X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Ftarget%2Ftgt_main.c;h=b767d1d3931940551b97a83a8c0c8c7892c80db2;hp=485232376eee655d55e46171fe68bdf13fc68e31;hb=e763467ebe00913e8d03f855dc4b918b95099931;hpb=d00c924149b5dd042e6fee09ef724013df1b4e14

diff --git a/lustre/target/tgt_main.c b/lustre/target/tgt_main.c
index 4852323..b767d1d 100644
--- a/lustre/target/tgt_main.c
+++ b/lustre/target/tgt_main.c
@@ -21,7 +21,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright (c) 2012, 2013, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
  */
 /*
  * lustre/target/tgt_main.c
@@ -37,6 +37,112 @@
 #include "tgt_internal.h"
 #include "../ptlrpc/ptlrpc_internal.h"
 
+/*
+ * Save cross-MDT lock in lut_slc_locks.
+ *
+ * Lock R/W count is not saved, but released in unlock (not canceled remotely),
+ * instead only a refcount is taken, so that the remote MDT where the object
+ * resides can detect conflict with this lock there.
+ *
+ * \param lut target
+ * \param lock cross-MDT lock to save
+ * \param transno when the transaction with this transno is committed, this lock
+ *		  can be canceled.
+ */
+void tgt_save_slc_lock(struct lu_target *lut, struct ldlm_lock *lock,
+		       __u64 transno)
+{
+	spin_lock(&lut->lut_slc_locks_guard);
+	lock_res_and_lock(lock);
+	if (ldlm_is_cbpending(lock)) {
+		/* if it was canceld by server, don't save, because remote MDT
+		 * will do Sync-on-Cancel. */
+		LDLM_LOCK_PUT(lock);
+	} else {
+		lock->l_transno = transno;
+		/* if this lock is in the list already, there are two operations
+		 * both use this lock, and save it after use, so for the second
+		 * one, just put the refcount. */
+		if (list_empty(&lock->l_slc_link))
+			list_add_tail(&lock->l_slc_link, &lut->lut_slc_locks);
+		else
+			LDLM_LOCK_PUT(lock);
+	}
+	unlock_res_and_lock(lock);
+	spin_unlock(&lut->lut_slc_locks_guard);
+}
+EXPORT_SYMBOL(tgt_save_slc_lock);
+
+/*
+ * Discard cross-MDT lock from lut_slc_locks.
+ *
+ * This is called upon BAST, just remove lock from lut_slc_locks and put lock
+ * refcount. The BAST will cancel this lock.
+ *
+ * \param lut target
+ * \param lock cross-MDT lock to discard
+ */
+void tgt_discard_slc_lock(struct lu_target *lut, struct ldlm_lock *lock)
+{
+	spin_lock(&lut->lut_slc_locks_guard);
+	lock_res_and_lock(lock);
+	/* may race with tgt_cancel_slc_locks() */
+	if (lock->l_transno != 0) {
+		LASSERT(!list_empty(&lock->l_slc_link));
+		LASSERT(ldlm_is_cbpending(lock));
+		list_del_init(&lock->l_slc_link);
+		lock->l_transno = 0;
+		LDLM_LOCK_PUT(lock);
+	}
+	unlock_res_and_lock(lock);
+	spin_unlock(&lut->lut_slc_locks_guard);
+}
+EXPORT_SYMBOL(tgt_discard_slc_lock);
+
+/*
+ * Cancel cross-MDT locks upon transaction commit.
+ *
+ * Remove cross-MDT locks from lut_slc_locks, cancel them and put lock refcount.
+ *
+ * \param lut target
+ * \param transno transaction with this number was committed.
+ */
+void tgt_cancel_slc_locks(struct lu_target *lut, __u64 transno)
+{
+	struct ldlm_lock *lock, *next;
+	LIST_HEAD(list);
+	struct lustre_handle lockh;
+
+	spin_lock(&lut->lut_slc_locks_guard);
+	list_for_each_entry_safe(lock, next, &lut->lut_slc_locks,
+				 l_slc_link) {
+		lock_res_and_lock(lock);
+		LASSERT(lock->l_transno != 0);
+		if (lock->l_transno > transno) {
+			unlock_res_and_lock(lock);
+			continue;
+		}
+		/* ouch, another operation is using it after it's saved */
+		if (lock->l_readers != 0 || lock->l_writers != 0) {
+			unlock_res_and_lock(lock);
+			continue;
+		}
+		/* set CBPENDING so that this lock won't be used again */
+		ldlm_set_cbpending(lock);
+		lock->l_transno = 0;
+		list_move(&lock->l_slc_link, &list);
+		unlock_res_and_lock(lock);
+	}
+	spin_unlock(&lut->lut_slc_locks_guard);
+
+	list_for_each_entry_safe(lock, next, &list, l_slc_link) {
+		list_del_init(&lock->l_slc_link);
+		ldlm_lock2handle(lock, &lockh);
+		ldlm_cli_cancel(&lockh, LCF_ASYNC);
+		LDLM_LOCK_PUT(lock);
+	}
+}
+
 int tgt_init(const struct lu_env *env, struct lu_target *lut,
 	     struct obd_device *obd, struct dt_device *dt,
 	     struct tgt_opc_slice *slice, int request_fail_id,
@@ -46,7 +152,9 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
 	struct lu_attr		 attr;
 	struct lu_fid		 fid;
 	struct dt_object	*o;
-	int			 rc = 0;
+	struct tg_grants_data	*tgd = &lut->lut_tgd;
+	struct obd_statfs	*osfs;
+	int i, rc = 0;
 
