/*
* Copy records from userland to the target mirror.
*
* The PFS is identified in the mirror structure. The passed ip is just
* some directory in the overall HAMMER filesystem and has nothing to
* do with the PFS. In fact, there might not even be a root directory for
* the PFS yet!
*/
int
hammer_ioc_mirror_write(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_mirror_rw *mirror)
{
union hammer_ioc_mrecord_any mrec;
struct hammer_cursor cursor;
u_int32_t localization;
int checkspace_count = 0;
int error;
int bytes;
char *uptr;
int seq;
localization = (u_int32_t)mirror->pfs_id << 16;
seq = trans->hmp->flusher.done;
/*
* Validate the mirror structure and relocalize the tracking keys.
*/
if (mirror->size < 0 || mirror->size > 0x70000000)
return(EINVAL);
mirror->key_beg.localization &= HAMMER_LOCALIZE_MASK;
mirror->key_beg.localization += localization;
mirror->key_end.localization &= HAMMER_LOCALIZE_MASK;
mirror->key_end.localization += localization;
mirror->key_cur.localization &= HAMMER_LOCALIZE_MASK;
mirror->key_cur.localization += localization;
/*
* Set up our tracking cursor for the loop. The tracking cursor
* is used to delete records that are no longer present on the
* master. The last handled record at key_cur must be skipped.
*/
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
cursor.key_beg = mirror->key_cur;
cursor.key_end = mirror->key_end;
cursor.flags |= HAMMER_CURSOR_BACKEND;
error = hammer_btree_first(&cursor);
if (error == 0)
cursor.flags |= HAMMER_CURSOR_ATEDISK;
if (error == ENOENT)
error = 0;
/*
* Loop until our input buffer has been exhausted.
*/
while (error == 0 &&
mirror->count + sizeof(mrec.head) <= mirror->size) {
/*
* Don't blow out the buffer cache. Leave room for frontend
* cache as well.
*
* WARNING: See warnings in hammer_unlock_cursor() function.
*/
while (hammer_flusher_meta_halflimit(trans->hmp) ||
hammer_flusher_undo_exhausted(trans, 2)) {
hammer_unlock_cursor(&cursor);
hammer_flusher_wait(trans->hmp, seq);
hammer_lock_cursor(&cursor);
seq = hammer_flusher_async_one(trans->hmp);
}
/*
* If there is insufficient free space it may be due to
* reserved bigblocks, which flushing might fix.
*/
if (hammer_checkspace(trans->hmp, HAMMER_CHKSPC_MIRROR)) {
if (++checkspace_count == 10) {
error = ENOSPC;
break;
}
hammer_unlock_cursor(&cursor);
hammer_flusher_wait(trans->hmp, seq);
hammer_lock_cursor(&cursor);
seq = hammer_flusher_async(trans->hmp, NULL);
}
/*
* Acquire and validate header
*/
if ((bytes = mirror->size - mirror->count) > sizeof(mrec))
bytes = sizeof(mrec);
uptr = (char *)mirror->ubuf + mirror->count;
error = copyin(uptr, &mrec, bytes);
if (error)
break;
if (mrec.head.signature != HAMMER_IOC_MIRROR_SIGNATURE) {
error = EINVAL;
break;
}
if (mrec.head.rec_size < sizeof(mrec.head) ||
mrec.head.rec_size > sizeof(mrec) + HAMMER_XBUFSIZE ||
mirror->count + mrec.head.rec_size > mirror->size) {
error = EINVAL;
break;
}
switch(mrec.head.type & HAMMER_MRECF_TYPE_MASK) {
case HAMMER_MREC_TYPE_SKIP:
if (mrec.head.rec_size != sizeof(mrec.skip))
error = EINVAL;
if (error == 0)
error = hammer_ioc_mirror_write_skip(&cursor, &mrec.skip, mirror, localization);
break;
case HAMMER_MREC_TYPE_REC:
if (mrec.head.rec_size < sizeof(mrec.rec))
error = EINVAL;
if (error == 0)
error = hammer_ioc_mirror_write_rec(&cursor, &mrec.rec, mirror, localization, uptr + sizeof(mrec.rec));
break;
case HAMMER_MREC_TYPE_REC_NODATA:
case HAMMER_MREC_TYPE_REC_BADCRC:
/*
* Records with bad data payloads are ignored XXX.
* Records with no data payload have to be skipped
* (they shouldn't have been written in the first
* place).
