X-Git-Url: https://git.whamcloud.com/?a=blobdiff_plain;f=lustre%2Flmv%2Flmv_obd.c;h=2ac04cab27769b916c291ef12bece81604c67173;hb=a25adb47c7b7eeb68a922e2647d74eeff3401c6a;hp=b9c4825f48b5e08a4505527952306c3fb451ebbc;hpb=edee0234fef8d9e2147a3a10914bbef69a912d69;p=fs%2Flustre-release.git diff --git a/lustre/lmv/lmv_obd.c b/lustre/lmv/lmv_obd.c index b9c4825..2ac04ca 100644 --- a/lustre/lmv/lmv_obd.c +++ b/lustre/lmv/lmv_obd.c @@ -1,7 +1,7 @@ /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- * vim:expandtab:shiftwidth=8:tabstop=8: * - * Copyright (C) 2002, 2003 Cluster File Systems, Inc. + * Copyright (C) 2002, 2003, 2004, 2005, 2006 Cluster File Systems, Inc. * * This file is part of Lustre, http://www.lustre.org. * @@ -29,26 +29,47 @@ #include #include #include +#include #include #include +#include #else #include #endif + #include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include -#include +#include +#include +#include +#include +#include +#include +#include +#include +#include #include "lmv_internal.h" +/* not defined for liblustre building */ +#if !defined(ATOMIC_INIT) +#define ATOMIC_INIT(val) { (val) } +#endif + +/* object cache. */ +cfs_mem_cache_t *obj_cache; +atomic_t obj_cache_count = ATOMIC_INIT(0); + +static void lmv_activate_target(struct lmv_obd *lmv, + struct lmv_tgt_desc *tgt, + int activate) +{ + if (tgt->ltd_active == activate) + return; + + tgt->ltd_active = activate; + lmv->desc.ld_active_tgt_count += (activate ? 1 : -1); +} + /* Error codes: * * -EINVAL : UUID can't be found in the LMV's target list @@ -58,8 +79,8 @@ static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid, int activate) { - struct obd_device *obd; struct lmv_tgt_desc *tgt; + struct obd_device *obd; int i, rc = 0; ENTRY; @@ -72,110 +93,140 @@ static int lmv_set_mdc_active(struct lmv_obd *lmv, struct obd_uuid *uuid, continue; CDEBUG(D_INFO, "lmv idx %d is %s conn "LPX64"\n", - i, tgt->uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie); - if (strncmp(uuid->uuid, tgt->uuid.uuid, sizeof uuid->uuid) == 0) + i, tgt->ltd_uuid.uuid, tgt->ltd_exp->exp_handle.h_cookie); + + if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) break; } if (i == lmv->desc.ld_tgt_count) - GOTO(out, rc = -EINVAL); + GOTO(out_lmv_lock, rc = -EINVAL); obd = class_exp2obd(tgt->ltd_exp); if (obd == NULL) - GOTO(out, rc = -ENOTCONN); + GOTO(out_lmv_lock, rc = -ENOTCONN); CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LMV idx %d\n", obd->obd_name, obd->obd_uuid.uuid, obd->obd_minor, obd, obd->obd_type->typ_name, i); LASSERT(strcmp(obd->obd_type->typ_name, LUSTRE_MDC_NAME) == 0); - if (tgt->active == activate) { + if (tgt->ltd_active == activate) { CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd, activate ? "" : "in"); - GOTO(out, rc); + GOTO(out_lmv_lock, rc); } - CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in"); + CDEBUG(D_INFO, "Marking OBD %p %sactive\n", + obd, activate ? "" : "in"); - tgt->active = activate; - if (activate) - lmv->desc.ld_active_tgt_count++; - else - lmv->desc.ld_active_tgt_count--; + lmv_activate_target(lmv, tgt, activate); EXIT; - out: + + out_lmv_lock: spin_unlock(&lmv->lmv_lock); return rc; } +static int lmv_set_mdc_data(struct lmv_obd *lmv, struct obd_uuid *uuid, + struct obd_connect_data *data) +{ + struct lmv_tgt_desc *tgt; + int i; + ENTRY; + + LASSERT(data != NULL); + + spin_lock(&lmv->lmv_lock); + for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) { + if (tgt->ltd_exp == NULL) + continue; + + if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) { + lmv->datas[tgt->ltd_idx] = *data; + break; + } + } + spin_unlock(&lmv->lmv_lock); + RETURN(0); +} + static int lmv_notify(struct obd_device *obd, struct obd_device *watched, - int active) + enum obd_notify_event ev, void *data) { - int rc; + struct lmv_obd *lmv = &obd->u.lmv; struct obd_uuid *uuid; + int rc = 0; + ENTRY; if (strcmp(watched->obd_type->typ_name, LUSTRE_MDC_NAME)) { CERROR("unexpected notification of %s %s!\n", watched->obd_type->typ_name, watched->obd_name); - return -EINVAL; + RETURN(-EINVAL); } - uuid = &watched->u.cli.cl_import->imp_target_uuid; - /* Set MDC as active before notifying the observer, so the - * observer can use the MDC normally. - */ - rc = lmv_set_mdc_active(&obd->u.lmv, uuid, active); - if (rc) { - CERROR("%sactivation of %s failed: %d\n", - active ? "" : "de", uuid->uuid, rc); - RETURN(rc); - } + uuid = &watched->u.cli.cl_target_uuid; + if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) { + /* + * Set MDC as active before notifying the observer, so the + * observer can use the MDC normally. + */ + rc = lmv_set_mdc_active(lmv, uuid, + ev == OBD_NOTIFY_ACTIVE); + if (rc) { + CERROR("%sactivation of %s failed: %d\n", + ev == OBD_NOTIFY_ACTIVE ? "" : "de", + uuid->uuid, rc); + RETURN(rc); + } + } else if (ev == OBD_NOTIFY_OCD) { + struct obd_connect_data *conn_data = + &watched->u.cli.cl_import->imp_connect_data; + + /* Set connect data to desired target, update + * exp_connect_flags. */ + rc = lmv_set_mdc_data(lmv, uuid, conn_data); + if (rc) { + CERROR("can't set connect data to target %s, rc %d\n", + uuid->uuid, rc); + RETURN(rc); + } + /* + * XXX: Make sure that ocd_connect_flags from all targets are + * the same. Otherwise one of MDTs runs wrong version or + * something like this. --umka + */ + obd->obd_self_export->exp_connect_flags = + conn_data->ocd_connect_flags; + } +#if 0 + else if (ev == OBD_NOTIFY_DISCON) { + /* For disconnect event, flush fld cache for failout MDS case. */ + fld_client_flush(&lmv->lmv_fld); + } +#endif + /* Pass the notification up the chain. */ if (obd->obd_observer) - /* Pass the notification up the chain. */ - rc = obd_notify(obd->obd_observer, watched, active); + rc = obd_notify(obd->obd_observer, watched, ev, data); RETURN(rc); } -int lmv_attach(struct obd_device *dev, obd_count len, void *data) +/* this is fake connect function. Its purpose is to initialize lmv and say + * caller that everything is okay. Real connection will be performed later. */ +static int lmv_connect(const struct lu_env *env, + struct lustre_handle *conn, struct obd_device *obd, + struct obd_uuid *cluuid, struct obd_connect_data *data) { - struct lprocfs_static_vars lvars; - int rc; - ENTRY; - - lprocfs_init_vars(lmv, &lvars); - rc = lprocfs_obd_attach(dev, lvars.obd_vars); - if (rc == 0) { #ifdef __KERNEL__ - struct proc_dir_entry *entry; - - entry = create_proc_entry("target_obd", 0444, dev->obd_proc_entry); - if (entry == NULL) - RETURN(-ENOMEM); - /* entry->proc_fops = &lmv_proc_target_fops; */ - entry->data = dev; + struct proc_dir_entry *lmv_proc_dir; #endif - } - RETURN (rc); -} - -int lmv_detach(struct obd_device *dev) -{ - return lprocfs_obd_detach(dev); -} - -/* This is fake connect function. Its purpose is to initialize lmv and - * say caller that everything is okay. Real connection will be performed - * later. */ -static int lmv_connect(struct lustre_handle *conn, struct obd_device *obd, - struct obd_uuid *cluuid) -{ struct lmv_obd *lmv = &obd->u.lmv; struct obd_export *exp; - int rc; + int rc = 0; ENTRY; rc = class_connect(conn, obd, cluuid); @@ -185,7 +236,8 @@ static int lmv_connect(struct lustre_handle *conn, struct obd_device *obd, } exp = class_conn2export(conn); - /* We don't want to actually do the underlying connections more than + + /* we don't want to actually do the underlying connections more than * once, so keep track. */ lmv->refcount++; if (lmv->refcount > 1) { @@ -193,14 +245,41 @@ static int lmv_connect(struct lustre_handle *conn, struct obd_device *obd, RETURN(0); } - lmv->cluuid = *cluuid; - lmv->connected = 0; lmv->exp = exp; + lmv->connected = 0; + lmv->cluuid = *cluuid; - RETURN(0); + if (data) + lmv->conn_data = *data; + +#ifdef __KERNEL__ + lmv_proc_dir = lprocfs_register("target_obds", obd->obd_proc_entry, + NULL, NULL); + if (IS_ERR(lmv_proc_dir)) { + CERROR("could not register /proc/fs/lustre/%s/%s/target_obds.", + obd->obd_type->typ_name, obd->obd_name); + lmv_proc_dir = NULL; + } +#endif + + /* all real clients should perform actual connection right away, because + * it is possible, that LMV will not have opportunity to connect targets + * and MDC stuff will be called directly, for instance while reading + * ../mdc/../kbytesfree procfs file, etc. */ + if (data->ocd_connect_flags & OBD_CONNECT_REAL) + rc = lmv_check_connect(obd); + +#ifdef __KERNEL__ + if (rc) { + if (lmv_proc_dir) + lprocfs_remove(&lmv_proc_dir); + } +#endif + + RETURN(rc); } -void lmv_set_timeouts(struct obd_device *obd) +static void lmv_set_timeouts(struct obd_device *obd) { struct lmv_tgt_desc *tgts; struct lmv_obd *lmv; @@ -216,113 +295,361 @@ void lmv_set_timeouts(struct obd_device *obd) for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) { if (tgts->ltd_exp == NULL) continue; - obd_set_info(tgts->ltd_exp, strlen("inter_mds"), - "inter_mds", 0, NULL); + + obd_set_info_async(tgts->ltd_exp, strlen("inter_mds"), + "inter_mds", 0, NULL, NULL); } } -/* Performs a check if passed obd is connected. If no - connect it. */ -int lmv_check_connect(struct obd_device *obd) { +static int lmv_init_ea_size(struct obd_export *exp, int easize, + int def_easize, int cookiesize) +{ + struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct obd_uuid *cluuid; - struct lmv_tgt_desc *tgts; - struct obd_export *exp; - int rc, rc2, i; + int i, rc = 0, change = 0; + ENTRY; - if (lmv->connected) - return 0; - - lmv->connected = 1; - cluuid = &lmv->cluuid; - exp = lmv->exp; - - CDEBUG(D_OTHER, "time to connect %s to %s\n", - cluuid->uuid, obd->obd_name); + if (lmv->max_easize < easize) { + lmv->max_easize = easize; + change = 1; + } + if (lmv->max_def_easize < def_easize) { + lmv->max_def_easize = def_easize; + change = 1; + } + if (lmv->max_cookiesize < cookiesize) { + lmv->max_cookiesize = cookiesize; + change = 1; + } + if (change == 0) + RETURN(0); - for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) { - struct obd_device *tgt_obd; - struct obd_uuid lmv_osc_uuid = { "LMV_OSC_UUID" }; - struct lustre_handle conn = {0, }; + if (lmv->connected == 0) + RETURN(0); - LASSERT(tgts != NULL); + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { + if (lmv->tgts[i].ltd_exp == NULL) { + CWARN("%s: NULL export for %d\n", obd->obd_name, i); + continue; + } - tgt_obd = class_find_client_obd(&tgts->uuid, LUSTRE_MDC_NAME, - &obd->obd_uuid); - if (!tgt_obd) { - CERROR("Target %s not attached\n", tgts->uuid.uuid); - GOTO(out_disc, rc = -EINVAL); + rc = md_init_ea_size(lmv->tgts[i].ltd_exp, easize, def_easize, + cookiesize); + if (rc) { + CERROR("obd_init_ea_size() failed on MDT target %d, " + "error %d.