Whamcloud - gitweb
LU-911 osd: zerocopy methods in ldiskfs osd
[fs/lustre-release.git] / lustre / osd-ldiskfs / osd_lproc.c
index 6f1bbac..2bc3f37 100644 (file)
@@ -28,6 +28,8 @@
 /*
  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2011, Whamcloud, Inc.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 
 #ifdef LPROCFS
 
-static const char *osd_counter_names[LPROC_OSD_NR] = {
+void osd_brw_stats_update(struct osd_device *osd, struct osd_iobuf *iobuf)
+{
+        struct brw_stats *s = &osd->od_brw_stats;
+        unsigned long    *last_block = NULL;
+        struct page     **pages = iobuf->dr_pages;
+        struct page      *last_page = NULL;
+        unsigned long     discont_pages = 0;
+        unsigned long     discont_blocks = 0;
+        unsigned long    *blocks = iobuf->dr_blocks;
+        int               i, nr_pages = iobuf->dr_npages;
+        int               blocks_per_page;
+        int               rw = iobuf->dr_rw;
+
+        if (unlikely(nr_pages == 0))
+                return;
+
+        blocks_per_page = CFS_PAGE_SIZE >> osd_sb(osd)->s_blocksize_bits;
+
+        lprocfs_oh_tally_log2(&s->hist[BRW_R_PAGES+rw], nr_pages);
+
+        while (nr_pages-- > 0) {
+                if (last_page && (*pages)->index != (last_page->index + 1))
+                        discont_pages++;
+                last_page = *pages;
+                pages++;
+                for (i = 0; i < blocks_per_page; i++) {
+                        if (last_block && *blocks != (*last_block + 1))
+                                discont_blocks++;
+                        last_block = blocks++;
+                }
+        }
+
+        lprocfs_oh_tally(&s->hist[BRW_R_DISCONT_PAGES+rw], discont_pages);
+        lprocfs_oh_tally(&s->hist[BRW_R_DISCONT_BLOCKS+rw], discont_blocks);
+}
+
+#define pct(a, b) (b ? a * 100 / b : 0)
+
+static void display_brw_stats(struct seq_file *seq, char *name, char *units,
+        struct obd_histogram *read, struct obd_histogram *write, int scale)
+{
+        unsigned long read_tot, write_tot, r, w, read_cum = 0, write_cum = 0;
+        int i;
+
+        seq_printf(seq, "\n%26s read      |     write\n", " ");
+        seq_printf(seq, "%-22s %-5s %% cum %% |  %-11s %% cum %%\n",
+                   name, units, units);
+
+        read_tot = lprocfs_oh_sum(read);
+        write_tot = lprocfs_oh_sum(write);
+        for (i = 0; i < OBD_HIST_MAX; i++) {
+                r = read->oh_buckets[i];
+                w = write->oh_buckets[i];
+                read_cum += r;
+                write_cum += w;
+                if (read_cum == 0 && write_cum == 0)
+                        continue;
+
+                if (!scale)
+                        seq_printf(seq, "%u", i);
+                else if (i < 10)
+                        seq_printf(seq, "%u", scale << i);
+                else if (i < 20)
+                        seq_printf(seq, "%uK", scale << (i-10));
+                else
+                        seq_printf(seq, "%uM", scale << (i-20));
+
+                seq_printf(seq, ":\t\t%10lu %3lu %3lu   | %4lu %3lu %3lu\n",
+                           r, pct(r, read_tot), pct(read_cum, read_tot),
+                           w, pct(w, write_tot), pct(write_cum, write_tot));
+
+                if (read_cum == read_tot && write_cum == write_tot)
+                        break;
+        }
+}
+
+static void brw_stats_show(struct seq_file *seq, struct brw_stats *brw_stats)
+{
+        struct timeval now;
+
+        /* this sampling races with updates */
+        cfs_gettimeofday(&now);
+        seq_printf(seq, "snapshot_time:         %lu.%lu (secs.usecs)\n",
+                   now.tv_sec, now.tv_usec);
+
+        display_brw_stats(seq, "pages per bulk r/w", "rpcs",
+                          &brw_stats->hist[BRW_R_PAGES],
+                          &brw_stats->hist[BRW_W_PAGES], 1);
+
+        display_brw_stats(seq, "discontiguous pages", "rpcs",
+                          &brw_stats->hist[BRW_R_DISCONT_PAGES],
+                          &brw_stats->hist[BRW_W_DISCONT_PAGES], 0);
+
+        display_brw_stats(seq, "discontiguous blocks", "rpcs",
