/* * Copyright (c) 2002-2014 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * * This file contains Original Code and/or Modifications of Original Code * as defined in and that are subject to the Apple Public Source License * Version 2.0 (the 'License'). You may not use this file except in * compliance with the License. The rights granted to you under the License * may not be used to create, or enable the creation or redistribution of, * unlawful or unlicensed copies of an Apple operating system, or to * circumvent, violate, or enable the circumvention or violation of, any * terms of an Apple operating system software license agreement. * * Please obtain a copy of the License at * http://www.opensource.apple.com/apsl/ and read it before using this file. * * The Original Code and all software distributed under the License are * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. * Please see the License for the specific language governing rights and * limitations under the License. * * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ // // This file implements a simple write-ahead journaling layer. // In theory any file system can make use of it by calling these // functions when the fs wants to modify meta-data blocks. See // vfs_journal.h for a more detailed description of the api and // data structures. // // Dominic Giampaolo (dbg@apple.com) // #ifdef KERNEL #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* OSAddAtomic */ kern_return_t thread_terminate(thread_t); /* * Set sysctl vfs.generic.jnl.kdebug.trim=1 to enable KERNEL_DEBUG_CONSTANT * logging of trim-related calls within the journal. (They're * disabled by default because there can be a lot of these events, * and we don't want to overwhelm the kernel debug buffer. If you * want to watch these events in particular, just set the sysctl.) */ static int jnl_kdebug = 0; SYSCTL_DECL(_vfs_generic); SYSCTL_NODE(_vfs_generic, OID_AUTO, jnl, CTLFLAG_RW|CTLFLAG_LOCKED, 0, "Journal"); SYSCTL_NODE(_vfs_generic_jnl, OID_AUTO, kdebug, CTLFLAG_RW|CTLFLAG_LOCKED, 0, "Journal kdebug"); SYSCTL_INT(_vfs_generic_jnl_kdebug, OID_AUTO, trim, CTLFLAG_RW|CTLFLAG_LOCKED, &jnl_kdebug, 0, "Enable kdebug logging for journal TRIM"); #define DBG_JOURNAL_FLUSH FSDBG_CODE(DBG_JOURNAL, 1) #define DBG_JOURNAL_TRIM_ADD FSDBG_CODE(DBG_JOURNAL, 2) #define DBG_JOURNAL_TRIM_REMOVE FSDBG_CODE(DBG_JOURNAL, 3) #define DBG_JOURNAL_TRIM_REMOVE_PENDING FSDBG_CODE(DBG_JOURNAL, 4) #define DBG_JOURNAL_TRIM_REALLOC FSDBG_CODE(DBG_JOURNAL, 5) #define DBG_JOURNAL_TRIM_FLUSH FSDBG_CODE(DBG_JOURNAL, 6) #define DBG_JOURNAL_TRIM_UNMAP FSDBG_CODE(DBG_JOURNAL, 7) /* * Cap the journal max size to 2GB. On HFS, it will attempt to occupy * a full allocation block if the current size is smaller than the allocation * block on which it resides. Once we hit the exabyte filesystem range, then * it will use 2GB allocation blocks. As a result, make the cap 2GB. */ #define MAX_JOURNAL_SIZE 0x80000000U #include /* DTRACE_IO1 */ #else #include #include #include #include #include #include #include #include #include #include "compat.h" #endif /* KERNEL */ #include "vfs_journal.h" #include #if 0 #undef KERNEL_DEBUG #define KERNEL_DEBUG KERNEL_DEBUG_CONSTANT #endif #ifndef CONFIG_HFS_TRIM #define CONFIG_HFS_TRIM 0 #endif #if JOURNALING // // By default, we grow the list of extents to trim by 4K at a time. // We'll opt to flush a transaction if it contains at least // JOURNAL_FLUSH_TRIM_EXTENTS extents to be trimmed (even if the number // of modified blocks is small). // enum { JOURNAL_DEFAULT_TRIM_BYTES = 4096, JOURNAL_DEFAULT_TRIM_EXTENTS = JOURNAL_DEFAULT_TRIM_BYTES / sizeof(dk_extent_t), JOURNAL_FLUSH_TRIM_EXTENTS = JOURNAL_DEFAULT_TRIM_EXTENTS * 15 / 16 }; unsigned int jnl_trim_flush_limit = JOURNAL_FLUSH_TRIM_EXTENTS; SYSCTL_UINT (_kern, OID_AUTO, jnl_trim_flush, CTLFLAG_RW, &jnl_trim_flush_limit, 0, "number of trimmed extents to cause a journal flush"); /* XXX next prototype should be from libsa/stdlib.h> but conflicts libkern */ __private_extern__ void qsort( void * array, size_t nmembers, size_t member_size, int (*)(const void *, const void *)); // number of bytes to checksum in a block_list_header // NOTE: this should be enough to clear out the header // fields as well as the first entry of binfo[] #define BLHDR_CHECKSUM_SIZE 32 static void lock_condition(journal *jnl, boolean_t *condition, const char *condition_name); static void wait_condition(journal *jnl, boolean_t *condition, const char *condition_name); static void unlock_condition(journal *jnl, boolean_t *condition); static void finish_end_thread(transaction *tr); static void write_header_thread(journal *jnl); static int finish_end_transaction(transaction *tr, errno_t (*callback)(void*), void *callback_arg); static int end_transaction(transaction *tr, int force_it, errno_t (*callback)(void*), void *callback_arg, boolean_t drop_lock, boolean_t must_wait); static void abort_transaction(journal *jnl, transaction *tr); static void dump_journal(journal *jnl); static __inline__ void lock_oldstart(journal *jnl); static __inline__ void unlock_oldstart(journal *jnl); static __inline__ void lock_flush(journal *jnl); static __inline__ void unlock_flush(journal *jnl); // // 3105942 - Coalesce writes to the same block on journal replay // typedef struct bucket { off_t block_num; uint32_t jnl_offset; uint32_t block_size; int32_t cksum; } bucket; #define STARTING_BUCKETS 256 static int add_block(journal *jnl, struct bucket **buf_ptr, off_t block_num, size_t size, size_t offset, int32_t cksum, int *num_buckets_ptr, int *num_full_ptr); static int grow_table(struct bucket **buf_ptr, int num_buckets, int new_size); static int lookup_bucket(struct bucket **buf_ptr, off_t block_num, int num_full); static int do_overlap(journal *jnl, struct bucket **buf_ptr, int blk_index, off_t block_num, size_t size, size_t offset, int32_t cksum, int *num_buckets_ptr, int *num_full_ptr); static int insert_block(journal *jnl, struct bucket **buf_ptr, int blk_index, off_t num, size_t size, size_t offset, int32_t cksum, int *num_buckets_ptr, int *num_full_ptr, int overwriting); #define CHECK_JOURNAL(jnl) \ do { \ if (jnl == NULL) { \ panic("%s:%d: null journal ptr?\n", __FILE__, __LINE__); \ } \ if (jnl->jdev == NULL) { \ panic("%s:%d: jdev is null!\n", __FILE__, __LINE__); \ } \ if (jnl->fsdev == NULL) { \ panic("%s:%d: fsdev is null!\n", __FILE__, __LINE__); \ } \ if (jnl->jhdr->magic != JOURNAL_HEADER_MAGIC) { \ panic("%s:%d: jhdr magic corrupted (0x%x != 0x%x)\n", \ __FILE__, __LINE__, jnl->jhdr->magic, JOURNAL_HEADER_MAGIC); \ } \ if ( jnl->jhdr->start <= 0 \ || jnl->jhdr->start > jnl->jhdr->size) { \ panic("%s:%d: jhdr start looks bad (0x%llx max size 0x%llx)\n", \ __FILE__, __LINE__, jnl->jhdr->start, jnl->jhdr->size); \ } \ if ( jnl->jhdr->end <= 0 \ || jnl->jhdr->end > jnl->jhdr->size) { \ panic("%s:%d: jhdr end looks bad (0x%llx max size 0x%llx)\n", \ __FILE__, __LINE__, jnl->jhdr->end, jnl->jhdr->size); \ } \ } while(0) #define CHECK_TRANSACTION(tr) \ do { \ if (tr == NULL) { \ panic("%s:%d: null transaction ptr?\n", __FILE__, __LINE__); \ } \ if (tr->jnl == NULL) { \ panic("%s:%d: null tr->jnl ptr?\n", __FILE__, __LINE__); \ } \ if (tr->blhdr != (block_list_header *)tr->tbuffer) { \ panic("%s:%d: blhdr (%p) != tbuffer (%p)\n", __FILE__, __LINE__, tr->blhdr, tr->tbuffer); \ } \ if (tr->total_bytes < 0) { \ panic("%s:%d: tr total_bytes looks bad: %d\n", __FILE__, __LINE__, tr->total_bytes); \ } \ if (tr->journal_start < 0) { \ panic("%s:%d: tr journal start looks bad: 0x%llx\n", __FILE__, __LINE__, tr->journal_start); \ } \ if (tr->journal_end < 0) { \ panic("%s:%d: tr journal end looks bad: 0x%llx\n", __FILE__, __LINE__, tr->journal_end); \ } \ if (tr->blhdr && (tr->blhdr->max_blocks <= 0 || tr->blhdr->max_blocks > (tr->jnl->jhdr->size/tr->jnl->jhdr->jhdr_size))) { \ panic("%s:%d: tr blhdr max_blocks looks bad: %d\n", __FILE__, __LINE__, tr->blhdr->max_blocks); \ } \ } while(0) // // this isn't a great checksum routine but it will do for now. // we use it to checksum the journal header and the block list // headers that are at the start of each transaction. // static unsigned int calc_checksum(char *ptr, int len) { int i; unsigned int cksum=0; // this is a lame checksum but for now it'll do for(i = 0; i < len; i++, ptr++) { cksum = (cksum << 8) ^ (cksum + *(unsigned char *)ptr); } return (~cksum); } // // Journal Locking // lck_grp_attr_t * jnl_group_attr; lck_attr_t * jnl_lock_attr; lck_grp_t * jnl_mutex_group; void journal_init(void) { jnl_lock_attr = lck_attr_alloc_init(); jnl_group_attr = lck_grp_attr_alloc_init(); jnl_mutex_group = lck_grp_alloc_init("jnl-mutex", jnl_group_attr); } __inline__ void journal_lock(journal *jnl) { lck_mtx_lock(&jnl->jlock); if (jnl->owner) { panic ("jnl: owner is %p, expected NULL\n", jnl->owner); } jnl->owner = current_thread(); } __inline__ void journal_unlock(journal *jnl) { jnl->owner = NULL; lck_mtx_unlock(&jnl->jlock); } static __inline__ void lock_flush(journal *jnl) { lck_mtx_lock(&jnl->flock); } static __inline__ void unlock_flush(journal *jnl) { lck_mtx_unlock(&jnl->flock); } static __inline__ void lock_oldstart(journal *jnl) { lck_mtx_lock(&jnl->old_start_lock); } static __inline__ void unlock_oldstart(journal *jnl) { lck_mtx_unlock(&jnl->old_start_lock); } #define JNL_WRITE 0x0001 #define JNL_READ 0x0002 #define JNL_HEADER 0x8000 // // This function sets up a fake buf and passes it directly to the // journal device strategy routine (so that it won't get cached in // the block cache. // // It also handles range checking the i/o so that we don't write // outside the journal boundaries and it will wrap the i/o back // to the beginning if necessary (skipping over the journal header) // static size_t do_journal_io(journal *jnl, off_t *offset, void *data, size_t len, int direction) { int err, curlen=len; size_t io_sz = 0; buf_t bp; off_t max_iosize; struct bufattr *bap; if (*offset < 0 || *offset > jnl->jhdr->size) { panic("jnl: do_jnl_io: bad offset 0x%llx (max 0x%llx)\n", *offset, jnl->jhdr->size); } if (direction & JNL_WRITE) max_iosize = jnl->max_write_size; else if (direction & JNL_READ) max_iosize = jnl->max_read_size; else max_iosize = 128 * 1024; again: bp = alloc_io_buf(jnl->jdev, 1); if (*offset + (off_t)curlen > jnl->jhdr->size && *offset != 0 && jnl->jhdr->size != 0) { if (*offset == jnl->jhdr->size) { *offset = jnl->jhdr->jhdr_size; } else { curlen = (off_t)jnl->jhdr->size - *offset; } } if (curlen > max_iosize) { curlen = max_iosize; } if (curlen <= 0) { panic("jnl: do_jnl_io: curlen == %d, offset 0x%llx len %zd\n", curlen, *offset, len); } if (*offset == 0 && (direction & JNL_HEADER) == 0) { panic("jnl: request for i/o to jnl-header without JNL_HEADER flag set! (len %d, data %p)\n", curlen, data); } /* * As alluded to in the block comment at the top of the function, we use a "fake" iobuf * here and issue directly to the disk device that the journal protects since we don't * want this to enter the block cache. As a result, we lose the ability to mark it * as a metadata buf_t for the layers below us that may care. If we were to * simply attach the B_META flag into the b_flags this may confuse things further * since this is an iobuf, not a metadata buffer. * * To address this, we use the extended bufattr struct embedded in the bp. * Explicitly mark the buf here as a metadata buffer in its bufattr flags. */ bap = &bp->b_attr; bap->ba_flags |= BA_META; if (direction & JNL_READ) buf_setflags(bp, B_READ); else { /* * don't have to set any flags */ vnode_startwrite(jnl->jdev); } buf_setsize(bp, curlen); buf_setcount(bp, curlen); buf_setdataptr(bp, (uintptr_t)data); buf_setblkno(bp, (daddr64_t) ((jnl->jdev_offset + *offset) / (off_t)jnl->jhdr->jhdr_size)); buf_setlblkno(bp, (daddr64_t) ((jnl->jdev_offset + *offset) / (off_t)jnl->jhdr->jhdr_size)); if ((direction & JNL_WRITE) && (jnl->flags & JOURNAL_DO_FUA_WRITES)) { buf_markfua(bp); } DTRACE_IO1(journal__start, buf_t, bp); err = VNOP_STRATEGY(bp); if (!err) { err = (int)buf_biowait(bp); } DTRACE_IO1(journal__done, buf_t, bp); free_io_buf(bp); if (err) { printf("jnl: %s: do_jnl_io: strategy err 0x%x\n", jnl->jdev_name, err); return 0; } *offset += curlen; io_sz += curlen; if (io_sz != len) { // handle wrap-around data = (char *)data + curlen; curlen = len - io_sz; if (*offset >= jnl->jhdr->size) { *offset = jnl->jhdr->jhdr_size; } goto again; } return io_sz; } static size_t read_journal_data(journal *jnl, off_t *offset, void *data, size_t len) { return do_journal_io(jnl, offset, data, len, JNL_READ); } static size_t write_journal_data(journal *jnl, off_t *offset, void *data, size_t len) { return do_journal_io(jnl, offset, data, len, JNL_WRITE); } static size_t read_journal_header(journal *jnl, void *data, size_t len) { off_t hdr_offset = 0; return do_journal_io(jnl, &hdr_offset, data, len, JNL_READ|JNL_HEADER); } static int write_journal_header(journal *jnl, int updating_start, uint32_t sequence_num) { static int num_err_prints = 0; int ret=0; off_t jhdr_offset = 0; struct vfs_context context; context.vc_thread = current_thread(); context.vc_ucred = NOCRED; // // Flush the track cache if we're not doing force-unit-access // writes. // if (!updating_start && (jnl->flags & JOURNAL_DO_FUA_WRITES) == 0) { ret = VNOP_IOCTL(jnl->jdev, DKIOCSYNCHRONIZECACHE, NULL, FWRITE, &context); } if (ret != 0) { // // Only print this error if it's a different error than the // previous one, or if it's the first time for this device // or if the total number of printfs is less than 25. We // allow for up to 25 printfs to insure that some make it // into the on-disk syslog. Otherwise if we only printed // one, it's possible it would never make it to the syslog // for the root volume and that makes debugging hard. // if ( ret != jnl->last_flush_err || (jnl->flags & JOURNAL_FLUSHCACHE_ERR) == 0 || num_err_prints++ < 25) { printf("jnl: %s: flushing fs disk buffer returned 0x%x\n", jnl->jdev_name, ret); jnl->flags |= JOURNAL_FLUSHCACHE_ERR; jnl->last_flush_err = ret; } } jnl->jhdr->sequence_num = sequence_num; jnl->jhdr->checksum = 0; jnl->jhdr->checksum = calc_checksum((char *)jnl->jhdr, JOURNAL_HEADER_CKSUM_SIZE); if (do_journal_io(jnl, &jhdr_offset, jnl->header_buf, jnl->jhdr->jhdr_size, JNL_WRITE|JNL_HEADER) != (size_t)jnl->jhdr->jhdr_size) { printf("jnl: %s: write_journal_header: error writing the journal header!\n", jnl->jdev_name); jnl->flags |= JOURNAL_INVALID; return -1; } // If we're not doing force-unit-access writes, then we // have to flush after writing the journal header so that // a future transaction doesn't sneak out to disk before // the header does and thus overwrite data that the old // journal header refers to. Saw this exact case happen // on an IDE bus analyzer with Larry Barras so while it // may seem obscure, it's not. // if (updating_start && (jnl->flags & JOURNAL_DO_FUA_WRITES) == 0) { VNOP_IOCTL(jnl->jdev, DKIOCSYNCHRONIZECACHE, NULL, FWRITE, &context); } return 0; } // // this is a work function used to free up transactions that // completed. they can't be free'd from buffer_flushed_callback // because it is called from deep with the disk driver stack // and thus can't do something that would potentially cause // paging. it gets called by each of the journal api entry // points so stuff shouldn't hang around for too long. // static void free_old_stuff(journal *jnl) { transaction *tr, *next; block_list_header *blhdr=NULL, *next_blhdr=NULL; if (jnl->tr_freeme == NULL) return; lock_oldstart(jnl); tr = jnl->tr_freeme; jnl->tr_freeme = NULL; unlock_oldstart(jnl); for(; tr; tr=next) { for (blhdr = tr->blhdr; blhdr; blhdr = next_blhdr) { next_blhdr = (block_list_header *)((long)blhdr->binfo[0].bnum); blhdr->binfo[0].bnum = 0xdeadc0de; kmem_free(kernel_map, (vm_offset_t)blhdr, tr->tbuffer_size); KERNEL_DEBUG(0xbbbbc01c, jnl, tr, tr->tbuffer_size, 0, 0); } next = tr->next; FREE_ZONE(tr, sizeof(transaction), M_JNL_TR); } } // // This is our callback that lets us know when a buffer has been // flushed to disk. It's called from deep within the driver stack // and thus is quite limited in what it can do. Notably, it can // not initiate any new i/o's or allocate/free memory. // static void buffer_flushed_callback(struct buf *bp, void *arg) { transaction *tr; journal *jnl; transaction *ctr, *prev=NULL, *next; size_t i; int bufsize, amt_flushed, total_bytes; //printf("jnl: buf flush: bp @ 0x%x l/blkno %qd/%qd vp 0x%x tr @ 0x%x\n", // bp, buf_lblkno(bp), buf_blkno(bp), buf_vnode(bp), arg); // snarf out the bits we want bufsize = buf_size(bp); tr = (transaction *)arg; // then we've already seen it if (tr == NULL) { return; } CHECK_TRANSACTION(tr); jnl = tr->jnl; CHECK_JOURNAL(jnl); amt_flushed = tr->num_killed; total_bytes = tr->total_bytes; // update the number of blocks that have been flushed. // this buf may represent more than one block so take // that into account. // // OSAddAtomic() returns the value of tr->num_flushed before the add // amt_flushed += OSAddAtomic(bufsize, &tr->num_flushed); // if this transaction isn't done yet, just return as // there is nothing to do. // // NOTE: we are careful to not reference anything through // the tr pointer after doing the OSAddAtomic(). if // this if statement fails then we are the last one // and then it's ok to dereference "tr". // if ((amt_flushed + bufsize) < total_bytes) { return; } // this will single thread checking the transaction lock_oldstart(jnl); if (tr->total_bytes == (int)0xfbadc0de) { // then someone beat us to it... unlock_oldstart(jnl); return; } // mark this so that we're the owner of dealing with the // cleanup for this transaction tr->total_bytes = 0xfbadc0de; if (jnl->flags & JOURNAL_INVALID) goto transaction_done; //printf("jnl: tr 0x%x (0x%llx 0x%llx) in jnl 0x%x completed.