\n", i, rc); + break; } + } + RETURN(rc); +} - /* for MDS: don't connect to yourself */ - if (obd_uuid_equals(&tgts->uuid, cluuid)) { - CDEBUG(D_OTHER, "don't connect back to %s\n", - cluuid->uuid); - tgts->ltd_exp = NULL; - continue; +#define MAX_STRING_SIZE 128 + +int lmv_connect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt) +{ + struct lmv_obd *lmv = &obd->u.lmv; + struct obd_uuid *cluuid = &lmv->cluuid; + struct obd_connect_data *mdc_data = NULL; + struct obd_uuid lmv_mdc_uuid = { "LMV_MDC_UUID" }; + struct lustre_handle conn = {0, }; + struct obd_device *mdc_obd; + struct obd_export *mdc_exp; + struct lu_fld_target target; + int rc; +#ifdef __KERNEL__ + struct proc_dir_entry *lmv_proc_dir; +#endif + ENTRY; + + /* for MDS: don't connect to yourself */ + if (obd_uuid_equals(&tgt->ltd_uuid, cluuid)) { + CDEBUG(D_CONFIG, "don't connect back to %s\n", cluuid->uuid); + /* XXX - the old code didn't increment active tgt count. + * should we ? */ + RETURN(0); + } + + mdc_obd = class_find_client_obd(&tgt->ltd_uuid, LUSTRE_MDC_NAME, + &obd->obd_uuid); + if (!mdc_obd) { + CERROR("target %s not attached\n", tgt->ltd_uuid.uuid); + RETURN(-EINVAL); + } + + CDEBUG(D_CONFIG, "connect to %s(%s) - %s, %s FOR %s\n", + mdc_obd->obd_name, mdc_obd->obd_uuid.uuid, + tgt->ltd_uuid.uuid, obd->obd_uuid.uuid, + cluuid->uuid); + + if (!mdc_obd->obd_set_up) { + CERROR("target %s is not set up\n", tgt->ltd_uuid.uuid); + RETURN(-EINVAL); + } + + rc = obd_connect(NULL, &conn, mdc_obd, &lmv_mdc_uuid, + &lmv->conn_data); + if (rc) { + CERROR("target %s connect error %d\n", tgt->ltd_uuid.uuid, rc); + RETURN(rc); + } + + mdc_exp = class_conn2export(&conn); + + /* Init fid sequence client for this mdc. */ + rc = obd_fid_init(mdc_exp); + if (rc) + RETURN(rc); + + /* Add new FLD target. */ + target.ft_srv = NULL; + target.ft_exp = mdc_exp; + target.ft_idx = tgt->ltd_idx; + + fld_client_add_target(&lmv->lmv_fld, &target); + + mdc_data = &class_exp2cliimp(mdc_exp)->imp_connect_data; + + rc = obd_register_observer(mdc_obd, obd); + if (rc) { + obd_disconnect(mdc_exp); + CERROR("target %s register_observer error %d\n", + tgt->ltd_uuid.uuid, rc); + RETURN(rc); + } + + if (obd->obd_observer) { + /* tell the mds_lmv about the new target */ + rc = obd_notify(obd->obd_observer, mdc_exp->exp_obd, + OBD_NOTIFY_ACTIVE, (void *)(tgt - lmv->tgts)); + if (rc) { + obd_disconnect(mdc_exp); + RETURN(rc); + } + } + + tgt->ltd_active = 1; + tgt->ltd_exp = mdc_exp; + lmv->desc.ld_active_tgt_count++; + + /* copy connect data, it may be used later */ + lmv->datas[tgt->ltd_idx] = *mdc_data; + + md_init_ea_size(tgt->ltd_exp, lmv->max_easize, + lmv->max_def_easize, lmv->max_cookiesize); + + CDEBUG(D_CONFIG, "connected to %s(%s) successfully (%d)\n", + mdc_obd->obd_name, mdc_obd->obd_uuid.uuid, + atomic_read(&obd->obd_refcount)); + +#ifdef __KERNEL__ + lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds"); + if (lmv_proc_dir) { + struct proc_dir_entry *mdc_symlink; + char name[MAX_STRING_SIZE + 1]; + + LASSERT(mdc_obd->obd_type != NULL); + LASSERT(mdc_obd->obd_type->typ_name != NULL); + name[MAX_STRING_SIZE] = '\0'; + snprintf(name, MAX_STRING_SIZE, "../../../%s/%s", + mdc_obd->obd_type->typ_name, + mdc_obd->obd_name); + mdc_symlink = proc_symlink(mdc_obd->obd_name, + lmv_proc_dir, name); + if (mdc_symlink == NULL) { + CERROR("could not register LMV target " + "/proc/fs/lustre/%s/%s/target_obds/%s.", + obd->obd_type->typ_name, obd->obd_name, + mdc_obd->obd_name); + lprocfs_remove(&lmv_proc_dir); + lmv_proc_dir = NULL; } + } +#endif + RETURN(0); +} + +int lmv_add_target(struct obd_device *obd, struct obd_uuid *tgt_uuid) +{ + struct lmv_obd *lmv = &obd->u.lmv; + struct lmv_tgt_desc *tgt; + int rc = 0; + ENTRY; + + CDEBUG(D_CONFIG, "tgt_uuid: %s.\n", tgt_uuid->uuid); - CDEBUG(D_OTHER, "connect to %s(%s) - %s, %s FOR %s\n", - tgt_obd->obd_name, tgt_obd->obd_uuid.uuid, - tgts->uuid.uuid, obd->obd_uuid.uuid, - cluuid->uuid); + lmv_init_lock(lmv); - if (!tgt_obd->obd_set_up) { - CERROR("Target %s not set up\n", tgts->uuid.uuid); - GOTO(out_disc, rc = -EINVAL); + if (lmv->desc.ld_active_tgt_count >= LMV_MAX_TGT_COUNT) { + lmv_init_unlock(lmv); + CERROR("can't add %s, LMV module compiled for %d MDCs. " + "That many MDCs already configured.\n", + tgt_uuid->uuid, LMV_MAX_TGT_COUNT); + RETURN(-EINVAL); + } + if (lmv->desc.ld_tgt_count == 0) { + struct obd_device *mdc_obd; + + mdc_obd = class_find_client_obd(tgt_uuid, LUSTRE_MDC_NAME, + &obd->obd_uuid); + if (!mdc_obd) { + lmv_init_unlock(lmv); + CERROR("Target %s not attached\n", tgt_uuid->uuid); + RETURN(-EINVAL); } - - rc = obd_connect(&conn, tgt_obd, &lmv_osc_uuid); + + rc = obd_llog_init(obd, NULL, mdc_obd, 0, NULL, tgt_uuid); if (rc) { - CERROR("Target %s connect error %d\n", - tgts->uuid.uuid, rc); - GOTO(out_disc, rc); + lmv_init_unlock(lmv); + CERROR("lmv failed to setup llogging subsystems\n"); } - tgts->ltd_exp = class_conn2export(&conn); + } + spin_lock(&lmv->lmv_lock); + tgt = lmv->tgts + lmv->desc.ld_tgt_count++; + tgt->ltd_uuid = *tgt_uuid; + spin_unlock(&lmv->lmv_lock); - obd_init_ea_size(tgts->ltd_exp, lmv->max_easize, - lmv->max_cookiesize); - - rc = obd_register_observer(tgt_obd, obd); + if (lmv->connected) { + rc = lmv_connect_mdc(obd, tgt); if (rc) { - CERROR("Target %s register_observer error %d\n", - tgts->uuid.uuid, rc); - obd_disconnect(tgts->ltd_exp, 0); - GOTO(out_disc, rc); + spin_lock(&lmv->lmv_lock); + lmv->desc.ld_tgt_count--; + memset(tgt, 0, sizeof(*tgt)); + spin_unlock(&lmv->lmv_lock); + } else { + int easize = sizeof(struct lmv_stripe_md) + + lmv->desc.ld_tgt_count * + sizeof(struct lu_fid); + lmv_init_ea_size(obd->obd_self_export, easize, 0, 0); } + } - lmv->desc.ld_active_tgt_count++; - tgts->active = 1; - - CDEBUG(D_OTHER, "connected to %s(%s) successfully (%d)\n", - tgt_obd->obd_name, tgt_obd->obd_uuid.uuid, - atomic_read(&obd->obd_refcount)); + lmv_init_unlock(lmv); + RETURN(rc); +} + +/* performs a check if passed obd is connected. If no - connect it. */ +int lmv_check_connect(struct obd_device *obd) +{ + struct lmv_obd *lmv = &obd->u.lmv; + struct lmv_tgt_desc *tgt; + int i, rc, easize; + ENTRY; + + if (lmv->connected) + RETURN(0); + + lmv_init_lock(lmv); + if (lmv->connected) { + lmv_init_unlock(lmv); + RETURN(0); } - lmv_set_timeouts(obd); + if (lmv->desc.ld_tgt_count == 0) { + CERROR("%s: no targets configured.\n", obd->obd_name); + RETURN(-EINVAL); + } + + CDEBUG(D_CONFIG, "time to connect %s to %s\n", + lmv->cluuid.uuid, obd->obd_name); + + LASSERT(lmv->tgts != NULL); - class_export_put(exp); - return 0; + for (i = 0, tgt = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgt++) { + rc = lmv_connect_mdc(obd, tgt); + if (rc) + GOTO(out_disc, rc); + } + + lmv_set_timeouts(obd); + class_export_put(lmv->exp); + lmv->connected = 1; + easize = lmv_get_easize(lmv); + lmv_init_ea_size(obd->obd_self_export, easize, 0, 0); + lmv_init_unlock(lmv); + RETURN(0); out_disc: while (i-- > 0) { - struct obd_uuid uuid; - --tgts; - --lmv->desc.ld_active_tgt_count; - tgts->active = 0; - /* save for CERROR below; (we know it's terminated) */ - uuid = tgts->uuid; - rc2 = obd_disconnect(tgts->ltd_exp, 0); - if (rc2) - CERROR("error: LMV target %s disconnect on MDT idx %d: " - "error %d\n", uuid.uuid, i, rc2); + int rc2; + --tgt; + tgt->ltd_active = 0; + if (tgt->ltd_exp) { + --lmv->desc.ld_active_tgt_count; + rc2 = obd_disconnect(tgt->ltd_exp); + if (rc2) { + CERROR("error: LMV target %s disconnect on " + "MDC idx %d: error %d\n", + tgt->ltd_uuid.uuid, i, rc2); + } + } + } + class_disconnect(lmv->exp); + lmv_init_unlock(lmv); + RETURN(rc); +} + +static int lmv_disconnect_mdc(struct obd_device *obd, struct lmv_tgt_desc *tgt) +{ +#ifdef __KERNEL__ + struct proc_dir_entry *lmv_proc_dir; +#endif + struct lmv_obd *lmv = &obd->u.lmv; + struct obd_device *mdc_obd; + int rc; + ENTRY; + + LASSERT(tgt != NULL); + LASSERT(obd != NULL); + + mdc_obd = class_exp2obd(tgt->ltd_exp); + + if (mdc_obd) + mdc_obd->obd_no_recov = obd->obd_no_recov; + +#ifdef __KERNEL__ + lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds"); + if (lmv_proc_dir) { + struct proc_dir_entry *mdc_symlink; + + mdc_symlink = lprocfs_srch(lmv_proc_dir, mdc_obd->obd_name); + if (mdc_symlink) { + lprocfs_remove(&mdc_symlink); + } else { + CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing\n", + obd->obd_type->typ_name, obd->obd_name, + mdc_obd->obd_name); + } + } +#endif + rc = obd_fid_fini(tgt->ltd_exp); + if (rc) + CERROR("Can't finanize fids factory\n"); + + CDEBUG(D_OTHER, "Disconnected from %s(%s) successfully\n", + tgt->ltd_exp->exp_obd->obd_name, + tgt->ltd_exp->exp_obd->obd_uuid.uuid); + + obd_register_observer(tgt->ltd_exp->exp_obd, NULL); + rc = obd_disconnect(tgt->ltd_exp); + if (rc) { + if (tgt->ltd_active) { + CERROR("Target %s disconnect error %d\n", + tgt->ltd_uuid.uuid, rc); + } } - class_disconnect(exp, 0); - RETURN (rc); + + lmv_activate_target(lmv, tgt, 0); + tgt->ltd_exp = NULL; + RETURN(0); } -static int lmv_disconnect(struct obd_export *exp, int flags) +static int lmv_disconnect(struct obd_export *exp) { struct obd_device *obd = class_exp2obd(exp); +#ifdef __KERNEL__ + struct proc_dir_entry *lmv_proc_dir; +#endif struct lmv_obd *lmv = &obd->u.lmv; int rc, i; ENTRY; @@ -338,45 +665,29 @@ static int lmv_disconnect(struct obd_export *exp, int flags) for (i = 0; i < lmv->desc.ld_tgt_count; i++) { if (lmv->tgts[i].ltd_exp == NULL) continue; + lmv_disconnect_mdc(obd, &lmv->tgts[i]); + } - if (obd->obd_no_recov) { - /* Pass it on to our clients. - * XXX This should be an argument to disconnect, - * XXX not a back-door flag on the OBD. Ah well. - */ - struct obd_device *mdc_obd; - mdc_obd = class_exp2obd(lmv->tgts[i].ltd_exp); - if (mdc_obd) - mdc_obd->obd_no_recov = 1; - } - - CDEBUG(D_OTHER, "disconnected from %s(%s) successfully\n", - lmv->tgts[i].ltd_exp->exp_obd->obd_name, - lmv->tgts[i].ltd_exp->exp_obd->obd_uuid.uuid); - - obd_register_observer(lmv->tgts[i].ltd_exp->exp_obd, NULL); - - rc = obd_disconnect(lmv->tgts[i].ltd_exp, flags); - if (rc) { - if (lmv->tgts[i].active) { - CERROR("Target %s disconnect error %d\n", - lmv->tgts[i].uuid.uuid, rc); - } - rc = 0; - } - if (lmv->tgts[i].active) { - lmv->desc.ld_active_tgt_count--; - lmv->tgts[i].active = 0; - } - lmv->tgts[i].ltd_exp = NULL; +#ifdef __KERNEL__ + lmv_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds"); + if (lmv_proc_dir) { + lprocfs_remove(&lmv_proc_dir); + } else { + CERROR("/proc/fs/lustre/%s/%s/target_obds missing\n", + obd->obd_type->typ_name, obd->obd_name); } +#endif out_local: - /* this is the case when no real connection is established by - * lmv_check_connect(). */ + /* + * This is the case when no real connection is established by + * lmv_check_connect(). + */ if (!lmv->connected) class_export_put(exp); - rc = class_disconnect(exp, 0); + rc = class_disconnect(exp); + if (lmv->refcount == 0) + lmv->connected = 0; RETURN(rc); } @@ -386,22 +697,23 @@ static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp, struct obd_device *obddev = class_exp2obd(exp); struct lmv_obd *lmv = &obddev->u.lmv; int i, rc = 0, set = 0; - ENTRY; if (lmv->desc.ld_tgt_count == 0) RETURN(-ENOTTY); - + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { int err; - err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, - len, karg, uarg); + if (lmv->tgts[i].ltd_exp == NULL) + continue; + + err = obd_iocontrol(cmd, lmv->tgts[i].ltd_exp, len, karg, uarg); if (err) { - if (lmv->tgts[i].active) { + if (lmv->tgts[i].ltd_active) { CERROR("error: iocontrol MDC %s on MDT" "idx %d: err = %d\n", - lmv->tgts[i].uuid.uuid, i, err); + lmv->tgts[i].ltd_uuid.uuid, i, err); if (!rc) rc = err; } @@ -414,283 +726,616 @@ static int lmv_iocontrol(unsigned int cmd, struct obd_export *exp, RETURN(rc); } -static int lmv_setup(struct obd_device *obd, obd_count len, void *buf) +enum MDS_POLICY { + CHAR_TYPE, + NID_TYPE +}; + +static int lmv_all_chars_policy(int count, const char *name, + int len) +{ + unsigned int c = 0; + + while (len > 0) + c += name[--len]; + c = c % count; + return c; +} + +static int lmv_nid_policy(struct lmv_obd *lmv) +{ + struct obd_import *imp = class_exp2cliimp(lmv->tgts[0].ltd_exp); + __u32 id; + /* + * XXX Hack: to get nid we assume that underlying obd device is mdc. + */ + id = imp->imp_connection->c_self ^ (imp->imp_connection->c_self >> 32); + return id % lmv->desc.ld_tgt_count; +} + +static int lmv_choose_mds(struct lmv_obd *lmv, struct md_op_data *op_data, + int type) +{ + switch (type) { + case CHAR_TYPE: + return lmv_all_chars_policy(lmv->desc.ld_tgt_count, + op_data->op_name, + op_data->op_namelen); + case NID_TYPE: + return lmv_nid_policy(lmv); + + default: + break; + } + + CERROR("unsupport type %d \n", type); + return -EINVAL; +} + +/* This is _inode_ placement policy function (not name). */ +static int lmv_placement_policy(struct obd_device *obd, + struct md_op_data *op_data, + mdsno_t *mds) { - struct lustre_cfg *lcfg = buf; - struct lmv_desc *desc; struct lmv_obd *lmv = &obd->u.lmv; - struct obd_uuid *uuids; - struct lmv_tgt_desc *tgts; - int i; - int rc = 0; + struct lmv_obj *obj; + int rc; ENTRY; - if (lcfg->lcfg_inllen1 < 1) { - CERROR("LMV setup requires a descriptor\n"); - RETURN(-EINVAL); + LASSERT(mds != NULL); + + /* + * Allocate new fid on target according to operation type and parent + * home mds. + */ + obj = lmv_obj_grab(obd, &op_data->op_fid1); + if (obj != NULL || op_data->op_name == NULL || + op_data->op_opc != LUSTRE_OPC_MKDIR) { + /* + * Allocate fid for non-dir or for null name or for case parent + * dir is split. + */ + if (obj) { + lmv_obj_put(obj); + + /* + * If we have this flag turned on, and we see that + * parent dir is split, this means, that caller did not + * notice split yet. This is race and we would like to + * let caller know that. + */ + if (op_data->op_bias & MDS_CHECK_SPLIT) + RETURN(-ERESTART); + } + + /* + * Allocate new fid on same mds where parent fid is located and + * where operation