tcwg_bmk-build.sh: Rework handling of base-artifacts/annex/bmk-data

To avoid committing unwanted files into git (e.g., raw benchmarking
data) we implement "annex" support.  Files are base-artifacts/annex
can be either symlinks to directories or regular files containing
an rsync-able url.

Here we convert directory symlinks into tarballs, upload to bkp-01
and replace symlinks with files pointing to their uploaded location.

In round-robin.sh:reset_artifacts() we do the opposite: download and
extract tarball into a temporary directory, and replace the file with
a symlink to that directory.

Change-Id: I78062a5147c944f18d6cd6eb5129f0d9fb74098c

1 files changed, 20 insertions, 42 deletions

diff --git a/tcwg_bmk-build.sh b/tcwg_bmk-build.sh
index f9bb375d..59ba6683 100755
--- a/tcwg_bmk-build.sh
+++ b/tcwg_bmk-build.sh
@@ -188,7 +188,6 @@ benchmark ()
     set -euf -o pipefail
 
     local bmk_flags="$2"
-    local results_id_file="$3"
 
     sanity_check_pwd
 
@@ -286,10 +285,10 @@ benchmark ()
 	*) echo "ERROR: Unknown hw_tag $hw_tag"; exit 1 ;;
     esac
 
-    # we create a tmpdir in /tmp
+    # Create directory for tcwg-benchmark to upload results to
     local results_dir
-    results_dir="$(mktemp -d -p /tmp)"
-    chmod g+wx $results_dir
+    results_dir="$(mktemp -d)"
+    chmod go+wx "$results_dir"
 
     # Trigger benchmarking job and capture its console output.
     # Ignore exit code of the trigger command to detect various failure
@@ -315,17 +314,6 @@ benchmark ()
 		${bmk_branch+-p bmk_branch="$bmk_branch"} \
 	| tee $run_step_artifacts/benchmark-build.log || true
 
-    # copy bmk results into artifacts/annex/bmk-data/
-    rm -rf ${rr[top_artifacts]}/annex/bmk-data/
-    mkdir -p ${rr[top_artifacts]}/annex/bmk-data/
-    rsync -rav $results_dir/ ${rr[top_artifacts]}/annex/bmk-data/
-
-    # make sure we are deleting a /tmp/xxx directory before forcing with sudo,
-    # and ignore error if any
-    if [[ "$results_dir" =~ ^/tmp/ ]]; then
-      sudo -n rm -rf "$results_dir" || true
-    fi
-
     local build_num
     build_num=$(head -n1 $run_step_artifacts/benchmark-build.log \
 		    | sed -e "s/Started.*#\([0-9]\+\).*/\1/")
@@ -370,8 +358,9 @@ benchmark ()
 	    | tee $run_step_artifacts/benchmark.log
     done
 
-    echo "$hw_tag/${rr[ci_project]}/${rr[ci_config]}-${rr[mode]}/$build_num" \
-			     > "$results_id_file"
+    rm -rf "${rr[top_artifacts]}/annex"
+    mkdir "${rr[top_artifacts]}/annex"
+    ln -s "$results_dir" "${rr[top_artifacts]}/annex/bmk-data"
 
     return $build_ret
     )
@@ -402,26 +391,19 @@ compare_results ()
 	sudo /usr/lib/linux-tools/install-armhf-perf-workaround.sh
     fi
 
-    local new_results ref_results
-    new_results="${rr[top_artifacts]}/annex/bmk-data"
-    assert_with_msg "Benchmarking succeeded, but annex/bmk-data results are missing" \
-		    [ -d $new_results ]
+    local new_results="${rr[top_artifacts]}/annex/bmk-data"
+    local ref_results="base-artifacts/annex/bmk-data"
 
-    if [ ! -f base-artifacts/results_id ] ; then
-	# base-artifacts has no reference results. It doesn't have any results_id.
-	# This can happen on initialization of base-artifacts (update_baseline=init)
-	# or in corner-cases like forced build failing during toolchain build.
+    assert_with_msg "Benchmarking succeeded, but no annex/bmk-data results" \
+		    [ -d "$new_results" ]
+
+    if ! [ -d "$ref_results" ]; then
+	# base-artifacts has no reference results.
+	# This can happen on init build (update_baseline=init).
 	# In such cases we compare results to themselves just as an exercise.
-	ref_results=$new_results
-    elif [ -d base-artifacts/annex/bmk-data ] ; then
-	# If bmk results in part of the baseline, use it.
-	# Usefule for bisect, during bisect, the first successful run, needs to be used as a reference.
-	# It updates the local baseline, but will not push the results to bkp-01.
-	# Use ref_results from baseline if exists.
-	ref_results="base-artifacts/annex/bmk-data"
-    else
-        # reference results to be retrieved before applying the comparison
-	ref_results="bkp-01.tcwglab:/home/tcwg-benchmark/results-$(cat base-artifacts/results_id)"
+	ref_results="$new_results"
+	assert_with_msg "No reference results" \
+			[ "${rr[update_baseline]}" = "init" ]
     fi
 
     # Compare vs previous run
@@ -510,7 +492,6 @@ compare_results ()
 }
 
 # Exit with code 0 if no regression compared to base-artifacts/.
-# Inspect build results ./results and performance results in ./results_id.
 no_regression_p ()
 {
     (
@@ -528,16 +509,13 @@ no_regression_p ()
         return 0
     fi
 
-    assert_with_msg "Benchmarking succeeded, but results_id is missing" \
-		    [ -f $run_step_top_artifacts/results_id ]
-
     # Make sure there is no stray results.regression file, which we use
     # as failure marker.
     assert ! [ -f $run_step_artifacts/results.regressions ]
 
     # At this stage,
-    # - new/results_id should exist (score>0)
-    # - ref/results_id might not exist, (ex: baseline score<0)
+    # - artifacts/annex/bmk-data should exist (score>0)
+    # - base-artifacts/annex/bmk-data should exist if update_baseline!=init
     # we call compare_results to generate the csv, and check metric regressions.
     compare_results "$metric_id"
 
@@ -586,7 +564,7 @@ case "${rr[toolchain]}" in
 	run_step skip_on_fail -3 build_bmk_llvm
 	;;
 esac
-run_step skip_on_fail 1 benchmark -- "$cflags" ${rr[top_artifacts]}/results_id
+run_step skip_on_fail 1 benchmark -- "$cflags"
 run_step reset_on_fail x check_regression
 
 trap "" EXIT