X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Flov%2Flov_obd.c;h=f293cfd042b30dc14fb847bc3892c8f0eac2040b;hp=14935b74dcf62542e2a0de6d943d933f52b1975e;hb=a25adb47c7b7eeb68a922e2647d74eeff3401c6a;hpb=b5b6224b227ca816fc40739741f0c37364db6677 diff --git a/lustre/lov/lov_obd.c b/lustre/lov/lov_obd.c index 14935b7..f293cfd 100644 --- a/lustre/lov/lov_obd.c +++ b/lustre/lov/lov_obd.c @@ -1,25 +1,29 @@ /* -*- 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) 2002-2006 Cluster File Systems, Inc. * Author: Phil Schwan * Peter Braam * Mike Shaver + * Nathan Rutman * - * This file is part of Lustre, http://www.lustre.org. + * This file is part of the Lustre file system, http://www.lustre.org + * Lustre is a trademark of Cluster File Systems, Inc. * - * Lustre is free software; you can redistribute it and/or - * modify it under the terms of version 2 of the GNU General Public - * License as published by the Free Software Foundation. + * You may have signed or agreed to another license before downloading + * this software. If so, you are bound by the terms and conditions + * of that agreement, and the following does not apply to you. See the + * LICENSE file included with this distribution for more information. * - * Lustre is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the - * GNU General Public License for more details. + * If you did not agree to a different license, then this copy of Lustre + * is open source software; you can redistribute it and/or modify it + * under the terms of version 2 of the GNU General Public License as + * published by the Free Software Foundation. * - * You should have received a copy of the GNU General Public License - * along with Lustre; if not, write to the Free Software - * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. + * In either case, Lustre is distributed in the hope that it will be + * useful, but WITHOUT ANY WARRANTY; without even the implied warranty + * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * license text for more details. */ #ifndef EXPORT_SYMTAB @@ -27,165 +31,159 @@ #endif #define DEBUG_SUBSYSTEM S_LOV #ifdef __KERNEL__ -#include -#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 #include "lov_internal.h" -static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off, - int stripeno, obd_off *obd_off); -struct lov_lock_handles { - struct portals_handle llh_handle; - atomic_t llh_refcount; - int llh_stripe_count; - struct lustre_handle llh_handles[0]; -}; - -static void lov_llh_addref(void *llhp) -{ - struct lov_lock_handles *llh = llhp; - - atomic_inc(&llh->llh_refcount); - CDEBUG(D_INFO, "GETting llh %p : new refcount %d\n", llh, - atomic_read(&llh->llh_refcount)); -} - -static struct lov_lock_handles *lov_llh_new(struct lov_stripe_md *lsm) +/* Keep a refcount of lov->tgt usage to prevent racing with addition/deletion. + Any function that expects lov_tgts to remain stationary must take a ref. */ +void lov_getref(struct obd_device *obd) { - struct lov_lock_handles *llh; + struct lov_obd *lov = &obd->u.lov; - OBD_ALLOC(llh, sizeof *llh + - sizeof(*llh->llh_handles) * lsm->lsm_stripe_count); - if (llh == NULL) { - CERROR("out of memory\n"); - return NULL; - } - atomic_set(&llh->llh_refcount, 2); - llh->llh_stripe_count = lsm->lsm_stripe_count; - INIT_LIST_HEAD(&llh->llh_handle.h_link); - class_handle_hash(&llh->llh_handle, lov_llh_addref); - return llh; + /* nobody gets through here until lov_putref is done */ + mutex_down(&lov->lov_lock); + atomic_inc(&lov->lov_refcount); + mutex_up(&lov->lov_lock); + return; } -static struct lov_lock_handles *lov_handle2llh(struct lustre_handle *handle) -{ - ENTRY; - LASSERT(handle != NULL); - RETURN(class_handle2object(handle->cookie)); -} +static void __lov_del_obd(struct obd_device *obd, __u32 index); -static void lov_llh_put(struct lov_lock_handles *llh) +void lov_putref(struct obd_device *obd) { - CDEBUG(D_INFO, "PUTting llh %p : new refcount %d\n", llh, - atomic_read(&llh->llh_refcount) - 1); - LASSERT(atomic_read(&llh->llh_refcount) > 0 && - atomic_read(&llh->llh_refcount) < 0x5a5a); - if (atomic_dec_and_test(&llh->llh_refcount)) { - LASSERT(list_empty(&llh->llh_handle.h_link)); - OBD_FREE(llh, sizeof *llh + - sizeof(*llh->llh_handles) * llh->llh_stripe_count); + struct lov_obd *lov = &obd->u.lov; + mutex_down(&lov->lov_lock); + /* ok to dec to 0 more than once -- ltd_exp's will be null */ + if (atomic_dec_and_test(&lov->lov_refcount) && lov->lov_death_row) { + int i; + CDEBUG(D_CONFIG, "destroying %d lov targets\n", + lov->lov_death_row); + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_reap) + continue; + /* Disconnect and delete from list */ + __lov_del_obd(obd, i); + lov->lov_death_row--; + } } + mutex_up(&lov->lov_lock); } -static void lov_llh_destroy(struct lov_lock_handles *llh) -{ - class_handle_unhash(&llh->llh_handle); - lov_llh_put(llh); -} - -/* obd methods */ #define MAX_STRING_SIZE 128 -static int lov_connect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt, - int activate) +static int lov_connect_obd(struct obd_device *obd, __u32 index, int activate, + struct obd_connect_data *data) { struct lov_obd *lov = &obd->u.lov; - struct obd_uuid *tgt_uuid = &tgt->uuid; + struct obd_uuid tgt_uuid = lov->lov_tgts[index]->ltd_uuid; struct obd_device *tgt_obd; struct obd_uuid lov_osc_uuid = { "LOV_OSC_UUID" }; struct lustre_handle conn = {0, }; - struct proc_dir_entry *lov_proc_dir; + struct obd_import *imp; +#ifdef __KERNEL__ + cfs_proc_dir_entry_t *lov_proc_dir; +#endif int rc; ENTRY; - tgt_obd = class_find_client_obd(tgt_uuid, LUSTRE_OSC_NAME, + if (!lov->lov_tgts[index]) + RETURN(-EINVAL); + + tgt_obd = class_find_client_obd(&tgt_uuid, LUSTRE_OSC_NAME, &obd->obd_uuid); if (!tgt_obd) { - CERROR("Target %s not attached\n", tgt_uuid->uuid); + CERROR("Target %s not attached\n", obd_uuid2str(&tgt_uuid)); RETURN(-EINVAL); } - if (!tgt_obd->obd_set_up) { - CERROR("Target %s not set up\n", tgt_uuid->uuid); + CERROR("Target %s not set up\n", obd_uuid2str(&tgt_uuid)); RETURN(-EINVAL); } + if (data && (data->ocd_connect_flags & OBD_CONNECT_INDEX)) + data->ocd_index = index; + + /* + * Divine LOV knows that OBDs under it are OSCs. + */ + imp = tgt_obd->u.cli.cl_import; + if (activate) { tgt_obd->obd_no_recov = 0; - ptlrpc_activate_import(tgt_obd->u.cli.cl_import); + /* FIXME this is probably supposed to be + ptlrpc_set_import_active. Horrible naming. */ + ptlrpc_activate_import(imp); } - if (tgt_obd->u.cli.cl_import->imp_invalid) { + if (imp->imp_invalid) { CERROR("not connecting OSC %s; administratively " - "disabled\n", tgt_uuid->uuid); + "disabled\n", obd_uuid2str(&tgt_uuid)); rc = obd_register_observer(tgt_obd, obd); if (rc) { CERROR("Target %s register_observer error %d; " "will not be able to reactivate\n", - tgt_uuid->uuid, rc); + obd_uuid2str(&tgt_uuid), rc); } RETURN(0); } - rc = obd_connect(&conn, tgt_obd, &lov_osc_uuid); + rc = obd_connect(NULL, &conn, tgt_obd, &lov_osc_uuid, data); if (rc) { - CERROR("Target %s connect error %d\n", tgt_uuid->uuid, rc); + CERROR("Target %s connect error %d\n", + obd_uuid2str(&tgt_uuid), rc); RETURN(rc); } - tgt->ltd_exp = class_conn2export(&conn); + lov->lov_tgts[index]->ltd_exp = class_conn2export(&conn); + if (!lov->lov_tgts[index]->ltd_exp) { + CERROR("Target %s: null export!\n", obd_uuid2str(&tgt_uuid)); + RETURN(-ENODEV); + } rc = obd_register_observer(tgt_obd, obd); if (rc) { CERROR("Target %s register_observer error %d\n", - tgt_uuid->uuid, rc); - obd_disconnect(tgt->ltd_exp, 0); - tgt->ltd_exp = NULL; + obd_uuid2str(&tgt_uuid), rc); + obd_disconnect(lov->lov_tgts[index]->ltd_exp); + lov->lov_tgts[index]->ltd_exp = NULL; RETURN(rc); } - tgt->active = 1; - lov->desc.ld_active_tgt_count++; + lov->lov_tgts[index]->ltd_reap = 0; + if (activate) { + lov->lov_tgts[index]->ltd_active = 1; + lov->desc.ld_active_tgt_count++; + } + CDEBUG(D_CONFIG, "Connected tgt idx %d %s (%s) %sactive\n", index, + obd_uuid2str(&tgt_uuid), tgt_obd->obd_name, activate ? "":"in"); +#ifdef __KERNEL__ lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds"); if (lov_proc_dir) { struct obd_device *osc_obd = class_conn2obd(&conn); - struct proc_dir_entry *osc_symlink; - char name[MAX_STRING_SIZE + 1]; + cfs_proc_dir_entry_t *osc_symlink; + char name[MAX_STRING_SIZE]; LASSERT(osc_obd != NULL); - LASSERT(osc_obd->obd_type != NULL); + LASSERT(osc_obd->obd_magic == OBD_DEVICE_MAGIC); LASSERT(osc_obd->obd_type->typ_name != NULL); - name[MAX_STRING_SIZE] = '\0'; snprintf(name, MAX_STRING_SIZE, "../../../%s/%s", osc_obd->obd_type->typ_name, osc_obd->obd_name); @@ -196,81 +194,83 @@ static int lov_connect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt, "/proc/fs/lustre/%s/%s/target_obds/%s.", obd->obd_type->typ_name, obd->obd_name, osc_obd->obd_name); - lprocfs_remove(lov_proc_dir); - lov_proc_dir = NULL; + lprocfs_remove(&lov_proc_dir); } } +#endif + + rc = qos_add_tgt(obd, index); + if (rc) + CERROR("qos_add_tgt failed %d\n", rc); RETURN(0); } -static int lov_connect(struct lustre_handle *conn, struct obd_device *obd, - struct obd_uuid *cluuid) +static int lov_connect(const struct lu_env *env, + struct lustre_handle *conn, struct obd_device *obd, + struct obd_uuid *cluuid, struct obd_connect_data *data) { struct lov_obd *lov = &obd->u.lov; struct lov_tgt_desc *tgt; - struct obd_export *exp; - int rc, rc2, i; + int i, rc; ENTRY; + CDEBUG(D_CONFIG, "connect #%d\n", lov->lov_connects); + rc = class_connect(conn, obd, cluuid); if (rc) RETURN(rc); - exp = class_conn2export(conn); + /* Why should there ever be more than 1 connect? */ + lov->lov_connects++; + LASSERT(lov->lov_connects == 1); - /* We don't want to actually do the underlying connections more than - * once, so keep track. */ - lov->refcount++; - if (lov->refcount > 1) { - class_export_put(exp); - RETURN(0); - } + memset(&lov->lov_ocd, 0, sizeof(lov->lov_ocd)); + if (data) + lov->lov_ocd = *data; - for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) { - if (obd_uuid_empty(&tgt->uuid)) + lov_getref(obd); + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + tgt = lov->lov_tgts[i]; + if (!tgt || obd_uuid_empty(&tgt->ltd_uuid)) continue; - rc = lov_connect_obd(obd, tgt, 0); - if (rc) - GOTO(out_disc, rc); + /* Flags will be lowest common denominator */ + rc = lov_connect_obd(obd, i, tgt->ltd_activate, &lov->lov_ocd); + if (rc) { + CERROR("%s: lov connect tgt %d failed: %d\n", + obd->obd_name, i, rc); + continue; + } } + lov_putref(obd); - class_export_put(exp); - RETURN (0); - - out_disc: - while (i-- > 0) { - struct obd_uuid uuid; - --tgt; - --lov->desc.ld_active_tgt_count; - tgt->active = 0; - /* save for CERROR below; (we know it's terminated) */ - uuid = tgt->uuid; - rc2 = obd_disconnect(tgt->ltd_exp, 0); - if (rc2) - CERROR("error: LOV target %s disconnect on OST idx %d: " - "rc = %d\n", uuid.uuid, i, rc2); - } - class_disconnect(exp, 0); - RETURN (rc); + RETURN(0); } -static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt, - int flags) +static int lov_disconnect_obd(struct obd_device *obd, __u32 index) { - struct proc_dir_entry *lov_proc_dir; - struct obd_device *osc_obd = class_exp2obd(tgt->ltd_exp); + cfs_proc_dir_entry_t *lov_proc_dir; struct lov_obd *lov = &obd->u.lov; + struct obd_device *osc_obd = + class_exp2obd(lov->lov_tgts[index]->ltd_exp); int rc; ENTRY; + CDEBUG(D_CONFIG, "%s: disconnecting target %s\n", + obd->obd_name, osc_obd->obd_name); + + if (lov->lov_tgts[index]->ltd_active) { + lov->lov_tgts[index]->ltd_active = 0; + lov->desc.ld_active_tgt_count--; + } + lov_proc_dir = lprocfs_srch(obd->obd_proc_entry, "target_obds"); if (lov_proc_dir) { - struct proc_dir_entry *osc_symlink; + cfs_proc_dir_entry_t *osc_symlink; osc_symlink = lprocfs_srch(lov_proc_dir, osc_obd->obd_name); if (osc_symlink) { - lprocfs_remove(osc_symlink); + lprocfs_remove(&osc_symlink); } else { CERROR("/proc/fs/lustre/%s/%s/target_obds/%s missing.", obd->obd_type->typ_name, obd->obd_name, @@ -278,58 +278,65 @@ static int lov_disconnect_obd(struct obd_device *obd, struct lov_tgt_desc *tgt, } } - if (obd->obd_no_recov) { + if (osc_obd) { /* 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. */ - if (osc_obd) - osc_obd->obd_no_recov = 1; + osc_obd->obd_force = obd->obd_force; + osc_obd->obd_fail = obd->obd_fail; + osc_obd->obd_no_recov = obd->obd_no_recov; } - obd_register_observer(tgt->ltd_exp->exp_obd, NULL); + obd_register_observer(osc_obd, NULL); - rc = obd_disconnect(tgt->ltd_exp, flags); + rc = obd_disconnect(lov->lov_tgts[index]->ltd_exp); if (rc) { - if (tgt->active) { - CERROR("Target %s disconnect error %d\n", - tgt->uuid.uuid, rc); - } + CERROR("Target %s disconnect error %d\n", + lov_uuid2str(lov, index), rc); rc = 0; } - if (tgt->active) { - tgt->active = 0; - lov->desc.ld_active_tgt_count--; - } + qos_del_tgt(obd, index); - tgt->ltd_exp = NULL; + lov->lov_tgts[index]->ltd_exp = NULL; RETURN(0); } -static int lov_disconnect(struct obd_export *exp, int flags) +static int lov_del_target(struct obd_device *obd, __u32 index, + struct obd_uuid *uuidp, int gen); + +static int lov_disconnect(struct obd_export *exp) { struct obd_device *obd = class_exp2obd(exp); struct lov_obd *lov = &obd->u.lov; - struct lov_tgt_desc *tgt; - int rc, i; + int i, rc; ENTRY; - if (!lov->tgts) - goto out_local; + if (!lov->lov_tgts) + goto out; /* Only disconnect the underlying layers on the final disconnect. */ - lov->refcount--; - if (lov->refcount != 0) - goto out_local; + lov->lov_connects--; + if (lov->lov_connects != 0) { + /* why should there be more than 1 connect? */ + CERROR("disconnect #%d\n", lov->lov_connects); + goto out; + } - for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) { - if (tgt->ltd_exp) - lov_disconnect_obd(obd, tgt, flags); + /* Let's hold another reference so lov_del_obd doesn't spin through + putref every time */ + lov_getref(obd); + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + if (lov->lov_tgts[i] && lov->lov_tgts[i]->ltd_exp) { + /* Disconnection is the last we know about an obd */ + lov_del_target(obd, i, 0, lov->lov_tgts[i]->ltd_gen); + } } + lov_putref(obd); - out_local: - rc = class_disconnect(exp, 0); +out: + rc = class_disconnect(exp); /* bz 9811 */ RETURN(rc); } @@ -339,10 +346,10 @@ static int lov_disconnect(struct obd_export *exp, int flags) * -ENOTCONN: The UUID is found, but the target connection is bad (!) * -EBADF : The UUID is found, but the OBD is the wrong type (!) */ -static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid, +static int lov_set_osc_active(struct obd_device *obd, struct obd_uuid *uuid, int activate) { - struct obd_device *obd; + struct lov_obd *lov = &obd->u.lov; struct lov_tgt_desc *tgt; int i, rc = 0; ENTRY; @@ -350,189 +357,113 @@ static int lov_set_osc_active(struct lov_obd *lov, struct obd_uuid *uuid, CDEBUG(D_INFO, "Searching in lov %p for uuid %s (activate=%d)\n", lov, uuid->uuid, activate); - spin_lock(&lov->lov_lock); - for (i = 0, tgt = lov->tgts; i < lov->desc.ld_tgt_count; i++, tgt++) { + lov_getref(obd); + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + tgt = lov->lov_tgts[i]; + if (!tgt || !tgt->ltd_exp) + continue; + CDEBUG(D_INFO, "lov 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, obd_uuid2str(&tgt->ltd_uuid), + tgt->ltd_exp->exp_handle.h_cookie); + if (obd_uuid_equals(uuid, &tgt->ltd_uuid)) break; } if (i == lov->desc.ld_tgt_count) GOTO(out, rc = -EINVAL); - obd = class_exp2obd(tgt->ltd_exp); - if (obd == NULL) { - /* This can happen if OST failure races with node shutdown */ - GOTO(out, rc = -ENOTCONN); - } - - CDEBUG(D_INFO, "Found OBD %s=%s device %d (%p) type %s at LOV 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_OSC_NAME) == 0); - - if (tgt->active == activate) { - CDEBUG(D_INFO, "OBD %p already %sactive!\n", obd, + if (lov->lov_tgts[i]->ltd_active == activate) { + CDEBUG(D_INFO, "OSC %s already %sactive!