LU-15216 lmv: improve MDT QOS space balance

When MDTs are not balanced, QOS code tries to keep subdirectory
creation local to the same MDT when it is deep in the directory
tree, to avoid creating too many remote directories, but the
existing weight to stay on the parent MDT until 50% of other MDTs
is too radical, and causes mkdirs to be "stuck" on the same MDT.

* remove "lq_threshold_rr" from above calculation because the check
  in ltd_qos_is_usable() handles this, so use only "dir_depth".
* the factor is changed to "16 / (dir_depth + 10)", then it's less
  likely to stick to the parent MDT for top levels, while more
  likely to stay on the parent MDT for low levels:
  depth=0 -> 160%, depth=4 -> 114%, depth=6 -> 100%,
  depth=8 -> 88%, depth=12 -> 72%
* rename lli_depth to lli_dir_depth to make usage more clear.

Signed-off-by: Lai Siyao <lai.siyao@whamcloud.com>
Change-Id: Iec6b77919b630d4baee6d54bee7bdb8ca9fb8574
Reviewed-on: https://review.whamcloud.com/45544
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Hongchao Zhang <hongchao@whamcloud.com>

diff --git a/lustre/llite/dir.c b/lustre/llite/dir.c
index 6294d23..4006417 100644 (file)
--- a/lustre/llite/dir.c
+++ b/lustre/llite/dir.c
@@ -511,7 +511,7 @@ static int ll_dir_setdirstripe(struct dentry *dparent, struct lmv_user_md *lump,
        if (IS_ERR(op_data))
                RETURN(PTR_ERR(op_data));
 
-       op_data->op_dir_depth = ll_i2info(parent)->lli_depth;
+       op_data->op_dir_depth = ll_i2info(parent)->lli_dir_depth;
 
        if (ll_sbi_has_encrypt(sbi) &&
            (IS_ENCRYPTED(parent) ||

diff --git a/lustre/llite/llite_internal.h b/lustre/llite/llite_internal.h
index 1fc1222..d066f98 100644 (file)
--- a/lustre/llite/llite_internal.h
+++ b/lustre/llite/llite_internal.h
@@ -170,7 +170,7 @@ struct ll_inode_info {
                         * -- I am the owner of dir statahead. */
                        pid_t                           lli_opendir_pid;
                        /* directory depth to ROOT */
-                       unsigned short                  lli_depth;
+                       unsigned short                  lli_dir_depth;
                        /* stat will try to access statahead entries or start
                         * statahead if this flag is set, and this flag will be
                         * set upon dir open, and cleared when dir is closed,

diff --git a/lustre/llite/llite_lib.c b/lustre/llite/llite_lib.c
index 86c6611..3872cd4 100644 (file)
--- a/lustre/llite/llite_lib.c
+++ b/lustre/llite/llite_lib.c
@@ -2695,8 +2695,9 @@ void ll_update_dir_depth(struct inode *dir, struct inode *inode)
                return;
 
        lli = ll_i2info(inode);
-       lli->lli_depth = ll_i2info(dir)->lli_depth + 1;
-       CDEBUG(D_INODE, DFID" depth %hu\n", PFID(&lli->lli_fid), lli->lli_depth);
+       lli->lli_dir_depth = ll_i2info(dir)->lli_dir_depth + 1;
+       CDEBUG(D_INODE, DFID" depth %hu\n",
+              PFID(&lli->lli_fid), lli->lli_dir_depth);
 }
 
 void ll_truncate_inode_pages_final(struct inode *inode)

diff --git a/lustre/llite/namei.c b/lustre/llite/namei.c
index 6f372c7..80b8460 100644 (file)
--- a/lustre/llite/namei.c
+++ b/lustre/llite/namei.c
@@ -1495,7 +1495,7 @@ static void ll_qos_mkdir_prep(struct md_op_data *op_data, struct inode *dir)
        struct ll_inode_info *lli = ll_i2info(dir);
        struct lmv_stripe_md *lsm;
 
-       op_data->op_dir_depth = lli->lli_depth;
+       op_data->op_dir_depth = lli->lli_dir_depth;
 
        /* parent directory is striped */
        if (unlikely(lli->lli_lsm_md))
@@ -1524,11 +1524,11 @@ static void ll_qos_mkdir_prep(struct md_op_data *op_data, struct inode *dir)
 
        if (lsm->lsm_md_max_inherit != LMV_INHERIT_NONE &&
            (lsm->lsm_md_max_inherit == LMV_INHERIT_UNLIMITED ||
-            lsm->lsm_md_max_inherit >= lli->lli_depth)) {
+            lsm->lsm_md_max_inherit >= lli->lli_dir_depth)) {
                op_data->op_flags |= MF_QOS_MKDIR;
                if (lsm->lsm_md_max_inherit_rr != LMV_INHERIT_RR_NONE &&
                    (lsm->lsm_md_max_inherit_rr == LMV_INHERIT_RR_UNLIMITED ||
-                    lsm->lsm_md_max_inherit_rr >= lli->lli_depth))
+                    lsm->lsm_md_max_inherit_rr >= lli->lli_dir_depth))
                        op_data->op_flags |= MF_RR_MKDIR;
                CDEBUG(D_INODE, DFID" requests qos mkdir %#x\n",
                       PFID(&lli->lli_fid), op_data->op_flags);

diff --git a/lustre/lmv/lmv_obd.c b/lustre/lmv/lmv_obd.c
index 38a2c97..1a1e699 100644 (file)
--- a/lustre/lmv/lmv_obd.c
+++ b/lustre/lmv/lmv_obd.c
@@ -1501,10 +1501,11 @@ static struct lu_tgt_desc *lmv_locate_tgt_qos(struct lmv_obd *lmv, __u32 mdt,
 
        /* if current MDT has above-average space, within range of the QOS
         * threshold, stay on the same MDT to avoid creating needless remote
-        * MDT directories. It's more likely for low level directories.
+        * MDT directories. It's more likely for low level directories
+        * "16 / (dir_depth + 10)" is the factor to make it more unlikely for
+        * top level directories, while more likely for low levels.
         */
-       rand = total_avail * (256 - lmv->lmv_qos.lq_threshold_rr) /
-              (total_usable * 256 * (1 + dir_depth / 4));
+       rand = total_avail * 16 / (total_usable * (dir_depth + 10));
        if (cur && cur->ltd_qos.ltq_avail >= rand) {
                tgt = cur;
                GOTO(unlock, tgt);