diff --git a/server/filestore.go b/server/filestore.go
index 618ffc0f..1f4479e5 100644
--- a/server/filestore.go
+++ b/server/filestore.go
@@ -1916,8 +1916,9 @@ func (fs *fileStore) storeRawMsg(subj string, hdr, msg []byte, seq uint64, ts in
 	}
 
 	// Per subject max check needed.
+	mmp := uint64(fs.cfg.MaxMsgsPer)
 	var psmc uint64
-	psmax := fs.cfg.MaxMsgsPer > 0 && len(subj) > 0
+	psmax := mmp > 0 && len(subj) > 0
 	if psmax {
 		if info, ok := fs.psim[subj]; ok {
 			psmc = info.total
@@ -1928,13 +1929,12 @@ func (fs *fileStore) storeRawMsg(subj string, hdr, msg []byte, seq uint64, ts in
 	// Check if we are discarding new messages when we reach the limit.
 	if fs.cfg.Discard == DiscardNew {
 		var asl bool
-		if psmax && psmc >= uint64(fs.cfg.MaxMsgsPer) {
+		if psmax && psmc >= mmp {
 			// If we are instructed to discard new per subject, this is an error.
 			if fs.cfg.DiscardNewPer {
 				return ErrMaxMsgsPerSubject
 			}
-			fseq, err = fs.firstSeqForSubj(subj)
-			if err != nil {
+			if fseq, err = fs.firstSeqForSubj(subj); err != nil {
 				return err
 			}
 			asl = true
@@ -1990,12 +1990,25 @@ func (fs *fileStore) storeRawMsg(subj string, hdr, msg []byte, seq uint64, ts in
 
 	// Enforce per message limits.
 	// We snapshotted psmc before our actual write, so >= comparison needed.
-	if psmax && psmc >= uint64(fs.cfg.MaxMsgsPer) {
+	if psmax && psmc >= mmp {
 		// We may have done this above.
 		if fseq == 0 {
 			fseq, _ = fs.firstSeqForSubj(subj)
 		}
-		fs.removeMsg(fseq, false, false)
+		if ok, _ := fs.removeMsg(fseq, false, false); ok {
+			// Make sure we are below the limit.
+			if psmc--; psmc >= mmp {
+				for info, ok := fs.psim[subj]; ok && info.total > mmp; info, ok = fs.psim[subj] {
+					if seq, _ := fs.firstSeqForSubj(subj); seq > 0 {
+						if ok, _ := fs.removeMsg(seq, false, false); !ok {
+							break
+						}
+					} else {
+						break
+					}
+				}
+			}
+		}
 	}
 
 	// Limits checks and enforcement.
@@ -5317,23 +5330,29 @@ func (mb *msgBlock) removeSeqPerSubject(subj string, seq uint64, smp *StoreMsg)
 		return
 	}
 
-	// Here what we are removing is the first message.
-	// If we only have one message left we can simply assign it to last.
+	// Only one left.
 	if ss.Msgs == 1 {
-		ss.First = ss.Last
+		if seq != ss.First {
+			ss.Last = ss.First
+		} else {
+			ss.First = ss.Last
+		}
 		return
 	}
 
+	// Recalculate first.
 	// TODO(dlc) - Might want to optimize this.
-	var smv StoreMsg
-	if smp == nil {
-		smp = &smv
-	}
-	for tseq := seq + 1; tseq <= ss.Last; tseq++ {
-		if sm, _ := mb.cacheLookup(tseq, smp); sm != nil {
-			if sm.subj == subj {
-				ss.First = tseq
-				return
+	if seq == ss.First {
+		var smv StoreMsg
+		if smp == nil {
+			smp = &smv
+		}
+		for tseq := seq + 1; tseq <= ss.Last; tseq++ {
+			if sm, _ := mb.cacheLookup(tseq, smp); sm != nil {
+				if sm.subj == subj {
+					ss.First = tseq
+					return
+				}
 			}
 		}
 	}
@@ -5443,9 +5462,6 @@ func (mb *msgBlock) loadPerSubjectInfo() ([]byte, error) {
 // Helper to make sure fss loaded if we are tracking.
 // Lock should be held
 func (mb *msgBlock) ensurePerSubjectInfoLoaded() error {
-	// Clear
-	mb.fssNeedsWrite = false
-
 	if mb.fss != nil || mb.noTrack {
 		return nil
 	}
@@ -5536,6 +5552,7 @@ func (mb *msgBlock) readPerSubjectInfo(hasLock bool) error {
 		fss[subj] = &SimpleState{Msgs: msgs, First: first, Last: last}
 	}
 	mb.fss = fss
+	mb.fssNeedsWrite = false
 
