static
int
hammer_ioc_get_data(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_data *data)
{
struct hammer_cursor cursor;
int bytes;
int error;
/* XXX cached inode ? */
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
if (error)
goto failed;
cursor.key_beg = data->elm;
cursor.flags |= HAMMER_CURSOR_BACKEND;
error = hammer_btree_lookup(&cursor);
if (error == 0) {
error = hammer_btree_extract_data(&cursor);
if (error == 0) {
data->leaf = *cursor.leaf;
bytes = cursor.leaf->data_len;
if (bytes > data->size)
bytes = data->size;
error = copyout(cursor.data, data->ubuf, bytes);
}
}
failed:
hammer_done_cursor(&cursor);
return (error);
}
/*
* Iterate PFS ondisk data.
* This function essentially does the same as hammer_load_pseudofs()
* except that this function only retrieves PFS data without touching
* hammer_pfs_rb_tree at all.
*
* NOTE: The ip used for ioctl is not necessarily related to the PFS
* since this ioctl only requires PFS id (or upper 16 bits of ip localization).
*
* NOTE: The API was changed in DragonFly 4.7, due to design issues
* this ioctl and libhammer (which is the only caller of this ioctl
* within DragonFly source, but no longer maintained by anyone) had.
*/
int
hammer_ioc_scan_pseudofs(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_pseudofs_rw *pfs)
{
struct hammer_cursor cursor;
hammer_inode_t dip;
uint32_t localization;
int error;
if ((error = hammer_pfs_autodetect(pfs, ip)) != 0)
return(error);
localization = pfs_to_lo(pfs->pfs_id);
pfs->bytes = sizeof(struct hammer_pseudofs_data);
pfs->version = HAMMER_IOC_PSEUDOFS_VERSION;
dip = hammer_get_inode(trans, NULL, HAMMER_OBJID_ROOT, HAMMER_MAX_TID,
HAMMER_DEF_LOCALIZATION, 0, &error);
error = hammer_init_cursor(trans, &cursor,
(dip ? &dip->cache[1] : NULL), dip);
if (error)
goto fail;
cursor.key_beg.localization = HAMMER_DEF_LOCALIZATION |
HAMMER_LOCALIZE_MISC;
cursor.key_beg.obj_id = HAMMER_OBJID_ROOT;
cursor.key_beg.create_tid = 0;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.rec_type = HAMMER_RECTYPE_PFS;
cursor.key_beg.obj_type = 0;
cursor.key_beg.key = localization;
cursor.asof = HAMMER_MAX_TID;
cursor.flags |= HAMMER_CURSOR_ASOF;
error = hammer_ip_lookup(&cursor);
if (error == 0) {
error = hammer_ip_resolve_data(&cursor);
if (error == 0) {
if (pfs->ondisk)
copyout(cursor.data, pfs->ondisk, cursor.leaf->data_len);
localization = cursor.leaf->base.key;
pfs->pfs_id = lo_to_pfs(localization);
}
}
hammer_done_cursor(&cursor);
fail:
if (dip)
hammer_rel_inode(dip, 0);
return(error);
}
/*
* Retrieve the PFS hammer cleanup utility config record. This is
* different (newer than) the PFS config.
*/
static
int
hammer_ioc_get_config(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_config *config)
{
struct hammer_cursor cursor;
int error;
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_CONFIG;
cursor.key_beg.localization = ip->obj_localization | HAMMER_LOCALIZE_INODE;
cursor.key_beg.key = 0; /* config space page 0 */
cursor.asof = HAMMER_MAX_TID;
cursor.flags |= HAMMER_CURSOR_ASOF;
error = hammer_btree_lookup(&cursor);
if (error == 0) {
error = hammer_btree_extract_data(&cursor);
if (error == 0)
config->config = cursor.data->config;
}
/* error can be ENOENT */
config->head.error = error;
hammer_done_cursor(&cursor);
return(0);
}
/*
* Rollback the specified PFS to (trunc_tid - 1), removing everything
* greater or equal to trunc_tid. The PFS must not have been in no-mirror
* mode or the MIRROR_FILTERED scan will not work properly.
*
* This is typically used to remove any partial syncs when upgrading a
* slave to a master. It can theoretically also be used to rollback
* any PFS, including PFS#0, BUT ONLY TO POINTS THAT HAVE NOT YET BEEN
* PRUNED, and to points that are older only if they are on a retained
* (pruning softlink) boundary.
*
* Rollbacks destroy information. If you don't mind inode numbers changing
* a better way would be to cpdup a snapshot back onto the master.
*/
static
int
hammer_pfs_rollback(hammer_transaction_t trans,
hammer_pseudofs_inmem_t pfsm,
hammer_tid_t trunc_tid)
{
struct hammer_cmirror cmirror;
struct hammer_cursor cursor;
struct hammer_base_elm key_cur;
int error;
int seq;
bzero(&cmirror, sizeof(cmirror));
bzero(&key_cur, sizeof(key_cur));
key_cur.localization = HAMMER_MIN_LOCALIZATION | pfsm->localization;
key_cur.obj_id = HAMMER_MIN_OBJID;
key_cur.key = HAMMER_MIN_KEY;
key_cur.create_tid = 1;
key_cur.rec_type = HAMMER_MIN_RECTYPE;
seq = trans->hmp->flusher.done;
retry:
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
if (error) {
hammer_done_cursor(&cursor);
goto failed;
}
cursor.key_beg = key_cur;
cursor.key_end.localization = HAMMER_MAX_LOCALIZATION |
pfsm->localization;
cursor.key_end.obj_id = HAMMER_MAX_OBJID;
cursor.key_end.key = HAMMER_MAX_KEY;
cursor.key_end.create_tid = HAMMER_MAX_TID;
cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
cursor.flags |= HAMMER_CURSOR_BACKEND;
/*
* Do an optimized scan of only records created or modified
* >= trunc_tid, so we can fix up those records. We must
* still check the TIDs but this greatly reduces the size of
* the scan.
*/
cursor.flags |= HAMMER_CURSOR_MIRROR_FILTERED;
cursor.cmirror = &cmirror;
cmirror.mirror_tid = trunc_tid;
error = hammer_btree_first(&cursor);
while (error == 0) {
/*
* Abort the rollback.
*/
if (error == 0) {
error = hammer_signal_check(trans->hmp);
if (error)
break;
}
/*
* We only care about leafs. Internal nodes can be returned
* in mirror-filtered mode (they are used to generate SKIP
* mrecords), but we don't need them for this code.
*
* WARNING: See warnings in hammer_unlock_cursor() function.
