[PATCH 0/4] NFSv4.1 DRC Rewrite Version 7

[PATCH 0/4] NFSv4.1 DRC Rewrite Version 7

[andros]

NFSv4.1 DRC Rewrite Version 7

These patches apply against git://linux-nfs.org/~bfields/linux: for-2.6.32
branch and continue to rewrite the NFSv4.1 Sessions DRC.

Responded to comments on Version 6.

1) The bound on the fore channel per-session DRC size are rewritten.
Instead of just using a maximum number of slots to bound the size,
The server gives the client the number of ca_maxresponsesize_cached slots it
requests bounded by NFSD_SLOT_CACHE_SIZE, NFSD_MAX_MEM_PER_SESSION and by
nfsd_drc_max_mem. Do not allow more than NFSD_MAX_SLOTS_PER_SESSION.

This allows clients to tailor a session to usage.
For example, an I/O session (READ/WRITE/COMMIT only) can have a much smaller
ca_maxresponsesize_cached (for only WRITE/COMMIt compound responses) and a lot
larger ca_maxresponses to service a large in-flight data window.

2) the page-based DRC is replaced with a buffer based DRC with each
slot table entry (struct nfsd4_slot + cache) allocated separately.
This allocation prepares us for slot size re-negotiation via the SEQUENCE
operation target and high slot id arguments.

Testing:

NFSv4.1 mount: pynfs tests - including the SEQUENCE replay cache tests.
connectathon tests.

NFSv4.0 mount: connectathon tests.

-->Andy

0001-nfsd41-bound-forechannel-drc-size-by-memory-usage.patch
0002-nfsd41-use-session-maxreqs-for-sequence-target-and.patch
0003-nfsd41-replace-nfserr_resource-in-pure-nfs41-respon.patch
0004-nfsd41-replace-page-based-DRC-with-buffer-based-DRC.patch

[PATCH 1/4] nfsd41: bound forechannel drc size by memory usage

[andros]

From: Andy Adamson <andros@netapp.com>

By using the requested ca_maxresponsesize_cached * ca_maxresponses to bound
a forechannel drc request size, clients can tailor a session to usage.

For example, an I/O session (READ/WRITE only) can have a much smaller
ca_maxresponsesize_cached (for only WRITE compound responses) and a lot larger
ca_maxresponses to service a large in-flight data window.

Signed-off-by: Andy Adamson <andros@netapp.com>
---
 fs/nfsd/nfs4state.c        |   66 ++++++++++++++++++++++++++++++++------------
 include/linux/nfsd/state.h |    8 ++++-
 2 files changed, 54 insertions(+), 20 deletions(-)

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index b44a2cf..02b3ddd 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -414,34 +414,64 @@ gen_sessionid(struct nfsd4_session *ses)
 }
 
 /*
- * Give the client the number of slots it requests bound by
- * NFSD_MAX_SLOTS_PER_SESSION and by nfsd_drc_max_mem.
+ * The protocol defines ca_maxresponssize_cached to include the size of
+ * the rpc header, but all we need to cache is the data starting after
+ * the end of the initial SEQUENCE operation--the rest we regenerate
+ * each time.  Therefore we can advertise a ca_maxresponssize_cached
+ * value that is the number of bytes in our cache plus a few additional
+ * bytes.  In order to stay on the safe side, and not promise more than
+ * we can cache, those additional bytes must be the minimum possible: 24
+ * bytes of rpc header (xid through accept state, with AUTH_NULL
+ * verifier), 12 for the compound header (with zero-length tag), and 44
+ * for the SEQUENCE op response:
+ */
+#define NFSD_MIN_HDR_SEQ_SZ  (24 + 12 + 44)
+
+/*
+ * Give the client the number of ca_maxresponsesize_cached slots it
+ * requests, of size bounded by NFSD_SLOT_CACHE_SIZE,
+ * NFSD_MAX_MEM_PER_SESSION, and nfsd_drc_max_mem. Do not allow more
+ * than NFSD_MAX_SLOTS_PER_SESSION.
  *
- * If we run out of reserved DRC memory we should (up to a point) re-negotiate
- * active sessions and reduce their slot usage to make rooom for new
- * connections. For now we just fail the create session.
+ * If we run out of reserved DRC memory we should (up to a point)
+ * re-negotiate active sessions and reduce their slot usage to make
+ * rooom for new connections. For now we just fail the create session.
  */
-static int set_forechannel_maxreqs(struct nfsd4_channel_attrs *fchan)
+static int set_forechannel_drc_size(struct nfsd4_channel_attrs *fchan)
 {
-	int mem;
+	int mem, size = fchan->maxresp_cached;
 
 	if (fchan->maxreqs < 1)
 		return nfserr_inval;
-	else if (fchan->maxreqs > NFSD_MAX_SLOTS_PER_SESSION)
-		fchan->maxreqs = NFSD_MAX_SLOTS_PER_SESSION;
 