 	ENTRY;
 
@@ -56,6 +164,10 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
 	lut->lut_bottom = dt;
 	lut->lut_last_rcvd = NULL;
 	lut->lut_client_bitmap = NULL;
+	atomic_set(&lut->lut_num_clients, 0);
+	atomic_set(&lut->lut_client_generation, 0);
+	lut->lut_reply_data = NULL;
+	lut->lut_reply_bitmap = NULL;
 	obd->u.obt.obt_lut = lut;
 	obd->u.obt.obt_magic = OBT_MAGIC;
 
@@ -67,17 +179,51 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
 	/* sptlrcp variables init */
 	rwlock_init(&lut->lut_sptlrpc_lock);
 	sptlrpc_rule_set_init(&lut->lut_sptlrpc_rset);
-	lut->lut_mds_capa = 1;
-	lut->lut_oss_capa = 1;
 
 	spin_lock_init(&lut->lut_flags_lock);
 	lut->lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
 
+	spin_lock_init(&lut->lut_slc_locks_guard);
+	INIT_LIST_HEAD(&lut->lut_slc_locks);
+
 	/* last_rcvd initialization is needed by replayable targets only */
 	if (!obd->obd_replayable)
 		RETURN(0);
 
+	/* initialize grant and statfs data in target */
+	dt_conf_get(env, lut->lut_bottom, &lut->lut_dt_conf);
+
+	/* statfs data */
+	spin_lock_init(&tgd->tgd_osfs_lock);
+	tgd->tgd_osfs_age = ktime_get_seconds() - 1000;
+	tgd->tgd_osfs_unstable = 0;
+	tgd->tgd_statfs_inflight = 0;
+	tgd->tgd_osfs_inflight = 0;
+
+	/* grant data */
+	spin_lock_init(&tgd->tgd_grant_lock);
+	tgd->tgd_tot_dirty = 0;
+	tgd->tgd_tot_granted = 0;
+	tgd->tgd_tot_pending = 0;
+	tgd->tgd_grant_compat_disable = 0;
+
+	/* populate cached statfs data */
+	osfs = &tgt_th_info(env)->tti_u.osfs;
+	rc = tgt_statfs_internal(env, lut, osfs, 0, NULL);
+	if (rc != 0) {
+		CERROR("%s: can't get statfs data, rc %d\n", tgt_name(lut),
+			rc);
+		GOTO(out, rc);
+	}
+	if (!is_power_of_2(osfs->os_bsize)) {
+		CERROR("%s: blocksize (%d) is not a power of 2\n",
+			tgt_name(lut), osfs->os_bsize);
+		GOTO(out, rc = -EPROTO);
+	}
+	tgd->tgd_blockbits = fls(osfs->os_bsize) - 1;
+
 	spin_lock_init(&lut->lut_translock);
+	spin_lock_init(&lut->lut_client_bitmap_lock);
 
