X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=blobdiff_plain;f=lustre%2Fmdt%2Fmdt_recovery.c;h=f9260a388c3405eadf2fe16df762e86931e1cbfb;hp=4e03c7f04e595fd3d6d35246c908d4d1a1e38977;hb=8daba6a7381a2fb8cc933f7e9486f60e659465d4;hpb=57e224145f7c390e33c295d95c735934d0c54883 diff --git a/lustre/mdt/mdt_recovery.c b/lustre/mdt/mdt_recovery.c index 4e03c7f..f9260a3 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, 2014, Intel Corporation. */ /* * This file is part of Lustre, http://www.lustre.org/ @@ -41,17 +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, - int need_sync); - struct lu_buf *mdt_buf(const struct lu_env *env, void *area, ssize_t len) { struct lu_buf *buf; @@ -78,909 +71,50 @@ 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_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_lsd); - - CDEBUG(D_INFO, "read last_rcvd header rc = %d:\n" - "uuid = %s\n" - "last_transno = "LPU64"\n", - rc, mdt->mdt_lsd.lsd_uuid, - mdt->mdt_lsd.lsd_last_transno); - return rc; -} - -static inline int mdt_last_rcvd_header_write(const struct lu_env *env, - struct mdt_device *mdt, - int need_sync) -{ - struct mdt_thread_info *mti; - struct thandle *th; - int rc; - ENTRY; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - - if (mti->mti_exp) { - spin_lock(&mti->mti_exp->exp_lock); - mti->mti_exp->exp_need_sync = need_sync; - spin_unlock(&mti->mti_exp->exp_lock); - } - 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)); - - mti->mti_off = 0; - lsd_cpu_to_le(&mdt->mdt_lsd, &mti->mti_lsd); - - if (need_sync && mti->mti_exp) - mdt_trans_add_cb(th, lut_cb_client, mti->mti_exp); - - rc = dt_record_write(env, mdt->mdt_last_rcvd, - mdt_buf_const(env, &mti->mti_lsd, - sizeof(mti->mti_lsd)), - &mti->mti_off, th); - - mdt_trans_stop(env, mdt, th); - - CDEBUG(D_INFO, "write last_rcvd header rc = %d:\n" - "uuid = %s\nlast_transno = "LPU64"\n", - rc, mdt->mdt_lsd.lsd_uuid, mdt->mdt_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) -{ - 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_last_rcvd, - mdt_buf(env, tmp, sizeof(*tmp)), off); - if (rc == 0) - lcd_le_to_cpu(tmp, lcd); - - CDEBUG(D_INFO, "read lcd @%d rc = %d:\n" - "uuid = %s\n" - "last_transno = "LPU64"\n" - "last_xid = "LPU64"\n" - "last_result = %u\n" - "last_data = %u\n" - "last_close_transno = "LPU64"\n" - "last_close_xid = "LPU64"\n" - "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_last_rcvd, - mdt_buf_const(env, tmp, sizeof(*tmp)), off, th); - - CDEBUG(D_INFO, "write lcd @%d rc = %d:\n" - "uuid = %s\n" - "last_transno = "LPU64"\n" - "last_xid = "LPU64"\n" - "last_result = %u\n" - "last_data = %u\n" - "last_close_transno = "LPU64"\n" - "last_close_xid = "LPU64"\n" - "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_clients_data_init(const struct lu_env *env, - struct mdt_device *mdt, - unsigned long last_size) -{ - struct lr_server_data *lsd = &mdt->mdt_lsd; - struct lsd_client_data *lcd = NULL; - struct obd_device *obd = mdt2obd_dev(mdt); - loff_t off; - int cl_idx; - int rc = 0; - ENTRY; - - /* When we do a clean MDS shutdown, we save the last_transno into - * the header. If we find clients with higher last_transno values - * then those clients may need recovery done. */ - LASSERT(atomic_read(&obd->obd_req_replay_clients) == 0); - for (cl_idx = 0, off = lsd->lsd_client_start; - off < last_size; cl_idx++) { - __u64 last_transno; - struct obd_export *exp; - - if (!lcd) { - OBD_ALLOC_PTR(lcd); - if (!lcd) - RETURN(-ENOMEM); - } - - off = lsd->lsd_client_start + - cl_idx * lsd->lsd_client_size; - - rc = mdt_last_rcvd_read(env, mdt, lcd, &off); - if (rc) { - CERROR("error reading MDS %s idx %d, off %llu: rc %d\n", - LAST_RCVD, cl_idx, off, rc); - rc = 0; - break; /* read error shouldn't cause startup to fail */ - } - - if (lcd->lcd_uuid[0] == '\0') { - CDEBUG(D_INFO, "skipping zeroed client at offset %d\n", - cl_idx); - continue; - } - - last_transno = lcd_last_transno(lcd); - - /* These exports are cleaned up by mdt_obd_disconnect(), so - * they need to be set up like real exports as - * mdt_obd_connect() does. - */ - CDEBUG(D_HA, "RCVRNG CLIENT uuid: %s idx: %d lr: "LPU64 - " srv lr: "LPU64" lx: "LPU64"\n", lcd->lcd_uuid, cl_idx, - last_transno, lsd->lsd_last_transno, - lcd_last_xid(lcd)); - - exp = class_new_export(obd, (struct obd_uuid *)lcd->lcd_uuid); - if (IS_ERR(exp)) { - if (PTR_ERR(exp) == -EALREADY) { - /* export already exists, zero out this one */ - lcd->lcd_uuid[0] = '\0'; - } else - GOTO(err_client, rc = PTR_ERR(exp)); - } else { - struct mdt_thread_info *mti; - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(mti != NULL); - mti->mti_exp = exp; - exp->exp_mdt_data.med_lcd = lcd; - rc = mdt_client_add(env, mdt, 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; - lcd = NULL; - 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_transno_lock); - mdt->mdt_last_transno = max(last_transno, - mdt->mdt_last_transno); - spin_unlock(&mdt->mdt_transno_lock); - } - -err_client: - if (lcd) - OBD_FREE_PTR(lcd); - RETURN(rc); -} - -static int mdt_server_data_init(const struct lu_env *env, - struct mdt_device *mdt, - struct lustre_sb_info *lsi) -{ - struct lr_server_data *lsd = &mdt->mdt_lsd; - struct lsd_client_data *lcd = NULL; - struct obd_device *obd = mdt2obd_dev(mdt); - 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; - ENTRY; - - /* ensure padding in the struct is the correct size */ - CLASSERT(offsetof(struct lr_server_data, lsd_padding) + - sizeof(lsd->lsd_padding) == LR_SERVER_SIZE); - CLASSERT(offsetof(struct lsd_client_data, lcd_padding) + - sizeof(lcd->lcd_padding) == LR_CLIENT_SIZE); - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(mti != NULL); - la = &mti->mti_attr.ma_attr; - - obj = mdt->mdt_last_rcvd; - rc = obj->do_ops->do_attr_get(env, mdt->mdt_last_rcvd, la, BYPASS_CAPA); - if (rc) - RETURN(rc); - - last_rcvd_size = (unsigned long)la->la_size; - - if (last_rcvd_size == 0) { - LCONSOLE_WARN("%s: new disk, initializing\n", obd->obd_name); - - memcpy(lsd->lsd_uuid, obd->obd_uuid.uuid, - sizeof(lsd->lsd_uuid)); - lsd->lsd_last_transno = 0; - lsd->lsd_mount_count = 0; - lsd->lsd_server_size = LR_SERVER_SIZE; - lsd->lsd_client_start = LR_CLIENT_START; - lsd->lsd_client_size = LR_CLIENT_SIZE; - lsd->lsd_feature_rocompat = OBD_ROCOMPAT_LOVOBJID; - lsd->lsd_feature_incompat = OBD_INCOMPAT_MDT | - OBD_INCOMPAT_COMMON_LR; - } else { - LCONSOLE_WARN("%s: used disk, loading\n", obd->obd_name); - rc = mdt_last_rcvd_header_read(env, mdt); - if (rc) { - CERROR("error reading MDS %s: rc %d\n", LAST_RCVD, rc); - GOTO(out, rc); - } - if (strcmp(lsd->lsd_uuid, obd->obd_uuid.uuid) != 0) { - LCONSOLE_ERROR_MSG(0x157, "Trying to start OBD %s using" - "the wrong disk %s. Were the /dev/ " - "assignments rearranged?