X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fofd%2Fofd_io.c;h=777a332b379325d575409afd36c5bb317f19b8bc;hp=afda149f6e22f75e180a437bdf7382372f9624c0;hb=f003d3eebbad43e7f8371af663e67c6e56bf5d9a;hpb=14cbce4b70833ee259427fa5a2ba826b75bb5c58 diff --git a/lustre/ofd/ofd_io.c b/lustre/ofd/ofd_io.c index afda149..777a332 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) 2012, 2013, Intel Corporation. + * Copyright (c) 2012, 2015, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -35,16 +31,415 @@ * * 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" +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, int niocount, + struct lu_attr *la, struct obdo *oa, int niocount, struct niobuf_remote *rnb, int *nr_local, struct niobuf_local *lnb, char *jobid) { @@ -63,11 +458,16 @@ static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp, 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)); + lnb + j, 0); if (unlikely(rc < 0)) GOTO(buf_put, rc); LASSERT(rc <= PTLRPC_MAX_BRW_PAGES); @@ -79,8 +479,7 @@ static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp, } 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); @@ -99,6 +498,29 @@ 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, const struct lu_fid *fid, struct lu_attr *la, struct obdo *oa, @@ -114,27 +536,62 @@ 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); @@ -148,6 +605,15 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, 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); + } + } + /* 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); @@ -156,8 +622,7 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, *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)); + rnb + i, lnb + j, 1); if (unlikely(rc < 0)) GOTO(err, rc); LASSERT(rc <= PTLRPC_MAX_BRW_PAGES); @@ -166,10 +631,6 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, lnb[j+k].lnb_flags = rnb[i].rnb_flags; if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED)) lnb[j+k].lnb_rc = -ENOSPC; - - /* 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; @@ -187,8 +648,9 @@ static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp, 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); + ofd_grant_commit(exp, oa->o_grant_used, 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. @@ -197,11 +659,29 @@ out: return rc; } +/** + * 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); @@ -218,10 +698,8 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp, } if (tgt_ses_req(tsi) == NULL) { /* echo client case */ - LASSERT(oti != NULL); info = ofd_info_init(env, exp); - ofd_oti2info(info, oti); - jobid = oti->oti_jobid; + jobid = NULL; } else { info = tsi2ofd_info(tsi); jobid = tsi->tsi_jobid; @@ -229,7 +707,7 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp, LASSERT(oa != NULL); - if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT)) { + if (OBD_FAIL_CHECK(OBD_FAIL_SRV_ENOENT)) { struct ofd_seq *oseq; oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi)); @@ -254,24 +732,15 @@ int ofd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp, LASSERT(obj->ioo_bufcnt > 0); if (cmd == OBD_BRW_WRITE) { - rc = ofd_auth_capa(exp, fid, ostid_seq(&oa->o_oi), - capa, CAPA_OPC_OSS_WRITE); - if (rc == 0) { - 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); - } + 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, fid, ostid_seq(&oa->o_oi), - capa, CAPA_OPC_OSS_READ); - if (rc == 0) { - ofd_grant_prepare_read(env, exp, oa); - rc = ofd_preprw_read(env, exp, ofd, fid, - &info->fti_attr, obj->ioo_bufcnt, - rnb, nr_local, lnb, jobid); - 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); @@ -280,6 +749,21 @@ 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, const struct lu_fid *fid, int objcount, int niocount, @@ -306,6 +790,21 @@ 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, @@ -332,7 +831,7 @@ 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) @@ -354,6 +853,12 @@ 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, @@ -370,21 +875,32 @@ 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: @@ -397,6 +913,17 @@ struct ofd_soft_sync_callback { 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) { @@ -411,6 +938,22 @@ static void ofd_cb_soft_sync(struct lu_env *env, struct thandle *th, 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; @@ -425,9 +968,8 @@ 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; - CFS_INIT_LIST_HEAD(&dcb->dcb_linkage); - strncpy(dcb->dcb_name, "ofd_cb_soft_sync", MAX_COMMIT_CB_STR_LEN); - dcb->dcb_name[MAX_COMMIT_CB_STR_LEN - 1] = '\0'; + 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) { @@ -438,22 +980,44 @@ static int ofd_soft_sync_cb_add(struct thandle *th, struct obd_export *exp) 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] granted grant space consumed for the bulk I/O + * \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 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) + int niocount, struct niobuf_local *lnb, + unsigned long granted, int old_rc) { - struct ofd_thread_info *info = ofd_info(env); - struct ofd_object *fo; - struct dt_object *o; - 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; + struct ofd_object *fo; + struct dt_object *o; + struct thandle *th; + int rc = 0; + int retries = 0; + int i; + bool soft_sync = false; + bool cb_registered = false; + bool fake_write = false; ENTRY; @@ -481,6 +1045,28 @@ 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_WRITE)) { + struct niobuf_local *last = &lnb[niocount - 1]; + __u64 file_size = last->lnb_file_offset + last->lnb_len; + __u64 valid = la->la_valid; + + la->la_valid = LA_SIZE; + la->la_size = 0; + rc = dt_attr_get(env, o, la); + if (rc < 0 && rc != -ENOENT) + GOTO(out, rc); + + if (file_size < la->la_size) + file_size = la->la_size; + + /* dirty inode by setting file size */ + la->la_valid = valid | LA_SIZE; + la->la_size = file_size; + + fake_write = true; + } + retry: th = ofd_trans_create(env, ofd); if (IS_ERR(th)) @@ -501,9 +1087,11 @@ retry: 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) - GOTO(out_stop, rc); + if (likely(!fake_write)) { + rc = dt_declare_write_commit(env, o, lnb, niocount, th); + if (rc) + GOTO(out_stop, rc); + } if (la->la_valid) { /* update [mac]time if needed */ @@ -516,18 +1104,20 @@ retry: if (rc) GOTO(out_stop, rc); - rc = dt_write_commit(env, o, lnb, niocount, th); - 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); + } 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 */ - rc = 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 @@ -541,6 +1131,11 @@ out_stop: cb_registered = true; } + if (rc == 0 && granted > 0) { + if (ofd_grant_commit_cb_add(th, exp, granted) == 0) + granted = 0; + } + ofd_trans_stop(env, ofd, th, rc); if (rc == -ENOSPC && retries++ < 3) { CDEBUG(D_INODE, "retry after force commit, retries:%d\n", @@ -561,15 +1156,37 @@ out: ofd_object_put(env, fo); /* second put is pair to object_get in ofd_preprw_write */ ofd_object_put(env, fo); - ofd_grant_commit(env, info->fti_exp, old_rc); + if (granted > 0) + ofd_grant_commit(exp, granted, old_rc); 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 = ofd_info(env); struct ofd_mod_data *fmd; @@ -603,7 +1220,8 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, } rc = ofd_commitrw_write(env, exp, ofd, fid, &info->fti_attr, - ff, objcount, npages, lnb, old_rc); + ff, objcount, npages, lnb, + oa->o_grant_used, old_rc); if (rc == 0) obdo_from_la(oa, &info->fti_attr, OFD_VALID_FLAGS | LA_GID | LA_UID); @@ -643,8 +1261,8 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, 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); } } @@ -657,7 +1275,5 @@ int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp, rc = -EPROTO; } - if (oti != NULL) - ofd_info2oti(info, oti); RETURN(rc); }