will be sent. In case of split dir, ->op_fid1 + * and ->op_mds here will contain fid and mds of slave directory + * object (assigned by caller). + */ + *mds = op_data->op_mds; + rc = 0; + +#if 0 + /* XXX: This should be removed later wehn we sure it is not + * needed. */ + rc = lmv_fld_lookup(lmv, &op_data->op_fid1, mds); + if (rc) + GOTO(out, rc); +#endif + } else { + /* + * Parent directory is not split and we want to create a + * directory in it. Let's calculate where to place it according + * to name. + */ + *mds = lmv_choose_mds(lmv, op_data, NID_TYPE); + rc = 0; + } + EXIT; +#if 0 +out: +#endif + if (rc) { + CERROR("Can't choose MDS, err = %d\n", rc); + } else { + LASSERT(*mds < lmv->desc.ld_tgt_count); } - if (lcfg->lcfg_inllen2 < 1) { - CERROR("LMV setup requires an OST UUID list\n"); - RETURN(-EINVAL); + return rc; +} + +int __lmv_fid_alloc(struct lmv_obd *lmv, struct lu_fid *fid, + mdsno_t mds) +{ + struct lmv_tgt_desc *tgt = &lmv->tgts[mds]; + int rc; + ENTRY; + + /* New seq alloc and FLD setup should be atomic. */ + down(&tgt->ltd_fid_sem); + + /* Asking underlaying tgt layer to allocate new fid. */ + rc = obd_fid_alloc(tgt->ltd_exp, fid, NULL); + if (rc > 0) { + LASSERT(fid_is_sane(fid)); + + /* Client switches to new sequence, setup FLD. */ + rc = fld_client_create(&lmv->lmv_fld, fid_seq(fid), + mds, NULL); + if (rc) { + CERROR("Can't create fld entry, rc %d\n", rc); + /* Delete just allocated fid sequence */ + obd_fid_delete(tgt->ltd_exp, NULL); + } } + up(&tgt->ltd_fid_sem); + RETURN(rc); +} - desc = (struct lmv_desc *)lcfg->lcfg_inlbuf1; - if (sizeof(*desc) > lcfg->lcfg_inllen1) { - CERROR("descriptor size wrong: %d > %d\n", - (int)sizeof(*desc), lcfg->lcfg_inllen1); - RETURN(-EINVAL); +int lmv_fid_alloc(struct obd_export *exp, struct lu_fid *fid, + struct md_op_data *op_data) +{ + struct obd_device *obd = class_exp2obd(exp); + struct lmv_obd *lmv = &obd->u.lmv; + mdsno_t mds; + int rc; + ENTRY; + + LASSERT(op_data != NULL); + LASSERT(fid != NULL); + + rc = lmv_placement_policy(obd, op_data, &mds); + if (rc) { + CERROR("Can't get target for allocating fid, " + "rc %d\n", rc); + RETURN(rc); + } + + rc = __lmv_fid_alloc(lmv, fid, mds); + if (rc) { + CERROR("Can't alloc new fid, rc %d\n", rc); + RETURN(rc); + } + + RETURN(rc); +} + +static int lmv_fid_delete(struct obd_export *exp, const struct lu_fid *fid) +{ + ENTRY; + + LASSERT(exp && fid); + if (lmv_obj_delete(exp, fid)) { + CDEBUG(D_OTHER, "lmv object "DFID" is destroyed.\n", + PFID(fid)); } + RETURN(0); +} - uuids = (struct obd_uuid *)lcfg->lcfg_inlbuf2; - if (sizeof(*uuids) * desc->ld_tgt_count != lcfg->lcfg_inllen2) { - CERROR("UUID array size wrong: %u * %u != %u\n", - sizeof(*uuids), desc->ld_tgt_count, lcfg->lcfg_inllen2); +static int lmv_setup(struct obd_device *obd, struct lustre_cfg *lcfg) +{ + struct lmv_obd *lmv = &obd->u.lmv; + struct lprocfs_static_vars lvars; + struct lmv_desc *desc; + int rc, i = 0; + ENTRY; + + if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) { + CERROR("LMV setup requires a descriptor\n"); RETURN(-EINVAL); } - lmv->bufsize = sizeof(struct lmv_tgt_desc) * desc->ld_tgt_count; - OBD_ALLOC(lmv->tgts, lmv->bufsize); - if (lmv->tgts == NULL) { - CERROR("Out of memory\n"); + desc = (struct lmv_desc *)lustre_cfg_buf(lcfg, 1); + if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) { + CERROR("descriptor size wrong: %d > %d\n", + (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1)); RETURN(-EINVAL); } - lmv->desc = *desc; - spin_lock_init(&lmv->lmv_lock); - - for (i = 0, tgts = lmv->tgts; i < desc->ld_tgt_count; i++, tgts++) - tgts->uuid = uuids[i]; - - lmv->max_easize = sizeof(struct ll_fid) * - desc->ld_tgt_count + sizeof(struct mea); - + lmv->tgts_size = LMV_MAX_TGT_COUNT * sizeof(struct lmv_tgt_desc); + + OBD_ALLOC(lmv->tgts, lmv->tgts_size); + if (lmv->tgts == NULL) + RETURN(-ENOMEM); + + for (i = 0; i < LMV_MAX_TGT_COUNT; i++) { + sema_init(&lmv->tgts[i].ltd_fid_sem, 1); + lmv->tgts[i].ltd_idx = i; + } + + lmv->datas_size = LMV_MAX_TGT_COUNT * sizeof(struct obd_connect_data); + + OBD_ALLOC(lmv->datas, lmv->datas_size); + if (lmv->datas == NULL) + GOTO(out_free_tgts, rc = -ENOMEM); + + obd_str2uuid(&lmv->desc.ld_uuid, desc->ld_uuid.uuid); + lmv->desc.ld_tgt_count = 0; + lmv->desc.ld_active_tgt_count = 0; lmv->max_cookiesize = 0; + lmv->max_def_easize = 0; + lmv->max_easize = 0; + + spin_lock_init(&lmv->lmv_lock); + sema_init(&lmv->init_sem, 1); + + rc = lmv_obj_setup(obd); + if (rc) { + CERROR("Can't setup LMV object manager, " + "error %d.\n", rc); + GOTO(out_free_datas, rc); + } + + lprocfs_init_vars(lmv, &lvars); + lprocfs_obd_setup(obd, lvars.obd_vars); +#ifdef LPROCFS + { + struct proc_dir_entry *entry; + + entry = create_proc_entry("target_obd_status", 0444, + obd->obd_proc_entry); + if (entry != NULL) { + entry->proc_fops = &lmv_proc_target_fops; + entry->data = obd; + } + } +#endif + rc = fld_client_init(&lmv->lmv_fld, obd->obd_name, + LUSTRE_CLI_FLD_HASH_DHT); + if (rc) { + CERROR("can't init FLD, err %d\n", + rc); + GOTO(out_free_datas, rc); + } + + RETURN(0); + +out_free_datas: + OBD_FREE(lmv->datas, lmv->datas_size); + lmv->datas = NULL; +out_free_tgts: + OBD_FREE(lmv->tgts, lmv->tgts_size); + lmv->tgts = NULL; + return rc; +} + +static int lmv_cleanup(struct obd_device *obd) +{ + struct lmv_obd *lmv = &obd->u.lmv; + ENTRY; + + fld_client_fini(&lmv->lmv_fld); + lprocfs_obd_cleanup(obd); + lmv_obj_cleanup(obd); + OBD_FREE(lmv->datas, lmv->datas_size); + OBD_FREE(lmv->tgts, lmv->tgts_size); + + RETURN(0); +} + +static int lmv_process_config(struct obd_device *obd, obd_count len, void *buf) +{ + struct lustre_cfg *lcfg = buf; + struct obd_uuid tgt_uuid; + int rc; + ENTRY; + + switch(lcfg->lcfg_command) { + case LCFG_ADD_MDC: + if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(tgt_uuid.uuid)) + GOTO(out, rc = -EINVAL); + obd_str2uuid(&tgt_uuid, lustre_cfg_string(lcfg, 1)); + rc = lmv_add_target(obd, &tgt_uuid); + GOTO(out, rc); + default: { + CERROR("Unknown command: %d\n", lcfg->lcfg_command); + GOTO(out, rc = -EINVAL); + } + } +out: RETURN(rc); } static int lmv_statfs(struct obd_device *obd, struct obd_statfs *osfs, - unsigned long max_age) + __u64 max_age) { struct lmv_obd *lmv = &obd->u.lmv; - struct obd_statfs temp; + struct obd_statfs *temp; int rc = 0, i; ENTRY; - + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - + + OBD_ALLOC(temp, sizeof(*temp)); + if (temp == NULL) + RETURN(-ENOMEM); + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { - rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, &temp, max_age); + if (lmv->tgts[i].ltd_exp == NULL) + continue; + + rc = obd_statfs(lmv->tgts[i].ltd_exp->exp_obd, temp, max_age); if (rc) { - CERROR("can't stat MDS #%d (%s)\n", i, - lmv->tgts[i].ltd_exp->exp_obd->obd_name); - RETURN(rc); + CERROR("can't stat MDS #%d (%s), error %d\n", i, + lmv->tgts[i].ltd_exp->exp_obd->obd_name, + rc); + GOTO(out_free_temp, rc); } if (i == 0) { - memcpy(osfs, &temp, sizeof(temp)); + *osfs = *temp; } else { - osfs->os_bavail += temp.os_bavail; - osfs->os_blocks += temp.os_blocks; - osfs->os_ffree += temp.os_ffree; - osfs->os_files += temp.os_files; + osfs->os_bavail += temp->os_bavail; + osfs->os_blocks += temp->os_blocks; + osfs->os_ffree += temp->os_ffree; + osfs->os_files += temp->os_files; } } - RETURN(rc); -} -static int lmv_cleanup(struct obd_device *obd, int flags) -{ - struct lmv_obd *lmv = &obd->u.lmv; - ENTRY; - lmv_cleanup_objs(obd); - OBD_FREE(lmv->tgts, lmv->bufsize); - RETURN(0); + EXIT; +out_free_temp: + OBD_FREE(temp, sizeof(*temp)); + return rc; } -static int lmv_getstatus(struct obd_export *exp, struct ll_fid *fid) +static int lmv_getstatus(struct obd_export *exp, + struct lu_fid *fid, + struct obd_capa **pc) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; int rc; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - rc = md_getstatus(lmv->tgts[0].ltd_exp, fid); - fid->mds = 0; + + rc = md_getstatus(lmv->tgts[0].ltd_exp, fid, pc); + RETURN(rc); } -static int lmv_getattr(struct obd_export *exp, struct ll_fid *fid, - unsigned long valid, unsigned int ea_size, - struct ptlrpc_request **request) +static int lmv_getxattr(struct obd_export *exp, const struct lu_fid *fid, + struct obd_capa *oc, obd_valid valid, const char *name, + const char *input, int input_size, int output_size, + int flags, struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - int rc, i = fid->mds; - struct lmv_obj *obj; + struct obd_export *tgt_exp; + int rc; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - obj = lmv_grab_obj(obd, fid, 0); - CDEBUG(D_OTHER, "GETATTR for %lu/%lu/%lu %s\n", - (unsigned long) fid->mds, - (unsigned long) fid->id, - (unsigned long) fid->generation, - obj ? "(splitted)" : ""); - - LASSERT(fid->mds < lmv->desc.ld_tgt_count); - rc = md_getattr(lmv->tgts[i].ltd_exp, fid, - valid, ea_size, request); - if (rc == 0 && obj) { - /* we have to loop over dirobjs here and gather attrs - * for all the slaves */ -#warning "attrs gathering here" - } - lmv_put_obj(obj); + + tgt_exp = lmv_find_export(lmv, fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_getxattr(tgt_exp, fid, oc, valid, name, input, input_size, + output_size, flags, request); + RETURN(rc); } -static int lmv_change_cbdata(struct obd_export *exp, - struct ll_fid *fid, - ldlm_iterator_t it, void *data) +static int lmv_setxattr(struct obd_export *exp, const struct lu_fid *fid, + struct obd_capa *oc, obd_valid valid, const char *name, + const char *input, int input_size, int output_size, + int flags, struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - int rc = 0; + struct obd_export *tgt_exp; + int rc; ENTRY; - + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu\n", - (unsigned long) fid->mds, - (unsigned long) fid->id, - (unsigned long) fid->generation); - LASSERT(fid->mds < lmv->desc.ld_tgt_count); - rc = md_change_cbdata(lmv->tgts[fid->mds].ltd_exp, fid, it, data); + + tgt_exp = lmv_find_export(lmv, fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_setxattr(tgt_exp, fid, oc, valid, name, + input, input_size, output_size, flags, request); + RETURN(rc); } -static int lmv_change_cbdata_name(struct obd_export *exp, struct ll_fid *pfid, - char *name, int len, struct ll_fid *cfid, - ldlm_iterator_t it, void *data) +static int lmv_getattr(struct obd_export *exp, const struct lu_fid *fid, + struct obd_capa *oc, obd_valid valid, int ea_size, + struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; struct lmv_obj *obj; - int rc = 0, mds; + int rc, i; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - LASSERT(pfid->mds < lmv->desc.ld_tgt_count); - LASSERT(cfid->mds < lmv->desc.ld_tgt_count); - CDEBUG(D_OTHER, "CBDATA for %lu/%lu/%lu:%*s -> %lu/%lu/%lu\n", - (unsigned long) pfid->mds, (unsigned long) pfid->id, - (unsigned long) pfid->generation, len, name, - (unsigned long) cfid->mds, (unsigned long) cfid->id, - (unsigned long) cfid->generation); - - /* this is default mds for directory name belongs to */ - mds = pfid->mds; - obj = lmv_grab_obj(obd, pfid, 0); + + tgt_exp = lmv_find_export(lmv, fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_getattr(tgt_exp, fid, oc, valid, ea_size, request); + if (rc) + RETURN(rc); + + obj = lmv_obj_grab(obd, fid); + + CDEBUG(D_OTHER, "GETATTR for "DFID" %s\n", PFID(fid), + obj ? "(split)" : ""); + + /* + * If object is split, then we loop over all the slaves and gather size + * attribute. In ideal world we would have to gather also mds field from + * all slaves, as object is spread over the cluster and this is + * definitely interesting information and it is not good to loss it, + * but... + */ if (obj) { - /* directory is splitted. look for right mds for this name */ - mds = raw_name2idx(obj->objcount, name, len); - lmv_put_obj(obj); + struct mdt_body *body; + + if (*request == NULL) { + lmv_obj_put(obj); + RETURN(rc); + } + + body = lustre_msg_buf((*request)->rq_repmsg, REPLY_REC_OFF, + sizeof(*body)); + LASSERT(body != NULL); + LASSERT(lustre_rep_swabbed(*request, REPLY_REC_OFF)); + + lmv_obj_lock(obj); + + for (i = 0; i < obj->lo_objcount; i++) { + if (lmv->tgts[i].ltd_exp == NULL) { + CWARN("%s: NULL export for %d\n", + obd->obd_name, i); + continue; + } + + /* skip master obj. */ + if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) + continue; + + lmv_update_body(body, &obj->lo_inodes[i]); + } + + lmv_obj_unlock(obj); + lmv_obj_put(obj); } - rc = md_change_cbdata(lmv->tgts[mds].ltd_exp, cfid, it, data); + RETURN(rc); } -static int lmv_valid_attrs(struct obd_export *exp, struct ll_fid *fid) +static int lmv_change_cbdata(struct obd_export *exp, const struct lu_fid *fid, + ldlm_iterator_t it, void *data) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - int rc = 0; + int i, rc; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - CDEBUG(D_OTHER, "validate %lu/%lu/%lu\n", (unsigned long) fid->mds, - (unsigned long) fid->id, (unsigned long) fid->generation); - LASSERT(fid->mds < lmv->desc.ld_tgt_count); - rc = md_valid_attrs(lmv->tgts[fid->mds].ltd_exp, fid); - RETURN(rc); + + CDEBUG(D_OTHER, "CBDATA for "DFID"\n", PFID(fid)); + + /* + * With CMD every object can have two locks in different namespaces: + * lookup lock in space of mds storing direntry and update/open lock in + * space of mds storing inode. + */ + for (i = 0; i < lmv->desc.ld_tgt_count; i++) + md_change_cbdata(lmv->tgts[i].ltd_exp, fid, it, data); + + RETURN(0); } -int lmv_close(struct obd_export *exp, struct obdo *obdo, - struct obd_client_handle *och, - struct ptlrpc_request **request) +static int lmv_close(struct obd_export *exp, + struct md_op_data *op_data, + struct md_open_data *mod, + struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - int rc, i = obdo->o_mds; + struct obd_export *tgt_exp; + int rc; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - LASSERT(i < lmv->desc.ld_tgt_count); - CDEBUG(D_OTHER, "CLOSE %lu/%lu/%lu\n", (unsigned long) obdo->o_mds, - (unsigned long) obdo->o_id, (unsigned long) obdo->o_generation); - rc = md_close(lmv->tgts[i].ltd_exp, obdo, och, request); + + tgt_exp = lmv_find_export(lmv, &op_data->op_fid1); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + CDEBUG(D_OTHER, "CLOSE "DFID"\n", PFID(&op_data->op_fid1)); + rc = md_close(tgt_exp, op_data, mod, request); RETURN(rc); } -int lmv_get_mea_and_update_object(struct obd_export *exp, struct ll_fid *fid) +/* + * Called in the case MDS returns -ERESTART on create on open, what means that + * directory is split and its LMV presentation object has to be updated. + */ +int lmv_handle_split(struct obd_export *exp, const struct lu_fid *fid) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; struct ptlrpc_request *req = NULL; + struct obd_export *tgt_exp; + struct lmv_obj *obj; struct lustre_md md; - unsigned long valid; int mealen, rc; + __u64 valid; + ENTRY; md.mea = NULL; - mealen = MEA_SIZE_LMV(lmv); - - valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA; + mealen = lmv_get_easize(lmv); + + valid = OBD_MD_FLEASIZE | OBD_MD_FLDIREA | OBD_MD_MEA; + + tgt_exp = lmv_find_export(lmv, fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); /* time to update mea of parent fid */ - rc = md_getattr(lmv->tgts[fid->mds].ltd_exp, fid, - valid, mealen, &req); + rc = md_getattr(tgt_exp, fid, NULL, valid, mealen, &req); if (rc) { CERROR("md_getattr() failed, error %d\n", rc); GOTO(cleanup, rc); } - rc = mdc_req2lustre_md(exp, req, 0, NULL, &md); + rc = md_get_lustre_md(tgt_exp, req, 1, NULL, exp, &md); if (rc) { - CERROR("mdc_req2lustre_md() failed, error %d\n", rc); + CERROR("mdc_get_lustre_md() failed, error %d\n", rc); GOTO(cleanup, rc); } if (md.mea == NULL) GOTO(cleanup, rc = -ENODATA); - rc = lmv_create_obj_from_attrs(exp, fid, md.mea); - obd_free_memmd(exp, (struct lov_stripe_md **) &md.mea); + obj = lmv_obj_create(exp, fid, md.mea); + if (IS_ERR(obj)) + rc = PTR_ERR(obj); + else + lmv_obj_put(obj); + + obd_free_memmd(exp, (struct lov_stripe_md **)&md.mea); + EXIT; cleanup: if (req) ptlrpc_req_finished(req); - RETURN(rc); + return rc; } -int lmv_create(struct obd_export *exp, struct mdc_op_data *op_data, +int lmv_create(struct obd_export *exp, struct md_op_data *op_data, const void *data, int datalen, int mode, __u32 uid, - __u32 gid, __u64 rdev, struct ptlrpc_request **request) + __u32 gid, __u32 cap_effective, __u64 rdev, + struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct mds_body *mds_body; + struct obd_export *tgt_exp; struct lmv_obj *obj; - int rc, mds; + int rc, loop = 0; ENTRY; rc = lmv_check_connect(obd); @@ -700,396 +1345,752 @@ int lmv_create(struct obd_export *exp, struct mdc_op_data *op_data, if (!lmv->desc.ld_active_tgt_count) RETURN(-EIO); repeat: - obj = lmv_grab_obj(obd, &op_data->fid1, 0); + ++loop; + LASSERT(loop <= 2); + obj = lmv_obj_grab(obd, &op_data->op_fid1); if (obj) { - mds = raw_name2idx(obj->objcount, op_data->name, - op_data->namelen); - op_data->fid1 = obj->objs[mds].fid; - lmv_put_obj(obj); - } - - CDEBUG(D_OTHER, "CREATE '%*s' on %lu/%lu/%lu\n", - op_data->namelen, op_data->name, - (unsigned long) op_data->fid1.mds, - (unsigned long) op_data->fid1.id, - (unsigned long) op_data->fid1.generation); - rc = md_create(lmv->tgts[op_data->fid1.mds].ltd_exp, op_data, data, - datalen, mode, uid, gid, rdev, request); + int mea_idx; + + mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount, + op_data->op_name, op_data->op_namelen); + op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid; + op_data->op_bias &= ~MDS_CHECK_SPLIT; + op_data->op_mds = obj->lo_inodes[mea_idx].li_mds; + tgt_exp = lmv_get_export(lmv, op_data->op_mds); + lmv_obj_put(obj); + } else { + struct lmv_tgt_desc *tgt; + + tgt = lmv_find_target(lmv, &op_data->op_fid1); + op_data->op_bias |= MDS_CHECK_SPLIT; + op_data->op_mds = tgt->ltd_idx; + tgt_exp = tgt->ltd_exp; + } + + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = lmv_fid_alloc(exp, &op_data->op_fid2, op_data); + if (rc == -ERESTART) + goto repeat; + else if (rc) + RETURN(rc); + + CDEBUG(D_OTHER, "CREATE '%*s' on "DFID"\n", op_data->op_namelen, + op_data->op_name, PFID(&op_data->op_fid1)); + + op_data->op_flags |= MF_MDC_CANCEL_FID1; + rc = md_create(tgt_exp, op_data, data, datalen, mode, uid, gid, + cap_effective, rdev, request); if (rc == 0) { if (*request == NULL) - RETURN(rc); - mds_body = lustre_msg_buf((*request)->rq_repmsg, 0, - sizeof(*mds_body)); - LASSERT(mds_body != NULL); - CDEBUG(D_OTHER, "created. id = %lu, generation = %lu, mds = %d\n", - (unsigned long) mds_body->fid1.id, - (unsigned long) mds_body->fid1.generation, - op_data->fid1.mds); - LASSERT(mds_body->valid & OBD_MD_MDS || - mds_body->mds == op_data->fid1.mds); + RETURN(rc); + CDEBUG(D_OTHER, "created - "DFID"\n", PFID(&op_data->op_fid1)); } else if (rc == -ERESTART) { - /* directory got splitted. time to update local object - * and repeat the request with proper MDS */ - rc = lmv_get_mea_and_update_object(exp, &op_data->fid1); + LASSERT(*request != NULL); + DEBUG_REQ(D_WARNING|D_RPCTRACE, *request, + "Got -ERESTART during create!\n"); + ptlrpc_req_finished(*request); + *request = NULL; + + /* + * Directory got split. Time to update local object and repeat + * the request with proper MDS. + */ + rc = lmv_handle_split(exp, &op_data->op_fid1); if (rc == 0) { - ptlrpc_req_finished(*request); + rc = lmv_alloc_slave_fids(obd, &op_data->op_fid1, + op_data, &op_data->op_fid2); + if (rc) + RETURN(rc); goto repeat; } } RETURN(rc); } -int lmv_done_writing(struct obd_export *exp, struct obdo *obdo) +static int lmv_done_writing(struct obd_export *exp, + struct md_op_data *op_data, + struct md_open_data *mod) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; int rc; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - /* FIXME: choose right MDC here */ - CWARN("this method isn't implemented yet\n"); - rc = md_done_writing(lmv->tgts[0].ltd_exp, obdo); + tgt_exp = lmv_find_export(lmv, &op_data->op_fid1); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_done_writing(tgt_exp, op_data, mod); RETURN(rc); } -int lmv_enqueue_slaves(struct obd_export *exp, int locktype, - struct lookup_intent *it, int lockmode, - struct mdc_op_data *data, struct lustre_handle *lockh, - void *lmm, int lmmsize, - ldlm_completion_callback cb_completion, - ldlm_blocking_callback cb_blocking, void *cb_data) +static int +lmv_enqueue_slaves(struct obd_export *exp, struct ldlm_enqueue_info *einfo, + struct lookup_intent *it, struct md_op_data *op_data, + struct lustre_handle *lockh, void *lmm, int lmmsize) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct mea *mea = data->mea1; - struct mdc_op_data data2; - int i, rc, mds; + struct lmv_stripe_md *mea = op_data->op_mea1; + struct md_op_data *op_data2; + struct obd_export *tgt_exp; + int i, rc = 0; ENTRY; + OBD_ALLOC_PTR(op_data2); + if (op_data2 == NULL) + RETURN(-ENOMEM); + LASSERT(mea != NULL); for (i = 0; i < mea->mea_count; i++) { - if (lmv->tgts[i].ltd_exp == NULL) + memset(op_data2, 0, sizeof(*op_data2)); + op_data2->op_fid1 = mea->mea_ids[i]; + op_data2->op_bias = 0; + + tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1); + if (IS_ERR(tgt_exp)) + GOTO(cleanup, rc = PTR_ERR(tgt_exp)); + + if (tgt_exp == NULL) continue; - memset(&data2, 0, sizeof(data2)); - data2.fid1 = mea->mea_fids[i]; - mds = data2.fid1.mds; - rc = md_enqueue(lmv->tgts[mds].ltd_exp, locktype, it, lockmode, - &data2, lockh + i, lmm, lmmsize, cb_completion, - cb_blocking, cb_data); - CDEBUG(D_OTHER, "take lock on slave %lu/%lu/%lu -> %d/%d\n", - (unsigned long) mea->mea_fids[i].mds, - (unsigned long) mea->mea_fids[i].id, - (unsigned long) mea->mea_fids[i].generation, - rc, it->d.lustre.it_status); + rc = md_enqueue(tgt_exp, einfo, it, op_data2, + lockh + i, lmm, lmmsize, 0); + + CDEBUG(D_OTHER, "take lock on slave "DFID" -> %d/%d\n", + PFID(&mea->mea_ids[i]), rc, it->d.lustre.it_status); + if (rc) GOTO(cleanup, rc); + if (it->d.lustre.it_data) { struct ptlrpc_request *req; - req = (struct ptlrpc_request *) it->d.lustre.it_data; + req = (struct ptlrpc_request *)it->d.lustre.it_data; ptlrpc_req_finished(req); } - + if (it->d.lustre.it_status) GOTO(cleanup, rc = it->d.lustre.it_status); } - RETURN(0); - + + EXIT; cleanup: - /* drop all taken locks */ - while (--i >= 0) { - if (lockh[i].cookie) - ldlm_lock_decref(lockh + i, lockmode); - lockh[i].cookie = 0; + OBD_FREE_PTR(op_data2); + + if (rc != 0) { + /* drop all taken locks */ + while (--i >= 0) { + if (lockh[i].cookie) + ldlm_lock_decref(lockh + i, einfo->ei_mode); + lockh[i].cookie = 0; + } } - RETURN(rc); + return rc; +} + +static int +lmv_enqueue_remote(struct obd_export *exp, struct ldlm_enqueue_info *einfo, + struct lookup_intent *it, struct md_op_data *op_data, + struct lustre_handle *lockh, void *lmm, int lmmsize, + int extra_lock_flags) +{ + struct ptlrpc_request *req = it->d.lustre.it_data; + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + struct lustre_handle plock; + struct obd_export *tgt_exp; + struct md_op_data *rdata; + struct lu_fid fid_copy; + struct mdt_body *body; + int rc = 0, pmode; + ENTRY; + + body = lustre_msg_buf(req->rq_repmsg, + DLM_REPLY_REC_OFF, sizeof(*body)); + LASSERT(body != NULL); + LASSERT(lustre_rep_swabbed(req, DLM_REPLY_REC_OFF)); + + if (!