[Devel] [PATCH VZ9 09/20] fuse: a protocol to reenable optimizations after replication finished

Alexey Kuznetsov kuznet at virtuozzo.com
Fri Oct 6 13:43:07 MSK 2023


When a replication starts we have to disable some optimizations,
including kernel cs offload and fanout submissions. The problem is that
our protocol does not notify clients and other CSes when replication
finishes and optimizations could be reenabled.

This patch suggests some protocol to allow to recover disabled
optimizations. The idea is to add new flags (PCS_CS_IO_CLEAR)
to write messages. If this write passes through replicating CS,
it clears the flag. Then all the nodes, both client and other
cses in chain, can see this flag in responses. If it is set
this means replication is actually finished.

Well, this is not a good idea to add such things to data path,
but it is small thing. Advantage: we do not have to load mds
with additional expensive protocol details.

Signed-off-by: Alexey Kuznetsov <kuznet at acronis.com>
---
 fs/fuse/kio/pcs/fuse_ktrace_prot.h |  2 +-
 fs/fuse/kio/pcs/pcs_cs.c           | 14 +++++++-------
 fs/fuse/kio/pcs/pcs_cs_accel.c     |  6 +++++-
 fs/fuse/kio/pcs/pcs_map.c          | 26 ++++++++++++++++++++++----
 fs/fuse/kio/pcs/pcs_map.h          |  6 ++++--
 5 files changed, 39 insertions(+), 15 deletions(-)

diff --git a/fs/fuse/kio/pcs/fuse_ktrace_prot.h b/fs/fuse/kio/pcs/fuse_ktrace_prot.h
index dc1ade9..3c985f9 100644
--- a/fs/fuse/kio/pcs/fuse_ktrace_prot.h
+++ b/fs/fuse/kio/pcs/fuse_ktrace_prot.h
@@ -36,7 +36,7 @@ struct fuse_tr_iotimes_hdr
 	__u16	type;
 	__u8	cses;
 	__u8	__pad;
-	__u32	__pad1;
+	__u32	chid;
 };
 
 struct fuse_tr_iotimes_cs
diff --git a/fs/fuse/kio/pcs/pcs_cs.c b/fs/fuse/kio/pcs/pcs_cs.c
index 59b752d..d298be9 100644
--- a/fs/fuse/kio/pcs/pcs_cs.c
+++ b/fs/fuse/kio/pcs/pcs_cs.c
@@ -299,11 +299,7 @@ void cs_log_io_times(struct pcs_int_request * ireq, struct pcs_msg * resp, unsig
 			struct pcs_int_request * parent = ireq->iochunk.parent_N;
 
 			parent->iochunk.fo.io_times[idx].csid = resp->rpc->peer_id.val;
-			/* XXX kio does not implement flow detection (for now) and does
-			 * not use flag PCS_CS_IO_SEQ. So, use it here to indicate
-			 * performed fanout.
-			 */
-			parent->iochunk.fo.io_times[idx].misc = h->sync.misc | PCS_CS_IO_SEQ;
+			parent->iochunk.fo.io_times[idx].misc = h->sync.misc;
 			parent->iochunk.fo.io_times[idx].ts_net = h->sync.ts_net;
 			parent->iochunk.fo.io_times[idx].ts_io = h->sync.ts_io;
 
@@ -340,6 +336,8 @@ void cs_log_io_times(struct pcs_int_request * ireq, struct pcs_msg * resp, unsig
 			th->ino = ireq->dentry->fileinfo.attr.id;
 			th->type = h->hdr.type;
 			th->cses = 1;
+			th->__pad = 0;
+			th->chid = (unsigned int)h->uid;
 
 			ch->csid = resp->rpc->peer_id.val;
 			ch->misc = h->sync.misc;
@@ -765,7 +763,9 @@ static void do_cs_submit(struct pcs_cs *cs, struct pcs_int_request *ireq)
 	if ((ireq->dentry->fileinfo.attr.attrib & PCS_FATTR_IMMEDIATE_WRITE) || ireq->dentry->no_write_delay)
 		ioh->sync.misc |= PCS_CS_IO_SYNC;
 	if (ireq->flags & IREQ_F_FANOUT)
-		ioh->sync.misc = PCS_CS_IO_FANOUT;
+		ioh->sync.misc |= PCS_CS_IO_FANOUT;
+	if (pcs_cs_fanout(storage_version))
+		ioh->sync.misc |= PCS_CS_IO_CLEAR;
 