\n", // tr, tr->journal_start, tr->journal_end, jnl); // find this entry in the old_start[] index and mark it completed for(i = 0; i < sizeof(jnl->old_start)/sizeof(jnl->old_start[0]); i++) { if ((off_t)(jnl->old_start[i] & ~(0x8000000000000000ULL)) == tr->journal_start) { jnl->old_start[i] &= ~(0x8000000000000000ULL); break; } } if (i >= sizeof(jnl->old_start)/sizeof(jnl->old_start[0])) { panic("jnl: buffer_flushed: did not find tr w/start @ %lld (tr %p, jnl %p)\n", tr->journal_start, tr, jnl); } // if we are here then we need to update the journal header // to reflect that this transaction is complete if (tr->journal_start == jnl->active_start) { jnl->active_start = tr->journal_end; tr->journal_start = tr->journal_end = (off_t)0; } // go through the completed_trs list and try to coalesce // entries, restarting back at the beginning if we have to. for (ctr = jnl->completed_trs; ctr; prev=ctr, ctr=next) { if (ctr->journal_start == jnl->active_start) { jnl->active_start = ctr->journal_end; if (prev) { prev->next = ctr->next; } if (ctr == jnl->completed_trs) { jnl->completed_trs = ctr->next; } next = jnl->completed_trs; // this starts us over again ctr->next = jnl->tr_freeme; jnl->tr_freeme = ctr; ctr = NULL; } else if (tr->journal_end == ctr->journal_start) { ctr->journal_start = tr->journal_start; next = jnl->completed_trs; // this starts us over again ctr = NULL; tr->journal_start = tr->journal_end = (off_t)0; } else if (tr->journal_start == ctr->journal_end) { ctr->journal_end = tr->journal_end; next = ctr->next; tr->journal_start = tr->journal_end = (off_t)0; } else if (ctr->next && ctr->journal_end == ctr->next->journal_start) { // coalesce the next entry with this one and link the next // entry in at the head of the tr_freeme list next = ctr->next; // temporarily use the "next" variable ctr->journal_end = next->journal_end; ctr->next = next->next; next->next = jnl->tr_freeme; // link in the next guy at the head of the tr_freeme list jnl->tr_freeme = next; next = jnl->completed_trs; // this starts us over again ctr = NULL; } else { next = ctr->next; } } // if this is true then we didn't merge with anyone // so link ourselves in at the head of the completed // transaction list. if (tr->journal_start != 0) { // put this entry into the correct sorted place // in the list instead of just at the head. // prev = NULL; for (ctr = jnl->completed_trs; ctr && tr->journal_start > ctr->journal_start; prev=ctr, ctr=ctr->next) { // just keep looping } if (ctr == NULL && prev == NULL) { jnl->completed_trs = tr; tr->next = NULL; } else if (ctr == jnl->completed_trs) { tr->next = jnl->completed_trs; jnl->completed_trs = tr; } else { tr->next = prev->next; prev->next = tr; } } else { // if we're here this tr got merged with someone else so // put it on the list to be free'd tr->next = jnl->tr_freeme; jnl->tr_freeme = tr; } transaction_done: unlock_oldstart(jnl); unlock_condition(jnl, &jnl->asyncIO); } #include #define SWAP16(x) OSSwapInt16(x) #define SWAP32(x) OSSwapInt32(x) #define SWAP64(x) OSSwapInt64(x) static void swap_journal_header(journal *jnl) { jnl->jhdr->magic = SWAP32(jnl->jhdr->magic); jnl->jhdr->endian = SWAP32(jnl->jhdr->endian); jnl->jhdr->start = SWAP64(jnl->jhdr->start); jnl->jhdr->end = SWAP64(jnl->jhdr->end); jnl->jhdr->size = SWAP64(jnl->jhdr->size); jnl->jhdr->blhdr_size = SWAP32(jnl->jhdr->blhdr_size); jnl->jhdr->checksum = SWAP32(jnl->jhdr->checksum); jnl->jhdr->jhdr_size = SWAP32(jnl->jhdr->jhdr_size); jnl->jhdr->sequence_num = SWAP32(jnl->jhdr->sequence_num); } static void swap_block_list_header(journal *jnl, block_list_header *blhdr) { int i; blhdr->max_blocks = SWAP16(blhdr->max_blocks); blhdr->num_blocks = SWAP16(blhdr->num_blocks); blhdr->bytes_used = SWAP32(blhdr->bytes_used); blhdr->checksum = SWAP32(blhdr->checksum); blhdr->flags = SWAP32(blhdr->flags); if (blhdr->num_blocks >= ((jnl->jhdr->blhdr_size / sizeof(block_info)) - 1)) { printf("jnl: %s: blhdr num blocks looks suspicious (%d / blhdr size %d). not swapping.\n", jnl->jdev_name, blhdr->num_blocks, jnl->jhdr->blhdr_size); return; } for(i = 0; i < blhdr->num_blocks; i++) { blhdr->binfo[i].bnum = SWAP64(blhdr->binfo[i].bnum); blhdr->binfo[i].u.bi.bsize = SWAP32(blhdr->binfo[i].u.bi.bsize); blhdr->binfo[i].u.bi.b.cksum = SWAP32(blhdr->binfo[i].u.bi.b.cksum); } } static int update_fs_block(journal *jnl, void *block_ptr, off_t fs_block, size_t bsize) { int ret; struct buf *oblock_bp=NULL; // first read the block we want. ret = buf_meta_bread(jnl->fsdev, (daddr64_t)fs_block, bsize, NOCRED, &oblock_bp); if (ret != 0) { printf("jnl: %s: update_fs_block: error reading fs block # %lld! (ret %d)\n", jnl->jdev_name, fs_block, ret); if (oblock_bp) { buf_brelse(oblock_bp); oblock_bp = NULL; } // let's try to be aggressive here and just re-write the block oblock_bp = buf_getblk(jnl->fsdev, (daddr64_t)fs_block, bsize, 0, 0, BLK_META); if (oblock_bp == NULL) { printf("jnl: %s: update_fs_block: buf_getblk() for %lld failed! failing update.\n", jnl->jdev_name, fs_block); return -1; } } // make sure it's the correct size. if (buf_size(oblock_bp) != bsize) { buf_brelse(oblock_bp); return -1; } // copy the journal data over top of it memcpy((char *)buf_dataptr(oblock_bp), block_ptr, bsize); if ((ret = VNOP_BWRITE(oblock_bp)) != 0) { printf("jnl: %s: update_fs_block: failed to update block %lld (ret %d)\n", jnl->jdev_name, fs_block,ret); return ret; } // and now invalidate it so that if someone else wants to read // it in a different size they'll be able to do it. ret = buf_meta_bread(jnl->fsdev, (daddr64_t)fs_block, bsize, NOCRED, &oblock_bp); if (oblock_bp) { buf_markinvalid(oblock_bp); buf_brelse(oblock_bp); } return 0; } static int grow_table(struct bucket **buf_ptr, int num_buckets, int new_size) { struct bucket *newBuf; int current_size = num_buckets, i; // return if newsize is less than the current size if (new_size < num_buckets) { return current_size; } if ((MALLOC(newBuf, struct bucket *, new_size*sizeof(struct bucket), M_TEMP, M_WAITOK)) == NULL) { printf("jnl: grow_table: no memory to expand coalesce buffer!\n"); return -1; } // printf("jnl: lookup_bucket: expanded co_buf to %d elems\n", new_size); // copy existing elements bcopy(*buf_ptr, newBuf, num_buckets*sizeof(struct bucket)); // initialize the new ones for(i = num_buckets; i < new_size; i++) { newBuf[i].block_num = (off_t)-1; } // free the old container FREE(*buf_ptr, M_TEMP); // reset the buf_ptr *buf_ptr = newBuf; return new_size; } static int lookup_bucket(struct bucket **buf_ptr, off_t block_num, int num_full) { int lo, hi, index, matches, i; if (num_full == 0) { return 0; // table is empty, so insert at index=0 } lo = 0; hi = num_full - 1; index = -1; // perform binary search for block_num do { int mid = (hi - lo)/2 + lo; off_t this_num = (*buf_ptr)[mid].block_num; if (block_num == this_num) { index = mid; break; } if (block_num < this_num) { hi = mid; continue; } if (block_num > this_num) { lo = mid + 1; continue; } } while (lo < hi); // check if lo and hi converged on the match if (block_num == (*buf_ptr)[hi].block_num) { index = hi; } // if no existing entry found, find index for new one if (index == -1) { index = (block_num < (*buf_ptr)[hi].block_num) ? hi : hi + 1; } else { // make sure that we return the right-most index in the case of multiple matches matches = 0; i = index + 1; while (i < num_full && block_num == (*buf_ptr)[i].block_num) { matches++; i++; } index += matches; } return index; } static int insert_block(journal *jnl, struct bucket **buf_ptr, int blk_index, off_t num, size_t size, size_t offset, int32_t cksum, int *num_buckets_ptr, int *num_full_ptr, int overwriting) { if (!overwriting) { // grow the table if we're out of space if (*num_full_ptr >= *num_buckets_ptr) { int new_size = *num_buckets_ptr * 2; int grow_size = grow_table(buf_ptr, *num_buckets_ptr, new_size); if (grow_size < new_size) { printf("jnl: %s: add_block: grow_table returned an error!\n", jnl->jdev_name); return -1; } *num_buckets_ptr = grow_size; //update num_buckets to reflect the new size } // if we're not inserting at the end, we need to bcopy if (blk_index != *num_full_ptr) { bcopy( (*buf_ptr)+(blk_index), (*buf_ptr)+(blk_index+1), (*num_full_ptr-blk_index)*sizeof(struct bucket) ); } (*num_full_ptr)++; // increment only if we're not overwriting } // sanity check the values we're about to add if ((off_t)offset >= jnl->jhdr->size) { offset = jnl->jhdr->jhdr_size + (offset - jnl->jhdr->size); } if (size <= 0) { panic("jnl: insert_block: bad size in insert_block (%zd)\n", size); } (*buf_ptr)[blk_index].block_num = num; (*buf_ptr)[blk_index].block_size = size; (*buf_ptr)[blk_index].jnl_offset = offset; (*buf_ptr)[blk_index].cksum = cksum; return blk_index; } static int do_overlap(journal *jnl, struct bucket **buf_ptr, int blk_index, off_t block_num, size_t size, __unused size_t offset, int32_t cksum, int *num_buckets_ptr, int *num_full_ptr) { int num_to_remove, index, i, overwrite, err; size_t jhdr_size = jnl->jhdr->jhdr_size, new_offset; off_t overlap, block_start, block_end; block_start = block_num*jhdr_size; block_end = block_start + size; overwrite = (block_num == (*buf_ptr)[blk_index].block_num && size >= (*buf_ptr)[blk_index].block_size); // first, eliminate any overlap with the previous entry if (blk_index != 0 && !overwrite) { off_t prev_block_start = (*buf_ptr)[blk_index-1].block_num*jhdr_size; off_t prev_block_end = prev_block_start + (*buf_ptr)[blk_index-1].block_size; overlap = prev_block_end - block_start; if (overlap > 0) { if (overlap % jhdr_size != 0) { panic("jnl: do_overlap: overlap with previous entry not a multiple of %zd\n", jhdr_size); } // if the previous entry completely overlaps this one, we need to break it into two pieces. if (prev_block_end > block_end) { off_t new_num = block_end / jhdr_size; size_t new_size = prev_block_end - block_end; new_offset = (*buf_ptr)[blk_index-1].jnl_offset + (block_end - prev_block_start); err = insert_block(jnl, buf_ptr, blk_index, new_num, new_size, new_offset, cksum, num_buckets_ptr, num_full_ptr, 0); if (err < 0) { panic("jnl: do_overlap: error inserting during pre-overlap\n"); } } // Regardless, we need to truncate the previous entry to the beginning of the overlap (*buf_ptr)[blk_index-1].block_size = block_start - prev_block_start; (*buf_ptr)[blk_index-1].cksum = 0; // have to blow it away because there's no way to check it } } // then, bail out fast if there's no overlap with the entries that follow if (!overwrite && block_end <= (off_t)((*buf_ptr)[blk_index].block_num*jhdr_size)) { return 0; // no overlap, no overwrite } else if (overwrite && (blk_index + 1 >= *num_full_ptr || block_end <= (off_t)((*buf_ptr)[blk_index+1].block_num*jhdr_size))) { (*buf_ptr)[blk_index].cksum = cksum; // update this return 1; // simple overwrite } // Otherwise, find all cases of total and partial overlap. We use the special // block_num of -2 to designate entries that are completely overlapped and must // be eliminated. The block_num, size, and jnl_offset of partially overlapped // entries must be adjusted to keep the array consistent. index = blk_index; num_to_remove = 0; while (index < *num_full_ptr && block_end > (off_t)((*buf_ptr)[index].block_num*jhdr_size)) { if (block_end >= (off_t)(((*buf_ptr)[index].block_num*jhdr_size + (*buf_ptr)[index].block_size))) { (*buf_ptr)[index].block_num = -2; // mark this for deletion num_to_remove++; } else { overlap = block_end - (*buf_ptr)[index].block_num*jhdr_size; if (overlap > 0) { if (overlap % jhdr_size != 0) { panic("jnl: do_overlap: overlap of %lld is not multiple of %zd\n", overlap, jhdr_size); } // if we partially overlap this entry, adjust its block number, jnl offset, and size (*buf_ptr)[index].block_num += (overlap / jhdr_size); // make sure overlap is multiple of jhdr_size, or round up (*buf_ptr)[index].cksum = 0; new_offset = (*buf_ptr)[index].jnl_offset + overlap; // check for wrap-around if ((off_t)new_offset >= jnl->jhdr->size) { new_offset = jhdr_size + (new_offset - jnl->jhdr->size); } (*buf_ptr)[index].jnl_offset = new_offset; (*buf_ptr)[index].block_size -= overlap; // sanity check for negative value if ((*buf_ptr)[index].block_size <= 0) { panic("jnl: do_overlap: after overlap, new block size is invalid (%u)\n", (*buf_ptr)[index].block_size); // return -1; // if above panic is removed, return -1 for error } } } index++; } // bcopy over any completely overlapped entries, starting at the right (where the above loop broke out) index--; // start with the last index used within the above loop while (index >= blk_index) { if ((*buf_ptr)[index].block_num == -2) { if (index == *num_full_ptr-1) { (*buf_ptr)[index].block_num = -1; // it's the last item in the table... just mark as free } else { bcopy( (*buf_ptr)+(index+1), (*buf_ptr)+(index), (*num_full_ptr - (index + 1)) * sizeof(struct bucket) ); } (*num_full_ptr)--; } index--; } // eliminate any stale entries at the end of the table for(i = *num_full_ptr; i < (*num_full_ptr + num_to_remove); i++) { (*buf_ptr)[i].block_num = -1; } return 0; // if we got this far, we need to insert the entry into the table (rather than overwrite) } // PR-3105942: Coalesce writes to the same block in journal replay // We coalesce writes by maintaining a dynamic sorted array of physical disk blocks // to be replayed and the corresponding location in the journal which contains // the most recent data for those blocks. The array is "played" once the all the // blocks in the journal have been coalesced. The code for the case of conflicting/ // overlapping writes to a single block is the most dense. Because coalescing can // disrupt the existing time-ordering of blocks in the journal playback, care // is taken to catch any overlaps and keep the array consistent. static int add_block(journal *jnl, struct bucket **buf_ptr, off_t block_num, size_t size, __unused size_t offset, int32_t cksum, int *num_buckets_ptr, int *num_full_ptr) { int blk_index, overwriting; // on return from lookup_bucket(), blk_index is the index into the table where block_num should be // inserted (or the index of the elem to overwrite). blk_index = lookup_bucket( buf_ptr, block_num, *num_full_ptr); // check if the index is within bounds (if we're adding this block to the end of // the table, blk_index will be equal to num_full) if (blk_index < 0 || blk_index > *num_full_ptr) { //printf("jnl: add_block: trouble adding block to co_buf\n"); return -1; } // else printf("jnl: add_block: adding block 0x%llx at i=%d\n", block_num, blk_index); // Determine whether we're overwriting an existing entry by checking for overlap overwriting = do_overlap(jnl, buf_ptr, blk_index, block_num, size, offset, cksum, num_buckets_ptr, num_full_ptr); if (overwriting < 0) { return -1; // if we got an error, pass it along } // returns the index, or -1 on error blk_index = insert_block(jnl, buf_ptr, blk_index, block_num, size, offset, cksum, num_buckets_ptr, num_full_ptr, overwriting); return blk_index; } static int replay_journal(journal *jnl) { int i, bad_blocks=0; unsigned int orig_checksum, checksum, check_block_checksums = 0; size_t ret; size_t max_bsize = 0; /* protected by block_ptr */ block_list_header *blhdr; off_t offset, txn_start_offset=0, blhdr_offset, orig_jnl_start; char *buff, *block_ptr=NULL; struct bucket *co_buf; int num_buckets = STARTING_BUCKETS, num_full, check_past_jnl_end = 1, in_uncharted_territory=0; uint32_t last_sequence_num = 0; int replay_retry_count = 0; // wrap the start ptr if it points to the very end of the journal if (jnl->jhdr->start == jnl->jhdr->size) { jnl->jhdr->start = jnl->jhdr->jhdr_size; } if (jnl->jhdr->end == jnl->jhdr->size) { jnl->jhdr->end = jnl->jhdr->jhdr_size; } if (jnl->jhdr->start == jnl->jhdr->end) { return 0; } orig_jnl_start = jnl->jhdr->start; // allocate memory for the header_block. we'll read each blhdr into this if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&buff, jnl->jhdr->blhdr_size)) { printf("jnl: %s: replay_journal: no memory for block buffer! (%d bytes)\n", jnl->jdev_name, jnl->jhdr->blhdr_size); return -1; } // allocate memory for the coalesce buffer if ((MALLOC(co_buf, struct bucket *, num_buckets*sizeof(struct bucket), M_TEMP, M_WAITOK)) == NULL) { printf("jnl: %s: replay_journal: no memory for coalesce buffer!\n", jnl->jdev_name); return -1; } restart_replay: // initialize entries for(i = 0; i < num_buckets; i++) { co_buf[i].block_num = -1; } num_full = 0; // empty at first printf("jnl: %s: replay_journal: from: %lld to: %lld (joffset 0x%llx)\n", jnl->jdev_name, jnl->jhdr->start, jnl->jhdr->end, jnl->jdev_offset); while (check_past_jnl_end || jnl->jhdr->start != jnl->jhdr->end) { offset = blhdr_offset = jnl->jhdr->start; ret = read_journal_data(jnl, &offset, buff, jnl->jhdr->blhdr_size); if (ret != (size_t)jnl->jhdr->blhdr_size) { printf("jnl: %s: replay_journal: Could not read block list header block @ 0x%llx!