\n", - obd->obd_uuid.uuid, lsd->lsd_uuid); - GOTO(out, rc = -EINVAL); - } - } - mount_count = lsd->lsd_mount_count; - - ldd = lsi->lsi_ldd; - - if (ldd->ldd_flags & LDD_F_IAM_DIR) - lsd->lsd_feature_incompat |= OBD_INCOMPAT_IAM_DIR; - - lsd->lsd_feature_compat = OBD_COMPAT_MDT; - lsd->lsd_feature_incompat |= OBD_INCOMPAT_FID; - - spin_lock(&mdt->mdt_transno_lock); - mdt->mdt_last_transno = lsd->lsd_last_transno; - spin_unlock(&mdt->mdt_transno_lock); - - CDEBUG(D_INODE, "========BEGIN DUMPING LAST_RCVD========\n"); - CDEBUG(D_INODE, "%s: server last_transno: "LPU64"\n", - obd->obd_name, mdt->mdt_last_transno); - CDEBUG(D_INODE, "%s: server mount_count: "LPU64"\n", - obd->obd_name, mount_count + 1); - CDEBUG(D_INODE, "%s: server data size: %u\n", - obd->obd_name, lsd->lsd_server_size); - CDEBUG(D_INODE, "%s: per-client data start: %u\n", - obd->obd_name, lsd->lsd_client_start); - CDEBUG(D_INODE, "%s: per-client data size: %u\n", - obd->obd_name, lsd->lsd_client_size); - CDEBUG(D_INODE, "%s: last_rcvd size: %lu\n", - obd->obd_name, last_rcvd_size); - CDEBUG(D_INODE, "%s: last_rcvd clients: %lu\n", obd->obd_name, - last_rcvd_size <= lsd->lsd_client_start ? 0 : - (last_rcvd_size - lsd->lsd_client_start) / - lsd->lsd_client_size); - CDEBUG(D_INODE, "========END DUMPING LAST_RCVD========\n"); - - if (!lsd->lsd_server_size || !lsd->lsd_client_start || - !lsd->lsd_client_size) { - CERROR("Bad last_rcvd contents!\n"); - GOTO(out, rc = -EINVAL); - } - - rc = mdt_clients_data_init(env, mdt, last_rcvd_size); - if (rc) - GOTO(err_client, rc); - - spin_lock(&mdt->mdt_transno_lock); - /* obd_last_committed is used for compatibility - * with other lustre recovery code */ - obd->obd_last_committed = mdt->mdt_last_transno; - spin_unlock(&mdt->mdt_transno_lock); - - mdt->mdt_mount_count++; - lsd->lsd_mount_count = mdt->mdt_mount_count; - - /* save it, so mount count and last_transno is current */ - rc = mdt_server_data_update(env, mdt, (mti->mti_exp && - mti->mti_exp->exp_need_sync)); - if (rc) - GOTO(err_client, rc); - - RETURN(0); - -err_client: - class_disconnect_exports(obd); -out: - return rc; -} - -static int mdt_server_data_update(const struct lu_env *env, - struct mdt_device *mdt, - int need_sync) -{ - int rc = 0; - ENTRY; - - CDEBUG(D_SUPER, "MDS mount_count is "LPU64", last_transno is "LPU64"\n", - mdt->mdt_mount_count, mdt->mdt_last_transno); - - spin_lock(&mdt->mdt_transno_lock); - mdt->mdt_lsd.lsd_last_transno = mdt->mdt_last_transno; - spin_unlock(&mdt->mdt_transno_lock); - - /* - * This may be called from difficult reply handler and - * mdt->mdt_last_rcvd may be NULL that time. - */ - if (mdt->mdt_last_rcvd != NULL) - rc = mdt_last_rcvd_header_write(env, mdt, need_sync); - RETURN(rc); -} - -int mdt_client_new(const struct lu_env *env, struct mdt_device *mdt) -{ - unsigned long *bitmap = mdt->mdt_client_bitmap; - struct mdt_thread_info *mti; - struct mdt_export_data *med; - struct lsd_client_data *lcd; - struct lr_server_data *lsd = &mdt->mdt_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); - - med = &mti->mti_exp->exp_mdt_data; - lcd = med->med_lcd; - - LASSERT(bitmap != NULL); - if (!strcmp(med->med_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 - */ - spin_lock(&mdt->mdt_client_bitmap_lock); - cl_idx = 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); - spin_unlock(&mdt->mdt_client_bitmap_lock); - RETURN(-EOVERFLOW); - } - set_bit(cl_idx, bitmap); - spin_unlock(&mdt->mdt_client_bitmap_lock); - - CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n", - cl_idx, med->med_lcd->lcd_uuid); - - med->med_lr_idx = cl_idx; - med->med_lr_off = lsd->lsd_client_start + - (cl_idx * lsd->lsd_client_size); - init_mutex(&med->med_lcd_lock); - - LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off); - - /* Write new client data. */ - off = med->med_lr_off; - 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, mti->mti_exp); - spin_lock(&mti->mti_exp->exp_lock); - mti->mti_exp->exp_need_sync = 1; - spin_unlock(&mti->mti_exp->exp_lock); - - rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th); - CDEBUG(D_INFO, "wrote client lcd at idx %u off %llu (len "LPSZ")\n", - cl_idx, med->med_lr_off, sizeof(*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 mdt_export_data *med; - unsigned long *bitmap = mdt->mdt_client_bitmap; - struct obd_device *obd = mdt2obd_dev(mdt); - struct lr_server_data *lsd = &mdt->mdt_lsd; - int rc = 0; - ENTRY; - - mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); - LASSERT(mti != NULL); - - med = &mti->mti_exp->exp_mdt_data; - - LASSERT(bitmap != NULL); - LASSERTF(cl_idx >= 0, "%d\n", cl_idx); - - if (!strcmp(med->med_lcd->lcd_uuid, obd->obd_uuid.uuid)) - RETURN(0); - - spin_lock(&mdt->mdt_client_bitmap_lock); - if (test_and_set_bit(cl_idx, bitmap)) { - CERROR("MDS client %d: bit already set in bitmap!!\n", - cl_idx); - LBUG(); - } - spin_unlock(&mdt->mdt_client_bitmap_lock); - - CDEBUG(D_INFO, "client at idx %d with UUID '%s' added\n", - cl_idx, med->med_lcd->lcd_uuid); - - med->med_lr_idx = cl_idx; - med->med_lr_off = lsd->lsd_client_start + - (cl_idx * lsd->lsd_client_size); - init_mutex(&med->med_lcd_lock); - - LASSERTF(med->med_lr_off > 0, "med_lr_off = %llu\n", med->med_lr_off); - - RETURN(rc); -} - -int mdt_client_del(const struct lu_env *env, struct mdt_device *mdt) -{ - struct mdt_thread_info *mti; - struct mdt_export_data *med; - struct lsd_client_data *lcd; - struct obd_device *obd = mdt2obd_dev(mdt); - struct obd_export *exp; - struct thandle *th; - int need_sync; - 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; - med = &exp->exp_mdt_data; - lcd = med->med_lcd; - if (!lcd) - RETURN(0); - - /* XXX: If lcd_uuid were a real obd_uuid, I could use obd_uuid_equals */ - if (!strcmp(lcd->lcd_uuid, obd->obd_uuid.uuid)) - GOTO(free, 0); - - CDEBUG(D_INFO, "freeing client at idx %u, offset %lld\n", - med->med_lr_idx, med->med_lr_off); - - off = med->med_lr_off; - - /* - * Don't clear med_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", - med->med_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 (!test_bit(med->med_lr_idx, mdt->mdt_client_bitmap)) { - CERROR("MDT client %u: bit already clear in bitmap!!\n", - med->med_lr_idx); - LBUG(); - } - - /* Don't force sync on disconnect if aborting recovery, - * or it does num_clients * num_osts. b=17194 */ - need_sync = (!exp->exp_libclient || exp->exp_need_sync) && - !(exp->exp_flags & OBD_OPT_ABORT_RECOV); - - /* - * This may be called from difficult reply handler path and - * mdt->mdt_last_rcvd may be NULL that time. - */ - if (mdt->mdt_last_rcvd != NULL) { - mdt_trans_credit_init(env, mdt, MDT_TXN_LAST_RCVD_WRITE_OP); - - spin_lock(&exp->exp_lock); - exp->exp_need_sync = need_sync; - spin_unlock(&exp->exp_lock); - - th = mdt_trans_start(env, mdt); - if (IS_ERR(th)) - GOTO(free, rc = PTR_ERR(th)); - - if (need_sync) { - /* - * Until this operations will be committed the sync - * is needed for this export. - */ - mdt_trans_add_cb(th, lut_cb_client, exp); - } - - mutex_down(&med->med_lcd_lock); - memset(lcd, 