Skip to contentSkip to conversation

lustre: OpenSFS updates April 18, 2022

PATCH New James Simmons 1 participant 8 messages Apr 19, 2022 Networking

lore
Share & tools
lore
#0 lore raw

cover lustre: OpenSFS updates April 18, 2022

Backport patches from the OpenSFS tree to the native Linux client as of April 18, 2022.

Alexander Zarochentsev (1):
lustre: ptlrpc: unregister reply buffer on rq_err

Andreas Dilger (1):
lustre: lmv: change default hash back to fnv_1a_64

Chris Horn (2):
lnet: only update gateway NI status on discovery
lnet: ln_api_mutex deadlocks

John L. Hammond (1):
lustre: lov: remove lo_trunc_stripeno

Patrick Farrell (1):
lustre: llite: Fix use of uninitialized fields

Shaun Tancheff (1):
lustre: clio: Disable lockless for DIO with O_APPEND

 fs/lustre/llite/file.c                  |  2 ++
 fs/lustre/llite/rw.c                    | 18 ++++++++++-----
 fs/lustre/lov/lov_cl_internal.h         | 17 ++++++--------
 fs/lustre/lov/lov_io.c                  | 40 +++++++++++++++++++++++----------
 fs/lustre/lov/lov_lock.c                | 35 ++++++++++++++++-------------
 fs/lustre/lov/lov_object.c              |  2 --
 fs/lustre/ptlrpc/client.c               |  5 +++++
 include/uapi/linux/lustre/lustre_user.h |  2 +-
 net/lnet/lnet/api-ni.c                  | 40 ++++++++++++++++++++++++++++-----
 net/lnet/lnet/config.c                  |  2 +-
 net/lnet/lnet/lib-move.c                | 18 +++++++++++----
 net/lnet/lnet/peer.c                    | 11 ++++++---
 net/lnet/lnet/router.c                  |  2 +-
 net/lnet/lnet/router_proc.c             |  2 +-
 14 files changed, 135 insertions(+), 61 deletions(-)
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#1 lore raw

PATCH 1/7 lustre: ptlrpc: unregister reply buffer on rq_err

From: Alexander Zarochentsev <[email protected]>

Unregister reply buffer on rq_err and prevent a late reply from modifying request flags in INTERPRET state.

HPE-bug-id: LUS-10717

Fixes b06c1d17e488 ("lustre: mgc: do not ignore target registration failure")

WC-bug-id: https://jira.whamcloud.com/browse/LU-15435
Lustre-commit: d8012811cc6ff9c7f ("LU-15435 ptlrpc: unregister reply buffer on rq_err")

Reviewed-on: https://review.whamcloud.com/46132

---
 fs/lustre/ptlrpc/client.c | 5 +++++
 1 file changed, 5 insertions(+)
fs/lustre/ptlrpc/client.c +5 -0
index ec0cd5f..685d6e2 100644
@@ -1857,6 +1857,11 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
 		}
 
 		if (req->rq_err) {
+			if (!ptlrpc_unregister_reply(req, 1)) {
+				ptlrpc_unregister_bulk(req, 1);
+				continue;
+			}
+
 			spin_lock(&req->rq_lock);
 			req->rq_replied = 0;
 			spin_unlock(&req->rq_lock);
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#2 lore raw

PATCH 2/7 lustre: llite: Fix use of uninitialized fields

In reply to James Simmons

From: Patrick Farrell <[email protected]>

We use data from ci_rw to set io_start_index and io_end_index, which is a problem for mmap because mmap does not use ci_rw.

When ci_rand_read is set or readahead is disabled, we use these values to decide how much data to read.

ci_rw is uninitialized, and if the values are non-zero, we may try to read data beyond the locks we took for our I/O.

If there is no lock (either because there was never one or it was cancelled), this results in an LBUG in osc_req_attr_set when it verifies the pages are covered by a lock.

