Whamcloud - gitweb
LU-2360 ptlrpc: a few fixes for req_history read
[fs/lustre-release.git] / lustre / ptlrpc / lproc_ptlrpc.c
index 50a981f..ab7d146 100644 (file)
@@ -111,6 +111,7 @@ struct ll_rpc_opcode {
         { OBD_PING,         "obd_ping" },
         { OBD_LOG_CANCEL,   "llog_origin_handle_cancel" },
         { OBD_QC_CALLBACK,  "obd_quota_callback" },
+       { OBD_IDX_READ,     "dt_index_read" },
         { LLOG_ORIGIN_HANDLE_CREATE,     "llog_origin_handle_create" },
         { LLOG_ORIGIN_HANDLE_NEXT_BLOCK, "llog_origin_handle_next_block" },
         { LLOG_ORIGIN_HANDLE_READ_HEADER,"llog_origin_handle_read_header" },
@@ -302,14 +303,14 @@ ptlrpc_lprocfs_write_req_history_max(struct file *file, const char *buffer,
         if (val > cfs_num_physpages/(2 * bufpages))
                 return -ERANGE;
 
-       cfs_spin_lock(&svc->srv_lock);
+       spin_lock(&svc->srv_lock);
 
        if (val == 0)
                svc->srv_hist_nrqbds_cpt_max = 0;
        else
                svc->srv_hist_nrqbds_cpt_max = max(1, (val / svc->srv_ncpts));
 
-       cfs_spin_unlock(&svc->srv_lock);
+       spin_unlock(&svc->srv_lock);
 
        return count;
 }
@@ -338,15 +339,15 @@ ptlrpc_lprocfs_wr_threads_min(struct file *file, const char *buffer,
        if (val / svc->srv_ncpts < PTLRPC_NTHRS_INIT)
                return -ERANGE;
 
-       cfs_spin_lock(&svc->srv_lock);
+       spin_lock(&svc->srv_lock);
        if (val > svc->srv_nthrs_cpt_limit * svc->srv_ncpts) {
-               cfs_spin_unlock(&svc->srv_lock);
+               spin_unlock(&svc->srv_lock);
                return -ERANGE;
        }
 
        svc->srv_nthrs_cpt_init = val / svc->srv_ncpts;
 
-       cfs_spin_unlock(&svc->srv_lock);
+       spin_unlock(&svc->srv_lock);
 
        return count;
 }
@@ -391,15 +392,15 @@ ptlrpc_lprocfs_wr_threads_max(struct file *file, const char *buffer,
        if (val / svc->srv_ncpts < PTLRPC_NTHRS_INIT)
                return -ERANGE;
 
-       cfs_spin_lock(&svc->srv_lock);
+       spin_lock(&svc->srv_lock);
        if (val < svc->srv_nthrs_cpt_init * svc->srv_ncpts) {
-               cfs_spin_unlock(&svc->srv_lock);
+               spin_unlock(&svc->srv_lock);
                return -ERANGE;
        }
 
        svc->srv_nthrs_cpt_limit = val / svc->srv_ncpts;
 
-       cfs_spin_unlock(&svc->srv_lock);
+       spin_unlock(&svc->srv_lock);
 
        return count;
 }
@@ -427,7 +428,10 @@ ptlrpc_lprocfs_svc_req_history_seek(struct ptlrpc_service_part *svcpt,
                  * Since the service history is LRU (i.e. culled reqs will
                  * be near the head), we shouldn't have to do long
                  * re-scans */
-                LASSERT (srhi->srhi_seq == srhi->srhi_req->rq_history_seq);
+               LASSERTF(srhi->srhi_seq == srhi->srhi_req->rq_history_seq,
+                        "%s:%d: seek seq "LPU64", request seq "LPU64"\n",
+                        svcpt->scp_service->srv_name, svcpt->scp_cpt,
+                        srhi->srhi_seq, srhi->srhi_req->rq_history_seq);
                LASSERTF(!cfs_list_empty(&svcpt->scp_hist_reqs),
                         "%s:%d: seek offset "LPU64", request seq "LPU64", "
                         "last culled "LPU64"\n",
@@ -453,15 +457,55 @@ ptlrpc_lprocfs_svc_req_history_seek(struct ptlrpc_service_part *svcpt,
         return -ENOENT;
 }
 