-	mem = fchan->maxreqs * NFSD_SLOT_CACHE_SIZE;
+	if (size < NFSD_MIN_HDR_SEQ_SZ)
+		size = NFSD_MIN_HDR_SEQ_SZ;
+	size -= NFSD_MIN_HDR_SEQ_SZ;
+	if (size > NFSD_SLOT_CACHE_SIZE)
+		size = NFSD_SLOT_CACHE_SIZE;
+
+	/* bound the maxreqs by NFSD_MAX_MEM_PER_SESSION */
+	mem = fchan->maxreqs * size;
+	if (mem > NFSD_MAX_MEM_PER_SESSION) {
+		fchan->maxreqs = NFSD_MAX_MEM_PER_SESSION / size;
+		if (fchan->maxreqs > NFSD_MAX_SLOTS_PER_SESSION)
+			fchan->maxreqs = NFSD_MAX_SLOTS_PER_SESSION;
+		mem = fchan->maxreqs * size;
+	}
 
 	spin_lock(&nfsd_drc_lock);
-	if (mem + nfsd_drc_mem_used > nfsd_drc_max_mem)
-		mem = ((nfsd_drc_max_mem - nfsd_drc_mem_used) /
-				NFSD_SLOT_CACHE_SIZE) * NFSD_SLOT_CACHE_SIZE;
+	/* bound the total session drc memory ussage */
+	if (mem + nfsd_drc_mem_used > nfsd_drc_max_mem) {
+		fchan->maxreqs = (nfsd_drc_max_mem - nfsd_drc_mem_used) / size;
+		mem = fchan->maxreqs * size;
+	}
 	nfsd_drc_mem_used += mem;
 	spin_unlock(&nfsd_drc_lock);
 
-	fchan->maxreqs = mem / NFSD_SLOT_CACHE_SIZE;
 	if (fchan->maxreqs == 0)
 		return nfserr_resource;
+
+	fchan->maxresp_cached = size + NFSD_MIN_HDR_SEQ_SZ;
 	return 0;
 }
 
@@ -466,9 +496,6 @@ static int init_forechannel_attrs(struct svc_rqst *rqstp,
 		fchan->maxresp_sz = maxcount;
 	session_fchan->maxresp_sz = fchan->maxresp_sz;
 
-	session_fchan->maxresp_cached = NFSD_SLOT_CACHE_SIZE;
-	fchan->maxresp_cached = session_fchan->maxresp_cached;
-
 	/* Use the client's maxops if possible */
 	if (fchan->maxops > NFSD_MAX_OPS_PER_COMPOUND)
 		fchan->maxops = NFSD_MAX_OPS_PER_COMPOUND;
@@ -478,9 +505,12 @@ static int init_forechannel_attrs(struct svc_rqst *rqstp,
 	 * recover pages from existing sessions. For now fail session
 	 * creation.
 	 */
-	status = set_forechannel_maxreqs(fchan);
+	status = set_forechannel_drc_size(fchan);
 
+	session_fchan->maxresp_cached = fchan->maxresp_cached;
 	session_fchan->maxreqs = fchan->maxreqs;
+
+	dprintk("%s status %d\n", __func__, status);
 	return status;
 }
 
diff --git a/include/linux/nfsd/state.h b/include/linux/nfsd/state.h
index fb0c404..ff0b771 100644
--- a/include/linux/nfsd/state.h
+++ b/include/linux/nfsd/state.h
@@ -92,13 +92,17 @@ struct nfs4_cb_conn {
 	struct rpc_cred	*	cb_cred;
 };
 
-/* Maximum number of slots per session. 128 is useful for long haul TCP */
-#define NFSD_MAX_SLOTS_PER_SESSION	128
+/* Maximum number of slots per session. 160 is useful for long haul TCP */
+#define NFSD_MAX_SLOTS_PER_SESSION     160
 /* Maximum number of pages per slot cache entry */
 #define NFSD_PAGES_PER_SLOT	1
 #define NFSD_SLOT_CACHE_SIZE		PAGE_SIZE
 /* Maximum number of operations per session compound */
 #define NFSD_MAX_OPS_PER_COMPOUND	16
+/* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */
+#define NFSD_CACHE_SIZE_SLOTS_PER_SESSION	32
+#define NFSD_MAX_MEM_PER_SESSION  \
+		(NFSD_CACHE_SIZE_SLOTS_PER_SESSION * NFSD_SLOT_CACHE_SIZE)
 
 struct nfsd4_cache_entry {
 	__be32		ce_status;
-- 
1.6.2.5

[PATCH 2/4] nfsd41: use session maxreqs for sequence target and highest slotid

[andros]

From: Andy Adamson <andros@netapp.com>

This fixes a bug in the sequence operation reply.