\n", jnl->jdev_name, offset); bad_blocks = 1; goto bad_txn_handling; } blhdr = (block_list_header *)buff; orig_checksum = blhdr->checksum; blhdr->checksum = 0; if (jnl->flags & JOURNAL_NEED_SWAP) { // calculate the checksum based on the unswapped data // because it is done byte-at-a-time. orig_checksum = (unsigned int)SWAP32(orig_checksum); checksum = calc_checksum((char *)blhdr, BLHDR_CHECKSUM_SIZE); swap_block_list_header(jnl, blhdr); } else { checksum = calc_checksum((char *)blhdr, BLHDR_CHECKSUM_SIZE); } // // XXXdbg - if these checks fail, we should replay as much // we can in the hopes that it will still leave the // drive in a better state than if we didn't replay // anything // if (checksum != orig_checksum) { if (check_past_jnl_end && in_uncharted_territory) { if (blhdr_offset != jnl->jhdr->end) { printf("jnl: %s: Extra txn replay stopped @ %lld / 0x%llx\n", jnl->jdev_name, blhdr_offset, blhdr_offset); } check_past_jnl_end = 0; jnl->jhdr->end = blhdr_offset; continue; } printf("jnl: %s: replay_journal: bad block list header @ 0x%llx (checksum 0x%x != 0x%x)\n", jnl->jdev_name, blhdr_offset, orig_checksum, checksum); if (blhdr_offset == orig_jnl_start) { // if there's nothing in the journal at all, just bail out altogether. goto bad_replay; } bad_blocks = 1; goto bad_txn_handling; } if ( (last_sequence_num != 0) && (blhdr->binfo[0].u.bi.b.sequence_num != 0) && (blhdr->binfo[0].u.bi.b.sequence_num != last_sequence_num) && (blhdr->binfo[0].u.bi.b.sequence_num != last_sequence_num+1)) { txn_start_offset = jnl->jhdr->end = blhdr_offset; if (check_past_jnl_end) { check_past_jnl_end = 0; printf("jnl: %s: 2: extra replay stopped @ %lld / 0x%llx (seq %d < %d)\n", jnl->jdev_name, blhdr_offset, blhdr_offset, blhdr->binfo[0].u.bi.b.sequence_num, last_sequence_num); continue; } printf("jnl: %s: txn sequence numbers out of order in txn @ %lld / %llx! (%d < %d)\n", jnl->jdev_name, blhdr_offset, blhdr_offset, blhdr->binfo[0].u.bi.b.sequence_num, last_sequence_num); bad_blocks = 1; goto bad_txn_handling; } last_sequence_num = blhdr->binfo[0].u.bi.b.sequence_num; if (blhdr_offset >= jnl->jhdr->end && jnl->jhdr->start <= jnl->jhdr->end) { if (last_sequence_num == 0) { check_past_jnl_end = 0; printf("jnl: %s: pre-sequence-num-enabled txn's - can not go further than end (%lld %lld).\n", jnl->jdev_name, jnl->jhdr->start, jnl->jhdr->end); if (jnl->jhdr->start != jnl->jhdr->end) { jnl->jhdr->start = jnl->jhdr->end; } continue; } printf("jnl: %s: examining extra transactions starting @ %lld / 0x%llx\n", jnl->jdev_name, blhdr_offset, blhdr_offset); } if ( blhdr->max_blocks <= 0 || blhdr->max_blocks > (jnl->jhdr->size/jnl->jhdr->jhdr_size) || blhdr->num_blocks <= 0 || blhdr->num_blocks > blhdr->max_blocks) { printf("jnl: %s: replay_journal: bad looking journal entry: max: %d num: %d\n", jnl->jdev_name, blhdr->max_blocks, blhdr->num_blocks); bad_blocks = 1; goto bad_txn_handling; } max_bsize = 0; for (i = 1; i < blhdr->num_blocks; i++) { if (blhdr->binfo[i].bnum < 0 && blhdr->binfo[i].bnum != (off_t)-1) { printf("jnl: %s: replay_journal: bogus block number 0x%llx\n", jnl->jdev_name, blhdr->binfo[i].bnum); bad_blocks = 1; goto bad_txn_handling; } if ((size_t)blhdr->binfo[i].u.bi.bsize > max_bsize) { max_bsize = blhdr->binfo[i].u.bi.bsize; } } if (blhdr->flags & BLHDR_CHECK_CHECKSUMS) { check_block_checksums = 1; if (kmem_alloc(kernel_map, (vm_offset_t *)&block_ptr, max_bsize)) { goto bad_replay; } } else { block_ptr = NULL; } if (blhdr->flags & BLHDR_FIRST_HEADER) { txn_start_offset = blhdr_offset; } //printf("jnl: replay_journal: adding %d blocks in journal entry @ 0x%llx to co_buf\n", // blhdr->num_blocks-1, jnl->jhdr->start); bad_blocks = 0; for (i = 1; i < blhdr->num_blocks; i++) { int size, ret_val; off_t number; size = blhdr->binfo[i].u.bi.bsize; number = blhdr->binfo[i].bnum; // don't add "killed" blocks if (number == (off_t)-1) { //printf("jnl: replay_journal: skipping killed fs block (index %d)\n", i); } else { if (check_block_checksums) { int32_t disk_cksum; off_t block_offset; block_offset = offset; // read the block so we can check the checksum ret = read_journal_data(jnl, &block_offset, block_ptr, size); if (ret != (size_t)size) { printf("jnl: %s: replay_journal: Could not read journal entry data @ offset 0x%llx!\n", jnl->jdev_name, offset); bad_blocks = 1; goto bad_txn_handling; } disk_cksum = calc_checksum(block_ptr, size); // there is no need to swap the checksum from disk because // it got swapped when the blhdr was read in. if (blhdr->binfo[i].u.bi.b.cksum != 0 && disk_cksum != blhdr->binfo[i].u.bi.b.cksum) { printf("jnl: %s: txn starting at %lld (%lld) @ index %3d bnum %lld (%d) with disk cksum != blhdr cksum (0x%.8x 0x%.8x)\n", jnl->jdev_name, txn_start_offset, blhdr_offset, i, number, size, disk_cksum, blhdr->binfo[i].u.bi.b.cksum); printf("jnl: 0x%.8x 0x%.8x 0x%.8x 0x%.8x 0x%.8x 0x%.8x 0x%.8x 0x%.8x\n", *(int *)&block_ptr[0*sizeof(int)], *(int *)&block_ptr[1*sizeof(int)], *(int *)&block_ptr[2*sizeof(int)], *(int *)&block_ptr[3*sizeof(int)], *(int *)&block_ptr[4*sizeof(int)], *(int *)&block_ptr[5*sizeof(int)], *(int *)&block_ptr[6*sizeof(int)], *(int *)&block_ptr[7*sizeof(int)]); bad_blocks = 1; goto bad_txn_handling; } } // add this bucket to co_buf, coalescing where possible // printf("jnl: replay_journal: adding block 0x%llx\n", number); ret_val = add_block(jnl, &co_buf, number, size, (size_t) offset, blhdr->binfo[i].u.bi.b.cksum, &num_buckets, &num_full); if (ret_val == -1) { printf("jnl: %s: replay_journal: trouble adding block to co_buf\n", jnl->jdev_name); goto bad_replay; } // else printf("jnl: replay_journal: added block 0x%llx at i=%d\n", number); } // increment offset offset += size; // check if the last block added puts us off the end of the jnl. // if so, we need to wrap to the beginning and take any remainder // into account // if (offset >= jnl->jhdr->size) { offset = jnl->jhdr->jhdr_size + (offset - jnl->jhdr->size); } } if (block_ptr) { kmem_free(kernel_map, (vm_offset_t)block_ptr, max_bsize); block_ptr = NULL; } bad_txn_handling: if (bad_blocks) { /* Journal replay got error before it found any valid * transations, abort replay */ if (txn_start_offset == 0) { printf("jnl: %s: no known good txn start offset! aborting journal replay.\n", jnl->jdev_name); goto bad_replay; } /* Repeated error during journal replay, abort replay */ if (replay_retry_count == 3) { printf("jnl: %s: repeated errors replaying journal! aborting journal replay.\n", jnl->jdev_name); goto bad_replay; } replay_retry_count++; /* There was an error replaying the journal (possibly * EIO/ENXIO from the device). So retry replaying all * the good transactions that we found before getting * the error. */ jnl->jhdr->start = orig_jnl_start; jnl->jhdr->end = txn_start_offset; check_past_jnl_end = 0; last_sequence_num = 0; printf("jnl: %s: restarting journal replay (%lld - %lld)!\n", jnl->jdev_name, jnl->jhdr->start, jnl->jhdr->end); goto restart_replay; } jnl->jhdr->start += blhdr->bytes_used; if (jnl->jhdr->start >= jnl->jhdr->size) { // wrap around and skip the journal header block jnl->jhdr->start = (jnl->jhdr->start % jnl->jhdr->size) + jnl->jhdr->jhdr_size; } if (jnl->jhdr->start == jnl->jhdr->end) { in_uncharted_territory = 1; } } if (jnl->jhdr->start != jnl->jhdr->end) { printf("jnl: %s: start %lld != end %lld. resetting end.\n", jnl->jdev_name, jnl->jhdr->start, jnl->jhdr->end); jnl->jhdr->end = jnl->jhdr->start; } //printf("jnl: replay_journal: replaying %d blocks\n", num_full); /* * make sure it's at least one page in size, so * start max_bsize at PAGE_SIZE */ for (i = 0, max_bsize = PAGE_SIZE; i < num_full; i++) { if (co_buf[i].block_num == (off_t)-1) continue; if (co_buf[i].block_size > max_bsize) max_bsize = co_buf[i].block_size; } /* * round max_bsize up to the nearest PAGE_SIZE multiple */ if (max_bsize & (PAGE_SIZE - 1)) { max_bsize = (max_bsize + PAGE_SIZE) & ~(PAGE_SIZE - 1); } if (kmem_alloc(kernel_map, (vm_offset_t *)&block_ptr, max_bsize)) { goto bad_replay; } // Replay the coalesced entries in the co-buf for(i = 0; i < num_full; i++) { size_t size = co_buf[i].block_size; off_t jnl_offset = (off_t) co_buf[i].jnl_offset; off_t number = co_buf[i].block_num; // printf("replaying co_buf[%d]: block 0x%llx, size 0x%x, jnl_offset 0x%llx\n", i, co_buf[i].block_num, // co_buf[i].block_size, co_buf[i].jnl_offset); if (number == (off_t)-1) { // printf("jnl: replay_journal: skipping killed fs block\n"); } else { // do journal read, and set the phys. block ret = read_journal_data(jnl, &jnl_offset, block_ptr, size); if (ret != size) { printf("jnl: %s: replay_journal: Could not read journal entry data @ offset 0x%llx!\n", jnl->jdev_name, offset); goto bad_replay; } if (update_fs_block(jnl, block_ptr, number, size) != 0) { goto bad_replay; } } } // done replaying; update jnl header if (write_journal_header(jnl, 1, jnl->jhdr->sequence_num) != 0) { goto bad_replay; } printf("jnl: %s: journal replay done.\n", jnl->jdev_name); // free block_ptr if (block_ptr) { kmem_free(kernel_map, (vm_offset_t)block_ptr, max_bsize); block_ptr = NULL; } // free the coalesce buffer FREE(co_buf, M_TEMP); co_buf = NULL; kmem_free(kernel_map, (vm_offset_t)buff, jnl->jhdr->blhdr_size); return 0; bad_replay: if (block_ptr) { kmem_free(kernel_map, (vm_offset_t)block_ptr, max_bsize); } if (co_buf) { FREE(co_buf, M_TEMP); } kmem_free(kernel_map, (vm_offset_t)buff, jnl->jhdr->blhdr_size); return -1; } #define DEFAULT_TRANSACTION_BUFFER_SIZE (128*1024) #define MAX_TRANSACTION_BUFFER_SIZE (3072*1024) // XXXdbg - so I can change it in the debugger int def_tbuffer_size = 0; // // This function sets the size of the tbuffer and the // size of the blhdr. It assumes that jnl->jhdr->size // and jnl->jhdr->jhdr_size are already valid. // static void size_up_tbuffer(journal *jnl, int tbuffer_size, int phys_blksz) { // // one-time initialization based on how much memory // there is in the machine. // if (def_tbuffer_size == 0) { if (max_mem < (256*1024*1024)) { def_tbuffer_size = DEFAULT_TRANSACTION_BUFFER_SIZE; } else if (max_mem < (512*1024*1024)) { def_tbuffer_size = DEFAULT_TRANSACTION_BUFFER_SIZE * 2; } else if (max_mem < (1024*1024*1024)) { def_tbuffer_size = DEFAULT_TRANSACTION_BUFFER_SIZE * 3; } else { def_tbuffer_size = DEFAULT_TRANSACTION_BUFFER_SIZE * (max_mem / (256*1024*1024)); } } // size up the transaction buffer... can't be larger than the number // of blocks that can fit in a block_list_header block. if (tbuffer_size == 0) { jnl->tbuffer_size = def_tbuffer_size; } else { // make sure that the specified tbuffer_size isn't too small if (tbuffer_size < jnl->jhdr->blhdr_size * 2) { tbuffer_size = jnl->jhdr->blhdr_size * 2; } // and make sure it's an even multiple of the block size if ((tbuffer_size % jnl->jhdr->jhdr_size) != 0) { tbuffer_size -= (tbuffer_size % jnl->jhdr->jhdr_size); } jnl->tbuffer_size = tbuffer_size; } if (jnl->tbuffer_size > (jnl->jhdr->size / 2)) { jnl->tbuffer_size = (jnl->jhdr->size / 2); } if (jnl->tbuffer_size > MAX_TRANSACTION_BUFFER_SIZE) { jnl->tbuffer_size = MAX_TRANSACTION_BUFFER_SIZE; } jnl->jhdr->blhdr_size = (jnl->tbuffer_size / jnl->jhdr->jhdr_size) * sizeof(block_info); if (jnl->jhdr->blhdr_size < phys_blksz) { jnl->jhdr->blhdr_size = phys_blksz; } else if ((jnl->jhdr->blhdr_size % phys_blksz) != 0) { // have to round up so we're an even multiple of the physical block size jnl->jhdr->blhdr_size = (jnl->jhdr->blhdr_size + (phys_blksz - 1)) & ~(phys_blksz - 1); } } static void get_io_info(struct vnode *devvp, size_t phys_blksz, journal *jnl, struct vfs_context *context) { off_t readblockcnt; off_t writeblockcnt; off_t readmaxcnt=0, tmp_readmaxcnt; off_t writemaxcnt=0, tmp_writemaxcnt; off_t readsegcnt, writesegcnt; int32_t features; if (VNOP_IOCTL(devvp, DKIOCGETFEATURES, (caddr_t)&features, 0, context) == 0) { if (features & DK_FEATURE_FORCE_UNIT_ACCESS) { const char *name = vnode_getname_printable(devvp); jnl->flags |= JOURNAL_DO_FUA_WRITES; printf("jnl: %s: enabling FUA writes (features 0x%x)\n", name, features); vnode_putname_printable(name); } if (features & DK_FEATURE_UNMAP) { jnl->flags |= JOURNAL_USE_UNMAP; } } // // First check the max read size via several different mechanisms... // VNOP_IOCTL(devvp, DKIOCGETMAXBYTECOUNTREAD, (caddr_t)&readmaxcnt, 0, context); if (VNOP_IOCTL(devvp, DKIOCGETMAXBLOCKCOUNTREAD, (caddr_t)&readblockcnt, 0, context) == 0) { tmp_readmaxcnt = readblockcnt * phys_blksz; if (readmaxcnt == 0 || (readblockcnt > 0 && tmp_readmaxcnt < readmaxcnt)) { readmaxcnt = tmp_readmaxcnt; } } if (VNOP_IOCTL(devvp, DKIOCGETMAXSEGMENTCOUNTREAD, (caddr_t)&readsegcnt, 0, context)) { readsegcnt = 0; } if (readsegcnt > 0 && (readsegcnt * PAGE_SIZE) < readmaxcnt) { readmaxcnt = readsegcnt * PAGE_SIZE; } if (readmaxcnt == 0) { readmaxcnt = 128 * 1024; } else if (readmaxcnt > UINT32_MAX) { readmaxcnt = UINT32_MAX; } // // Now check the max writes size via several different mechanisms... // VNOP_IOCTL(devvp, DKIOCGETMAXBYTECOUNTWRITE, (caddr_t)&writemaxcnt, 0, context); if (VNOP_IOCTL(devvp, DKIOCGETMAXBLOCKCOUNTWRITE, (caddr_t)&writeblockcnt, 0, context) == 0) { tmp_writemaxcnt = writeblockcnt * phys_blksz; if (writemaxcnt == 0 || (writeblockcnt > 0 && tmp_writemaxcnt < writemaxcnt)) { writemaxcnt = tmp_writemaxcnt; } } if (VNOP_IOCTL(devvp, DKIOCGETMAXSEGMENTCOUNTWRITE, (caddr_t)&writesegcnt, 0, context)) { writesegcnt = 0; } if (writesegcnt > 0 && (writesegcnt * PAGE_SIZE) < writemaxcnt) { writemaxcnt = writesegcnt * PAGE_SIZE; } if (writemaxcnt == 0) { writemaxcnt = 128 * 1024; } else if (writemaxcnt > UINT32_MAX) { writemaxcnt = UINT32_MAX; } jnl->max_read_size = readmaxcnt; jnl->max_write_size = writemaxcnt; // printf("jnl: %s: max read/write: %lld k / %lld k\n", // jnl->jdev_name ? jnl->jdev_name : "unknown", // jnl->max_read_size/1024, jnl->max_write_size/1024); } journal * journal_create(struct vnode *jvp, off_t offset, off_t journal_size, struct vnode *fsvp, size_t min_fs_blksz, int32_t flags, int32_t tbuffer_size, void (*flush)(void *arg), void *arg, struct mount *fsmount) { journal *jnl; uint32_t phys_blksz, new_txn_base; u_int32_t min_size; struct vfs_context context; const char *jdev_name; /* * Cap the journal max size to 2GB. On HFS, it will attempt to occupy * a full allocation block if the current size is smaller than the allocation * block on which it resides. Once we hit the exabyte filesystem range, then * it will use 2GB allocation blocks. As a result, make the cap 2GB. */ context.vc_thread = current_thread(); context.vc_ucred = FSCRED; jdev_name = vnode_getname_printable(jvp); /* Get the real physical block size. */ if (VNOP_IOCTL(jvp, DKIOCGETBLOCKSIZE, (caddr_t)&phys_blksz, 0, &context)) { goto cleanup_jdev_name; } if (journal_size < (256*1024) || journal_size > (MAX_JOURNAL_SIZE)) { printf("jnl: %s: create: journal size %lld looks bogus.\n", jdev_name, journal_size); goto cleanup_jdev_name; } min_size = phys_blksz * (phys_blksz / sizeof(block_info)); /* Reject journals that are too small given the sector size of the device */ if (journal_size < min_size) { printf("jnl: %s: create: journal size (%lld) too small given sector size of (%u)\n", jdev_name, journal_size, phys_blksz); goto cleanup_jdev_name; } if (phys_blksz > min_fs_blksz) { printf("jnl: %s: create: error: phys blksize %u bigger than min fs blksize %zd\n", jdev_name, phys_blksz, min_fs_blksz); goto cleanup_jdev_name; } if ((journal_size % phys_blksz) != 0) { printf("jnl: %s: create: journal size 0x%llx is not an even multiple of block size 0x%ux\n", jdev_name, journal_size, phys_blksz); goto cleanup_jdev_name; } MALLOC_ZONE(jnl, struct journal *, sizeof(struct journal), M_JNL_JNL, M_WAITOK); memset(jnl, 0, sizeof(*jnl)); jnl->jdev = jvp; jnl->jdev_offset = offset; jnl->fsdev = fsvp; jnl->flush = flush; jnl->flush_arg = arg; jnl->flags = (flags & JOURNAL_OPTION_FLAGS_MASK); jnl->jdev_name = jdev_name; lck_mtx_init(&jnl->old_start_lock, jnl_mutex_group, jnl_lock_attr); // Keep a point to the mount around for use in IO throttling. jnl->fsmount = fsmount; // XXX: This lock discipline looks correct based on dounmount(), but it // doesn't seem to be documented anywhere. mount_ref(fsmount, 0); get_io_info(jvp, phys_blksz, jnl, &context); if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&jnl->header_buf, phys_blksz)) { printf("jnl: %s: create: could not allocate space for header buffer (%u bytes)\n", jdev_name, phys_blksz); goto bad_kmem_alloc; } jnl->header_buf_size = phys_blksz; jnl->jhdr = (journal_header *)jnl->header_buf; memset(jnl->jhdr, 0, sizeof(journal_header)); // we have to set this up here so that do_journal_io() will work jnl->jhdr->jhdr_size = phys_blksz; // // We try and read the journal header to see if there is already one // out there. If there is, it's possible that it has transactions // in it that we might replay if we happen to pick a sequence number // that is a little less than the old one, there is a crash and the // last txn written ends right at the start of a txn from the previous // incarnation of this file system. If all that happens we would // replay the transactions from the old file system and that would // destroy your disk. Although it is extremely unlikely for all those // conditions to happen, the probability is non-zero and the result is // severe - you lose your file system. Therefore if we find a valid // journal header and the sequence number is non-zero we write junk // over the entire journal so that there is no way we will encounter // any old transactions. This is slow but should be a rare event // since most tools erase the journal. // if ( read_journal_header(jnl, jnl->jhdr, phys_blksz) == phys_blksz && jnl->jhdr->magic == JOURNAL_HEADER_MAGIC && jnl->jhdr->sequence_num != 0) { new_txn_base = (jnl->jhdr->sequence_num + (journal_size / phys_blksz) + (random() % 16384)) & 0x00ffffff; printf("jnl: %s: create: avoiding old sequence number 0x%x (0x%x)\n", jdev_name, jnl->jhdr->sequence_num, new_txn_base); #if 0 int i; off_t pos=0; for(i = 1; i < journal_size / phys_blksz; i++) { pos = i*phys_blksz; // we don't really care what data we write just so long // as it's not a valid transaction header. since we have // the header_buf sitting around we'll use that. write_journal_data(jnl, &pos, jnl->header_buf, phys_blksz); } printf("jnl: create: done clearing journal (i=%d)\n", i); #endif } else { new_txn_base = random() & 0x00ffffff; } memset(jnl->header_buf, 0, phys_blksz); jnl->jhdr->magic = JOURNAL_HEADER_MAGIC; jnl->jhdr->endian = ENDIAN_MAGIC; jnl->jhdr->start = phys_blksz; // start at block #1, block #0 is for the jhdr itself jnl->jhdr->end = phys_blksz; jnl->jhdr->size = journal_size; jnl->jhdr->jhdr_size = phys_blksz; size_up_tbuffer(jnl, tbuffer_size, phys_blksz); jnl->active_start = jnl->jhdr->start; // XXXdbg - for testing you can force the journal to wrap around // jnl->jhdr->start = jnl->jhdr->size - (phys_blksz*3); // jnl->jhdr->end = jnl->jhdr->size - (phys_blksz*3); jnl->jhdr->sequence_num = new_txn_base; lck_mtx_init(&jnl->jlock, jnl_mutex_group, jnl_lock_attr); lck_mtx_init(&jnl->flock, jnl_mutex_group, jnl_lock_attr); lck_rw_init(&jnl->trim_lock, jnl_mutex_group, jnl_lock_attr); jnl->flushing = FALSE; jnl->asyncIO = FALSE; jnl->flush_aborted = FALSE; jnl->writing_header = FALSE; jnl->async_trim = NULL; jnl->sequence_num = jnl->jhdr->sequence_num; if (write_journal_header(jnl, 1, jnl->jhdr->sequence_num) != 0) { printf("jnl: %s: journal_create: failed to write journal header.