+                          &brw_stats->hist[BRW_R_DISCONT_BLOCKS],
+                          &brw_stats->hist[BRW_W_DISCONT_BLOCKS], 0);
+
+        display_brw_stats(seq, "disk fragmented I/Os", "ios",
+                          &brw_stats->hist[BRW_R_DIO_FRAGS],
+                          &brw_stats->hist[BRW_W_DIO_FRAGS], 0);
+
+        display_brw_stats(seq, "disk I/Os in flight", "ios",
+                          &brw_stats->hist[BRW_R_RPC_HIST],
+                          &brw_stats->hist[BRW_W_RPC_HIST], 0);
+
+        display_brw_stats(seq, "I/O time (1/1000s)", "ios",
+                          &brw_stats->hist[BRW_R_IO_TIME],
+                          &brw_stats->hist[BRW_W_IO_TIME], 1000 / CFS_HZ);
+
+        display_brw_stats(seq, "disk I/O size", "ios",
+                          &brw_stats->hist[BRW_R_DISK_IOSIZE],
+                          &brw_stats->hist[BRW_W_DISK_IOSIZE], 1);
+}
+
+#undef pct
+
+static int osd_brw_stats_seq_show(struct seq_file *seq, void *v)
+{
+        struct osd_device *osd = seq->private;
+
+        brw_stats_show(seq, &osd->od_brw_stats);
+
+        return 0;
+}
+
+static ssize_t osd_brw_stats_seq_write(struct file *file, const char *buf,
+                                       size_t len, loff_t *off)
+{
+        struct seq_file *seq = file->private_data;
+        struct osd_device *osd = seq->private;
+        int i;
+
+        for (i = 0; i < BRW_LAST; i++)
+                lprocfs_oh_clear(&osd->od_brw_stats.hist[i]);
+
+        return len;
+}
+
+LPROC_SEQ_FOPS(osd_brw_stats);
+
+static int osd_stats_init(struct osd_device *osd)
+{
+        int i, result;
+        ENTRY;
+
+        for (i = 0; i < BRW_LAST; i++)
+                cfs_spin_lock_init(&osd->od_brw_stats.hist[i].oh_lock);
+
+        osd->od_stats = lprocfs_alloc_stats(LPROC_OSD_LAST, 0);
+        if (osd->od_stats != NULL) {
+                result = lprocfs_register_stats(osd->od_proc_entry, "stats",
+                                                osd->od_stats);
+                if (result)
+                        GOTO(out, result);
+
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_GET_PAGE,
+                                     LPROCFS_CNTR_AVGMINMAX|LPROCFS_CNTR_STDDEV,
+                                     "get_page", "usec");
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_NO_PAGE,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "get_page_failures", "num");
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_CACHE_ACCESS,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "cache_access", "pages");
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_CACHE_HIT,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "cache_hit", "pages");
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_CACHE_MISS,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "cache_miss", "pages");
+#if OSD_THANDLE_STATS
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_THANDLE_STARTING,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "thandle starting", "usec");
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_THANDLE_OPEN,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "thandle open", "usec");
+                lprocfs_counter_init(osd->od_stats, LPROC_OSD_THANDLE_CLOSING,
+                                     LPROCFS_CNTR_AVGMINMAX,
+                                     "thandle closing", "usec");
+#endif
+                lprocfs_seq_create(osd->od_proc_entry, "brw_stats",
+                                   0444, &osd_brw_stats_fops, osd);
+        } else
+                result = -ENOMEM;
+
+out:
+        RETURN(result);
+}
+
+static const char *osd_counter_names[] = {
 #if OSD_THANDLE_STATS
         [LPROC_OSD_THANDLE_STARTING] = "thandle starting",
         [LPROC_OSD_THANDLE_OPEN]     = "thandle open",
@@ -85,6 +277,9 @@ int osd_procfs_init(struct osd_device *osd, const char *name)
         rc = lu_time_init(&osd->od_stats,
                           osd->od_proc_entry,
                           osd_counter_names, ARRAY_SIZE(osd_counter_names));
+
+        rc = osd_stats_init(osd);
+
         EXIT;
 out:
         if (rc)