The sequence operation returns the highest slotid it will accept in the future
in sr_highest_slotid, and the highest slotid it prefers the client to use.
Since we do not re-negotiate the session slot table yet, these should both
always be set to the session ca_maxrequests.

Signed-off-by: Andy Adamson <andros@netapp.com>
---
 fs/nfsd/nfs4state.c |    6 +++++-
 1 files changed, 5 insertions(+), 1 deletions(-)

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index 02b3ddd..ec074e7 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -1133,7 +1133,6 @@ nfsd4_replay_cache_entry(struct nfsd4_compoundres *resp,
 	 * session inactivity timer fires and a solo sequence operation
 	 * is sent (lease renewal).
 	 */
-	seq->maxslots = resp->cstate.session->se_fchannel.maxreqs;
 
 	/* Either returns 0 or nfserr_retry_uncached */
 	status = nfsd4_enc_sequence_replay(resp->rqstp->rq_argp, resp);
@@ -1497,6 +1496,11 @@ nfsd4_sequence(struct svc_rqst *rqstp,
 	slot = &session->se_slots[seq->slotid];
 	dprintk("%s: slotid %d\n", __func__, seq->slotid);
 
+	/* We do not negotiate the number of slots yet, so set the
+	 * maxslots to the session maxreqs which is used to encode
+	 * sr_highest_slotid and the sr_target_slot id to maxslots */
+	seq->maxslots = session->se_fchannel.maxreqs;
+
 	status = check_slot_seqid(seq->seqid, slot->sl_seqid, slot->sl_inuse);
 	if (status == nfserr_replay_cache) {
 		cstate->slot = slot;
-- 
1.6.2.5

[PATCH 3/4] nfsd41: replace nfserr_resource in pure nfs41 responses

[andros]

From: Andy Adamson <andros@netapp.com>

nfserr_resource is not a legal error for NFSv4.1. Replace it with
nfserr_serverfault for EXCHANGE_ID and CREATE_SESSION processing.

We will also need to map nfserr_resource to other errors in routines shared
by NFSv4.0 and NFSv4.1

Signed-off-by: Andy Adamson <andros@netapp.com>
---
 fs/nfsd/nfs4state.c |    6 +++---
 1 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index ec074e7..c9a45f4 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -469,7 +469,7 @@ static int set_forechannel_drc_size(struct nfsd4_channel_attrs *fchan)
 	spin_unlock(&nfsd_drc_lock);
 
 	if (fchan->maxreqs == 0)
-		return nfserr_resource;
+		return nfserr_serverfault;
 
 	fchan->maxresp_cached = size + NFSD_MIN_HDR_SEQ_SZ;
 	return 0;
@@ -519,7 +519,7 @@ alloc_init_session(struct svc_rqst *rqstp, struct nfs4_client *clp,
 		   struct nfsd4_create_session *cses)
 {
 	struct nfsd4_session *new, tmp;
-	int idx, status = nfserr_resource, slotsize;
+	int idx, status = nfserr_serverfault, slotsize;
 
 	memset(&tmp, 0, sizeof(tmp));
 
@@ -1282,7 +1282,7 @@ out_new:
 	/* Normal case */
 	new = create_client(exid->clname, dname);
 	if (new == NULL) {
-		status = nfserr_resource;
+		status = nfserr_serverfault;
 		goto out;
 	}
 
-- 
1.6.2.5

[PATCH 4/4] nfsd41: replace page based DRC with buffer based DRC

[andros]

From: Andy Adamson <andros@netapp.com>

Use NFSD_SLOT_CACHE_SIZE size buffers for sessions DRC instead of holding nfsd
pages in cache.

Connectathon testing has shown that 1024 bytes for encoded compound operation
responses past the sequence operation is sufficient, 512 bytes is a little too
small. Set NFSD_SLOT_CACHE_SIZE to 1024.

Allocate memory for the session DRC in the CREATE_SESSION operation
to guarantee that the memory resource is available for caching responses.
Allocate each slot individually in preparation for slot table size negotiation.

Remove struct nfsd4_cache_entry and helper functions for the old page-based
DRC.

The iov_len calculation in nfs4svc_encode_compoundres is now always
correct.  Replay is now done in nfsd4_sequence under the state lock, so
the session ref count is only bumped on non-replay. Clean up the
nfs4svc_encode_compoundres session logic.

The nfsd4_compound_state statp pointer is also not used.
Remove nfsd4_set_statp().

Move useful nfsd4_cache_entry fields into nfsd4_slot.