0, sizeof *lcd); - - rc = mdt_last_rcvd_write(env, mdt, lcd, &off, th); - mutex_up(&med->med_lcd_lock); - mdt_trans_stop(env, mdt, th); - } - - CDEBUG(rc == 0 ? D_INFO : D_ERROR, "Zeroing out client idx %u in " - "%s %ssync rc %d\n", med->med_lr_idx, LAST_RCVD, - need_sync ? "" : "a", rc); - - spin_lock(&mdt->mdt_client_bitmap_lock); - clear_bit(med->med_lr_idx, mdt->mdt_client_bitmap); - spin_unlock(&mdt->mdt_client_bitmap_lock); - - /* - * Make sure the server's last_transno is up to date. Do this - * after the client is freed so we know all the client's - * transactions have been committed. - */ - mdt_server_data_update(env, mdt, need_sync); - - EXIT; -free: - OBD_FREE_PTR(lcd); - med->med_lcd = NULL; - return 0; -} - /* * last_rcvd & last_committed update callbacks */ -static int mdt_last_rcvd_update(struct mdt_thread_info *mti, - struct thandle *th) -{ - struct mdt_device *mdt = mti->mti_mdt; - struct ptlrpc_request *req = mdt_info_req(mti); - struct mdt_export_data *med; - struct lsd_client_data *lcd; - loff_t off; - int err; - __s32 rc = th->th_result; - - ENTRY; - LASSERT(req); - LASSERT(req->rq_export); - LASSERT(mdt); - med = &req->rq_export->exp_mdt_data; - LASSERT(med); - lcd = med->med_lcd; - /* if the export has already been failed, we have no last_rcvd slot */ - if (req->rq_export->exp_failed) { - CWARN("commit transaction for disconnected client %s: rc %d\n", - req->rq_export->exp_client_uuid.uuid, rc); - if (rc == 0) - rc = -ENOTCONN; - RETURN(rc); - } - - off = med->med_lr_off; - LASSERT(ergo(mti->mti_transno == 0, rc != 0)); - mutex_down(&med->med_lcd_lock); - 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) - lcd->lcd_last_close_transno = mti->mti_transno; - lcd->lcd_last_close_xid = req->rq_xid; - lcd->lcd_last_close_result = rc; - } else { - /* VBR: save versions in last_rcvd for reconstruct. */ - __u64 *pre_versions = lustre_msg_get_versions(req->rq_repmsg); - if (pre_versions) { - lcd->lcd_pre_versions[0] = pre_versions[0]; - lcd->lcd_pre_versions[1] = pre_versions[1]; - lcd->lcd_pre_versions[2] = pre_versions[2]; - lcd->lcd_pre_versions[3] = pre_versions[3]; - } - if (mti->mti_transno != 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? - * also there is bug - intent_dispostion is __u64, - * see struct ldlm_reply->lock_policy_res1; */ - lcd->lcd_last_data = mti->mti_opdata; - } - - if (off <= 0) { - CERROR("client idx %d has offset %lld\n", med->med_lr_idx, off); - err = -EINVAL; - } else { - err = mdt_last_rcvd_write(mti->mti_env, mdt, lcd, &off, th); - } - mutex_up(&med->med_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 mdt_device *mdt = cookie; - - param->tp_credits += mdt_trans_credit_get(env, mdt, - MDT_TXN_LAST_RCVD_WRITE_OP); - return 0; -} - -/* Set new object versions */ -static void mdt_versions_set(struct mdt_thread_info *info) -{ - int i; - for (i = 0; i < PTLRPC_NUM_VERSIONS; i++) - if (info->mti_mos[i] != NULL) - mo_version_set(info->mti_env, - mdt_object_child(info->mti_mos[i]), - info->mti_transno); -} - -/* Update last_rcvd records with latests transaction data */ +/* This callback notifies MDT that transaction was done. This is needed by + * mdt_save_lock() only. It is similar to new target code and will be removed + * as mdt_save_lock() will be converted to use target structures */ 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_has_trans) { - /* XXX: currently