pNFS/flexfiles: add layoutstats tracking
authorTrond Myklebust <trond.myklebust@primarydata.com>
Tue, 23 Jun 2015 11:51:59 +0000 (19:51 +0800)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 24 Jun 2015 14:53:10 +0000 (10:53 -0400)
Reviewed-by: Jeff Layton <jeff.layton@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/flexfilelayout/flexfilelayout.c
fs/nfs/flexfilelayout/flexfilelayout.h

index 0f1410c94827e95d1985eba0cd4b7f42d42712ee..8ac31fbf040a6b1fe9666145693c3cbaed25dc02 100644 (file)
@@ -412,6 +412,130 @@ ff_layout_get_lseg_count(struct nfs4_ff_layout_segment *fls)
        return 1;
 }
 
+static void
+nfs4_ff_start_busy_timer(struct nfs4_ff_busy_timer *timer)
+{
+       ktime_t old, new;
+
+       /*
+        * Note: careful here!
+        * If the counter is zero, then we must not increment it until after
+        * we've set the start_time.
+        * If we were instead to use atomic_inc_return(), then another
+        * request might come in, bump, and then call end_busy_timer()
+        * before we've set the timer->start_time.
+        */
+       old = timer->start_time;
+       if (atomic_inc_not_zero(&timer->n_ops) == 0) {
+               new = ktime_get();
+               cmpxchg(&timer->start_time.tv64, old.tv64, new.tv64);
+               atomic_inc(&timer->n_ops);
+       }
+}
+
+static ktime_t
+nfs4_ff_end_busy_timer(struct nfs4_ff_busy_timer *timer)
+{
+       ktime_t start, now;
+
+       now = ktime_get();
+       start.tv64 = xchg(&timer->start_time.tv64, now.tv64);
+       atomic_dec(&timer->n_ops);
+       return ktime_sub(now, start);
+}
+
+static ktime_t
+nfs4_ff_layout_calc_completion_time(struct rpc_task *task)
+{
+       return ktime_sub(ktime_get(), task->tk_start);
+}
+
+static void
+nfs4_ff_layoutstat_start_io(struct nfs4_ff_layoutstat *layoutstat)
+{
+       nfs4_ff_start_busy_timer(&layoutstat->busy_timer);
+}
+
+static void
+nfs4_ff_layout_stat_io_update_requested(struct nfs4_ff_layoutstat *layoutstat,
+               __u64 requested)
+{
+       struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
+
+       iostat->ops_requested++;
+       iostat->bytes_requested += requested;
+}
+
+static void
+nfs4_ff_layout_stat_io_update_completed(struct nfs4_ff_layoutstat *layoutstat,
+               __u64 requested,
+               __u64 completed,
+               ktime_t time_completed)
+{
+       struct nfs4_ff_io_stat *iostat = &layoutstat->io_stat;
+       ktime_t timer;
+
+       iostat->ops_completed++;
+       iostat->bytes_completed += completed;
+       iostat->bytes_not_delivered += requested - completed;
+
+       timer = nfs4_ff_end_busy_timer(&layoutstat->busy_timer);
+       iostat->total_busy_time =
+                       ktime_add(iostat->total_busy_time, timer);
+       iostat->aggregate_completion_time =
+                       ktime_add(iostat->aggregate_completion_time, time_completed);
+}
+
+static void
+nfs4_ff_layout_stat_io_start_read(struct nfs4_ff_layout_mirror *mirror,
+               __u64 requested)
+{
+       spin_lock(&mirror->lock);
+       nfs4_ff_layoutstat_start_io(&mirror->read_stat);
+       nfs4_ff_layout_stat_io_update_requested(&mirror->read_stat, requested);
+       spin_unlock(&mirror->lock);
+}
+
+static void
+nfs4_ff_layout_stat_io_end_read(struct rpc_task *task,
+               struct nfs4_ff_layout_mirror *mirror,
+               __u64 requested,
+               __u64 completed)
+{
+       spin_lock(&mirror->lock);
+       nfs4_ff_layout_stat_io_update_completed(&mirror->read_stat,
+                       requested, completed,
+                       nfs4_ff_layout_calc_completion_time(task));
+       spin_unlock(&mirror->lock);
+}
+
+static void
+nfs4_ff_layout_stat_io_start_write(struct nfs4_ff_layout_mirror *mirror,
+               __u64 requested)
+{
+       spin_lock(&mirror->lock);
+       nfs4_ff_layoutstat_start_io(&mirror->write_stat);
+       nfs4_ff_layout_stat_io_update_requested(&mirror->write_stat, requested);
+       spin_unlock(&mirror->lock);
+}
+
+static void
+nfs4_ff_layout_stat_io_end_write(struct rpc_task *task,
+               struct nfs4_ff_layout_mirror *mirror,
+               __u64 requested,
+               __u64 completed,
+               enum nfs3_stable_how committed)
+{
+       if (committed == NFS_UNSTABLE)
+               requested = completed = 0;
+
+       spin_lock(&mirror->lock);
+       nfs4_ff_layout_stat_io_update_completed(&mirror->write_stat,
+                       requested, completed,
+                       nfs4_ff_layout_calc_completion_time(task));
+       spin_unlock(&mirror->lock);
+}
+
 static int
 ff_layout_alloc_commit_info(struct pnfs_layout_segment *lseg,
                            struct nfs_commit_info *cinfo,
@@ -906,6 +1030,10 @@ ff_layout_reset_to_mds(struct pnfs_layout_segment *lseg, int idx)
 static int ff_layout_read_prepare_common(struct rpc_task *task,
                                         struct nfs_pgio_header *hdr)
 {
+       nfs4_ff_layout_stat_io_start_read(
+                       FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
+                       hdr->args.count);
+
        if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
                rpc_exit(task, -EIO);
                return -EIO;
@@ -959,15 +1087,15 @@ static void ff_layout_read_prepare_v4(struct rpc_task *task, void *data)
 {
        struct nfs_pgio_header *hdr = data;
 
