X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fofd%2Fofd_io.c;h=8b793629777dcf8ca2a3f88d727abb31ce0036d4;hp=bbaf0860f9d8880bbfc9ee1b1c34662a29b0427a;hb=a21ce928aa641cbfacb9dde74e4be1cd1b658fb0;hpb=e9389613eb29297f7b5a6b6cd896bafc7a5551f8

diff --git a/lustre/ofd/ofd_io.c b/lustre/ofd/ofd_io.c
index bbaf086..8b79362 100644
--- a/lustre/ofd/ofd_io.c
+++ b/lustre/ofd/ofd_io.c
@@ -23,11 +23,10 @@
  * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2012, 2016, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
  *
  * lustre/ofd/ofd_io.c
  *
@@ -42,6 +41,7 @@
 
 #include <linux/kthread.h>
 #include "ofd_internal.h"
+#include <lustre_nodemap.h>
 
 struct ofd_inconsistency_item {
 	struct list_head	 oii_list;
@@ -146,6 +146,17 @@ static void ofd_inconsistency_verify_one(const struct lu_env *env,
 	OBD_FREE_PTR(oii);
 }
 
+struct oivm_args {
+	struct ofd_device	*od_ofd;
+	struct lu_env		od_env;
+	struct lfsck_req_local	od_lrl;
+	struct completion	*od_started;
+};
+
+#ifndef TASK_IDLE
+#define TASK_IDLE TASK_INTERRUPTIBLE
+#endif
+
 /**
  * Verification thread to check parent FID consistency.
  *
@@ -157,54 +168,39 @@ static void ofd_inconsistency_verify_one(const struct lu_env *env,
  * \retval		0 on successful thread termination
  * \retval		negative value if thread can't start
  */
-static int ofd_inconsistency_verification_main(void *args)
+static int ofd_inconsistency_verification_main(void *_args)
 {
-	struct lu_env env;
-	struct ofd_device *ofd = args;
-	struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread;
+	struct oivm_args *args = _args;
+	struct lu_env *env = &args->od_env;
+	struct ofd_device *ofd = args->od_ofd;
 	struct ofd_inconsistency_item *oii;
-	struct lfsck_req_local *lrl = NULL;
-	struct l_wait_info lwi = { 0 };
-	int rc;
+	struct lfsck_req_local *lrl = &args->od_lrl;
 	ENTRY;
 
-	rc = lu_env_init(&env, LCT_DT_THREAD);
-	spin_lock(&ofd->ofd_inconsistency_lock);
-	thread_set_flags(thread, rc ? SVC_STOPPED : SVC_RUNNING);
-	wake_up_all(&thread->t_ctl_waitq);
-	spin_unlock(&ofd->ofd_inconsistency_lock);
-	if (rc)
-		RETURN(rc);
-
-	OBD_ALLOC_PTR(lrl);
-	if (unlikely(!lrl))
-		GOTO(out_unlocked, rc = -ENOMEM);
-
 	lrl->lrl_event = LEL_PAIRS_VERIFY_LOCAL;
 	lrl->lrl_active = LFSCK_TYPE_LAYOUT;
+	complete(args->od_started);
 
 	spin_lock(&ofd->ofd_inconsistency_lock);
-	while (1) {
-		if (unlikely(!thread_is_running(thread)))
-			break;
+	while (({set_current_state(TASK_IDLE);
+		 !kthread_should_stop(); })) {
 
 		while (!list_empty(&ofd->ofd_inconsistency_list)) {
+			__set_current_state(TASK_RUNNING);
 			oii = list_entry(ofd->ofd_inconsistency_list.next,
 					 struct ofd_inconsistency_item,
 					 oii_list);
 			list_del_init(&oii->oii_list);
 			spin_unlock(&ofd->ofd_inconsistency_lock);
-			ofd_inconsistency_verify_one(&env, ofd, oii, lrl);
+			ofd_inconsistency_verify_one(env, ofd, oii, lrl);
 			spin_lock(&ofd->ofd_inconsistency_lock);
 		}
 
 		spin_unlock(&ofd->ofd_inconsistency_lock);
-		l_wait_event(thread->t_ctl_waitq,
-			     !list_empty(&ofd->ofd_inconsistency_list) ||
-			     !thread_is_running(thread),
-			     &lwi);
+		schedule();
 		spin_lock(&ofd->ofd_inconsistency_lock);
 	}
+	__set_current_state(TASK_RUNNING);
 
 	while (!list_empty(&ofd->ofd_inconsistency_list)) {
 		struct ofd_object *fo;
@@ -216,28 +212,20 @@ static int ofd_inconsistency_verification_main(void *args)
 		fo = oii->oii_obj;
 		spin_unlock(&ofd->ofd_inconsistency_lock);
 
-		ofd_write_lock(&env, fo);
+		ofd_write_lock(env, fo);
 		fo->ofo_pfid_checking = 0;
-		ofd_write_unlock(&env, fo);
+		ofd_write_unlock(env, fo);
 
-		ofd_object_put(&env, fo);
+		ofd_object_put(env, fo);
 		OBD_FREE_PTR(oii);
 		spin_lock(&ofd->ofd_inconsistency_lock);
 	}
 
