static inline void displace_large_file(reiserfs_blocknr_hint_t *hint)
{
if ( TEST_OPTION(displace_based_on_dirid, hint->th->t_super))
hint->search_start = hint->beg + keyed_hash((char *)(&INODE_PKEY(hint->inode)->k_dir_id), 4) % (hint->end - hint->beg);
else
hint->search_start = hint->beg + keyed_hash((char *)(&INODE_PKEY(hint->inode)->k_objectid), 4) % (hint->end - hint->beg);
}
static inline void displace_new_packing_locality (reiserfs_blocknr_hint_t *hint)
{
struct reiserfs_key * key = &hint->key;
hint->th->displace_new_blocks = 0;
hint->search_start = hint->beg + keyed_hash((char*)(&key->k_objectid),4) % (hint->end - hint->beg);
}
// if root directory is empty - we set default - Yura's - hash and
// warn about it
// FIXME: we look for only one name in a directory. If tea and yura
// bith have the same value - we ask user to send report to the
// mailing list
__u32 find_hash_out (struct super_block * s)
{
int retval;
struct inode * inode;
struct cpu_key key;
INITIALIZE_PATH (path);
struct reiserfs_dir_entry de;
__u32 hash = DEFAULT_HASH;
inode = s->s_root->d_inode;
do { // Some serious "goto"-hater was there ;)
u32 teahash, r5hash, yurahash;
make_cpu_key (&key, inode, ~0, TYPE_DIRENTRY, 3);
retval = search_by_entry_key (s, &key, &path, &de);
if (retval == IO_ERROR) {
pathrelse (&path);
return UNSET_HASH ;
}
if (retval == NAME_NOT_FOUND)
de.de_entry_num --;
set_de_name_and_namelen (&de);
if (deh_offset( &(de.de_deh[de.de_entry_num]) ) == DOT_DOT_OFFSET) {
/* allow override in this case */
if (reiserfs_rupasov_hash(s)) {
hash = YURA_HASH ;
}
reiserfs_warning("reiserfs: FS seems to be empty, autodetect "
"is using the default hash\n");
break;
}
r5hash=GET_HASH_VALUE (r5_hash (de.de_name, de.de_namelen));
teahash=GET_HASH_VALUE (keyed_hash (de.de_name, de.de_namelen));
yurahash=GET_HASH_VALUE (yura_hash (de.de_name, de.de_namelen));
if ( ( (teahash == r5hash) && (GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num]))) == r5hash) ) ||
( (teahash == yurahash) && (yurahash == GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num])))) ) ||
( (r5hash == yurahash) && (yurahash == GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num])))) ) ) {
reiserfs_warning("reiserfs: Unable to automatically detect hash"
"function for device %s\n"
"please mount with -o hash={tea,rupasov,r5}\n", kdevname (s->s_dev));
hash = UNSET_HASH;
break;
}
if (GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num])) ) == yurahash)
hash = YURA_HASH;
else if (GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num])) ) == teahash)
hash = TEA_HASH;
else if (GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num])) ) == r5hash)
hash = R5_HASH;
else {
reiserfs_warning("reiserfs: Unrecognised hash function for "
"device %s\n", kdevname (s->s_dev));
hash = UNSET_HASH;
}
} while (0);
pathrelse (&path);
return hash;
}
static inline void hundredth_slices (reiserfs_blocknr_hint_t * hint)
{
struct reiserfs_key * key = &hint->key;
b_blocknr_t slice_start;
slice_start = (keyed_hash((char*)(&key->k_dir_id),4) % 100) * (hint->end / 100);
if ( slice_start > hint->search_start || slice_start + (hint->end / 100) <= hint->search_start) {
hint->search_start = slice_start;
}
}
static inline void hash_formatted_node(reiserfs_blocknr_hint_t *hint)
{
char * hash_in;
if (!hint->inode)
hash_in = (char*)&hint->key.k_dir_id;
else if ( TEST_OPTION(displace_based_on_dirid, hint->th->t_super))
hash_in = (char *)(&INODE_PKEY(hint->inode)->k_dir_id);
else
hash_in = (char *)(&INODE_PKEY(hint->inode)->k_objectid);
hint->search_start = hint->beg + keyed_hash(hash_in, 4) % (hint->end - hint->beg);
}
/* Keyed 32-bit hash function using TEA in a Davis-Meyer function */
static unsigned long get_third_component (const char * name, int len)
{
unsigned long res;
if (!len || (len == 1 && name[0] == '.'))
