/*
* Reblock a B-Tree internal node. The parent must be adjusted to point to
* the new copy of the internal node, and the node's children's parent
* pointers must also be adjusted to point to the new copy.
*
* elm is a pointer to the parent element pointing at cursor.node.
*/
static int
hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
hammer_cursor_t cursor, hammer_btree_elm_t elm)
{
struct hammer_node_lock lockroot;
hammer_node_t onode;
hammer_node_t nnode;
int error;
hammer_node_lock_init(&lockroot, cursor->node);
error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
if (error)
goto done;
/*
* Don't supply a hint when allocating the leaf. Fills are done
* from the leaf upwards.
*/
onode = cursor->node;
nnode = hammer_alloc_btree(cursor->trans, 0, &error);
if (nnode == NULL)
goto done;
hammer_lock_ex(&nnode->lock);
hammer_modify_node_noundo(cursor->trans, nnode);
hammer_move_node(cursor, elm, onode, nnode);
/*
* Clean up.
*
* The new node replaces the current node in the cursor. The cursor
* expects it to be locked so leave it locked. Discard onode.
*/
hammer_cursor_replaced_node(onode, nnode);
hammer_delete_node(cursor->trans, onode);
if (hammer_debug_general & 0x4000) {
hdkprintf("%08x %016jx -> %016jx\n",
(elm ? elm->base.localization : -1),
(intmax_t)onode->node_offset,
(intmax_t)nnode->node_offset);
}
hammer_modify_node_done(nnode);
cursor->node = nnode;
hammer_unlock(&onode->lock);
hammer_rel_node(onode);
done:
hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
return (error);
}
/*
* Remove a volume.
*/
int
hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_volume *ioc)
{
hammer_mount_t hmp = trans->hmp;
struct mount *mp = hmp->mp;
struct hammer_volume_ondisk ondisk;
hammer_volume_t volume;
int64_t total_bigblocks, empty_bigblocks;
int vol_no;
int error = 0;
if (mp->mnt_flag & MNT_RDONLY) {
hmkprintf(hmp, "Cannot del volume from read-only HAMMER filesystem\n");
return (EINVAL);
}
if (hammer_lock_ex_try(&hmp->volume_lock) != 0) {
hmkprintf(hmp, "Another volume operation is in progress!\n");
return (EAGAIN);
}
if (hmp->nvolumes <= 1) {
hammer_unlock(&hmp->volume_lock);
hmkprintf(hmp, "No HAMMER volume to delete\n");
return (EINVAL);
}
/*
* find volume by volname
*/
volume = NULL;
HAMMER_VOLUME_NUMBER_FOREACH(hmp, vol_no) {
volume = hammer_get_volume(hmp, vol_no, &error);
KKASSERT(volume != NULL && error == 0);
if (strcmp(volume->vol_name, ioc->device_name) == 0) {
break;
}
hammer_rel_volume(volume, 0);
volume = NULL;
}
int
hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_volume *ioc)
{
hammer_mount_t hmp = trans->hmp;
struct mount *mp = hmp->mp;
struct hammer_volume_ondisk ondisk;
hammer_volume_t volume;
int64_t total_bigblocks, empty_bigblocks;
int free_vol_no = 0;
int error;
if (mp->mnt_flag & MNT_RDONLY) {
hmkprintf(hmp, "Cannot add volume to read-only HAMMER filesystem\n");
return (EINVAL);
}
if (hammer_lock_ex_try(&hmp->volume_lock) != 0) {
hmkprintf(hmp, "Another volume operation is in progress!\n");
return (EAGAIN);
}
if (hmp->nvolumes >= HAMMER_MAX_VOLUMES) {
hammer_unlock(&hmp->volume_lock);
hmkprintf(hmp, "Max number of HAMMER volumes exceeded\n");
return (EINVAL);
}
/*
* Find an unused volume number.
*/
while (free_vol_no < HAMMER_MAX_VOLUMES &&
hammer_volume_number_test(hmp, free_vol_no)) {
++free_vol_no;
}
if (free_vol_no >= HAMMER_MAX_VOLUMES) {
hmkprintf(hmp, "Max number of HAMMER volumes exceeded\n");
error = EINVAL;
goto end;
}
error = hammer_format_volume_header(hmp, ioc, &ondisk, free_vol_no);
if (error)
goto end;
error = hammer_install_volume(hmp, ioc->device_name, NULL, &ondisk);
if (error)
goto end;
hammer_sync_lock_sh(trans);
hammer_lock_ex(&hmp->blkmap_lock);
volume = hammer_get_volume(hmp, free_vol_no, &error);
KKASSERT(volume != NULL && error == 0);
error = hammer_format_freemap(trans, volume);
KKASSERT(error == 0);
error = hammer_count_bigblocks(hmp, volume,
&total_bigblocks, &empty_bigblocks);
KKASSERT(error == 0);
KKASSERT(total_bigblocks == empty_bigblocks);
hammer_rel_volume(volume, 0);
++hmp->nvolumes;
error = hammer_update_volumes_header(trans,
total_bigblocks, empty_bigblocks);
KKASSERT(error == 0);
hammer_unlock(&hmp->blkmap_lock);
hammer_sync_unlock(trans);
KKASSERT(error == 0);
end:
hammer_unlock(&hmp->volume_lock);
if (error)
hmkprintf(hmp, "An error occurred: %d\n", error);
return (error);
}
/*
* HAMMER version 4+ REDO support.
*
* REDO records are used to improve fsync() performance. Instead of having
* to go through a complete double-flush cycle involving at least two disk
* synchronizations the fsync need only flush UNDO/REDO FIFO buffers through
* the related REDO records, which is a single synchronization requiring
* no track seeking. If a recovery becomes necessary the recovery code
* will generate logical data writes based on the REDO records encountered.
