DAOS-14317 vos: initial changes for the phase2 object pre-load (#15001)

- Introduced new durable format 'vos_obj_p2_df' for the md-on-ssd phase2
  object, at most 4 evict-able bucket IDs could be stored.

- Changed vos_obj_hold() & vos_obj_release() to pin or unpin object
  respectively.

- Changed the private data of VOS dkey/akey/value trees from 'vos_pool' to
  'vos_object', the private data will be used for allocating/reserving from
  the evict-able bucket.

- Move the vos_obj_hold() call from vos_update_end() to vos_update_begin()
  for the phase2 pool, reserve value from the object evict-able bucket.

TODO:
- Reorg GC code to pre-load objects before starting transaction.
- Reorg DTX commit code to pre-load objects before starting transaction.
- Reorg multiple-objects tx code to pre-load objects.
- Improve engine scheduler to take both non-evict-able & evict-able zones
  space pressure into account.

Signed-off-by: Niu Yawei <[email protected]>

src/common/btree.c

@@ -945,8 +945,12 @@ btr_root_alloc(struct btr_context *tcx)
 	struct btr_instance	*tins = &tcx->tc_tins;
 	struct btr_root		*root;
 
-	tins->ti_root_off = umem_zalloc(btr_umm(tcx),
-					sizeof(struct btr_root));
+	if (btr_ops(tcx)->to_node_alloc != NULL)
+		tins->ti_root_off = btr_ops(tcx)->to_node_alloc(&tcx->tc_tins,
+								sizeof(struct btr_root));
+	else
+		tins->ti_root_off = umem_zalloc(btr_umm(tcx), sizeof(struct btr_root));
+
 	if (UMOFF_IS_NULL(tins->ti_root_off))
 		return btr_umm(tcx)->umm_nospc_rc;
 
@@ -3884,6 +3888,7 @@ btr_tree_destroy(struct btr_context *tcx, void *args, bool *destroyed)
 		tcx->tc_tins.ti_root_off, tcx->tc_order);
 
 	root = tcx->tc_tins.ti_root;
+	tcx->tc_tins.ti_destroy = 1;
 	if (root && !UMOFF_IS_NULL(root->tr_node)) {
 		/* destroy the root and all descendants */
 		rc = btr_node_destroy(tcx, root->tr_node, args, &empty);

src/include/daos/btree.h

@@ -429,6 +429,8 @@ struct btr_instance {
 	struct btr_root			*ti_root;
 	/** Customized operations for the tree */
 	btr_ops_t			*ti_ops;
+	/** The context is used for tree destroy */
+	unsigned int			 ti_destroy : 1;
 };
 
 /**

src/include/daos/mem.h

@@ -451,8 +451,6 @@ typedef void
 umem_cache_wait_cb_t(void *arg, uint64_t chkpt_tx, uint64_t *committed_tx);
 
 /**
- * Write all dirty pages before @wal_tx to MD blob. (XXX: not yet implemented)
- *
  * This function can yield internally, it is called by checkpoint service of upper level stack.
  *
  * \param[in]		store		The umem store

src/include/daos_srv/evtree.h

@@ -1,5 +1,5 @@
 /**
- * (C) Copyright 2017-2023 Intel Corporation.
+ * (C) Copyright 2017-2024 Intel Corporation.
  *
  * SPDX-License-Identifier: BSD-2-Clause-Patent
  */
@@ -70,6 +70,10 @@ struct evt_desc_cbs {
 					  struct evt_desc *desc,
 					  daos_size_t nob, void *args);
 	void		 *dc_bio_free_args;
+	/**
+	 * Argument for allocation.
+	 */
+	void		 *dc_alloc_arg;
 	/**
 	 * Availability check, it is for data tracked by DTX undo log.
 	 * It is optional, EVTree always treats data extent is available if

src/tests/ftest/util/telemetry_utils.py

@@ -424,7 +424,7 @@ class TelemetryUtils():
         ENGINE_NVME_INTEL_VENDOR_METRICS
     ENGINE_MEM_USAGE_METRICS = [
         "engine_mem_vos_dtx_cmt_ent_48",
-        "engine_mem_vos_vos_obj_360",
+        "engine_mem_vos_vos_obj_384",
         "engine_mem_vos_vos_lru_size",
         "engine_mem_dtx_dtx_leader_handle_360"]
     ENGINE_MEM_TOTAL_USAGE_METRICS = [

src/vos/evtree.c

@@ -1443,8 +1443,9 @@ evt_node_alloc(struct evt_context *tcx, unsigned int flags,
 	struct evt_node         *nd;
 	umem_off_t		 nd_off;
 	bool                     leaf = (flags & EVT_NODE_LEAF);
+	struct vos_object	*obj = tcx->tc_desc_cbs.dc_alloc_arg;
 