*/
if (mrec.head.rec_size < sizeof(mrec.rec))
error = EINVAL;
break;
case HAMMER_MREC_TYPE_PASS:
if (mrec.head.rec_size != sizeof(mrec.rec))
error = EINVAL;
if (error == 0)
error = hammer_ioc_mirror_write_pass(&cursor, &mrec.rec, mirror, localization);
break;
default:
error = EINVAL;
break;
}
/*
* Retry the current record on deadlock, otherwise setup
* for the next loop.
*/
if (error == EDEADLK) {
while (error == EDEADLK) {
hammer_sync_lock_sh(trans);
hammer_recover_cursor(&cursor);
error = hammer_cursor_upgrade(&cursor);
hammer_sync_unlock(trans);
}
} else {
if (error == EALREADY)
error = 0;
if (error == 0) {
mirror->count +=
HAMMER_HEAD_DOALIGN(mrec.head.rec_size);
}
}
}
hammer_done_cursor(&cursor);
/*
* cumulative error
*/
if (error) {
mirror->head.flags |= HAMMER_IOC_HEAD_ERROR;
mirror->head.error = error;
}
/*
* ioctls don't update the RW data structure if an error is returned,
* always return 0.
*/
return(0);
}
/*
* Reblock the B-Tree (leaf) node, record, and/or data if necessary.
*
* XXX We have no visibility into internal B-Tree nodes at the moment,
* only leaf nodes.
*/
static int
hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
hammer_cursor_t cursor, hammer_btree_elm_t elm)
{
hammer_mount_t hmp;
hammer_off_t tmp_offset;
hammer_node_ondisk_t ondisk;
struct hammer_btree_leaf_elm leaf;
int error;
int bytes;
int cur;
int iocflags;
error = 0;
hmp = cursor->trans->hmp;
/*
* Reblock data. Note that data embedded in a record is reblocked
* by the record reblock code. Data processing only occurs at leaf
* nodes and for RECORD element types.
*/
if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
goto skip;
if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
return(0);
tmp_offset = elm->leaf.data_offset;
if (tmp_offset == 0)
goto skip;
if (error)
goto skip;
/*
* NOTE: Localization restrictions may also have been set-up, we can't
* just set the match flags willy-nilly here.
*/
switch (elm->leaf.base.rec_type) {
case HAMMER_RECTYPE_INODE:
case HAMMER_RECTYPE_SNAPSHOT:
case HAMMER_RECTYPE_CONFIG:
iocflags = HAMMER_IOC_DO_INODES;
break;
case HAMMER_RECTYPE_EXT:
case HAMMER_RECTYPE_FIX:
case HAMMER_RECTYPE_PFS:
case HAMMER_RECTYPE_DIRENTRY:
iocflags = HAMMER_IOC_DO_DIRS;
break;
case HAMMER_RECTYPE_DATA:
case HAMMER_RECTYPE_DB:
iocflags = HAMMER_IOC_DO_DATA;
break;
default:
iocflags = 0;
break;
}
if (reblock->head.flags & iocflags) {
++reblock->data_count;
reblock->data_byte_count += elm->leaf.data_len;
bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
if (hammer_debug_general & 0x4000)
kprintf("D %6d/%d\n", bytes, reblock->free_level);
if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
bytes >= reblock->free_level) {
/*
* This is nasty, the uncache code may have to get
* vnode locks and because of that we can't hold
* the cursor locked.
*
* WARNING: See warnings in hammer_unlock_cursor()
* function.
*/
leaf = elm->leaf;
hammer_unlock_cursor(cursor);
hammer_io_direct_uncache(hmp, &leaf);
hammer_lock_cursor(cursor);
/*
* elm may have become stale or invalid, reload it.
* ondisk variable is temporary only. Note that
* cursor->node and thus cursor->node->ondisk may
* also changed.
*/
ondisk = cursor->node->ondisk;
elm = &ondisk->elms[cursor->index];
if (cursor->flags & HAMMER_CURSOR_RETEST) {
kprintf("hammer: debug: retest on "
"reblocker uncache\n");
error = EDEADLK;
} else if (ondisk->type != HAMMER_BTREE_TYPE_LEAF ||
cursor->index >= ondisk->count) {
kprintf("hammer: debug: shifted on "
"reblocker uncache\n");
error = EDEADLK;
} else if (bcmp(&elm->leaf, &leaf, sizeof(leaf))) {
kprintf("hammer: debug: changed on "
"reblocker uncache\n");
error = EDEADLK;
}
if (error == 0)
error = hammer_cursor_upgrade(cursor);
if (error == 0) {
KKASSERT(cursor->index < ondisk->count);
error = hammer_reblock_data(reblock,
cursor, elm);
}
if (error == 0) {
++reblock->data_moves;
reblock->data_byte_moves += elm->leaf.data_len;
}
}
}
skip:
/*
* Reblock a B-Tree internal or leaf node. A leaf node is reblocked
* on initial entry only (element 0). An internal node is reblocked
* when entered upward from its first leaf node only (also element 0).