* That is, the recovery code will UNDO any partial meta-data/data writes
* at the raw disk block level and then REDO the data writes at the logical
* level.
*/
int
hammer_generate_redo(hammer_transaction_t trans, hammer_inode_t ip,
hammer_off_t file_off, u_int32_t flags,
void *base, int len)
{
hammer_mount_t hmp;
hammer_volume_t root_volume;
hammer_blockmap_t undomap;
hammer_buffer_t buffer = NULL;
hammer_fifo_redo_t redo;
hammer_fifo_tail_t tail;
hammer_off_t next_offset;
int error;
int bytes;
int n;
/*
* Setup
*/
hmp = trans->hmp;
root_volume = trans->rootvol;
undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
/*
* No undo recursion when modifying the root volume
*/
hammer_modify_volume(NULL, root_volume, NULL, 0);
hammer_lock_ex(&hmp->undo_lock);
/* undo had better not roll over (loose test) */
if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
panic("hammer: insufficient undo FIFO space!");
/*
* Loop until the undo for the entire range has been laid down.
* Loop at least once (len might be 0 as a degenerate case).
*/
for (;;) {
/*
* Fetch the layout offset in the UNDO FIFO, wrap it as
* necessary.
*/
if (undomap->next_offset == undomap->alloc_offset) {
undomap->next_offset =
HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
}
next_offset = undomap->next_offset;
/*
* This is a tail-chasing FIFO, when we hit the start of a new
* buffer we don't have to read it in.
*/
if ((next_offset & HAMMER_BUFMASK) == 0) {
redo = hammer_bnew(hmp, next_offset, &error, &buffer);
hammer_format_undo(redo, hmp->undo_seqno ^ 0x40000000);
} else {
redo = hammer_bread(hmp, next_offset, &error, &buffer);
}
if (error)
break;
hammer_modify_buffer(NULL, buffer, NULL, 0);
/*
* Calculate how big a media structure fits up to the next
* alignment point and how large a data payload we can
* accomodate.
*
* If n calculates to 0 or negative there is no room for
* anything but a PAD.
*/
bytes = HAMMER_UNDO_ALIGN -
((int)next_offset & HAMMER_UNDO_MASK);
n = bytes -
(int)sizeof(struct hammer_fifo_redo) -
(int)sizeof(struct hammer_fifo_tail);
/*
* If available space is insufficient for any payload
* we have to lay down a PAD.
*
* The minimum PAD is 8 bytes and the head and tail will
* overlap each other in that case. PADs do not have
* sequence numbers or CRCs.
*
* A PAD may not start on a boundary. That is, every
* 512-byte block in the UNDO/REDO FIFO must begin with
* a record containing a sequence number.
*/
if (n <= 0) {
KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
tail = (void *)((char *)redo + bytes - sizeof(*tail));
if ((void *)redo != (void *)tail) {
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_PAD;
tail->tail_size = bytes;
}
redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
redo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
redo->head.hdr_size = bytes;
/* NO CRC OR SEQ NO */
undomap->next_offset += bytes;
hammer_modify_buffer_done(buffer);
hammer_stats_redo += bytes;
continue;
}
/*
* When generating an inode-related REDO record we track
* the point in the UNDO/REDO FIFO containing the inode's
* earliest REDO record. See hammer_generate_redo_sync().
*
* redo_fifo_next is cleared when an inode is staged to
* the backend and then used to determine how to reassign
* redo_fifo_start after the inode flush completes.
*/
if (ip) {
redo->redo_objid = ip->obj_id;
redo->redo_localization = ip->obj_localization;
if ((ip->flags & HAMMER_INODE_RDIRTY) == 0) {
ip->redo_fifo_start = next_offset;
if (RB_INSERT(hammer_redo_rb_tree,
&hmp->rb_redo_root, ip)) {
panic("hammer_generate_redo: "
"cannot insert inode %p on "
"redo FIFO", ip);
}
ip->flags |= HAMMER_INODE_RDIRTY;
}
if (ip->redo_fifo_next == 0)
ip->redo_fifo_next = next_offset;
} else {
redo->redo_objid = 0;
redo->redo_localization = 0;
}
/*
* Calculate the actual payload and recalculate the size
* of the media structure as necessary. If no data buffer
* is supplied there is no payload.
*/
if (base == NULL) {
n = 0;
} else if (n > len) {
n = len;
}
bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
~HAMMER_HEAD_ALIGN_MASK) +
(int)sizeof(struct hammer_fifo_redo) +
(int)sizeof(struct hammer_fifo_tail);
if (hammer_debug_general & 0x0080) {
kprintf("redo %016llx %d %d\n",
(long long)next_offset, bytes, n);
}
redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
redo->head.hdr_type = HAMMER_HEAD_TYPE_REDO;
redo->head.hdr_size = bytes;
redo->head.hdr_seq = hmp->undo_seqno++;
redo->head.hdr_crc = 0;
redo->redo_mtime = trans->time;
redo->redo_offset = file_off;
redo->redo_flags = flags;
/*
* Incremental payload. If no payload we throw the entire
* len into redo_data_bytes and will not loop.
*/
if (base) {
redo->redo_data_bytes = n;
bcopy(base, redo + 1, n);
len -= n;
base = (char *)base + n;
file_off += n;
} else {
redo->redo_data_bytes = len;
file_off += len;
len = 0;
}
tail = (void *)((char *)redo + bytes - sizeof(*tail));
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_REDO;
tail->tail_size = bytes;
KKASSERT(bytes >= sizeof(redo->head));
redo->head.hdr_crc = crc32(redo, HAMMER_FIFO_HEAD_CRCOFF) ^
crc32(&redo->head + 1, bytes - sizeof(redo->head));
undomap->next_offset += bytes;
hammer_stats_redo += bytes;
/*
* Before we finish off the buffer we have to deal with any
* junk between the end of the media structure we just laid
* down and the UNDO alignment boundary. We do this by laying
* down a dummy PAD. Even though we will probably overwrite
* it almost immediately we have to do this so recovery runs
* can iterate the UNDO space without having to depend on
* the indices in the volume header.