Signed-off-by: Andy Adamson <andros@netapp.com
---
 fs/nfsd/nfs4state.c        |  211 +++++++++++++-------------------------------
 fs/nfsd/nfs4xdr.c          |   17 ++--
 fs/nfsd/nfssvc.c           |    4 -
 include/linux/nfsd/state.h |   27 ++----
 include/linux/nfsd/xdr4.h  |    5 +-
 5 files changed, 79 insertions(+), 185 deletions(-)

diff --git a/fs/nfsd/nfs4state.c b/fs/nfsd/nfs4state.c
index c9a45f4..46e9ac5 100644
--- a/fs/nfsd/nfs4state.c
+++ b/fs/nfsd/nfs4state.c
@@ -514,12 +514,23 @@ static int init_forechannel_attrs(struct svc_rqst *rqstp,
 	return status;
 }
 
+static void
+free_session_slots(struct nfsd4_session *ses)
+{
+	int i;
+
+	for (i = 0; i < ses->se_fchannel.maxreqs; i++)
+		kfree(ses->se_slots[i]);
+}
+
 static int
 alloc_init_session(struct svc_rqst *rqstp, struct nfs4_client *clp,
 		   struct nfsd4_create_session *cses)
 {
 	struct nfsd4_session *new, tmp;
-	int idx, status = nfserr_serverfault, slotsize;
+	struct nfsd4_slot *sp;
+	int idx, slotsize, cachesize, i;
+	int status;
 
 	memset(&tmp, 0, sizeof(tmp));
 
@@ -530,14 +541,27 @@ alloc_init_session(struct svc_rqst *rqstp, struct nfs4_client *clp,
 	if (status)
 		goto out;
 
-	/* allocate struct nfsd4_session and slot table in one piece */
-	slotsize = tmp.se_fchannel.maxreqs * sizeof(struct nfsd4_slot);
+	BUILD_BUG_ON(NFSD_MAX_SLOTS_PER_SESSION * sizeof(struct nfsd4_slot)
+		     + sizeof(struct nfsd4_session) > PAGE_SIZE);
+
+	status = nfserr_serverfault;
+	/* allocate struct nfsd4_session and slot table pointers in one piece */
+	slotsize = tmp.se_fchannel.maxreqs * sizeof(struct nfsd4_slot *);
 	new = kzalloc(sizeof(*new) + slotsize, GFP_KERNEL);
 	if (!new)
 		goto out;
 
 	memcpy(new, &tmp, sizeof(*new));
 
+	/* allocate each struct nfsd4_slot and data cache in one piece */
+	cachesize = new->se_fchannel.maxresp_cached - NFSD_MIN_HDR_SEQ_SZ;
+	for (i = 0; i < new->se_fchannel.maxreqs; i++) {
+		sp = kzalloc(sizeof(*sp) + cachesize, GFP_KERNEL);
+		if (!sp)
+			goto out_free;
+		new->se_slots[i] = sp;
+	}
+
 	new->se_client = clp;
 	gen_sessionid(new);
 	idx = hash_sessionid(&new->se_sessionid);
@@ -554,6 +578,10 @@ alloc_init_session(struct svc_rqst *rqstp, struct nfs4_client *clp,
 	status = nfs_ok;
 out:
 	return status;
+out_free:
+	free_session_slots(new);
+	kfree(new);
+	goto out;
 }
 
 /* caller must hold sessionid_lock */
@@ -596,22 +624,16 @@ release_session(struct nfsd4_session *ses)
 	nfsd4_put_session(ses);
 }
 
-static void nfsd4_release_respages(struct page **respages, short resused);
-
 void
 free_session(struct kref *kref)
 {
 	struct nfsd4_session *ses;
-	int i;
 
 	ses = container_of(kref, struct nfsd4_session, se_ref);
-	for (i = 0; i < ses->se_fchannel.maxreqs; i++) {
-		struct nfsd4_cache_entry *e = &ses->se_slots[i].sl_cache_entry;
-		nfsd4_release_respages(e->ce_respages, e->ce_resused);
-	}
 	spin_lock(&nfsd_drc_lock);
 	nfsd_drc_mem_used -= ses->se_fchannel.maxreqs * NFSD_SLOT_CACHE_SIZE;
 	spin_unlock(&nfsd_drc_lock);
+	free_session_slots(ses);
 	kfree(ses);
 }
 
