X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fofd%2Fofd_io.c;h=892abb8d62ab36fa99da6d2c9375f3e9a50ecdcc;hp=a75e5cd75cfeaeb0681e6d915b574b3f05180a31;hb=490ea83a486c0c44274f4d4f6f77d8e0e16835d5;hpb=a67ea1c51daad1a2c774bdceeabbfac927d348e6 diff --git a/lustre/ofd/ofd_io.c b/lustre/ofd/ofd_io.c index a75e5cd..892abb8 100644 --- a/lustre/ofd/ofd_io.c +++ b/lustre/ofd/ofd_io.c @@ -15,11 +15,7 @@ * * You should have received a copy of the GNU General Public License * version 2 along with this program; If not, see - * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf - * - * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, - * CA 95054 USA or visit www.sun.com if you need additional information or - * have any questions. + * http://www.gnu.org/licenses/gpl-2.0.html * * GPL HEADER END */ @@ -27,7 +23,7 @@ * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2011, 2012, Whamcloud, Inc. + * Copyright (c) 2012, 2014 Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -35,17 +31,417 @@ * * lustre/ofd/ofd_io.c * - * Author: Alex Tomas + * This file provides functions to handle IO requests from clients and + * also LFSCK routines to check parent file identifier (PFID) consistency. + * + * Author: Alexey Zhuravlev + * Author: Fan Yong */ #define DEBUG_SUBSYSTEM S_FILTER +#include #include "ofd_internal.h" -static int ofd_preprw_read(const struct lu_env *env, struct ofd_device *ofd, - struct lu_fid *fid, struct lu_attr *la, int niocount, +struct ofd_inconsistency_item { + struct list_head oii_list; + struct ofd_object *oii_obj; + struct lu_fid oii_pfid; +}; + +/** + * Verify single object for parent FID consistency. + * + * Part of LFSCK processing which checks single object PFID stored in extended + * attribute (XATTR) against real FID of MDT parent object received by LFSCK. + * This verifies that the OST object is being referenced by only a single MDT + * object. + * + * \param[in] env execution environment + * \param[in] ofd OFD device + * \param[in] oii object-related local data + * \param[in] lr LFSCK request data + */ +static void ofd_inconsistency_verify_one(const struct lu_env *env, + struct ofd_device *ofd, + struct ofd_inconsistency_item *oii, + struct lfsck_request *lr) +{ + struct ofd_object *fo = oii->oii_obj; + struct lu_fid *pfid = &fo->ofo_pfid; + int rc; + + LASSERT(fo->ofo_pfid_checking); + LASSERT(!fo->ofo_pfid_verified); + + lr->lr_fid = fo->ofo_header.loh_fid; /* OST-object itself FID. */ + lr->lr_fid2 = oii->oii_pfid; /* client given PFID. */ + lr->lr_fid3 = *pfid; /* OST local stored PFID. */ + + rc = lfsck_in_notify(env, ofd->ofd_osd, lr, NULL); + ofd_write_lock(env, fo); + switch (lr->lr_status) { + case LPVS_INIT: + LASSERT(rc <= 0); + + if (rc < 0) + CDEBUG(D_LFSCK, "%s: fail to verify OST local stored " + "PFID xattr for "DFID", the client given PFID " + DFID", OST local stored PFID "DFID": rc = %d\n", + ofd_name(ofd), PFID(&fo->ofo_header.loh_fid), + PFID(&oii->oii_pfid), PFID(pfid), rc); + else + fo->ofo_pfid_verified = 1; + break; + case LPVS_INCONSISTENT: + LASSERT(rc != 0); + + ofd->ofd_inconsistency_self_detected++; + if (rc < 0) + CDEBUG(D_LFSCK, "%s: fail to verify the client given " + "PFID for "DFID", the client given PFID "DFID + ", local stored PFID "DFID": rc = %d\n", + ofd_name(ofd), PFID(&fo->ofo_header.loh_fid), + PFID(&oii->oii_pfid), PFID(pfid), rc); + else + CDEBUG(D_LFSCK, "%s: both the client given PFID and " + "the OST local stored PFID are stale for the " + "OST-object "DFID", client given PFID is "DFID + ", local stored PFID is "DFID"\n", + ofd_name(ofd), PFID(&fo->ofo_header.loh_fid), + PFID(&oii->oii_pfid), PFID(pfid)); + break; + case LPVS_INCONSISTENT_TOFIX: + ofd->ofd_inconsistency_self_detected++; + if (rc == 0) { + ofd->ofd_inconsistency_self_repaired++; + CDEBUG(D_LFSCK, "%s: fixed the staled OST PFID xattr " + "for "DFID", with the client given PFID "DFID + ", the old stored PFID "DFID"\n", + ofd_name(ofd), PFID(&fo->ofo_header.loh_fid), + PFID(&oii->oii_pfid), PFID(pfid)); + } else if (rc < 0) { + CDEBUG(D_LFSCK, "%s: fail to fix the OST PFID xattr " + "for "DFID", client given PFID "DFID", local " + "stored PFID "DFID": rc = %d\n", + ofd_name(ofd), PFID(&fo->ofo_header.loh_fid), + PFID(&oii->oii_pfid), PFID(pfid), rc); + } + *pfid = oii->oii_pfid; + fo->ofo_pfid_verified = 1; + break; + default: + break; + } + fo->ofo_pfid_checking = 0; + ofd_write_unlock(env, fo); + + lu_object_put(env, &fo->ofo_obj.do_lu); + OBD_FREE_PTR(oii); +} + +/** + * Verification thread to check parent FID consistency. + * + * Kernel