*/
cursor.flags |= HAMMER_CURSOR_ATEDISK;
if (cursor.node->ondisk->type == HAMMER_BTREE_TYPE_LEAF) {
key_cur = cursor.node->ondisk->elms[cursor.index].base;
error = hammer_pfs_delete_at_cursor(&cursor, trunc_tid);
}
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 (error == 0)
error = hammer_btree_iterate(&cursor);
}
if (error == ENOENT)
error = 0;
hammer_done_cursor(&cursor);
if (error == EDEADLK)
goto retry;
failed:
return(error);
}
/* corresponds to hammer_vop_readdir */
int hammerfs_readdir(struct file *file, void *dirent, filldir_t filldir)
{
struct dentry *de = file->f_dentry;
struct hammer_transaction trans;
struct hammer_cursor cursor;
struct hammer_inode *ip = (struct hammer_inode *)de->d_inode->i_private;
hammer_base_elm_t base;
int r = 0;
int error;
int dtype;
printk(KERN_INFO "hammerfs_readdir(file->f_pos=%lld)\n", file->f_pos);
/*
* Handle artificial entries
*/
if (file->f_pos == 0)
r = filldir(dirent, ".", 1, file->f_pos++,
de->d_inode->i_ino, DT_DIR);
if (!r && file->f_pos == 1) {
if (de->d_parent->d_inode)
r = filldir(dirent, "..", 2, file->f_pos++,
de->d_parent->d_inode->i_ino,
DT_DIR);
else
r = filldir(dirent, "..", 2, file->f_pos++,
de->d_inode->i_ino,
DT_DIR);
}
if (!r) {
hammer_simple_transaction(&trans, ip->hmp);
/*
* Key range (begin and end inclusive) to scan. Directory keys
* directly translate to a 64 bit 'seek' position.
*/
hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
cursor.key_beg.localization = ip->obj_localization +
HAMMER_LOCALIZE_MISC;
cursor.key_beg.obj_id = ip->obj_id;
cursor.key_beg.create_tid = 0;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
cursor.key_beg.obj_type = 0;
cursor.key_beg.key = file->f_pos;
cursor.key_end = cursor.key_beg;
cursor.key_end.key = HAMMER_MAX_KEY;
cursor.asof = ip->obj_asof;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE |
HAMMER_CURSOR_ASOF;
error = hammer_ip_first(&cursor);
while (error == 0) {
error = hammer_ip_resolve_data(&cursor);
if (error)
break;
base = &cursor.leaf->base;
KKASSERT(cursor.leaf->data_len > HAMMER_ENTRY_NAME_OFF);
if (base->obj_id != de->d_inode->i_ino)
panic("readdir: bad record at %p", cursor.node);
/*
* Convert pseudo-filesystems into softlinks
*/
dtype = hammerfs_get_itype(cursor.leaf->base.obj_type);
r = filldir(dirent, (void *)cursor.data->entry.name,
cursor.leaf->data_len - HAMMER_ENTRY_NAME_OFF,
file->f_pos, cursor.data->entry.obj_id, dtype);
if (r)
break;
file->f_pos++;
error = hammer_ip_next(&cursor);
}
hammer_done_cursor(&cursor);
}
/* done: */
/*hammer_done_transaction(&trans);*/
/* return(1); */ /* TODO */
/*failed:
return(1); */
return 0;
}
/* corresponds to hammer_vop_nresolve */
struct dentry *hammerfs_inode_lookup(struct inode *parent_inode,
struct dentry *dentry,
struct nameidata *nameidata)
{
struct hammer_transaction trans;
struct super_block *sb;
struct inode *inode;
hammer_inode_t dip;
hammer_inode_t ip;
hammer_tid_t asof;
struct hammer_cursor cursor;
int64_t namekey;
u_int32_t max_iterations;
int64_t obj_id;
int nlen;
int flags;
int error;
u_int32_t localization;
printk(KERN_INFO "hammerfs_inode_lookup(parent_inode->i_ino=%lu, dentry->d_name.name=%s)\n",
parent_inode->i_ino, dentry->d_name.name);
sb = parent_inode->i_sb;
dip = (hammer_inode_t)parent_inode->i_private;
asof = dip->obj_asof;
localization = dip->obj_localization; /* for code consistency */
nlen = dentry->d_name.len;
flags = dip->flags & HAMMER_INODE_RO;
hammer_simple_transaction(&trans, dip->hmp);
/*
* Calculate the namekey and setup the key range for the scan. This
* works kinda like a chained hash table where the lower 32 bits
* of the namekey synthesize the chain.
*
* The key range is inclusive of both key_beg and key_end.
*/
namekey = hammer_directory_namekey(dip, dentry->d_name.name, nlen,
&max_iterations);
error = hammer_init_cursor(&trans, &cursor, &dip->cache[1], dip);
cursor.key_beg.localization = dip->obj_localization +
HAMMER_LOCALIZE_MISC;
cursor.key_beg.obj_id = dip->obj_id;
cursor.key_beg.key = namekey;
cursor.key_beg.create_tid = 0;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
cursor.key_beg.obj_type = 0;
cursor.key_end = cursor.key_beg;
cursor.key_end.key += max_iterations;
cursor.asof = asof;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE | HAMMER_CURSOR_ASOF;
/*
* Scan all matching records (the chain), locate the one matching
* the requested path component.
*
* The hammer_ip_*() functions merge in-memory records with on-disk
* records for the purposes of the search.
*/
obj_id = 0;
localization = HAMMER_DEF_LOCALIZATION;
if (error == 0) {
error = hammer_ip_first(&cursor);
while (error == 0) {
error = hammer_ip_resolve_data(&cursor);
if (error)
break;
if (nlen == cursor.leaf->data_len -
HAMMER_ENTRY_NAME_OFF &&
bcmp(dentry->d_name.name,
cursor.data->entry.name, nlen) == 0) {
obj_id = cursor.data->entry.obj_id;
localization = cursor.data->entry.localization;
break;
}
error = hammer_ip_next(&cursor);
}
}
hammer_done_cursor(&cursor);
if (error == 0) {
ip = hammer_get_inode(&trans, dip, obj_id,
asof, localization,
flags, &error);
if (error == 0)
error = hammerfs_get_inode(sb, ip, &inode);
/* hammer_rel_inode(ip, 0); */
else
ip = NULL;
if (error == 0)
d_add(dentry, inode);
goto done;
}
done:
/*hammer_done_transaction(&trans);*/
return NULL;
}
/*
* Retrieve the PFS hammer cleanup utility config record. This is
* different (newer than) the PFS config.