return DOT_OFFSET;
if (len == 2 && name[0] == '.' && name[1] == '.')
return DOT_DOT_OFFSET;
res = keyed_hash (name, len);
res = GET_HASH_VALUE(res);
if (res == 0)
res = 128;
return res + MAX_GEN_NUMBER;
}
// if root directory is empty - we set default - Yura's - hash and
// warn about it
// FIXME: we look for only one name in a directory. If tea and yura
// bith have the same value - we ask user to send report to the
// mailing list
__u32 find_hash_out (struct super_block * s)
{
int retval;
struct inode * inode;
struct cpu_key key;
INITIALIZE_PATH (path);
struct reiserfs_dir_entry de;
__u32 hash = DEFAULT_HASH;
inode = s->s_root->d_inode;
while (1) {
make_cpu_key (&key, inode, ~0, TYPE_DIRENTRY, 3);
retval = search_by_entry_key (s, &key, &path, &de);
if (retval == IO_ERROR) {
pathrelse (&path);
return UNSET_HASH ;
}
if (retval == NAME_NOT_FOUND)
de.de_entry_num --;
set_de_name_and_namelen (&de);
if (deh_offset( &(de.de_deh[de.de_entry_num]) ) == DOT_DOT_OFFSET) {
/* allow override in this case */
if (reiserfs_rupasov_hash(s)) {
hash = YURA_HASH ;
}
reiserfs_warning("reiserfs: FS seems to be empty, autodetect "
"is using the default hash\n");
break;
}
if (GET_HASH_VALUE(yura_hash (de.de_name, de.de_namelen)) ==
GET_HASH_VALUE(keyed_hash (de.de_name, de.de_namelen))) {
reiserfs_warning ("reiserfs: Could not detect hash function "
"please mount with -o hash={tea,rupasov,r5}\n") ;
hash = UNSET_HASH ;
break;
}
if (GET_HASH_VALUE( deh_offset(&(de.de_deh[de.de_entry_num])) ) ==
GET_HASH_VALUE (yura_hash (de.de_name, de.de_namelen)))
hash = YURA_HASH;
else
hash = TEA_HASH;
break;
}
pathrelse (&path);
return hash;
}
static inline int old_hashed_relocation (reiserfs_blocknr_hint_t * hint)
{
b_blocknr_t border;
u32 hash_in;
if (hint->formatted_node || hint->inode == NULL) {
return 0;
}
hash_in = le32_to_cpu((INODE_PKEY(hint->inode))->k_dir_id);
border = hint->beg + (u32) keyed_hash(((char *) (&hash_in)), 4) % (hint->end - hint->beg - 1);
if (border > hint->search_start)
hint->search_start = border;
return 1;
}
static void inline new_hashed_relocation (reiserfs_blocknr_hint_t * hint)
{
char * hash_in;
if (hint->formatted_node) {
hash_in = (char*)&hint->key.k_dir_id;
} else {
if (!hint->inode) {
//hint->search_start = hint->beg;
hash_in = (char*)&hint->key.k_dir_id;
} else if ( TEST_OPTION(displace_based_on_dirid, hint->th->t_super))
hash_in = (char *)(&INODE_PKEY(hint->inode)->k_dir_id);
else
hash_in = (char *)(&INODE_PKEY(hint->inode)->k_objectid);
}
hint->search_start = hint->beg + keyed_hash(hash_in, 4) % (hint->end - hint->beg);
}
static int bmap_hash_id(struct super_block *s, u32 id) {
char * hash_in = NULL;
unsigned long hash;
unsigned bm;
if (id <= 2) {
bm = 1;
} else {
hash_in = (char *)(&id);
hash = keyed_hash(hash_in, 4);
bm = hash % SB_BMAP_NR(s);
if (!bm)
bm = 1;
}
/* this can only be true when SB_BMAP_NR = 1 */
if (bm >= SB_BMAP_NR(s))
bm = 0;
return bm;
}
/*
* If root directory is empty - we set default - Yura's - hash and warn
* about it.
