X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fptlrpc%2Fniobuf.c;h=bd8f0d4e46810578e05ca45d847bb2788bd5f28a;hb=db49fbf00d24edc8347006241f314fb0f82f1b63;hp=ee7163c01b6f3c4fdccfa293328c204b263d8d4f;hpb=3af10d510078ddf050fbda2ef80bdc6920d89f8d;p=fs%2Flustre-release.git diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c index ee7163c..bd8f0d4 100644 --- a/lustre/ptlrpc/niobuf.c +++ b/lustre/ptlrpc/niobuf.c @@ -23,20 +23,21 @@ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2016, Intel Corporation. + * Copyright (c) 2012, 2017, 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 +#include #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 @@ -60,7 +61,7 @@ static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len, md.threshold = (ack == LNET_ACK_REQ) ? 2 : 1; md.options = PTLRPC_MD_OPTIONS; md.user_ptr = cbid; - md.eq_handle = ptlrpc_eq_h; + md.handler = ptlrpc_handler; LNetInvalidateMDHandle(&md.bulk_handle); if (bulk_cookie) { @@ -74,7 +75,7 @@ static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len, ack = LNET_NOACK_REQ; } - rc = LNetMDBind (md, LNET_UNLINK, mdh); + rc = LNetMDBind(&md, LNET_UNLINK, mdh); if (unlikely(rc != 0)) { CERROR ("LNetMDBind failed: %d\n", rc); LASSERT (rc == -ENOMEM); @@ -84,6 +85,8 @@ static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len, CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %lld, offset %u\n", len, portal, xid, offset); + percpu_ref_get(&ptlrpc_pending); + rc = LNetPut(self, *mdh, ack, peer_id, portal, xid, offset, 0); if (unlikely(rc != 0)) { @@ -100,12 +103,15 @@ static int ptl_send_buf(struct lnet_handle_md *mdh, void *base, int len, RETURN (0); } -static void mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, int count) +#define mdunlink_iterate_helper(mds, count) \ + __mdunlink_iterate_helper(mds, count, false) +static void __mdunlink_iterate_helper(struct lnet_handle_md *bd_mds, + int count, bool discard) { int i; for (i = 0; i < count; i++) - LNetMDUnlink(bd_mds[i]); + __LNetMDUnlink(bd_mds[i], discard); } #ifdef HAVE_SERVER_SUPPORT @@ -167,7 +173,6 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) RETURN(0); /* NB no locking required until desc is on the network */ - LASSERT(desc->bd_md_count == 0); LASSERT(ptlrpc_is_bulk_op_active(desc->bd_type)); LASSERT(desc->bd_cbid.cbid_fn == server_bulk_callback); @@ -189,12 +194,11 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) * off high bits to get bulk count for this RPC. LU-1431 */ 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_refs = total_md; desc->bd_failure = 0; md.user_ptr = &desc->bd_cbid; - md.eq_handle = ptlrpc_eq_h; + md.handler = ptlrpc_handler; md.threshold = 2; /* SENT and ACK/REPLY */ for (posted_md = 0; posted_md < total_md; mbits++) { @@ -204,7 +208,7 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) * 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]); + 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); @@ -215,14 +219,12 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) } break; } + percpu_ref_get(&ptlrpc_pending); - /* LU-6441: last md is not sent and desc->bd_md_count == 1 */ - if (OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB3, - CFS_FAIL_ONCE) && - total_md > 1 && posted_md == total_md - 1) { - posted_md++; - continue; - } + /* 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 (ptlrpc_is_bulk_put_source(desc->bd_type)) @@ -231,7 +233,7 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) desc->bd_portal, mbits, 0, 0); else rc = LNetGet(self_nid, desc->bd_mds[posted_md], - peer_id, desc->bd_portal, mbits, 0); + peer_id, desc->bd_portal, mbits, 0, false); posted_md++; if (rc != 0) { @@ -248,9 +250,9 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) * 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; + desc->bd_refs -= total_md - posted_md; spin_unlock(&desc->bd_lock); - LASSERT(desc->bd_md_count >= 0); + LASSERT(desc->bd_refs >= 