-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
+/*
+ * GPL HEADER START
*
- * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
*
- * This file is part of Lustre, http://www.lustre.org.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
*
- * Lustre is free software; you can redistribute it and/or
- * modify it under the terms of version 2 of the GNU General Public
- * License as published by the Free Software Foundation.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
*
- * Lustre is distributed in the hope that it will be useful,
- * but WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
- * GNU General Public License for more details.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
*
- * You should have received a copy of the GNU General Public License
- * along with Lustre; if not, write to the Free Software
- * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
*
+ * GPL HEADER END
+ */
+/*
+ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
+ * Use is subject to license terms.
+ *
+ * Copyright (c) 2012, 2013, Intel Corporation.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
*/
#define DEBUG_SUBSYSTEM S_RPC
-#ifndef __KERNEL__
-#include <liblustre.h>
-#endif
-#include <linux/obd_class.h>
-#include <linux/obd_support.h>
-#include <linux/lustre_net.h>
-#include <linux/lustre_lib.h>
-#include <linux/obd.h>
-#include <linux/lustre_sec.h>
+#include <obd_support.h>
+#include <lustre_net.h>
+#include <lustre_lib.h>
+#include <obd.h>
+#include <obd_class.h>
#include "ptlrpc_internal.h"
-static int ptl_send_buf (ptl_handle_md_t *mdh, void *base, int len,
- ptl_ack_req_t ack, struct ptlrpc_cb_id *cbid,
- struct ptlrpc_connection *conn, int portal, __u64 xid)
+/**
+ * Helper function. Sends \a len bytes from \a base at offset \a offset
+ * over \a conn connection to portal \a portal.
+ * Returns 0 on success or error code.
+ */
+static int ptl_send_buf (lnet_handle_md_t *mdh, void *base, int len,
+ lnet_ack_req_t ack, struct ptlrpc_cb_id *cbid,
+ struct ptlrpc_connection *conn, int portal, __u64 xid,
+ unsigned int offset)
{
int rc;
- ptl_md_t md;
- char str[PTL_NALFMT_SIZE];
+ lnet_md_t md;
ENTRY;
LASSERT (portal != 0);
LASSERT (conn != NULL);
- CDEBUG (D_INFO, "conn=%p ni %s id %s on %s\n",
- conn, conn->c_peer.peer_ni->pni_name,
- ptlrpc_id2str(&conn->c_peer, str),
- conn->c_peer.peer_ni->pni_name);
+ CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer));
md.start = base;
md.length = len;
- md.threshold = (ack == PTL_ACK_REQ) ? 2 : 1;
+ md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1;
md.options = PTLRPC_MD_OPTIONS;
md.user_ptr = cbid;
- md.eq_handle = conn->c_peer.peer_ni->pni_eq_h;
+ md.eq_handle = ptlrpc_eq_h;
- if (ack == PTL_ACK_REQ &&
- OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_ACK | OBD_FAIL_ONCE)) {
+ if (unlikely(ack == LNET_ACK_REQ &&
+ OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_ACK, OBD_FAIL_ONCE))){
/* don't ask for the ack to simulate failing client */
- ack = PTL_NOACK_REQ;
- obd_fail_loc |= OBD_FAIL_ONCE | OBD_FAILED;
+ ack = LNET_NOACK_REQ;
}
- rc = PtlMDBind (conn->c_peer.peer_ni->pni_ni_h, md,
- PTL_UNLINK, mdh);
- if (rc != PTL_OK) {
- CERROR ("PtlMDBind failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
+ rc = LNetMDBind (md, LNET_UNLINK, mdh);
+ if (unlikely(rc != 0)) {
+ CERROR ("LNetMDBind failed: %d\n", rc);
+ LASSERT (rc == -ENOMEM);
RETURN (-ENOMEM);
}
- CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
- len, portal, xid);
+ CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n",
+ len, portal, xid, offset);
- rc = PtlPut (*mdh, ack, conn->c_peer.peer_id, portal, 0, xid, 0, 0);
- if (rc != PTL_OK) {
- int rc2;
+ rc = LNetPut (conn->c_self, *mdh, ack,
+ conn->c_peer, portal, xid, offset, 0);
+ if (unlikely(rc != 0)) {
+ int rc2;
/* We're going to get an UNLINK event when I unlink below,
* which will complete just like any other failed send, so
* I fall through and return success here! */
- CERROR("PtlPut(%s, %d, "LPD64") failed: %d\n",
- ptlrpc_id2str(&conn->c_peer, str),
- portal, xid, rc);
- rc2 = PtlMDUnlink(*mdh);
- LASSERTF(rc2 == PTL_OK, "rc2 = %d\n", rc2);
+ CERROR("LNetPut(%s, %d, "LPD64") failed: %d\n",
+ libcfs_id2str(conn->c_peer), portal, xid, rc);
+ rc2 = LNetMDUnlink(*mdh);
+ LASSERTF(rc2 == 0, "rc2 = %d\n", rc2);
}
RETURN (0);
}
-int ptlrpc_start_bulk_transfer (struct ptlrpc_bulk_desc *desc)
+static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count)
{
- int rc;
- int rc2;
- struct ptlrpc_peer *peer;
- ptl_md_t md;
- __u64 xid;
- char str[PTL_NALFMT_SIZE];
- ENTRY;
+ int i;
- if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_PUT_NET))
- RETURN(0);
+ for (i = 0; i < count; i++)
+ LNetMDUnlink(bd_mds[i]);
+}
- /* NB no locking required until desc is on the network */
- LASSERT (!desc->bd_network_rw);
- LASSERT (desc->bd_type == BULK_PUT_SOURCE ||
- desc->bd_type == BULK_GET_SINK);
- desc->bd_success = 0;
- peer = &desc->bd_export->exp_connection->c_peer;
-
- md.user_ptr = &desc->bd_cbid;
- md.eq_handle = peer->peer_ni->pni_eq_h;
- md.threshold = 2; /* SENT and ACK/REPLY */
- md.options = PTLRPC_MD_OPTIONS;
- ptlrpc_fill_bulk_md(&md, desc);
-
- LASSERT (desc->bd_cbid.cbid_fn == server_bulk_callback);
- LASSERT (desc->bd_cbid.cbid_arg == desc);
-
- /* NB total length may be 0 for a read past EOF, so we send a 0
- * length bulk, since the client expects a bulk event. */
-
- rc = PtlMDBind(peer->peer_ni->pni_ni_h, md,
- PTL_UNLINK, &desc->bd_md_h);
- if (rc != PTL_OK) {
- CERROR("PtlMDBind failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- RETURN(-ENOMEM);
- }
+#ifdef HAVE_SERVER_SUPPORT
+/**
+ * Prepare bulk descriptor for specified incoming request \a req that
+ * can fit \a npages * pages. \a type is bulk type. \a portal is where
+ * the bulk to be sent. Used on server-side after request was already
+ * received.
+ * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
+ * error.