 	OBD_ALLOC(lut->lut_client_bitmap, LR_MAX_CLIENTS >> 3);
 	if (lut->lut_client_bitmap == NULL)
@@ -95,56 +241,167 @@ int tgt_init(const struct lu_env *env, struct lu_target *lut,
 		rc = PTR_ERR(o);
 		CERROR("%s: cannot open LAST_RCVD: rc = %d\n", tgt_name(lut),
 		       rc);
-		GOTO(out_bitmap, rc);
+		GOTO(out_put, rc);
 	}
 
 	lut->lut_last_rcvd = o;
 	rc = tgt_server_data_init(env, lut);
 	if (rc < 0)
-		GOTO(out_obj, rc);
+		GOTO(out_put, rc);
 
 	/* prepare transactions callbacks */
 	lut->lut_txn_cb.dtc_txn_start = tgt_txn_start_cb;
 	lut->lut_txn_cb.dtc_txn_stop = tgt_txn_stop_cb;
-	lut->lut_txn_cb.dtc_txn_commit = NULL;
 	lut->lut_txn_cb.dtc_cookie = lut;
 	lut->lut_txn_cb.dtc_tag = LCT_DT_THREAD | LCT_MD_THREAD;
 	INIT_LIST_HEAD(&lut->lut_txn_cb.dtc_linkage);
 
 	dt_txn_callback_add(lut->lut_bottom, &lut->lut_txn_cb);
+	lut->lut_bottom->dd_lu_dev.ld_site->ls_tgt = lut;
+
+	/* reply_data is supported by MDT targets only for now */
+	if (strncmp(obd->obd_type->typ_name, LUSTRE_MDT_NAME, 3) != 0)
+		RETURN(0);
+
+	OBD_ALLOC(lut->lut_reply_bitmap,
+		  LUT_REPLY_SLOTS_MAX_CHUNKS * sizeof(unsigned long *));
+	if (lut->lut_reply_bitmap == NULL)
+		GOTO(out, rc = -ENOMEM);
+
+	memset(&attr, 0, sizeof(attr));
+	attr.la_valid = LA_MODE;
+	attr.la_mode = S_IFREG | S_IRUGO | S_IWUSR;
+	dof.dof_type = dt_mode_to_dft(S_IFREG);
+
+	lu_local_obj_fid(&fid, REPLY_DATA_OID);
+
+	o = dt_find_or_create(env, lut->lut_bottom, &fid, &dof, &attr);
+	if (IS_ERR(o)) {
+		rc = PTR_ERR(o);
+		CERROR("%s: cannot open REPLY_DATA: rc = %d\n", tgt_name(lut),
+		       rc);
+		GOTO(out, rc);
+	}
+	lut->lut_reply_data = o;
+
+	rc = tgt_reply_data_init(env, lut);
+	if (rc < 0)
+		GOTO(out, rc);
+
+	atomic_set(&lut->lut_sync_count, 0);
 
