X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Fofd%2Fofd_io.c;h=5e7801f07d9897a58da7a2ce627692668f99e3e4;hb=HEAD;hp=3703b4a04196510db68847b04fb4cec55dcec772;hpb=5999c0b881e8d9f2b431c842bdd1717115a82b7c;p=fs%2Flustre-release.git diff --git a/lustre/ofd/ofd_io.c b/lustre/ofd/ofd_io.c index 3703b4a..54cfe02 100644 --- a/lustre/ofd/ofd_io.c +++ b/lustre/ofd/ofd_io.c @@ -23,11 +23,10 @@ * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. * - * Copyright (c) 2012, 2016, Intel Corporation. + * Copyright (c) 2012, 2017, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ - * Lustre is a trademark of Sun Microsystems, Inc. * * lustre/ofd/ofd_io.c * @@ -147,6 +146,17 @@ static void ofd_inconsistency_verify_one(const struct lu_env *env, OBD_FREE_PTR(oii); } +struct oivm_args { + struct ofd_device *od_ofd; + struct lu_env od_env; + struct lfsck_req_local od_lrl; + struct completion *od_started; +}; + +#ifndef TASK_IDLE +#define TASK_IDLE TASK_INTERRUPTIBLE +#endif + /** * Verification thread to check parent FID consistency. * @@ -158,87 +168,64 @@ static void ofd_inconsistency_verify_one(const struct lu_env *env, * \retval 0 on successful thread termination * \retval negative value if thread can't start */ -static int ofd_inconsistency_verification_main(void *args) +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 oivm_args *args = _args; + struct lu_env *env = &args->od_env; + struct ofd_device *ofd = args->od_ofd; struct ofd_inconsistency_item *oii; - struct lfsck_req_local *lrl = NULL; - struct l_wait_info lwi = { 0 }; - int rc; + struct lfsck_req_local *lrl = &args->od_lrl; ENTRY; - rc = lu_env_init(&env, LCT_DT_THREAD); - spin_lock(&ofd->ofd_inconsistency_lock); - thread_set_flags(thread, rc ? SVC_STOPPED : SVC_RUNNING); - wake_up_all(&thread->t_ctl_waitq); - spin_unlock(&ofd->ofd_inconsistency_lock); - if (rc) - RETURN(rc); - - OBD_ALLOC_PTR(lrl); - if (unlikely(!lrl)) - GOTO(out_unlocked, rc = -ENOMEM); - lrl->lrl_event = LEL_PAIRS_VERIFY_LOCAL; lrl->lrl_active = LFSCK_TYPE_LAYOUT; + complete(args->od_started); spin_lock(&ofd->ofd_inconsistency_lock); - while (1) { - if (unlikely(!thread_is_running(thread))) - break; + while (({set_current_state(TASK_IDLE); + !kthread_should_stop(); })) { while (!list_empty(&ofd->ofd_inconsistency_list)) { - oii = list_entry(ofd->ofd_inconsistency_list.next, - struct ofd_inconsistency_item, - oii_list); + __set_current_state(TASK_RUNNING); + oii = list_first_entry(&ofd->ofd_inconsistency_list, + 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, lrl); + ofd_inconsistency_verify_one(env, ofd, oii, lrl); 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); + schedule(); spin_lock(&ofd->ofd_inconsistency_lock); } + __set_current_state(TASK_RUNNING); 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); + oii = list_first_entry(&ofd->ofd_inconsistency_list, + 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); + ofd_write_lock(env, fo); fo->ofo_pfid_checking = 0; - ofd_write_unlock(&env, fo); + ofd_write_unlock(env, fo); - ofd_object_put(&env, fo); + ofd_object_put(env, fo); OBD_FREE_PTR(oii); spin_lock(&ofd->ofd_inconsistency_lock); } - OBD_FREE_PTR(lrl); - - 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; + lu_env_fini(&args->od_env); + OBD_FREE_PTR(args); + return 0; } /** @@ -253,32 +240,50 @@ out: */ 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; + struct oivm_args *args; + DECLARE_COMPLETION_ONSTACK(started); int rc; - spin_lock(&ofd->ofd_inconsistency_lock); - if (unlikely(thread_is_running(thread))) { - spin_unlock(&ofd->ofd_inconsistency_lock); - + if (ofd->ofd_inconsistency_task) return -EALREADY; + + OBD_ALLOC_PTR(args); + if (!args) + return -ENOMEM; + rc = lu_env_init(&args->od_env, LCT_DT_THREAD); + if (rc) { + OBD_FREE_PTR(args); + return rc; } - thread_set_flags(thread, 0); - spin_unlock(&ofd->ofd_inconsistency_lock); - task = kthread_run(ofd_inconsistency_verification_main, ofd, - "inconsistency_verification"); + args->od_ofd = ofd; + args->od_started = &started; + task = kthread_create(ofd_inconsistency_verification_main, args, + "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); + spin_lock(&ofd->ofd_inconsistency_lock); + if (ofd->ofd_inconsistency_task) + rc = -EALREADY; + else + ofd->ofd_inconsistency_task = task; + spin_unlock(&ofd->ofd_inconsistency_lock); + + if (rc) + kthread_stop(task); + else { + wake_up_process(task); + wait_for_completion(&started); + } + } + if (rc) { + lu_env_fini(&args->od_env); + OBD_FREE_PTR(args); } return rc; @@ -294,22 +299,16 @@ int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd) */ int ofd_stop_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; spin_lock(&ofd->ofd_inconsistency_lock); - if (thread_is_init(thread) || thread_is_stopped(thread)) { - spin_unlock(&ofd->ofd_inconsistency_lock); + task = ofd->ofd_inconsistency_task; + ofd->ofd_inconsistency_task = NULL; + spin_unlock(&ofd->ofd_inconsistency_lock); + if (!task) 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); + kthread_stop(task); return 0; } @@ -357,9 +356,9 @@ static void ofd_add_inconsistency_item(const struct lu_env *env, if (list_empty(&ofd->ofd_inconsistency_list)) wakeup = true; list_add_tail(&oii->oii_list, &ofd->ofd_inconsistency_list); + if (wakeup && ofd->ofd_inconsistency_task) + wake_up_process(ofd->ofd_inconsistency_task); 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 @@ -440,47 +439,131 @@ int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo, int ofd_verify_layout_version(const struct lu_env *env, struct ofd_object *fo, const struct obdo *oa) { - __u32 layout_version; int rc; ENTRY; - if (unlikely(OBD_FAIL_CHECK(OBD_FAIL_OST_SKIP_LV_CHECK))) + if (unlikely(CFS_FAIL_CHECK(OBD_FAIL_OST_SKIP_LV_CHECK))) GOTO(out, rc = 0); rc = ofd_object_ff_load(env, fo); if (rc < 0) { if (rc == -ENODATA) - rc = -EINPROGRESS; + rc = 0; GOTO(out, rc); } - layout_version = fo->ofo_ff.ff_layout_version; - if (oa->o_layout_version >= layout_version && - oa->o_layout_version <= layout_version + fo->ofo_ff.ff_range) - GOTO(out, rc = 0); - - /* normal traffic, decide if to return ESTALE or EINPROGRESS */ - layout_version &= ~LU_LAYOUT_RESYNC; - - /* this update is not legitimate */ - if ((oa->o_layout_version & ~LU_LAYOUT_RESYNC) <= layout_version) + /** + * this update is not legitimate, whose layout version is older than + * that on the disk. + */ + if (ofd_layout_version_less(oa->o_layout_version, + fo->ofo_ff.ff_layout_version)) GOTO(out, rc = -ESTALE); - /* layout version may not be transmitted yet */ - if ((oa->o_layout_version & ~LU_LAYOUT_RESYNC) > layout_version) - GOTO(out, rc = -EINPROGRESS); - - EXIT; - out: - CDEBUG(D_INODE, DFID " verify layout version: %u vs. %u/%u, rc: %d\n", + CDEBUG(D_INODE, DFID " verify layout version: %u vs. %u/%u: rc = %d\n", PFID(lu_object_fid(&fo->ofo_obj.do_lu)), oa->o_layout_version, fo->ofo_ff.ff_layout_version, fo->ofo_ff.ff_range, rc); - return rc; + RETURN(rc); } +/* + * Handle multiple attrs at once: + * + * Lazy ATIME update to refresh atime every ofd_atime_diff + * seconds so that external scanning tool can see it actual + * within that period and be able to identify accessed files + * + * Update enc flag on OST object + * OSD layer will find out if this is really necessary. + */ +static void ofd_handle_attrs(const struct lu_env *env, struct ofd_device *ofd, + struct ofd_object *fo, struct obdo *oa) +{ + bool need_atime = (oa->o_valid & OBD_MD_FLATIME); + bool need_encfl = (oa->o_valid & OBD_MD_FLFLAGS && + oa->o_flags & LUSTRE_ENCRYPT_FL); + struct lu_attr *la; + struct dt_object *o; + struct thandle *th; + int rc; + + if (need_atime && ofd->ofd_atime_diff == 0) + need_atime = false; + + if (need_encfl && CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) + need_encfl = false; + + if (!need_atime && !need_encfl) + return; + + la = &ofd_info(env)->fti_attr2; + o = ofd_object_child(fo); + + if (need_atime) { + if (unlikely(fo->ofo_atime_ondisk == 0)) { + rc = dt_attr_get(env, o, la); + if (unlikely(rc)) { + need_atime = false; + GOTO(trans, rc); + } + LASSERT(la->la_valid & LA_ATIME); + if (la->la_atime == 0) + la->la_atime = la->la_mtime; + fo->ofo_atime_ondisk = la->la_atime; + } + if (oa->o_atime - fo->ofo_atime_ondisk < ofd->ofd_atime_diff) { + need_atime = false; + GOTO(trans, rc = 0); + } + + /* atime hasn't been updated too long, update it */ + fo->ofo_atime_ondisk = oa->o_atime; + } + +trans: + if (!need_atime && !need_encfl) + return; + + th = ofd_trans_create(env, ofd); + if (IS_ERR(th)) { + CERROR("%s: cannot create transaction: rc = %d\n", + ofd_name(ofd), (int)PTR_ERR(th)); + return; + } + + la->la_valid = 0; + if (need_atime) { + la->la_valid |= LA_ATIME; + la->la_atime = fo->ofo_atime_ondisk; + } + if (need_encfl) { + la->la_valid |= LA_FLAGS; + la->la_flags = LUSTRE_ENCRYPT_FL; + } + + rc = dt_declare_attr_set(env, o, la, th); + if (rc) + GOTO(out_tx, rc); + + rc = dt_trans_start_local(env, ofd->ofd_osd , th); + if (rc) { + CERROR("%s: cannot start transaction: rc = %d\n", + ofd_name(ofd), rc); + GOTO(out_tx, rc); + } + + ofd_read_lock(env, fo); + if (ofd_object_exists(fo)) + rc = dt_attr_set(env, o, la, th); + ofd_read_unlock(env, fo); + +out_tx: + ofd_trans_stop(env, ofd, th, rc); +} + /** * Prepare buffers for read request processing. * @@ -506,11 +589,13 @@ 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, char *jobid) + struct niobuf_local *lnb) { struct ofd_object *fo; int i, j, rc, tot_bytes = 0; enum dt_bufs_type dbt = DT_BUFS_TYPE_READ; + int maxlnb = *nr_local; + __u64 begin, end; ENTRY; LASSERT(env != NULL); @@ -520,48 +605,73 @@ static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp, RETURN(PTR_ERR(fo)); LASSERT(fo != NULL); - ofd_read_lock(env, fo); - if (!ofd_object_exists(fo)) - GOTO(unlock, rc = -ENOENT); + ofd_info(env)->fti_obj = fo; - 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); - } + ofd_handle_attrs(env, ofd, fo, oa); + + if (!ofd_object_exists(fo)) + GOTO(obj_put, rc = -ENOENT); if (ptlrpc_connection_is_local(exp->exp_connection)) dbt |= DT_BUFS_TYPE_LOCAL; + begin = -1; + end = 0; + for (*nr_local = 0, i = 0, j = 0; i < niocount; i++) { + begin = min_t(__u64, begin, rnb[i].rnb_offset); + end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len); + + if (CFS_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF)) + rnb[i].rnb_len = 100 * 1024 * 1024; + rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i, - lnb + j, dbt); + lnb + j, maxlnb, dbt); 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; + maxlnb -= rc; LASSERT(j <= PTLRPC_MAX_BRW_PAGES); tot_bytes += rnb[i].rnb_len; } - LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES); - rc = dt_attr_get(env, ofd_object_child(fo), la); - if (unlikely(rc)) - GOTO(buf_put, rc); + ofd_read_lock(env, fo); + if (!ofd_object_exists(fo)) + GOTO(unlock, rc = -ENOENT); + 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); + } + + LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES); rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local); if (unlikely(rc)) - GOTO(buf_put, rc); + GOTO(unlock, rc); + ofd_read_unlock(env, fo); + + ofd_access(env, ofd, + &(struct lu_fid) { + .f_seq = oa->o_parent_seq, + .f_oid = oa->o_parent_oid, + .f_ver = oa->o_stripe_idx, + }, + begin, end, + tot_bytes, + niocount, + READ); - ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, tot_bytes); RETURN(0); -buf_put: - dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); unlock: ofd_read_unlock(env, fo); +buf_put: + dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); +obj_put: ofd_object_put(env, fo); return rc; } @@ -594,19 +704,21 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, struct lu_attr *la, struct obdo *oa, int objcount, struct obd_ioobj *obj, struct niobuf_remote *rnb, int *nr_local, - struct niobuf_local *lnb, char *jobid) + struct niobuf_local *lnb) { struct ofd_object *fo; int i, j, k, rc = 0, tot_bytes = 0; enum dt_bufs_type dbt = DT_BUFS_TYPE_WRITE; + int maxlnb = *nr_local; + __u64 begin, end; ENTRY; LASSERT(env != NULL); LASSERT(objcount == 1); if (unlikely(exp->exp_obd->obd_recovering)) { - u64 seq = fid_seq(fid); - u64 oid = fid_oid(fid); + u64 seq = ostid_seq(&oa->o_oi); + u64 oid = ostid_id(&oa->o_oi); struct ofd_seq *oseq; oseq = ofd_seq_load(env, ofd, seq); @@ -624,19 +736,11 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, 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)); + sync = ofd_seq_is_exhausted(ofd, oa); + if (sync < 0) { mutex_unlock(&oseq->os_create_lock); ofd_seq_put(env, oseq); - GOTO(out, rc = -EINVAL); + GOTO(out, rc = sync); } while (diff > 0) { @@ -644,7 +748,8 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, int count = ofd_precreate_batch(ofd, diff); rc = ofd_precreate_objects(env, ofd, next_id, - oseq, count, sync); + oseq, count, sync, + false); if (rc < 0) { mutex_unlock(&oseq->os_create_lock); ofd_seq_put(env, oseq); @@ -660,54 +765,43 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, ofd_seq_put(env, oseq); } + /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some + * space back if possible, we have to do this outside of the lock as + * grant preparation may need to sync whole fs thus wait for all the + * transactions to complete. */ + tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt); + 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); + ofd_info(env)->fti_obj = fo; + if (!ofd_object_exists(fo)) { 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); } - 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); - } - } - - /* need to verify layout version */ - if (oa->o_valid & OBD_MD_LAYOUT_VERSION) { - rc = ofd_verify_layout_version(env, fo, oa); - if (rc) { - ofd_read_unlock(env, fo); - ofd_object_put(env, fo); - GOTO(out, rc); - } - - oa->o_valid &= ~OBD_MD_LAYOUT_VERSION; - } - - /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some - * space back if possible */ - tgt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt); - if (ptlrpc_connection_is_local(exp->exp_connection)) dbt |= DT_BUFS_TYPE_LOCAL; + begin = -1; + end = 0; + /* parse remote buffers to local buffers and prepare the latter */ for (*nr_local = 0, i = 0, j = 0; i < obj->ioo_bufcnt; i++) { + begin = min_t(__u64, begin, rnb[i].rnb_offset); + end = max_t(__u64, end, rnb[i].rnb_offset + rnb[i].rnb_len); + + if (CFS_FAIL_CHECK(OBD_FAIL_OST_2BIG_NIOBUF)) + rnb[i].rnb_len += PAGE_SIZE; rc = dt_bufs_get(env, ofd_object_child(fo), - rnb + i, lnb + j, dbt); + rnb + i, lnb + j, maxlnb, dbt); if (unlikely(rc < 0)) - GOTO(err, rc); + GOTO(err_nolock, rc); LASSERT(rc <= PTLRPC_MAX_BRW_PAGES); /* correct index for local buffers to continue with */ for (k = 0; k < rc; k++) { @@ -718,20 +812,55 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, } j += rc; *nr_local += rc; + maxlnb -= rc; LASSERT(j <= PTLRPC_MAX_BRW_PAGES); tot_bytes += rnb[i].rnb_len; } LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES); + ofd_read_lock(env, fo); + if (!ofd_object_exists(fo)) { + CERROR("%s: BRW to missing obj "DOSTID": rc = -ENOENT\n", + exp->exp_obd->obd_name, POSTID(&obj->ioo_oid)); + GOTO(err, rc = -ENOENT); + } + + if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) { + rc = ofd_verify_ff(env, fo, oa); + if (rc != 0) + GOTO(err, rc); + } + + /* need to verify layout version */ + if (oa->o_valid & OBD_MD_LAYOUT_VERSION) { + rc = ofd_verify_layout_version(env, fo, oa); + if (rc) + GOTO(err, rc); + } + rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local); if (unlikely(rc != 0)) GOTO(err, rc); - ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, tot_bytes); + ofd_read_unlock(env, fo); + + ofd_access(env, ofd, + &(struct lu_fid) { + .f_seq = oa->o_parent_seq, + .f_oid = oa->o_parent_oid, + .f_ver = oa->o_stripe_idx, + }, + begin, end, + tot_bytes, + obj->ioo_bufcnt, + WRITE); + RETURN(0); + err: - dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); ofd_read_unlock(env, fo); +err_nolock: + dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local); ofd_object_put(env, fo); /* tgt_grant_prepare_write() was called, so we must commit */ tgt_grant_commit(exp, oa->o_grant_used, rc); @@ -770,28 +899,24 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp, 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 (*nr_local > PTLRPC_MAX_BRW_PAGES) { - CERROR("%s: bulk has too many pages %d, which exceeds the" - "maximum pages per RPC of %d\n", + 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; } LASSERT(oa != NULL); - if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) { + if (CFS_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) { struct ofd_seq *oseq; oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi)); @@ -818,13 +943,11 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp, if (cmd == OBD_BRW_WRITE) { 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); + objcount, obj, rnb, nr_local, lnb); } else if (cmd == OBD_BRW_READ) { tgt_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); + obj->ioo_bufcnt, rnb, nr_local, lnb); } else { CERROR("%s: wrong cmd %d received!\n", exp->exp_obd->obd_name, cmd); @@ -859,16 +982,15 @@ ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd, LASSERT(niocount > 0); - fo = ofd_object_find(env, ofd, fid); - if (IS_ERR(fo)) - RETURN(PTR_ERR(fo)); + fo = ofd_info(env)->fti_obj; LASSERT(fo != NULL); - LASSERT(ofd_object_exists(fo)); + /* + * there is no guarantee the object still exists as the client + * behind this RPC could have been evicted allowing concurrent + * OST_DESTROY to remove the object. + */ dt_bufs_put(env, ofd_object_child(fo), lnb, niocount); - ofd_read_unlock(env, fo); - ofd_object_put(env, fo); - /* second put is pair to object_get in ofd_preprw_read */ ofd_object_put(env, fo); RETURN(0); @@ -894,13 +1016,14 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, struct ofd_object *ofd_obj, struct lu_attr *la, struct obdo *oa) { - struct ofd_thread_info *info = ofd_info(env); - struct filter_fid *ff = &info->fti_mds_fid; - __u64 valid = la->la_valid; - struct thandle *th; - struct dt_object *dt_obj; - int fl = 0; - int rc; + struct ofd_thread_info *info = ofd_info(env); + struct filter_fid *ff = &info->fti_mds_fid; + __u64 valid = la->la_valid; + __u32 flags = la->la_flags; + struct thandle *th; + struct dt_object *dt_obj; + int fl = 0; + int rc; ENTRY; @@ -910,16 +1033,18 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, LASSERT(dt_obj != NULL); la->la_valid &= LA_UID | LA_GID | LA_PROJID; + if (oa->o_valid & OBD_MD_FLFLAGS && oa->o_flags & LUSTRE_ENCRYPT_FL && + !CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NO_ENCFLAG)) { + la->la_valid |= LA_FLAGS; + la->la_flags = LUSTRE_ENCRYPT_FL; + } rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */); if (rc != 0) GOTO(out, rc); - fl = ofd_object_ff_update(env, ofd_obj, oa, ff); - if (fl < 0) - GOTO(out, rc = fl); - - if (!la->la_valid && !fl) + if (!la->la_valid && !(oa->o_valid & + (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION))) /* no attributes to set */ GOTO(out, rc = 0); @@ -933,55 +1058,71 @@ ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd, GOTO(out_tx, rc); } - if (fl) { - if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1)) - ff->ff_parent.f_oid = cpu_to_le32(1UL << 31); - else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2)) - le32_add_cpu(&ff->ff_parent.f_oid, -1); - + if (oa->o_valid & (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | + OBD_MD_LAYOUT_VERSION)) { rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th); if (rc) GOTO(out_tx, rc); } - /* We don't need a transno for this operation which will be re-executed * anyway when the OST_WRITE (with a transno assigned) is replayed */ rc = dt_trans_start_local(env, ofd->ofd_osd , th); if (rc) GOTO(out_tx, rc); + ofd_read_lock(env, ofd_obj); + + rc = ofd_attr_handle_id(env, ofd_obj, la, 0 /* !is_setattr */); + if (rc != 0) + GOTO(out_unlock, rc); + + if (!la->la_valid && !