LU-10777 dom: disable read-on-open with resend

The read-on-open can fill more data on reply buffer than
client allocated, this causes buffer re-allocation followed
by resend. Meanwhile FIO read test shows that such resends
perform worse than separate READ RPC. For example:
FIO 8k read is ~50% better without buffer re-allocation
with resend. Considering that there is parameter on MDC
'mdc_dom_min_repsize' to control read-on-open inline buffer
size, there is no sense to keep 'reallocation+resend'
option on MDT. Patch removes it.

Signed-off-by: Mikhail Pershin <mpershin@whamcloud.com>
Change-Id: I7eb9d64f5551789e93b1f7676f61c0e7a5149f76
Reviewed-on: https://review.whamcloud.com/34700
Reviewed-by: Patrick Farrell <pfarrell@whamcloud.com>
Tested-by: Jenkins
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>

diff --git a/lustre/mdt/mdt_io.c b/lustre/mdt/mdt_io.c
index 3999e26..633c36d 100644 (file)
--- a/lustre/mdt/mdt_io.c
+++ b/lustre/mdt/mdt_io.c
@@ -1396,7 +1396,6 @@ int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
        struct niobuf_remote *rnb = NULL;
        struct niobuf_local *lnb;
        int rc;
-       int max_reply_len;
        loff_t offset;
        unsigned int len, copied = 0;
        int lnbs, nr_local, i;
@@ -1433,20 +1432,13 @@ int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
        if (mbo->mbo_dom_size == 0)
                RETURN(0);
 
-       /* check the maximum size available in reply */
-       max_reply_len =
-               req->rq_rqbd->rqbd_svcpt->scp_service->srv_max_reply_size;
-
-       CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d/%d\n",
-              mbo->mbo_dom_size, max_reply_len, req->rq_reqmsg->lm_repsize,
-              req->rq_replen);
+       CDEBUG(D_INFO, "File size %llu, reply sizes %d/%d\n",
+              mbo->mbo_dom_size, req->rq_reqmsg->lm_repsize, req->rq_replen);
        len = req->rq_reqmsg->lm_repsize - req->rq_replen;
-       max_reply_len -= req->rq_replen;
 
        /* NB: at this moment we have the following sizes:
         * - req->rq_replen: used data in reply
         * - req->rq_reqmsg->lm_repsize: total allocated reply buffer at client
-        * - max_reply_len: maximum reply size allowed by protocol
         *
         * Ideal case when file size fits in allocated reply buffer,
         * that mean we can return whole data in reply. We can also fit more
@@ -1458,20 +1450,12 @@ int mdt_dom_read_on_open(struct mdt_thread_info *mti, struct mdt_device *mdt,
         *
         * At the moment the following strategy is used:
         * 1) try to fit into the buffer we have
-        * 2) respond with bigger buffer so client will re-allocate it and
-        *    resend (up to srv_max_reply_size value).
-        * 3) return just file tail otherwise.
+        * 2) return just file tail otherwise.
         */
        if (mbo->mbo_dom_size <= len) {
                /* can fit whole data */
                len = mbo->mbo_dom_size;
                offset = 0;
-       } else if (mbo->mbo_dom_size <= max_reply_len) {
-               /* It is worth to make this tunable ON/OFF because this will
-                * cause buffer re-allocation and resend
-                */
-               len = mbo->mbo_dom_size;
-               offset = 0;
        } else {
                int tail, pgbits;
 

diff --git a/lustre/tests/sanity.sh b/lustre/tests/sanity.sh
index 4f302fe..0c9fc6a 100755 (executable)
--- a/lustre/tests/sanity.sh
+++ b/lustre/tests/sanity.sh
@@ -17554,61 +17554,6 @@ test_271d() {
 }
 run_test 271d "DoM: read on open (1K file in reply buffer)"
 
-test_271e() {
-       [ $MDS1_VERSION -lt $(version_code 2.10.57) ] &&
-               skip "Need MDS version at least 2.10.57"
-
-       local dom=$DIR/$tdir/dom
-       local tmp=$TMP/${tfile}.data
-       trap "cleanup_271def_tests $tmp" EXIT
-
-       mkdir -p $DIR/$tdir
-
-       $LFS setstripe -E 1024K -L mdt $DIR/$tdir
-
-       local mdtidx=$($LFS getstripe --mdt-index $DIR/$tdir)
-
-       cancel_lru_locks mdc
-       dd if=/dev/urandom of=$tmp bs=30K count=1
-       dd if=$tmp of=$dom bs=30K count=1
-       cancel_lru_locks mdc
-       cat /etc/hosts >> $tmp
-       lctl set_param -n mdc.*.stats=clear
-
-       echo "Append to the same page"
-       cat /etc/hosts >> $dom
-
-       local num=$(get_mdc_stats $mdtidx ost_read)
-       local ra=$(get_mdc_stats $mdtidx req_active)
-       local rw=$(get_mdc_stats $mdtidx req_waittime)
-
-       [ -z $num ] || error "$num READ RPC occured"
-       # Reply buffer can be adjusted for larger buffer by resend
-       echo "... DONE with $((ra - rw)) resends"
-
-       # compare content
-       cmp $tmp $dom || error "file miscompare"
-
-       cancel_lru_locks mdc
-       lctl set_param -n mdc.*.stats=clear
-
-       echo "Open and read file"
-       cat $dom > /dev/null
-       local num=$(get_mdc_stats $mdtidx ost_read)
-       local ra=$(get_mdc_stats $mdtidx req_active)
-       local rw=$(get_mdc_stats $mdtidx req_waittime)
-
-       [ -z $num ] || error "$num READ RPC occured"
-       # Reply buffer can be adjusted for larger buffer by resend
-       echo "... DONE with $((ra - rw)) resends"
-
-       # compare content
-       cmp $tmp $dom || error "file miscompare"
-
-       return 0
-}
-run_test 271e "DoM: read on open (30K file with reply buffer adjusting)"
-
 test_271f() {
        [ $MDS1_VERSION -lt $(version_code 2.10.57) ] &&
                skip "Need MDS version at least 2.10.57"