static void determine_search_start(reiserfs_blocknr_hint_t *hint,
int amount_needed)
{
struct super_block *s = hint->th->t_super;
int unfm_hint;
hint->beg = 0;
hint->end = SB_BLOCK_COUNT(s) - 1;
/* This is former border algorithm. Now with tunable border offset */
if (concentrating_formatted_nodes(s))
set_border_in_hint(s, hint);
#ifdef DISPLACE_NEW_PACKING_LOCALITIES
/* whenever we create a new directory, we displace it. At first we will
hash for location, later we might look for a moderately empty place for
it */
if (displacing_new_packing_localities(s)
&& hint->th->displace_new_blocks) {
displace_new_packing_locality(hint);
/* we do not continue determine_search_start,
* if new packing locality is being displaced */
return;
}
#endif
/* all persons should feel encouraged to add more special cases here and
* test them */
if (displacing_large_files(s) && !hint->formatted_node
&& this_blocknr_allocation_would_make_it_a_large_file(hint)) {
displace_large_file(hint);
return;
}
/* if none of our special cases is relevant, use the left neighbor in the
tree order of the new node we are allocating for */
if (hint->formatted_node && TEST_OPTION(hashed_formatted_nodes,s)) {
hash_formatted_node(hint);
return;
}
unfm_hint = get_left_neighbor(hint);
/* Mimic old block allocator behaviour, that is if VFS allowed for preallocation,
new blocks are displaced based on directory ID. Also, if suggested search_start
is less than last preallocated block, we start searching from it, assuming that
HDD dataflow is faster in forward direction */
if ( TEST_OPTION(old_way, s)) {
if (!hint->formatted_node) {
if ( !reiserfs_hashed_relocation(s))
old_way(hint);
else if (!reiserfs_no_unhashed_relocation(s))
old_hashed_relocation(hint);
if ( hint->inode && hint->search_start < REISERFS_I(hint->inode)->i_prealloc_block)
hint->search_start = REISERFS_I(hint->inode)->i_prealloc_block;
}
return;
}
/* This is an approach proposed by Hans */
if ( TEST_OPTION(hundredth_slices, s) && ! (displacing_large_files(s) && !hint->formatted_node)) {
hundredth_slices(hint);
return;
}
/* old_hashed_relocation only works on unformatted */
if (!unfm_hint && !hint->formatted_node &&
TEST_OPTION(old_hashed_relocation, s))
{
old_hashed_relocation(hint);
}
/* new_hashed_relocation works with both formatted/unformatted nodes */
if ((!unfm_hint || hint->formatted_node) &&
TEST_OPTION(new_hashed_relocation, s))
{
new_hashed_relocation(hint);
}
/* dirid grouping works only on unformatted nodes */
if (!unfm_hint && !hint->formatted_node && TEST_OPTION(dirid_groups,s))
{
dirid_groups(hint);
}
#ifdef DISPLACE_NEW_PACKING_LOCALITIES
if (hint->formatted_node && TEST_OPTION(dirid_groups,s))
{
dirid_groups(hint);
}
#endif
/* oid grouping works only on unformatted nodes */
if (!unfm_hint && !hint->formatted_node && TEST_OPTION(oid_groups,s))
{
oid_groups(hint);
}
return;
}
static int show_super(struct seq_file *m, void *unused)
{
struct super_block *sb = m->private;
struct reiserfs_sb_info *r = REISERFS_SB(sb);
seq_printf(m, "state: \t%s\n"
"mount options: \t%s%s%s%s%s%s%s%s%s%s%s\n"
"gen. counter: \t%i\n"
"s_disk_reads: \t%i\n"
"s_disk_writes: \t%i\n"
"s_fix_nodes: \t%i\n"
"s_do_balance: \t%i\n"
"s_unneeded_left_neighbor: \t%i\n"
"s_good_search_by_key_reada: \t%i\n"
"s_bmaps: \t%i\n"
"s_bmaps_without_search: \t%i\n"
"s_direct2indirect: \t%i\n"
"s_indirect2direct: \t%i\n"
"\n"
"max_hash_collisions: \t%i\n"
"breads: \t%lu\n"
"bread_misses: \t%lu\n"
"search_by_key: \t%lu\n"
"search_by_key_fs_changed: \t%lu\n"
"search_by_key_restarted: \t%lu\n"
"insert_item_restarted: \t%lu\n"
"paste_into_item_restarted: \t%lu\n"
"cut_from_item_restarted: \t%lu\n"
"delete_solid_item_restarted: \t%lu\n"
"delete_item_restarted: \t%lu\n"
"leaked_oid: \t%lu\n"
"leaves_removable: \t%lu\n",
SF(s_mount_state) == REISERFS_VALID_FS ?
"REISERFS_VALID_FS" : "REISERFS_ERROR_FS",
reiserfs_r5_hash(sb) ? "FORCE_R5 " : "",
reiserfs_rupasov_hash(sb) ? "FORCE_RUPASOV " : "",
reiserfs_tea_hash(sb) ? "FORCE_TEA " : "",
reiserfs_hash_detect(sb) ? "DETECT_HASH " : "",
reiserfs_no_border(sb) ? "NO_BORDER " : "BORDER ",
reiserfs_no_unhashed_relocation(sb) ?
