X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fost%2Fost_handler.c;h=405d69cbdbbbbaba86da89d8835f0eeac8ac9276;hp=28f3debeaca3ba41fffea1de1aa423d7c56f44cc;hb=7eb9b5de99d4205fc4e83549728df8a53176ca70;hpb=c139827b023c4ced7db36060832550ff9eeaec03 diff --git a/lustre/ost/ost_handler.c b/lustre/ost/ost_handler.c index 28f3deb..405d69c 100644 --- a/lustre/ost/ost_handler.c +++ b/lustre/ost/ost_handler.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -39,9 +39,6 @@ * Author: Phil Schwan */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_OST #include @@ -51,11 +48,15 @@ #include #include #include +#include +#include #include #include #include #include +#include #include "ost_internal.h" +#include static int oss_num_threads; CFS_MODULE_PARM(oss_num_threads, "i", int, 0444, @@ -69,10 +70,27 @@ static int oss_num_create_threads; CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444, "number of OSS create threads to start"); +static char *oss_cpts; +CFS_MODULE_PARM(oss_cpts, "s", charp, 0444, + "CPU partitions OSS threads should run on"); + +static char *oss_io_cpts; +CFS_MODULE_PARM(oss_io_cpts, "s", charp, 0444, + "CPU partitions OSS IO threads should run on"); + +/* + * this page is allocated statically when module is initializing + * it is used to simulate data corruptions, see ost_checksum_bulk() + * for details. as the original pages provided by the layers below + * can be remain in the internal cache, we do not want to modify + * them. + */ +static struct page *ost_page_to_corrupt = NULL; + /** * Do not return server-side uid/gid to remote client */ -static void ost_drop_id(struct obd_export *exp, struct obdo *oa) +static void ost_drop_id(struct obd_export *exp, struct obdo *oa) { if (exp_connect_rmtclient(exp)) { oa->o_uid = -1; @@ -81,6 +99,78 @@ static void ost_drop_id(struct obd_export *exp, struct obdo *oa) } } +/** + * Validate oa from client. + * If the request comes from 2.0 clients, currently only RSVD seq and IDIF + * req are valid. + * a. objects in Single MDT FS seq = FID_SEQ_OST_MDT0, oi_id != 0 + * b. Echo objects(seq = 2), old echo client still use oi_id/oi_seq to + * pack ost_id. Because non-zero oi_seq will make it diffcult to tell + * whether this is oi_fid or real ostid. So it will check + * OBD_CONNECT_FID, then convert the ostid to FID for old client. + * c. Old FID-disable osc will send IDIF. + * d. new FID-enable osc/osp will send normal FID. + * + * And also oi_id/f_oid should always start from 1. oi_id/f_oid = 0 will + * be used for LAST_ID file, and only being accessed inside OST now. + */ +static int ost_validate_obdo(struct obd_export *exp, struct obdo *oa, + struct obd_ioobj *ioobj) +{ + int rc = 0; + + if (unlikely(!(exp_connect_flags(exp) & OBD_CONNECT_FID) && + fid_seq_is_echo(oa->o_oi.oi.oi_seq) && oa != NULL)) { + /* Sigh 2.[123] client still sends echo req with oi_id = 0 + * during create, and we will reset this to 1, since this + * oi_id is basically useless in the following create process, + * but oi_id == 0 will make it difficult to tell whether it is + * real FID or ost_id. */ + oa->o_oi.oi_fid.f_oid = oa->o_oi.oi.oi_id ?: 1; + oa->o_oi.oi_fid.f_seq = FID_SEQ_ECHO; + oa->o_oi.oi_fid.f_ver = 0; + } else { + if (unlikely((oa == NULL) || ostid_id(&oa->o_oi) == 0)) + GOTO(out, rc = -EPROTO); + + /* Note: this check might be forced in 2.5 or 2.6, i.e. + * all of the requests are required to setup FLGROUP */ + if (unlikely(!(oa->o_valid & OBD_MD_FLGROUP))) { + ostid_set_seq_mdt0(&oa->o_oi); + if (ioobj) + ostid_set_seq_mdt0(&ioobj->ioo_oid); + oa->o_valid |= OBD_MD_FLGROUP; + } + + if (unlikely(!(fid_seq_is_idif(ostid_seq(&oa->o_oi)) || + fid_seq_is_mdt0(ostid_seq(&oa->o_oi)) || + fid_seq_is_norm(ostid_seq(&oa->o_oi)) || + fid_seq_is_echo(ostid_seq(&oa->o_oi))))) + GOTO(out, rc = -EPROTO); + } + + if (ioobj != NULL) { + unsigned max_brw = ioobj_max_brw_get(ioobj); + + if (unlikely((max_brw & (max_brw - 1)) != 0)) { + CERROR("%s: client %s sent bad ioobj max %u for "DOSTID + ": rc = -EPROTO\n", exp->exp_obd->obd_name, + obd_export_nid2str(exp), max_brw, + POSTID(&oa->o_oi)); + GOTO(out, rc = -EPROTO); + } + ioobj->ioo_oid = oa->o_oi; + } + +out: + if (rc != 0) + CERROR("%s: client %s sent bad object "DOSTID": rc = %d\n", + exp->exp_obd->obd_name, obd_export_nid2str(exp), + oa ? ostid_seq(&oa->o_oi) : -1, + oa ? ostid_id(&oa->o_oi) : -1, rc); + return rc; +} + void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) { struct oti_req_ack_lock *ack_lock; @@ -89,8 +179,12 @@ void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req) if (oti == NULL) return; - if (req->rq_repmsg) + if (req->rq_repmsg) { + __u64 versions[PTLRPC_NUM_VERSIONS] = { 0 }; lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno); + versions[0] = oti->oti_pre_version; + lustre_msg_set_versions(req->rq_repmsg, versions); + } req->rq_transno = oti->oti_transno; /* XXX 4 == entries in oti_ack_locks??? */ @@ -115,8 +209,12 @@ static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, if (body == NULL) RETURN(-EFAULT); - if (body->oa.o_id == 0) - RETURN(-EPROTO); + if (ostid_id(&body->oa.o_oi) == 0) + RETURN(-EPROTO); + + rc = ost_validate_obdo(exp, &body->oa, NULL); + if (rc) + RETURN(rc); /* If there's a DLM request, cancel the locks mentioned in it*/ if (req_capsule_field_present(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT)) { @@ -151,7 +249,8 @@ static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); /* Do the destroy and set the reply status accordingly */ - req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL, capa); + req->rq_status = obd_destroy(req->rq_svc_thread->t_env, exp, + &repbody->oa, NULL, oti, NULL, capa); RETURN(0); } @@ -161,22 +260,26 @@ static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req, */ static int ost_lock_get(struct obd_export *exp, struct obdo *oa, __u64 start, __u64 count, struct lustre_handle *lh, - int mode, int flags) + int mode, __u64 flags) { - struct ldlm_res_id res_id = { .name = { oa->o_id, 0, oa->o_gr, 0} }; + struct ldlm_res_id res_id; ldlm_policy_data_t policy; __u64 end = start + count; ENTRY; LASSERT(!lustre_handle_is_used(lh)); - LASSERT((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) == - (OBD_MD_FLID | OBD_MD_FLGROUP)); + /* o_id and o_gr are used for localizing resource, if client miss to set + * them, do not trigger ASSERTION. */ + if (unlikely((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) != + (OBD_MD_FLID | OBD_MD_FLGROUP))) + RETURN(-EPROTO); if (!(oa->o_valid & OBD_MD_FLFLAGS) || !(oa->o_flags & OBD_FL_SRVLOCK)) RETURN(0); + ostid_build_res_name(&oa->o_oi, &res_id); CDEBUG(D_INODE, "OST-side extent lock.\n"); policy.l_extent.start = start & CFS_PAGE_MASK; @@ -191,7 +294,8 @@ static int ost_lock_get(struct obd_export *exp, struct obdo *oa, RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id, LDLM_EXTENT, &policy, mode, &flags, ldlm_blocking_ast, ldlm_completion_ast, - ldlm_glimpse_ast, NULL, 0, NULL, lh)); + ldlm_glimpse_ast, NULL, 0, LVB_T_NONE, + NULL, lh)); } /* Helper function: release lock, if any. */ @@ -207,8 +311,9 @@ static void ost_lock_put(struct obd_export *exp, static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) { struct ost_body *body, *repbody; - struct obd_info oinfo = { { { 0 } } }; + struct obd_info *oinfo; struct lustre_handle lh = { 0 }; + struct lustre_capa *capa = NULL; int rc; ENTRY; @@ -216,6 +321,18 @@ static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) if (body == NULL) RETURN(-EFAULT); + rc = ost_validate_obdo(exp, &body->oa, NULL); + if (rc) + RETURN(rc); + + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); + if (capa == NULL) { + CERROR("Missing capability for OST GETATTR"); + RETURN(-EFAULT); + } + } + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); @@ -223,25 +340,25 @@ static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req) repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); repbody->oa = body->oa; - rc = ost_lock_get(exp, &body->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0); + rc = ost_lock_get(exp, &repbody->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0); if (rc) RETURN(rc); - oinfo.oi_oa = &repbody->oa; - if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) { - oinfo.oi_capa = req_capsule_client_get(&req->rq_pill, - &RMF_CAPA1); - if (oinfo.oi_capa == NULL) { - CERROR("Missing capability for OST GETATTR"); - RETURN (-EFAULT); - } - } + OBD_ALLOC_PTR(oinfo); + if (!oinfo) + GOTO(unlock, rc = -ENOMEM); + oinfo->oi_oa = &repbody->oa; + oinfo->oi_capa = capa; - req->rq_status = obd_getattr(exp, &oinfo); - ost_lock_put(exp, &lh, LCK_PR); + req->rq_status = obd_getattr(req->rq_svc_thread->t_env, exp, oinfo); + + OBD_FREE_PTR(oinfo); ost_drop_id(exp, &repbody->oa); - RETURN(0); + +unlock: + ost_lock_put(exp, &lh, LCK_PR); + RETURN(rc); } static int ost_statfs(struct ptlrpc_request *req) @@ -256,11 +373,16 @@ static int ost_statfs(struct ptlrpc_request *req) osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS); - req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs, - cfs_time_current_64() - CFS_HZ, 0); + req->rq_status = obd_statfs(req->rq_svc_thread->t_env, req->rq_export, + osfs, + cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS), + 0); if (req->rq_status != 0) CERROR("ost: statfs failed: rc %d\n", req->rq_status); + if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_EINPROGRESS)) + req->rq_status = -EINPROGRESS; + RETURN(0); } @@ -275,15 +397,20 @@ static int ost_create(struct obd_export *exp, struct ptlrpc_request *req, if (body == NULL) RETURN(-EFAULT); + rc = ost_validate_obdo(req->rq_export, &body->oa, NULL); + if (rc) + RETURN(rc); + rc = req_capsule_server_pack(&req->rq_pill); if (rc) RETURN(rc); repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - oti->oti_logcookies = &repbody->oa.o_lcookie; + repbody->oa = body->oa; + oti->oti_logcookies = &body->oa.o_lcookie; - req->rq_status = obd_create(exp, &repbody->oa, NULL, oti); + req->rq_status = obd_create(req->rq_svc_thread->t_env, exp, + &repbody->oa, NULL, oti); //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0); RETURN(0); } @@ -291,10 +418,10 @@ static int ost_create(struct obd_export *exp, struct ptlrpc_request *req, static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { - struct obd_info oinfo = { { { 0 } } }; struct ost_body *body, *repbody; - int rc, flags = 0; + __u64 flags = 0; struct lustre_handle lh = {0,}; + int rc; ENTRY; /* check that we do support OBD_CONNECT_TRUNCLOCK. */ @@ -304,11 +431,11 @@ static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, if (body == NULL) RETURN(-EFAULT); - oinfo.oi_oa = &body->oa; - oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size; - oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks; + rc = ost_validate_obdo(exp, &body->oa, NULL); + if (rc) + RETURN(rc); - if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != + if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) != (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) RETURN(-EPROTO); @@ -318,41 +445,61 @@ static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req, /* standard truncate optimization: if file body is completely * destroyed, don't send data back to the server. */ - if (oinfo.oi_oa->o_size == 0) - flags |= LDLM_AST_DISCARD_DATA; + if (body->oa.o_size == 0) + flags |= LDLM_FL_AST_DISCARD_DATA; repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); - rc = ost_lock_get(exp, oinfo.oi_oa, oinfo.oi_oa->o_size, - oinfo.oi_oa->o_blocks, &lh, LCK_PW, flags); + repbody->oa = body->oa; + + rc = ost_lock_get(exp, &repbody->oa, repbody->oa.o_size, + repbody->oa.o_blocks, &lh, LCK_PW, flags); if (rc == 0) { - if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS && - oinfo.oi_oa->o_flags == OBD_FL_SRVLOCK) + struct obd_info *oinfo; + struct lustre_capa *capa = NULL; + + if (repbody->oa.o_valid & OBD_MD_FLFLAGS && + repbody->oa.o_flags == OBD_FL_SRVLOCK) /* * If OBD_FL_SRVLOCK is the only bit set in * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling * through filter_setattr() to filter_iocontrol(). */ - oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS; + repbody->oa.o_valid &= ~OBD_MD_FLFLAGS; - if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) { - oinfo.oi_capa = req_capsule_client_get(&req->rq_pill, - &RMF_CAPA1); - if (oinfo.oi_capa == NULL) { + if (repbody->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, + &RMF_CAPA1); + if (capa == NULL) { CERROR("Missing capability for OST PUNCH"); - RETURN (-EFAULT); + GOTO(unlock, rc = -EFAULT); } } - req->rq_status = obd_punch(exp, &oinfo, oti, NULL); + + OBD_ALLOC_PTR(oinfo); + if (!oinfo) + GOTO(unlock, rc = -ENOMEM); + oinfo->oi_oa = &repbody->oa; + oinfo->oi_policy.l_extent.start = oinfo->oi_oa->o_size; + oinfo->oi_policy.l_extent.end = oinfo->oi_oa->o_blocks; + oinfo->oi_capa = capa; + oinfo->oi_flags = OBD_FL_PUNCH; + + req->rq_status = obd_punch(req->rq_svc_thread->t_env, exp, + oinfo, oti, NULL); + OBD_FREE_PTR(oinfo); +unlock: ost_lock_put(exp, &lh, LCK_PW); } - repbody->oa = *oinfo.oi_oa; + ost_drop_id(exp, &repbody->oa); RETURN(rc); } -static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req) +static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req, + struct obd_trans_info *oti) { struct ost_body *body, *repbody; + struct obd_info *oinfo; struct lustre_capa *capa = NULL; int rc; ENTRY; @@ -361,6 +508,10 @@ static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req) if (body == NULL) RETURN(-EFAULT); + rc = ost_validate_obdo(exp, &body->oa, NULL); + if (rc) + RETURN(rc); + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); if (capa == NULL) { @@ -374,9 +525,20 @@ static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req) RETURN(rc); repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); - memcpy(&repbody->oa, &body->oa, sizeof(body->oa)); - req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size, - repbody->oa.o_blocks, capa); + repbody->oa = body->oa; + + OBD_ALLOC_PTR(oinfo); + if (!oinfo) + RETURN(-ENOMEM); + + oinfo->oi_oa = &repbody->oa; + oinfo->oi_capa = capa; + oinfo->oi_jobid = oti->oti_jobid; + req->rq_status = obd_sync(req->rq_svc_thread->t_env, exp, oinfo, + repbody->oa.o_size, repbody->oa.o_blocks, + NULL); + OBD_FREE_PTR(oinfo); + ost_drop_id(exp, &repbody->oa); RETURN(0); } @@ -385,89 +547,133 @@ static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ost_body *body, *repbody; + struct obd_info *oinfo; + struct lustre_capa *capa = NULL; int rc; - struct obd_info oinfo = { { { 0 } } }; ENTRY; body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) RETURN(-EFAULT); - rc = req_capsule_server_pack(&req->rq_pill); + rc = ost_validate_obdo(req->rq_export, &body->oa, NULL); if (rc) RETURN(rc); - repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); - repbody->oa = body->oa; + rc = req_capsule_server_pack(&req->rq_pill); + if (rc) + RETURN(rc); - oinfo.oi_oa = &repbody->oa; - if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) { - oinfo.oi_capa = req_capsule_client_get(&req->rq_pill, - &RMF_CAPA1); - if (oinfo.oi_capa == NULL) { + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { + capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); + if (capa == NULL) { CERROR("Missing capability for OST SETATTR"); RETURN (-EFAULT); } } - req->rq_status = obd_setattr(exp, &oinfo, oti); + + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); + repbody->oa = body->oa; + + OBD_ALLOC_PTR(oinfo); + if (!oinfo) + RETURN(-ENOMEM); + oinfo->oi_oa = &repbody->oa; + oinfo->oi_capa = capa; + + req->rq_status = obd_setattr(req->rq_svc_thread->t_env, exp, oinfo, + oti); + + OBD_FREE_PTR(oinfo); + ost_drop_id(exp, &repbody->oa); RETURN(0); } -static int ost_bulk_timeout(void *data) -{ - ENTRY; - /* We don't fail the connection here, because having the export - * killed makes the (vital) call to commitrw very sad. - */ - RETURN(1); -} - static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc, - cksum_type_t cksum_type) + cksum_type_t cksum_type) { - __u32 cksum; - int i; - - cksum = init_checksum(cksum_type); - for (i = 0; i < desc->bd_iov_count; i++) { - struct page *page = desc->bd_iov[i].kiov_page; - int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK; - char *ptr = kmap(page) + off; - int len = desc->bd_iov[i].kiov_len; - - /* corrupt the data before we compute the checksum, to - * simulate a client->OST data error */ - if (i == 0 && opc == OST_WRITE && - OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) - memcpy(ptr, "bad3", min(4, len)); - cksum = compute_checksum(cksum, ptr, len, cksum_type); - /* corrupt the data after we compute the checksum, to - * simulate an OST->client data error */ - if (i == 0 && opc == OST_READ && - OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) { - memcpy(ptr, "bad4", min(4, len)); - /* nobody should use corrupted page again */ - ClearPageUptodate(page); - } - kunmap(page); - } - - return cksum; + struct cfs_crypto_hash_desc *hdesc; + unsigned int bufsize; + int i, err; + unsigned char cfs_alg = cksum_obd2cfs(cksum_type); + __u32 cksum; + + hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0); + if (IS_ERR(hdesc)) { + CERROR("Unable to initialize checksum hash %s\n", + cfs_crypto_hash_name(cfs_alg)); + return PTR_ERR(hdesc); + } + CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg)); + for (i = 0; i < desc->bd_iov_count; i++) { + + /* corrupt the data before we compute the checksum, to + * simulate a client->OST data error */ + if (i == 0 && opc == OST_WRITE && + OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) { + int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK; + int len = desc->bd_iov[i].kiov_len; + struct page *np = ost_page_to_corrupt; + char *ptr = kmap(desc->bd_iov[i].kiov_page) + off; + + if (np) { + char *ptr2 = kmap(np) + off; + + memcpy(ptr2, ptr, len); + memcpy(ptr2, "bad3", min(4, len)); + kunmap(np); + desc->bd_iov[i].kiov_page = np; + } else { + CERROR("can't alloc page for corruption\n"); + } + } + cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page, + desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK, + desc->bd_iov[i].kiov_len); + + /* corrupt the data after we compute the checksum, to + * simulate an OST->client data error */ + if (i == 0 && opc == OST_READ && + OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) { + int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK; + int len = desc->bd_iov[i].kiov_len; + struct page *np = ost_page_to_corrupt; + char *ptr = kmap(desc->bd_iov[i].kiov_page) + off; + + if (np) { + char *ptr2 = kmap(np) + off; + + memcpy(ptr2, ptr, len); + memcpy(ptr2, "bad4", min(4, len)); + kunmap(np); + desc->bd_iov[i].kiov_page = np; + } else { + CERROR("can't alloc page for corruption\n"); + } + } + } + + bufsize = 4; + err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize); + if (err) + cfs_crypto_hash_final(hdesc, NULL, NULL); + + return cksum; } static int ost_brw_lock_get(int mode, struct obd_export *exp, struct obd_ioobj *obj, struct niobuf_remote *nb, struct lustre_handle *lh) { - int flags = 0; + __u64 flags = 0; int nrbufs = obj->ioo_bufcnt; struct ldlm_res_id res_id; ldlm_policy_data_t policy; int i; ENTRY; - osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id); + ostid_build_res_name(&obj->ioo_oid, &res_id); LASSERT(mode == LCK_PR || mode == LCK_PW); LASSERT(!lustre_handle_is_used(lh)); @@ -486,7 +692,8 @@ static int ost_brw_lock_get(int mode, struct obd_export *exp, RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id, LDLM_EXTENT, &policy, mode, &flags, ldlm_blocking_ast, ldlm_completion_ast, - ldlm_glimpse_ast, NULL, 0, NULL, lh)); + ldlm_glimpse_ast, NULL, 0, LVB_T_NONE, + NULL, lh)); } static void ost_brw_lock_put(int mode, @@ -502,122 +709,41 @@ static void ost_brw_lock_put(int mode, EXIT; } -struct ost_prolong_data { - struct obd_export *opd_exp; - ldlm_policy_data_t opd_policy; - struct obdo *opd_oa; - ldlm_mode_t opd_mode; - int opd_lock_match; - int opd_timeout; -}; - -static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data) +/* Allocate thread local buffers if needed */ +static struct ost_thread_local_cache *ost_tls_get(struct ptlrpc_request *r) { - struct ost_prolong_data *opd = data; - - LASSERT(lock->l_resource->lr_type == LDLM_EXTENT); - - if (lock->l_req_mode != lock->l_granted_mode) { - /* scan granted locks only */ - return LDLM_ITER_STOP; - } - - if (lock->l_export != opd->opd_exp) { - /* prolong locks only for given client */ - return LDLM_ITER_CONTINUE; - } - - if (!(lock->l_granted_mode & opd->opd_mode)) { - /* we aren't interesting in all type of locks */ - return LDLM_ITER_CONTINUE; - } - - if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start || - lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) { - /* the request doesn't cross the lock, skip it */ - return LDLM_ITER_CONTINUE; - } - - /* Fill the obdo with the matched lock handle. - * XXX: it is possible in some cases the IO RPC is covered by several - * locks, even for the write case, so it may need to be a lock list. */ - if (opd->opd_oa && !(opd->opd_oa->o_valid & OBD_MD_FLHANDLE)) { - opd->opd_oa->o_handle.cookie = lock->l_handle.h_cookie; - opd->opd_oa->o_valid |= OBD_MD_FLHANDLE; - } - - if (!