\n", jdev_name); goto bad_write; } goto journal_create_complete; bad_write: kmem_free(kernel_map, (vm_offset_t)jnl->header_buf, phys_blksz); bad_kmem_alloc: jnl->jhdr = NULL; FREE_ZONE(jnl, sizeof(struct journal), M_JNL_JNL); mount_drop(fsmount, 0); cleanup_jdev_name: vnode_putname_printable(jdev_name); jnl = NULL; journal_create_complete: return jnl; } journal * journal_open(struct vnode *jvp, off_t offset, off_t journal_size, struct vnode *fsvp, size_t min_fs_blksz, int32_t flags, int32_t tbuffer_size, void (*flush)(void *arg), void *arg, struct mount *fsmount) { journal *jnl; uint32_t orig_blksz=0; uint32_t phys_blksz; u_int32_t min_size = 0; int orig_checksum, checksum; struct vfs_context context; const char *jdev_name = vnode_getname_printable(jvp); context.vc_thread = current_thread(); context.vc_ucred = FSCRED; /* Get the real physical block size. */ if (VNOP_IOCTL(jvp, DKIOCGETBLOCKSIZE, (caddr_t)&phys_blksz, 0, &context)) { goto cleanup_jdev_name; } if (phys_blksz > min_fs_blksz) { printf("jnl: %s: open: error: phys blksize %u bigger than min fs blksize %zd\n", jdev_name, phys_blksz, min_fs_blksz); goto cleanup_jdev_name; } if (journal_size < (256*1024) || journal_size > (1024*1024*1024)) { printf("jnl: %s: open: journal size %lld looks bogus.\n", jdev_name, journal_size); goto cleanup_jdev_name; } min_size = phys_blksz * (phys_blksz / sizeof(block_info)); /* Reject journals that are too small given the sector size of the device */ if (journal_size < min_size) { printf("jnl: %s: open: journal size (%lld) too small given sector size of (%u)\n", jdev_name, journal_size, phys_blksz); goto cleanup_jdev_name; } if ((journal_size % phys_blksz) != 0) { printf("jnl: %s: open: journal size 0x%llx is not an even multiple of block size 0x%x\n", jdev_name, journal_size, phys_blksz); goto cleanup_jdev_name; } MALLOC_ZONE(jnl, struct journal *, sizeof(struct journal), M_JNL_JNL, M_WAITOK); memset(jnl, 0, sizeof(*jnl)); jnl->jdev = jvp; jnl->jdev_offset = offset; jnl->fsdev = fsvp; jnl->flush = flush; jnl->flush_arg = arg; jnl->flags = (flags & JOURNAL_OPTION_FLAGS_MASK); jnl->jdev_name = jdev_name; lck_mtx_init(&jnl->old_start_lock, jnl_mutex_group, jnl_lock_attr); /* We need a reference to the mount to later pass to the throttling code for * IO accounting. */ jnl->fsmount = fsmount; mount_ref(fsmount, 0); get_io_info(jvp, phys_blksz, jnl, &context); if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&jnl->header_buf, phys_blksz)) { printf("jnl: %s: create: could not allocate space for header buffer (%u bytes)\n", jdev_name, phys_blksz); goto bad_kmem_alloc; } jnl->header_buf_size = phys_blksz; jnl->jhdr = (journal_header *)jnl->header_buf; memset(jnl->jhdr, 0, sizeof(journal_header)); // we have to set this up here so that do_journal_io() will work jnl->jhdr->jhdr_size = phys_blksz; if (read_journal_header(jnl, jnl->jhdr, phys_blksz) != phys_blksz) { printf("jnl: %s: open: could not read %u bytes for the journal header.\n", jdev_name, phys_blksz); goto bad_journal; } orig_checksum = jnl->jhdr->checksum; jnl->jhdr->checksum = 0; if (jnl->jhdr->magic == SWAP32(JOURNAL_HEADER_MAGIC)) { // do this before the swap since it's done byte-at-a-time orig_checksum = SWAP32(orig_checksum); checksum = calc_checksum((char *)jnl->jhdr, JOURNAL_HEADER_CKSUM_SIZE); swap_journal_header(jnl); jnl->flags |= JOURNAL_NEED_SWAP; } else { checksum = calc_checksum((char *)jnl->jhdr, JOURNAL_HEADER_CKSUM_SIZE); } if (jnl->jhdr->magic != JOURNAL_HEADER_MAGIC && jnl->jhdr->magic != OLD_JOURNAL_HEADER_MAGIC) { printf("jnl: %s: open: journal magic is bad (0x%x != 0x%x)\n", jnl->jdev_name, jnl->jhdr->magic, JOURNAL_HEADER_MAGIC); goto bad_journal; } // only check if we're the current journal header magic value if (jnl->jhdr->magic == JOURNAL_HEADER_MAGIC) { if (orig_checksum != checksum) { printf("jnl: %s: open: journal checksum is bad (0x%x != 0x%x)\n", jdev_name, orig_checksum, checksum); //goto bad_journal; } } // XXXdbg - convert old style magic numbers to the new one if (jnl->jhdr->magic == OLD_JOURNAL_HEADER_MAGIC) { jnl->jhdr->magic = JOURNAL_HEADER_MAGIC; } if (phys_blksz != (size_t)jnl->jhdr->jhdr_size && jnl->jhdr->jhdr_size != 0) { /* * The volume has probably been resized (such that we had to adjust the * logical sector size), or copied to media with a different logical * sector size. * * Temporarily change the device's logical block size to match the * journal's header size. This will allow us to replay the journal * safely. If the replay succeeds, we will update the journal's header * size (later in this function). */ orig_blksz = phys_blksz; phys_blksz = jnl->jhdr->jhdr_size; VNOP_IOCTL(jvp, DKIOCSETBLOCKSIZE, (caddr_t)&phys_blksz, FWRITE, &context); printf("jnl: %s: open: temporarily switched block size from %u to %u\n", jdev_name, orig_blksz, phys_blksz); } if ( jnl->jhdr->start <= 0 || jnl->jhdr->start > jnl->jhdr->size || jnl->jhdr->start > 1024*1024*1024) { printf("jnl: %s: open: jhdr start looks bad (0x%llx max size 0x%llx)\n", jdev_name, jnl->jhdr->start, jnl->jhdr->size); goto bad_journal; } if ( jnl->jhdr->end <= 0 || jnl->jhdr->end > jnl->jhdr->size || jnl->jhdr->end > 1024*1024*1024) { printf("jnl: %s: open: jhdr end looks bad (0x%llx max size 0x%llx)\n", jdev_name, jnl->jhdr->end, jnl->jhdr->size); goto bad_journal; } if (jnl->jhdr->size < (256*1024) || jnl->jhdr->size > 1024*1024*1024) { printf("jnl: %s: open: jhdr size looks bad (0x%llx)\n", jdev_name, jnl->jhdr->size); goto bad_journal; } // XXXdbg - can't do these checks because hfs writes all kinds of // non-uniform sized blocks even on devices that have a block size // that is larger than 512 bytes (i.e. optical media w/2k blocks). // therefore these checks will fail and so we just have to punt and // do more relaxed checking... // XXXdbg if ((jnl->jhdr->start % jnl->jhdr->jhdr_size) != 0) { if ((jnl->jhdr->start % 512) != 0) { printf("jnl: %s: open: journal start (0x%llx) not a multiple of 512?\n", jdev_name, jnl->jhdr->start); goto bad_journal; } //XXXdbg if ((jnl->jhdr->end % jnl->jhdr->jhdr_size) != 0) { if ((jnl->jhdr->end % 512) != 0) { printf("jnl: %s: open: journal end (0x%llx) not a multiple of block size (0x%x)?\n", jdev_name, jnl->jhdr->end, jnl->jhdr->jhdr_size); goto bad_journal; } // take care of replaying the journal if necessary if (flags & JOURNAL_RESET) { printf("jnl: %s: journal start/end pointers reset! (jnl %p; s 0x%llx e 0x%llx)\n", jdev_name, jnl, jnl->jhdr->start, jnl->jhdr->end); jnl->jhdr->start = jnl->jhdr->end; } else if (replay_journal(jnl) != 0) { printf("jnl: %s: journal_open: Error replaying the journal!\n", jdev_name); goto bad_journal; } /* * When we get here, we know that the journal is empty (jnl->jhdr->start == * jnl->jhdr->end). If the device's logical block size was different from * the journal's header size, then we can now restore the device's logical * block size and update the journal's header size to match. * * Note that we also adjust the journal's start and end so that they will * be aligned on the new block size. We pick a new sequence number to * avoid any problems if a replay found previous transactions using the old * journal header size. (See the comments in journal_create(), above.) */ if (orig_blksz != 0) { VNOP_IOCTL(jvp, DKIOCSETBLOCKSIZE, (caddr_t)&orig_blksz, FWRITE, &context); phys_blksz = orig_blksz; orig_blksz = 0; jnl->jhdr->jhdr_size = phys_blksz; jnl->jhdr->start = phys_blksz; jnl->jhdr->end = phys_blksz; jnl->jhdr->sequence_num = (jnl->jhdr->sequence_num + (journal_size / phys_blksz) + (random() % 16384)) & 0x00ffffff; if (write_journal_header(jnl, 1, jnl->jhdr->sequence_num)) { printf("jnl: %s: open: failed to update journal header size\n", jdev_name); goto bad_journal; } } // make sure this is in sync! jnl->active_start = jnl->jhdr->start; jnl->sequence_num = jnl->jhdr->sequence_num; // set this now, after we've replayed the journal size_up_tbuffer(jnl, tbuffer_size, phys_blksz); // TODO: Does this need to change if the device's logical block size changed? if ((off_t)(jnl->jhdr->blhdr_size/sizeof(block_info)-1) > (jnl->jhdr->size/jnl->jhdr->jhdr_size)) { printf("jnl: %s: open: jhdr size and blhdr size are not compatible (0x%llx, %d, %d)\n", jdev_name, jnl->jhdr->size, jnl->jhdr->blhdr_size, jnl->jhdr->jhdr_size); goto bad_journal; } lck_mtx_init(&jnl->jlock, jnl_mutex_group, jnl_lock_attr); lck_mtx_init(&jnl->flock, jnl_mutex_group, jnl_lock_attr); lck_rw_init(&jnl->trim_lock, jnl_mutex_group, jnl_lock_attr); goto journal_open_complete; bad_journal: if (orig_blksz != 0) { phys_blksz = orig_blksz; VNOP_IOCTL(jvp, DKIOCSETBLOCKSIZE, (caddr_t)&orig_blksz, FWRITE, &context); printf("jnl: %s: open: restored block size after error\n", jdev_name); } kmem_free(kernel_map, (vm_offset_t)jnl->header_buf, phys_blksz); bad_kmem_alloc: FREE_ZONE(jnl, sizeof(struct journal), M_JNL_JNL); mount_drop(fsmount, 0); cleanup_jdev_name: vnode_putname_printable(jdev_name); jnl = NULL; journal_open_complete: return jnl; } int journal_is_clean(struct vnode *jvp, off_t offset, off_t journal_size, struct vnode *fsvp, size_t min_fs_block_size) { journal jnl; uint32_t phys_blksz; int ret; int orig_checksum, checksum; struct vfs_context context; const char *jdev_name = vnode_getname_printable(jvp); context.vc_thread = current_thread(); context.vc_ucred = FSCRED; /* Get the real physical block size. */ if (VNOP_IOCTL(jvp, DKIOCGETBLOCKSIZE, (caddr_t)&phys_blksz, 0, &context)) { printf("jnl: %s: is_clean: failed to get device block size.\n", jdev_name); ret = EINVAL; goto cleanup_jdev_name; } if (phys_blksz > (uint32_t)min_fs_block_size) { printf("jnl: %s: is_clean: error: phys blksize %d bigger than min fs blksize %zd\n", jdev_name, phys_blksz, min_fs_block_size); ret = EINVAL; goto cleanup_jdev_name; } if (journal_size < (256*1024) || journal_size > (MAX_JOURNAL_SIZE)) { printf("jnl: %s: is_clean: journal size %lld looks bogus.\n", jdev_name, journal_size); ret = EINVAL; goto cleanup_jdev_name; } if ((journal_size % phys_blksz) != 0) { printf("jnl: %s: is_clean: journal size 0x%llx is not an even multiple of block size 0x%x\n", jdev_name, journal_size, phys_blksz); ret = EINVAL; goto cleanup_jdev_name; } memset(&jnl, 0, sizeof(jnl)); if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&jnl.header_buf, phys_blksz)) { printf("jnl: %s: is_clean: could not allocate space for header buffer (%d bytes)\n", jdev_name, phys_blksz); ret = ENOMEM; goto cleanup_jdev_name; } jnl.header_buf_size = phys_blksz; get_io_info(jvp, phys_blksz, &jnl, &context); jnl.jhdr = (journal_header *)jnl.header_buf; memset(jnl.jhdr, 0, sizeof(journal_header)); jnl.jdev = jvp; jnl.jdev_offset = offset; jnl.fsdev = fsvp; // we have to set this up here so that do_journal_io() will work jnl.jhdr->jhdr_size = phys_blksz; if (read_journal_header(&jnl, jnl.jhdr, phys_blksz) != (unsigned)phys_blksz) { printf("jnl: %s: is_clean: could not read %d bytes for the journal header.\n", jdev_name, phys_blksz); ret = EINVAL; goto get_out; } orig_checksum = jnl.jhdr->checksum; jnl.jhdr->checksum = 0; if (jnl.jhdr->magic == SWAP32(JOURNAL_HEADER_MAGIC)) { // do this before the swap since it's done byte-at-a-time orig_checksum = SWAP32(orig_checksum); checksum = calc_checksum((char *)jnl.jhdr, JOURNAL_HEADER_CKSUM_SIZE); swap_journal_header(&jnl); jnl.flags |= JOURNAL_NEED_SWAP; } else { checksum = calc_checksum((char *)jnl.jhdr, JOURNAL_HEADER_CKSUM_SIZE); } if (jnl.jhdr->magic != JOURNAL_HEADER_MAGIC && jnl.jhdr->magic != OLD_JOURNAL_HEADER_MAGIC) { printf("jnl: %s: is_clean: journal magic is bad (0x%x != 0x%x)\n", jdev_name, jnl.jhdr->magic, JOURNAL_HEADER_MAGIC); ret = EINVAL; goto get_out; } if (orig_checksum != checksum) { printf("jnl: %s: is_clean: journal checksum is bad (0x%x != 0x%x)\n", jdev_name, orig_checksum, checksum); ret = EINVAL; goto get_out; } // // if the start and end are equal then the journal is clean. // otherwise it's not clean and therefore an error. // if (jnl.jhdr->start == jnl.jhdr->end) { ret = 0; } else { ret = EBUSY; // so the caller can differentiate an invalid journal from a "busy" one } get_out: kmem_free(kernel_map, (vm_offset_t)jnl.header_buf, phys_blksz); cleanup_jdev_name: vnode_putname_printable(jdev_name); return ret; } void journal_close(journal *jnl) { volatile off_t *start, *end; int counter=0; CHECK_JOURNAL(jnl); // set this before doing anything that would block so that // we start tearing things down properly. // jnl->flags |= JOURNAL_CLOSE_PENDING; if (jnl->owner != current_thread()) { journal_lock(jnl); } wait_condition(jnl, &jnl->flushing, "journal_close"); // // only write stuff to disk if the journal is still valid // if ((jnl->flags & JOURNAL_INVALID) == 0) { if (jnl->active_tr) { /* * "journal_end_transaction" will fire the flush asynchronously */ journal_end_transaction(jnl); } // flush any buffered transactions if (jnl->cur_tr) { transaction *tr = jnl->cur_tr; jnl->cur_tr = NULL; /* * "end_transaction" will wait for any in-progress flush to complete * before flushing "cur_tr" synchronously("must_wait" == TRUE) */ end_transaction(tr, 1, NULL, NULL, FALSE, TRUE); } /* * if there was an "active_tr", make sure we wait for * it to flush if there was no "cur_tr" to process */ wait_condition(jnl, &jnl->flushing, "journal_close"); //start = &jnl->jhdr->start; start = &jnl->active_start; end = &jnl->jhdr->end; while (*start != *end && counter++ < 5000) { //printf("jnl: close: flushing the buffer cache (start 0x%llx end 0x%llx)\n", *start, *end); if (jnl->flush) { jnl->flush(jnl->flush_arg); } tsleep((caddr_t)jnl, PRIBIO, "jnl_close", 2); } if (*start != *end) { printf("jnl: %s: close: buffer flushing didn't seem to flush out all the transactions! (0x%llx - 0x%llx)\n", jnl->jdev_name, *start, *end); } // make sure this is in sync when we close the journal jnl->jhdr->start = jnl->active_start; // if this fails there's not much we can do at this point... write_journal_header(jnl, 1, jnl->sequence_num); } else { // if we're here the journal isn't valid any more. // so make sure we don't leave any locked blocks lying around printf("jnl: %s: close: journal %p, is invalid. aborting outstanding transactions\n", jnl->jdev_name, jnl); if (jnl->active_tr || jnl->cur_tr) { transaction *tr; if (jnl->active_tr) { tr = jnl->active_tr; jnl->active_tr = NULL; } else { tr = jnl->cur_tr; jnl->cur_tr = NULL; } abort_transaction(jnl, tr); if (jnl->active_tr || jnl->cur_tr) { panic("jnl: %s: close: jnl @ %p had both an active and cur tr\n", jnl->jdev_name, jnl); } } } wait_condition(jnl, &jnl->asyncIO, "journal_close"); free_old_stuff(jnl); kmem_free(kernel_map, (vm_offset_t)jnl->header_buf, jnl->header_buf_size); jnl->jhdr = (void *)0xbeefbabe; // Release reference on the mount if (jnl->fsmount) mount_drop(jnl->fsmount, 0); vnode_putname_printable(jnl->jdev_name); journal_unlock(jnl); lck_mtx_destroy(&jnl->old_start_lock, jnl_mutex_group); lck_mtx_destroy(&jnl->jlock, jnl_mutex_group); lck_mtx_destroy(&jnl->flock, jnl_mutex_group); FREE_ZONE(jnl, sizeof(struct journal), M_JNL_JNL); } static void dump_journal(journal *jnl) { transaction *ctr; printf("journal for dev %s:", jnl->jdev_name); printf(" jdev_offset %.8llx\n", jnl->jdev_offset); printf(" magic: 0x%.8x\n", jnl->jhdr->magic); printf(" start: 0x%.8llx\n", jnl->jhdr->start); printf(" end: 0x%.8llx\n", jnl->jhdr->end); printf(" size: 0x%.8llx\n", jnl->jhdr->size); printf(" blhdr size: %d\n", jnl->jhdr->blhdr_size); printf(" jhdr size: %d\n", jnl->jhdr->jhdr_size); printf(" chksum: 0x%.8x\n", jnl->jhdr->checksum); printf(" completed transactions:\n"); for (ctr = jnl->completed_trs; ctr; ctr = ctr->next) { printf(" 0x%.8llx - 0x%.8llx\n", ctr->journal_start, ctr->journal_end); } } static off_t free_space(journal *jnl) { off_t free_space_offset; if (jnl->jhdr->start < jnl->jhdr->end) { free_space_offset = jnl->jhdr->size - (jnl->jhdr->end - jnl->jhdr->start) - jnl->jhdr->jhdr_size; } else if (jnl->jhdr->start > jnl->jhdr->end) { free_space_offset = jnl->jhdr->start - jnl->jhdr->end; } else { // journal is completely empty free_space_offset = jnl->jhdr->size - jnl->jhdr->jhdr_size; } return free_space_offset; } // // The journal must be locked on entry to this function. // The "desired_size" is in bytes. // static int check_free_space(journal *jnl, int desired_size, boolean_t *delayed_header_write, uint32_t sequence_num) { size_t i; int counter=0; //printf("jnl: check free space (desired 0x%x, avail 0x%Lx)\n", // desired_size, free_space(jnl)); if (delayed_header_write) *delayed_header_write = FALSE; while (1) { int old_start_empty; // make sure there's space in the journal to hold this transaction if (free_space(jnl) > desired_size && jnl->old_start[0] == 0) { break; } if (counter++ == 5000) { dump_journal(jnl); panic("jnl: check_free_space: buffer flushing isn't working " "(jnl @ %p s %lld e %lld f %lld [active start %lld]).\n", jnl, jnl->jhdr->start, jnl->jhdr->end, free_space(jnl), jnl->active_start); } if (counter > 7500) { printf("jnl: %s: check_free_space: giving up waiting for free space.\n", jnl->jdev_name); return ENOSPC; } // // here's where we lazily bump up jnl->jhdr->start. we'll consume // entries until there is enough space for the next transaction. // old_start_empty = 1; lock_oldstart(jnl); for (i = 0; i < sizeof(jnl->old_start)/sizeof(jnl->old_start[0]); i++) { int lcl_counter; lcl_counter = 0; while (jnl->old_start[i] & 0x8000000000000000LL) { if (lcl_counter++ > 10000) { panic("jnl: check_free_space: tr starting @ 0x%llx not flushing (jnl %p).