X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_recovery.c;h=2bccbf64bcdfe965057697362a0dce00242bf115;hp=6279111fde7893861faea3ce621aa53f8dd14c50;hb=f5ecf580adc319fd9abe41bd79389b50015d5c13;hpb=bf111a4513db2713d5c43460feb5d520a27c00d6 diff --git a/lustre/mdt/mdt_recovery.c b/lustre/mdt/mdt_recovery.c index 6279111..2bccbf6 100644 --- a/lustre/mdt/mdt_recovery.c +++ b/lustre/mdt/mdt_recovery.c @@ -1,6 +1,4 @@ -/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*- - * vim:expandtab:shiftwidth=8:tabstop=8: - * +/* * GPL HEADER START * * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. @@ -26,8 +24,10 @@ * GPL HEADER END */ /* - * Copyright 2008 Sun Microsystems, Inc. All rights reserved + * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved. * Use is subject to license terms. + * + * Copyright (c) 2011, 2013, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -41,16 +41,10 @@ * Author: Pershin Mike */ -#ifndef EXPORT_SYMTAB -# define EXPORT_SYMTAB -#endif #define DEBUG_SUBSYSTEM S_MDS #include "mdt_internal.h" -static int mdt_server_data_update(const struct lu_env *env, - struct mdt_device *mdt); - struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len) { struct lu_buf *buf; @@ -77,163 +71,10 @@ const struct lu_buf *mdt_buf_const(const struct lu_env *env, return buf; } -static inline int mdt_trans_credit_get(const struct lu_env *env, - struct mdt_device *mdt, - enum mdt_txn_op op) -{ - struct dt_device *dev = mdt->mdt_bottom; - int cr; - switch (op) { - case MDT_TXN_CAPA_KEYS_WRITE_OP: - case MDT_TXN_LAST_RCVD_WRITE_OP: - cr = dev->dd_ops->dt_credit_get(env, - dev, - DTO_WRITE_BLOCK); - break; - default: - LBUG(); - } - return cr; -} - -void mdt_trans_credit_init(const struct lu_env *env, - struct mdt_device *mdt, - enum mdt_txn_op op) -{ - struct mdt_thread_info *mti; - struct txn_param *p; - int cr; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - p = &mti->mti_txn_param; - - cr = mdt_trans_credit_get(env, mdt, op); - txn_param_init(p, cr); -} - -struct thandle* mdt_trans_start(const struct lu_env *env, - struct mdt_device *mdt) -{ - struct mdt_thread_info *mti; - struct txn_param *p; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - p = &mti->mti_txn_param; - - /* export can require sync operations */ - if (mti->mti_exp != NULL) - p->tp_sync = mti->mti_exp->exp_need_sync; - - return mdt->mdt_bottom->dd_ops->dt_trans_start(env, mdt->mdt_bottom, p); -} - void mdt_trans_stop(const struct lu_env *env, struct mdt_device *mdt, struct thandle *th) { - mdt->mdt_bottom->dd_ops->dt_trans_stop(env, th); -} - -static inline int mdt_last_rcvd_header_read(const struct lu_env *env, - struct mdt_device *mdt) -{ - struct mdt_thread_info *mti; - int rc; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - - mti->mti_off = 0; - rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd, - mdt_buf(env, &mti->mti_lsd, sizeof(mti->mti_lsd)), - &mti->mti_off); - if (rc == 0) - lsd_le_to_cpu(&mti->mti_lsd, &mdt->mdt_lut.lut_lsd); - - CDEBUG(D_INFO, "read last_rcvd header rc = %d, uuid = %s, " - "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid, - mdt->mdt_lut.lut_lsd.lsd_last_transno); - return rc; -} - -static int mdt_last_rcvd_header_write(const struct lu_env *env, - struct mdt_device *mdt, - struct thandle *th) -{ - struct mdt_thread_info *mti; - int rc; - ENTRY; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - - mti->mti_off = 0; - lsd_cpu_to_le(&mdt->mdt_lut.lut_lsd, &mti->mti_lsd); - - rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd, - mdt_buf_const(env, &mti->mti_lsd, - sizeof(mti->mti_lsd)), - &mti->mti_off, th); - - CDEBUG(D_INFO, "write last_rcvd header rc = %d, uuid = %s, " - "last_transno = "LPU64"\n", rc, mdt->mdt_lut.lut_lsd.lsd_uuid, - mdt->mdt_lut.lut_lsd.lsd_last_transno); - - RETURN(rc); -} - -static int mdt_last_rcvd_read(const struct lu_env *env, struct mdt_device *mdt, - struct lsd_client_data *lcd, loff_t *off, - int index) -{ - struct mdt_thread_info *mti; - struct lsd_client_data *tmp; - int rc; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - tmp = &mti->mti_lcd; - rc = dt_record_read(env, mdt->mdt_lut.lut_last_rcvd, - mdt_buf(env, tmp, sizeof(*tmp)), off); - if (rc == 0) { - check_lcd(mdt2obd_dev(mdt)->obd_name, index, tmp); - lcd_le_to_cpu(tmp, lcd); - } - - CDEBUG(D_INFO, "read lcd @%d rc = %d, uuid = %s, last_transno = "LPU64 - ", last_xid = "LPU64", last_result = %u, last_data = %u, " - "last_close_transno = "LPU64", last_close_xid = "LPU64", " - "last_close_result = %u\n", (int)(*off - sizeof(*tmp)), - rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid, - lcd->lcd_last_result, lcd->lcd_last_data, - lcd->lcd_last_close_transno, lcd->lcd_last_close_xid, - lcd->lcd_last_close_result); - return rc; -} - -static int mdt_last_rcvd_write(const struct lu_env *env, - struct mdt_device *mdt, - struct lsd_client_data *lcd, - loff_t *off, struct thandle *th) -{ - struct mdt_thread_info *mti; - struct lsd_client_data *tmp; - int rc; - - LASSERT(th != NULL); - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - tmp = &mti->mti_lcd; - - lcd_cpu_to_le(lcd, tmp); - - rc = dt_record_write(env, mdt->mdt_lut.lut_last_rcvd, - mdt_buf_const(env, tmp, sizeof(*tmp)), off, th); - - CDEBUG(D_INFO, "write lcd @%d rc = %d, uuid = %s, last_transno = "LPU64 - ", last_xid = "LPU64", last_result = %u, last_data = %u, " - "last_close_transno = "LPU64", last_close_xid = "LPU64" ," - "last_close_result = %u\n", (int)(*off - sizeof(*tmp)), - rc, lcd->lcd_uuid, lcd->lcd_last_transno, lcd->lcd_last_xid, - lcd->lcd_last_result, lcd->lcd_last_data, - lcd->lcd_last_close_transno, lcd->lcd_last_close_xid, - lcd->lcd_last_close_result); - return rc; + dt_trans_stop(env, mdt->mdt_bottom, th); } static int mdt_clients_data_init(const struct lu_env *env, @@ -265,7 +106,7 @@ static int mdt_clients_data_init(const struct lu_env *env, off = lsd->lsd_client_start + cl_idx * lsd->lsd_client_size; - rc = mdt_last_rcvd_read(env, mdt, lcd, &off, cl_idx); + rc = tgt_client_data_read(env, &mdt->mdt_lut, lcd, &off, cl_idx); if (rc) { CERROR("error reading MDS %s idx %d, off %llu: rc %d\n", LAST_RCVD, cl_idx, off, rc); @@ -305,25 +146,25 @@ static int mdt_clients_data_init(const struct lu_env *env, mti->mti_exp = exp; /* copy on-disk lcd to the export */ *exp->exp_target_data.ted_lcd = *lcd; - rc = mdt_client_add(env, mdt, cl_idx); + rc = tgt_client_add(env, exp, cl_idx); /* can't fail existing */ LASSERTF(rc == 0, "rc = %d\n", rc); /* VBR: set export last committed version */ exp->exp_last_committed = last_transno; - cfs_spin_lock(&exp->exp_lock); - exp->exp_connecting = 0; - exp->exp_in_recovery = 0; - cfs_spin_unlock(&exp->exp_lock); - obd->obd_max_recoverable_clients++; - class_export_put(exp); - - CDEBUG(D_OTHER, "client at idx %d has last_transno="LPU64"\n", - cl_idx, last_transno); - /* protect __u64 value update */ - cfs_spin_lock(&mdt->mdt_lut.lut_translock); - mdt->mdt_lut.lut_last_transno = max(last_transno, - mdt->mdt_lut.lut_last_transno); - cfs_spin_unlock(&mdt->mdt_lut.lut_translock); + spin_lock(&exp->exp_lock); + exp->exp_connecting = 0; + exp->exp_in_recovery = 0; + spin_unlock(&exp->exp_lock); + obd->obd_max_recoverable_clients++; + class_export_put(exp); + + CDEBUG(D_OTHER, "client at idx %d has last_transno ="LPU64"\n", + cl_idx, last_transno); + /* protect __u64 value update */ + spin_lock(&mdt->mdt_lut.lut_translock); + mdt->mdt_lut.lut_last_transno = max(last_transno, + mdt->mdt_lut.lut_last_transno); + spin_unlock(&mdt->mdt_lut.lut_translock); } err_client: @@ -341,7 +182,6 @@ static int mdt_server_data_init(const struct lu_env *env, struct mdt_thread_info *mti; struct dt_object *obj; struct lu_attr *la; - struct lustre_disk_data *ldd; unsigned long last_rcvd_size; __u64 mount_count; int rc; @@ -358,7 +198,7 @@ static int mdt_server_data_init(const struct lu_env *env, la = &mti->mti_attr.ma_attr; obj = mdt->mdt_lut.lut_last_rcvd; - rc = obj->do_ops->do_attr_get(env, obj, la, BYPASS_CAPA); + rc = dt_attr_get(env, obj, la, BYPASS_CAPA); if (rc) RETURN(rc); @@ -376,11 +216,12 @@ static int mdt_server_data_init(const struct lu_env *env, lsd->lsd_client_size = LR_CLIENT_SIZE; lsd->lsd_feature_compat = OBD_COMPAT_MDT; lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID; - lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT | - OBD_INCOMPAT_COMMON_LR; + lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT | + OBD_INCOMPAT_COMMON_LR | + OBD_INCOMPAT_MULTI_OI; } else { LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name); - rc = mdt_last_rcvd_header_read(env, mdt); + rc = tgt_server_data_read(env, &mdt->mdt_lut); if (rc) { CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc); GOTO(out, rc); @@ -398,8 +239,6 @@ static int mdt_server_data_init(const struct lu_env *env, } mount_count = lsd->lsd_mount_count; - ldd = lsi->lsi_ldd; - if (lsd->lsd_feature_incompat & ~MDT_INCOMPAT_SUPP) { CERROR("%s: unsupported incompat filesystem feature(s) %x\n", obd->obd_name, @@ -415,24 +254,25 @@ static int mdt_server_data_init(const struct lu_env *env, } /** Interop: evict all clients at first boot with 1.8 last_rcvd */ if (!(lsd->lsd_feature_compat & OBD_COMPAT_20)) { - LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, remove all" - " clients for interop needs\n", obd->obd_name); - simple_truncate(lsi->lsi_srv_mnt->mnt_sb->s_root, - lsi->lsi_srv_mnt, LAST_RCVD, - lsd->lsd_client_start); - last_rcvd_size = lsd->lsd_client_start; + if (last_rcvd_size > lsd->lsd_client_start) { + LCONSOLE_WARN("Mounting %s at first time on 1.8 FS, " + "remove all clients for interop needs\n", + obd->obd_name); + rc = tgt_truncate_last_rcvd(env, &mdt->mdt_lut, + lsd->lsd_client_start); + if (rc) + GOTO(out, rc); + last_rcvd_size = lsd->lsd_client_start; + } /** set 2.0 flag to upgrade/downgrade between 1.8 and 2.0 */ lsd->lsd_feature_compat |= OBD_COMPAT_20; } - if (ldd->ldd_flags & LDD_F_IAM_DIR) - lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR; - lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID; - cfs_spin_lock(&mdt->mdt_lut.lut_translock); - mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno; - cfs_spin_unlock(&mdt->mdt_lut.lut_translock); + spin_lock(&mdt->mdt_lut.lut_translock); + mdt->mdt_lut.lut_last_transno = lsd->lsd_last_transno; + spin_unlock(&mdt->mdt_lut.lut_translock); CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n"); CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n", @@ -463,17 +303,18 @@ static int mdt_server_data_init(const struct lu_env *env, if (rc) GOTO(err_client, rc); - cfs_spin_lock(&mdt->mdt_lut.lut_translock); - /* obd_last_committed is used for compatibility - * with other lustre recovery code */ - obd->obd_last_committed = mdt->mdt_lut.lut_last_transno; - cfs_spin_unlock(&mdt->mdt_lut.lut_translock); + spin_lock(&mdt->mdt_lut.lut_translock); + /* obd_last_committed is used for compatibility + * with other lustre recovery code */ + obd->obd_last_committed = mdt->mdt_lut.lut_last_transno; + spin_unlock(&mdt->mdt_lut.lut_translock); - mdt->mdt_lut.lut_mount_count = mount_count + 1; - lsd->lsd_mount_count = mdt->mdt_lut.lut_mount_count; + obd->u.obt.obt_mount_count = mount_count + 1; + obd->u.obt.obt_instance = (__u32)obd->u.obt.obt_mount_count; + lsd->lsd_mount_count = obd->u.obt.obt_mount_count; /* save it, so mount count and last_transno is current */ - rc = mdt_server_data_update(env, mdt); + rc = tgt_server_data_update(env, &mdt->mdt_lut, 0); if (rc) GOTO(err_client, rc); @@ -485,233 +326,6 @@ out: return rc; } -static int mdt_server_data_update(const struct lu_env *env, - struct mdt_device *mdt) -{ - struct mdt_thread_info *mti; - struct thandle *th; - int rc; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - - mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP); - th = mdt_trans_start(env, mdt); - if (IS_ERR(th)) - RETURN(PTR_ERR(th)); - - CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n", - mdt->mdt_lut.lut_mount_count, mdt->mdt_lut.lut_last_transno); - - cfs_spin_lock(&mdt->mdt_lut.lut_translock); - mdt->mdt_lut.lut_lsd.lsd_last_transno = mdt->mdt_lut.lut_last_transno; - cfs_spin_unlock(&mdt->mdt_lut.lut_translock); - - rc = mdt_last_rcvd_header_write(env, mdt, th); - mdt_trans_stop(env, mdt, th); - return rc; -} - - -int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt) -{ - unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap; - struct mdt_thread_info *mti; - struct tg_export_data *ted; - struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd; - struct obd_device *obd = mdt2obd_dev(mdt); - struct thandle *th; - loff_t off; - int rc; - int cl_idx; - ENTRY; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(mti != NULL); - - ted = &mti->mti_exp->exp_target_data; - - LASSERT(bitmap != NULL); - if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid)) - RETURN(0); - - /* the bitmap operations can handle cl_idx > sizeof(long) * 8, so - * there's no need for extra complication here - */ - cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock); - cl_idx = cfs_find_first_zero_bit(bitmap, LR_MAX_CLIENTS); - if (cl_idx >= LR_MAX_CLIENTS || - OBD_FAIL_CHECK(OBD_FAIL_MDS_CLIENT_ADD)) { - CERROR("no room for %u clients - fix LR_MAX_CLIENTS\n", - cl_idx); - cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock); - RETURN(-EOVERFLOW); - } - cfs_set_bit(cl_idx, bitmap); - cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock); - - CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n", - cl_idx, ted->ted_lcd->lcd_uuid); - - ted->ted_lr_idx = cl_idx; - ted->ted_lr_off = lsd->lsd_client_start + - (cl_idx * lsd->lsd_client_size); - cfs_init_mutex(&ted->ted_lcd_lock); - - LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off); - - /* Write new client data. */ - off = ted->ted_lr_off; - - if (OBD_FAIL_CHECK(OBD_FAIL_TGT_CLIENT_ADD)) - RETURN(-ENOSPC); - - mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP); - - th = mdt_trans_start(env, mdt); - if (IS_ERR(th)) - RETURN(PTR_ERR(th)); - - /* - * Until this operations will be committed the sync is needed - * for this export. This should be done _after_ starting the - * transaction so that many connecting clients will not bring - * server down with lots of sync writes. - */ - mdt_trans_add_cb(th, lut_cb_client, class_export_cb_get(mti->mti_exp)); - cfs_spin_lock(&mti->mti_exp->exp_lock); - mti->mti_exp->exp_need_sync = 1; - cfs_spin_unlock(&mti->mti_exp->exp_lock); - - rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th); - CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len %u)\n", - cl_idx, ted->ted_lr_off, (int)sizeof(*(ted->ted_lcd))); - mdt_trans_stop(env, mdt, th); - - RETURN(rc); -} - -/* Add client data to the MDS. We use a bitmap to locate a free space - * in the last_rcvd file if cl_off is -1 (i.e. a new client). - * Otherwise, we just have to read the data from the last_rcvd file and - * we know its offset. - * - * It should not be possible to fail adding an existing client - otherwise - * mdt_init_server_data() callsite needs to be fixed. - */ -int mdt_client_add(const struct lu_env *env, - struct mdt_device *mdt, int cl_idx) -{ - struct mdt_thread_info *mti; - struct tg_export_data *ted; - unsigned long *bitmap = mdt->mdt_lut.lut_client_bitmap; - struct obd_device *obd = mdt2obd_dev(mdt); - struct lr_server_data *lsd = &mdt->mdt_lut.lut_lsd; - int rc = 0; - ENTRY; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(mti != NULL); - - ted = &mti->mti_exp->exp_target_data; - - LASSERT(bitmap != NULL); - LASSERTF(cl_idx >= 0, "%d\n", cl_idx); - - if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid)) - RETURN(0); - - cfs_spin_lock(&mdt->mdt_lut.lut_client_bitmap_lock); - if (cfs_test_and_set_bit(cl_idx, bitmap)) { - CERROR("MDS client %d: bit already set in bitmap!!