WC-bug-id: https://jira.whamcloud.com/browse/LU-15637
Lustre-commit: 9884f37985c1108fb ("LU-15637 llite: Fix use of uninitialized fields")

Reviewed-on: https://review.whamcloud.com/46776

---
 fs/lustre/llite/rw.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)
fs/lustre/llite/rw.c +13 -5
index b8cffde..0ddd920 100644
@@ -1627,6 +1627,8 @@ int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
 	struct ll_readahead_state *ras = NULL;
 	struct cl_2queue *queue = &io->ci_queue;
 	struct ll_sb_info *sbi = ll_i2sbi(inode);
+	struct vvp_io *vio = vvp_env_io(env);
+	bool mmap = !vio->vui_ra_valid;
 	struct cl_sync_io *anchor = NULL;
 	pgoff_t ra_start_index = 0;
 	pgoff_t io_start_index;
@@ -1644,12 +1646,11 @@ int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
 	uptodate = vpg->vpg_defer_uptodate;
 
 	if (ll_readahead_enabled(sbi) && !vpg->vpg_ra_updated && ras) {
-		struct vvp_io *vio = vvp_env_io(env);
 		enum ras_update_flags flags = 0;
 
 		if (uptodate)
 			flags |= LL_RAS_HIT;
-		if (!vio->vui_ra_valid)
+		if (mmap)
 			flags |= LL_RAS_MMAP;
 		ras_update(sbi, inode, ras, vvp_index(vpg), flags, io);
 	}
@@ -1667,9 +1668,16 @@ int ll_io_read_page(const struct lu_env *env, struct cl_io *io,
 		cl_page_list_add(&queue->c2_qin, page, true);
 	}
 
-	io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
-	io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
-				io->u.ci_rw.crw_count - 1);
+	/* mmap does not set the ci_rw fields */
+	if (!mmap) {
+		io_start_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos);
+		io_end_index = cl_index(io->ci_obj, io->u.ci_rw.crw_pos +
+					io->u.ci_rw.crw_count - 1);
+	} else {
+		io_start_index = vvp_index(vpg);
+		io_end_index = vvp_index(vpg);
+	}
+
 	if (ll_readahead_enabled(sbi) && ras && !io->ci_rand_read) {
 		pgoff_t skip_index = 0;
 
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#3 lore raw

PATCH 3/7 lustre: lov: remove lo_trunc_stripeno

In reply to James Simmons

From: "John L. Hammond" <[email protected]>

Remove the lo_trunc_stripeno member of struct lov_layout_raid0 and add an lis_trunc_stripe_index array to struct lov_io. This makes the truncate stripe index information belong to the IO and not to the concurrently accessed object. This is needed because we do not have locking that protects it from its initialization in lov_io_iter_init() to its use in lov_lock_sub_init(). Also remove the unused lo_write_lock member of struct lov_object.

Fixes d83ed47d35 ("lustre: lov: correctly set OST obj size")

WC-bug-id: https://jira.whamcloud.com/browse/LU-15702
Lustre-commit: 42a6d1fdb6818f1b3 ("LU-15702 lov: remove lo_trunc_stripeno")

Reviewed-on: https://review.whamcloud.com/46940

---
 fs/lustre/lov/lov_cl_internal.h | 17 +++++++----------
 fs/lustre/lov/lov_io.c          | 40 ++++++++++++++++++++++++++++------------
 fs/lustre/lov/lov_lock.c        | 35 +++++++++++++++++++----------------
 fs/lustre/lov/lov_object.c      |  2 --
 4 files changed, 54 insertions(+), 40 deletions(-)
fs/lustre/lov/lov_cl_internal.h +7 -10
index 42fd10a..6b96543 100644
@@ -175,11 +175,6 @@ struct lov_comp_layout_entry_ops {
 struct lov_layout_raid0 {
 	unsigned int		lo_nr;
 	/**
-	 * record the stripe no before the truncate size, used for setting OST
-	 * object size for truncate. LU-14128.
-	 */
-	int			lo_trunc_stripeno;
-	/**
 	 * When this is true, lov_object::lo_attr contains
 	 * valid up to date attributes for a top-level
 	 * object. This field is reset to 0 when attributes of
@@ -325,11 +320,6 @@ struct lov_object {
 			 */
 			int		lo_preferred_mirror;
 			/**
-			 * For FLR: the lock to protect access to
-			 * lo_preferred_mirror.
-			 */
-			spinlock_t	lo_write_lock;
-			/**
 			 * For FLR: Number of (valid) mirrors.
 			 */
 			unsigned int	lo_mirror_count;
@@ -562,6 +552,13 @@ struct lov_io {
 	loff_t			lis_io_endpos;
 