-	nd_off = umem_zalloc(evt_umm(tcx), evt_node_size(tcx, leaf));
+	nd_off = vos_obj_alloc(evt_umm(tcx), obj, evt_node_size(tcx, leaf), true);
 	if (UMOFF_IS_NULL(nd_off))
 		return -DER_NOSPACE;
 
@@ -3249,8 +3250,9 @@ evt_common_insert(struct evt_context *tcx, struct evt_node *nd,
 	}
 
 	if (leaf) {
-		umem_off_t desc_off;
-		uint32_t   csum_buf_size = 0;
+		umem_off_t		 desc_off;
+		uint32_t		 csum_buf_size = 0;
+		struct vos_object	*obj = tcx->tc_desc_cbs.dc_alloc_arg;
 
 		if (ci_is_valid(&ent->ei_csum))
 			csum_buf_size = ci_csums_len(ent->ei_csum);
@@ -3263,7 +3265,7 @@ evt_common_insert(struct evt_context *tcx, struct evt_node *nd,
 			D_DEBUG(DB_TRACE, "Allocating an extra %d bytes "
 						"for checksum", csum_buf_size);
 		}
-		desc_off = umem_zalloc(evt_umm(tcx), desc_size);
+		desc_off = vos_obj_alloc(evt_umm(tcx), obj, desc_size, true);
 		if (UMOFF_IS_NULL(desc_off))
 			return -DER_NOSPACE;
 

src/vos/tests/vts_io.c

@@ -898,7 +898,7 @@ io_update_and_fetch_dkey(struct io_test_args *arg, daos_epoch_t update_epoch,
 static inline int
 hold_obj(struct vos_container *cont, daos_unit_oid_t oid, daos_epoch_range_t *epr,
 	 daos_epoch_t bound, uint64_t flags, uint32_t intent, struct vos_object **obj_p,
-	 struct vos_ts_set *ts_set)
+	 struct vos_ts_set *ts_set, struct umem_instance *umm)
 {
 	int rc;
 
@@ -908,7 +908,16 @@ hold_obj(struct vos_container *cont, daos_unit_oid_t oid, daos_epoch_range_t *ep
 
 	if (flags & VOS_OBJ_CREATE) {
 		assert_ptr_not_equal(*obj_p, NULL);
+
+		if (umm != NULL) {
+			rc = umem_tx_begin(umm, NULL);
+			assert_rc_equal(rc, 0);
+		}
+
 		rc = vos_obj_incarnate(*obj_p, epr, bound, flags, intent, ts_set);
+
+		if (umm != NULL)
+			rc = umem_tx_end(umm, rc);
 	}
 
 	return rc;
@@ -926,7 +935,8 @@ hold_objects(struct vos_object **objs, daos_handle_t *coh, daos_unit_oid_t *oid,
 	hold_flags |= VOS_OBJ_VISIBLE;
 	for (i = start; i < end; i++) {
 		rc = hold_obj(vos_hdl2cont(*coh), *oid, &epr, 0, hold_flags,
-			      no_create ? DAOS_INTENT_DEFAULT : DAOS_INTENT_UPDATE, &objs[i], 0);
+			      no_create ? DAOS_INTENT_DEFAULT : DAOS_INTENT_UPDATE,
+			      &objs[i], 0, NULL);
 		if (rc != exp_rc)
 			return 1;
 	}
@@ -1006,82 +1016,72 @@ io_obj_cache_test(void **state)
 
 	ummg = vos_cont2umm(vos_hdl2cont(ctx->tc_co_hdl));
 	umml = vos_cont2umm(vos_hdl2cont(l_coh));
-	rc = umem_tx_begin(ummg, NULL);
-	assert_rc_equal(rc, 0);
 
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0,
-		      VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &objs[0], 0);
+		      VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &objs[0], 0, ummg);
 	assert_rc_equal(rc, 0);
 