-	OBD_FREE_PTR(lrl);
-
-	GOTO(out, rc = 0);
-
-out_unlocked:
-	spin_lock(&ofd->ofd_inconsistency_lock);
-out:
-	thread_set_flags(thread, SVC_STOPPED);
-	wake_up_all(&thread->t_ctl_waitq);
 	spin_unlock(&ofd->ofd_inconsistency_lock);
-	lu_env_fini(&env);
 
-	return rc;
+	lu_env_fini(&args->od_env);
+	OBD_FREE_PTR(args);
+	return 0;
 }
 
 /**
@@ -252,32 +240,50 @@ out:
  */
 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
 {
-	struct ptlrpc_thread	*thread = &ofd->ofd_inconsistency_thread;
-	struct l_wait_info	 lwi	= { 0 };
 	struct task_struct	*task;
+	struct oivm_args	*args;
+	DECLARE_COMPLETION_ONSTACK(started);
 	int			 rc;
 
-	spin_lock(&ofd->ofd_inconsistency_lock);
-	if (unlikely(thread_is_running(thread))) {
-		spin_unlock(&ofd->ofd_inconsistency_lock);
-
+	if (ofd->ofd_inconsistency_task)
 		return -EALREADY;
+
+	OBD_ALLOC_PTR(args);
+	if (!args)
+		return -ENOMEM;
+	rc = lu_env_init(&args->od_env, LCT_DT_THREAD);
+	if (rc) {
+		OBD_FREE_PTR(args);
+		return rc;
 	}
 
-	thread_set_flags(thread, 0);
-	spin_unlock(&ofd->ofd_inconsistency_lock);
-	task = kthread_run(ofd_inconsistency_verification_main, ofd,
-			   "inconsistency_verification");
+	args->od_ofd = ofd;
+	args->od_started = &started;
+	task = kthread_create(ofd_inconsistency_verification_main, args,
+			      "inconsistency_verification");
 	if (IS_ERR(task)) {
 		rc = PTR_ERR(task);
 		CERROR("%s: cannot start self_repair thread: rc = %d\n",
 		       ofd_name(ofd), rc);
 	} else {
 		rc = 0;
-		l_wait_event(thread->t_ctl_waitq,
-			     thread_is_running(thread) ||
-			     thread_is_stopped(thread),
-			     &lwi);
+		spin_lock(&ofd->ofd_inconsistency_lock);
+		if (ofd->ofd_inconsistency_task)
+			rc = -EALREADY;
+		else
+			ofd->ofd_inconsistency_task = task;
+		spin_unlock(&ofd->ofd_inconsistency_lock);
+
+		if (rc)
+			kthread_stop(task);
+		else {
+			wake_up_process(task);
+			wait_for_completion(&started);
+		}
+	}
+	if (rc) {
+		lu_env_fini(&args->od_env);
+		OBD_FREE_PTR(args);
 	}
 
 	return rc;
@@ -293,22 +299,16 @@ int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd)
  */
 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd)
 {
-	struct ptlrpc_thread	*thread = &ofd->ofd_inconsistency_thread;
-	struct l_wait_info	 lwi	= { 0 };
+	struct task_struct *task;
 
 	spin_lock(&ofd->ofd_inconsistency_lock);
-	if (thread_is_init(thread) || thread_is_stopped(thread)) {
-		spin_unlock(&ofd->ofd_inconsistency_lock);
+	task = ofd->ofd_inconsistency_task;
+	ofd->ofd_inconsistency_task = NULL;
+	spin_unlock(&ofd->ofd_inconsistency_lock);
 
+	if (!task)
 		return -EALREADY;
-	}
-
-	thread_set_flags(thread, SVC_STOPPING);
-	spin_unlock(&ofd->ofd_inconsistency_lock);
-	wake_up_all(&thread->t_ctl_waitq);
-	l_wait_event(thread->t_ctl_waitq,
-		     thread_is_stopped(thread),
-		     &lwi);
+	kthread_stop(task);
 
 	return 0;
 }
@@ -356,9 +356,9 @@ static void ofd_add_inconsistency_item(const struct lu_env *env,
 	if (list_empty(&ofd->ofd_inconsistency_list))
 		wakeup = true;
 	list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list);
+	if (wakeup && ofd->ofd_inconsistency_task)
+		wake_up_process(ofd->ofd_inconsistency_task);
 	spin_unlock(&ofd->ofd_inconsistency_lock);
-	if (wakeup)
-		wake_up_all(&ofd->ofd_inconsistency_thread.t_ctl_waitq);
 
 	/* XXX: When the found inconsistency exceeds some threshold,
 	 *	we can trigger the LFSCK to scan part of the system
@@ -426,6 +426,126 @@ int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
 }
 
