/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */ // vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4: #ident "$Id$" /*====== This file is part of PerconaFT. Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved. PerconaFT is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License, version 2, as published by the Free Software Foundation. PerconaFT 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. You should have received a copy of the GNU General Public License along with PerconaFT. If not, see . ---------------------------------------- PerconaFT is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License, version 3, as published by the Free Software Foundation. PerconaFT 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 Affero General Public License for more details. You should have received a copy of the GNU Affero General Public License along with PerconaFT. If not, see . ======= */ #ident "Copyright (c) 2006, 2015, Percona and/or its affiliates. All rights reserved." /* rollback and rollforward routines. */ #include #include "ft/ft-ops.h" #include "ft/ft.h" #include "ft/log_header.h" #include "ft/logger/log-internal.h" #include "ft/txn/rollback-apply.h" #include "ft/txn/xids.h" // functionality provided by roll.c is exposed by an autogenerated // header file, logheader.h // // this (poorly) explains the absence of "roll.h" // these flags control whether or not we send commit messages for // various operations // When a transaction is committed, should we send a FT_COMMIT message // for each FT_INSERT message sent earlier by the transaction? #define TOKU_DO_COMMIT_CMD_INSERT 0 // When a transaction is committed, should we send a FT_COMMIT message // for each FT_DELETE_ANY message sent earlier by the transaction? #define TOKU_DO_COMMIT_CMD_DELETE 1 // When a transaction is committed, should we send a FT_COMMIT message // for each FT_UPDATE message sent earlier by the transaction? #define TOKU_DO_COMMIT_CMD_UPDATE 0 int toku_commit_fdelete (FILENUM filenum, TOKUTXN txn, LSN UU(oplsn)) //oplsn is the lsn of the commit { int r; CACHEFILE cf; CACHETABLE ct = txn->logger->ct; // Try to get the cachefile for this filenum. A missing file on recovery // is not an error, but a missing file outside of recovery is. r = toku_cachefile_of_filenum(ct, filenum, &cf); if (r == ENOENT) { assert(txn->for_recovery); r = 0; goto done; } assert_zero(r); // bug fix for #4718 // bug was introduced in with fix for #3590 // Before Maxwell (and fix for #3590), // the recovery log was fsynced after the xcommit was loged but // before we processed rollback entries and before we released // the row locks (in the lock tree). Due to performance concerns, // the fsync was moved to after the release of row locks, which comes // after processing rollback entries. As a result, we may be unlinking a file // here as part of a transactoin that may abort if we do not fsync the log. // So, we fsync the log here. if (txn->logger) { toku_logger_fsync_if_lsn_not_fsynced(txn->logger, txn->do_fsync_lsn); } // Mark the cachefile as unlink on close. There are two ways for close // to be eventually called on the cachefile: // // - when this txn completes, it will release a reference on the // ft and close it, UNLESS it was pinned by checkpoint // - if the cf was pinned by checkpoint, an unpin will release the // final reference and call close. it must be the final reference // since this txn has exclusive access to dictionary (by the // directory row lock for its dname) and we would not get this // far if there were other live handles. toku_cachefile_unlink_on_close(cf); done: return r; } int toku_rollback_fdelete (FILENUM UU(filenum), TOKUTXN