X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;ds=sidebyside;f=lustre%2Fptlrpc%2Fniobuf.c;h=4720592e5ebf4477880615f5a2b3f2f2c4638728;hb=917655fc2938b90a9c246dd2d58408c42aa1658d;hp=a695e5789538164d975c7cc1f8fbe3bf24c1e8f3;hpb=294e507d34de0d7990ec507334f71ff1d1a7db01;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c index a695e57..4720592 100644 --- a/lustre/ptlrpc/niobuf.c +++ b/lustre/ptlrpc/niobuf.c @@ -15,11 +15,7 @@ * * 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 - * - * 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. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2013, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -35,73 +31,77 @@ */ #define DEBUG_SUBSYSTEM S_RPC -#ifndef __KERNEL__ -#include -#endif #include #include #include #include #include #include "ptlrpc_internal.h" +#include /* for CFS_FAIL_PTLRPC_OST_BULK_CB2 */ /** * 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) +static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len, + enum lnet_ack_req ack, struct ptlrpc_cb_id *cbid, + lnet_nid_t self, struct lnet_process_id peer_id, + int portal, __u64 xid, unsigned int offset, + struct lnet_handle_md *bulk_cookie) { - int rc; - lnet_md_t md; - ENTRY; + int rc; + struct lnet_md md; + ENTRY; - LASSERT (portal != 0); - LASSERT (conn != NULL); - CDEBUG (D_INFO, "conn=%p id %s\n", conn, libcfs_id2str(conn->c_peer)); - md.start = base; - md.length = len; - md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1; - md.options = PTLRPC_MD_OPTIONS; - md.user_ptr = cbid; - md.eq_handle = ptlrpc_eq_h; + LASSERT (portal != 0); + CDEBUG (D_INFO, "peer_id %s\n", libcfs_id2str(peer_id)); + md.start = base; + md.length = len; + md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1; + md.options = PTLRPC_MD_OPTIONS; + md.user_ptr = cbid; + md.eq_handle = ptlrpc_eq_h; + LNetInvalidateMDHandle(&md.bulk_handle); - 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 = LNET_NOACK_REQ; - } + if (bulk_cookie) { + md.bulk_handle = *bulk_cookie; + md.options |= LNET_MD_BULK_HANDLE; + } - rc = LNetMDBind (md, LNET_UNLINK, mdh); - if (unlikely(rc != 0)) { - CERROR ("LNetMDBind failed: %d\n", rc); - LASSERT (rc == -ENOMEM); - RETURN (-ENOMEM); - } + 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 = LNET_NOACK_REQ; + } - CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64", offset %u\n", - len, portal, xid, offset); - - 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("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); - } + rc = LNetMDBind (md, LNET_UNLINK, mdh); + if (unlikely(rc != 0)) { + CERROR ("LNetMDBind failed: %d\n", rc); + LASSERT (rc == -ENOMEM); + RETURN (-ENOMEM); + } - RETURN (0); + CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n", + len, portal, xid, offset); + + rc = LNetPut(self, *mdh, ack, + peer_id, 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("LNetPut(%s, %d, %lld) failed: %d\n", + libcfs_id2str(peer_id), portal, xid, rc); + rc2 = LNetMDUnlink(*mdh); + LASSERTF(rc2 == 0, "rc2 = %d\n", rc2); + } + + RETURN (0); } -static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count) +static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count) { int i; @@ -112,23 +112,26 @@ static void mdunlink_iterate_helper(lnet_handle_md_t *bd_mds, int count) #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 + * can fit \a nfrags * 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) + unsigned nfrags, unsigned max_brw, + unsigned int type, + unsigned portal, + const struct ptlrpc_bulk_frag_ops + *ops) { struct obd_export *exp = req->rq_export; struct ptlrpc_bulk_desc *desc; ENTRY; - LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK); + LASSERT(ptlrpc_is_bulk_op_active(type)); - desc = ptlrpc_new_bulk(npages, max_brw, type, portal); + desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops); if (desc == NULL) RETURN(NULL); @@ -152,12 +155,13 @@ EXPORT_SYMBOL(ptlrpc_prep_bulk_exp); int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) { struct obd_export *exp = desc->bd_export; - struct ptlrpc_connection *conn = exp->exp_connection; + lnet_nid_t self_nid; + struct lnet_process_id peer_id; int rc = 0; - __u64 xid; + __u64 mbits; int posted_md; int total_md; - lnet_md_t md; + struct lnet_md md; ENTRY; if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_PUT_NET)) @@ -165,20 +169,27 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) /* 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(ptlrpc_is_bulk_op_active(desc->bd_type)); LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback); LASSERT(desc->bd_cbid.cbid_arg == desc); + /* + * Multi-Rail: get the preferred