 /**
+ * FLR: verify the layout version of object.
+ *
+ * \param[in] env	execution environment
+ * \param[in] fo	OFD object
+ * \param[in] oa	OBDO structure with layout version
+ *
+ * \retval		0 on successful verification
+ * \retval		-EINPROGRESS layout version is in transfer
+ * \retval		-ESTALE the layout version on client is stale
+ */
+int ofd_verify_layout_version(const struct lu_env *env,
+			      struct ofd_object *fo, const struct obdo *oa)
+{
+	__u32 layout_version;
+	int rc;
+	ENTRY;
+
+	if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_SKIP_LV_CHECK)))
+		GOTO(out, rc = 0);
+
+	rc = ofd_object_ff_load(env, fo);
+	if (rc < 0) {
+		if (rc == -ENODATA)
+			rc = -EINPROGRESS;
+		GOTO(out, rc);
+	}
+
+	layout_version = fo->ofo_ff.ff_layout_version;
+	if (oa->o_layout_version >= layout_version &&
+	    oa->o_layout_version <= layout_version + fo->ofo_ff.ff_range)
+		GOTO(out, rc = 0);
+
+	/* normal traffic, decide if to return ESTALE or EINPROGRESS */
+	layout_version &= ~LU_LAYOUT_RESYNC;
+
+	/* this update is not legitimate */
+	if ((oa->o_layout_version & ~LU_LAYOUT_RESYNC) <= layout_version)
+		GOTO(out, rc = -ESTALE);
+
+	/* layout version may not be transmitted yet */
+	if ((oa->o_layout_version & ~LU_LAYOUT_RESYNC) > layout_version)
+		GOTO(out, rc = -EINPROGRESS);
+
+	EXIT;
+
+out:
+	CDEBUG(D_INODE, DFID " verify layout version: %u vs. %u/%u, rc: %d\n",
+	       PFID(lu_object_fid(&fo->ofo_obj.do_lu)),
+	       oa->o_layout_version, fo->ofo_ff.ff_layout_version,
+	       fo->ofo_ff.ff_range, rc);
+	return rc;
+
+}
+
+/*
+ * Lazy ATIME update to refresh atime every ofd_atime_diff
+ * seconds so that external scanning tool can see it actual
+ * within that period and be able to identify accessed files
+ */
+static void ofd_handle_atime(const struct lu_env *env, struct ofd_device *ofd,
+			     struct ofd_object *fo, time64_t atime)
+{
+	struct lu_attr *la;
+	struct dt_object *o;
+	struct thandle *th;
+	int rc;
+
+	if (ofd->ofd_atime_diff == 0)
+		return;
+
+	la = &ofd_info(env)->fti_attr2;
+	o = ofd_object_child(fo);
+
+	if (unlikely(fo->ofo_atime_ondisk == 0)) {
+		rc = dt_attr_get(env, o, la);
+		if (unlikely(rc))
+			return;
+		LASSERT(la->la_valid & LA_ATIME);
+		if (la->la_atime == 0)
+			la->la_atime = la->la_mtime;
+		fo->ofo_atime_ondisk = la->la_atime;
+	}
+	if (atime - fo->ofo_atime_ondisk < ofd->ofd_atime_diff)
+		return;
+
+	/* atime hasn't been updated too long, update it */
+	fo->ofo_atime_ondisk = atime;
+
+	th = ofd_trans_create(env, ofd);
+	if (IS_ERR(th)) {
+		CERROR("%s: cannot create transaction: rc = %d\n",
+		       ofd_name(ofd), (int)PTR_ERR(th));
+		return;
+	}
+
+	la->la_valid = LA_ATIME;
+	rc = dt_declare_attr_set(env, o, la, th);
+	if (rc)
+		GOTO(out_tx, rc);
+
+	rc = dt_trans_start_local(env, ofd->ofd_osd , th);
+	if (rc) {
+		CERROR("%s: cannot start transaction: rc = %d\n",
+		       ofd_name(ofd), rc);
+		GOTO(out_tx, rc);
+	}
+
+	ofd_read_lock(env, fo);
+	if (ofd_object_exists(fo)) {
+		la->la_atime = fo->ofo_atime_ondisk;
+		rc = dt_attr_set(env, o, la, th);
+	}
+
+	ofd_read_unlock(env, fo);
+
+out_tx:
+	ofd_trans_stop(env, ofd, th, rc);
+}
+
+/**
  * Prepare buffers for read request processing.
  *
  * This function converts remote buffers from client to local buffers
@@ -450,10 +570,13 @@ static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
 			   struct ofd_device *ofd, const struct lu_fid *fid,
 			   struct lu_attr *la, struct obdo *oa, int niocount,
 			   struct niobuf_remote *rnb, int *nr_local,
-			   struct niobuf_local *lnb, char *jobid)
+			   struct niobuf_local *lnb)
 {
-	struct ofd_object	*fo;
-	int			 i, j, rc, tot_bytes = 0;
+	struct ofd_object *fo;
+	int i, j, rc, tot_bytes = 0;
+	enum dt_bufs_type dbt = DT_BUFS_TYPE_READ;
+	int maxlnb = *nr_local;
+	__u64 begin, end;
 
 	ENTRY;
 	LASSERT(env != NULL);
@@ -463,6 +586,11 @@ static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
 		RETURN(PTR_ERR(fo));
 	LASSERT(fo != NULL);
 
+	ofd_info(env)->fti_obj = fo;
+
+	if (oa->o_valid & OBD_MD_FLATIME)
+		ofd_handle_atime(env, ofd, fo, oa->o_atime);
+
 	ofd_read_lock(env, fo);
 	if (!ofd_object_exists(fo))
 		GOTO(unlock, rc = -ENOENT);
@@ -473,30 +601,49 @@ static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
 			GOTO(unlock, rc);
 	}
 