+ */
+struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
+ unsigned npages, unsigned max_brw,
+ unsigned type, unsigned portal)
+{
+ struct obd_export *exp = req->rq_export;
+ struct ptlrpc_bulk_desc *desc;
- /* Client's bulk and reply matchbits are the same */
- xid = desc->bd_req->rq_xid;
- CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d on %s "
- "nid %s pid %d xid "LPX64"\n", desc->bd_iov_count,
- desc->bd_nob, desc->bd_portal, peer->peer_ni->pni_name,
- ptlrpc_id2str(peer, str), peer->peer_id.pid, xid);
+ ENTRY;
+ LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
- /* Network is about to get at the memory */
- desc->bd_network_rw = 1;
+ desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
+ if (desc == NULL)
+ RETURN(NULL);
- if (desc->bd_type == BULK_PUT_SOURCE)
- rc = PtlPut (desc->bd_md_h, PTL_ACK_REQ, peer->peer_id,
- desc->bd_portal, 0, xid, 0, 0);
- else
- rc = PtlGet (desc->bd_md_h, peer->peer_id,
- desc->bd_portal, 0, xid, 0);
-
- if (rc != PTL_OK) {
- /* Can't send, so we unlink the MD bound above. The UNLINK
- * event this creates will signal completion with failure,
- * so we return SUCCESS here! */
- CERROR("Transfer(%s, %d, "LPX64") failed: %d\n",
- ptlrpc_id2str(peer, str),
- desc->bd_portal, xid, rc);
- rc2 = PtlMDUnlink(desc->bd_md_h);
- LASSERT (rc2 == PTL_OK);
- }
+ desc->bd_export = class_export_get(exp);
+ desc->bd_req = req;
- RETURN(0);
-}
+ desc->bd_cbid.cbid_fn = server_bulk_callback;
+ desc->bd_cbid.cbid_arg = desc;
-void ptlrpc_abort_bulk (struct ptlrpc_bulk_desc *desc)
-{
- /* Server side bulk abort. Idempotent. Not thread-safe (i.e. only
- * serialises with completion callback) */
- struct l_wait_info lwi;
- int rc;
+ /* NB we don't assign rq_bulk here; server-side requests are
+ * re-used, and the handler frees the bulk desc explicitly. */
- LASSERT (!in_interrupt ()); /* might sleep */
+ return desc;
+}
+EXPORT_SYMBOL(ptlrpc_prep_bulk_exp);
- if (!ptlrpc_bulk_active(desc)) /* completed or */
- return; /* never started */
-
- /* The unlink ensures the callback happens ASAP and is the last
- * one. If it fails, it must be because completion just happened,
- * but we must still l_wait_event() in this case, to give liblustre
- * a chance to run server_bulk_callback()*/
+/**
+ * Starts bulk transfer for descriptor \a desc on the server.
+ * Returns 0 on success or error code.
+ */
+int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc)
+{
+ struct obd_export *exp = desc->bd_export;
+ struct ptlrpc_connection *conn = exp->exp_connection;
+ int rc = 0;
+ __u64 xid;
+ int posted_md;
+ int total_md;
+ lnet_md_t md;
+ ENTRY;
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET))
+ RETURN(0);
+
+ /* NB no locking required until desc is on the network */
+ LASSERT(desc->bd_md_count == 0);
+ LASSERT(desc->bd_type == BULK_PUT_SOURCE ||
+ desc->bd_type == BULK_GET_SINK);
+
+ LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback);
+ LASSERT(desc->bd_cbid.cbid_arg == desc);
+
+ /* NB total length may be 0 for a read past EOF, so we send 0
+ * length bulks, since the client expects bulk events.
+ *
+ * The client may not need all of the bulk XIDs for the RPC. The RPC
+ * used the XID of the highest bulk XID needed, and the server masks
+ * off high bits to get bulk count for this RPC. LU-1431 */
+ xid = desc->bd_req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
+ total_md = desc->bd_req->rq_xid - xid + 1;
+
+ desc->bd_md_count = total_md;
+ desc->bd_failure = 0;
+
+ md.user_ptr = &desc->bd_cbid;
+ md.eq_handle = ptlrpc_eq_h;
+ md.threshold = 2; /* SENT and ACK/REPLY */
+
+ for (posted_md = 0; posted_md < total_md; xid++) {
+ md.options = PTLRPC_MD_OPTIONS;
+
+ /* NB it's assumed that source and sink buffer frags are
+ * page-aligned. Otherwise we'd have to send client bulk
+ * sizes over and split server buffer accordingly */
+ ptlrpc_fill_bulk_md(&md, desc, posted_md);
+ rc = LNetMDBind(md, LNET_UNLINK, &desc->bd_mds[posted_md]);
+ if (rc != 0) {
+ CERROR("%s: LNetMDBind failed for MD %u: rc = %d\n",
+ exp->exp_obd->obd_name, posted_md, rc);
+ LASSERT(rc == -ENOMEM);
+ if (posted_md == 0) {
+ desc->bd_md_count = 0;
+ RETURN(-ENOMEM);
+ }
+ break;
+ }
+ /* Network is about to get at the memory */
+ if (desc->bd_type == BULK_PUT_SOURCE)
+ rc = LNetPut(conn->c_self, desc->bd_mds[posted_md],
+ LNET_ACK_REQ, conn->c_peer,
+ desc->bd_portal, xid, 0, 0);
+ else
+ rc = LNetGet(conn->c_self, desc->bd_mds[posted_md],
+ conn->c_peer, desc->bd_portal, xid, 0);
+
+ posted_md++;
+ if (rc != 0) {
+ CERROR("%s: failed bulk transfer with %s:%u x"LPU64": "
+ "rc = %d\n", exp->exp_obd->obd_name,
+ libcfs_id2str(conn->c_peer), desc->bd_portal,
+ xid, rc);
+ break;
+ }
+ }
+
+ if (rc != 0) {
+ /* Can't send, so we unlink the MD bound above. The UNLINK
+ * event this creates will signal completion with failure,
+ * so we return SUCCESS here! */
+ spin_lock(&desc->bd_lock);
+ desc->bd_md_count -= total_md - posted_md;
+ spin_unlock(&desc->bd_lock);
+ LASSERT(desc->bd_md_count >= 0);
+
+ mdunlink_iterate_helper(desc->bd_mds, posted_md);
+ RETURN(0);
+ }
+
+ CDEBUG(D_NET, "Transferring %u pages %u bytes via portal %d "
+ "id %s xid "LPX64"-"LPX64"\n", desc->bd_iov_count,
+ desc->bd_nob, desc->bd_portal, libcfs_id2str(conn->c_peer),
+ xid - posted_md, xid - 1);
+
+ RETURN(0);
+}
+EXPORT_SYMBOL(ptlrpc_start_bulk_transfer);
- PtlMDUnlink (desc->bd_md_h);
+/**
+ * Server side bulk abort. Idempotent. Not thread-safe (i.e. only
+ * serialises with completion callback)
+ */
+void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
+{
+ struct l_wait_info lwi;
+ int rc;
+
+ LASSERT(!in_interrupt()); /* might sleep */
+
+ if (!ptlrpc_server_bulk_active(desc)) /* completed or */
+ return; /* never started */
+
+ /* We used to poison the pages with 0xab here because we did not want to
+ * send any meaningful data over the wire for evicted clients (bug 9297)
+ * However, this is no longer safe now that we use the page cache on the
+ * OSS (bug 20560) */
+
+ /* The unlink ensures the callback happens ASAP and is the last
+ * one. If it fails, it must be because completion just happened,
+ * but we must still l_wait_event() in this case, to give liblustre
+ * a chance to run server_bulk_callback()*/
+ mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_count);
+
+ for (;;) {
+ /* Network access will complete in finite time but the HUGE
+ * timeout lets us CWARN for visibility of sluggish NALs */
+ lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
+ cfs_time_seconds(1), NULL, NULL);
+ rc = l_wait_event(desc->bd_waitq,
+ !ptlrpc_server_bulk_active(desc), &lwi);
+ if (rc == 0)
+ return;
+
+ LASSERT(rc == -ETIMEDOUT);
+ CWARN("Unexpectedly long timeout: desc %p\n", desc);
+ }
+}
+EXPORT_SYMBOL(ptlrpc_abort_bulk);
+#endif /* HAVE_SERVER_SUPPORT */
- for (;;) {
- /* Network access will complete in finite time but the HUGE
- * timeout lets us CWARN for visibility of sluggish NALs */
- lwi = LWI_TIMEOUT (300 * HZ, NULL, NULL);
- rc = l_wait_event(desc->bd_waitq,
- !ptlrpc_bulk_active(desc), &lwi);
- if (rc == 0)
- return;
+/**
+ * Register bulk at the sender for later transfer.