\n", jnl->old_start[i], jnl); } unlock_oldstart(jnl); if (jnl->flush) { jnl->flush(jnl->flush_arg); } tsleep((caddr_t)jnl, PRIBIO, "check_free_space1", 1); lock_oldstart(jnl); } if (jnl->old_start[i] == 0) { continue; } old_start_empty = 0; jnl->jhdr->start = jnl->old_start[i]; jnl->old_start[i] = 0; if (free_space(jnl) > desired_size) { if (delayed_header_write) *delayed_header_write = TRUE; else { unlock_oldstart(jnl); write_journal_header(jnl, 1, sequence_num); lock_oldstart(jnl); } break; } } unlock_oldstart(jnl); // if we bumped the start, loop and try again if (i < sizeof(jnl->old_start)/sizeof(jnl->old_start[0])) { continue; } else if (old_start_empty) { // // if there is nothing in old_start anymore then we can // bump the jhdr->start to be the same as active_start // since it is possible there was only one very large // transaction in the old_start array. if we didn't do // this then jhdr->start would never get updated and we // would wind up looping until we hit the panic at the // start of the loop. // jnl->jhdr->start = jnl->active_start; if (delayed_header_write) *delayed_header_write = TRUE; else write_journal_header(jnl, 1, sequence_num); continue; } // if the file system gave us a flush function, call it to so that // it can flush some blocks which hopefully will cause some transactions // to complete and thus free up space in the journal. if (jnl->flush) { jnl->flush(jnl->flush_arg); } // wait for a while to avoid being cpu-bound (this will // put us to sleep for 10 milliseconds) tsleep((caddr_t)jnl, PRIBIO, "check_free_space2", 1); } return 0; } /* * Allocate a new active transaction. */ static errno_t journal_allocate_transaction(journal *jnl) { transaction *tr; boolean_t was_vm_privileged; if (jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) { /* * the disk driver can allocate memory on this path... * if we block waiting for memory, and there is enough pressure to * cause us to try and create a new swap file, we may end up deadlocking * due to waiting for the journal on the swap file creation path... * by making ourselves vm_privileged, we give ourselves the best chance * of not blocking */ was_vm_privileged = set_vm_privilege(TRUE); } MALLOC_ZONE(tr, transaction *, sizeof(transaction), M_JNL_TR, M_WAITOK); memset(tr, 0, sizeof(transaction)); tr->tbuffer_size = jnl->tbuffer_size; if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&tr->tbuffer, tr->tbuffer_size)) { FREE_ZONE(tr, sizeof(transaction), M_JNL_TR); jnl->active_tr = NULL; return ENOMEM; } if ((jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) && (was_vm_privileged == FALSE)) set_vm_privilege(FALSE); // journal replay code checksum check depends on this. memset(tr->tbuffer, 0, BLHDR_CHECKSUM_SIZE); // Fill up the rest of the block with unimportant bytes (0x5a 'Z' chosen for visibility) memset(tr->tbuffer + BLHDR_CHECKSUM_SIZE, 0x5a, jnl->jhdr->blhdr_size - BLHDR_CHECKSUM_SIZE); tr->blhdr = (block_list_header *)tr->tbuffer; tr->blhdr->max_blocks = (jnl->jhdr->blhdr_size / sizeof(block_info)) - 1; tr->blhdr->num_blocks = 1; // accounts for this header block tr->blhdr->bytes_used = jnl->jhdr->blhdr_size; tr->blhdr->flags = BLHDR_CHECK_CHECKSUMS | BLHDR_FIRST_HEADER; tr->sequence_num = ++jnl->sequence_num; tr->num_blhdrs = 1; tr->total_bytes = jnl->jhdr->blhdr_size; tr->jnl = jnl; jnl->active_tr = tr; return 0; } int journal_start_transaction(journal *jnl) { int ret; CHECK_JOURNAL(jnl); free_old_stuff(jnl); if (jnl->flags & JOURNAL_INVALID) { return EINVAL; } if (jnl->owner == current_thread()) { if (jnl->active_tr == NULL) { panic("jnl: start_tr: active_tr is NULL (jnl @ %p, owner %p, current_thread %p\n", jnl, jnl->owner, current_thread()); } jnl->nested_count++; return 0; } journal_lock(jnl); if (jnl->nested_count != 0 || jnl->active_tr != NULL) { panic("jnl: start_tr: owner %p, nested count %d, active_tr %p jnl @ %p\n", jnl->owner, jnl->nested_count, jnl->active_tr, jnl); } jnl->nested_count = 1; #if JOE // make sure there's room in the journal if (free_space(jnl) < jnl->tbuffer_size) { KERNEL_DEBUG(0xbbbbc030 | DBG_FUNC_START, jnl, 0, 0, 0, 0); // this is the call that really waits for space to free up // as well as updating jnl->jhdr->start if (check_free_space(jnl, jnl->tbuffer_size, NULL, jnl->sequence_num) != 0) { printf("jnl: %s: start transaction failed: no space\n", jnl->jdev_name); ret = ENOSPC; goto bad_start; } KERNEL_DEBUG(0xbbbbc030 | DBG_FUNC_END, jnl, 0, 0, 0, 0); } #endif // if there's a buffered transaction, use it. if (jnl->cur_tr) { jnl->active_tr = jnl->cur_tr; jnl->cur_tr = NULL; return 0; } ret = journal_allocate_transaction(jnl); if (ret) { goto bad_start; } // printf("jnl: start_tr: owner 0x%x new tr @ 0x%x\n", jnl->owner, jnl->active_tr); return 0; bad_start: jnl->nested_count = 0; journal_unlock(jnl); return ret; } int journal_modify_block_start(journal *jnl, struct buf *bp) { transaction *tr; CHECK_JOURNAL(jnl); free_old_stuff(jnl); if (jnl->flags & JOURNAL_INVALID) { return EINVAL; } // XXXdbg - for debugging I want this to be true. later it may // not be necessary. if ((buf_flags(bp) & B_META) == 0) { panic("jnl: modify_block_start: bp @ %p is not a meta-data block! (jnl %p)\n", bp, jnl); } tr = jnl->active_tr; CHECK_TRANSACTION(tr); if (jnl->owner != current_thread()) { panic("jnl: modify_block_start: called w/out a transaction! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } //printf("jnl: mod block start (bp 0x%x vp 0x%x l/blkno %qd/%qd bsz %d; total bytes %d)\n", // bp, buf_vnode(bp), buf_lblkno(bp), buf_blkno(bp), buf_size(bp), tr->total_bytes); // can't allow blocks that aren't an even multiple of the // underlying block size. if ((buf_size(bp) % jnl->jhdr->jhdr_size) != 0) { uint32_t phys_blksz, bad=0; if (VNOP_IOCTL(jnl->jdev, DKIOCGETBLOCKSIZE, (caddr_t)&phys_blksz, 0, vfs_context_kernel())) { bad = 1; } else if (phys_blksz != (uint32_t)jnl->jhdr->jhdr_size) { if (phys_blksz < 512) { panic("jnl: mod block start: phys blksz %d is too small (%d, %d)\n", phys_blksz, buf_size(bp), jnl->jhdr->jhdr_size); } if ((buf_size(bp) % phys_blksz) != 0) { bad = 1; } else if (phys_blksz < (uint32_t)jnl->jhdr->jhdr_size) { jnl->jhdr->jhdr_size = phys_blksz; } else { // the phys_blksz is now larger... need to realloc the jhdr char *new_header_buf; printf("jnl: %s: phys blksz got bigger (was: %d/%d now %d)\n", jnl->jdev_name, jnl->header_buf_size, jnl->jhdr->jhdr_size, phys_blksz); if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&new_header_buf, phys_blksz)) { printf("jnl: modify_block_start: %s: create: phys blksz change (was %d, now %d) but could not allocate space for new header\n", jnl->jdev_name, jnl->jhdr->jhdr_size, phys_blksz); bad = 1; } else { memcpy(new_header_buf, jnl->header_buf, jnl->header_buf_size); memset(&new_header_buf[jnl->header_buf_size], 0x18, (phys_blksz - jnl->header_buf_size)); kmem_free(kernel_map, (vm_offset_t)jnl->header_buf, jnl->header_buf_size); jnl->header_buf = new_header_buf; jnl->header_buf_size = phys_blksz; jnl->jhdr = (journal_header *)jnl->header_buf; jnl->jhdr->jhdr_size = phys_blksz; } } } else { bad = 1; } if (bad) { panic("jnl: mod block start: bufsize %d not a multiple of block size %d\n", buf_size(bp), jnl->jhdr->jhdr_size); return -1; } } // make sure that this transaction isn't bigger than the whole journal if (tr->total_bytes+buf_size(bp) >= (jnl->jhdr->size - jnl->jhdr->jhdr_size)) { panic("jnl: transaction too big (%d >= %lld bytes, bufsize %d, tr %p bp %p)\n", tr->total_bytes, (tr->jnl->jhdr->size - jnl->jhdr->jhdr_size), buf_size(bp), tr, bp); return -1; } // if the block is dirty and not already locked we have to write // it out before we muck with it because it has data that belongs // (presumably) to another transaction. // if ((buf_flags(bp) & (B_DELWRI | B_LOCKED)) == B_DELWRI) { if (buf_flags(bp) & B_ASYNC) { panic("modify_block_start: bp @ %p has async flag set!\n", bp); } if (bp->b_shadow_ref) panic("modify_block_start: dirty bp @ %p has shadows!\n", bp); // this will cause it to not be buf_brelse()'d buf_setflags(bp, B_NORELSE); VNOP_BWRITE(bp); } buf_setflags(bp, B_LOCKED); return 0; } int journal_modify_block_abort(journal *jnl, struct buf *bp) { transaction *tr; block_list_header *blhdr; int i; CHECK_JOURNAL(jnl); free_old_stuff(jnl); tr = jnl->active_tr; // // if there's no active transaction then we just want to // call buf_brelse() and return since this is just a block // that happened to be modified as part of another tr. // if (tr == NULL) { buf_brelse(bp); return 0; } if (jnl->flags & JOURNAL_INVALID) { /* Still need to buf_brelse(). Callers assume we consume the bp. */ buf_brelse(bp); return EINVAL; } CHECK_TRANSACTION(tr); if (jnl->owner != current_thread()) { panic("jnl: modify_block_abort: called w/out a transaction! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } // printf("jnl: modify_block_abort: tr 0x%x bp 0x%x\n", jnl->active_tr, bp); // first check if it's already part of this transaction for (blhdr = tr->blhdr; blhdr; blhdr = (block_list_header *)((long)blhdr->binfo[0].bnum)) { for (i = 1; i < blhdr->num_blocks; i++) { if (bp == blhdr->binfo[i].u.bp) { break; } } if (i < blhdr->num_blocks) { break; } } // // if blhdr is null, then this block has only had modify_block_start // called on it as part of the current transaction. that means that // it is ok to clear the LOCKED bit since it hasn't actually been // modified. if blhdr is non-null then modify_block_end was called // on it and so we need to keep it locked in memory. // if (blhdr == NULL) { buf_clearflags(bp, B_LOCKED); } buf_brelse(bp); return 0; } int journal_modify_block_end(journal *jnl, struct buf *bp, void (*func)(buf_t bp, void *arg), void *arg) { int i = 1; int tbuffer_offset=0; block_list_header *blhdr, *prev=NULL; transaction *tr; CHECK_JOURNAL(jnl); free_old_stuff(jnl); if (jnl->flags & JOURNAL_INVALID) { /* Still need to buf_brelse(). Callers assume we consume the bp. */ buf_brelse(bp); return EINVAL; } tr = jnl->active_tr; CHECK_TRANSACTION(tr); if (jnl->owner != current_thread()) { panic("jnl: modify_block_end: called w/out a transaction! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } //printf("jnl: mod block end: (bp 0x%x vp 0x%x l/blkno %qd/%qd bsz %d, total bytes %d)\n", // bp, buf_vnode(bp), buf_lblkno(bp), buf_blkno(bp), buf_size(bp), tr->total_bytes); if ((buf_flags(bp) & B_LOCKED) == 0) { panic("jnl: modify_block_end: bp %p not locked! jnl @ %p\n", bp, jnl); } // first check if it's already part of this transaction for (blhdr = tr->blhdr; blhdr; prev = blhdr, blhdr = (block_list_header *)((long)blhdr->binfo[0].bnum)) { tbuffer_offset = jnl->jhdr->blhdr_size; for (i = 1; i < blhdr->num_blocks; i++) { if (bp == blhdr->binfo[i].u.bp) { break; } if (blhdr->binfo[i].bnum != (off_t)-1) { tbuffer_offset += buf_size(blhdr->binfo[i].u.bp); } else { tbuffer_offset += blhdr->binfo[i].u.bi.bsize; } } if (i < blhdr->num_blocks) { break; } } if (blhdr == NULL && prev && (prev->num_blocks+1) <= prev->max_blocks && (prev->bytes_used+buf_size(bp)) <= (uint32_t)tr->tbuffer_size) { blhdr = prev; } else if (blhdr == NULL) { block_list_header *nblhdr; if (prev == NULL) { panic("jnl: modify block end: no way man, prev == NULL?!?, jnl %p, bp %p\n", jnl, bp); } // we got to the end of the list, didn't find the block and there's // no room in the block_list_header pointed to by prev // we allocate another tbuffer and link it in at the end of the list // through prev->binfo[0].bnum. that's a skanky way to do things but // avoids having yet another linked list of small data structures to manage. if (kmem_alloc_kobject(kernel_map, (vm_offset_t *)&nblhdr, tr->tbuffer_size)) { panic("jnl: end_tr: no space for new block tr @ %p (total bytes: %d)!\n", tr, tr->total_bytes); } // journal replay code checksum check depends on this. memset(nblhdr, 0, BLHDR_CHECKSUM_SIZE); // Fill up the rest of the block with unimportant bytes memset(nblhdr + BLHDR_CHECKSUM_SIZE, 0x5a, jnl->jhdr->blhdr_size - BLHDR_CHECKSUM_SIZE); // initialize the new guy nblhdr->max_blocks = (jnl->jhdr->blhdr_size / sizeof(block_info)) - 1; nblhdr->num_blocks = 1; // accounts for this header block nblhdr->bytes_used = jnl->jhdr->blhdr_size; nblhdr->flags = BLHDR_CHECK_CHECKSUMS; tr->num_blhdrs++; tr->total_bytes += jnl->jhdr->blhdr_size; // then link him in at the end prev->binfo[0].bnum = (off_t)((long)nblhdr); // and finally switch to using the new guy blhdr = nblhdr; tbuffer_offset = jnl->jhdr->blhdr_size; i = 1; } if ((i+1) > blhdr->max_blocks) { panic("jnl: modify_block_end: i = %d, max_blocks %d\n", i, blhdr->max_blocks); } // if this is true then this is a new block we haven't seen if (i >= blhdr->num_blocks) { int bsize; vnode_t vp; vp = buf_vnode(bp); vnode_ref(vp); bsize = buf_size(bp); blhdr->binfo[i].bnum = (off_t)(buf_blkno(bp)); blhdr->binfo[i].u.bp = bp; KERNEL_DEBUG_CONSTANT(0x3018004, VM_KERNEL_ADDRPERM(vp), blhdr->binfo[i].bnum, bsize, 0, 0); if (func) { void (*old_func)(buf_t, void *)=NULL, *old_arg=NULL; buf_setfilter(bp, func, arg, &old_func, &old_arg); if (old_func != NULL && old_func != func) { panic("jnl: modify_block_end: old func %p / arg %p (func %p)", old_func, old_arg, func); } } blhdr->bytes_used += bsize; tr->total_bytes += bsize; blhdr->num_blocks++; } buf_bdwrite(bp); return 0; } int journal_kill_block(journal *jnl, struct buf *bp) { int i; int bflags; block_list_header *blhdr; transaction *tr; CHECK_JOURNAL(jnl); free_old_stuff(jnl); if (jnl->flags & JOURNAL_INVALID) { return EINVAL; } tr = jnl->active_tr; CHECK_TRANSACTION(tr); if (jnl->owner != current_thread()) { panic("jnl: modify_block_end: called w/out a transaction! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } bflags = buf_flags(bp); if ( !(bflags & B_LOCKED)) panic("jnl: modify_block_end: called with bp not B_LOCKED"); /* * bp must be BL_BUSY and B_LOCKED * first check if it's already part of this transaction */ for (blhdr = tr->blhdr; blhdr; blhdr = (block_list_header *)((long)blhdr->binfo[0].bnum)) { for (i = 1; i < blhdr->num_blocks; i++) { if (bp == blhdr->binfo[i].u.bp) { vnode_t vp; buf_clearflags(bp, B_LOCKED); // this undoes the vnode_ref() in journal_modify_block_end() vp = buf_vnode(bp); vnode_rele_ext(vp, 0, 1); // if the block has the DELWRI and FILTER bits sets, then // things are seriously weird. if it was part of another // transaction then journal_modify_block_start() should // have force it to be written. // //if ((bflags & B_DELWRI) && (bflags & B_FILTER)) { // panic("jnl: kill block: this defies all logic! bp 0x%x\n", bp); //} else { tr->num_killed += buf_size(bp); //} blhdr->binfo[i].bnum = (off_t)-1; blhdr->binfo[i].u.bp = NULL; blhdr->binfo[i].u.bi.bsize = buf_size(bp); buf_markinvalid(bp); buf_brelse(bp); break; } } if (i < blhdr->num_blocks) { break; } } return 0; } /* ;________________________________________________________________________________ ; ; Routine: journal_trim_set_callback ; ; Function: Provide the journal with a routine to be called back when a ; TRIM has (or would have) been issued to the device. That ; is, the transaction has been flushed to the device, and the ; blocks freed by the transaction are now safe for reuse. ; ; CAUTION: If the journal becomes invalid (eg., due to an I/O ; error when trying to write to the journal), this callback ; will stop getting called, even if extents got freed before ; the journal became invalid! ; ; Input Arguments: ; jnl - The journal structure for the filesystem. ; callback - The function to call when the TRIM is complete. ; arg - An argument to be passed to callback. ;________________________________________________________________________________ */ __private_extern__ void journal_trim_set_callback(journal *jnl, jnl_trim_callback_t callback, void *arg) { jnl->trim_callback = callback; jnl->trim_callback_arg = arg; } /* ;________________________________________________________________________________ ; ; Routine: journal_trim_realloc ; ; Function: Increase the amount of memory allocated for the list of extents ; to be unmapped (trimmed). This routine will be called when ; adding an extent to the list, and the list already occupies ; all of the space allocated to it. This routine returns ENOMEM ; if unable to allocate more space, or 0 if the extent list was ; grown successfully. ; ; Input Arguments: ; trim - The trim list to be resized. ; ; Output: ; (result) - ENOMEM or 0. ; ; Side effects: ; The allocated_count and extents fields of tr->trim are updated ; if the function returned 0. ;________________________________________________________________________________ */ static int trim_realloc(journal *jnl, struct jnl_trim_list *trim) { void *new_extents; uint32_t new_allocated_count; boolean_t was_vm_privileged; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REALLOC | DBG_FUNC_START, VM_KERNEL_ADDRPERM(trim), 0, trim->allocated_count, trim->extent_count, 0); new_allocated_count = trim->allocated_count + JOURNAL_DEFAULT_TRIM_EXTENTS; if (jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) { /* * if we block waiting for memory, and there is enough pressure to * cause us to try and create a new swap file, we may end up deadlocking * due to waiting for the journal on the swap file creation path... * by making ourselves vm_privileged, we give ourselves the best chance * of not blocking */ was_vm_privileged = set_vm_privilege(TRUE); } new_extents = kalloc(new_allocated_count * sizeof(dk_extent_t)); if ((jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) && (was_vm_privileged == FALSE)) set_vm_privilege(FALSE); if (new_extents == NULL) { printf("jnl: trim_realloc: unable to grow extent list!