+/*
+ * ptlrpc history sequence is used as "position" of seq_file, in some case,
+ * seq_read() will increase "position" to indicate reading the next
+ * element, however, low bits of history sequence are reserved for CPT id
+ * (check the details from comments before ptlrpc_req_add_history), which
+ * means seq_read() might change CPT id of history sequence and never
+ * finish reading of requests on a CPT. To make it work, we have to shift
+ * CPT id to high bits and timestamp to low bits, so seq_read() will only
+ * increase timestamp which can correctly indicate the next position.
+ */
+
+/* convert seq_file pos to cpt */
+#define PTLRPC_REQ_POS2CPT(svc, pos)                   \
+       ((svc)->srv_cpt_bits == 0 ? 0 :                 \
+        (__u64)(pos) >> (64 - (svc)->srv_cpt_bits))
+
+/* make up seq_file pos from cpt */
+#define PTLRPC_REQ_CPT2POS(svc, cpt)                   \
+       ((svc)->srv_cpt_bits == 0 ? 0 :                 \
+        (cpt) << (64 - (svc)->srv_cpt_bits))
+
+/* convert sequence to position */
+#define PTLRPC_REQ_SEQ2POS(svc, seq)                   \
+       ((svc)->srv_cpt_bits == 0 ? (seq) :             \
+        ((seq) >> (svc)->srv_cpt_bits) |               \
+        ((seq) << (64 - (svc)->srv_cpt_bits)))
+
+/* convert position to sequence */
+#define PTLRPC_REQ_POS2SEQ(svc, pos)                   \
+       ((svc)->srv_cpt_bits == 0 ? (pos) :             \
+        ((__u64)(pos) << (svc)->srv_cpt_bits) |        \
+        ((__u64)(pos) >> (64 - (svc)->srv_cpt_bits)))
+
 static void *
 ptlrpc_lprocfs_svc_req_history_start(struct seq_file *s, loff_t *pos)
 {
        struct ptlrpc_service           *svc = s->private;
        struct ptlrpc_service_part      *svcpt;
        struct ptlrpc_srh_iterator      *srhi;
+       unsigned int                    cpt;
        int                             rc;
        int                             i;
 
+       if (sizeof(loff_t) != sizeof(__u64)) { /* can't support */
+               CWARN("Failed to read request history because size of loff_t "
+                     "%d can't match size of u64\n", (int)sizeof(loff_t));
+               return NULL;
+       }
+
        OBD_ALLOC(srhi, sizeof(*srhi));
        if (srhi == NULL)
                return NULL;
@@ -469,14 +513,21 @@ ptlrpc_lprocfs_svc_req_history_start(struct seq_file *s, loff_t *pos)
        srhi->srhi_seq = 0;
        srhi->srhi_req = NULL;
 
-       ptlrpc_service_for_each_part(svcpt, i, svc) {
-               srhi->srhi_idx = i;
+       cpt = PTLRPC_REQ_POS2CPT(svc, *pos);
 
-               cfs_spin_lock(&svcpt->scp_lock);
-               rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi, *pos);
-               cfs_spin_unlock(&svcpt->scp_lock);
+       ptlrpc_service_for_each_part(svcpt, i, svc) {
+               if (i < cpt) /* skip */
+                       continue;
+               if (i > cpt) /* make up the lowest position for this CPT */
+                       *pos = PTLRPC_REQ_CPT2POS(svc, i);
+
+               spin_lock(&svcpt->scp_lock);
+               rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi,
+                               PTLRPC_REQ_POS2SEQ(svc, *pos));
+               spin_unlock(&svcpt->scp_lock);
                if (rc == 0) {
-                       *pos = srhi->srhi_seq;
+                       *pos = PTLRPC_REQ_SEQ2POS(svc, srhi->srhi_seq);
+                       srhi->srhi_idx = i;
                        return srhi;
                }
        }
@@ -501,28 +552,32 @@ ptlrpc_lprocfs_svc_req_history_next(struct seq_file *s,
        struct ptlrpc_service           *svc = s->private;
        struct ptlrpc_srh_iterator      *srhi = iter;
        struct ptlrpc_service_part      *svcpt;
-       int                             rc = 0;
+       __u64                           seq;
+       int                             rc;
        int                             i;
 
        for (i = srhi->srhi_idx; i < svc->srv_ncpts; i++) {
                svcpt = svc->srv_parts[i];
 