\n", - cl_idx); - LBUG(); - } - cfs_spin_unlock(&mdt->mdt_lut.lut_client_bitmap_lock); - - CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n", - cl_idx, ted->ted_lcd->lcd_uuid); - - ted->ted_lr_idx = cl_idx; - ted->ted_lr_off = lsd->lsd_client_start + - (cl_idx * lsd->lsd_client_size); - cfs_init_mutex(&ted->ted_lcd_lock); - - LASSERTF(ted->ted_lr_off > 0, "ted_lr_off = %llu\n", ted->ted_lr_off); - - RETURN(rc); -} - -int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt) -{ - struct mdt_thread_info *mti; - struct tg_export_data *ted; - struct obd_device *obd = mdt2obd_dev(mdt); - struct obd_export *exp; - struct thandle *th; - loff_t off; - int rc = 0; - ENTRY; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(mti != NULL); - - exp = mti->mti_exp; - ted = &exp->exp_target_data; - if (!ted->ted_lcd) - RETURN(0); - - /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */ - if (!strcmp(ted->ted_lcd->lcd_uuid, obd->obd_uuid.uuid)) - GOTO(free, 0); - - CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n", - ted->ted_lr_idx, ted->ted_lr_off); - - off = ted->ted_lr_off; - - /* - * Don't clear ted_lr_idx here as it is likely also unset. At worst we - * leak a client slot that will be cleaned on the next recovery. - */ - if (off <= 0) { - CERROR("client idx %d has offset %lld\n", - ted->ted_lr_idx, off); - GOTO(free, rc = -EINVAL); - } - - /* - * Clear the bit _after_ zeroing out the client so we don't race with - * mdt_client_add and zero out new clients. - */ - if (!cfs_test_bit(ted->ted_lr_idx, mdt->mdt_lut.lut_client_bitmap)) { - CERROR("MDT client %u: bit already clear in bitmap!!\n", - ted->ted_lr_idx); - LBUG(); - } - - /* Make sure the server's last_transno is up to date. - * This should be done before zeroing client slot so last_transno will - * be in server data or in client data in case of failure */ - mdt_server_data_update(env, mdt); - - mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP); - th = mdt_trans_start(env, mdt); - if (IS_ERR(th)) - GOTO(free, rc = PTR_ERR(th)); - - cfs_mutex_down(&ted->ted_lcd_lock); - memset(ted->ted_lcd->lcd_uuid, 0, sizeof ted->ted_lcd->lcd_uuid); - rc = mdt_last_rcvd_write(env, mdt, ted->ted_lcd, &off, th); - cfs_mutex_up(&ted->ted_lcd_lock); - mdt_trans_stop(env, mdt, th); - - CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in " - "%s, rc %d\n", ted->ted_lr_idx, LAST_RCVD, rc); - RETURN(0); -free: - return 0; -} - /* * last_rcvd & last_committed update callbacks */ @@ -733,15 +347,15 @@ static int mdt_last_rcvd_update(struct mdt_thread_info *mti, ted = &req->rq_export->exp_target_data; LASSERT(ted); - cfs_mutex_down(&ted->ted_lcd_lock); - lcd = ted->ted_lcd; - /* if the export has already been disconnected, we have no last_rcvd slot, - * update server data with latest transno then */ - if (lcd == NULL) { - cfs_mutex_up(&ted->ted_lcd_lock); + mutex_lock(&ted->ted_lcd_lock); + lcd = ted->ted_lcd; + /* if the export has already been disconnected, we have no last_rcvd + * slot, update server data with latest transno then */ + if (lcd == NULL) { + mutex_unlock(&ted->ted_lcd_lock); CWARN("commit transaction for disconnected client %s: rc %d\n", req->rq_export->exp_client_uuid.uuid, rc); - err = mdt_last_rcvd_header_write(mti->mti_env, mdt, th); + err = tgt_server_data_write(mti->mti_env, &mdt->mdt_lut, th); RETURN(err); } @@ -749,8 +363,23 @@ static int mdt_last_rcvd_update(struct mdt_thread_info *mti, LASSERT(ergo(mti->mti_transno == 0, rc != 0)); if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE || lustre_msg_get_opc(req->rq_reqmsg) == MDS_DONE_WRITING) { - if (mti->mti_transno != 0) + if (mti->mti_transno != 0) { + if (lcd->lcd_last_close_transno > mti->mti_transno) { + CERROR("Trying to overwrite bigger transno:" + "on-disk: "LPU64", new: "LPU64" " + "replay: %d. see LU-617.