*
* This dummy PAD will be overwritten on the next undo so
* we do not adjust undomap->next_offset.
*/
bytes = HAMMER_UNDO_ALIGN -
((int)undomap->next_offset & HAMMER_UNDO_MASK);
if (bytes != HAMMER_UNDO_ALIGN) {
KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
redo = (void *)(tail + 1);
tail = (void *)((char *)redo + bytes - sizeof(*tail));
if ((void *)redo != (void *)tail) {
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_PAD;
tail->tail_size = bytes;
}
redo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
redo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
redo->head.hdr_size = bytes;
/* NO CRC OR SEQ NO */
}
hammer_modify_buffer_done(buffer);
if (len == 0)
break;
}
hammer_modify_volume_done(root_volume);
hammer_unlock(&hmp->undo_lock);
if (buffer)
hammer_rel_buffer(buffer, 0);
/*
* Make sure the nominal undo span contains at least one REDO_SYNC,
* otherwise the REDO recovery will not be triggered.
*/
if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
flags != HAMMER_REDO_SYNC) {
hammer_generate_redo_sync(trans);
}
return(error);
}
/*
* Delete snapshot transaction id(s) from the list of snapshots.
*/
static
int
hammer_ioc_del_snapshot(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_snapshot *snap)
{
hammer_mount_t hmp = ip->hmp;
struct hammer_cursor cursor;
int error;
/*
* Validate structure
*/
if (snap->count > HAMMER_SNAPS_PER_IOCTL)
return (EINVAL);
if (snap->index >= snap->count)
return (EINVAL);
hammer_lock_ex(&hmp->snapshot_lock);
again:
/*
* Look for keys starting after the previous iteration, or at
* the beginning if snap->count is 0.
*/
error = hammer_init_cursor(trans, &cursor, &ip->cache[0], NULL);
if (error) {
hammer_done_cursor(&cursor);
return(error);
}
cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
cursor.key_beg.create_tid = 0;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.obj_type = 0;
cursor.key_beg.rec_type = HAMMER_RECTYPE_SNAPSHOT;
cursor.key_beg.localization = ip->obj_localization | HAMMER_LOCALIZE_INODE;
cursor.asof = HAMMER_MAX_TID;
cursor.flags |= HAMMER_CURSOR_ASOF;
while (snap->index < snap->count) {
cursor.key_beg.key = (int64_t)snap->snaps[snap->index].tid;
error = hammer_btree_lookup(&cursor);
if (error)
break;
error = hammer_btree_extract_leaf(&cursor);
if (error)
break;
error = hammer_delete_at_cursor(&cursor, HAMMER_DELETE_DESTROY,
0, 0, 0, NULL);
if (error == EDEADLK) {
hammer_done_cursor(&cursor);
goto again;
}
if (error)
break;
++snap->index;
}
snap->head.error = error;
hammer_done_cursor(&cursor);
hammer_unlock(&hmp->snapshot_lock);
return(0);
}
/*
* Add a snapshot transaction id(s) to the list of snapshots.
*
* NOTE: Records are created with an allocated TID. If a flush cycle
* is in progress the record may be synced in the current flush
* cycle and the volume header will reflect the allocation of the
* TID, but the synchronization point may not catch up to the
* TID until the next flush cycle.
*/
static
int
hammer_ioc_add_snapshot(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_snapshot *snap)
{
hammer_mount_t hmp = ip->hmp;
struct hammer_btree_leaf_elm leaf;
struct hammer_cursor cursor;
int error;
/*
* Validate structure
*/
if (snap->count > HAMMER_SNAPS_PER_IOCTL)
return (EINVAL);
if (snap->index >= snap->count)
return (EINVAL);
hammer_lock_ex(&hmp->snapshot_lock);
again:
/*
* Look for keys starting after the previous iteration, or at
* the beginning if snap->count is 0.
*/
error = hammer_init_cursor(trans, &cursor, &ip->cache[0], NULL);
if (error) {
hammer_done_cursor(&cursor);
return(error);
}
cursor.asof = HAMMER_MAX_TID;
cursor.flags |= HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF;
bzero(&leaf, sizeof(leaf));
leaf.base.obj_id = HAMMER_OBJID_ROOT;
leaf.base.rec_type = HAMMER_RECTYPE_SNAPSHOT;
leaf.base.create_tid = hammer_alloc_tid(hmp, 1);
leaf.base.btype = HAMMER_BTREE_TYPE_RECORD;
leaf.base.localization = ip->obj_localization | HAMMER_LOCALIZE_INODE;
leaf.data_len = sizeof(struct hammer_snapshot_data);
while (snap->index < snap->count) {
leaf.base.key = (int64_t)snap->snaps[snap->index].tid;
cursor.key_beg = leaf.base;
error = hammer_btree_lookup(&cursor);
if (error == 0) {
error = EEXIST;
break;
}
/*
* NOTE: Must reload key_beg after an ASOF search because
* the create_tid may have been modified during the
* search.