* FIXME: we look for only one name in a directory. If tea and yura both
* have the same value - we ask user to send report to the mailing list
*/
uint32_t find_hash_out(struct reiserfs_mount *rmp)
{
int retval;
struct cpu_key key;
INITIALIZE_PATH(path);
struct reiserfs_node *ip;
struct reiserfs_sb_info *sbi;
struct reiserfs_dir_entry de;
uint32_t hash = DEFAULT_HASH;
get_root_node(rmp, &ip);
if (!ip)
return (UNSET_HASH);
sbi = rmp->rm_reiserfs;
do {
uint32_t teahash, r5hash, yurahash;
reiserfs_log(LOG_DEBUG, "make_cpu_key\n");
make_cpu_key(&key, ip, ~0, TYPE_DIRENTRY, 3);
reiserfs_log(LOG_DEBUG, "search_by_entry_key for "
"key(objectid=%d,dirid=%d)\n",
key.on_disk_key.k_objectid, key.on_disk_key.k_dir_id);
retval = search_by_entry_key(sbi, &key, &path, &de);
if (retval == IO_ERROR) {
pathrelse(&path);
return (UNSET_HASH);
}
if (retval == NAME_NOT_FOUND)
de.de_entry_num--;
reiserfs_log(LOG_DEBUG, "name found\n");
set_de_name_and_namelen(&de);
if (deh_offset(&(de.de_deh[de.de_entry_num])) == DOT_DOT_OFFSET) {
/* Allow override in this case */
if (reiserfs_rupasov_hash(sbi)) {
hash = YURA_HASH;
}
reiserfs_log(LOG_DEBUG,
"FS seems to be empty, autodetect "
"is using the default hash");
break;
}
r5hash = GET_HASH_VALUE(r5_hash(de.de_name, de.de_namelen));
teahash = GET_HASH_VALUE(keyed_hash(de.de_name,
de.de_namelen));
yurahash = GET_HASH_VALUE(yura_hash(de.de_name, de.de_namelen));
if (((teahash == r5hash) &&
(GET_HASH_VALUE(
deh_offset(&(de.de_deh[de.de_entry_num]))) == r5hash)) ||
((teahash == yurahash) &&
(yurahash ==
GET_HASH_VALUE(
deh_offset(&(de.de_deh[de.de_entry_num]))))) ||
((r5hash == yurahash) &&
(yurahash ==
GET_HASH_VALUE(
deh_offset(&(de.de_deh[de.de_entry_num])))))) {
reiserfs_log(LOG_ERR,
"unable to automatically detect hash "
"function. Please mount with -o "
"hash={tea,rupasov,r5}");
hash = UNSET_HASH;
break;
}
if (GET_HASH_VALUE(
deh_offset(&(de.de_deh[de.de_entry_num]))) == yurahash) {
reiserfs_log(LOG_DEBUG, "detected YURA hash\n");
hash = YURA_HASH;
} else if (GET_HASH_VALUE(
deh_offset(&(de.de_deh[de.de_entry_num]))) == teahash) {
reiserfs_log(LOG_DEBUG, "detected TEA hash\n");
hash = TEA_HASH;
} else if (GET_HASH_VALUE(
deh_offset(&(de.de_deh[de.de_entry_num]))) == r5hash) {
reiserfs_log(LOG_DEBUG, "detected R5 hash\n");
hash = R5_HASH;
} else {
reiserfs_log(LOG_WARNING, "unrecognised hash function");
hash = UNSET_HASH;
}
} while (0);
pathrelse(&path);
return (hash);
}
id = sa->id_i;
} else {
nonce = sa->n_i;
sk = sa->sk_p_r;
id = sa->id_r;
}
IF_TRACE({
TRACE((PLOGLOC, "SK\n"));
plogdump(PLOG_DEBUG, PLOGLOC, 0, sk->v, sk->l);
TRACE((PLOGLOC, "ID\n"));
plogdump(PLOG_DEBUG, PLOGLOC, 0, id->v, id->l);
});
/* prf(SK, ID) */
prf_output = keyed_hash(prf, sk, id);
if (!prf_output)
goto end;
IF_TRACE({
TRACE((PLOGLOC, "prf(SK, ID)\n"));
plogdump(PLOG_DEBUG, PLOGLOC, 0, prf_output->v, prf_output->l);
});
/* octets = message | N | prf(SK, ID) */
octets = rc_vmalloc(message->l + nonce->l + prf_output->l);
if (!octets)
goto end;
p = (uint8_t *)octets->v;
VCONCAT(octets, p, message);