+ * Returns 0 on success or error code.
+ */
+int ptlrpc_register_bulk(struct ptlrpc_request *req)
+{
+ struct ptlrpc_bulk_desc *desc = req->rq_bulk;
+ lnet_process_id_t peer;
+ int rc = 0;
+ int rc2;
+ int posted_md;
+ int total_md;
+ __u64 xid;
+ lnet_handle_me_t me_h;
+ lnet_md_t md;
+ ENTRY;
+
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET))
+ RETURN(0);
- LASSERT(rc == -ETIMEDOUT);
- CWARN("Unexpectedly long timeout: desc %p\n", desc);
- }
+ /* NB no locking required until desc is on the network */
+ LASSERT(desc->bd_nob > 0);
+ LASSERT(desc->bd_md_count == 0);
+ LASSERT(desc->bd_md_max_brw <= PTLRPC_BULK_OPS_COUNT);
+ LASSERT(desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
+ LASSERT(desc->bd_req != NULL);
+ LASSERT(desc->bd_type == BULK_PUT_SINK ||
+ desc->bd_type == BULK_GET_SOURCE);
+
+ /* cleanup the state of the bulk for it will be reused */
+ if (req->rq_resend || req->rq_send_state == LUSTRE_IMP_REPLAY)
+ desc->bd_nob_transferred = 0;
+ else
+ LASSERT(desc->bd_nob_transferred == 0);
+
+ desc->bd_failure = 0;
+
+ peer = desc->bd_import->imp_connection->c_peer;
+
+ LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback);
+ LASSERT(desc->bd_cbid.cbid_arg == desc);
+
+ /* An XID is only used for a single request from the client.
+ * For retried bulk transfers, a new XID will be allocated in
+ * in ptlrpc_check_set() if it needs to be resent, so it is not
+ * using the same RDMA match bits after an error.
+ *
+ * For multi-bulk RPCs, rq_xid is the last XID needed for bulks. The
+ * first bulk XID is power-of-two aligned before rq_xid. LU-1431 */
+ xid = req->rq_xid & ~((__u64)desc->bd_md_max_brw - 1);
+ LASSERTF(!(desc->bd_registered &&
+ req->rq_send_state != LUSTRE_IMP_REPLAY) ||
+ xid != desc->bd_last_xid,
+ "registered: %d rq_xid: "LPU64" bd_last_xid: "LPU64"\n",
+ desc->bd_registered, xid, desc->bd_last_xid);
+
+ total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV;
+ desc->bd_registered = 1;
+ desc->bd_last_xid = xid;
+ desc->bd_md_count = total_md;
+ md.user_ptr = &desc->bd_cbid;
+ md.eq_handle = ptlrpc_eq_h;
+ md.threshold = 1; /* PUT or GET */
+
+ for (posted_md = 0; posted_md < total_md; posted_md++, xid++) {
+ md.options = PTLRPC_MD_OPTIONS |
+ ((desc->bd_type == BULK_GET_SOURCE) ?
+ LNET_MD_OP_GET : LNET_MD_OP_PUT);
+ ptlrpc_fill_bulk_md(&md, desc, posted_md);
+
+ rc = LNetMEAttach(desc->bd_portal, peer, xid, 0,
+ LNET_UNLINK, LNET_INS_AFTER, &me_h);
+ if (rc != 0) {
+ CERROR("%s: LNetMEAttach failed x"LPU64"/%d: rc = %d\n",
+ desc->bd_import->imp_obd->obd_name, xid,
+ posted_md, rc);
+ break;
+ }
+
+ /* About to let the network at it... */
+ rc = LNetMDAttach(me_h, md, LNET_UNLINK,
+ &desc->bd_mds[posted_md]);
+ if (rc != 0) {
+ CERROR("%s: LNetMDAttach failed x"LPU64"/%d: rc = %d\n",
+ desc->bd_import->imp_obd->obd_name, xid,
+ posted_md, rc);
+ rc2 = LNetMEUnlink(me_h);
+ LASSERT(rc2 == 0);
+ break;
+ }
+ }
+
+ if (rc != 0) {
+ LASSERT(rc == -ENOMEM);
+ spin_lock(&desc->bd_lock);
+ desc->bd_md_count -= total_md - posted_md;
+ spin_unlock(&desc->bd_lock);
+ LASSERT(desc->bd_md_count >= 0);
+ mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
+ req->rq_status = -ENOMEM;
+ RETURN(-ENOMEM);
+ }
+
+ /* Set rq_xid to matchbits of the final bulk so that server can
+ * infer the number of bulks that were prepared */
+ req->rq_xid = --xid;
+ LASSERTF(desc->bd_last_xid == (req->rq_xid & PTLRPC_BULK_OPS_MASK),
+ "bd_last_xid = x"LPU64", rq_xid = x"LPU64"\n",
+ desc->bd_last_xid, req->rq_xid);
+
+ spin_lock(&desc->bd_lock);
+ /* Holler if peer manages to touch buffers before he knows the xid */
+ if (desc->bd_md_count != total_md)
+ CWARN("%s: Peer %s touched %d buffers while I registered\n",
+ desc->bd_import->imp_obd->obd_name, libcfs_id2str(peer),
+ total_md - desc->bd_md_count);
+ spin_unlock(&desc->bd_lock);
+
+ CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, "
+ "xid x"LPX64"-"LPX64", portal %u\n", desc->bd_md_count,
+ desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
+ desc->bd_iov_count, desc->bd_nob,
+ desc->bd_last_xid, req->rq_xid, desc->bd_portal);
+
+ RETURN(0);
}
+EXPORT_SYMBOL(ptlrpc_register_bulk);
-int ptlrpc_register_bulk (struct ptlrpc_request *req)
+/**
+ * Disconnect a bulk desc from the network. Idempotent. Not
+ * thread-safe (i.e. only interlocks with completion callback).
+ * Returns 1 on success or 0 if network unregistration failed for whatever
+ * reason.