*
* This is kinda a hack.
*/
static
int
hammer_ioc_set_config(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_config *config)
{
struct hammer_btree_leaf_elm leaf;
struct hammer_cursor cursor;
hammer_mount_t hmp = ip->hmp;
int error;
again:
error = hammer_init_cursor(trans, &cursor, &ip->cache[0], NULL);
if (error) {
hammer_done_cursor(&cursor);
return(error);
}
bzero(&leaf, sizeof(leaf));
leaf.base.obj_id = HAMMER_OBJID_ROOT;
leaf.base.rec_type = HAMMER_RECTYPE_CONFIG;
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.base.key = 0; /* page 0 */
leaf.data_len = sizeof(struct hammer_config_data);
cursor.key_beg = leaf.base;
cursor.asof = HAMMER_MAX_TID;
cursor.flags |= HAMMER_CURSOR_BACKEND | HAMMER_CURSOR_ASOF;
error = hammer_btree_lookup(&cursor);
if (error == 0) {
error = hammer_btree_extract_data(&cursor);
error = hammer_delete_at_cursor(&cursor, HAMMER_DELETE_DESTROY,
0, 0, 0, NULL);
if (error == EDEADLK) {
hammer_done_cursor(&cursor);
goto again;
}
}
if (error == ENOENT)
error = 0;
if (error == 0) {
/*
* 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,
&config->config,
HAMMER_CREATE_MODE_SYS);
if (error == EDEADLK) {
hammer_done_cursor(&cursor);
goto again;
}
}
config->head.error = error;
hammer_done_cursor(&cursor);
return(0);
}
// corresponds to hammer_vop_strategy_read
int hammerfs_readpage(struct file *file, struct page *page)
{
void *page_addr;
hammer_mount_t hmp;
struct buffer_head *bh;
struct super_block *sb;
struct hammer_transaction trans;
struct hammer_cursor cursor;
struct inode *inode;
struct hammer_inode *ip;
hammer_base_elm_t base;
hammer_off_t disk_offset;
int64_t rec_offset;
int64_t file_offset;
int error = 0;
int boff;
int roff;
int n;
int i=0;
int block_num;
int block_offset;
int bytes_read;
int64_t sb_offset;
hammer_off_t zone2_offset;
int vol_no;
hammer_volume_t volume;
printk ("hammerfs_readpage(page->index=%d)\n", (int) page->index);
inode = file->f_path.dentry->d_inode;
ip = (struct hammer_inode *)inode->i_private;
sb = inode->i_sb;
hmp = (hammer_mount_t)sb->s_fs_info;
hammer_simple_transaction(&trans, ip->hmp);
hammer_init_cursor(&trans, &cursor, &ip->cache[1], ip);
file_offset = page->index * PAGE_SIZE;
if (file_offset > inode->i_size) {
error = -ENOSPC;
goto done;
}
SetPageUptodate (page);
page_addr = kmap (page);
if(!page_addr) {
error = -ENOSPC;
goto failed;
}
/*
* Key range (begin and end inclusive) to scan. Note that the key's
* stored in the actual records represent BASE+LEN, not BASE. The
* first record containing bio_offset will have a key > bio_offset.
*/
cursor.key_beg.localization = ip->obj_localization +
HAMMER_LOCALIZE_MISC;
cursor.key_beg.obj_id = ip->obj_id;
cursor.key_beg.create_tid = 0;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.obj_type = 0;
cursor.key_beg.key = file_offset + 1;
cursor.asof = ip->obj_asof;
cursor.flags |= HAMMER_CURSOR_ASOF;
cursor.key_end = cursor.key_beg;
KKASSERT(ip->ino_data.obj_type == HAMMER_OBJTYPE_REGFILE);
cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
cursor.key_end.rec_type = HAMMER_RECTYPE_DATA;
cursor.key_end.key = 0x7FFFFFFFFFFFFFFFLL;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
error = hammer_ip_first(&cursor);
boff = 0;
while(error == 0) {
/*
* Get the base file offset of the record. The key for
* data records is (base + bytes) rather then (base).
*/
base = &cursor.leaf->base;
rec_offset = base->key - cursor.leaf->data_len;
/*
* Calculate the gap, if any, and zero-fill it.
*
* n is the offset of the start of the record verses our
* current seek offset in the bio.
*/
n = (int)(rec_offset - (file_offset + boff));
if (n > 0) {
if (n > PAGE_SIZE - boff)
n = PAGE_SIZE - boff;
bzero((char *)page_addr + boff, n);
boff += n;
n = 0;
}
/*
* Calculate the data offset in the record and the number
* of bytes we can copy.
*
* There are two degenerate cases. First, boff may already
* be at bp->b_bufsize. Secondly, the data offset within
* the record may exceed the record's size.
*/
roff = -n;
rec_offset += roff;
n = cursor.leaf->data_len - roff;
if (n <= 0) {
printk("hammerfs_readpage: bad n=%d roff=%d\n", n, roff);
n = 0;
} else if (n > PAGE_SIZE - boff) {
n = PAGE_SIZE - boff;
}
/*
* Deal with cached truncations. This cool bit of code
* allows truncate()/ftruncate() to avoid having to sync
* the file.
*
* If the frontend is truncated then all backend records are
* subject to the frontend's truncation.
*
* If the backend is truncated then backend records on-disk
* (but not in-memory) are subject to the backend's
* truncation. In-memory records owned by the backend
* represent data written after the truncation point on the
* backend and must not be truncated.
*
* Truncate operations deal with frontend buffer cache
* buffers and frontend-owned in-memory records synchronously.
*/
if (ip->flags & HAMMER_INODE_TRUNCATED) {
if (hammer_cursor_ondisk(&cursor) ||
cursor.iprec->flush_state == HAMMER_FST_FLUSH) {
if (ip->trunc_off <= rec_offset)
n = 0;
else if (ip->trunc_off < rec_offset + n)
n = (int)(ip->trunc_off - rec_offset);
}
}
if (ip->sync_flags & HAMMER_INODE_TRUNCATED) {
if (hammer_cursor_ondisk(&cursor)) {
if (ip->sync_trunc_off <= rec_offset)
n = 0;
else if (ip->sync_trunc_off < rec_offset + n)
n = (int)(ip->sync_trunc_off - rec_offset);
}
}
/*
* Calculate the data offset in the record and the number
* of bytes we can copy.