\n"); /* * Since we could be called when allocating space previously marked * to be trimmed, we need to empty out the list to be safe. */ trim->extent_count = 0; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REALLOC | DBG_FUNC_END, ENOMEM, 0, trim->allocated_count, 0, 0); return ENOMEM; } /* Copy the old extent list to the newly allocated list. */ if (trim->extents != NULL) { memmove(new_extents, trim->extents, trim->allocated_count * sizeof(dk_extent_t)); kfree(trim->extents, trim->allocated_count * sizeof(dk_extent_t)); } trim->allocated_count = new_allocated_count; trim->extents = new_extents; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REALLOC | DBG_FUNC_END, 0, 0, new_allocated_count, trim->extent_count, 0); return 0; } /* ;________________________________________________________________________________ ; ; Routine: trim_search_extent ; ; Function: Search the given extent list to see if any of its extents ; overlap the given extent. ; ; Input Arguments: ; trim - The trim list to be searched. ; offset - The first byte of the range to be searched for. ; length - The number of bytes of the extent being searched for. ; overlap_start - start of the overlapping extent ; overlap_len - length of the overlapping extent ; ; Output: ; (result) - TRUE if one or more extents overlap, FALSE otherwise. ;________________________________________________________________________________ */ static int trim_search_extent(struct jnl_trim_list *trim, uint64_t offset, uint64_t length, uint64_t *overlap_start, uint64_t *overlap_len) { uint64_t end = offset + length; uint32_t lower = 0; /* Lowest index to search */ uint32_t upper = trim->extent_count; /* Highest index to search + 1 */ uint32_t middle; /* A binary search over the extent list. */ while (lower < upper) { middle = (lower + upper) / 2; if (trim->extents[middle].offset >= end) upper = middle; else if (trim->extents[middle].offset + trim->extents[middle].length <= offset) lower = middle + 1; else { if (overlap_start) { *overlap_start = trim->extents[middle].offset; } if (overlap_len) { *overlap_len = trim->extents[middle].length; } return TRUE; } } return FALSE; } /* ;________________________________________________________________________________ ; ; Routine: journal_trim_add_extent ; ; Function: Keep track of extents that have been freed as part of this ; transaction. If the underlying device supports TRIM (UNMAP), ; then those extents will be trimmed/unmapped once the ; transaction has been written to the journal. (For example, ; SSDs can support trim/unmap and avoid having to recopy those ; blocks when doing wear leveling, and may reuse the same ; phsyical blocks for different logical blocks.) ; ; HFS also uses this, in combination with journal_trim_set_callback, ; to add recently freed extents to its free extent cache, but ; only after the transaction that freed them is committed to ; disk. (This reduces the chance of overwriting live data in ; a way that causes data loss if a transaction never gets ; written to the journal.) ; ; Input Arguments: ; jnl - The journal for the volume containing the byte range. ; offset - The first byte of the range to be trimmed. ; length - The number of bytes of the extent being trimmed. ;________________________________________________________________________________ */ __private_extern__ int journal_trim_add_extent(journal *jnl, uint64_t offset, uint64_t length) { uint64_t end; transaction *tr; dk_extent_t *extent; uint32_t insert_index; uint32_t replace_count; CHECK_JOURNAL(jnl); /* TODO: Is it OK to manipulate the trim list even if JOURNAL_INVALID is set? I think so... */ if (jnl->flags & JOURNAL_INVALID) { return EINVAL; } tr = jnl->active_tr; CHECK_TRANSACTION(tr); if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_ADD | DBG_FUNC_START, VM_KERNEL_ADDRPERM(jnl), offset, length, tr->trim.extent_count, 0); if (jnl->owner != current_thread()) { panic("jnl: trim_add_extent: called w/out a transaction! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } free_old_stuff(jnl); end = offset + length; /* * Find the range of existing extents that can be combined with the * input extent. We start by counting the number of extents that end * strictly before the input extent, then count the number of extents * that overlap or are contiguous with the input extent. */ extent = tr->trim.extents; insert_index = 0; while (insert_index < tr->trim.extent_count && extent->offset + extent->length < offset) { ++insert_index; ++extent; } replace_count = 0; while (insert_index + replace_count < tr->trim.extent_count && extent->offset <= end) { ++replace_count; ++extent; } /* * If none of the existing extents can be combined with the input extent, * then just insert it in the list (before item number insert_index). */ if (replace_count == 0) { /* If the list was already full, we need to grow it. */ if (tr->trim.extent_count == tr->trim.allocated_count) { if (trim_realloc(jnl, &tr->trim) != 0) { printf("jnl: trim_add_extent: out of memory!"); if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_ADD | DBG_FUNC_END, ENOMEM, 0, 0, tr->trim.extent_count, 0); return ENOMEM; } } /* Shift any existing extents with larger offsets. */ if (insert_index < tr->trim.extent_count) { memmove(&tr->trim.extents[insert_index+1], &tr->trim.extents[insert_index], (tr->trim.extent_count - insert_index) * sizeof(dk_extent_t)); } tr->trim.extent_count++; /* Store the new extent in the list. */ tr->trim.extents[insert_index].offset = offset; tr->trim.extents[insert_index].length = length; /* We're done. */ if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_ADD | DBG_FUNC_END, 0, 0, 0, tr->trim.extent_count, 0); return 0; } /* * Update extent number insert_index to be the union of the input extent * and all of the replaced extents. */ if (tr->trim.extents[insert_index].offset < offset) offset = tr->trim.extents[insert_index].offset; extent = &tr->trim.extents[insert_index + replace_count - 1]; if (extent->offset + extent->length > end) end = extent->offset + extent->length; tr->trim.extents[insert_index].offset = offset; tr->trim.extents[insert_index].length = end - offset; /* * If we were replacing more than one existing extent, then shift any * extents with larger offsets, and update the count of extents. * * We're going to leave extent #insert_index alone since it was just updated, above. * We need to move extents from index (insert_index + replace_count) through the end of * the list by (replace_count - 1) positions so that they overwrite extent #(insert_index + 1). */ if (replace_count > 1 && (insert_index + replace_count) < tr->trim.extent_count) { memmove(&tr->trim.extents[insert_index + 1], &tr->trim.extents[insert_index + replace_count], (tr->trim.extent_count - insert_index - replace_count) * sizeof(dk_extent_t)); } tr->trim.extent_count -= replace_count - 1; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_ADD | DBG_FUNC_END, 0, 0, 0, tr->trim.extent_count, 0); return 0; } /* * journal_trim_extent_overlap * * Return 1 if there are any pending TRIMs that overlap with the given offset and length * Return 0 otherwise. */ int journal_trim_extent_overlap (journal *jnl, uint64_t offset, uint64_t length, uint64_t *end) { transaction *tr = NULL; int overlap = 0; uint64_t overlap_start; uint64_t overlap_len; tr = jnl->active_tr; CHECK_TRANSACTION(tr); /* * There are two lists that need to be examined for potential overlaps: * * The first is the current transaction. Since this function requires that * a transaction be active when this is called, this is the "active_tr" * pointer in the journal struct. This has a trimlist pointer which needs * to be searched. */ overlap = trim_search_extent (&tr->trim, offset, length, &overlap_start, &overlap_len); if (overlap == 0) { /* * The second is the async trim list, which is only done if the current * transaction group (active transaction) did not overlap with our target * extent. This async trim list is the set of all previously * committed transaction groups whose I/Os are now in-flight. We need to hold the * trim lock in order to search this list. If we grab the list before the * TRIM has completed, then we will compare it. If it is grabbed AFTER the * TRIM has completed, then the pointer will be zeroed out and we won't have * to check anything. */ lck_rw_lock_shared (&jnl->trim_lock); if (jnl->async_trim != NULL) { overlap = trim_search_extent(jnl->async_trim, offset, length, &overlap_start, &overlap_len); } lck_rw_unlock_shared (&jnl->trim_lock); } if (overlap) { /* compute the end (min) of the overlapping range */ if ( (overlap_start + overlap_len) < (offset + length)) { *end = (overlap_start + overlap_len); } else { *end = (offset + length); } } return overlap; } /* * journal_request_immediate_flush * * FS requests that the journal flush immediately upon the * active transaction's completion. * * Returns 0 if operation succeeds * Returns EPERM if we failed to leave hint */ int journal_request_immediate_flush (journal *jnl) { transaction *tr = NULL; /* * Is a transaction still in process? You must do * this while there are txns open */ tr = jnl->active_tr; if (tr != NULL) { CHECK_TRANSACTION(tr); tr->flush_on_completion = TRUE; } else { return EPERM; } return 0; } /* ;________________________________________________________________________________ ; ; Routine: trim_remove_extent ; ; Function: Indicate that a range of bytes, some of which may have previously ; been passed to journal_trim_add_extent, is now allocated. ; Any overlapping ranges currently in the journal's trim list will ; be removed. If the underlying device supports TRIM (UNMAP), then ; these extents will not be trimmed/unmapped when the transaction ; is written to the journal. ; ; HFS also uses this to prevent newly allocated space from being ; added to its free extent cache (if some portion of the newly ; allocated space was recently freed). ; ; Input Arguments: ; trim - The trim list to update. ; offset - The first byte of the range to be trimmed. ; length - The number of bytes of the extent being trimmed. ;________________________________________________________________________________ */ static int trim_remove_extent(journal *jnl, struct jnl_trim_list *trim, uint64_t offset, uint64_t length) { u_int64_t end; dk_extent_t *extent; u_int32_t keep_before; u_int32_t keep_after; end = offset + length; /* * Find any existing extents that start before or end after the input * extent. These extents will be modified if they overlap the input * extent. Other extents between them will be deleted. */ extent = trim->extents; keep_before = 0; while (keep_before < trim->extent_count && extent->offset < offset) { ++keep_before; ++extent; } keep_after = keep_before; if (keep_after > 0) { /* See if previous extent extends beyond both ends of input extent. */ --keep_after; --extent; } while (keep_after < trim->extent_count && (extent->offset + extent->length) <= end) { ++keep_after; ++extent; } /* * When we get here, the first keep_before extents (0 .. keep_before-1) * start before the input extent, and extents (keep_after .. extent_count-1) * end after the input extent. We'll need to keep, all of those extents, * but possibly modify #(keep_before-1) and #keep_after to remove the portion * that overlaps with the input extent. */ /* * Does the input extent start after and end before the same existing * extent? If so, we have to "punch a hole" in that extent and convert * it to two separate extents. */ if (keep_before > keep_after) { /* If the list was already full, we need to grow it. */ if (trim->extent_count == trim->allocated_count) { if (trim_realloc(jnl, trim) != 0) { printf("jnl: trim_remove_extent: out of memory!"); return ENOMEM; } } /* * Make room for a new extent by shifting extents #keep_after and later * down by one extent. When we're done, extents #keep_before and * #keep_after will be identical, and we can fall through to removing * the portion that overlaps the input extent. */ memmove(&trim->extents[keep_before], &trim->extents[keep_after], (trim->extent_count - keep_after) * sizeof(dk_extent_t)); ++trim->extent_count; ++keep_after; /* * Fall through. We now have the case where the length of extent * #(keep_before - 1) needs to be updated, and the start of extent * #(keep_after) needs to be updated. */ } /* * May need to truncate the end of extent #(keep_before - 1) if it overlaps * the input extent. */ if (keep_before > 0) { extent = &trim->extents[keep_before - 1]; if (extent->offset + extent->length > offset) { extent->length = offset - extent->offset; } } /* * May need to update the start of extent #(keep_after) if it overlaps the * input extent. */ if (keep_after < trim->extent_count) { extent = &trim->extents[keep_after]; if (extent->offset < end) { extent->length = extent->offset + extent->length - end; extent->offset = end; } } /* * If there were whole extents that overlapped the input extent, get rid * of them by shifting any following extents, and updating the count. */ if (keep_after > keep_before && keep_after < trim->extent_count) { memmove(&trim->extents[keep_before], &trim->extents[keep_after], (trim->extent_count - keep_after) * sizeof(dk_extent_t)); } trim->extent_count -= keep_after - keep_before; return 0; } /* ;________________________________________________________________________________ ; ; Routine: journal_trim_remove_extent ; ; Function: Make note of a range of bytes, some of which may have previously ; been passed to journal_trim_add_extent, is now in use on the ; volume. The given bytes will be not be trimmed as part of ; this transaction, or a pending trim of a transaction being ; asynchronously flushed. ; ; Input Arguments: ; jnl - The journal for the volume containing the byte range. ; offset - The first byte of the range to be trimmed. ; length - The number of bytes of the extent being trimmed. ;________________________________________________________________________________ */ __private_extern__ int journal_trim_remove_extent(journal *jnl, uint64_t offset, uint64_t length) { int error = 0; transaction *tr; CHECK_JOURNAL(jnl); /* TODO: Is it OK to manipulate the trim list even if JOURNAL_INVALID is set? I think so... */ if (jnl->flags & JOURNAL_INVALID) { return EINVAL; } tr = jnl->active_tr; CHECK_TRANSACTION(tr); if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REMOVE | DBG_FUNC_START, VM_KERNEL_ADDRPERM(jnl), offset, length, tr->trim.extent_count, 0); if (jnl->owner != current_thread()) { panic("jnl: trim_remove_extent: called w/out a transaction! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } free_old_stuff(jnl); error = trim_remove_extent(jnl, &tr->trim, offset, length); if (error == 0) { int found = FALSE; /* * See if a pending trim has any extents that overlap with the * one we were given. */ lck_rw_lock_shared(&jnl->trim_lock); if (jnl->async_trim != NULL) found = trim_search_extent(jnl->async_trim, offset, length, NULL, NULL); lck_rw_unlock_shared(&jnl->trim_lock); if (found) { /* * There was an overlap, so avoid trimming the extent we * just allocated. (Otherwise, it might get trimmed after * we've written to it, which will cause that data to be * corrupted.) */ uint32_t async_extent_count = 0; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REMOVE_PENDING | DBG_FUNC_START, VM_KERNEL_ADDRPERM(jnl), offset, length, 0, 0); lck_rw_lock_exclusive(&jnl->trim_lock); if (jnl->async_trim != NULL) { error = trim_remove_extent(jnl, jnl->async_trim, offset, length); async_extent_count = jnl->async_trim->extent_count; } lck_rw_unlock_exclusive(&jnl->trim_lock); if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REMOVE_PENDING | DBG_FUNC_END, error, 0, 0, async_extent_count, 0); } } if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_REMOVE | DBG_FUNC_END, error, 0, 0, tr->trim.extent_count, 0); return error; } static int journal_trim_flush(journal *jnl, transaction *tr) { int errno = 0; boolean_t was_vm_privileged; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_FLUSH | DBG_FUNC_START, VM_KERNEL_ADDRPERM(jnl), tr, 0, tr->trim.extent_count, 0); if (jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) { /* * the disk driver can allocate memory on this path... * if we block waiting for memory, and there is enough pressure to * cause us to try and create a new swap file, we may end up deadlocking * due to waiting for the journal on the swap file creation path... * by making ourselves vm_privileged, we give ourselves the best chance * of not blocking */ was_vm_privileged = set_vm_privilege(TRUE); } lck_rw_lock_shared(&jnl->trim_lock); if (tr->trim.extent_count > 0) { dk_unmap_t unmap; bzero(&unmap, sizeof(unmap)); if (CONFIG_HFS_TRIM && (jnl->flags & JOURNAL_USE_UNMAP)) { unmap.extents = tr->trim.extents; unmap.extentsCount = tr->trim.extent_count; if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_UNMAP | DBG_FUNC_START, VM_KERNEL_ADDRPERM(jnl), tr, 0, tr->trim.extent_count, 0); errno = VNOP_IOCTL(jnl->fsdev, DKIOCUNMAP, (caddr_t)&unmap, FWRITE, vfs_context_kernel()); if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_UNMAP | DBG_FUNC_END, errno, 0, 0, 0, 0); } /* * Call back into the file system to tell them that we have * trimmed some extents and that they can now be reused. * * CAUTION: If the journal becomes invalid (eg., due to an I/O * error when trying to write to the journal), this callback * will stop getting called, even if extents got freed before * the journal became invalid! */ if (jnl->trim_callback) jnl->trim_callback(jnl->trim_callback_arg, tr->trim.extent_count, tr->trim.extents); } lck_rw_unlock_shared(&jnl->trim_lock); if ((jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) && (was_vm_privileged == FALSE)) set_vm_privilege(FALSE); /* * If the transaction we're flushing was the async transaction, then * tell the current transaction that there is no pending trim * any more. * * NOTE: Since we released the lock, another thread could have * removed one or more extents from our list. That's not a * problem since any writes to the re-allocated blocks * would get sent to the device after the DKIOCUNMAP. */ lck_rw_lock_exclusive(&jnl->trim_lock); if (jnl->async_trim == &tr->trim) jnl->async_trim = NULL; lck_rw_unlock_exclusive(&jnl->trim_lock); /* * By the time we get here, no other thread can discover the address * of "tr", so it is safe for us to manipulate tr->trim without * holding any locks. */ if (tr->trim.extents) { kfree(tr->trim.extents, tr->trim.allocated_count * sizeof(dk_extent_t)); tr->trim.allocated_count = 0; tr->trim.extent_count = 0; tr->trim.extents = NULL; } if (jnl_kdebug) KERNEL_DEBUG_CONSTANT(DBG_JOURNAL_TRIM_FLUSH | DBG_FUNC_END, errno, 0, 0, 0, 0); return errno; } static int journal_binfo_cmp(const void *a, const void *b) { const block_info *bi_a = (const struct block_info *)a; const block_info *bi_b = (const struct block_info *)b; daddr64_t res; if (bi_a->bnum == (off_t)-1) { return 1; } if (bi_b->bnum == (off_t)-1) { return -1; } // don't have to worry about negative block // numbers so this is ok to do. // res = (buf_blkno(bi_a->u.bp) - buf_blkno(bi_b->u.bp)); return (int)res; } /* * End a transaction. If the transaction is small enough, and we're not forcing * a write to disk, the "active" transaction becomes the "current" transaction, * and will be reused for the next transaction that is started (group commit). * * If the transaction gets written to disk (because force_it is true, or no * group commit, or the transaction is sufficiently full), the blocks get * written into the journal first, then the are written asynchronously. When * those async writes complete, the transaction can be freed and removed from * the journal. * * An optional callback can be supplied. If given, it is called after the * the blocks have been written to the journal, but before the async writes * of those blocks to their normal on-disk locations. This is used by * journal_relocate so that the location of the journal can be changed and * flushed to disk before the blocks get written to their normal locations. * Note that the callback is only called if the transaction gets written to * the journal during this end_transaction call; you probably want to set the * force_it flag. * * Inputs: * tr Transaction to add to the journal * force_it If true, force this transaction to the on-disk journal immediately. * callback See description above. Pass NULL for no callback. * callback_arg Argument passed to callback routine. * * Result * 0 No errors * -1 An error occurred. The journal is marked invalid. */ static int end_transaction(transaction *tr, int force_it, errno_t (*callback)(void*), void *callback_arg, boolean_t drop_lock, boolean_t must_wait) { block_list_header *blhdr=NULL, *next=NULL; int i, ret_val = 0; errno_t errno; journal *jnl = tr->jnl; struct buf *bp; size_t tbuffer_offset; boolean_t drop_lock_early; if (jnl->cur_tr) { panic("jnl: jnl @ %p already has cur_tr %p, new tr: %p\n", jnl, jnl->cur_tr, tr); } // if there weren't any modified blocks in the transaction // just save off the transaction pointer and return. if (tr->total_bytes == jnl->jhdr->blhdr_size) { jnl->cur_tr = tr; goto done; } // if our transaction buffer isn't very full, just hang // on to it and don't actually flush anything. this is // what is known as "group commit". we will flush the // transaction buffer if it's full or if we have more than // one of them so we don't start hogging too much memory. // // We also check the device supports UNMAP/TRIM, and if so, // the number of extents waiting to be trimmed. If it is // small enough, then keep accumulating more (so we can // reduce the overhead of trimming). If there was a prior // trim error, then we stop issuing trims for this // volume, so we can also coalesce transactions. // if ( force_it == 0 && (jnl->flags & JOURNAL_NO_GROUP_COMMIT) == 0 && tr->num_blhdrs < 3 && (tr->total_bytes <= ((tr->tbuffer_size*tr->num_blhdrs) - tr->tbuffer_size/8)) && (!(jnl->flags & JOURNAL_USE_UNMAP) || (tr->trim.extent_count < jnl_trim_flush_limit))) { jnl->cur_tr = tr; goto done; } KERNEL_DEBUG(0xbbbbc018|DBG_FUNC_START, jnl, tr, drop_lock, must_wait, 0); lock_condition(jnl, &jnl->flushing, "end_transaction"); /* * if the previous 'finish_end_transaction' was being run * asynchronously, it could have encountered a condition * that caused it to mark the journal invalid... if that * occurred while we were waiting for it to finish, we * need to notice and abort the current transaction */ if ((jnl->flags & JOURNAL_INVALID) || jnl->flush_aborted == TRUE) { unlock_condition(jnl, &jnl->flushing); abort_transaction(jnl, tr); ret_val = -1; KERNEL_DEBUG(0xbbbbc018|DBG_FUNC_END, jnl, tr, ret_val, 0, 0); goto done; } /* * Store a pointer to this transaction's trim list so that * future transactions can find it. * * Note: if there are no extents in the trim list, then don't * bother saving the pointer since nothing can add new extents * to the list (and other threads/transactions only care if * there is a trim pending). */ lck_rw_lock_exclusive(&jnl->trim_lock); if (jnl->async_trim != NULL) panic("jnl: end_transaction: async_trim already non-NULL!"); if (tr->trim.extent_count > 0) jnl->async_trim = &tr->trim; lck_rw_unlock_exclusive(&jnl->trim_lock); /* * snapshot the transaction sequence number while we are still behind * the journal lock since it will be bumped upon the start of the * next transaction group which may overlap the current journal flush... * we pass the snapshot into write_journal_header during the journal * flush so that it can write the correct version in the header... * because we hold the 'flushing' condition variable for the duration * of the journal flush, 'saved_sequence_num' remains stable */ jnl->saved_sequence_num = jnl->sequence_num; /* * if we're here we're going to flush the transaction buffer to disk. * 'check_free_space' will not return untl there is enough free * space for this transaction in the journal and jnl->old_start[0] * is avaiable for use */ KERNEL_DEBUG(0xbbbbc030 | DBG_FUNC_START, jnl, 0, 0, 0, 0); check_free_space(jnl, tr->total_bytes, &tr->delayed_header_write, jnl->saved_sequence_num); KERNEL_DEBUG(0xbbbbc030 | DBG_FUNC_END, jnl, tr->delayed_header_write, 0, 0, 0); // range check the end index if (jnl->jhdr->end <= 0 || jnl->jhdr->end > jnl->jhdr->size) { panic("jnl: end_transaction: end is bogus 0x%llx (sz 0x%llx)\n", jnl->jhdr->end, jnl->jhdr->size); } if (tr->delayed_header_write == TRUE) { thread_t thread = THREAD_NULL; lock_condition(jnl, &jnl->writing_header, "end_transaction"); /* * fire up a thread to write the journal header * asynchronously... when it finishes, it will call * unlock_condition... we can overlap the preparation of * the log and buffers during this time */ kernel_thread_start((thread_continue_t)write_header_thread, jnl, &thread); } else jnl->write_header_failed = FALSE; // this transaction starts where the current journal ends tr->journal_start = jnl->jhdr->end; lock_oldstart(jnl); /* * Because old_start is locked above, we can cast away the volatile qualifier before passing it to memcpy. * slide everyone else down and put our latest guy in the last * entry in the old_start array */ memcpy(__CAST_AWAY_QUALIFIER(&jnl->old_start[0], volatile, void *), __CAST_AWAY_QUALIFIER(&jnl->old_start[1], volatile, void *), sizeof(jnl->old_start)-sizeof(jnl->old_start[0])); jnl->old_start[sizeof(jnl->old_start)/sizeof(jnl->old_start[0]) - 1] = tr->journal_start | 0x8000000000000000LL; unlock_oldstart(jnl); for (blhdr = tr->blhdr; blhdr; blhdr = next) { char *blkptr; buf_t sbp; int32_t bsize; tbuffer_offset = jnl->jhdr->blhdr_size; for (i = 1; i < blhdr->num_blocks; i++) { if (blhdr->binfo[i].bnum != (off_t)-1) { void (*func)(buf_t, void *); void *arg; bp = blhdr->binfo[i].u.bp; if (bp == NULL) { panic("jnl: inconsistent binfo (NULL bp w/bnum %lld; jnl @ %p, tr %p)\n", blhdr->binfo[i].bnum, jnl, tr); } /* * acquire the bp here so that we can safely * mess around with its data. buf_acquire() * will return EAGAIN if the buffer was busy, * so loop trying again. */ do { errno = buf_acquire(bp, BAC_REMOVE, 0, 0); } while (errno == EAGAIN); if (errno) panic("could not acquire bp %p (err %d)\n", bp, errno); if ((buf_flags(bp) & (B_LOCKED|B_DELWRI)) != (B_LOCKED|B_DELWRI)) { if (jnl->flags & JOURNAL_CLOSE_PENDING) { buf_clearflags(bp, B_LOCKED); buf_brelse(bp); /* * this is an odd case that appears to happen occasionally * make sure we mark this block as no longer valid * so that we don't process it in "finish_end_transaction" since * the bp that is recorded in our array no longer belongs * to us (normally we substitute a shadow bp to be processed * issuing a 'buf_bawrite' on a stale buf_t pointer leads * to all kinds of problems. */ blhdr->binfo[i].bnum = (off_t)-1; continue; } else { panic("jnl: end_tr: !!!DANGER!!! bp %p flags (0x%x) not LOCKED & DELWRI\n", bp, buf_flags(bp)); } } bsize = buf_size(bp); buf_setfilter(bp, NULL, NULL, &func, &arg); blkptr = (char *)&((char *)blhdr)[tbuffer_offset]; sbp = buf_create_shadow_priv(bp, FALSE, (uintptr_t)blkptr, 0, 0); if (sbp == NULL) panic("jnl: buf_create_shadow returned NULL"); /* * copy the data into the transaction buffer... */ memcpy(blkptr, (char *)buf_dataptr(bp), bsize); buf_clearflags(bp, B_LOCKED); buf_markclean(bp); buf_drop(bp); /* * adopt the shadow buffer for this block */ if (func) { /* * transfer FS hook function to the * shadow buffer... it will get called * in finish_end_transaction */ buf_setfilter(sbp, func, arg, NULL, NULL); } blhdr->binfo[i].u.bp = sbp; } else { // bnum == -1, only true if a block was "killed" bsize = blhdr->binfo[i].u.bi.bsize; } tbuffer_offset += bsize; } next = (block_list_header *)((long)blhdr->binfo[0].bnum); } /* * if callback != NULL, we don't want to drop the journal * lock, or complete end_transaction asynchronously, since * the caller is expecting the callback to run in the calling * context * * if drop_lock == FALSE, we can't complete end_transaction * asynchronously */ if (callback) drop_lock_early = FALSE; else drop_lock_early = drop_lock; if (drop_lock_early == FALSE) must_wait = TRUE; if (drop_lock_early == TRUE) { journal_unlock(jnl); drop_lock = FALSE; } if (must_wait == TRUE) ret_val = finish_end_transaction(tr, callback, callback_arg); else { thread_t thread = THREAD_NULL; /* * fire up a thread to complete processing this transaction * asynchronously... when it finishes, it will call * unlock_condition */ kernel_thread_start((thread_continue_t)finish_end_thread, tr, &thread); } KERNEL_DEBUG(0xbbbbc018|DBG_FUNC_END, jnl, tr, ret_val, 0, 0); done: if (drop_lock == TRUE) { journal_unlock(jnl); } return (ret_val); } static void finish_end_thread(transaction *tr) { proc_set_task_policy(current_task(), current_thread(), TASK_POLICY_INTERNAL, TASK_POLICY_IOPOL, IOPOL_PASSIVE); finish_end_transaction(tr, NULL, NULL); thread_deallocate(current_thread()); thread_terminate(current_thread()); } static void write_header_thread(journal *jnl) { proc_set_task_policy(current_task(), current_thread(), TASK_POLICY_INTERNAL, TASK_POLICY_IOPOL, IOPOL_PASSIVE); if (write_journal_header(jnl, 1, jnl->saved_sequence_num)) jnl->write_header_failed = TRUE; else jnl->write_header_failed = FALSE; unlock_condition(jnl, &jnl->writing_header); thread_deallocate(current_thread()); thread_terminate(current_thread()); } static int finish_end_transaction(transaction *tr, errno_t (*callback)(void*), void *callback_arg) { int i, amt; int ret = 0; off_t end; journal *jnl = tr->jnl; buf_t bp, *bparray; vnode_t vp; block_list_header *blhdr=NULL, *next=NULL; size_t tbuffer_offset; int bufs_written = 0; int ret_val = 0; KERNEL_DEBUG(0xbbbbc028|DBG_FUNC_START, jnl, tr, 0, 0, 0); end = jnl->jhdr->end; for (blhdr = tr->blhdr; blhdr; blhdr = (block_list_header *)((long)blhdr->binfo[0].bnum)) { boolean_t was_vm_privileged; amt = blhdr->bytes_used; blhdr->binfo[0].u.bi.b.sequence_num = tr->sequence_num; blhdr->checksum = 0; blhdr->checksum = calc_checksum((char *)blhdr, BLHDR_CHECKSUM_SIZE); if (jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) { /* * if we block waiting for memory, and there is enough pressure to * cause us to try and create a new swap file, we may end up deadlocking * due to waiting for the journal on the swap file creation path... * by making ourselves vm_privileged, we give ourselves the best chance * of not blocking */ was_vm_privileged = set_vm_privilege(TRUE); } if (kmem_alloc(kernel_map, (vm_offset_t *)&bparray, blhdr->num_blocks * sizeof(struct buf *))) { panic("can't allocate %zd bytes for bparray\n", blhdr->num_blocks * sizeof(struct buf *)); } if ((jnl->fsmount->mnt_kern_flag & MNTK_SWAP_MOUNT) && (was_vm_privileged == FALSE)) set_vm_privilege(FALSE); tbuffer_offset = jnl->jhdr->blhdr_size; for (i = 1; i < blhdr->num_blocks; i++) { void (*func)(buf_t, void *); void *arg; int32_t bsize; /* * finish preparing the shadow buf_t before * calculating the individual block checksums */ if (blhdr->binfo[i].bnum != (off_t)-1) { daddr64_t blkno; daddr64_t lblkno; bp = blhdr->binfo[i].u.bp; vp = buf_vnode(bp); blkno = buf_blkno(bp); lblkno = buf_lblkno(bp); if (vp == NULL && lblkno == blkno) { printf("jnl: %s: end_tr: bad news! bp @ %p w/null vp and l/blkno = %qd/%qd. aborting the transaction (tr %p jnl %p).\n", jnl->jdev_name, bp, lblkno, blkno, tr, jnl); ret_val = -1; goto bad_journal; } // if the lblkno is the same as blkno and this bp isn't // associated with the underlying file system device then // we need to call bmap() to get the actual physical block. // if ((lblkno == blkno) && (vp != jnl->fsdev)) { off_t f_offset; size_t contig_bytes; if (VNOP_BLKTOOFF(vp, lblkno, &f_offset)) { printf("jnl: %s: end_tr: vnop_blktooff failed @ %p, jnl %p\n", jnl->jdev_name, bp, jnl); ret_val = -1; goto bad_journal; } if (VNOP_BLOCKMAP(vp, f_offset, buf_count(bp), &blkno, &contig_bytes, NULL, 0, NULL)) { printf("jnl: %s: end_tr: can't blockmap the bp @ %p, jnl %p\n", jnl->jdev_name, bp, jnl); ret_val = -1; goto bad_journal; } if ((uint32_t)contig_bytes < buf_count(bp)) { printf("jnl: %s: end_tr: blk not physically contiguous on disk@ %p, jnl %p\n", jnl->jdev_name, bp, jnl); ret_val = -1; goto bad_journal; } buf_setblkno(bp, blkno); } // update this so we write out the correct physical block number! blhdr->binfo[i].bnum = (off_t)(blkno); /* * pick up the FS hook function (if any) and prepare * to fire this buffer off in the next pass */ buf_setfilter(bp, buffer_flushed_callback, tr, &func, &arg); if (func) { /* * call the hook function supplied by the filesystem... * this needs to happen BEFORE cacl_checksum in case * the FS morphs the data in the buffer */ func(bp, arg); } bparray[i] = bp; bsize = buf_size(bp); blhdr->binfo[i].u.bi.bsize = bsize; blhdr->binfo[i].u.bi.b.cksum = calc_checksum(&((char *)blhdr)[tbuffer_offset], bsize); } else { bparray[i] = NULL; bsize = blhdr->binfo[i].u.bi.bsize; blhdr->binfo[i].u.bi.b.cksum = 0; } tbuffer_offset += bsize; } /* * if we fired off the journal_write_header asynchronously in * 'end_transaction', we need to wait for its completion * before writing the actual journal data */ wait_condition(jnl, &jnl->writing_header, "finish_end_transaction"); if (jnl->write_header_failed == FALSE) ret = write_journal_data(jnl, &end, blhdr, amt); else ret_val = -1; /* * put the bp pointers back so that we can * make the final pass on them */ for (i = 1; i < blhdr->num_blocks; i++) blhdr->binfo[i].u.bp = bparray[i]; kmem_free(kernel_map, (vm_offset_t)bparray, blhdr->num_blocks * sizeof(struct buf *)); if (ret_val == -1) goto bad_journal; if (ret != amt) { printf("jnl: %s: end_transaction: only wrote %d of %d bytes to the journal!\n", jnl->jdev_name, ret, amt); ret_val = -1; goto bad_journal; } } jnl->jhdr->end = end; // update where the journal now ends tr->journal_end = end; // the transaction ends here too if (tr->journal_start == 0 || tr->journal_end == 0) { panic("jnl: end_transaction: bad tr journal start/end: 0x%llx 0x%llx\n", tr->journal_start, tr->journal_end); } if (write_journal_header(jnl, 0, jnl->saved_sequence_num) != 0) { ret_val = -1; goto bad_journal; } /* * If the caller supplied a callback, call it now that the blocks have been * written to the journal. This is used by journal_relocate so, for example, * the file system can change its pointer to the new journal. */ if (callback != NULL && callback(callback_arg) != 0) { ret_val = -1; goto bad_journal; } // // Send a DKIOCUNMAP for the extents trimmed by this transaction, and // free up the extent list. // journal_trim_flush(jnl, tr); // the buffer_flushed_callback will only be called for the // real blocks that get flushed so we have to account for // the block_list_headers here. // tr->num_flushed = tr->num_blhdrs * jnl->jhdr->blhdr_size; lock_condition(jnl, &jnl->asyncIO, "finish_end_transaction"); // // setup for looping through all the blhdr's. // for (blhdr = tr->blhdr; blhdr; blhdr = next) { uint16_t num_blocks; /* * grab this info ahead of issuing the buf_bawrites... * once the last one goes out, its possible for blhdr * to be freed (especially if we get preempted) before * we do the last check of num_blocks or * grab the next blhdr pointer... */ next = (block_list_header *)((long)blhdr->binfo[0].bnum); num_blocks = blhdr->num_blocks; /* * we can re-order the buf ptrs because everything is written out already */ qsort(&blhdr->binfo[1], num_blocks-1, sizeof(block_info), journal_binfo_cmp); /* * need to make sure that the loop issuing the buf_bawrite's * does not touch blhdr once the last buf_bawrite has been * issued... at that point, we no longer have a legitmate * reference on the associated storage since it will be * released upon the completion of that last buf_bawrite */ for (i = num_blocks-1; i >= 1; i--) { if (blhdr->binfo[i].bnum != (off_t)-1) break; num_blocks--; } for (i = 1; i < num_blocks; i++) { if ((bp = blhdr->binfo[i].u.bp)) { vp = buf_vnode(bp); buf_bawrite(bp); // this undoes the vnode_ref() in journal_modify_block_end() vnode_rele_ext(vp, 0, 1); bufs_written++; } } } if (bufs_written == 0) { /* * since we didn't issue any buf_bawrite's, there is no * async trigger to cause the memory associated with this * transaction to be freed... so, move it to the garbage * list now */ lock_oldstart(jnl); tr->next = jnl->tr_freeme; jnl->tr_freeme = tr; unlock_oldstart(jnl); unlock_condition(jnl, &jnl->asyncIO); } //printf("jnl: end_tr: tr @ 0x%x, jnl-blocks: 0x%llx - 0x%llx. exit!