thread to check consistency of parent FID for any + * new item added for checking by ofd_add_inconsistency_item(). + * + * \param[in] args OFD device + * + * \retval 0 on successful thread termination + * \retval negative value if thread can't start + */ +static int ofd_inconsistency_verification_main(void *args) +{ + struct lu_env env; + struct ofd_device *ofd = args; + struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread; + struct ofd_inconsistency_item *oii; + struct lfsck_request *lr = NULL; + struct l_wait_info lwi = { 0 }; + int rc; + ENTRY; + + rc = lu_env_init(&env, LCT_DT_THREAD); + spin_lock(&ofd->ofd_inconsistency_lock); + thread_set_flags(thread, rc != 0 ? SVC_STOPPED : SVC_RUNNING); + wake_up_all(&thread->t_ctl_waitq); + spin_unlock(&ofd->ofd_inconsistency_lock); + if (rc != 0) + RETURN(rc); + + OBD_ALLOC_PTR(lr); + if (unlikely(lr == NULL)) + GOTO(out_unlocked, rc = -ENOMEM); + + lr->lr_event = LE_PAIRS_VERIFY; + lr->lr_active = LFSCK_TYPE_LAYOUT; + + spin_lock(&ofd->ofd_inconsistency_lock); + while (1) { + if (unlikely(!thread_is_running(thread))) + break; + + while (!list_empty(&ofd->ofd_inconsistency_list)) { + oii = list_entry(ofd->ofd_inconsistency_list.next, + struct ofd_inconsistency_item, + oii_list); + list_del_init(&oii->oii_list); + spin_unlock(&ofd->ofd_inconsistency_lock); + ofd_inconsistency_verify_one(&env, ofd, oii, lr); + spin_lock(&ofd->ofd_inconsistency_lock); + } + + spin_unlock(&ofd->ofd_inconsistency_lock); + l_wait_event(thread->t_ctl_waitq, + !list_empty(&ofd->ofd_inconsistency_list) || + !thread_is_running(thread), + &lwi); + spin_lock(&ofd->ofd_inconsistency_lock); + } + + while (!list_empty(&ofd->ofd_inconsistency_list)) { + struct ofd_object *fo; + + oii = list_entry(ofd->ofd_inconsistency_list.next, + struct ofd_inconsistency_item, + oii_list); + list_del_init(&oii->oii_list); + fo = oii->oii_obj; + spin_unlock(&ofd->ofd_inconsistency_lock); + + ofd_write_lock(&env, fo); + fo->ofo_pfid_checking = 0; + ofd_write_unlock(&env, fo); + + lu_object_put(&env, &fo->ofo_obj.do_lu); + OBD_FREE_PTR(oii); + spin_lock(&ofd->ofd_inconsistency_lock); + } + + OBD_FREE_PTR(lr); + + GOTO(out, rc = 0); + +out_unlocked: + spin_lock(&ofd->ofd_inconsistency_lock); +out: + thread_set_flags(thread, SVC_STOPPED); + wake_up_all(&thread->t_ctl_waitq); + spin_unlock(&ofd->ofd_inconsistency_lock); + lu_env_fini(&env); + + return rc; +} + +/** + * Start parent FID verification thread. + * + * See ofd_inconsistency_verification_main(). + * + * \param[in] ofd OFD device + * + * \retval 0 on successful start of thread + * \retval negative value on error + */ +int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd) +{ + struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread; + struct l_wait_info lwi = { 0 }; + struct task_struct *task; + int rc; + + spin_lock(&ofd->ofd_inconsistency_lock); + if (unlikely(thread_is_running(thread))) { + spin_unlock(&ofd->ofd_inconsistency_lock); + + return -EALREADY; + } + + thread_set_flags(thread, 0); + spin_unlock(&ofd->ofd_inconsistency_lock); + task = kthread_run(ofd_inconsistency_verification_main, ofd, + "inconsistency_verification"); + if (IS_ERR(task)) { + rc = PTR_ERR(task); + CERROR("%s: cannot start self_repair thread: rc = %d\n", + ofd_name(ofd), rc); + } else { + rc = 0; + l_wait_event(thread->t_ctl_waitq, + thread_is_running(thread) || + thread_is_stopped(thread), + &lwi); + } + + return rc; +} + +/** + * Stop parent FID verification thread. + * + * \param[in] ofd OFD device + * + * \retval 0 on successful start of thread + * \retval -EALREADY if thread is already stopped + */ +int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd) +{ + struct ptlrpc_thread *thread = &ofd->ofd_inconsistency_thread; + struct l_wait_info lwi = { 0 }; + + spin_lock(&ofd->ofd_inconsistency_lock); + if (thread_is_init(thread) || thread_is_stopped(thread)) { + spin_unlock(&ofd->ofd_inconsistency_lock); + + return -EALREADY; + } + + thread_set_flags(thread, SVC_STOPPING); + spin_unlock(&ofd->ofd_inconsistency_lock); + wake_up_all(&thread->t_ctl_waitq); + l_wait_event(thread->t_ctl_waitq, + thread_is_stopped(thread), + &lwi); + + return 0; +} + +/** + * Add new item for parent FID verification. + * + * Prepare new verification item and pass it to the dedicated + * verification thread for further processing. + * + * \param[in] env execution environment + * \param[in] fo OFD object + * \param[in] oa OBDO structure with PFID + */ +static void ofd_add_inconsistency_item(const struct lu_env *env, + struct ofd_object *fo, struct