self and peer NIDs from the + * request, so they are based on the route taken by the + * message. + */ + self_nid = desc->bd_req->rq_self; + peer_id = desc->bd_req->rq_source; + /* 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 + * The client may not need all of the bulk mbits for the RPC. The RPC + * used the mbits of the highest bulk mbits 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; + mbits = desc->bd_req->rq_mbits & ~((__u64)desc->bd_md_max_brw - 1); + total_md = desc->bd_req->rq_mbits - mbits + 1; desc->bd_md_count = total_md; desc->bd_failure = 0; @@ -187,7 +198,7 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) md.eq_handle = ptlrpc_eq_h; md.threshold = 2; /* SENT and ACK/REPLY */ - for (posted_md = 0; posted_md < total_md; xid++) { + for (posted_md = 0; posted_md < total_md; mbits++) { md.options = PTLRPC_MD_OPTIONS; /* NB it's assumed that source and sink buffer frags are @@ -205,21 +216,27 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) } break; } + + /* sanity.sh 224c: lets skip last md */ + if (posted_md == desc->bd_md_max_brw - 1) + OBD_FAIL_CHECK_RESET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3, + CFS_FAIL_PTLRPC_OST_BULK_CB2); + /* 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); + if (ptlrpc_is_bulk_put_source(desc->bd_type)) + rc = LNetPut(self_nid, desc->bd_mds[posted_md], + LNET_ACK_REQ, peer_id, + desc->bd_portal, mbits, 0, 0); else - rc = LNetGet(conn->c_self, desc->bd_mds[posted_md], - conn->c_peer, desc->bd_portal, xid, 0); + rc = LNetGet(self_nid, desc->bd_mds[posted_md], + peer_id, desc->bd_portal, mbits, 0, false); posted_md++; if (rc != 0) { - CERROR("%s: failed bulk transfer with %s:%u x"LPU64": " + CERROR("%s: failed bulk transfer with %s:%u x%llu: " "rc = %d\n", exp->exp_obd->obd_name, - libcfs_id2str(conn->c_peer), desc->bd_portal, - xid, rc); + libcfs_id2str(peer_id), desc->bd_portal, + mbits, rc); break; } } @@ -238,13 +255,12 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) } 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); + "id %s mbits %#llx-%#llx\n", desc->bd_iov_count, + desc->bd_nob, desc->bd_portal, libcfs_id2str(peer_id), + mbits - posted_md, mbits - 1); RETURN(0); } -EXPORT_SYMBOL(ptlrpc_start_bulk_transfer); /** * Server side bulk abort. Idempotent. Not thread-safe (i.e. only @@ -269,7 +285,7 @@ void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc) * 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); + mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw); for (;;) { /* Network access will complete in finite time but the HUGE @@ -285,7 +301,6 @@ void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc) CWARN("Unexpectedly long timeout: desc %p\n", desc); } } -EXPORT_SYMBOL(ptlrpc_abort_bulk); #endif /* HAVE_SERVER_SUPPORT */ /** @@ -295,14 +310,14 @@ EXPORT_SYMBOL(ptlrpc_abort_bulk); int ptlrpc_register_bulk(struct ptlrpc_request *req) { struct ptlrpc_bulk_desc *desc = req->rq_bulk; - lnet_process_id_t peer; + struct lnet_process_id peer; int rc = 0; int rc2; int posted_md; int total_md; - __u64 xid; - lnet_handle_me_t me_h; - lnet_md_t md; + __u64 mbits; + struct lnet_handle_me me_h; + struct lnet_md md; ENTRY; if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_GET_NET)) @@ -314,14 +329,16 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) 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); + LASSERT(ptlrpc_is_bulk_op_passive(desc->bd_type)); /* 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); + else if (desc->bd_nob_transferred != 0) + /* If the network failed after an RPC was sent, this condition + * could happen. Rather than assert (was here before), return + * an EIO error. */ + RETURN(-EIO); desc->bd_failure = 0; @@ -330,39 +347,42 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) 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); + total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV; + /* rq_mbits is matchbits of the final bulk */ + mbits = req->rq_mbits - total_md + 1; + + LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK), + "first mbits = x%llu, last mbits = x%llu\n", + mbits, req->rq_mbits); 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); + mbits != desc->bd_last_mbits, + "registered: %d rq_mbits: %llu bd_last_mbits: %llu\n", + desc->bd_registered, mbits, desc->bd_last_mbits); - total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV; desc->bd_registered = 1; - desc->bd_last_xid = xid; + desc->bd_last_mbits = mbits; 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++) { + for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) { md.options = PTLRPC_MD_OPTIONS | - ((desc->bd_type == BULK_GET_SOURCE) ? + (ptlrpc_is_bulk_op_get(desc->bd_type) ? LNET_MD_OP_GET : LNET_MD_OP_PUT); ptlrpc_fill_bulk_md(&md, desc, posted_md); - rc = LNetMEAttach(desc->bd_portal, peer, xid, 0, + if (posted_md > 0 && posted_md + 1 == total_md && + OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) { + rc = -ENOMEM; + } else { + rc = LNetMEAttach(desc->bd_portal, peer, mbits, 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, + CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n", + desc->bd_import->imp_obd->obd_name, mbits, posted_md, rc); break; } @@ -371,8 +391,8 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) 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, + CERROR("%s: LNetMDAttach failed x%llu/%d: rc = %d\n", + desc->bd_import->imp_obd->obd_name, mbits, posted_md, rc); rc2 = LNetMEUnlink(me_h); LASSERT(rc2 == 0); @@ -388,18 +408,12 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) LASSERT(desc->bd_md_count >= 0); mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw); req->rq_status = -ENOMEM; + desc->bd_registered = 0; 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 */ + /* Holler if peer manages to touch buffers before he knows the mbits */ 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), @@ -407,14 +421,13 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) 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", + "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count, + ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink", desc->bd_iov_count, desc->bd_nob, - desc->bd_last_xid, req->rq_xid, desc->bd_portal); + desc->bd_last_mbits, req->rq_mbits, desc->bd_portal); RETURN(0); } -EXPORT_SYMBOL(ptlrpc_register_bulk); /** * Disconnect a bulk desc from the network. Idempotent. Not @@ -433,8 +446,8 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) /* 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; + async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0) + req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK; if (ptlrpc_client_bulk_active(req) == 0) /* completed or */ RETURN(1); /* never registered */ @@ -450,43 +463,42 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) 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); + /* Move to "Unregistering" phase as bulk was not unlinked yet. */ + ptlrpc_rqphase_move(req, RQ_PHASE_UNREG_BULK); - /* Do not wait for unlink to finish. */ - if (async) - RETURN(0); + /* Do not wait for unlink to finish. */ + if (async) + RETURN(0); - for (;;) { -#ifdef __KERNEL__ + 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; -#endif - /* 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); - } + /* + * 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); + } - LASSERT(rc == -ETIMEDOUT); - DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p", - desc); - } - RETURN(0); + LASSERT(rc == -ETIMEDOUT); + DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p", + desc); + } + RETURN(0); } -EXPORT_SYMBOL(ptlrpc_unregister_bulk); static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) { 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() - + int service_time = max_t(int, ktime_get_real_seconds() - req->rq_arrival_time.tv_sec, 1); if (!(flags & PTLRPC_REPLY_EARLY) && @@ -508,37 +520,53 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) } /* 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 - 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); - } + /* Report service time estimate for future client reqs, but report 0 + * (to be ignored by client) if it's an error reply during recovery. + * b=15815 + */ + 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 { + time64_t timeout; + + if (req->rq_export && req->rq_reqmsg != NULL && + (flags & PTLRPC_REPLY_EARLY) && + lustre_msg_get_flags(req->rq_reqmsg) & + (MSG_REPLAY | MSG_REQ_REPLAY_DONE | MSG_LOCK_REPLAY_DONE)) { + struct obd_device *exp_obd = req->rq_export->exp_obd; + + timeout = ktime_get_real_seconds() - + req->rq_arrival_time.tv_sec + + min_t(time64_t, at_extra, + exp_obd->obd_recovery_timeout / 4); + } else { + timeout = at_get(&svcpt->scp_at_estimate); + } + lustre_msg_set_timeout(req->rq_repmsg, timeout); + } + + 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=%x/%x len=%d\n", + flags, lustre_msg_get_flags(req->rq_reqmsg), + lustre_msg_get_magic(req->rq_reqmsg), + lustre_msg_get_magic(req->rq_repmsg), req->rq_replen); + } } /** * Send request reply from request \a req reply buffer. * \a flags defines reply types - * Returns 0 on sucess or error code + * Returns 0 on success or error code */ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) { - struct ptlrpc_reply_state *rs = req->rq_reply_state; - struct ptlrpc_connection *conn; - int rc; + 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). The reply generated by sptlrpc layer (e.g. @@ -567,16 +595,6 @@ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) 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; @@ -605,21 +623,20 @@ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) if (unlikely(rc)) goto out; - req->rq_sent = cfs_time_current_sec(); + req->rq_sent = ktime_get_real_seconds(); - rc = ptl_send_buf (&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len, - (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); + rc = ptl_send_buf(&rs->rs_md_h, rs->rs_repbuf, rs->rs_repdata_len, + (rs->rs_difficult && !rs->rs_no_ack) ? + LNET_ACK_REQ : LNET_NOACK_REQ, + &rs->rs_cb_id, req->rq_self, req->rq_source, + ptlrpc_req2svc(req)->srv_rep_portal, + req->rq_xid, req->rq_reply_off, NULL); out: 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) { @@ -628,7 +645,6 @@ int ptlrpc_reply (struct ptlrpc_request *req) else return (ptlrpc_send_reply(req, 0)); } -EXPORT_SYMBOL(ptlrpc_reply); /** * For request \a req send an error reply back. Create empty @@ -656,13 +672,11 @@ int ptlrpc_send_error(struct ptlrpc_request *req, int may_be_difficult) rc = ptlrpc_send_reply(req, may_be_difficult); RETURN(rc); } -EXPORT_SYMBOL(ptlrpc_send_error); int ptlrpc_error(struct ptlrpc_request *req) { return ptlrpc_send_error(req, 0); } -EXPORT_SYMBOL(ptlrpc_error); /** * Send request \a request. @@ -672,14 +686,18 @@ EXPORT_SYMBOL(ptlrpc_error); */ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) { - int rc; - int rc2; - int mpflag = 0; - struct ptlrpc_connection *connection; - lnet_handle_me_t reply_me_h; - lnet_md_t reply_md; - struct obd_device *obd = request->rq_import->imp_obd; - ENTRY; + int rc; + int rc2; + int mpflag = 0; + struct lnet_handle_md bulk_cookie; + struct ptlrpc_connection *connection; + struct lnet_handle_me reply_me_h; + struct lnet_md reply_md; + struct obd_import *imp = request->rq_import; + struct obd_device *obd = imp->imp_obd; + ENTRY; + + LNetInvalidateMDHandle(&bulk_cookie); if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC)) RETURN(0); @@ -690,12 +708,13 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) /* If this is a re-transmit, we're required to have disengaged * cleanly from the previous attempt */ LASSERT(!request->rq_receiving_reply); + LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) && + (imp->imp_state == LUSTRE_IMP_FULL))); - if (request->rq_import->imp_obd && - request->rq_import->imp_obd->obd_fail) { - CDEBUG(D_HA, "muting rpc for failed imp obd %s\n", - request->rq_import->imp_obd->obd_name); - /* this prevents us from waiting in ptlrpc_queue_wait */ + 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); @@ -703,32 +722,79 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) RETURN(-ENODEV); } - connection = request->rq_import->imp_connection; - - lustre_msg_set_handle(request->rq_reqmsg, - &request->rq_import->imp_remote_handle); - lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST); - lustre_msg_set_conn_cnt(request->rq_reqmsg, - request->rq_import->imp_conn_cnt); - lustre_msghdr_set_flags(request->rq_reqmsg, - request->rq_import->imp_msghdr_flags); - - if (request->rq_resend) - lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT); + 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); + + /* If it's the first time to resend the request for EINPROGRESS, + * we need to allocate a new XID (see after_reply()), it's different + * from the resend for reply timeout. */ + if (request->rq_nr_resend != 0 && + list_empty(&request->rq_unreplied_list)) { + __u64 min_xid = 0; + /* resend for EINPROGRESS, allocate new xid to avoid reply + * reconstruction */ + spin_lock(&imp->imp_lock); + ptlrpc_assign_next_xid_nolock(request); + min_xid = ptlrpc_known_replied_xid(imp); + spin_unlock(&imp->imp_lock); + + lustre_msg_set_last_xid(request->rq_reqmsg, min_xid); + DEBUG_REQ(D_RPCTRACE, request, + "Allocating new XID for resend on EINPROGRESS"); + } - if (request->rq_memalloc) - mpflag = cfs_memory_pressure_get_and_set(); + if (request->rq_bulk != NULL) { + ptlrpc_set_bulk_mbits(request); + lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits); + } - rc = sptlrpc_cli_wrap_request(request); - if (rc) - GOTO(out, rc); + if (list_empty(&request->rq_unreplied_list) || + request->rq_xid <= imp->imp_known_replied_xid) { + DEBUG_REQ(D_ERROR, request, + "xid=%llu, replied=%llu, list_empty=%d", + request->rq_xid, imp->imp_known_replied_xid, + list_empty(&request->rq_unreplied_list)); + LBUG(); + } - /* bulk register should