*/
cursor.flags &= ~HAMMER_CURSOR_ASOF;
cursor.key_beg = leaf.base;
error = hammer_create_at_cursor(&cursor, &leaf,
&snap->snaps[snap->index],
HAMMER_CREATE_MODE_SYS);
if (error == EDEADLK) {
hammer_done_cursor(&cursor);
goto again;
}
cursor.flags |= HAMMER_CURSOR_ASOF;
if (error)
break;
++snap->index;
}
snap->head.error = error;
hammer_done_cursor(&cursor);
hammer_unlock(&hmp->snapshot_lock);
return(0);
}
/*
* Set version info
*/
static
int
hammer_ioc_set_version(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_version *ver)
{
hammer_mount_t hmp = trans->hmp;
struct hammer_cursor cursor;
hammer_volume_t volume;
int error;
int over = hmp->version;
/*
* Generally do not allow downgrades. However, version 4 can
* be downgraded to version 3.
*/
if (ver->cur_version < hmp->version) {
if (!(ver->cur_version == 3 && hmp->version == 4))
return(EINVAL);
}
if (ver->cur_version == hmp->version)
return(0);
if (ver->cur_version > HAMMER_VOL_VERSION_MAX)
return(EINVAL);
if (hmp->ronly)
return(EROFS);
/*
* Update the root volume header and the version cached in
* the hammer_mount structure.
*/
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
if (error)
goto failed;
hammer_lock_ex(&hmp->flusher.finalize_lock);
hammer_sync_lock_ex(trans);
hmp->version = ver->cur_version;
/*
* If upgrading from version < 4 to version >= 4 the UNDO FIFO
* must be reinitialized.
*/
if (over < HAMMER_VOL_VERSION_FOUR &&
ver->cur_version >= HAMMER_VOL_VERSION_FOUR) {
hkprintf("upgrade undo to version 4\n");
error = hammer_upgrade_undo_4(trans);
if (error)
goto failed;
}
/*
* Adjust the version in the volume header
*/
volume = hammer_get_root_volume(hmp, &error);
KKASSERT(error == 0);
hammer_modify_volume_field(cursor.trans, volume, vol_version);
volume->ondisk->vol_version = ver->cur_version;
hammer_modify_volume_done(volume);
hammer_rel_volume(volume, 0);
hammer_sync_unlock(trans);
hammer_unlock(&hmp->flusher.finalize_lock);
failed:
ver->head.error = error;
hammer_done_cursor(&cursor);
return(0);
}
/*
* Generate UNDO record(s) for the block of data at the specified zone1
* or zone2 offset.
*
* The recovery code will execute UNDOs in reverse order, allowing overlaps.
* All the UNDOs are executed together so if we already laid one down we
* do not have to lay another one down for the same range.
*
* For HAMMER version 4+ UNDO a 512 byte boundary is enforced and a PAD
* will be laid down for any unused space. UNDO FIFO media structures
* will implement the hdr_seq field (it used to be reserved01), and
* both flush and recovery mechanics will be very different.
*
* WARNING! See also hammer_generate_redo() in hammer_redo.c
*/
int
hammer_generate_undo(hammer_transaction_t trans,
hammer_off_t zone_off, void *base, int len)
{
hammer_mount_t hmp;
hammer_volume_t root_volume;
hammer_blockmap_t undomap;
hammer_buffer_t buffer = NULL;
hammer_fifo_undo_t undo;
hammer_fifo_tail_t tail;
hammer_off_t next_offset;
int error;
int bytes;
int n;
hmp = trans->hmp;
/*
* A SYNC record may be required before we can lay down a general
* UNDO. This ensures that the nominal recovery span contains
* at least one SYNC record telling the recovery code how far
* out-of-span it must go to run the REDOs.
*/
if ((hmp->flags & HAMMER_MOUNT_REDO_SYNC) == 0 &&
hmp->version >= HAMMER_VOL_VERSION_FOUR) {
hammer_generate_redo_sync(trans);
}
/*
* Enter the offset into our undo history. If there is an existing
* undo we do not have to generate a new one.
*/
if (hammer_enter_undo_history(hmp, zone_off, len) == EALREADY)
return(0);
root_volume = trans->rootvol;
undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
/* no undo recursion */
hammer_modify_volume_noundo(NULL, root_volume);
hammer_lock_ex(&hmp->undo_lock);
/* undo had better not roll over (loose test) */
if (hammer_undo_space(trans) < len + HAMMER_BUFSIZE*3)
panic("hammer: insufficient undo FIFO space!");
/*
* Loop until the undo for the entire range has been laid down.
*/
while (len) {
/*
* Fetch the layout offset in the UNDO FIFO, wrap it as
* necessary.
*/
if (undomap->next_offset == undomap->alloc_offset) {
undomap->next_offset =
HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
}
next_offset = undomap->next_offset;
/*
* This is a tail-chasing FIFO, when we hit the start of a new
* buffer we don't have to read it in.
*/
if ((next_offset & HAMMER_BUFMASK) == 0) {
undo = hammer_bnew(hmp, next_offset, &error, &buffer);
hammer_format_undo(undo, hmp->undo_seqno ^ 0x40000000);
} else {
undo = hammer_bread(hmp, next_offset, &error, &buffer);
}
if (error)
break;
/* no undo recursion */
hammer_modify_buffer_noundo(NULL, buffer);
/*
* Calculate how big a media structure fits up to the next
* alignment point and how large a data payload we can
* accomodate.
*
* If n calculates to 0 or negative there is no room for
* anything but a PAD.
*/
bytes = HAMMER_UNDO_ALIGN -
((int)next_offset & HAMMER_UNDO_MASK);
n = bytes -
(int)sizeof(struct hammer_fifo_undo) -
(int)sizeof(struct hammer_fifo_tail);
/*
* If available space is insufficient for any payload
* we have to lay down a PAD.
*
* The minimum PAD is 8 bytes and the head and tail will
* overlap each other in that case. PADs do not have
* sequence numbers or CRCs.