UU(txn), LSN UU(oplsn)) //oplsn is the lsn of the abort { //Rolling back an fdelete is an no-op. return 0; } int toku_commit_fcreate (FILENUM UU(filenum), BYTESTRING UU(bs_fname), TOKUTXN UU(txn), LSN UU(oplsn)) { return 0; } int toku_rollback_fcreate (FILENUM filenum, BYTESTRING UU(bs_fname), TOKUTXN txn, LSN UU(oplsn)) { int r; CACHEFILE cf; CACHETABLE ct = txn->logger->ct; // Try to get the cachefile for this filenum. A missing file on recovery // is not an error, but a missing file outside of recovery is. r = toku_cachefile_of_filenum(ct, filenum, &cf); if (r == ENOENT) { r = 0; goto done; } assert_zero(r); // Mark the cachefile as unlink on close. There are two ways for close // to be eventually called on the cachefile: // // - when this txn completes, it will release a reference on the // ft and close it, UNLESS it was pinned by checkpoint // - if the cf was pinned by checkpoint, an unpin will release the // final reference and call close. it must be the final reference // since this txn has exclusive access to dictionary (by the // directory row lock for its dname) and we would not get this // far if there were other live handles. toku_cachefile_unlink_on_close(cf); toku_cachefile_skip_log_recover_on_close(cf); done: return 0; } int toku_commit_frename(BYTESTRING /* old_name */, BYTESTRING /* new_iname */, TOKUTXN /* txn */, LSN UU(oplsn)) { return 0; } int toku_rollback_frename(BYTESTRING old_iname, BYTESTRING new_iname, TOKUTXN txn, LSN UU(oplsn)) { assert(txn); assert(txn->logger); assert(txn->logger->ct); CACHETABLE cachetable = txn->logger->ct; toku_struct_stat stat; bool old_exist = true; bool new_exist = true; std::unique_ptr old_iname_full( toku_cachetable_get_fname_in_cwd(cachetable, old_iname.data), &toku_free); std::unique_ptr new_iname_full( toku_cachetable_get_fname_in_cwd(cachetable, new_iname.data), &toku_free); if (toku_stat(old_iname_full.get(), &stat, toku_uninstrumented) == -1) { if (ENOENT == errno) old_exist = false; else return 1; } if (toku_stat(new_iname_full.get(), &stat, toku_uninstrumented) == -1) { if (ENOENT == errno || ENAMETOOLONG == errno) new_exist = false; else return 1; } // Both old and new files can exist if: // - rename() is not completed // - fcreate was replayed during recovery // 'Stalled cachefiles' container cachefile_list::m_stale_fileid contains // closed but not yet evicted cachefiles and the key of this container is // fs-dependent file id - (device id, inode number) pair. To preserve the // new cachefile // file's id and keep it in 'stalled cachefiles' container the old file is // removed // and the new file is renamed. if (old_exist && new_exist && (toku_os_delete(old_iname_full.get()) == -1 || toku_os_rename(new_iname_full.get(), old_iname_full.get()) == -1 || toku_fsync_directory(new_iname_full.get()) == -1 || toku_fsync_directory(old_iname_full.get()) == -1)) return 1; if (!old_exist && new_exist && (!toku_create_subdirs_if_needed(old_iname_full.get()) || toku_os_rename(new_iname_full.get(), old_iname_full.get()) == -1 || toku_fsync_directory(new_iname_full.get()) == -1 || toku_fsync_directory(old_iname_full.get()) == -1)) return 1; // it's ok if both files do not exist on recovery if (!old_exist && !new_exist) assert(txn->for_recovery); CACHEFILE cf; int r = toku_cachefile_of_iname_in_env(cachetable, new_iname.data, &cf); if (r != ENOENT) { char *old_fname_in_cf = toku_cachefile_fname_in_env(cf); toku_cachefile_set_fname_in_env(cf, toku_xstrdup(old_iname.data)); toku_free(old_fname_in_cf); // There is at least one case when fclose logging cause error: // 1) start transaction // 2) create ft 'a'(write "fcreate" in recovery log) // 3) rename ft 'a' to 