+ */
+int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
{
- struct ptlrpc_bulk_desc *desc = req->rq_bulk;
- struct ptlrpc_peer *peer;
- int rc;
- int rc2;
- ptl_handle_me_t me_h;
- ptl_md_t md;
- ENTRY;
+ struct ptlrpc_bulk_desc *desc = req->rq_bulk;
+ struct l_wait_info lwi;
+ int rc;
+ ENTRY;
+
+ LASSERT(!in_interrupt()); /* might sleep */
+
+ /* Let's setup deadline for reply unlink. */
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
+ async && req->rq_bulk_deadline == 0)
+ req->rq_bulk_deadline = cfs_time_current_sec() + LONG_UNLINK;
+
+ if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
+ RETURN(1); /* never registered */
+
+ LASSERT(desc->bd_req == req); /* bd_req NULL until registered */
+
+ /* the unlink ensures the callback happens ASAP and is the last
+ * one. If it fails, it must be because completion just happened,
+ * but we must still l_wait_event() in this case to give liblustre
+ * a chance to run client_bulk_callback() */
+ mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw);
+
+ if (ptlrpc_client_bulk_active(req) == 0) /* completed or */
+ RETURN(1); /* never registered */
+
+ /* Move to "Unregistering" phase as bulk was not unlinked yet. */
+ ptlrpc_rqphase_move(req, RQ_PHASE_UNREGISTERING);
- if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_BULK_GET_NET))
+ /* Do not wait for unlink to finish. */
+ if (async)
RETURN(0);
- /* NB no locking required until desc is on the network */
- LASSERT (desc->bd_nob > 0);
- LASSERT (!desc->bd_network_rw);
- LASSERT (desc->bd_iov_count <= PTLRPC_MAX_BRW_PAGES);
- LASSERT (desc->bd_req != NULL);
- LASSERT (desc->bd_type == BULK_PUT_SINK ||
- desc->bd_type == BULK_GET_SOURCE);
-
- desc->bd_success = 0;
-
- peer = &desc->bd_import->imp_connection->c_peer;
-
- md.user_ptr = &desc->bd_cbid;
- md.eq_handle = peer->peer_ni->pni_eq_h;
- md.threshold = 1; /* PUT or GET */
- md.options = PTLRPC_MD_OPTIONS |
- ((desc->bd_type == BULK_GET_SOURCE) ?
- PTL_MD_OP_GET : PTL_MD_OP_PUT);
- ptlrpc_fill_bulk_md(&md, desc);
-
- LASSERT (desc->bd_cbid.cbid_fn == client_bulk_callback);
- LASSERT (desc->bd_cbid.cbid_arg == desc);
-
- /* XXX Registering the same xid on retried bulk makes my head
- * explode trying to understand how the original request's bulk
- * might interfere with the retried request -eeb */
- LASSERT (!desc->bd_registered || req->rq_xid != desc->bd_last_xid);
- desc->bd_registered = 1;
- desc->bd_last_xid = req->rq_xid;
-
- rc = PtlMEAttach(peer->peer_ni->pni_ni_h, desc->bd_portal,
- desc->bd_import->imp_connection->c_peer.peer_id,
- req->rq_xid, 0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
- if (rc != PTL_OK) {
- CERROR("PtlMEAttach failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- RETURN (-ENOMEM);
- }
+ for (;;) {
+ /* The wq argument is ignored by user-space wait_event macros */
+ wait_queue_head_t *wq = (req->rq_set != NULL) ?
+ &req->rq_set->set_waitq :
+ &req->rq_reply_waitq;
+ /* Network access will complete in finite time but the HUGE
+ * timeout lets us CWARN for visibility of sluggish NALs */
+ lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
+ cfs_time_seconds(1), NULL, NULL);
+ rc = l_wait_event(*wq, !ptlrpc_client_bulk_active(req), &lwi);
+ if (rc == 0) {
+ ptlrpc_rqphase_move(req, req->rq_next_phase);
+ RETURN(1);
+ }
- /* About to let the network at it... */
- desc->bd_network_rw = 1;
- rc = PtlMDAttach(me_h, md, PTL_UNLINK, &desc->bd_md_h);
- if (rc != PTL_OK) {
- CERROR("PtlMDAttach failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- desc->bd_network_rw = 0;
- rc2 = PtlMEUnlink (me_h);
- LASSERT (rc2 == PTL_OK);
- RETURN (-ENOMEM);
+ LASSERT(rc == -ETIMEDOUT);
+ DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p",
+ desc);
}
-
- CDEBUG(D_NET, "Setup bulk %s buffers: %u pages %u bytes, xid "LPX64", "
- "portal %u on %s\n",
- desc->bd_type == BULK_GET_SOURCE ? "get-source" : "put-sink",
- desc->bd_iov_count, desc->bd_nob,
- req->rq_xid, desc->bd_portal, peer->peer_ni->pni_name);
RETURN(0);
}
+EXPORT_SYMBOL(ptlrpc_unregister_bulk);
-void ptlrpc_unregister_bulk (struct ptlrpc_request *req)
+static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags)
{
- /* Disconnect a bulk desc from the network. Idempotent. Not
- * thread-safe (i.e. only interlocks with completion callback). */
- struct ptlrpc_bulk_desc *desc = req->rq_bulk;
- wait_queue_head_t *wq;
- struct l_wait_info lwi;
- int rc;
-
- LASSERT (!in_interrupt ()); /* might sleep */
-
- if (!ptlrpc_bulk_active(desc)) /* completed or */
- return; /* never registered */
-
- LASSERT (desc->bd_req == req); /* bd_req NULL until registered */
-
- /* the unlink ensures the callback happens ASAP and is the last
- * one. If it fails, it must be because completion just happened,
- * but we must still l_wait_event() in this case to give liblustre
- * a chance to run client_bulk_callback() */
-
- PtlMDUnlink (desc->bd_md_h);
-
- if (req->rq_set != NULL)
- wq = &req->rq_set->set_waitq;
+ struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt;
+ struct ptlrpc_service *svc = svcpt->scp_service;
+ int service_time = max_t(int, cfs_time_current_sec() -
+ req->rq_arrival_time.tv_sec, 1);
+
+ if (!(flags & PTLRPC_REPLY_EARLY) &&
+ (req->rq_type != PTL_RPC_MSG_ERR) &&
+ (req->rq_reqmsg != NULL) &&
+ !(lustre_msg_get_flags(req->rq_reqmsg) &
+ (MSG_RESENT | MSG_REPLAY |
+ MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE))) {
+ /* early replies, errors and recovery requests don't count
+ * toward our service time estimate */
+ int oldse = at_measured(&svcpt->scp_at_estimate, service_time);
+
+ if (oldse != 0) {
+ DEBUG_REQ(D_ADAPTTO, req,
+ "svc %s changed estimate from %d to %d",
+ svc->srv_name, oldse,
+ at_get(&svcpt->scp_at_estimate));
+ }
+ }
+ /* Report actual service time for client latency calc */
+ lustre_msg_set_service_time(req->rq_repmsg, service_time);
+ /* Report service time estimate for future client reqs, but report 0
+ * (to be ignored by client) if it's a error reply during recovery.