*
* A PAD may not start on a boundary. That is, every
* 512-byte block in the UNDO/REDO FIFO must begin with
* a record containing a sequence number.
*/
if (n <= 0) {
KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
KKASSERT(((int)next_offset & HAMMER_UNDO_MASK) != 0);
tail = (void *)((char *)undo + bytes - sizeof(*tail));
if ((void *)undo != (void *)tail) {
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_PAD;
tail->tail_size = bytes;
}
undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
undo->head.hdr_size = bytes;
/* NO CRC OR SEQ NO */
undomap->next_offset += bytes;
hammer_modify_buffer_done(buffer);
hammer_stats_undo += bytes;
continue;
}
/*
* Calculate the actual payload and recalculate the size
* of the media structure as necessary.
*/
if (n > len) {
n = len;
bytes = ((n + HAMMER_HEAD_ALIGN_MASK) &
~HAMMER_HEAD_ALIGN_MASK) +
(int)sizeof(struct hammer_fifo_undo) +
(int)sizeof(struct hammer_fifo_tail);
}
if (hammer_debug_general & 0x0080) {
kprintf("undo %016llx %d %d\n",
(long long)next_offset, bytes, n);
}
undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
undo->head.hdr_type = HAMMER_HEAD_TYPE_UNDO;
undo->head.hdr_size = bytes;
undo->head.hdr_seq = hmp->undo_seqno++;
undo->head.hdr_crc = 0;
undo->undo_offset = zone_off;
undo->undo_data_bytes = n;
bcopy(base, undo + 1, n);
tail = (void *)((char *)undo + bytes - sizeof(*tail));
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_UNDO;
tail->tail_size = bytes;
KKASSERT(bytes >= sizeof(undo->head));
undo->head.hdr_crc = crc32(undo, HAMMER_FIFO_HEAD_CRCOFF) ^
crc32(&undo->head + 1, bytes - sizeof(undo->head));
undomap->next_offset += bytes;
hammer_stats_undo += bytes;
/*
* Before we finish off the buffer we have to deal with any
* junk between the end of the media structure we just laid
* down and the UNDO alignment boundary. We do this by laying
* down a dummy PAD. Even though we will probably overwrite
* it almost immediately we have to do this so recovery runs
* can iterate the UNDO space without having to depend on
* the indices in the volume header.
*
* This dummy PAD will be overwritten on the next undo so
* we do not adjust undomap->next_offset.
*/
bytes = HAMMER_UNDO_ALIGN -
((int)undomap->next_offset & HAMMER_UNDO_MASK);
if (bytes != HAMMER_UNDO_ALIGN) {
KKASSERT(bytes >= sizeof(struct hammer_fifo_tail));
undo = (void *)(tail + 1);
tail = (void *)((char *)undo + bytes - sizeof(*tail));
if ((void *)undo != (void *)tail) {
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_PAD;
tail->tail_size = bytes;
}
undo->head.hdr_signature = HAMMER_HEAD_SIGNATURE;
undo->head.hdr_type = HAMMER_HEAD_TYPE_PAD;
undo->head.hdr_size = bytes;
/* NO CRC OR SEQ NO */
}
hammer_modify_buffer_done(buffer);
/*
* Adjust for loop
*/
len -= n;
base = (char *)base + n;
zone_off += n;
}
hammer_modify_volume_done(root_volume);
hammer_unlock(&hmp->undo_lock);
if (buffer)
hammer_rel_buffer(buffer, 0);
return(error);
}
/*
* HAMMER version 4+ conversion support.
*
* Convert a HAMMER version < 4 UNDO FIFO area to a 4+ UNDO FIFO area.
* The 4+ UNDO FIFO area is backwards compatible. The conversion is
* needed to initialize the sequence space and place headers on the
* new 512-byte undo boundary.
*/
int
hammer_upgrade_undo_4(hammer_transaction_t trans)
{
hammer_mount_t hmp;
hammer_volume_t root_volume;
hammer_blockmap_t undomap;
hammer_buffer_t buffer = NULL;
hammer_fifo_head_t head;
hammer_fifo_tail_t tail;
hammer_off_t next_offset;
u_int32_t seqno;
int error;
int bytes;
hmp = trans->hmp;
root_volume = trans->rootvol;
/* no undo recursion */
hammer_lock_ex(&hmp->undo_lock);
hammer_modify_volume_noundo(NULL, root_volume);
/*
* Adjust the in-core undomap and the on-disk undomap.
*/
next_offset = HAMMER_ZONE_ENCODE(HAMMER_ZONE_UNDO_INDEX, 0);
undomap = &hmp->blockmap[HAMMER_ZONE_UNDO_INDEX];
undomap->next_offset = next_offset;
undomap->first_offset = next_offset;
undomap = &root_volume->ondisk->vol0_blockmap[HAMMER_ZONE_UNDO_INDEX];
undomap->next_offset = next_offset;
undomap->first_offset = next_offset;
/*
* Loop over the entire UNDO space creating DUMMY entries. Sequence
* numbers are assigned.
*/
seqno = 0;
bytes = HAMMER_UNDO_ALIGN;
while (next_offset != undomap->alloc_offset) {
head = hammer_bnew(hmp, next_offset, &error, &buffer);
if (error)
break;
hammer_modify_buffer_noundo(NULL, buffer);
tail = (void *)((char *)head + bytes - sizeof(*tail));
head->hdr_signature = HAMMER_HEAD_SIGNATURE;
head->hdr_type = HAMMER_HEAD_TYPE_DUMMY;
head->hdr_size = bytes;
head->hdr_seq = seqno;
head->hdr_crc = 0;
tail = (void *)((char *)head + bytes - sizeof(*tail));
tail->tail_signature = HAMMER_TAIL_SIGNATURE;
tail->tail_type = HAMMER_HEAD_TYPE_DUMMY;
tail->tail_size = bytes;
head->hdr_crc = crc32(head, HAMMER_FIFO_HEAD_CRCOFF) ^
crc32(head + 1, bytes - sizeof(*head));
hammer_modify_buffer_done(buffer);
hammer_stats_undo += bytes;
next_offset += HAMMER_UNDO_ALIGN;
++seqno;
}
/*
* The sequence number will be the next sequence number to lay down.