 	msg->size = ioh->hdr.len;
 	msg->rpc = NULL;
@@ -817,7 +817,7 @@ static inline int eligible_for_fanout(struct pcs_int_request * ireq)
 	return (cs_enable_fanout && pcs_cs_fanout(atomic_read(&ireq->cc->storage_version)) &&
 		ireq->iochunk.csl->nsrv <= PCS_MAP_MAX_FO_CS &&
 		ireq->iochunk.cs_index + 1 < ireq->iochunk.csl->nsrv &&
-		!(ireq->iochunk.csl->flags & CS_FL_REPLICATING));
+		!test_bit(CSL_SF_HAS_REPLICATING, &ireq->iochunk.csl->state_flags));
 }
 
 void pcs_cs_submit(struct pcs_cs *cs, struct pcs_int_request *ireq)
diff --git a/fs/fuse/kio/pcs/pcs_cs_accel.c b/fs/fuse/kio/pcs/pcs_cs_accel.c
index 3b75571..5d44862 100644
--- a/fs/fuse/kio/pcs/pcs_cs_accel.c
+++ b/fs/fuse/kio/pcs/pcs_cs_accel.c
@@ -415,6 +415,8 @@ static void __pcs_csa_final_completion(struct pcs_aio_req *areq)
 				th->ino = ireq->dentry->fileinfo.attr.id;
 				th->type = PCS_CS_READ_RESP;
 				th->cses = 1;
+				th->__pad = 0;
+				th->chid = (unsigned int)ireq->iochunk.map->id;
 
 				ch->csid = ireq->iochunk.csl->cs[ireq->iochunk.cs_index].info.id.val | PCS_NODE_ALT_MASK;
 				ch->misc = ktime_to_us(ireq->ts_sent);
@@ -756,7 +758,6 @@ static void __complete_acr_work(struct work_struct * w)
 
 		WARN_ON(!(ireq->flags & IREQ_F_FANOUT));
 		parent->iochunk.fo.io_times[idx] = ireq->iochunk.acr.io_times[idx];
-		parent->iochunk.fo.io_times[idx].misc |= PCS_CS_IO_SEQ;
 	} else {
 		struct fuse_conn * fc = container_of(ireq->cc, struct pcs_fuse_cluster, cc)->fc;
 
@@ -782,6 +783,8 @@ static void __complete_acr_work(struct work_struct * w)
 				th->ino = ireq->dentry->fileinfo.attr.id;
 				th->type = PCS_CS_WRITE_AL_RESP;
 				th->cses = n;
+				th->__pad = 0;
+				th->chid = (unsigned int)ireq->iochunk.map->id;
 
 				for (i = 0; i < n; i++, ch++)
 					*ch = ireq->iochunk.acr.io_times[i];
@@ -818,6 +821,7 @@ static void __pcs_csa_write_final_completion(struct pcs_accel_write_req *areq)
 		th->ts_net = 0;
 		th->ts_io = ktime_to_us(ktime_get()) - th->misc;
 		th->misc &= PCS_CS_TS_MASK;
+		th->misc |= PCS_CS_IO_CLEAR | PCS_CS_IO_FANOUT;
 	}
 
 	csa_complete_acr(ireq);
diff --git a/fs/fuse/kio/pcs/pcs_map.c b/fs/fuse/kio/pcs/pcs_map.c
index 29ed4fa..59833eb 100644
--- a/fs/fuse/kio/pcs/pcs_map.c
+++ b/fs/fuse/kio/pcs/pcs_map.c
@@ -959,7 +959,7 @@ struct pcs_cs_list* cslist_alloc( struct pcs_cs_set *css, struct pcs_cs_info *re
 	atomic_set(&cs_list->refcnt, 1);
 	atomic_set(&cs_list->seq_read_in_flight, 0);
 	cs_list->read_index = -1;
-	cs_list->flags = 0;
+	cs_list->state_flags = 0;
 	cs_list->serno = atomic64_inc_return(&css->csl_serno_gen);
 	cs_list->blacklist = 0;
 	cs_list->read_timeout = (read_tout * HZ) / 1000;
@@ -980,7 +980,7 @@ struct pcs_cs_list* cslist_alloc( struct pcs_cs_set *css, struct pcs_cs_info *re
 
 		if (cs_list->cs[i].info.flags & CS_FL_REPLICATING) {
 			__set_bit(i, &cs_list->blacklist);
-			cs_list->flags |= CS_FL_REPLICATING;
+			set_bit(CSL_SF_HAS_REPLICATING, &cs_list->state_flags);
 			cs_list->blacklist_expires = jiffies + PCS_REPLICATION_BLACKLIST_TIMEOUT;
 		}
 