-	*nr_local = 0;
-	for (i = 0, j = 0; i < niocount; i++) {
+	if (ptlrpc_connection_is_local(exp->exp_connection))
+		dbt |= DT_BUFS_TYPE_LOCAL;
+
+	begin = -1;
+	end = 0;
+
+	for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) {
+		begin = min_t(__u64, begin, rnb[i].rnb_offset);
+		end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
+
+		if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
+			rnb[i].rnb_len = 100 * 1024 * 1024;
+
 		rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
-				 lnb + j, 0);
+				 lnb + j, maxlnb, dbt);
 		if (unlikely(rc < 0))
 			GOTO(buf_put, rc);
 		LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 		/* correct index for local buffers to continue with */
 		j += rc;
 		*nr_local += rc;
+		maxlnb -= rc;
 		LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 		tot_bytes += rnb[i].rnb_len;
 	}
 
 	LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
-	rc = dt_attr_get(env, ofd_object_child(fo), la);
-	if (unlikely(rc))
-		GOTO(buf_put, rc);
-
 	rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
 	if (unlikely(rc))
 		GOTO(buf_put, rc);
+	ofd_read_unlock(env, fo);
+
+	ofd_access(env, ofd,
+		&(struct lu_fid) {
+			.f_seq = oa->o_parent_seq,
+			.f_oid = oa->o_parent_oid,
+			.f_ver = oa->o_stripe_idx,
+		},
+		begin, end,
+		tot_bytes,
+		niocount,
+		READ);
 
-	ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes);
 	RETURN(0);
 
 buf_put:
@@ -535,18 +682,21 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
 			    struct lu_attr *la, struct obdo *oa,
 			    int objcount, struct obd_ioobj *obj,
 			    struct niobuf_remote *rnb, int *nr_local,
-			    struct niobuf_local *lnb, char *jobid)
+			    struct niobuf_local *lnb)
 {
-	struct ofd_object	*fo;
-	int			 i, j, k, rc = 0, tot_bytes = 0;
+	struct ofd_object *fo;
+	int i, j, k, rc = 0, tot_bytes = 0;
+	enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE;
+	int maxlnb = *nr_local;
+	__u64 begin, end;
 
 	ENTRY;
 	LASSERT(env != NULL);
 	LASSERT(objcount == 1);
 
 	if (unlikely(exp->exp_obd->obd_recovering)) {
-		u64 seq = fid_seq(fid);
-		u64 oid = fid_oid(fid);
+		u64 seq = ostid_seq(&oa->o_oi);
+		u64 oid = ostid_id(&oa->o_oi);
 		struct ofd_seq *oseq;
 
 		oseq = ofd_seq_load(env, ofd, seq);
@@ -600,63 +750,103 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
 		ofd_seq_put(env, oseq);
 	}
 
+	/* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
+	 * space back if possible, we have to do this outside of the lock as
+	 * grant preparation may need to sync whole fs thus wait for all the
+	 * transactions to complete. */
+	tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
+
 	fo = ofd_object_find(env, ofd, fid);
 	if (IS_ERR(fo))
 		GOTO(out, rc = PTR_ERR(fo));
 	LASSERT(fo != NULL);
 
-	ofd_read_lock(env, fo);
+	ofd_info(env)->fti_obj = fo;
+
 	if (!ofd_object_exists(fo)) {
 		CERROR("%s: BRW to missing obj "DOSTID"\n",
 		       exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
-		ofd_read_unlock(env, fo);
 		ofd_object_put(env, fo);
 		GOTO(out, rc = -ENOENT);
 	}
 
-	if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
-		rc = ofd_verify_ff(env, fo, oa);
-		if (rc != 0) {
-			ofd_read_unlock(env, fo);
-			ofd_object_put(env, fo);
-			GOTO(out, rc);
-		}
-	}
+	if (ptlrpc_connection_is_local(exp->exp_connection))
+		dbt |= DT_BUFS_TYPE_LOCAL;
 
-	/* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
-	 * space back if possible */
-	tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
+	begin = -1;
+	end = 0;
 
 	/* parse remote buffers to local buffers and prepare the latter */
-	*nr_local = 0;
-	for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
+	for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
+		begin = min_t(__u64, begin, rnb[i].rnb_offset);
+		end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len);
+
+		if (OBD_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF))
+			rnb[i].rnb_len += PAGE_SIZE;
 		rc = dt_bufs_get(env, ofd_object_child(fo),
-				 rnb + i, lnb + j, 1);
+				 rnb + i, lnb + j, maxlnb, dbt);
 		if (unlikely(rc < 0))
-			GOTO(err, rc);
+			GOTO(err_nolock, rc);
 		LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
 		/* correct index for local buffers to continue with */
 		for (k = 0; k < rc; k++) {
 			lnb[j+k].lnb_flags = rnb[i].rnb_flags;
+			lnb[j+k].lnb_flags &= ~OBD_BRW_LOCALS;
 			if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
 				lnb[j+k].lnb_rc = -ENOSPC;
 		}
 		j += rc;
 		*nr_local += rc;
+		maxlnb -= rc;
 		LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
 		tot_bytes += rnb[i].rnb_len;
 	}
 	LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
 