(body->valid & OBD_MD_MDS)) + RETURN(0); + + CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID" -> "DFID"\n", + LL_IT2STR(it), PFID(&op_data->op_fid1), PFID(&body->fid1)); + + /* We got LOOKUP lock, but we really need attrs */ + pmode = it->d.lustre.it_lock_mode; + LASSERT(pmode != 0); + memcpy(&plock, lockh, sizeof(plock)); + it->d.lustre.it_lock_mode = 0; + it->d.lustre.it_data = NULL; + fid_copy = body->fid1; + + it->d.lustre.it_disposition &= ~DISP_ENQ_COMPLETE; + ptlrpc_req_finished(req); + + tgt_exp = lmv_find_export(lmv, &fid_copy); + if (IS_ERR(tgt_exp)) + GOTO(out, rc = PTR_ERR(tgt_exp)); + + OBD_ALLOC_PTR(rdata); + if (rdata == NULL) + GOTO(out, rc = -ENOMEM); + + rdata->op_fid1 = fid_copy; + rdata->op_bias = MDS_CROSS_REF; + + rc = md_enqueue(tgt_exp, einfo, it, rdata, lockh, + lmm, lmmsize, extra_lock_flags); + OBD_FREE_PTR(rdata); + EXIT; +out: + ldlm_lock_decref(&plock, pmode); + return rc; } -int lmv_enqueue(struct obd_export *exp, int lock_type, - struct lookup_intent *it, int lock_mode, - struct mdc_op_data *data, struct lustre_handle *lockh, - void *lmm, int lmmsize, - ldlm_completion_callback cb_completion, - ldlm_blocking_callback cb_blocking, void *cb_data) +static int +lmv_enqueue(struct obd_export *exp, struct ldlm_enqueue_info *einfo, + struct lookup_intent *it, struct md_op_data *op_data, + struct lustre_handle *lockh, void *lmm, int lmmsize, + int extra_lock_flags) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp = NULL; struct lmv_obj *obj; - int rc, mds; + int rc; ENTRY; rc = lmv_check_connect(obd); if (rc) RETURN(rc); - if (it->it_op == IT_UNLINK) { - rc = lmv_enqueue_slaves(exp, lock_type, it, lock_mode, - data, lockh, lmm, lmmsize, - cb_completion, cb_blocking, cb_data); + if (op_data->op_mea1 && it->it_op == IT_UNLINK) { + rc = lmv_enqueue_slaves(exp, einfo, it, op_data, + lockh, lmm, lmmsize); RETURN(rc); } - if (data->namelen) { - obj = lmv_grab_obj(obd, &data->fid1, 0); + if (op_data->op_namelen) { + obj = lmv_obj_grab(obd, &op_data->op_fid1); if (obj) { - /* directory is splitted. look for - * right mds for this name */ - mds = raw_name2idx(obj->objcount, (char *)data->name, - data->namelen); - data->fid1 = obj->objs[mds].fid; - lmv_put_obj(obj); + int mea_idx; + + /* directory is split. look for right mds for this + * name */ + mea_idx = raw_name2idx(obj->lo_hashtype, + obj->lo_objcount, + (char *)op_data->op_name, + op_data->op_namelen); + op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid; + tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds); + lmv_obj_put(obj); } } - CDEBUG(D_OTHER, "ENQUEUE '%s' on %lu/%lu\n", - LL_IT2STR(it), (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation); - rc = md_enqueue(lmv->tgts[data->fid1.mds].ltd_exp, lock_type, it, - lock_mode, data, lockh, lmm, lmmsize, cb_completion, - cb_blocking, cb_data); + if (tgt_exp == NULL) + tgt_exp = lmv_find_export(lmv, &op_data->op_fid1); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + CDEBUG(D_OTHER, "ENQUEUE '%s' on "DFID"\n", LL_IT2STR(it), + PFID(&op_data->op_fid1)); + + rc = md_enqueue(tgt_exp, einfo, it, op_data, lockh, + lmm, lmmsize, extra_lock_flags); + + if (rc == 0 && it->it_op == IT_OPEN) + rc = lmv_enqueue_remote(exp, einfo, it, op_data, lockh, + lmm, lmmsize, extra_lock_flags); RETURN(rc); } -int lmv_getattr_name(struct obd_export *exp, struct ll_fid *fid, - char *filename, int namelen, unsigned long valid, - unsigned int ea_size, struct ptlrpc_request **request) +static int +lmv_getattr_name(struct obd_export *exp, const struct lu_fid *fid, + struct obd_capa *oc, const char *filename, int namelen, + obd_valid valid, int ea_size, struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct ll_fid rfid = *fid; - int rc, mds = fid->mds; - struct mds_body *body; + struct lu_fid rid = *fid; + struct obd_export *tgt_exp; + struct mdt_body *body; struct lmv_obj *obj; + int rc, loop = 0; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); + repeat: - obj = lmv_grab_obj(obd, fid, 0); + ++loop; + LASSERT(loop <= 2); + obj = lmv_obj_grab(obd, &rid); if (obj) { - /* directory is splitted. look for right mds for this name */ - mds = raw_name2idx(obj->objcount, filename, namelen - 1); - rfid = obj->objs[mds].fid; - lmv_put_obj(obj); - } - CDEBUG(D_OTHER, "getattr_name for %*s on %lu/%lu/%lu -> %lu/%lu/%lu\n", - namelen, filename, (unsigned long) fid->mds, - (unsigned long) fid->id, (unsigned long) fid->generation, - (unsigned long) rfid.mds, (unsigned long) rfid.id, - (unsigned long) rfid.generation); - rc = md_getattr_name(lmv->tgts[mds].ltd_exp, &rfid, filename, namelen, - valid, ea_size, request); + int mea_idx; + + /* Directory is split. Look for right mds for this name */ + mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount, + filename, namelen - 1); + rid = obj->lo_inodes[mea_idx].li_fid; + tgt_exp = lmv_get_export(lmv, obj->lo_inodes[mea_idx].li_mds); + lmv_obj_put(obj); + valid &= ~OBD_MD_FLCKSPLIT; + } else { + tgt_exp = lmv_find_export(lmv, &rid); + valid |= OBD_MD_FLCKSPLIT; + } + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + CDEBUG(D_OTHER, "getattr_name for %*s on "DFID" -> "DFID"\n", + namelen, filename, PFID(fid), PFID(&rid)); + + rc = md_getattr_name(tgt_exp, &rid, oc, filename, namelen, valid, + ea_size, request); if (rc == 0) { - /* this could be cross-node reference. in this case all - * we have right now is mds/ino/generation triple. we'd - * like to find other attributes */ - body = lustre_msg_buf((*request)->rq_repmsg, 0, sizeof(*body)); + body = lustre_msg_buf((*request)->rq_repmsg, + REQ_REC_OFF, sizeof(*body)); LASSERT(body != NULL); + LASSERT(lustre_rep_swabbed(*request, REQ_REC_OFF)); + if (body->valid & OBD_MD_MDS) { struct ptlrpc_request *req = NULL; - rfid = body->fid1; - CDEBUG(D_OTHER, "request attrs for %lu/%lu/%lu\n", - (unsigned long) rfid.mds, - (unsigned long) rfid.id, - (unsigned long) rfid.generation); - rc = md_getattr_name(lmv->tgts[rfid.mds].ltd_exp, &rfid, - NULL, 1, valid, ea_size, &req); + + rid = body->fid1; + CDEBUG(D_OTHER, "request attrs for "DFID"\n", + PFID(&rid)); + + tgt_exp = lmv_find_export(lmv, &rid); + if (IS_ERR(tgt_exp)) { + ptlrpc_req_finished(*request); + RETURN(PTR_ERR(tgt_exp)); + } + + rc = md_getattr_name(tgt_exp, &rid, NULL, NULL, 1, + valid | OBD_MD_FLCROSSREF, + ea_size, &req); ptlrpc_req_finished(*request); *request = req; } } else if (rc == -ERESTART) { - /* directory got splitted. time to update local object - * and repeat the request with proper MDS */ - rc = lmv_get_mea_and_update_object(exp, &rfid); - if (rc == 0) { - ptlrpc_req_finished(*request); + LASSERT(*request != NULL); + DEBUG_REQ(D_WARNING|D_RPCTRACE, *request, + "Got -ERESTART during getattr!\n"); + ptlrpc_req_finished(*request); + *request = NULL; + + /* + * Directory got split. Time to update local object and repeat + * the request with proper MDS. + */ + rc = lmv_handle_split(exp, &rid); + if (rc == 0) goto repeat; - } } RETURN(rc); } +#define md_op_data_fid(op_data, fl) \ + (fl == MF_MDC_CANCEL_FID1 ? &op_data->op_fid1 : \ + fl == MF_MDC_CANCEL_FID2 ? &op_data->op_fid2 : \ + fl == MF_MDC_CANCEL_FID3 ? &op_data->op_fid3 : \ + fl == MF_MDC_CANCEL_FID4 ? &op_data->op_fid4 : \ + NULL) + +/* @tgt_exp is the export the metadata request is sent. + * @fid_exp is the export the cancel should be sent for the current fid. + * if @fid_exp is NULL, the export is found for the current fid. + * @op_data keeps the current fid, which is pointed through @flag. + * @mode, @bits -- lock match parameters. */ +static int lmv_early_cancel(struct lmv_obd *lmv, struct obd_export *tgt_exp, + struct obd_export *fid_exp, + struct md_op_data *op_data, + ldlm_mode_t mode, int bits, int flag) +{ + struct lu_fid *fid = md_op_data_fid(op_data, flag); + ldlm_policy_data_t policy = {{0}}; + int rc = 0; + ENTRY; + + if (!fid_is_sane(fid)) + RETURN(0); + + if (fid_exp == NULL) + fid_exp = lmv_find_export(lmv, fid); + + if (tgt_exp == fid_exp) { + /* The export is the same as on the target server, cancel + * will be sent along with the main metadata operation. */ + op_data->op_flags |= flag; + RETURN(0); + } + + policy.l_inodebits.bits = bits; + rc = md_cancel_unused(fid_exp, fid, &policy, mode, LDLM_FL_ASYNC, NULL); + RETURN(rc); +} + +#ifdef EARLY_CANCEL_FOR_STRIPED_DIR_IS_READY +/* Check if the fid in @op_data pointed to by flag is of the same export(s) + * as @tgt_exp. Early cancels will be sent later by mdc code, otherwise, call + * md_cancel_unused for child export(s). */ +static int lmv_early_cancel_stripes(struct obd_export *exp, + struct obd_export *tgt_exp, + struct md_op_data *op_data, + ldlm_mode_t mode, int bits, int flag) +{ + struct lu_fid *fid = md_op_data_fid(op_data, flag); + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *st_exp; + struct lmv_obj *obj; + int rc = 0; + ENTRY; + + if (!fid_is_sane(fid)) + RETURN(0); + + obj = lmv_obj_grab(obd, fid); + if (obj) { + ldlm_policy_data_t policy = {{0}}; + struct lu_fid *st_fid; + int i; + + policy.l_inodebits.bits = bits; + for (i = 0; i < obj->lo_objcount; i++) { + st_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds); + st_fid = &obj->lo_inodes[i].li_fid; + if (tgt_exp != st_exp) { + rc = md_cancel_unused(st_exp, st_fid, &policy, + mode, LDLM_FL_ASYNC, + NULL); + if (rc) + break; + } else { + /* Some export matches to @tgt_exp, do cancel + * for its fid in mdc */ + *fid = *st_fid; + op_data->op_flags |= flag; + } + } + lmv_obj_put(obj); + } else { + rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, + mode, bits, flag); + } + RETURN(rc); +} +#endif /* - * llite passes fid of an target inode in data->fid1 and - * fid of directory in data->fid2 + * llite passes fid of an target inode in op_data->op_fid1 and id of directory in + * op_data->op_fid2 */ -int lmv_link(struct obd_export *exp, struct mdc_op_data *data, - struct ptlrpc_request **request) +static int lmv_link(struct obd_export *exp, struct md_op_data *op_data, + struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; struct lmv_obj *obj; - int rc; + int rc, loop = 0; + mdsno_t mds; ENTRY; + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - if (data->namelen != 0) { - /* usual link request */ - obj = lmv_grab_obj(obd, &data->fid1, 0); + +repeat: + ++loop; + LASSERT(loop <= 2); + if (op_data->op_namelen != 0) { + int mea_idx; + + /* Usual link request */ + obj = lmv_obj_grab(obd, &op_data->op_fid2); if (obj) { - rc = raw_name2idx(obj->objcount, data->name, - data->namelen); - data->fid1 = obj->objs[rc].fid; - lmv_put_obj(obj); + mea_idx = raw_name2idx(obj->lo_hashtype, + obj->lo_objcount, + op_data->op_name, + op_data->op_namelen); + op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid; + mds = obj->lo_inodes[mea_idx].li_mds; + lmv_obj_put(obj); + } else { + rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds); + if (rc) + RETURN(rc); } - CDEBUG(D_OTHER,"link %u/%u/%u:%*s to %u/%u/%u mds %d\n", - (unsigned) data->fid2.mds, (unsigned) data->fid2.id, - (unsigned) data->fid2.generation, data->namelen, - data->name, (unsigned) data->fid1.mds, - (unsigned) data->fid1.id, - (unsigned) data->fid1.generation, data->fid1.mds); + + CDEBUG(D_OTHER,"link "DFID":%*s to "DFID"\n", + PFID(&op_data->op_fid2), op_data->op_namelen, + op_data->op_name, PFID(&op_data->op_fid1)); } else { + rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds); + if (rc) + RETURN(rc); + /* request from MDS to acquire i_links for inode by fid1 */ - CDEBUG(D_OTHER, "inc i_nlinks for %u/%u/%u\n", - (unsigned) data->fid1.mds, (unsigned) data->fid1.id, - (unsigned) data->fid1.generation); + CDEBUG(D_OTHER, "inc i_nlinks for "DFID"\n", + PFID(&op_data->op_fid1)); + } + + CDEBUG(D_OTHER, "forward to MDS #"LPU64" ("DFID")\n", + mds, PFID(&op_data->op_fid1)); + + op_data->op_fsuid = current->fsuid; + op_data->op_fsgid = current->fsgid; + op_data->op_cap = current->cap_effective; + + tgt_exp = lmv->tgts[mds].ltd_exp; + if (op_data->op_namelen) { + op_data->op_flags |= MF_MDC_CANCEL_FID2; + /* Cancel UPDATE lock on child (fid1). */ + rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX, + MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID1); + } + if (rc == 0) + rc = md_link(tgt_exp, op_data, request); + if (rc == -ERESTART) { + LASSERT(*request != NULL); + DEBUG_REQ(D_WARNING|D_RPCTRACE, *request, + "Got -ERESTART during link!