"NO_UNHASHED_RELOCATION " : "",
reiserfs_hashed_relocation(sb) ? "UNHASHED_RELOCATION " : "",
reiserfs_test4(sb) ? "TEST4 " : "",
have_large_tails(sb) ? "TAILS " : have_small_tails(sb) ?
"SMALL_TAILS " : "NO_TAILS ",
replay_only(sb) ? "REPLAY_ONLY " : "",
convert_reiserfs(sb) ? "CONV " : "",
atomic_read(&r->s_generation_counter),
SF(s_disk_reads), SF(s_disk_writes), SF(s_fix_nodes),
SF(s_do_balance), SF(s_unneeded_left_neighbor),
SF(s_good_search_by_key_reada), SF(s_bmaps),
SF(s_bmaps_without_search), SF(s_direct2indirect),
SF(s_indirect2direct), SFP(max_hash_collisions), SFP(breads),
SFP(bread_miss), SFP(search_by_key),
SFP(search_by_key_fs_changed), SFP(search_by_key_restarted),
SFP(insert_item_restarted), SFP(paste_into_item_restarted),
SFP(cut_from_item_restarted),
SFP(delete_solid_item_restarted), SFP(delete_item_restarted),
SFP(leaked_oid), SFP(leaves_removable));
return 0;
}
int reiserfs_super_in_proc( char *buffer, char **start, off_t offset,
int count, int *eof, void *data )
{
struct super_block *sb;
struct reiserfs_sb_info *r;
int len = 0;
sb = procinfo_prologue( ( kdev_t ) ( long ) data );
if( sb == NULL )
return -ENOENT;
r = &sb->u.reiserfs_sb;
len += sprintf( &buffer[ len ],
"state: \t%s\n"
"mount options: \t%s%s%s%s%s%s%s%s%s%s%s%s\n"
"gen. counter: \t%i\n"
"s_kmallocs: \t%i\n"
"s_disk_reads: \t%i\n"
"s_disk_writes: \t%i\n"
"s_fix_nodes: \t%i\n"
"s_do_balance: \t%i\n"
"s_unneeded_left_neighbor: \t%i\n"
"s_good_search_by_key_reada: \t%i\n"
"s_bmaps: \t%i\n"
"s_bmaps_without_search: \t%i\n"
"s_direct2indirect: \t%i\n"
"s_indirect2direct: \t%i\n"
"\n"
"max_hash_collisions: \t%i\n"
"breads: \t%lu\n"
"bread_misses: \t%lu\n"
"search_by_key: \t%lu\n"
"search_by_key_fs_changed: \t%lu\n"
"search_by_key_restarted: \t%lu\n"
"insert_item_restarted: \t%lu\n"
"paste_into_item_restarted: \t%lu\n"
"cut_from_item_restarted: \t%lu\n"
"delete_solid_item_restarted: \t%lu\n"
"delete_item_restarted: \t%lu\n"
"leaked_oid: \t%lu\n"
"leaves_removable: \t%lu\n",
SF( s_mount_state ) == REISERFS_VALID_FS ?
"REISERFS_VALID_FS" : "REISERFS_ERROR_FS",
reiserfs_r5_hash( sb ) ? "FORCE_R5 " : "",
reiserfs_rupasov_hash( sb ) ? "FORCE_RUPASOV " : "",
reiserfs_tea_hash( sb ) ? "FORCE_TEA " : "",
reiserfs_hash_detect( sb ) ? "DETECT_HASH " : "",
reiserfs_no_border( sb ) ? "NO_BORDER " : "BORDER ",
reiserfs_no_unhashed_relocation( sb ) ? "NO_UNHASHED_RELOCATION " : "",
reiserfs_hashed_relocation( sb ) ? "UNHASHED_RELOCATION " : "",
reiserfs_test4( sb ) ? "TEST4 " : "",
have_large_tails( sb ) ? "TAILS " : have_small_tails(sb)?"SMALL_TAILS ":"NO_TAILS ",
replay_only( sb ) ? "REPLAY_ONLY " : "",
reiserfs_dont_log( sb ) ? "DONT_LOG " : "LOG ",
convert_reiserfs( sb ) ? "CONV " : "",
atomic_read( &r -> s_generation_counter ),
SF( s_kmallocs ),
SF( s_disk_reads ),
SF( s_disk_writes ),
SF( s_fix_nodes ),
SF( s_do_balance ),
SF( s_unneeded_left_neighbor ),
SF( s_good_search_by_key_reada ),
SF( s_bmaps ),
SF( s_bmaps_without_search ),
SF( s_direct2indirect ),
SF( s_indirect2direct ),
SFP( max_hash_collisions ),
SFP( breads ),
SFP( bread_miss ),
SFP( search_by_key ),
SFP( search_by_key_fs_changed ),
SFP( search_by_key_restarted ),
SFP( insert_item_restarted ),
SFP( paste_into_item_restarted ),
SFP( cut_from_item_restarted ),
SFP( delete_solid_item_restarted ),
SFP( delete_item_restarted ),
SFP( leaked_oid ),
SFP( leaves_removable ) );
procinfo_epilogue( sb );
return reiserfs_proc_tail( len, buffer, start, offset, count, eof );
}
/*
** 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;
}