+ * (bz15815) */
+ if (req->rq_type == PTL_RPC_MSG_ERR &&
+ (req->rq_export == NULL || req->rq_export->exp_obd->obd_recovering))
+ lustre_msg_set_timeout(req->rq_repmsg, 0);
else
- wq = &req->rq_reply_waitq;
-
- for (;;) {
- /* Network access will complete in finite time but the HUGE
- * timeout lets us CWARN for visibility of sluggish NALs */
- lwi = LWI_TIMEOUT (300 * HZ, NULL, NULL);
- rc = l_wait_event(*wq, !ptlrpc_bulk_active(desc), &lwi);
- if (rc == 0)
- return;
-
- LASSERT (rc == -ETIMEDOUT);
- CWARN("Unexpectedly long timeout: desc %p\n", desc);
+ lustre_msg_set_timeout(req->rq_repmsg,
+ at_get(&svcpt->scp_at_estimate));
+
+ if (req->rq_reqmsg &&
+ !(lustre_msghdr_get_flags(req->rq_reqmsg) & MSGHDR_AT_SUPPORT)) {
+ CDEBUG(D_ADAPTTO, "No early reply support: flags=%#x "
+ "req_flags=%#x magic=%d:%x/%x len=%d\n",
+ flags, lustre_msg_get_flags(req->rq_reqmsg),
+ lustre_msg_is_v1(req->rq_reqmsg),
+ lustre_msg_get_magic(req->rq_reqmsg),
+ lustre_msg_get_magic(req->rq_repmsg), req->rq_replen);
}
}
-int ptlrpc_send_reply (struct ptlrpc_request *req, int may_be_difficult)
+/**
+ * Send request reply from request \a req reply buffer.
+ * \a flags defines reply types
+ * Returns 0 on sucess or error code
+ */
+int ptlrpc_send_reply(struct ptlrpc_request *req, int flags)
{
- struct ptlrpc_service *svc = req->rq_rqbd->rqbd_srv_ni->sni_service;
struct ptlrpc_reply_state *rs = req->rq_reply_state;
struct ptlrpc_connection *conn;
int rc;
/* We must already have a reply buffer (only ptlrpc_error() may be
- * called without one). We usually also have a request buffer which
- * is either the actual (swabbed) incoming request, or a saved copy
- * if this is a req saved in target_queue_final_reply(). but this
- * will not be true since some security handling may skip the reqmsg
- * setting and prepare reply under normal ptlrpc layer */
+ * called without one). The reply generated by sptlrpc layer (e.g.
+ * error notify, etc.) might have NULL rq->reqmsg; Otherwise we must
+ * have a request buffer which is either the actual (swabbed) incoming
+ * request, or a saved copy if this is a req saved in
+ * target_queue_final_reply().
+ */
+ LASSERT (req->rq_no_reply == 0);
+ LASSERT (req->rq_reqbuf != NULL);
LASSERT (rs != NULL);
+ LASSERT ((flags & PTLRPC_REPLY_MAYBE_DIFFICULT) || !rs->rs_difficult);
LASSERT (req->rq_repmsg != NULL);
- LASSERT (may_be_difficult || !rs->rs_difficult);
LASSERT (req->rq_repmsg == rs->rs_msg);
LASSERT (rs->rs_cb_id.cbid_fn == reply_out_callback);
LASSERT (rs->rs_cb_id.cbid_arg == rs);
- LASSERT (req->rq_repmsg != NULL);
+ /* There may be no rq_export during failover */
+
+ if (unlikely(req->rq_export && req->rq_export->exp_obd &&
+ req->rq_export->exp_obd->obd_fail)) {
+ /* Failed obd's only send ENODEV */
+ req->rq_type = PTL_RPC_MSG_ERR;
+ req->rq_status = -ENODEV;
+ CDEBUG(D_HA, "sending ENODEV from failed obd %d\n",
+ req->rq_export->exp_obd->obd_minor);
+ }
+
+ /* In order to keep interoprability with the client (< 2.3) which
+ * doesn't have pb_jobid in ptlrpc_body, We have to shrink the
+ * ptlrpc_body in reply buffer to ptlrpc_body_v2, otherwise, the
+ * reply buffer on client will be overflow.
+ *
+ * XXX Remove this whenver we drop the interoprability with such client.
+ */
+ req->rq_replen = lustre_shrink_msg(req->rq_repmsg, 0,
+ sizeof(struct ptlrpc_body_v2), 1);
+
if (req->rq_type != PTL_RPC_MSG_ERR)
req->rq_type = PTL_RPC_MSG_REPLY;
- req->rq_repmsg->type = req->rq_type;
- req->rq_repmsg->status = req->rq_status;
- req->rq_repmsg->opc = req->rq_reqmsg ? req->rq_reqmsg->opc : 0;
+ lustre_msg_set_type(req->rq_repmsg, req->rq_type);
+ lustre_msg_set_status(req->rq_repmsg,
+ ptlrpc_status_hton(req->rq_status));
+ lustre_msg_set_opc(req->rq_repmsg,
+ req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : 0);
- if (req->rq_export == NULL)
- conn = ptlrpc_get_connection(&req->rq_peer, NULL);
+ target_pack_pool_reply(req);
+
+ ptlrpc_at_set_reply(req, flags);
+
+ if (req->rq_export == NULL || req->rq_export->exp_connection == NULL)
+ conn = ptlrpc_connection_get(req->rq_peer, req->rq_self, NULL);
else
conn = ptlrpc_connection_addref(req->rq_export->exp_connection);
- atomic_inc (&svc->srv_outstanding_replies);
+ if (unlikely(conn == NULL)) {
+ CERROR("not replying on NULL connection\n"); /* bug 9635 */
+ return -ENOTCONN;
+ }
+ ptlrpc_rs_addref(rs); /* +1 ref for the network */
- rc = svcsec_authorize(req);
- if (rc) {
- CERROR("Error wrap reply message "LPX64"\n", req->rq_xid);
+ rc = sptlrpc_svc_wrap_reply(req);
+ if (unlikely(rc))
goto out;
- }
+
+ req->rq_sent = cfs_time_current_sec();
rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len,
- rs->rs_difficult ? PTL_ACK_REQ : PTL_NOACK_REQ,
- &rs->rs_cb_id, conn,
- svc->srv_rep_portal, req->rq_xid);
+ (rs->rs_difficult && !rs->rs_no_ack) ?
+ LNET_ACK_REQ : LNET_NOACK_REQ,
+ &rs->rs_cb_id, conn,
+ ptlrpc_req2svc(req)->srv_rep_portal,
+ req->rq_xid, req->rq_reply_off);
out:
- if (rc != 0) {
- atomic_dec (&svc->srv_outstanding_replies);
-
- if (!rs->rs_difficult) {
- /* Callers other than target_send_reply() expect me
- * to clean up on a comms error */
- lustre_free_reply_state (rs);
- req->rq_reply_state = NULL;
- req->rq_repmsg = NULL;
- }
- }
- ptlrpc_put_connection(conn);
+ if (unlikely(rc != 0))
+ ptlrpc_req_drop_rs(req);
+ ptlrpc_connection_put(conn);
return rc;
}
+EXPORT_SYMBOL(ptlrpc_send_reply);
int ptlrpc_reply (struct ptlrpc_request *req)
{
- return (ptlrpc_send_reply (req, 0));
+ if (req->rq_no_reply)
+ return 0;
+ else
+ return (ptlrpc_send_reply(req, 0));
}
+EXPORT_SYMBOL(ptlrpc_reply);
-int ptlrpc_error(struct ptlrpc_request *req)
+/**
+ * For request \a req send an error reply back. Create empty
+ * reply buffers if necessary.
