Whamcloud - gitweb
LU-4420 fld: retry fld rpc even for ESHUTDOWN
[fs/lustre-release.git] / lustre / fld / fld_request.c
index 93d1c1e..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)
@@ -282,33 +238,32 @@ EXPORT_SYMBOL(fld_client_del_target);
 #ifdef LPROCFS
 static int fld_client_proc_init(struct lu_client_fld *fld)
 {
-        int rc;
-        ENTRY;
-
-        fld->lcf_proc_dir = lprocfs_register(fld->lcf_name,
-                                             fld_type_proc_dir,
-                                             NULL, NULL);
+       int rc;
+       ENTRY;
 
-        if (IS_ERR(fld->lcf_proc_dir)) {
-                CERROR("%s: LProcFS failed in fld-init\n",
-                       fld->lcf_name);
-                rc = PTR_ERR(fld->lcf_proc_dir);
-                RETURN(rc);
-        }
+       fld->lcf_proc_dir = lprocfs_seq_register(fld->lcf_name,
+                                                fld_type_proc_dir,
+                                                NULL, NULL);
+       if (IS_ERR(fld->lcf_proc_dir)) {
+               CERROR("%s: LProcFS failed in fld-init\n",
+                      fld->lcf_name);
+               rc = PTR_ERR(fld->lcf_proc_dir);
+               RETURN(rc);
+       }
 
-        rc = lprocfs_add_vars(fld->lcf_proc_dir,
-                              fld_client_proc_list, fld);
-        if (rc) {
-                CERROR("%s: Can't init FLD proc, rc %d\n",
-                       fld->lcf_name, rc);
-                GOTO(out_cleanup, rc);
-        }
+       rc = lprocfs_seq_add_vars(fld->lcf_proc_dir,
+                                 fld_client_proc_list, fld);
+       if (rc) {
+               CERROR("%s: Can't init FLD proc, rc %d\n",
+                      fld->lcf_name, rc);
+               GOTO(out_cleanup, rc);
+       }
 
-        RETURN(0);
+       RETURN(0);
 
 out_cleanup:
-        fld_client_proc_fini(fld);
-        return rc;
+       fld_client_proc_fini(fld);
+       return rc;
 }
 
 void fld_client_proc_fini(struct lu_client_fld *fld)
@@ -362,7 +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);
@@ -398,10 +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);
@@ -419,55 +373,101 @@ void fld_client_fini(struct lu_client_fld *fld)
 EXPORT_SYMBOL(fld_client_fini);
 
 int fld_client_rpc(struct obd_export *exp,
-                   struct lu_seq_range *range, __u32 fld_op)
+                  struct lu_seq_range *range, __u32 fld_op,
+                  struct ptlrpc_request **reqp)
 {
-       struct ptlrpc_request *req;
+       struct ptlrpc_request *req = NULL;
        struct lu_seq_range   *prange;
        __u32                 *op;
-       int                    rc;
+       int                    rc = 0;
        struct obd_import     *imp;
        ENTRY;
 
        LASSERT(exp != NULL);
 
+again:
        imp = class_exp2cliimp(exp);
-       req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY, LUSTRE_MDS_VERSION,
-                                       FLD_QUERY);
-        if (req == NULL)
-                RETURN(-ENOMEM);
-
-        op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
-        *op = fld_op;
+       switch (fld_op) {
+       case FLD_QUERY:
+               req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_QUERY,
+                                               LUSTRE_MDS_VERSION, FLD_QUERY);
+               if (req == NULL)
+                       RETURN(-ENOMEM);
+
+               /* XXX: only needed when talking to old server(< 2.6), it should
+                * be removed when < 2.6 server is not supported */
+               op = req_capsule_client_get(&req->rq_pill, &RMF_FLD_OPC);
+               *op = FLD_LOOKUP;
+
+               /* For MDS_MDS seq lookup, it will always use LWP connection,
+                * but LWP will be evicted after restart, so cause the error.
+                * so we will set no_delay for seq lookup request, once the
+                * request fails because of the eviction. always retry here */
+               if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
+                       req->rq_allow_replay = 1;
+                       req->rq_no_delay = 1;
+               }
+               break;
+       case FLD_READ:
+               req = ptlrpc_request_alloc_pack(imp, &RQF_FLD_READ,
+                                               LUSTRE_MDS_VERSION, FLD_READ);
+               if (req == NULL)
+                       RETURN(-ENOMEM);
+
+               req_capsule_set_size(&req->rq_pill, &RMF_GENERIC_DATA,
+                                    RCL_SERVER, PAGE_CACHE_SIZE);
+               break;
+       default:
+               rc = -EINVAL;
+               break;
+       }
 