 	/**
+	 * Record the stripe index before the truncate size, used for setting
+	 * OST object size for truncate. LU-14128. lis_trunc_stripe_index[i]
+	 * refers to lov_object.u.composite.lo_entries[i].
+	 */
+	int *lis_trunc_stripe_index;
+
+	/**
 	 * starting position within a file, for the current io loop iteration
 	 * (stripe), used by ci_io_loop().
 	 */
fs/lustre/lov/lov_io.c +28 -12
index 904bafd..38dacd35 100644
@@ -782,6 +782,7 @@ static int lov_io_iter_init(const struct lu_env *env,
 {
 	struct lov_io *lio = cl2lov_io(env, ios);
 	struct lov_stripe_md *lsm = lio->lis_object->lo_lsm;
+	bool is_trunc = cl_io_is_trunc(ios->cis_io);
 	struct lov_io_sub *sub;
 	struct lu_extent ext;
 	int rc = 0;
@@ -790,6 +791,16 @@ static int lov_io_iter_init(const struct lu_env *env,
 	ext.e_start = lio->lis_pos;
 	ext.e_end = lio->lis_endpos;
 
+	if (is_trunc) {
+		int count = lio->lis_object->u.composite.lo_entry_count;
+
+		lio->lis_trunc_stripe_index = kcalloc(count,
+						      sizeof(lio->lis_trunc_stripe_index[0]),
+						      GFP_NOFS);
+		if (!lio->lis_trunc_stripe_index)
+			return -ENOMEM;
+	}
+
 	lov_foreach_io_layout(index, lio, &ext) {
 		struct lov_layout_entry *le = lov_entry(lio->lis_object, index);
 		struct lov_layout_raid0 *r0 = &le->lle_raid0;
@@ -798,7 +809,8 @@ static int lov_io_iter_init(const struct lu_env *env,
 		u64 start;
 		u64 end;
 
-		r0->lo_trunc_stripeno = -1;
+		if (is_trunc)
+			lio->lis_trunc_stripe_index[index] = -1;
 
 		CDEBUG(D_VFSTRACE, "component[%d] flags %#x\n",
 		       index, lsm->lsm_entries[index]->lsme_flags);
@@ -832,8 +844,7 @@ static int lov_io_iter_init(const struct lu_env *env,
 				continue;
 			}
 
-			if (cl_io_is_trunc(ios->cis_io) &&
-			    !tested_trunc_stripe) {
+			if (is_trunc && !tested_trunc_stripe) {
 				int prev;
 				u64 tr_start;
 
@@ -848,20 +859,22 @@ static int lov_io_iter_init(const struct lu_env *env,
 				if (ext.e_start <
 				    lsm->lsm_entries[index]->lsme_extent.e_start) {
 					/* need previous stripe involvement */
-					r0->lo_trunc_stripeno = prev;
+					lio->lis_trunc_stripe_index[index] = prev;
 				} else {
 					tr_start = ext.e_start;
 					tr_start = lov_do_div64(tr_start,
 								stripe_width(lsm, index));
 					/* tr_start %= stripe_swidth */
-					if (tr_start == stripe * lsm->lsm_entries[index]->lsme_stripe_size)
-						r0->lo_trunc_stripeno = prev;
+					if (tr_start ==
+					    stripe * lsm->lsm_entries[index]->lsme_stripe_size)
+						lio->lis_trunc_stripe_index[index] = prev;
 				}
 			}
 
 			/* if the last stripe is the trunc stripeno */
-			if (r0->lo_trunc_stripeno == stripe)
-				r0->lo_trunc_stripeno = -1;
+			if (is_trunc &&
+			    lio->lis_trunc_stripe_index[index] == stripe)
+				lio->lis_trunc_stripe_index[index] = -1;
 
 			sub = lov_sub_get(env, lio,
 					  lov_comp_index(index, stripe));
@@ -875,10 +888,10 @@ static int lov_io_iter_init(const struct lu_env *env,
 		if (rc != 0)
 			break;
 