'b'(write "frename" in recovery log) // 4) abort transaction: // a) rollback rename ft (renames 'b' to 'a') // b) rollback create ft (removes 'a'): // invokes toku_cachefile_unlink_on_close - lazy unlink on file // close, // it just sets corresponding flag in cachefile object // c) write "unlink" for 'a' in recovery log // (when transaction is aborted all locks are released, // when file lock is released the file is closed and unlinked if // corresponding flag is set in cachefile object) // 5) crash // // After this we have the following records in recovery log: // - create ft 'a', // - rename 'a' to 'b', // - unlink 'a' // // On recovery: // - create 'a' // - rename 'a' to 'b' // - unlink 'a' - as 'a' file does not exist we have crash on assert // here // // There is no need to write "unlink" in recovery log in (4a) because // 'a' will be removed // on transaction rollback on recovery. toku_cachefile_skip_log_recover_on_close(cf); } return 0; } int find_ft_from_filenum (const FT &ft, const FILENUM &filenum); int find_ft_from_filenum (const FT &ft, const FILENUM &filenum) { FILENUM thisfnum = toku_cachefile_filenum(ft->cf); if (thisfnum.fileidfilenum.fileid) return +1; return 0; } // Input arg reset_root_xid_that_created true means that this operation has changed the definition of this dictionary. // (Example use is for schema change committed with txn that inserted cmdupdatebroadcast message.) // The oplsn argument is ZERO_LSN for normal operation. When this function is called for recovery, it has the LSN of // the operation (insert, delete, update, etc). static int do_insertion (enum ft_msg_type type, FILENUM filenum, BYTESTRING key, BYTESTRING *data, TOKUTXN txn, LSN oplsn, bool reset_root_xid_that_created) { int r = 0; //printf("%s:%d committing insert %s %s\n", __FILE__, __LINE__, key.data, data.data); FT ft = nullptr; r = txn->open_fts.find_zero(filenum, &ft, NULL); if (r == DB_NOTFOUND) { assert(txn->for_recovery); r = 0; goto done; } assert(r==0); if (oplsn.lsn != 0) { // if we are executing the recovery algorithm LSN treelsn = toku_ft_checkpoint_lsn(ft); if (oplsn.lsn <= treelsn.lsn) { // if operation was already applied to tree ... r = 0; // ... do not apply it again. goto done; } } DBT key_dbt,data_dbt; XIDS xids; xids = toku_txn_get_xids(txn); { const DBT *kdbt = key.len > 0 ? toku_fill_dbt(&key_dbt, key.data, key.len) : toku_init_dbt(&key_dbt); const DBT *vdbt = data ? toku_fill_dbt(&data_dbt, data->data, data->len) : toku_init_dbt(&data_dbt); ft_msg msg(kdbt, vdbt, type, ZERO_MSN, xids); TXN_MANAGER txn_manager = toku_logger_get_txn_manager(txn->logger); txn_manager_state txn_state_for_gc(txn_manager); TXNID oldest_referenced_xid_estimate = toku_txn_manager_get_oldest_referenced_xid_estimate(txn_manager); txn_gc_info gc_info(&txn_state_for_gc, oldest_referenced_xid_estimate, // no messages above us, we can implicitly promote uxrs based on this xid oldest_referenced_xid_estimate, !txn->for_recovery); toku_ft_root_put_msg(ft, msg, &gc_info); if (reset_root_xid_that_created) { TXNID new_root_xid_that_created = toku_xids_get_outermost_xid(xids); toku_reset_root_xid_that_created(ft, new_root_xid_that_created); } } done: return r; } static int do_nothing_with_filenum(TOKUTXN UU(txn), FILENUM UU(filenum)) { return 0; } int toku_commit_cmdinsert (FILENUM filenum, BYTESTRING UU(key), TOKUTXN txn, LSN UU(oplsn)) { #if TOKU_DO_COMMIT_CMD_INSERT return do_insertion (FT_COMMIT_ANY, filenum, key, 0, txn, oplsn, false); #else return do_nothing_with_filenum(txn, filenum); #endif } int toku_rollback_cmdinsert (FILENUM filenum, BYTESTRING key, TOKUTXN txn, LSN oplsn) { return do_insertion (FT_ABORT_ANY, filenum, key, 