\n", uuid->uuid, activate ? "" : "in"); GOTO(out, rc); } - CDEBUG(D_INFO, "Marking OBD %p %sactive\n", obd, activate ? "" : "in"); + CDEBUG(D_CONFIG, "Marking OSC %s %sactive\n", obd_uuid2str(uuid), + activate ? "" : "in"); - tgt->active = activate; + lov->lov_tgts[i]->ltd_active = activate; if (activate) lov->desc.ld_active_tgt_count++; else lov->desc.ld_active_tgt_count--; + /* remove any old qos penalty */ + lov->lov_tgts[i]->ltd_qos.ltq_penalty = 0; - EXIT; out: - spin_unlock(&lov->lov_lock); - return rc; -} - -static int lov_notify(struct obd_device *obd, struct obd_device *watched, - int active) -{ - int rc; - struct obd_uuid *uuid; - - if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) { - CERROR("unexpected notification of %s %s!\n", - watched->obd_type->typ_name, - watched->obd_name); - return -EINVAL; - } - uuid = &watched->u.cli.cl_import->imp_target_uuid; - - /* Set OSC as active before notifying the observer, so the - * observer can use the OSC normally. - */ - rc = lov_set_osc_active(&obd->u.lov, uuid, active); - if (rc) { - CERROR("%sactivation of %s failed: %d\n", - active ? "" : "de", uuid->uuid, rc); - RETURN(rc); - } - - if (obd->obd_observer) - /* Pass the notification up the chain. */ - rc = obd_notify(obd->obd_observer, watched, active); - + lov_putref(obd); RETURN(rc); } -int lov_attach(struct obd_device *dev, obd_count len, void *data) -{ - struct lprocfs_static_vars lvars; - int rc; - - lprocfs_init_vars(lov, &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) { - rc = -ENOMEM; - } else { - entry->proc_fops = &lov_proc_target_fops; - entry->data = dev; - } -#endif - } - return rc; -} - -int lov_detach(struct obd_device *dev) -{ - return lprocfs_obd_detach(dev); -} - -static int lov_setup(struct obd_device *obd, obd_count len, void *buf) +static int lov_notify(struct obd_device *obd, struct obd_device *watched, + enum obd_notify_event ev, void *data) { - struct lustre_cfg *lcfg = buf; - struct lov_desc *desc; - struct lov_obd *lov = &obd->u.lov; - int count; + int rc = 0; ENTRY; - if (lcfg->lcfg_inllen1 < 1) { - CERROR("LOV setup requires a descriptor\n"); - RETURN(-EINVAL); - } + if (ev == OBD_NOTIFY_ACTIVE || ev == OBD_NOTIFY_INACTIVE) { + struct obd_uuid *uuid; - desc = (struct lov_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); - } + LASSERT(watched); - /* Because of 64-bit divide/mod operations only work with a 32-bit - * divisor in a 32-bit kernel, we cannot support a stripe width - * of 4GB or larger on 32-bit CPUs. - */ + if (strcmp(watched->obd_type->typ_name, LUSTRE_OSC_NAME)) { + CERROR("unexpected notification of %s %s!\n", + watched->obd_type->typ_name, + watched->obd_name); + RETURN(-EINVAL); + } + uuid = &watched->u.cli.cl_target_uuid; - count = desc->ld_default_stripe_count; - if (count && (count * desc->ld_default_stripe_size) > ~0UL) { - CERROR("LOV: stripe width "LPU64"x%u > %lu on 32-bit system\n", - desc->ld_default_stripe_size, count, ~0UL); - RETURN(-EINVAL); + /* Set OSC as active before notifying the observer, so the + * observer can use the OSC normally. + */ + rc = lov_set_osc_active(obd, uuid, ev == OBD_NOTIFY_ACTIVE); + if (rc) { + CERROR("%sactivation of %s failed: %d\n", + (ev == OBD_NOTIFY_ACTIVE) ? "" : "de", + obd_uuid2str(uuid), rc); + RETURN(rc); + } } - if (desc->ld_tgt_count) { - lov->bufsize = desc->ld_tgt_count * sizeof(struct lov_tgt_desc); - OBD_ALLOC(lov->tgts, lov->bufsize); - if (lov->tgts == NULL) { - lov->bufsize = 0; - CERROR("couldn't allocate %d bytes for target table.\n", - lov->bufsize); - RETURN(-EINVAL); + /* Pass the notification up the chain. */ + if (watched) { + rc = obd_notify_observer(obd, watched, ev, data); + } else { + /* NULL watched means all osc's in the lov (only for syncs) */ + struct lov_obd *lov = &obd->u.lov; + struct obd_device *tgt_obd; + int i; + lov_getref(obd); + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + if (!lov->lov_tgts[i]) + continue; + tgt_obd = class_exp2obd(lov->lov_tgts[i]->ltd_exp); + rc = obd_notify_observer(obd, tgt_obd, ev, data); + if (rc) { + CERROR("%s: notify %s of %s failed %d\n", + obd->obd_name, + obd->obd_observer->obd_name, + tgt_obd->obd_name, rc); + break; + } } - desc->ld_tgt_count = 0; + lov_putref(obd); } - desc->ld_active_tgt_count = 0; - CDEBUG(D_CONFIG, "tgts: %p bufsize: 0x%x\n", lov->tgts, lov->bufsize); - lov->desc = *desc; - spin_lock_init(&lov->lov_lock); - - RETURN(0); -} - -static int lov_cleanup(struct obd_device *obd, int flags) -{ - struct lov_obd *lov = &obd->u.lov; - - OBD_FREE(lov->tgts, lov->bufsize); - RETURN(0); + RETURN(rc); } -static int -lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen) +static int lov_add_target(struct obd_device *obd, struct obd_uuid *uuidp, + __u32 index, int gen, int active) { struct lov_obd *lov = &obd->u.lov; struct lov_tgt_desc *tgt; - struct obd_export *exp_observer; - __u32 bufsize; - __u32 size = 2; - obd_id params[2]; - char name[32] = "CATLIST"; - int rc, old_count; + int rc; ENTRY; - CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n", - uuidp->uuid, index, gen); - - if (index < 0) { - CERROR("request to add OBD %s at invalid index: %d\n", - uuidp->uuid, index); - RETURN(-EINVAL); - } + CDEBUG(D_CONFIG, "uuid:%s idx:%d gen:%d active:%d\n", + uuidp->uuid, index, gen, active); if (gen <= 0) { CERROR("request to add OBD %s with invalid generation: %d\n", @@ -540,282 +471,454 @@ lov_add_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen) RETURN(-EINVAL); } - bufsize = sizeof(struct lov_tgt_desc) * (index + 1); - if (bufsize > lov->bufsize) { - OBD_ALLOC(tgt, bufsize); - if (tgt == NULL) { - CERROR("couldn't allocate %d bytes for new table.\n", - bufsize); + mutex_down(&lov->lov_lock); + + if ((index < lov->lov_tgt_size) && (lov->lov_tgts[index] != NULL)) { + tgt = lov->lov_tgts[index]; + CERROR("UUID %s already assigned at LOV target index %d\n", + obd_uuid2str(&tgt->ltd_uuid), index); + mutex_up(&lov->lov_lock); + RETURN(-EEXIST); + } + + if (index >= lov->lov_tgt_size) { + /* We need to reallocate the lov target array. */ + struct lov_tgt_desc **newtgts, **old = NULL; + __u32 newsize, oldsize = 0; + + newsize = max(lov->lov_tgt_size, (__u32)2); + while (newsize < index + 1) + newsize = newsize << 1; + OBD_ALLOC(newtgts, sizeof(*newtgts) * newsize); + if (newtgts == NULL) { + mutex_up(&lov->lov_lock); RETURN(-ENOMEM); } - memset(tgt, 0, bufsize); - if (lov->tgts) { - memcpy(tgt, lov->tgts, lov->bufsize); - OBD_FREE(lov->tgts, lov->bufsize); + if (lov->lov_tgt_size) { + memcpy(newtgts, lov->lov_tgts, sizeof(*newtgts) * + lov->lov_tgt_size); + old = lov->lov_tgts; + oldsize = lov->lov_tgt_size; } - lov->tgts = tgt; - lov->bufsize = bufsize; - CDEBUG(D_CONFIG, "tgts: %p bufsize: %d\n", - lov->tgts, lov->bufsize); + lov->lov_tgts = newtgts; + lov->lov_tgt_size = newsize; +#ifdef __KERNEL__ + smp_rmb(); +#endif + if (old) + OBD_FREE(old, sizeof(*old) * oldsize); + + CDEBUG(D_CONFIG, "tgts: %p size: %d\n", + lov->lov_tgts, lov->lov_tgt_size); } - tgt = lov->tgts + index; - if (!obd_uuid_empty(&tgt->uuid)) { - CERROR("OBD already assigned at LOV target index %d\n", - index); - RETURN(-EEXIST); + + OBD_ALLOC_PTR(tgt); + if (!tgt) { + mutex_up(&lov->lov_lock); + RETURN(-ENOMEM); } - tgt->uuid = *uuidp; + memset(tgt, 0, sizeof(*tgt)); + tgt->ltd_uuid = *uuidp; /* XXX - add a sanity check on the generation number. */ tgt->ltd_gen = gen; - - old_count = lov->desc.ld_tgt_count; + tgt->ltd_index = index; + tgt->ltd_activate = active; + lov->lov_tgts[index] = tgt; if (index >= lov->desc.ld_tgt_count) lov->desc.ld_tgt_count = index + 1; + mutex_up(&lov->lov_lock); - CDEBUG(D_CONFIG, "idx: %d ltd_gen: %d ld_tgt_count: %d\n", + CDEBUG(D_CONFIG, "idx=%d ltd_gen=%d ld_tgt_count=%d\n", index, tgt->ltd_gen, lov->desc.ld_tgt_count); - if (lov->refcount == 0) + if (lov->lov_connects == 0) { + /* lov_connect hasn't been called yet. We'll do the + lov_connect_obd on this target when that fn first runs, + because we don't know the connect flags yet. */ RETURN(0); - - if (tgt->ltd_exp) { - struct obd_device *osc_obd; - - osc_obd = class_exp2obd(tgt->ltd_exp); - if (osc_obd) - osc_obd->obd_no_recov = 0; } - rc = lov_connect_obd(obd, tgt, 1); - if (rc || !obd->obd_observer) - RETURN(rc); - - /* tell the mds_lov about the new target */ - obd_llog_finish(obd->obd_observer, &obd->obd_observer->obd_llogs, - old_count); - obd_llog_cat_initialize(obd->obd_observer, - &obd->obd_observer->obd_llogs, - lov->desc.ld_tgt_count, name); + lov_getref(obd); - params[0] = index; - rc = obd_get_info(tgt->ltd_exp, strlen("last_id"), "last_id", &size, - ¶ms[1]); + rc = lov_connect_obd(obd, index, active, &lov->lov_ocd); if (rc) GOTO(out, rc); - exp_observer = obd->obd_observer->obd_self_export; - rc = obd_set_info(exp_observer, strlen("next_id"),"next_id", 2, params); - if (rc) - GOTO(out, rc); + rc = lov_notify(obd, tgt->ltd_exp->exp_obd, + active ? OBD_NOTIFY_ACTIVE : OBD_NOTIFY_INACTIVE, + (void *)&index); - rc = lov_notify(obd, tgt->ltd_exp->exp_obd, 1); - GOTO(out, rc); - out: - if (rc && tgt->ltd_exp != NULL) - lov_disconnect_obd(obd, tgt, 0); - return rc; +out: + if (rc) { + CERROR("add failed (%d), deleting %s\n", rc, + obd_uuid2str(&tgt->ltd_uuid)); + lov_del_target(obd, index, 0, 0); + } + lov_putref(obd); + RETURN(rc); } -static int -lov_del_obd(struct obd_device *obd, struct obd_uuid *uuidp, int index, int gen) +/* Schedule a target for deletion */ +static int lov_del_target(struct obd_device *obd, __u32 index, + struct obd_uuid *uuidp, int gen) { struct lov_obd *lov = &obd->u.lov; - struct lov_tgt_desc *tgt; int count = lov->desc.ld_tgt_count; int rc = 0; ENTRY; - CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d\n", - uuidp->uuid, index, gen); - if (index >= count) { CERROR("LOV target index %d >= number of LOV OBDs %d.\n", index, count); RETURN(-EINVAL); } - tgt = lov->tgts + index; + lov_getref(obd); - if (obd_uuid_empty(&tgt->uuid)) { + if (!lov->lov_tgts[index]) { CERROR("LOV target at index %d is not setup.\n", index); - RETURN(-EINVAL); + GOTO(out, rc = -EINVAL); } - if (strncmp(uuidp->uuid, tgt->uuid.uuid, sizeof uuidp->uuid) != 0) { + if (uuidp && !obd_uuid_equals(uuidp, &lov->lov_tgts[index]->ltd_uuid)) { CERROR("LOV target UUID %s at index %d doesn't match %s.\n", - tgt->uuid.uuid, index, uuidp->uuid); - RETURN(-EINVAL); - } - - if (tgt->ltd_exp) { - struct obd_device *osc_obd; - - osc_obd = class_exp2obd(tgt->ltd_exp); - if (osc_obd) { - osc_obd->obd_no_recov = 1; - rc = obd_llog_finish(osc_obd, &osc_obd->obd_llogs, 1); - if (rc) - CERROR("osc_llog_finish error: %d\n", rc); - } - lov_disconnect_obd(obd, tgt, 0); + lov_uuid2str(lov, index), index, + obd_uuid2str(uuidp)); + GOTO(out, rc = -EINVAL); } - /* XXX - right now there is a dependency on ld_tgt_count being the - * maximum tgt index for computing the mds_max_easize. So we can't - * shrink it. */ - - /* lt_gen = 0 will mean it will not match the gen of any valid loi */ - memset(tgt, 0, sizeof(*tgt)); - CDEBUG(D_CONFIG, "uuid: %s idx: %d gen: %d exp: %p active: %d\n", - tgt->uuid.uuid, index, tgt->ltd_gen, tgt->ltd_exp, tgt->active); + lov_uuid2str(lov, index), index, + lov->lov_tgts[index]->ltd_gen, lov->lov_tgts[index]->ltd_exp, + lov->lov_tgts[index]->ltd_active); + + lov->lov_tgts[index]->ltd_reap = 1; + lov->lov_death_row++; + /* we really delete it from lov_putref */ +out: + lov_putref(obd); RETURN(rc); } -static int lov_process_config(struct obd_device *obd, obd_count len, void *buf) +/* We are holding lov_lock */ +static void __lov_del_obd(struct obd_device *obd, __u32 index) { - struct lustre_cfg *lcfg = buf; - struct obd_uuid obd_uuid; - int cmd; - int index; - int gen; - int rc = 0; - ENTRY; - - switch(cmd = lcfg->lcfg_command) { - case LCFG_LOV_ADD_OBD: - case LCFG_LOV_DEL_OBD: { - if (lcfg->lcfg_inllen1 > sizeof(obd_uuid.uuid)) - GOTO(out, rc = -EINVAL); + struct lov_obd *lov = &obd->u.lov; + struct obd_device *osc_obd; + struct lov_tgt_desc *tgt = lov->lov_tgts[index]; - obd_str2uuid(&obd_uuid, lcfg->lcfg_inlbuf1); + LASSERT(tgt); + LASSERT(tgt->ltd_reap); - if (sscanf(lcfg->lcfg_inlbuf2, "%d", &index) != 1) - GOTO(out, rc = -EINVAL); - if (sscanf(lcfg->lcfg_inlbuf3, "%d", &gen) != 1) - GOTO(out, rc = -EINVAL); - if (cmd == LCFG_LOV_ADD_OBD) - rc = lov_add_obd(obd, &obd_uuid, index, gen); - else - rc = lov_del_obd(obd, &obd_uuid, index, gen); - GOTO(out, rc); - } - default: { - CERROR("Unknown command: %d\n", lcfg->lcfg_command); - GOTO(out, rc = -EINVAL); + osc_obd = class_exp2obd(tgt->ltd_exp); - } - } -out: - RETURN(rc); -} + CDEBUG(D_CONFIG, "Removing tgt %s : %s\n", + lov_uuid2str(lov, index), + osc_obd ? osc_obd->obd_name : ""); -/* compute object size given "stripeno" and the ost size */ -static obd_size lov_stripe_size(struct lov_stripe_md *lsm, obd_size ost_size, - int stripeno) -{ - unsigned long ssize = lsm->lsm_stripe_size; - unsigned long swidth = ssize * lsm->lsm_stripe_count; - unsigned long stripe_size; - obd_size lov_size; + if (tgt->ltd_exp) + lov_disconnect_obd(obd, index); - if (ost_size == 0) - return 0; + /* XXX - right now there is a dependency on ld_tgt_count being the + * maximum tgt index for computing the mds_max_easize. So we can't + * shrink it. */ - /* do_div(a, b) returns a % b, and a = a / b */ - stripe_size = do_div(ost_size, ssize); - if (stripe_size) - lov_size = ost_size * swidth + stripeno * ssize + stripe_size; - else - lov_size = (ost_size - 1) * swidth + (stripeno + 1) * ssize; + lov->lov_tgts[index] = NULL; + OBD_FREE_PTR(tgt); - return lov_size; + /* Manual cleanup - no cleanup logs to clean up the osc's. We must + do it ourselves. And we can't do it from lov_cleanup, + because we just lost our only reference to it. */ + if (osc_obd) + class_manual_cleanup(osc_obd); } -static void lov_merge_attrs(struct obdo *tgt, struct obdo *src, obd_flag valid, - struct lov_stripe_md *lsm, int stripeno, int *set) +void lov_fix_desc(struct lov_desc *desc) { - valid &= src->o_valid; + if (desc->ld_default_stripe_size < PTLRPC_MAX_BRW_SIZE) { + LCONSOLE_WARN("Increasing default stripe size to min %u\n", + PTLRPC_MAX_BRW_SIZE); + desc->ld_default_stripe_size = PTLRPC_MAX_BRW_SIZE; + } else if (desc->ld_default_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) { + desc->ld_default_stripe_size &= ~(LOV_MIN_STRIPE_SIZE - 1); + LCONSOLE_WARN("Changing default stripe size to "LPU64" (a " + "multiple of %u)\n", + desc->ld_default_stripe_size,LOV_MIN_STRIPE_SIZE); + } - if (*set) { - if (valid & OBD_MD_FLSIZE) { - /* this handles sparse files properly */ - obd_size lov_size; + if (desc->ld_default_stripe_count == 0) + desc->ld_default_stripe_count = 1; - lov_size = lov_stripe_size(lsm, src->o_size, stripeno); - if (lov_size > tgt->o_size) - tgt->o_size = lov_size; - } - if (valid & OBD_MD_FLBLOCKS) - tgt->o_blocks += src->o_blocks; - if (valid & OBD_MD_FLBLKSZ) - tgt->o_blksize += src->o_blksize; - if (valid & OBD_MD_FLCTIME && tgt->o_ctime < src->o_ctime) - tgt->o_ctime = src->o_ctime; - if (valid & OBD_MD_FLMTIME && tgt->o_mtime < src->o_mtime) - tgt->o_mtime = src->o_mtime; - } else { - memcpy(tgt, src, sizeof(*tgt)); - tgt->o_id = lsm->lsm_object_id; - if (valid & OBD_MD_FLSIZE) - tgt->o_size = lov_stripe_size(lsm,src->o_size,stripeno); - *set = 1; + /* from lov_setstripe */ + if ((desc->ld_pattern != 0) && + (desc->ld_pattern != LOV_PATTERN_RAID0)) { + LCONSOLE_WARN("Unknown stripe pattern: %#x\n",desc->ld_pattern); + desc->ld_pattern = 0; } -} -#ifndef log2 -#define log2(n) ffz(~(n)) -#endif + if (desc->ld_qos_maxage == 0) + desc->ld_qos_maxage = QOS_DEFAULT_MAXAGE; +} -static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa, - struct lov_stripe_md **ea, - struct obd_trans_info *oti) +static int lov_setup(struct obd_device *obd, struct lustre_cfg *lcfg) { - struct lov_obd *lov; - struct obdo *tmp_oa; - struct obd_uuid *ost_uuid = NULL; - int rc = 0, i; + struct lprocfs_static_vars lvars; + struct lov_desc *desc; + struct lov_obd *lov = &obd->u.lov; + int count; ENTRY; - LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS && - src_oa->o_flags == OBD_FL_DELORPHAN); + if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) { + CERROR("LOV setup requires a descriptor\n"); + RETURN(-EINVAL); + } - lov = &export->exp_obd->u.lov; + desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1); - tmp_oa = obdo_alloc(); - if (tmp_oa == NULL) - RETURN(-ENOMEM); + 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); + } - if (src_oa->o_valid & OBD_MD_FLINLINE) { - ost_uuid = (struct obd_uuid *)src_oa->o_inline; - CDEBUG(D_HA, "clearing orphans only for %s\n", - ost_uuid->uuid); + if (desc->ld_magic != LOV_DESC_MAGIC) { + if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) { + CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n", + obd->obd_name, desc); + lustre_swab_lov_desc(desc); + } else { + CERROR("%s: Bad lov desc magic: %#x\n", + obd->obd_name, desc->ld_magic); + RETURN(-EINVAL); + } } + lov_fix_desc(desc); + + /* Because of 64-bit divide/mod operations only work with a 32-bit + * divisor in a 32-bit kernel, we cannot support a stripe width + * of 4GB or larger on 32-bit CPUs. */ + count = desc->ld_default_stripe_count; + if ((count > 0 ? count : desc->ld_tgt_count) * + desc->ld_default_stripe_size > 0xffffffff) { + CERROR("LOV: stripe width "LPU64"x%u > 4294967295 bytes\n", + desc->ld_default_stripe_size, count); + RETURN(-EINVAL); + } + + desc->ld_active_tgt_count = 0; + lov->desc = *desc; + lov->lov_tgt_size = 0; + sema_init(&lov->lov_lock, 1); + atomic_set(&lov->lov_refcount, 0); + INIT_LIST_HEAD(&lov->lov_qos.lq_oss_list); + init_rwsem(&lov->lov_qos.lq_rw_sem); + lov->lov_qos.lq_dirty = 1; + lov->lov_qos.lq_dirty_rr = 1; + lov->lov_qos.lq_reset = 1; + /* Default priority is toward free space balance */ + lov->lov_qos.lq_prio_free = 232; + + lprocfs_init_vars(lov, &lvars); + lprocfs_obd_setup(obd, lvars.obd_vars); +#ifdef LPROCFS + { + cfs_proc_dir_entry_t *entry; + + entry = create_proc_entry("target_obd", 0444, + obd->obd_proc_entry); + if (entry != NULL) { + entry->proc_fops = &lov_proc_target_fops; + entry->data = obd; + } + } +#endif + + RETURN(0); +} + +static int lov_precleanup(struct obd_device *obd, enum obd_cleanup_stage stage) +{ + int rc = 0; + ENTRY; + + switch (stage) { + case OBD_CLEANUP_EARLY: { + struct lov_obd *lov = &obd->u.lov; + int i; + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) + continue; + obd_precleanup(class_exp2obd(lov->lov_tgts[i]->ltd_exp), + OBD_CLEANUP_EARLY); + } + break; + } + case OBD_CLEANUP_EXPORTS: + break; + case OBD_CLEANUP_SELF_EXP: + rc = obd_llog_finish(obd, 0); + if (rc != 0) + CERROR("failed to cleanup llogging subsystems\n"); + break; + case OBD_CLEANUP_OBD: + break; + } + RETURN(rc); +} + +static int lov_cleanup(struct obd_device *obd) +{ + struct lov_obd *lov = &obd->u.lov; + + lprocfs_obd_cleanup(obd); + if (lov->lov_tgts) { + int i; + for (i = 0; i < lov->desc.ld_tgt_count; i++) { + if (lov->lov_tgts[i]) { + /* Inactive targets may never have connected */ + if (lov->lov_tgts[i]->ltd_active || + atomic_read(&lov->lov_refcount)) + /* We should never get here - these + should have been removed in the + disconnect. */ + CERROR("lov tgt %d not cleaned!" + " deathrow=%d, lovrc=%d\n", + i, lov->lov_death_row, + atomic_read(&lov->lov_refcount)); + lov_del_target(obd, i, 0, 0); + } + } + OBD_FREE(lov->lov_tgts, sizeof(*lov->lov_tgts) * + lov->lov_tgt_size); + lov->lov_tgt_size = 0; + } + + if (lov->lov_qos.lq_rr_size) + OBD_FREE(lov->lov_qos.lq_rr_array, lov->lov_qos.lq_rr_size); + + RETURN(0); +} + +static int lov_process_config(struct obd_device *obd, obd_count len, void *buf) +{ + struct lustre_cfg *lcfg = buf; + struct obd_uuid obd_uuid; + int cmd; + int rc = 0; + ENTRY; + + switch(cmd = lcfg->lcfg_command) { + case LCFG_LOV_ADD_OBD: + case LCFG_LOV_ADD_INA: + case LCFG_LOV_DEL_OBD: { + __u32 index; + int gen; + /* lov_modify_tgts add 0:lov_mdsA 1:ost1_UUID 2:0 3:1 */ + if (LUSTRE_CFG_BUFLEN(lcfg, 1) > sizeof(obd_uuid.uuid)) + GOTO(out, rc = -EINVAL); + + obd_str2uuid(&obd_uuid, lustre_cfg_buf(lcfg, 1)); + + if (sscanf(lustre_cfg_buf(lcfg, 2), "%d", &index) != 1) + GOTO(out, rc = -EINVAL); + if (sscanf(lustre_cfg_buf(lcfg, 3), "%d", &gen) != 1) + GOTO(out, rc = -EINVAL); + if (cmd == LCFG_LOV_ADD_OBD) + rc = lov_add_target(obd, &obd_uuid, index, gen, 1); + else if (cmd == LCFG_LOV_ADD_INA) + rc = lov_add_target(obd, &obd_uuid, index, gen, 0); + else + rc = lov_del_target(obd, index, &obd_uuid, gen); + GOTO(out, rc); + } + case LCFG_PARAM: { + struct lprocfs_static_vars lvars; + struct lov_desc *desc = &(obd->u.lov.desc); + + if (!desc) + GOTO(out, rc = -EINVAL); + + lprocfs_init_vars(lov, &lvars); + + rc = class_process_proc_param(PARAM_LOV, lvars.obd_vars, + lcfg, obd); + GOTO(out, rc); + } + default: { + CERROR("Unknown command: %d\n", lcfg->lcfg_command); + GOTO(out, rc = -EINVAL); + + } + } +out: + RETURN(rc); +} + +#ifndef log2 +#define log2(n) ffz(~(n)) +#endif + +static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa, + struct lov_stripe_md **ea, + struct obd_trans_info *oti) +{ + struct lov_obd *lov; + struct obdo *tmp_oa; + struct obd_uuid *ost_uuid = NULL; + int rc = 0, i; + ENTRY; + + LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS && + src_oa->o_flags == OBD_FL_DELORPHAN); + + lov = &export->exp_obd->u.lov; + + OBDO_ALLOC(tmp_oa); + if (tmp_oa == NULL) + RETURN(-ENOMEM); + + if (src_oa->o_valid & OBD_MD_FLINLINE) { + ost_uuid = (struct obd_uuid *)src_oa->o_inline; + CDEBUG(D_HA, "clearing orphans only for %s\n", + ost_uuid->uuid); + } + + lov_getref(export->exp_obd); for (i = 0; i < lov->desc.ld_tgt_count; i++) { struct lov_stripe_md obj_md; struct lov_stripe_md *obj_mdp = &obj_md; + struct lov_tgt_desc *tgt; int err; + tgt = lov->lov_tgts[i]; + if (!tgt) + continue; + /* if called for a specific target, we don't care if it is not active. */ - if (lov->tgts[i].active == 0 && ost_uuid == NULL) { + if (!lov->lov_tgts[i]->ltd_active == 0 && ost_uuid == NULL) { CDEBUG(D_HA, "lov idx %d inactive\n", i); continue; } - if (ost_uuid && !obd_uuid_equals(ost_uuid, &lov->tgts[i].uuid)) + if (ost_uuid && !obd_uuid_equals(ost_uuid, &tgt->ltd_uuid)) continue; + CDEBUG(D_CONFIG,"Clear orphans for %d:%s\n", i, + obd_uuid2str(ost_uuid)); + memcpy(tmp_oa, src_oa, sizeof(*tmp_oa)); + LASSERT(lov->lov_tgts[i]->ltd_exp); /* XXX: LOV STACKING: use real "obj_mdp" sub-data */ - err = obd_create(lov->tgts[i].ltd_exp, tmp_oa, &obj_mdp, oti); + err = obd_create(lov->lov_tgts[i]->ltd_exp, + tmp_oa, &obj_mdp, oti); if (err) /* This export will be disabled until it is recovered, and then orphan recovery will be completed. */ @@ -825,27 +928,68 @@ static int lov_clear_orphans(struct obd_export *export, struct obdo *src_oa, if (ost_uuid) break; } - obdo_free(tmp_oa); + lov_putref(export->exp_obd); + + OBDO_FREE(tmp_oa); RETURN(rc); } -#define LOV_CREATE_RESEED_INTERVAL 1000 +static int lov_recreate(struct obd_export *exp, struct obdo *src_oa, + struct lov_stripe_md **ea, struct obd_trans_info *oti) +{ + struct lov_stripe_md *obj_mdp, *lsm; + struct lov_obd *lov = &exp->exp_obd->u.lov; + unsigned ost_idx; + int rc, i; + ENTRY; + + LASSERT(src_oa->o_valid & OBD_MD_FLFLAGS && + src_oa->o_flags & OBD_FL_RECREATE_OBJS); + + OBD_ALLOC(obj_mdp, sizeof(*obj_mdp)); + if (obj_mdp == NULL) + RETURN(-ENOMEM); + + ost_idx = src_oa->o_nlink; + lsm = *ea; + if (lsm == NULL) + GOTO(out, rc = -EINVAL); + if (ost_idx >= lov->desc.ld_tgt_count || + !lov->lov_tgts[ost_idx]) + GOTO(out, rc = -EINVAL); + + for (i = 0; i < lsm->lsm_stripe_count; i++) { + if (lsm->lsm_oinfo[i]->loi_ost_idx == ost_idx) { + if (lsm->lsm_oinfo[i]->loi_id != src_oa->o_id) + GOTO(out, rc = -EINVAL); + break; + } + } + if (i == lsm->lsm_stripe_count) + GOTO(out, rc = -EINVAL); + + rc = obd_create(lov->lov_tgts[ost_idx]->ltd_exp, src_oa, &obj_mdp, oti); +out: + OBD_FREE(obj_mdp, sizeof(*obj_mdp)); + RETURN(rc); +} /* the LOV expects oa->o_id to be set to the LOV object id */ static int lov_create(struct obd_export *exp, struct obdo *src_oa, struct lov_stripe_md **ea, struct obd_trans_info *oti) { - static int ost_start_idx, ost_start_count; struct lov_obd *lov; - struct lov_stripe_md *lsm; - struct lov_oinfo *loi = NULL; - struct obdo *tmp_oa, *ret_oa; - struct llog_cookie *cookies = NULL; - unsigned ost_count, ost_idx; - int set = 0, obj_alloc = 0, cookie_sent = 0, rc = 0, i; + struct obd_info oinfo; + struct lov_request_set *set = NULL; + struct lov_request *req; + struct obd_statfs osfs; + __u64 maxage; + int rc = 0; ENTRY; LASSERT(ea != NULL); + if (exp == NULL) + RETURN(-EINVAL); if ((src_oa->o_valid & OBD_MD_FLFLAGS) && src_oa->o_flags == OBD_FL_DELORPHAN) { @@ -853,1410 +997,711 @@ static int lov_create(struct obd_export *exp, struct obdo *src_oa, RETURN(rc); } - if (exp == NULL) - RETURN(-EINVAL); - lov = &exp->exp_obd->u.lov; - if (!lov->desc.ld_active_tgt_count) RETURN(-EIO); /* Recreate a specific object id at the given OST index */ if ((src_oa->o_valid & OBD_MD_FLFLAGS) && (src_oa->o_flags & OBD_FL_RECREATE_OBJS)) { - struct lov_stripe_md obj_md; - struct lov_stripe_md *obj_mdp = &obj_md; - - ost_idx = src_oa->o_nlink; - lsm = *ea; - if (lsm == NULL) - RETURN(-EINVAL); - if (ost_idx >= lov->desc.ld_tgt_count) - RETURN(-EINVAL); - for (i = 0; i < lsm->lsm_stripe_count; i++) { - if (lsm->lsm_oinfo[i].loi_ost_idx == ost_idx) { - if (lsm->lsm_oinfo[i].loi_id != src_oa->o_id || - lsm->lsm_oinfo[i].loi_gr != src_oa->o_gr) { - RETURN(-EINVAL); - } - break; - } - } - if (i == lsm->lsm_stripe_count) - RETURN(-EINVAL); - - rc = obd_create(lov->tgts[ost_idx].ltd_exp, src_oa, - &obj_mdp, oti); + rc = lov_recreate(exp, src_oa, ea, oti); RETURN(rc); } - ret_oa = obdo_alloc(); - if (!ret_oa) - RETURN(-ENOMEM); - - tmp_oa = obdo_alloc(); - if (!tmp_oa) - GOTO(out_oa, rc = -ENOMEM); - - lsm = *ea; - if (lsm == NULL) { - int stripes; - ost_count = lov_get_stripecnt(lov, 0); - - /* If the MDS file was truncated up to some size, stripe over - * enough OSTs to allow the file to be created at that size. */ - if (src_oa->o_valid & OBD_MD_FLSIZE) { - stripes=((src_oa->o_size+LUSTRE_STRIPE_MAXBYTES)>>12)-1; - do_div(stripes, (__u32)(LUSTRE_STRIPE_MAXBYTES >> 12)); - - if (stripes > lov->desc.ld_active_tgt_count) - RETURN(-EFBIG); - if (stripes < ost_count) - stripes = ost_count; - } else { - stripes = ost_count; - } - - rc = lov_alloc_memmd(&lsm, stripes, lov->desc.ld_pattern ? - lov->desc.ld_pattern : LOV_PATTERN_RAID0); - if (rc < 0) - GOTO(out_tmp, rc); - - rc = 0; - } - - ost_count = lov->desc.ld_tgt_count; - - LASSERT(src_oa->o_gr > 0); - LASSERT(src_oa->o_valid & OBD_MD_FLID); - lsm->lsm_object_id = src_oa->o_id; - lsm->lsm_object_gr = src_oa->o_gr; - if (!lsm->lsm_stripe_size) - lsm->lsm_stripe_size = lov->desc.ld_default_stripe_size; - if (!lsm->lsm_pattern) { - lsm->lsm_pattern = lov->desc.ld_pattern ? - lov->desc.ld_pattern : LOV_PATTERN_RAID0; - } - - if (*ea == NULL || lsm->lsm_oinfo[0].loi_ost_idx >= ost_count) { - if (--ost_start_count <= 0) { - ost_start_idx = ll_insecure_random_int(); - ost_start_count = LOV_CREATE_RESEED_INTERVAL; - } else if (lsm->lsm_stripe_count >= - lov->desc.ld_active_tgt_count) { - /* If we allocate from all of the stripes, make the - * next file start on the next OST. */ - ++ost_start_idx; - } - ost_idx = ost_start_idx % ost_count; - } else { - ost_idx = lsm->lsm_oinfo[0].loi_ost_idx; - } - - CDEBUG(D_INODE, "allocating %d subobjs for objid "LPX64" at idx %d\n", - lsm->lsm_stripe_count, lsm->lsm_object_id, ost_idx); - - /* XXX LOV STACKING: need to figure out how many real OSCs */ - if (oti && (src_oa->o_valid & OBD_MD_FLCOOKIE)) { - oti_alloc_cookies(oti, lsm->lsm_stripe_count); - if (!oti->oti_logcookies) - GOTO(out_cleanup, rc = -ENOMEM); - cookies = oti->oti_logcookies; - } - - loi = lsm->lsm_oinfo; - for (i = 0; i < ost_count; i++, ost_idx = (ost_idx + 1) % ost_count) { - struct lov_stripe_md obj_md; - struct lov_stripe_md *obj_mdp = &obj_md; - int err; - - ++ost_start_idx; - if (lov->tgts[ost_idx].active == 0) { - if (!obd_uuid_empty(&lov->tgts[ost_idx].uuid)) - CDEBUG(D_HA, "lov idx %d inactive\n", ost_idx); - continue; - } - - /* create data objects with "parent" OA */ - memcpy(tmp_oa, src_oa, sizeof(*tmp_oa)); - - /* XXX When we start creating objects on demand, we need to - * make sure that we always create the object on the - * stripe which holds the existing file size. - */ - if (src_oa->o_valid & OBD_MD_FLSIZE) { - if (lov_stripe_offset(lsm, src_oa->o_size, i, - &tmp_oa->o_size) < 0 && - tmp_oa->o_size) - tmp_oa->o_size--; - - CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n", - i, tmp_oa->o_size, src_oa->o_size); - } + maxage = cfs_time_shift_64(-lov->desc.ld_qos_maxage); + obd_statfs_rqset(exp->exp_obd, &osfs, maxage); + rc = lov_prep_create_set(exp, &oinfo, ea, src_oa, oti, &set); + if (rc) + RETURN(rc); + list_for_each_entry(req, &set->set_list, rq_link) { /* XXX: LOV STACKING: use real "obj_mdp" sub-data */ - err = obd_create(lov->tgts[ost_idx].ltd_exp, tmp_oa, &obj_mdp, - oti); - if (err) { - if (lov->tgts[ost_idx].active) { - CERROR("error creating objid "LPX64" sub-object" - " on OST idx %d/%d: rc = %d\n", - src_oa->o_id, ost_idx, - lsm->lsm_stripe_count, err); - if (err > 0) { - CERROR("obd_create returned invalid " - "err %d\n", err); - err = -EIO; - } - } - if (!rc) - rc = err; - continue; - } - if (oti->oti_objid) - oti->oti_objid[ost_idx] = tmp_oa->o_id; - loi->loi_id = tmp_oa->o_id; - loi->loi_gr = tmp_oa->o_gr; - loi->loi_ost_idx = ost_idx; - loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen; - CDEBUG(D_INODE, "objid "LPX64" has subobj "LPX64" at " - "idx %d gen %d\n", lsm->lsm_object_id, loi->loi_id, - ost_idx, loi->loi_ost_gen); - - lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, - obj_alloc, &set); - loi_init(loi); - - if (cookies) - ++oti->oti_logcookies; - if (tmp_oa->o_valid & OBD_MD_FLCOOKIE) - ++cookie_sent; - ++obj_alloc; - ++loi; - - /* If we have allocated enough objects, we are OK */ - if (obj_alloc == lsm->lsm_stripe_count) - GOTO(out_done, rc = 0); - } - - if (obj_alloc == 0) { - if (rc == 0) - rc = -EIO; - GOTO(out_cleanup, rc); - } - - /* If we were passed specific striping params, then a failure to - * meet those requirements is an error, since we can't reallocate - * that memory (it might be part of a larger array or something). - * - * We can only get here if lsm_stripe_count was originally > 1. - */ - if (*ea != NULL) { - CERROR("can't lstripe objid "LPX64": have %u want %u, rc %d\n", - lsm->lsm_object_id, obj_alloc, lsm->lsm_stripe_count,rc); - if (rc == 0) - rc = -EFBIG; - GOTO(out_cleanup, rc); - } else { - struct lov_stripe_md *lsm_new; - /* XXX LOV STACKING call into osc for sizes */ - unsigned oldsize, newsize; - - if (oti && cookies && cookie_sent) { - oldsize = lsm->lsm_stripe_count * sizeof(*cookies); - newsize = obj_alloc * sizeof(*cookies); - - oti_alloc_cookies(oti, obj_alloc); - if (oti->oti_logcookies) { - memcpy(oti->oti_logcookies, cookies, newsize); - OBD_FREE(cookies, oldsize); - cookies = oti->oti_logcookies; - } else { - CWARN("'leaking' %d bytes\n", oldsize-newsize); - } - } - - CWARN("using fewer stripes for object "LPX64": old %u new %u\n", - lsm->lsm_object_id, lsm->lsm_stripe_count, obj_alloc); - oldsize = lov_stripe_md_size(lsm->lsm_stripe_count); - newsize = lov_stripe_md_size(obj_alloc); - OBD_ALLOC(lsm_new, newsize); - if (lsm_new != NULL) { - memcpy(lsm_new, lsm, newsize); - lsm_new->lsm_stripe_count = obj_alloc; - OBD_FREE(lsm, oldsize); - lsm = lsm_new; - } else { - CWARN("'leaking' %d bytes\n", oldsize - newsize); - } - rc = 0; - } - EXIT; - out_done: - *ea = lsm; - if (src_oa->o_valid & OBD_MD_FLSIZE && - ret_oa->o_size != src_oa->o_size) { - CERROR("original size "LPU64" isn't new object size "LPU64"\n", - src_oa->o_size, ret_oa->o_size); - LBUG(); - } - ret_oa->o_id = src_oa->o_id; - ret_oa->o_gr = src_oa->o_gr; - ret_oa->o_valid |= OBD_MD_FLGROUP; - memcpy(src_oa, ret_oa, sizeof(*src_oa)); - - out_tmp: - obdo_free(tmp_oa); - out_oa: - obdo_free(ret_oa); - if (oti && cookies) { - oti->oti_logcookies = cookies; - if (!cookie_sent) { - oti_free_cookies(oti); - src_oa->o_valid &= ~OBD_MD_FLCOOKIE; - } else { - src_oa->o_valid |= OBD_MD_FLCOOKIE; - } + rc = obd_create(lov->lov_tgts[req->rq_idx]->ltd_exp, + req->rq_oi.oi_oa, &req->rq_oi.oi_md, oti); + lov_update_create_set(set, req, rc); } + rc = lov_fini_create_set(set, ea); RETURN(rc); - - out_cleanup: - while (obj_alloc-- > 0) { - struct obd_export *sub_exp; - int err; - - --loi; - sub_exp = lov->tgts[loi->loi_ost_idx].ltd_exp; - /* destroy already created objects here */ - memcpy(tmp_oa, src_oa, sizeof(*tmp_oa)); - tmp_oa->o_id = loi->loi_id; - - err = obd_destroy(sub_exp, tmp_oa, NULL, oti); - if (err) - CERROR("Failed to uncreate objid "LPX64" subobj "LPX64 - " on OST idx %d: rc = %d\n", src_oa->o_id, - loi->loi_id, loi->loi_ost_idx, err); - } - if (*ea == NULL) - obd_free_memmd(exp, &lsm); - goto out_tmp; -} - -static int lov_revalidate_policy(struct lov_obd *lov, struct lov_stripe_md *lsm) -{ - static int last_idx = 0; - struct lov_tgt_desc *tgt; - int i, count; - - /* XXX - we should do something clever and take lsm - * into account but just do round robin for now. */ - - /* last_idx must always be less that count because - * ld_tgt_count currently cannot shrink. */ - count = lov->desc.ld_tgt_count; - - for (i = last_idx, tgt = lov->tgts + i; i < count; i++, tgt++) - if (tgt->active) - RETURN(last_idx = i); - - for (i = 0, tgt = lov->tgts; i < last_idx; i++, tgt++) - if (tgt->active) - RETURN(last_idx = i); - - RETURN(-EIO); } -#define lsm_bad_magic(LSMP) \ -({ \ - struct lov_stripe_md *_lsm__ = (LSMP); \ - int _ret__ = 0; \ - if (!