-       if (ff_layout_read_prepare_common(task, hdr))
-               return;
-
        if (ff_layout_setup_sequence(hdr->ds_clp,
                                     &hdr->args.seq_args,
                                     &hdr->res.seq_res,
                                     task))
                return;
 
+       if (ff_layout_read_prepare_common(task, hdr))
+               return;
+
        if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
                        hdr->args.lock_context, FMODE_READ) == -EIO)
                rpc_exit(task, -EIO); /* lost lock, terminate I/O */
@@ -979,6 +1107,10 @@ static void ff_layout_read_call_done(struct rpc_task *task, void *data)
 
        dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 
+       nfs4_ff_layout_stat_io_end_read(task,
+                       FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
+                       hdr->args.count, hdr->res.count);
+
        if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
            task->tk_status == 0) {
                nfs4_sequence_done(task, &hdr->res.seq_res);
@@ -1080,6 +1212,10 @@ static int ff_layout_commit_done_cb(struct rpc_task *task,
 static int ff_layout_write_prepare_common(struct rpc_task *task,
                                          struct nfs_pgio_header *hdr)
 {
+       nfs4_ff_layout_stat_io_start_write(
+                       FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
+                       hdr->args.count);
+
        if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
                rpc_exit(task, -EIO);
                return -EIO;
@@ -1113,15 +1249,15 @@ static void ff_layout_write_prepare_v4(struct rpc_task *task, void *data)
 {
        struct nfs_pgio_header *hdr = data;
 
-       if (ff_layout_write_prepare_common(task, hdr))
-               return;
-
        if (ff_layout_setup_sequence(hdr->ds_clp,
                                     &hdr->args.seq_args,
                                     &hdr->res.seq_res,
                                     task))
                return;
 
+       if (ff_layout_write_prepare_common(task, hdr))
+               return;
+
        if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
                        hdr->args.lock_context, FMODE_WRITE) == -EIO)
                rpc_exit(task, -EIO); /* lost lock, terminate I/O */
@@ -1131,6 +1267,11 @@ static void ff_layout_write_call_done(struct rpc_task *task, void *data)
 {
        struct nfs_pgio_header *hdr = data;
 
+       nfs4_ff_layout_stat_io_end_write(task,
+                       FF_LAYOUT_COMP(hdr->lseg, hdr->pgio_mirror_idx),
+                       hdr->args.count, hdr->res.count,
+                       hdr->res.verf->committed);
+
        if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
            task->tk_status == 0) {
                nfs4_sequence_done(task, &hdr->res.seq_res);
@@ -1149,8 +1290,17 @@ static void ff_layout_write_count_stats(struct rpc_task *task, void *data)
            &NFS_CLIENT(hdr->inode)->cl_metrics[NFSPROC4_CLNT_WRITE]);
 }
 
+static void ff_layout_commit_prepare_common(struct rpc_task *task,
+               struct nfs_commit_data *cdata)
+{
+       nfs4_ff_layout_stat_io_start_write(
+                       FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
+                       0);
+}
+
 static void ff_layout_commit_prepare_v3(struct rpc_task *task, void *data)
 {
+       ff_layout_commit_prepare_common(task, data);
        rpc_call_start(task);
 }
 