-		if (r0->lo_trunc_stripeno != -1) {
-			stripe = r0->lo_trunc_stripeno;
+		if (is_trunc && lio->lis_trunc_stripe_index[index] != -1) {
+			stripe = lio->lis_trunc_stripe_index[index];
 			if (unlikely(!r0->lo_sub[stripe])) {
-				r0->lo_trunc_stripeno = -1;
+				lio->lis_trunc_stripe_index[index] = -1;
 				continue;
 			}
 			sub = lov_sub_get(env, lio,
@@ -892,7 +905,7 @@ static int lov_io_iter_init(const struct lu_env *env,
 			 * read get wrong kms.
 			 */
 			if (!list_empty(&sub->sub_linkage)) {
-				r0->lo_trunc_stripeno = -1;
+				lio->lis_trunc_stripe_index[index] = -1;
 				continue;
 			}
 
@@ -1091,6 +1104,9 @@ static void lov_io_iter_fini(const struct lu_env *env,
 	struct lov_io *lio = cl2lov_io(env, ios);
 	int rc;
 
+	kfree(lio->lis_trunc_stripe_index);
+	lio->lis_trunc_stripe_index = NULL;
+
 	rc = lov_io_call(env, lio, lov_io_iter_fini_wrapper);
 	LASSERT(rc == 0);
 	while (!list_empty(&lio->lis_active))
fs/lustre/lov/lov_lock.c +19 -16
index d137614..313c09a 100644
@@ -115,6 +115,8 @@ static struct lov_lock *lov_lock_sub_init(const struct lu_env *env,
 					  struct cl_lock *lock)
 {
 	struct lov_object *lov = cl2lov(obj);
+	struct lov_io *lio = lov_env_io(env);
+	bool is_trunc = cl_io_is_trunc(io);
 	struct lov_lock *lovlck;
 	struct lu_extent ext;
 	int result = 0;
@@ -124,6 +126,8 @@ static struct lov_lock *lov_lock_sub_init(const struct lu_env *env,
 	u64 start;
 	u64 end;
 
+	LASSERT(ergo(is_trunc, lio->lis_trunc_stripe_index != NULL));
+
 	ext.e_start = cl_offset(obj, lock->cll_descr.cld_start);
 	if (lock->cll_descr.cld_end == CL_PAGE_EOF)
 		ext.e_end = OBD_OBJECT_EOF;
@@ -131,16 +135,16 @@ static struct lov_lock *lov_lock_sub_init(const struct lu_env *env,
 		ext.e_end = cl_offset(obj, lock->cll_descr.cld_end + 1);
 
 	nr = 0;
-	lov_foreach_io_layout(index, lov_env_io(env), &ext) {
+	lov_foreach_io_layout(index, lio, &ext) {
 		struct lov_layout_raid0 *r0 = lov_r0(lov, index);
 
 		for (i = 0; i < r0->lo_nr; i++) {
 			if (likely(r0->lo_sub[i])) { /* spare layout */
-				if (lov_stripe_intersects(lov->lo_lsm, index, i,
-							  &ext, &start, &end))
-					nr++;
-				else if (cl_io_is_trunc(io) &&
-					 r0->lo_trunc_stripeno == i)
+				if (lov_stripe_intersects(lov->lo_lsm, index,
+							  i, &ext, &start,
+							  &end) ||
+				    (is_trunc &&
+				     i == lio->lis_trunc_stripe_index[index]))
 					nr++;
 			}
 		}
@@ -162,24 +166,23 @@ static struct lov_lock *lov_lock_sub_init(const struct lu_env *env,
 		struct lov_layout_raid0 *r0 = lov_r0(lov, index);
 
 		for (i = 0; i < r0->lo_nr; ++i) {
-			struct lov_lock_sub *lls = &lovlck->lls_sub[nr];
-			struct cl_lock_descr *descr = &lls->sub_lock.cll_descr;
-			bool intersect = false;
+			struct lov_lock_sub *lls;
+			struct cl_lock_descr *descr;
 
 			if (unlikely(!r0->lo_sub[i]))
 				continue;
 