*/
disk_offset = cursor.leaf->data_offset + roff;
// move this to hammerfs_direct_io_read
zone2_offset = hammer_blockmap_lookup(hmp, disk_offset, &error);
vol_no = HAMMER_VOL_DECODE(zone2_offset);
volume = hammer_get_volume(hmp, vol_no, &error);
// n is the number of bytes we should read, sb_offset the
// offset on disk
sb_offset = volume->ondisk->vol_buf_beg + (zone2_offset & HAMMER_OFF_SHORT_MASK);
while(n > 0 && boff != PAGE_SIZE) {
block_num = sb_offset / BLOCK_SIZE;
block_offset = sb_offset % BLOCK_SIZE;
// the minimum between what is available and what we can maximally provide
bytes_read = min(BLOCK_SIZE - (int )block_offset, PAGE_SIZE - (int )boff);
bh = sb_bread(sb, block_num + i);
if(!bh) {
error = -ENOMEM;
goto failed;
}
memcpy((char*)page_addr + roff, (char*)bh->b_data + boff + block_offset, bytes_read);
brelse(bh);
n -= bytes_read;
boff += bytes_read;
roff += bytes_read;
}
/*
* Iterate until we have filled the request.
*/
if (boff == PAGE_SIZE)
break;
error = hammer_ip_next(&cursor);
}
hammer_done_cursor(&cursor);
hammer_done_transaction(&trans);
failed:
if (PageLocked (page))
unlock_page (page);
kunmap (page);
done:
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);
}
static
int
hammer_ioc_gethistory(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_history *hist)
{
struct hammer_cursor cursor;
hammer_btree_elm_t elm;
int error;
/*
* Validate the structure and initialize for return.
*/
if (hist->beg_tid > hist->end_tid)
return(EINVAL);
if (hist->head.flags & HAMMER_IOC_HISTORY_ATKEY) {
if (hist->key > hist->nxt_key)
return(EINVAL);
}
hist->obj_id = ip->obj_id;
hist->count = 0;
hist->nxt_tid = hist->end_tid;
hist->head.flags &= ~HAMMER_IOC_HISTORY_NEXT_TID;
hist->head.flags &= ~HAMMER_IOC_HISTORY_NEXT_KEY;
hist->head.flags &= ~HAMMER_IOC_HISTORY_EOF;
hist->head.flags &= ~HAMMER_IOC_HISTORY_UNSYNCED;
if ((ip->flags & HAMMER_INODE_MODMASK) &
~(HAMMER_INODE_ATIME | HAMMER_INODE_MTIME)) {
hist->head.flags |= HAMMER_IOC_HISTORY_UNSYNCED;
}
/*
* Setup the cursor. We can't handle undeletable records
* (create_tid of 0) at the moment. A create_tid of 0 has
* a special meaning and cannot be specified in the cursor.
*/
error = hammer_init_cursor(trans, &cursor, &ip->cache[0], NULL);
if (error) {
hammer_done_cursor(&cursor);
return(error);
}
cursor.key_beg.obj_id = hist->obj_id;
cursor.key_beg.create_tid = hist->beg_tid;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.obj_type = 0;
if (cursor.key_beg.create_tid == HAMMER_MIN_TID)
cursor.key_beg.create_tid = 1;
cursor.key_end.obj_id = hist->obj_id;
cursor.key_end.create_tid = hist->end_tid;
cursor.key_end.delete_tid = 0;
cursor.key_end.obj_type = 0;
cursor.flags |= HAMMER_CURSOR_END_EXCLUSIVE;
if (hist->head.flags & HAMMER_IOC_HISTORY_ATKEY) {
/*
* key-range within the file. For a regular file the
* on-disk key represents BASE+LEN, not BASE, so the
* first possible record containing the offset 'key'
* has an on-disk key of (key + 1).
*/
cursor.key_beg.key = hist->key;
cursor.key_end.key = HAMMER_MAX_KEY;
cursor.key_beg.localization = ip->obj_localization |
HAMMER_LOCALIZE_MISC;
cursor.key_end.localization = ip->obj_localization |
HAMMER_LOCALIZE_MISC;
switch(ip->ino_data.obj_type) {
case HAMMER_OBJTYPE_REGFILE:
++cursor.key_beg.key;
cursor.key_beg.rec_type = HAMMER_RECTYPE_DATA;
break;
case HAMMER_OBJTYPE_DIRECTORY:
cursor.key_beg.rec_type = HAMMER_RECTYPE_DIRENTRY;
cursor.key_beg.localization = ip->obj_localization |
hammer_dir_localization(ip);
cursor.key_end.localization = ip->obj_localization |
hammer_dir_localization(ip);
break;
case HAMMER_OBJTYPE_DBFILE:
cursor.key_beg.rec_type = HAMMER_RECTYPE_DB;
break;
default:
error = EINVAL;
break;
}
cursor.key_end.rec_type = cursor.key_beg.rec_type;
} else {
/*
* The inode itself.
*/
cursor.key_beg.key = 0;
cursor.key_end.key = 0;
cursor.key_beg.rec_type = HAMMER_RECTYPE_INODE;
cursor.key_end.rec_type = HAMMER_RECTYPE_INODE;
cursor.key_beg.localization = ip->obj_localization |
HAMMER_LOCALIZE_INODE;
cursor.key_end.localization = ip->obj_localization |
HAMMER_LOCALIZE_INODE;
}
error = hammer_btree_first(&cursor);
while (error == 0) {
elm = &cursor.node->ondisk->elms[cursor.index];
add_history(ip, hist, elm);
if (hist->head.flags & (HAMMER_IOC_HISTORY_NEXT_TID |
HAMMER_IOC_HISTORY_NEXT_KEY |
HAMMER_IOC_HISTORY_EOF)) {
break;
}
error = hammer_btree_iterate(&cursor);
}
if (error == ENOENT) {
hist->head.flags |= HAMMER_IOC_HISTORY_EOF;
error = 0;
}
hammer_done_cursor(&cursor);
return(error);
}
int
hammer_ioc_prune(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_prune *prune)
{
struct hammer_cursor cursor;
hammer_btree_leaf_elm_t elm;
struct hammer_ioc_prune_elm *copy_elms;
struct hammer_ioc_prune_elm *user_elms;
int error;
int isdir;
int elm_array_size;
int seq;
if (prune->nelms < 0 || prune->nelms > HAMMER_MAX_PRUNE_ELMS)
return(EINVAL);
if ((prune->key_beg.localization | prune->key_end.localization) &
HAMMER_LOCALIZE_PSEUDOFS_MASK) {
return(EINVAL);
}
if (prune->key_beg.localization > prune->key_end.localization)
return(EINVAL);
if (prune->key_beg.localization == prune->key_end.localization) {
if (prune->key_beg.obj_id > prune->key_end.obj_id)
return(EINVAL);
/* key-space limitations - no check needed */
}
if ((prune->head.flags & HAMMER_IOC_PRUNE_ALL) && prune->nelms)
return(EINVAL);
/* 22 EINVAL */
prune->key_cur.localization = (prune->key_end.localization &
HAMMER_LOCALIZE_MASK) +
ip->obj_localization;
prune->key_cur.obj_id = prune->key_end.obj_id;
prune->key_cur.key = HAMMER_MAX_KEY;
/*
* Copy element array from userland
*/
elm_array_size = sizeof(*copy_elms) * prune->nelms;
user_elms = prune->elms;
copy_elms = kmalloc(elm_array_size, M_TEMP, M_WAITOK);
if ((error = copyin(user_elms, copy_elms, elm_array_size)) != 0)
goto failed;
prune->elms = copy_elms;
seq = trans->hmp->flusher.done;
/*
* Scan backwards. Retries typically occur if a deadlock is detected.