-        prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
-        *prange = *range;
+       if (rc != 0)
+               RETURN(rc);
 
-        ptlrpc_request_set_replen(req);
+       prange = req_capsule_client_get(&req->rq_pill, &RMF_FLD_MDFLD);
+       *prange = *range;
+       ptlrpc_request_set_replen(req);
         req->rq_request_portal = FLD_REQUEST_PORTAL;
+       req->rq_reply_portal = MDC_REPLY_PORTAL;
         ptlrpc_at_set_req_timeout(req);
 
-       if (fld_op == FLD_LOOKUP &&
-           imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS)
-               req->rq_allow_replay = 1;
-
-        if (fld_op != FLD_LOOKUP)
-                mdc_get_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
-        fld_enter_request(&exp->exp_obd->u.cli);
-        rc = ptlrpc_queue_wait(req);
-        fld_exit_request(&exp->exp_obd->u.cli);
-        if (fld_op != FLD_LOOKUP)
-                mdc_put_rpc_lock(exp->exp_obd->u.cli.cl_rpc_lock, NULL);
-        if (rc)
-                GOTO(out_req, rc);
+       obd_get_request_slot(&exp->exp_obd->u.cli);
+       rc = ptlrpc_queue_wait(req);
+       obd_put_request_slot(&exp->exp_obd->u.cli);
+       if (rc != 0) {
+               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);
+       }
 
-        prange = req_capsule_server_get(&req->rq_pill, &RMF_FLD_MDFLD);
-        if (prange == NULL)
-                GOTO(out_req, rc = -EFAULT);
-        *range = *prange;
-        EXIT;
+       if (fld_op == FLD_QUERY) {
+               prange = req_capsule_server_get(&req->rq_pill,
+                                               &RMF_FLD_MDFLD);
+               if (prange == NULL)
+                       GOTO(out_req, rc = -EFAULT);
+               *range = *prange;
+       }
+
+       EXIT;
 out_req:
-        ptlrpc_req_finished(req);
-        return rc;
+       if (rc != 0 || reqp == NULL) {
+               ptlrpc_req_finished(req);
+               req = NULL;
+       }
+
+       if (reqp != NULL)
+               *reqp = req;
+
+       return rc;
 }
 
 int fld_client_lookup(struct lu_client_fld *fld, seqno_t seq, mdsno_t *mds,
@@ -504,7 +504,7 @@ int fld_client_lookup(struct lu_client_fld *fld, seqno_t seq, mdsno_t *mds,
        } else
 #endif
        {
-               rc = fld_client_rpc(target->ft_exp, &res, FLD_LOOKUP);
+               rc = fld_client_rpc(target->ft_exp, &res, FLD_QUERY, NULL);
        }
 
        if (rc == 0) {
@@ -528,9 +528,9 @@ struct proc_dir_entry *fld_type_proc_dir;
 
 static int __init fld_mod_init(void)
 {
-       fld_type_proc_dir = lprocfs_register(LUSTRE_FLD_NAME,
-                                            proc_lustre_root,
-                                            NULL, NULL);
+       fld_type_proc_dir = lprocfs_seq_register(LUSTRE_FLD_NAME,
+                                                proc_lustre_root,
+                                                NULL, NULL);
        if (IS_ERR(fld_type_proc_dir))
                return PTR_ERR(fld_type_proc_dir);