+	ofd_read_lock(env, fo);
+	if (!ofd_object_exists(fo)) {
+		CERROR("%s: BRW to missing obj "DOSTID": rc = -ENOENT\n",
+		       exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
+		GOTO(err, rc = -ENOENT);
+	}
+
+	if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
+		rc = ofd_verify_ff(env, fo, oa);
+		if (rc != 0)
+			GOTO(err, rc);
+	}
+
+	/* need to verify layout version */
+	if (oa->o_valid & OBD_MD_LAYOUT_VERSION) {
+		rc = ofd_verify_layout_version(env, fo, oa);
+		if (rc)
+			GOTO(err, rc);
+		oa->o_valid &= ~OBD_MD_LAYOUT_VERSION;
+	}
+
 	rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
 	if (unlikely(rc != 0))
 		GOTO(err, rc);
 
-	ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes);
+	ofd_read_unlock(env, fo);
+
+	ofd_access(env, ofd,
+		&(struct lu_fid) {
+			.f_seq = oa->o_parent_seq,
+			.f_oid = oa->o_parent_oid,
+			.f_ver = oa->o_stripe_idx,
+		},
+		begin, end,
+		tot_bytes,
+		obj->ioo_bufcnt,
+		WRITE);
+
 	RETURN(0);
+
 err:
-	dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 	ofd_read_unlock(env, fo);
+err_nolock:
+	dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
 	ofd_object_put(env, fo);
 	/* tgt_grant_prepare_write() was called, so we must commit */
 	tgt_grant_commit(exp, oa->o_grant_used, rc);
@@ -695,23 +885,19 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 	struct tgt_session_info	*tsi = tgt_ses_info(env);
 	struct ofd_device	*ofd = ofd_exp(exp);
 	struct ofd_thread_info	*info;
-	char			*jobid;
 	const struct lu_fid	*fid = &oa->o_oi.oi_fid;
 	int			 rc = 0;
 
 	if (*nr_local > PTLRPC_MAX_BRW_PAGES) {
-		CERROR("%s: bulk has too many pages %d, which exceeds the"
-		       "maximum pages per RPC of %d\n",
+		CERROR("%s: bulk has too many pages %d, which exceeds the maximum pages per RPC of %d\n",
 		       exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES);
 		RETURN(-EPROTO);
 	}
 
 	if (tgt_ses_req(tsi) == NULL) { /* echo client case */
 		info = ofd_info_init(env, exp);
-		jobid = NULL;
 	} else {
 		info = tsi2ofd_info(tsi);
-		jobid = tsi->tsi_jobid;
 	}
 
 	LASSERT(oa != NULL);
@@ -743,13 +929,11 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
 	if (cmd == OBD_BRW_WRITE) {
 		la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
 		rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa,
-				      objcount, obj, rnb, nr_local, lnb, jobid);
+				      objcount, obj, rnb, nr_local, lnb);
 	} else if (cmd == OBD_BRW_READ) {
 		tgt_grant_prepare_read(env, exp, oa);
 		rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa,
-				     obj->ioo_bufcnt, rnb, nr_local, lnb,
-				     jobid);
-		obdo_from_la(oa, &info->fti_attr, LA_ATIME);
+				     obj->ioo_bufcnt, rnb, nr_local, lnb);
 	} else {
 		CERROR("%s: wrong cmd %d received!\n",
 		       exp->exp_obd->obd_name, cmd);
@@ -784,16 +968,11 @@ ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
 
 	LASSERT(niocount > 0);
 
-	fo = ofd_object_find(env, ofd, fid);
-	if (IS_ERR(fo))
-		RETURN(PTR_ERR(fo));
+	fo = ofd_info(env)->fti_obj;
 	LASSERT(fo != NULL);
 	LASSERT(ofd_object_exists(fo));
 	dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
 
-	ofd_read_unlock(env, fo);
-	ofd_object_put(env, fo);
-	/* second put is pair to object_get in ofd_preprw_read */
 	ofd_object_put(env, fo);
 
 	RETURN(0);
@@ -809,7 +988,7 @@ ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
  * \param[in] ofd	OFD device
  * \param[in] ofd_obj	OFD object
  * \param[in] la	object attributes
- * \param[in] ff	parent FID
+ * \param[in] oa	obdo
  *
  * \retval		0 on successful attributes update
  * \retval		negative value on error
@@ -817,14 +996,15 @@ ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
 static int
 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
 		   struct ofd_object *ofd_obj, struct lu_attr *la,
-		   struct filter_fid *ff)
+		   struct obdo *oa)
 {
 	struct ofd_thread_info	*info = ofd_info(env);
+	struct filter_fid	*ff = &info->fti_mds_fid;
 	__u64			 valid = la->la_valid;
-	int			 rc;
 	struct thandle		*th;
 	struct dt_object	*dt_obj;
-	int			 ff_needed = 0;
+	int			 fl = 0;
+	int			 rc;
 
 	ENTRY;
 