obdo *oa) +{ + struct ofd_device *ofd = ofd_obj2dev(fo); + struct ofd_inconsistency_item *oii; + bool wakeup = false; + + OBD_ALLOC_PTR(oii); + if (oii == NULL) + return; + + INIT_LIST_HEAD(&oii->oii_list); + lu_object_get(&fo->ofo_obj.do_lu); + oii->oii_obj = fo; + oii->oii_pfid.f_seq = oa->o_parent_seq; + oii->oii_pfid.f_oid = oa->o_parent_oid; + oii->oii_pfid.f_stripe_idx = oa->o_stripe_idx; + + spin_lock(&ofd->ofd_inconsistency_lock); + if (fo->ofo_pfid_checking || fo->ofo_pfid_verified) { + spin_unlock(&ofd->ofd_inconsistency_lock); + OBD_FREE_PTR(oii); + + return; + } + + fo->ofo_pfid_checking = 1; + if (list_empty(&ofd->ofd_inconsistency_list)) + wakeup = true; + list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list); + spin_unlock(&ofd->ofd_inconsistency_lock); + if (wakeup) + wake_up_all(&ofd->ofd_inconsistency_thread.t_ctl_waitq); + + /* XXX: When the found inconsistency exceeds some threshold, + * we can trigger the LFSCK to scan part of the system + * or the whole system, which depends on how to define + * the threshold, a simple way maybe like that: define + * the absolute value of how many inconsisteny allowed + * to be repaired via self detect/repair mechanism, if + * exceeded, then trigger the LFSCK to scan the layout + * inconsistency within the whole system. */ +} + +/** + * Verify parent FID of an object. + * + * Check the parent FID is sane and start extended + * verification procedure otherwise. + * + * \param[in] env execution environment + * \param[in] fo OFD object + * \param[in] oa OBDO structure with PFID + * + * \retval 0 on successful verification + * \retval -EINPROGRESS if PFID is being repaired + * \retval -EPERM if PFID was verified but still insane + */ +int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo, + struct obdo *oa) +{ + struct lu_fid *pfid = &fo->ofo_pfid; + int rc = 0; + ENTRY; + + if (fid_is_sane(pfid)) { + if (likely(oa->o_parent_seq == pfid->f_seq && + oa->o_parent_oid == pfid->f_oid && + oa->o_stripe_idx == pfid->f_stripe_idx)) + RETURN(0); + + if (fo->ofo_pfid_verified) + RETURN(-EPERM); + } + + /* The OST-object may be inconsistent, and we need further verification. + * To avoid block the RPC service thread, return -EINPROGRESS to client + * and make it retry later. */ + if (fo->ofo_pfid_checking) + RETURN(-EINPROGRESS); + + rc = ofd_object_ff_load(env, fo); + if (rc == -ENODATA) + RETURN(0); + + if (rc < 0) + RETURN(rc); + + if (likely(oa->o_parent_seq == pfid->f_seq && + oa->o_parent_oid == pfid->f_oid && + oa->o_stripe_idx == pfid->f_stripe_idx)) + RETURN(0); + + /* Push it to the dedicated thread for further verification. */ + ofd_add_inconsistency_item(env, fo, oa); + + RETURN(-EINPROGRESS); +} + +/** + * Prepare buffers for read request processing. + * + * This function converts remote buffers from client to local buffers + * and prepares the latter. + * + * \param[in] env execution environment + * \param[in] exp OBD export of client + * \param[in] ofd OFD device + * \param[in] fid FID of object + * \param[in] la object attributes + * \param[in] oa OBDO structure from client + * \param[in] niocount number of remote buffers + * \param[in] rnb remote buffers + * \param[in] nr_local number of local buffers + * \param[in] lnb local buffers + * \param[in] jobid job ID name + * + * \retval 0 on successful prepare + * \retval negative value on error + */ +static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp, + struct ofd_device *ofd, const struct lu_fid *fid, + struct lu_attr *la, struct obdo *oa, int niocount, struct niobuf_remote *rnb, int *nr_local, - struct niobuf_local *lnb) + struct niobuf_local *lnb, char *jobid) { struct ofd_object *fo; int i, j, rc, tot_bytes = 0; @@ -62,30 +458,36 @@ static int ofd_preprw_read(const struct lu_env *env, struct ofd_device *ofd, if (!ofd_object_exists(fo)) GOTO(unlock, rc = -ENOENT); - /* parse remote buffers to local buffers and prepare the latter */ + if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) { + rc = ofd_verify_ff(env, fo, oa); + if (rc != 0) + GOTO(unlock, rc); + } + + *nr_local = 0; for (i = 0, j = 0; i < niocount; i++) { rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i, - lnb + j, 0, ofd_object_capa(env, fo)); - LASSERT(rc > 0); + lnb + j, 0); + if (unlikely(rc < 0)) + GOTO(buf_put, rc); LASSERT(rc <= PTLRPC_MAX_BRW_PAGES); /* correct index for local buffers to continue with */ j += rc; + *nr_local += rc; LASSERT(j <= PTLRPC_MAX_BRW_PAGES); tot_bytes += rnb[i].rnb_len; } - *nr_local = j; LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES); - rc = dt_attr_get(env, ofd_object_child(fo), la, - ofd_object_capa(env, fo)); + rc = dt_attr_get(env, ofd_object_child(fo), la); if (unlikely(rc)) GOTO(buf_put, rc); rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local); if (unlikely(rc)) GOTO(buf_put, rc); - lprocfs_counter_add(ofd_obd(ofd)->obd_stats, - LPROC_OFD_READ_BYTES, tot_bytes); + + ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes); RETURN(0); buf_put: @@ -96,13 +498,35 @@ unlock: return rc; } +/** + * Prepare buffers for write request processing. + * + * This function converts remote buffers from client to local buffers + * and prepares the latter. If there is recovery in progress and required + * object is missing then it can be re-created before write. + * + * \param[in] env execution environment + * \param[in] exp OBD export of client + * \param[in] ofd OFD device + * \param[in] fid FID of object + * \param[in] la object attributes + * \param[in] oa OBDO structure from client + * \param[in] objcount always 1 + * \param[in] obj object data + * \param[in] rnb remote buffers + * \param[in] nr_local number of local buffers + * \param[in] lnb local buffers + * \param[in] jobid job ID name + * + * \retval 0 on successful prepare + * \retval negative value on error + */ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, - struct ofd_device *ofd, struct lu_fid *fid, + struct ofd_device *ofd, const struct lu_fid *fid, struct lu_attr *la, struct obdo *oa, int objcount, struct obd_ioobj *obj, struct niobuf_remote *rnb, int *nr_local, - struct niobuf_local *lnb, - struct obd_trans_info *oti) + struct niobuf_local *lnb, char *jobid) { struct ofd_object *fo; int i, j, k, rc = 0, tot_bytes = 0; @@ -112,132 +536,215 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, LASSERT(objcount == 1); if (unlikely(exp->exp_obd->obd_recovering)) { - struct ofd_thread_info *info = ofd_info(env); - - /* copied from ofd_precreate_object */ - /* XXX this should be consolidated to use the same code - * instead of a copy, due to the ongoing risk of bugs. */ - memset(&info->fti_attr, 0, sizeof(info->fti_attr)); - info->fti_attr.la_valid = LA_TYPE | LA_MODE; - info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666; - info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME; - /* Initialize a/c/m time so any client timestamp will always - * be newer and update the inode. ctime = 0 is also handled - * specially in osd_inode_setattr(). See LU-221, LU-1042 */ - info->fti_attr.la_atime = 0; - info->fti_attr.la_mtime = 0; - info->fti_attr.la_ctime = 0; - - fo = ofd_object_find_or_create(env, ofd, fid, &info->fti_attr); - } else { - fo = ofd_object_find(env, ofd, fid); + u64 seq = fid_seq(fid); + u64 oid = fid_oid(fid); + struct ofd_seq *oseq; + + oseq = ofd_seq_load(env, ofd, seq); + if (IS_ERR(oseq)) { + CERROR("%s: Can't find FID Sequence "LPX64": rc = %d\n", + ofd_name(ofd), seq, (int)PTR_ERR(oseq)); + GOTO(out, rc = -EINVAL); + } + + if (oid > ofd_seq_last_oid(oseq)) { + int sync = 0; + int diff; + + mutex_lock(&oseq->os_create_lock); + diff = oid - ofd_seq_last_oid(oseq); + + /* Do sync create if the seq is about to used up */ + if (fid_seq_is_idif(seq) || fid_seq_is_mdt0(seq)) { + if (unlikely(oid >= IDIF_MAX_OID - 1)) + sync = 1; + } else if (fid_seq_is_norm(seq)) { + if (unlikely(oid >= + LUSTRE_DATA_SEQ_MAX_WIDTH - 1)) + sync = 1; + } else { + CERROR("%s : invalid o_seq "DOSTID"\n", + ofd_name(ofd), POSTID(&oa->o_oi)); + mutex_unlock(&oseq->os_create_lock); + ofd_seq_put(env, oseq); + GOTO(out, rc = -EINVAL); + } + + while (diff > 0) { + u64 next_id = ofd_seq_last_oid(oseq) + 1; + int count = ofd_precreate_batch(ofd, diff); + + rc = ofd_precreate_objects(env, ofd, next_id, + oseq, count, sync); + if (rc < 0) { + mutex_unlock(&oseq->os_create_lock); + ofd_seq_put(env, oseq); + GOTO(out, rc); + } + + diff -= rc; + } + + mutex_unlock(&oseq->os_create_lock); + } + + ofd_seq_put(env, oseq); } + fo = ofd_object_find(env, ofd, fid); if (IS_ERR(fo)) GOTO(out, rc = PTR_ERR(fo)); LASSERT(fo != NULL); ofd_read_lock(env, fo); if (!ofd_object_exists(fo)) { - CERROR("%s: BRW to missing obj "LPU64"/"LPU64"\n", - exp->exp_obd->obd_name, obj->ioo_id, obj->ioo_seq); + CERROR("%s: BRW to missing obj "DOSTID"\n", + exp->exp_obd->obd_name, POSTID(&obj->ioo_oid)); ofd_read_unlock(env, fo); ofd_object_put(env, fo); GOTO(out, rc = -ENOENT); } - /* Always sync if syncjournal parameter is set */ - oti->oti_sync_write = ofd->ofd_syncjournal; + if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) { + rc = ofd_verify_ff(env, fo, oa); + if (rc != 0) { + ofd_read_unlock(env, fo); + ofd_object_put(env, fo); + GOTO(out, rc); + } + } /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some * space back if possible */ ofd_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt); /* parse remote buffers to local buffers and prepare the latter */ + *nr_local = 0; for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) { rc = dt_bufs_get(env, ofd_object_child(fo), - rnb + i, lnb + j, 1, - ofd_object_capa(env, fo)); - LASSERT(rc > 0); + rnb + i, lnb + j, 1); + if (unlikely(rc < 0)) + GOTO(err, rc); LASSERT(rc <= PTLRPC_MAX_BRW_PAGES); /* correct index for local buffers to continue with */ for (k = 0; k < rc; k++) { lnb[j+k].lnb_flags = rnb[i].rnb_flags; if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED)) lnb[j+k].lnb_rc = -ENOSPC; - if (!(rnb[i].rnb_flags & OBD_BRW_ASYNC)) - oti->oti_sync_write = 1; + + /* remote client can't break through quota */ + if (exp_connect_rmtclient(exp)) + lnb[j+k].lnb_flags &= ~OBD_BRW_NOQUOTA; } j += rc; + *nr_local += rc; LASSERT(j <= PTLRPC_MAX_BRW_PAGES); tot_bytes += rnb[i].rnb_len; } - *nr_local = j; LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES); - lprocfs_counter_add(ofd_obd(ofd)->obd_stats, - LPROC_OFD_WRITE_BYTES, tot_bytes); rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local); - if (unlikely(rc != 0)) { - dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); - ofd_read_unlock(env, fo); - /* ofd_grant_prepare_write() was called, so we must commit */ - ofd_grant_commit(env, exp, rc); - } + if (unlikely(rc != 0)) + GOTO(err, rc); - RETURN(rc); + ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes); + RETURN(0); +err: + dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); + ofd_read_unlock(env, fo); + ofd_object_put(env, fo); + /* ofd_grant_prepare_write() was called, so we must commit */ + ofd_grant_commit(env, exp, rc); out: /* let's still process incoming grant information packed in the oa, * but without enforcing grant since we won't proceed with the write. * Just like a read request actually. */ ofd_grant_prepare_read(env, exp, oa); - RETURN(rc); + return rc; } -int ofd_preprw(const struct lu_env* env, int cmd, struct obd_export *exp, +/** + * Prepare bulk IO requests for processing. + * + * This function does initial checks of IO and calls corresponding + * functions for read/write processing. + * + * \param[in] env execution environment + * \param[in] cmd IO type (read/write) + * \param[in] exp OBD export of client + * \param[in] oa OBDO structure from request + * \param[in] objcount always 1 + * \param[in] obj object data + * \param[in] rnb remote buffers + * \param[in] nr_local number of local buffers + * \param[in] lnb local buffers + * + * \retval 0 on successful prepare + * \retval negative value on error + */ +int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp, struct obdo *oa, int objcount, struct obd_ioobj *obj, struct niobuf_remote *rnb, int *nr_local, - struct niobuf_local *lnb, struct obd_trans_info *oti, - struct lustre_capa *capa) + struct niobuf_local *lnb) { + struct tgt_session_info *tsi = tgt_ses_info(env); struct ofd_device *ofd = ofd_exp(exp); struct ofd_thread_info *info; + char *jobid; + const struct lu_fid *fid = &oa->o_oi.oi_fid; int rc = 0; - if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT) && - ofd->ofd_destroys_in_progress == 0) { - /* don't fail lookups for orphan recovery, it causes - * later LBUGs when objects still exist during precreate */ - CDEBUG(D_INFO, "*** obd_fail_loc=%x ***\n",OBD_FAIL_OST_ENOENT); - RETURN(-ENOENT); + if (*nr_local > PTLRPC_MAX_BRW_PAGES) { + CERROR("%s: bulk has too many pages %d, which exceeds the" + "maximum pages per RPC of %d\n", + exp->exp_obd->obd_name, *nr_local, PTLRPC_MAX_BRW_PAGES); + RETURN(-EPROTO); + } + + if (tgt_ses_req(tsi) == NULL) { /* echo client case */ + info = ofd_info_init(env, exp); + jobid = NULL; + } else { + info = tsi2ofd_info(tsi); + jobid = tsi->tsi_jobid; } - info = ofd_info_init(env, exp); + LASSERT(oa != NULL); + + if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT)) { + struct ofd_seq *oseq; + + oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi)); + if (IS_ERR(oseq)) { + CERROR("%s: Can not find seq for "DOSTID + ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi), + PTR_ERR(oseq)); + RETURN(-EINVAL); + } + + if (oseq->os_destroys_in_progress == 0) { + /* don't fail lookups for orphan recovery, it causes + * later LBUGs when objects still exist during + * precreate */ + ofd_seq_put(env, oseq); + RETURN(-ENOENT); + } + ofd_seq_put(env, oseq); + } LASSERT(objcount == 1); LASSERT(obj->ioo_bufcnt > 0); - fid_ostid_unpack(&info->fti_fid, &oa->o_oi, 0); if (cmd == OBD_BRW_WRITE) { - rc = ofd_auth_capa(exp, &info->fti_fid, oa->o_seq, - capa, CAPA_OPC_OSS_WRITE); - if (rc == 0) { - LASSERT(oa != NULL); - la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR); - rc = ofd_preprw_write(env, exp, ofd, &info->fti_fid, - &info->fti_attr, oa, objcount, - obj, rnb, nr_local, lnb, oti); - } + la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR); + rc = ofd_preprw_write(env, exp, ofd, fid, &info->fti_attr, oa, + objcount, obj, rnb, nr_local, lnb, jobid); } else if (cmd == OBD_BRW_READ) { - rc = ofd_auth_capa(exp, &info->fti_fid, oa->o_seq, - capa, CAPA_OPC_OSS_READ); - if (rc == 0) { - ofd_grant_prepare_read(env, exp, oa); - rc = ofd_preprw_read(env, ofd, &info->fti_fid, - &info->fti_attr, obj->ioo_bufcnt, - rnb, nr_local, lnb); - obdo_from_la(oa, &info->fti_attr, LA_ATIME); - } + ofd_grant_prepare_read(env, exp, oa); + rc = ofd_preprw_read(env, exp, ofd, fid, &info->fti_attr, oa, + obj->ioo_bufcnt, rnb, nr_local, lnb, + jobid); + obdo_from_la(oa, &info->fti_attr, LA_ATIME); } else { CERROR("%s: wrong cmd %d received!\n", exp->exp_obd->obd_name, cmd); @@ -246,9 +753,24 @@ int ofd_preprw(const struct lu_env* env, int cmd, struct obd_export *exp, RETURN(rc); } +/** + * Drop reference on local buffers for read bulk IO. + * + * This will free all local buffers use by this read request. + * + * \param[in] env execution environment + * \param[in] ofd OFD device + * \param[in] fid object FID + * \param[in] objcount always 1 + * \param[in] niocount number of local buffers + * \param[in] lnb local buffers + * + * \retval 0 on successful execution + * \retval negative value on error + */ static int ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd, - struct lu_fid *fid, int objcount, int niocount, + const struct lu_fid *fid, int objcount, int niocount, struct niobuf_local *lnb) { struct ofd_object *fo; @@ -272,6 +794,21 @@ ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd, RETURN(0); } +/** + * Set attributes of object during write bulk IO processing. + * + * Change object attributes and write parent FID into extended + * attributes when needed. + * + * \param[in] env execution environment + * \param[in] ofd OFD device + * \param[in] ofd_obj OFD object + * \param[in] la object attributes + * \param[in] ff parent FID + * + * \retval 0 on successful attributes update + * \retval negative value on error + */ static int ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, struct ofd_object *ofd_obj, struct lu_attr *la, @@ -298,7 +835,7 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, GOTO(out, rc); if (ff != NULL) { - rc = ofd_object_ff_check(env, ofd_obj); + rc = ofd_object_ff_load(env, ofd_obj); if (rc == -ENODATA) ff_needed = 1; else if (rc < 0) @@ -320,6 +857,12 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, } if (ff_needed) { + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1)) + ff->ff_parent.f_oid = cpu_to_le32(1UL << 31); + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2)) + ff->ff_parent.f_oid = + cpu_to_le32(le32_to_cpu(ff->ff_parent.f_oid) - 1); + info->fti_buf.lb_buf = ff; info->fti_buf.lb_len = sizeof(*ff); rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf, @@ -336,21 +879,32 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, /* set uid/gid */ if (la->la_valid) { - rc = dt_attr_set(env, dt_obj, la, th, - ofd_object_capa(env, ofd_obj)); + rc = dt_attr_set(env, dt_obj, la, th); if (rc) GOTO(out_tx, rc); } /* set filter fid EA */ if (ff_needed) { - rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, - 0, th, BYPASS_CAPA); - if (rc) + if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID)) GOTO(out_tx, rc); + + rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, + 0, th); + if (rc == 0) { + ofd_obj->ofo_pfid.f_seq = le64_to_cpu(ff->ff_parent.f_seq); + ofd_obj->ofo_pfid.f_oid = le32_to_cpu(ff->ff_parent.f_oid); + /* Currently, the filter_fid::ff_parent::f_ver is not + * the real parent MDT-object's FID::f_ver, instead it + * is the OST-object index in its parent MDT-object's + * layout EA. */ + ofd_obj->ofo_pfid.f_stripe_idx = + le32_to_cpu(ff->ff_parent.f_stripe_idx); + } } - EXIT; + GOTO(out_tx, rc); + out_tx: dt_trans_stop(env, ofd->ofd_osd, th); out: @@ -358,12 +912,103 @@ out: return rc; } +struct ofd_soft_sync_callback { + struct dt_txn_commit_cb ossc_cb; + struct obd_export *ossc_exp; +}; + +/** + * Callback function for "soft