\n", + lcd->lcd_last_close_transno, + mti->mti_transno, req_is_replay(req)); + if (req_is_replay(req)) { + spin_lock(&req->rq_export->exp_lock); + req->rq_export->exp_vbr_failed = 1; + spin_unlock(&req->rq_export->exp_lock); + } + mutex_unlock(&ted->ted_lcd_lock); + RETURN(req_is_replay(req) ? -EOVERFLOW : 0); + } lcd->lcd_last_close_transno = mti->mti_transno; + } lcd->lcd_last_close_xid = req->rq_xid; lcd->lcd_last_close_result = rc; } else { @@ -762,8 +391,23 @@ static int mdt_last_rcvd_update(struct mdt_thread_info *mti, lcd->lcd_pre_versions[2] = pre_versions[2]; lcd->lcd_pre_versions[3] = pre_versions[3]; } - if (mti->mti_transno != 0) + if (mti->mti_transno != 0) { + if (lcd->lcd_last_transno > mti->mti_transno) { + CERROR("Trying to overwrite bigger transno:" + "on-disk: "LPU64", new: "LPU64" " + "replay: %d. see LU-617.\n", + lcd->lcd_last_transno, + mti->mti_transno, req_is_replay(req)); + if (req_is_replay(req)) { + spin_lock(&req->rq_export->exp_lock); + req->rq_export->exp_vbr_failed = 1; + spin_unlock(&req->rq_export->exp_lock); + } + mutex_unlock(&ted->ted_lcd_lock); + RETURN(req_is_replay(req) ? -EOVERFLOW : 0); + } lcd->lcd_last_transno = mti->mti_transno; + } lcd->lcd_last_xid = req->rq_xid; lcd->lcd_last_result = rc; /*XXX: save intent_disposition in mdt_thread_info? @@ -772,37 +416,74 @@ static int mdt_last_rcvd_update(struct mdt_thread_info *mti, lcd->lcd_last_data = mti->mti_opdata; } - if (off <= 0) { - CERROR("client idx %d has offset %lld\n", ted->ted_lr_idx, off); - err = -EINVAL; - } else { - err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th); - } - cfs_mutex_up(&ted->ted_lcd_lock); - RETURN(err); + if ((mti->mti_exp->exp_connect_flags & OBD_CONNECT_LIGHTWEIGHT) != 0) { + /* Although lightweight (LW) connections have no slot in + * last_rcvd, we still want to maintain the in-memory + * lsd_client_data structure in order to properly handle reply + * reconstruction. */ + struct lu_target *tg = &mdt->mdt_lut; + bool update = false; + + mutex_unlock(&ted->ted_lcd_lock); + err = 0; + + /* All operations performed by LW clients are synchronous and + * we store the committed transno in the last_rcvd header */ + spin_lock(&tg->lut_translock); + if (mti->mti_transno > tg->lut_lsd.lsd_last_transno) { + tg->lut_lsd.lsd_last_transno = mti->mti_transno; + update = true; + } + spin_unlock(&tg->lut_translock); + + if (update) + err = tgt_server_data_write(mti->mti_env, tg, th); + } else if (off <= 0) { + CERROR("%s: client idx %d has offset %lld\n", + mdt2obd_dev(mdt)->obd_name, ted->ted_lr_idx, off); + mutex_unlock(&ted->ted_lcd_lock); + err = -EINVAL; + } else { + err = tgt_client_data_write(mti->mti_env, &mdt->mdt_lut, lcd, + &off, th); + mutex_unlock(&ted->ted_lcd_lock); + } + RETURN(err); } extern struct lu_context_key mdt_thread_key; /* add credits for last_rcvd update */ static int mdt_txn_start_cb(const struct lu_env *env, - struct txn_param *param, void *cookie) + struct thandle *th, void *cookie) { - struct mdt_device *mdt = cookie; + struct mdt_device *mdt = cookie; + struct mdt_thread_info *mti; + int rc; + ENTRY; - param->tp_credits += mdt_trans_credit_get(env, mdt, - MDT_TXN_LAST_RCVD_WRITE_OP); - return 0; -} + mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); -/* Set new object versions */ -static void mdt_version_set(struct mdt_thread_info *info) -{ - if (info->mti_mos != NULL) { - mo_version_set(info->mti_env, mdt_object_child(info->mti_mos), - info->mti_transno); - info->mti_mos = NULL; - } + LASSERT(mdt->mdt_lut.lut_last_rcvd); + if (mti->mti_exp == NULL) + RETURN(0); + + rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd, + sizeof(struct lsd_client_data), + mti->mti_exp->exp_target_data.ted_lr_off, + th); + if (rc) + return rc; + + rc = dt_declare_record_write(env, mdt->mdt_lut.lut_last_rcvd, + sizeof(struct lr_server_data), 0, th); + if (rc) + return rc; + + if (mti->mti_mos != NULL) + rc = dt_declare_version_set(env, mdt_obj2dt(mti->mti_mos), th); + + return rc; } /* Update last_rcvd records with latests transaction data */ @@ -810,20 +491,14 @@ static int mdt_txn_stop_cb(const struct lu_env *env, struct thandle *txn, void *cookie) { struct mdt_device *mdt = cookie; - struct mdt_txn_info *txi; struct mdt_thread_info *mti; struct