+ */
+int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult)
{
int rc;
ENTRY;
+ if (req->rq_no_reply)
+ RETURN(0);
+
if (!req->rq_repmsg) {
- rc = lustre_pack_reply(req, 0, NULL, NULL);
+ rc = lustre_pack_reply(req, 1, NULL, NULL);
if (rc)
RETURN(rc);
}
- req->rq_type = PTL_RPC_MSG_ERR;
+ if (req->rq_status != -ENOSPC && req->rq_status != -EACCES &&
+ req->rq_status != -EPERM && req->rq_status != -ENOENT &&
+ req->rq_status != -EINPROGRESS && req->rq_status != -EDQUOT)
+ req->rq_type = PTL_RPC_MSG_ERR;
- rc = ptlrpc_send_reply (req, 0);
+ rc = ptlrpc_send_reply(req, may_be_difficult);
RETURN(rc);
}
+EXPORT_SYMBOL(ptlrpc_send_error);
-int ptl_send_rpc(struct ptlrpc_request *request)
+int ptlrpc_error(struct ptlrpc_request *req)
+{
+ return ptlrpc_send_error(req, 0);
+}
+EXPORT_SYMBOL(ptlrpc_error);
+
+/**
+ * Send request \a request.
+ * if \a noreply is set, don't expect any reply back and don't set up
+ * reply buffers.
+ * Returns 0 on success or error code.
+ */
+int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
{
int rc;
int rc2;
+ int mpflag = 0;
struct ptlrpc_connection *connection;
- unsigned long flags;
- ptl_handle_me_t reply_me_h;
- ptl_md_t reply_md;
+ lnet_handle_me_t reply_me_h;
+ lnet_md_t reply_md;
+ struct obd_import *imp = request->rq_import;
+ struct obd_device *obd = imp->imp_obd;
ENTRY;
- LASSERT (request->rq_type == PTL_RPC_MSG_REQUEST);
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
+ RETURN(0);
+
+ LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST);
+ LASSERT(request->rq_wait_ctx == 0);
/* If this is a re-transmit, we're required to have disengaged
* cleanly from the previous attempt */
- LASSERT (!request->rq_receiving_reply);
+ LASSERT(!request->rq_receiving_reply);
+ LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
+ (imp->imp_state == LUSTRE_IMP_FULL)));
+
+ if (unlikely(obd != NULL && obd->obd_fail)) {
+ CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
+ obd->obd_name);
+ /* this prevents us from waiting in ptlrpc_queue_wait */
+ spin_lock(&request->rq_lock);
+ request->rq_err = 1;
+ spin_unlock(&request->rq_lock);
+ request->rq_status = -ENODEV;
+ RETURN(-ENODEV);
+ }
- connection = request->rq_import->imp_connection;
+ connection = imp->imp_connection;
+
+ lustre_msg_set_handle(request->rq_reqmsg,
+ &imp->imp_remote_handle);
+ lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
+ lustre_msg_set_conn_cnt(request->rq_reqmsg,
+ imp->imp_conn_cnt);
+ lustre_msghdr_set_flags(request->rq_reqmsg,
+ imp->imp_msghdr_flags);
+
+ /** For enabled AT all request should have AT_SUPPORT in the
+ * FULL import state when OBD_CONNECT_AT is set */
+ LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
+ (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
+ !(imp->imp_connect_data.ocd_connect_flags &
+ OBD_CONNECT_AT));
+
+ if (request->rq_resend) {
+ lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
+ if (request->rq_resend_cb != NULL)
+ request->rq_resend_cb(request, &request->rq_async_args);
+ }
+ if (request->rq_memalloc)
+ mpflag = cfs_memory_pressure_get_and_set();
+
+ rc = sptlrpc_cli_wrap_request(request);
+ if (rc)
+ GOTO(out, rc);
+ /* bulk register should be done after wrap_request() */
if (request->rq_bulk != NULL) {
rc = ptlrpc_register_bulk (request);
if (rc != 0)
- RETURN(rc);
+ GOTO(out, rc);
}
- request->rq_reqmsg->handle = request->rq_import->imp_remote_handle;
- request->rq_reqmsg->type = PTL_RPC_MSG_REQUEST;
- request->rq_reqmsg->conn_cnt = request->rq_import->imp_conn_cnt;
-
- /* wrap_request might need to refresh gss cred, if this is called
- * in ptlrpcd then the whole daemon thread will be waiting on
- * gss negotiate rpc. FIXME
- */
- rc = ptlrpcs_cli_wrap_request(request);
- if (rc)
- GOTO(cleanup_bulk, rc);
-
- LASSERT (request->rq_replen != 0);
- if (request->rq_repbuf == NULL) {
- rc = ptlrpcs_cli_alloc_repbuf(request, request->rq_replen);
- if (rc)
- GOTO(cleanup_bulk, rc);
- }
+ if (!noreply) {
+ LASSERT (request->rq_replen != 0);
+ if (request->rq_repbuf == NULL) {
+ LASSERT(request->rq_repdata == NULL);
+ LASSERT(request->rq_repmsg == NULL);
+ rc = sptlrpc_cli_alloc_repbuf(request,
+ request->rq_replen);
+ if (rc) {
+ /* this prevents us from looping in
+ * ptlrpc_queue_wait */
+ spin_lock(&request->rq_lock);
+ request->rq_err = 1;
+ spin_unlock(&request->rq_lock);
+ request->rq_status = rc;
+ GOTO(cleanup_bulk, rc);
+ }
+ } else {
+ request->rq_repdata = NULL;
+ request->rq_repmsg = NULL;
+ }
- rc = PtlMEAttach(connection->c_peer.peer_ni->pni_ni_h,
- request->rq_reply_portal, /* XXX FIXME bug 249 */
- connection->c_peer.peer_id, request->rq_xid, 0,
- PTL_UNLINK, PTL_INS_AFTER, &reply_me_h);
- if (rc != PTL_OK) {
- CERROR("PtlMEAttach failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- GOTO(cleanup_repmsg, rc = -ENOMEM);
+ rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/
+ connection->c_peer, request->rq_xid, 0,
+ LNET_UNLINK, LNET_INS_AFTER, &reply_me_h);
+ if (rc != 0) {
+ CERROR("LNetMEAttach failed: %d\n", rc);
+ LASSERT (rc == -ENOMEM);
+ GOTO(cleanup_bulk, rc = -ENOMEM);
+ }
}
- spin_lock_irqsave (&request->rq_lock, flags);
- /* If the MD attach succeeds, there _will_ be a reply_in callback */
- request->rq_receiving_reply = 1;
- /* Clear any flags that may be present from previous sends. */
+ spin_lock(&request->rq_lock);
+ /* We are responsible for unlinking the reply buffer */
+ request->rq_reply_unlinked = noreply;
+ request->rq_receiving_reply = !noreply;
+ /* Clear any flags that may be present from previous sends. */
+ request->rq_req_unlinked = 0;
request->rq_replied = 0;
request->rq_err = 0;
request->rq_timedout = 0;