*/
retry:
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
if (error) {
hammer_done_cursor(&cursor);
goto failed;
}
cursor.key_beg.localization = (prune->key_beg.localization &
HAMMER_LOCALIZE_MASK) +
ip->obj_localization;
cursor.key_beg.obj_id = prune->key_beg.obj_id;
cursor.key_beg.key = HAMMER_MIN_KEY;
cursor.key_beg.create_tid = 1;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
cursor.key_beg.obj_type = 0;
cursor.key_end.localization = prune->key_cur.localization;
cursor.key_end.obj_id = prune->key_cur.obj_id;
cursor.key_end.key = prune->key_cur.key;
cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
cursor.key_end.delete_tid = 0;
cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
cursor.key_end.obj_type = 0;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
cursor.flags |= HAMMER_CURSOR_BACKEND;
/*
* This flag allows the B-Tree code to clean up loose ends. At
* the moment (XXX) it also means we have to hold the sync lock
* through the iteration.
*/
cursor.flags |= HAMMER_CURSOR_PRUNING;
hammer_sync_lock_sh(trans);
error = hammer_btree_last(&cursor);
hammer_sync_unlock(trans);
while (error == 0) {
/*
* Check for work
*/
elm = &cursor.node->ondisk->elms[cursor.index].leaf;
prune->key_cur = elm->base;
/*
* Yield to more important tasks
*/
if ((error = hammer_signal_check(trans->hmp)) != 0)
break;
if (prune->stat_oldest_tid > elm->base.create_tid)
prune->stat_oldest_tid = elm->base.create_tid;
if (hammer_debug_general & 0x0200) {
kprintf("check %016llx %016llx cre=%016llx del=%016llx\n",
(long long)elm->base.obj_id,
(long long)elm->base.key,
(long long)elm->base.create_tid,
(long long)elm->base.delete_tid);
}
if (prune_should_delete(prune, elm)) {
if (hammer_debug_general & 0x0200) {
kprintf("check %016llx %016llx: DELETE\n",
(long long)elm->base.obj_id,
(long long)elm->base.key);
}
/*
* NOTE: This can return EDEADLK
*
* Acquiring the sync lock guarantees that the
* operation will not cross a synchronization
* boundary (see the flusher).
*
* We dont need to track inodes or next_tid when
* we are destroying deleted records.
*/
isdir = (elm->base.rec_type == HAMMER_RECTYPE_DIRENTRY);
hammer_sync_lock_sh(trans);
error = hammer_delete_at_cursor(&cursor,
HAMMER_DELETE_DESTROY,
cursor.trans->tid,
cursor.trans->time32,
0, &prune->stat_bytes);
hammer_sync_unlock(trans);
if (error)
break;
if (isdir)
++prune->stat_dirrecords;
else
++prune->stat_rawrecords;
/*
* The current record might now be the one after
* the one we deleted, set ATEDISK to force us
* to skip it (since we are iterating backwards).
*/
cursor.flags |= HAMMER_CURSOR_ATEDISK;
} else {
/*
* Nothing to delete, but we may have to check other
* things.
*/
prune_check_nlinks(&cursor, elm);
cursor.flags |= HAMMER_CURSOR_ATEDISK;
if (hammer_debug_general & 0x0100) {
kprintf("check %016llx %016llx: SKIP\n",
(long long)elm->base.obj_id,
(long long)elm->base.key);
}
}
++prune->stat_scanrecords;
/*
* 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);
}
hammer_sync_lock_sh(trans);
error = hammer_btree_iterate_reverse(&cursor);
hammer_sync_unlock(trans);
}
if (error == ENOENT)
error = 0;
hammer_done_cursor(&cursor);
if (error == EDEADLK)
goto retry;
if (error == EINTR) {
prune->head.flags |= HAMMER_IOC_HEAD_INTR;
error = 0;
}
failed:
prune->key_cur.localization &= HAMMER_LOCALIZE_MASK;
prune->elms = user_elms;
kfree(copy_elms, M_TEMP);
return(error);
}
int
hammer_ioc_reblock(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_reblock *reblock)
{
struct hammer_cursor cursor;
hammer_btree_elm_t elm;
int checkspace_count;
int error;
int seq;
int slop;
/*
* A fill level <= 20% is considered an emergency. free_level is
* inverted from fill_level.
*/
if (reblock->free_level >= HAMMER_LARGEBLOCK_SIZE * 8 / 10)
slop = HAMMER_CHKSPC_EMERGENCY;
else
slop = HAMMER_CHKSPC_REBLOCK;
if ((reblock->key_beg.localization | reblock->key_end.localization) &
HAMMER_LOCALIZE_PSEUDOFS_MASK) {
return(EINVAL);
}
if (reblock->key_beg.obj_id >= reblock->key_end.obj_id)
return(EINVAL);
if (reblock->free_level < 0)
return(EINVAL);
reblock->key_cur = reblock->key_beg;
reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
reblock->key_cur.localization += ip->obj_localization;
checkspace_count = 0;
seq = trans->hmp->flusher.done;
retry:
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
if (error) {
hammer_done_cursor(&cursor);
goto failed;
}
cursor.key_beg.localization = reblock->key_cur.localization;
cursor.key_beg.obj_id = reblock->key_cur.obj_id;
cursor.key_beg.key = HAMMER_MIN_KEY;
cursor.key_beg.create_tid = 1;
cursor.key_beg.delete_tid = 0;
cursor.key_beg.rec_type = HAMMER_MIN_RECTYPE;
cursor.key_beg.obj_type = 0;
cursor.key_end.localization = (reblock->key_end.localization &
HAMMER_LOCALIZE_MASK) +
ip->obj_localization;
cursor.key_end.obj_id = reblock->key_end.obj_id;
cursor.key_end.key = HAMMER_MAX_KEY;
cursor.key_end.create_tid = HAMMER_MAX_TID - 1;
cursor.key_end.delete_tid = 0;
cursor.key_end.rec_type = HAMMER_MAX_RECTYPE;
cursor.key_end.obj_type = 0;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
cursor.flags |= HAMMER_CURSOR_BACKEND;
cursor.flags |= HAMMER_CURSOR_NOSWAPCACHE;
/*
* This flag allows the btree scan code to return internal nodes,
* so we can reblock them in addition to the leafs. Only specify it
* if we intend to reblock B-Tree nodes.