@@ -1158,10 +1308,30 @@ static void ff_layout_commit_prepare_v4(struct rpc_task *task, void *data)
 {
        struct nfs_commit_data *wdata = data;
 
-       ff_layout_setup_sequence(wdata->ds_clp,
+       if (ff_layout_setup_sequence(wdata->ds_clp,
                                 &wdata->args.seq_args,
                                 &wdata->res.seq_res,
-                                task);
+                                task))
+               return;
+       ff_layout_commit_prepare_common(task, data);
+}
+
+static void ff_layout_commit_done(struct rpc_task *task, void *data)
+{
+       struct nfs_commit_data *cdata = data;
+       struct nfs_page *req;
+       __u64 count = 0;
+
+       if (task->tk_status == 0) {
+               list_for_each_entry(req, &cdata->pages, wb_list)
+                       count += req->wb_bytes;
+       }
+
+       nfs4_ff_layout_stat_io_end_write(task,
+                       FF_LAYOUT_COMP(cdata->lseg, cdata->ds_commit_index),
+                       count, count, NFS_FILE_SYNC);
+
+       pnfs_generic_write_commit_done(task, data);
 }
 
 static void ff_layout_commit_count_stats(struct rpc_task *task, void *data)
@@ -1202,14 +1372,14 @@ static const struct rpc_call_ops ff_layout_write_call_ops_v4 = {
 
 static const struct rpc_call_ops ff_layout_commit_call_ops_v3 = {
        .rpc_call_prepare = ff_layout_commit_prepare_v3,
-       .rpc_call_done = pnfs_generic_write_commit_done,
+       .rpc_call_done = ff_layout_commit_done,
        .rpc_count_stats = ff_layout_commit_count_stats,
        .rpc_release = pnfs_generic_commit_release,
 };
 
 static const struct rpc_call_ops ff_layout_commit_call_ops_v4 = {
        .rpc_call_prepare = ff_layout_commit_prepare_v4,
-       .rpc_call_done = pnfs_generic_write_commit_done,
+       .rpc_call_done = ff_layout_commit_done,
        .rpc_count_stats = ff_layout_commit_count_stats,
        .rpc_release = pnfs_generic_commit_release,
 };
@@ -1253,7 +1423,6 @@ ff_layout_read_pagelist(struct nfs_pgio_header *hdr)
        fh = nfs4_ff_layout_select_ds_fh(lseg, idx);
        if (fh)
                hdr->args.fh = fh;
-
        /*
         * Note that if we ever decide to split across DSes,
         * then we may need to handle dense-like offsets.
@@ -1382,6 +1551,7 @@ static int ff_layout_initiate_commit(struct nfs_commit_data *data, int how)
        fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
        if (fh)
                data->args.fh = fh;
+
        return nfs_initiate_commit(ds_clnt, data, ds->ds_clp->rpc_ops,
                                   vers == 3 ? &ff_layout_commit_call_ops_v3 :
                                               &ff_layout_commit_call_ops_v4,
index 275e4b735247f49224e9ccbe39f3e9726eae0b7a..f7493f7cf13c44bcbf39c9a0bd66db9ae63fc9d6 100644 (file)
@@ -41,6 +41,26 @@ struct nfs4_ff_layout_ds_err {
        struct nfs4_deviceid            deviceid;
 };
 
+struct nfs4_ff_io_stat {
+       __u64                           ops_requested;
+       __u64                           bytes_requested;
+       __u64                           ops_completed;
+       __u64                           bytes_completed;
+       __u64                           bytes_not_delivered;
+       ktime_t                         total_busy_time;
+       ktime_t                         aggregate_completion_time;
+};
+
+struct nfs4_ff_busy_timer {
+       ktime_t start_time;
+       atomic_t n_ops;
+};
+
+struct nfs4_ff_layoutstat {
+       struct nfs4_ff_io_stat io_stat;
+       struct nfs4_ff_busy_timer busy_timer;
+};
+
 struct nfs4_ff_layout_mirror {
        u32                             ds_count;
        u32                             efficiency;
@@ -52,6 +72,8 @@ struct nfs4_ff_layout_mirror {
        u32                             gid;
        struct rpc_cred                 *cred;
        spinlock_t                      lock;
+       struct nfs4_ff_layoutstat       read_stat;
+       struct nfs4_ff_layoutstat       write_stat;
 };
 
 struct nfs4_ff_layout_segment {