-               srhi->srhi_idx = i;
+               if (i > srhi->srhi_idx) { /* reset iterator for a new CPT */
+                       srhi->srhi_req = NULL;
+                       seq = srhi->srhi_seq = 0;
+               } else { /* the next sequence */
+                       seq = srhi->srhi_seq + (1 << svc->srv_cpt_bits);
+               }
 
-               cfs_spin_lock(&svcpt->scp_lock);
-               rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi, *pos + 1);
-               cfs_spin_unlock(&svcpt->scp_lock);
-               if (rc == 0)
-                       break;
+               spin_lock(&svcpt->scp_lock);
+               rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi, seq);
+               spin_unlock(&svcpt->scp_lock);
+               if (rc == 0) {
+                       *pos = PTLRPC_REQ_SEQ2POS(svc, srhi->srhi_seq);
+                       srhi->srhi_idx = i;
+                       return srhi;
+               }
        }
 
-        if (rc != 0) {
-                OBD_FREE(srhi, sizeof(*srhi));
-                return NULL;
-        }
-
-        *pos = srhi->srhi_seq;
-        return srhi;
+       OBD_FREE(srhi, sizeof(*srhi));
+       return NULL;
 }
 
 /* common ost/mdt so_req_printer */
@@ -567,7 +622,7 @@ static int ptlrpc_lprocfs_svc_req_history_show(struct seq_file *s, void *iter)
 
        svcpt = svc->srv_parts[srhi->srhi_idx];
 
-       cfs_spin_lock(&svcpt->scp_lock);
+       spin_lock(&svcpt->scp_lock);
 
        rc = ptlrpc_lprocfs_svc_req_history_seek(svcpt, srhi, srhi->srhi_seq);
 
@@ -593,7 +648,7 @@ static int ptlrpc_lprocfs_svc_req_history_show(struct seq_file *s, void *iter)
                        svc->srv_ops.so_req_printer(s, srhi->srhi_req);
         }
 
-       cfs_spin_unlock(&svcpt->scp_lock);
+       spin_unlock(&svcpt->scp_lock);
        return rc;
 }
 
@@ -701,9 +756,9 @@ static int ptlrpc_lprocfs_wr_hp_ratio(struct file *file, const char *buffer,
        if (val < 0)
                return -ERANGE;
 
-       cfs_spin_lock(&svc->srv_lock);
+       spin_lock(&svc->srv_lock);
        svc->srv_hpreq_ratio = val;
-       cfs_spin_unlock(&svc->srv_lock);
+       spin_unlock(&svc->srv_lock);
 
        return count;
 }
@@ -865,8 +920,8 @@ int lprocfs_wr_evict_client(struct file *file, const char *buffer,
          * the proc entries under the being destroyed export{}, so I have
          * to drop the lock at first here.
          * - jay, jxiong@clusterfs.com */
-        class_incref(obd, __FUNCTION__, cfs_current());
         LPROCFS_EXIT();
+       class_incref(obd, __FUNCTION__, cfs_current());
 
         if (strncmp(tmpbuf, "nid:", 4) == 0)
                 obd_export_evict_by_nid(obd, tmpbuf + 4);
@@ -875,8 +930,8 @@ int lprocfs_wr_evict_client(struct file *file, const char *buffer,
         else
                 obd_export_evict_by_uuid(obd, tmpbuf);
 
+       class_decref(obd, __FUNCTION__, cfs_current());
         LPROCFS_ENTRY();
-        class_decref(obd, __FUNCTION__, cfs_current());
 
 out:
         OBD_FREE(kbuf, BUFLEN);
@@ -1008,12 +1063,12 @@ int lprocfs_wr_pinger_recov(struct file *file, const char *buffer,
                 return -ERANGE;
 
         LPROCFS_CLIMP_CHECK(obd);
-        cfs_spin_lock(&imp->imp_lock);
-        imp->imp_no_pinger_recover = !val;
-        cfs_spin_unlock(&imp->imp_lock);
-        LPROCFS_CLIMP_EXIT(obd);
+       spin_lock(&imp->imp_lock);
+       imp->imp_no_pinger_recover = !val;
+       spin_unlock(&imp->imp_lock);
+       LPROCFS_CLIMP_EXIT(obd);
 
-        return count;
+       return count;
 
 }
 EXPORT_SYMBOL(lprocfs_wr_pinger_recov);