-			intersect = lov_stripe_intersects(lov->lo_lsm, index, i,
-							  &ext, &start, &end);
-			if (intersect)
-				goto init_sublock;
-
-			if (cl_io_is_trunc(io) && i == r0->lo_trunc_stripeno)
+			if (lov_stripe_intersects(lov->lo_lsm, index, i, &ext,
+						  &start, &end) ||
+			    (is_trunc &&
+			     i == lio->lis_trunc_stripe_index[index]))
 				goto init_sublock;
 
 			continue;
-
 init_sublock:
+			LASSERT(nr < lovlck->lls_nr);
+			lls = &lovlck->lls_sub[nr];
+			descr = &lls->sub_lock.cll_descr;
 			LASSERT(!descr->cld_obj);
 			descr->cld_obj = lovsub2cl(r0->lo_sub[i]);
 			descr->cld_start = cl_index(descr->cld_obj, start);
fs/lustre/lov/lov_object.c +0 -2
index ff0f7fa..d9eaf15 100644
@@ -214,7 +214,6 @@ static int lov_init_raid0(const struct lu_env *env, struct lov_device *dev,
 
 	spin_lock_init(&r0->lo_sub_lock);
 	r0->lo_nr = lse->lsme_stripe_count;
-	r0->lo_trunc_stripeno = -1;
 
 	flags = memalloc_nofs_save();
 	r0->lo_sub = kvmalloc_array(r0->lo_nr, sizeof(r0->lo_sub[0]),
@@ -641,7 +640,6 @@ static int lov_init_composite(const struct lu_env *env, struct lov_device *dev,
 
 	entry_count = lsm->lsm_entry_count;
 
-	spin_lock_init(&comp->lo_write_lock);
 	comp->lo_flags = lsm->lsm_flags;
 	comp->lo_mirror_count = lsm->lsm_mirror_count + 1;
 	comp->lo_entry_count = lsm->lsm_entry_count;
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#4 lore raw

PATCH 4/7 lustre: lmv: change default hash back to fnv_1a_64

In reply to James Simmons

From: Andreas Dilger <[email protected]>

Until performance issue is resolved, change the default directory hash type from 'crush' back to 'fnv_1a_64'.

Fixes 92fb134e43a0 ("lustre: lmv: change default hash type to crush")

WC-bug-id: https://jira.whamcloud.com/browse/LU-15692
Lustre-commit: 0090b6f6f6cfd65fc ("LU-15692 lmv: change default hash back to fnv_1a_64")

Reviewed-on: https://review.whamcloud.com/46950

---
 include/uapi/linux/lustre/lustre_user.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
include/uapi/linux/lustre/lustre_user.h +1 -1
index 3017148..fa01c28 100644
@@ -707,7 +707,7 @@ static __attribute__((unused)) const char *mdt_hash_name[] = {
 	"crush",
 };
 
-#define LMV_HASH_TYPE_DEFAULT LMV_HASH_TYPE_CRUSH
+#define LMV_HASH_TYPE_DEFAULT LMV_HASH_TYPE_FNV_1A_64
 
 /* Right now only the lower part(0-16bits) of lmv_hash_type is being used,
  * and the higher part will be the flag to indicate the status of object,
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#5 lore raw

PATCH 5/7 lnet: only update gateway NI status on discovery

In reply to James Simmons

From: Chris Horn <[email protected]>

Move the NI status from DOWN to UP only when receiving a discovery PING. The discovery PING should be the only message which should update the NI status since it's used as the gateway NI keep alive mechanism.

This is done to avoid the following scenario:

The gateway itself can push its updates to the peers which have removed it from its routing table. The peers would respond to the PUSH with an ACK, the ACK will bring the gateway's NI status to up. Therefore other peers which have avoid_asym_router_failure=1 will have their route status remain up even though the symmetrical route is gone.

Note: there is no way for the gateway to differentiate between a keep alive discovery and a manually triggered discovery or ping. However, this a narrow case which will not be handled.

net_last_alive converted to use ktime_get_seconds() instead of ktime_get_real_seconds() since the NTP adjustment is not needed.