(lock->l_flags & LDLM_FL_AST_SENT)) { - /* ignore locks not being cancelled */ - return LDLM_ITER_CONTINUE; + struct ost_thread_local_cache *tls = + (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data); + + /* In normal mode of operation an I/O request is serviced only + * by ll_ost_io threads each of them has own tls buffers allocated by + * ost_io_thread_init(). + * During recovery, an I/O request may be queued until any of the ost + * service threads process it. Not necessary it should be one of + * ll_ost_io threads. In that case we dynamically allocating tls + * buffers for the request service time. */ + if (unlikely(tls == NULL)) { + LASSERT(r->rq_export->exp_in_recovery); + OBD_ALLOC_PTR(tls); + if (tls != NULL) { + tls->temporary = 1; + r->rq_svc_thread->t_data = tls; + } } - - CDEBUG(D_DLMTRACE,"refresh lock: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", - lock->l_resource->lr_name.name[0], - lock->l_resource->lr_name.name[1], - opd->opd_policy.l_extent.start, opd->opd_policy.l_extent.end); - /* OK. this is a possible lock the user holds doing I/O - * let's refresh eviction timer for it */ - ldlm_refresh_waiting_lock(lock, opd->opd_timeout); - opd->opd_lock_match = 1; - - return LDLM_ITER_CONTINUE; + return tls; } -static int ost_rw_prolong_locks(struct ptlrpc_request *req, struct obd_ioobj *obj, - struct niobuf_remote *nb, struct obdo *oa, - ldlm_mode_t mode) +/* Free thread local buffers if they were allocated only for servicing + * this one request */ +static void ost_tls_put(struct ptlrpc_request *r) { - struct ldlm_res_id res_id; - int nrbufs = obj->ioo_bufcnt; - struct ost_prolong_data opd = { 0 }; - ENTRY; + struct ost_thread_local_cache *tls = + (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data); - osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id); - - opd.opd_mode = mode; - opd.opd_exp = req->rq_export; - opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK; - opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset + - nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK; - - /* prolong locks for the current service time of the corresponding - * portal (= OST_IO_PORTAL) */ - opd.opd_timeout = AT_OFF ? obd_timeout / 2: - max(at_est2timeout(at_get(&req->rq_rqbd-> - rqbd_service->srv_at_estimate)), ldlm_timeout); - - CDEBUG(D_INFO,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", - res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start, - opd.opd_policy.l_extent.end); - - if (oa->o_valid & OBD_MD_FLHANDLE) { - struct ldlm_lock *lock; - - lock = ldlm_handle2lock(&oa->o_handle); - if (lock != NULL) { - ost_prolong_locks_iter(lock, &opd); - if (opd.opd_lock_match) { - LDLM_LOCK_PUT(lock); - RETURN(1); - } - - /* Check if the lock covers the whole IO region, - * otherwise iterate through the resource. */ - if (lock->l_policy_data.l_extent.end >= - opd.opd_policy.l_extent.end && - lock->l_policy_data.l_extent.start <= - opd.opd_policy.l_extent.start) { - LDLM_LOCK_PUT(lock); - RETURN(0); - } - LDLM_LOCK_PUT(lock); - } + if (unlikely(tls->temporary)) { + OBD_FREE_PTR(tls); + r->rq_svc_thread->t_data = NULL; } - - opd.opd_oa = oa; - ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id, - ost_prolong_locks_iter, &opd); - RETURN(opd.opd_lock_match); } static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) @@ -633,6 +759,7 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) struct lustre_handle lockh = { 0 }; int niocount, npages, nob = 0, rc, i; int no_reply = 0; + struct ost_thread_local_cache *tls; ENTRY; req->rq_bulk_read = 1; @@ -668,6 +795,10 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) if (ioo == NULL) GOTO(out, rc = -EFAULT); + rc = ost_validate_obdo(exp, &body->oa, ioo); + if (rc) + RETURN(rc); + niocount = ioo->ioo_bufcnt; remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); if (remote_nb == NULL) @@ -685,47 +816,44 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) if (rc) GOTO(out, rc); - /* - * Per-thread array of struct niobuf_{local,remote}'s was allocated by - * ost_thread_init(). - */ - local_nb = ost_tls(req)->local; + tls = ost_tls_get(req); + if (tls == NULL) + GOTO(out_bulk, rc = -ENOMEM); + local_nb = tls->local; rc = ost_brw_lock_get(LCK_PR, exp, ioo, remote_nb, &lockh); if (rc != 0) - GOTO(out_bulk, rc); + GOTO(out_tls, rc); + + /* + * If getting the lock took more time than + * client was willing to wait, drop it. b=11330 + */ + if (cfs_time_current_sec() > req->rq_deadline || + OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { + no_reply = 1; + CERROR("Dropping timed-out read from %s because locking" + "object "DOSTID" took %ld seconds (limit was %ld).\n", + libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid), + cfs_time_current_sec() - req->rq_arrival_time.tv_sec, + req->rq_deadline - req->rq_arrival_time.tv_sec); + GOTO(out_lock, rc = -ETIMEDOUT); + } - /* - * If getting the lock took more time than - * client was willing to wait, drop it. b=11330 - */ - if (cfs_time_current_sec() > req->rq_deadline || - OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { - no_reply = 1; - CERROR("Dropping timed-out read from %s because locking" - "object "LPX64" took %ld seconds (limit was %ld).\n", - libcfs_id2str(req->rq_peer), ioo->ioo_id, - cfs_time_current_sec() - req->rq_arrival_time.tv_sec, - req->rq_deadline - req->rq_arrival_time.tv_sec); - GOTO(out_lock, rc = -ETIMEDOUT); - } + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); + memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); npages = OST_THREAD_POOL_SIZE; - rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1, ioo, - remote_nb, &npages, local_nb, oti, capa); + rc = obd_preprw(req->rq_svc_thread->t_env, OBD_BRW_READ, exp, + &repbody->oa, 1, ioo, remote_nb, &npages, local_nb, + oti, capa); if (rc != 0) GOTO(out_lock, rc); - desc = ptlrpc_prep_bulk_exp(req, npages, - BULK_PUT_SOURCE, OST_BULK_PORTAL); - if (desc == NULL) - GOTO(out_lock, rc = -ENOMEM); - - if (!lustre_handle_is_used(&lockh)) - /* no needs to try to prolong lock if server is asked - * to handle locking (= OBD_BRW_SRVLOCK) */ - ost_rw_prolong_locks(req, ioo, remote_nb, &body->oa, - LCK_PW | LCK_PR); + desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), + BULK_PUT_SOURCE, OST_BULK_PORTAL); + if (desc == NULL) + GOTO(out_commitrw, rc = -ENOMEM); nob = 0; for (i = 0; i < npages; i++) { @@ -739,9 +867,9 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) nob += page_rc; if (page_rc != 0) { /* some data! */ LASSERT (local_nb[i].page != NULL); - ptlrpc_prep_bulk_page(desc, local_nb[i].page, - local_nb[i].offset & ~CFS_PAGE_MASK, - page_rc); + ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].page, + local_nb[i].lnb_page_offset, + page_rc); } if (page_rc != local_nb[i].len) { /* short read */ @@ -753,110 +881,43 @@ static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti) } if (body->oa.o_valid & OBD_MD_FLCKSUM) { - cksum_type_t cksum_type = OBD_CKSUM_CRC32; - - if (body->oa.o_valid & OBD_MD_FLFLAGS) - cksum_type = cksum_type_unpack(body->oa.o_flags); - body->oa.o_flags = cksum_type_pack(cksum_type); - body->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS; - body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ, cksum_type); - CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum); + cksum_type_t cksum_type = + cksum_type_unpack(repbody->oa.o_valid & OBD_MD_FLFLAGS ? + repbody->oa.o_flags : 0); + repbody->oa.o_flags = cksum_type_pack(cksum_type); + repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS; + repbody->oa.o_cksum = ost_checksum_bulk(desc, OST_READ,cksum_type); + CDEBUG(D_PAGE, "checksum at read origin: %x\n", + repbody->oa.o_cksum); } else { - body->oa.o_valid = 0; + repbody->oa.o_valid = 0; } /* We're finishing using body->oa as an input variable */ /* Check if client was evicted while we were doing i/o before touching network */ if (rc == 0) { - /* Check if there is eviction in progress, and if so, wait for - * it to finish */ - if (unlikely(cfs_atomic_read(&exp->exp_obd-> - obd_evict_inprogress))) { - lwi = LWI_INTR(NULL, NULL); - rc = l_wait_event(exp->exp_obd-> - obd_evict_inprogress_waitq, - !cfs_atomic_read(&exp->exp_obd-> - obd_evict_inprogress), - &lwi); - } - /* Check if client was evicted or tried to reconnect already */ - if (exp->exp_failed || exp->exp_abort_active_req) - rc = -ENOTCONN; - else { - rc = sptlrpc_svc_wrap_bulk(req, desc); - if (rc == 0) - rc = ptlrpc_start_bulk_transfer(desc); - } - - if (rc == 0) { - time_t start = cfs_time_current_sec(); - do { - long timeoutl = req->rq_deadline - - cfs_time_current_sec(); - cfs_duration_t timeout = timeoutl <= 0 ? - CFS_TICK : cfs_time_seconds(timeoutl); - lwi = LWI_TIMEOUT_INTERVAL(timeout, - cfs_time_seconds(1), - ost_bulk_timeout, - desc); - rc = l_wait_event(desc->bd_waitq, - !ptlrpc_server_bulk_active(desc) || - exp->exp_failed || - exp->exp_abort_active_req, - &lwi); - LASSERT(rc == 0 || rc == -ETIMEDOUT); - /* Wait again if we changed deadline */ - } while ((rc == -ETIMEDOUT) && - (req->rq_deadline > cfs_time_current_sec())); - - if (rc == -ETIMEDOUT) { - DEBUG_REQ(D_ERROR, req, - "timeout on bulk PUT after %ld%+lds", - req->rq_deadline - start, - cfs_time_current_sec() - - req->rq_deadline); - ptlrpc_abort_bulk(desc); - } else if (exp->exp_failed) { - DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT"); - rc = -ENOTCONN; - ptlrpc_abort_bulk(desc); - } else if (exp->exp_abort_active_req) { - DEBUG_REQ(D_ERROR, req, "Reconnect on bulk PUT"); - /* we don't reply anyway */ - rc = -ETIMEDOUT; - ptlrpc_abort_bulk(desc); - } else if (!desc->bd_success || - desc->bd_nob_transferred != desc->bd_nob) { - DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)", - desc->bd_success ? - "truncated" : "network error on", - desc->bd_nob_transferred, - desc->bd_nob); - /* XXX should this be a different errno? */ - rc = -ETIMEDOUT; - } - } else { - DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc); - } + if (likely(!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) + rc = target_bulk_io(exp, desc, &lwi); no_reply = rc != 0; } +out_commitrw: /* Must commit after prep above in all cases */ - rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, ioo, - remote_nb, npages, local_nb, oti, rc); + rc = obd_commitrw(req->rq_svc_thread->t_env, OBD_BRW_READ, exp, + &repbody->oa, 1, ioo, remote_nb, npages, local_nb, + oti, rc); - if (rc == 0) { - repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); - memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); + if (rc == 0) ost_drop_id(exp, &repbody->oa); - } out_lock: ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh); +out_tls: + ost_tls_put(req); out_bulk: - if (desc) - ptlrpc_free_bulk(desc); + if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)) + ptlrpc_free_bulk_nopin(desc); out: LASSERT(rc <= 0); if (rc == 0) { @@ -872,17 +933,75 @@ out: } else { /* reply out callback would free */ ptlrpc_req_drop_rs(req); - CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - " - "client will retry\n", - exp->exp_obd->obd_name, - exp->exp_client_uuid.uuid, - exp->exp_connection->c_remote_uuid.uuid, - libcfs_id2str(req->rq_peer)); + LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), " + "client will retry: rc %d\n", + exp->exp_obd->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), rc); + } + /* send a bulk after reply to simulate a network delay or reordering + * by a router */ + if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) { + cfs_waitq_t waitq; + struct l_wait_info lwi1; + + CDEBUG(D_INFO, "reorder BULK\n"); + cfs_waitq_init(&waitq); + + lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL); + l_wait_event(waitq, 0, &lwi1); + rc = target_bulk_io(exp, desc, &lwi); + ptlrpc_free_bulk_nopin(desc); } RETURN(rc); } +static void ost_warn_on_cksum(struct ptlrpc_request *req, + struct ptlrpc_bulk_desc *desc, + struct niobuf_local *local_nb, int npages, + obd_count client_cksum, obd_count server_cksum, + int mmap) +{ + struct obd_export *exp = req->rq_export; + struct ost_body *body; + char *router; + char *via; + + body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); + LASSERT (body != NULL); + + if (req->rq_peer.nid == desc->bd_sender) { + via = router = ""; + } else { + via = " via "; + router = libcfs_nid2str(desc->bd_sender); + } + + if (mmap) { + CDEBUG_LIMIT(D_INFO, "client csum %x, server csum %x\n", + client_cksum, server_cksum); + return; + } + + LCONSOLE_ERROR_MSG(0x168, "BAD WRITE CHECKSUM: %s from %s%s%s inode " + DFID" object "DOSTID" extent ["LPU64"-"LPU64 + "]: client csum %x, server csum %x\n", + exp->exp_obd->obd_name, libcfs_id2str(req->rq_peer), + via, router, + body->oa.o_valid & OBD_MD_FLFID ? + body->oa.o_parent_seq : (__u64)0, + body->oa.o_valid & OBD_MD_FLFID ? + body->oa.o_parent_oid : 0, + body->oa.o_valid & OBD_MD_FLFID ? + body->oa.o_parent_ver : 0, + POSTID(&body->oa.o_oi), + local_nb[0].lnb_file_offset, + local_nb[npages-1].lnb_file_offset + + local_nb[npages-1].len - 1, + client_cksum, server_cksum); +} + static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) { struct ptlrpc_bulk_desc *desc = NULL; @@ -899,8 +1018,9 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) int rc, i, j; obd_count client_cksum = 0, server_cksum = 0; cksum_type_t cksum_type = OBD_CKSUM_CRC32; - int no_reply = 0; + int no_reply = 0, mmap = 0; __u32 o_uid = 0, o_gid = 0; + struct ost_thread_local_cache *tls; ENTRY; req->rq_bulk_write = 1; @@ -913,32 +1033,22 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) /* pause before transaction has been started */ OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4); - /* Check if there is eviction in progress, and if so, wait for it to - * finish */ - if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) { - lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes - rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq, - !