* Further revisits of the internal node (index > 0) are ignored.
*/
tmp_offset = cursor->node->node_offset;
if (cursor->index == 0 &&
error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
++reblock->btree_count;
bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
if (hammer_debug_general & 0x4000)
kprintf("B %6d/%d\n", bytes, reblock->free_level);
if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
bytes >= reblock->free_level) {
error = hammer_cursor_upgrade(cursor);
if (error == 0) {
if (cursor->parent) {
KKASSERT(cursor->parent_index <
cursor->parent->ondisk->count);
elm = &cursor->parent->ondisk->elms[cursor->parent_index];
} else {
elm = NULL;
}
switch(cursor->node->ondisk->type) {
case HAMMER_BTREE_TYPE_LEAF:
error = hammer_reblock_leaf_node(
reblock, cursor, elm);
break;
case HAMMER_BTREE_TYPE_INTERNAL:
error = hammer_reblock_int_node(
reblock, cursor, elm);
break;
default:
panic("Illegal B-Tree node type");
}
}
if (error == 0) {
++reblock->btree_moves;
}
}
}
hammer_cursor_downgrade(cursor);
return(error);
}
/*
* NOTE: THIS CODE HAS BEEN REMOVED! Pruning no longer attempts to realign
* adjacent records because it seriously interferes with every
* mirroring algorithm I could come up with.
*
* This means that historical accesses beyond the first snapshot
* softlink should be on snapshot boundaries only. Historical
* accesses from "now" to the first snapshot softlink continue to
* be fine-grained.
*
* NOTE: It also looks like there's a bug in the removed code. It is believed
* that create_tid can sometimes get set to 0xffffffffffffffff. Just as
* well we no longer try to do this fancy shit. Probably the attempt to
* correct the rhb is blowing up the cursor's indexing or addressing mapping.
*
* Align the record to cover any gaps created through the deletion of
* records within the pruning space. If we were to just delete the records
* there would be gaps which in turn would cause a snapshot that is NOT on
* a pruning boundary to appear corrupt to the user. Forcing alignment
* of the create_tid and delete_tid for retained records 'reconnects'
* the previously contiguous space, making it contiguous again after the
* deletions.
*
* The use of a reverse iteration allows us to safely align the records and
* related elements without creating temporary overlaps. XXX we should
* add ordering dependancies for record buffers to guarantee consistency
* during recovery.
*/
static int
realign_prune(struct hammer_ioc_prune *prune,
hammer_cursor_t cursor, int realign_cre, int realign_del)
{
struct hammer_ioc_prune_elm *scan;
hammer_btree_elm_t elm;
hammer_tid_t delta;
hammer_tid_t tid;
int error;
hammer_cursor_downgrade(cursor);
elm = &cursor->node->ondisk->elms[cursor->index];
++prune->stat_realignments;
/*
* Align the create_tid. By doing a reverse iteration we guarantee
* that all records after our current record have already been
* aligned, allowing us to safely correct the right-hand-boundary
* (because no record to our right is otherwise exactly matching
* will have a create_tid to the left of our aligned create_tid).
*/
error = 0;
if (realign_cre >= 0) {
scan = &prune->elms[realign_cre];
delta = (elm->leaf.base.create_tid - scan->beg_tid) %
scan->mod_tid;
if (delta) {
tid = elm->leaf.base.create_tid - delta + scan->mod_tid;
/* can EDEADLK */
error = hammer_btree_correct_rhb(cursor, tid + 1);
if (error == 0) {
error = hammer_btree_extract(cursor,
HAMMER_CURSOR_GET_LEAF);
}
if (error == 0) {
/* can EDEADLK */
error = hammer_cursor_upgrade(cursor);
}
if (error == 0) {
hammer_modify_node(cursor->trans, cursor->node,
&elm->leaf.base.create_tid,
sizeof(elm->leaf.base.create_tid));
elm->leaf.base.create_tid = tid;
hammer_modify_node_done(cursor->node);
}
}
}
/*
* Align the delete_tid. This only occurs if the record is historical
* was deleted at some point. Realigning the delete_tid does not
* move the record within the B-Tree but may cause it to temporarily
* overlap a record that has not yet been pruned.
