
>>>> + pcd = (struct confchg_data *)malloc(sizeof(struct confchg_data));
>>>> + memset(pcd, 0, sizeof(struct confchg_data));
>>>> +
>>>> + memcpy(&pcd->cpg_group, &cpg_group, sizeof(struct cpg_name));
>>>> + pcd->left_list[0].nodeid = left_pi->nodeid;
>>>> + pcd->left_list[0].pid = left_pi->pid;
>>>> + pcd->left_list[0].reason = CONFCHG_CPG_REASON_NODEDOWN;
>>>> + pcd->left_list_entries = 1;
>>>> +
>>>> + qb_map_put(group_htab, pcd->cpg_group.value, pcd);
>>>> +
>>>> + list_init(&pcd->list);
>>>> + list_add(&pcd->list, &confchg_data_list_head);
>>>> + }
>>>
>>> This could be a bit simplified:
>>>
>>> pcd = (struct confchg_data *)qb_map_get(group_map,
>>> cpg_group.value);
>>> if (pcd == NULL) {
>>> pcd = (struct confchg_data *)calloc(1, sizeof(struct confchg_data));
>>> memcpy(&pcd->cpg_group, &cpg_group, sizeof(struct cpg_name));
>>> qb_map_put(group_map, pcd->cpg_group.value, pcd);
>>> }
>>> size = pcd->left_list_entries;
>>> pcd->left_list[size].nodeid = left_pi->nodeid;
>>> pcd->left_list[size].pid = left_pi->pid;
>>> pcd->left_list[size].reason = CONFCHG_CPG_REASON_NODEDOWN;
>>> pcd->left_list_entries++;
>>> list_del (&left_pi->list);
>>> free (left_pi);
>>
>> Thanks, It looks cool!
>>
>>>
>>>> +
>>>> + list_del (&left_pi->list);
>>>> + free (left_pi);
>>>> + }
>>>> + }
>>>> +
>>>> + /* send only one confchg event per cpg group */
>>>> + for (iter = confchg_data_list_head.next; iter != &confchg_data_list_head; ) {
>>>> + pcd = list_entry(iter, struct confchg_data, list);
>>>> + iter = iter->next;
>>>> +
>>>> + marshall_to_mar_cpg_name_t(&group, &pcd->cpg_group);
>>>> +
>>>> + log_printf (LOG_DEBUG, "left_list_entries:%d", pcd->left_list_entries);
>>>> + for (i=0; i<pcd->left_list_entries; i++) {
>>>> + log_printf (LOG_DEBUG, "left_list[%d] group:%d, ip:%s, pid:%d",
>>>> + i, pcd->cpg_group.value,
>>>> + (char*)api->totem_ifaces_print(pcd->left_list[i].nodeid),
>>>> + pcd->left_list[i].pid);
>>>> + }
>>>> +
>>>> + /* send confchg event */
>>>> + notify_lib_joinlist(&group, NULL,
>>>> + 0, NULL,
>>>> + pcd->left_list_entries,
>>>> + pcd->left_list,
>>>> + MESSAGE_RES_CPG_CONFCHG_CALLBACK);
>>>> +
>>>> + free(pcd);
>>>> }
>>>> +
>>>> + qb_map_destroy(group_htab);
>>>> }
>>>>
>>>> static void downlist_messages_delete (void)
>>>> --
>>>> 1.7.6.4
>>> _______________________________________________
>>> discuss mailing list
>>> discuss@xxxxxxxxxxxx
>>> http://lists.corosync.org/mailman/listinfo/discuss
>>>
>>
>>
>>
>
> _______________________________________________
> discuss mailing list
> discuss@xxxxxxxxxxxx
> http://lists.corosync.org/mailman/listinfo/discuss
>
--
Yunkai Zhang
Work at Taobao
_______________________________________________
discuss mailing list
discuss@xxxxxxxxxxxx
http://lists.corosync.org/mailman/listinfo/discuss
