Branch b1_4_mountconf

b=4482
Problem with the mds notify method of online ost addition - lov_notify
can't notify the mds until the mds is set up, due to a check of obd_set_up
in obd_notify.  But obd_set_up isn't set until
mds_postrecov/mds_postsetup/mds_setup returns to class_setup.

(This can be seen by adding ost's before adding the mdt - the other
order doesn't have the problem because there are no osc's in the lov
during mds_postrecov at that point, so no notifies.)

So I hacked in an additional flag check to obd_notify that happens only to
be set for the mds at this point in it's life: obd_async_recov.  Don't hurt
me.

diff --git a/lustre/include/linux/lustre_disk.h b/lustre/include/linux/lustre_disk.h
index d751c6b..3dec274 100644 (file)
--- a/lustre/include/linux/lustre_disk.h
+++ b/lustre/include/linux/lustre_disk.h
@@ -69,7 +69,7 @@ static inline char *mt_str(enum ldd_mount_type mt)
 }
 
 #ifndef MTI_NIDS_MAX  /* match lustre_idl.h */
-#define MTI_NIDS_MAX 10
+#define MTI_NIDS_MAX 64
 #endif
 
 #define LDD_INCOMPAT_SUPP 0
@@ -91,7 +91,7 @@ struct lustre_disk_data {
                                            svname */
         __u16      ldd_mgsnid_count;
         __u16      ldd_failnid_count;   /* server failover nid count */
-        lnet_nid_t ldd_mgsnid[MTI_NIDS_MAX];  /* mgmt nid list; lmd can 
+        lnet_nid_t ldd_mgsnid[MTI_NIDS_MAX];  /* mgs nid list; lmd can 
                                                  override */
         lnet_nid_t ldd_failnid[MTI_NIDS_MAX]; /* server failover nids */
         char       ldd_mount_opts[2048]; /* target fs mount opts */
@@ -119,7 +119,7 @@ static inline int sv_make_name(__u32 flags, __u16 index, char *fs, char *name)
                         (flags & LDD_F_SV_TYPE_MDT) ? "MDT" : "OST",  
                         index);
         } else if (flags & LDD_F_SV_TYPE_MGS) {
-                sprintf(name, "MGMT");
+                sprintf(name, "MGS");
         } else {
                 CERROR("unknown server type %#x\n", flags);
                 return 1;

diff --git a/lustre/include/linux/lustre_idl.h b/lustre/include/linux/lustre_idl.h
index cadfb03..71f19fe 100644 (file)
--- a/lustre/include/linux/lustre_idl.h
+++ b/lustre/include/linux/lustre_idl.h
@@ -990,7 +990,9 @@ typedef enum {
 
 #define MTI_NAME_MAXLEN 64
 #define MTI_UUID_MAXLEN MTI_NAME_MAXLEN + 5
-#define MTI_NIDS_MAX 10 /* match lustre_disk.h */
+/* each host can have multiple nids, and multiple failover hosts, and I don't
+   want to run out of room... */
+#define MTI_NIDS_MAX 64 /* match lustre_disk.h */
 
 struct mgs_target_info {
         char             mti_fsname[MTI_NAME_MAXLEN];

diff --git a/lustre/include/linux/obd_class.h b/lustre/include/linux/obd_class.h
index eed0753..cebe4e4 100644 (file)
--- a/lustre/include/linux/obd_class.h
+++ b/lustre/include/linux/obd_class.h
@@ -1069,9 +1069,14 @@ static inline int obd_notify(struct obd_device *obd,
                              void *data)
 {
         OBD_CHECK_DEV(obd);
-        if (!obd->obd_set_up) {
-                CERROR("obd %s not set up\n", obd->obd_name);
-                return -EINVAL;
+
+        /* the check for async_recov is a complete hack - I'm hereby
+           overloading the meaning to also mean "this was called from
+           mds_postsetup".  I know that my mds is able to handle notifies
+           by this point, and it needs to get them to execute mds_postrecov. */ 
+        if (!obd->obd_set_up && !obd->obd_async_recov) {
+                CERROR("obd %s not set up, notifying anyhow\n", obd->obd_name);
+                return -EAGAIN;
         }
 
         if (!OBP(obd, notify)) {

diff --git a/lustre/mds/handler.c b/lustre/mds/handler.c
index 91a5f88..09386fd 100644 (file)
--- a/lustre/mds/handler.c
+++ b/lustre/mds/handler.c
@@ -2115,10 +2115,11 @@ int mds_postrecov(struct obd_device *obd)
 
         /* FIXME Does target_finish_recovery really need this to block? */
         /* Notify the LOV, which will in turn call mds_notify for each tgt */
+        /* This means that we have to fool obd_notify to think we're obd_set_up
+           during mds_lov_connect. */
         obd_notify(obd->u.mds.mds_osc_obd, NULL, 
                    obd->obd_async_recov ? OBD_NOTIFY_SYNC_NONBLOCK :
                    OBD_NOTIFY_SYNC, NULL);
-        //mds_lov_start_synchronize(obd, NULL, NULL, obd->obd_async_recov);
 