request->rq_net_err = 0;
request->rq_resend = 0;
- request->rq_ptlrpcs_restart = 0;
request->rq_restart = 0;
- spin_unlock_irqrestore (&request->rq_lock, flags);
-
- reply_md.start = request->rq_repbuf;
- reply_md.length = request->rq_repbuf_len;
- reply_md.threshold = 1;
- reply_md.options = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
- reply_md.user_ptr = &request->rq_reply_cbid;
- reply_md.eq_handle = connection->c_peer.peer_ni->pni_eq_h;
-
- rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK,
- &request->rq_reply_md_h);
- if (rc != PTL_OK) {
- CERROR("PtlMDAttach failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- GOTO(cleanup_me, rc -ENOMEM);
+ request->rq_reply_truncated = 0;
+ spin_unlock(&request->rq_lock);
+
+ if (!noreply) {
+ reply_md.start = request->rq_repbuf;
+ reply_md.length = request->rq_repbuf_len;
+ /* Allow multiple early replies */
+ reply_md.threshold = LNET_MD_THRESH_INF;
+ /* Manage remote for early replies */
+ reply_md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT |
+ LNET_MD_MANAGE_REMOTE |
+ LNET_MD_TRUNCATE; /* allow to make EOVERFLOW error */;
+ reply_md.user_ptr = &request->rq_reply_cbid;
+ reply_md.eq_handle = ptlrpc_eq_h;
+
+ /* We must see the unlink callback to set rq_reply_unlinked,
+ * so we can't auto-unlink */
+ rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
+ &request->rq_reply_md_h);
+ if (rc != 0) {
+ CERROR("LNetMDAttach failed: %d\n", rc);
+ LASSERT (rc == -ENOMEM);
+ spin_lock(&request->rq_lock);
+ /* ...but the MD attach didn't succeed... */
+ request->rq_receiving_reply = 0;
+ spin_unlock(&request->rq_lock);
+ GOTO(cleanup_me, rc = -ENOMEM);
+ }
+
+ CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
+ ", portal %u\n",
+ request->rq_repbuf_len, request->rq_xid,
+ request->rq_reply_portal);
}
- CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
- ", portal %u on %s\n",
- request->rq_replen, request->rq_xid,
- request->rq_reply_portal,
- connection->c_peer.peer_ni->pni_name);
+ /* add references on request for request_out_callback */
+ ptlrpc_request_addref(request);
+ if (obd != NULL && obd->obd_svc_stats != NULL)
+ lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
+ atomic_read(&imp->imp_inflight));
+
+ OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
+
+ do_gettimeofday(&request->rq_sent_tv);
+ request->rq_sent = cfs_time_current_sec();
+ /* We give the server rq_timeout secs to process the req, and
+ add the network latency for our local timeout. */
+ request->rq_deadline = request->rq_sent + request->rq_timeout +
+ ptlrpc_at_get_net_latency(request);
- ptlrpc_request_addref(request); /* +1 ref for the SENT callback */
+ ptlrpc_pinger_sending_on_import(imp);
- request->rq_sent = LTIME_S(CURRENT_TIME);
- rc = ptl_send_buf(&request->rq_req_md_h,
+ DEBUG_REQ(D_INFO, request, "send flg=%x",
+ lustre_msg_get_flags(request->rq_reqmsg));
+ rc = ptl_send_buf(&request->rq_req_md_h,
request->rq_reqbuf, request->rq_reqdata_len,
- PTL_NOACK_REQ, &request->rq_req_cbid,
+ LNET_NOACK_REQ, &request->rq_req_cbid,
connection,
request->rq_request_portal,
- request->rq_xid);
- if (rc == 0) {
- ptlrpc_lprocfs_rpc_sent(request);
- RETURN(rc);
- }
+ request->rq_xid, 0);
+ if (likely(rc == 0))
+ GOTO(out, rc);
- ptlrpc_req_finished (request); /* drop callback ref */
+ request->rq_req_unlinked = 1;
+ ptlrpc_req_finished(request);
+ if (noreply)
+ GOTO(out, rc);
cleanup_me:
/* MEUnlink is safe; the PUT didn't even get off the ground, and
* nobody apart from the PUT's target has the right nid+XID to
* access the reply buffer. */
- rc2 = PtlMEUnlink(reply_me_h);
- LASSERT (rc2 == PTL_OK);
+ rc2 = LNetMEUnlink(reply_me_h);
+ LASSERT (rc2 == 0);
/* UNLINKED callback called synchronously */
- LASSERT (!request->rq_receiving_reply);
-
- cleanup_repmsg:
- ptlrpcs_cli_free_repbuf(request);
+ LASSERT(!request->rq_receiving_reply);
cleanup_bulk:
- if (request->rq_bulk != NULL)
- ptlrpc_unregister_bulk(request);
-
+ /* We do sync unlink here as there was no real transfer here so
+ * the chance to have long unlink to sluggish net is smaller here. */
+ ptlrpc_unregister_bulk(request, 0);
+ out:
+ if (request->rq_memalloc)
+ cfs_memory_pressure_restore(mpflag);
return rc;
}
+EXPORT_SYMBOL(ptl_send_rpc);
-int ptlrpc_register_rqbd (struct ptlrpc_request_buffer_desc *rqbd)
+/**
+ * Register request buffer descriptor for request receiving.
+ */
+int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd)
{
- struct ptlrpc_srv_ni *srv_ni = rqbd->rqbd_srv_ni;
- struct ptlrpc_service *service = srv_ni->sni_service;
- static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
- int rc;
- ptl_md_t md;
- ptl_handle_me_t me_h;
+ struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service;
+ static lnet_process_id_t match_id = {LNET_NID_ANY, LNET_PID_ANY};
+ int rc;
+ lnet_md_t md;
+ lnet_handle_me_t me_h;
- CDEBUG(D_NET, "PtlMEAttach: portal %d on %s\n",
- service->srv_req_portal, srv_ni->sni_ni->pni_name);
+ CDEBUG(D_NET, "LNetMEAttach: portal %d\n",
+ service->srv_req_portal);
- if (OBD_FAIL_CHECK_ONCE(OBD_FAIL_PTLRPC_RQBD))
+ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_RQBD))
return (-ENOMEM);
- rc = PtlMEAttach(srv_ni->sni_ni->pni_ni_h, service->srv_req_portal,
- match_id, 0, ~0, PTL_UNLINK, PTL_INS_AFTER, &me_h);
- if (rc != PTL_OK) {
- CERROR("PtlMEAttach failed: %d\n", rc);
+ /* NB: CPT affinity service should use new LNet flag LNET_INS_LOCAL,
+ * which means buffer can only be attached on local CPT, and LND
+ * threads can find it by grabbing a local lock */
+ rc = LNetMEAttach(service->srv_req_portal,
+ match_id, 0, ~0, LNET_UNLINK,
+ rqbd->rqbd_svcpt->scp_cpt >= 0 ?