\n", // tr, tr->journal_start, tr->journal_end); bad_journal: if (ret_val == -1) { /* * 'flush_aborted' is protected by the flushing condition... we need to * set it before dropping the condition so that it will be * noticed in 'end_transaction'... we add this additional * aborted condition so that we can drop the 'flushing' condition * before grabbing the journal lock... this avoids a deadlock * in 'end_transaction' which is holding the journal lock while * waiting for the 'flushing' condition to clear... * everyone else will notice the JOURNAL_INVALID flag */ jnl->flush_aborted = TRUE; unlock_condition(jnl, &jnl->flushing); journal_lock(jnl); jnl->flags |= JOURNAL_INVALID; jnl->old_start[sizeof(jnl->old_start)/sizeof(jnl->old_start[0]) - 1] &= ~0x8000000000000000LL; abort_transaction(jnl, tr); // cleans up list of extents to be trimmed journal_unlock(jnl); } else unlock_condition(jnl, &jnl->flushing); KERNEL_DEBUG(0xbbbbc028|DBG_FUNC_END, jnl, tr, bufs_written, ret_val, 0); return (ret_val); } static void lock_condition(journal *jnl, boolean_t *condition, const char *condition_name) { KERNEL_DEBUG(0xbbbbc020|DBG_FUNC_START, jnl, condition, 0, 0, 0); lock_flush(jnl); while (*condition == TRUE) msleep(condition, &jnl->flock, PRIBIO, condition_name, NULL); *condition = TRUE; unlock_flush(jnl); KERNEL_DEBUG(0xbbbbc020|DBG_FUNC_END, jnl, condition, 0, 0, 0); } static void wait_condition(journal *jnl, boolean_t *condition, const char *condition_name) { if (*condition == FALSE) return; KERNEL_DEBUG(0xbbbbc02c|DBG_FUNC_START, jnl, condition, 0, 0, 0); lock_flush(jnl); while (*condition == TRUE) msleep(condition, &jnl->flock, PRIBIO, condition_name, NULL); unlock_flush(jnl); KERNEL_DEBUG(0xbbbbc02c|DBG_FUNC_END, jnl, condition, 0, 0, 0); } static void unlock_condition(journal *jnl, boolean_t *condition) { lock_flush(jnl); *condition = FALSE; wakeup(condition); unlock_flush(jnl); } static void abort_transaction(journal *jnl, transaction *tr) { block_list_header *blhdr, *next; // for each block list header, iterate over the blocks then // free up the memory associated with the block list. // // find each of the primary blocks (i.e. the list could // contain a mix of shadowed and real buf_t's depending // on when the abort condition was detected) and mark them // clean and locked in the cache... this at least allows // the FS a consistent view between it's incore data structures // and the meta-data held in the cache // KERNEL_DEBUG(0xbbbbc034|DBG_FUNC_START, jnl, tr, 0, 0, 0); for (blhdr = tr->blhdr; blhdr; blhdr = next) { int i; for (i = 1; i < blhdr->num_blocks; i++) { buf_t bp, tbp, sbp; vnode_t bp_vp; errno_t errno; if (blhdr->binfo[i].bnum == (off_t)-1) continue; tbp = blhdr->binfo[i].u.bp; bp_vp = buf_vnode(tbp); buf_setfilter(tbp, NULL, NULL, NULL, NULL); if (buf_shadow(tbp)) sbp = tbp; else sbp = NULL; if (bp_vp) { errno = buf_meta_bread(bp_vp, buf_lblkno(tbp), buf_size(tbp), NOCRED, &bp); if (errno == 0) { if (sbp == NULL && bp != tbp && (buf_flags(tbp) & B_LOCKED)) { panic("jnl: abort_tr: got back a different bp! (bp %p should be %p, jnl %p\n", bp, tbp, jnl); } /* * once the journal has been marked INVALID and aborted, * NO meta data can be written back to the disk, so * mark the buf_t clean and make sure it's locked in the cache * note: if we found a shadow, the real buf_t needs to be relocked */ buf_setflags(bp, B_LOCKED); buf_markclean(bp); buf_brelse(bp); KERNEL_DEBUG(0xbbbbc034|DBG_FUNC_NONE, jnl, tr, bp, 0, 0); /* * this undoes the vnode_ref() in journal_modify_block_end() */ vnode_rele_ext(bp_vp, 0, 1); } else { printf("jnl: %s: abort_tr: could not find block %lld vp %p!\n", jnl->jdev_name, blhdr->binfo[i].bnum, tbp); if (bp) { buf_brelse(bp); } } } if (sbp) buf_brelse(sbp); } next = (block_list_header *)((long)blhdr->binfo[0].bnum); // we can free blhdr here since we won't need it any more blhdr->binfo[0].bnum = 0xdeadc0de; kmem_free(kernel_map, (vm_offset_t)blhdr, tr->tbuffer_size); } /* * If the transaction we're aborting was the async transaction, then * tell the current transaction that there is no pending trim * any more. */ lck_rw_lock_exclusive(&jnl->trim_lock); if (jnl->async_trim == &tr->trim) jnl->async_trim = NULL; lck_rw_unlock_exclusive(&jnl->trim_lock); if (tr->trim.extents) { kfree(tr->trim.extents, tr->trim.allocated_count * sizeof(dk_extent_t)); } tr->trim.allocated_count = 0; tr->trim.extent_count = 0; tr->trim.extents = NULL; tr->tbuffer = NULL; tr->blhdr = NULL; tr->total_bytes = 0xdbadc0de; FREE_ZONE(tr, sizeof(transaction), M_JNL_TR); KERNEL_DEBUG(0xbbbbc034|DBG_FUNC_END, jnl, tr, 0, 0, 0); } int journal_end_transaction(journal *jnl) { int ret; transaction *tr; CHECK_JOURNAL(jnl); free_old_stuff(jnl); if ((jnl->flags & JOURNAL_INVALID) && jnl->owner == NULL) { return 0; } if (jnl->owner != current_thread()) { panic("jnl: end_tr: I'm not the owner! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } jnl->nested_count--; if (jnl->nested_count > 0) { return 0; } else if (jnl->nested_count < 0) { panic("jnl: jnl @ %p has negative nested count (%d). bad boy.\n", jnl, jnl->nested_count); } if (jnl->flags & JOURNAL_INVALID) { if (jnl->active_tr) { if (jnl->cur_tr != NULL) { panic("jnl: journal @ %p has active tr (%p) and cur tr (%p)\n", jnl, jnl->active_tr, jnl->cur_tr); } tr = jnl->active_tr; jnl->active_tr = NULL; abort_transaction(jnl, tr); } journal_unlock(jnl); return EINVAL; } tr = jnl->active_tr; CHECK_TRANSACTION(tr); // clear this out here so that when check_free_space() calls // the FS flush function, we don't panic in journal_flush() // if the FS were to call that. note: check_free_space() is // called from end_transaction(). // jnl->active_tr = NULL; /* Examine the force-journal-flush state in the active txn */ if (tr->flush_on_completion == TRUE) { /* * If the FS requested it, disallow group commit and force the * transaction out to disk immediately. */ ret = end_transaction(tr, 1, NULL, NULL, TRUE, TRUE); } else { /* in the common path we can simply use the double-buffered journal */ ret = end_transaction(tr, 0, NULL, NULL, TRUE, FALSE); } return ret; } /* * Flush the contents of the journal to the disk. * * Input: * wait_for_IO - * If TRUE, wait to write in-memory journal to the disk * consistently, and also wait to write all asynchronous * metadata blocks to its corresponding locations * consistently on the disk. This means that the journal * is empty at this point and does not contain any * transactions. This is overkill in normal scenarios * but is useful whenever the metadata blocks are required * to be consistent on-disk instead of just the journal * being consistent; like before live verification * and live volume resizing. * * If FALSE, only wait to write in-memory journal to the * disk consistently. This means that the journal still * contains uncommitted transactions and the file system * metadata blocks in the journal transactions might be * written asynchronously to the disk. But there is no * guarantee that they are written to the disk before * returning to the caller. Note that this option is * sufficient for file system data integrity as it * guarantees consistent journal content on the disk. */ int journal_flush(journal *jnl, boolean_t wait_for_IO) { boolean_t drop_lock = FALSE; CHECK_JOURNAL(jnl); free_old_stuff(jnl); if (jnl->flags & JOURNAL_INVALID) { return -1; } KERNEL_DEBUG(DBG_JOURNAL_FLUSH | DBG_FUNC_START, jnl, 0, 0, 0, 0); if (jnl->owner != current_thread()) { journal_lock(jnl); drop_lock = TRUE; } // if we're not active, flush any buffered transactions if (jnl->active_tr == NULL && jnl->cur_tr) { transaction *tr = jnl->cur_tr; jnl->cur_tr = NULL; if (wait_for_IO) { wait_condition(jnl, &jnl->flushing, "journal_flush"); wait_condition(jnl, &jnl->asyncIO, "journal_flush"); } /* * "end_transction" will wait for any current async flush * to complete, before flushing "cur_tr"... because we've * specified the 'must_wait' arg as TRUE, it will then * synchronously flush the "cur_tr" */ end_transaction(tr, 1, NULL, NULL, drop_lock, TRUE); // force it to get flushed } else { if (drop_lock == TRUE) { journal_unlock(jnl); } /* Because of pipelined journal, the journal transactions * might be in process of being flushed on another thread. * If there is nothing to flush currently, we should * synchronize ourselves with the pipelined journal thread * to ensure that all inflight transactions, if any, are * flushed before we return success to caller. */ wait_condition(jnl, &jnl->flushing, "journal_flush"); } if (wait_for_IO) { wait_condition(jnl, &jnl->asyncIO, "journal_flush"); } KERNEL_DEBUG(DBG_JOURNAL_FLUSH | DBG_FUNC_END, jnl, 0, 0, 0, 0); return 0; } int journal_active(journal *jnl) { if (jnl->flags & JOURNAL_INVALID) { return -1; } return (jnl->active_tr == NULL) ? 0 : 1; } void * journal_owner(journal *jnl) { return jnl->owner; } int journal_uses_fua(journal *jnl) { if (jnl->flags & JOURNAL_DO_FUA_WRITES) return 1; return 0; } /* * Relocate the journal. * * You provide the new starting offset and size for the journal. You may * optionally provide a new tbuffer_size; passing zero defaults to not * changing the tbuffer size except as needed to fit within the new journal * size. * * You must have already started a transaction. The transaction may contain * modified blocks (such as those needed to deallocate the old journal, * allocate the new journal, and update the location and size of the journal * in filesystem-private structures). Any transactions prior to the active * transaction will be flushed to the old journal. The new journal will be * initialized, and the blocks from the active transaction will be written to * the new journal. * * The caller will need to update the structures that identify the location * and size of the journal. These updates should be made in the supplied * callback routine. These updates must NOT go into a transaction. You should * force these updates to the media before returning from the callback. In the * even of a crash, either the old journal will be found, with an empty journal, * or the new journal will be found with the contents of the active transaction. * * Upon return from the callback, the blocks from the active transaction are * written to their normal locations on disk. * * (Remember that we have to ensure that blocks get committed to the journal * before being committed to their normal locations. But the blocks don't count * as committed until the new journal is pointed at.) * * Upon return, there is still an active transaction: newly allocated, and * with no modified blocks. Call journal_end_transaction as normal. You may * modifiy additional blocks before calling journal_end_transaction, and those * blocks will (eventually) go to the relocated journal. * * Inputs: * jnl The (opened) journal to relocate. * offset The new journal byte offset (from start of the journal device). * journal_size The size, in bytes, of the new journal. * tbuffer_size The new desired transaction buffer size. Pass zero to keep * the same size as the current journal. The size will be * modified as needed to fit the new journal. * callback Routine called after the new journal has been initialized, * and the active transaction written to the new journal, but * before the blocks are written to their normal locations. * Pass NULL for no callback. * callback_arg An argument passed to the callback routine. * * Result: * 0 No errors * EINVAL The offset is not block aligned * EINVAL The journal_size is not a multiple of the block size * EINVAL The journal is invalid * (any) An error returned by journal_flush. * */ int journal_relocate(journal *jnl, off_t offset, off_t journal_size, int32_t tbuffer_size, errno_t (*callback)(void *), void *callback_arg) { int ret; transaction *tr; size_t i = 0; /* * Sanity check inputs, and adjust the size of the transaction buffer. */ if ((offset % jnl->jhdr->jhdr_size) != 0) { printf("jnl: %s: relocate: offset 0x%llx is not an even multiple of block size 0x%x\n", jnl->jdev_name, offset, jnl->jhdr->jhdr_size); return EINVAL; } if ((journal_size % jnl->jhdr->jhdr_size) != 0) { printf("jnl: %s: relocate: journal size 0x%llx is not an even multiple of block size 0x%x\n", jnl->jdev_name, journal_size, jnl->jhdr->jhdr_size); return EINVAL; } CHECK_JOURNAL(jnl); /* Guarantee we own the active transaction. */ if (jnl->flags & JOURNAL_INVALID) { return EINVAL; } if (jnl->owner != current_thread()) { panic("jnl: relocate: Not the owner! jnl %p, owner %p, curact %p\n", jnl, jnl->owner, current_thread()); } if (tbuffer_size == 0) tbuffer_size = jnl->tbuffer_size; size_up_tbuffer(jnl, tbuffer_size, jnl->jhdr->jhdr_size); /* * Flush any non-active transactions. We have to temporarily hide the * active transaction to make journal_flush flush out non-active but * current (unwritten) transactions. */ tr = jnl->active_tr; CHECK_TRANSACTION(tr); jnl->active_tr = NULL; ret = journal_flush(jnl, TRUE); jnl->active_tr = tr; if (ret) { return ret; } wait_condition(jnl, &jnl->flushing, "end_transaction"); /* * At this point, we have completely flushed the contents of the current * journal to disk (and have asynchronously written all of the txns to * their actual desired locations). As a result, we can (and must) clear * out the old_start array. If we do not, then if the last written transaction * started at the beginning of the journal (starting 1 block into the * journal file) it could confuse the buffer_flushed callback. This is * because we're about to reset the start/end pointers of the journal header * below. */ lock_oldstart(jnl); for (i = 0; i < sizeof (jnl->old_start) / sizeof(jnl->old_start[0]); i++) { jnl->old_start[i] = 0; } unlock_oldstart(jnl); /* Update the journal's offset and size in memory. */ jnl->jdev_offset = offset; jnl->jhdr->start = jnl->jhdr->end = jnl->jhdr->jhdr_size; jnl->jhdr->size = journal_size; jnl->active_start = jnl->jhdr->start; /* * Force the active transaction to be written to the new journal. Call the * supplied callback after the blocks have been written to the journal, but * before they get written to their normal on-disk locations. */ jnl->active_tr = NULL; ret = end_transaction(tr, 1, callback, callback_arg, FALSE, TRUE); if (ret) { printf("jnl: %s: relocate: end_transaction failed (%d)\n", jnl->jdev_name, ret); goto bad_journal; } /* * Create a new, empty transaction to be the active transaction. This way * our caller can use journal_end_transaction as usual. */ ret = journal_allocate_transaction(jnl); if (ret) { printf("jnl: %s: relocate: could not allocate new transaction (%d)\n", jnl->jdev_name, ret); goto bad_journal; } return 0; bad_journal: jnl->flags |= JOURNAL_INVALID; abort_transaction(jnl, tr); return ret; } #else // !JOURNALING - so provide stub functions int journal_uses_fua(__unused journal *jnl) { return 0; } journal * journal_create(__unused struct vnode *jvp, __unused off_t offset, __unused off_t journal_size, __unused struct vnode *fsvp, __unused size_t min_fs_blksz, __unused int32_t flags, __unused int32_t tbuffer_size, __unused void (*flush)(void *arg), __unused void *arg, __unused struct mount *fsmount) { return NULL; } journal * journal_open(__unused struct vnode *jvp, __unused off_t offset, __unused off_t journal_size, __unused struct vnode *fsvp, __unused size_t min_fs_blksz, __unused int32_t flags, __unused int32_t tbuffer_size, __unused void (*flush)(void *arg), __unused void *arg, __unused struct mount *fsmount) { return NULL; } int journal_modify_block_start(__unused journal *jnl, __unused struct buf *bp) { return EINVAL; } int journal_modify_block_end(__unused journal *jnl, __unused struct buf *bp, __unused void (*func)(struct buf *bp, void *arg), __unused void *arg) { return EINVAL; } int journal_kill_block(__unused journal *jnl, __unused struct buf *bp) { return EINVAL; } int journal_relocate(__unused journal *jnl, __unused off_t offset, __unused off_t journal_size, __unused int32_t tbuffer_size, __unused errno_t (*callback)(void *), __unused void *callback_arg) { return EINVAL; } void journal_close(__unused journal *jnl) { } int journal_start_transaction(__unused journal *jnl) { return EINVAL; } int journal_end_transaction(__unused journal *jnl) { return EINVAL; } int journal_flush(__unused journal *jnl, __unused boolean_t wait_for_IO) { return EINVAL; } int journal_is_clean(__unused struct vnode *jvp, __unused off_t offset, __unused off_t journal_size, __unused struct vnode *fsvp, __unused size_t min_fs_block_size) { return 0; } void * journal_owner(__unused journal *jnl) { return NULL; } void journal_lock(__unused journal *jnl) { return; } void journal_unlock(__unused journal *jnl) { return; } __private_extern__ int journal_trim_add_extent(__unused journal *jnl, __unused uint64_t offset, __unused uint64_t length) { return 0; } int journal_request_immediate_flush(__unused journal *jnl) { return 0; } __private_extern__ int journal_trim_remove_extent(__unused journal *jnl, __unused uint64_t offset, __unused uint64_t length) { return 0; } int journal_trim_extent_overlap(__unused journal *jnl, __unused uint64_t offset, __unused uint64_t length, __unused uint64_t *end) { return 0; } #endif // !JOURNALING