0, txn, oplsn, false); } int toku_commit_cmdupdate(FILENUM filenum, BYTESTRING UU(key), TOKUTXN txn, LSN UU(oplsn)) { #if TOKU_DO_COMMIT_CMD_UPDATE return do_insertion(FT_COMMIT_ANY, filenum, key, 0, txn, oplsn, false); #else return do_nothing_with_filenum(txn, filenum); #endif } int toku_rollback_cmdupdate(FILENUM filenum, BYTESTRING key, TOKUTXN txn, LSN oplsn) { return do_insertion(FT_ABORT_ANY, filenum, key, 0, txn, oplsn, false); } int toku_commit_cmdupdatebroadcast(FILENUM filenum, bool is_resetting_op, TOKUTXN txn, LSN oplsn) { // if is_resetting_op, reset root_xid_that_created in // relevant ft. bool reset_root_xid_that_created = (is_resetting_op ? true : false); const enum ft_msg_type msg_type = (is_resetting_op ? FT_COMMIT_BROADCAST_ALL : FT_COMMIT_BROADCAST_TXN); BYTESTRING nullkey = { 0, NULL }; return do_insertion(msg_type, filenum, nullkey, 0, txn, oplsn, reset_root_xid_that_created); } int toku_rollback_cmdupdatebroadcast(FILENUM filenum, bool UU(is_resetting_op), TOKUTXN txn, LSN oplsn) { BYTESTRING nullkey = { 0, NULL }; return do_insertion(FT_ABORT_BROADCAST_TXN, filenum, nullkey, 0, txn, oplsn, false); } int toku_commit_cmddelete (FILENUM filenum, BYTESTRING key, TOKUTXN txn, LSN oplsn) { #if TOKU_DO_COMMIT_CMD_DELETE return do_insertion (FT_COMMIT_ANY, filenum, key, 0, txn, oplsn, false); #else key = key; oplsn = oplsn; return do_nothing_with_filenum(txn, filenum); #endif } int toku_rollback_cmddelete (FILENUM filenum, BYTESTRING key, TOKUTXN txn, LSN oplsn) { return do_insertion (FT_ABORT_ANY, filenum, key, 0, txn, oplsn, false); } static int toku_apply_rollinclude (TXNID_PAIR xid, uint64_t num_nodes, BLOCKNUM spilled_head, BLOCKNUM spilled_tail, TOKUTXN txn, LSN oplsn, apply_rollback_item func) { int r = 0; struct roll_entry *item; BLOCKNUM next_log = spilled_tail; uint64_t last_sequence = num_nodes; bool found_head = false; assert(next_log.b != ROLLBACK_NONE.b); while (next_log.b != ROLLBACK_NONE.b) { //pin log ROLLBACK_LOG_NODE log; toku_get_and_pin_rollback_log(txn, next_log, &log); toku_rollback_verify_contents(log, xid, last_sequence - 1); last_sequence = log->sequence; toku_maybe_prefetch_previous_rollback_log(txn, log); while ((item=log->newest_logentry)) { log->newest_logentry = item->prev; r = func(txn, item, oplsn); if (r!=0) return r; } if (next_log.b == spilled_head.b) { assert(!found_head); found_head = true; assert(log->sequence == 0); } next_log = log->previous; { //Clean up transaction structure to prevent //toku_txn_close from double-freeing spilled_tail = next_log; if (found_head) { assert(next_log.b == ROLLBACK_NONE.b); spilled_head = next_log; } } toku_rollback_log_unpin_and_remove(txn, log); } return r; } int toku_commit_rollinclude (TXNID_PAIR xid, uint64_t num_nodes, BLOCKNUM spilled_head, BLOCKNUM spilled_tail, TOKUTXN txn, LSN oplsn) { int r; r = toku_apply_rollinclude(xid, num_nodes, spilled_head, spilled_tail, txn, oplsn, toku_commit_rollback_item); return r; } int toku_rollback_rollinclude (TXNID_PAIR xid, uint64_t num_nodes, BLOCKNUM spilled_head, BLOCKNUM spilled_tail, TOKUTXN txn, LSN oplsn) { int r; r = toku_apply_rollinclude(xid, num_nodes, spilled_head, spilled_tail, txn, oplsn, toku_abort_rollback_item); return r; } int toku_commit_load (FILENUM old_filenum, BYTESTRING UU(new_iname), TOKUTXN txn, LSN UU(oplsn)) { int r; CACHEFILE old_cf; CACHETABLE ct = txn->logger->ct; // To commit a dictionary load, we delete the old file // // Try to get the cachefile for the old filenum. A missing file on recovery // is not an error, but a missing file outside of recovery is. r = toku_cachefile_of_filenum(ct, old_filenum, &old_cf); if (r == ENOENT) { invariant(txn->for_recovery); r = 