/*
* 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_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);
}
/*
* 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);
}
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);
}