*/
hmp->undo_seqno = seqno;
kprintf("version upgrade seqno start %08x\n", seqno);
hammer_modify_volume_done(root_volume);
hammer_unlock(&hmp->undo_lock);
if (buffer)
hammer_rel_buffer(buffer, 0);
return (error);
}
/*
* Reblock a B-Tree internal node. The parent must be adjusted to point to
* the new copy of the internal node, and the node's children's parent
* pointers must also be adjusted to point to the new copy.
*
* elm is a pointer to the parent element pointing at cursor.node.
*/
static int
hammer_reblock_int_node(struct hammer_ioc_reblock *reblock,
hammer_cursor_t cursor, hammer_btree_elm_t elm)
{
struct hammer_node_lock lockroot;
hammer_node_t onode;
hammer_node_t nnode;
int error;
int i;
hammer_node_lock_init(&lockroot, cursor->node);
error = hammer_btree_lock_children(cursor, 1, &lockroot, NULL);
if (error)
goto done;
onode = cursor->node;
nnode = hammer_alloc_btree(cursor->trans, 0, &error);
if (nnode == NULL)
goto done;
/*
* Move the node. Adjust the parent's pointer to us first.
*/
hammer_lock_ex(&nnode->lock);
hammer_modify_node_noundo(cursor->trans, nnode);
bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
if (elm) {
/*
* We are not the root of the B-Tree
*/
hammer_modify_node(cursor->trans, cursor->parent,
&elm->internal.subtree_offset,
sizeof(elm->internal.subtree_offset));
elm->internal.subtree_offset = nnode->node_offset;
hammer_modify_node_done(cursor->parent);
} else {
/*
* We are the root of the B-Tree
*/
hammer_volume_t volume;
volume = hammer_get_root_volume(cursor->trans->hmp, &error);
KKASSERT(error == 0);
hammer_modify_volume_field(cursor->trans, volume,
vol0_btree_root);
volume->ondisk->vol0_btree_root = nnode->node_offset;
hammer_modify_volume_done(volume);
hammer_rel_volume(volume, 0);
}
/*
* Now adjust our children's pointers to us.
*/
for (i = 0; i < nnode->ondisk->count; ++i) {
elm = &nnode->ondisk->elms[i];
error = btree_set_parent(cursor->trans, nnode, elm);
if (error)
panic("reblock internal node: fixup problem");
}
/*
* Clean up.
*
* The new node replaces the current node in the cursor. The cursor
* expects it to be locked so leave it locked. Discard onode.
*/
hammer_cursor_replaced_node(onode, nnode);
hammer_delete_node(cursor->trans, onode);
if (hammer_debug_general & 0x4000) {
kprintf("REBLOCK INODE %016llx -> %016llx\n",
(long long)onode->node_offset,
(long long)nnode->node_offset);
}
hammer_modify_node_done(nnode);
cursor->node = nnode;
hammer_unlock(&onode->lock);
hammer_rel_node(onode);
done:
hammer_btree_unlock_children(cursor->trans->hmp, &lockroot, NULL);
return (error);
}
/*
* Reblock a B-Tree leaf node. The parent must be adjusted to point to
* the new copy of the leaf node.
*
* elm is a pointer to the parent element pointing at cursor.node.
*/
static int
hammer_reblock_leaf_node(struct hammer_ioc_reblock *reblock,
hammer_cursor_t cursor, hammer_btree_elm_t elm)
{
hammer_node_t onode;
hammer_node_t nnode;
int error;
/*
* Don't supply a hint when allocating the leaf. Fills are done
* from the leaf upwards.
*/
onode = cursor->node;
nnode = hammer_alloc_btree(cursor->trans, 0, &error);
if (nnode == NULL)
return (error);
/*
* Move the node
*/
hammer_lock_ex(&nnode->lock);
hammer_modify_node_noundo(cursor->trans, nnode);
bcopy(onode->ondisk, nnode->ondisk, sizeof(*nnode->ondisk));
if (elm) {
/*
* We are not the root of the B-Tree
*/
hammer_modify_node(cursor->trans, cursor->parent,
&elm->internal.subtree_offset,
sizeof(elm->internal.subtree_offset));
elm->internal.subtree_offset = nnode->node_offset;
hammer_modify_node_done(cursor->parent);
} else {
/*
* We are the root of the B-Tree
*/
hammer_volume_t volume;
volume = hammer_get_root_volume(cursor->trans->hmp, &error);
KKASSERT(error == 0);
hammer_modify_volume_field(cursor->trans, volume,
vol0_btree_root);
volume->ondisk->vol0_btree_root = nnode->node_offset;
hammer_modify_volume_done(volume);
hammer_rel_volume(volume, 0);
}
hammer_cursor_replaced_node(onode, nnode);
hammer_delete_node(cursor->trans, onode);
if (hammer_debug_general & 0x4000) {
kprintf("REBLOCK LNODE %016llx -> %016llx\n",
(long long)onode->node_offset,
(long long)nnode->node_offset);
}
hammer_modify_node_done(nnode);
cursor->node = nnode;
hammer_unlock(&onode->lock);
hammer_rel_node(onode);
return (error);
}
int
hammer_ioc_volume_add(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_volume *ioc)
{
struct hammer_mount *hmp = trans->hmp;
struct mount *mp = hmp->mp;
hammer_volume_t volume;
int error;
if (mp->mnt_flag & MNT_RDONLY) {
kprintf("Cannot add volume to read-only HAMMER filesystem\n");
return (EINVAL);
}
if (hmp->nvolumes + 1 >= HAMMER_MAX_VOLUMES) {
kprintf("Max number of HAMMER volumes exceeded\n");
return (EINVAL);
}
if (hammer_lock_ex_try(&hmp->volume_lock) != 0) {
kprintf("Another volume operation is in progress!\n");
return (EAGAIN);
}
/*
* Find an unused volume number.