0); mdunlink_iterate_helper(desc->bd_mds, posted_md); RETURN(0); @@ -270,9 +272,6 @@ int ptlrpc_start_bulk_transfer(struct ptlrpc_bulk_desc *desc) */ 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 */ @@ -285,21 +284,24 @@ void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc) /* 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_max_brw); + * but we must still wait_event_idle_timeout() in this case, to give + * us a chance to run server_bulk_callback() + */ + __mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw, true); 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) + int seconds = PTLRPC_REQ_LONG_UNLINK; + + while (seconds > 0 && + wait_event_idle_timeout(desc->bd_waitq, + !ptlrpc_server_bulk_active(desc), + cfs_time_seconds(1)) == 0) + seconds -= 1; + if (seconds > 0) return; - LASSERT(rc == -ETIMEDOUT); CWARN("Unexpectedly long timeout: desc %p\n", desc); } } @@ -312,13 +314,12 @@ void ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc) int ptlrpc_register_bulk(struct ptlrpc_request *req) { struct ptlrpc_bulk_desc *desc = req->rq_bulk; - struct lnet_process_id peer; + struct lnet_processid peer; int rc = 0; - int rc2; int posted_md; int total_md; __u64 mbits; - struct lnet_handle_me me_h; + struct lnet_me *me; struct lnet_md md; ENTRY; @@ -327,7 +328,6 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) /* 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); @@ -344,14 +344,16 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) desc->bd_failure = 0; - peer = desc->bd_import->imp_connection->c_peer; + peer.pid = desc->bd_import->imp_connection->c_peer.pid; + lnet_nid4_to_nid(desc->bd_import->imp_connection->c_peer.nid, + &peer.nid); LASSERT(desc->bd_cbid.cbid_fn == client_bulk_callback); LASSERT(desc->bd_cbid.cbid_arg == desc); - total_md = (desc->bd_iov_count + LNET_MAX_IOV - 1) / LNET_MAX_IOV; + total_md = desc->bd_md_count; /* rq_mbits is matchbits of the final bulk */ - mbits = req->rq_mbits - total_md + 1; + mbits = req->rq_mbits - desc->bd_md_count + 1; LASSERTF(mbits == (req->rq_mbits & PTLRPC_BULK_OPS_MASK), "first mbits = x%llu, last mbits = x%llu\n", @@ -364,35 +366,41 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) desc->bd_registered = 1; desc->bd_last_mbits = mbits; - desc->bd_md_count = total_md; + desc->bd_refs = total_md; md.user_ptr = &desc->bd_cbid; - md.eq_handle = ptlrpc_eq_h; + md.handler = ptlrpc_handler; md.threshold = 1; /* PUT or GET */ - for (posted_md = 0; posted_md < total_md; posted_md++, mbits++) { + for (posted_md = 0; posted_md < desc->bd_md_count; + posted_md++, mbits++) { md.options = PTLRPC_MD_OPTIONS | (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, mbits, 0, - LNET_UNLINK, LNET_INS_AFTER, &me_h); + if (posted_md > 0 && posted_md + 1 == desc->bd_md_count && + OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_ATTACH)) { + rc = -ENOMEM; + } else { + me = LNetMEAttach(desc->bd_portal, &peer, mbits, 0, + LNET_UNLINK, LNET_INS_AFTER); + rc = PTR_ERR_OR_ZERO(me); + } if (rc != 0) { CERROR("%s: LNetMEAttach failed x%llu/%d: rc = %d\n", desc->bd_import->imp_obd->obd_name, mbits, posted_md, rc); break; } + percpu_ref_get(&ptlrpc_pending); /* About to let the network at it... */ - rc = LNetMDAttach(me_h, md, LNET_UNLINK, + rc = LNetMDAttach(me, &md, LNET_UNLINK, &desc->bd_mds[posted_md]); if (rc != 0) { 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); break; } } @@ -400,24 +408,26 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) if (rc != 0) { LASSERT(rc == -ENOMEM); spin_lock(&desc->bd_lock); - desc->bd_md_count -= total_md - posted_md; + desc->bd_refs -= total_md - posted_md; spin_unlock(&desc->bd_lock); - LASSERT(desc->bd_md_count >= 0); + LASSERT(desc->bd_refs >= 0); mdunlink_iterate_helper(desc->bd_mds, desc->bd_md_max_brw); req->rq_status = -ENOMEM; + desc->bd_registered = 0; RETURN(-ENOMEM); } spin_lock(&desc->bd_lock); /* Holler if peer manages to touch