From: Thomas Bertschinger <bertschinger@lanl.gov>
Date: Tue, 25 Jul 2023 16:03:47 +0000 (-0400)
Subject: LU-16981 lod: update llc_stripe_count after ost inactive
X-Git-Tag: 2.15.59~135
X-Git-Url: https://git.whamcloud.com/?a=commitdiff_plain;h=78336aa166f4a7a0128a5891c747eecf26ff9565;p=fs%2Flustre-release.git

LU-16981 lod: update llc_stripe_count after ost inactive

If an OST gets deactivated while lod_ost_alloc_qos() is trying to
allocate stripes for a file create, then normally this is caught and
EAGAIN is returned which causes the lod_comp->llc_stripe_count to
get updated to accurately reflect the stripe count. But there is a
race condition and if the OST is deactivated after the call to
ltd_qos_is_usable() but before the stripes are allocated, then
updating the stripe count never occurred.

This causes an LBUG later in lod_striped_create() because fewer
stripes are allocated than the number in llc_stripe_count so it
finds a stripe that is NULL.

The solution is to properly update lod_comp->llc_stripe_count when
the number of stripes created is less than expected.

Fixes: ced540165ef5 ("LU-16623 lod: handle object allocation consistently")
Test-Parameters: testlist=sanity env=ONLY=27V,ONLY_REPEAT=100
Signed-off-by: Thomas Bertschinger <bertschinger@lanl.gov>
Change-Id: Ia1264f24904fed00454b3bc3c0d6c7b9b947737f
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/51759
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Sergey Cheremencev <scherementsev@ddn.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
Reviewed-by: Andreas Dilger <adilger@whamcloud.com>
---

diff --git a/lustre/lod/lod_qos.c b/lustre/lod/lod_qos.c
index 69190ba..2244cbb 100644
--- a/lustre/lod/lod_qos.c
+++ b/lustre/lod/lod_qos.c
@@ -1730,6 +1730,8 @@ static int lod_ost_alloc_qos(const struct lu_env *env, struct lod_object *lo,
 		set_bit(LQ_DIRTY, &lod->lod_ost_descs.ltd_qos.lq_flags);
 		clear_bit(LQ_SAME_SPACE, &lod->lod_ost_descs.ltd_qos.lq_flags);
 		rc = -EAGAIN;
+	} else if (nfound < lod_comp->llc_stripe_count) {
+		lod_comp->llc_stripe_count = nfound;
 	}
 
 	/* If there are enough OSTs, a component with overstriping requessted
diff --git a/lustre/tests/sanity.sh b/lustre/tests/sanity.sh
index 2623127..f3dc745 100755
--- a/lustre/tests/sanity.sh
+++ b/lustre/tests/sanity.sh
@@ -3505,6 +3505,45 @@ test_27U() {
 }
 run_test 27U "append pool and stripe count work with composite default layout"
 
+test_27V() {
+	[ $PARALLEL == "yes" ] && skip "skip parallel run"
+	(( $OSTCOUNT >= 4 )) || skip_env "needs >= 4 OSTs"
+
+	local dir=$DIR/$tdir
+	local osp_param=osp.$FSNAME-OST0000-osc-MDT0000.max_create_count
+	local lod_param=lod.$FSNAME-MDT0000-mdtlov.qos_threshold_rr
+	local saved_max=$(do_facet mds1 $LCTL get_param -n $osp_param)
+	local saved_qos=$(do_facet mds1 $LCTL get_param -n $lod_param)
+	local pid
+
+	stack_trap "do_facet mds1 $LCTL set_param $osp_param=$saved_max"
+
+	do_facet mds1 $LCTL set_param $lod_param=0
+	stack_trap "do_facet mds1 $LCTL set_param $lod_param=$saved_qos"
+
+	$LFS setdirstripe --mdt-count=1 --mdt-index=0 $dir
+	stack_trap "rm -rf $dir"
+
+	# exercise race in LU-16981 with deactivating OST while creating a file
+	(
+		while true; do
+			do_facet mds1 $LCTL set_param $osp_param=0 > /dev/null
+			sleep 0.1
+			do_facet mds1 \
+				$LCTL set_param $osp_param=$saved_max > /dev/null
+		done
+	) &
+
+	pid=$!
+	stack_trap "kill -9 $pid"
+
+	# errors here are OK so ignore them (just don't want to crash)
+	$LFS setstripe -c -1 $dir/f.{1..200} 2> /dev/null
+
+	return 0
+}
+run_test 27V "creating widely striped file races with deactivating OST"
+
 # createtest also checks that device nodes are created and
 # then visible correctly (#2091)
 test_28() { # bug 2091