+ LNET_INS_LOCAL : LNET_INS_AFTER, &me_h);
+ if (rc != 0) {
+ CERROR("LNetMEAttach failed: %d\n", rc);
return (-ENOMEM);
}
md.start = rqbd->rqbd_buffer;
md.length = service->srv_buf_size;
md.max_size = service->srv_max_req_size;
- md.threshold = PTL_MD_THRESH_INF;
- md.options = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT | PTL_MD_MAX_SIZE;
+ md.threshold = LNET_MD_THRESH_INF;
+ md.options = PTLRPC_MD_OPTIONS | LNET_MD_OP_PUT | LNET_MD_MAX_SIZE;
md.user_ptr = &rqbd->rqbd_cbid;
- md.eq_handle = srv_ni->sni_ni->pni_eq_h;
-
- rc = PtlMDAttach(me_h, md, PTL_UNLINK, &rqbd->rqbd_md_h);
- if (rc == PTL_OK)
+ md.eq_handle = ptlrpc_eq_h;
+
+ rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h);
+ if (rc == 0)
return (0);
- CERROR("PtlMDAttach failed: %d; \n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- rc = PtlMEUnlink (me_h);
- LASSERT (rc == PTL_OK);
+ CERROR("LNetMDAttach failed: %d; \n", rc);
+ LASSERT (rc == -ENOMEM);
+ rc = LNetMEUnlink (me_h);
+ LASSERT (rc == 0);
rqbd->rqbd_refcount = 0;
-
- return (-ENOMEM);
-}
-
-static int rawrpc_timedout(void *data)
-{
- struct ptlrpc_request *req = (struct ptlrpc_request *) data;
- unsigned long flags;
-
- spin_lock_irqsave(&req->rq_lock, flags);
- if (!req->rq_replied)
- req->rq_timedout = 1;
- spin_unlock_irqrestore(&req->rq_lock, flags);
-
- return 1;
-}
-
-/* to make things as simple as possible */
-static int rawrpc_check_reply(struct ptlrpc_request *req)
-{
- unsigned long flags;
- int rc;
-
- spin_lock_irqsave (&req->rq_lock, flags);
- rc = req->rq_replied || req->rq_net_err || req->rq_err ||
- req->rq_resend || req->rq_restart;
- spin_unlock_irqrestore(&req->rq_lock, flags);
- return rc;
-}
-
-/*
- * Construct a fake ptlrpc_request to do the work, in order to
- * user the existing callback/wakeup facilities
- */
-int ptlrpc_do_rawrpc(struct obd_import *imp,
- char *reqbuf, int reqlen,
- char *repbuf, int *replenp,
- int timeout)
-{
- struct ptlrpc_connection *conn;
- struct ptlrpc_request request; /* just a fake one */
- ptl_handle_me_t reply_me_h;
- ptl_md_t reply_md, req_md;
- struct l_wait_info lwi;
- unsigned long irq_flags;
- int rc;
- ENTRY;
- LASSERT(imp);
- class_import_get(imp);
- if (imp->imp_state == LUSTRE_IMP_CLOSED) {
- CDEBUG(D_SEC, "raw rpc on closed imp(=>%s)? send anyway\n",
- imp->imp_target_uuid.uuid);
- }
-
- conn = imp->imp_connection;
-
- /* initialize request */
- memset(&request, 0, sizeof(request));
- request.rq_req_cbid.cbid_fn = request_out_callback;
- request.rq_req_cbid.cbid_arg = &request;
- request.rq_reply_cbid.cbid_fn = reply_in_callback;
- request.rq_reply_cbid.cbid_arg = &request;
- request.rq_reqbuf = reqbuf;
- request.rq_reqbuf_len = reqlen;
- request.rq_repbuf = repbuf;
- request.rq_repbuf_len = *replenp;
- request.rq_set = NULL;
- spin_lock_init(&request.rq_lock);
- init_waitqueue_head(&request.rq_reply_waitq);
- atomic_set(&request.rq_refcount, 1000000); /* never be droped */
- request.rq_xid = ptlrpc_next_xid();
-
- /* add into sending list */
- spin_lock_irqsave(&imp->imp_lock, irq_flags);
- list_add_tail(&request.rq_list, &imp->imp_rawrpc_list);
- spin_unlock_irqrestore(&imp->imp_lock, irq_flags);
-
- /* prepare reply buffer */
- rc = PtlMEAttach(conn->c_peer.peer_ni->pni_ni_h,
- imp->imp_client->cli_reply_portal,
- conn->c_peer.peer_id, request.rq_xid, 0, PTL_UNLINK,
- PTL_INS_AFTER, &reply_me_h);
- if (rc != PTL_OK) {
- CERROR("PtlMEAttach failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- GOTO(cleanup_imp, rc = -ENOMEM);
- }
-
- spin_lock_irqsave(&request.rq_lock, irq_flags);
- request.rq_receiving_reply = 1;
- spin_unlock_irqrestore(&request.rq_lock, irq_flags);
-
- reply_md.start = repbuf;
- reply_md.length = *replenp;
- reply_md.threshold = 1;
- reply_md.options = PTLRPC_MD_OPTIONS | PTL_MD_OP_PUT;
- reply_md.user_ptr = &request.rq_reply_cbid;
- reply_md.eq_handle = conn->c_peer.peer_ni->pni_eq_h;
-
- rc = PtlMDAttach(reply_me_h, reply_md, PTL_UNLINK,
- &request.rq_reply_md_h);
- if (rc != PTL_OK) {
- CERROR("PtlMDAttach failed: %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- GOTO(cleanup_me, rc = -ENOMEM);
- }
-
- /* prepare request buffer */
- req_md.start = reqbuf;
- req_md.length = reqlen;
- req_md.threshold = 1;
- req_md.options = PTLRPC_MD_OPTIONS;
- req_md.user_ptr = &request.rq_req_cbid;
- req_md.eq_handle = conn->c_peer.peer_ni->pni_eq_h;
-
- rc = PtlMDBind(conn->c_peer.peer_ni->pni_ni_h,
- req_md, PTL_UNLINK, &request.rq_req_md_h);
- if (rc != PTL_OK) {
- CERROR("PtlMDBind failed %d\n", rc);
- LASSERT (rc == PTL_NO_SPACE);
- GOTO(cleanup_me, rc = -ENOMEM);
- }
-
- rc = PtlPut(request.rq_req_md_h, PTL_NOACK_REQ, conn->c_peer.peer_id,
- imp->imp_client->cli_request_portal,
- 0, request.rq_xid, 0, 0);
- if (rc != PTL_OK) {
- CERROR("PtlPut failed %d\n", rc);
- GOTO(cleanup_md, rc);
- }
-
- lwi = LWI_TIMEOUT(timeout * HZ, rawrpc_timedout, &request);
- l_wait_event(request.rq_reply_waitq,
- rawrpc_check_reply(&request), &lwi);
-
- ptlrpc_unregister_reply(&request);
-
- if (request.rq_err || request.rq_resend || request.rq_intr ||
- request.rq_timedout || !request.rq_replied) {
- CERROR("secinit rpc error: err %d, resend %d, "
- "intr %d, timeout %d, replied %d\n",
- request.rq_err, request.rq_resend, request.rq_intr,
- request.rq_timedout, request.rq_replied);
- rc = -EINVAL;
- } else {
- *replenp = request.rq_nob_received;
- rc = 0;
- }
- GOTO(cleanup_imp, rc);
-
-cleanup_md:
- PtlMDUnlink(request.rq_req_md_h);
-cleanup_me:
- PtlMEUnlink(reply_me_h);
-cleanup_imp:
- spin_lock_irqsave(&imp->imp_lock, irq_flags);
- list_del_init(&request.rq_list);
- spin_unlock_irqrestore(&imp->imp_lock, irq_flags);
-
- class_import_put(imp);
- RETURN(rc);
+ return (-ENOMEM);
}
git://git.whamcloud.com / fs / lustre-release.git / blobdiff