@@ -839,15 +1019,8 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
 	if (rc != 0)
 		GOTO(out, rc);
 
-	if (ff != NULL) {
-		rc = ofd_object_ff_load(env, ofd_obj);
-		if (rc == -ENODATA)
-			ff_needed = 1;
-		else if (rc < 0)
-			GOTO(out, rc);
-	}
-
-	if (!la->la_valid && !ff_needed)
+	if (!la->la_valid && !(oa->o_valid &
+	    (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
 		/* no attributes to set */
 		GOTO(out, rc = 0);
 
@@ -861,46 +1034,66 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
 			GOTO(out_tx, rc);
 	}
 
-	if (ff_needed) {
-		if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
-			ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
-		else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
-			le32_add_cpu(&ff->ff_parent.f_oid, -1);
-
-		info->fti_buf.lb_buf = ff;
-		info->fti_buf.lb_len = sizeof(*ff);
+	if (oa->o_valid & (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT |
+			   OBD_MD_LAYOUT_VERSION)) {
 		rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
 					  XATTR_NAME_FID, 0, th);
 		if (rc)
 			GOTO(out_tx, rc);
 	}
-
 	/* We don't need a transno for this operation which will be re-executed
 	 * anyway when the OST_WRITE (with a transno assigned) is replayed */
 	rc = dt_trans_start_local(env, ofd->ofd_osd , th);
 	if (rc)
 		GOTO(out_tx, rc);
 
-	/* set uid/gid/pid */
+	ofd_read_lock(env, ofd_obj);
+
+	rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */);
+	if (rc != 0)
+		GOTO(out_unlock, rc);
+
+	if (!la->la_valid && !(oa->o_valid &
+	    (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION)))
+		/* no attributes to set */
+		GOTO(out_unlock, rc = 0);
+
+
+
+	/* set uid/gid/projid */
 	if (la->la_valid) {
 		rc = dt_attr_set(env, dt_obj, la, th);
 		if (rc)
-			GOTO(out_tx, rc);
-	}
-
-	/* set filter fid EA */
-	if (ff_needed) {
-		if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
-			GOTO(out_tx, rc);
-
-		rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID,
-				  0, th);
-		if (!rc)
-			filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
+			GOTO(out_unlock, rc);
 	}
 
-	GOTO(out_tx, rc);
-
+	fl = ofd_object_ff_update(env, ofd_obj, oa, ff);
+	if (fl <= 0)
+		GOTO(out_unlock, rc = fl);
+
+	/* set filter fid EA.
+	 * FIXME: it holds read lock of ofd object to modify the XATTR_NAME_FID
+	 * while the write lock should be held. However, it should work because
+	 * write RPCs only modify ff_{parent,layout} and those information will
+	 * be the same from all the write RPCs. The reason that fl is not used
+	 * in dt_xattr_set() is to allow this race. */
+	if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
+		GOTO(out_unlock, rc);
+	if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
+		ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
+	else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
+		le32_add_cpu(&ff->ff_parent.f_oid, -1);
+
+	info->fti_buf.lb_buf = ff;
+	info->fti_buf.lb_len = sizeof(*ff);
+	rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th);
+	if (rc == 0)
+		filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff));
+
+	GOTO(out_unlock, rc);
+
+out_unlock:
+	ofd_read_unlock(env, ofd_obj);
 out_tx:
 	dt_trans_stop(env, ofd->ofd_osd, th);
 out:
@@ -1004,10 +1197,11 @@ static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp)
 static int
 ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
 		   struct ofd_device *ofd, const struct lu_fid *fid,
-		   struct lu_attr *la, struct filter_fid *ff, int objcount,
+		   struct lu_attr *la, struct obdo *oa, int objcount,
 		   int niocount, struct niobuf_local *lnb,
 		   unsigned long granted, int old_rc)
 {
+	struct ofd_thread_info *info = ofd_info(env);
 	struct filter_export_data *fed = &exp->exp_filter_data;
 	struct ofd_object *fo;
 	struct dt_object *o;
@@ -1015,7 +1209,7 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
 	int rc = 0;
 	int rc2 = 0;
 	int retries = 0;
-	int i;
+	int i, restart = 0;
 	bool soft_sync = false;
 	bool cb_registered = false;
 	bool fake_write = false;
@@ -1024,15 +1218,16 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
 
 	LASSERT(objcount == 1);
 
-	fo = ofd_object_find(env, ofd, fid);
+	fo = ofd_info(env)->fti_obj;
 	LASSERT(fo != NULL);
-	LASSERT(ofd_object_exists(fo));
 
 	o = ofd_object_child(fo);
 	LASSERT(o != NULL);
 
 	if (old_rc)
 		GOTO(out, rc = old_rc);
+	if (!ofd_object_exists(fo))
+		GOTO(out, rc = -ENOENT);
 