 	/** Hold object for discard */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_DISCARD,
-		      DAOS_INTENT_DISCARD, &obj1, 0);
+		      DAOS_INTENT_DISCARD, &obj1, 0, ummg);
 	assert_rc_equal(rc, 0);
 	/** Second discard should fail */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_DISCARD,
-		      DAOS_INTENT_DISCARD, &obj2, 0);
+		      DAOS_INTENT_DISCARD, &obj2, 0, ummg);
 	assert_rc_equal(rc, -DER_BUSY);
 	/** Should prevent simultaneous aggregation */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_AGGREGATE,
-		     DAOS_INTENT_PURGE, &obj2, 0);
+		     DAOS_INTENT_PURGE, &obj2, 0, ummg);
 	assert_rc_equal(rc, -DER_BUSY);
 	/** Should prevent simultaneous hold for create as well */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0,
-				   VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &obj2, 0);
+				   VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &obj2,
+				   0, ummg);
 	assert_rc_equal(rc, -DER_UPDATE_AGAIN);
 
 	/** Need to be able to hold for read though or iteration won't work */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_VISIBLE,
-		      DAOS_INTENT_DEFAULT, &obj2, 0);
+		      DAOS_INTENT_DEFAULT, &obj2, 0, ummg);
 	vos_obj_release(obj2, 0, false);
 	vos_obj_release(obj1, VOS_OBJ_DISCARD, false);
 
 	/** Hold object for aggregation */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_AGGREGATE,
-		      DAOS_INTENT_PURGE, &obj1, 0);
+		      DAOS_INTENT_PURGE, &obj1, 0, ummg);
 	assert_rc_equal(rc, 0);
 	/** Discard should fail */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_DISCARD,
-		      DAOS_INTENT_DISCARD, &obj2, 0);
+		      DAOS_INTENT_DISCARD, &obj2, 0, ummg);
 	assert_rc_equal(rc, -DER_BUSY);
 	/** Second aggregation should fail */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_AGGREGATE,
-		      DAOS_INTENT_PURGE, &obj2, 0);
+		      DAOS_INTENT_PURGE, &obj2, 0, ummg);
 	assert_rc_equal(rc, -DER_BUSY);
 	/** Simultaneous create should work */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0,
-		      VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &obj2, 0);
+		      VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &obj2, 0, ummg);
 	assert_rc_equal(rc, 0);
 	vos_obj_release(obj2, 0, false);
 
 	/** Need to be able to hold for read though or iteration won't work */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_VISIBLE,
-		      DAOS_INTENT_DEFAULT, &obj2, 0);
+		      DAOS_INTENT_DEFAULT, &obj2, 0, ummg);
 	vos_obj_release(obj2, 0, false);
 	vos_obj_release(obj1, VOS_OBJ_AGGREGATE, false);
 
 	/** Now that other one is done, this should work */
 	rc = hold_obj(vos_hdl2cont(ctx->tc_co_hdl), oids[0], &epr, 0, VOS_OBJ_DISCARD,
-		      DAOS_INTENT_DISCARD, &obj2, 0);
+		      DAOS_INTENT_DISCARD, &obj2, 0, ummg);
 	assert_rc_equal(rc, 0);
 	vos_obj_release(obj2, VOS_OBJ_DISCARD, false);
 
-	rc = umem_tx_end(ummg, 0);
-	assert_rc_equal(rc, 0);
-
 	vos_obj_release(objs[0], 0, false);
 
-	rc = umem_tx_begin(umml, NULL);
-	assert_rc_equal(rc, 0);
-
 	rc = hold_obj(vos_hdl2cont(l_coh), oids[1], &epr, 0,
-		      VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &objs[0], 0);
+		      VOS_OBJ_CREATE | VOS_OBJ_VISIBLE, DAOS_INTENT_UPDATE, &objs[0], 0, umml);
 	assert_rc_equal(rc, 0);
 	vos_obj_release(objs[0], 0, false);
 
-	rc = umem_tx_end(umml, 0);
-	assert_rc_equal(rc, 0);
-
 	rc = hold_objects(objs, &ctx->tc_co_hdl, &oids[0], 0, 10, true, 0);
 	assert_int_equal(rc, 0);
 
@@ -1091,7 +1091,7 @@ io_obj_cache_test(void **state)
 	rc = hold_objects(objs, &l_coh, &oids[1], 10, 15, true, 0);
 	assert_int_equal(rc, 0);
 	rc = hold_obj(vos_hdl2cont(l_coh), oids[1], &epr, 0, VOS_OBJ_VISIBLE,
-		      DAOS_INTENT_DEFAULT, &objs[16], 0);
+		      DAOS_INTENT_DEFAULT, &objs[16], 0, NULL);
 	assert_rc_equal(rc, 0);
 
 	vos_obj_release(objs[16], 0, false);

src/vos/vos_aggregate.c

@@ -984,7 +984,7 @@ reserve_segment(struct vos_object *obj, struct agg_io_context *io,
 