*/
if (error == 0 && realign_del >= 0) {
scan = &prune->elms[realign_del];
delta = (elm->leaf.base.delete_tid - scan->beg_tid) %
scan->mod_tid;
if (delta) {
error = hammer_btree_extract(cursor,
HAMMER_CURSOR_GET_LEAF);
if (error == 0) {
hammer_modify_node(cursor->trans, cursor->node,
&elm->leaf.base.delete_tid,
sizeof(elm->leaf.base.delete_tid));
elm->leaf.base.delete_tid =
elm->leaf.base.delete_tid -
delta + scan->mod_tid;
hammer_modify_node_done(cursor->node);
}
}
}
return (error);
}
/*
* Reblock the B-Tree (leaf) node, record, and/or data if necessary.
*
* XXX We have no visibility into internal B-Tree nodes at the moment,
* only leaf nodes.
*/
static int
hammer_reblock_helper(struct hammer_ioc_reblock *reblock,
hammer_cursor_t cursor, hammer_btree_elm_t elm)
{
hammer_mount_t hmp;
hammer_off_t tmp_offset;
struct hammer_btree_leaf_elm leaf;
int error;
int bytes;
int cur;
int iocflags;
error = 0;
hmp = cursor->trans->hmp;
/*
* Reblock data. Note that data embedded in a record is reblocked
* by the record reblock code. Data processing only occurs at leaf
* nodes and for RECORD element types.
*/
if (cursor->node->ondisk->type != HAMMER_BTREE_TYPE_LEAF)
goto skip;
if (elm->leaf.base.btype != HAMMER_BTREE_TYPE_RECORD)
return(0);
tmp_offset = elm->leaf.data_offset;
if (tmp_offset == 0)
goto skip;
if (error)
goto skip;
/*
* NOTE: Localization restrictions may also have been set-up, we can't
* just set the match flags willy-nilly here.
*/
switch(elm->leaf.base.rec_type) {
case HAMMER_RECTYPE_INODE:
iocflags = HAMMER_IOC_DO_INODES;
break;
case HAMMER_RECTYPE_EXT:
case HAMMER_RECTYPE_FIX:
case HAMMER_RECTYPE_PFS:
case HAMMER_RECTYPE_DIRENTRY:
iocflags = HAMMER_IOC_DO_DIRS;
break;
case HAMMER_RECTYPE_DATA:
case HAMMER_RECTYPE_DB:
iocflags = HAMMER_IOC_DO_DATA;
break;
default:
iocflags = 0;
break;
}
if (reblock->head.flags & iocflags) {
++reblock->data_count;
reblock->data_byte_count += elm->leaf.data_len;
bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
if (hammer_debug_general & 0x4000)
kprintf("D %6d/%d\n", bytes, reblock->free_level);
if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
bytes >= reblock->free_level) {
/*
* This is nasty, the uncache code may have to get
* vnode locks and because of that we can't hold
* the cursor locked.
*/
leaf = elm->leaf;
hammer_unlock_cursor(cursor, 0);
hammer_io_direct_uncache(hmp, &leaf);
hammer_lock_cursor(cursor, 0);
if (cursor->flags & HAMMER_CURSOR_RETEST) {
kprintf("hammer: retest after uncache\n");
error = EDEADLK;
} else {
KKASSERT(bcmp(&elm->leaf, &leaf, sizeof(leaf)) == 0);
}
if (error == 0)
error = hammer_cursor_upgrade(cursor);
if (error == 0) {
error = hammer_reblock_data(reblock,
cursor, elm);
}
if (error == 0) {
++reblock->data_moves;
reblock->data_byte_moves += elm->leaf.data_len;
}
}
}
skip:
/*
* Reblock a B-Tree internal or leaf node.
*/
tmp_offset = cursor->node->node_offset;
if (cursor->index == 0 &&
error == 0 && (reblock->head.flags & HAMMER_IOC_DO_BTREE)) {
++reblock->btree_count;
bytes = hammer_blockmap_getfree(hmp, tmp_offset, &cur, &error);
if (hammer_debug_general & 0x4000)
kprintf("B %6d/%d\n", bytes, reblock->free_level);
if (error == 0 && (cur == 0 || reblock->free_level == 0) &&
bytes >= reblock->free_level) {
error = hammer_cursor_upgrade(cursor);
if (error == 0) {
if (cursor->parent)
elm = &cursor->parent->ondisk->elms[cursor->parent_index];
else
elm = NULL;
switch(cursor->node->ondisk->type) {
case HAMMER_BTREE_TYPE_LEAF:
error = hammer_reblock_leaf_node(
reblock, cursor, elm);
break;
case HAMMER_BTREE_TYPE_INTERNAL:
error = hammer_reblock_int_node(
reblock, cursor, elm);
break;
default:
panic("Illegal B-Tree node type");
}
}
if (error == 0) {
++reblock->btree_moves;
}
}
}
hammer_cursor_downgrade(cursor);
return(error);
}