 	/*
 	 * The first write to each object must set some attributes.  It is
@@ -1040,14 +1235,14 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp,
 	 * dt_declare_write_commit() since quota enforcement is now handled in
 	 * declare phases.
 	 */
-	rc = ofd_write_attr_set(env, ofd, fo, la, ff);
+	rc = ofd_write_attr_set(env, ofd, fo, la, oa);
 	if (rc)
 		GOTO(out, rc);
 
 	la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
 
 	/* do fake write, to simulate the write case for performance testing */
-	if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) {
+	if (OBD_FAIL_CHECK_QUIET(OBD_FAIL_OST_FAKE_RW)) {
 		struct niobuf_local *last = &lnb[niocount - 1];
 		__u64 file_size = last->lnb_file_offset + last->lnb_len;
 		__u64 valid = la->la_valid;
@@ -1073,7 +1268,7 @@ retry:
 	if (IS_ERR(th))
 		GOTO(out, rc = PTR_ERR(th));
 
-	th->th_sync |= ofd->ofd_syncjournal;
+	th->th_sync |= ofd->ofd_sync_journal;
 	if (th->th_sync == 0) {
 		for (i = 0; i < niocount; i++) {
 			if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
@@ -1094,6 +1289,10 @@ retry:
 			GOTO(out_stop, rc);
 	}
 
+	/* don't update atime on disk if it is older */
+	if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk)
+		la->la_valid &= ~LA_ATIME;
+
 	if (la->la_valid) {
 		/* update [mac]time if needed */
 		rc = dt_declare_attr_set(env, o, la, th);
@@ -1105,21 +1304,35 @@ retry:
 	if (rc)
 		GOTO(out_stop, rc);
 
-	if (likely(!fake_write)) {
-		rc = dt_write_commit(env, o, lnb, niocount, th);
-		if (rc)
-			GOTO(out_stop, rc);
-	}
+	ofd_read_lock(env, fo);
+	if (!ofd_object_exists(fo))
+		GOTO(out_unlock, rc = -ENOENT);
 
-	if (la->la_valid) {
+	/* Don't update timestamps if this write is older than a
+	 * setattr which modifies the timestamps. b=10150 */
+	if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) {
 		rc = dt_attr_set(env, o, la, th);
 		if (rc)
-			GOTO(out_stop, rc);
+			GOTO(out_unlock, rc);
+		if (la->la_valid & LA_ATIME)
+			fo->ofo_atime_ondisk = la->la_atime;
+	}
+
+	if (likely(!fake_write)) {
+		OBD_FAIL_TIMEOUT_ORSET(OBD_FAIL_OST_WR_ATTR_DELAY,
+				       OBD_FAIL_ONCE, cfs_fail_val);
+		rc = dt_write_commit(env, o, lnb, niocount, th, oa->o_size);
+		if (rc) {
+			restart = th->th_restart_tran;
+			GOTO(out_unlock, rc);
+		}
 	}
 
 	/* get attr to return */
 	rc = dt_attr_get(env, o, la);
 
+out_unlock:
+	ofd_read_unlock(env, fo);
 out_stop:
 	/* Force commit to make the just-deleted blocks
 	 * reusable. LU-456 */
@@ -1137,7 +1350,7 @@ out_stop:
 			granted = 0;
 	}
 
-	rc2 = ofd_trans_stop(env, ofd, th, rc);
+	rc2 = ofd_trans_stop(env, ofd, th, restart ? 0 : rc);
 	if (!rc)
 		rc = rc2;
 	if (rc == -ENOSPC && retries++ < 3) {
@@ -1146,6 +1359,16 @@ out_stop:
 		goto retry;
 	}
 
+	if (restart) {
+		retries++;
+		restart = 0;
+		if (retries % 10000 == 0)
+			CERROR("%s: restart IO write too many times: %d\n",
+				ofd_name(ofd), retries);
+		CDEBUG(D_INODE, "retry transaction, retries:%d\n",
+		       retries);
+		goto retry;
+	}
 	if (!soft_sync)
 		/* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */
 		atomic_set(&fed->fed_soft_sync_count, 0);
@@ -1155,9 +1378,6 @@ out_stop:
 
 out:
 	dt_bufs_put(env, o, lnb, niocount);
-	ofd_read_unlock(env, fo);
-	ofd_object_put(env, fo);
-	/* second put is pair to object_get in ofd_preprw_write */
 	ofd_object_put(env, fo);
 	if (granted > 0)
 		tgt_grant_commit(exp, granted, old_rc);
@@ -1189,41 +1409,72 @@ out:
 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 		 struct obdo *oa, int objcount, struct obd_ioobj *obj,
 		 struct niobuf_remote *rnb, int npages,
-		 struct niobuf_local *lnb, int old_rc)
+		 struct niobuf_local *lnb, int old_rc, int nob, ktime_t kstart)
 {
-	struct ofd_thread_info	*info = ofd_info(env);
-	struct ofd_mod_data	*fmd;
-	__u64			 valid;
-	struct ofd_device	*ofd = ofd_exp(exp);
-	struct filter_fid	*ff = NULL;
-	const struct lu_fid	*fid = &oa->o_oi.oi_fid;
-	int			 rc = 0;
+	struct tgt_session_info	*tsi = tgt_ses_info(env);
+	struct ofd_thread_info *info = ofd_info(env);
+	struct ofd_device *ofd = ofd_exp(exp);
+	const struct lu_fid *fid = &oa->o_oi.oi_fid;
+	struct ldlm_namespace *ns = ofd->ofd_namespace;
+	struct ldlm_resource *rs = NULL;
+	char *jobid;
+	__u64 valid;
+	int rc = 0;
+	int root_squash = 0;
 