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress), - &lwi); - } - if (exp->exp_failed) - GOTO(out, rc = -ENOTCONN); - /* ost_body, ioobj & noibuf_remote are verified and swabbed in * ost_rw_hpreq_check(). */ body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); if (body == NULL) GOTO(out, rc = -EFAULT); - if ((body->oa.o_flags & OBD_BRW_MEMALLOC) && - (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)) - libcfs_memory_pressure_set(); - objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, RCL_CLIENT) / sizeof(*ioo); ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); if (ioo == NULL) GOTO(out, rc = -EFAULT); + + rc = ost_validate_obdo(exp, &body->oa, ioo); + if (rc) + RETURN(rc); + for (niocount = i = 0; i < objcount; i++) niocount += ioo[i].ioo_bufcnt; @@ -952,6 +1062,10 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb))) GOTO(out, rc = -EFAULT); + if ((remote_nb[0].flags & OBD_BRW_MEMALLOC) && + (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self)) + memory_pressure_set(); + if (body->oa.o_valid & OBD_MD_FLOSSCAPA) { capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1); if (capa == NULL) { @@ -965,38 +1079,32 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) rc = req_capsule_server_pack(&req->rq_pill); if (rc != 0) GOTO(out, rc); - OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val); + CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val); rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS); - /* - * Per-thread array of struct niobuf_{local,remote}'s was allocated by - * ost_thread_init(). - */ - local_nb = ost_tls(req)->local; + tls = ost_tls_get(req); + if (tls == NULL) + GOTO(out_bulk, rc = -ENOMEM); + local_nb = tls->local; rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh); if (rc != 0) - GOTO(out_bulk, rc); - - /* - * If getting the lock took more time than - * client was willing to wait, drop it. b=11330 - */ - if (cfs_time_current_sec() > req->rq_deadline || - OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { - no_reply = 1; - CERROR("Dropping timed-out write from %s because locking " - "object "LPX64" took %ld seconds (limit was %ld).\n", - libcfs_id2str(req->rq_peer), ioo->ioo_id, - cfs_time_current_sec() - req->rq_arrival_time.tv_sec, - req->rq_deadline - req->rq_arrival_time.tv_sec); - GOTO(out_lock, rc = -ETIMEDOUT); - } - - if (!lustre_handle_is_used(&lockh)) - /* no needs to try to prolong lock if server is asked - * to handle locking (= OBD_BRW_SRVLOCK) */ - ost_rw_prolong_locks(req, ioo, remote_nb,&body->oa, LCK_PW); + GOTO(out_tls, rc); + + /* + * If getting the lock took more time than + * client was willing to wait, drop it. b=11330 + */ + if (cfs_time_current_sec() > req->rq_deadline || + OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) { + no_reply = 1; + CERROR("Dropping timed-out write from %s because locking " + "object "DOSTID" took %ld seconds (limit was %ld).\n", + libcfs_id2str(req->rq_peer), POSTID(&ioo->ioo_oid), + cfs_time_current_sec() - req->rq_arrival_time.tv_sec, + req->rq_deadline - req->rq_arrival_time.tv_sec); + GOTO(out_lock, rc = -ETIMEDOUT); + } /* obd_preprw clobbers oa->valid, so save what we need */ if (body->oa.o_valid & OBD_MD_FLCKSUM) { @@ -1004,6 +1112,8 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) if (body->oa.o_valid & OBD_MD_FLFLAGS) cksum_type = cksum_type_unpack(body->oa.o_flags); } + if (body->oa.o_valid & OBD_MD_FLFLAGS && body->oa.o_flags & OBD_FL_MMAP) + mmap = 1; /* Because we already sync grant info with client when reconnect, * grant info will be cleared for resent req, then fed_grant and @@ -1017,86 +1127,37 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) o_uid = body->oa.o_uid; o_gid = body->oa.o_gid; } + + repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); + memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); + npages = OST_THREAD_POOL_SIZE; - rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount, - ioo, remote_nb, &npages, local_nb, oti, capa); + rc = obd_preprw(req->rq_svc_thread->t_env, OBD_BRW_WRITE, exp, + &repbody->oa, objcount, ioo, remote_nb, &npages, + local_nb, oti, capa); if (rc != 0) GOTO(out_lock, rc); - desc = ptlrpc_prep_bulk_exp(req, npages, - BULK_GET_SINK, OST_BULK_PORTAL); - if (desc == NULL) - GOTO(out_lock, rc = -ENOMEM); - - /* NB Having prepped, we must commit... */ + desc = ptlrpc_prep_bulk_exp(req, npages, ioobj_max_brw_get(ioo), + BULK_GET_SINK, OST_BULK_PORTAL); + if (desc == NULL) + GOTO(skip_transfer, rc = -ENOMEM); - for (i = 0; i < npages; i++) - ptlrpc_prep_bulk_page(desc, local_nb[i].page, - local_nb[i].offset & ~CFS_PAGE_MASK, - local_nb[i].len); + /* NB Having prepped, we must commit... */ + for (i = 0; i < npages; i++) + ptlrpc_prep_bulk_page_nopin(desc, local_nb[i].page, + local_nb[i].lnb_page_offset, + local_nb[i].len); rc = sptlrpc_svc_prep_bulk(req, desc); if (rc != 0) GOTO(out_lock, rc); - /* Check if client was evicted or tried to reconnect while we - * were doing i/o before touching network */ - if (desc->bd_export->exp_failed || - desc->bd_export->exp_abort_active_req) - rc = -ENOTCONN; - else - rc = ptlrpc_start_bulk_transfer(desc); - if (rc == 0) { - time_t start = cfs_time_current_sec(); - do { - long timeoutl = req->rq_deadline - - cfs_time_current_sec(); - cfs_duration_t timeout = timeoutl <= 0 ? - CFS_TICK : cfs_time_seconds(timeoutl); - lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1), - ost_bulk_timeout, desc); - rc = l_wait_event(desc->bd_waitq, - !ptlrpc_server_bulk_active(desc) || - desc->bd_export->exp_failed || - desc->bd_export->exp_abort_active_req, - &lwi); - LASSERT(rc == 0 || rc == -ETIMEDOUT); - /* Wait again if we changed deadline */ - } while ((rc == -ETIMEDOUT) && - (req->rq_deadline > cfs_time_current_sec())); - - if (rc == -ETIMEDOUT) { - DEBUG_REQ(D_ERROR, req, - "timeout on bulk GET after %ld%+lds", - req->rq_deadline - start, - cfs_time_current_sec() - - req->rq_deadline); - ptlrpc_abort_bulk(desc); - } else if (desc->bd_export->exp_failed) { - DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET"); - rc = -ENOTCONN; - ptlrpc_abort_bulk(desc); - } else if (desc->bd_export->exp_abort_active_req) { - DEBUG_REQ(D_ERROR, req, "Reconnect on bulk GET"); - /* we don't reply anyway */ - rc = -ETIMEDOUT; - ptlrpc_abort_bulk(desc); - } else if (!desc->bd_success) { - DEBUG_REQ(D_ERROR, req, "network error on bulk GET"); - /* XXX should this be a different errno? */ - rc = -ETIMEDOUT; - } else { - rc = sptlrpc_svc_unwrap_bulk(req, desc); - } - } else { - DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc); - } + rc = target_bulk_io(exp, desc, &lwi); no_reply = rc != 0; - repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY); - memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa)); - - if (unlikely(client_cksum != 0 && rc == 0)) { +skip_transfer: + if (client_cksum != 0 && rc == 0) { static int cksum_counter; repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS; repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL; @@ -1105,9 +1166,10 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) repbody->oa.o_cksum = server_cksum; cksum_counter++; if (unlikely(client_cksum != server_cksum)) { - CERROR("client csum %x, server csum %x\n", - client_cksum, server_cksum); + ost_warn_on_cksum(req, desc, local_nb, npages, + client_cksum, server_cksum, mmap); cksum_counter = 0; + } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){ CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n", cksum_counter, libcfs_id2str(req->rq_peer), @@ -1116,8 +1178,9 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) } /* Must commit after prep above in all cases */ - rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo, - remote_nb, npages, local_nb, oti, rc); + rc = obd_commitrw(req->rq_svc_thread->t_env, OBD_BRW_WRITE, exp, + &repbody->oa, objcount, ioo, remote_nb, npages, + local_nb, oti, rc); if (rc == -ENOTCONN) /* quota acquire process has been given up because * either the client has been evicted or the client @@ -1136,47 +1199,6 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) */ repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME); - if (unlikely(client_cksum != server_cksum && rc == 0)) { - int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type); - char *msg; - char *via; - char *router; - - if (new_cksum == server_cksum) - msg = "changed in transit before arrival at OST"; - else if (new_cksum == client_cksum) - msg = "initial checksum before message complete"; - else - msg = "changed in transit AND after initial checksum"; - - if (req->rq_peer.nid == desc->bd_sender) { - via = router = ""; - } else { - via = " via "; - router = libcfs_nid2str(desc->bd_sender); - } - - LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from " - "%s%s%s inum "LPU64"/"LPU64" object " - LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n", - exp->exp_obd->obd_name, msg, - libcfs_id2str(req->rq_peer), - via, router, - body->oa.o_valid & OBD_MD_FLFID ? - body->oa.o_fid : (__u64)0, - body->oa.o_valid & OBD_MD_FLFID ? - body->oa.o_generation :(__u64)0, - body->oa.o_id, - body->oa.o_valid & OBD_MD_FLGROUP ? - body->oa.o_gr : (__u64)0, - local_nb[0].offset, - local_nb[npages-1].offset + - local_nb[npages-1].len - 1 ); - CERROR("client csum %x, original server csum %x, " - "server csum now %x\n", - client_cksum, server_cksum, new_cksum); - } - if (rc == 0) { int nob = 0; @@ -1201,9 +1223,11 @@ static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti) out_lock: ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh); +out_tls: + ost_tls_put(req); out_bulk: if (desc) - ptlrpc_free_bulk(desc); + ptlrpc_free_bulk_nopin(desc); out: if (rc == 0) { oti_to_request(oti, req); @@ -1217,14 +1241,13 @@ out: } else { /* reply out callback would free */ ptlrpc_req_drop_rs(req); - CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - " - "client will retry\n", - exp->exp_obd->obd_name, - exp->exp_client_uuid.uuid, - exp->exp_connection->c_remote_uuid.uuid, - libcfs_id2str(req->rq_peer)); + LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s), " + "client will retry: rc %d\n", + exp->exp_obd->obd_name, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp), rc); } - libcfs_memory_pressure_clr(); + memory_pressure_clr(); RETURN(rc); } @@ -1300,17 +1323,96 @@ static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req) /* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to * a struct ost_body * value */ - rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL); + rc = obd_set_info_async(req->rq_svc_thread->t_env, exp, keylen, + key, vallen, val, NULL); out: lustre_msg_set_status(req->rq_repmsg, 0); RETURN(rc); } +struct locked_region { + cfs_list_t list; + struct lustre_handle lh; +}; + +static int lock_region(struct obd_export *exp, struct obdo *oa, + unsigned long long begin, unsigned long long end, + cfs_list_t *locked) +{ + struct locked_region *region = NULL; + int rc; + + LASSERT(begin <= end); + OBD_ALLOC_PTR(region); + if (region == NULL) + return -ENOMEM; + + rc = ost_lock_get(exp, oa, begin, end - begin, ®ion->lh, LCK_PR, 0); + if (rc) { + OBD_FREE_PTR(region); + return rc; + } + + CDEBUG(D_OTHER, "ost lock [%llu,%llu], lh=%p\n", + begin, end, ®ion->lh); + cfs_list_add(®ion->list, locked); + + return 0; +} + +static int lock_zero_regions(struct obd_export *exp, struct obdo *oa, + struct ll_user_fiemap *fiemap, + cfs_list_t *locked) +{ + __u64 begin = fiemap->fm_start; + unsigned int i; + int rc = 0; + struct ll_fiemap_extent *fiemap_start = fiemap->fm_extents; + ENTRY; + + CDEBUG(D_OTHER, "extents count %u\n", fiemap->fm_mapped_extents); + for (i = 0; i < fiemap->fm_mapped_extents; i++) { + if (fiemap_start[i].fe_logical > begin) { + CDEBUG(D_OTHER, "ost lock [%llu,%llu]\n", + begin, fiemap_start[i].fe_logical); + rc = lock_region(exp, oa, begin, + fiemap_start[i].fe_logical, locked); + if (rc) + RETURN(rc); + } + + begin = fiemap_start[i].fe_logical + fiemap_start[i].fe_length; + } + + if (begin < (fiemap->fm_start + fiemap->fm_length)) { + CDEBUG(D_OTHER, "ost lock [%llu,%llu]\n", + begin, fiemap->fm_start + fiemap->fm_length); + rc = lock_region(exp, oa, begin, + fiemap->fm_start + fiemap->fm_length, locked); + } + + RETURN(rc); +} + +static void unlock_zero_regions(struct obd_export *exp, cfs_list_t *locked) +{ + struct locked_region *entry, *temp; + cfs_list_for_each_entry_safe(entry, temp, locked, list) { + CDEBUG(D_OTHER, "ost unlock lh=%p\n", &entry->lh); + ost_lock_put(exp, &entry->lh, LCK_PR); + cfs_list_del(&entry->list); + OBD_FREE_PTR(entry); + } +} + static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req) { void *key, *reply; int keylen, replylen, rc = 0; struct req_capsule *pill = &req->rq_pill; + cfs_list_t locked = CFS_LIST_HEAD_INIT(locked); + struct ll_fiemap_info_key *fm_key = NULL; + struct ll_user_fiemap *fiemap; ENTRY; /* this common part for get_info rpc */ @@ -1321,29 +1423,71 @@ static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req) } keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT); - rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL); + if (KEY_IS(KEY_FIEMAP)) { + fm_key = key; + rc = ost_validate_obdo(exp, &fm_key->oa, NULL); + if (rc) + RETURN(rc); + } + + rc = obd_get_info(req->rq_svc_thread->t_env, exp, keylen, key, + &replylen, NULL, NULL); if (rc) - RETURN(rc); + RETURN(rc); req_capsule_set_size(pill, &RMF_GENERIC_DATA, RCL_SERVER, replylen); rc = req_capsule_server_pack(pill); if (rc) - RETURN(rc); + RETURN(rc); reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA); if (reply == NULL) - RETURN(-ENOMEM); + RETURN(-ENOMEM); + + if (KEY_IS(KEY_LAST_FID)) { + void *val; + int vallen; + + req_capsule_extend(pill, &RQF_OST_GET_INFO_LAST_FID); + val = req_capsule_client_get(pill, &RMF_SETINFO_VAL); + vallen = req_capsule_get_size(pill, &RMF_SETINFO_VAL, + RCL_CLIENT); + if (val != NULL && vallen > 0 && replylen >= vallen) { + memcpy(reply, val, vallen); + } else { + CERROR("%s: invalid req val %p vallen %d replylen %d\n", + exp->exp_obd->obd_name, val, vallen, replylen); + RETURN(-EINVAL); + } + } + + /* call again to fill in the reply buffer */ + rc = obd_get_info(req->rq_svc_thread->t_env, exp, keylen, key, + &replylen, reply, NULL); + + /* LU-3219: Lock the sparse areas to make sure dirty flushed back + * from client, then call fiemap again. */ + if (KEY_IS(KEY_FIEMAP) && (fm_key->oa.o_valid & OBD_MD_FLFLAGS) && + (fm_key->oa.o_flags & OBD_FL_SRVLOCK)) { + fiemap = (struct ll_user_fiemap *)reply; + fm_key = key; + + rc = lock_zero_regions(exp, &fm_key->oa, fiemap, &locked); + if (rc == 0 && !cfs_list_empty(&locked)) + rc = obd_get_info(req->rq_svc_thread->t_env, exp, + keylen, key, &replylen, reply, NULL); + unlock_zero_regions(exp, &locked); + if (rc) + RETURN(rc); + } + + lustre_msg_set_status(req->rq_repmsg, 0); - /* call again to fill in the reply buffer */ - rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL); - - lustre_msg_set_status(req->rq_repmsg, 0); RETURN(rc); } -#ifdef HAVE_QUOTA_SUPPORT static int ost_handle_quotactl(struct ptlrpc_request *req) { struct obd_quotactl *oqctl, *repoqc; @@ -1380,35 +1524,11 @@ static int ost_handle_quotacheck(struct ptlrpc_request *req) if (rc) RETURN(-ENOMEM); - req->rq_status = obd_quotacheck(req->rq_export, oqctl); - RETURN(0); + /* deprecated, not used any more */ + req->rq_status = -EOPNOTSUPP; + RETURN(-EOPNOTSUPP); } -static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req) -{ - struct quota_adjust_qunit *oqaq, *repoqa; - struct lustre_quota_ctxt *qctxt; - int rc; - ENTRY; - - qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt; - oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT); - if (oqaq == NULL) - GOTO(out, rc = -EPROTO); - - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - GOTO(out, rc); - - repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT); - req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt); - *repoqa = *oqaq; - - out: - RETURN(rc); -} -#endif - static int ost_llog_handle_connect(struct obd_export *exp, struct ptlrpc_request *req) { @@ -1421,14 +1541,14 @@ static int ost_llog_handle_connect(struct obd_export *exp, RETURN(rc); } -#define ost_init_sec_none(reply, exp) \ -do { \ - reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \ - OBD_CONNECT_RMT_CLIENT_FORCE | \ - OBD_CONNECT_OSS_CAPA); \ - cfs_spin_lock(&exp->exp_lock); \ - exp->exp_connect_flags = reply->ocd_connect_flags; \ - cfs_spin_unlock(&exp->exp_lock); \ +#define ost_init_sec_none(reply, exp) \ +do { \ + reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \ + OBD_CONNECT_RMT_CLIENT_FORCE | \ + OBD_CONNECT_OSS_CAPA); \ + spin_lock(&exp->exp_lock); \ + *exp_connect_flags_ptr(exp) = reply->ocd_connect_flags; \ + spin_unlock(&exp->exp_lock); \ } while (0) static int ost_init_sec_level(struct ptlrpc_request *req) @@ -1525,9 +1645,9 @@ static int ost_init_sec_level(struct ptlrpc_request *req) if (!filter->fo_fl_oss_capa) reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA; - cfs_spin_lock(&exp->exp_lock); - exp->exp_connect_flags = reply->ocd_connect_flags; - cfs_spin_unlock(&exp->exp_lock); + spin_lock(&exp->exp_lock); + *exp_connect_flags_ptr(exp) = reply->ocd_connect_flags; + spin_unlock(&exp->exp_lock); } break; default: @@ -1560,14 +1680,14 @@ static int ost_connect_check_sptlrpc(struct ptlrpc_request *req) } if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) { - cfs_read_lock(&filter->fo_sptlrpc_lock); - sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset, - req->rq_sp_from, - req->rq_peer.nid, - &flvr); - cfs_read_unlock(&filter->fo_sptlrpc_lock); + read_lock(&filter->fo_sptlrpc_lock); + sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset, + req->rq_sp_from, + req->rq_peer.nid, + &flvr); + read_unlock(&filter->fo_sptlrpc_lock); - cfs_spin_lock(&exp->exp_lock); + spin_lock(&exp->exp_lock); exp->exp_sp_peer = req->rq_sp_from; exp->exp_flvr = flvr; @@ -1581,7 +1701,7 @@ static int ost_connect_check_sptlrpc(struct ptlrpc_request *req) rc = -EACCES; } - cfs_spin_unlock(&exp->exp_lock); + spin_unlock(&exp->exp_lock); } else { if (exp->exp_sp_peer != req->rq_sp_from) { CERROR("RPC source %s doesn't match %s\n", @@ -1596,6 +1716,63 @@ static int ost_connect_check_sptlrpc(struct ptlrpc_request *req) return rc; } +/* Ensure that data and metadata are synced to the disk when lock is cancelled + * (if requested) */ +int ost_blocking_ast(struct ldlm_lock *lock, struct ldlm_lock_desc *desc, + void *data, int flag) +{ + struct lu_env env; + __u32 sync_lock_cancel = 0; + __u32 len = sizeof(sync_lock_cancel); + int rc = 0; + + ENTRY; + + rc = lu_env_init(&env, LCT_DT_THREAD); + if (unlikely(rc != 0)) + RETURN(rc); + + rc = obd_get_info(&env, lock->l_export, sizeof(KEY_SYNC_LOCK_CANCEL), + KEY_SYNC_LOCK_CANCEL, &len, &sync_lock_cancel, NULL); + if (rc == 0 && flag == LDLM_CB_CANCELING && + (lock->l_granted_mode & (LCK_PW|LCK_GROUP)) && + (sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL || + (sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL && + lock->l_flags & LDLM_FL_CBPENDING))) { + struct obd_info *oinfo; + struct obdo *oa; + int rc; + + OBD_ALLOC_PTR(oinfo); + if (!oinfo) + GOTO(out_env, rc = -ENOMEM); + OBDO_ALLOC(oa); + if (!oa) { + OBD_FREE_PTR(oinfo); + GOTO(out_env, rc = -ENOMEM); + } + + ostid_res_name_to_id(&oa->o_oi, &lock->l_resource->lr_name); + oa->o_valid = OBD_MD_FLID|OBD_MD_FLGROUP; + oinfo->oi_oa = oa; + oinfo->oi_capa = BYPASS_CAPA; + + rc = obd_sync(&env, lock->l_export, oinfo, + lock->l_policy_data.l_extent.start, + lock->l_policy_data.l_extent.end, NULL); + if (rc) + CERROR("Error %d syncing data on lock cancel\n", rc); + + OBDO_FREE(oa); + OBD_FREE_PTR(oinfo); + } + + rc = ldlm_server_blocking_ast(lock, desc, data, flag); +out_env: + lu_env_fini(&env); + RETURN(rc); +} + static int ost_filter_recovery_request(struct ptlrpc_request *req, struct obd_device *obd, int *process) { @@ -1618,7 +1795,7 @@ static int ost_filter_recovery_request(struct ptlrpc_request *req, RETURN(0); default: - DEBUG_REQ(D_ERROR, req, "not permitted during recovery"); + DEBUG_REQ(D_WARNING, req, "not permitted during recovery"); *process = -EAGAIN; RETURN(0); } @@ -1653,11 +1830,8 @@ int ost_msg_check_version(struct lustre_msg *msg) case OST_SYNC: case OST_SET_INFO: case OST_GET_INFO: -#ifdef HAVE_QUOTA_SUPPORT case OST_QUOTACHECK: case OST_QUOTACTL: - case OST_QUOTA_ADJUST_QUNIT: -#endif rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION); if (rc) CERROR("bad opc %u version %08x, expecting %08x\n", @@ -1686,6 +1860,10 @@ int ost_msg_check_version(struct lustre_msg *msg) lustre_msg_get_version(msg), LUSTRE_LOG_VERSION); break; + case OST_QUOTA_ADJUST_QUNIT: + rc = -ENOTSUPP; + CERROR("Quota adjust is deprecated as of 2.4.0\n"); + break; default: CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg)); rc = -ENOTSUPP; @@ -1693,6 +1871,104 @@ int ost_msg_check_version(struct lustre_msg *msg) return rc; } +struct ost_prolong_data { + struct ptlrpc_request *opd_req; + struct obd_export *opd_exp; + struct obdo *opd_oa; + struct ldlm_res_id opd_resid; + struct ldlm_extent opd_extent; + ldlm_mode_t opd_mode; + unsigned int opd_locks; + int opd_timeout; +}; + +/* prolong locks for the current service time of the corresponding + * portal (= OST_IO_PORTAL) + */ +static inline int prolong_timeout(struct ptlrpc_request *req) +{ + struct ptlrpc_service_part *svcpt = req->rq_rqbd->rqbd_svcpt; + + if (AT_OFF) + return obd_timeout / 2; + + return max(at_est2timeout(at_get(&svcpt->scp_at_estimate)), + ldlm_timeout); +} + +static void ost_prolong_lock_one(struct ost_prolong_data *opd, + struct ldlm_lock *lock) +{ + LASSERT(lock->l_export == opd->opd_exp); + + if (lock->l_flags & LDLM_FL_DESTROYED) /* lock already cancelled */ + return; + + /* XXX: never try to grab resource lock here because we're inside + * exp_bl_list_lock; in ldlm_lockd.c to handle waiting list we take + * res lock and then exp_bl_list_lock. */ + + if (!