         /* quota recovery */
         lquota_recovery(quota_interface, obd);

diff --git a/lustre/mds/mds_lov.c b/lustre/mds/mds_lov.c
index 4c0c4a0..7726861 100644 (file)
--- a/lustre/mds/mds_lov.c
+++ b/lustre/mds/mds_lov.c
@@ -368,7 +368,7 @@ int mds_lov_connect(struct obd_device *obd, char * lov_name)
          * set_nextid().  The class driver can help us here, because
          * it can use the obd_recovering flag to determine when the
          * the OBD is full available. */
-        if (!obd->obd_recovering)
+        if (!obd->obd_recovering) 
                 rc = mds_postrecov(obd);
         RETURN(rc);
 
@@ -749,6 +749,7 @@ int mds_notify(struct obd_device *obd, struct obd_device *watched,
         ENTRY;
 
         switch (ev) {
+        /* We only handle these: */
         case OBD_NOTIFY_ACTIVE:
         case OBD_NOTIFY_SYNC:
         case OBD_NOTIFY_SYNC_NONBLOCK:
@@ -759,7 +760,7 @@ int mds_notify(struct obd_device *obd, struct obd_device *watched,
 
         CDEBUG(D_WARNING, "notify %s ev=%d\n", watched->obd_name, ev);
 
-        if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) {
+        if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME) != 0) {
                 CERROR("unexpected notification of %s %s!\n",
                        watched->obd_type->typ_name, watched->obd_name);
                 RETURN(-EINVAL);

diff --git a/lustre/utils/mkfs_lustre.c b/lustre/utils/mkfs_lustre.c
index 1fff034..54db5fb 100644 (file)
--- a/lustre/utils/mkfs_lustre.c
+++ b/lustre/utils/mkfs_lustre.c
@@ -189,31 +189,6 @@ static void run_command_out()
         }
 }
 
-#if 0
-static int lnet_setup = 0;
-static int lnet_start()
-{
-        ptl_initialize(0, NULL);
-        if (access("/proc/sys/lnet", X_OK) != 0) {
-                fprintf(stderr, "%s: The LNET module must be loaded to "
-                        "determine local NIDs\n", progname);
-                return 1;
-        }
-        if (jt_ptl_get_nids(NULL) == -ENETDOWN) {
-                char *cmd[]={"network", "up"};
-                jt_ptl_network(2, cmd);
-                lnet_setup++;
-        }
-        return 0;
-}
-
-static void lnet_stop()
-{
-        char *cmd[]={"network", "down"};
-        if (--lnet_setup == 0)
-                jt_ptl_network(2, cmd);
-}
-#endif
 
 /*============ disk dev functions ===================*/
 
@@ -910,7 +885,7 @@ int parse_opts(int argc, char *const argv[], struct mkfs_opts *mop,
                         char *s1 = optarg, *s2;
                         if (IS_MGS(&mop->mo_ldd)) {
                                 badopt(long_opt[longidx].name, 
-                                       "non-MGMT MDT,OST");
+                                       "non-MGS MDT,OST");
                                 return 1;
                         }
                         while ((s2 = strsep(&s1, ","))) {
@@ -1081,46 +1056,6 @@ int main(int argc, char *const argv[])
                 mop.mo_ldd.ldd_flags |= LDD_F_SV_TYPE_MGS;
         }
 
-#if 0
-        if (IS_MGS(&mop.mo_ldd) && (mop.mo_ldd.ldd_mgsnid_count == 0)) {
-                int i;
-                __u64 *nids;
-                
-                vprint("No mgs nids specified, using all local nids\n");
-                ret = lnet_start();
-                if (ret)
-                        goto out;
-                i = jt_ptl_get_nids(&nids);
-                if (i < 0) {
-                        fprintf(stderr, "%s: Can't find local nids "
-                                "(is the lnet module loaded?)\n", progname);
-                } else {
-                        if (i > 0) {
-                                if (i > MTI_NIDS_MAX) 
-                                        i = MTI_NIDS_MAX;
-                                vprint("Adding %d local nids for MGS\n", i);
-                                memcpy(mop.mo_ldd.ldd_mgsnid, nids,
-                                       sizeof(mop.mo_ldd.ldd_mgsnid));
-                                free(nids);
-                        }
-                        mop.mo_ldd.ldd_mgsnid_count = i;
-                }
-        }
-
-        if (IS_MGS(&mop.mo_ldd) && mop.mo_ldd.ldd_failnid_count) {
-                /* Add failover nids to mgsnids if we start an MGS
-                   (MDT must have all possible MGS nids for failover.) */
-                int i = 0, j = mop.mo_ldd.ldd_mgsnid_count;
-                while (i < mop.mo_ldd.ldd_failnid_count) {
-                        if (j >= MTI_NIDS_MAX) 
-                                break;
-                        mop.mo_ldd.ldd_mgsnid[j++] =
-                                mop.mo_ldd.ldd_failnid[i++];
-                }
-                mop.mo_ldd.ldd_mgsnid_count = j;
-        }
-#endif
-        
         if (!IS_MGS(&mop.mo_ldd) && (mop.mo_ldd.ldd_mgsnid_count == 0)) {
                 fatal();
                 fprintf(stderr, "Must specify either --mgs or --mgsnid\n");