WC-bug-id: https://jira.whamcloud.com/browse/LU-13714
Lustre-commit: 3e3f70eb1ec95f32d ("LU-13714 lnet: only update gateway NI status on discovery")

Reviewed-on: https://review.whamcloud.com/39176

---
 net/lnet/lnet/config.c      |  2 +-
 net/lnet/lnet/lib-move.c    | 16 ++++++++++++----
 net/lnet/lnet/router.c      |  2 +-
 net/lnet/lnet/router_proc.c |  2 +-
 4 files changed, 15 insertions(+), 7 deletions(-)
net/lnet/lnet/config.c +1 -1
index f499c91..da3d20e 100644
@@ -350,7 +350,7 @@ struct lnet_net *
 	spin_lock_init(&net->net_lock);
 
 	net->net_id = net_id;
-	net->net_last_alive = ktime_get_real_seconds();
+	net->net_last_alive = ktime_get_seconds();
 
 	net->net_sel_priority = LNET_MAX_SELECTION_PRIORITY;
 
net/lnet/lnet/lib-move.c +12 -4
index 3ad13d0..0b3986e 100644
@@ -4250,6 +4250,7 @@ void lnet_monitor_thr_stop(void)
 	u32 type;
 	int rc = 0;
 	int cpt;
+	time64_t now = ktime_get_seconds();
 
 	LASSERT(!in_interrupt());
 
@@ -4301,11 +4302,18 @@ void lnet_monitor_thr_stop(void)
 		return -EPROTO;
 	}
 
-	if (the_lnet.ln_routing &&
-	    ni->ni_net->net_last_alive != ktime_get_real_seconds()) {
+	/* Only update net_last_alive for incoming GETs on the reserved portal
+	 * (i.e. incoming lnet/discovery pings).
+	 * This avoids situations where the router's own traffic results in NI
+	 * status changes
+	 */
+	if (the_lnet.ln_routing && type == LNET_MSG_GET &&
+	    hdr->msg.get.ptl_index == LNET_RESERVED_PORTAL &&
+	    !lnet_islocalnid(&src_nid) &&
+	    ni->ni_net->net_last_alive != now) {
 		lnet_ni_lock(ni);
 		spin_lock(&ni->ni_net->net_lock);
-		ni->ni_net->net_last_alive = ktime_get_real_seconds();
+		ni->ni_net->net_last_alive = now;
 		spin_unlock(&ni->ni_net->net_lock);
 		push = lnet_ni_set_status_locked(ni, LNET_NI_STATUS_UP);
 		lnet_ni_unlock(ni);
@@ -4480,7 +4488,7 @@ void lnet_monitor_thr_stop(void)
 		}
 	}
 
-	lpni->lpni_last_alive = ktime_get_seconds();
+	lpni->lpni_last_alive = now;
 
 	msg->msg_rxpeer = lpni;
 	msg->msg_rxni = ni;
net/lnet/lnet/router.c +1 -1
index beded3e..60ae15d 100644
@@ -1044,7 +1044,7 @@ int lnet_get_rtr_pool_cfg(int cpt, struct lnet_ioctl_pool_cfg *pool_cfg)
 
 	timeout = router_ping_timeout + alive_router_check_interval;
 
-	now = ktime_get_real_seconds();
+	now = ktime_get_seconds();
 	list_for_each_entry(net, &the_lnet.ln_nets, net_list) {
 		if (net->net_lnd->lnd_type == LOLND)
 			continue;
net/lnet/lnet/router_proc.c +1 -1
index a53d6fa..f231da1 100644
@@ -663,7 +663,7 @@ static int proc_lnet_nis(struct ctl_table *table, int write,
 		if (ni) {
 			struct lnet_tx_queue *tq;
 			char *stat;
-			time64_t now = ktime_get_real_seconds();
+			time64_t now = ktime_get_seconds();
 			time64_t last_alive = -1;
 			int i;
 			int j;
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#6 lore raw

PATCH 6/7 lnet: ln_api_mutex deadlocks

In reply to James Simmons

From: Chris Horn <[email protected]>

LNetNIFini() acquires the ln_api_mutex and holds onto it throughout various shutdown routines. Meanwhile, LND threads (via lnet_nid2peerni_locked()) or the discovery thread (via lnet_peer_data_present()) may need to acquire this mutex in order to progress.

Address these potential deadlocks by setting the_lnet.ln_state to LNET_STATE_STOPPING earlier in LNetNIFini(), and release the mutex prior to any call into LND module or before any wait.

LNetNIInit() is modified to return -ESHUTDOWN if it finds that there is a concurrent shutdown in progress.