@@ -968,116 +990,31 @@ out_err:
 	return;
 }
 
-void
-nfsd4_set_statp(struct svc_rqst *rqstp, __be32 *statp)
-{
-	struct nfsd4_compoundres *resp = rqstp->rq_resp;
-
-	resp->cstate.statp = statp;
-}
-
-/*
- * Dereference the result pages.
- */
-static void
-nfsd4_release_respages(struct page **respages, short resused)
-{
-	int i;
-
-	dprintk("--> %s\n", __func__);
-	for (i = 0; i < resused; i++) {
-		if (!respages[i])
-			continue;
-		put_page(respages[i]);
-		respages[i] = NULL;
-	}
-}
-
-static void
-nfsd4_copy_pages(struct page **topages, struct page **frompages, short count)
-{
-	int i;
-
-	for (i = 0; i < count; i++) {
-		topages[i] = frompages[i];
-		if (!topages[i])
-			continue;
-		get_page(topages[i]);
-	}
-}
-
 /*
- * Cache the reply pages up to NFSD_PAGES_PER_SLOT + 1, clearing the previous
- * pages. We add a page to NFSD_PAGES_PER_SLOT for the case where the total
- * length of the XDR response is less than se_fmaxresp_cached
- * (NFSD_PAGES_PER_SLOT * PAGE_SIZE) but the xdr_buf pages is used for a
- * of the reply (e.g. readdir).
- *
- * Store the base and length of the rq_req.head[0] page
- * of the NFSv4.1 data, just past the rpc header.
+ * Cache a reply. nfsd4_check_drc_limit() has bounded the cache size.
  */
 void
 nfsd4_store_cache_entry(struct nfsd4_compoundres *resp)
 {
-	struct nfsd4_cache_entry *entry = &resp->cstate.slot->sl_cache_entry;
-	struct svc_rqst *rqstp = resp->rqstp;
-	struct kvec *resv = &rqstp->rq_res.head[0];
-
-	dprintk("--> %s entry %p\n", __func__, entry);
+	struct nfsd4_slot *slot = resp->cstate.slot;
+	unsigned int base;
 
-	nfsd4_release_respages(entry->ce_respages, entry->ce_resused);
-	entry->ce_opcnt = resp->opcnt;
-	entry->ce_status = resp->cstate.status;
+	dprintk("--> %s slot %p\n", __func__, slot);
 
-	/*
-	 * Don't need a page to cache just the sequence operation - the slot
-	 * does this for us!
-	 */
+	slot->sl_opcnt = resp->opcnt;
+	slot->sl_status = resp->cstate.status;
 
 	if (nfsd4_not_cached(resp)) {
-		entry->ce_resused = 0;
-		entry->ce_rpchdrlen = 0;
-		dprintk("%s Just cache SEQUENCE. ce_cachethis %d\n", __func__,
-			resp->cstate.slot->sl_cache_entry.ce_cachethis);
+		slot->sl_datalen = 0;
 		return;
 	}
-	entry->ce_resused = rqstp->rq_resused;
-	if (entry->ce_resused > NFSD_PAGES_PER_SLOT + 1)
-		entry->ce_resused = NFSD_PAGES_PER_SLOT + 1;
-	nfsd4_copy_pages(entry->ce_respages, rqstp->rq_respages,
-			 entry->ce_resused);
-	entry->ce_datav.iov_base = resp->cstate.statp;
-	entry->ce_datav.iov_len = resv->iov_len - ((char *)resp->cstate.statp -
-				(char *)page_address(rqstp->rq_respages[0]));
-	/* Current request rpc header length*/
-	entry->ce_rpchdrlen = (char *)resp->cstate.statp -
-				(char *)page_address(rqstp->rq_respages[0]);
-}
-
-/*
- * We keep the rpc header, but take the nfs reply from the replycache.
- */
-static int
-nfsd41_copy_replay_data(struct nfsd4_compoundres *resp,
-			struct nfsd4_cache_entry *entry)
-{
-	struct svc_rqst *rqstp = resp->rqstp;
-	struct kvec *resv = &resp->rqstp->rq_res.head[0];
-	int len;
-
-	/* Current request rpc header length*/
-	len = (char *)resp->cstate.statp -
-			(char *)page_address(rqstp->rq_respages[0]);
-	if (entry->ce_datav.iov_len + len > PAGE_SIZE) {
-		dprintk("%s v41 cached reply too large (%Zd).\n", __func__,
-			entry->ce_datav.iov_len);
-		return 0;
-	}
-	/* copy the cached reply nfsd data past the current rpc header */
-	memcpy((char *)resv->iov_base + len, entry->ce_datav.iov_base,
-		entry->ce_datav.iov_len);
-	resv->iov_len = len + entry->ce_datav.iov_len;
-	return 1;
+	slot->sl_datalen = (char *)resp->p - (char *)resp->cstate.datap;
+	base = (char *)resp->cstate.datap -
+					(char *)resp->xbuf->head[0].iov_base;
+	if (read_bytes_from_xdr_buf(resp->xbuf, base, slot->sl_data,
+				    slot->sl_datalen))
+		WARN("%s: sessions DRC could not cache compound\n", __func__);
+	return;
 }
 