(lock->l_flags & LDLM_FL_AST_SENT)) + /* ignore locks not being cancelled */ + return; + + LDLM_DEBUG(lock, + "refreshed for req x"LPU64" ext("LPU64"->"LPU64") to %ds.\n", + opd->opd_req->rq_xid, opd->opd_extent.start, + opd->opd_extent.end, opd->opd_timeout); + + /* OK. this is a possible lock the user holds doing I/O + * let's refresh eviction timer for it */ + ldlm_refresh_waiting_lock(lock, opd->opd_timeout); + ++opd->opd_locks; +} + +static void ost_prolong_locks(struct ost_prolong_data *data) +{ + struct obd_export *exp = data->opd_exp; + struct obdo *oa = data->opd_oa; + struct ldlm_lock *lock; + ENTRY; + + if (oa->o_valid & OBD_MD_FLHANDLE) { + /* mostly a request should be covered by only one lock, try + * fast path. */ + lock = ldlm_handle2lock(&oa->o_handle); + if (lock != NULL) { + /* Fast path to check if the lock covers the whole IO + * region exclusively. */ + if (lock->l_granted_mode == LCK_PW && + ldlm_extent_contain(&lock->l_policy_data.l_extent, + &data->opd_extent)) { + /* bingo */ + ost_prolong_lock_one(data, lock); + LDLM_LOCK_PUT(lock); + RETURN_EXIT; + } + LDLM_LOCK_PUT(lock); + } + } + + + spin_lock_bh(&exp->exp_bl_list_lock); + cfs_list_for_each_entry(lock, &exp->exp_bl_list, l_exp_list) { + LASSERT(lock->l_flags & LDLM_FL_AST_SENT); + LASSERT(lock->l_resource->lr_type == LDLM_EXTENT); + + if (!ldlm_res_eq(&data->opd_resid, &lock->l_resource->lr_name)) + continue; + + if (!ldlm_extent_overlap(&lock->l_policy_data.l_extent, + &data->opd_extent)) + continue; + + ost_prolong_lock_one(data, lock); + } + spin_unlock_bh(&exp->exp_bl_list_lock); + + EXIT; +} + /** * Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does * not. @@ -1702,57 +1978,34 @@ static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req, { struct niobuf_remote *nb; struct obd_ioobj *ioo; - struct ost_body *body; - int objcount, niocount; - int mode, opc, i; - __u64 start, end; + int mode, opc; + struct ldlm_extent ext; ENTRY; opc = lustre_msg_get_opc(req->rq_reqmsg); LASSERT(opc == OST_READ || opc == OST_WRITE); - /* As the request may be covered by several locks, do not look at - * o_handle, look at the RPC IO region. */ - body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); - if (body == NULL) - RETURN(0); - - objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, - RCL_CLIENT) / sizeof(*ioo); ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); - if (ioo == NULL) - RETURN(0); - - for (niocount = i = 0; i < objcount; i++) - niocount += ioo[i].ioo_bufcnt; + LASSERT(ioo != NULL); nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); - if (nb == NULL || - niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE, - RCL_CLIENT) / sizeof(*nb))) - RETURN(0); + LASSERT(nb != NULL); + + ext.start = nb->offset; + nb += ioo->ioo_bufcnt - 1; + ext.end = nb->offset + nb->len - 1; + + LASSERT(lock->l_resource != NULL); + if (!ostid_res_name_eq(&ioo->ioo_oid, &lock->l_resource->lr_name)) + RETURN(0); mode = LCK_PW; if (opc == OST_READ) mode |= LCK_PR; - - start = nb[0].offset & CFS_PAGE_MASK; - end = (nb[ioo->ioo_bufcnt - 1].offset + - nb[ioo->ioo_bufcnt - 1].len - 1) | ~CFS_PAGE_MASK; - - LASSERT(lock->l_resource != NULL); - if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_gr, - &lock->l_resource->lr_name)) - RETURN(0); - if (!(lock->l_granted_mode & mode)) RETURN(0); - if (lock->l_policy_data.l_extent.end < start || - lock->l_policy_data.l_extent.start > end) - RETURN(0); - - RETURN(1); + RETURN(ldlm_extent_overlap(&lock->l_policy_data.l_extent, &ext)); } /** @@ -1767,74 +2020,62 @@ static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req, */ static int ost_rw_hpreq_check(struct ptlrpc_request *req) { - struct niobuf_remote *nb; - struct obd_ioobj *ioo; + struct obd_device *obd = req->rq_export->exp_obd; struct ost_body *body; - int objcount, niocount; - int mode, opc, i; + struct obd_ioobj *ioo; + struct niobuf_remote *nb; + struct ost_prolong_data opd = { 0 }; + int mode, opc; ENTRY; + /* + * Use LASSERT to do sanity check because malformed RPCs should have + * been filtered out in ost_hpreq_handler(). + */ opc = lustre_msg_get_opc(req->rq_reqmsg); LASSERT(opc == OST_READ || opc == OST_WRITE); body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); - if (body == NULL) - RETURN(-EFAULT); + LASSERT(body != NULL); - objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, - RCL_CLIENT) / sizeof(*ioo); ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ); - if (ioo == NULL) - RETURN(-EFAULT); + LASSERT(ioo != NULL); - for (niocount = i = 0; i < objcount; i++) - niocount += ioo[i].ioo_bufcnt; nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE); - if (nb == NULL || - niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE, - RCL_CLIENT) / sizeof(*nb))) - RETURN(-EFAULT); - if (niocount != 0 && (nb[0].flags & OBD_BRW_SRVLOCK)) - RETURN(-EFAULT); + LASSERT(nb != NULL); + LASSERT(!(nb->flags & OBD_BRW_SRVLOCK)); + + ostid_build_res_name(&ioo->ioo_oid, &opd.opd_resid); + opd.opd_req = req; mode = LCK_PW; if (opc == OST_READ) mode |= LCK_PR; - RETURN(ost_rw_prolong_locks(req, ioo, nb, &body->oa, mode)); -} + opd.opd_mode = mode; + opd.opd_exp = req->rq_export; + opd.opd_oa = &body->oa; + opd.opd_extent.start = nb->offset; + nb += ioo->ioo_bufcnt - 1; + opd.opd_extent.end = nb->offset + nb->len - 1; + opd.opd_timeout = prolong_timeout(req); -static int ost_punch_prolong_locks(struct ptlrpc_request *req, struct obdo *oa) -{ - struct ldlm_res_id res_id = { .name = { oa->o_id } }; - struct ost_prolong_data opd = { 0 }; - __u64 start, end; - ENTRY; + DEBUG_REQ(D_RPCTRACE, req, + "%s %s: refresh rw locks: " LPU64"/"LPU64" ("LPU64"->"LPU64")\n", + obd->obd_name, cfs_current()->comm, + opd.opd_resid.name[0], opd.opd_resid.name[1], + opd.opd_extent.start, opd.opd_extent.end); - start = oa->o_size; - end = start + oa->o_blocks; + ost_prolong_locks(&opd); - opd.opd_mode = LCK_PW; - opd.opd_exp = req->rq_export; - opd.opd_policy.l_extent.start = start & CFS_PAGE_MASK; - if (oa->o_blocks == OBD_OBJECT_EOF || end < start) - opd.opd_policy.l_extent.end = OBD_OBJECT_EOF; - else - opd.opd_policy.l_extent.end = end | ~CFS_PAGE_MASK; - - /* prolong locks for the current service time of the corresponding - * portal (= OST_IO_PORTAL) */ - opd.opd_timeout = AT_OFF ? obd_timeout / 2: - max(at_est2timeout(at_get(&req->rq_rqbd-> - rqbd_service->srv_at_estimate)), ldlm_timeout); - - CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", - res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start, - opd.opd_policy.l_extent.end); - - opd.opd_oa = oa; - ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id, - ost_prolong_locks_iter, &opd); - RETURN(opd.opd_lock_match); + CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n", + obd->obd_name, opd.opd_locks, req); + + RETURN(opd.opd_locks > 0); +} + +static void ost_rw_hpreq_fini(struct ptlrpc_request *req) +{ + (void)ost_rw_hpreq_check(req); } /** @@ -1847,12 +2088,12 @@ static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req, ENTRY; body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); - if (body == NULL) - RETURN(0); /* can't return -EFAULT here */ + LASSERT(body != NULL); if (body->oa.o_valid & OBD_MD_FLHANDLE && body->oa.o_handle.cookie == lock->l_handle.h_cookie) RETURN(1); + RETURN(0); } @@ -1861,30 +2102,68 @@ static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req, */ static int ost_punch_hpreq_check(struct ptlrpc_request *req) { + struct obd_device *obd = req->rq_export->exp_obd; struct ost_body *body; + struct obdo *oa; + struct ost_prolong_data opd = { 0 }; + __u64 start, end; + ENTRY; body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY); - if (body == NULL) - RETURN(-EFAULT); + LASSERT(body != NULL); - LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) || - !(body->oa.o_flags & OBD_FL_SRVLOCK)); + oa = &body->oa; + LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS) || + !(oa->o_flags & OBD_FL_SRVLOCK)); - RETURN(ost_punch_prolong_locks(req, &body->oa)); + start = oa->o_size; + end = start + oa->o_blocks; + + opd.opd_req = req; + opd.opd_mode = LCK_PW; + opd.opd_exp = req->rq_export; + opd.opd_oa = oa; + opd.opd_extent.start = start; + opd.opd_extent.end = end; + if (oa->o_blocks == OBD_OBJECT_EOF) + opd.opd_extent.end = OBD_OBJECT_EOF; + opd.opd_timeout = prolong_timeout(req); + + ostid_build_res_name(&oa->o_oi, &opd.opd_resid); + + CDEBUG(D_DLMTRACE, + "%s: refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n", + obd->obd_name, + opd.opd_resid.name[0], opd.opd_resid.name[1], + opd.opd_extent.start, opd.opd_extent.end); + + ost_prolong_locks(&opd); + + CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n", + obd->obd_name, opd.opd_locks, req); + + RETURN(opd.opd_locks > 0); +} + +static void ost_punch_hpreq_fini(struct ptlrpc_request *req) +{ + (void)ost_punch_hpreq_check(req); } struct ptlrpc_hpreq_ops ost_hpreq_rw = { - .hpreq_lock_match = ost_rw_hpreq_lock_match, - .hpreq_check = ost_rw_hpreq_check, + .hpreq_lock_match = ost_rw_hpreq_lock_match, + .hpreq_check = ost_rw_hpreq_check, + .hpreq_fini = ost_rw_hpreq_fini }; struct ptlrpc_hpreq_ops ost_hpreq_punch = { - .hpreq_lock_match = ost_punch_hpreq_lock_match, - .hpreq_check = ost_punch_hpreq_check, + .hpreq_lock_match = ost_punch_hpreq_lock_match, + .hpreq_check = ost_punch_hpreq_check, + .hpreq_fini = ost_punch_hpreq_fini }; /** Assign high priority operations to the request if needed. */ -static int ost_hpreq_handler(struct ptlrpc_request *req) +static int ost_io_hpreq_handler(struct ptlrpc_request *req) { ENTRY; if (req->rq_export) { @@ -1895,6 +2174,7 @@ static int ost_hpreq_handler(struct ptlrpc_request *req) struct niobuf_remote *nb; struct obd_ioobj *ioo; int objcount, niocount; + int rc; int i; /* RPCs on the H-P queue can be inspected before @@ -1917,6 +2197,7 @@ static int ost_hpreq_handler(struct ptlrpc_request *req) CERROR("Missing/short ost_body\n"); RETURN(-EFAULT); } + objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ, RCL_CLIENT) / @@ -1937,6 +2218,12 @@ static int ost_hpreq_handler(struct ptlrpc_request *req) RETURN(-EFAULT); } + rc = ost_validate_obdo(req->rq_export, &body->oa, ioo); + if (rc) { + CERROR("invalid object ids\n"); + RETURN(rc); + } + for (niocount = i = 0; i < objcount; i++) { if (ioo[i].ioo_bufcnt == 0) { CERROR("ioo[%d] has zero bufcnt\n", i); @@ -1982,31 +2269,35 @@ static int