/*
** We pre-allocate 8 blocks. Pre-allocation is used for files > 16 KB only.
** This lowers fragmentation on large files by grabbing a contiguous set of
** blocks at once. It also limits the number of times the bitmap block is
** logged by making X number of allocation changes in a single transaction.
**
** We are using a border to divide the disk into two parts. The first part
** is used for tree blocks, which have a very high turnover rate (they
** are constantly allocated then freed)
**
** The second part of the disk is for the unformatted nodes of larger files.
** Putting them away from the tree blocks lowers fragmentation, and makes
** it easier to group files together. There are a number of different
** allocation schemes being tried right now, each is documented below.
**
** A great deal of the allocator's speed comes because reiserfs_get_block
** sends us the block number of the last unformatted node in the file. Once
** a given block is allocated past the border, we don't collide with the
** blocks near the search_start again.
**
*/
int reiserfs_new_unf_blocknrs2 (struct reiserfs_transaction_handle *th,
struct inode * p_s_inode,
unsigned long * free_blocknrs,
unsigned long search_start)
{
int ret=0, blks_gotten=0;
unsigned long border = 0;
unsigned long bstart = 0;
unsigned long hash_in, hash_out;
unsigned long saved_search_start=search_start;
int allocated[PREALLOCATION_SIZE];
int blks;
if (!reiserfs_no_border(th->t_super)) {
/* we default to having the border at the 10% mark of the disk. This
** is an arbitrary decision and it needs tuning. It also needs a limit
** to prevent it from taking too much space on huge drives.
*/
bstart = (SB_BLOCK_COUNT(th->t_super) / 10);
}
if (!reiserfs_no_unhashed_relocation(th->t_super)) {
/* this is a very simple first attempt at preventing too much grouping
** around the border value. Since k_dir_id is never larger than the
** highest allocated oid, it is far from perfect, and files will tend
** to be grouped towards the start of the border
*/
border = le32_to_cpu(INODE_PKEY(p_s_inode)->k_dir_id) % (SB_BLOCK_COUNT(th->t_super) - bstart - 1) ;
} else {
/* why would we want to delcare a local variable to this if statement
** name border????? -chris
** unsigned long border = 0;
*/
if (!reiserfs_hashed_relocation(th->t_super)) {
hash_in = le32_to_cpu((INODE_PKEY(p_s_inode))->k_dir_id);
/* I wonder if the CPU cost of the
hash will obscure the layout
effect? Of course, whether that
effect is good or bad we don't
know.... :-) */
hash_out = keyed_hash(((char *) (&hash_in)), 4);
border = hash_out % (SB_BLOCK_COUNT(th->t_super) - bstart - 1) ;
}
}
border += bstart ;
allocated[0] = 0 ; /* important. Allows a check later on to see if at
* least one block was allocated. This prevents false
* no disk space returns
*/
if ( (p_s_inode->i_size < 4 * 4096) ||
!(S_ISREG(p_s_inode->i_mode)) )
{
if ( search_start < border
|| (
/* allow us to test whether it is a
good idea to prevent files from
getting too far away from their
packing locality by some unexpected
means. This might be poor code for
directories whose files total
larger than 1/10th of the disk, and
it might be good code for
suffering from old insertions when the disk
was almost full. */
/* changed from !reiserfs_test3(th->t_super), which doesn't
** seem like a good idea. Think about adding blocks to
** a large file. If you've allocated 10% of the disk
** in contiguous blocks, you start over at the border value
** for every new allocation. This throws away all the
** information sent in about the last block that was allocated
** in the file. Not a good general case at all.
** -chris
*/
reiserfs_test4(th->t_super) &&
(search_start > border + (SB_BLOCK_COUNT(th->t_super) / 10))
)
)
search_start=border;
ret = do_reiserfs_new_blocknrs(th, free_blocknrs, search_start,
1/*amount_needed*/,
0/*use reserved blocks for root */,
1/*for_formatted*/,
0/*for prealloc */) ;
return ret;
}
/* take a block off the prealloc list and return it -Hans */
if (p_s_inode->u.reiserfs_i.i_prealloc_count > 0) {
p_s_inode->u.reiserfs_i.i_prealloc_count--;
*free_blocknrs = p_s_inode->u.reiserfs_i.i_prealloc_block++;
/* if no more preallocated blocks, remove inode from list */
if (! p_s_inode->u.reiserfs_i.i_prealloc_count) {
list_del(&p_s_inode->u.reiserfs_i.i_prealloc_list);
}
return ret;
}
/* else get a new preallocation for the file */
reiserfs_discard_prealloc (th, p_s_inode);