HPE-bug-id: LUS-10681
WC-bug-id: https://jira.whamcloud.com/browse/LU-15616
Lustre-commit: 22de0bd145b649768 ("LU-15616 lnet: ln_api_mutex deadlocks")

Reviewed-on: https://review.whamcloud.com/46727

---
 net/lnet/lnet/api-ni.c   | 40 +++++++++++++++++++++++++++++++++++-----
 net/lnet/lnet/lib-move.c |  2 ++
 net/lnet/lnet/peer.c     | 11 ++++++++---
 3 files changed, 45 insertions(+), 8 deletions(-)
net/lnet/lnet/api-ni.c +35 -5
index 1978905..44d5014 100644
@@ -2244,14 +2244,16 @@ static void lnet_push_target_fini(void)
 		islo = ni->ni_net->net_lnd->lnd_type == LOLND;
 
 		LASSERT(!in_interrupt());
-		/* Holding the mutex makes it safe for lnd_shutdown
+		/* Holding the LND mutex makes it safe for lnd_shutdown
 		 * to call module_put(). Module unload cannot finish
 		 * until lnet_unregister_lnd() completes, and that
-		 * requires the mutex.
+		 * requires the LND mutex.
 		 */
+		mutex_unlock(&the_lnet.ln_api_mutex);
 		mutex_lock(&the_lnet.ln_lnd_mutex);
 		net->net_lnd->lnd_shutdown(ni);
 		mutex_unlock(&the_lnet.ln_lnd_mutex);
+		mutex_lock(&the_lnet.ln_api_mutex);
 
 		if (!islo)
 			CDEBUG(D_LNI, "Removed LNI %s\n",
@@ -2323,7 +2325,8 @@ static void lnet_push_target_fini(void)
 	/* NB called holding the global mutex */
 
 	/* All quiet on the API front */
-	LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING);
+	LASSERT(the_lnet.ln_state == LNET_STATE_RUNNING ||
+		the_lnet.ln_state == LNET_STATE_STOPPING);
 	LASSERT(!the_lnet.ln_refcount);
 
 	lnet_net_lock(LNET_LOCK_EX);
@@ -2823,6 +2826,11 @@ void lnet_lib_exit(void)
 
 	CDEBUG(D_OTHER, "refs %d\n", the_lnet.ln_refcount);
 
+	if (the_lnet.ln_state == LNET_STATE_STOPPING) {
+		mutex_unlock(&the_lnet.ln_api_mutex);
+		return -ESHUTDOWN;
+	}
+
 	if (the_lnet.ln_refcount > 0) {
 		rc = the_lnet.ln_refcount++;
 		mutex_unlock(&the_lnet.ln_api_mutex);
@@ -2968,6 +2976,10 @@ void lnet_lib_exit(void)
 	} else {
 		LASSERT(!the_lnet.ln_niinit_self);
 
+		lnet_net_lock(LNET_LOCK_EX);
+		the_lnet.ln_state = LNET_STATE_STOPPING;
+		lnet_net_unlock(LNET_LOCK_EX);
+
 		lnet_fault_fini();
 		lnet_router_debugfs_fini();
 		lnet_monitor_thr_stop();
@@ -3433,6 +3445,10 @@ static int lnet_handle_legacy_ip2nets(char *ip2nets,
 	lnet_set_tune_defaults(tun);
 
 	mutex_lock(&the_lnet.ln_api_mutex);
+	if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+		rc = -ESHUTDOWN;
+		goto out;
+	}
 	while ((net = list_first_entry_or_null(&net_head,
 					       struct lnet_net,
 					       net_list)) != NULL) {
@@ -3498,8 +3514,10 @@ int lnet_dyn_add_ni(struct lnet_ioctl_config_ni *conf)
 	lnet_set_tune_defaults(tun);
 
 	mutex_lock(&the_lnet.ln_api_mutex);
-
-	rc = lnet_add_net_common(net, tun);
+	if (the_lnet.ln_state != LNET_STATE_RUNNING)
+		rc = -ESHUTDOWN;
+	else
+		rc = lnet_add_net_common(net, tun);
 
 	mutex_unlock(&the_lnet.ln_api_mutex);
 
@@ -3522,6 +3540,10 @@ int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
 		return -EINVAL;
 
 	mutex_lock(&the_lnet.ln_api_mutex);
+	if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+		rc = -ESHUTDOWN;
+		goto unlock_api_mutex;
+	}
 
 	lnet_net_lock(0);
 