ost_hpreq_handler(struct ptlrpc_request *req) /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */ int ost_handle(struct ptlrpc_request *req) { - struct obd_trans_info trans_info = { 0, }; - struct obd_trans_info *oti = &trans_info; - int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0; - struct obd_device *obd = NULL; - ENTRY; - - LASSERT(current->journal_info == NULL); - - /* primordial rpcs don't affect server recovery */ - switch (lustre_msg_get_opc(req->rq_reqmsg)) { - case SEC_CTX_INIT: - case SEC_CTX_INIT_CONT: - case SEC_CTX_FINI: - GOTO(out, rc = 0); - } - - req_capsule_init(&req->rq_pill, req, RCL_SERVER); - - if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) { - int recovering; - - if (!class_connected_export(req->rq_export)) { - CDEBUG(D_HA,"operation %d on unconnected OST from %s\n", - lustre_msg_get_opc(req->rq_reqmsg), - libcfs_id2str(req->rq_peer)); + struct obd_trans_info trans_info = { 0, }; + struct obd_trans_info *oti = &trans_info; + int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0; + struct obd_device *obd = NULL; + __u32 opc = lustre_msg_get_opc(req->rq_reqmsg); + ENTRY; + + /* OST module is kept between remounts, but the last reference + * to specific module (say, osd or ofd) kills all related keys + * from the environment. so we have to refill it until the root + * cause is fixed properly */ + lu_env_refill(req->rq_svc_thread->t_env); + + LASSERT(current->journal_info == NULL); + + /* primordial rpcs don't affect server recovery */ + switch (opc) { + case SEC_CTX_INIT: + case SEC_CTX_INIT_CONT: + case SEC_CTX_FINI: + GOTO(out, rc = 0); + } + + req_capsule_init(&req->rq_pill, req, RCL_SERVER); + + if (opc != OST_CONNECT) { + if (!class_connected_export(req->rq_export)) { + CDEBUG(D_HA,"operation %d on unconnected OST from %s\n", + opc, libcfs_id2str(req->rq_peer)); req->rq_status = -ENOTCONN; GOTO(out, rc = -ENOTCONN); } @@ -2014,10 +2305,7 @@ int ost_handle(struct ptlrpc_request *req) obd = req->rq_export->exp_obd; /* Check for aborted recovery. */ - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - recovering = obd->obd_recovering; - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); - if (recovering) { + if (obd->obd_recovering) { rc = ost_filter_recovery_request(req, obd, &should_process); if (rc || !should_process) @@ -2036,7 +2324,11 @@ int ost_handle(struct ptlrpc_request *req) if (rc) RETURN(rc); - switch (lustre_msg_get_opc(req->rq_reqmsg)) { + if (req && req->rq_reqmsg && req->rq_export && + (exp_connect_flags(req->rq_export) & OBD_CONNECT_JOBSTATS)) + oti->oti_jobid = lustre_msg_get_jobid(req->rq_reqmsg); + + switch (opc) { case OST_CONNECT: { CDEBUG(D_INODE, "connect\n"); req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT); @@ -2095,18 +2387,18 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE); CDEBUG(D_INODE, "write\n"); /* req->rq_request_portal would be nice, if it was set */ - if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ - CERROR("%s: deny write request from %s to portal %u\n", - req->rq_export->exp_obd->obd_name, - obd_export_nid2str(req->rq_export), - req->rq_rqbd->rqbd_service->srv_req_portal); + if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) { + CERROR("%s: deny write request from %s to portal %u\n", + req->rq_export->exp_obd->obd_name, + obd_export_nid2str(req->rq_export), + ptlrpc_req2svc(req)->srv_req_portal); GOTO(out, rc = -EPROTO); } if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET)) RETURN(0); if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC)) GOTO(out, rc = -ENOSPC); - if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS)) + if (OBD_FAIL_TIMEOUT(OBD_FAIL_OST_EROFS, 1)) GOTO(out, rc = -EROFS); rc = ost_brw_write(req, oti); LASSERT(current->journal_info == NULL); @@ -2116,11 +2408,11 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ); CDEBUG(D_INODE, "read\n"); /* req->rq_request_portal would be nice, if it was set */ - if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){ - CERROR("%s: deny read request from %s to portal %u\n", - req->rq_export->exp_obd->obd_name, - obd_export_nid2str(req->rq_export), - req->rq_rqbd->rqbd_service->srv_req_portal); + if (ptlrpc_req2svc(req)->srv_req_portal != OST_IO_PORTAL) { + CERROR("%s: deny read request from %s to portal %u\n", + req->rq_export->exp_obd->obd_name, + obd_export_nid2str(req->rq_export), + ptlrpc_req2svc(req)->srv_req_portal); GOTO(out, rc = -EPROTO); } if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET)) @@ -2150,7 +2442,7 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_OST_SYNC); if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET)) RETURN(0); - rc = ost_sync(req->rq_export, req); + rc = ost_sync(req->rq_export, req, oti); break; case OST_SET_INFO: DEBUG_REQ(D_INODE, req, "set_info"); @@ -2162,7 +2454,6 @@ int ost_handle(struct ptlrpc_request *req) req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC); rc = ost_get_info(req->rq_export, req); break; -#ifdef HAVE_QUOTA_SUPPORT case OST_QUOTACHECK: CDEBUG(D_INODE, "quotacheck\n"); req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK); @@ -2177,12 +2468,6 @@ int ost_handle(struct ptlrpc_request *req) RETURN(0); rc = ost_handle_quotactl(req); break; - case OST_QUOTA_ADJUST_QUNIT: - CDEBUG(D_INODE, "quota_adjust_qunit\n"); - req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT); - rc = ost_handle_quota_adjust_qunit(req); - break; -#endif case OBD_PING: DEBUG_REQ(D_INODE, req, "ping"); req_capsule_set(&req->rq_pill, &RQF_OBD_PING); @@ -2198,51 +2483,37 @@ int ost_handle(struct ptlrpc_request *req) if (rc) RETURN(rc); RETURN(ptlrpc_reply(req)); - case OBD_LOG_CANCEL: - CDEBUG(D_INODE, "log cancel\n"); - req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET)) - RETURN(0); - rc = llog_origin_handle_cancel(req); - if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP)) - RETURN(0); - req->rq_status = rc; - rc = req_capsule_server_pack(&req->rq_pill); - if (rc) - RETURN(rc); - RETURN(ptlrpc_reply(req)); - case LDLM_ENQUEUE: - CDEBUG(D_INODE, "enqueue\n"); - req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE)) - RETURN(0); - rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast, - ldlm_server_blocking_ast, - ldlm_server_glimpse_ast); - fail = OBD_FAIL_OST_LDLM_REPLY_NET; - break; - case LDLM_CONVERT: - CDEBUG(D_INODE, "convert\n"); - req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT)) - RETURN(0); - rc = ldlm_handle_convert(req); - break; - case LDLM_CANCEL: - CDEBUG(D_INODE, "cancel\n"); - req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); - if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL)) - RETURN(0); - rc = ldlm_handle_cancel(req); - break; + case LDLM_ENQUEUE: + CDEBUG(D_INODE, "enqueue\n"); + req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE); + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE_NET)) + RETURN(0); + rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast, + ost_blocking_ast, + ldlm_server_glimpse_ast); + fail = OBD_FAIL_OST_LDLM_REPLY_NET; + break; + case LDLM_CONVERT: + CDEBUG(D_INODE, "convert\n"); + req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT); + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT_NET)) + RETURN(0); + rc = ldlm_handle_convert(req); + break; + case LDLM_CANCEL: + CDEBUG(D_INODE, "cancel\n"); + req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL); + if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL_NET)) + RETURN(0); + rc = ldlm_handle_cancel(req); + break; case LDLM_BL_CALLBACK: case LDLM_CP_CALLBACK: CDEBUG(D_INODE, "callback\n"); CERROR("callbacks should not happen on OST\n"); /* fall through */ default: - CERROR("Unexpected opcode %d\n", - lustre_msg_get_opc(req->rq_reqmsg)); + CERROR("Unexpected opcode %d\n", opc); req->rq_status = -ENOTSUPP; rc = ptlrpc_error(req); RETURN(rc); @@ -2252,7 +2523,7 @@ int ost_handle(struct ptlrpc_request *req) EXIT; /* If we're DISCONNECTing, the export_data is already freed */ - if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT) + if (!rc && opc != OST_DISCONNECT) target_committed_to_req(req); out: @@ -2263,10 +2534,11 @@ out: return 0; } EXPORT_SYMBOL(ost_handle); + /* - * free per-thread pool created by ost_thread_init(). + * free per-thread pool created by ost_io_thread_init(). */ -static void ost_thread_done(struct ptlrpc_thread *thread) +static void ost_io_thread_done(struct ptlrpc_thread *thread) { struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local * Storage */ @@ -2277,7 +2549,7 @@ static void ost_thread_done(struct ptlrpc_thread *thread) /* * be prepared to handle partially-initialized pools (because this is - * called from ost_thread_init() for cleanup. + * called from ost_io_thread_init() for cleanup. */ tls = thread->t_data; if (tls != NULL) { @@ -2290,7 +2562,7 @@ static void ost_thread_done(struct ptlrpc_thread *thread) /* * initialize per-thread page pool (bug 5137). */ -static int ost_thread_init(struct ptlrpc_thread *thread) +static int ost_io_thread_init(struct ptlrpc_thread *thread) { struct ost_thread_local_cache *tls; @@ -2298,7 +2570,6 @@ static int ost_thread_init(struct ptlrpc_thread *thread) LASSERT(thread != NULL); LASSERT(thread->t_data == NULL); - LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id); OBD_ALLOC_PTR(tls); if (tls == NULL) @@ -2309,17 +2580,17 @@ static int ost_thread_init(struct ptlrpc_thread *thread) #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000) +static struct cfs_cpt_table *ost_io_cptable; + /* Sigh - really, this is an OSS, the _server_, not the _target_ */ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) { - struct ost_obd *ost = &obd->u.ost; - struct lprocfs_static_vars lvars; - int oss_min_threads; - int oss_max_threads; - int oss_min_create_threads; - int oss_max_create_threads; - int rc; - ENTRY; + static struct ptlrpc_service_conf svc_conf; + struct ost_obd *ost = &obd->u.ost; + struct lprocfs_static_vars lvars; + nodemask_t *mask; + int rc; + ENTRY; rc = cfs_cleanup_group_info(); if (rc) @@ -2328,101 +2599,246 @@ static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg) lprocfs_ost_init_vars(&lvars); lprocfs_obd_setup(obd, lvars.obd_vars); - cfs_sema_init(&ost->ost_health_sem, 1); - - if (oss_num_threads) { - /* If oss_num_threads is set, it is the min and the max. */ - if (oss_num_threads > OSS_THREADS_MAX) - oss_num_threads = OSS_THREADS_MAX; - if (oss_num_threads < OSS_THREADS_MIN) - oss_num_threads = OSS_THREADS_MIN; - oss_max_threads = oss_min_threads = oss_num_threads; - } else { - /* Base min threads on memory and cpus */ - oss_min_threads = - cfs_num_possible_cpus() * CFS_NUM_CACHEPAGES >> - (27 - CFS_PAGE_SHIFT); - if (oss_min_threads < OSS_THREADS_MIN) - oss_min_threads = OSS_THREADS_MIN; - /* Insure a 4x range for dynamic threads */ - if (oss_min_threads > OSS_THREADS_MAX / 4) - oss_min_threads = OSS_THREADS_MAX / 4; - oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1); - } - - ost->ost_service = - ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE, - OST_MAXREPSIZE, OST_REQUEST_PORTAL, - OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR, - ost_handle, LUSTRE_OSS_NAME, - obd->obd_proc_entry, target_print_req, - oss_min_threads, oss_max_threads, - "ll_ost", LCT_DT_THREAD, NULL); - if (ost->ost_service == NULL) { - CERROR("failed to start service\n"); - GOTO(out_lprocfs, rc = -ENOMEM); - } - - rc = ptlrpc_start_threads(obd, ost->ost_service); - if (rc) - GOTO(out_service, rc = -EINVAL); - - if (oss_num_create_threads) { - if (oss_num_create_threads > OSS_MAX_CREATE_THREADS) - oss_num_create_threads = OSS_MAX_CREATE_THREADS; - if (oss_num_create_threads < OSS_MIN_CREATE_THREADS) - oss_num_create_threads = OSS_MIN_CREATE_THREADS; - oss_min_create_threads = oss_max_create_threads = - oss_num_create_threads; - } else { - oss_min_create_threads = OSS_MIN_CREATE_THREADS; - oss_max_create_threads = OSS_MAX_CREATE_THREADS; + mutex_init(&ost->ost_health_mutex); + + svc_conf = (typeof(svc_conf)) { + .psc_name = LUSTRE_OSS_NAME, + .