buffers before he knows the mbits */ - if (desc->bd_md_count != total_md) + if (desc->bd_refs != 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); + desc->bd_import->imp_obd->obd_name, libcfs_idstr(&peer), + total_md - desc->bd_refs); spin_unlock(&desc->bd_lock); - CDEBUG(D_NET, "Setup %u bulk %s buffers: %u pages %u bytes, " - "mbits x%#llx-%#llx, portal %u\n", desc->bd_md_count, + CDEBUG(D_NET, + "Setup %u bulk %s buffers: %u pages %u bytes, mbits x%#llx-%#llx, portal %u\n", + desc->bd_refs, ptlrpc_is_bulk_op_get(desc->bd_type) ? "get-source" : "put-sink", desc->bd_iov_count, desc->bd_nob, desc->bd_last_mbits, req->rq_mbits, desc->bd_portal); @@ -434,16 +444,18 @@ int ptlrpc_register_bulk(struct ptlrpc_request *req) int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) { struct ptlrpc_bulk_desc *desc = req->rq_bulk; - struct l_wait_info lwi; - int rc; ENTRY; LASSERT(!in_interrupt()); /* might sleep */ + if (desc) + desc->bd_registered = 0; + /* Let's setup deadline for reply unlink. */ if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) && async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0) - req->rq_bulk_deadline = ktime_get_real_seconds() + LONG_UNLINK; + req->rq_bulk_deadline = ktime_get_real_seconds() + + PTLRPC_REQ_LONG_UNLINK; if (ptlrpc_client_bulk_active(req) == 0) /* completed or */ RETURN(1); /* never registered */ @@ -452,8 +464,9 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) /* 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() */ + * but we must still wait_event_idle_timeout() in this case to give + * us 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 */ @@ -471,30 +484,37 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async) 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); - } - - LASSERT(rc == -ETIMEDOUT); - DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p", - desc); - } - RETURN(0); + /* + * Network access will complete in finite time but the HUGE + * timeout lets us CWARN for visibility of sluggish NALs. + */ + int seconds = PTLRPC_REQ_LONG_UNLINK; + + while (seconds > 0 && + wait_event_idle_timeout(*wq, + !ptlrpc_client_bulk_active(req), + cfs_time_seconds(1)) == 0) + seconds -= 1; + if (seconds > 0) { + ptlrpc_rqphase_move(req, req->rq_next_phase); + RETURN(1); + } + + DEBUG_REQ(D_WARNING, req, "Unexpectedly long timeout: desc %p", + desc); + } + RETURN(0); } 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, ktime_get_real_seconds() - - req->rq_arrival_time.tv_sec, 1); + timeout_t service_timeout; + service_timeout = clamp_t(timeout_t, ktime_get_real_seconds() - + req->rq_arrival_time.tv_sec, 1, + (AT_OFF ? obd_timeout * 3 / 2 : at_max)); if (!(flags & PTLRPC_REPLY_EARLY) && (req->rq_type != PTL_RPC_MSG_ERR) && (req->rq_reqmsg != NULL) && @@ -502,8 +522,10 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) (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); + * toward our service time estimate + */ + timeout_t oldse = at_measured(&svcpt->scp_at_estimate, + service_timeout); if (oldse != 0) { DEBUG_REQ(D_ADAPTTO, req, @@ -513,7 +535,7 @@ 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); + lustre_msg_set_service_timeout(req->rq_repmsg, service_timeout); /* 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 @@ -523,7 +545,7 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) req->rq_export->exp_obd->obd_recovering)) { lustre_msg_set_timeout(req->rq_repmsg, 0); } else { - time64_t timeout; + timeout_t timeout; if (req->rq_export && req->rq_reqmsg != NULL && (flags & PTLRPC_REPLY_EARLY) && @@ -533,7 +555,7 @@ static void ptlrpc_at_set_reply(struct ptlrpc_request *req, int flags) timeout = ktime_get_real_seconds() - req->rq_arrival_time.tv_sec + - min_t(time64_t, at_extra, + min_t(timeout_t, at_extra, exp_obd->obd_recovery_timeout / 4); } else { timeout = at_get(&svcpt->scp_at_estimate); @@ -589,16 +611,6 @@ 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; @@ -634,7 +646,8 @@ int ptlrpc_send_reply(struct ptlrpc_request *req, int flags) 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); + req->rq_rep_mbits ? req->rq_rep_mbits : req->rq_xid, + req->rq_reply_off, NULL); out: if (unlikely(rc != 0)) ptlrpc_req_drop_rs(req); @@ -691,11 +704,13 @@ int ptlrpc_error(struct ptlrpc_request *req) int ptl_send_rpc(struct ptlrpc_request *request, int noreply) { int rc; - int rc2; + __u32 opc; int mpflag = 0; + bool rep_mbits = false; struct lnet_handle_md bulk_cookie; + struct lnet_processid peer; struct ptlrpc_connection *connection; - struct lnet_handle_me reply_me_h; + struct lnet_me *reply_me = NULL; struct lnet_md reply_md; struct obd_import *imp = request->rq_import; struct obd_device *obd = imp->imp_obd; @@ -703,28 +718,34 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) LNetInvalidateMDHandle(&bulk_cookie); - if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC)) - RETURN(0); + if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC)) + RETURN(0); + + if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DELAY_RECOV) && + lustre_msg_get_opc(request->rq_reqmsg) == MDS_CONNECT && + strcmp(obd->obd_type->typ_name, LUSTRE_OSP_NAME) == 0)) { + RETURN(0); + } - LASSERT(request->rq_type == PTL_RPC_MSG_REQUEST); - LASSERT(request->rq_wait_ctx == 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); + /* 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))); + (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); + 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); - } + request->rq_status = -ENODEV; + RETURN(-ENODEV); + } connection = imp->imp_connection; @@ -750,20 +771,26 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) 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"); + DEBUG_REQ(D_RPCTRACE, request, + "Allocating new XID for resend on EINPROGRESS"); } - if (request->rq_bulk != NULL) { - ptlrpc_set_bulk_mbits(request); + opc = lustre_msg_get_opc(request->rq_reqmsg); + if (opc != OST_CONNECT && opc != MDS_CONNECT && + opc != MGS_CONNECT && OCD_HAS_FLAG(&imp->imp_connect_data, FLAGS2)) + rep_mbits = imp->imp_connect_data.ocd_connect_flags2 & + OBD_CONNECT2_REP_MBITS; + + if ((request->rq_bulk != NULL) || rep_mbits) { + ptlrpc_set_mbits(request); lustre_msg_set_mbits(request->rq_reqmsg, request->rq_mbits); } 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\n", 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(); } @@ -773,15 +800,15 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) 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)); + 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(); + if (request->rq_memalloc) + mpflag = memalloc_noreclaim_save(); rc = sptlrpc_cli_wrap_request(request); if (rc) @@ -791,7 +818,7 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) if (request->rq_bulk != NULL) { rc = ptlrpc_register_bulk (request); if (rc != 0) - GOTO(out, rc); + 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 @@ -800,36 +827,47 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) bulk_cookie = request->rq_bulk->bd_mds[0]; } - 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 */ + 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 = 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); - } - } + request->rq_status = rc; + GOTO(cleanup_bulk, rc); + } + } else { + request->rq_repdata = NULL; + request->rq_repmsg = NULL; + } + + peer.pid = connection->c_peer.pid; + lnet_nid4_to_nid(connection->c_peer.nid, &peer.nid); + if (request->rq_bulk && + OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) { + reply_me = ERR_PTR(-ENOMEM); + } else { + reply_me = LNetMEAttach(request->rq_reply_portal, + &peer, + rep_mbits ? request->rq_mbits : + request->rq_xid, + 0, LNET_UNLINK, LNET_INS_AFTER); + } + + if (IS_ERR(reply_me)) { + rc = PTR_ERR(reply_me); + CERROR("LNetMEAttach failed: %d\n", rc); + LASSERT(rc == -ENOMEM); + GOTO(cleanup_bulk, rc = -ENOMEM); + } + } spin_lock(&request->rq_lock); /* We are responsible for unlinking the reply buffer */ @@ -837,65 +875,65 @@ int ptl_send_rpc(struct ptlrpc_request *request, int 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_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_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; + 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.handler = ptlrpc_handler; /* 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); + rc = LNetMDAttach(reply_me, &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); - } + GOTO(cleanup_bulk, rc = -ENOMEM); + } + percpu_ref_get(&ptlrpc_pending); - CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %llu" - ", portal %u\n", - request->rq_repbuf_len, request->rq_xid, - request->rq_reply_portal); - } + CDEBUG(D_NET, + "Setup reply buffer: %u bytes, xid %llu, portal %u\n", + request->rq_repbuf_len, request->rq_xid, + request->rq_reply_portal); + } - /* add references on request for request_out_callback */ - ptlrpc_request_addref(request); + /* 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)); + atomic_read(&imp->imp_inflight)); OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5); 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(imp); + * add the network latency for our local timeout. + */ + request->rq_deadline = request->rq_sent + request->rq_timeout + + ptlrpc_at_get_net_latency(request); - DEBUG_REQ(D_INFO, request, "send flg=%x", + 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, @@ -907,23 +945,19 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) 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); + ptlrpc_req_finished(request); + if (noreply) + GOTO(out, rc); + + LNetMDUnlink(request->rq_reply_md_h); + + /* 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. */ - ptlrpc_unregister_bulk(request, 0); + /* 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 (rc == -ENOMEM) { /* set rq_sent so that this request is treated @@ -932,7 +966,7 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) } if (request->rq_memalloc) - cfs_memory_pressure_restore(mpflag); + memalloc_noreclaim_restore(mpflag); return rc; } @@ -944,13 +978,13 @@ 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 struct lnet_process_id match_id = { - .nid = LNET_NID_ANY, + static struct lnet_processid match_id = { + .nid = LNET_ANY_NID, .pid = LNET_PID_ANY }; int rc; struct lnet_md md; - struct lnet_handle_me me_h; + struct lnet_me *me; CDEBUG(D_NET, "LNetMEAttach: portal %d\n", service->srv_req_portal); @@ -961,35 +995,36 @@ int ptlrpc_register_rqbd(struct ptlrpc_request_buffer_desc *rqbd) /* 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, + me = 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); - } - - LASSERT(rqbd->rqbd_refcount == 0); - rqbd->rqbd_refcount = 1; - - md.start = rqbd->rqbd_buffer; - md.length = service->srv_buf_size; - md.max_size = service->srv_max_req_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 = ptlrpc_eq_h; + LNET_INS_LOCAL : LNET_INS_AFTER); + if (IS_ERR(me)) { + CERROR("LNetMEAttach failed: %ld\n", PTR_ERR(me)); + return -ENOMEM; + } - rc = LNetMDAttach(me_h, md, LNET_UNLINK, &rqbd->rqbd_md_h); - if (rc == 0) - return (0); + LASSERT(rqbd->rqbd_refcount == 0); + rqbd->rqbd_refcount = 1; + + md.start = rqbd->rqbd_buffer; + md.length = service->srv_buf_size; + md.max_size = service->srv_max_req_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.handler = ptlrpc_handler; + + rc = LNetMDAttach(me, &md, LNET_UNLINK, &rqbd->rqbd_md_h); + if (rc == 0) { + percpu_ref_get(&ptlrpc_pending); + return 0; + } - CERROR("LNetMDAttach failed: %d; \n", rc); - LASSERT (rc == -ENOMEM); - rc = LNetMEUnlink (me_h); - LASSERT (rc == 0); - rqbd->rqbd_refcount = 0; + CERROR("ptlrpc: LNetMDAttach failed: rc = %d\n", rc); + LASSERT(rc == -ENOMEM); + LASSERT(rc == 0); + rqbd->rqbd_refcount = 0; - return (-ENOMEM); + return -ENOMEM; }