@@ -3615,6 +3637,10 @@ int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
 		return rc == 0 ? -EINVAL : rc;
 
 	mutex_lock(&the_lnet.ln_api_mutex);
+	if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+		rc = -ESHUTDOWN;
+		goto out_unlock_clean;
+	}
 
 	if (rc > 1) {
 		rc = -EINVAL; /* only add one network per call */
@@ -3668,6 +3694,10 @@ int lnet_dyn_del_ni(struct lnet_ioctl_config_ni *conf)
 		return -EINVAL;
 
 	mutex_lock(&the_lnet.ln_api_mutex);
+	if (the_lnet.ln_state != LNET_STATE_RUNNING) {
+		rc = -ESHUTDOWN;
+		goto out;
+	}
 
 	lnet_net_lock(0);
 
net/lnet/lnet/lib-move.c +2 -0
index 0b3986e..0496bf5 100644
@@ -3872,7 +3872,9 @@ void lnet_monitor_thr_stop(void)
 	complete(&the_lnet.ln_mt_wait_complete);
 
 	/* block until monitor thread signals that it's done */
+	mutex_unlock(&the_lnet.ln_api_mutex);
 	wait_for_completion(&the_lnet.ln_mt_signal);
+	mutex_lock(&the_lnet.ln_api_mutex);
 	LASSERT(the_lnet.ln_mt_state == LNET_MT_STATE_SHUTDOWN);
 
 	/* perform cleanup tasks */
net/lnet/lnet/peer.c +8 -3
index 98f71dd..714326a 100644
@@ -3244,12 +3244,15 @@ static int lnet_peer_deletion(struct lnet_peer *lp)
 	if (lp->lp_state & LNET_PEER_MARK_DELETED)
 		return 0;
 
-	if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
-		return -ESHUTDOWN;
-
 	spin_unlock(&lp->lp_lock);
 
 	mutex_lock(&the_lnet.ln_api_mutex);
+	if (the_lnet.ln_state != LNET_STATE_RUNNING ||
+	    the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING) {
+		mutex_unlock(&the_lnet.ln_api_mutex);
+		spin_lock(&lp->lp_lock);
+		return -ESHUTDOWN;
+	}
 
 	lnet_net_lock(LNET_LOCK_EX);
 	/* remove the peer from the discovery work
@@ -3929,8 +3932,10 @@ void lnet_peer_discovery_stop(void)
 	else
 		wake_up(&the_lnet.ln_dc_waitq);
 
+	mutex_unlock(&the_lnet.ln_api_mutex);
 	wait_event(the_lnet.ln_dc_waitq,
 		   the_lnet.ln_dc_state == LNET_DC_STATE_SHUTDOWN);
+	mutex_lock(&the_lnet.ln_api_mutex);
 
 	LASSERT(list_empty(&the_lnet.ln_dc_request));
 	LASSERT(list_empty(&the_lnet.ln_dc_working));
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org
#7 lore raw

PATCH 7/7 lustre: clio: Disable lockless for DIO with O_APPEND

In reply to James Simmons

From: Shaun Tancheff <[email protected]>

Lockless O_DIRECT with O_APPEND can allow interleaved / racy appends from concurrent I/O.

Disable lockless I/O when O_APPEND is set

HPE-bug-id: LUS-9776
WC-bug-id: https://jira.whamcloud.com/browse/LU-15670
Lustre-commit: 649d638467c037579 ("LU-15670 clio: Disable lockless for DIO with O_APPEND")

Reviewed-on: https://review.whamcloud.com/46890

---
 fs/lustre/llite/file.c | 2 ++
 1 file changed, 2 insertions(+)
fs/lustre/llite/file.c +2 -0
index 4855156..1ac3e4f 100644
@@ -1673,6 +1673,8 @@ static void ll_heat_add(struct inode *inode, enum cl_io_type iot,
 
 	io = vvp_env_thread_io(env);
 	if (file->f_flags & O_DIRECT) {
+		if (file->f_flags & O_APPEND)
+			dio_lock = 1;
 		if (!is_sync_kiocb(args->u.normal.via_iocb))
 			is_aio = true;
 
Signature · 6 lines
1.8.3.1

_______________________________________________
lustre-devel mailing list
[email protected]
http://lists.lustre.org/listinfo.cgi/lustre-devel-lustre.org