_lsm__) { \ - CERROR("LOV requires striping ea\n"); \ - _ret__ = 1; \ - } else if (_lsm__->lsm_magic != LOV_MAGIC) { \ - CERROR("LOV striping magic bad %#x != %#x\n", \ - _lsm__->lsm_magic, LOV_MAGIC); \ - _ret__ = 1; \ - } \ - _ret__; \ -}) +#define ASSERT_LSM_MAGIC(lsmp) \ +do { \ + LASSERT((lsmp) != NULL); \ + LASSERTF(((lsmp)->lsm_magic == LOV_MAGIC || \ + (lsmp)->lsm_magic == LOV_MAGIC_JOIN), "%p->lsm_magic=%x\n", \ + (lsmp), (lsmp)->lsm_magic); \ +} while (0) static int lov_destroy(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *lsm, struct obd_trans_info *oti) + struct lov_stripe_md *lsm, struct obd_trans_info *oti, + struct obd_export *md_exp) { - struct obdo tmp; + struct lov_request_set *set; + struct obd_info oinfo; + struct lov_request *req; + struct list_head *pos; struct lov_obd *lov; - struct lov_oinfo *loi; - int rc = 0, i; + int rc = 0, err; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); if (!exp || !exp->exp_obd) RETURN(-ENODEV); + if (oa->o_valid & OBD_MD_FLCOOKIE) { + LASSERT(oti); + LASSERT(oti->oti_logcookies); + } + lov = &exp->exp_obd->u.lov; - loi = lsm->lsm_oinfo; - for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) { + rc = lov_prep_destroy_set(exp, &oinfo, oa, lsm, oti, &set); + if (rc) + RETURN(rc); + + list_for_each (pos, &set->set_list) { int err; + req = list_entry(pos, struct lov_request, rq_link); - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - /* Orphan clean up will (someday) fix this up. */ - if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) - oti->oti_logcookies++; - continue; - } + if (oa->o_valid & OBD_MD_FLCOOKIE) + oti->oti_logcookies = set->set_cookies + req->rq_stripe; - memcpy(&tmp, oa, sizeof(tmp)); - tmp.o_id = loi->loi_id; - err = obd_destroy(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, - NULL, oti); - if (err && lov->tgts[loi->loi_ost_idx].active) { - CDEBUG(D_INODE, "error: destroying objid "LPX64" subobj " + err = obd_destroy(lov->lov_tgts[req->rq_idx]->ltd_exp, + req->rq_oi.oi_oa, NULL, oti, NULL); + err = lov_update_common_set(set, req, err); + if (err) { + CERROR("error: destroying objid "LPX64" subobj " LPX64" on OST idx %d: rc = %d\n", - oa->o_id, loi->loi_id, loi->loi_ost_idx, err); + oa->o_id, req->rq_oi.oi_oa->o_id, + req->rq_idx, err); if (!rc) rc = err; } } - RETURN(rc); + + if (rc == 0) { + LASSERT(lsm_op_find(lsm->lsm_magic) != NULL); + rc = lsm_op_find(lsm->lsm_magic)->lsm_destroy(lsm, oa, md_exp); + } + err = lov_fini_destroy_set(set); + RETURN(rc ? rc : err); } -static int lov_getattr(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *lsm) +static int lov_getattr(struct obd_export *exp, struct obd_info *oinfo) { - struct obdo tmp; + struct lov_request_set *set; + struct lov_request *req; + struct list_head *pos; struct lov_obd *lov; - struct lov_oinfo *loi; - int i, rc = 0, set = 0; + int err = 0, rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); if (!exp || !exp->exp_obd) RETURN(-ENODEV); lov = &exp->exp_obd->u.lov; - CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n", - lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size); - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { - int err; + rc = lov_prep_getattr_set(exp, oinfo, &set); + if (rc) + RETURN(rc); - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - continue; - } + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx " - "%u\n", oa->o_id, i, loi->loi_id, loi->loi_ost_idx); - /* create data objects with "parent" OA */ - memcpy(&tmp, oa, sizeof(tmp)); - tmp.o_id = loi->loi_id; + "%u\n", oinfo->oi_oa->o_id, req->rq_stripe, + req->rq_oi.oi_oa->o_id, req->rq_idx); - err = obd_getattr(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, - NULL); + rc = obd_getattr(lov->lov_tgts[req->rq_idx]->ltd_exp, + &req->rq_oi); + err = lov_update_common_set(set, req, rc); if (err) { - if (lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: getattr objid "LPX64" subobj " - LPX64" on OST idx %d: rc = %d\n", - oa->o_id, loi->loi_id, loi->loi_ost_idx, - err); - RETURN(err); - } - } else { - lov_merge_attrs(oa, &tmp, tmp.o_valid, lsm, i, &set); + CERROR("error: getattr objid "LPX64" subobj " + LPX64" on OST idx %d: rc = %d\n", + oinfo->oi_oa->o_id, req->rq_oi.oi_oa->o_id, + req->rq_idx, err); + break; } } - if (!set) - rc = -EIO; + + rc = lov_fini_getattr_set(set); + if (err) + rc = err; RETURN(rc); } -static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, void *data, - int rc) +static int lov_getattr_interpret(struct ptlrpc_request_set *rqset, + void *data, int rc) { - struct lov_getattr_async_args *aa = data; - struct lov_stripe_md *lsm = aa->aa_lsm; - struct obdo *oa = aa->aa_oa; - struct obdo *obdos = aa->aa_obdos; - struct lov_oinfo *loi; - int i; - int set = 0; + struct lov_request_set *lovset = (struct lov_request_set *)data; + int err; ENTRY; - if (rc == 0) { - /* NB all stripe requests succeeded to get here */ - - loi = lsm->lsm_oinfo; - for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) { - if (obdos[i].o_valid == 0) /* inactive stripe */ - continue; - - lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm, - i, &set); - } - - if (!set) { - CERROR ("No stripes had valid attrs\n"); - rc = -EIO; - } - } - - OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos)); - RETURN (rc); + /* don't do attribute merge if this aysnc op failed */ + if (rc) + lovset->set_completes = 0; + err = lov_fini_getattr_set(lovset); + RETURN(rc ? rc : err); } -static int lov_getattr_async(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *lsm, +static int lov_getattr_async(struct obd_export *exp, struct obd_info *oinfo, struct ptlrpc_request_set *rqset) { - struct obdo *obdos; + struct lov_request_set *lovset; struct lov_obd *lov; - struct lov_oinfo *loi; - struct lov_getattr_async_args *aa; - int i, rc = 0, set = 0; + struct list_head *pos; + struct lov_request *req; + int rc = 0, err; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); if (!exp || !exp->exp_obd) RETURN(-ENODEV); lov = &exp->exp_obd->u.lov; - OBD_ALLOC (obdos, lsm->lsm_stripe_count * sizeof (*obdos)); - if (obdos == NULL) - RETURN(-ENOMEM); + rc = lov_prep_getattr_set(exp, oinfo, &lovset); + if (rc) + RETURN(rc); CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n", - lsm->lsm_object_id, lsm->lsm_stripe_count, lsm->lsm_stripe_size); - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { - int err; - - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - /* leaves obdos[i].obd_valid unset */ - continue; - } + oinfo->oi_md->lsm_object_id, oinfo->oi_md->lsm_stripe_count, + oinfo->oi_md->lsm_stripe_size); - CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at " - "idx %u gen %d\n", oa->o_id, i, loi->loi_id, - loi->loi_ost_idx, loi->loi_ost_gen); + list_for_each (pos, &lovset->set_list) { + req = list_entry(pos, struct lov_request, rq_link); - /* create data objects with "parent" OA */ - memcpy(&obdos[i], oa, sizeof(obdos[i])); - obdos[i].o_id = loi->loi_id; - - err = obd_getattr_async(lov->tgts[loi->loi_ost_idx].ltd_exp, - &obdos[i], NULL, rqset); - if (err) { + CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx " + "%u\n", oinfo->oi_oa->o_id, req->rq_stripe, + req->rq_oi.oi_oa->o_id, req->rq_idx); + rc = obd_getattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp, + &req->rq_oi, rqset); + if (rc) { CERROR("error: getattr objid "LPX64" subobj " LPX64" on OST idx %d: rc = %d\n", - oa->o_id, loi->loi_id, loi->loi_ost_idx, - err); - GOTO(out_obdos, rc = err); + oinfo->oi_oa->o_id, req->rq_oi.oi_oa->o_id, + req->rq_idx, rc); + GOTO(out, rc); } - set = 1; - } - if (!set) - GOTO (out_obdos, rc = -EIO); - - LASSERT (rqset->set_interpret == NULL); - rqset->set_interpret = lov_getattr_interpret; - LASSERT (sizeof (rqset->set_args) >= sizeof (*aa)); - aa = (struct lov_getattr_async_args *)&rqset->set_args; - aa->aa_lsm = lsm; - aa->aa_oa = oa; - aa->aa_obdos = obdos; - aa->aa_lov = lov; - GOTO(out, rc = 0); - -out_obdos: - OBD_FREE (obdos, lsm->lsm_stripe_count * sizeof (*obdos)); + } + + if (!list_empty(&rqset->set_requests)) { + LASSERT(rc == 0); + LASSERT (rqset->set_interpret == NULL); + rqset->set_interpret = lov_getattr_interpret; + rqset->set_arg = (void *)lovset; + RETURN(rc); + } out: - RETURN(rc); + if (rc) + lovset->set_completes = 0; + err = lov_fini_getattr_set(lovset); + RETURN(rc ? rc : err); } - -static int lov_setattr(struct obd_export *exp, struct obdo *src_oa, - struct lov_stripe_md *lsm, struct obd_trans_info *oti) +static int lov_setattr(struct obd_export *exp, struct obd_info *oinfo, + struct obd_trans_info *oti) { - struct obdo *tmp_oa, *ret_oa; + struct lov_request_set *set; struct lov_obd *lov; - struct lov_oinfo *loi; - int rc = 0, i, set = 0; + struct list_head *pos; + struct lov_request *req; + int err = 0, rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); if (!exp || !exp->exp_obd) RETURN(-ENODEV); - /* for now, we only expect time updates here */ - LASSERT(!(src_oa->o_valid & ~(OBD_MD_FLID|OBD_MD_FLTYPE | OBD_MD_FLMODE| - OBD_MD_FLATIME | OBD_MD_FLMTIME | - OBD_MD_FLCTIME | OBD_MD_FLFLAGS | - OBD_MD_FLSIZE | OBD_MD_FLGROUP))); - - LASSERT(!(src_oa->o_valid & OBD_MD_FLGROUP) || src_oa->o_gr > 0); - - ret_oa = obdo_alloc(); - if (!ret_oa) - RETURN(-ENOMEM); - - tmp_oa = obdo_alloc(); - if (!tmp_oa) - GOTO(out_oa, rc = -ENOMEM); - + /* for now, we only expect the following updates here */ + LASSERT(!(oinfo->oi_oa->o_valid & ~(OBD_MD_FLID | OBD_MD_FLTYPE | + OBD_MD_FLMODE | OBD_MD_FLATIME | + OBD_MD_FLMTIME | OBD_MD_FLCTIME | + OBD_MD_FLFLAGS | OBD_MD_FLSIZE | + OBD_MD_FLGROUP | OBD_MD_FLUID | + OBD_MD_FLGID | OBD_MD_FLINLINE | + OBD_MD_FLFID | OBD_MD_FLGENER))); lov = &exp->exp_obd->u.lov; - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { - int err; - - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - continue; - } - - memcpy(tmp_oa, src_oa, sizeof(*tmp_oa)); - tmp_oa->o_id = loi->loi_id; - LASSERT(!(tmp_oa->o_valid & OBD_MD_FLGROUP) || tmp_oa->o_gr>0); - - if (src_oa->o_valid & OBD_MD_FLSIZE) { - if (lov_stripe_offset(lsm, src_oa->o_size, i, - &tmp_oa->o_size) < 0 && - tmp_oa->o_size) - tmp_oa->o_size--; + rc = lov_prep_setattr_set(exp, oinfo, oti, &set); + if (rc) + RETURN(rc); - CDEBUG(D_INODE, "stripe %d has size "LPU64"/"LPU64"\n", - i, tmp_oa->o_size, src_oa->o_size); - } + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); - err = obd_setattr(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp_oa, - NULL, NULL); + rc = obd_setattr(lov->lov_tgts[req->rq_idx]->ltd_exp, + &req->rq_oi, NULL); + err = lov_update_setattr_set(set, req, rc); if (err) { - if (lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: setattr objid "LPX64" subobj " - LPX64" on OST idx %d: rc = %d\n", - src_oa->o_id, loi->loi_id, - loi->loi_ost_idx, err); - if (!rc) - rc = err; - } - continue; - } - lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, i, &set); - } - if (!set && !rc) - rc = -EIO; - - ret_oa->o_id = src_oa->o_id; - memcpy(src_oa, ret_oa, sizeof(*src_oa)); - GOTO(out_tmp, rc); -out_tmp: - obdo_free(tmp_oa); -out_oa: - obdo_free(ret_oa); - return rc; -} - -/* we have an offset in file backed by an lov and want to find out where - * that offset lands in our given stripe of the file. for the easy - * case where the offset is within the stripe, we just have to scale the - * offset down to make it relative to the stripe instead of the lov. - * - * the harder case is what to do when the offset doesn't intersect the - * stripe. callers will want start offsets clamped ahead to the start - * of the nearest stripe in the file. end offsets similarly clamped to the - * nearest ending byte of a stripe in the file: - * - * all this function does is move offsets to the nearest region of the - * stripe, and it does its work "mod" the full length of all the stripes. - * consider a file with 3 stripes: - * - * S E - * --------------------------------------------------------------------- - * | 0 | 1 | 2 | 0 | 1 | 2 | - * --------------------------------------------------------------------- - * - * to find stripe 1's offsets for S and E, it divides by the full stripe - * width and does its math in the context of a single set of stripes: - * - * S E - * ----------------------------------- - * | 0 | 1 | 2 | - * ----------------------------------- - * - * it'll notice that E is outside stripe 1 and clamp it to the end of the - * stripe, then multiply it back out by lov_off to give the real offsets in - * the stripe: - * - * S E - * --------------------------------------------------------------------- - * | 1 | 1 | 1 | 1 | 1 | 1 | - * --------------------------------------------------------------------- - * - * it would have done similarly and pulled S forward to the start of a 1 - * stripe if, say, S had landed in a 0 stripe. - * - * this rounding isn't always correct. consider an E lov offset that lands - * on a 0 stripe, the "mod stripe width" math will pull it forward to the - * start of a 1 stripe, when in fact it wanted to be rounded back to the end - * of a previous 1 stripe. this logic is handled by callers and this is why: - * - * this function returns < 0 when the offset was "before" the stripe and - * was moved forward to the start of the stripe in question; 0 when it - * falls in the stripe and no shifting was done; > 0 when the offset - * was outside the stripe and was pulled back to its final byte. */ -static int lov_stripe_offset(struct lov_stripe_md *lsm, obd_off lov_off, - int stripeno, obd_off *obd_off) -{ - unsigned long ssize = lsm->lsm_stripe_size; - unsigned long swidth = ssize * lsm->lsm_stripe_count; - unsigned long stripe_off, this_stripe; - int ret = 0; - - if (lov_off == OBD_OBJECT_EOF) { - *obd_off = OBD_OBJECT_EOF; - return 0; - } - - /* do_div(a, b) returns a % b, and a = a / b */ - stripe_off = do_div(lov_off, swidth); - - this_stripe = stripeno * ssize; - if (stripe_off < this_stripe) { - stripe_off = 0; - ret = -1; - } else { - stripe_off -= this_stripe; - - if (stripe_off >= ssize) { - stripe_off = ssize; - ret = 1; - } - } - - *obd_off = lov_off * ssize + stripe_off; - return ret; -} - -/* Given a whole-file size and a stripe number, give the file size which - * corresponds to the individual object of that stripe. - * - * This behaves basically in the same was as lov_stripe_offset, except that - * file sizes falling before the beginning of a stripe are clamped to the end - * of the previous stripe, not the beginning of the next: - * - * S - * --------------------------------------------------------------------- - * | 0 | 1 | 2 | 0 | 1 | 2 | - * --------------------------------------------------------------------- - * - * if clamped to stripe 2 becomes: - * - * S - * --------------------------------------------------------------------- - * | 0 | 1 | 2 | 0 | 1 | 2 | - * --------------------------------------------------------------------- - */ -static obd_off lov_size_to_stripe(struct lov_stripe_md *lsm, obd_off file_size, - int stripeno) -{ - unsigned long ssize = lsm->lsm_stripe_size; - unsigned long swidth = ssize * lsm->lsm_stripe_count; - unsigned long stripe_off, this_stripe; - - if (file_size == OBD_OBJECT_EOF) - return OBD_OBJECT_EOF; - - /* do_div(a, b) returns a % b, and a = a / b */ - stripe_off = do_div(file_size, swidth); - - this_stripe = stripeno * ssize; - if (stripe_off < this_stripe) { - /* Move to end of previous stripe, or zero */ - if (file_size > 0) { - file_size--; - stripe_off = ssize; - } else { - stripe_off = 0; - } - } else { - stripe_off -= this_stripe; - - if (stripe_off >= ssize) { - /* Clamp to end of this stripe */ - stripe_off = ssize; + CERROR("error: setattr objid "LPX64" subobj " + LPX64" on OST idx %d: rc = %d\n", + set->set_oi->oi_oa->o_id, + req->rq_oi.oi_oa->o_id, req->rq_idx, err); + if (!rc) + rc = err; } } - - return (file_size * ssize + stripe_off); + err = lov_fini_setattr_set(set); + if (!rc) + rc = err; + RETURN(rc); } -/* given an extent in an lov and a stripe, calculate the extent of the stripe - * that is contained within the lov extent. this returns true if the given - * stripe does intersect with the lov extent. */ -static int lov_stripe_intersects(struct lov_stripe_md *lsm, int stripeno, - obd_off start, obd_off end, - obd_off *obd_start, obd_off *obd_end) +static int lov_setattr_interpret(struct ptlrpc_request_set *rqset, + void *data, int rc) { - int start_side = 0, end_side = 0; - - switch (lsm->lsm_pattern) { - case LOV_PATTERN_RAID0: - start_side = lov_stripe_offset(lsm, start, stripeno, obd_start); - end_side = lov_stripe_offset(lsm, end, stripeno, obd_end); - break; - case LOV_PATTERN_CMOBD: - *obd_start = start; - *obd_end = end; - start_side = end_side = 0; - break; - default: - LBUG(); - } - - CDEBUG(D_INODE, "["LPU64"->"LPU64"] -> [(%d) "LPU64"->"LPU64" (%d)]\n", - start, end, start_side, *obd_start, *obd_end, end_side); - - /* this stripe doesn't intersect the file extent when neither - * start or the end intersected the stripe and obd_start and - * obd_end got rounded up to the save value. */ - if (start_side != 0 && end_side != 0 && *obd_start == *obd_end) - return 0; - - /* as mentioned in the lov_stripe_offset commentary, end - * might have been shifted in the wrong direction. This - * happens when an end offset is before the stripe when viewed - * through the "mod stripe size" math. we detect it being shifted - * in the wrong direction and touch it up. - * interestingly, this can't underflow since end must be > start - * if we passed through the previous check. - * (should we assert for that somewhere?) */ - if (end_side != 0) - (*obd_end)--; + struct lov_request_set *lovset = (struct lov_request_set *)data; + int err; + ENTRY; - return 1; + if (rc) + lovset->set_completes = 0; + err = lov_fini_setattr_set(lovset); + RETURN(rc ? rc : err); } -/* compute which stripe number "lov_off" will be written into */ -static int lov_stripe_number(struct lov_stripe_md *lsm, obd_off lov_off) +/* If @oti is given, the request goes from MDS and responses from OSTs are not + needed. Otherwise, a client is waiting for responses. */ +static int lov_setattr_async(struct obd_export *exp, struct obd_info *oinfo, + struct obd_trans_info *oti, + struct ptlrpc_request_set *rqset) { - unsigned long ssize = lsm->lsm_stripe_size; - unsigned long swidth = ssize * lsm->lsm_stripe_count; - unsigned long stripe_off; + struct lov_request_set *set; + struct lov_request *req; + struct list_head *pos; + struct lov_obd *lov; + int rc = 0; + ENTRY; - if (lsm->lsm_pattern == LOV_PATTERN_CMOBD) - return 0; + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); + if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) { + LASSERT(oti); + LASSERT(oti->oti_logcookies); + } - stripe_off = do_div(lov_off, swidth); + if (!exp || !exp->exp_obd) + RETURN(-ENODEV); - return stripe_off / ssize; -} + lov = &exp->exp_obd->u.lov; + rc = lov_prep_setattr_set(exp, oinfo, oti, &set); + if (rc) + RETURN(rc); -static int lov_revalidate_md(struct obd_export *exp, struct obdo *src_oa, - struct lov_stripe_md *ea, - struct obd_trans_info *oti) -{ - struct obd_export *osc_exp; - struct lov_obd *lov = &exp->exp_obd->u.lov; - struct lov_stripe_md *lsm = ea; - struct lov_stripe_md obj_md; - struct lov_stripe_md *obj_mdp = &obj_md; - struct lov_oinfo *loi; - struct obdo *tmp_oa; - int ost_idx, updates = 0, i; - ENTRY; + CDEBUG(D_INFO, "objid "LPX64": %ux%u byte stripes\n", + oinfo->oi_md->lsm_object_id, oinfo->oi_md->lsm_stripe_count, + oinfo->oi_md->lsm_stripe_size); - tmp_oa = obdo_alloc(); - if (tmp_oa == NULL) - RETURN(-ENOMEM); + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); - loi = lsm->lsm_oinfo; - for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) { - int rc; - if (!obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) - continue; + if (oinfo->oi_oa->o_valid & OBD_MD_FLCOOKIE) + oti->oti_logcookies = set->set_cookies + req->rq_stripe; - ost_idx = lov_revalidate_policy(lov, lsm); - if (ost_idx < 0) { - /* FIXME: punt for now. */ - CERROR("lov_revalidate_policy failed; no active " - "OSCs?\n"); - continue; - } + CDEBUG(D_INFO, "objid "LPX64"[%d] has subobj "LPX64" at idx " + "%u\n", oinfo->oi_oa->o_id, req->rq_stripe, + req->rq_oi.oi_oa->o_id, req->rq_idx); - /* create a new object */ - memcpy(tmp_oa, src_oa, sizeof(*tmp_oa)); - /* XXX: LOV STACKING: use real "obj_mdp" sub-data */ - osc_exp = lov->tgts[ost_idx].ltd_exp; - rc = obd_create(osc_exp, tmp_oa, &obj_mdp, oti); + rc = obd_setattr_async(lov->lov_tgts[req->rq_idx]->ltd_exp, + &req->rq_oi, oti, rqset); if (rc) { - CERROR("error creating new subobj at idx %d; " - "rc = %d\n", ost_idx, rc); - continue; + CERROR("error: setattr objid "LPX64" subobj " + LPX64" on OST idx %d: rc = %d\n", + set->set_oi->oi_oa->o_id, + req->rq_oi.oi_oa->o_id, + req->rq_idx, rc); + break; } - if (oti->oti_objid) - oti->oti_objid[ost_idx] = tmp_oa->o_id; - loi->loi_id = tmp_oa->o_id; - loi->loi_ost_idx = ost_idx; - loi->loi_ost_gen = lov->tgts[ost_idx].ltd_gen; - CDEBUG(D_INODE, "replacing objid "LPX64" subobj "LPX64 - " with idx %d gen %d.\n", lsm->lsm_object_id, - loi->loi_id, ost_idx, loi->loi_ost_gen); - updates = 1; } - /* If we got an error revalidating an entry there's no need to - * cleanup up objects we allocated here because the bad entry - * still points to a deleted OST. */ + /* If we are not waiting for responses on async requests, return. */ + if (rc || !rqset || list_empty(&rqset->set_requests)) { + int err; + if (rc) + set->set_completes = 0; + err = lov_fini_setattr_set(set); + RETURN(rc ? rc : err); + } + + LASSERT(rqset->set_interpret == NULL); + rqset->set_interpret = lov_setattr_interpret; + rqset->set_arg = (void *)set; + + RETURN(0); +} + +static int lov_punch_interpret(struct ptlrpc_request_set *rqset, + void *data, int rc) +{ + struct lov_request_set *lovset = (struct lov_request_set *)data; + int err; + ENTRY; - obdo_free(tmp_oa); - RETURN(updates); + if (rc) + lovset->set_completes = 0; + err = lov_fini_punch_set(lovset); + RETURN(rc ? rc : err); } /* FIXME: maybe we'll just make one node the authoritative attribute node, then * we can send this 'punch' to just the authoritative node and the nodes * that the punch will affect. */ -static int lov_punch(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *lsm, - obd_off start, obd_off end, struct obd_trans_info *oti) +static int lov_punch(struct obd_export *exp, struct obd_info *oinfo, + struct obd_trans_info *oti, + struct ptlrpc_request_set *rqset) { - struct obdo tmp; + struct lov_request_set *set; struct lov_obd *lov; - struct lov_oinfo *loi; - int rc = 0, i; + struct list_head *pos; + struct lov_request *req; + int rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); if (!exp || !exp->exp_obd) RETURN(-ENODEV); lov = &exp->exp_obd->u.lov; - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { - obd_off starti, endi; - int err; + rc = lov_prep_punch_set(exp, oinfo, oti, &set); + if (rc) + RETURN(rc); - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - continue; + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); + + rc = obd_punch(lov->lov_tgts[req->rq_idx]->ltd_exp, + &req->rq_oi, NULL, rqset); + if (rc) { + CERROR("error: punch objid "LPX64" subobj "LPX64 + " on OST idx %d: rc = %d\n", + set->set_oi->oi_oa->o_id, + req->rq_oi.oi_oa->o_id, req->rq_idx, rc); + break; } + } - if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi)) - continue; + if (rc || list_empty(&rqset->set_requests)) { + int err; + err = lov_fini_punch_set(set); + RETURN(rc ? rc : err); + } - /* create data objects with "parent" OA */ - memcpy(&tmp, oa, sizeof(tmp)); - tmp.o_id = loi->loi_id; + LASSERT(rqset->set_interpret == NULL); + rqset->set_interpret = lov_punch_interpret; + rqset->set_arg = (void *)set; - err = obd_punch(lov->tgts[loi->loi_ost_idx].ltd_exp, &tmp, NULL, - starti, endi, NULL); - if (err) { - if (lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: punch objid "LPX64" subobj "LPX64 - " on OST idx %d: rc = %d\n", oa->o_id, - loi->loi_id, loi->loi_ost_idx, err); - } - if (!rc) - rc = err; - } else { - loi->loi_kms = loi->loi_rss = starti; - } - } - RETURN(rc); + RETURN(0); } static int lov_sync(struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *lsm, obd_off start, obd_off end) + struct lov_stripe_md *lsm, obd_off start, obd_off end, + void *capa) { - struct obdo *tmp; + struct lov_request_set *set; + struct obd_info oinfo; struct lov_obd *lov; - struct lov_oinfo *loi; - int rc = 0, i; + struct list_head *pos; + struct lov_request *req; + int err = 0, rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); if (!exp->exp_obd) RETURN(-ENODEV); - tmp = obdo_alloc(); - if (!tmp) - RETURN(-ENOMEM); - lov = &exp->exp_obd->u.lov; - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { - obd_off starti, endi; - int err; - - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - continue; - } - - if (!lov_stripe_intersects(lsm, i, start, end, &starti, &endi)) - continue; + rc = lov_prep_sync_set(exp, &oinfo, oa, lsm, start, end, &set); + if (rc) + RETURN(rc); - memcpy(tmp, oa, sizeof(*tmp)); - tmp->o_id = loi->loi_id; + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); - err = obd_sync(lov->tgts[loi->loi_ost_idx].ltd_exp, tmp, NULL, - starti, endi); + rc = obd_sync(lov->lov_tgts[req->rq_idx]->ltd_exp, + req->rq_oi.oi_oa, NULL, + req->rq_oi.oi_policy.l_extent.start, + req->rq_oi.oi_policy.l_extent.end, capa); + err = lov_update_common_set(set, req, rc); if (err) { - if (lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: fsync objid "LPX64" subobj "LPX64 - " on OST idx %d: rc = %d\n", oa->o_id, - loi->loi_id, loi->loi_ost_idx, err); - } + CERROR("error: fsync objid "LPX64" subobj "LPX64 + " on OST idx %d: rc = %d\n", + set->set_oi->oi_oa->o_id, + req->rq_oi.oi_oa->o_id, req->rq_idx, rc); if (!rc) rc = err; } } - - obdo_free(tmp); + err = lov_fini_sync_set(set); + if (!rc) + rc = err; RETURN(rc); } -static int lov_brw_check(struct lov_obd *lov, struct obdo *oa, - struct lov_stripe_md *lsm, +static int lov_brw_check(struct lov_obd *lov, struct obd_info *lov_oinfo, obd_count oa_bufs, struct brw_page *pga) { + struct obd_info oinfo = { { { 0 } } }; int i, rc = 0; - ENTRY; + + oinfo.oi_oa = lov_oinfo->oi_oa; /* The caller just wants to know if there's a chance that this * I/O can succeed */ for (i = 0; i < oa_bufs; i++) { - int stripe = lov_stripe_number(lsm, pga[i].off); - int ost = lsm->lsm_oinfo[stripe].loi_ost_idx; + int stripe = lov_stripe_number(lov_oinfo->oi_md, pga[i].off); + int ost = lov_oinfo->oi_md->lsm_oinfo[stripe]->loi_ost_idx; obd_off start, end; - if (!lov_stripe_intersects(lsm, i, pga[i].off, + if (!lov_stripe_intersects(lov_oinfo->oi_md, i, pga[i].off, pga[i].off + pga[i].count, &start, &end)) continue; - if (lov->tgts[ost].active == 0) { + if (!lov->lov_tgts[ost] || !lov->lov_tgts[ost]->ltd_active) { CDEBUG(D_HA, "lov idx %d inactive\n", ost); - RETURN(-EIO); + return -EIO; } - rc = obd_brw(OBD_BRW_CHECK, lov->tgts[ost].ltd_exp, oa, - NULL, 1, &pga[i], NULL); + + rc = obd_brw(OBD_BRW_CHECK, lov->lov_tgts[ost]->ltd_exp, &oinfo, + 1, &pga[i], NULL); if (rc) break; } - RETURN(rc); + return rc; } -static int lov_brw(int cmd, struct obd_export *exp, struct obdo *src_oa, - struct lov_stripe_md *lsm, obd_count oa_bufs, - struct brw_page *pga, struct obd_trans_info *oti) +static int lov_brw(int cmd, struct obd_export *exp, struct obd_info *oinfo, + obd_count oa_bufs, struct brw_page *pga, + struct obd_trans_info *oti) { - struct { - int bufct; - int index; - int subcount; - struct lov_stripe_md lsm; - int ost_idx; - } *stripeinfo, *si, *si_last; - struct obdo *ret_oa = NULL, *tmp_oa = NULL; - struct lov_obd *lov; - struct brw_page *ioarr; - struct lov_oinfo *loi; - int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count, set = 0; + struct lov_request_set *set; + struct lov_request *req; + struct list_head *pos; + struct lov_obd *lov = &exp->exp_obd->u.lov; + int err, rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); - - lov = &exp->exp_obd->u.lov; + ASSERT_LSM_MAGIC(oinfo->oi_md); if (cmd == OBD_BRW_CHECK) { - rc = lov_brw_check(lov, src_oa, lsm, oa_bufs, pga); + rc = lov_brw_check(lov, oinfo, oa_bufs, pga); RETURN(rc); } - OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo)); - if (!stripeinfo) - RETURN(-ENOMEM); - - OBD_ALLOC(where, sizeof(*where) * oa_bufs); - if (!where) - GOTO(out_sinfo, rc = -ENOMEM); - - OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs); - if (!ioarr) - GOTO(out_where, rc = -ENOMEM); - - if (src_oa) { - ret_oa = obdo_alloc(); - if (!ret_oa) - GOTO(out_ioarr, rc = -ENOMEM); - - tmp_oa = obdo_alloc(); - if (!tmp_oa) - GOTO(out_oa, rc = -ENOMEM); - } - - for (i = 0; i < oa_bufs; i++) { - where[i] = lov_stripe_number(lsm, pga[i].off); - stripeinfo[where[i]].bufct++; - } - - for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo; - i < stripe_count; i++, loi++, si_last = si, si++) { - if (i > 0) - si->index = si_last->index + si_last->bufct; - si->lsm.lsm_object_id = loi->loi_id; - si->lsm.lsm_object_gr = lsm->lsm_object_gr; - si->ost_idx = loi->loi_ost_idx; - } - - for (i = 0; i < oa_bufs; i++) { - int which = where[i]; - int shift; - - shift = stripeinfo[which].index + stripeinfo[which].subcount; - LASSERT(shift < oa_bufs); - ioarr[shift] = pga[i]; - lov_stripe_offset(lsm, pga[i].off, which, - &ioarr[shift].off); - stripeinfo[which].subcount++; - } - - for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) { - int shift = si->index; - - if (lov->tgts[si->ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx); - GOTO(out_oa, rc = -EIO); - } - - if (si->bufct) { - LASSERT(shift < oa_bufs); - if (src_oa) - memcpy(tmp_oa, src_oa, sizeof(*tmp_oa)); + rc = lov_prep_brw_set(exp, oinfo, oa_bufs, pga, oti, &set); + if (rc) + RETURN(rc); - tmp_oa->o_id = si->lsm.lsm_object_id; - rc = obd_brw(cmd, lov->tgts[si->ost_idx].ltd_exp, - tmp_oa, &si->lsm, si->bufct, - &ioarr[shift], oti); - if (rc) - GOTO(out_ioarr, rc); + list_for_each (pos, &set->set_list) { + struct obd_export *sub_exp; + struct brw_page *sub_pga; + req = list_entry(pos, struct lov_request, rq_link); - lov_merge_attrs(ret_oa, tmp_oa, tmp_oa->o_valid, lsm, - i, &set); - } + sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp; + sub_pga = set->set_pga + req->rq_pgaidx; + rc = obd_brw(cmd, sub_exp, &req->rq_oi, req->rq_oabufs, + sub_pga, oti); + if (rc) + break; + lov_update_common_set(set, req, rc); } - ret_oa->o_id = src_oa->o_id; - memcpy(src_oa, ret_oa, sizeof(*src_oa)); - - GOTO(out_oa, rc); - out_oa: - if (tmp_oa) - obdo_free(tmp_oa); - if (ret_oa) - obdo_free(ret_oa); - out_ioarr: - OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs); - out_where: - OBD_FREE(where, sizeof(*where) * oa_bufs); - out_sinfo: - OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo)); - return rc; + err = lov_fini_brw_set(set); + if (!rc) + rc = err; + RETURN(rc); } static int lov_brw_interpret(struct ptlrpc_request_set *reqset, void *data, int rc) { - struct