(oa->o_valid & + (OBD_MD_FLFID | OBD_MD_FLOSTLAYOUT | OBD_MD_LAYOUT_VERSION))) + /* no attributes to set */ + GOTO(out_unlock, rc = 0); + + + /* set uid/gid/projid */ if (la->la_valid) { rc = dt_attr_set(env, dt_obj, la, th); if (rc) - GOTO(out_tx, rc); + GOTO(out_unlock, rc); } + fl = ofd_object_ff_update(env, ofd_obj, oa, ff); + if (fl <= 0) + GOTO(out_unlock, rc = fl); + /* set filter fid EA. * FIXME: it holds read lock of ofd object to modify the XATTR_NAME_FID * while the write lock should be held. However, it should work because * write RPCs only modify ff_{parent,layout} and those information will * be the same from all the write RPCs. The reason that fl is not used * in dt_xattr_set() is to allow this race. */ - if (fl) { - if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID)) - GOTO(out_tx, rc); - - info->fti_buf.lb_buf = ff; - info->fti_buf.lb_len = sizeof(*ff); - rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, - 0, th); - if (rc == 0) - filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff)); - } - - GOTO(out_tx, rc); - + if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID)) + GOTO(out_unlock, rc); + if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1)) + ff->ff_parent.f_oid = cpu_to_le32(1UL << 31); + else if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2)) + le32_add_cpu(&ff->ff_parent.f_oid, -1); + + info->fti_buf.lb_buf = ff; + info->fti_buf.lb_len = sizeof(*ff); + rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID, 0, th); + if (rc == 0) + filter_fid_le_to_cpu(&ofd_obj->ofo_ff, ff, sizeof(*ff)); + + GOTO(out_unlock, rc); + +out_unlock: + ofd_read_unlock(env, ofd_obj); out_tx: dt_trans_stop(env, ofd->ofd_osd, th); out: la->la_valid = valid; + la->la_flags = flags; return rc; } @@ -1046,7 +1187,7 @@ static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *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)); + strscpy(dcb->dcb_name, "ofd_cb_soft_sync", sizeof(dcb->dcb_name)); rc = dt_trans_cb_add(th, dcb); if (rc) { @@ -1085,6 +1226,7 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp, int niocount, struct niobuf_local *lnb, unsigned long granted, int old_rc) { + struct ofd_thread_info *info = ofd_info(env); struct filter_export_data *fed = &exp->exp_filter_data; struct ofd_object *fo; struct dt_object *o; @@ -1092,7 +1234,7 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp, int rc = 0; int rc2 = 0; int retries = 0; - int i; + int i, restart = 0; bool soft_sync = false; bool cb_registered = false; bool fake_write = false; @@ -1101,15 +1243,16 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp, LASSERT(objcount == 1); - fo = ofd_object_find(env, ofd, fid); + fo = ofd_info(env)->fti_obj; LASSERT(fo != NULL); - LASSERT(ofd_object_exists(fo)); o = ofd_object_child(fo); LASSERT(o != NULL); if (old_rc) GOTO(out, rc = old_rc); + if (!ofd_object_exists(fo)) + GOTO(out, rc = -ENOENT); /* * The first write to each object must set some attributes. It is @@ -1124,7 +1267,7 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp, la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME; /* do fake write, to simulate the write case for performance testing */ - if (OBD_FAIL_CHECK(OBD_FAIL_OST_FAKE_RW)) { + if (CFS_FAIL_CHECK_QUIET(OBD_FAIL_OST_FAKE_RW)) { struct niobuf_local *last = &lnb[niocount - 1]; __u64 file_size = last->lnb_file_offset + last->lnb_len; __u64 valid = la->la_valid; @@ -1146,11 +1289,13 @@ ofd_commitrw_write(const struct lu_env *env, struct obd_export *exp, } retry: + CFS_FAIL_TIMEOUT(OBD_FAIL_OFD_COMMITRW_DELAY, cfs_fail_val); + th = ofd_trans_create(env, ofd); if (IS_ERR(th)) GOTO(out, rc = PTR_ERR(th)); - th->th_sync |= ofd->ofd_syncjournal; + th->th_sync |= ofd->ofd_sync_journal; if (th->th_sync == 0) { for (i = 0; i < niocount; i++) { if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) { @@ -1162,7 +1307,7 @@ retry: } } - if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET)) + if (CFS_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET)) GOTO(out_stop, rc = -EINPROGRESS); if (likely(!fake_write)) { @@ -1171,6 +1316,10 @@ retry: GOTO(out_stop, rc); } + /* don't update atime on disk if it is older */ + if (la->la_valid & LA_ATIME && la->la_atime <= fo->ofo_atime_ondisk) + la->la_valid &= ~LA_ATIME; + if (la->la_valid) { /* update [mac]time if needed */ rc = dt_declare_attr_set(env, o, la, th); @@ -1182,21 +1331,35 @@ retry: if (rc) GOTO(out_stop, rc); - if (likely(!fake_write)) { - rc = dt_write_commit(env, o, lnb, niocount, th); - if (rc) - GOTO(out_stop, rc); - } + ofd_read_lock(env, fo); + if (!ofd_object_exists(fo)) + GOTO(out_unlock, rc = -ENOENT); - if (la->la_valid) { + /* Don't update timestamps if this write is older than a + * setattr which modifies the timestamps. b=10150 */ + if (la->la_valid && tgt_fmd_check(exp, fid, info->fti_xid)) { rc = dt_attr_set(env, o, la, th); if (rc) - GOTO(out_stop, rc); + GOTO(out_unlock, rc); + if (la->la_valid & LA_ATIME) + fo->ofo_atime_ondisk = la->la_atime; + } + + if (likely(!fake_write)) { + CFS_FAIL_TIMEOUT_ORSET(OBD_FAIL_OST_WR_ATTR_DELAY, + CFS_FAIL_ONCE, cfs_fail_val); + rc = dt_write_commit(env, o, lnb, niocount, th, oa->o_size); + if (rc) { + restart = th->th_restart_tran; + GOTO(out_unlock, rc); + } } /* get attr to return */ rc = dt_attr_get(env, o, la); +out_unlock: + ofd_read_unlock(env, fo); out_stop: /* Force commit to make the just-deleted blocks * reusable. LU-456 */ @@ -1214,7 +1377,7 @@ out_stop: granted = 0; } - rc2 = ofd_trans_stop(env, ofd, th, rc); + rc2 = ofd_trans_stop(env, ofd, th, restart ? 0 : rc); if (!rc) rc = rc2; if (rc == -ENOSPC && retries++ < 3) { @@ -1223,6 +1386,16 @@ out_stop: goto retry; } + if (restart) { + retries++; + restart = 0; + if (retries % 10000 == 0) + CERROR("%s: restart IO write too many times: %d\n", + ofd_name(ofd), retries); + CDEBUG(D_INODE, "retry transaction, retries:%d\n", + retries); + goto retry; + } if (!soft_sync) /* reset fed_soft_sync_count upon non-SOFT_SYNC RPC */ atomic_set(&fed->fed_soft_sync_count, 0); @@ -1232,9 +1405,6 @@ out_stop: out: dt_bufs_put(env, o, lnb, niocount); - ofd_read_unlock(env, fo); - ofd_object_put(env, fo); - /* second put is pair to object_get in ofd_preprw_write */ ofd_object_put(env, fo); if (granted > 0) tgt_grant_commit(exp, granted, old_rc); @@ -1266,33 +1436,73 @@ out: 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, int old_rc) + struct niobuf_local *lnb, int old_rc, int nob, ktime_t kstart) { - struct ofd_thread_info *info = ofd_info(env); - struct ofd_mod_data *fmd; - __u64 valid; - struct ofd_device *ofd = ofd_exp(exp); - const struct lu_fid *fid = &oa->o_oi.oi_fid; - int rc = 0; + struct tgt_session_info *tsi = tgt_ses_info(env); + struct ofd_thread_info *info = ofd_info(env); + struct ofd_device *ofd = ofd_exp(exp); + const struct lu_fid *fid = &oa->o_oi.oi_fid; + struct ldlm_namespace *ns = ofd->ofd_namespace; + struct ldlm_resource *rs = NULL; + char *jobid; + __u64 valid; + int rc = 0; + int root_squash = 0; LASSERT(npages > 0); + if (tgt_ses_req(tsi) == NULL) { /* echo client case */ + jobid = NULL; + } else { + jobid = tsi->tsi_jobid; + } + if (cmd == OBD_BRW_WRITE) { struct lu_nodemap *nodemap; + __u32 mapped_uid, mapped_gid, mapped_projid; + + /* doing this before the commit operation places the counter + * update almost immediately after reply to the client, which + * gives reasonable time stats and lets us use the actual + * bytes of i/o (rather than requested) + */ + ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE_BYTES, jobid, nob); + ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE, jobid, + ktime_us_delta(ktime_get(), kstart)); + + mapped_uid = oa->o_uid; + mapped_gid = oa->o_gid; + mapped_projid = oa->o_projid; + nodemap = nodemap_get_from_exp(exp); + if (!IS_ERR(nodemap)) { + mapped_uid = nodemap_map_id(nodemap, NODEMAP_UID, + NODEMAP_FS_TO_CLIENT, + oa->o_uid); + mapped_gid = nodemap_map_id(nodemap, NODEMAP_GID, + NODEMAP_FS_TO_CLIENT, + oa->o_gid); + mapped_projid = nodemap_map_id(nodemap, NODEMAP_PROJID, + NODEMAP_FS_TO_CLIENT, + oa->o_projid); + } else if (old_rc == 0) { + old_rc = PTR_ERR(nodemap); + } + + if (!IS_ERR_OR_NULL(nodemap)) { + /* do not bypass quota enforcement if squashed uid */ + if (unlikely(mapped_uid == nodemap->nm_squash_uid)) { + int idx; + + for (idx = 0; idx < npages; idx++) + lnb[idx].lnb_flags &= + ~OBD_BRW_SYS_RESOURCE; + root_squash = 1; + } + nodemap_putref(nodemap); + } - /* Don't update timestamps if this write is older than a - * setattr which modifies the timestamps. b=10150 */ - - /* XXX when we start having persistent reservations this needs - * to be changed to ofd_fmd_get() to create the fmd if it - * doesn't already exist so we can store the reservation handle - * there. */ - valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID; - 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; - ofd_fmd_put(exp, fmd); + valid = OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLPROJID | + OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME; la_from_obdo(&info->fti_attr, oa, valid); rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr, @@ -1330,40 +1540,45 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, oa->o_flags = OBD_FL_NO_PRJQUOTA; } + if (lnb[0].lnb_flags & OBD_BRW_ROOT_PRJQUOTA) + oa->o_flags |= OBD_FL_ROOT_PRJQUOTA; + + if (root_squash) + oa->o_flags |= OBD_FL_ROOT_SQUASH; + oa->o_valid |= OBD_MD_FLFLAGS; oa->o_valid |= OBD_MD_FLALLQUOTA; } - /* Convert back to client IDs. LU-9671. - * nodemap_get_from_exp() may fail due to nodemap deactivated, - * server ID will be returned back to client in that case. */ - nodemap = nodemap_get_from_exp(exp); - if (nodemap != NULL && !IS_ERR(nodemap)) { - oa->o_uid = nodemap_map_id(nodemap, NODEMAP_UID, - NODEMAP_FS_TO_CLIENT, - oa->o_uid); - oa->o_gid = nodemap_map_id(nodemap, NODEMAP_GID, - NODEMAP_FS_TO_CLIENT, - oa->o_gid); - nodemap_putref(nodemap); - } - } else if (cmd == OBD_BRW_READ) { - struct ldlm_namespace *ns = ofd->ofd_namespace; - - /* If oa != NULL then ofd_preprw_read updated the inode - * atime and we should update the lvb so that other glimpses - * will also get the updated value. bug 5972 */ - if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) { - struct ldlm_resource *rs = NULL; - + /** + * Update LVB after writing finish for server lock, see + * comments in ldlm_lock_decref_internal(), If this is a + * local lock on a server namespace and this was the last + * reference, lock will be destroyed directly thus there + * is no chance for ldlm_request_cancel() to update lvb. + */ + if (rc == 0 && (rnb[0].rnb_flags & OBD_BRW_SRVLOCK)) { ost_fid_build_resid(fid, &info->fti_resid); - rs = ldlm_resource_get(ns, NULL, &info->fti_resid, + rs = ldlm_resource_get(ns, &info->fti_resid, LDLM_EXTENT, 0); if (!IS_ERR(rs)) { ldlm_res_lvbo_update(rs, NULL, 1); ldlm_resource_putref(rs); } } + + /* Convert back to client IDs. LU-9671. + * nodemap_get_from_exp() may fail due to nodemap deactivated, + * server ID will be returned back to client in that case. */ + oa->o_uid = mapped_uid; + oa->o_gid = mapped_gid; + oa->o_projid = mapped_projid; + } else if (cmd == OBD_BRW_READ) { + /* see comment on LPROC_OFD_STATS_WRITE_BYTES usage above */ + ofd_counter_incr(exp, LPROC_OFD_STATS_READ_BYTES, jobid, nob); + ofd_counter_incr(exp, LPROC_OFD_STATS_READ, jobid, + ktime_us_delta(ktime_get(), kstart)); + rc = ofd_commitrw_read(env, ofd, fid, objcount, npages, lnb); if (old_rc)