\n"); + ptlrpc_req_finished(*request); + *request = NULL; + + /* + * Directory got split. Time to update local object and repeat + * the request with proper MDS. + */ + rc = lmv_handle_split(exp, &op_data->op_fid2); + if (rc == 0) + goto repeat; } - - rc = md_link(lmv->tgts[data->fid1.mds].ltd_exp, data, request); + RETURN(rc); } -int lmv_rename(struct obd_export *exp, struct mdc_op_data *data, - const char *old, int oldlen, const char *new, int newlen, - struct ptlrpc_request **request) +static int lmv_rename(struct obd_export *exp, struct md_op_data *op_data, + const char *old, int oldlen, const char *new, int newlen, + struct ptlrpc_request **request) { + struct obd_export *tgt_exp = NULL, *src_exp; struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; + int rc, mea_idx, loop = 0; struct lmv_obj *obj; - int rc, mds; + mdsno_t mds1, mds2; ENTRY; - CDEBUG(D_OTHER, "rename %*s in %lu/%lu/%lu to %*s in %lu/%lu/%lu\n", - oldlen, old, (unsigned long) data->fid1.mds, - (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation, - newlen, new, (unsigned long) data->fid2.mds, - (unsigned long) data->fid2.id, - (unsigned long) data->fid2.generation); - if (!fid_equal(&data->fid1, &data->fid2)) - CWARN("cross-node rename %lu/%lu/%lu:%*s to %lu/%lu/%lu:%*s\n", - (unsigned long) data->fid1.mds, - (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation, oldlen, old, - (unsigned long) data->fid2.mds, - (unsigned long) data->fid2.id, - (unsigned long) data->fid2.generation, newlen, new); + CDEBUG(D_OTHER, "rename %*s in "DFID" to %*s in "DFID"\n", + oldlen, old, PFID(&op_data->op_fid1), + newlen, new, PFID(&op_data->op_fid2)); rc = lmv_check_connect(obd); if (rc) RETURN(rc); if (oldlen == 0) { - /* MDS with old dir entry is asking another MDS - * to create name there */ + /* + * MDS with old dir entry is asking another MDS to create name + * there. + */ CDEBUG(D_OTHER, - "create %*s(%d/%d) in %lu/%lu/%lu pointing to %lu/%lu/%lu\n", - newlen, new, oldlen, newlen, - (unsigned long) data->fid2.mds, - (unsigned long) data->fid2.id, - (unsigned long) data->fid2.generation, - (unsigned long) data->fid1.mds, - (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation); - mds = data->fid2.mds; + "create %*s(%d/%d) in "DFID" pointing " + "to "DFID"\n", newlen, new, oldlen, newlen, + PFID(&op_data->op_fid2), PFID(&op_data->op_fid1)); + + rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds1); + if (rc) + RETURN(rc); + + /* + * Target directory can be split, sowe should forward request to + * the right MDS. + */ + obj = lmv_obj_grab(obd, &op_data->op_fid2); + if (obj) { + mea_idx = raw_name2idx(obj->lo_hashtype, + obj->lo_objcount, + (char *)new, newlen); + op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid; + CDEBUG(D_OTHER, "Parent obj "DFID"\n", + PFID(&op_data->op_fid2)); + lmv_obj_put(obj); + } goto request; } - obj = lmv_grab_obj(obd, &data->fid1, 0); +repeat: + ++loop; + LASSERT(loop <= 2); + obj = lmv_obj_grab(obd, &op_data->op_fid1); if (obj) { - /* directory is already splitted, so we have to forward - * request to the right MDS */ - mds = raw_name2idx(obj->objcount, (char *)old, oldlen); - data->fid1 = obj->objs[mds].fid; - CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds, - (unsigned long) obj->objs[mds].fid.mds, - (unsigned long) obj->objs[mds].fid.id, - (unsigned long) obj->objs[mds].fid.generation); - } - lmv_put_obj(obj); - - obj = lmv_grab_obj(obd, &data->fid2, 0); + /* + * directory is already split, so we have to forward request to + * the right MDS. + */ + mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount, + (char *)old, oldlen); + op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid; + mds1 = obj->lo_inodes[mea_idx].li_mds; + CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid1)); + lmv_obj_put(obj); + } else { + rc = lmv_fld_lookup(lmv, &op_data->op_fid1, &mds1); + if (rc) + RETURN(rc); + } + + obj = lmv_obj_grab(obd, &op_data->op_fid2); if (obj) { - /* directory is already splitted, so we have to forward - * request to the right MDS */ - mds = raw_name2idx(obj->objcount, (char *)new, newlen); - data->fid2 = obj->objs[mds].fid; - CDEBUG(D_OTHER, "forward to MDS #%u (%lu/%lu/%lu)\n", mds, - (unsigned long) obj->objs[mds].fid.mds, - (unsigned long) obj->objs[mds].fid.id, - (unsigned long) obj->objs[mds].fid.generation); - } - lmv_put_obj(obj); - - mds = data->fid1.mds; + /* + * Directory is already split, so we have to forward request to + * the right MDS. + */ + mea_idx = raw_name2idx(obj->lo_hashtype, obj->lo_objcount, + (char *)new, newlen); + + mds2 = obj->lo_inodes[mea_idx].li_mds; + op_data->op_fid2 = obj->lo_inodes[mea_idx].li_fid; + CDEBUG(D_OTHER, "Parent obj "DFID"\n", PFID(&op_data->op_fid2)); + lmv_obj_put(obj); + } else { + rc = lmv_fld_lookup(lmv, &op_data->op_fid2, &mds2); + if (rc) + RETURN(rc); + } request: - rc = md_rename(lmv->tgts[mds].ltd_exp, data, old, oldlen, - new, newlen, request); + op_data->op_fsuid = current->fsuid; + op_data->op_fsgid = current->fsgid; + op_data->op_cap = current->cap_effective; + + src_exp = lmv_get_export(lmv, mds1); + tgt_exp = lmv_get_export(lmv, mds2); + if (oldlen) { + /* LOOKUP lock on src child (fid3) should also be cancelled for + * src_exp in mdc_rename. */ + op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3; + + /* Cancel UPDATE locks on tgt parent (fid2), tgt_exp is its + * own export. */ + rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, LCK_EX, + MDS_INODELOCK_UPDATE, MF_MDC_CANCEL_FID2); + + /* Cancel LOOKUP locks on tgt child (fid4) for parent tgt_exp.*/ + if (rc == 0) + rc = lmv_early_cancel(lmv, src_exp, tgt_exp, op_data, + LCK_EX, MDS_INODELOCK_LOOKUP, + MF_MDC_CANCEL_FID4); + + /* XXX: the case when child is a striped dir is not supported. + * Only the master stripe has all locks cancelled early. */ + /* Cancel all the locks on tgt child (fid4). */ + if (rc == 0) + rc = lmv_early_cancel(lmv, src_exp, NULL, op_data, + LCK_EX, MDS_INODELOCK_FULL, + MF_MDC_CANCEL_FID4); + } + + if (rc == 0) + rc = md_rename(src_exp, op_data, old, oldlen, + new, newlen, request); + if (rc == -ERESTART) { + LASSERT(*request != NULL); + DEBUG_REQ(D_WARNING|D_RPCTRACE, *request, + "Got -ERESTART during rename!\n"); + ptlrpc_req_finished(*request); + *request = NULL; + + /* + * Directory got split. Time to update local object and repeat + * the request with proper MDS. + */ + rc = lmv_handle_split(exp, &op_data->op_fid1); + if (rc == 0) + goto repeat; + } RETURN(rc); } -int lmv_setattr(struct obd_export *exp, struct mdc_op_data *data, - struct iattr *iattr, void *ea, int ealen, void *ea2, int ea2len, - struct ptlrpc_request **request) +static int lmv_setattr(struct obd_export *exp, struct md_op_data *op_data, + void *ea, int ealen, void *ea2, int ea2len, + struct ptlrpc_request **request, + struct md_open_data **mod) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - int rc = 0, i = data->fid1.mds; struct ptlrpc_request *req; - struct mds_body *mds_body; + struct obd_export *tgt_exp; struct lmv_obj *obj; + int rc = 0, i; ENTRY; rc = lmv_check_connect(obd); if (rc) RETURN(rc); - obj = lmv_grab_obj(obd, &data->fid1, 0); - CDEBUG(D_OTHER, "SETATTR for %lu/%lu/%lu, valid 0x%x%s\n", - (unsigned long) data->fid1.mds, - (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation, iattr->ia_valid, - obj ? ", splitted" : ""); + obj = lmv_obj_grab(obd, &op_data->op_fid1); + + CDEBUG(D_OTHER, "SETATTR for "DFID", valid 0x%x%s\n", + PFID(&op_data->op_fid1), op_data->op_attr.ia_valid, + obj ? ", split" : ""); + + op_data->op_flags |= MF_MDC_CANCEL_FID1; if (obj) { - for (i = 0; i < obj->objcount; i++) { - data->fid1 = obj->objs[i].fid; - rc = md_setattr(lmv->tgts[i].ltd_exp, data, iattr, ea, - ealen, ea2, ea2len, &req); - LASSERT(rc == 0); - if (fid_equal(&obj->fid, &obj->objs[i].fid)) { - /* this is master object and this request - * should be returned back to llite */ + for (i = 0; i < obj->lo_objcount; i++) { + op_data->op_fid1 = obj->lo_inodes[i].li_fid; + + tgt_exp = lmv_get_export(lmv, obj->lo_inodes[i].li_mds); + if (IS_ERR(tgt_exp)) { + rc = PTR_ERR(tgt_exp); + break; + } + + rc = md_setattr(tgt_exp, op_data, ea, ealen, + ea2, ea2len, &req, mod); + + if (lu_fid_eq(&obj->lo_fid, &obj->lo_inodes[i].li_fid)) { + /* + * this is master object and this request should + * be returned back to llite. + */ *request = req; } else { ptlrpc_req_finished(req); } + + if (rc) + break; } - lmv_put_obj(obj); + lmv_obj_put(obj); } else { - LASSERT(data->fid1.mds < lmv->desc.ld_tgt_count); - rc = md_setattr(lmv->tgts[i].ltd_exp, data, iattr, ea, ealen, - ea2, ea2len, request); - if (rc == 0) { - mds_body = lustre_msg_buf((*request)->rq_repmsg, 0, - sizeof(*mds_body)); - LASSERT(mds_body != NULL); - LASSERT(mds_body->mds == i); - } + tgt_exp = lmv_find_export(lmv, &op_data->op_fid1); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_setattr(tgt_exp, op_data, ea, ealen, ea2, + ea2len, request, mod); } RETURN(rc); } -int lmv_sync(struct obd_export *exp, struct ll_fid *fid, - struct ptlrpc_request **request) +static int lmv_sync(struct obd_export *exp, const struct lu_fid *fid, + struct obd_capa *oc, struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; int rc; ENTRY; @@ -1097,12 +2098,19 @@ int lmv_sync(struct obd_export *exp, struct ll_fid *fid, if (rc) RETURN(rc); - rc = md_sync(lmv->tgts[0].ltd_exp, fid, request); + tgt_exp = lmv_find_export(lmv, fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_sync(tgt_exp, fid, oc, request); RETURN(rc); } -int lmv_dirobj_blocking_ast(struct ldlm_lock *lock, - struct ldlm_lock_desc *desc, void *data, int flag) +/* main purpose of LMV blocking ast is to remove split directory LMV + * presentation object (struct lmv_obj) attached to the lock being revoked. */ +int lmv_blocking_ast(struct ldlm_lock *lock, + struct ldlm_lock_desc *desc, + void *data, int flag) { struct lustre_handle lockh; struct lmv_obj *obj; @@ -1121,17 +2129,16 @@ int lmv_dirobj_blocking_ast(struct ldlm_lock *lock, case LDLM_CB_CANCELING: /* time to drop cached attrs for dirobj */ obj = lock->l_ast_data; - if (!obj) - break; - - CDEBUG(D_OTHER, "cancel %s on %lu/%lu, master %lu/%lu/%lu\n", - lock->l_resource->lr_name.name[3] == 1 ? - "LOOKUP" : "UPDATE", - (unsigned long) lock->l_resource->lr_name.name[0], - (unsigned long) lock->l_resource->lr_name.name[1], - (unsigned long) obj->fid.mds, - (unsigned long) obj->fid.id, - (unsigned long) obj->fid.generation); + if (obj) { + CDEBUG(D_OTHER, "cancel %s on "LPU64"/"LPU64 + ", master "DFID"\n", + lock->l_resource->lr_name.name[3] == 1 ? + "LOOKUP" : "UPDATE", + lock->l_resource->lr_name.name[0], + lock->l_resource->lr_name.name[1], + PFID(&obj->lo_fid)); + lmv_obj_put(obj); + } break; default: LBUG(); @@ -1139,303 +2146,337 @@ int lmv_dirobj_blocking_ast(struct ldlm_lock *lock, RETURN(0); } -void lmv_remove_dots(struct page *page) +static void lmv_hash_adjust(__u32 *hash, __u32 hash_adj) { - char *kaddr = page_address(page); - unsigned limit = PAGE_CACHE_SIZE; - unsigned offs, rec_len; - struct ext2_dir_entry_2 *p; + __u32 val; - for (offs = 0; offs <= limit - EXT2_DIR_REC_LEN(1); offs += rec_len) { - p = (struct ext2_dir_entry_2 *)(kaddr + offs); - rec_len = le16_to_cpu(p->rec_len); + val = le32_to_cpu(*hash); + if (val < hash_adj) + val += MAX_HASH_SIZE; + if (val != DIR_END_OFF) + *hash = cpu_to_le32(val - hash_adj); +} - if ((p->name_len == 1 && p->name[0] == '.') || - (p->name_len == 2 && p->name[0] == '.' && p->name[1] == '.')) - p->inode = 0; - } +static __u32 lmv_node_rank(struct obd_export *exp, const struct lu_fid *fid) +{ + __u64 id; + struct obd_import *imp; + + /* + * XXX Hack: to get nid we assume that underlying obd device is mdc. + */ + imp = class_exp2cliimp(exp); + id = imp->imp_connection->c_self + fid_flatten(fid); + + CDEBUG(D_INFO, "node rank: %llx "DFID" %llx %llx\n", + imp->imp_connection->c_self, PFID(fid), id, id ^ (id >> 32)); + + return id ^ (id >> 32); } -int lmv_readpage(struct obd_export *exp, struct ll_fid *mdc_fid, - __u64 offset, struct page *page, - struct ptlrpc_request **request) +static int lmv_readpage(struct obd_export *exp, const struct lu_fid *fid, + struct obd_capa *oc, __u64 offset64, struct page *page, + struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct ll_fid rfid = *mdc_fid; + struct obd_export *tgt_exp; + struct lu_fid rid = *fid; struct lmv_obj *obj; - int rc, i; + __u32 offset0; + __u32 offset; + __u32 hash_adj = 0; + __u32 rank = 0; + __u32 seg_size = 0; + int tgt = 0; + int tgt0 = 0; + int rc; + int nr = 0; ENTRY; + offset0 = offset = offset64; + /* + * Check that offset is representable by 32bit number. + */ + LASSERT((__u64)offset == offset64); + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - LASSERT(mdc_fid->mds < lmv->desc.ld_tgt_count); - CDEBUG(D_OTHER, "READPAGE at %llu from %lu/%lu/%lu\n", - offset, (unsigned long) rfid.mds, - (unsigned long) rfid.id, - (unsigned long) rfid.generation); + CDEBUG(D_INFO, "READPAGE at %x from "DFID"\n", offset, PFID(&rid)); - obj = lmv_grab_obj(obd, mdc_fid, 0); + obj = lmv_obj_grab(obd, fid); if (obj) { - /* find dirobj containing page with requested offset */ - /* FIXME: what about protecting cached attrs here? */ - for (i = 0; i < obj->objcount; i++) { - if (offset < obj->objs[i].size) - break; - offset -= obj->objs[i].size; - } - rfid = obj->objs[i].fid; - CDEBUG(D_OTHER, "forward to %lu/%lu/%lu with offset %lu\n", - (unsigned long) rfid.mds, - (unsigned long) rfid.id, - (unsigned long) rfid.generation, - (unsigned long) offset); - } - rc = md_readpage(lmv->tgts[rfid.mds].ltd_exp, &rfid, offset, page, request); - if (rc == 0 && !fid_equal(&rfid, mdc_fid)) { - /* this page isn't from master object. to avoid - * ./.. duplication in directory, we have to remove them - * from all slave objects */ - lmv_remove_dots(page); - } - - lmv_put_obj(obj); + struct lmv_inode *loi; + + lmv_obj_lock(obj); + + nr = obj->lo_objcount; + LASSERT(nr > 0); + seg_size = MAX_HASH_SIZE / nr; + loi = obj->lo_inodes; + rank = lmv_node_rank(lmv_get_export(lmv, loi[0].li_mds), + fid) % nr; + tgt0 = (offset / seg_size) % nr; + tgt = (tgt0 + rank) % nr; + + if (tgt < tgt0) + /* + * Wrap around. + * + * Last segment has unusual length due to division + * rounding. + */ + hash_adj = MAX_HASH_SIZE - seg_size * nr; + else + hash_adj = 0; - RETURN(rc); + hash_adj += rank * seg_size; + + CDEBUG(D_INFO, "hash_adj: %x %x %x/%x -> %x/%x\n", + rank, hash_adj, offset, tgt0, offset + hash_adj, tgt); + + offset = (offset + hash_adj) % MAX_HASH_SIZE; + rid = obj->lo_inodes[tgt].li_fid; + tgt_exp = lmv_get_export(lmv, loi[tgt].li_mds); + + CDEBUG(D_INFO, "forward to "DFID" with offset %lu i %d\n", + PFID(&rid), (unsigned long)offset, tgt); + } else + tgt_exp = lmv_find_export(lmv, &rid); + + if (IS_ERR(tgt_exp)) + GOTO(cleanup, rc = PTR_ERR(tgt_exp)); + + rc = md_readpage(tgt_exp, &rid, oc, offset, page, request); + if (rc) + GOTO(cleanup, rc); + if (obj) { + struct lu_dirpage *dp; + struct lu_dirent *ent; + + dp = cfs_kmap(page); + + lmv_hash_adjust(&dp->ldp_hash_start, hash_adj); + lmv_hash_adjust(&dp->ldp_hash_end, hash_adj); + LASSERT(cpu_to_le32(dp->ldp_hash_start) <= offset0); + + for (ent = lu_dirent_start(dp); ent != NULL; + ent = lu_dirent_next(ent)) + lmv_hash_adjust(&ent->lde_hash, hash_adj); + + if (tgt0 != nr - 1) { + __u32 end; + + end = le32_to_cpu(dp->ldp_hash_end); + if (end == DIR_END_OFF) { + dp->ldp_hash_end = cpu_to_le32(seg_size * + (tgt0 + 1)); + CDEBUG(D_INFO, ""DFID" reset end %x tgt %d\n", + PFID(&rid), + le32_to_cpu(dp->ldp_hash_end), tgt); + } + } + cfs_kunmap(page); + } + /* + * Here we could remove "." and ".." from all pages which at not from + * master. But MDS has only "." and ".." for master dir. + */ + EXIT; +cleanup: + if (obj) { + lmv_obj_unlock(obj); + lmv_obj_put(obj); + } + return rc; } -int lmv_unlink_slaves(struct obd_export *exp, struct mdc_op_data *data, - struct ptlrpc_request **req) +static int lmv_unlink_slaves(struct obd_export *exp, + struct md_op_data *op_data, + struct ptlrpc_request **req) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct mea *mea = data->mea1; - struct mdc_op_data data2; - int i, rc = 0, mds; + struct lmv_stripe_md *mea = op_data->op_mea1; + struct md_op_data *op_data2; + struct obd_export *tgt_exp; + int i, rc = 0; ENTRY; + OBD_ALLOC_PTR(op_data2); + if (op_data2 == NULL) + RETURN(-ENOMEM); + + op_data2->op_mode = S_IFDIR; + op_data2->op_fsuid = current->fsuid; + op_data2->op_fsgid = current->fsgid; + op_data2->op_bias = 0; + LASSERT(mea != NULL); for (i = 0; i < mea->mea_count; i++) { - if (lmv->tgts[i].ltd_exp == NULL) + memset(op_data2, 0, sizeof(*op_data2)); + op_data2->op_fid1 = mea->mea_ids[i]; + tgt_exp = lmv_find_export(lmv, &op_data2->op_fid1); + if (IS_ERR(tgt_exp)) + GOTO(out_free_op_data2, rc = PTR_ERR(tgt_exp)); + + if (tgt_exp == NULL) continue; - memset(&data2, 0, sizeof(data2)); - data2.fid1 = mea->mea_fids[i]; - data2.create_mode = MDS_MODE_DONT_LOCK | S_IFDIR; - mds = data2.fid1.mds; - rc = md_unlink(lmv->tgts[mds].ltd_exp, &data2, req); - CDEBUG(D_OTHER, "unlink slave %lu/%lu/%lu -> %d\n", - (unsigned long) mea->mea_fids[i].mds, - (unsigned long) mea->mea_fids[i].id, - (unsigned long) mea->mea_fids[i].generation, rc); + rc = md_unlink(tgt_exp, op_data2, req); + + CDEBUG(D_OTHER, "unlink slave "DFID" -> %d\n", + PFID(&mea->mea_ids[i]), rc); + if (*req) { ptlrpc_req_finished(*req); *req = NULL; } if (rc) - break; + GOTO(out_free_op_data2, rc); } - RETURN(rc); + + EXIT; +out_free_op_data2: + OBD_FREE_PTR(op_data2); + return rc; } -int lmv_unlink(struct obd_export *exp, struct mdc_op_data *data, - struct ptlrpc_request **request) +static int lmv_unlink(struct obd_export *exp, struct md_op_data *op_data, + struct ptlrpc_request **request) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - int rc, i = 0; + struct obd_export *tgt_exp = NULL; + struct lmv_obj *obj; + int rc, loop = 0; ENTRY; - + rc = lmv_check_connect(obd); if (rc) RETURN(rc); - if (data->namelen == 0 && data->mea1 != NULL) { + if (op_data->op_namelen == 0 && op_data->op_mea1 != NULL) { /* mds asks to remove slave objects */ - rc = lmv_unlink_slaves(exp, data, request); + rc = lmv_unlink_slaves(exp, op_data, request); RETURN(rc); - } else if (data->namelen != 0) { - struct lmv_obj *obj; - obj = lmv_grab_obj(obd, &data->fid1, 0); + } + +repeat: + ++loop; + LASSERT(loop <= 2); + if (op_data->op_namelen != 0) { + int mea_idx; + + obj = lmv_obj_grab(obd, &op_data->op_fid1); if (obj) { - i = raw_name2idx(obj->objcount, data->name, - data->namelen); - data->fid1 = obj->objs[i].fid; - lmv_put_obj(obj); + mea_idx = raw_name2idx(obj->lo_hashtype, + obj->lo_objcount, + op_data->op_name, + op_data->op_namelen); + op_data->op_bias &= ~MDS_CHECK_SPLIT; + op_data->op_fid1 = obj->lo_inodes[mea_idx].li_fid; + tgt_exp = lmv_get_export(lmv, + obj->lo_inodes[mea_idx].li_mds); + lmv_obj_put(obj); + CDEBUG(D_OTHER, "unlink '%*s' in "DFID" -> %u\n", + op_data->op_namelen, op_data->op_name, + PFID(&op_data->op_fid1), mea_idx); } - CDEBUG(D_OTHER, "unlink '%*s' in %lu/%lu/%lu -> %u\n", - data->namelen, data->name, - (unsigned long) data->fid1.mds, - (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation, i); } else { - CDEBUG(D_OTHER, "drop i_nlink on %lu/%lu/%lu\n", - (unsigned long) data->fid1.mds, - (unsigned long) data->fid1.id, - (unsigned long) data->fid1.generation); + CDEBUG(D_OTHER, "drop i_nlink on "DFID"\n", + PFID(&op_data->op_fid1)); + } + if (tgt_exp == NULL) { + tgt_exp = lmv_find_export(lmv, &op_data->op_fid1); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + op_data->op_bias |= MDS_CHECK_SPLIT; + } + + op_data->op_fsuid = current->fsuid; + op_data->op_fsgid = current->fsgid; + op_data->op_cap = current->cap_effective; + + /* If child's fid is given, cancel unused locks for it if it is from + * another export than parent. */ + if (op_data->op_namelen) { + /* LOOKUP lock for child (fid3) should also be cancelled on + * parent tgt_exp in mdc_unlink(). */ + op_data->op_flags |= MF_MDC_CANCEL_FID1 | MF_MDC_CANCEL_FID3; + + /* XXX: the case when child is a striped dir is not supported. + * Only the master stripe has all locks cancelled early. */ + /* Cancel FULL locks on child (fid3). */ + rc = lmv_early_cancel(lmv, tgt_exp, NULL, op_data, LCK_EX, + MDS_INODELOCK_FULL, MF_MDC_CANCEL_FID3); + } + if (rc == 0) + rc = md_unlink(tgt_exp, op_data, request); + if (rc == -ERESTART) { + LASSERT(*request != NULL); + DEBUG_REQ(D_WARNING|D_RPCTRACE, *request, + "Got -ERESTART during unlink!\n"); + ptlrpc_req_finished(*request); + *request = NULL; + + /* + * Directory got split. Time to update local object and repeat + * the request with proper MDS. + */ + rc = lmv_handle_split(exp, &op_data->op_fid1); + if (rc == 0) + goto repeat; } - rc = md_unlink(lmv->tgts[data->fid1.mds].ltd_exp, data, request); RETURN(rc); } -struct obd_device *lmv_get_real_obd(struct obd_export *exp, - char *name, int len) +static int lmv_llog_init(struct obd_device *obd, struct obd_llogs* llogs, + struct obd_device *tgt, int count, + struct llog_catid *logid, struct obd_uuid *uuid) { - struct obd_device *obd = exp->exp_obd; - struct lmv_obd *lmv = &obd->u.lmv; + struct llog_ctxt *ctxt; int rc; ENTRY; - rc = lmv_check_connect(obd); - if (rc) - RETURN(ERR_PTR(rc)); - obd = lmv->tgts[0].ltd_exp->exp_obd; - EXIT; - return obd; -} - -int lmv_init_ea_size(struct obd_export *exp, int easize, int cookiesize) -{ - struct obd_device *obd = exp->exp_obd; - struct lmv_obd *lmv = &obd->u.lmv; - int i, rc = 0, change = 0; - ENTRY; - - if (lmv->max_easize < easize) { - lmv->max_easize = easize; - change = 1; - } - if (lmv->max_cookiesize < cookiesize) { - lmv->max_cookiesize = cookiesize; - change = 1; + rc = llog_setup(obd, llogs, LLOG_CONFIG_REPL_CTXT, tgt, 0, NULL, + &llog_client_ops); + if (rc == 0) { + ctxt = llog_get_context(obd, LLOG_CONFIG_REPL_CTXT); + ctxt->loc_imp = tgt->u.cli.cl_import; } - if (change == 0) - RETURN(0); - - if (lmv->connected == 0) - RETURN(0); - for (i = 0; i < lmv->desc.ld_tgt_count; i++) { - rc = obd_init_ea_size(lmv->tgts[i].ltd_exp, easize, cookiesize); - if (rc) { - CERROR("obd_init_ea_size() failed on MDT target %d, " - "error %d.\n", i, rc); - break; - } - } RETURN(rc); } -int lmv_obd_create_single(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md **ea, struct obd_trans_info *oti) +static int lmv_llog_finish(struct obd_device *obd, int count) { - struct obd_device *obd = exp->exp_obd; - struct lmv_obd *lmv = &obd->u.lmv; - struct lov_stripe_md obj_md; - struct lov_stripe_md *obj_mdp = &obj_md; - int rc = 0; + int rc; ENTRY; - rc = lmv_check_connect(obd); - if (rc) - RETURN(rc); - - LASSERT(ea == NULL); - LASSERT(oa->o_mds < lmv->desc.ld_tgt_count); - - rc = obd_create(lmv->tgts[oa->o_mds].ltd_exp, oa, &obj_mdp, oti); - + rc = llog_cleanup(llog_get_context(obd, LLOG_CONFIG_REPL_CTXT)); RETURN(rc); } -/* - * to be called from MDS only - */ -int lmv_obd_create(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md **ea, struct obd_trans_info *oti) +static int lmv_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage) { - struct obd_device *obd = exp->exp_obd; - struct lmv_obd *lmv = &obd->u.lmv; - struct mea *mea; - int i, c, rc = 0; - struct ll_fid mfid; - ENTRY; - - rc = lmv_check_connect(obd); - if (rc) - RETURN(rc); - - LASSERT(oa != NULL); - - if (ea == NULL) { - rc = lmv_obd_create_single(exp, oa, NULL, oti); - RETURN(rc); - } - - if (*ea == NULL) { - rc = obd_alloc_diskmd(exp, (struct lov_mds_md **)ea); - if (rc < 0) { - CERROR("obd_alloc_diskmd() failed, error %d\n", - rc); - RETURN(rc); - } - - if (*ea == NULL) - RETURN(-EINVAL); - } - - rc = 0; - mfid.id = oa->o_id; - mfid.generation = oa->o_generation; - - mea = (struct mea *)*ea; - if (!mea->mea_count || mea->mea_count > lmv->desc.ld_tgt_count) - mea->mea_count = lmv->desc.ld_tgt_count; - - mea->mea_master = -1; - - for (i = 0, c = 0; c < mea->mea_count && - i < lmv->desc.ld_tgt_count; i++) { - struct lov_stripe_md obj_md; - struct lov_stripe_md *obj_mdp = &obj_md; - - if (lmv->tgts[i].ltd_exp == NULL) { - /* this is master MDS */ - mea->mea_fids[c].id = mfid.id; - mea->mea_fids[c].generation = mfid.generation; - mea->mea_fids[c].mds = i; - mea->mea_master = i; - c++; - continue; - } - - /* "master" MDS should always be part of stripped dir, so scan - for it. */ - if (mea->mea_master == -1 && c == mea->mea_count - 1) - continue; - - oa->o_valid = OBD_MD_FLGENER | OBD_MD_FLTYPE | OBD_MD_FLMODE - | OBD_MD_FLUID | OBD_MD_FLGID | OBD_MD_FLID; - - rc = obd_create(lmv->tgts[c].ltd_exp, oa, &obj_mdp, oti); - if (rc) { - CERROR("obd_create() failed on MDT target %d, " - "error %d\n", c, rc); - RETURN(rc); - } + int rc = 0; - mea->mea_fids[c].id = oa->o_id; - mea->mea_fids[c].generation = oa->o_generation; - mea->mea_fids[c].mds = i; - c++; - CDEBUG(D_OTHER, "dirobj at mds %d: "LPU64"/%u\n", - i, oa->o_id, oa->o_generation); + switch (stage) { + case OBD_CLEANUP_EARLY: + /* XXX: here should be calling obd_precleanup() down to + * stack. */ + break; + case OBD_CLEANUP_SELF_EXP: + rc = obd_llog_finish(obd, 0); + if (rc != 0) + CERROR("failed to cleanup llogging subsystems\n"); + break; + default: + break; } - LASSERT(c == mea->mea_count); - CDEBUG(D_OTHER, "%d dirobjects created\n", (int) mea->mea_count); - RETURN(rc); } @@ -1444,6 +2485,7 @@ static int lmv_get_info(struct obd_export *exp, __u32 keylen, { struct obd_device *obd; struct lmv_obd *lmv; + int rc = 0; ENTRY; obd = class_exp2obd(exp); @@ -1454,37 +2496,57 @@ static int lmv_get_info(struct obd_export *exp, __u32 keylen, } lmv = &obd->u.lmv; - if (keylen == 6 && memcmp(key, "mdsize", 6) == 0) { - __u32 *mdsize = val; - *vallen = sizeof(__u32); - *mdsize = sizeof(struct ll_fid) * lmv->desc.ld_tgt_count - + sizeof(struct mea); - RETURN(0); - } else if (keylen == 6 && memcmp(key, "mdsnum", 6) == 0) { - struct obd_uuid *cluuid = &lmv->cluuid; + if (keylen >= strlen("remote_flag") && !strcmp(key, "remote_flag")) { struct lmv_tgt_desc *tgts; - __u32 *mdsnum = val; int i; - for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; i++, tgts++) { - if (obd_uuid_equals(&tgts->uuid, cluuid)) { - *vallen = sizeof(__u32); - *mdsnum = i; - RETURN(0); + rc = lmv_check_connect(obd); + if (rc) + RETURN(rc); + + LASSERT(*vallen == sizeof(__u32)); + for (i = 0, tgts = lmv->tgts; i < lmv->desc.ld_tgt_count; + i++, tgts++) { + + /* all tgts should be connected when this get called. */ + if (!tgts || !tgts->ltd_exp) { + CERROR("target not setup?