*/
if (reblock->head.flags & HAMMER_IOC_DO_BTREE)
cursor.flags |= HAMMER_CURSOR_REBLOCKING;
error = hammer_btree_first(&cursor);
while (error == 0) {
/*
* Internal or Leaf node
*/
KKASSERT(cursor.index < cursor.node->ondisk->count);
elm = &cursor.node->ondisk->elms[cursor.index];
reblock->key_cur.obj_id = elm->base.obj_id;
reblock->key_cur.localization = elm->base.localization;
/*
* Yield to more important tasks
*/
if ((error = hammer_signal_check(trans->hmp)) != 0)
break;
/*
* If there is insufficient free space it may be due to
* reserved bigblocks, which flushing might fix.
*
* We must force a retest in case the unlocked cursor is
* moved to the end of the leaf, or moved to an internal
* node.
*
* WARNING: See warnings in hammer_unlock_cursor() function.
*/
if (hammer_checkspace(trans->hmp, slop)) {
if (++checkspace_count == 10) {
error = ENOSPC;
break;
}
hammer_unlock_cursor(&cursor);
cursor.flags |= HAMMER_CURSOR_RETEST;
hammer_flusher_wait(trans->hmp, seq);
hammer_lock_cursor(&cursor);
seq = hammer_flusher_async(trans->hmp, NULL);
goto skip;
}
/*
* Acquiring the sync_lock prevents the operation from
* crossing a synchronization boundary.
*
* NOTE: cursor.node may have changed on return.
*
* WARNING: See warnings in hammer_unlock_cursor() function.
*/
hammer_sync_lock_sh(trans);
error = hammer_reblock_helper(reblock, &cursor, elm);
hammer_sync_unlock(trans);
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);
}
/*
* Setup for iteration, our cursor flags may be modified by
* other threads while we are unlocked.
*/
cursor.flags |= HAMMER_CURSOR_ATEDISK;
/*
* We allocate data buffers, which atm we don't track
* dirty levels for because we allow the kernel to write
* them. But if we allocate too many we can still deadlock
* the buffer cache.
*
* WARNING: See warnings in hammer_unlock_cursor() function.
* (The cursor's node and element may change!)
*/
if (bd_heatup()) {
hammer_unlock_cursor(&cursor);
bwillwrite(HAMMER_XBUFSIZE);
hammer_lock_cursor(&cursor);
}
/* XXX vm_wait_nominal(); */
skip:
if (error == 0) {
error = hammer_btree_iterate(&cursor);
}
}
if (error == ENOENT)
error = 0;
hammer_done_cursor(&cursor);
if (error == EWOULDBLOCK) {
hammer_flusher_sync(trans->hmp);
goto retry;
}
if (error == EDEADLK)
goto retry;
if (error == EINTR) {
reblock->head.flags |= HAMMER_IOC_HEAD_INTR;
error = 0;
}
failed:
reblock->key_cur.localization &= HAMMER_LOCALIZE_MASK;
return(error);
}
/*
* 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);
}
/*
* Retrieve as many snapshot ids as possible or until the array is
* full, starting after the last transaction id passed in. If count
* is 0 we retrieve starting at the beginning.
*
* NOTE: Because the b-tree key field is signed but transaction ids
* are unsigned the returned list will be signed-sorted instead
* of unsigned sorted. The Caller must still sort the aggregate
* results.
*/
static
int
hammer_ioc_get_snapshot(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_snapshot *snap)
{
struct hammer_cursor cursor;
int error;
/*
* Validate structure
*/
if (snap->index != 0)
return (EINVAL);
if (snap->count > HAMMER_SNAPS_PER_IOCTL)
return (EINVAL);
/*
* 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;
if (snap->count == 0)
cursor.key_beg.key = HAMMER_MIN_KEY;
else
cursor.key_beg.key = (int64_t)snap->snaps[snap->count - 1].tid + 1;
cursor.key_end = cursor.key_beg;
cursor.key_end.key = HAMMER_MAX_KEY;
cursor.asof = HAMMER_MAX_TID;
cursor.flags |= HAMMER_CURSOR_END_EXCLUSIVE | HAMMER_CURSOR_ASOF;
snap->count = 0;
error = hammer_btree_first(&cursor);
while (error == 0 && snap->count < HAMMER_SNAPS_PER_IOCTL) {
error = hammer_btree_extract_leaf(&cursor);
if (error)
break;
if (cursor.leaf->base.rec_type == HAMMER_RECTYPE_SNAPSHOT) {
error = hammer_btree_extract_data(&cursor);
snap->snaps[snap->count] = cursor.data->snap;
/*
* The snap data tid should match the key but might
* not due to a bug in the HAMMER v3 conversion code.
*
* This error will work itself out over time but we
* have to force a match or the snapshot will not
* be deletable.
*/
if (cursor.data->snap.tid !=
(hammer_tid_t)cursor.leaf->base.key) {
hkprintf("lo=%08x snapshot key "
"0x%016jx data mismatch 0x%016jx\n",
cursor.key_beg.localization,
(uintmax_t)cursor.data->snap.tid,
cursor.leaf->base.key);
hkprintf("Probably left over from the "
"original v3 conversion, hammer "
"cleanup should get it eventually\n");
snap->snaps[snap->count].tid =
cursor.leaf->base.key;
}
++snap->count;
}
error = hammer_btree_iterate(&cursor);
}
if (error == ENOENT) {
snap->head.flags |= HAMMER_IOC_SNAPSHOT_EOF;
error = 0;
}
snap->head.error = error;
hammer_done_cursor(&cursor);
return(0);
}
/*
* All B-Tree records within the specified key range which also conform
* to the transaction id range are returned. Mirroring code keeps track
* of the last transaction id fully scanned and can efficiently pick up
* where it left off if interrupted.
*
* 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.