ptlrpc_request *req; - /* transno in two contexts - for commit_cb and for thread */ - txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key); mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); req = mdt_info_req(mti); - if (mti->mti_mdt == NULL || req == NULL || mti->mti_no_need_trans) { - txi->txi_transno = 0; - mti->mti_no_need_trans = 0; - return 0; - } + if (mti->mti_mdt == NULL || req == NULL) + return 0; if (mti->mti_has_trans) { /* XXX: currently there are allowed cases, but the wrong cases @@ -834,11 +509,12 @@ static int mdt_txn_stop_cb(const struct lu_env *env, } mti->mti_has_trans = 1; - cfs_spin_lock(&mdt->mdt_lut.lut_translock); + spin_lock(&mdt->mdt_lut.lut_translock); if (txn->th_result != 0) { if (mti->mti_transno != 0) { - CERROR("Replay transno "LPU64" failed: rc %i\n", - mti->mti_transno, txn->th_result); + CERROR("Replay transno "LPU64" failed: rc %d\n", + mti->mti_transno, txn->th_result); + return 0; } } else if (mti->mti_transno == 0) { mti->mti_transno = ++ mdt->mdt_lut.lut_last_transno; @@ -847,13 +523,16 @@ static int mdt_txn_stop_cb(const struct lu_env *env, if (mti->mti_transno > mdt->mdt_lut.lut_last_transno) mdt->mdt_lut.lut_last_transno = mti->mti_transno; } - cfs_spin_unlock(&mdt->mdt_lut.lut_translock); + spin_unlock(&mdt->mdt_lut.lut_translock); /* sometimes the reply message has not been successfully packed */ LASSERT(req != NULL && req->rq_repmsg != NULL); /** VBR: set new versions */ - if (txn->th_result == 0) - mdt_version_set(mti); + if (txn->th_result == 0 && mti->mti_mos != NULL) { + dt_version_set(env, mdt_obj2dt(mti->mti_mos), + mti->mti_transno, txn); + mti->mti_mos = NULL; + } /* filling reply data */ CDEBUG(D_INODE, "transno = "LPU64", last_committed = "LPU64"\n", @@ -861,41 +540,17 @@ static int mdt_txn_stop_cb(const struct lu_env *env, req->rq_transno = mti->mti_transno; lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno); - /* save transno for the commit callback */ - txi->txi_transno = mti->mti_transno; - - /* add separate commit callback for transaction handling because we need - * export as parameter */ - mdt_trans_add_cb(txn, lut_cb_last_committed, - class_export_cb_get(mti->mti_exp)); - + /* if can't add callback, do sync write */ + txn->th_sync |= !!tgt_last_commit_cb_add(txn, &mdt->mdt_lut, + mti->mti_exp, + mti->mti_transno); return mdt_last_rcvd_update(mti, txn); } -/* commit callback, need to update last_committed value */ -static int mdt_txn_commit_cb(const struct lu_env *env, - struct thandle *txn, void *cookie) -{ - struct mdt_device *mdt = cookie; - struct mdt_txn_info *txi; - int i; - - txi = lu_context_key_get(&txn->th_ctx, &mdt_txn_key); - - /* iterate through all additional callbacks */ - for (i = 0; i < txi->txi_cb_count; i++) { - txi->txi_cb[i].lut_cb_func(&mdt->mdt_lut, txi->txi_transno, - txi->txi_cb[i].lut_cb_data, 0); - } - return 0; -} - int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt, struct obd_device *obd, struct lustre_sb_info *lsi) { - struct lu_fid fid; - struct dt_object *o; int rc = 0; ENTRY; @@ -905,7 +560,7 @@ int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt, /* prepare transactions callbacks */ mdt->mdt_txn_cb.dtc_txn_start = mdt_txn_start_cb; mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb; - mdt->mdt_txn_cb.dtc_txn_commit = mdt_txn_commit_cb; + mdt->mdt_txn_cb.dtc_txn_commit = NULL; mdt->mdt_txn_cb.dtc_cookie = mdt; mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD; CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage); @@ -913,28 +568,8 @@ int mdt_fs_setup(const struct lu_env *env, struct mdt_device *mdt, dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb); rc = mdt_server_data_init(env, mdt, lsi); - if (rc) - RETURN(rc); - o = dt_store_open(env, mdt->mdt_bottom, "", CAPA_KEYS, &fid); - if (!IS_ERR(o)) { - mdt->mdt_ck_obj = o; - rc = mdt_capa_keys_init(env, mdt); - if (rc) - GOTO(put_ck_object, rc); - } else { - rc = PTR_ERR(o); - CERROR("cannot open %s: rc = %d\n", CAPA_KEYS, rc); - GOTO(disconnect_exports, rc); - } - RETURN(0); - -put_ck_object: - lu_object_put(env, &o->do_lu); - mdt->mdt_ck_obj = NULL; -disconnect_exports: - class_disconnect_exports(obd); - return rc; + RETURN(rc); } void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt) @@ -952,14 +587,14 @@ void mdt_fs_cleanup(const struct lu_env *env, struct mdt_device *mdt) /* reconstruction code */ static void mdt_steal_ack_locks(struct ptlrpc_request *req) { + struct ptlrpc_service_part *svcpt; struct obd_export *exp = req->rq_export; cfs_list_t *tmp; struct ptlrpc_reply_state *oldrep; - struct ptlrpc_service *svc; int i; /* CAVEAT EMPTOR: spinlock order */ - cfs_spin_lock(&exp->exp_lock); + spin_lock(&exp->exp_lock); cfs_list_for_each (tmp, &exp->exp_outstanding_replies) { oldrep = cfs_list_entry(tmp, struct ptlrpc_reply_state, rs_exp_list); @@ -973,16 +608,16 @@ static void mdt_steal_ack_locks(struct ptlrpc_request *req) lustre_msg_get_opc(req->rq_reqmsg), oldrep->rs_opc); - svc = oldrep->rs_service; - cfs_spin_lock (&svc->srv_lock); + svcpt = oldrep->rs_svcpt; + spin_lock(&svcpt->scp_rep_lock); cfs_list_del_init (&oldrep->rs_exp_list); - CWARN("Stealing %d locks from rs %p x"LPD64".t"LPD64 - " o%d NID %s\n", - oldrep->rs_nlocks, oldrep, - oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc, - libcfs_nid2str(exp->exp_connection->c_peer.nid)); + CDEBUG(D_HA, "Stealing %d locks from rs %p x"LPD64".t"LPD64 + " o%d NID %s\n", + oldrep->rs_nlocks, oldrep, + oldrep->rs_xid, oldrep->rs_transno, oldrep->rs_opc, + libcfs_nid2str(exp->exp_connection->c_peer.nid)); for (i = 0; i < oldrep->rs_nlocks; i++) ptlrpc_save_lock(req, &oldrep->rs_locks[i], @@ -990,14 +625,14 @@ static void mdt_steal_ack_locks(struct ptlrpc_request *req) oldrep->rs_nlocks = 0; DEBUG_REQ(D_HA, req, "stole locks for"); - cfs_spin_lock(&oldrep->rs_lock); - ptlrpc_schedule_difficult_reply (oldrep); - cfs_spin_unlock(&oldrep->rs_lock); - - cfs_spin_unlock (&svc->srv_lock); - break; - } - cfs_spin_unlock(&exp->exp_lock); + spin_lock(&oldrep->rs_lock); + ptlrpc_schedule_difficult_reply(oldrep); + spin_unlock(&oldrep->rs_lock); + + spin_unlock(&svcpt->scp_rep_lock); + break; + } + spin_unlock(&exp->exp_lock); } /** @@ -1080,7 +715,7 @@ static void mdt_reconstruct_create(struct mdt_thread_info *mti, body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY); mti->mti_attr.ma_need = MA_INODE; mti->mti_attr.ma_valid = 0; - rc = mo_attr_get(mti->mti_env, mdt_object_child(child), &mti->mti_attr); + rc = mdt_attr_get_complex(mti, child, &mti->mti_attr); if (rc == -EREMOTE) { /* object was created on remote server */ req->rq_status = rc; @@ -1120,7 +755,7 @@ static void mdt_reconstruct_setattr(struct mdt_thread_info *mti, } mti->mti_attr.ma_need = MA_INODE; mti->mti_attr.ma_valid = 0; - mo_attr_get(mti->mti_env, mdt_object_child(obj), &mti->mti_attr); + mdt_attr_get_complex(mti, obj, &mti->mti_attr); mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr, mdt_object_fid(obj)); if (mti->mti_ioepoch && (mti->mti_ioepoch->flags & MF_EPOCH_OPEN)) { @@ -1129,17 +764,17 @@ static void mdt_reconstruct_setattr(struct mdt_thread_info *mti, repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY); repbody->ioepoch = obj->mot_ioepoch; - cfs_spin_lock(&med->med_open_lock); - cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) { - if (mfd->mfd_xid == req->rq_xid) - break; - } - LASSERT(&mfd->mfd_list != &med->med_open_head); - cfs_spin_unlock(&med->med_open_lock); - repbody->handle.cookie = mfd->mfd_handle.h_cookie; - } - - mdt_object_put(mti->mti_env, obj); + spin_lock(&med->med_open_lock); + cfs_list_for_each_entry(mfd, &med->med_open_head, mfd_list) { + if (mfd->mfd_xid == req->rq_xid) + break; + } + LASSERT(&mfd->mfd_list != &med->med_open_head); + spin_unlock(&med->med_open_lock); + repbody->handle.cookie = mfd->mfd_handle.h_cookie; + } + + mdt_object_put(mti->mti_env, obj); } typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,