there are allowed cases, but the wrong cases - * are also possible, so better check is needed here */ - CDEBUG(D_INFO, "More than one transaction "LPU64"\n", - mti->mti_transno); - return 0; - } - - mti->mti_has_trans = 1; - spin_lock(&mdt->mdt_transno_lock); - if (txn->th_result != 0) { - if (mti->mti_transno != 0) { - CERROR("Replay transno "LPU64" failed: rc %i\n", - mti->mti_transno, txn->th_result); - } - } else if (mti->mti_transno == 0) { - mti->mti_transno = ++ mdt->mdt_last_transno; - } else { - /* should be replay */ - if (mti->mti_transno > mdt->mdt_last_transno) - mdt->mdt_last_transno = mti->mti_transno; - } - spin_unlock(&mdt->mdt_transno_lock); - /* sometimes the reply message has not been successfully packed */ - LASSERT(req != NULL && req->rq_repmsg != NULL); + struct mdt_thread_info *mti; - /** VBR: set new versions */ - if (txn->th_result == 0) - mdt_versions_set(mti); + mti = lu_context_key_get(&env->le_ctx, &mdt_thread_key); + LASSERT(mti); - /* filling reply data */ - CDEBUG(D_INODE, "transno = %llu, last_committed = %llu\n", - mti->mti_transno, req->rq_export->exp_obd->obd_last_committed); - - req->rq_transno = mti->mti_transno; - lustre_msg_set_transno(req->rq_repmsg, mti->mti_transno); - lustre_msg_set_last_xid(req->rq_repmsg, - lcd_last_xid(req->rq_export->exp_mdt_data.med_lcd)); - /* 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, mti->mti_exp); - - 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; + if (mti->mti_has_trans) + CDEBUG(D_INFO, "More than one transaction\n"); + else + mti->mti_has_trans = 1; + 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 obd_device *obd, struct lustre_sb_info *lsi) { - struct lu_fid fid; - struct dt_object *o; - int rc = 0; - ENTRY; + int rc = 0; - if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP)) - RETURN(-ENOENT); + ENTRY; - /* 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_cookie = mdt; - mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD; - CFS_INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage); + if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FS_SETUP)) + RETURN(-ENOENT); - dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb); + /* prepare transactions callbacks */ + mdt->mdt_txn_cb.dtc_txn_start = NULL; + mdt->mdt_txn_cb.dtc_txn_stop = mdt_txn_stop_cb; + mdt->mdt_txn_cb.dtc_txn_commit = NULL; + mdt->mdt_txn_cb.dtc_cookie = NULL; + mdt->mdt_txn_cb.dtc_tag = LCT_MD_THREAD; + INIT_LIST_HEAD(&mdt->mdt_txn_cb.dtc_linkage); - 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); + dt_txn_callback_add(mdt->mdt_bottom, &mdt->mdt_txn_cb); -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) @@ -998,38 +132,37 @@ 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 obd_export *exp = req->rq_export; - struct list_head *tmp; - struct ptlrpc_reply_state *oldrep; - struct ptlrpc_service *svc; - int i; - - /* CAVEAT EMPTOR: spinlock order */ - spin_lock(&exp->exp_lock); - list_for_each (tmp, &exp->exp_outstanding_replies) { - oldrep = list_entry(tmp, struct ptlrpc_reply_state,rs_exp_list); + struct ptlrpc_service_part *svcpt; + struct obd_export *exp = req->rq_export; + struct list_head *tmp; + struct ptlrpc_reply_state *oldrep; + int i; + + /* CAVEAT EMPTOR: spinlock order */ + spin_lock(&exp->exp_lock); + list_for_each(tmp, &exp->exp_outstanding_replies) { + oldrep = list_entry(tmp, struct ptlrpc_reply_state, + rs_exp_list); if (oldrep->rs_xid != req->rq_xid) continue; - if (lustre_msg_get_opc(oldrep->rs_msg) != - lustre_msg_get_opc(req->rq_reqmsg)) + if (oldrep->rs_opc != lustre_msg_get_opc(req->rq_reqmsg)) CERROR ("Resent req xid "LPU64" has mismatched opc: " "new %d old %d\n", req->rq_xid, lustre_msg_get_opc(req->rq_reqmsg), - lustre_msg_get_opc(oldrep->rs_msg)); + oldrep->rs_opc); - svc = oldrep->rs_service; - spin_lock (&svc->srv_lock); + svcpt = oldrep->rs_svcpt; + spin_lock(&svcpt->scp_rep_lock); - list_del_init (&oldrep->rs_exp_list); + 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, - lustre_msg_get_opc(oldrep->rs_msg), - 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], @@ -1037,21 +170,21 @@ static void mdt_steal_ack_locks(struct ptlrpc_request *req) oldrep->rs_nlocks = 0; DEBUG_REQ(D_HA, req, "stole locks for"); - spin_lock(&oldrep->rs_lock); - ptlrpc_schedule_difficult_reply (oldrep); - spin_unlock(&oldrep->rs_lock); - - spin_unlock (&svc->srv_lock); - break; - } - 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); } /** * VBR: restore versions */ -void mdt_vbr_reconstruct(struct ptlrpc_request *req, - struct lsd_client_data *lcd) +static void mdt_vbr_reconstruct(struct ptlrpc_request *req, + struct lsd_client_data *lcd) { __u64 pre_versions[4] = {0}; pre_versions[0] = lcd->lcd_pre_versions[0]; @@ -1061,14 +194,12 @@ void mdt_vbr_reconstruct(struct ptlrpc_request *req, lustre_msg_set_versions(req->rq_repmsg, pre_versions); } -void mdt_req_from_lcd(struct ptlrpc_request *req, - struct lsd_client_data *lcd) +void mdt_req_from_lcd(struct ptlrpc_request *req, struct lsd_client_data *lcd) { DEBUG_REQ(D_HA, req, "restoring transno "LPD64"/status %d", lcd->lcd_last_transno, lcd->lcd_last_result); - if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE || - lustre_msg_get_opc(req->rq_repmsg) == MDS_DONE_WRITING) { + if (lustre_msg_get_opc(req->rq_reqmsg) == MDS_CLOSE) { req->rq_transno = lcd->lcd_last_close_transno; req->rq_status = lcd->lcd_last_close_result; } else { @@ -1090,9 +221,29 @@ void mdt_reconstruct_generic(struct mdt_thread_info *mti, struct mdt_lock_handle *lhc) { struct ptlrpc_request *req = mdt_info_req(mti); - struct mdt_export_data *med = &req->rq_export->exp_mdt_data; + struct tg_export_data *ted = &req->rq_export->exp_target_data; - return mdt_req_from_lcd(req, med->med_lcd); + return mdt_req_from_lcd(req, ted->ted_lcd); +} + +/** + * Generate fake attributes for a non-existing object + * + * While the client was waiting for the reply, the original transaction + * got committed and corresponding rep-ack lock got released, then another + * client was able to destroy the object. But we still need to send some + * attributes back. So we fake them and set nlink=0, so the client will + * be able to detect a non-existing object and drop it from the cache + * immediately. + * + * \param[out] ma attributes to fill + */ +static void mdt_fake_ma(struct md_attr *ma) +{ + ma->ma_valid = MA_INODE; + memset(&ma->ma_attr, 0, sizeof(ma->ma_attr)); + ma->ma_attr.la_valid = LA_NLINK; + ma->ma_attr.la_mode = S_IFREG; } static void mdt_reconstruct_create(struct mdt_thread_info *mti, @@ -1100,42 +251,47 @@ static void mdt_reconstruct_create(struct mdt_thread_info *mti, { struct ptlrpc_request *req = mdt_info_req(mti); struct obd_export *exp = req->rq_export; - struct mdt_export_data *med = &exp->exp_mdt_data; + struct tg_export_data *ted = &exp->exp_target_data; struct mdt_device *mdt = mti->mti_mdt; struct mdt_object *child; struct mdt_body *body; int rc; - mdt_req_from_lcd(req, med->med_lcd); + mdt_req_from_lcd(req, ted->ted_lcd); if (req->rq_status) return; - /* if no error, so child was created with requested fid */ - child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2); - if (IS_ERR(child)) { - rc = PTR_ERR(child); - LCONSOLE_WARN("Child "DFID" lookup error %d." - " Evicting client %s with export %s.