*/
int free_vol_no = 0;
while (free_vol_no < HAMMER_MAX_VOLUMES &&
RB_LOOKUP(hammer_vol_rb_tree, &hmp->rb_vols_root, free_vol_no)) {
++free_vol_no;
}
if (free_vol_no >= HAMMER_MAX_VOLUMES) {
kprintf("Max number of HAMMER volumes exceeded\n");
hammer_unlock(&hmp->volume_lock);
return (EINVAL);
}
struct vnode *devvp = NULL;
error = hammer_setup_device(&devvp, ioc->device_name, 0);
if (error)
goto end;
KKASSERT(devvp);
error = hammer_format_volume_header(
hmp,
devvp,
hmp->rootvol->ondisk->vol_name,
free_vol_no,
hmp->nvolumes+1,
ioc->vol_size,
ioc->boot_area_size,
ioc->mem_area_size);
hammer_close_device(&devvp, 0);
if (error)
goto end;
error = hammer_install_volume(hmp, ioc->device_name, NULL);
if (error)
goto end;
hammer_sync_lock_sh(trans);
hammer_lock_ex(&hmp->blkmap_lock);
++hmp->nvolumes;
/*
* Set each volumes new value of the vol_count field.
*/
for (int vol_no = 0; vol_no < HAMMER_MAX_VOLUMES; ++vol_no) {
volume = hammer_get_volume(hmp, vol_no, &error);
if (volume == NULL && error == ENOENT) {
/*
* Skip unused volume numbers
*/
error = 0;
continue;
}
KKASSERT(volume != NULL && error == 0);
hammer_modify_volume_field(trans, volume, vol_count);
volume->ondisk->vol_count = hmp->nvolumes;
hammer_modify_volume_done(volume);
/*
* Only changes to the header of the root volume
* are automatically flushed to disk. For all
* other volumes that we modify we do it here.
*
* No interlock is needed, volume buffers are not
* messed with by bioops.
*/
if (volume != trans->rootvol && volume->io.modified) {
hammer_crc_set_volume(volume->ondisk);
hammer_io_flush(&volume->io, 0);
}
hammer_rel_volume(volume, 0);
}
volume = hammer_get_volume(hmp, free_vol_no, &error);
KKASSERT(volume != NULL && error == 0);
struct bigblock_stat stat;
error = hammer_format_freemap(trans, volume, &stat);
KKASSERT(error == 0);
/*
* Increase the total number of bigblocks and update stat/vstat totals.
*/
hammer_modify_volume_field(trans, trans->rootvol,
vol0_stat_bigblocks);
trans->rootvol->ondisk->vol0_stat_bigblocks += stat.total_bigblocks;
hammer_modify_volume_done(trans->rootvol);
/*
* Bigblock count changed so recompute the total number of blocks.
*/
mp->mnt_stat.f_blocks = trans->rootvol->ondisk->vol0_stat_bigblocks *
(HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
mp->mnt_vstat.f_blocks = trans->rootvol->ondisk->vol0_stat_bigblocks *
(HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
/*
* Increase the number of free bigblocks
* (including the copy in hmp)
*/
hammer_modify_volume_field(trans, trans->rootvol,
vol0_stat_freebigblocks);
trans->rootvol->ondisk->vol0_stat_freebigblocks += stat.total_free_bigblocks;
hmp->copy_stat_freebigblocks =
trans->rootvol->ondisk->vol0_stat_freebigblocks;
hammer_modify_volume_done(trans->rootvol);
hammer_rel_volume(volume, 0);
hammer_unlock(&hmp->blkmap_lock);
hammer_sync_unlock(trans);
KKASSERT(error == 0);
end:
hammer_unlock(&hmp->volume_lock);
if (error)
kprintf("An error occurred: %d\n", error);
return (error);
}
/*
* Remove a volume.