\n"); + continue; } + + if (!obd_get_info(tgts->ltd_exp, keylen, key, + vallen, val)) + RETURN(0); + } + RETURN(-EINVAL); + } else if (KEY_IS(KEY_MAX_EASIZE) || KEY_IS(KEY_CONN_DATA)) { + rc = lmv_check_connect(obd); + if (rc) + RETURN(rc); + + /* forwarding this request to first MDS, it should know LOV + * desc. */ + rc = obd_get_info(lmv->tgts[0].ltd_exp, keylen, key, + vallen, val); + if (!rc && KEY_IS(KEY_CONN_DATA)) { + exp->exp_connect_flags = + ((struct obd_connect_data *)val)->ocd_connect_flags; } - LASSERT(0); + RETURN(rc); } CDEBUG(D_IOCTL, "invalid key\n"); RETURN(-EINVAL); } -int lmv_set_info(struct obd_export *exp, obd_count keylen, - void *key, obd_count vallen, void *val) +int lmv_set_info_async(struct obd_export *exp, obd_count keylen, + void *key, obd_count vallen, void *val, + struct ptlrpc_request_set *set) { - struct obd_device *obd; - struct lmv_obd *lmv; + struct lmv_tgt_desc *tgt; + struct obd_device *obd; + struct lmv_obd *lmv; + int rc = 0; ENTRY; obd = class_exp2obd(exp); @@ -1495,27 +2557,25 @@ int lmv_set_info(struct obd_export *exp, obd_count keylen, } lmv = &obd->u.lmv; - if (keylen >= strlen("client") && strcmp(key, "client") == 0) { - struct lmv_tgt_desc *tgts; - int i, rc; + if (KEY_IS(KEY_READ_ONLY) || KEY_IS(KEY_FLUSH_CTX) || + KEY_IS(KEY_INIT_RECOV_BACKUP)) { + int i, err = 0; - rc = lmv_check_connect(obd); - if (rc) - RETURN(rc); + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { + tgt = &lmv->tgts[i]; - for (i = 0, tgts = lmv->tgts; - i < lmv->desc.ld_tgt_count; i++, tgts++) { - rc = obd_set_info(tgts->ltd_exp, keylen, key, vallen, val); - if (rc) - RETURN(rc); + if (!tgt->ltd_exp) + continue; + + err = obd_set_info_async(tgt->ltd_exp, + keylen, key, vallen, val, set); + if (err && rc == 0) + rc = err; } - RETURN(0); - } else if (keylen >= strlen("inter_mds") && strcmp(key, "inter_mds") == 0) { - lmv->server_timeout = 1; - lmv_set_timeouts(obd); - RETURN(0); + + RETURN(rc); } - + RETURN(-EINVAL); } @@ -1524,11 +2584,11 @@ int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp, { struct obd_device *obd = class_exp2obd(exp); struct lmv_obd *lmv = &obd->u.lmv; - int mea_size; + struct lmv_stripe_md *meap, *lsmp; + int mea_size, i; ENTRY; - mea_size = sizeof(struct ll_fid) * - lmv->desc.ld_tgt_count + sizeof(struct mea); + mea_size = lmv_get_easize(lmv); if (!lmmp) RETURN(mea_size); @@ -1547,106 +2607,287 @@ int lmv_packmd(struct obd_export *exp, struct lov_mds_md **lmmp, if (!lsm) RETURN(mea_size); -#warning "MEA packing/convertation must be here! -bzzz" - memcpy(*lmmp, lsm, mea_size); + lsmp = (struct lmv_stripe_md *)lsm; + meap = (struct lmv_stripe_md *)*lmmp; + + if (lsmp->mea_magic != MEA_MAGIC_LAST_CHAR && + lsmp->mea_magic != MEA_MAGIC_ALL_CHARS) + RETURN(-EINVAL); + + meap->mea_magic = cpu_to_le32(lsmp->mea_magic); + meap->mea_count = cpu_to_le32(lsmp->mea_count); + meap->mea_master = cpu_to_le32(lsmp->mea_master); + + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { + meap->mea_ids[i] = meap->mea_ids[i]; + fid_cpu_to_le(&meap->mea_ids[i], &meap->mea_ids[i]); + } + RETURN(mea_size); } -int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **mem_tgt, - struct lov_mds_md *disk_src, int mdsize) +int lmv_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp, + struct lov_mds_md *lmm, int lmm_size) { struct obd_device *obd = class_exp2obd(exp); + struct lmv_stripe_md **tmea = (struct lmv_stripe_md **)lsmp; + struct lmv_stripe_md *mea = (struct lmv_stripe_md *)lmm; struct lmv_obd *lmv = &obd->u.lmv; - struct mea **tmea = (struct mea **) mem_tgt; - struct mea *mea = (void *) disk_src; - int mea_size; + int mea_size, i; + __u32 magic; ENTRY; - mea_size = sizeof(struct ll_fid) * - lmv->desc.ld_tgt_count + sizeof(struct mea); - if (mem_tgt == NULL) + mea_size = lmv_get_easize(lmv); + if (lsmp == NULL) return mea_size; - if (*mem_tgt != NULL && disk_src == NULL) { + if (*lsmp != NULL && lmm == NULL) { OBD_FREE(*tmea, mea_size); + *lsmp = NULL; RETURN(0); } - LASSERT(mea_size == mdsize); + LASSERT(mea_size == lmm_size); OBD_ALLOC(*tmea, mea_size); if (*tmea == NULL) RETURN(-ENOMEM); - if (!disk_src) + if (!lmm) RETURN(mea_size); -#warning "MEA unpacking/convertation must be here! -bzzz" - memcpy(*tmea, mea, mdsize); + if (mea->mea_magic == MEA_MAGIC_LAST_CHAR || + mea->mea_magic == MEA_MAGIC_ALL_CHARS || + mea->mea_magic == MEA_MAGIC_HASH_SEGMENT) + { + magic = le32_to_cpu(mea->mea_magic); + } else { + /* old mea is not handled here */ + LBUG(); + } + + (*tmea)->mea_magic = magic; + (*tmea)->mea_count = le32_to_cpu(mea->mea_count); + (*tmea)->mea_master = le32_to_cpu(mea->mea_master); + + for (i = 0; i < (*tmea)->mea_count; i++) { + (*tmea)->mea_ids[i] = mea->mea_ids[i]; + fid_le_to_cpu(&(*tmea)->mea_ids[i], &(*tmea)->mea_ids[i]); + } RETURN(mea_size); } -int lmv_brw(int rw, struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *ea, obd_count oa_bufs, - struct brw_page *pgarr, struct obd_trans_info *oti) +static int lmv_cancel_unused(struct obd_export *exp, + const struct lu_fid *fid, + ldlm_policy_data_t *policy, + ldlm_mode_t mode, int flags, void *opaque) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + int rc = 0, err, i; + ENTRY; + + LASSERT(fid != NULL); + + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { + if (!lmv->tgts[i].ltd_exp || !lmv->tgts[i].ltd_active) + continue; + + err = md_cancel_unused(lmv->tgts[i].ltd_exp, fid, + policy, mode, flags, opaque); + if (!rc) + rc = err; + } + RETURN(rc); +} + +int lmv_set_lock_data(struct obd_export *exp, __u64 *lockh, void *data) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + + ENTRY; + RETURN(md_set_lock_data(lmv->tgts[0].ltd_exp, lockh, data)); +} + +ldlm_mode_t lmv_lock_match(struct obd_export *exp, int flags, + const struct lu_fid *fid, ldlm_type_t type, + ldlm_policy_data_t *policy, ldlm_mode_t mode, + struct lustre_handle *lockh) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + ldlm_mode_t rc; + int i; + ENTRY; + + CDEBUG(D_OTHER, "lock match for "DFID"\n", PFID(fid)); + + /* with CMD every object can have two locks in different namespaces: + * lookup lock in space of mds storing direntry and update/open lock in + * space of mds storing inode. Thus we check all targets, not only that + * one fid was created in. */ + for (i = 0; i < lmv->desc.ld_tgt_count; i++) { + rc = md_lock_match(lmv->tgts[i].ltd_exp, flags, fid, + type, policy, mode, lockh); + if (rc) + RETURN(rc); + } + + RETURN(0); +} + +int lmv_get_lustre_md(struct obd_export *exp, struct ptlrpc_request *req, + int offset, struct obd_export *dt_exp, + struct obd_export *md_exp, struct lustre_md *md) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + int rc; + + ENTRY; + rc = md_get_lustre_md(lmv->tgts[0].ltd_exp, req, offset, dt_exp, md_exp, + md); + RETURN(rc); +} + +int lmv_free_lustre_md(struct obd_export *exp, struct lustre_md *md) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + + ENTRY; + if (md->mea) + obd_free_memmd(exp, (struct lov_stripe_md**)&md->mea); + RETURN(md_free_lustre_md(lmv->tgts[0].ltd_exp, md)); +} + +int lmv_set_open_replay_data(struct obd_export *exp, + struct obd_client_handle *och, + struct ptlrpc_request *open_req) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; + + ENTRY; + + tgt_exp = lmv_find_export(lmv, &och->och_fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + RETURN(md_set_open_replay_data(tgt_exp, och, open_req)); +} + +int lmv_clear_open_replay_data(struct obd_export *exp, + struct obd_client_handle *och) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; + ENTRY; + + tgt_exp = lmv_find_export(lmv, &och->och_fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + RETURN(md_clear_open_replay_data(tgt_exp, och)); +} + +static int lmv_get_remote_perm(struct obd_export *exp, + const struct lu_fid *fid, + struct obd_capa *oc, + struct ptlrpc_request **request) +{ + struct obd_device *obd = exp->exp_obd; + struct lmv_obd *lmv = &obd->u.lmv; + struct obd_export *tgt_exp; + int rc; + + ENTRY; + + rc = lmv_check_connect(obd); + if (rc) + RETURN(rc); + + tgt_exp = lmv_find_export(lmv, fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_get_remote_perm(tgt_exp, fid, oc, request); + + RETURN(rc); +} + +static int lmv_renew_capa(struct obd_export *exp, struct obd_capa *oc, + renew_capa_cb_t cb) { struct obd_device *obd = exp->exp_obd; struct lmv_obd *lmv = &obd->u.lmv; - struct mea *mea = (struct mea *) ea; - int err; - - LASSERT(oa != NULL); - LASSERT(ea != NULL); - LASSERT(pgarr != NULL); - LASSERT(oa->o_mds < lmv->desc.ld_tgt_count); - - oa->o_gr = mea->mea_fids[oa->o_mds].generation; - oa->o_id = mea->mea_fids[oa->o_mds].id; - oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP; - err = obd_brw(rw, lmv->tgts[oa->o_mds].ltd_exp, oa, - NULL, oa_bufs, pgarr, oti); - RETURN(err); + struct obd_export *tgt_exp; + int rc; + ENTRY; + + rc = lmv_check_connect(obd); + if (rc) + RETURN(rc); + + tgt_exp = lmv_find_export(lmv, &oc->c_capa.lc_fid); + if (IS_ERR(tgt_exp)) + RETURN(PTR_ERR(tgt_exp)); + + rc = md_renew_capa(tgt_exp, oc, cb); + RETURN(rc); } struct obd_ops lmv_obd_ops = { .o_owner = THIS_MODULE, - .o_attach = lmv_attach, - .o_detach = lmv_detach, .o_setup = lmv_setup, .o_cleanup = lmv_cleanup, + .o_precleanup = lmv_precleanup, + .o_process_config = lmv_process_config, .o_connect = lmv_connect, .o_disconnect = lmv_disconnect, .o_statfs = lmv_statfs, + .o_llog_init = lmv_llog_init, + .o_llog_finish = lmv_llog_finish, .o_get_info = lmv_get_info, - .o_set_info = lmv_set_info, - .o_create = lmv_obd_create, + .o_set_info_async = lmv_set_info_async, .o_packmd = lmv_packmd, .o_unpackmd = lmv_unpackmd, - .o_brw = lmv_brw, - .o_init_ea_size = lmv_init_ea_size, .o_notify = lmv_notify, .o_iocontrol = lmv_iocontrol, + .o_fid_delete = lmv_fid_delete }; struct md_ops lmv_md_ops = { .m_getstatus = lmv_getstatus, - .m_getattr = lmv_getattr, .m_change_cbdata = lmv_change_cbdata, - .m_change_cbdata_name = lmv_change_cbdata_name, .m_close = lmv_close, .m_create = lmv_create, .m_done_writing = lmv_done_writing, .m_enqueue = lmv_enqueue, + .m_getattr = lmv_getattr, + .m_getxattr = lmv_getxattr, .m_getattr_name = lmv_getattr_name, .m_intent_lock = lmv_intent_lock, .m_link = lmv_link, .m_rename = lmv_rename, .m_setattr = lmv_setattr, + .m_setxattr = lmv_setxattr, .m_sync = lmv_sync, .m_readpage = lmv_readpage, .m_unlink = lmv_unlink, - .m_get_real_obd = lmv_get_real_obd, - .m_valid_attrs = lmv_valid_attrs, + .m_init_ea_size = lmv_init_ea_size, + .m_cancel_unused = lmv_cancel_unused, + .m_set_lock_data = lmv_set_lock_data, + .m_lock_match = lmv_lock_match, + .m_get_lustre_md = lmv_get_lustre_md, + .m_free_lustre_md = lmv_free_lustre_md, + .m_set_open_replay_data = lmv_set_open_replay_data, + .m_clear_open_replay_data = lmv_clear_open_replay_data, + .m_get_remote_perm = lmv_get_remote_perm, + .m_renew_capa = lmv_renew_capa }; int __init lmv_init(void) @@ -1654,16 +2895,34 @@ int __init lmv_init(void) struct lprocfs_static_vars lvars; int rc; + obj_cache = cfs_mem_cache_create("lmv_objects", + sizeof(struct lmv_obj), + 0, 0); + if (!obj_cache) { + CERROR("error allocating lmv objects cache\n"); + return -ENOMEM; + } + lprocfs_init_vars(lmv, &lvars); rc = class_register_type(&lmv_obd_ops, &lmv_md_ops, - lvars.module_vars, OBD_LMV_DEVICENAME); - RETURN(rc); + lvars.module_vars, LUSTRE_LMV_NAME, NULL); + if (rc) + cfs_mem_cache_destroy(obj_cache); + + return rc; } #ifdef __KERNEL__ static void lmv_exit(void) { - class_unregister_type(OBD_LMV_DEVICENAME); + int rc; + + class_unregister_type(LUSTRE_LMV_NAME); + + rc = cfs_mem_cache_destroy(obj_cache); + LASSERTF(rc == 0, + "can't free lmv objects cache, %d object(s)" + "still in use\n", atomic_read(&obj_cache_count)); } MODULE_AUTHOR("Cluster File Systems, Inc. ");