/* this uses the last preallocated block as the search_start. discard
** prealloc does not zero out this number.
*/
if (search_start <= p_s_inode->u.reiserfs_i.i_prealloc_block) {
search_start = p_s_inode->u.reiserfs_i.i_prealloc_block;
}
/* doing the compare again forces search_start to be >= the border,
** even if the file already had prealloction done. This seems extra,
** and should probably be removed
*/
if ( search_start < border ) search_start=border;
/* If the disk free space is already below 10% we should
** start looking for the free blocks from the beginning
** of the partition, before the border line.
*/
if ( SB_FREE_BLOCKS(th->t_super) <= (SB_BLOCK_COUNT(th->t_super) / 10) ) {
search_start=saved_search_start;
}
*free_blocknrs = 0;
blks = PREALLOCATION_SIZE-1;
for (blks_gotten=0; blks_gotten<PREALLOCATION_SIZE; blks_gotten++) {
ret = do_reiserfs_new_blocknrs(th, free_blocknrs, search_start,
1/*amount_needed*/,
0/*for root reserved*/,
1/*for_formatted*/,
(blks_gotten > 0)/*must_be_contiguous*/) ;
/* if we didn't find a block this time, adjust blks to reflect
** the actual number of blocks allocated
*/
if (ret != CARRY_ON) {
blks = blks_gotten > 0 ? (blks_gotten - 1) : 0 ;
break ;
}
allocated[blks_gotten]= *free_blocknrs;
#ifdef CONFIG_REISERFS_CHECK
if ( (blks_gotten>0) && (allocated[blks_gotten] - allocated[blks_gotten-1]) != 1 ) {
/* this should be caught by new_blocknrs now, checking code */
reiserfs_warning("yura-1, reiserfs_new_unf_blocknrs2: pre-allocated not contiguous set of blocks!\n") ;
reiserfs_free_block(th, allocated[blks_gotten]);
blks = blks_gotten-1;
break;
}
#endif
if (blks_gotten==0) {
p_s_inode->u.reiserfs_i.i_prealloc_block = *free_blocknrs;
}
search_start = *free_blocknrs;
*free_blocknrs = 0;
}
p_s_inode->u.reiserfs_i.i_prealloc_count = blks;
*free_blocknrs = p_s_inode->u.reiserfs_i.i_prealloc_block;
p_s_inode->u.reiserfs_i.i_prealloc_block++;
/* if inode has preallocated blocks, link him to list */
if (p_s_inode->u.reiserfs_i.i_prealloc_count) {
list_add(&p_s_inode->u.reiserfs_i.i_prealloc_list,
&SB_JOURNAL(th->t_super)->j_prealloc_list);
}
/* we did actually manage to get 1 block */
if (ret != CARRY_ON && allocated[0] > 0) {
return CARRY_ON ;
}
/* NO_MORE_UNUSED_CONTIGUOUS_BLOCKS should only mean something to
** the preallocation code. The rest of the filesystem asks for a block
** and should either get it, or know the disk is full. The code
** above should never allow ret == NO_MORE_UNUSED_CONTIGUOUS_BLOCK,
** as it doesn't send for_prealloc = 1 to do_reiserfs_new_blocknrs
** unless it has already successfully allocated at least one block.
** Just in case, we translate into a return value the rest of the
** filesystem can understand.
**
** It is an error to change this without making the
** rest of the filesystem understand NO_MORE_UNUSED_CONTIGUOUS_BLOCKS
** If you consider it a bug to return NO_DISK_SPACE here, fix the rest
** of the fs first.
*/
if (ret == NO_MORE_UNUSED_CONTIGUOUS_BLOCKS) {
#ifdef CONFIG_REISERFS_CHECK
reiserfs_warning("reiser-2015: this shouldn't happen, may cause false out of disk space error");
#endif
return NO_DISK_SPACE;
}
return ret;
}