\n", - PFID(mdt_object_fid(child)), rc, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp)); - mdt_export_evict(exp); - EXIT; - return; - } + /* if no error, so child was created with requested fid */ + child = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid2); + if (IS_ERR(child)) { + rc = PTR_ERR(child); + LCONSOLE_WARN("cannot lookup child "DFID": rc = %d; " + "evicting client %s with export %s\n", + PFID(mti->mti_rr.rr_fid2), rc, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp)); + mdt_export_evict(exp); + RETURN_EXIT; + } 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); - if (rc == -EREMOTE) { - /* object was created on remote server */ - req->rq_status = rc; - body->valid |= OBD_MD_MDS; - } - mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr, - mdt_object_fid(child)); - mdt_object_put(mti->mti_env, child); + rc = mdt_attr_get_complex(mti, child, &mti->mti_attr); + if (rc == -ENOENT) { + mdt_fake_ma(&mti->mti_attr); + } else if (rc == -EREMOTE) { + /* object was created on remote server */ + if (!mdt_is_dne_client(exp)) + /* Return -EIO for old client */ + rc = -EIO; + + req->rq_status = rc; + body->mbo_valid |= OBD_MD_MDS; + } + mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr, + mdt_object_fid(child)); + mdt_object_put(mti->mti_env, child); } static void mdt_reconstruct_setattr(struct mdt_thread_info *mti, @@ -1147,46 +303,35 @@ static void mdt_reconstruct_setattr(struct mdt_thread_info *mti, struct mdt_device *mdt = mti->mti_mdt; struct mdt_object *obj; struct mdt_body *body; + int rc; - mdt_req_from_lcd(req, med->med_lcd); + mdt_req_from_lcd(req, med->med_ted.ted_lcd); if (req->rq_status) return; body = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY); obj = mdt_object_find(mti->mti_env, mdt, mti->mti_rr.rr_fid1); - if (IS_ERR(obj)) { - int rc = PTR_ERR(obj); - LCONSOLE_WARN(""DFID" lookup error %d." - " Evicting client %s with export %s.\n", - PFID(mdt_object_fid(obj)), rc, - obd_uuid2str(&exp->exp_client_uuid), - obd_export_nid2str(exp)); - mdt_export_evict(exp); - EXIT; - return; - } + if (IS_ERR(obj)) { + rc = PTR_ERR(obj); + LCONSOLE_WARN("cannot lookup "DFID": rc = %d; " + "evicting client %s with export %s\n", + PFID(mti->mti_rr.rr_fid1), rc, + obd_uuid2str(&exp->exp_client_uuid), + obd_export_nid2str(exp)); + mdt_export_evict(exp); + RETURN_EXIT; + } + 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); + + rc = mdt_attr_get_complex(mti, obj, &mti->mti_attr); + if (rc == -ENOENT) + mdt_fake_ma(&mti->mti_attr); mdt_pack_attr2body(mti, body, &mti->mti_attr.ma_attr, mdt_object_fid(obj)); - if (mti->mti_epoch && (mti->mti_epoch->flags & MF_EPOCH_OPEN)) { - struct mdt_file_data *mfd; - struct mdt_body *repbody; - - repbody = req_capsule_server_get(mti->mti_pill, &RMF_MDT_BODY); - repbody->ioepoch = obj->mot_ioepoch; - spin_lock(&med->med_open_lock); - 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); + mdt_object_put(mti->mti_env, obj); } typedef void (*mdt_reconstructor)(struct mdt_thread_info *mti,