 	RETURN(0);
-out_obj:
-	lu_object_put(env, &lut->lut_last_rcvd->do_lu);
-	lut->lut_last_rcvd = NULL;
-out_bitmap:
-	OBD_FREE(lut->lut_client_bitmap, LR_MAX_CLIENTS >> 3);
+
+out:
+	dt_txn_callback_del(lut->lut_bottom, &lut->lut_txn_cb);
+out_put:
+	obd->u.obt.obt_magic = 0;
+	obd->u.obt.obt_lut = NULL;
+	if (lut->lut_last_rcvd != NULL) {
+		dt_object_put(env, lut->lut_last_rcvd);
+		lut->lut_last_rcvd = NULL;
+	}
+	if (lut->lut_client_bitmap != NULL)
+		OBD_FREE(lut->lut_client_bitmap, LR_MAX_CLIENTS >> 3);
 	lut->lut_client_bitmap = NULL;
+	if (lut->lut_reply_data != NULL)
+		dt_object_put(env, lut->lut_reply_data);
+	lut->lut_reply_data = NULL;
+	if (lut->lut_reply_bitmap != NULL) {
+		for (i = 0; i < LUT_REPLY_SLOTS_MAX_CHUNKS; i++) {
+			if (lut->lut_reply_bitmap[i] != NULL)
+				OBD_FREE_LARGE(lut->lut_reply_bitmap[i],
+				    BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
+				    sizeof(long));
+			lut->lut_reply_bitmap[i] = NULL;
+		}
+		OBD_FREE(lut->lut_reply_bitmap,
+			 LUT_REPLY_SLOTS_MAX_CHUNKS * sizeof(unsigned long *));
+	}
+	lut->lut_reply_bitmap = NULL;
 	return rc;
 }
 EXPORT_SYMBOL(tgt_init);
 
 void tgt_fini(const struct lu_env *env, struct lu_target *lut)
 {
+	int i;
+	int rc;
 	ENTRY;
 
+	if (lut->lut_lsd.lsd_feature_incompat & OBD_INCOMPAT_MULTI_RPCS &&
+	    atomic_read(&lut->lut_num_clients) == 0) {
+		/* Clear MULTI RPCS incompatibility flag that prevents previous
+		 * Lustre versions to mount a target with reply_data file */
+		lut->lut_lsd.lsd_feature_incompat &= ~OBD_INCOMPAT_MULTI_RPCS;
+		rc = tgt_server_data_update(env, lut, 1);
+		if (rc < 0)
+			CERROR("%s: unable to clear MULTI RPCS "
+			       "incompatibility flag\n",
+			       lut->lut_obd->obd_name);
+	}
+
 	sptlrpc_rule_set_free(&lut->lut_sptlrpc_rset);
 
+	if (lut->lut_reply_data != NULL)
+		dt_object_put(env, lut->lut_reply_data);
+	lut->lut_reply_data = NULL;
+	if (lut->lut_reply_bitmap != NULL) {
+		for (i = 0; i < LUT_REPLY_SLOTS_MAX_CHUNKS; i++) {
+			if (lut->lut_reply_bitmap[i] != NULL)
+				OBD_FREE_LARGE(lut->lut_reply_bitmap[i],
+				    BITS_TO_LONGS(LUT_REPLY_SLOTS_PER_CHUNK) *
+				    sizeof(long));
+			lut->lut_reply_bitmap[i] = NULL;
+		}
+		OBD_FREE(lut->lut_reply_bitmap,
+			 LUT_REPLY_SLOTS_MAX_CHUNKS * sizeof(unsigned long *));
+	}
+	lut->lut_reply_bitmap = NULL;
 	if (lut->lut_client_bitmap) {
 		OBD_FREE(lut->lut_client_bitmap, LR_MAX_CLIENTS >> 3);
 		lut->lut_client_bitmap = NULL;
 	}
 	if (lut->lut_last_rcvd) {
 		dt_txn_callback_del(lut->lut_bottom, &lut->lut_txn_cb);
-		lu_object_put(env, &lut->lut_last_rcvd->do_lu);
+		dt_object_put(env, lut->lut_last_rcvd);
 		lut->lut_last_rcvd = NULL;
 	}
 	EXIT;
 }
 EXPORT_SYMBOL(tgt_fini);
 