0; goto done; } lazy_assert(r == 0); // bug fix for #4718 // bug was introduced in with fix for #3590 // Before Maxwell (and fix for #3590), // the recovery log was fsynced after the xcommit was loged but // before we processed rollback entries and before we released // the row locks (in the lock tree). Due to performance concerns, // the fsync was moved to after the release of row locks, which comes // after processing rollback entries. As a result, we may be unlinking a file // here as part of a transactoin that may abort if we do not fsync the log. // So, we fsync the log here. if (txn->logger) { toku_logger_fsync_if_lsn_not_fsynced(txn->logger, txn->do_fsync_lsn); } // TODO: Zardosht // Explain why this condition is valid, because I forget. if (!toku_cachefile_is_unlink_on_close(old_cf)) { toku_cachefile_unlink_on_close(old_cf); } done: return r; } int toku_rollback_load (FILENUM UU(old_filenum), BYTESTRING new_iname, TOKUTXN txn, LSN UU(oplsn)) { int r; CACHEFILE new_cf; CACHETABLE ct = txn->logger->ct; // To rollback a dictionary load, we delete the new file. // Try to get the cachefile for the new fname. char *fname_in_env = fixup_fname(&new_iname); r = toku_cachefile_of_iname_in_env(ct, fname_in_env, &new_cf); if (r == ENOENT) { // It's possible the new iname was never created, so just try to // unlink it if it's there and ignore the error if it's not. char *fname_in_cwd = toku_cachetable_get_fname_in_cwd(ct, fname_in_env); r = unlink(fname_in_cwd); assert(r == 0 || get_error_errno() == ENOENT); toku_free(fname_in_cwd); r = 0; } else { assert_zero(r); toku_cachefile_unlink_on_close(new_cf); } toku_free(fname_in_env); return r; } //2954 int toku_commit_hot_index (FILENUMS UU(hot_index_filenums), TOKUTXN UU(txn), LSN UU(oplsn)) { // nothing return 0; } int toku_rollback_hot_index (FILENUMS UU(hot_index_filenums), TOKUTXN UU(txn), LSN UU(oplsn)) { return 0; } int toku_commit_dictionary_redirect (FILENUM UU(old_filenum), FILENUM UU(new_filenum), TOKUTXN UU(txn), LSN UU(oplsn)) //oplsn is the lsn of the commit { //Redirect only has meaning during normal operation (NOT during recovery). if (!txn->for_recovery) { //NO-OP } return 0; } int toku_rollback_dictionary_redirect (FILENUM old_filenum, FILENUM new_filenum, TOKUTXN txn, LSN UU(oplsn)) //oplsn is the lsn of the abort { int r = 0; //Redirect only has meaning during normal operation (NOT during recovery). if (!txn->for_recovery) { CACHEFILE new_cf = NULL; r = toku_cachefile_of_filenum(txn->logger->ct, new_filenum, &new_cf); assert(r == 0); FT CAST_FROM_VOIDP(new_ft, toku_cachefile_get_userdata(new_cf)); CACHEFILE old_cf = NULL; r = toku_cachefile_of_filenum(txn->logger->ct, old_filenum, &old_cf); assert(r == 0); FT CAST_FROM_VOIDP(old_ft, toku_cachefile_get_userdata(old_cf)); //Redirect back from new to old. r = toku_dictionary_redirect_abort(old_ft, new_ft, txn); assert(r==0); } return r; } int toku_commit_change_fdescriptor(FILENUM filenum, BYTESTRING UU(old_descriptor), TOKUTXN txn, LSN UU(oplsn)) { return do_nothing_with_filenum(txn, filenum); } int toku_rollback_change_fdescriptor(FILENUM filenum, BYTESTRING old_descriptor, TOKUTXN txn, LSN UU(oplsn)) { CACHEFILE cf; int r; r = toku_cachefile_of_filenum(txn->logger->ct, filenum, &cf); if (r == ENOENT) { //Missing file on recovered transaction is not an error assert(txn->for_recovery); r = 0; goto done; } // file must be open, because the txn that created it opened it and // noted it, assert(r == 0); FT ft; ft = NULL; r = txn->open_fts.find_zero(filenum, &ft, NULL); assert(r == 0); DESCRIPTOR_S d; toku_fill_dbt(&d.dbt, old_descriptor.data, old_descriptor.len); toku_ft_update_descriptor(ft, &d); done: return r; }