 	// Make sure we run the cache expire timer.
 	if len(mb.fss) > 0 {
@@ -5613,10 +5630,11 @@ func (mb *msgBlock) close(sync bool) {
 	}
 
 	// Check if we are tracking by subject.
-	if mb.fss != nil {
+	if len(mb.fss) > 0 && mb.fssNeedsWrite {
 		mb.writePerSubjectInfo()
-		mb.fss = nil
 	}
+	mb.fss = nil
+	mb.fssNeedsWrite = false
 
 	// Close cache
 	mb.clearCacheAndOffset()
diff --git a/server/norace_test.go b/server/norace_test.go
index 104af425..9db3c5d0 100644
--- a/server/norace_test.go
+++ b/server/norace_test.go
@@ -23,6 +23,7 @@ import (
 	"context"
 	"encoding/binary"
 	"encoding/json"
+	"errors"
 	"fmt"
 	"io"
 	"math/rand"
@@ -6273,3 +6274,80 @@ func TestNoRaceJetStreamClusterConsumerInfoSpeed(t *testing.T) {
 
 	checkNumPending(toKeep)
 }
+
+func TestNoRaceJetStreamKVAccountWithServerRestarts(t *testing.T) {
+	c := createJetStreamClusterExplicit(t, "R3S", 3)
+	defer c.shutdown()
+
+	nc, js := jsClientConnect(t, c.randomServer())
+	defer nc.Close()
+
+	_, err := js.CreateKeyValue(&nats.KeyValueConfig{
+		Bucket:   "TEST",
+		Replicas: 3,
+	})
+	require_NoError(t, err)
+
+	npubs := 10_000
+	par := 8
+	iter := 2
+	nsubjs := 250
+
+	wg := sync.WaitGroup{}
+	putKeys := func() {
+		wg.Add(1)
+		go func() {
+			defer wg.Done()
+			nc, js := jsClientConnect(t, c.randomServer())
+			defer nc.Close()
+			kv, err := js.KeyValue("TEST")
+			require_NoError(t, err)
+
+			for i := 0; i < npubs; i++ {
+				subj := fmt.Sprintf("KEY-%d", rand.Intn(nsubjs))
+				if _, err := kv.PutString(subj, "hello"); err != nil {
+					nc, js := jsClientConnect(t, c.randomServer())
+					defer nc.Close()
+					kv, err = js.KeyValue("TEST")
+					require_NoError(t, err)
+				}
+			}
+		}()
+	}
+
+	restartServers := func() {
+		time.Sleep(2 * time.Second)
+		// Rotate through and restart the servers.
+		for _, server := range c.servers {
+			server.Shutdown()
+			restarted := c.restartServer(server)
+			checkFor(t, time.Second, 200*time.Millisecond, func() error {
+				hs := restarted.healthz(&HealthzOptions{
+					JSEnabled:    true,
+					JSServerOnly: true,
+				})
+				if hs.Error != _EMPTY_ {
+					return errors.New(hs.Error)
+				}
+				return nil
+			})
+		}
+		c.waitOnLeader()
+		c.waitOnStreamLeader(globalAccountName, "KV_TEST")
+	}
+
+	for n := 0; n < iter; n++ {
+		for i := 0; i < par; i++ {
+			putKeys()
+		}
+		restartServers()
+	}
+	wg.Wait()
+
+	nc, js = jsClientConnect(t, c.randomServer())
+	defer nc.Close()
+
+	si, err := js.StreamInfo("KV_TEST")
+	require_NoError(t, err)
+	require_True(t, si.State.NumSubjects == uint64(nsubjs))
+}