be done after wrap_request() */ - if (request->rq_bulk != NULL) { - rc = ptlrpc_register_bulk (request); - if (rc != 0) - GOTO(out, rc); - } + /** 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) + GOTO(cleanup_bulk, rc); + /* + * All the mds in the request will have the same cpt + * encoded in the cookie. So we can just get the first + * one. + */ + bulk_cookie = request->rq_bulk->bd_mds[0]; + } if (!noreply) { LASSERT (request->rq_replen != 0); @@ -743,13 +809,13 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) 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; - } + request->rq_status = rc; + GOTO(cleanup_bulk, rc); + } + } else { + request->rq_repdata = NULL; + request->rq_repmsg = NULL; + } rc = LNetMEAttach(request->rq_reply_portal,/*XXX FIXME bug 249*/ connection->c_peer, request->rq_xid, 0, @@ -762,18 +828,18 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) } spin_lock(&request->rq_lock); - /* If the MD attach succeeds, there _will_ be a reply_in callback */ - request->rq_receiving_reply = !noreply; - /* We are responsible for unlinking the reply buffer */ - request->rq_must_unlink = !noreply; - /* Clear any flags that may be present from previous sends. */ + /* 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_restart = 0; - request->rq_reply_truncate = 0; + request->rq_reply_truncated = 0; spin_unlock(&request->rq_lock); if (!noreply) { @@ -788,8 +854,8 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) reply_md.user_ptr = &request->rq_reply_cbid; reply_md.eq_handle = ptlrpc_eq_h; - /* We must see the unlink callback to unset rq_must_unlink, - so we can't auto-unlink */ + /* 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) { @@ -802,7 +868,7 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) GOTO(cleanup_me, rc = -ENOMEM); } - CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64 + CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu" ", portal %u\n", request->rq_repbuf_len, request->rq_xid, request->rq_reply_portal); @@ -810,53 +876,63 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) /* add references on request for request_out_callback */ ptlrpc_request_addref(request); - if (obd->obd_svc_stats != NULL) - lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR, - atomic_read(&request->rq_import->imp_inflight)); + 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_arrival_time); - request->rq_sent = cfs_time_current_sec(); + request->rq_sent_ns = ktime_get_real(); + request->rq_sent = ktime_get_real_seconds(); /* 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_pinger_sending_on_import(request->rq_import); + ptlrpc_pinger_sending_on_import(imp); - 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, - LNET_NOACK_REQ, &request->rq_req_cbid, - connection, - request->rq_request_portal, - request->rq_xid, 0); - if (rc == 0) - GOTO(out, rc); + DEBUG_REQ(D_INFO, request, "send flags=%x", + lustre_msg_get_flags(request->rq_reqmsg)); + rc = ptl_send_buf(&request->rq_req_md_h, + request->rq_reqbuf, request->rq_reqdata_len, + LNET_NOACK_REQ, &request->rq_req_cbid, + LNET_NID_ANY, connection->c_peer, + request->rq_request_portal, + request->rq_xid, 0, &bulk_cookie); + if (likely(rc == 0)) + GOTO(out, rc); + 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 = LNetMEUnlink(reply_me_h); - LASSERT (rc2 == 0); - /* UNLINKED callback called synchronously */ - LASSERT(!request->rq_receiving_reply); + /* 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 = LNetMEUnlink(reply_me_h); + LASSERT (rc2 == 0); + /* UNLINKED callback called synchronously */ + LASSERT(!request->rq_receiving_reply); cleanup_bulk: - /* 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. */ + /* 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); + if (request->rq_bulk != NULL) + request->rq_bulk->bd_registered = 0; out: - if (request->rq_memalloc) - cfs_memory_pressure_restore(mpflag); - return rc; + if (rc == -ENOMEM) { + /* set rq_sent so that this request is treated + * as a delayed send in the upper layers */ + request->rq_sent = ktime_get_real_seconds(); + } + + if (request->rq_memalloc) + cfs_memory_pressure_restore(mpflag); + + return rc; } EXPORT_SYMBOL(ptl_send_rpc); @@ -865,11 +941,14 @@ EXPORT_SYMBOL(ptl_send_rpc); */ int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd) { - 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; + struct ptlrpc_service *service = rqbd->rqbd_svcpt->scp_service; + static struct lnet_process_id match_id = { + .nid = LNET_NID_ANY, + .pid = LNET_PID_ANY + }; + int rc; + struct lnet_md md; + struct lnet_handle_me me_h; CDEBUG(D_NET, "LNetMEAttach: portal %d\n", service->srv_req_portal);