 	LASSERT(npages > 0);
 
-	if (cmd == OBD_BRW_WRITE) {
-		/* Don't update timestamps if this write is older than a
-		 * setattr which modifies the timestamps. b=10150 */
-
-		/* XXX when we start having persistent reservations this needs
-		 * to be changed to ofd_fmd_get() to create the fmd if it
-		 * doesn't already exist so we can store the reservation handle
-		 * there. */
-		valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID;
-		fmd = ofd_fmd_find(exp, fid);
-		if (!fmd || fmd->fmd_mactime_xid < info->fti_xid)
-			valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
-				 OBD_MD_FLCTIME;
-		ofd_fmd_put(exp, fmd);
-		la_from_obdo(&info->fti_attr, oa, valid);
+	if (tgt_ses_req(tsi) == NULL) { /* echo client case */
+		jobid = NULL;
+	} else {
+		jobid = tsi->tsi_jobid;
+	}
 
-		if (oa->o_valid & OBD_MD_FLFID) {
-			ff = &info->fti_mds_fid;
-			ofd_prepare_fidea(ff, oa);
+	if (cmd == OBD_BRW_WRITE) {
+		struct lu_nodemap *nodemap;
+		__u32 mapped_uid, mapped_gid, mapped_projid;
+
+		/* doing this before the commit operation places the counter
+		 * update almost immediately after reply to the client, which
+		 * gives reasonable time stats and lets us use the actual
+		 * bytes of i/o (rather than requested)
+		 */
+		ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE_BYTES, jobid, nob);
+		ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid,
+				 ktime_us_delta(ktime_get(), kstart));
+
+		nodemap = nodemap_get_from_exp(exp);
+		if (IS_ERR(nodemap))
+			RETURN(PTR_ERR(nodemap));
+		mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID,
+					    NODEMAP_FS_TO_CLIENT,
+					    oa->o_uid);
+		mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID,
+					    NODEMAP_FS_TO_CLIENT,
+					    oa->o_gid);
+		mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID,
+					       NODEMAP_FS_TO_CLIENT,
+					       oa->o_projid);
+
+		if (!IS_ERR_OR_NULL(nodemap)) {
+			/* do not bypass quota enforcement if squashed uid */
+			if (unlikely(mapped_uid == nodemap->nm_squash_uid)) {
+				int idx;
+
+				for (idx = 0; idx < npages; idx++)
+					lnb[idx].lnb_flags &=
+						~OBD_BRW_SYS_RESOURCE;
+				root_squash = 1;
+			}
+			nodemap_putref(nodemap);
 		}
 
+		valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID |
+			OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
+		la_from_obdo(&info->fti_attr, oa, valid);
+
 		rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr,
-					ff, objcount, npages, lnb,
+					oa, objcount, npages, lnb,
 					oa->o_grant_used, old_rc);
 		if (rc == 0)
 			obdo_from_la(oa, &info->fti_attr,
@@ -1257,18 +1508,21 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 					oa->o_flags = OBD_FL_NO_PRJQUOTA;
 			}
 
+			if (root_squash)
+				oa->o_flags |= OBD_FL_ROOT_SQUASH;
+
 			oa->o_valid |= OBD_MD_FLFLAGS;
 			oa->o_valid |= OBD_MD_FLALLQUOTA;
 		}
-	} else if (cmd == OBD_BRW_READ) {
-		struct ldlm_namespace *ns = ofd->ofd_namespace;
-
-		/* If oa != NULL then ofd_preprw_read updated the inode
-		 * atime and we should update the lvb so that other glimpses
-		 * will also get the updated value. bug 5972 */
-		if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
-			 struct ldlm_resource *rs = NULL;
 
+		/**
+		 * Update LVB after writing finish for server lock, see
+		 * comments in ldlm_lock_decref_internal(), If this is a
+		 * local lock on a server namespace and this was the last
+		 * reference, lock will be destroyed directly thus there
+		 * is no chance for ldlm_request_cancel() to update lvb.
+		 */
+		if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) {
 			ost_fid_build_resid(fid, &info->fti_resid);
 			rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
 					       LDLM_EXTENT, 0);
@@ -1277,6 +1531,19 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
 				ldlm_resource_putref(rs);
 			}
 		}
+
+		/* Convert back to client IDs. LU-9671.
+		 * nodemap_get_from_exp() may fail due to nodemap deactivated,
+		 * server ID will be returned back to client in that case. */
+		oa->o_uid = mapped_uid;
+		oa->o_gid = mapped_gid;
+		oa->o_projid = mapped_projid;
+	} else if (cmd == OBD_BRW_READ) {
+		/* see comment on LPROC_OFD_STATS_WRITE_BYTES usage above */
+		ofd_counter_incr(exp, LPROC_OFD_STATS_READ_BYTES, jobid, nob);
+		ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid,
+			 ktime_us_delta(ktime_get(), kstart));
+
 		rc = ofd_commitrw_read(env, ofd, fid, objcount,
 				       npages, lnb);
 		if (old_rc)