 	if (vos_io_scm(vos_obj2pool(obj), DAOS_IOD_ARRAY, size, VOS_IOS_AGGREGATION)) {
 		/** Store on SCM */
-		off = vos_reserve_scm(obj->obj_cont, io->ic_rsrvd_scm, size);
+		off = vos_reserve_scm(obj->obj_cont, io->ic_rsrvd_scm, size, obj);
 		if (UMOFF_IS_NULL(off)) {
 			now = daos_gettime_coarse();
 			if (now - obj->obj_cont->vc_agg_nospc_ts > VOS_NOSPC_ERROR_INTVL) {

src/vos/vos_gc.c

@@ -1,5 +1,5 @@
 /**
- * (C) Copyright 2019-2023 Intel Corporation.
+ * (C) Copyright 2019-2024 Intel Corporation.
  *
  * SPDX-License-Identifier: BSD-2-Clause-Patent
  */
@@ -115,7 +115,7 @@ gc_drain_evt(struct vos_gc *gc, struct vos_pool *pool, daos_handle_t coh,
 	daos_handle_t	    toh;
 	int		    rc;
 
-	vos_evt_desc_cbs_init(&cbs, pool, coh);
+	vos_evt_desc_cbs_init(&cbs, pool, coh, NULL);
 	rc = evt_open(root, &pool->vp_uma, &cbs, &toh);
 	if (rc == -DER_NONEXIST) {
 		*empty = true;

src/vos/vos_internal.h

@@ -1258,7 +1258,7 @@ vos_bio_addr_free(struct vos_pool *pool, bio_addr_t *addr, daos_size_t nob);
 
 void
 vos_evt_desc_cbs_init(struct evt_desc_cbs *cbs, struct vos_pool *pool,
-		      daos_handle_t coh);
+		      daos_handle_t coh, struct vos_object *obj);
 
 int
 vos_tx_begin(struct dtx_handle *dth, struct umem_instance *umm, bool is_sysdb);
@@ -1312,7 +1312,7 @@ vos_dedup_invalidate(struct vos_pool *pool);
 
 umem_off_t
 vos_reserve_scm(struct vos_container *cont, struct umem_rsrvd_act *rsrvd_scm,
-		daos_size_t size);
+		daos_size_t size, struct vos_object *obj);
 int
 vos_publish_scm(struct umem_instance *umm, struct umem_rsrvd_act *rsrvd_scm, bool publish);
 int
@@ -1329,6 +1329,12 @@ vos_pool2umm(struct vos_pool *pool)
 	return &pool->vp_umm;
 }
 
+static inline struct umem_store *
+vos_pool2store(struct vos_pool *pool)
+{
+	return &pool->vp_umm.umm_pool->up_store;
+}
+
 static inline struct umem_instance *
 vos_cont2umm(struct vos_container *cont)
 {
@@ -1844,4 +1850,42 @@ vos_io_scm(struct vos_pool *pool, daos_iod_type_t type, daos_size_t size, enum v
 int
 vos_insert_oid(struct dtx_handle *dth, struct vos_container *cont, daos_unit_oid_t *oid);
 
+static inline bool
+vos_pool_is_p2(struct vos_pool *pool)
+{
+	struct umem_store	*store = vos_pool2store(pool);
+
+	return store->store_type == DAOS_MD_BMEM_V2;
+}
+
+static inline umem_off_t
+vos_obj_alloc(struct umem_instance *umm, struct vos_object *obj, size_t size, bool zeroing)
+{
+
+	if (obj != NULL && vos_pool_is_p2(vos_obj2pool(obj))) {
+		D_ASSERT(obj->obj_bkt_allot == 1);
+		if (zeroing)
+			return umem_zalloc_from_bucket(umm, size, obj->obj_bkt_ids[0]);
+
+		return umem_alloc_from_bucket(umm, size, obj->obj_bkt_ids[0]);
+	}
+
+	if (zeroing)
+		return umem_zalloc(umm, size);
+
+	return umem_alloc(umm, size);
+}
+
+static inline umem_off_t
+vos_obj_reserve(struct umem_instance *umm, struct vos_object *obj,
+		struct umem_rsrvd_act *rsrvd_scm, daos_size_t size)
+{
+	if (obj != NULL && vos_pool_is_p2(vos_obj2pool(obj))) {
+		D_ASSERT(obj->obj_bkt_allot == 1);
+		return umem_reserve_from_bucket(umm, rsrvd_scm, size, obj->obj_bkt_ids[0]);
+	}
+
+	return umem_reserve(umm, rsrvd_scm, size);
+}
+
 #endif /* __VOS_INTERNAL_H__ */