+static struct kmem_cache *tgt_thread_kmem;
+static struct kmem_cache *tgt_session_kmem;
+static struct lu_kmem_descr tgt_caches[] = {
+	{
+		.ckd_cache = &tgt_thread_kmem,
+		.ckd_name  = "tgt_thread_kmem",
+		.ckd_size  = sizeof(struct tgt_thread_info),
+	},
+	{
+		.ckd_cache = &tgt_session_kmem,
+		.ckd_name  = "tgt_session_kmem",
+		.ckd_size  = sizeof(struct tgt_session_info)
+	},
+	{
+		.ckd_cache = NULL
+	}
+};
+
+
 /* context key constructor/destructor: tg_key_init, tg_key_fini */
-LU_KEY_INIT(tgt, struct tgt_thread_info);
+static void *tgt_key_init(const struct lu_context *ctx,
+				  struct lu_context_key *key)
+{
+	struct tgt_thread_info *thread;
+
+	OBD_SLAB_ALLOC_PTR_GFP(thread, tgt_thread_kmem, GFP_NOFS);
+	if (thread == NULL)
+		return ERR_PTR(-ENOMEM);
+
+	return thread;
+}
 
 static void tgt_key_fini(const struct lu_context *ctx,
 			 struct lu_context_key *key, void *data)
@@ -161,7 +418,7 @@ static void tgt_key_fini(const struct lu_context *ctx,
 	if (args->ta_args != NULL)
 		OBD_FREE(args->ta_args, sizeof(args->ta_args[0]) *
 					args->ta_alloc_args);
-	OBD_FREE_PTR(info);
+	OBD_SLAB_FREE_PTR(info, tgt_thread_kmem);
 }
 
 static void tgt_key_exit(const struct lu_context *ctx,
@@ -180,12 +437,28 @@ struct lu_context_key tgt_thread_key = {
 	.lct_fini = tgt_key_fini,
 	.lct_exit = tgt_key_exit,
 };
-EXPORT_SYMBOL(tgt_thread_key);
 
 LU_KEY_INIT_GENERIC(tgt);
 
-/* context key constructor/destructor: tgt_ses_key_init, tgt_ses_key_fini */
-LU_KEY_INIT_FINI(tgt_ses, struct tgt_session_info);
+static void *tgt_ses_key_init(const struct lu_context *ctx,
+			      struct lu_context_key *key)
+{
+	struct tgt_session_info *session;
+
+	OBD_SLAB_ALLOC_PTR_GFP(session, tgt_session_kmem, GFP_NOFS);
+	if (session == NULL)
+		return ERR_PTR(-ENOMEM);
+
+	return session;
+}
+
+static void tgt_ses_key_fini(const struct lu_context *ctx,
+			     struct lu_context_key *key, void *data)
+{
+	struct tgt_session_info *session = data;
+
+	OBD_SLAB_FREE_PTR(session, tgt_session_kmem);
+}
 
 /* context key: tgt_session_key */
 struct lu_context_key tgt_session_key = {
@@ -208,25 +481,37 @@ struct page *tgt_page_to_corrupt;
 
 int tgt_mod_init(void)
 {
+	int	result;
 	ENTRY;
 
-	tgt_page_to_corrupt = alloc_page(GFP_IOFS);
+	result = lu_kmem_init(tgt_caches);
+	if (result != 0)
+		RETURN(result);
+
+	tgt_page_to_corrupt = alloc_page(GFP_KERNEL);
 
 	tgt_key_init_generic(&tgt_thread_key, NULL);
 	lu_context_key_register_many(&tgt_thread_key, NULL);
 
 	tgt_ses_key_init_generic(&tgt_session_key, NULL);
 	lu_context_key_register_many(&tgt_session_key, NULL);
+	barrier_init();
+
+	update_info_init();
 
 	RETURN(0);
 }
 
 void tgt_mod_exit(void)
 {
+	barrier_fini();
 	if (tgt_page_to_corrupt != NULL)
-		page_cache_release(tgt_page_to_corrupt);
+		put_page(tgt_page_to_corrupt);
 
 	lu_context_key_degister(&tgt_thread_key);
 	lu_context_key_degister(&tgt_session_key);
+	update_info_fini();
+
+	lu_kmem_fini(tgt_caches);
 }