*/
int
hammer_ioc_mirror_read(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_mirror_rw *mirror)
{
struct hammer_cmirror cmirror;
struct hammer_cursor cursor;
union hammer_ioc_mrecord_any mrec;
hammer_btree_leaf_elm_t elm;
const int crc_start = HAMMER_MREC_CRCOFF;
char *uptr;
int error;
int data_len;
int bytes;
int eatdisk;
int mrec_flags;
u_int32_t localization;
u_int32_t rec_crc;
localization = (u_int32_t)mirror->pfs_id << 16;
if ((mirror->key_beg.localization | mirror->key_end.localization) &
HAMMER_LOCALIZE_PSEUDOFS_MASK) {
return(EINVAL);
}
if (hammer_btree_cmp(&mirror->key_beg, &mirror->key_end) > 0)
return(EINVAL);
mirror->key_cur = mirror->key_beg;
mirror->key_cur.localization &= HAMMER_LOCALIZE_MASK;
mirror->key_cur.localization += localization;
bzero(&mrec, sizeof(mrec));
bzero(&cmirror, sizeof(cmirror));
/*
* Make CRC errors non-fatal (at least on data), causing an EDOM
* error instead of EIO.
*/
trans->flags |= HAMMER_TRANSF_CRCDOM;
retry:
error = hammer_init_cursor(trans, &cursor, NULL, NULL);
if (error) {
hammer_done_cursor(&cursor);
goto failed;
}
cursor.key_beg = mirror->key_cur;
cursor.key_end = mirror->key_end;
cursor.key_end.localization &= HAMMER_LOCALIZE_MASK;
cursor.key_end.localization += localization;
cursor.flags |= HAMMER_CURSOR_END_INCLUSIVE;
cursor.flags |= HAMMER_CURSOR_BACKEND;
/*
* This flag filters the search to only return elements whos create
* or delete TID is >= mirror_tid. The B-Tree uses the mirror_tid
* field stored with internal and leaf nodes to shortcut the scan.
*/
cursor.flags |= HAMMER_CURSOR_MIRROR_FILTERED;
cursor.cmirror = &cmirror;
cmirror.mirror_tid = mirror->tid_beg;
error = hammer_btree_first(&cursor);
while (error == 0) {
/*
* Yield to more important tasks
*/
if (error == 0) {
error = hammer_signal_check(trans->hmp);
if (error)
break;
}
/*
* An internal node can be returned in mirror-filtered
* mode and indicates that the scan is returning a skip
* range in the cursor->cmirror structure.
*/
uptr = (char *)mirror->ubuf + mirror->count;
if (cursor.node->ondisk->type == HAMMER_BTREE_TYPE_INTERNAL) {
/*
* Check space
*/
mirror->key_cur = cmirror.skip_beg;
bytes = sizeof(mrec.skip);
if (mirror->count + HAMMER_HEAD_DOALIGN(bytes) >
mirror->size) {
break;
}
/*
* Fill mrec
*/
mrec.head.signature = HAMMER_IOC_MIRROR_SIGNATURE;
mrec.head.type = HAMMER_MREC_TYPE_SKIP;
mrec.head.rec_size = bytes;
mrec.skip.skip_beg = cmirror.skip_beg;
mrec.skip.skip_end = cmirror.skip_end;
mrec.head.rec_crc = crc32(&mrec.head.rec_size,
bytes - crc_start);
error = copyout(&mrec, uptr, bytes);
eatdisk = 0;
goto didwrite;
}
/*
* Leaf node. In full-history mode we could filter out
* elements modified outside the user-requested TID range.
*
* However, such elements must be returned so the writer
* can compare them against the target to determine what
* needs to be deleted on the target, particular for
* no-history mirrors.
*/
KKASSERT(cursor.node->ondisk->type == HAMMER_BTREE_TYPE_LEAF);
elm = &cursor.node->ondisk->elms[cursor.index].leaf;
mirror->key_cur = elm->base;
/*
* If the record was created after our end point we just
* ignore it.
*/
if (elm->base.create_tid > mirror->tid_end) {
error = 0;
bytes = 0;
eatdisk = 1;
goto didwrite;
}
/*
* Determine if we should generate a PASS or a REC. PASS
* records are records without any data payload. Such
* records will be generated if the target is already expected
* to have the record, allowing it to delete the gaps.
*
* A PASS record is also used to perform deletions on the
* target.
*
* Such deletions are needed if the master or files on the
* master are no-history, or if the slave is so far behind
* the master has already been pruned.
*/
if (elm->base.create_tid < mirror->tid_beg) {
bytes = sizeof(mrec.rec);
if (mirror->count + HAMMER_HEAD_DOALIGN(bytes) >
mirror->size) {
break;
}
/*
* Fill mrec.
*/
mrec.head.signature = HAMMER_IOC_MIRROR_SIGNATURE;
mrec.head.type = HAMMER_MREC_TYPE_PASS;
mrec.head.rec_size = bytes;
mrec.rec.leaf = *elm;
mrec.head.rec_crc = crc32(&mrec.head.rec_size,
bytes - crc_start);
error = copyout(&mrec, uptr, bytes);
eatdisk = 1;
goto didwrite;
}
/*
* The core code exports the data to userland.
*
* CRC errors on data are reported but passed through,
* but the data must be washed by the user program.
*
* If userland just wants the btree records it can
* request that bulk data not be returned. This is
* use during mirror-stream histogram generation.
*/
mrec_flags = 0;
data_len = (elm->data_offset) ? elm->data_len : 0;
if (data_len &&
(mirror->head.flags & HAMMER_IOC_MIRROR_NODATA)) {
data_len = 0;
mrec_flags |= HAMMER_MRECF_NODATA;
}
if (data_len) {
error = hammer_btree_extract(&cursor,
HAMMER_CURSOR_GET_DATA);
if (error) {
if (error != EDOM)
break;
mrec_flags |= HAMMER_MRECF_CRC_ERROR |
HAMMER_MRECF_DATA_CRC_BAD;
}
}
bytes = sizeof(mrec.rec) + data_len;
if (mirror->count + HAMMER_HEAD_DOALIGN(bytes) > mirror->size)
break;
/*
* Construct the record for userland and copyout.
*
* The user is asking for a snapshot, if the record was
* deleted beyond the user-requested ending tid, the record
* is not considered deleted from the point of view of
* userland and delete_tid is cleared.