@@ -1005,7 +1005,7 @@ struct pcs_cs_list* cslist_alloc( struct pcs_cs_set *css, struct pcs_cs_info *re
 		cs->mds_flags = cs_list->cs[i].info.flags;
 		if (cs->mds_flags & CS_FL_LOCAL) {
 			set_bit(CS_SF_LOCAL, &cs->state);
-			cs_list->flags |= CSL_FL_HAS_LOCAL;
+			set_bit(CSL_SF_HAS_LOCAL, &cs_list->state_flags);
 		} else if (test_bit(CS_SF_LOCAL, &cs->state))
 			clear_bit(CS_SF_LOCAL, &cs->state);
 
@@ -1961,7 +1961,7 @@ static int pcs_cslist_submit_read(struct pcs_int_request *ireq, struct pcs_cs_li
 		    (now > selected + PCS_MAP_MIN_REBALANCE_TIMEOUT &&
 		     (!is_seq || get_io_locality(cc) < 0 ||
 		      (!csl_seq &&
-		       !(test_bit(CS_SF_LOCAL, &cs->state)) && (csl->flags & CSL_FL_HAS_LOCAL))))) {
+		       !(test_bit(CS_SF_LOCAL, &cs->state)) && test_bit(CSL_SF_HAS_LOCAL, &csl->state_flags))))) {
 			i = -1;
 			WRITE_ONCE(csl->read_index, -1);
 		}
@@ -2832,6 +2832,24 @@ static int commit_sync_info(struct pcs_int_request *req,
 			if (max_iolat < srec->sync.ts_io)
 				max_iolat = srec->sync.ts_io;
 		}
+
+		/* If we got CLEAR response to a write to csl replication has already stopped */
+		if (test_bit(CSL_SF_HAS_REPLICATING, &csl->state_flags) && (h->sync.misc & PCS_CS_IO_CLEAR) &&
+		    (h->hdr.type == PCS_CS_WRITE_RESP || h->hdr.type == PCS_CS_WRITE_AL_RESP))
+			clear_bit(CSL_SF_HAS_REPLICATING, &csl->state_flags);
+	} else {
+		/* In case we did successful read on would-be replicating CS resync the state */
+		if (test_bit(CSL_SF_HAS_REPLICATING, &csl->state_flags) &&
+		    test_bit(CS_SF_REPLICATING, &csl->cs[req->iochunk.cs_index].cslink.cs->state)) {
+			int idx;
+			clear_bit(CS_SF_REPLICATING, &csl->cs[req->iochunk.cs_index].cslink.cs->state);
+			for (idx = csl->nsrv - 1; idx >= 0; idx++) {
+				if (test_bit(CS_SF_REPLICATING, &csl->cs[idx].cslink.cs->state))
+					break;
+			}
+			if (idx < 0)
+				clear_bit(CSL_SF_HAS_REPLICATING, &csl->state_flags);
+		}
 	}
 	pcs_fuse_stat_io_count(req, resp, max_iolat, net_lat);
 	cs_log_io_times(req, resp, max_iolat);
diff --git a/fs/fuse/kio/pcs/pcs_map.h b/fs/fuse/kio/pcs/pcs_map.h
index 8cc7dfe..6d77138 100644
--- a/fs/fuse/kio/pcs/pcs_map.h
+++ b/fs/fuse/kio/pcs/pcs_map.h
@@ -112,9 +112,11 @@ struct pcs_cs_list
 	unsigned long		blacklist;		/* Atomic bit field */
 	abs_time_t		blacklist_expires;	/* volatile blacklist stamp */
 	abs_time_t		select_stamp;		/* volatile read hint stamp */
+	/* Semi-immutable. No locking, but can be modified unlocked */
+	unsigned long		state_flags;
+#define CSL_SF_HAS_LOCAL	0
+#define CSL_SF_HAS_REPLICATING	1
 	/* members below are immutable accross cslist life time */
-#define CSL_FL_HAS_LOCAL	1
-	unsigned int		flags;
 	u64                     serno;
 	int			read_timeout;
 	int			write_timeout;
-- 
1.8.3.1



More information about the Devel mailing list