lov_brw_async_args *aa = data; - struct lov_stripe_md *lsm = aa->aa_lsm; - obd_count oa_bufs = aa->aa_oa_bufs; - struct obdo *oa = aa->aa_oa; - struct obdo *obdos = aa->aa_obdos; - struct brw_page *ioarr = aa->aa_ioarr; - struct lov_oinfo *loi; - int i, set = 0; + struct lov_request_set *lovset = (struct lov_request_set *)data; ENTRY; - if (rc == 0) { - /* NB all stripe requests succeeded to get here */ - - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++) { - if (obdos[i].o_valid == 0) /* inactive stripe */ - continue; - - lov_merge_attrs(oa, &obdos[i], obdos[i].o_valid, lsm, - i, &set); - } - - if (!set) { - CERROR("No stripes had valid attrs\n"); - rc = -EIO; - } + if (rc) { + lovset->set_completes = 0; + lov_fini_brw_set(lovset); + } else { + rc = lov_fini_brw_set(lovset); } - oa->o_id = lsm->lsm_object_id; - OBD_FREE(obdos, lsm->lsm_stripe_count * sizeof(*obdos)); - OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs); RETURN(rc); } -static int lov_brw_async(int cmd, struct obd_export *exp, struct obdo *oa, - struct lov_stripe_md *lsm, obd_count oa_bufs, - struct brw_page *pga, struct ptlrpc_request_set *set, - struct obd_trans_info *oti) +static int lov_brw_async(int cmd, struct obd_export *exp, + struct obd_info *oinfo, obd_count oa_bufs, + struct brw_page *pga, struct obd_trans_info *oti, + struct ptlrpc_request_set *set) { - struct { - int bufct; - int index; - int subcount; - struct lov_stripe_md lsm; - int ost_idx; - } *stripeinfo, *si, *si_last; - struct lov_obd *lov; - struct brw_page *ioarr; - struct obdo *obdos = NULL; - struct lov_oinfo *loi; - struct lov_brw_async_args *aa; - int rc = 0, i, *where, stripe_count = lsm->lsm_stripe_count; + struct lov_request_set *lovset; + struct lov_request *req; + struct list_head *pos; + struct lov_obd *lov = &exp->exp_obd->u.lov; + int rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); - - lov = &exp->exp_obd->u.lov; + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); if (cmd == OBD_BRW_CHECK) { - rc = lov_brw_check(lov, oa, lsm, oa_bufs, pga); + rc = lov_brw_check(lov, oinfo, oa_bufs, pga); RETURN(rc); } - OBD_ALLOC(stripeinfo, stripe_count * sizeof(*stripeinfo)); - if (!stripeinfo) - RETURN(-ENOMEM); - - OBD_ALLOC(where, sizeof(*where) * oa_bufs); - if (!where) - GOTO(out_sinfo, rc = -ENOMEM); - - if (oa) { - OBD_ALLOC(obdos, sizeof(*obdos) * stripe_count); - if (!obdos) - GOTO(out_where, rc = -ENOMEM); - } - - OBD_ALLOC(ioarr, sizeof(*ioarr) * oa_bufs); - if (!ioarr) - GOTO(out_obdos, rc = -ENOMEM); - - for (i = 0; i < oa_bufs; i++) { - where[i] = lov_stripe_number(lsm, pga[i].off); - stripeinfo[where[i]].bufct++; - } - - for (i = 0, loi = lsm->lsm_oinfo, si_last = si = stripeinfo; - i < stripe_count; i++, loi++, si_last = si, si++) { - if (i > 0) - si->index = si_last->index + si_last->bufct; - si->lsm.lsm_object_id = loi->loi_id; - si->ost_idx = loi->loi_ost_idx; - - if (oa) { - memcpy(&obdos[i], oa, sizeof(*obdos)); - obdos[i].o_id = si->lsm.lsm_object_id; - } - } - - for (i = 0; i < oa_bufs; i++) { - int which = where[i]; - int shift; - - shift = stripeinfo[which].index + stripeinfo[which].subcount; - LASSERT(shift < oa_bufs); - ioarr[shift] = pga[i]; - lov_stripe_offset(lsm, pga[i].off, which, - &ioarr[shift].off); - stripeinfo[which].subcount++; - } - - for (i = 0, si = stripeinfo; i < stripe_count; i++, si++) { - int shift = si->index; - - if (si->bufct == 0) - continue; - - if (lov->tgts[si->ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", si->ost_idx); - GOTO(out_ioarr, rc = -EIO); - } + rc = lov_prep_brw_set(exp, oinfo, oa_bufs, pga, oti, &lovset); + if (rc) + RETURN(rc); - LASSERT(shift < oa_bufs); + list_for_each (pos, &lovset->set_list) { + struct obd_export *sub_exp; + struct brw_page *sub_pga; + req = list_entry(pos, struct lov_request, rq_link); - rc = obd_brw_async(cmd, lov->tgts[si->ost_idx].ltd_exp, - &obdos[i], &si->lsm, si->bufct, - &ioarr[shift], set, oti); + sub_exp = lov->lov_tgts[req->rq_idx]->ltd_exp; + sub_pga = lovset->set_pga + req->rq_pgaidx; + rc = obd_brw_async(cmd, sub_exp, &req->rq_oi, req->rq_oabufs, + sub_pga, oti, set); if (rc) - GOTO(out_ioarr, rc); + GOTO(out, rc); + lov_update_common_set(lovset, req, rc); } LASSERT(rc == 0); LASSERT(set->set_interpret == NULL); set->set_interpret = (set_interpreter_func)lov_brw_interpret; - LASSERT(sizeof(set->set_args) >= sizeof(struct lov_brw_async_args)); - aa = (struct lov_brw_async_args *)&set->set_args; - aa->aa_lsm = lsm; - aa->aa_obdos = obdos; - aa->aa_oa = oa; - aa->aa_ioarr = ioarr; - aa->aa_oa_bufs = oa_bufs; - - /* Don't free ioarr or obdos - that's done in lov_brw_interpret */ - GOTO(out_where, rc); - - out_ioarr: - OBD_FREE(ioarr, sizeof(*ioarr) * oa_bufs); - out_obdos: - OBD_FREE(obdos, stripe_count * sizeof(*obdos)); - out_where: - OBD_FREE(where, sizeof(*where) * oa_bufs); - out_sinfo: - OBD_FREE(stripeinfo, stripe_count * sizeof(*stripeinfo)); - return rc; -} + set->set_arg = (void *)lovset; -struct lov_async_page *lap_from_cookie(void *cookie) -{ - struct lov_async_page *lap = cookie; - if (lap->lap_magic != LAP_MAGIC) - return ERR_PTR(-EINVAL); - return lap; -}; + RETURN(rc); +out: + lov_fini_brw_set(lovset); + RETURN(rc); +} static int lov_ap_make_ready(void *data, int cmd) { - struct lov_async_page *lap = lap_from_cookie(data); - /* XXX should these assert? */ - if (IS_ERR(lap)) - return -EINVAL; + struct lov_async_page *lap = LAP_FROM_COOKIE(data); return lap->lap_caller_ops->ap_make_ready(lap->lap_caller_data, cmd); } + static int lov_ap_refresh_count(void *data, int cmd) { - struct lov_async_page *lap = lap_from_cookie(data); - if (IS_ERR(lap)) - return -EINVAL; + struct lov_async_page *lap = LAP_FROM_COOKIE(data); return lap->lap_caller_ops->ap_refresh_count(lap->lap_caller_data, cmd); } + static void lov_ap_fill_obdo(void *data, int cmd, struct obdo *oa) { - struct lov_async_page *lap = lap_from_cookie(data); - /* XXX should these assert? */ - if (IS_ERR(lap)) - return; + struct lov_async_page *lap = LAP_FROM_COOKIE(data); lap->lap_caller_ops->ap_fill_obdo(lap->lap_caller_data, cmd, oa); /* XXX woah, shouldn't we be altering more here? size? */ oa->o_id = lap->lap_loi_id; + oa->o_gr = lap->lap_loi_gr; + oa->o_valid |= OBD_MD_FLGROUP; + oa->o_stripe_idx = lap->lap_stripe; +} + +static void lov_ap_update_obdo(void *data, int cmd, struct obdo *oa, + obd_valid valid) +{ + struct lov_async_page *lap = LAP_FROM_COOKIE(data); + + lap->lap_caller_ops->ap_update_obdo(lap->lap_caller_data, cmd,oa,valid); } -static void lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc) +static int lov_ap_completion(void *data, int cmd, struct obdo *oa, int rc) { - struct lov_async_page *lap = lap_from_cookie(data); - if (IS_ERR(lap)) - return; + struct lov_async_page *lap = LAP_FROM_COOKIE(data); /* in a raid1 regime this would down a count of many ios * in flight, onl calling the caller_ops completion when all * the raid1 ios are complete */ - lap->lap_caller_ops->ap_completion(lap->lap_caller_data, cmd, oa, rc); + rc = lap->lap_caller_ops->ap_completion(lap->lap_caller_data,cmd,oa,rc); + return rc; +} + +static struct obd_capa *lov_ap_lookup_capa(void *data, int cmd) +{ + struct lov_async_page *lap = LAP_FROM_COOKIE(data); + return lap->lap_caller_ops->ap_lookup_capa(lap->lap_caller_data, cmd); } static struct obd_async_page_ops lov_async_page_ops = { .ap_make_ready = lov_ap_make_ready, .ap_refresh_count = lov_ap_refresh_count, .ap_fill_obdo = lov_ap_fill_obdo, + .ap_update_obdo = lov_ap_update_obdo, .ap_completion = lov_ap_completion, + .ap_lookup_capa = lov_ap_lookup_capa, }; -static int lov_prep_async_page(struct obd_export *exp, - struct lov_stripe_md *lsm, - struct lov_oinfo *loi, struct page *page, - obd_off offset, struct obd_async_page_ops *ops, - void *data, void **res) +int lov_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm, + struct lov_oinfo *loi, cfs_page_t *page, + obd_off offset, struct obd_async_page_ops *ops, + void *data, void **res) { struct lov_obd *lov = &exp->exp_obd->u.lov; struct lov_async_page *lap; - int rc, stripe; + int rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); - LASSERT(loi == NULL); - - stripe = lov_stripe_number(lsm, offset); - loi = &lsm->lsm_oinfo[stripe]; - - if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) - RETURN(-EIO); - if (lov->tgts[loi->loi_ost_idx].active == 0) - RETURN(-EIO); - if (lov->tgts[loi->loi_ost_idx].ltd_exp == NULL) { - CERROR("ltd_exp == NULL, but OST idx %d doesn't appear to be " - "deleted or inactive.\n", loi->loi_ost_idx); - RETURN(-EIO); + if (!page) { + int i = 0; + /* Find an existing osc so we can get it's stupid sizeof(*oap). + Only because of this layering limitation will a client + mount with no osts fail */ + while (!lov->lov_tgts || !lov->lov_tgts[i] || + !lov->lov_tgts[i]->ltd_exp) { + i++; + if (i >= lov->desc.ld_tgt_count) + RETURN(-ENOMEDIUM); + } + rc = size_round(sizeof(*lap)) + + obd_prep_async_page(lov->lov_tgts[i]->ltd_exp, NULL, + NULL, NULL, 0, NULL, NULL, NULL); + RETURN(rc); } + ASSERT_LSM_MAGIC(lsm); + LASSERT(loi == NULL); - OBD_ALLOC(lap, sizeof(*lap)); - if (lap == NULL) - RETURN(-ENOMEM); - - lap->lap_magic = LAP_MAGIC; + lap = *res; + lap->lap_magic = LOV_AP_MAGIC; lap->lap_caller_ops = ops; lap->lap_caller_data = data; - /* FIXME handle multiple oscs after landing b_raid1 */ - lap->lap_stripe = stripe; - switch (lsm->lsm_pattern) { - case LOV_PATTERN_RAID0: - lov_stripe_offset(lsm, offset, lap->lap_stripe, - &lap->lap_sub_offset); - break; - case LOV_PATTERN_CMOBD: - lap->lap_sub_offset = offset; - break; - default: - LBUG(); - } + /* for now only raid 0 which passes through */ + lap->lap_stripe = lov_stripe_number(lsm, offset); + lov_stripe_offset(lsm, offset, lap->lap_stripe, &lap->lap_sub_offset); + loi = lsm->lsm_oinfo[lap->lap_stripe]; /* so the callback doesn't need the lsm */ lap->lap_loi_id = loi->loi_id; + lap->lap_loi_gr = lsm->lsm_object_gr; + LASSERT(lsm->lsm_object_gr > 0); + + lap->lap_sub_cookie = (void *)lap + size_round(sizeof(*lap)); - rc = obd_prep_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp, + rc = obd_prep_async_page(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm, loi, page, lap->lap_sub_offset, &lov_async_page_ops, lap, &lap->lap_sub_cookie); - if (rc) { - OBD_FREE(lap, sizeof(*lap)); + if (rc) RETURN(rc); - } CDEBUG(D_CACHE, "lap %p page %p cookie %p off "LPU64"\n", lap, page, lap->lap_sub_cookie, offset); - *res = lap; RETURN(0); } @@ -2272,16 +1717,13 @@ static int lov_queue_async_io(struct obd_export *exp, LASSERT(loi == NULL); - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); - lap = lap_from_cookie(cookie); - if (IS_ERR(lap)) - RETURN(PTR_ERR(lap)); + lap = LAP_FROM_COOKIE(cookie); - loi = &lsm->lsm_oinfo[lap->lap_stripe]; + loi = lsm->lsm_oinfo[lap->lap_stripe]; - rc = obd_queue_async_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, + rc = obd_queue_async_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm, loi, lap->lap_sub_cookie, cmd, off, count, brw_flags, async_flags); RETURN(rc); @@ -2298,16 +1740,13 @@ static int lov_set_async_flags(struct obd_export *exp, LASSERT(loi == NULL); - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); - lap = lap_from_cookie(cookie); - if (IS_ERR(lap)) - RETURN(PTR_ERR(lap)); + lap = LAP_FROM_COOKIE(cookie); - loi = &lsm->lsm_oinfo[lap->lap_stripe]; + loi = lsm->lsm_oinfo[lap->lap_stripe]; - rc = obd_set_async_flags(lov->tgts[loi->loi_ost_idx].ltd_exp, + rc = obd_set_async_flags(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm, loi, lap->lap_sub_cookie, async_flags); RETURN(rc); } @@ -2325,17 +1764,14 @@ static int lov_queue_group_io(struct obd_export *exp, LASSERT(loi == NULL); - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); - lap = lap_from_cookie(cookie); - if (IS_ERR(lap)) - RETURN(PTR_ERR(lap)); + lap = LAP_FROM_COOKIE(cookie); - loi = &lsm->lsm_oinfo[lap->lap_stripe]; + loi = lsm->lsm_oinfo[lap->lap_stripe]; - rc = obd_queue_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, lsm, loi, - oig, lap->lap_sub_cookie, cmd, off, count, + rc = obd_queue_group_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm, + loi, oig, lap->lap_sub_cookie, cmd, off, count, brw_flags, async_flags); RETURN(rc); } @@ -2353,17 +1789,17 @@ static int lov_trigger_group_io(struct obd_export *exp, LASSERT(loi == NULL); - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); - loi = lsm->lsm_oinfo; - for (i = 0; i < lsm->lsm_stripe_count; i++, loi++) { - if (lov->tgts[loi->loi_ost_idx].active == 0) { + for (i = 0; i < lsm->lsm_stripe_count; i++) { + loi = lsm->lsm_oinfo[i]; + if (!lov->lov_tgts[loi->loi_ost_idx] || + !lov->lov_tgts[loi->loi_ost_idx]->ltd_active) { CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); continue; } - err = obd_trigger_group_io(lov->tgts[loi->loi_ost_idx].ltd_exp, + err = obd_trigger_group_io(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm, loi, oig); if (rc == 0 && err != 0) rc = err; @@ -2381,278 +1817,120 @@ static int lov_teardown_async_page(struct obd_export *exp, LASSERT(loi == NULL); - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); - lap = lap_from_cookie(cookie); - if (IS_ERR(lap)) - RETURN(PTR_ERR(lap)); + lap = LAP_FROM_COOKIE(cookie); - loi = &lsm->lsm_oinfo[lap->lap_stripe]; + loi = lsm->lsm_oinfo[lap->lap_stripe]; - rc = obd_teardown_async_page(lov->tgts[loi->loi_ost_idx].ltd_exp, + rc = obd_teardown_async_page(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, lsm, loi, lap->lap_sub_cookie); if (rc) { CERROR("unable to teardown sub cookie %p: %d\n", lap->lap_sub_cookie, rc); RETURN(rc); } - OBD_FREE(lap, sizeof(*lap)); RETURN(rc); } -static int lov_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm, - __u32 type, ldlm_policy_data_t *policy, __u32 mode, - int *flags, void *bl_cb, void *cp_cb, void *gl_cb, - void *data,__u32 lvb_len, void *lvb_swabber, - struct lustre_handle *lockh) +static int lov_enqueue_interpret(struct ptlrpc_request_set *rqset, + void *data, int rc) { - struct lov_lock_handles *lov_lockh = NULL; - struct lustre_handle *lov_lockhp; + struct lov_request_set *lovset = (struct lov_request_set *)data; + ENTRY; + rc = lov_fini_enqueue_set(lovset, lovset->set_ei->ei_mode, rc, rqset); + RETURN(rc); +} + +static int lov_enqueue(struct obd_export *exp, struct obd_info *oinfo, + struct ldlm_enqueue_info *einfo, + struct ptlrpc_request_set *rqset) +{ + ldlm_mode_t mode = einfo->ei_mode; + struct lov_request_set *set; + struct lov_request *req; + struct list_head *pos; struct lov_obd *lov; - struct lov_oinfo *loi; - char submd_buf[sizeof(struct lov_stripe_md) + sizeof(struct lov_oinfo)]; - struct lov_stripe_md *submd = (void *)submd_buf; ldlm_error_t rc; - int i, save_flags = *flags; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + LASSERT(oinfo); + ASSERT_LSM_MAGIC(oinfo->oi_md); + LASSERT(mode == (mode & -mode)); /* we should never be asked to replay a lock this way. */ - LASSERT((*flags & LDLM_FL_REPLAY) == 0); + LASSERT((oinfo->oi_flags & LDLM_FL_REPLAY) == 0); if (!exp || !exp->exp_obd) RETURN(-ENODEV); - if (lsm->lsm_stripe_count > 1) { - lov_lockh = lov_llh_new(lsm); - if (lov_lockh == NULL) - RETURN(-ENOMEM); - - lockh->cookie = lov_lockh->llh_handle.h_cookie; - lov_lockhp = lov_lockh->llh_handles; - } else { - lov_lockhp = lockh; - } - lov = &exp->exp_obd->u.lov; - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++, lov_lockhp++) { - ldlm_policy_data_t sub_ext; - obd_off start, end; - - if (!lov_stripe_intersects(lsm, i, policy->l_extent.start, - policy->l_extent.end, &start, - &end)) - continue; - - sub_ext.l_extent.start = start; - sub_ext.l_extent.end = end; - sub_ext.l_extent.gid = policy->l_extent.gid; - - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - continue; - } - - /* XXX LOV STACKING: submd should be from the subobj */ - submd->lsm_object_id = loi->loi_id; - submd->lsm_object_gr = lsm->lsm_object_gr; - submd->lsm_stripe_count = 0; - submd->lsm_oinfo->loi_kms_valid = loi->loi_kms_valid; - submd->lsm_oinfo->loi_rss = loi->loi_rss; - submd->lsm_oinfo->loi_kms = loi->loi_kms; - submd->lsm_oinfo->loi_blocks = loi->loi_blocks; - loi->loi_mtime = submd->lsm_oinfo->loi_mtime; - /* XXX submd is not fully initialized here */ - *flags = save_flags; - rc = obd_enqueue(lov->tgts[loi->loi_ost_idx].ltd_exp, submd, - type, &sub_ext, mode, flags, bl_cb, cp_cb, - gl_cb, data, lvb_len, lvb_swabber, lov_lockhp); - - /* XXX FIXME: This unpleasantness doesn't belong here at *all*. - * It belongs in the OSC, except that the OSC doesn't have - * access to the real LOI -- it gets a copy, that we created - * above, and that copy can be arbitrarily out of date. - * - * The LOV API is due for a serious rewriting anyways, and this - * can be addressed then. */ - if (rc == ELDLM_OK) { - struct ldlm_lock *lock = ldlm_handle2lock(lov_lockhp); - __u64 tmp = submd->lsm_oinfo->loi_rss; - - LASSERT(lock != NULL); - loi->loi_rss = tmp; - loi->loi_blocks = submd->lsm_oinfo->loi_blocks; - /* Extend KMS up to the end of this lock and no further - * A lock on [x,y] means a KMS of up to y + 1 bytes! */ - if (tmp > lock->l_policy_data.l_extent.end) - tmp = lock->l_policy_data.l_extent.end + 1; - if (tmp >= loi->loi_kms) { - CDEBUG(D_INODE, "lock acquired, setting rss=" - LPU64", kms="LPU64"\n", loi->loi_rss, - tmp); - loi->loi_kms = tmp; - loi->loi_kms_valid = 1; - } else { - CDEBUG(D_INODE, "lock acquired, setting rss=" - LPU64"; leaving kms="LPU64", end="LPU64 - "\n", loi->loi_rss, loi->loi_kms, - lock->l_policy_data.l_extent.end); - } - ldlm_lock_allow_match(lock); - LDLM_LOCK_PUT(lock); - } else if (rc == ELDLM_LOCK_ABORTED && - save_flags & LDLM_FL_HAS_INTENT) { - memset(lov_lockhp, 0, sizeof(*lov_lockhp)); - loi->loi_rss = submd->lsm_oinfo->loi_rss; - loi->loi_blocks = submd->lsm_oinfo->loi_blocks; - CDEBUG(D_INODE, "glimpsed, setting rss="LPU64"; leaving" - " kms="LPU64"\n", loi->loi_rss, loi->loi_kms); - } else { - memset(lov_lockhp, 0, sizeof(*lov_lockhp)); - if (lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: enqueue objid "LPX64" subobj " - LPX64" on OST idx %d: rc = %d\n", - lsm->lsm_object_id, loi->loi_id, - loi->loi_ost_idx, rc); - GOTO(out_locks, rc); - } - } - } - if (lsm->lsm_stripe_count > 1) - lov_llh_put(lov_lockh); - RETURN(ELDLM_OK); - - out_locks: - while (loi--, lov_lockhp--, i-- > 0) { - struct lov_stripe_md submd; - int err; + rc = lov_prep_enqueue_set(exp, oinfo, einfo, &set); + if (rc) + RETURN(rc); - if (lov_lockhp->cookie == 0) - continue; + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); - /* XXX LOV STACKING: submd should be from the subobj */ - submd.lsm_object_id = loi->loi_id; - submd.lsm_object_gr = lsm->lsm_object_gr; - submd.lsm_stripe_count = 0; - err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd, - mode, lov_lockhp); - if (err && lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: cancelling objid "LPX64" on OST " - "idx %d after enqueue error: rc = %d\n", - loi->loi_id, loi->loi_ost_idx, err); - } + rc = obd_enqueue(lov->lov_tgts[req->rq_idx]->ltd_exp, + &req->rq_oi, einfo, rqset); + if (rc != ELDLM_OK) + GOTO(out, rc); } - if (lsm->lsm_stripe_count > 1) { - lov_llh_destroy(lov_lockh); - lov_llh_put(lov_lockh); + if (rqset && !list_empty(&rqset->set_requests)) { + LASSERT(rc == 0); + LASSERT(rqset->set_interpret == NULL); + rqset->set_interpret = lov_enqueue_interpret; + rqset->set_arg = (void *)set; + RETURN(rc); } - return rc; +out: + rc = lov_fini_enqueue_set(set, mode, rc, rqset); + RETURN(rc); } static int lov_match(struct obd_export *exp, struct lov_stripe_md *lsm, __u32 type, ldlm_policy_data_t *policy, __u32 mode, int *flags, void *data, struct lustre_handle *lockh) { - struct lov_lock_handles *lov_lockh = NULL; + struct lov_request_set *set; + struct obd_info oinfo; + struct lov_request *req; + struct list_head *pos; + struct lov_obd *lov = &exp->exp_obd->u.lov; struct lustre_handle *lov_lockhp; - struct lov_obd *lov; - struct lov_oinfo *loi; - struct lov_stripe_md submd; - ldlm_error_t rc = 0; - int i; + int lov_flags, rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); + LASSERT((*flags & LDLM_FL_TEST_LOCK) || mode == (mode & -mode)); if (!exp || !exp->exp_obd) RETURN(-ENODEV); - if (lsm->lsm_stripe_count > 1) { - lov_lockh = lov_llh_new(lsm); - if (lov_lockh == NULL) - RETURN(-ENOMEM); - - lockh->cookie = lov_lockh->llh_handle.h_cookie; - lov_lockhp = lov_lockh->llh_handles; - } else { - lov_lockhp = lockh; - } - lov = &exp->exp_obd->u.lov; - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++, lov_lockhp++) { - ldlm_policy_data_t sub_ext; - obd_off start, end; - int lov_flags; - - if (!lov_stripe_intersects(lsm, i, policy->l_extent.start, - policy->l_extent.end, &start, &end)) - continue; - - sub_ext.l_extent.start = start; - sub_ext.l_extent.end = end; + rc = lov_prep_match_set(exp, &oinfo, lsm, policy, mode, lockh, &set); + if (rc) + RETURN(rc); - if (obd_uuid_empty(&lov->tgts[loi->loi_ost_idx].uuid)) { - CDEBUG(D_HA, "lov idx %d deleted\n", loi->loi_ost_idx); - continue; - } - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); - rc = -EIO; - break; - } + list_for_each (pos, &set->set_list) { + ldlm_policy_data_t sub_policy; + req = list_entry(pos, struct lov_request, rq_link); + lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe; + LASSERT(lov_lockhp); - /* XXX LOV STACKING: submd should be from the subobj */ - submd.lsm_object_id = loi->loi_id; - submd.lsm_object_gr = lsm->lsm_object_gr; - submd.lsm_stripe_count = 0; lov_flags = *flags; - /* XXX submd is not fully initialized here */ - rc = obd_match(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd, - type, &sub_ext, mode, &lov_flags, data, - lov_lockhp); - if (rc != 1) - break; - } - if (rc == 1) { - if (lsm->lsm_stripe_count > 1) { - if (*flags & LDLM_FL_TEST_LOCK) - lov_llh_destroy(lov_lockh); - lov_llh_put(lov_lockh); - } - RETURN(1); - } - - while (loi--, lov_lockhp--, i-- > 0) { - struct lov_stripe_md submd; - int err; - - if (lov_lockhp->cookie == 0) - continue; - - /* XXX LOV STACKING: submd should be from the subobj */ - submd.lsm_object_id = loi->loi_id; - submd.lsm_object_gr = lsm->lsm_object_gr; - submd.lsm_stripe_count = 0; - err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd, - mode, lov_lockhp); - if (err && lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: cancelling objid "LPX64" on OST " - "idx %d after match failure: rc = %d\n", - loi->loi_id, loi->loi_ost_idx, err); - } - } + sub_policy.l_extent = req->rq_oi.oi_policy.l_extent; - if (lsm->lsm_stripe_count > 1) { - lov_llh_destroy(lov_lockh); - lov_llh_put(lov_lockh); + rc = obd_match(lov->lov_tgts[req->rq_idx]->ltd_exp, + req->rq_oi.oi_md, type, &sub_policy, + mode, &lov_flags, data, lov_lockhp); + rc = lov_update_match_set(set, req, rc); + if (rc <= 0) + break; } + lov_fini_match_set(set, mode, *flags); RETURN(rc); } @@ -2661,12 +1939,10 @@ static int lov_change_cbdata(struct obd_export *exp, void *data) { struct lov_obd *lov; - struct lov_oinfo *loi; int rc = 0, i; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); if (!exp || !exp->exp_obd) RETURN(-ENODEV); @@ -2674,17 +1950,19 @@ static int lov_change_cbdata(struct obd_export *exp, LASSERT(lsm->lsm_object_gr > 0); lov = &exp->exp_obd->u.lov; - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { + for (i = 0; i < lsm->lsm_stripe_count; i++) { struct lov_stripe_md submd; - if (lov->tgts[loi->loi_ost_idx].active == 0) { - CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); + struct lov_oinfo *loi = lsm->lsm_oinfo[i]; + + if (!lov->lov_tgts[loi->loi_ost_idx]) { + CDEBUG(D_HA, "lov idx %d NULL \n", loi->loi_ost_idx); continue; } - + submd.lsm_object_id = loi->loi_id; submd.lsm_object_gr = lsm->lsm_object_gr; submd.lsm_stripe_count = 0; - rc = obd_change_cbdata(lov->tgts[loi->loi_ost_idx].ltd_exp, + rc = obd_change_cbdata(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, &submd, it, data); } RETURN(rc); @@ -2693,111 +1971,96 @@ static int lov_change_cbdata(struct obd_export *exp, static int lov_cancel(struct obd_export *exp, struct lov_stripe_md *lsm, __u32 mode, struct lustre_handle *lockh) { - struct lov_lock_handles *lov_lockh = NULL; + struct lov_request_set *set; + struct obd_info oinfo; + struct lov_request *req; + struct list_head *pos; + struct lov_obd *lov = &exp->exp_obd->u.lov; struct lustre_handle *lov_lockhp; - struct lov_obd *lov; - struct lov_oinfo *loi; - int rc = 0, i; + int err = 0, rc = 0; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); if (!exp || !exp->exp_obd) RETURN(-ENODEV); LASSERT(lsm->lsm_object_gr > 0); - LASSERT(lockh); - if (lsm->lsm_stripe_count > 1) { - lov_lockh = lov_handle2llh(lockh); - if (!lov_lockh) { - CERROR("LOV: invalid lov lock handle %p\n", lockh); - RETURN(-EINVAL); - } - - lov_lockhp = lov_lockh->llh_handles; - } else { - lov_lockhp = lockh; - } - lov = &exp->exp_obd->u.lov; - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++, lov_lockhp++) { - struct lov_stripe_md submd; - int err; + rc = lov_prep_cancel_set(exp, &oinfo, lsm, mode, lockh, &set); + if (rc) + RETURN(rc); - if (lov_lockhp->cookie == 0) { - CDEBUG(D_HA, "lov idx %d subobj "LPX64" no lock?\n", - loi->loi_ost_idx, loi->loi_id); - continue; - } + list_for_each (pos, &set->set_list) { + req = list_entry(pos, struct lov_request, rq_link); + lov_lockhp = set->set_lockh->llh_handles + req->rq_stripe; - /* XXX LOV STACKING: submd should be from the subobj */ - submd.lsm_object_id = loi->loi_id; - submd.lsm_object_gr = lsm->lsm_object_gr; - submd.lsm_stripe_count = 0; - err = obd_cancel(lov->tgts[loi->loi_ost_idx].ltd_exp, &submd, - mode, lov_lockhp); - if (err) { - if (lov->tgts[loi->loi_ost_idx].active) { - CERROR("error: cancel objid "LPX64" subobj " - LPX64" on OST idx %d: rc = %d\n", - lsm->lsm_object_id, - loi->loi_id, loi->loi_ost_idx, err); - if (!rc) - rc = err; - } + rc = obd_cancel(lov->lov_tgts[req->rq_idx]->ltd_exp, + req->rq_oi.oi_md, mode, lov_lockhp); + rc = lov_update_common_set(set, req, rc); + if (rc) { + CERROR("error: cancel objid "LPX64" subobj " + LPX64" on OST idx %d: rc = %d\n", + lsm->lsm_object_id, + req->rq_oi.oi_md->lsm_object_id, + req->rq_idx, rc); + err = rc; } - } - if (lsm->lsm_stripe_count > 1) - lov_llh_destroy(lov_lockh); - if (lov_lockh != NULL) - lov_llh_put(lov_lockh); - RETURN(rc); + } + lov_fini_cancel_set(set); + RETURN(err); } static int lov_cancel_unused(struct obd_export *exp, - struct lov_stripe_md *lsm, int flags, void *opaque) + struct lov_stripe_md *lsm, + int flags, void *opaque) { struct lov_obd *lov; - struct lov_oinfo *loi; int rc = 0, i; ENTRY; + if (!exp || !exp->exp_obd) + RETURN(-ENODEV); + lov = &exp->exp_obd->u.lov; if (lsm == NULL) { for (i = 0; i < lov->desc.ld_tgt_count; i++) { - int err = obd_cancel_unused(lov->tgts[i].ltd_exp, NULL, - flags, opaque); + int err; + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp) + continue; + + err = obd_cancel_unused(lov->lov_tgts[i]->ltd_exp, NULL, + flags, opaque); if (!rc) rc = err; } RETURN(rc); } - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); - - if (!exp || !exp->exp_obd) - RETURN(-ENODEV); + ASSERT_LSM_MAGIC(lsm); LASSERT(lsm->lsm_object_gr > 0); - - for (i = 0,loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; i++,loi++) { + for (i = 0; i < lsm->lsm_stripe_count; i++) { struct lov_stripe_md submd; + struct lov_oinfo *loi = lsm->lsm_oinfo[i]; int err; - if (lov->tgts[loi->loi_ost_idx].active == 0) + if (!lov->lov_tgts[loi->loi_ost_idx]) { + CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx); + continue; + } + + if (!lov->lov_tgts[loi->loi_ost_idx]->ltd_active) CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); submd.lsm_object_id = loi->loi_id; submd.lsm_object_gr = lsm->lsm_object_gr; submd.lsm_stripe_count = 0; - err = obd_cancel_unused(lov->tgts[loi->loi_ost_idx].ltd_exp, + err = obd_cancel_unused(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, &submd, flags, opaque); - if (err && lov->tgts[loi->loi_ost_idx].active) { + if (err && lov->lov_tgts[loi->loi_ost_idx]->ltd_active) { CERROR("error: cancel unused objid "LPX64" subobj "LPX64 " on OST idx %d: rc = %d\n", lsm->lsm_object_id, loi->loi_id, loi->loi_ost_idx, err); @@ -2808,78 +2071,138 @@ static int lov_cancel_unused(struct obd_export *exp, RETURN(rc); } -#define LOV_U64_MAX ((__u64)~0ULL) -#define LOV_SUM_MAX(tot, add) \ - do { \ - if ((tot) + (add) < (tot)) \ - (tot) = LOV_U64_MAX; \ - else \ - (tot) += (add); \ - } while(0) +static int lov_join_lru(struct obd_export *exp, + struct lov_stripe_md *lsm, int join) +{ + struct lov_obd *lov; + int i, count = 0; + ENTRY; + + ASSERT_LSM_MAGIC(lsm); + if (!exp || !exp->exp_obd) + RETURN(-ENODEV); + + lov = &exp->exp_obd->u.lov; + for (i = 0; i < lsm->lsm_stripe_count; i++) { + struct lov_stripe_md submd; + struct lov_oinfo *loi = lsm->lsm_oinfo[i]; + int rc = 0; + + if (!lov->lov_tgts[loi->loi_ost_idx]) { + CDEBUG(D_HA, "lov idx %d NULL\n", loi->loi_ost_idx); + continue; + } + + if (!lov->lov_tgts[loi->loi_ost_idx]->ltd_active) + CDEBUG(D_HA, "lov idx %d inactive\n", loi->loi_ost_idx); + + submd.lsm_object_id = loi->loi_id; + submd.lsm_object_gr = lsm->lsm_object_gr; + submd.lsm_stripe_count = 0; + rc = obd_join_lru(lov->lov_tgts[loi->loi_ost_idx]->ltd_exp, + &submd, join); + if (rc < 0) { + CERROR("join lru failed. objid: "LPX64" subobj: "LPX64 + " ostidx: %d rc: %d\n", lsm->lsm_object_id, + loi->loi_id, loi->loi_ost_idx, rc); + return rc; + } else { + count += rc; + } + } + RETURN(count); +} + +static int lov_statfs_interpret(struct ptlrpc_request_set *rqset, + void *data, int rc) +{ + struct lov_request_set *lovset = (struct lov_request_set *)data; + int err; + ENTRY; + + if (rc) + lovset->set_completes = 0; + + err = lov_fini_statfs_set(lovset); + RETURN(rc ? rc : err); +} + +static int lov_statfs_async(struct obd_device *obd, struct obd_info *oinfo, + __u64 max_age, struct ptlrpc_request_set *rqset) +{ + struct lov_request_set *set; + struct lov_request *req; + struct list_head *pos; + struct lov_obd *lov; + int rc = 0; + ENTRY; + + LASSERT(oinfo != NULL); + LASSERT(oinfo->oi_osfs != NULL); + + lov = &obd->u.lov; + rc = lov_prep_statfs_set(obd, oinfo, &set); + if (rc) + RETURN(rc); + + list_for_each (pos, &set->set_list) { + struct obd_device *osc_obd; + + req = list_entry(pos, struct lov_request, rq_link); + + osc_obd = class_exp2obd(lov->lov_tgts[req->rq_idx]->ltd_exp); + rc = obd_statfs_async(osc_obd, &req->rq_oi, max_age, rqset); + if (rc) + break; + } + + if (rc || list_empty(&rqset->set_requests)) { + int err; + if (rc) + set->set_completes = 0; + err = lov_fini_statfs_set(set); + RETURN(rc ? rc : err); + } + + LASSERT(rqset->set_interpret == NULL); + rqset->set_interpret = lov_statfs_interpret; + rqset->set_arg = (void *)set; + RETURN(0); +} static int lov_statfs(struct obd_device *obd, struct obd_statfs *osfs, - unsigned long max_age) + __u64 max_age) { struct lov_obd *lov = &obd->u.lov; struct obd_statfs lov_sfs; int set = 0; - int rc = 0; + int rc = 0, err; int i; ENTRY; - /* We only get block data from the OBD */ for (i = 0; i < lov->desc.ld_tgt_count; i++) { - int err; - if (!lov->tgts[i].active) { + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) { CDEBUG(D_HA, "lov idx %d inactive\n", i); continue; } - err = obd_statfs(class_exp2obd(lov->tgts[i].ltd_exp), &lov_sfs, - max_age); + err = obd_statfs(class_exp2obd(lov->lov_tgts[i]->ltd_exp), + &lov_sfs, max_age); if (err) { - if (lov->tgts[i].active && !rc) + if (lov->lov_tgts[i]->ltd_active && !rc) rc = err; continue; } - - if (!set) { - memcpy(osfs, &lov_sfs, sizeof(lov_sfs)); - set = 1; - } else { - osfs->os_bfree += lov_sfs.os_bfree; - osfs->os_bavail += lov_sfs.os_bavail; - osfs->os_blocks += lov_sfs.os_blocks; - /* XXX not sure about this one - depends on policy. - * - could be minimum if we always stripe on all OBDs - * (but that would be wrong for any other policy, - * if one of the OBDs has no more objects left) - * - could be sum if we stripe whole objects - * - could be average, just to give a nice number - * - * To give a "reasonable" (if not wholly accurate) - * number, we divide the total number of free objects - * by expected stripe count (watch out for