 /*
@@ -1095,14 +1032,14 @@ nfsd4_enc_sequence_replay(struct nfsd4_compoundargs *args,
 	struct nfsd4_slot *slot = resp->cstate.slot;
 
 	dprintk("--> %s resp->opcnt %d cachethis %u \n", __func__,
-		resp->opcnt, resp->cstate.slot->sl_cache_entry.ce_cachethis);
+		resp->opcnt, resp->cstate.slot->sl_cachethis);
 
 	/* Encode the replayed sequence operation */
 	op = &args->ops[resp->opcnt - 1];
 	nfsd4_encode_operation(resp, op);
 
 	/* Return nfserr_retry_uncached_rep in next operation. */
-	if (args->opcnt > 1 && slot->sl_cache_entry.ce_cachethis == 0) {
+	if (args->opcnt > 1 && slot->sl_cachethis == 0) {
 		op = &args->ops[resp->opcnt++];
 		op->status = nfserr_retry_uncached_rep;
 		nfsd4_encode_operation(resp, op);
@@ -1111,57 +1048,29 @@ nfsd4_enc_sequence_replay(struct nfsd4_compoundargs *args,
 }
 
 /*
- * Keep the first page of the replay. Copy the NFSv4.1 data from the first
- * cached page.  Replace any futher replay pages from the cache.
+ * The sequence operation is not cached because we can use the slot and
+ * session values.
  */
 __be32
 nfsd4_replay_cache_entry(struct nfsd4_compoundres *resp,
 			 struct nfsd4_sequence *seq)
 {
-	struct nfsd4_cache_entry *entry = &resp->cstate.slot->sl_cache_entry;
+	struct nfsd4_slot *slot = resp->cstate.slot;
 	__be32 status;
 
-	dprintk("--> %s entry %p\n", __func__, entry);
-
-	/*
-	 * If this is just the sequence operation, we did not keep
-	 * a page in the cache entry because we can just use the
-	 * slot info stored in struct nfsd4_sequence that was checked
-	 * against the slot in nfsd4_sequence().
-	 *
-	 * This occurs when seq->cachethis is FALSE, or when the client
-	 * session inactivity timer fires and a solo sequence operation
-	 * is sent (lease renewal).
-	 */
+	dprintk("--> %s slot %p\n", __func__, slot);
 
 	/* Either returns 0 or nfserr_retry_uncached */
 	status = nfsd4_enc_sequence_replay(resp->rqstp->rq_argp, resp);
 	if (status == nfserr_retry_uncached_rep)
 		return status;
 
-	if (!nfsd41_copy_replay_data(resp, entry)) {
-		/*
-		 * Not enough room to use the replay rpc header, send the
-		 * cached header. Release all the allocated result pages.
-		 */
-		svc_free_res_pages(resp->rqstp);
-		nfsd4_copy_pages(resp->rqstp->rq_respages, entry->ce_respages,
-			entry->ce_resused);
-	} else {
-		/* Release all but the first allocated result page */
-
-		resp->rqstp->rq_resused--;
-		svc_free_res_pages(resp->rqstp);
-
-		nfsd4_copy_pages(&resp->rqstp->rq_respages[1],
-				 &entry->ce_respages[1],
-				 entry->ce_resused - 1);
-	}
+	/* The sequence operation has been encoded, cstate->datap set. */
+	memcpy(resp->cstate.datap, slot->sl_data, slot->sl_datalen);
 
-	resp->rqstp->rq_resused = entry->ce_resused;
-	resp->opcnt = entry->ce_opcnt;
-	resp->cstate.iovlen = entry->ce_datav.iov_len + entry->ce_rpchdrlen;
-	status = entry->ce_status;
+	resp->opcnt = slot->sl_opcnt;
+	resp->p = resp->cstate.datap + XDR_QUADLEN(slot->sl_datalen);
+	status = slot->sl_status;
 
 	return status;
 }
@@ -1493,7 +1402,7 @@ nfsd4_sequence(struct svc_rqst *rqstp,
 	if (seq->slotid >= session->se_fchannel.maxreqs)
 		goto out;
 
-	slot = &session->se_slots[seq->slotid];
+	slot = session->se_slots[seq->slotid];
 	dprintk("%s: slotid %d\n", __func__, seq->slotid);
 