sync" update. + * + * Reset fed_soft_sync_count upon committing the "soft_sync" update. + * See ofd_soft_sync_cb_add() below for more details on soft sync. + * + * \param[in] env execution environment + * \param[in] th transaction handle + * \param[in] cb callback data + * \param[in] err error code + */ +static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th, + struct dt_txn_commit_cb *cb, int err) +{ + struct ofd_soft_sync_callback *ossc; + + ossc = container_of(cb, struct ofd_soft_sync_callback, ossc_cb); + + CDEBUG(D_INODE, "export %p soft sync count is reset\n", ossc->ossc_exp); + atomic_set(&ossc->ossc_exp->exp_filter_data.fed_soft_sync_count, 0); + + class_export_cb_put(ossc->ossc_exp); + OBD_FREE_PTR(ossc); +} + +/** + * Add callback for "soft sync" processing. + * + * The "soft sync" mechanism does asynchronous commit when OBD_BRW_SOFT_SYNC + * flag is set in client buffers. The intention is for this operation to + * commit pages belonging to a client which has "too many" outstanding + * unstable pages in its cache. See LU-2139 for details. + * + * This function adds callback to be called when commit is done. + * + * \param[in] th transaction handle + * \param[in] exp OBD export of client + * + * \retval 0 on successful callback adding + * \retval negative value on error + */ +static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp) +{ + struct ofd_soft_sync_callback *ossc; + struct dt_txn_commit_cb *dcb; + int rc; + + OBD_ALLOC_PTR(ossc); + if (ossc == NULL) + return -ENOMEM; + + ossc->ossc_exp = class_export_cb_get(exp); + + dcb = &ossc->ossc_cb; + dcb->dcb_func = ofd_cb_soft_sync; + INIT_LIST_HEAD(&dcb->dcb_linkage); + strlcpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name)); + + rc = dt_trans_cb_add(th, dcb); + if (rc) { + class_export_cb_put(exp); + OBD_FREE_PTR(ossc); + } + + return rc; +} + +/** + * Commit bulk IO buffers to the storage. + * + * This function finalizes write IO processing by writing data to the disk. + * That write can be synchronous or asynchronous depending on buffers flags. + * + * \param[in] env execution environment + * \param[in] exp OBD export of client + * \param[in] ofd OFD device + * \param[in] fid FID of object + * \param[in] la object attributes + * \param[in] ff parent FID of object + * \param[in] objcount always 1 + * \param[in] niocount number of local buffers + * \param[in] lnb local buffers + * \param[in] old_rc result of processing at this point + * + * \retval 0 on successful commit + * \retval negative value on error + */ static int -ofd_commitrw_write(const struct lu_env *env, struct ofd_device *ofd, - struct lu_fid *fid, struct lu_attr *la, - struct filter_fid *ff, int objcount, - int niocount, struct niobuf_local *lnb, - struct obd_trans_info *oti, int old_rc) +ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp, + struct ofd_device *ofd, const struct lu_fid *fid, + struct lu_attr *la, struct filter_fid *ff, int objcount, + int niocount, struct niobuf_local *lnb, int old_rc) { struct ofd_thread_info *info = ofd_info(env); struct ofd_object *fo; @@ -371,6 +1016,10 @@ ofd_commitrw_write(const struct lu_env *env, struct ofd_device *ofd, struct thandle *th; int rc = 0; int retries = 0; + int i; + struct filter_export_data *fed = &exp->exp_filter_data; + bool soft_sync = false; + bool cb_registered = false; ENTRY; @@ -403,7 +1052,20 @@ retry: if (IS_ERR(th)) GOTO(out, rc = PTR_ERR(th)); - th->th_sync |= oti->oti_sync_write; + th->th_sync |= ofd->ofd_syncjournal; + if (th->th_sync == 0) { + for (i = 0; i < niocount; i++) { + if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) { + th->th_sync = 1; + break; + } + if (lnb[i].lnb_flags & OBD_BRW_SOFT_SYNC) + soft_sync = true; + } + } + + if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET)) + GOTO(out_stop, rc = -EINPROGRESS); rc = dt_declare_write_commit(env, o, lnb, niocount, th); if (rc) @@ -425,13 +1087,13 @@ retry: GOTO(out_stop, rc); if (la->la_valid) { - rc = dt_attr_set(env, o, la, th, ofd_object_capa(env, fo)); + rc = dt_attr_set(env, o, la, th); if (rc) GOTO(out_stop, rc); } /* get attr to return */ - dt_attr_get(env, o, la, ofd_object_capa(env, fo)); + rc = dt_attr_get(env, o, la); out_stop: /* Force commit to make the just-deleted blocks @@ -439,6 +1101,12 @@ out_stop: if (rc == -ENOSPC) th->th_sync = 1; + /* do this before trans stop in case commit has finished */ + if (!th->th_sync && soft_sync && !cb_registered) { + ofd_soft_sync_cb_add(th, exp); + cb_registered = true; + } + ofd_trans_stop(env, ofd, th, rc); if (rc == -ENOSPC && retries++ < 3) { CDEBUG(D_INODE, "retry after force commit, retries:%d\n", @@ -446,6 +1114,13 @@ out_stop: goto retry; } + if (!soft_sync) + /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */ + atomic_set(&fed->fed_soft_sync_count, 0); + else if (atomic_inc_return(&fed->fed_soft_sync_count) == + ofd->ofd_soft_sync_limit) + dt_commit_async(env, ofd->ofd_osd); + out: dt_bufs_put(env, o, lnb, niocount); ofd_read_unlock(env, fo); @@ -456,25 +1131,43 @@ out: RETURN(rc); } +/** + * Commit bulk IO to the storage. + * + * This is companion function to the ofd_preprw(). It finishes bulk IO + * request processing by committing buffers to the storage (WRITE) and/or + * freeing those buffers (read/write). See ofd_commitrw_read() and + * ofd_commitrw_write() for details about each type of IO. + * + * \param[in] env execution environment + * \param[in] cmd IO type (READ/WRITE) + * \param[in] exp OBD export of client + * \param[in] oa OBDO structure from client + * \param[in] objcount always 1 + * \param[in] obj object data + * \param[in] rnb remote buffers + * \param[in] npages number of local buffers + * \param[in] lnb local buffers + * \param[in] old_rc result of processing at this point + * + * \retval 0 on successful commit + * \retval negative value on error + */ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, struct obdo *oa, int objcount, struct obd_ioobj *obj, struct niobuf_remote *rnb, int npages, - struct niobuf_local *lnb, struct obd_trans_info *oti, - int old_rc) + struct niobuf_local *lnb, int old_rc) { - struct ofd_thread_info *info; + struct ofd_thread_info *info = ofd_info(env); struct ofd_mod_data *fmd; __u64 valid; struct ofd_device *ofd = ofd_exp(exp); struct filter_fid *ff = NULL; + const struct lu_fid *fid = &oa->o_oi.oi_fid; int rc = 0; - info = ofd_info(env); - ofd_oti2info(info, oti); - LASSERT(npages > 0); - fid_ostid_unpack(&info->fti_fid, &oa->o_oi, 0); if (cmd == OBD_BRW_WRITE) { /* Don't update timestamps if this write is older than a * setattr which modifies the timestamps. b=10150 */ @@ -484,7 +1177,7 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, * doesn't already exist so we can store the reservation handle * there. */ valid = OBD_MD_FLUID | OBD_MD_FLGID; - fmd = ofd_fmd_find(exp, &info->fti_fid); + fmd = ofd_fmd_find(exp, fid); if (!fmd || fmd->fmd_mactime_xid < info->fti_xid) valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME; @@ -496,21 +1189,35 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, ofd_prepare_fidea(ff, oa); } - rc = ofd_commitrw_write(env, ofd, &info->fti_fid, - &info->fti_attr, ff, objcount, npages, - lnb, oti, old_rc); + rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr, + ff, objcount, npages, lnb, old_rc); if (rc == 0) obdo_from_la(oa, &info->fti_attr, OFD_VALID_FLAGS | LA_GID | LA_UID); else obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID); - if (ofd_grant_prohibit(exp, ofd)) - /* Trick to prevent clients from waiting for bulk write - * in flight since they won't get any grant in the reply - * anyway so they had better firing the sync write RPC - * straight away */ + /* don't report overquota flag if we failed before reaching + * commit */ + if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) { + /* return the overquota flags to client */ + if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) { + if (oa->o_valid & OBD_MD_FLFLAGS) + oa->o_flags |= OBD_FL_NO_USRQUOTA; + else + oa->o_flags = OBD_FL_NO_USRQUOTA; + } + + if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) { + if (oa->o_valid & OBD_MD_FLFLAGS) + oa->o_flags |= OBD_FL_NO_GRPQUOTA; + else + oa->o_flags = OBD_FL_NO_GRPQUOTA; + } + + oa->o_valid |= OBD_MD_FLFLAGS; oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA; + } } else if (cmd == OBD_BRW_READ) { struct ldlm_namespace *ns = ofd->ofd_namespace; @@ -520,16 +1227,16 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) { struct ldlm_resource *rs = NULL; - ofd_build_resid(&info->fti_fid, &info->fti_resid); + ost_fid_build_resid(fid, &info->fti_resid); rs = ldlm_resource_get(ns, NULL, &info->fti_resid, LDLM_EXTENT, 0); - if (rs != NULL) { - ns->ns_lvbo->lvbo_update(rs, NULL, 1); + if (!IS_ERR(rs)) { + ldlm_res_lvbo_update(rs, NULL, 1); ldlm_resource_putref(rs); } } - rc = ofd_commitrw_read(env, ofd, &info->fti_fid, objcount, - npages, lnb); + rc = ofd_commitrw_read(env, ofd, fid, objcount, + npages, lnb); if (old_rc) rc = old_rc; } else { @@ -537,6 +1244,5 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, rc = -EPROTO; } - ofd_info2oti(info, oti); RETURN(rc); }