overflow). - */ - LOV_SUM_MAX(osfs->os_files, lov_sfs.os_files); - LOV_SUM_MAX(osfs->os_ffree, lov_sfs.os_ffree); - } + lov_update_statfs(class_exp2obd(lov->lov_tgts[i]->ltd_exp), + osfs, &lov_sfs, set); + set++; } - if (set) { - __u32 expected_stripes = lov->desc.ld_default_stripe_count ? - lov->desc.ld_default_stripe_count : - lov->desc.ld_active_tgt_count; + err = lov_fini_statfs(obd, osfs, set); + qos_update(lov); - if (osfs->os_files != LOV_U64_MAX) - do_div(osfs->os_files, expected_stripes); - if (osfs->os_ffree != LOV_U64_MAX) - do_div(osfs->os_ffree, expected_stripes); - } else if (!rc) - rc = -EIO; - - RETURN(rc); + RETURN(rc ? rc : err); } static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len, @@ -2893,13 +2216,11 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len, switch (cmd) { case OBD_IOC_LOV_GET_CONFIG: { - struct obd_ioctl_data *data = karg; - struct lov_tgt_desc *tgtdesc; + struct obd_ioctl_data *data; struct lov_desc *desc; char *buf = NULL; __u32 *genp; - buf = NULL; len = 0; if (obd_ioctl_getdata(&buf, &len, (void *)uarg)) RETURN(-EINVAL); @@ -2907,17 +2228,17 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len, data = (struct obd_ioctl_data *)buf; if (sizeof(*desc) > data->ioc_inllen1) { - OBD_FREE(buf, len); + obd_ioctl_freedata(buf, len); RETURN(-EINVAL); } if (sizeof(uuidp->uuid) * count > data->ioc_inllen2) { - OBD_FREE(buf, len); + obd_ioctl_freedata(buf, len); RETURN(-EINVAL); } if (sizeof(__u32) * count > data->ioc_inllen3) { - OBD_FREE(buf, len); + obd_ioctl_freedata(buf, len); RETURN(-EINVAL); } @@ -2926,11 +2247,12 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len, uuidp = (struct obd_uuid *)data->ioc_inlbuf2; genp = (__u32 *)data->ioc_inlbuf3; - tgtdesc = lov->tgts; /* the uuid will be empty for deleted OSTs */ - for (i = 0; i < count; i++, uuidp++, genp++, tgtdesc++) { - obd_str2uuid(uuidp, tgtdesc->uuid.uuid); - *genp = tgtdesc->ltd_gen; + for (i = 0; i < count; i++, uuidp++, genp++) { + if (!lov->lov_tgts[i]) + continue; + *uuidp = lov->lov_tgts[i]->ltd_uuid; + *genp = lov->lov_tgts[i]->ltd_gen; } rc = copy_to_user((void *)uarg, buf, len); @@ -2950,28 +2272,36 @@ static int lov_iocontrol(unsigned int cmd, struct obd_export *exp, int len, break; default: { int set = 0; + if (count == 0) RETURN(-ENOTTY); + rc = 0; for (i = 0; i < count; i++) { int err; - /* OST was deleted */ - if (obd_uuid_empty(&lov->tgts[i].uuid)) + /* OST was disconnected */ + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp) continue; - err = obd_iocontrol(cmd, lov->tgts[i].ltd_exp, + err = obd_iocontrol(cmd, lov->lov_tgts[i]->ltd_exp, len, karg, uarg); - if (err) { - if (lov->tgts[i].active) { - CERROR("error: iocontrol OSC %s on OST" - "idx %d: err = %d\n", - lov->tgts[i].uuid.uuid, i, err); + if (err == -ENODATA && cmd == OBD_IOC_POLL_QUOTACHECK) { + RETURN(err); + } else if (err) { + if (lov->lov_tgts[i]->ltd_active) { + CDEBUG(err == -ENOTTY ? + D_IOCTL : D_WARNING, + "iocontrol OSC %s on OST " + "idx %d cmd %x: err = %d\n", + lov_uuid2str(lov, i), + i, cmd, err); if (!rc) rc = err; } - } else + } else { set = 1; + } } if (!set && !rc) rc = -EIO; @@ -2986,24 +2316,26 @@ static int lov_get_info(struct obd_export *exp, __u32 keylen, { struct obd_device *obddev = class_exp2obd(exp); struct lov_obd *lov = &obddev->u.lov; - int i; + int i, rc; ENTRY; if (!vallen || !val) RETURN(-EFAULT); - if (keylen > strlen("lock_to_stripe") && - strcmp(key, "lock_to_stripe") == 0) { + lov_getref(obddev); + + if (KEY_IS("lock_to_stripe")) { struct { char name[16]; struct ldlm_lock *lock; struct lov_stripe_md *lsm; } *data = key; + struct ldlm_res_id *res_id = &data->lock->l_resource->lr_name; struct lov_oinfo *loi; __u32 *stripe = val; if (*vallen < sizeof(*stripe)) - RETURN(-EFAULT); + GOTO(out, rc = -EFAULT); *vallen = sizeof(*stripe); /* XXX This is another one of those bits that will need to @@ -3012,171 +2344,181 @@ static int lov_get_info(struct obd_export *exp, __u32 keylen, /* XXX - it's assumed all the locks for deleted OSTs have * been cancelled. Also, the export for deleted OSTs will * be NULL and won't match the lock's export. */ - for (i = 0, loi = data->lsm->lsm_oinfo; - i < data->lsm->lsm_stripe_count; - i++, loi++) { - if (lov->tgts[loi->loi_ost_idx].ltd_exp == - data->lock->l_conn_export) { + for (i = 0; i < data->lsm->lsm_stripe_count; i++) { + loi = data->lsm->lsm_oinfo[i]; + if (!lov->lov_tgts[loi->loi_ost_idx]) + continue; + if (lov->lov_tgts[loi->loi_ost_idx]->ltd_exp == + data->lock->l_conn_export && + loi->loi_id == res_id->name[0] && + loi->loi_gr == res_id->name[2]) { *stripe = i; - RETURN(0); + GOTO(out, rc = 0); } } - RETURN(-ENXIO); - } else if (keylen >= strlen("size_to_stripe") && - strcmp(key, "size_to_stripe") == 0) { - struct { - int stripe_number; - __u64 size; - struct lov_stripe_md *lsm; - } *data = val; - - if (*vallen < sizeof(*data)) - RETURN(-EFAULT); - - data->size = lov_size_to_stripe(data->lsm, data->size, - data->stripe_number); - RETURN(0); - } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) { + LDLM_ERROR(data->lock, "lock on inode without such object"); + dump_lsm(D_ERROR, data->lsm); + GOTO(out, rc = -ENXIO); + } else if (KEY_IS("last_id")) { obd_id *ids = val; - int rc, size = sizeof(obd_id); + int size = sizeof(obd_id); for (i = 0; i < lov->desc.ld_tgt_count; i++) { - if (!lov->tgts[i].active) + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_active) continue; - rc = obd_get_info(lov->tgts[i].ltd_exp, keylen, key, - &size, &(ids[i])); + rc = obd_get_info(lov->lov_tgts[i]->ltd_exp, + keylen, key, &size, ids + i); if (rc != 0) - RETURN(rc); + GOTO(out, rc); } - RETURN(0); - } else if (keylen >= strlen("lovdesc") && strcmp(key, "lovdesc") == 0) { + GOTO(out, rc = 0); + } else if (KEY_IS(KEY_LOVDESC)) { struct lov_desc *desc_ret = val; *desc_ret = lov->desc; - RETURN(0); + GOTO(out, rc = 0); } - RETURN(-EINVAL); + rc = -EINVAL; +out: + lov_putref(obddev); + RETURN(rc); } -static int lov_set_info(struct obd_export *exp, obd_count keylen, - void *key, obd_count vallen, void *val) +static int lov_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 *obddev = class_exp2obd(exp); struct lov_obd *lov = &obddev->u.lov; - int i, rc = 0; + int i, rc = 0, err; + int no_set = !set; + int incr = 0, check_uuid = 0, do_inactive = 0; ENTRY; -#define KEY_IS(str) \ - (keylen == strlen(str) && memcmp(key, str, keylen) == 0) + if (no_set) { + set = ptlrpc_prep_set(); + if (!set) + RETURN(-ENOMEM); + } - if (KEY_IS("next_id")) { - if (vallen != lov->desc.ld_tgt_count) + if (KEY_IS(KEY_NEXT_ID)) { + if (vallen != lov->desc.ld_tgt_count * sizeof(obd_id)) RETURN(-EINVAL); - for (i = 0; i < lov->desc.ld_tgt_count; i++) { - int er; + vallen = sizeof(obd_id); + incr = sizeof(obd_id); + do_inactive = 1; + } else if (KEY_IS("checksum")) { + do_inactive = 1; + } else if (KEY_IS("unlinked")) { + check_uuid = val ? 1 : 0; + } else if (KEY_IS("evict_by_nid")) { + /* use defaults: do_inactive = incr = 0; */ + } + + lov_getref(obddev); + + for (i = 0; i < lov->desc.ld_tgt_count; i++, val = (char *)val + incr) { + /* OST was disconnected */ + if (!lov->lov_tgts[i] || !lov->lov_tgts[i]->ltd_exp) + continue; - /* OST was deleted */ - if (obd_uuid_empty(&lov->tgts[i].uuid)) - continue; + /* OST is inactive and we don't want inactive OSCs */ + if (!lov->lov_tgts[i]->ltd_active && !do_inactive) + continue; - /* initialize all OSCs, even inactive ones */ + if (KEY_IS(KEY_MDS_CONN)) { + struct mds_group_info *mgi; - er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, - sizeof(obd_id), ((obd_id*)val) + i); - if (!rc) - rc = er; + LASSERT(vallen == sizeof(*mgi)); + mgi = (struct mds_group_info *)val; + + /* Only want a specific OSC */ + if (mgi->uuid && !obd_uuid_equals(mgi->uuid, + &lov->lov_tgts[i]->ltd_uuid)) + continue; + + err = obd_set_info_async(lov->lov_tgts[i]->ltd_exp, + keylen, key, sizeof(int), + &mgi->group, set); + } else { + /* Only want a specific OSC */ + if (check_uuid && + !obd_uuid_equals(val, &lov->lov_tgts[i]->ltd_uuid)) + continue; + + err = obd_set_info_async(lov->lov_tgts[i]->ltd_exp, + keylen, key, vallen, val, set); } - RETURN(rc); - } - if (KEY_IS("growth_count")) { - if (vallen != sizeof(int)) - RETURN(-EINVAL); - } else if (KEY_IS("mds_conn")) { - if (vallen != sizeof(__u32)) - RETURN(-EINVAL); - } else if (KEY_IS("unlinked") || KEY_IS("unrecovery")) { - if (vallen != 0) - RETURN(-EINVAL); - } else { - RETURN(-EINVAL); + + if (!rc) + rc = err; } - for (i = 0; i < lov->desc.ld_tgt_count; i++) { - int er; - - /* OST was deleted */ - if (obd_uuid_empty(&lov->tgts[i].uuid)) - continue; - - if (!val && !lov->tgts[i].active) - continue; - - er = obd_set_info(lov->tgts[i].ltd_exp, keylen, key, vallen, - val); + lov_putref(obddev); + if (no_set) { + err = ptlrpc_set_wait(set); if (!rc) - rc = er; + rc = err; + ptlrpc_set_destroy(set); } RETURN(rc); -#undef KEY_IS - } -/* Merge rss if kms == 0 - * - * Even when merging RSS, we will take the KMS value if it's larger. - * This prevents getattr from stomping on dirty cached pages which - * extend the file size. */ -__u64 lov_merge_size(struct lov_stripe_md *lsm, int kms) +static int lov_checkmd(struct obd_export *exp, struct obd_export *md_exp, + struct lov_stripe_md *lsm) { - struct lov_oinfo *loi; - __u64 size = 0; - int i; - - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++) { - obd_size lov_size, tmpsize; - - tmpsize = loi->loi_kms; - if (kms == 0 && loi->loi_rss > tmpsize) - tmpsize = loi->loi_rss; + int rc; + ENTRY; - lov_size = lov_stripe_size(lsm, tmpsize, i); - if (lov_size > size) - size = lov_size; - } + if (!lsm) + RETURN(0); + LASSERT(md_exp); + LASSERT(lsm_op_find(lsm->lsm_magic) != NULL); + rc = lsm_op_find(lsm->lsm_magic)->lsm_revalidate(lsm, md_exp->exp_obd); - return size; + RETURN(rc); } -EXPORT_SYMBOL(lov_merge_size); -/* Merge blocks */ -__u64 lov_merge_blocks(struct lov_stripe_md *lsm) +int lov_test_and_clear_async_rc(struct lov_stripe_md *lsm) { - struct lov_oinfo *loi; - __u64 blocks = 0; - int i; + int i, rc = 0; + ENTRY; - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++) { - blocks += loi->loi_blocks; + for (i = 0; i < lsm->lsm_stripe_count; i++) { + struct lov_oinfo *loi = lsm->lsm_oinfo[i]; + if (loi->loi_ar.ar_rc && !rc) + rc = loi->loi_ar.ar_rc; + loi->loi_ar.ar_rc = 0; } - return blocks; + RETURN(rc); } -EXPORT_SYMBOL(lov_merge_blocks); +EXPORT_SYMBOL(lov_test_and_clear_async_rc); + -__u64 lov_merge_mtime(struct lov_stripe_md *lsm, __u64 current_time) +static int lov_extent_calc(struct obd_export *exp, struct lov_stripe_md *lsm, + int cmd, __u64 *offset) { - struct lov_oinfo *loi; - int i; + __u32 ssize = lsm->lsm_stripe_size; + __u64 start; - for (i = 0, loi = lsm->lsm_oinfo; i < lsm->lsm_stripe_count; - i++, loi++) { - if (loi->loi_mtime > current_time) - current_time = loi->loi_mtime; + start = *offset; + do_div(start, ssize); + start = start * ssize; + + CDEBUG(D_DLMTRACE, "offset "LPU64", stripe %u, start "LPU64 + ", end "LPU64"\n", *offset, ssize, start, + start + ssize - 1); + if (cmd == OBD_CALC_STRIPE_END) { + *offset = start + ssize - 1; + } else if (cmd == OBD_CALC_STRIPE_START) { + *offset = start; + } else { + LBUG(); } - return current_time; + + RETURN(0); } -EXPORT_SYMBOL(lov_merge_mtime); + #if 0 struct lov_multi_wait { @@ -3197,8 +2539,7 @@ int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm, int rc = 0, i; ENTRY; - if (lsm_bad_magic(lsm)) - RETURN(-EINVAL); + ASSERT_LSM_MAGIC(lsm); if (!exp || !exp->exp_obd) RETURN(-ENODEV); @@ -3225,7 +2566,6 @@ int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm, i++, loi++, lov_lockhp++) { struct ldlm_lock *lock; struct obd_device *obd; - unsigned long irqflags; lock = ldlm_handle2lock(lov_lockhp); if (lock == NULL) { @@ -3243,9 +2583,9 @@ int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm, if (obd != NULL) imp = obd->u.cli.cl_import; if (imp != NULL) { - spin_lock_irqsave(&imp->imp_lock, irqflags); + spin_lock(&imp->imp_lock); queues[i].generation = imp->imp_generation; - spin_unlock_irqrestore(&imp->imp_lock, irqflags); + spin_unlock(&imp->imp_lock); } } @@ -3268,45 +2608,42 @@ int lov_complete_many(struct obd_export *exp, struct lov_stripe_md *lsm, } #endif -void lov_increase_kms(struct obd_export *exp, struct lov_stripe_md *lsm, - obd_off size) +void lov_stripe_lock(struct lov_stripe_md *md) { - struct lov_oinfo *loi; - int stripe = 0; - __u64 kms; - ENTRY; - - if (size > 0) - stripe = lov_stripe_number(lsm, size - 1); - kms = lov_size_to_stripe(lsm, size, stripe); - loi = &(lsm->lsm_oinfo[stripe]); + LASSERT(md->lsm_lock_owner != cfs_current()); + spin_lock(&md->lsm_lock); + LASSERT(md->lsm_lock_owner == NULL); + md->lsm_lock_owner = cfs_current(); +} +EXPORT_SYMBOL(lov_stripe_lock); - CDEBUG(D_INODE, "stripe %d KMS %sincreasing "LPU64"->"LPU64"\n", - stripe, kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms); - if (kms > loi->loi_kms) - loi->loi_kms = kms; - EXIT; +void lov_stripe_unlock(struct lov_stripe_md *md) +{ + LASSERT(md->lsm_lock_owner == cfs_current()); + md->lsm_lock_owner = NULL; + spin_unlock(&md->lsm_lock); } -EXPORT_SYMBOL(lov_increase_kms); +EXPORT_SYMBOL(lov_stripe_unlock); struct obd_ops lov_obd_ops = { .o_owner = THIS_MODULE, - .o_attach = lov_attach, - .o_detach = lov_detach, .o_setup = lov_setup, + .o_precleanup = lov_precleanup, .o_cleanup = lov_cleanup, - o_process_config: lov_process_config, + .o_process_config = lov_process_config, .o_connect = lov_connect, .o_disconnect = lov_disconnect, .o_statfs = lov_statfs, + .o_statfs_async = lov_statfs_async, .o_packmd = lov_packmd, .o_unpackmd = lov_unpackmd, - o_revalidate_md: lov_revalidate_md, + .o_checkmd = lov_checkmd, .o_create = lov_create, .o_destroy = lov_destroy, .o_getattr = lov_getattr, .o_getattr_async = lov_getattr_async, .o_setattr = lov_setattr, + .o_setattr_async = lov_setattr_async, .o_brw = lov_brw, .o_brw_async = lov_brw_async, .o_prep_async_page = lov_prep_async_page, @@ -3315,6 +2652,8 @@ struct obd_ops lov_obd_ops = { .o_queue_group_io = lov_queue_group_io, .o_trigger_group_io = lov_trigger_group_io, .o_teardown_async_page = lov_teardown_async_page, + .o_merge_lvb = lov_merge_lvb, + .o_adjust_kms = lov_adjust_kms, .o_punch = lov_punch, .o_sync = lov_sync, .o_enqueue = lov_enqueue, @@ -3322,36 +2661,66 @@ struct obd_ops lov_obd_ops = { .o_change_cbdata = lov_change_cbdata, .o_cancel = lov_cancel, .o_cancel_unused = lov_cancel_unused, + .o_join_lru = lov_join_lru, .o_iocontrol = lov_iocontrol, .o_get_info = lov_get_info, - .o_set_info = lov_set_info, + .o_set_info_async = lov_set_info_async, + .o_extent_calc = lov_extent_calc, .o_llog_init = lov_llog_init, .o_llog_finish = lov_llog_finish, .o_notify = lov_notify, }; +static quota_interface_t *quota_interface; +extern quota_interface_t lov_quota_interface; + +cfs_mem_cache_t *lov_oinfo_slab; + int __init lov_init(void) { struct lprocfs_static_vars lvars; - int rc; + int rc, rc2; ENTRY; + lov_oinfo_slab = cfs_mem_cache_create("lov_oinfo", + sizeof(struct lov_oinfo), + 0, SLAB_HWCACHE_ALIGN); + if (lov_oinfo_slab == NULL) + return -ENOMEM; lprocfs_init_vars(lov, &lvars); + + request_module("lquota"); + quota_interface = PORTAL_SYMBOL_GET(lov_quota_interface); + init_obd_quota_ops(quota_interface, &lov_obd_ops); + rc = class_register_type(&lov_obd_ops, NULL, lvars.module_vars, - OBD_LOV_DEVICENAME); + LUSTRE_LOV_NAME, NULL); + if (rc) { + if (quota_interface) + PORTAL_SYMBOL_PUT(lov_quota_interface); + rc2 = cfs_mem_cache_destroy(lov_oinfo_slab); + LASSERT(rc2 == 0); + } + RETURN(rc); } #ifdef __KERNEL__ static void /*__exit*/ lov_exit(void) { - class_unregister_type(OBD_LOV_DEVICENAME); + int rc; + + if (quota_interface) + PORTAL_SYMBOL_PUT(lov_quota_interface); + + class_unregister_type(LUSTRE_LOV_NAME); + rc = cfs_mem_cache_destroy(lov_oinfo_slab); + LASSERT(rc == 0); } MODULE_AUTHOR("Cluster File Systems, Inc. "); MODULE_DESCRIPTION("Lustre Logical Object Volume OBD driver"); MODULE_LICENSE("GPL"); -module_init(lov_init); -module_exit(lov_exit); +cfs_module(lov, LUSTRE_VERSION_STRING, lov_init, lov_exit); #endif