Whamcloud - gitweb
LU-4420 fld: retry fld rpc even for ESHUTDOWN
[fs/lustre-release.git] / lustre / fld / fld_request.c
index 3c2d48f..4b402cb 100644 (file)
@@ -45,7 +45,7 @@
 #ifdef __KERNEL__
 # include <libcfs/libcfs.h>
 # include <linux/module.h>
-# include <asm/div64.h>
+# include <linux/math64.h>
 #else /* __KERNEL__ */
 # include <liblustre.h>
 # include <libcfs/list.h>
 #include <lustre_mdc.h>
 #include "fld_internal.h"
 
-/* TODO: these 3 functions are copies of flow-control code from mdc_lib.c
- * It should be common thing. The same about mdc RPC lock */
-static int fld_req_avail(struct client_obd *cli, struct mdc_cache_waiter *mcw)
-{
-        int rc;
-        ENTRY;
-        client_obd_list_lock(&cli->cl_loi_list_lock);
-        rc = cfs_list_empty(&mcw->mcw_entry);
-        client_obd_list_unlock(&cli->cl_loi_list_lock);
-        RETURN(rc);
-};
-
-static void fld_enter_request(struct client_obd *cli)
-{
-       struct mdc_cache_waiter mcw;
-       struct l_wait_info lwi = { 0 };
-
-       client_obd_list_lock(&cli->cl_loi_list_lock);
-       if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
-               cfs_list_add_tail(&mcw.mcw_entry, &cli->cl_cache_waiters);
-               init_waitqueue_head(&mcw.mcw_waitq);
-               client_obd_list_unlock(&cli->cl_loi_list_lock);
-               l_wait_event(mcw.mcw_waitq, fld_req_avail(cli, &mcw), &lwi);
-       } else {
-               cli->cl_r_in_flight++;
-               client_obd_list_unlock(&cli->cl_loi_list_lock);
-       }
-}
-
-static void fld_exit_request(struct client_obd *cli)
-{
-       cfs_list_t *l, *tmp;
-       struct mdc_cache_waiter *mcw;
-
-       client_obd_list_lock(&cli->cl_loi_list_lock);
-       cli->cl_r_in_flight--;
-       cfs_list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
-
-               if (cli->cl_r_in_flight >= cli->cl_max_rpcs_in_flight) {
-                       /* No free request slots anymore */
-                       break;
-               }
-
-               mcw = cfs_list_entry(l, struct mdc_cache_waiter, mcw_entry);
-               cfs_list_del_init(&mcw->mcw_entry);
-               cli->cl_r_in_flight++;
-               wake_up(&mcw->mcw_waitq);
-       }
-       client_obd_list_unlock(&cli->cl_loi_list_lock);
-}
-
 static int fld_rrb_hash(struct lu_client_fld *fld,
                         seqno_t seq)
 {
@@ -135,16 +84,25 @@ fld_rrb_scan(struct lu_client_fld *fld, seqno_t seq)
        else
                hash = 0;
 
-        cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
+again:
+       list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
                 if (target->ft_idx == hash)
                         RETURN(target);
         }
 
+       if (hash != 0) {
+               /* It is possible the remote target(MDT) are not connected to
+                * with client yet, so we will refer this to MDT0, which should
+                * be connected during mount */
+               hash = 0;
+               goto again;
+       }
+
         CERROR("%s: Can't find target by hash %d (seq "LPX64"). "
                "Targets (%d):\n", fld->lcf_name, hash, seq,
                fld->lcf_count);
 