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR, + .psc_buf = { + .bc_nbufs = OST_NBUFS, + .bc_buf_size = OST_BUFSIZE, + .bc_req_max_size = OST_MAXREQSIZE, + .bc_rep_max_size = OST_MAXREPSIZE, + .bc_req_portal = OST_REQUEST_PORTAL, + .bc_rep_portal = OSC_REPLY_PORTAL, + }, + .psc_thr = { + .tc_thr_name = "ll_ost", + .tc_thr_factor = OSS_THR_FACTOR, + .tc_nthrs_init = OSS_NTHRS_INIT, + .tc_nthrs_base = OSS_NTHRS_BASE, + .tc_nthrs_max = OSS_NTHRS_MAX, + .tc_nthrs_user = oss_num_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_DT_THREAD, + }, + .psc_cpt = { + .cc_pattern = oss_cpts, + }, + .psc_ops = { + .so_req_handler = ost_handle, + .so_req_printer = target_print_req, + .so_hpreq_handler = ptlrpc_hpreq_handler, + }, + }; + ost->ost_service = ptlrpc_register_service(&svc_conf, + obd->obd_proc_entry); + if (IS_ERR(ost->ost_service)) { + rc = PTR_ERR(ost->ost_service); + CERROR("failed to start service: %d\n", rc); + GOTO(out_lprocfs, rc); } - ost->ost_create_service = - ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE, - OST_MAXREPSIZE, OST_CREATE_PORTAL, - OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR, - ost_handle, "ost_create", - obd->obd_proc_entry, target_print_req, - oss_min_create_threads, oss_max_create_threads, - "ll_ost_creat", LCT_DT_THREAD, NULL); - if (ost->ost_create_service == NULL) { - CERROR("failed to start OST create service\n"); - GOTO(out_service, rc = -ENOMEM); + memset(&svc_conf, 0, sizeof(svc_conf)); + svc_conf = (typeof(svc_conf)) { + .psc_name = "ost_create", + .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR, + .psc_buf = { + .bc_nbufs = OST_NBUFS, + .bc_buf_size = OST_BUFSIZE, + .bc_req_max_size = OST_MAXREQSIZE, + .bc_rep_max_size = OST_MAXREPSIZE, + .bc_req_portal = OST_CREATE_PORTAL, + .bc_rep_portal = OSC_REPLY_PORTAL, + }, + .psc_thr = { + .tc_thr_name = "ll_ost_create", + .tc_thr_factor = OSS_CR_THR_FACTOR, + .tc_nthrs_init = OSS_CR_NTHRS_INIT, + .tc_nthrs_base = OSS_CR_NTHRS_BASE, + .tc_nthrs_max = OSS_CR_NTHRS_MAX, + .tc_nthrs_user = oss_num_create_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_DT_THREAD, + }, + .psc_cpt = { + .cc_pattern = oss_cpts, + }, + .psc_ops = { + .so_req_handler = ost_handle, + .so_req_printer = target_print_req, + }, + }; + ost->ost_create_service = ptlrpc_register_service(&svc_conf, + obd->obd_proc_entry); + if (IS_ERR(ost->ost_create_service)) { + rc = PTR_ERR(ost->ost_create_service); + CERROR("failed to start OST create service: %d\n", rc); + GOTO(out_service, rc); } - rc = ptlrpc_start_threads(obd, ost->ost_create_service); - if (rc) - GOTO(out_create, rc = -EINVAL); - - ost->ost_io_service = - ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE, - OST_MAXREPSIZE, OST_IO_PORTAL, - OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR, - ost_handle, "ost_io", - obd->obd_proc_entry, target_print_req, - oss_min_threads, oss_max_threads, - "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler); - if (ost->ost_io_service == NULL) { - CERROR("failed to start OST I/O service\n"); - GOTO(out_create, rc = -ENOMEM); + mask = cfs_cpt_table->ctb_nodemask; + /* event CPT feature is disabled in libcfs level by set partition + * number to 1, we still want to set node affinity for io service */ + if (cfs_cpt_number(cfs_cpt_table) == 1 && nodes_weight(*mask) > 1) { + int cpt = 0; + int i; + + ost_io_cptable = cfs_cpt_table_alloc(nodes_weight(*mask)); + for_each_node_mask(i, *mask) { + if (ost_io_cptable == NULL) { + CWARN("OSS failed to create CPT table\n"); + break; + } + + rc = cfs_cpt_set_node(ost_io_cptable, cpt++, i); + if (!rc) { + CWARN("OSS Failed to set node %d for" + "IO CPT table\n", i); + cfs_cpt_table_free(ost_io_cptable); + ost_io_cptable = NULL; + break; + } + } + } + + memset(&svc_conf, 0, sizeof(svc_conf)); + svc_conf = (typeof(svc_conf)) { + .psc_name = "ost_io", + .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR, + .psc_buf = { + .bc_nbufs = OST_NBUFS, + .bc_buf_size = OST_IO_BUFSIZE, + .bc_req_max_size = OST_IO_MAXREQSIZE, + .bc_rep_max_size = OST_IO_MAXREPSIZE, + .bc_req_portal = OST_IO_PORTAL, + .bc_rep_portal = OSC_REPLY_PORTAL, + }, + .psc_thr = { + .tc_thr_name = "ll_ost_io", + .tc_thr_factor = OSS_THR_FACTOR, + .tc_nthrs_init = OSS_NTHRS_INIT, + .tc_nthrs_base = OSS_NTHRS_BASE, + .tc_nthrs_max = OSS_NTHRS_MAX, + .tc_nthrs_user = oss_num_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_DT_THREAD, + }, + .psc_cpt = { + .cc_cptable = ost_io_cptable, + .cc_pattern = ost_io_cptable == NULL ? + oss_io_cpts : NULL, + }, + .psc_ops = { + .so_thr_init = ost_io_thread_init, + .so_thr_done = ost_io_thread_done, + .so_req_handler = ost_handle, + .so_hpreq_handler = ost_io_hpreq_handler, + .so_req_printer = target_print_req, + }, + }; + ost->ost_io_service = ptlrpc_register_service(&svc_conf, + obd->obd_proc_entry); + if (IS_ERR(ost->ost_io_service)) { + rc = PTR_ERR(ost->ost_io_service); + CERROR("failed to start OST I/O service: %d\n", rc); + ost->ost_io_service = NULL; + GOTO(out_create, rc); } - ost->ost_io_service->srv_init = ost_thread_init; - ost->ost_io_service->srv_done = ost_thread_done; - ost->ost_io_service->srv_cpu_affinity = 1; - rc = ptlrpc_start_threads(obd, ost->ost_io_service); - if (rc) - GOTO(out_io, rc = -EINVAL); - - ping_evictor_start(); - - RETURN(0); + memset(&svc_conf, 0, sizeof(svc_conf)); + svc_conf = (typeof(svc_conf)) { + .psc_name = "ost_seq", + .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR, + .psc_buf = { + .bc_nbufs = OST_NBUFS, + .bc_buf_size = OST_BUFSIZE, + .bc_req_max_size = OST_MAXREQSIZE, + .bc_rep_max_size = OST_MAXREPSIZE, + .bc_req_portal = SEQ_DATA_PORTAL, + .bc_rep_portal = OSC_REPLY_PORTAL, + }, + .psc_thr = { + .tc_thr_name = "ll_ost_seq", + .tc_thr_factor = OSS_CR_THR_FACTOR, + .tc_nthrs_init = OSS_CR_NTHRS_INIT, + .tc_nthrs_base = OSS_CR_NTHRS_BASE, + .tc_nthrs_max = OSS_CR_NTHRS_MAX, + .tc_nthrs_user = oss_num_create_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_DT_THREAD, + }, + + .psc_cpt = { + .cc_pattern = oss_cpts, + }, + .psc_ops = { + .so_req_handler = tgt_request_handle, + .so_req_printer = target_print_req, + .so_hpreq_handler = NULL, + }, + }; + ost->ost_seq_service = ptlrpc_register_service(&svc_conf, + obd->obd_proc_entry); + if (IS_ERR(ost->ost_seq_service)) { + rc = PTR_ERR(ost->ost_seq_service); + CERROR("failed to start OST seq service: %d\n", rc); + ost->ost_seq_service = NULL; + GOTO(out_io, rc); + } + +#if 0 + /* Object update service */ + memset(&svc_conf, 0, sizeof(svc_conf)); + svc_conf = (typeof(svc_conf)) { + .psc_name = "ost_out", + .psc_watchdog_factor = OSS_SERVICE_WATCHDOG_FACTOR, + .psc_buf = { + .bc_nbufs = OST_NBUFS, + .bc_buf_size = OUT_BUFSIZE, + .bc_req_max_size = OUT_MAXREQSIZE, + .bc_rep_max_size = OUT_MAXREPSIZE, + .bc_req_portal = OUT_PORTAL, + .bc_rep_portal = OSC_REPLY_PORTAL, + }, + /* + * We'd like to have a mechanism to set this on a per-device + * basis, but alas... + */ + .psc_thr = { + .tc_thr_name = "ll_ost_out", + .tc_thr_factor = OSS_CR_THR_FACTOR, + .tc_nthrs_init = OSS_CR_NTHRS_INIT, + .tc_nthrs_base = OSS_CR_NTHRS_BASE, + .tc_nthrs_max = OSS_CR_NTHRS_MAX, + .tc_nthrs_user = oss_num_create_threads, + .tc_cpu_affinity = 1, + .tc_ctx_tags = LCT_DT_THREAD, + }, + .psc_cpt = { + .cc_pattern = oss_cpts, + }, + .psc_ops = { + .so_req_handler = tgt_request_handle, + .so_req_printer = target_print_req, + .so_hpreq_handler = NULL, + }, + }; + ost->ost_out_service = ptlrpc_register_service(&svc_conf, + obd->obd_proc_entry); + if (IS_ERR(ost->ost_out_service)) { + rc = PTR_ERR(ost->ost_out_service); + CERROR("failed to start out service: %d\n", rc); + ost->ost_out_service = NULL; + GOTO(out_seq, rc); + } +#endif + ping_evictor_start(); + RETURN(0); out_io: - ptlrpc_unregister_service(ost->ost_io_service); - ost->ost_io_service = NULL; + ptlrpc_unregister_service(ost->ost_io_service); + ost->ost_io_service = NULL; out_create: ptlrpc_unregister_service(ost->ost_create_service); ost->ost_create_service = NULL; @@ -2436,42 +2852,51 @@ out_lprocfs: static int ost_cleanup(struct obd_device *obd) { - struct ost_obd *ost = &obd->u.ost; - int err = 0; - ENTRY; - - ping_evictor_stop(); + struct ost_obd *ost = &obd->u.ost; + int err = 0; + ENTRY; + + ping_evictor_stop(); + + /* there is no recovery for OST OBD, all recovery is controlled by + * obdfilter OBD */ + LASSERT(obd->obd_recovering == 0); + mutex_lock(&ost->ost_health_mutex); + ptlrpc_unregister_service(ost->ost_service); + ptlrpc_unregister_service(ost->ost_create_service); + ptlrpc_unregister_service(ost->ost_io_service); + ptlrpc_unregister_service(ost->ost_seq_service); +#if 0 + ptlrpc_unregister_service(ost->ost_out_service); +#endif + ost->ost_service = NULL; + ost->ost_create_service = NULL; + ost->ost_io_service = NULL; + ost->ost_seq_service = NULL; + ost->ost_out_service = NULL; - cfs_spin_lock_bh(&obd->obd_processing_task_lock); - if (obd->obd_recovering) { - target_cancel_recovery_timer(obd); - obd->obd_recovering = 0; - } - cfs_spin_unlock_bh(&obd->obd_processing_task_lock); + mutex_unlock(&ost->ost_health_mutex); - cfs_down(&ost->ost_health_sem); - ptlrpc_unregister_service(ost->ost_service); - ptlrpc_unregister_service(ost->ost_create_service); - ptlrpc_unregister_service(ost->ost_io_service); - ost->ost_service = NULL; - ost->ost_create_service = NULL; - cfs_up(&ost->ost_health_sem); + lprocfs_obd_cleanup(obd); - lprocfs_obd_cleanup(obd); + if (ost_io_cptable != NULL) { + cfs_cpt_table_free(ost_io_cptable); + ost_io_cptable = NULL; + } - RETURN(err); + RETURN(err); } -static int ost_health_check(struct obd_device *obd) +static int ost_health_check(const struct lu_env *env, struct obd_device *obd) { struct ost_obd *ost = &obd->u.ost; int rc = 0; - cfs_down(&ost->ost_health_sem); + mutex_lock(&ost->ost_health_mutex); rc |= ptlrpc_service_health_check(ost->ost_service); rc |= ptlrpc_service_health_check(ost->ost_create_service); rc |= ptlrpc_service_health_check(ost->ost_io_service); - cfs_up(&ost->ost_health_sem); + mutex_unlock(&ost->ost_health_mutex); /* * health_check to return 0 on healthy @@ -2503,6 +2928,8 @@ static int __init ost_init(void) int rc; ENTRY; + ost_page_to_corrupt = alloc_page(GFP_IOFS); + lprocfs_ost_init_vars(&lvars); rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars, LUSTRE_OSS_NAME, NULL); @@ -2519,6 +2946,9 @@ static int __init ost_init(void) static void /*__exit*/ ost_exit(void) { + if (ost_page_to_corrupt) + page_cache_release(ost_page_to_corrupt); + class_unregister_type(LUSTRE_OSS_NAME); }