 	/* We do not negotiate the number of slots yet, so set the
@@ -1506,7 +1415,7 @@ nfsd4_sequence(struct svc_rqst *rqstp,
 		cstate->slot = slot;
 		cstate->session = session;
 		/* Return the cached reply status and set cstate->status
-		 * for nfsd4_svc_encode_compoundres processing */
+		 * for nfsd4_proc_compound processing */
 		status = nfsd4_replay_cache_entry(resp, seq);
 		cstate->status = nfserr_replay_cache;
 		goto out;
@@ -1517,7 +1426,7 @@ nfsd4_sequence(struct svc_rqst *rqstp,
 	/* Success! bump slot seqid */
 	slot->sl_inuse = true;
 	slot->sl_seqid = seq->seqid;
-	slot->sl_cache_entry.ce_cachethis = seq->cachethis;
+	slot->sl_cachethis = seq->cachethis;
 
 	cstate->slot = slot;
 	cstate->session = session;
diff --git a/fs/nfsd/nfs4xdr.c b/fs/nfsd/nfs4xdr.c
index fdf632b..8fa38f5 100644
--- a/fs/nfsd/nfs4xdr.c
+++ b/fs/nfsd/nfs4xdr.c
@@ -3064,6 +3064,7 @@ nfsd4_encode_sequence(struct nfsd4_compoundres *resp, int nfserr,
 	WRITE32(0);
 
 	ADJUST_ARGS();
+	resp->cstate.datap = p; /* DRC cache data pointer */
 	return 0;
 }
 
@@ -3166,7 +3167,7 @@ static int nfsd4_check_drc_limit(struct nfsd4_compoundres *resp)
 		return status;
 
 	session = resp->cstate.session;
-	if (session == NULL || slot->sl_cache_entry.ce_cachethis == 0)
+	if (session == NULL || slot->sl_cachethis == 0)
 		return status;
 
 	if (resp->opcnt >= args->opcnt)
@@ -3291,6 +3292,7 @@ nfs4svc_encode_compoundres(struct svc_rqst *rqstp, __be32 *p, struct nfsd4_compo
 	/*
 	 * All that remains is to write the tag and operation count...
 	 */
+	struct nfsd4_compound_state *cs = &resp->cstate;
 	struct kvec *iov;
 	p = resp->tagp;
 	*p++ = htonl(resp->taglen);
@@ -3304,15 +3306,10 @@ nfs4svc_encode_compoundres(struct svc_rqst *rqstp, __be32 *p, struct nfsd4_compo
 		iov = &rqstp->rq_res.head[0];
 	iov->iov_len = ((char*)resp->p) - (char*)iov->iov_base;
 	BUG_ON(iov->iov_len > PAGE_SIZE);
-	if (nfsd4_has_session(&resp->cstate)) {
-		if (resp->cstate.status == nfserr_replay_cache &&
-				!nfsd4_not_cached(resp)) {
-			iov->iov_len = resp->cstate.iovlen;
-		} else {
-			nfsd4_store_cache_entry(resp);
-			dprintk("%s: SET SLOT STATE TO AVAILABLE\n", __func__);
-			resp->cstate.slot->sl_inuse = 0;
-		}
+	if (nfsd4_has_session(cs) && cs->status != nfserr_replay_cache) {
+		nfsd4_store_cache_entry(resp);
+		dprintk("%s: SET SLOT STATE TO AVAILABLE\n", __func__);
+		resp->cstate.slot->sl_inuse = false;
 		nfsd4_put_session(resp->cstate.session);
 	}
 	return 1;
diff --git a/fs/nfsd/nfssvc.c b/fs/nfsd/nfssvc.c
index 675d395..4472449 100644
--- a/fs/nfsd/nfssvc.c
+++ b/fs/nfsd/nfssvc.c
@@ -577,10 +577,6 @@ nfsd_dispatch(struct svc_rqst *rqstp, __be32 *statp)
 		+ rqstp->rq_res.head[0].iov_len;
 	rqstp->rq_res.head[0].iov_len += sizeof(__be32);
 
-	/* NFSv4.1 DRC requires statp */
-	if (rqstp->rq_vers == 4)
-		nfsd4_set_statp(rqstp, statp);
-
 	/* Now call the procedure handler, and encode NFS status. */
 	nfserr = proc->pc_func(rqstp, rqstp->rq_argp, rqstp->rq_resp);
 	nfserr = map_new_errors(rqstp->rq_vers, nfserr);
diff --git a/include/linux/nfsd/state.h b/include/linux/nfsd/state.h
index ff0b771..70ef5f4 100644
--- a/include/linux/nfsd/state.h
+++ b/include/linux/nfsd/state.h
@@ -94,30 +94,23 @@ struct nfs4_cb_conn {
 
 /* Maximum number of slots per session. 160 is useful for long haul TCP */
 #define NFSD_MAX_SLOTS_PER_SESSION     160
-/* Maximum number of pages per slot cache entry */
-#define NFSD_PAGES_PER_SLOT	1
-#define NFSD_SLOT_CACHE_SIZE		PAGE_SIZE
 /* Maximum number of operations per session compound */
 #define NFSD_MAX_OPS_PER_COMPOUND	16
+/* Maximum  session per slot cache size */
+#define NFSD_SLOT_CACHE_SIZE		1024
 /* Maximum number of NFSD_SLOT_CACHE_SIZE slots per session */
 #define NFSD_CACHE_SIZE_SLOTS_PER_SESSION	32
 #define NFSD_MAX_MEM_PER_SESSION  \
 		(NFSD_CACHE_SIZE_SLOTS_PER_SESSION * NFSD_SLOT_CACHE_SIZE)
 