-        cfs_list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
+       list_for_each_entry(target, &fld->lcf_targets, ft_chain) {
                 const char *srv_name = target->ft_srv != NULL  ?
                         target->ft_srv->lsf_name : "<null>";
                 const char *exp_name = target->ft_exp != NULL ?
@@ -227,7 +185,7 @@ int fld_client_add_target(struct lu_client_fld *fld,
                 RETURN(-ENOMEM);
 
        spin_lock(&fld->lcf_lock);
-       cfs_list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
+       list_for_each_entry(tmp, &fld->lcf_targets, ft_chain) {
                if (tmp->ft_idx == tar->ft_idx) {
                        spin_unlock(&fld->lcf_lock);
                         OBD_FREE_PTR(target);
@@ -243,8 +201,7 @@ int fld_client_add_target(struct lu_client_fld *fld,
         target->ft_srv = tar->ft_srv;
         target->ft_idx = tar->ft_idx;
 
-        cfs_list_add_tail(&target->ft_chain,
-                          &fld->lcf_targets);
+       list_add_tail(&target->ft_chain, &fld->lcf_targets);
 
         fld->lcf_count++;
        spin_unlock(&fld->lcf_lock);
@@ -260,11 +217,10 @@ int fld_client_del_target(struct lu_client_fld *fld, __u64 idx)
        ENTRY;
 
        spin_lock(&fld->lcf_lock);
-       cfs_list_for_each_entry_safe(target, tmp,
-                                    &fld->lcf_targets, ft_chain) {
+       list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
                if (target->ft_idx == idx) {
                        fld->lcf_count--;
-                       cfs_list_del(&target->ft_chain);
+                       list_del(&target->ft_chain);
                        spin_unlock(&fld->lcf_lock);
 
                         if (target->ft_exp != NULL)
@@ -361,7 +317,7 @@ int fld_client_init(struct lu_client_fld *fld,
        spin_lock_init(&fld->lcf_lock);
         fld->lcf_hash = &fld_hash[hash];
         fld->lcf_flags = LUSTRE_FLD_INIT;
-        CFS_INIT_LIST_HEAD(&fld->lcf_targets);
+       INIT_LIST_HEAD(&fld->lcf_targets);
 
         cache_size = FLD_CLIENT_CACHE_SIZE /
                 sizeof(struct fld_cache_entry);
@@ -397,10 +353,9 @@ void fld_client_fini(struct lu_client_fld *fld)
        ENTRY;
 
        spin_lock(&fld->lcf_lock);
-        cfs_list_for_each_entry_safe(target, tmp,
-                                     &fld->lcf_targets, ft_chain) {
+       list_for_each_entry_safe(target, tmp, &fld->lcf_targets, ft_chain) {
                 fld->lcf_count--;
-                cfs_list_del(&target->ft_chain);
+               list_del(&target->ft_chain);
                 if (target->ft_exp != NULL)
                         class_export_put(target->ft_exp);
                 OBD_FREE_PTR(target);
@@ -477,17 +432,18 @@ again:
        req->rq_reply_portal = MDC_REPLY_PORTAL;
         ptlrpc_at_set_req_timeout(req);
 
-       fld_enter_request(&exp->exp_obd->u.cli);
+       obd_get_request_slot(&exp->exp_obd->u.cli);
        rc = ptlrpc_queue_wait(req);
-       fld_exit_request(&exp->exp_obd->u.cli);
+       obd_put_request_slot(&exp->exp_obd->u.cli);
        if (rc != 0) {
-               if (rc == -EWOULDBLOCK) {
-                       /* For no_delay req(see above), EWOULDBLOCK means the
-                        * connection is being evicted, but this seq lookup
-                        * should not return error, since it would cause
-                        * unecessary failure of the application, instead
-                        * it should retry here */
+               if (rc == -EWOULDBLOCK || rc == -ESHUTDOWN) {
+                       /* For no_delay req(see above), EWOULDBLOCK and
+                        * ESHUTDOWN means the connection is being evicted,
+                        * but this seq lookup should not return error,
+                        * since it would cause unecessary failure of the
+                        * application, instead it should retry here */
                        ptlrpc_req_finished(req);
+                       rc = 0;
                        goto again;
                }
                GOTO(out_req, rc);