*/
int
hammer_ioc_volume_del(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_volume *ioc)
{
struct hammer_mount *hmp = trans->hmp;
struct mount *mp = hmp->mp;
hammer_volume_t volume;
int error = 0;
if (mp->mnt_flag & MNT_RDONLY) {
kprintf("Cannot del volume from read-only HAMMER filesystem\n");
return (EINVAL);
}
if (hammer_lock_ex_try(&hmp->volume_lock) != 0) {
kprintf("Another volume operation is in progress!\n");
return (EAGAIN);
}
volume = NULL;
/*
* find volume by volname
*/
for (int vol_no = 0; vol_no < HAMMER_MAX_VOLUMES; ++vol_no) {
volume = hammer_get_volume(hmp, vol_no, &error);
if (volume == NULL && error == ENOENT) {
/*
* Skip unused volume numbers
*/
error = 0;
continue;
}
KKASSERT(volume != NULL && error == 0);
if (strcmp(volume->vol_name, ioc->device_name) == 0) {
break;
}
hammer_rel_volume(volume, 0);
volume = NULL;
}
if (volume == NULL) {
kprintf("Couldn't find volume\n");
error = EINVAL;
goto end;
}
if (volume == trans->rootvol) {
kprintf("Cannot remove root-volume\n");
hammer_rel_volume(volume, 0);
error = EINVAL;
goto end;
}
/*
*
*/
hmp->volume_to_remove = volume->vol_no;
struct hammer_ioc_reblock reblock;
bzero(&reblock, sizeof(reblock));
reblock.key_beg.localization = HAMMER_MIN_LOCALIZATION;
reblock.key_beg.obj_id = HAMMER_MIN_OBJID;
reblock.key_end.localization = HAMMER_MAX_LOCALIZATION;
reblock.key_end.obj_id = HAMMER_MAX_OBJID;
reblock.head.flags = HAMMER_IOC_DO_FLAGS;
reblock.free_level = 0;
error = hammer_ioc_reblock(trans, ip, &reblock);
if (reblock.head.flags & HAMMER_IOC_HEAD_INTR) {
error = EINTR;
}
if (error) {
if (error == EINTR) {
kprintf("reblock was interrupted\n");
} else {
kprintf("reblock failed: %d\n", error);
}
hmp->volume_to_remove = -1;
hammer_rel_volume(volume, 0);
goto end;
}
/*
* Sync filesystem
*/
int count = 0;
while (hammer_flusher_haswork(hmp)) {
hammer_flusher_sync(hmp);
++count;
if (count >= 5) {
if (count == 5)
kprintf("HAMMER: flushing.");
else
kprintf(".");
tsleep(&count, 0, "hmrufl", hz);
}
if (count == 30) {
kprintf("giving up");
break;
}
}
kprintf("\n");
hammer_sync_lock_sh(trans);
hammer_lock_ex(&hmp->blkmap_lock);
/*
* We use stat later to update rootvol's bigblock stats
*/
struct bigblock_stat stat;
error = hammer_free_freemap(trans, volume, &stat);
if (error) {
kprintf("Failed to free volume. Volume not empty!\n");
hmp->volume_to_remove = -1;
hammer_rel_volume(volume, 0);
hammer_unlock(&hmp->blkmap_lock);
hammer_sync_unlock(trans);
goto end;
}
hmp->volume_to_remove = -1;
hammer_rel_volume(volume, 0);
/*
* Unload buffers
*/
RB_SCAN(hammer_buf_rb_tree, &hmp->rb_bufs_root, NULL,
hammer_unload_buffer, volume);
error = hammer_unload_volume(volume, NULL);
if (error == -1) {
kprintf("Failed to unload volume\n");
hammer_unlock(&hmp->blkmap_lock);
hammer_sync_unlock(trans);
goto end;
}
volume = NULL;
--hmp->nvolumes;
/*
* Set each volume's new value of the vol_count field.
*/
for (int vol_no = 0; vol_no < HAMMER_MAX_VOLUMES; ++vol_no) {
volume = hammer_get_volume(hmp, vol_no, &error);
if (volume == NULL && error == ENOENT) {
/*
* Skip unused volume numbers
*/
error = 0;
continue;
}
KKASSERT(volume != NULL && error == 0);
hammer_modify_volume_field(trans, volume, vol_count);
volume->ondisk->vol_count = hmp->nvolumes;
hammer_modify_volume_done(volume);
/*
* Only changes to the header of the root volume
* are automatically flushed to disk. For all
* other volumes that we modify we do it here.
*
* No interlock is needed, volume buffers are not
* messed with by bioops.
*/
if (volume != trans->rootvol && volume->io.modified) {
hammer_crc_set_volume(volume->ondisk);
hammer_io_flush(&volume->io, 0);
}
hammer_rel_volume(volume, 0);
}
/*
* Update the total number of bigblocks
*/
hammer_modify_volume_field(trans, trans->rootvol,
vol0_stat_bigblocks);
trans->rootvol->ondisk->vol0_stat_bigblocks -= stat.total_bigblocks;
hammer_modify_volume_done(trans->rootvol);
/*
* Update the number of free bigblocks
* (including the copy in hmp)
*/
hammer_modify_volume_field(trans, trans->rootvol,
vol0_stat_freebigblocks);
trans->rootvol->ondisk->vol0_stat_freebigblocks -= stat.total_free_bigblocks;
hmp->copy_stat_freebigblocks =
trans->rootvol->ondisk->vol0_stat_freebigblocks;
hammer_modify_volume_done(trans->rootvol);
/*
* Bigblock count changed so recompute the total number of blocks.
*/
mp->mnt_stat.f_blocks = trans->rootvol->ondisk->vol0_stat_bigblocks *
(HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
mp->mnt_vstat.f_blocks = trans->rootvol->ondisk->vol0_stat_bigblocks *
(HAMMER_LARGEBLOCK_SIZE / HAMMER_BUFSIZE);
hammer_unlock(&hmp->blkmap_lock);
hammer_sync_unlock(trans);
/*
* Erase the volume header of the removed device.
*
* This is to not accidentally mount the volume again.
*/
struct vnode *devvp = NULL;
error = hammer_setup_device(&devvp, ioc->device_name, 0);
if (error) {
kprintf("Failed to open device: %s\n", ioc->device_name);
goto end;
}
KKASSERT(devvp);
error = hammer_clear_volume_header(devvp);
if (error) {
kprintf("Failed to clear volume header of device: %s\n",
ioc->device_name);
goto end;
}
hammer_close_device(&devvp, 0);
KKASSERT(error == 0);
end:
hammer_unlock(&hmp->volume_lock);
return (error);
}
void
hammer_sync_unlock(hammer_transaction_t trans)
{
--trans->sync_lock_refs;
hammer_unlock(&trans->hmp->sync_lock);
}