-struct nfsd4_cache_entry {
-	__be32		ce_status;
-	struct kvec	ce_datav; /* encoded NFSv4.1 data in rq_res.head[0] */
-	struct page	*ce_respages[NFSD_PAGES_PER_SLOT + 1];
-	int		ce_cachethis;
-	short		ce_resused;
-	int		ce_opcnt;
-	int		ce_rpchdrlen;
-};
-
 struct nfsd4_slot {
-	bool				sl_inuse;
-	u32				sl_seqid;
-	struct nfsd4_cache_entry	sl_cache_entry;
+	bool	sl_inuse;
+	bool	sl_cachethis;
+	u16	sl_opcnt;
+	u32	sl_seqid;
+	__be32	sl_status;
+	u32	sl_datalen;
+	char	sl_data[];
 };
 
 struct nfsd4_channel_attrs {
@@ -159,7 +152,7 @@ struct nfsd4_session {
 	struct nfs4_sessionid	se_sessionid;
 	struct nfsd4_channel_attrs se_fchannel;
 	struct nfsd4_channel_attrs se_bchannel;
-	struct nfsd4_slot	se_slots[];	/* forward channel slots */
+	struct nfsd4_slot	*se_slots[];	/* forward channel slots */
 };
 
 static inline void
diff --git a/include/linux/nfsd/xdr4.h b/include/linux/nfsd/xdr4.h
index 3f71660..73164c2 100644
--- a/include/linux/nfsd/xdr4.h
+++ b/include/linux/nfsd/xdr4.h
@@ -51,7 +51,7 @@ struct nfsd4_compound_state {
 	/* For sessions DRC */
 	struct nfsd4_session	*session;
 	struct nfsd4_slot	*slot;
-	__be32			*statp;
+	__be32			*datap;
 	size_t			iovlen;
 	u32			minorversion;
 	u32			status;
@@ -472,8 +472,7 @@ static inline bool nfsd4_is_solo_sequence(struct nfsd4_compoundres *resp)
 
 static inline bool nfsd4_not_cached(struct nfsd4_compoundres *resp)
 {
-	return !resp->cstate.slot->sl_cache_entry.ce_cachethis ||
-			nfsd4_is_solo_sequence(resp);
+	return !resp->cstate.slot->sl_cachethis || nfsd4_is_solo_sequence(resp);
 }
 
 #define NFS4_SVC_XDRSIZE		sizeof(struct nfsd4_compoundargs)
-- 
1.6.2.5

Re: [PATCH 0/4] NFSv4.1 DRC Rewrite Version 7

[J. Bruce Fields]

On Fri, Aug 28, 2009 at 08:45:00AM -0400, andros@netapp.com wrote:
> NFSv4.1 DRC Rewrite Version 7
> 
> These patches apply against git://linux-nfs.org/~bfields/linux: for-2.6.32
> branch and continue to rewrite the NFSv4.1 Sessions DRC.

All applied.  Thanks, Andy!--b.

> 
> Responded to comments on Version 6.
> 
> 1) The bound on the fore channel per-session DRC size are rewritten.
> Instead of just using a maximum number of slots to bound the size,
> The server gives the client the number of ca_maxresponsesize_cached slots it
> requests bounded by NFSD_SLOT_CACHE_SIZE, NFSD_MAX_MEM_PER_SESSION and by
> nfsd_drc_max_mem. Do not allow more than NFSD_MAX_SLOTS_PER_SESSION.
> 
> This allows clients to tailor a session to usage.
> For example, an I/O session (READ/WRITE/COMMIT only) can have a much smaller
> ca_maxresponsesize_cached (for only WRITE/COMMIt compound responses) and a lot
> larger ca_maxresponses to service a large in-flight data window.
> 
> 2) the page-based DRC is replaced with a buffer based DRC with each
> slot table entry (struct nfsd4_slot + cache) allocated separately.
> This allocation prepares us for slot size re-negotiation via the SEQUENCE
> operation target and high slot id arguments.
> 
> Testing:
> 
> NFSv4.1 mount: pynfs tests - including the SEQUENCE replay cache tests.
> connectathon tests.
> 
> NFSv4.0 mount: connectathon tests.
> 
> -->Andy
> 
> 0001-nfsd41-bound-forechannel-drc-size-by-memory-usage.patch
> 0002-nfsd41-use-session-maxreqs-for-sequence-target-and.patch
> 0003-nfsd41-replace-nfserr_resource-in-pure-nfs41-respon.patch
> 0004-nfsd41-replace-page-based-DRC-with-buffer-based-DRC.patch
>