*/
mrec.head.signature = HAMMER_IOC_MIRROR_SIGNATURE;
mrec.head.type = HAMMER_MREC_TYPE_REC | mrec_flags;
mrec.head.rec_size = bytes;
mrec.rec.leaf = *elm;
if (elm->base.delete_tid > mirror->tid_end)
mrec.rec.leaf.base.delete_tid = 0;
rec_crc = crc32(&mrec.head.rec_size,
sizeof(mrec.rec) - crc_start);
if (data_len)
rec_crc = crc32_ext(cursor.data, data_len, rec_crc);
mrec.head.rec_crc = rec_crc;
error = copyout(&mrec, uptr, sizeof(mrec.rec));
if (data_len && error == 0) {
error = copyout(cursor.data, uptr + sizeof(mrec.rec),
data_len);
}
eatdisk = 1;
/*
* eatdisk controls whether we skip the current cursor
* position on the next scan or not. If doing a SKIP
* the cursor is already positioned properly for the next
* scan and eatdisk will be 0.
*/
didwrite:
if (error == 0) {
mirror->count += HAMMER_HEAD_DOALIGN(bytes);
if (eatdisk)
cursor.flags |= HAMMER_CURSOR_ATEDISK;
else
cursor.flags &= ~HAMMER_CURSOR_ATEDISK;
error = hammer_btree_iterate(&cursor);
}
}
if (error == ENOENT) {
mirror->key_cur = mirror->key_end;
error = 0;
}
hammer_done_cursor(&cursor);
if (error == EDEADLK)
goto retry;
if (error == EINTR) {
mirror->head.flags |= HAMMER_IOC_HEAD_INTR;
error = 0;
}
failed:
mirror->key_cur.localization &= HAMMER_LOCALIZE_MASK;
return(error);
}
int
hammer_ioc_dedup(hammer_transaction_t trans, hammer_inode_t ip,
struct hammer_ioc_dedup *dedup)
{
struct hammer_cursor cursor1, cursor2;
int error;
int seq;
/*
* Enforce hammer filesystem version requirements
*/
if (trans->hmp->version < HAMMER_VOL_VERSION_FIVE) {
kprintf("hammer: Filesystem must be upgraded to v5 "
"before you can run dedup\n");
return (EOPNOTSUPP); /* 95*/
}
/*
* Cursor1, return an error -> candidate goes to pass2 list
*/
error = hammer_init_cursor(trans, &cursor1, NULL, NULL);
if (error)
goto done_cursor;
cursor1.key_beg = dedup->elm1;
cursor1.flags |= HAMMER_CURSOR_BACKEND;
error = hammer_btree_lookup(&cursor1);
if (error)
goto done_cursor;
error = hammer_btree_extract(&cursor1, HAMMER_CURSOR_GET_LEAF |
HAMMER_CURSOR_GET_DATA);
if (error)
goto done_cursor;
/*
* Cursor2, return an error -> candidate goes to pass2 list
*/
error = hammer_init_cursor(trans, &cursor2, NULL, NULL);
if (error)
goto done_cursors;
cursor2.key_beg = dedup->elm2;
cursor2.flags |= HAMMER_CURSOR_BACKEND;
error = hammer_btree_lookup(&cursor2);
if (error)
goto done_cursors;
error = hammer_btree_extract(&cursor2, HAMMER_CURSOR_GET_LEAF |
HAMMER_CURSOR_GET_DATA);
if (error)
goto done_cursors;
/*
* Zone validation. We can't de-dup any of the other zones
* (BTREE or META) or bad things will happen.
*
* Return with error = 0, but set an INVALID_ZONE flag.
*/
error = validate_zone(cursor1.leaf->data_offset) +
validate_zone(cursor2.leaf->data_offset);
if (error) {
dedup->head.flags |= HAMMER_IOC_DEDUP_INVALID_ZONE;
error = 0;
goto done_cursors;
}
/*
* Comparison checks
*
* If zones don't match or data_len fields aren't the same
* we consider it to be a comparison failure.
*
* Return with error = 0, but set a CMP_FAILURE flag.
*/
if ((cursor1.leaf->data_offset & HAMMER_OFF_ZONE_MASK) !=
(cursor2.leaf->data_offset & HAMMER_OFF_ZONE_MASK)) {
dedup->head.flags |= HAMMER_IOC_DEDUP_CMP_FAILURE;
goto done_cursors;
}
if (cursor1.leaf->data_len != cursor2.leaf->data_len) {
dedup->head.flags |= HAMMER_IOC_DEDUP_CMP_FAILURE;
goto done_cursors;
}
/* byte-by-byte comparison to be sure */
if (bcmp(cursor1.data, cursor2.data, cursor1.leaf->data_len)) {
dedup->head.flags |= HAMMER_IOC_DEDUP_CMP_FAILURE;
goto done_cursors;
}
/*
* Upgrade both cursors together to an exclusive lock
*
* Return an error -> candidate goes to pass2 list
*/
hammer_sync_lock_sh(trans);
error = hammer_cursor_upgrade2(&cursor1, &cursor2);
if (error) {
hammer_sync_unlock(trans);
goto done_cursors;
}
error = hammer_blockmap_dedup(cursor1.trans,
cursor1.leaf->data_offset, cursor1.leaf->data_len);
if (error) {
if (error == ERANGE) {
/*
* Return with error = 0, but set an UNDERFLOW flag
*/
dedup->head.flags |= HAMMER_IOC_DEDUP_UNDERFLOW;
error = 0;
goto downgrade_cursors;
} else {
/*
* Return an error -> block goes to pass2 list
*/
goto downgrade_cursors;
}
}
/*
* The cursor2's cache must be invalidated before calling
* hammer_blockmap_free(), otherwise it will not be able to
* invalidate the underlying data buffer.
*/
hammer_cursor_invalidate_cache(&cursor2);
hammer_blockmap_free(cursor2.trans,
cursor2.leaf->data_offset, cursor2.leaf->data_len);
hammer_modify_node(cursor2.trans, cursor2.node,
&cursor2.leaf->data_offset, sizeof(hammer_off_t));
cursor2.leaf->data_offset = cursor1.leaf->data_offset;
hammer_modify_node_done(cursor2.node);
downgrade_cursors:
hammer_cursor_downgrade2(&cursor1, &cursor2);
hammer_sync_unlock(trans);
done_cursors:
hammer_done_cursor(&cursor2);
done_cursor:
hammer_done_cursor(&cursor1);
/*
* Avoid deadlocking the buffer cache
*/
seq = trans->hmp->flusher.done;
while (hammer_flusher_meta_halflimit(trans->hmp) ||
hammer_flusher_undo_exhausted(trans, 2)) {
hammer_flusher_wait(trans->hmp, seq);
seq = hammer_flusher_async_one(trans->hmp);
}
return (error);
}
